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 560902 : eqsel(PG_FUNCTION_ARGS)
301 : {
302 560902 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
303 : }
304 :
305 : /*
306 : * Common code for eqsel() and neqsel()
307 : */
308 : static double
309 593403 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
310 : {
311 593403 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
312 593403 : Oid operator = PG_GETARG_OID(1);
313 593403 : List *args = (List *) PG_GETARG_POINTER(2);
314 593403 : int varRelid = PG_GETARG_INT32(3);
315 593403 : 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 593403 : if (negate)
326 : {
327 32501 : operator = get_negator(operator);
328 32501 : 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 593403 : if (!get_restriction_variable(root, args, varRelid,
340 : &vardata, &other, &varonleft))
341 3618 : 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 589781 : if (IsA(other, Const))
349 225280 : selec = var_eq_const(&vardata, operator, collation,
350 225280 : ((Const *) other)->constvalue,
351 225280 : ((Const *) other)->constisnull,
352 : varonleft, negate);
353 : else
354 364501 : selec = var_eq_non_const(&vardata, operator, collation, other,
355 : varonleft, negate);
356 :
357 589781 : ReleaseVariableStats(vardata);
358 :
359 589781 : 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 255457 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
369 : Datum constval, bool constisnull,
370 : bool varonleft, bool negate)
371 : {
372 : double selec;
373 255457 : 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 255457 : if (constisnull)
382 273 : 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 255184 : if (HeapTupleIsValid(vardata->statsTuple))
389 : {
390 : Form_pg_statistic stats;
391 :
392 181403 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
393 181403 : 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 255184 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
404 : {
405 53652 : selec = 1.0 / vardata->rel->tuples;
406 : }
407 340812 : else if (HeapTupleIsValid(vardata->statsTuple) &&
408 139280 : statistic_proc_security_check(vardata,
409 139280 : (opfuncoid = get_opcode(oproid))))
410 139280 : {
411 : AttStatsSlot sslot;
412 139280 : 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 139280 : if (get_attstatsslot(&sslot, vardata->statsTuple,
423 : STATISTIC_KIND_MCV, InvalidOid,
424 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
425 : {
426 126256 : LOCAL_FCINFO(fcinfo, 2);
427 : FmgrInfo eqproc;
428 :
429 126256 : 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 126256 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
438 : NULL, NULL);
439 126256 : fcinfo->args[0].isnull = false;
440 126256 : fcinfo->args[1].isnull = false;
441 : /* be careful to apply operator right way 'round */
442 126256 : if (varonleft)
443 126230 : fcinfo->args[1].value = constval;
444 : else
445 26 : fcinfo->args[0].value = constval;
446 :
447 2110187 : for (i = 0; i < sslot.nvalues; i++)
448 : {
449 : Datum fresult;
450 :
451 2048470 : if (varonleft)
452 2048424 : fcinfo->args[0].value = sslot.values[i];
453 : else
454 46 : fcinfo->args[1].value = sslot.values[i];
455 2048470 : fcinfo->isnull = false;
456 2048470 : fresult = FunctionCallInvoke(fcinfo);
457 2048470 : if (!fcinfo->isnull && DatumGetBool(fresult))
458 : {
459 64539 : match = true;
460 64539 : break;
461 : }
462 : }
463 : }
464 : else
465 : {
466 : /* no most-common-value info available */
467 13024 : i = 0; /* keep compiler quiet */
468 : }
469 :
470 139280 : 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 64539 : 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 74741 : double sumcommon = 0.0;
486 : double otherdistinct;
487 :
488 1792304 : for (i = 0; i < sslot.nnumbers; i++)
489 1717563 : sumcommon += sslot.numbers[i];
490 74741 : selec = 1.0 - sumcommon - nullfrac;
491 74741 : 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 74741 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
499 74741 : sslot.nnumbers;
500 74741 : if (otherdistinct > 1)
501 34714 : 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 74741 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
508 0 : selec = sslot.numbers[sslot.nnumbers - 1];
509 : }
510 :
511 139280 : 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 62252 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
521 : }
522 :
523 : /* now adjust if we wanted <> rather than = */
524 255184 : if (negate)
525 25774 : selec = 1.0 - selec - nullfrac;
526 :
527 : /* result should be in range, but make sure... */
528 255184 : CLAMP_PROBABILITY(selec);
529 :
530 255184 : 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 364501 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
540 : Node *other,
541 : bool varonleft, bool negate)
542 : {
543 : double selec;
544 364501 : double nullfrac = 0.0;
545 : bool isdefault;
546 :
547 : /*
548 : * Grab the nullfrac for use below.
549 : */
550 364501 : if (HeapTupleIsValid(vardata->statsTuple))
551 : {
552 : Form_pg_statistic stats;
553 :
554 222030 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
555 222030 : 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 364501 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
566 : {
567 128087 : selec = 1.0 / vardata->rel->tuples;
568 : }
569 236414 : 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 113635 : selec = 1.0 - nullfrac;
585 113635 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
586 113635 : if (ndistinct > 1)
587 111106 : selec /= ndistinct;
588 :
589 : /*
590 : * Cross-check: selectivity should never be estimated as more than the
591 : * most common value's.
592 : */
593 113635 : if (get_attstatsslot(&sslot, vardata->statsTuple,
594 : STATISTIC_KIND_MCV, InvalidOid,
595 : ATTSTATSSLOT_NUMBERS))
596 : {
597 97848 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
598 483 : selec = sslot.numbers[0];
599 97848 : 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 122779 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
610 : }
611 :
612 : /* now adjust if we wanted <> rather than = */
613 364501 : if (negate)
614 5138 : selec = 1.0 - selec - nullfrac;
615 :
616 : /* result should be in range, but make sure... */
617 364501 : CLAMP_PROBABILITY(selec);
618 :
619 364501 : 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 32501 : neqsel(PG_FUNCTION_ARGS)
631 : {
632 32501 : 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 247734 : 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 247734 : 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 21382 : if (vardata->var && IsA(vardata->var, Var) &&
672 17355 : ((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 1684 : if (vardata->rel->pages == 0)
683 0 : return 1.0;
684 :
685 1684 : itemptr = (ItemPointer) DatumGetPointer(constval);
686 1684 : 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 1684 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
696 :
697 : /* If target is the last page, use half the density. */
698 1684 : if (block >= vardata->rel->pages - 1)
699 25 : 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 1684 : if (density > 0.0)
710 : {
711 1684 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
712 :
713 1684 : 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 1684 : 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 1684 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
731 1567 : selec -= (1.0 / vardata->rel->tuples);
732 :
733 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
734 1684 : if (isgt)
735 1552 : selec = 1.0 - selec;
736 :
737 1684 : CLAMP_PROBABILITY(selec);
738 1684 : return selec;
739 : }
740 :
741 : /* no stats available, so default result */
742 19698 : return DEFAULT_INEQ_SEL;
743 : }
744 226352 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
745 :
746 226352 : 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 226352 : 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 226352 : 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 226352 : selec = 1.0 - stats->stanullfrac - sumcommon;
772 :
773 226352 : if (hist_selec >= 0.0)
774 138480 : 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 87872 : selec *= 0.5;
782 : }
783 :
784 226352 : selec += mcv_selec;
785 :
786 : /* result should be in range, but make sure... */
787 226352 : CLAMP_PROBABILITY(selec);
788 :
789 226352 : 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 230713 : 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 230713 : mcv_selec = 0.0;
815 230713 : sumcommon = 0.0;
816 :
817 459638 : if (HeapTupleIsValid(vardata->statsTuple) &&
818 457575 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
819 228650 : get_attstatsslot(&sslot, vardata->statsTuple,
820 : STATISTIC_KIND_MCV, InvalidOid,
821 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
822 : {
823 128435 : 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 128435 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
833 : NULL, NULL);
834 128435 : fcinfo->args[0].isnull = false;
835 128435 : fcinfo->args[1].isnull = false;
836 : /* be careful to apply operator right way 'round */
837 128435 : if (varonleft)
838 128435 : fcinfo->args[1].value = constval;
839 : else
840 0 : fcinfo->args[0].value = constval;
841 :
842 3052174 : for (i = 0; i < sslot.nvalues; i++)
843 : {
844 : Datum fresult;
845 :
846 2923739 : if (varonleft)
847 2923739 : fcinfo->args[0].value = sslot.values[i];
848 : else
849 0 : fcinfo->args[1].value = sslot.values[i];
850 2923739 : fcinfo->isnull = false;
851 2923739 : fresult = FunctionCallInvoke(fcinfo);
852 2923739 : if (!fcinfo->isnull && DatumGetBool(fresult))
853 1199898 : mcv_selec += sslot.numbers[i];
854 2923739 : sumcommon += sslot.numbers[i];
855 : }
856 128435 : free_attstatsslot(&sslot);
857 : }
858 :
859 230713 : *sumcommonp = sumcommon;
860 230713 : 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 4361 : 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 6934 : if (HeapTupleIsValid(vardata->statsTuple) &&
910 5141 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
911 2568 : get_attstatsslot(&sslot, vardata->statsTuple,
912 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
913 : ATTSTATSSLOT_VALUES))
914 : {
915 2490 : *hist_size = sslot.nvalues;
916 2490 : if (sslot.nvalues >= min_hist_size)
917 : {
918 1354 : LOCAL_FCINFO(fcinfo, 2);
919 1354 : 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 1354 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
931 : NULL, NULL);
932 1354 : fcinfo->args[0].isnull = false;
933 1354 : fcinfo->args[1].isnull = false;
934 : /* be careful to apply operator right way 'round */
935 1354 : if (varonleft)
936 1354 : fcinfo->args[1].value = constval;
937 : else
938 0 : fcinfo->args[0].value = constval;
939 :
940 113402 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
941 : {
942 : Datum fresult;
943 :
944 112048 : if (varonleft)
945 112048 : fcinfo->args[0].value = sslot.values[i];
946 : else
947 0 : fcinfo->args[1].value = sslot.values[i];
948 112048 : fcinfo->isnull = false;
949 112048 : fresult = FunctionCallInvoke(fcinfo);
950 112048 : if (!fcinfo->isnull && DatumGetBool(fresult))
951 4848 : nmatch++;
952 : }
953 1354 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
954 : }
955 : else
956 1136 : result = -1;
957 2490 : free_attstatsslot(&sslot);
958 : }
959 : else
960 : {
961 1871 : *hist_size = 0;
962 1871 : result = -1;
963 : }
964 :
965 4361 : 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
987 845 : 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 845 : 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 845 : if (IsA(other, Const) &&
1009 845 : ((Const *) other)->constisnull)
1010 : {
1011 0 : ReleaseVariableStats(vardata);
1012 0 : return 0.0;
1013 : }
1014 :
1015 845 : if (IsA(other, Const))
1016 : {
1017 : /* Variable is being compared to a known non-null constant */
1018 845 : Datum constval = ((Const *) other)->constvalue;
1019 : FmgrInfo opproc;
1020 : double mcvsum;
1021 : double mcvsel;
1022 : double nullfrac;
1023 : int hist_size;
1024 :
1025 845 : fmgr_info(get_opcode(oproid), &opproc);
1026 :
1027 : /*
1028 : * Calculate the selectivity for the column's most common values.
1029 : */
1030 845 : 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 845 : selec = histogram_selectivity(&vardata, &opproc, collation,
1041 : constval, varonleft,
1042 : 10, 1, &hist_size);
1043 845 : if (selec < 0)
1044 : {
1045 : /* Nope, fall back on default */
1046 845 : 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 845 : if (selec < 0.0001)
1063 0 : selec = 0.0001;
1064 845 : else if (selec > 0.9999)
1065 0 : selec = 0.9999;
1066 :
1067 : /* Don't forget to account for nulls. */
1068 845 : if (HeapTupleIsValid(vardata.statsTuple))
1069 70 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
1070 : else
1071 775 : 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 845 : selec *= 1.0 - nullfrac - mcvsum;
1079 845 : 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 845 : ReleaseVariableStats(vardata);
1088 :
1089 : /* result should be in range, but make sure... */
1090 845 : CLAMP_PROBABILITY(selec);
1091 :
1092 845 : 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 229609 : 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 229609 : 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 458717 : if (HeapTupleIsValid(vardata->statsTuple) &&
1138 457946 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1139 228838 : get_attstatsslot(&sslot, vardata->statsTuple,
1140 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1141 : ATTSTATSSLOT_VALUES))
1142 : {
1143 141234 : if (sslot.nvalues > 1 &&
1144 282406 : sslot.stacoll == collation &&
1145 141172 : comparison_ops_are_compatible(sslot.staop, opoid))
1146 141082 : {
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 141082 : int lobound = 0; /* first possible slot to search */
1167 141082 : int hibound = sslot.nvalues; /* last+1 slot to search */
1168 141082 : 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 141082 : if (sslot.nvalues == 2)
1177 2689 : have_end = get_actual_variable_range(root,
1178 : vardata,
1179 : sslot.staop,
1180 : collation,
1181 : &sslot.values[0],
1182 2689 : &sslot.values[1]);
1183 :
1184 925576 : while (lobound < hibound)
1185 : {
1186 784494 : 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 784494 : if (probe == 0 && sslot.nvalues > 2)
1195 68609 : have_end = get_actual_variable_range(root,
1196 : vardata,
1197 : sslot.staop,
1198 : collation,
1199 : &sslot.values[0],
1200 : NULL);
1201 715885 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1202 48450 : have_end = get_actual_variable_range(root,
1203 : vardata,
1204 : sslot.staop,
1205 : collation,
1206 : NULL,
1207 48450 : &sslot.values[probe]);
1208 :
1209 784494 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1210 : collation,
1211 784494 : sslot.values[probe],
1212 : constval));
1213 784494 : if (isgt)
1214 50279 : ltcmp = !ltcmp;
1215 784494 : if (ltcmp)
1216 294938 : lobound = probe + 1;
1217 : else
1218 489556 : hibound = probe;
1219 : }
1220 :
1221 141082 : 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 60267 : histfrac = 0.0;
1232 : }
1233 80815 : else if (lobound >= sslot.nvalues)
1234 : {
1235 : /*
1236 : * Inverse case: constant is above upper histogram boundary.
1237 : */
1238 23494 : histfrac = 1.0;
1239 : }
1240 : else
1241 : {
1242 : /* We have values[i-1] <= constant <= values[i]. */
1243 57321 : int i = lobound;
1244 57321 : 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 57321 : if (i == 1 || isgt == iseq)
1267 : {
1268 : double otherdistinct;
1269 : bool isdefault;
1270 : AttStatsSlot mcvslot;
1271 :
1272 : /* Get estimated number of distinct values */
1273 23896 : otherdistinct = get_variable_numdistinct(vardata,
1274 : &isdefault);
1275 :
1276 : /* Subtract off the number of known MCVs */
1277 23896 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1278 : STATISTIC_KIND_MCV, InvalidOid,
1279 : ATTSTATSSLOT_NUMBERS))
1280 : {
1281 3032 : otherdistinct -= mcvslot.nnumbers;
1282 3032 : free_attstatsslot(&mcvslot);
1283 : }
1284 :
1285 : /* If result doesn't seem sane, leave eq_selec at 0 */
1286 23896 : if (otherdistinct > 1)
1287 23874 : 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 57321 : if (convert_to_scalar(constval, consttype, collation,
1296 : &val,
1297 57321 : sslot.values[i - 1], sslot.values[i],
1298 : vardata->vartype,
1299 : &low, &high))
1300 : {
1301 57321 : if (high <= low)
1302 : {
1303 : /* cope if bin boundaries appear identical */
1304 0 : binfrac = 0.5;
1305 : }
1306 57321 : else if (val <= low)
1307 11567 : binfrac = 0.0;
1308 45754 : else if (val >= high)
1309 2881 : binfrac = 1.0;
1310 : else
1311 : {
1312 42873 : 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 42873 : if (isnan(binfrac) ||
1321 42873 : 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 57321 : histfrac = (double) (i - 1) + binfrac;
1344 57321 : 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 57321 : if (i == 1)
1379 10151 : 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 57321 : if (isgt == iseq)
1387 18766 : 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 141082 : 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 141082 : if (have_end)
1406 79828 : CLAMP_PROBABILITY(hist_selec);
1407 : else
1408 : {
1409 61254 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1410 :
1411 61254 : if (hist_selec < cutoff)
1412 20662 : hist_selec = cutoff;
1413 40592 : else if (hist_selec > 1.0 - cutoff)
1414 15223 : hist_selec = 1.0 - cutoff;
1415 : }
1416 : }
1417 152 : 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 152 : LOCAL_FCINFO(fcinfo, 2);
1427 152 : int nmatch = 0;
1428 :
1429 152 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1430 : NULL, NULL);
1431 152 : fcinfo->args[0].isnull = false;
1432 152 : fcinfo->args[1].isnull = false;
1433 152 : fcinfo->args[1].value = constval;
1434 801750 : for (int i = 0; i < sslot.nvalues; i++)
1435 : {
1436 : Datum fresult;
1437 :
1438 801598 : fcinfo->args[0].value = sslot.values[i];
1439 801598 : fcinfo->isnull = false;
1440 801598 : fresult = FunctionCallInvoke(fcinfo);
1441 801598 : if (!fcinfo->isnull && DatumGetBool(fresult))
1442 1608 : nmatch++;
1443 : }
1444 152 : 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 152 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1454 :
1455 152 : if (hist_selec < cutoff)
1456 10 : hist_selec = cutoff;
1457 142 : else if (hist_selec > 1.0 - cutoff)
1458 10 : hist_selec = 1.0 - cutoff;
1459 : }
1460 : }
1461 :
1462 141234 : free_attstatsslot(&sslot);
1463 : }
1464 :
1465 229609 : return hist_selec;
1466 : }
1467 :
1468 : /*
1469 : * Common wrapper function for the selectivity estimators that simply
1470 : * invoke scalarineqsel().
1471 : */
1472 : static Datum
1473 38866 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1474 : {
1475 38866 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1476 38866 : Oid operator = PG_GETARG_OID(1);
1477 38866 : List *args = (List *) PG_GETARG_POINTER(2);
1478 38866 : int varRelid = PG_GETARG_INT32(3);
1479 38866 : 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 38866 : if (!get_restriction_variable(root, args, varRelid,
1492 : &vardata, &other, &varonleft))
1493 462 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1494 :
1495 : /*
1496 : * Can't do anything useful if the something is not a constant, either.
1497 : */
1498 38404 : if (!IsA(other, Const))
1499 : {
1500 2355 : ReleaseVariableStats(vardata);
1501 2355 : 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 36049 : if (((Const *) other)->constisnull)
1509 : {
1510 55 : ReleaseVariableStats(vardata);
1511 55 : PG_RETURN_FLOAT8(0.0);
1512 : }
1513 35994 : constval = ((Const *) other)->constvalue;
1514 35994 : consttype = ((Const *) other)->consttype;
1515 :
1516 : /*
1517 : * Force the var to be on the left to simplify logic in scalarineqsel.
1518 : */
1519 35994 : if (!varonleft)
1520 : {
1521 318 : operator = get_commutator(operator);
1522 318 : 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 318 : isgt = !isgt;
1529 : }
1530 :
1531 : /* The rest of the work is done by scalarineqsel(). */
1532 35994 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1533 : &vardata, constval, consttype);
1534 :
1535 35994 : ReleaseVariableStats(vardata);
1536 :
1537 35994 : PG_RETURN_FLOAT8((float8) selec);
1538 : }
1539 :
1540 : /*
1541 : * scalarltsel - Selectivity of "<" for scalars.
1542 : */
1543 : Datum
1544 12149 : scalarltsel(PG_FUNCTION_ARGS)
1545 : {
1546 12149 : return scalarineqsel_wrapper(fcinfo, false, false);
1547 : }
1548 :
1549 : /*
1550 : * scalarlesel - Selectivity of "<=" for scalars.
1551 : */
1552 : Datum
1553 3812 : scalarlesel(PG_FUNCTION_ARGS)
1554 : {
1555 3812 : return scalarineqsel_wrapper(fcinfo, false, true);
1556 : }
1557 :
1558 : /*
1559 : * scalargtsel - Selectivity of ">" for scalars.
1560 : */
1561 : Datum
1562 12362 : scalargtsel(PG_FUNCTION_ARGS)
1563 : {
1564 12362 : return scalarineqsel_wrapper(fcinfo, true, false);
1565 : }
1566 :
1567 : /*
1568 : * scalargesel - Selectivity of ">=" for scalars.
1569 : */
1570 : Datum
1571 10543 : scalargesel(PG_FUNCTION_ARGS)
1572 : {
1573 10543 : 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 47296 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1586 : {
1587 : VariableStatData vardata;
1588 : double selec;
1589 :
1590 47296 : examine_variable(root, arg, varRelid, &vardata);
1591 47296 : 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 23210 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1598 : BoolGetDatum(true), false, true, false);
1599 : }
1600 24086 : 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 11329 : selec = 0.3333333;
1610 : }
1611 : else
1612 : {
1613 : /* Otherwise, the default estimate is 0.5 */
1614 12757 : selec = 0.5;
1615 : }
1616 47296 : ReleaseVariableStats(vardata);
1617 47296 : return selec;
1618 : }
1619 :
1620 : /*
1621 : * booltestsel - Selectivity of BooleanTest Node.
1622 : */
1623 : Selectivity
1624 781 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1625 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1626 : {
1627 : VariableStatData vardata;
1628 : double selec;
1629 :
1630 781 : examine_variable(root, arg, varRelid, &vardata);
1631 :
1632 781 : if (HeapTupleIsValid(vardata.statsTuple))
1633 : {
1634 : Form_pg_statistic stats;
1635 : double freq_null;
1636 : AttStatsSlot sslot;
1637 :
1638 20 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1639 20 : freq_null = stats->stanullfrac;
1640 :
1641 20 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1642 : STATISTIC_KIND_MCV, InvalidOid,
1643 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1644 10 : && sslot.nnumbers > 0)
1645 10 : {
1646 : double freq_true;
1647 : double freq_false;
1648 :
1649 : /*
1650 : * Get first MCV frequency and derive frequency for true.
1651 : */
1652 10 : if (DatumGetBool(sslot.values[0]))
1653 0 : freq_true = sslot.numbers[0];
1654 : else
1655 10 : 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 10 : freq_false = 1.0 - freq_true - freq_null;
1662 :
1663 10 : 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 10 : case IS_TRUE:
1674 : /* select only TRUE values */
1675 10 : selec = freq_true;
1676 10 : 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 10 : 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 10 : switch (booltesttype)
1706 : {
1707 10 : case IS_UNKNOWN:
1708 : /* select only NULL values */
1709 10 : selec = freq_null;
1710 10 : 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 761 : switch (booltesttype)
1743 : {
1744 40 : case IS_UNKNOWN:
1745 40 : selec = DEFAULT_UNK_SEL;
1746 40 : break;
1747 90 : case IS_NOT_UNKNOWN:
1748 90 : selec = DEFAULT_NOT_UNK_SEL;
1749 90 : break;
1750 210 : case IS_TRUE:
1751 : case IS_NOT_FALSE:
1752 210 : selec = (double) clause_selectivity(root, arg,
1753 : varRelid,
1754 : jointype, sjinfo);
1755 210 : break;
1756 421 : case IS_FALSE:
1757 : case IS_NOT_TRUE:
1758 421 : selec = 1.0 - (double) clause_selectivity(root, arg,
1759 : varRelid,
1760 : jointype, sjinfo);
1761 421 : 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 781 : ReleaseVariableStats(vardata);
1771 :
1772 : /* result should be in range, but make sure... */
1773 781 : CLAMP_PROBABILITY(selec);
1774 :
1775 781 : return (Selectivity) selec;
1776 : }
1777 :
1778 : /*
1779 : * nulltestsel - Selectivity of NullTest Node.
1780 : */
1781 : Selectivity
1782 13894 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1783 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1784 : {
1785 : VariableStatData vardata;
1786 : double selec;
1787 :
1788 13894 : examine_variable(root, arg, varRelid, &vardata);
1789 :
1790 13894 : if (HeapTupleIsValid(vardata.statsTuple))
1791 : {
1792 : Form_pg_statistic stats;
1793 : double freq_null;
1794 :
1795 7628 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1796 7628 : freq_null = stats->stanullfrac;
1797 :
1798 7628 : switch (nulltesttype)
1799 : {
1800 5848 : case IS_NULL:
1801 :
1802 : /*
1803 : * Use freq_null directly.
1804 : */
1805 5848 : selec = freq_null;
1806 5848 : break;
1807 1780 : case IS_NOT_NULL:
1808 :
1809 : /*
1810 : * Select not unknown (not null) values. Calculate from
1811 : * freq_null.
1812 : */
1813 1780 : selec = 1.0 - freq_null;
1814 1780 : break;
1815 0 : default:
1816 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1817 : (int) nulltesttype);
1818 : return (Selectivity) 0; /* keep compiler quiet */
1819 : }
1820 : }
1821 6266 : else if (vardata.var && IsA(vardata.var, Var) &&
1822 5752 : ((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 89 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1829 : }
1830 : else
1831 : {
1832 : /*
1833 : * No ANALYZE stats available, so make a guess
1834 : */
1835 6177 : switch (nulltesttype)
1836 : {
1837 1618 : case IS_NULL:
1838 1618 : selec = DEFAULT_UNK_SEL;
1839 1618 : break;
1840 4559 : case IS_NOT_NULL:
1841 4559 : selec = DEFAULT_NOT_UNK_SEL;
1842 4559 : 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 13894 : ReleaseVariableStats(vardata);
1851 :
1852 : /* result should be in range, but make sure... */
1853 13894 : CLAMP_PROBABILITY(selec);
1854 :
1855 13894 : 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 105946 : strip_array_coercion(Node *node)
1868 : {
1869 : for (;;)
1870 : {
1871 106034 : if (node && IsA(node, ArrayCoerceExpr))
1872 88 : {
1873 1969 : 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 1969 : if (IsA(acoerce->elemexpr, RelabelType) &&
1880 88 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1881 88 : node = (Node *) acoerce->arg;
1882 : else
1883 : break;
1884 : }
1885 104065 : 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 105946 : return node;
1894 : }
1895 :
1896 : /*
1897 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1898 : */
1899 : Selectivity
1900 18482 : scalararraysel(PlannerInfo *root,
1901 : ScalarArrayOpExpr *clause,
1902 : bool is_join_clause,
1903 : int varRelid,
1904 : JoinType jointype,
1905 : SpecialJoinInfo *sjinfo)
1906 : {
1907 18482 : Oid operator = clause->opno;
1908 18482 : bool useOr = clause->useOr;
1909 18482 : bool isEquality = false;
1910 18482 : 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 18482 : leftop = (Node *) linitial(clause->args);
1924 18482 : rightop = (Node *) lsecond(clause->args);
1925 :
1926 : /* aggressively reduce both sides to constants */
1927 18482 : leftop = estimate_expression_value(root, leftop);
1928 18482 : rightop = estimate_expression_value(root, rightop);
1929 :
1930 : /* get nominal (after relabeling) element type of rightop */
1931 18482 : nominal_element_type = get_base_element_type(exprType(rightop));
1932 18482 : if (!OidIsValid(nominal_element_type))
1933 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1934 : /* get nominal collation, too, for generating constants */
1935 18482 : nominal_element_collation = exprCollation(rightop);
1936 :
1937 : /* look through any binary-compatible relabeling of rightop */
1938 18482 : 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 18482 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1945 18482 : if (OidIsValid(typentry->eq_opr))
1946 : {
1947 18480 : if (operator == typentry->eq_opr)
1948 16091 : isEquality = true;
1949 2389 : else if (get_negator(operator) == typentry->eq_opr)
1950 1918 : 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 18482 : if ((isEquality || isInequality) && !is_join_clause)
1960 : {
1961 18008 : s1 = scalararraysel_containment(root, leftop, rightop,
1962 : nominal_element_type,
1963 : isEquality, useOr, varRelid);
1964 18008 : if (s1 >= 0.0)
1965 95 : return s1;
1966 : }
1967 :
1968 : /*
1969 : * Look up the underlying operator's selectivity estimator. Punt if it
1970 : * hasn't got one.
1971 : */
1972 18387 : if (is_join_clause)
1973 1 : oprsel = get_oprjoin(operator);
1974 : else
1975 18386 : oprsel = get_oprrest(operator);
1976 18387 : if (!oprsel)
1977 2 : return (Selectivity) 0.5;
1978 18385 : 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 18385 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1989 16158 : isEquality = true;
1990 2227 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1991 1827 : 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 18385 : if (rightop && IsA(rightop, Const))
2006 14849 : {
2007 14899 : Datum arraydatum = ((Const *) rightop)->constvalue;
2008 14899 : 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 14899 : if (arrayisnull) /* qual can't succeed if null array */
2019 50 : return (Selectivity) 0.0;
2020 14874 : arrayval = DatumGetArrayTypeP(arraydatum);
2021 :
2022 : /*
2023 : * When the array contains a NULL constant, same as var_eq_const, we
2024 : * assume the operator is strict and nothing will match, thus return
2025 : * 0.0.
2026 : */
2027 14874 : if (!useOr && array_contains_nulls(arrayval))
2028 25 : return (Selectivity) 0.0;
2029 :
2030 14849 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
2031 : &elmlen, &elmbyval, &elmalign);
2032 14849 : deconstruct_array(arrayval,
2033 : ARR_ELEMTYPE(arrayval),
2034 : elmlen, elmbyval, elmalign,
2035 : &elem_values, &elem_nulls, &num_elems);
2036 :
2037 : /*
2038 : * For generic operators, we assume the probability of success is
2039 : * independent for each array element. But for "= ANY" or "<> ALL",
2040 : * if the array elements are distinct (which'd typically be the case)
2041 : * then the probabilities are disjoint, and we should just sum them.
2042 : *
2043 : * If we were being really tense we would try to confirm that the
2044 : * elements are all distinct, but that would be expensive and it
2045 : * doesn't seem to be worth the cycles; it would amount to penalizing
2046 : * well-written queries in favor of poorly-written ones. However, we
2047 : * do protect ourselves a little bit by checking whether the
2048 : * disjointness assumption leads to an impossible (out of range)
2049 : * probability; if so, we fall back to the normal calculation.
2050 : */
2051 14849 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2052 :
2053 57983 : for (i = 0; i < num_elems; i++)
2054 : {
2055 : List *args;
2056 : Selectivity s2;
2057 :
2058 43134 : args = list_make2(leftop,
2059 : makeConst(nominal_element_type,
2060 : -1,
2061 : nominal_element_collation,
2062 : elmlen,
2063 : elem_values[i],
2064 : elem_nulls[i],
2065 : elmbyval));
2066 43134 : if (is_join_clause)
2067 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2068 : clause->inputcollid,
2069 : PointerGetDatum(root),
2070 : ObjectIdGetDatum(operator),
2071 : PointerGetDatum(args),
2072 : Int16GetDatum(jointype),
2073 : PointerGetDatum(sjinfo)));
2074 : else
2075 43134 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2076 : clause->inputcollid,
2077 : PointerGetDatum(root),
2078 : ObjectIdGetDatum(operator),
2079 : PointerGetDatum(args),
2080 : Int32GetDatum(varRelid)));
2081 :
2082 43134 : if (useOr)
2083 : {
2084 37284 : s1 = s1 + s2 - s1 * s2;
2085 37284 : if (isEquality)
2086 36414 : s1disjoint += s2;
2087 : }
2088 : else
2089 : {
2090 5850 : s1 = s1 * s2;
2091 5850 : if (isInequality)
2092 5590 : s1disjoint += s2 - 1.0;
2093 : }
2094 : }
2095 :
2096 : /* accept disjoint-probability estimate if in range */
2097 14849 : if ((useOr ? isEquality : isInequality) &&
2098 14339 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2099 14299 : s1 = s1disjoint;
2100 : }
2101 3486 : else if (rightop && IsA(rightop, ArrayExpr) &&
2102 297 : !((ArrayExpr *) rightop)->multidims)
2103 292 : {
2104 297 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2105 : int16 elmlen;
2106 : bool elmbyval;
2107 : ListCell *l;
2108 :
2109 297 : get_typlenbyval(arrayexpr->element_typeid,
2110 : &elmlen, &elmbyval);
2111 :
2112 : /*
2113 : * We use the assumption of disjoint probabilities here too, although
2114 : * the odds of equal array elements are rather higher if the elements
2115 : * are not all constants (which they won't be, else constant folding
2116 : * would have reduced the ArrayExpr to a Const). In this path it's
2117 : * critical to have the sanity check on the s1disjoint estimate.
2118 : */
2119 297 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2120 :
2121 1087 : foreach(l, arrayexpr->elements)
2122 : {
2123 795 : Node *elem = (Node *) lfirst(l);
2124 : List *args;
2125 : Selectivity s2;
2126 :
2127 : /*
2128 : * When the array contains a NULL constant, same as var_eq_const,
2129 : * we assume the operator is strict and nothing will match, thus
2130 : * return 0.0.
2131 : */
2132 795 : if (!useOr && IsA(elem, Const) && ((Const *) elem)->constisnull)
2133 5 : return (Selectivity) 0.0;
2134 :
2135 : /*
2136 : * Theoretically, if elem isn't of nominal_element_type we should
2137 : * insert a RelabelType, but it seems unlikely that any operator
2138 : * estimation function would really care ...
2139 : */
2140 790 : args = list_make2(leftop, elem);
2141 790 : if (is_join_clause)
2142 3 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2143 : clause->inputcollid,
2144 : PointerGetDatum(root),
2145 : ObjectIdGetDatum(operator),
2146 : PointerGetDatum(args),
2147 : Int16GetDatum(jointype),
2148 : PointerGetDatum(sjinfo)));
2149 : else
2150 787 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2151 : clause->inputcollid,
2152 : PointerGetDatum(root),
2153 : ObjectIdGetDatum(operator),
2154 : PointerGetDatum(args),
2155 : Int32GetDatum(varRelid)));
2156 :
2157 790 : if (useOr)
2158 : {
2159 770 : s1 = s1 + s2 - s1 * s2;
2160 770 : if (isEquality)
2161 770 : s1disjoint += s2;
2162 : }
2163 : else
2164 : {
2165 20 : s1 = s1 * s2;
2166 20 : if (isInequality)
2167 20 : s1disjoint += s2 - 1.0;
2168 : }
2169 : }
2170 :
2171 : /* accept disjoint-probability estimate if in range */
2172 292 : if ((useOr ? isEquality : isInequality) &&
2173 292 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2174 292 : s1 = s1disjoint;
2175 : }
2176 : else
2177 : {
2178 : CaseTestExpr *dummyexpr;
2179 : List *args;
2180 : Selectivity s2;
2181 : int i;
2182 :
2183 : /*
2184 : * We need a dummy rightop to pass to the operator selectivity
2185 : * routine. It can be pretty much anything that doesn't look like a
2186 : * constant; CaseTestExpr is a convenient choice.
2187 : */
2188 3189 : dummyexpr = makeNode(CaseTestExpr);
2189 3189 : dummyexpr->typeId = nominal_element_type;
2190 3189 : dummyexpr->typeMod = -1;
2191 3189 : dummyexpr->collation = clause->inputcollid;
2192 3189 : args = list_make2(leftop, dummyexpr);
2193 3189 : if (is_join_clause)
2194 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2195 : clause->inputcollid,
2196 : PointerGetDatum(root),
2197 : ObjectIdGetDatum(operator),
2198 : PointerGetDatum(args),
2199 : Int16GetDatum(jointype),
2200 : PointerGetDatum(sjinfo)));
2201 : else
2202 3189 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2203 : clause->inputcollid,
2204 : PointerGetDatum(root),
2205 : ObjectIdGetDatum(operator),
2206 : PointerGetDatum(args),
2207 : Int32GetDatum(varRelid)));
2208 3189 : s1 = useOr ? 0.0 : 1.0;
2209 :
2210 : /*
2211 : * Arbitrarily assume 10 elements in the eventual array value (see
2212 : * also estimate_array_length). We don't risk an assumption of
2213 : * disjoint probabilities here.
2214 : */
2215 35079 : for (i = 0; i < 10; i++)
2216 : {
2217 31890 : if (useOr)
2218 31890 : s1 = s1 + s2 - s1 * s2;
2219 : else
2220 0 : s1 = s1 * s2;
2221 : }
2222 : }
2223 :
2224 : /* result should be in range, but make sure... */
2225 18330 : CLAMP_PROBABILITY(s1);
2226 :
2227 18330 : return s1;
2228 : }
2229 :
2230 : /*
2231 : * Estimate number of elements in the array yielded by an expression.
2232 : *
2233 : * Note: the result is integral, but we use "double" to avoid overflow
2234 : * concerns. Most callers will use it in double-type expressions anyway.
2235 : *
2236 : * Note: in some code paths root can be passed as NULL, resulting in
2237 : * slightly worse estimates.
2238 : */
2239 : double
2240 87464 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2241 : {
2242 : /* look through any binary-compatible relabeling of arrayexpr */
2243 87464 : arrayexpr = strip_array_coercion(arrayexpr);
2244 :
2245 87464 : if (arrayexpr && IsA(arrayexpr, Const))
2246 : {
2247 38719 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2248 38719 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2249 : ArrayType *arrayval;
2250 :
2251 38719 : if (arrayisnull)
2252 75 : return 0;
2253 38644 : arrayval = DatumGetArrayTypeP(arraydatum);
2254 38644 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2255 : }
2256 48745 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2257 516 : !((ArrayExpr *) arrayexpr)->multidims)
2258 : {
2259 516 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2260 : }
2261 48229 : else if (arrayexpr && root)
2262 : {
2263 : /* See if we can find any statistics about it */
2264 : VariableStatData vardata;
2265 : AttStatsSlot sslot;
2266 48209 : double nelem = 0;
2267 :
2268 48209 : examine_variable(root, arrayexpr, 0, &vardata);
2269 48209 : if (HeapTupleIsValid(vardata.statsTuple))
2270 : {
2271 : /*
2272 : * Found stats, so use the average element count, which is stored
2273 : * in the last stanumbers element of the DECHIST statistics.
2274 : * Actually that is the average count of *distinct* elements;
2275 : * perhaps we should scale it up somewhat?
2276 : */
2277 8023 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2278 : STATISTIC_KIND_DECHIST, InvalidOid,
2279 : ATTSTATSSLOT_NUMBERS))
2280 : {
2281 7928 : if (sslot.nnumbers > 0)
2282 7928 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2283 7928 : free_attstatsslot(&sslot);
2284 : }
2285 : }
2286 48209 : ReleaseVariableStats(vardata);
2287 :
2288 48209 : if (nelem > 0)
2289 7928 : return nelem;
2290 : }
2291 :
2292 : /* Else use a default guess --- this should match scalararraysel */
2293 40301 : return 10;
2294 : }
2295 :
2296 : /*
2297 : * rowcomparesel - Selectivity of RowCompareExpr Node.
2298 : *
2299 : * We estimate RowCompare selectivity by considering just the first (high
2300 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2301 : * this estimate could be refined by considering additional columns, it
2302 : * seems unlikely that we could do a lot better without multi-column
2303 : * statistics.
2304 : */
2305 : Selectivity
2306 220 : rowcomparesel(PlannerInfo *root,
2307 : RowCompareExpr *clause,
2308 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2309 : {
2310 : Selectivity s1;
2311 220 : Oid opno = linitial_oid(clause->opnos);
2312 220 : Oid inputcollid = linitial_oid(clause->inputcollids);
2313 : List *opargs;
2314 : bool is_join_clause;
2315 :
2316 : /* Build equivalent arg list for single operator */
2317 220 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2318 :
2319 : /*
2320 : * Decide if it's a join clause. This should match clausesel.c's
2321 : * treat_as_join_clause(), except that we intentionally consider only the
2322 : * leading columns and not the rest of the clause.
2323 : */
2324 220 : if (varRelid != 0)
2325 : {
2326 : /*
2327 : * Caller is forcing restriction mode (eg, because we are examining an
2328 : * inner indexscan qual).
2329 : */
2330 45 : is_join_clause = false;
2331 : }
2332 175 : else if (sjinfo == NULL)
2333 : {
2334 : /*
2335 : * It must be a restriction clause, since it's being evaluated at a
2336 : * scan node.
2337 : */
2338 155 : is_join_clause = false;
2339 : }
2340 : else
2341 : {
2342 : /*
2343 : * Otherwise, it's a join if there's more than one base relation used.
2344 : */
2345 20 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2346 : }
2347 :
2348 220 : if (is_join_clause)
2349 : {
2350 : /* Estimate selectivity for a join clause. */
2351 20 : s1 = join_selectivity(root, opno,
2352 : opargs,
2353 : inputcollid,
2354 : jointype,
2355 : sjinfo);
2356 : }
2357 : else
2358 : {
2359 : /* Estimate selectivity for a restriction clause. */
2360 200 : s1 = restriction_selectivity(root, opno,
2361 : opargs,
2362 : inputcollid,
2363 : varRelid);
2364 : }
2365 :
2366 220 : return s1;
2367 : }
2368 :
2369 : /*
2370 : * eqjoinsel - Join selectivity of "="
2371 : */
2372 : Datum
2373 216866 : eqjoinsel(PG_FUNCTION_ARGS)
2374 : {
2375 216866 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2376 216866 : Oid operator = PG_GETARG_OID(1);
2377 216866 : List *args = (List *) PG_GETARG_POINTER(2);
2378 :
2379 : #ifdef NOT_USED
2380 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2381 : #endif
2382 216866 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2383 216866 : Oid collation = PG_GET_COLLATION();
2384 : double selec;
2385 : double selec_inner;
2386 : VariableStatData vardata1;
2387 : VariableStatData vardata2;
2388 : double nd1;
2389 : double nd2;
2390 : bool isdefault1;
2391 : bool isdefault2;
2392 : Oid opfuncoid;
2393 : FmgrInfo eqproc;
2394 216866 : Oid hashLeft = InvalidOid;
2395 216866 : Oid hashRight = InvalidOid;
2396 : AttStatsSlot sslot1;
2397 : AttStatsSlot sslot2;
2398 216866 : Form_pg_statistic stats1 = NULL;
2399 216866 : Form_pg_statistic stats2 = NULL;
2400 216866 : bool have_mcvs1 = false;
2401 216866 : bool have_mcvs2 = false;
2402 216866 : bool *hasmatch1 = NULL;
2403 216866 : bool *hasmatch2 = NULL;
2404 216866 : int nmatches = 0;
2405 : bool get_mcv_stats;
2406 : bool join_is_reversed;
2407 : RelOptInfo *inner_rel;
2408 :
2409 216866 : get_join_variables(root, args, sjinfo,
2410 : &vardata1, &vardata2, &join_is_reversed);
2411 :
2412 216866 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2413 216866 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2414 :
2415 216866 : opfuncoid = get_opcode(operator);
2416 :
2417 216866 : memset(&sslot1, 0, sizeof(sslot1));
2418 216866 : memset(&sslot2, 0, sizeof(sslot2));
2419 :
2420 : /*
2421 : * There is no use in fetching one side's MCVs if we lack MCVs for the
2422 : * other side, so do a quick check to verify that both stats exist.
2423 : */
2424 584094 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2425 263729 : HeapTupleIsValid(vardata2.statsTuple) &&
2426 113367 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2427 : STATISTIC_KIND_MCV, InvalidOid,
2428 367228 : 0) &&
2429 56423 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2430 : STATISTIC_KIND_MCV, InvalidOid,
2431 : 0));
2432 :
2433 216866 : if (HeapTupleIsValid(vardata1.statsTuple))
2434 : {
2435 : /* note we allow use of nullfrac regardless of security check */
2436 150362 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2437 173152 : if (get_mcv_stats &&
2438 22790 : statistic_proc_security_check(&vardata1, opfuncoid))
2439 22790 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2440 : STATISTIC_KIND_MCV, InvalidOid,
2441 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2442 : }
2443 :
2444 216866 : if (HeapTupleIsValid(vardata2.statsTuple))
2445 : {
2446 : /* note we allow use of nullfrac regardless of security check */
2447 130307 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2448 153097 : if (get_mcv_stats &&
2449 22790 : statistic_proc_security_check(&vardata2, opfuncoid))
2450 22790 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2451 : STATISTIC_KIND_MCV, InvalidOid,
2452 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2453 : }
2454 :
2455 : /* Prepare info usable by both eqjoinsel_inner and eqjoinsel_semi */
2456 216866 : if (have_mcvs1 && have_mcvs2)
2457 : {
2458 22790 : fmgr_info(opfuncoid, &eqproc);
2459 22790 : hasmatch1 = (bool *) palloc0(sslot1.nvalues * sizeof(bool));
2460 22790 : hasmatch2 = (bool *) palloc0(sslot2.nvalues * sizeof(bool));
2461 :
2462 : /*
2463 : * If the MCV lists are long enough to justify hashing, try to look up
2464 : * hash functions for the join operator.
2465 : */
2466 22790 : if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
2467 1292 : (void) get_op_hash_functions(operator, &hashLeft, &hashRight);
2468 : }
2469 : else
2470 194076 : memset(&eqproc, 0, sizeof(eqproc)); /* silence uninit-var warnings */
2471 :
2472 : /* We need to compute the inner-join selectivity in all cases */
2473 216866 : selec_inner = eqjoinsel_inner(&eqproc, collation,
2474 : hashLeft, hashRight,
2475 : &vardata1, &vardata2,
2476 : nd1, nd2,
2477 : isdefault1, isdefault2,
2478 : &sslot1, &sslot2,
2479 : stats1, stats2,
2480 : have_mcvs1, have_mcvs2,
2481 : hasmatch1, hasmatch2,
2482 : &nmatches);
2483 :
2484 216866 : switch (sjinfo->jointype)
2485 : {
2486 205658 : case JOIN_INNER:
2487 : case JOIN_LEFT:
2488 : case JOIN_FULL:
2489 205658 : selec = selec_inner;
2490 205658 : break;
2491 11208 : case JOIN_SEMI:
2492 : case JOIN_ANTI:
2493 :
2494 : /*
2495 : * Look up the join's inner relation. min_righthand is sufficient
2496 : * information because neither SEMI nor ANTI joins permit any
2497 : * reassociation into or out of their RHS, so the righthand will
2498 : * always be exactly that set of rels.
2499 : */
2500 11208 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2501 :
2502 11208 : if (!join_is_reversed)
2503 5379 : selec = eqjoinsel_semi(&eqproc, collation,
2504 : hashLeft, hashRight,
2505 : false,
2506 : &vardata1, &vardata2,
2507 : nd1, nd2,
2508 : isdefault1, isdefault2,
2509 : &sslot1, &sslot2,
2510 : stats1, stats2,
2511 : have_mcvs1, have_mcvs2,
2512 : hasmatch1, hasmatch2,
2513 : &nmatches,
2514 : inner_rel);
2515 : else
2516 5829 : selec = eqjoinsel_semi(&eqproc, collation,
2517 : hashLeft, hashRight,
2518 : true,
2519 : &vardata2, &vardata1,
2520 : nd2, nd1,
2521 : isdefault2, isdefault1,
2522 : &sslot2, &sslot1,
2523 : stats2, stats1,
2524 : have_mcvs2, have_mcvs1,
2525 : hasmatch2, hasmatch1,
2526 : &nmatches,
2527 : inner_rel);
2528 :
2529 : /*
2530 : * We should never estimate the output of a semijoin to be more
2531 : * rows than we estimate for an inner join with the same input
2532 : * rels and join condition; it's obviously impossible for that to
2533 : * happen. The former estimate is N1 * Ssemi while the latter is
2534 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2535 : * this is worthwhile because of the shakier estimation rules we
2536 : * use in eqjoinsel_semi, particularly in cases where it has to
2537 : * punt entirely.
2538 : */
2539 11208 : selec = Min(selec, inner_rel->rows * selec_inner);
2540 11208 : break;
2541 0 : default:
2542 : /* other values not expected here */
2543 0 : elog(ERROR, "unrecognized join type: %d",
2544 : (int) sjinfo->jointype);
2545 : selec = 0; /* keep compiler quiet */
2546 : break;
2547 : }
2548 :
2549 216866 : free_attstatsslot(&sslot1);
2550 216866 : free_attstatsslot(&sslot2);
2551 :
2552 216866 : ReleaseVariableStats(vardata1);
2553 216866 : ReleaseVariableStats(vardata2);
2554 :
2555 216866 : if (hasmatch1)
2556 22790 : pfree(hasmatch1);
2557 216866 : if (hasmatch2)
2558 22790 : pfree(hasmatch2);
2559 :
2560 216866 : CLAMP_PROBABILITY(selec);
2561 :
2562 216866 : PG_RETURN_FLOAT8((float8) selec);
2563 : }
2564 :
2565 : /*
2566 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2567 : *
2568 : * In addition to computing the selectivity estimate, this will fill
2569 : * hasmatch1[], hasmatch2[], and *p_nmatches (if have_mcvs1 && have_mcvs2).
2570 : * We may be able to re-use that data in eqjoinsel_semi.
2571 : *
2572 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2573 : * that it's worth trying to distinguish them here.
2574 : */
2575 : static double
2576 216866 : eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
2577 : Oid hashLeft, Oid hashRight,
2578 : VariableStatData *vardata1, VariableStatData *vardata2,
2579 : double nd1, double nd2,
2580 : bool isdefault1, bool isdefault2,
2581 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2582 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2583 : bool have_mcvs1, bool have_mcvs2,
2584 : bool *hasmatch1, bool *hasmatch2,
2585 : int *p_nmatches)
2586 : {
2587 : double selec;
2588 :
2589 216866 : if (have_mcvs1 && have_mcvs2)
2590 22790 : {
2591 : /*
2592 : * We have most-common-value lists for both relations. Run through
2593 : * the lists to see which MCVs actually join to each other with the
2594 : * given operator. This allows us to determine the exact join
2595 : * selectivity for the portion of the relations represented by the MCV
2596 : * lists. We still have to estimate for the remaining population, but
2597 : * in a skewed distribution this gives us a big leg up in accuracy.
2598 : * For motivation see the analysis in Y. Ioannidis and S.
2599 : * Christodoulakis, "On the propagation of errors in the size of join
2600 : * results", Technical Report 1018, Computer Science Dept., University
2601 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2602 : */
2603 22790 : double nullfrac1 = stats1->stanullfrac;
2604 22790 : double nullfrac2 = stats2->stanullfrac;
2605 : double matchprodfreq,
2606 : matchfreq1,
2607 : matchfreq2,
2608 : unmatchfreq1,
2609 : unmatchfreq2,
2610 : otherfreq1,
2611 : otherfreq2,
2612 : totalsel1,
2613 : totalsel2;
2614 : int i,
2615 : nmatches;
2616 :
2617 : /* Fill the match arrays */
2618 22790 : eqjoinsel_find_matches(eqproc, collation,
2619 : hashLeft, hashRight,
2620 : false,
2621 : sslot1, sslot2,
2622 : sslot1->nvalues, sslot2->nvalues,
2623 : hasmatch1, hasmatch2,
2624 : p_nmatches, &matchprodfreq);
2625 22790 : nmatches = *p_nmatches;
2626 22790 : CLAMP_PROBABILITY(matchprodfreq);
2627 :
2628 : /* Sum up frequencies of matched and unmatched MCVs */
2629 22790 : matchfreq1 = unmatchfreq1 = 0.0;
2630 527671 : for (i = 0; i < sslot1->nvalues; i++)
2631 : {
2632 504881 : if (hasmatch1[i])
2633 235526 : matchfreq1 += sslot1->numbers[i];
2634 : else
2635 269355 : unmatchfreq1 += sslot1->numbers[i];
2636 : }
2637 22790 : CLAMP_PROBABILITY(matchfreq1);
2638 22790 : CLAMP_PROBABILITY(unmatchfreq1);
2639 22790 : matchfreq2 = unmatchfreq2 = 0.0;
2640 415840 : for (i = 0; i < sslot2->nvalues; i++)
2641 : {
2642 393050 : if (hasmatch2[i])
2643 235526 : matchfreq2 += sslot2->numbers[i];
2644 : else
2645 157524 : unmatchfreq2 += sslot2->numbers[i];
2646 : }
2647 22790 : CLAMP_PROBABILITY(matchfreq2);
2648 22790 : CLAMP_PROBABILITY(unmatchfreq2);
2649 :
2650 : /*
2651 : * Compute total frequency of non-null values that are not in the MCV
2652 : * lists.
2653 : */
2654 22790 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2655 22790 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2656 22790 : CLAMP_PROBABILITY(otherfreq1);
2657 22790 : CLAMP_PROBABILITY(otherfreq2);
2658 :
2659 : /*
2660 : * We can estimate the total selectivity from the point of view of
2661 : * relation 1 as: the known selectivity for matched MCVs, plus
2662 : * unmatched MCVs that are assumed to match against random members of
2663 : * relation 2's non-MCV population, plus non-MCV values that are
2664 : * assumed to match against random members of relation 2's unmatched
2665 : * MCVs plus non-MCV values.
2666 : */
2667 22790 : totalsel1 = matchprodfreq;
2668 22790 : if (nd2 > sslot2->nvalues)
2669 4538 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2670 22790 : if (nd2 > nmatches)
2671 8471 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2672 8471 : (nd2 - nmatches);
2673 : /* Same estimate from the point of view of relation 2. */
2674 22790 : totalsel2 = matchprodfreq;
2675 22790 : if (nd1 > sslot1->nvalues)
2676 5018 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2677 22790 : if (nd1 > nmatches)
2678 7455 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2679 7455 : (nd1 - nmatches);
2680 :
2681 : /*
2682 : * Use the smaller of the two estimates. This can be justified in
2683 : * essentially the same terms as given below for the no-stats case: to
2684 : * a first approximation, we are estimating from the point of view of
2685 : * the relation with smaller nd.
2686 : */
2687 22790 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2688 : }
2689 : else
2690 : {
2691 : /*
2692 : * We do not have MCV lists for both sides. Estimate the join
2693 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2694 : * is plausible if we assume that the join operator is strict and the
2695 : * non-null values are about equally distributed: a given non-null
2696 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2697 : * of rel2, so total join rows are at most
2698 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2699 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2700 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2701 : * with MIN() is an upper bound. Using the MIN() means we estimate
2702 : * from the point of view of the relation with smaller nd (since the
2703 : * larger nd is determining the MIN). It is reasonable to assume that
2704 : * most tuples in this rel will have join partners, so the bound is
2705 : * probably reasonably tight and should be taken as-is.
2706 : *
2707 : * XXX Can we be smarter if we have an MCV list for just one side? It
2708 : * seems that if we assume equal distribution for the other side, we
2709 : * end up with the same answer anyway.
2710 : */
2711 194076 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2712 194076 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2713 :
2714 194076 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2715 194076 : if (nd1 > nd2)
2716 96226 : selec /= nd1;
2717 : else
2718 97850 : selec /= nd2;
2719 : }
2720 :
2721 216866 : return selec;
2722 : }
2723 :
2724 : /*
2725 : * eqjoinsel_semi --- eqjoinsel for semi join
2726 : *
2727 : * (Also used for anti join, which we are supposed to estimate the same way.)
2728 : * Caller has ensured that vardata1 is the LHS variable; however, eqproc
2729 : * is for the original join operator, which might now need to have the inputs
2730 : * swapped in order to apply correctly. Also, if have_mcvs1 && have_mcvs2
2731 : * then hasmatch1[], hasmatch2[], and *p_nmatches were filled by
2732 : * eqjoinsel_inner.
2733 : */
2734 : static double
2735 11208 : eqjoinsel_semi(FmgrInfo *eqproc, Oid collation,
2736 : Oid hashLeft, Oid hashRight,
2737 : bool op_is_reversed,
2738 : VariableStatData *vardata1, VariableStatData *vardata2,
2739 : double nd1, double nd2,
2740 : bool isdefault1, bool isdefault2,
2741 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2742 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2743 : bool have_mcvs1, bool have_mcvs2,
2744 : bool *hasmatch1, bool *hasmatch2,
2745 : int *p_nmatches,
2746 : RelOptInfo *inner_rel)
2747 : {
2748 : double selec;
2749 :
2750 : /*
2751 : * We clamp nd2 to be not more than what we estimate the inner relation's
2752 : * size to be. This is intuitively somewhat reasonable since obviously
2753 : * there can't be more than that many distinct values coming from the
2754 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2755 : * likewise) is that this is the only pathway by which restriction clauses
2756 : * applied to the inner rel will affect the join result size estimate,
2757 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2758 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2759 : * the selectivity of outer-rel restrictions.
2760 : *
2761 : * We can apply this clamping both with respect to the base relation from
2762 : * which the join variable comes (if there is just one), and to the
2763 : * immediate inner input relation of the current join.
2764 : *
2765 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2766 : * great, maybe, but it didn't come out of nowhere either. This is most
2767 : * helpful when the inner relation is empty and consequently has no stats.
2768 : */
2769 11208 : if (vardata2->rel)
2770 : {
2771 11203 : if (nd2 >= vardata2->rel->rows)
2772 : {
2773 9456 : nd2 = vardata2->rel->rows;
2774 9456 : isdefault2 = false;
2775 : }
2776 : }
2777 11208 : if (nd2 >= inner_rel->rows)
2778 : {
2779 9409 : nd2 = inner_rel->rows;
2780 9409 : isdefault2 = false;
2781 : }
2782 :
2783 11208 : if (have_mcvs1 && have_mcvs2)
2784 519 : {
2785 : /*
2786 : * We have most-common-value lists for both relations. Run through
2787 : * the lists to see which MCVs actually join to each other with the
2788 : * given operator. This allows us to determine the exact join
2789 : * selectivity for the portion of the relations represented by the MCV
2790 : * lists. We still have to estimate for the remaining population, but
2791 : * in a skewed distribution this gives us a big leg up in accuracy.
2792 : */
2793 519 : double nullfrac1 = stats1->stanullfrac;
2794 : double matchprodfreq,
2795 : matchfreq1,
2796 : uncertainfrac,
2797 : uncertain;
2798 : int i,
2799 : nmatches,
2800 : clamped_nvalues2;
2801 :
2802 : /*
2803 : * The clamping above could have resulted in nd2 being less than
2804 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2805 : * most common values in the relation will appear in the join input,
2806 : * and so compare to only the first nd2 members of the MCV list. Of
2807 : * course this is frequently wrong, but it's the best bet we can make.
2808 : */
2809 519 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2810 :
2811 : /*
2812 : * If we did not set clamped_nvalues2 to less than sslot2->nvalues,
2813 : * then the hasmatch1[] and hasmatch2[] match flags computed by
2814 : * eqjoinsel_inner are still perfectly applicable, so we need not
2815 : * re-do the matching work. Note that it does not matter if
2816 : * op_is_reversed: we'd get the same answers.
2817 : *
2818 : * If we did clamp, then a different set of sslot2 values is to be
2819 : * compared, so we have to re-do the matching.
2820 : */
2821 519 : if (clamped_nvalues2 != sslot2->nvalues)
2822 : {
2823 : /* Must re-zero the arrays */
2824 0 : memset(hasmatch1, 0, sslot1->nvalues * sizeof(bool));
2825 0 : memset(hasmatch2, 0, clamped_nvalues2 * sizeof(bool));
2826 : /* Re-fill the match arrays */
2827 0 : eqjoinsel_find_matches(eqproc, collation,
2828 : hashLeft, hashRight,
2829 : op_is_reversed,
2830 : sslot1, sslot2,
2831 : sslot1->nvalues, clamped_nvalues2,
2832 : hasmatch1, hasmatch2,
2833 : p_nmatches, &matchprodfreq);
2834 : }
2835 519 : nmatches = *p_nmatches;
2836 :
2837 : /* Sum up frequencies of matched MCVs */
2838 519 : matchfreq1 = 0.0;
2839 11613 : for (i = 0; i < sslot1->nvalues; i++)
2840 : {
2841 11094 : if (hasmatch1[i])
2842 9534 : matchfreq1 += sslot1->numbers[i];
2843 : }
2844 519 : CLAMP_PROBABILITY(matchfreq1);
2845 :
2846 : /*
2847 : * Now we need to estimate the fraction of relation 1 that has at
2848 : * least one join partner. We know for certain that the matched MCVs
2849 : * do, so that gives us a lower bound, but we're really in the dark
2850 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2851 : * assume all non-null rel1 rows have join partners, else assume for
2852 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2853 : * can discount the known-matched MCVs from the distinct-values counts
2854 : * before doing the division.
2855 : *
2856 : * Crude as the above is, it's completely useless if we don't have
2857 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2858 : * nd2 is default, punt and assume half of the uncertain rows have
2859 : * join partners.
2860 : */
2861 519 : if (!isdefault1 && !isdefault2)
2862 : {
2863 519 : nd1 -= nmatches;
2864 519 : nd2 -= nmatches;
2865 519 : if (nd1 <= nd2 || nd2 < 0)
2866 489 : uncertainfrac = 1.0;
2867 : else
2868 30 : uncertainfrac = nd2 / nd1;
2869 : }
2870 : else
2871 0 : uncertainfrac = 0.5;
2872 519 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2873 519 : CLAMP_PROBABILITY(uncertain);
2874 519 : selec = matchfreq1 + uncertainfrac * uncertain;
2875 : }
2876 : else
2877 : {
2878 : /*
2879 : * Without MCV lists for both sides, we can only use the heuristic
2880 : * about nd1 vs nd2.
2881 : */
2882 10689 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2883 :
2884 10689 : if (!isdefault1 && !isdefault2)
2885 : {
2886 7984 : if (nd1 <= nd2 || nd2 < 0)
2887 4036 : selec = 1.0 - nullfrac1;
2888 : else
2889 3948 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2890 : }
2891 : else
2892 2705 : selec = 0.5 * (1.0 - nullfrac1);
2893 : }
2894 :
2895 11208 : return selec;
2896 : }
2897 :
2898 : /*
2899 : * Identify matching MCVs for eqjoinsel_inner or eqjoinsel_semi.
2900 : *
2901 : * Inputs:
2902 : * eqproc: FmgrInfo for equality function to use (might be reversed)
2903 : * collation: OID of collation to use
2904 : * hashLeft, hashRight: OIDs of hash functions associated with equality op,
2905 : * or InvalidOid if we're not to use hashing
2906 : * op_is_reversed: indicates that eqproc compares right type to left type
2907 : * sslot1, sslot2: MCV values for the lefthand and righthand inputs
2908 : * nvalues1, nvalues2: number of values to be considered (can be less than
2909 : * sslotN->nvalues, but not more)
2910 : * Outputs:
2911 : * hasmatch1[], hasmatch2[]: pre-zeroed arrays of lengths nvalues1, nvalues2;
2912 : * entries are set to true if that MCV has a match on the other side
2913 : * *p_nmatches: receives number of MCV pairs that match
2914 : * *p_matchprodfreq: receives sum(sslot1->numbers[i] * sslot2->numbers[j])
2915 : * for matching MCVs
2916 : *
2917 : * Note that hashLeft is for the eqproc's left-hand input type, hashRight
2918 : * for its right, regardless of op_is_reversed.
2919 : *
2920 : * Note we assume that each MCV will match at most one member of the other
2921 : * MCV list. If the operator isn't really equality, there could be multiple
2922 : * matches --- but we don't look for them, both for speed and because the
2923 : * math wouldn't add up...
2924 : */
2925 : static void
2926 22790 : eqjoinsel_find_matches(FmgrInfo *eqproc, Oid collation,
2927 : Oid hashLeft, Oid hashRight,
2928 : bool op_is_reversed,
2929 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2930 : int nvalues1, int nvalues2,
2931 : bool *hasmatch1, bool *hasmatch2,
2932 : int *p_nmatches, double *p_matchprodfreq)
2933 : {
2934 22790 : LOCAL_FCINFO(fcinfo, 2);
2935 22790 : double matchprodfreq = 0.0;
2936 22790 : int nmatches = 0;
2937 :
2938 : /*
2939 : * Save a few cycles by setting up the fcinfo struct just once. Using
2940 : * FunctionCallInvoke directly also avoids failure if the eqproc returns
2941 : * NULL, though really equality functions should never do that.
2942 : */
2943 22790 : InitFunctionCallInfoData(*fcinfo, eqproc, 2, collation,
2944 : NULL, NULL);
2945 22790 : fcinfo->args[0].isnull = false;
2946 22790 : fcinfo->args[1].isnull = false;
2947 :
2948 22790 : if (OidIsValid(hashLeft) && OidIsValid(hashRight))
2949 1292 : {
2950 : /* Use a hash table to speed up the matching */
2951 1292 : LOCAL_FCINFO(hash_fcinfo, 1);
2952 : FmgrInfo hash_proc;
2953 : MCVHashContext hashContext;
2954 : MCVHashTable_hash *hashTable;
2955 : AttStatsSlot *statsProbe;
2956 : AttStatsSlot *statsHash;
2957 : bool *hasMatchProbe;
2958 : bool *hasMatchHash;
2959 : int nvaluesProbe;
2960 : int nvaluesHash;
2961 :
2962 : /* Make sure we build the hash table on the smaller array. */
2963 1292 : if (sslot1->nvalues >= sslot2->nvalues)
2964 : {
2965 1292 : statsProbe = sslot1;
2966 1292 : statsHash = sslot2;
2967 1292 : hasMatchProbe = hasmatch1;
2968 1292 : hasMatchHash = hasmatch2;
2969 1292 : nvaluesProbe = nvalues1;
2970 1292 : nvaluesHash = nvalues2;
2971 : }
2972 : else
2973 : {
2974 : /* We'll have to reverse the direction of use of the operator. */
2975 0 : op_is_reversed = !op_is_reversed;
2976 0 : statsProbe = sslot2;
2977 0 : statsHash = sslot1;
2978 0 : hasMatchProbe = hasmatch2;
2979 0 : hasMatchHash = hasmatch1;
2980 0 : nvaluesProbe = nvalues2;
2981 0 : nvaluesHash = nvalues1;
2982 : }
2983 :
2984 : /*
2985 : * Build the hash table on the smaller array, using the appropriate
2986 : * hash function for its data type.
2987 : */
2988 1292 : fmgr_info(op_is_reversed ? hashLeft : hashRight, &hash_proc);
2989 1292 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
2990 : NULL, NULL);
2991 1292 : hash_fcinfo->args[0].isnull = false;
2992 :
2993 1292 : hashContext.equal_fcinfo = fcinfo;
2994 1292 : hashContext.hash_fcinfo = hash_fcinfo;
2995 1292 : hashContext.op_is_reversed = op_is_reversed;
2996 1292 : hashContext.insert_mode = true;
2997 1292 : get_typlenbyval(statsHash->valuetype,
2998 : &hashContext.hash_typlen,
2999 : &hashContext.hash_typbyval);
3000 :
3001 1292 : hashTable = MCVHashTable_create(CurrentMemoryContext,
3002 : nvaluesHash,
3003 : &hashContext);
3004 :
3005 130492 : for (int i = 0; i < nvaluesHash; i++)
3006 : {
3007 129200 : bool found = false;
3008 129200 : MCVHashEntry *entry = MCVHashTable_insert(hashTable,
3009 129200 : statsHash->values[i],
3010 : &found);
3011 :
3012 : /*
3013 : * MCVHashTable_insert will only report "found" if the new value
3014 : * is equal to some previous one per datum_image_eq(). That
3015 : * probably shouldn't happen, since we're not expecting duplicates
3016 : * in the MCV list. If we do find a dup, just ignore it, leaving
3017 : * the hash entry's index pointing at the first occurrence. That
3018 : * matches the behavior that the non-hashed code path would have.
3019 : */
3020 129200 : if (likely(!found))
3021 129200 : entry->index = i;
3022 : }
3023 :
3024 : /*
3025 : * Prepare to probe the hash table. If the probe values are of a
3026 : * different data type, then we need to change hash functions. (This
3027 : * code relies on the assumption that since we defined SH_STORE_HASH,
3028 : * simplehash.h will never need to compute hash values for existing
3029 : * hash table entries.)
3030 : */
3031 1292 : hashContext.insert_mode = false;
3032 1292 : if (hashLeft != hashRight)
3033 : {
3034 0 : fmgr_info(op_is_reversed ? hashRight : hashLeft, &hash_proc);
3035 : /* Resetting hash_fcinfo is probably unnecessary, but be safe */
3036 0 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
3037 : NULL, NULL);
3038 0 : hash_fcinfo->args[0].isnull = false;
3039 : }
3040 :
3041 : /* Look up each probe value in turn. */
3042 130492 : for (int i = 0; i < nvaluesProbe; i++)
3043 : {
3044 129200 : MCVHashEntry *entry = MCVHashTable_lookup(hashTable,
3045 129200 : statsProbe->values[i]);
3046 :
3047 : /* As in the other code path, skip already-matched hash entries */
3048 129200 : if (entry != NULL && !hasMatchHash[entry->index])
3049 : {
3050 54341 : hasMatchHash[entry->index] = hasMatchProbe[i] = true;
3051 54341 : nmatches++;
3052 54341 : matchprodfreq += statsHash->numbers[entry->index] * statsProbe->numbers[i];
3053 : }
3054 : }
3055 :
3056 1292 : MCVHashTable_destroy(hashTable);
3057 : }
3058 : else
3059 : {
3060 : /* We're not to use hashing, so do it the O(N^2) way */
3061 : int index1,
3062 : index2;
3063 :
3064 : /* Set up to supply the values in the order the operator expects */
3065 21498 : if (op_is_reversed)
3066 : {
3067 0 : index1 = 1;
3068 0 : index2 = 0;
3069 : }
3070 : else
3071 : {
3072 21498 : index1 = 0;
3073 21498 : index2 = 1;
3074 : }
3075 :
3076 397179 : for (int i = 0; i < nvalues1; i++)
3077 : {
3078 375681 : fcinfo->args[index1].value = sslot1->values[i];
3079 :
3080 7984187 : for (int j = 0; j < nvalues2; j++)
3081 : {
3082 : Datum fresult;
3083 :
3084 7789691 : if (hasmatch2[j])
3085 2458852 : continue;
3086 5330839 : fcinfo->args[index2].value = sslot2->values[j];
3087 5330839 : fcinfo->isnull = false;
3088 5330839 : fresult = FunctionCallInvoke(fcinfo);
3089 5330839 : if (!fcinfo->isnull && DatumGetBool(fresult))
3090 : {
3091 181185 : hasmatch1[i] = hasmatch2[j] = true;
3092 181185 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
3093 181185 : nmatches++;
3094 181185 : break;
3095 : }
3096 : }
3097 : }
3098 : }
3099 :
3100 22790 : *p_nmatches = nmatches;
3101 22790 : *p_matchprodfreq = matchprodfreq;
3102 22790 : }
3103 :
3104 : /*
3105 : * Support functions for the hash tables used by eqjoinsel_find_matches
3106 : */
3107 : static uint32
3108 258400 : hash_mcv(MCVHashTable_hash *tab, Datum key)
3109 : {
3110 258400 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3111 258400 : FunctionCallInfo fcinfo = context->hash_fcinfo;
3112 : Datum fresult;
3113 :
3114 258400 : fcinfo->args[0].value = key;
3115 258400 : fcinfo->isnull = false;
3116 258400 : fresult = FunctionCallInvoke(fcinfo);
3117 : Assert(!fcinfo->isnull);
3118 258400 : return DatumGetUInt32(fresult);
3119 : }
3120 :
3121 : static bool
3122 54341 : mcvs_equal(MCVHashTable_hash *tab, Datum key0, Datum key1)
3123 : {
3124 54341 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3125 :
3126 54341 : if (context->insert_mode)
3127 : {
3128 : /*
3129 : * During the insertion step, any comparisons will be between two
3130 : * Datums of the hash table's data type, so if the given operator is
3131 : * cross-type it will be the wrong thing to use. Fortunately, we can
3132 : * use datum_image_eq instead. The MCV values should all be distinct
3133 : * anyway, so it's mostly pro-forma to compare them at all.
3134 : */
3135 0 : return datum_image_eq(key0, key1,
3136 0 : context->hash_typbyval, context->hash_typlen);
3137 : }
3138 : else
3139 : {
3140 54341 : FunctionCallInfo fcinfo = context->equal_fcinfo;
3141 : Datum fresult;
3142 :
3143 : /*
3144 : * Apply the operator the correct way around. Although simplehash.h
3145 : * doesn't document this explicitly, during lookups key0 is from the
3146 : * hash table while key1 is the probe value, so we should compare them
3147 : * in that order only if op_is_reversed.
3148 : */
3149 54341 : if (context->op_is_reversed)
3150 : {
3151 0 : fcinfo->args[0].value = key0;
3152 0 : fcinfo->args[1].value = key1;
3153 : }
3154 : else
3155 : {
3156 54341 : fcinfo->args[0].value = key1;
3157 54341 : fcinfo->args[1].value = key0;
3158 : }
3159 54341 : fcinfo->isnull = false;
3160 54341 : fresult = FunctionCallInvoke(fcinfo);
3161 54341 : return (!fcinfo->isnull && DatumGetBool(fresult));
3162 : }
3163 : }
3164 :
3165 : /*
3166 : * neqjoinsel - Join selectivity of "!="
3167 : */
3168 : Datum
3169 2938 : neqjoinsel(PG_FUNCTION_ARGS)
3170 : {
3171 2938 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3172 2938 : Oid operator = PG_GETARG_OID(1);
3173 2938 : List *args = (List *) PG_GETARG_POINTER(2);
3174 2938 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
3175 2938 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
3176 2938 : Oid collation = PG_GET_COLLATION();
3177 : float8 result;
3178 :
3179 2938 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
3180 1015 : {
3181 : /*
3182 : * For semi-joins, if there is more than one distinct value in the RHS
3183 : * relation then every non-null LHS row must find a row to join since
3184 : * it can only be equal to one of them. We'll assume that there is
3185 : * always more than one distinct RHS value for the sake of stability,
3186 : * though in theory we could have special cases for empty RHS
3187 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
3188 : * fraction of LHS that has the same value as the single RHS value).
3189 : *
3190 : * For anti-joins, if we use the same assumption that there is more
3191 : * than one distinct key in the RHS relation, then every non-null LHS
3192 : * row must be suppressed by the anti-join.
3193 : *
3194 : * So either way, the selectivity estimate should be 1 - nullfrac.
3195 : */
3196 : VariableStatData leftvar;
3197 : VariableStatData rightvar;
3198 : bool reversed;
3199 : HeapTuple statsTuple;
3200 : double nullfrac;
3201 :
3202 1015 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
3203 1015 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
3204 1015 : if (HeapTupleIsValid(statsTuple))
3205 838 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
3206 : else
3207 177 : nullfrac = 0.0;
3208 1015 : ReleaseVariableStats(leftvar);
3209 1015 : ReleaseVariableStats(rightvar);
3210 :
3211 1015 : result = 1.0 - nullfrac;
3212 : }
3213 : else
3214 : {
3215 : /*
3216 : * We want 1 - eqjoinsel() where the equality operator is the one
3217 : * associated with this != operator, that is, its negator.
3218 : */
3219 1923 : Oid eqop = get_negator(operator);
3220 :
3221 1923 : if (eqop)
3222 : {
3223 : result =
3224 1923 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
3225 : collation,
3226 : PointerGetDatum(root),
3227 : ObjectIdGetDatum(eqop),
3228 : PointerGetDatum(args),
3229 : Int16GetDatum(jointype),
3230 : PointerGetDatum(sjinfo)));
3231 : }
3232 : else
3233 : {
3234 : /* Use default selectivity (should we raise an error instead?) */
3235 0 : result = DEFAULT_EQ_SEL;
3236 : }
3237 1923 : result = 1.0 - result;
3238 : }
3239 :
3240 2938 : PG_RETURN_FLOAT8(result);
3241 : }
3242 :
3243 : /*
3244 : * scalarltjoinsel - Join selectivity of "<" for scalars
3245 : */
3246 : Datum
3247 270 : scalarltjoinsel(PG_FUNCTION_ARGS)
3248 : {
3249 270 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3250 : }
3251 :
3252 : /*
3253 : * scalarlejoinsel - Join selectivity of "<=" for scalars
3254 : */
3255 : Datum
3256 198 : scalarlejoinsel(PG_FUNCTION_ARGS)
3257 : {
3258 198 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3259 : }
3260 :
3261 : /*
3262 : * scalargtjoinsel - Join selectivity of ">" for scalars
3263 : */
3264 : Datum
3265 240 : scalargtjoinsel(PG_FUNCTION_ARGS)
3266 : {
3267 240 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3268 : }
3269 :
3270 : /*
3271 : * scalargejoinsel - Join selectivity of ">=" for scalars
3272 : */
3273 : Datum
3274 152 : scalargejoinsel(PG_FUNCTION_ARGS)
3275 : {
3276 152 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3277 : }
3278 :
3279 :
3280 : /*
3281 : * mergejoinscansel - Scan selectivity of merge join.
3282 : *
3283 : * A merge join will stop as soon as it exhausts either input stream.
3284 : * Therefore, if we can estimate the ranges of both input variables,
3285 : * we can estimate how much of the input will actually be read. This
3286 : * can have a considerable impact on the cost when using indexscans.
3287 : *
3288 : * Also, we can estimate how much of each input has to be read before the
3289 : * first join pair is found, which will affect the join's startup time.
3290 : *
3291 : * clause should be a clause already known to be mergejoinable. opfamily,
3292 : * cmptype, and nulls_first specify the sort ordering being used.
3293 : *
3294 : * The outputs are:
3295 : * *leftstart is set to the fraction of the left-hand variable expected
3296 : * to be scanned before the first join pair is found (0 to 1).
3297 : * *leftend is set to the fraction of the left-hand variable expected
3298 : * to be scanned before the join terminates (0 to 1).
3299 : * *rightstart, *rightend similarly for the right-hand variable.
3300 : */
3301 : void
3302 102076 : mergejoinscansel(PlannerInfo *root, Node *clause,
3303 : Oid opfamily, CompareType cmptype, bool nulls_first,
3304 : Selectivity *leftstart, Selectivity *leftend,
3305 : Selectivity *rightstart, Selectivity *rightend)
3306 : {
3307 : Node *left,
3308 : *right;
3309 : VariableStatData leftvar,
3310 : rightvar;
3311 : Oid opmethod;
3312 : int op_strategy;
3313 : Oid op_lefttype;
3314 : Oid op_righttype;
3315 : Oid opno,
3316 : collation,
3317 : lsortop,
3318 : rsortop,
3319 : lstatop,
3320 : rstatop,
3321 : ltop,
3322 : leop,
3323 : revltop,
3324 : revleop;
3325 : StrategyNumber ltstrat,
3326 : lestrat,
3327 : gtstrat,
3328 : gestrat;
3329 : bool isgt;
3330 : Datum leftmin,
3331 : leftmax,
3332 : rightmin,
3333 : rightmax;
3334 : double selec;
3335 :
3336 : /* Set default results if we can't figure anything out. */
3337 : /* XXX should default "start" fraction be a bit more than 0? */
3338 102076 : *leftstart = *rightstart = 0.0;
3339 102076 : *leftend = *rightend = 1.0;
3340 :
3341 : /* Deconstruct the merge clause */
3342 102076 : if (!is_opclause(clause))
3343 0 : return; /* shouldn't happen */
3344 102076 : opno = ((OpExpr *) clause)->opno;
3345 102076 : collation = ((OpExpr *) clause)->inputcollid;
3346 102076 : left = get_leftop((Expr *) clause);
3347 102076 : right = get_rightop((Expr *) clause);
3348 102076 : if (!right)
3349 0 : return; /* shouldn't happen */
3350 :
3351 : /* Look for stats for the inputs */
3352 102076 : examine_variable(root, left, 0, &leftvar);
3353 102076 : examine_variable(root, right, 0, &rightvar);
3354 :
3355 102076 : opmethod = get_opfamily_method(opfamily);
3356 :
3357 : /* Extract the operator's declared left/right datatypes */
3358 102076 : get_op_opfamily_properties(opno, opfamily, false,
3359 : &op_strategy,
3360 : &op_lefttype,
3361 : &op_righttype);
3362 : Assert(IndexAmTranslateStrategy(op_strategy, opmethod, opfamily, true) == COMPARE_EQ);
3363 :
3364 : /*
3365 : * Look up the various operators we need. If we don't find them all, it
3366 : * probably means the opfamily is broken, but we just fail silently.
3367 : *
3368 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3369 : * operator, regardless of which sort direction we are considering.
3370 : */
3371 102076 : switch (cmptype)
3372 : {
3373 102047 : case COMPARE_LT:
3374 102047 : isgt = false;
3375 102047 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3376 102047 : lestrat = IndexAmTranslateCompareType(COMPARE_LE, opmethod, opfamily, true);
3377 102047 : if (op_lefttype == op_righttype)
3378 : {
3379 : /* easy case */
3380 100879 : ltop = get_opfamily_member(opfamily,
3381 : op_lefttype, op_righttype,
3382 : ltstrat);
3383 100879 : leop = get_opfamily_member(opfamily,
3384 : op_lefttype, op_righttype,
3385 : lestrat);
3386 100879 : lsortop = ltop;
3387 100879 : rsortop = ltop;
3388 100879 : lstatop = lsortop;
3389 100879 : rstatop = rsortop;
3390 100879 : revltop = ltop;
3391 100879 : revleop = leop;
3392 : }
3393 : else
3394 : {
3395 1168 : ltop = get_opfamily_member(opfamily,
3396 : op_lefttype, op_righttype,
3397 : ltstrat);
3398 1168 : leop = get_opfamily_member(opfamily,
3399 : op_lefttype, op_righttype,
3400 : lestrat);
3401 1168 : lsortop = get_opfamily_member(opfamily,
3402 : op_lefttype, op_lefttype,
3403 : ltstrat);
3404 1168 : rsortop = get_opfamily_member(opfamily,
3405 : op_righttype, op_righttype,
3406 : ltstrat);
3407 1168 : lstatop = lsortop;
3408 1168 : rstatop = rsortop;
3409 1168 : revltop = get_opfamily_member(opfamily,
3410 : op_righttype, op_lefttype,
3411 : ltstrat);
3412 1168 : revleop = get_opfamily_member(opfamily,
3413 : op_righttype, op_lefttype,
3414 : lestrat);
3415 : }
3416 102047 : break;
3417 29 : case COMPARE_GT:
3418 : /* descending-order case */
3419 29 : isgt = true;
3420 29 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3421 29 : gtstrat = IndexAmTranslateCompareType(COMPARE_GT, opmethod, opfamily, true);
3422 29 : gestrat = IndexAmTranslateCompareType(COMPARE_GE, opmethod, opfamily, true);
3423 29 : if (op_lefttype == op_righttype)
3424 : {
3425 : /* easy case */
3426 29 : ltop = get_opfamily_member(opfamily,
3427 : op_lefttype, op_righttype,
3428 : gtstrat);
3429 29 : leop = get_opfamily_member(opfamily,
3430 : op_lefttype, op_righttype,
3431 : gestrat);
3432 29 : lsortop = ltop;
3433 29 : rsortop = ltop;
3434 29 : lstatop = get_opfamily_member(opfamily,
3435 : op_lefttype, op_lefttype,
3436 : ltstrat);
3437 29 : rstatop = lstatop;
3438 29 : revltop = ltop;
3439 29 : revleop = leop;
3440 : }
3441 : else
3442 : {
3443 0 : ltop = get_opfamily_member(opfamily,
3444 : op_lefttype, op_righttype,
3445 : gtstrat);
3446 0 : leop = get_opfamily_member(opfamily,
3447 : op_lefttype, op_righttype,
3448 : gestrat);
3449 0 : lsortop = get_opfamily_member(opfamily,
3450 : op_lefttype, op_lefttype,
3451 : gtstrat);
3452 0 : rsortop = get_opfamily_member(opfamily,
3453 : op_righttype, op_righttype,
3454 : gtstrat);
3455 0 : lstatop = get_opfamily_member(opfamily,
3456 : op_lefttype, op_lefttype,
3457 : ltstrat);
3458 0 : rstatop = get_opfamily_member(opfamily,
3459 : op_righttype, op_righttype,
3460 : ltstrat);
3461 0 : revltop = get_opfamily_member(opfamily,
3462 : op_righttype, op_lefttype,
3463 : gtstrat);
3464 0 : revleop = get_opfamily_member(opfamily,
3465 : op_righttype, op_lefttype,
3466 : gestrat);
3467 : }
3468 29 : break;
3469 0 : default:
3470 0 : goto fail; /* shouldn't get here */
3471 : }
3472 :
3473 102076 : if (!OidIsValid(lsortop) ||
3474 102076 : !OidIsValid(rsortop) ||
3475 102076 : !OidIsValid(lstatop) ||
3476 102076 : !OidIsValid(rstatop) ||
3477 102066 : !OidIsValid(ltop) ||
3478 102066 : !OidIsValid(leop) ||
3479 102066 : !OidIsValid(revltop) ||
3480 : !OidIsValid(revleop))
3481 10 : goto fail; /* insufficient info in catalogs */
3482 :
3483 : /* Try to get ranges of both inputs */
3484 102066 : if (!isgt)
3485 : {
3486 102037 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3487 : &leftmin, &leftmax))
3488 31684 : goto fail; /* no range available from stats */
3489 70353 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3490 : &rightmin, &rightmax))
3491 17423 : goto fail; /* no range available from stats */
3492 : }
3493 : else
3494 : {
3495 : /* need to swap the max and min */
3496 29 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3497 : &leftmax, &leftmin))
3498 24 : goto fail; /* no range available from stats */
3499 5 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3500 : &rightmax, &rightmin))
3501 0 : goto fail; /* no range available from stats */
3502 : }
3503 :
3504 : /*
3505 : * Now, the fraction of the left variable that will be scanned is the
3506 : * fraction that's <= the right-side maximum value. But only believe
3507 : * non-default estimates, else stick with our 1.0.
3508 : */
3509 52935 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3510 : rightmax, op_righttype);
3511 52935 : if (selec != DEFAULT_INEQ_SEL)
3512 52931 : *leftend = selec;
3513 :
3514 : /* And similarly for the right variable. */
3515 52935 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3516 : leftmax, op_lefttype);
3517 52935 : if (selec != DEFAULT_INEQ_SEL)
3518 52935 : *rightend = selec;
3519 :
3520 : /*
3521 : * Only one of the two "end" fractions can really be less than 1.0;
3522 : * believe the smaller estimate and reset the other one to exactly 1.0. If
3523 : * we get exactly equal estimates (as can easily happen with self-joins),
3524 : * believe neither.
3525 : */
3526 52935 : if (*leftend > *rightend)
3527 16339 : *leftend = 1.0;
3528 36596 : else if (*leftend < *rightend)
3529 20419 : *rightend = 1.0;
3530 : else
3531 16177 : *leftend = *rightend = 1.0;
3532 :
3533 : /*
3534 : * Also, the fraction of the left variable that will be scanned before the
3535 : * first join pair is found is the fraction that's < the right-side
3536 : * minimum value. But only believe non-default estimates, else stick with
3537 : * our own default.
3538 : */
3539 52935 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3540 : rightmin, op_righttype);
3541 52935 : if (selec != DEFAULT_INEQ_SEL)
3542 52935 : *leftstart = selec;
3543 :
3544 : /* And similarly for the right variable. */
3545 52935 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3546 : leftmin, op_lefttype);
3547 52935 : if (selec != DEFAULT_INEQ_SEL)
3548 52935 : *rightstart = selec;
3549 :
3550 : /*
3551 : * Only one of the two "start" fractions can really be more than zero;
3552 : * believe the larger estimate and reset the other one to exactly 0.0. If
3553 : * we get exactly equal estimates (as can easily happen with self-joins),
3554 : * believe neither.
3555 : */
3556 52935 : if (*leftstart < *rightstart)
3557 10564 : *leftstart = 0.0;
3558 42371 : else if (*leftstart > *rightstart)
3559 15175 : *rightstart = 0.0;
3560 : else
3561 27196 : *leftstart = *rightstart = 0.0;
3562 :
3563 : /*
3564 : * If the sort order is nulls-first, we're going to have to skip over any
3565 : * nulls too. These would not have been counted by scalarineqsel, and we
3566 : * can safely add in this fraction regardless of whether we believe
3567 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3568 : */
3569 52935 : if (nulls_first)
3570 : {
3571 : Form_pg_statistic stats;
3572 :
3573 5 : if (HeapTupleIsValid(leftvar.statsTuple))
3574 : {
3575 5 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3576 5 : *leftstart += stats->stanullfrac;
3577 5 : CLAMP_PROBABILITY(*leftstart);
3578 5 : *leftend += stats->stanullfrac;
3579 5 : CLAMP_PROBABILITY(*leftend);
3580 : }
3581 5 : if (HeapTupleIsValid(rightvar.statsTuple))
3582 : {
3583 5 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3584 5 : *rightstart += stats->stanullfrac;
3585 5 : CLAMP_PROBABILITY(*rightstart);
3586 5 : *rightend += stats->stanullfrac;
3587 5 : CLAMP_PROBABILITY(*rightend);
3588 : }
3589 : }
3590 :
3591 : /* Disbelieve start >= end, just in case that can happen */
3592 52935 : if (*leftstart >= *leftend)
3593 : {
3594 119 : *leftstart = 0.0;
3595 119 : *leftend = 1.0;
3596 : }
3597 52935 : if (*rightstart >= *rightend)
3598 : {
3599 798 : *rightstart = 0.0;
3600 798 : *rightend = 1.0;
3601 : }
3602 :
3603 52137 : fail:
3604 102076 : ReleaseVariableStats(leftvar);
3605 102076 : ReleaseVariableStats(rightvar);
3606 : }
3607 :
3608 :
3609 : /*
3610 : * matchingsel -- generic matching-operator selectivity support
3611 : *
3612 : * Use these for any operators that (a) are on data types for which we collect
3613 : * standard statistics, and (b) have behavior for which the default estimate
3614 : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3615 : * operators.
3616 : */
3617 :
3618 : Datum
3619 845 : matchingsel(PG_FUNCTION_ARGS)
3620 : {
3621 845 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3622 845 : Oid operator = PG_GETARG_OID(1);
3623 845 : List *args = (List *) PG_GETARG_POINTER(2);
3624 845 : int varRelid = PG_GETARG_INT32(3);
3625 845 : Oid collation = PG_GET_COLLATION();
3626 : double selec;
3627 :
3628 : /* Use generic restriction selectivity logic. */
3629 845 : selec = generic_restriction_selectivity(root, operator, collation,
3630 : args, varRelid,
3631 : DEFAULT_MATCHING_SEL);
3632 :
3633 845 : PG_RETURN_FLOAT8((float8) selec);
3634 : }
3635 :
3636 : Datum
3637 5 : matchingjoinsel(PG_FUNCTION_ARGS)
3638 : {
3639 : /* Just punt, for the moment. */
3640 5 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3641 : }
3642 :
3643 :
3644 : /*
3645 : * Helper routine for estimate_num_groups: add an item to a list of
3646 : * GroupVarInfos, but only if it's not known equal to any of the existing
3647 : * entries.
3648 : */
3649 : typedef struct
3650 : {
3651 : Node *var; /* might be an expression, not just a Var */
3652 : RelOptInfo *rel; /* relation it belongs to */
3653 : double ndistinct; /* # distinct values */
3654 : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3655 : } GroupVarInfo;
3656 :
3657 : static List *
3658 279943 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3659 : Node *var, VariableStatData *vardata)
3660 : {
3661 : GroupVarInfo *varinfo;
3662 : double ndistinct;
3663 : bool isdefault;
3664 : ListCell *lc;
3665 :
3666 279943 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3667 :
3668 : /*
3669 : * The nullingrels bits within the var could cause the same var to be
3670 : * counted multiple times if it's marked with different nullingrels. They
3671 : * could also prevent us from matching the var to the expressions in
3672 : * extended statistics (see estimate_multivariate_ndistinct). So strip
3673 : * them out first.
3674 : */
3675 279943 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3676 :
3677 345565 : foreach(lc, varinfos)
3678 : {
3679 66401 : varinfo = (GroupVarInfo *) lfirst(lc);
3680 :
3681 : /* Drop exact duplicates */
3682 66401 : if (equal(var, varinfo->var))
3683 779 : return varinfos;
3684 :
3685 : /*
3686 : * Drop known-equal vars, but only if they belong to different
3687 : * relations (see comments for estimate_num_groups). We aren't too
3688 : * fussy about the semantics of "equal" here.
3689 : */
3690 69927 : if (vardata->rel != varinfo->rel &&
3691 4111 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3692 : {
3693 210 : if (varinfo->ndistinct <= ndistinct)
3694 : {
3695 : /* Keep older item, forget new one */
3696 194 : return varinfos;
3697 : }
3698 : else
3699 : {
3700 : /* Delete the older item */
3701 16 : varinfos = foreach_delete_current(varinfos, lc);
3702 : }
3703 : }
3704 : }
3705 :
3706 279164 : varinfo = palloc_object(GroupVarInfo);
3707 :
3708 279164 : varinfo->var = var;
3709 279164 : varinfo->rel = vardata->rel;
3710 279164 : varinfo->ndistinct = ndistinct;
3711 279164 : varinfo->isdefault = isdefault;
3712 279164 : varinfos = lappend(varinfos, varinfo);
3713 279164 : return varinfos;
3714 : }
3715 :
3716 : /*
3717 : * estimate_num_groups - Estimate number of groups in a grouped query
3718 : *
3719 : * Given a query having a GROUP BY clause, estimate how many groups there
3720 : * will be --- ie, the number of distinct combinations of the GROUP BY
3721 : * expressions.
3722 : *
3723 : * This routine is also used to estimate the number of rows emitted by
3724 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3725 : * actually, we only use it for DISTINCT when there's no grouping or
3726 : * aggregation ahead of the DISTINCT.)
3727 : *
3728 : * Inputs:
3729 : * root - the query
3730 : * groupExprs - list of expressions being grouped by
3731 : * input_rows - number of rows estimated to arrive at the group/unique
3732 : * filter step
3733 : * pgset - NULL, or a List** pointing to a grouping set to filter the
3734 : * groupExprs against
3735 : *
3736 : * Outputs:
3737 : * estinfo - When passed as non-NULL, the function will set bits in the
3738 : * "flags" field in order to provide callers with additional information
3739 : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3740 : * bit if we used any default values in the estimation.
3741 : *
3742 : * Given the lack of any cross-correlation statistics in the system, it's
3743 : * impossible to do anything really trustworthy with GROUP BY conditions
3744 : * involving multiple Vars. We should however avoid assuming the worst
3745 : * case (all possible cross-product terms actually appear as groups) since
3746 : * very often the grouped-by Vars are highly correlated. Our current approach
3747 : * is as follows:
3748 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3749 : * independently of their content, and are ignored in the subsequent
3750 : * steps. This is mainly because tests like "col IS NULL" break the
3751 : * heuristic used in step 2 especially badly.
3752 : * 2. Reduce the given expressions to a list of unique Vars used. For
3753 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3754 : * It is clearly correct not to count the same Var more than once.
3755 : * It is also reasonable to treat f(x) the same as x: f() cannot
3756 : * increase the number of distinct values (unless it is volatile,
3757 : * which we consider unlikely for grouping), but it probably won't
3758 : * reduce the number of distinct values much either.
3759 : * As a special case, if a GROUP BY expression can be matched to an
3760 : * expressional index for which we have statistics, then we treat the
3761 : * whole expression as though it were just a Var.
3762 : * 3. If the list contains Vars of different relations that are known equal
3763 : * due to equivalence classes, then drop all but one of the Vars from each
3764 : * known-equal set, keeping the one with smallest estimated # of values
3765 : * (since the extra values of the others can't appear in joined rows).
3766 : * Note the reason we only consider Vars of different relations is that
3767 : * if we considered ones of the same rel, we'd be double-counting the
3768 : * restriction selectivity of the equality in the next step.
3769 : * 4. For Vars within a single source rel, we multiply together the numbers
3770 : * of values, clamp to the number of rows in the rel (divided by 10 if
3771 : * more than one Var), and then multiply by a factor based on the
3772 : * selectivity of the restriction clauses for that rel. When there's
3773 : * more than one Var, the initial product is probably too high (it's the
3774 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3775 : * helpful heuristic for not letting the estimate get out of hand. (The
3776 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3777 : * we multiply by to adjust for the restriction selectivity assumes that
3778 : * the restriction clauses are independent of the grouping, which may not
3779 : * be a valid assumption, but it's hard to do better.
3780 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3781 : * rel, and multiply the results together.
3782 : * Note that rels not containing grouped Vars are ignored completely, as are
3783 : * join clauses. Such rels cannot increase the number of groups, and we
3784 : * assume such clauses do not reduce the number either (somewhat bogus,
3785 : * but we don't have the info to do better).
3786 : */
3787 : double
3788 243045 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3789 : List **pgset, EstimationInfo *estinfo)
3790 : {
3791 243045 : List *varinfos = NIL;
3792 243045 : double srf_multiplier = 1.0;
3793 : double numdistinct;
3794 : ListCell *l;
3795 : int i;
3796 :
3797 : /* Zero the estinfo output parameter, if non-NULL */
3798 243045 : if (estinfo != NULL)
3799 202699 : memset(estinfo, 0, sizeof(EstimationInfo));
3800 :
3801 : /*
3802 : * We don't ever want to return an estimate of zero groups, as that tends
3803 : * to lead to division-by-zero and other unpleasantness. The input_rows
3804 : * estimate is usually already at least 1, but clamp it just in case it
3805 : * isn't.
3806 : */
3807 243045 : input_rows = clamp_row_est(input_rows);
3808 :
3809 : /*
3810 : * If no grouping columns, there's exactly one group. (This can't happen
3811 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3812 : * corner cases with set operations.)
3813 : */
3814 243045 : if (groupExprs == NIL || (pgset && *pgset == NIL))
3815 893 : return 1.0;
3816 :
3817 : /*
3818 : * Count groups derived from boolean grouping expressions. For other
3819 : * expressions, find the unique Vars used, treating an expression as a Var
3820 : * if we can find stats for it. For each one, record the statistical
3821 : * estimate of number of distinct values (total in its table, without
3822 : * regard for filtering).
3823 : */
3824 242152 : numdistinct = 1.0;
3825 :
3826 242152 : i = 0;
3827 521459 : foreach(l, groupExprs)
3828 : {
3829 279347 : Node *groupexpr = (Node *) lfirst(l);
3830 : double this_srf_multiplier;
3831 : VariableStatData vardata;
3832 : List *varshere;
3833 : ListCell *l2;
3834 :
3835 : /* is expression in this grouping set? */
3836 279347 : if (pgset && !list_member_int(*pgset, i++))
3837 221528 : continue;
3838 :
3839 : /*
3840 : * Set-returning functions in grouping columns are a bit problematic.
3841 : * The code below will effectively ignore their SRF nature and come up
3842 : * with a numdistinct estimate as though they were scalar functions.
3843 : * We compensate by scaling up the end result by the largest SRF
3844 : * rowcount estimate. (This will be an overestimate if the SRF
3845 : * produces multiple copies of any output value, but it seems best to
3846 : * assume the SRF's outputs are distinct. In any case, it's probably
3847 : * pointless to worry too much about this without much better
3848 : * estimates for SRF output rowcounts than we have today.)
3849 : */
3850 278679 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3851 278679 : if (srf_multiplier < this_srf_multiplier)
3852 130 : srf_multiplier = this_srf_multiplier;
3853 :
3854 : /* Short-circuit for expressions returning boolean */
3855 278679 : if (exprType(groupexpr) == BOOLOID)
3856 : {
3857 282 : numdistinct *= 2.0;
3858 282 : continue;
3859 : }
3860 :
3861 : /*
3862 : * If examine_variable is able to deduce anything about the GROUP BY
3863 : * expression, treat it as a single variable even if it's really more
3864 : * complicated.
3865 : *
3866 : * XXX This has the consequence that if there's a statistics object on
3867 : * the expression, we don't split it into individual Vars. This
3868 : * affects our selection of statistics in
3869 : * estimate_multivariate_ndistinct, because it's probably better to
3870 : * use more accurate estimate for each expression and treat them as
3871 : * independent, than to combine estimates for the extracted variables
3872 : * when we don't know how that relates to the expressions.
3873 : */
3874 278397 : examine_variable(root, groupexpr, 0, &vardata);
3875 278397 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3876 : {
3877 220080 : varinfos = add_unique_group_var(root, varinfos,
3878 : groupexpr, &vardata);
3879 220080 : ReleaseVariableStats(vardata);
3880 220080 : continue;
3881 : }
3882 58317 : ReleaseVariableStats(vardata);
3883 :
3884 : /*
3885 : * Else pull out the component Vars. Handle PlaceHolderVars by
3886 : * recursing into their arguments (effectively assuming that the
3887 : * PlaceHolderVar doesn't change the number of groups, which boils
3888 : * down to ignoring the possible addition of nulls to the result set).
3889 : */
3890 58317 : varshere = pull_var_clause(groupexpr,
3891 : PVC_RECURSE_AGGREGATES |
3892 : PVC_RECURSE_WINDOWFUNCS |
3893 : PVC_RECURSE_PLACEHOLDERS);
3894 :
3895 : /*
3896 : * If we find any variable-free GROUP BY item, then either it is a
3897 : * constant (and we can ignore it) or it contains a volatile function;
3898 : * in the latter case we punt and assume that each input row will
3899 : * yield a distinct group.
3900 : */
3901 58317 : if (varshere == NIL)
3902 : {
3903 538 : if (contain_volatile_functions(groupexpr))
3904 40 : return input_rows;
3905 498 : continue;
3906 : }
3907 :
3908 : /*
3909 : * Else add variables to varinfos list
3910 : */
3911 117642 : foreach(l2, varshere)
3912 : {
3913 59863 : Node *var = (Node *) lfirst(l2);
3914 :
3915 59863 : examine_variable(root, var, 0, &vardata);
3916 59863 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3917 59863 : ReleaseVariableStats(vardata);
3918 : }
3919 : }
3920 :
3921 : /*
3922 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3923 : * list.
3924 : */
3925 242112 : if (varinfos == NIL)
3926 : {
3927 : /* Apply SRF multiplier as we would do in the long path */
3928 375 : numdistinct *= srf_multiplier;
3929 : /* Round off */
3930 375 : numdistinct = ceil(numdistinct);
3931 : /* Guard against out-of-range answers */
3932 375 : if (numdistinct > input_rows)
3933 71 : numdistinct = input_rows;
3934 375 : if (numdistinct < 1.0)
3935 0 : numdistinct = 1.0;
3936 375 : return numdistinct;
3937 : }
3938 :
3939 : /*
3940 : * Group Vars by relation and estimate total numdistinct.
3941 : *
3942 : * For each iteration of the outer loop, we process the frontmost Var in
3943 : * varinfos, plus all other Vars in the same relation. We remove these
3944 : * Vars from the newvarinfos list for the next iteration. This is the
3945 : * easiest way to group Vars of same rel together.
3946 : */
3947 : do
3948 : {
3949 243892 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3950 243892 : RelOptInfo *rel = varinfo1->rel;
3951 243892 : double reldistinct = 1;
3952 243892 : double relmaxndistinct = reldistinct;
3953 243892 : int relvarcount = 0;
3954 243892 : List *newvarinfos = NIL;
3955 243892 : List *relvarinfos = NIL;
3956 :
3957 : /*
3958 : * Split the list of varinfos in two - one for the current rel, one
3959 : * for remaining Vars on other rels.
3960 : */
3961 243892 : relvarinfos = lappend(relvarinfos, varinfo1);
3962 282279 : for_each_from(l, varinfos, 1)
3963 : {
3964 38387 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3965 :
3966 38387 : if (varinfo2->rel == varinfo1->rel)
3967 : {
3968 : /* varinfos on current rel */
3969 35256 : relvarinfos = lappend(relvarinfos, varinfo2);
3970 : }
3971 : else
3972 : {
3973 : /* not time to process varinfo2 yet */
3974 3131 : newvarinfos = lappend(newvarinfos, varinfo2);
3975 : }
3976 : }
3977 :
3978 : /*
3979 : * Get the numdistinct estimate for the Vars of this rel. We
3980 : * iteratively search for multivariate n-distinct with maximum number
3981 : * of vars; assuming that each var group is independent of the others,
3982 : * we multiply them together. Any remaining relvarinfos after no more
3983 : * multivariate matches are found are assumed independent too, so
3984 : * their individual ndistinct estimates are multiplied also.
3985 : *
3986 : * While iterating, count how many separate numdistinct values we
3987 : * apply. We apply a fudge factor below, but only if we multiplied
3988 : * more than one such values.
3989 : */
3990 487889 : while (relvarinfos)
3991 : {
3992 : double mvndistinct;
3993 :
3994 243997 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3995 : &mvndistinct))
3996 : {
3997 345 : reldistinct *= mvndistinct;
3998 345 : if (relmaxndistinct < mvndistinct)
3999 335 : relmaxndistinct = mvndistinct;
4000 345 : relvarcount++;
4001 : }
4002 : else
4003 : {
4004 522070 : foreach(l, relvarinfos)
4005 : {
4006 278418 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
4007 :
4008 278418 : reldistinct *= varinfo2->ndistinct;
4009 278418 : if (relmaxndistinct < varinfo2->ndistinct)
4010 247421 : relmaxndistinct = varinfo2->ndistinct;
4011 278418 : relvarcount++;
4012 :
4013 : /*
4014 : * When varinfo2's isdefault is set then we'd better set
4015 : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
4016 : */
4017 278418 : if (estinfo != NULL && varinfo2->isdefault)
4018 17327 : estinfo->flags |= SELFLAG_USED_DEFAULT;
4019 : }
4020 :
4021 : /* we're done with this relation */
4022 243652 : relvarinfos = NIL;
4023 : }
4024 : }
4025 :
4026 : /*
4027 : * Sanity check --- don't divide by zero if empty relation.
4028 : */
4029 : Assert(IS_SIMPLE_REL(rel));
4030 243892 : if (rel->tuples > 0)
4031 : {
4032 : /*
4033 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
4034 : * fudge factor is because the Vars are probably correlated but we
4035 : * don't know by how much. We should never clamp to less than the
4036 : * largest ndistinct value for any of the Vars, though, since
4037 : * there will surely be at least that many groups.
4038 : */
4039 243028 : double clamp = rel->tuples;
4040 :
4041 243028 : if (relvarcount > 1)
4042 : {
4043 30797 : clamp *= 0.1;
4044 30797 : if (clamp < relmaxndistinct)
4045 : {
4046 28663 : clamp = relmaxndistinct;
4047 : /* for sanity in case some ndistinct is too large: */
4048 28663 : if (clamp > rel->tuples)
4049 65 : clamp = rel->tuples;
4050 : }
4051 : }
4052 243028 : if (reldistinct > clamp)
4053 25122 : reldistinct = clamp;
4054 :
4055 : /*
4056 : * Update the estimate based on the restriction selectivity,
4057 : * guarding against division by zero when reldistinct is zero.
4058 : * Also skip this if we know that we are returning all rows.
4059 : */
4060 243028 : if (reldistinct > 0 && rel->rows < rel->tuples)
4061 : {
4062 : /*
4063 : * Given a table containing N rows with n distinct values in a
4064 : * uniform distribution, if we select p rows at random then
4065 : * the expected number of distinct values selected is
4066 : *
4067 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
4068 : *
4069 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
4070 : *
4071 : * See "Approximating block accesses in database
4072 : * organizations", S. B. Yao, Communications of the ACM,
4073 : * Volume 20 Issue 4, April 1977 Pages 260-261.
4074 : *
4075 : * Alternatively, re-arranging the terms from the factorials,
4076 : * this may be written as
4077 : *
4078 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
4079 : *
4080 : * This form of the formula is more efficient to compute in
4081 : * the common case where p is larger than N/n. Additionally,
4082 : * as pointed out by Dell'Era, if i << N for all terms in the
4083 : * product, it can be approximated by
4084 : *
4085 : * n * (1 - ((N-p)/N)^(N/n))
4086 : *
4087 : * See "Expected distinct values when selecting from a bag
4088 : * without replacement", Alberto Dell'Era,
4089 : * http://www.adellera.it/investigations/distinct_balls/.
4090 : *
4091 : * The condition i << N is equivalent to n >> 1, so this is a
4092 : * good approximation when the number of distinct values in
4093 : * the table is large. It turns out that this formula also
4094 : * works well even when n is small.
4095 : */
4096 72784 : reldistinct *=
4097 72784 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
4098 72784 : rel->tuples / reldistinct));
4099 : }
4100 243028 : reldistinct = clamp_row_est(reldistinct);
4101 :
4102 : /*
4103 : * Update estimate of total distinct groups.
4104 : */
4105 243028 : numdistinct *= reldistinct;
4106 : }
4107 :
4108 243892 : varinfos = newvarinfos;
4109 243892 : } while (varinfos != NIL);
4110 :
4111 : /* Now we can account for the effects of any SRFs */
4112 241737 : numdistinct *= srf_multiplier;
4113 :
4114 : /* Round off */
4115 241737 : numdistinct = ceil(numdistinct);
4116 :
4117 : /* Guard against out-of-range answers */
4118 241737 : if (numdistinct > input_rows)
4119 52116 : numdistinct = input_rows;
4120 241737 : if (numdistinct < 1.0)
4121 0 : numdistinct = 1.0;
4122 :
4123 241737 : return numdistinct;
4124 : }
4125 :
4126 : /*
4127 : * Try to estimate the bucket size of the hash join inner side when the join
4128 : * condition contains two or more clauses by employing extended statistics.
4129 : *
4130 : * The main idea of this approach is that the distinct value generated by
4131 : * multivariate estimation on two or more columns would provide less bucket size
4132 : * than estimation on one separate column.
4133 : *
4134 : * IMPORTANT: It is crucial to synchronize the approach of combining different
4135 : * estimations with the caller's method.
4136 : *
4137 : * Return a list of clauses that didn't fetch any extended statistics.
4138 : */
4139 : List *
4140 344892 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
4141 : List *hashclauses,
4142 : Selectivity *innerbucketsize)
4143 : {
4144 : List *clauses;
4145 : List *otherclauses;
4146 : double ndistinct;
4147 :
4148 344892 : if (list_length(hashclauses) <= 1)
4149 : {
4150 : /*
4151 : * Nothing to do for a single clause. Could we employ univariate
4152 : * extended stat here?
4153 : */
4154 315339 : return hashclauses;
4155 : }
4156 :
4157 : /* "clauses" is the list of hashclauses we've not dealt with yet */
4158 29553 : clauses = list_copy(hashclauses);
4159 : /* "otherclauses" holds clauses we are going to return to caller */
4160 29553 : otherclauses = NIL;
4161 : /* current estimate of ndistinct */
4162 29553 : ndistinct = 1.0;
4163 59116 : while (clauses != NIL)
4164 : {
4165 : ListCell *lc;
4166 29563 : int relid = -1;
4167 29563 : List *varinfos = NIL;
4168 29563 : List *origin_rinfos = NIL;
4169 : double mvndistinct;
4170 : List *origin_varinfos;
4171 29563 : int group_relid = -1;
4172 29563 : RelOptInfo *group_rel = NULL;
4173 : ListCell *lc1,
4174 : *lc2;
4175 :
4176 : /*
4177 : * Find clauses, referencing the same single base relation and try to
4178 : * estimate such a group with extended statistics. Create varinfo for
4179 : * an approved clause, push it to otherclauses, if it can't be
4180 : * estimated here or ignore to process at the next iteration.
4181 : */
4182 90714 : foreach(lc, clauses)
4183 : {
4184 61151 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
4185 : Node *expr;
4186 : Relids relids;
4187 : GroupVarInfo *varinfo;
4188 :
4189 : /*
4190 : * Find the inner side of the join, which we need to estimate the
4191 : * number of buckets. Use outer_is_left because the
4192 : * clause_sides_match_join routine has called on hash clauses.
4193 : */
4194 122302 : relids = rinfo->outer_is_left ?
4195 61151 : rinfo->right_relids : rinfo->left_relids;
4196 122302 : expr = rinfo->outer_is_left ?
4197 61151 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
4198 :
4199 61151 : if (bms_get_singleton_member(relids, &relid) &&
4200 59531 : root->simple_rel_array[relid]->statlist != NIL)
4201 40 : {
4202 50 : bool is_duplicate = false;
4203 :
4204 : /*
4205 : * This inner-side expression references only one relation.
4206 : * Extended statistics on this clause can exist.
4207 : */
4208 50 : if (group_relid < 0)
4209 : {
4210 25 : RangeTblEntry *rte = root->simple_rte_array[relid];
4211 :
4212 25 : if (!rte || (rte->relkind != RELKIND_RELATION &&
4213 0 : rte->relkind != RELKIND_MATVIEW &&
4214 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
4215 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
4216 : {
4217 : /* Extended statistics can't exist in principle */
4218 0 : otherclauses = lappend(otherclauses, rinfo);
4219 0 : clauses = foreach_delete_current(clauses, lc);
4220 0 : continue;
4221 : }
4222 :
4223 25 : group_relid = relid;
4224 25 : group_rel = root->simple_rel_array[relid];
4225 : }
4226 25 : else if (group_relid != relid)
4227 : {
4228 : /*
4229 : * Being in the group forming state we don't need other
4230 : * clauses.
4231 : */
4232 0 : continue;
4233 : }
4234 :
4235 : /*
4236 : * We're going to add the new clause to the varinfos list. We
4237 : * might re-use add_unique_group_var(), but we don't do so for
4238 : * two reasons.
4239 : *
4240 : * 1) We must keep the origin_rinfos list ordered exactly the
4241 : * same way as varinfos.
4242 : *
4243 : * 2) add_unique_group_var() is designed for
4244 : * estimate_num_groups(), where a larger number of groups is
4245 : * worse. While estimating the number of hash buckets, we
4246 : * have the opposite: a lesser number of groups is worse.
4247 : * Therefore, we don't have to remove "known equal" vars: the
4248 : * removed var may valuably contribute to the multivariate
4249 : * statistics to grow the number of groups.
4250 : */
4251 :
4252 : /*
4253 : * Clear nullingrels to correctly match hash keys. See
4254 : * add_unique_group_var()'s comment for details.
4255 : */
4256 50 : expr = remove_nulling_relids(expr, root->outer_join_rels, NULL);
4257 :
4258 : /*
4259 : * Detect and exclude exact duplicates from the list of hash
4260 : * keys (like add_unique_group_var does).
4261 : */
4262 70 : foreach(lc1, varinfos)
4263 : {
4264 30 : varinfo = (GroupVarInfo *) lfirst(lc1);
4265 :
4266 30 : if (!equal(expr, varinfo->var))
4267 20 : continue;
4268 :
4269 10 : is_duplicate = true;
4270 10 : break;
4271 : }
4272 :
4273 50 : if (is_duplicate)
4274 : {
4275 : /*
4276 : * Skip exact duplicates. Adding them to the otherclauses
4277 : * list also doesn't make sense.
4278 : */
4279 10 : continue;
4280 : }
4281 :
4282 : /*
4283 : * Initialize GroupVarInfo. We only use it to call
4284 : * estimate_multivariate_ndistinct(), which doesn't care about
4285 : * ndistinct and isdefault fields. Thus, skip these fields.
4286 : */
4287 40 : varinfo = palloc0_object(GroupVarInfo);
4288 40 : varinfo->var = expr;
4289 40 : varinfo->rel = root->simple_rel_array[relid];
4290 40 : varinfos = lappend(varinfos, varinfo);
4291 :
4292 : /*
4293 : * Remember the link to RestrictInfo for the case the clause
4294 : * is failed to be estimated.
4295 : */
4296 40 : origin_rinfos = lappend(origin_rinfos, rinfo);
4297 : }
4298 : else
4299 : {
4300 : /* This clause can't be estimated with extended statistics */
4301 61101 : otherclauses = lappend(otherclauses, rinfo);
4302 : }
4303 :
4304 61141 : clauses = foreach_delete_current(clauses, lc);
4305 : }
4306 :
4307 29563 : if (list_length(varinfos) < 2)
4308 : {
4309 : /*
4310 : * Multivariate statistics doesn't apply to single columns except
4311 : * for expressions, but it has not been implemented yet.
4312 : */
4313 29553 : otherclauses = list_concat(otherclauses, origin_rinfos);
4314 29553 : list_free_deep(varinfos);
4315 29553 : list_free(origin_rinfos);
4316 29553 : continue;
4317 : }
4318 :
4319 : Assert(group_rel != NULL);
4320 :
4321 : /* Employ the extended statistics. */
4322 10 : origin_varinfos = varinfos;
4323 : for (;;)
4324 10 : {
4325 20 : bool estimated = estimate_multivariate_ndistinct(root,
4326 : group_rel,
4327 : &varinfos,
4328 : &mvndistinct);
4329 :
4330 20 : if (!estimated)
4331 10 : break;
4332 :
4333 : /*
4334 : * We've got an estimation. Use ndistinct value in a consistent
4335 : * way - according to the caller's logic (see
4336 : * final_cost_hashjoin).
4337 : */
4338 10 : if (ndistinct < mvndistinct)
4339 10 : ndistinct = mvndistinct;
4340 : Assert(ndistinct >= 1.0);
4341 : }
4342 :
4343 : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
4344 :
4345 : /* Collect unmatched clauses as otherclauses. */
4346 35 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
4347 : {
4348 25 : GroupVarInfo *vinfo = lfirst(lc1);
4349 :
4350 25 : if (!list_member_ptr(varinfos, vinfo))
4351 : /* Already estimated */
4352 25 : continue;
4353 :
4354 : /* Can't be estimated here - push to the returning list */
4355 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
4356 : }
4357 : }
4358 :
4359 29553 : *innerbucketsize = 1.0 / ndistinct;
4360 29553 : return otherclauses;
4361 : }
4362 :
4363 : /*
4364 : * Estimate hash bucket statistics when the specified expression is used
4365 : * as a hash key for the given number of buckets.
4366 : *
4367 : * This attempts to determine two values:
4368 : *
4369 : * 1. The frequency of the most common value of the expression (returns
4370 : * zero into *mcv_freq if we can't get that). This will be frequency
4371 : * relative to the entire underlying table.
4372 : *
4373 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
4374 : * divided by total number of tuples to be hashed.
4375 : *
4376 : * XXX This is really pretty bogus since we're effectively assuming that the
4377 : * distribution of hash keys will be the same after applying restriction
4378 : * clauses as it was in the underlying relation. However, we are not nearly
4379 : * smart enough to figure out how the restrict clauses might change the
4380 : * distribution, so this will have to do for now.
4381 : *
4382 : * We are passed the number of buckets the executor will use for the given
4383 : * input relation. If the data were perfectly distributed, with the same
4384 : * number of tuples going into each available bucket, then the bucketsize
4385 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
4386 : * only if (a) there are at least nbuckets distinct data values, and (b)
4387 : * we have a not-too-skewed data distribution. Otherwise the buckets will
4388 : * be nonuniformly occupied. If the other relation in the join has a key
4389 : * distribution similar to this one's, then the most-loaded buckets are
4390 : * exactly those that will be probed most often. Therefore, the "average"
4391 : * bucket size for costing purposes should really be taken as something close
4392 : * to the "worst case" bucket size. We try to estimate this by adjusting the
4393 : * fraction if there are too few distinct data values, and then clamping to
4394 : * at least the bucket size implied by the most common value's frequency.
4395 : *
4396 : * If no statistics are available, use a default estimate of 0.1. This will
4397 : * discourage use of a hash rather strongly if the inner relation is large,
4398 : * which is what we want. We do not want to hash unless we know that the
4399 : * inner rel is well-dispersed (or the alternatives seem much worse).
4400 : *
4401 : * The caller should also check that the mcv_freq is not so large that the
4402 : * most common value would by itself require an impractically large bucket.
4403 : * In a hash join, the executor can split buckets if they get too big, but
4404 : * obviously that doesn't help for a bucket that contains many duplicates of
4405 : * the same value.
4406 : */
4407 : void
4408 153461 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
4409 : Selectivity *mcv_freq,
4410 : Selectivity *bucketsize_frac)
4411 : {
4412 : VariableStatData vardata;
4413 : double estfract,
4414 : ndistinct;
4415 : bool isdefault;
4416 : AttStatsSlot sslot;
4417 :
4418 153461 : examine_variable(root, hashkey, 0, &vardata);
4419 :
4420 : /* Initialize *mcv_freq to "unknown" */
4421 153461 : *mcv_freq = 0.0;
4422 :
4423 : /* Look up the frequency of the most common value, if available */
4424 153461 : if (HeapTupleIsValid(vardata.statsTuple))
4425 : {
4426 100706 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4427 : STATISTIC_KIND_MCV, InvalidOid,
4428 : ATTSTATSSLOT_NUMBERS))
4429 : {
4430 : /*
4431 : * The first MCV stat is for the most common value.
4432 : */
4433 59950 : if (sslot.nnumbers > 0)
4434 59950 : *mcv_freq = sslot.numbers[0];
4435 59950 : free_attstatsslot(&sslot);
4436 : }
4437 40756 : else if (get_attstatsslot(&sslot, vardata.statsTuple,
4438 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
4439 : 0))
4440 : {
4441 : /*
4442 : * If there are no recorded MCVs, but we do have a histogram, then
4443 : * assume that ANALYZE determined that the column is unique.
4444 : */
4445 39306 : if (vardata.rel && vardata.rel->tuples > 0)
4446 39291 : *mcv_freq = 1.0 / vardata.rel->tuples;
4447 : }
4448 : }
4449 :
4450 : /* Get number of distinct values */
4451 153461 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4452 :
4453 : /*
4454 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4455 : * mcv_freq is known to be even higher than that, use it instead.
4456 : */
4457 153461 : if (isdefault)
4458 : {
4459 26805 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4460 26805 : ReleaseVariableStats(vardata);
4461 26805 : return;
4462 : }
4463 :
4464 : /*
4465 : * Adjust ndistinct to account for restriction clauses. Observe we are
4466 : * assuming that the data distribution is affected uniformly by the
4467 : * restriction clauses!
4468 : *
4469 : * XXX Possibly better way, but much more expensive: multiply by
4470 : * selectivity of rel's restriction clauses that mention the target Var.
4471 : */
4472 126656 : if (vardata.rel && vardata.rel->tuples > 0)
4473 : {
4474 126635 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
4475 126635 : ndistinct = clamp_row_est(ndistinct);
4476 : }
4477 :
4478 : /*
4479 : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
4480 : * number of buckets is less than the expected number of distinct values;
4481 : * otherwise it is 1/ndistinct.
4482 : */
4483 126656 : if (ndistinct > nbuckets)
4484 71 : estfract = 1.0 / nbuckets;
4485 : else
4486 126585 : estfract = 1.0 / ndistinct;
4487 :
4488 : /*
4489 : * Clamp the bucketsize fraction to be not less than the MCV frequency,
4490 : * since whichever bucket the MCV values end up in will have at least that
4491 : * size. This has no effect if *mcv_freq is still zero.
4492 : */
4493 126656 : estfract = Max(estfract, *mcv_freq);
4494 :
4495 126656 : *bucketsize_frac = (Selectivity) estfract;
4496 :
4497 126656 : ReleaseVariableStats(vardata);
4498 : }
4499 :
4500 : /*
4501 : * estimate_hashagg_tablesize
4502 : * estimate the number of bytes that a hash aggregate hashtable will
4503 : * require based on the agg_costs, path width and number of groups.
4504 : *
4505 : * We return the result as "double" to forestall any possible overflow
4506 : * problem in the multiplication by dNumGroups.
4507 : *
4508 : * XXX this may be over-estimating the size now that hashagg knows to omit
4509 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4510 : * grouping columns not in the hashed set are counted here even though hashagg
4511 : * won't store them. Is this a problem?
4512 : */
4513 : double
4514 2194 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4515 : const AggClauseCosts *agg_costs, double dNumGroups)
4516 : {
4517 : Size hashentrysize;
4518 :
4519 2194 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4520 2194 : path->pathtarget->width,
4521 2194 : agg_costs->transitionSpace);
4522 :
4523 : /*
4524 : * Note that this disregards the effect of fill-factor and growth policy
4525 : * of the hash table. That's probably ok, given that the default
4526 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4527 : * "double-in-size" growth policies here.
4528 : */
4529 2194 : return hashentrysize * dNumGroups;
4530 : }
4531 :
4532 :
4533 : /*-------------------------------------------------------------------------
4534 : *
4535 : * Support routines
4536 : *
4537 : *-------------------------------------------------------------------------
4538 : */
4539 :
4540 : /*
4541 : * Find the best matching ndistinct extended statistics for the given list of
4542 : * GroupVarInfos.
4543 : *
4544 : * Callers must ensure that the given GroupVarInfos all belong to 'rel' and
4545 : * the GroupVarInfos list does not contain any duplicate Vars or expressions.
4546 : *
4547 : * When statistics are found that match > 1 of the given GroupVarInfo, the
4548 : * *ndistinct parameter is set according to the ndistinct estimate and a new
4549 : * list is built with the matching GroupVarInfos removed, which is output via
4550 : * the *varinfos parameter before returning true. When no matching stats are
4551 : * found, false is returned and the *varinfos and *ndistinct parameters are
4552 : * left untouched.
4553 : */
4554 : static bool
4555 244017 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4556 : List **varinfos, double *ndistinct)
4557 : {
4558 : ListCell *lc;
4559 : int nmatches_vars;
4560 : int nmatches_exprs;
4561 244017 : Oid statOid = InvalidOid;
4562 : MVNDistinct *stats;
4563 244017 : StatisticExtInfo *matched_info = NULL;
4564 244017 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4565 :
4566 : /* bail out immediately if the table has no extended statistics */
4567 244017 : if (!rel->statlist)
4568 243546 : return false;
4569 :
4570 : /* look for the ndistinct statistics object matching the most vars */
4571 471 : nmatches_vars = 0; /* we require at least two matches */
4572 471 : nmatches_exprs = 0;
4573 1875 : foreach(lc, rel->statlist)
4574 : {
4575 : ListCell *lc2;
4576 1404 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
4577 1404 : int nshared_vars = 0;
4578 1404 : int nshared_exprs = 0;
4579 :
4580 : /* skip statistics of other kinds */
4581 1404 : if (info->kind != STATS_EXT_NDISTINCT)
4582 663 : continue;
4583 :
4584 : /* skip statistics with mismatching stxdinherit value */
4585 741 : if (info->inherit != rte->inh)
4586 25 : continue;
4587 :
4588 : /*
4589 : * Determine how many expressions (and variables in non-matched
4590 : * expressions) match. We'll then use these numbers to pick the
4591 : * statistics object that best matches the clauses.
4592 : */
4593 2267 : foreach(lc2, *varinfos)
4594 : {
4595 : ListCell *lc3;
4596 1551 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4597 : AttrNumber attnum;
4598 :
4599 : Assert(varinfo->rel == rel);
4600 :
4601 : /* simple Var, search in statistics keys directly */
4602 1551 : if (IsA(varinfo->var, Var))
4603 : {
4604 1246 : attnum = ((Var *) varinfo->var)->varattno;
4605 :
4606 : /*
4607 : * Ignore system attributes - we don't support statistics on
4608 : * them, so can't match them (and it'd fail as the values are
4609 : * negative).
4610 : */
4611 1246 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4612 10 : continue;
4613 :
4614 1236 : if (bms_is_member(attnum, info->keys))
4615 730 : nshared_vars++;
4616 :
4617 1236 : continue;
4618 : }
4619 :
4620 : /* expression - see if it's in the statistics object */
4621 550 : foreach(lc3, info->exprs)
4622 : {
4623 440 : Node *expr = (Node *) lfirst(lc3);
4624 :
4625 440 : if (equal(varinfo->var, expr))
4626 : {
4627 195 : nshared_exprs++;
4628 195 : break;
4629 : }
4630 : }
4631 : }
4632 :
4633 : /*
4634 : * The ndistinct extended statistics contain estimates for a minimum
4635 : * of pairs of columns which the statistics are defined on and
4636 : * certainly not single columns. Here we skip unless we managed to
4637 : * match to at least two columns.
4638 : */
4639 716 : if (nshared_vars + nshared_exprs < 2)
4640 331 : continue;
4641 :
4642 : /*
4643 : * Check if these statistics are a better match than the previous best
4644 : * match and if so, take note of the StatisticExtInfo.
4645 : *
4646 : * The statslist is sorted by statOid, so the StatisticExtInfo we
4647 : * select as the best match is deterministic even when multiple sets
4648 : * of statistics match equally as well.
4649 : */
4650 385 : if ((nshared_exprs > nmatches_exprs) ||
4651 295 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4652 : {
4653 365 : statOid = info->statOid;
4654 365 : nmatches_vars = nshared_vars;
4655 365 : nmatches_exprs = nshared_exprs;
4656 365 : matched_info = info;
4657 : }
4658 : }
4659 :
4660 : /* No match? */
4661 471 : if (statOid == InvalidOid)
4662 116 : return false;
4663 :
4664 : Assert(nmatches_vars + nmatches_exprs > 1);
4665 :
4666 355 : stats = statext_ndistinct_load(statOid, rte->inh);
4667 :
4668 : /*
4669 : * If we have a match, search it for the specific item that matches (there
4670 : * must be one), and construct the output values.
4671 : */
4672 355 : if (stats)
4673 : {
4674 : int i;
4675 355 : List *newlist = NIL;
4676 355 : MVNDistinctItem *item = NULL;
4677 : ListCell *lc2;
4678 355 : Bitmapset *matched = NULL;
4679 : AttrNumber attnum_offset;
4680 :
4681 : /*
4682 : * How much we need to offset the attnums? If there are no
4683 : * expressions, no offset is needed. Otherwise offset enough to move
4684 : * the lowest one (which is equal to number of expressions) to 1.
4685 : */
4686 355 : if (matched_info->exprs)
4687 125 : attnum_offset = (list_length(matched_info->exprs) + 1);
4688 : else
4689 230 : attnum_offset = 0;
4690 :
4691 : /* see what actually matched */
4692 1240 : foreach(lc2, *varinfos)
4693 : {
4694 : ListCell *lc3;
4695 : int idx;
4696 885 : bool found = false;
4697 :
4698 885 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4699 :
4700 : /*
4701 : * Process a simple Var expression, by matching it to keys
4702 : * directly. If there's a matching expression, we'll try matching
4703 : * it later.
4704 : */
4705 885 : if (IsA(varinfo->var, Var))
4706 : {
4707 730 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4708 :
4709 : /*
4710 : * Ignore expressions on system attributes. Can't rely on the
4711 : * bms check for negative values.
4712 : */
4713 730 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4714 5 : continue;
4715 :
4716 : /* Is the variable covered by the statistics object? */
4717 725 : if (!bms_is_member(attnum, matched_info->keys))
4718 100 : continue;
4719 :
4720 625 : attnum = attnum + attnum_offset;
4721 :
4722 : /* ensure sufficient offset */
4723 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4724 :
4725 625 : matched = bms_add_member(matched, attnum);
4726 :
4727 625 : found = true;
4728 : }
4729 :
4730 : /*
4731 : * XXX Maybe we should allow searching the expressions even if we
4732 : * found an attribute matching the expression? That would handle
4733 : * trivial expressions like "(a)" but it seems fairly useless.
4734 : */
4735 780 : if (found)
4736 625 : continue;
4737 :
4738 : /* expression - see if it's in the statistics object */
4739 155 : idx = 0;
4740 255 : foreach(lc3, matched_info->exprs)
4741 : {
4742 230 : Node *expr = (Node *) lfirst(lc3);
4743 :
4744 230 : if (equal(varinfo->var, expr))
4745 : {
4746 130 : AttrNumber attnum = -(idx + 1);
4747 :
4748 130 : attnum = attnum + attnum_offset;
4749 :
4750 : /* ensure sufficient offset */
4751 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4752 :
4753 130 : matched = bms_add_member(matched, attnum);
4754 :
4755 : /* there should be just one matching expression */
4756 130 : break;
4757 : }
4758 :
4759 100 : idx++;
4760 : }
4761 : }
4762 :
4763 : /* Find the specific item that exactly matches the combination */
4764 720 : for (i = 0; i < stats->nitems; i++)
4765 : {
4766 : int j;
4767 720 : MVNDistinctItem *tmpitem = &stats->items[i];
4768 :
4769 720 : if (tmpitem->nattributes != bms_num_members(matched))
4770 135 : continue;
4771 :
4772 : /* assume it's the right item */
4773 585 : item = tmpitem;
4774 :
4775 : /* check that all item attributes/expressions fit the match */
4776 1410 : for (j = 0; j < tmpitem->nattributes; j++)
4777 : {
4778 1055 : AttrNumber attnum = tmpitem->attributes[j];
4779 :
4780 : /*
4781 : * Thanks to how we constructed the matched bitmap above, we
4782 : * can just offset all attnums the same way.
4783 : */
4784 1055 : attnum = attnum + attnum_offset;
4785 :
4786 1055 : if (!bms_is_member(attnum, matched))
4787 : {
4788 : /* nah, it's not this item */
4789 230 : item = NULL;
4790 230 : break;
4791 : }
4792 : }
4793 :
4794 : /*
4795 : * If the item has all the matched attributes, we know it's the
4796 : * right one - there can't be a better one. matching more.
4797 : */
4798 585 : if (item)
4799 355 : break;
4800 : }
4801 :
4802 : /*
4803 : * Make sure we found an item. There has to be one, because ndistinct
4804 : * statistics includes all combinations of attributes.
4805 : */
4806 355 : if (!item)
4807 0 : elog(ERROR, "corrupt MVNDistinct entry");
4808 :
4809 : /* Form the output varinfo list, keeping only unmatched ones */
4810 1240 : foreach(lc, *varinfos)
4811 : {
4812 885 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4813 : ListCell *lc3;
4814 885 : bool found = false;
4815 :
4816 : /*
4817 : * Let's look at plain variables first, because it's the most
4818 : * common case and the check is quite cheap. We can simply get the
4819 : * attnum and check (with an offset) matched bitmap.
4820 : */
4821 885 : if (IsA(varinfo->var, Var))
4822 725 : {
4823 730 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4824 :
4825 : /*
4826 : * If it's a system attribute, we're done. We don't support
4827 : * extended statistics on system attributes, so it's clearly
4828 : * not matched. Just keep the expression and continue.
4829 : */
4830 730 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4831 : {
4832 5 : newlist = lappend(newlist, varinfo);
4833 5 : continue;
4834 : }
4835 :
4836 : /* apply the same offset as above */
4837 725 : attnum += attnum_offset;
4838 :
4839 : /* if it's not matched, keep the varinfo */
4840 725 : if (!bms_is_member(attnum, matched))
4841 100 : newlist = lappend(newlist, varinfo);
4842 :
4843 : /* The rest of the loop deals with complex expressions. */
4844 725 : continue;
4845 : }
4846 :
4847 : /*
4848 : * Process complex expressions, not just simple Vars.
4849 : *
4850 : * First, we search for an exact match of an expression. If we
4851 : * find one, we can just discard the whole GroupVarInfo, with all
4852 : * the variables we extracted from it.
4853 : *
4854 : * Otherwise we inspect the individual vars, and try matching it
4855 : * to variables in the item.
4856 : */
4857 255 : foreach(lc3, matched_info->exprs)
4858 : {
4859 230 : Node *expr = (Node *) lfirst(lc3);
4860 :
4861 230 : if (equal(varinfo->var, expr))
4862 : {
4863 130 : found = true;
4864 130 : break;
4865 : }
4866 : }
4867 :
4868 : /* found exact match, skip */
4869 155 : if (found)
4870 130 : continue;
4871 :
4872 25 : newlist = lappend(newlist, varinfo);
4873 : }
4874 :
4875 355 : *varinfos = newlist;
4876 355 : *ndistinct = item->ndistinct;
4877 355 : return true;
4878 : }
4879 :
4880 0 : return false;
4881 : }
4882 :
4883 : /*
4884 : * convert_to_scalar
4885 : * Convert non-NULL values of the indicated types to the comparison
4886 : * scale needed by scalarineqsel().
4887 : * Returns "true" if successful.
4888 : *
4889 : * XXX this routine is a hack: ideally we should look up the conversion
4890 : * subroutines in pg_type.
4891 : *
4892 : * All numeric datatypes are simply converted to their equivalent
4893 : * "double" values. (NUMERIC values that are outside the range of "double"
4894 : * are clamped to +/- HUGE_VAL.)
4895 : *
4896 : * String datatypes are converted by convert_string_to_scalar(),
4897 : * which is explained below. The reason why this routine deals with
4898 : * three values at a time, not just one, is that we need it for strings.
4899 : *
4900 : * The bytea datatype is just enough different from strings that it has
4901 : * to be treated separately.
4902 : *
4903 : * The several datatypes representing absolute times are all converted
4904 : * to Timestamp, which is actually an int64, and then we promote that to
4905 : * a double. Note this will give correct results even for the "special"
4906 : * values of Timestamp, since those are chosen to compare correctly;
4907 : * see timestamp_cmp.
4908 : *
4909 : * The several datatypes representing relative times (intervals) are all
4910 : * converted to measurements expressed in seconds.
4911 : */
4912 : static bool
4913 57321 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4914 : Datum lobound, Datum hibound, Oid boundstypid,
4915 : double *scaledlobound, double *scaledhibound)
4916 : {
4917 57321 : bool failure = false;
4918 :
4919 : /*
4920 : * Both the valuetypid and the boundstypid should exactly match the
4921 : * declared input type(s) of the operator we are invoked for. However,
4922 : * extensions might try to use scalarineqsel as estimator for operators
4923 : * with input type(s) we don't handle here; in such cases, we want to
4924 : * return false, not fail. In any case, we mustn't assume that valuetypid
4925 : * and boundstypid are identical.
4926 : *
4927 : * XXX The histogram we are interpolating between points of could belong
4928 : * to a column that's only binary-compatible with the declared type. In
4929 : * essence we are assuming that the semantics of binary-compatible types
4930 : * are enough alike that we can use a histogram generated with one type's
4931 : * operators to estimate selectivity for the other's. This is outright
4932 : * wrong in some cases --- in particular signed versus unsigned
4933 : * interpretation could trip us up. But it's useful enough in the
4934 : * majority of cases that we do it anyway. Should think about more
4935 : * rigorous ways to do it.
4936 : */
4937 57321 : switch (valuetypid)
4938 : {
4939 : /*
4940 : * Built-in numeric types
4941 : */
4942 52524 : case BOOLOID:
4943 : case INT2OID:
4944 : case INT4OID:
4945 : case INT8OID:
4946 : case FLOAT4OID:
4947 : case FLOAT8OID:
4948 : case NUMERICOID:
4949 : case OIDOID:
4950 : case REGPROCOID:
4951 : case REGPROCEDUREOID:
4952 : case REGOPEROID:
4953 : case REGOPERATOROID:
4954 : case REGCLASSOID:
4955 : case REGTYPEOID:
4956 : case REGCOLLATIONOID:
4957 : case REGCONFIGOID:
4958 : case REGDICTIONARYOID:
4959 : case REGROLEOID:
4960 : case REGNAMESPACEOID:
4961 : case REGDATABASEOID:
4962 52524 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4963 : &failure);
4964 52524 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4965 : &failure);
4966 52524 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4967 : &failure);
4968 52524 : return !failure;
4969 :
4970 : /*
4971 : * Built-in string types
4972 : */
4973 4797 : case CHAROID:
4974 : case BPCHAROID:
4975 : case VARCHAROID:
4976 : case TEXTOID:
4977 : case NAMEOID:
4978 : {
4979 4797 : char *valstr = convert_string_datum(value, valuetypid,
4980 : collid, &failure);
4981 4797 : char *lostr = convert_string_datum(lobound, boundstypid,
4982 : collid, &failure);
4983 4797 : char *histr = convert_string_datum(hibound, boundstypid,
4984 : collid, &failure);
4985 :
4986 : /*
4987 : * Bail out if any of the values is not of string type. We
4988 : * might leak converted strings for the other value(s), but
4989 : * that's not worth troubling over.
4990 : */
4991 4797 : if (failure)
4992 0 : return false;
4993 :
4994 4797 : convert_string_to_scalar(valstr, scaledvalue,
4995 : lostr, scaledlobound,
4996 : histr, scaledhibound);
4997 4797 : pfree(valstr);
4998 4797 : pfree(lostr);
4999 4797 : pfree(histr);
5000 4797 : return true;
5001 : }
5002 :
5003 : /*
5004 : * Built-in bytea type
5005 : */
5006 0 : case BYTEAOID:
5007 : {
5008 : /* We only support bytea vs bytea comparison */
5009 0 : if (boundstypid != BYTEAOID)
5010 0 : return false;
5011 0 : convert_bytea_to_scalar(value, scaledvalue,
5012 : lobound, scaledlobound,
5013 : hibound, scaledhibound);
5014 0 : return true;
5015 : }
5016 :
5017 : /*
5018 : * Built-in time types
5019 : */
5020 0 : case TIMESTAMPOID:
5021 : case TIMESTAMPTZOID:
5022 : case DATEOID:
5023 : case INTERVALOID:
5024 : case TIMEOID:
5025 : case TIMETZOID:
5026 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
5027 : &failure);
5028 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
5029 : &failure);
5030 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
5031 : &failure);
5032 0 : return !failure;
5033 :
5034 : /*
5035 : * Built-in network types
5036 : */
5037 0 : case INETOID:
5038 : case CIDROID:
5039 : case MACADDROID:
5040 : case MACADDR8OID:
5041 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
5042 : &failure);
5043 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
5044 : &failure);
5045 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
5046 : &failure);
5047 0 : return !failure;
5048 : }
5049 : /* Don't know how to convert */
5050 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
5051 0 : return false;
5052 : }
5053 :
5054 : /*
5055 : * Do convert_to_scalar()'s work for any numeric data type.
5056 : *
5057 : * On failure (e.g., unsupported typid), set *failure to true;
5058 : * otherwise, that variable is not changed.
5059 : */
5060 : static double
5061 157572 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
5062 : {
5063 157572 : switch (typid)
5064 : {
5065 0 : case BOOLOID:
5066 0 : return (double) DatumGetBool(value);
5067 10 : case INT2OID:
5068 10 : return (double) DatumGetInt16(value);
5069 21998 : case INT4OID:
5070 21998 : return (double) DatumGetInt32(value);
5071 0 : case INT8OID:
5072 0 : return (double) DatumGetInt64(value);
5073 0 : case FLOAT4OID:
5074 0 : return (double) DatumGetFloat4(value);
5075 45 : case FLOAT8OID:
5076 45 : return (double) DatumGetFloat8(value);
5077 0 : case NUMERICOID:
5078 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
5079 0 : return (double)
5080 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
5081 : value));
5082 135519 : case OIDOID:
5083 : case REGPROCOID:
5084 : case REGPROCEDUREOID:
5085 : case REGOPEROID:
5086 : case REGOPERATOROID:
5087 : case REGCLASSOID:
5088 : case REGTYPEOID:
5089 : case REGCOLLATIONOID:
5090 : case REGCONFIGOID:
5091 : case REGDICTIONARYOID:
5092 : case REGROLEOID:
5093 : case REGNAMESPACEOID:
5094 : case REGDATABASEOID:
5095 : /* we can treat OIDs as integers... */
5096 135519 : return (double) DatumGetObjectId(value);
5097 : }
5098 :
5099 0 : *failure = true;
5100 0 : return 0;
5101 : }
5102 :
5103 : /*
5104 : * Do convert_to_scalar()'s work for any character-string data type.
5105 : *
5106 : * String datatypes are converted to a scale that ranges from 0 to 1,
5107 : * where we visualize the bytes of the string as fractional digits.
5108 : *
5109 : * We do not want the base to be 256, however, since that tends to
5110 : * generate inflated selectivity estimates; few databases will have
5111 : * occurrences of all 256 possible byte values at each position.
5112 : * Instead, use the smallest and largest byte values seen in the bounds
5113 : * as the estimated range for each byte, after some fudging to deal with
5114 : * the fact that we probably aren't going to see the full range that way.
5115 : *
5116 : * An additional refinement is that we discard any common prefix of the
5117 : * three strings before computing the scaled values. This allows us to
5118 : * "zoom in" when we encounter a narrow data range. An example is a phone
5119 : * number database where all the values begin with the same area code.
5120 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
5121 : * so this is more likely to happen than you might think.)
5122 : */
5123 : static void
5124 4797 : convert_string_to_scalar(char *value,
5125 : double *scaledvalue,
5126 : char *lobound,
5127 : double *scaledlobound,
5128 : char *hibound,
5129 : double *scaledhibound)
5130 : {
5131 : int rangelo,
5132 : rangehi;
5133 : char *sptr;
5134 :
5135 4797 : rangelo = rangehi = (unsigned char) hibound[0];
5136 65125 : for (sptr = lobound; *sptr; sptr++)
5137 : {
5138 60328 : if (rangelo > (unsigned char) *sptr)
5139 11767 : rangelo = (unsigned char) *sptr;
5140 60328 : if (rangehi < (unsigned char) *sptr)
5141 5720 : rangehi = (unsigned char) *sptr;
5142 : }
5143 52433 : for (sptr = hibound; *sptr; sptr++)
5144 : {
5145 47636 : if (rangelo > (unsigned char) *sptr)
5146 624 : rangelo = (unsigned char) *sptr;
5147 47636 : if (rangehi < (unsigned char) *sptr)
5148 1875 : rangehi = (unsigned char) *sptr;
5149 : }
5150 : /* If range includes any upper-case ASCII chars, make it include all */
5151 4797 : if (rangelo <= 'Z' && rangehi >= 'A')
5152 : {
5153 1163 : if (rangelo > 'A')
5154 185 : rangelo = 'A';
5155 1163 : if (rangehi < 'Z')
5156 425 : rangehi = 'Z';
5157 : }
5158 : /* Ditto lower-case */
5159 4797 : if (rangelo <= 'z' && rangehi >= 'a')
5160 : {
5161 4357 : if (rangelo > 'a')
5162 5 : rangelo = 'a';
5163 4357 : if (rangehi < 'z')
5164 4309 : rangehi = 'z';
5165 : }
5166 : /* Ditto digits */
5167 4797 : if (rangelo <= '9' && rangehi >= '0')
5168 : {
5169 568 : if (rangelo > '0')
5170 495 : rangelo = '0';
5171 568 : if (rangehi < '9')
5172 9 : rangehi = '9';
5173 : }
5174 :
5175 : /*
5176 : * If range includes less than 10 chars, assume we have not got enough
5177 : * data, and make it include regular ASCII set.
5178 : */
5179 4797 : if (rangehi - rangelo < 9)
5180 : {
5181 0 : rangelo = ' ';
5182 0 : rangehi = 127;
5183 : }
5184 :
5185 : /*
5186 : * Now strip any common prefix of the three strings.
5187 : */
5188 9308 : while (*lobound)
5189 : {
5190 9308 : if (*lobound != *hibound || *lobound != *value)
5191 : break;
5192 4511 : lobound++, hibound++, value++;
5193 : }
5194 :
5195 : /*
5196 : * Now we can do the conversions.
5197 : */
5198 4797 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
5199 4797 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
5200 4797 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
5201 4797 : }
5202 :
5203 : static double
5204 14391 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
5205 : {
5206 14391 : int slen = strlen(value);
5207 : double num,
5208 : denom,
5209 : base;
5210 :
5211 14391 : if (slen <= 0)
5212 0 : return 0.0; /* empty string has scalar value 0 */
5213 :
5214 : /*
5215 : * There seems little point in considering more than a dozen bytes from
5216 : * the string. Since base is at least 10, that will give us nominal
5217 : * resolution of at least 12 decimal digits, which is surely far more
5218 : * precision than this estimation technique has got anyway (especially in
5219 : * non-C locales). Also, even with the maximum possible base of 256, this
5220 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
5221 : * overflow on any known machine.
5222 : */
5223 14391 : if (slen > 12)
5224 3804 : slen = 12;
5225 :
5226 : /* Convert initial characters to fraction */
5227 14391 : base = rangehi - rangelo + 1;
5228 14391 : num = 0.0;
5229 14391 : denom = base;
5230 121983 : while (slen-- > 0)
5231 : {
5232 107592 : int ch = (unsigned char) *value++;
5233 :
5234 107592 : if (ch < rangelo)
5235 132 : ch = rangelo - 1;
5236 107460 : else if (ch > rangehi)
5237 0 : ch = rangehi + 1;
5238 107592 : num += ((double) (ch - rangelo)) / denom;
5239 107592 : denom *= base;
5240 : }
5241 :
5242 14391 : return num;
5243 : }
5244 :
5245 : /*
5246 : * Convert a string-type Datum into a palloc'd, null-terminated string.
5247 : *
5248 : * On failure (e.g., unsupported typid), set *failure to true;
5249 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
5250 : *
5251 : * When using a non-C locale, we must pass the string through pg_strxfrm()
5252 : * before continuing, so as to generate correct locale-specific results.
5253 : */
5254 : static char *
5255 14391 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
5256 : {
5257 : char *val;
5258 : pg_locale_t mylocale;
5259 :
5260 14391 : switch (typid)
5261 : {
5262 0 : case CHAROID:
5263 0 : val = (char *) palloc(2);
5264 0 : val[0] = DatumGetChar(value);
5265 0 : val[1] = '\0';
5266 0 : break;
5267 4328 : case BPCHAROID:
5268 : case VARCHAROID:
5269 : case TEXTOID:
5270 4328 : val = TextDatumGetCString(value);
5271 4328 : break;
5272 10063 : case NAMEOID:
5273 : {
5274 10063 : NameData *nm = (NameData *) DatumGetPointer(value);
5275 :
5276 10063 : val = pstrdup(NameStr(*nm));
5277 10063 : break;
5278 : }
5279 0 : default:
5280 0 : *failure = true;
5281 0 : return NULL;
5282 : }
5283 :
5284 14391 : mylocale = pg_newlocale_from_collation(collid);
5285 :
5286 14391 : if (!mylocale->collate_is_c)
5287 : {
5288 : char *xfrmstr;
5289 : size_t xfrmlen;
5290 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
5291 :
5292 : /*
5293 : * XXX: We could guess at a suitable output buffer size and only call
5294 : * pg_strxfrm() twice if our guess is too small.
5295 : *
5296 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
5297 : * bogus data or set an error. This is not really a problem unless it
5298 : * crashes since it will only give an estimation error and nothing
5299 : * fatal.
5300 : *
5301 : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
5302 : * some cases, libc strxfrm() may return the wrong results, but that
5303 : * will only lead to an estimation error.
5304 : */
5305 66 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
5306 : #ifdef WIN32
5307 :
5308 : /*
5309 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
5310 : * of trying to allocate this much memory (and fail), just return the
5311 : * original string unmodified as if we were in the C locale.
5312 : */
5313 : if (xfrmlen == INT_MAX)
5314 : return val;
5315 : #endif
5316 66 : xfrmstr = (char *) palloc(xfrmlen + 1);
5317 66 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
5318 :
5319 : /*
5320 : * Some systems (e.g., glibc) can return a smaller value from the
5321 : * second call than the first; thus the Assert must be <= not ==.
5322 : */
5323 : Assert(xfrmlen2 <= xfrmlen);
5324 66 : pfree(val);
5325 66 : val = xfrmstr;
5326 : }
5327 :
5328 14391 : return val;
5329 : }
5330 :
5331 : /*
5332 : * Do convert_to_scalar()'s work for any bytea data type.
5333 : *
5334 : * Very similar to convert_string_to_scalar except we can't assume
5335 : * null-termination and therefore pass explicit lengths around.
5336 : *
5337 : * Also, assumptions about likely "normal" ranges of characters have been
5338 : * removed - a data range of 0..255 is always used, for now. (Perhaps
5339 : * someday we will add information about actual byte data range to
5340 : * pg_statistic.)
5341 : */
5342 : static void
5343 0 : convert_bytea_to_scalar(Datum value,
5344 : double *scaledvalue,
5345 : Datum lobound,
5346 : double *scaledlobound,
5347 : Datum hibound,
5348 : double *scaledhibound)
5349 : {
5350 0 : bytea *valuep = DatumGetByteaPP(value);
5351 0 : bytea *loboundp = DatumGetByteaPP(lobound);
5352 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
5353 : int rangelo,
5354 : rangehi,
5355 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
5356 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
5357 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
5358 : i,
5359 : minlen;
5360 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
5361 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
5362 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
5363 :
5364 : /*
5365 : * Assume bytea data is uniformly distributed across all byte values.
5366 : */
5367 0 : rangelo = 0;
5368 0 : rangehi = 255;
5369 :
5370 : /*
5371 : * Now strip any common prefix of the three strings.
5372 : */
5373 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
5374 0 : for (i = 0; i < minlen; i++)
5375 : {
5376 0 : if (*lostr != *histr || *lostr != *valstr)
5377 : break;
5378 0 : lostr++, histr++, valstr++;
5379 0 : loboundlen--, hiboundlen--, valuelen--;
5380 : }
5381 :
5382 : /*
5383 : * Now we can do the conversions.
5384 : */
5385 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
5386 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
5387 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
5388 0 : }
5389 :
5390 : static double
5391 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
5392 : int rangelo, int rangehi)
5393 : {
5394 : double num,
5395 : denom,
5396 : base;
5397 :
5398 0 : if (valuelen <= 0)
5399 0 : return 0.0; /* empty string has scalar value 0 */
5400 :
5401 : /*
5402 : * Since base is 256, need not consider more than about 10 chars (even
5403 : * this many seems like overkill)
5404 : */
5405 0 : if (valuelen > 10)
5406 0 : valuelen = 10;
5407 :
5408 : /* Convert initial characters to fraction */
5409 0 : base = rangehi - rangelo + 1;
5410 0 : num = 0.0;
5411 0 : denom = base;
5412 0 : while (valuelen-- > 0)
5413 : {
5414 0 : int ch = *value++;
5415 :
5416 0 : if (ch < rangelo)
5417 0 : ch = rangelo - 1;
5418 0 : else if (ch > rangehi)
5419 0 : ch = rangehi + 1;
5420 0 : num += ((double) (ch - rangelo)) / denom;
5421 0 : denom *= base;
5422 : }
5423 :
5424 0 : return num;
5425 : }
5426 :
5427 : /*
5428 : * Do convert_to_scalar()'s work for any timevalue data type.
5429 : *
5430 : * On failure (e.g., unsupported typid), set *failure to true;
5431 : * otherwise, that variable is not changed.
5432 : */
5433 : static double
5434 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5435 : {
5436 0 : switch (typid)
5437 : {
5438 0 : case TIMESTAMPOID:
5439 0 : return DatumGetTimestamp(value);
5440 0 : case TIMESTAMPTZOID:
5441 0 : return DatumGetTimestampTz(value);
5442 0 : case DATEOID:
5443 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
5444 0 : case INTERVALOID:
5445 : {
5446 0 : Interval *interval = DatumGetIntervalP(value);
5447 :
5448 : /*
5449 : * Convert the month part of Interval to days using assumed
5450 : * average month length of 365.25/12.0 days. Not too
5451 : * accurate, but plenty good enough for our purposes.
5452 : *
5453 : * This also works for infinite intervals, which just have all
5454 : * fields set to INT_MIN/INT_MAX, and so will produce a result
5455 : * smaller/larger than any finite interval.
5456 : */
5457 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5458 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5459 : }
5460 0 : case TIMEOID:
5461 0 : return DatumGetTimeADT(value);
5462 0 : case TIMETZOID:
5463 : {
5464 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5465 :
5466 : /* use GMT-equivalent time */
5467 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5468 : }
5469 : }
5470 :
5471 0 : *failure = true;
5472 0 : return 0;
5473 : }
5474 :
5475 :
5476 : /*
5477 : * get_restriction_variable
5478 : * Examine the args of a restriction clause to see if it's of the
5479 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5480 : * where "variable" could be either a Var or an expression in vars of a
5481 : * single relation. If so, extract information about the variable,
5482 : * and also indicate which side it was on and the other argument.
5483 : *
5484 : * Inputs:
5485 : * root: the planner info
5486 : * args: clause argument list
5487 : * varRelid: see specs for restriction selectivity functions
5488 : *
5489 : * Outputs: (these are valid only if true is returned)
5490 : * *vardata: gets information about variable (see examine_variable)
5491 : * *other: gets other clause argument, aggressively reduced to a constant
5492 : * *varonleft: set true if variable is on the left, false if on the right
5493 : *
5494 : * Returns true if a variable is identified, otherwise false.
5495 : *
5496 : * Note: if there are Vars on both sides of the clause, we must fail, because
5497 : * callers are expecting that the other side will act like a pseudoconstant.
5498 : */
5499 : bool
5500 647306 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5501 : VariableStatData *vardata, Node **other,
5502 : bool *varonleft)
5503 : {
5504 : Node *left,
5505 : *right;
5506 : VariableStatData rdata;
5507 :
5508 : /* Fail if not a binary opclause (probably shouldn't happen) */
5509 647306 : if (list_length(args) != 2)
5510 0 : return false;
5511 :
5512 647306 : left = (Node *) linitial(args);
5513 647306 : right = (Node *) lsecond(args);
5514 :
5515 : /*
5516 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5517 : * relations will be treated as pseudoconstants.
5518 : */
5519 647306 : examine_variable(root, left, varRelid, vardata);
5520 647306 : examine_variable(root, right, varRelid, &rdata);
5521 :
5522 : /*
5523 : * If one side is a variable and the other not, we win.
5524 : */
5525 647306 : if (vardata->rel && rdata.rel == NULL)
5526 : {
5527 576413 : *varonleft = true;
5528 576413 : *other = estimate_expression_value(root, rdata.var);
5529 : /* Assume we need no ReleaseVariableStats(rdata) here */
5530 576409 : return true;
5531 : }
5532 :
5533 70893 : if (vardata->rel == NULL && rdata.rel)
5534 : {
5535 66575 : *varonleft = false;
5536 66575 : *other = estimate_expression_value(root, vardata->var);
5537 : /* Assume we need no ReleaseVariableStats(*vardata) here */
5538 66575 : *vardata = rdata;
5539 66575 : return true;
5540 : }
5541 :
5542 : /* Oops, clause has wrong structure (probably var op var) */
5543 4318 : ReleaseVariableStats(*vardata);
5544 4318 : ReleaseVariableStats(rdata);
5545 :
5546 4318 : return false;
5547 : }
5548 :
5549 : /*
5550 : * get_join_variables
5551 : * Apply examine_variable() to each side of a join clause.
5552 : * Also, attempt to identify whether the join clause has the same
5553 : * or reversed sense compared to the SpecialJoinInfo.
5554 : *
5555 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5556 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5557 : * where we can't tell for sure, we default to assuming it's normal.
5558 : */
5559 : void
5560 217881 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5561 : VariableStatData *vardata1, VariableStatData *vardata2,
5562 : bool *join_is_reversed)
5563 : {
5564 : Node *left,
5565 : *right;
5566 :
5567 217881 : if (list_length(args) != 2)
5568 0 : elog(ERROR, "join operator should take two arguments");
5569 :
5570 217881 : left = (Node *) linitial(args);
5571 217881 : right = (Node *) lsecond(args);
5572 :
5573 217881 : examine_variable(root, left, 0, vardata1);
5574 217881 : examine_variable(root, right, 0, vardata2);
5575 :
5576 435571 : if (vardata1->rel &&
5577 217690 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
5578 70815 : *join_is_reversed = true; /* var1 is on RHS */
5579 293941 : else if (vardata2->rel &&
5580 146875 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
5581 304 : *join_is_reversed = true; /* var2 is on LHS */
5582 : else
5583 146762 : *join_is_reversed = false;
5584 217881 : }
5585 :
5586 : /* statext_expressions_load copies the tuple, so just pfree it. */
5587 : static void
5588 1385 : ReleaseDummy(HeapTuple tuple)
5589 : {
5590 1385 : pfree(tuple);
5591 1385 : }
5592 :
5593 : /*
5594 : * examine_variable
5595 : * Try to look up statistical data about an expression.
5596 : * Fill in a VariableStatData struct to describe the expression.
5597 : *
5598 : * Inputs:
5599 : * root: the planner info
5600 : * node: the expression tree to examine
5601 : * varRelid: see specs for restriction selectivity functions
5602 : *
5603 : * Outputs: *vardata is filled as follows:
5604 : * var: the input expression (with any phvs or binary relabeling stripped,
5605 : * if it is or contains a variable; but otherwise unchanged)
5606 : * rel: RelOptInfo for relation containing variable; NULL if expression
5607 : * contains no Vars (NOTE this could point to a RelOptInfo of a
5608 : * subquery, not one in the current query).
5609 : * statsTuple: the pg_statistic entry for the variable, if one exists;
5610 : * otherwise NULL.
5611 : * freefunc: pointer to a function to release statsTuple with.
5612 : * vartype: exposed type of the expression; this should always match
5613 : * the declared input type of the operator we are estimating for.
5614 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5615 : * commonly the same as the exposed type of the variable argument,
5616 : * but can be different in binary-compatible-type cases.
5617 : * isunique: true if we were able to match the var to a unique index, a
5618 : * single-column DISTINCT or GROUP-BY clause, implying its values are
5619 : * unique for this query. (Caution: this should be trusted for
5620 : * statistical purposes only, since we do not check indimmediate nor
5621 : * verify that the exact same definition of equality applies.)
5622 : * acl_ok: true if current user has permission to read all table rows from
5623 : * the column(s) underlying the pg_statistic entry. This is consulted by
5624 : * statistic_proc_security_check().
5625 : *
5626 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5627 : */
5628 : void
5629 2554435 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5630 : VariableStatData *vardata)
5631 : {
5632 : Node *basenode;
5633 : Relids varnos;
5634 : Relids basevarnos;
5635 : RelOptInfo *onerel;
5636 :
5637 : /* Make sure we don't return dangling pointers in vardata */
5638 17881045 : MemSet(vardata, 0, sizeof(VariableStatData));
5639 :
5640 : /* Save the exposed type of the expression */
5641 2554435 : vardata->vartype = exprType(node);
5642 :
5643 : /*
5644 : * PlaceHolderVars are transparent for the purpose of statistics lookup;
5645 : * they do not alter the value distribution of the underlying expression.
5646 : * However, they can obscure the structure, preventing us from recognizing
5647 : * matches to base columns, index expressions, or extended statistics. So
5648 : * strip them out first.
5649 : */
5650 2554435 : basenode = strip_all_phvs_deep(root, node);
5651 :
5652 : /*
5653 : * Look inside any binary-compatible relabeling. We need to handle nested
5654 : * RelabelType nodes here, because the prior stripping of PlaceHolderVars
5655 : * may have brought separate RelabelTypes into adjacency.
5656 : */
5657 2594041 : while (IsA(basenode, RelabelType))
5658 39606 : basenode = (Node *) ((RelabelType *) basenode)->arg;
5659 :
5660 : /* Fast path for a simple Var */
5661 2554435 : if (IsA(basenode, Var) &&
5662 669498 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5663 : {
5664 1817760 : Var *var = (Var *) basenode;
5665 :
5666 : /* Set up result fields other than the stats tuple */
5667 1817760 : vardata->var = basenode; /* return Var without phvs or relabeling */
5668 1817760 : vardata->rel = find_base_rel(root, var->varno);
5669 1817760 : vardata->atttype = var->vartype;
5670 1817760 : vardata->atttypmod = var->vartypmod;
5671 1817760 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5672 :
5673 : /* Try to locate some stats */
5674 1817760 : examine_simple_variable(root, var, vardata);
5675 :
5676 1817760 : return;
5677 : }
5678 :
5679 : /*
5680 : * Okay, it's a more complicated expression. Determine variable
5681 : * membership. Note that when varRelid isn't zero, only vars of that
5682 : * relation are considered "real" vars.
5683 : */
5684 736675 : varnos = pull_varnos(root, basenode);
5685 736675 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5686 :
5687 736675 : onerel = NULL;
5688 :
5689 736675 : if (bms_is_empty(basevarnos))
5690 : {
5691 : /* No Vars at all ... must be pseudo-constant clause */
5692 : }
5693 : else
5694 : {
5695 : int relid;
5696 :
5697 : /* Check if the expression is in vars of a single base relation */
5698 391295 : if (bms_get_singleton_member(basevarnos, &relid))
5699 : {
5700 387713 : if (varRelid == 0 || varRelid == relid)
5701 : {
5702 53864 : onerel = find_base_rel(root, relid);
5703 53864 : vardata->rel = onerel;
5704 53864 : node = basenode; /* strip any phvs or relabeling */
5705 : }
5706 : /* else treat it as a constant */
5707 : }
5708 : else
5709 : {
5710 : /* varnos has multiple relids */
5711 3582 : if (varRelid == 0)
5712 : {
5713 : /* treat it as a variable of a join relation */
5714 2642 : vardata->rel = find_join_rel(root, varnos);
5715 2642 : node = basenode; /* strip any phvs or relabeling */
5716 : }
5717 940 : else if (bms_is_member(varRelid, varnos))
5718 : {
5719 : /* ignore the vars belonging to other relations */
5720 845 : vardata->rel = find_base_rel(root, varRelid);
5721 845 : node = basenode; /* strip any phvs or relabeling */
5722 : /* note: no point in expressional-index search here */
5723 : }
5724 : /* else treat it as a constant */
5725 : }
5726 : }
5727 :
5728 736675 : bms_free(basevarnos);
5729 :
5730 736675 : vardata->var = node;
5731 736675 : vardata->atttype = exprType(node);
5732 736675 : vardata->atttypmod = exprTypmod(node);
5733 :
5734 736675 : if (onerel)
5735 : {
5736 : /*
5737 : * We have an expression in vars of a single relation. Try to match
5738 : * it to expressional index columns, in hopes of finding some
5739 : * statistics.
5740 : *
5741 : * Note that we consider all index columns including INCLUDE columns,
5742 : * since there could be stats for such columns. But the test for
5743 : * uniqueness needs to be warier.
5744 : *
5745 : * XXX it's conceivable that there are multiple matches with different
5746 : * index opfamilies; if so, we need to pick one that matches the
5747 : * operator we are estimating for. FIXME later.
5748 : */
5749 : ListCell *ilist;
5750 : ListCell *slist;
5751 :
5752 : /*
5753 : * The nullingrels bits within the expression could prevent us from
5754 : * matching it to expressional index columns or to the expressions in
5755 : * extended statistics. So strip them out first.
5756 : */
5757 53864 : if (bms_overlap(varnos, root->outer_join_rels))
5758 1577 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5759 :
5760 118909 : foreach(ilist, onerel->indexlist)
5761 : {
5762 67422 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5763 : ListCell *indexpr_item;
5764 : int pos;
5765 :
5766 67422 : indexpr_item = list_head(index->indexprs);
5767 67422 : if (indexpr_item == NULL)
5768 63470 : continue; /* no expressions here... */
5769 :
5770 5587 : for (pos = 0; pos < index->ncolumns; pos++)
5771 : {
5772 4012 : if (index->indexkeys[pos] == 0)
5773 : {
5774 : Node *indexkey;
5775 :
5776 3952 : if (indexpr_item == NULL)
5777 0 : elog(ERROR, "too few entries in indexprs list");
5778 3952 : indexkey = (Node *) lfirst(indexpr_item);
5779 3952 : if (indexkey && IsA(indexkey, RelabelType))
5780 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5781 3952 : if (equal(node, indexkey))
5782 : {
5783 : /*
5784 : * Found a match ... is it a unique index? Tests here
5785 : * should match has_unique_index().
5786 : */
5787 2907 : if (index->unique &&
5788 365 : index->nkeycolumns == 1 &&
5789 365 : pos == 0 &&
5790 365 : (index->indpred == NIL || index->predOK))
5791 365 : vardata->isunique = true;
5792 :
5793 : /*
5794 : * Has it got stats? We only consider stats for
5795 : * non-partial indexes, since partial indexes probably
5796 : * don't reflect whole-relation statistics; the above
5797 : * check for uniqueness is the only info we take from
5798 : * a partial index.
5799 : *
5800 : * An index stats hook, however, must make its own
5801 : * decisions about what to do with partial indexes.
5802 : */
5803 2907 : if (get_index_stats_hook &&
5804 0 : (*get_index_stats_hook) (root, index->indexoid,
5805 0 : pos + 1, vardata))
5806 : {
5807 : /*
5808 : * The hook took control of acquiring a stats
5809 : * tuple. If it did supply a tuple, it'd better
5810 : * have supplied a freefunc.
5811 : */
5812 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5813 0 : !vardata->freefunc)
5814 0 : elog(ERROR, "no function provided to release variable stats with");
5815 : }
5816 2907 : else if (index->indpred == NIL)
5817 : {
5818 2907 : vardata->statsTuple =
5819 5814 : SearchSysCache3(STATRELATTINH,
5820 : ObjectIdGetDatum(index->indexoid),
5821 2907 : Int16GetDatum(pos + 1),
5822 : BoolGetDatum(false));
5823 2907 : vardata->freefunc = ReleaseSysCache;
5824 :
5825 2907 : if (HeapTupleIsValid(vardata->statsTuple))
5826 : {
5827 : /*
5828 : * Test if user has permission to access all
5829 : * rows from the index's table.
5830 : *
5831 : * For simplicity, we insist on the whole
5832 : * table being selectable, rather than trying
5833 : * to identify which column(s) the index
5834 : * depends on.
5835 : *
5836 : * Note that for an inheritance child,
5837 : * permissions are checked on the inheritance
5838 : * root parent, and whole-table select
5839 : * privilege on the parent doesn't quite
5840 : * guarantee that the user could read all
5841 : * columns of the child. But in practice it's
5842 : * unlikely that any interesting security
5843 : * violation could result from allowing access
5844 : * to the expression index's stats, so we
5845 : * allow it anyway. See similar code in
5846 : * examine_simple_variable() for additional
5847 : * comments.
5848 : */
5849 2377 : vardata->acl_ok =
5850 2377 : all_rows_selectable(root,
5851 2377 : index->rel->relid,
5852 : NULL);
5853 : }
5854 : else
5855 : {
5856 : /* suppress leakproofness checks later */
5857 530 : vardata->acl_ok = true;
5858 : }
5859 : }
5860 2907 : if (vardata->statsTuple)
5861 2377 : break;
5862 : }
5863 1575 : indexpr_item = lnext(index->indexprs, indexpr_item);
5864 : }
5865 : }
5866 3952 : if (vardata->statsTuple)
5867 2377 : break;
5868 : }
5869 :
5870 : /*
5871 : * Search extended statistics for one with a matching expression.
5872 : * There might be multiple ones, so just grab the first one. In the
5873 : * future, we might consider the statistics target (and pick the most
5874 : * accurate statistics) and maybe some other parameters.
5875 : */
5876 57338 : foreach(slist, onerel->statlist)
5877 : {
5878 3714 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5879 3714 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5880 : ListCell *expr_item;
5881 : int pos;
5882 :
5883 : /*
5884 : * Stop once we've found statistics for the expression (either
5885 : * from extended stats, or for an index in the preceding loop).
5886 : */
5887 3714 : if (vardata->statsTuple)
5888 240 : break;
5889 :
5890 : /* skip stats without per-expression stats */
5891 3474 : if (info->kind != STATS_EXT_EXPRESSIONS)
5892 1788 : continue;
5893 :
5894 : /* skip stats with mismatching stxdinherit value */
5895 1686 : if (info->inherit != rte->inh)
5896 5 : continue;
5897 :
5898 1681 : pos = 0;
5899 2766 : foreach(expr_item, info->exprs)
5900 : {
5901 2471 : Node *expr = (Node *) lfirst(expr_item);
5902 :
5903 : Assert(expr);
5904 :
5905 : /* strip RelabelType before comparing it */
5906 2471 : if (expr && IsA(expr, RelabelType))
5907 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5908 :
5909 : /* found a match, see if we can extract pg_statistic row */
5910 2471 : if (equal(node, expr))
5911 : {
5912 : /*
5913 : * XXX Not sure if we should cache the tuple somewhere.
5914 : * Now we just create a new copy every time.
5915 : */
5916 1386 : vardata->statsTuple =
5917 1386 : statext_expressions_load(info->statOid, rte->inh, pos);
5918 :
5919 : /* Nothing to release if no data found */
5920 1386 : if (vardata->statsTuple != NULL)
5921 : {
5922 1385 : vardata->freefunc = ReleaseDummy;
5923 : }
5924 :
5925 : /*
5926 : * Test if user has permission to access all rows from the
5927 : * table.
5928 : *
5929 : * For simplicity, we insist on the whole table being
5930 : * selectable, rather than trying to identify which
5931 : * column(s) the statistics object depends on.
5932 : *
5933 : * Note that for an inheritance child, permissions are
5934 : * checked on the inheritance root parent, and whole-table
5935 : * select privilege on the parent doesn't quite guarantee
5936 : * that the user could read all columns of the child. But
5937 : * in practice it's unlikely that any interesting security
5938 : * violation could result from allowing access to the
5939 : * expression stats, so we allow it anyway. See similar
5940 : * code in examine_simple_variable() for additional
5941 : * comments.
5942 : */
5943 1386 : vardata->acl_ok = all_rows_selectable(root,
5944 : onerel->relid,
5945 : NULL);
5946 :
5947 1386 : break;
5948 : }
5949 :
5950 1085 : pos++;
5951 : }
5952 : }
5953 : }
5954 :
5955 736675 : bms_free(varnos);
5956 : }
5957 :
5958 : /*
5959 : * strip_all_phvs_deep
5960 : * Deeply strip all PlaceHolderVars in an expression.
5961 :
5962 : * As a performance optimization, we first use a lightweight walker to check
5963 : * for the presence of any PlaceHolderVars. The expensive mutator is invoked
5964 : * only if a PlaceHolderVar is found, avoiding unnecessary memory allocation
5965 : * and tree copying in the common case where no PlaceHolderVars are present.
5966 : */
5967 : static Node *
5968 2554435 : strip_all_phvs_deep(PlannerInfo *root, Node *node)
5969 : {
5970 : /* If there are no PHVs anywhere, we needn't work hard */
5971 2554435 : if (root->glob->lastPHId == 0)
5972 2528769 : return node;
5973 :
5974 25666 : if (!contain_placeholder_walker(node, NULL))
5975 22186 : return node;
5976 3480 : return strip_all_phvs_mutator(node, NULL);
5977 : }
5978 :
5979 : /*
5980 : * contain_placeholder_walker
5981 : * Lightweight walker to check if an expression contains any
5982 : * PlaceHolderVars
5983 : */
5984 : static bool
5985 28982 : contain_placeholder_walker(Node *node, void *context)
5986 : {
5987 28982 : if (node == NULL)
5988 109 : return false;
5989 28873 : if (IsA(node, PlaceHolderVar))
5990 3480 : return true;
5991 :
5992 25393 : return expression_tree_walker(node, contain_placeholder_walker, context);
5993 : }
5994 :
5995 : /*
5996 : * strip_all_phvs_mutator
5997 : * Mutator to deeply strip all PlaceHolderVars
5998 : */
5999 : static Node *
6000 9383 : strip_all_phvs_mutator(Node *node, void *context)
6001 : {
6002 9383 : if (node == NULL)
6003 34 : return NULL;
6004 9349 : if (IsA(node, PlaceHolderVar))
6005 : {
6006 : /* Strip it and recurse into its contained expression */
6007 3600 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6008 :
6009 3600 : return strip_all_phvs_mutator((Node *) phv->phexpr, context);
6010 : }
6011 :
6012 5749 : return expression_tree_mutator(node, strip_all_phvs_mutator, context);
6013 : }
6014 :
6015 : /*
6016 : * examine_simple_variable
6017 : * Handle a simple Var for examine_variable
6018 : *
6019 : * This is split out as a subroutine so that we can recurse to deal with
6020 : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
6021 : *
6022 : * We already filled in all the fields of *vardata except for the stats tuple.
6023 : */
6024 : static void
6025 1825838 : examine_simple_variable(PlannerInfo *root, Var *var,
6026 : VariableStatData *vardata)
6027 : {
6028 1825838 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
6029 :
6030 : Assert(IsA(rte, RangeTblEntry));
6031 :
6032 1825838 : if (get_relation_stats_hook &&
6033 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
6034 : {
6035 : /*
6036 : * The hook took control of acquiring a stats tuple. If it did supply
6037 : * a tuple, it'd better have supplied a freefunc.
6038 : */
6039 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6040 0 : !vardata->freefunc)
6041 0 : elog(ERROR, "no function provided to release variable stats with");
6042 : }
6043 1825838 : else if (rte->rtekind == RTE_RELATION)
6044 : {
6045 : /*
6046 : * Plain table or parent of an inheritance appendrel, so look up the
6047 : * column in pg_statistic
6048 : */
6049 1734802 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6050 : ObjectIdGetDatum(rte->relid),
6051 1734802 : Int16GetDatum(var->varattno),
6052 1734802 : BoolGetDatum(rte->inh));
6053 1734802 : vardata->freefunc = ReleaseSysCache;
6054 :
6055 1734802 : if (HeapTupleIsValid(vardata->statsTuple))
6056 : {
6057 : /*
6058 : * Test if user has permission to read all rows from this column.
6059 : *
6060 : * This requires that the user has the appropriate SELECT
6061 : * privileges and that there are no securityQuals from security
6062 : * barrier views or RLS policies. If that's not the case, then we
6063 : * only permit leakproof functions to be passed pg_statistic data
6064 : * in vardata, otherwise the functions might reveal data that the
6065 : * user doesn't have permission to see --- see
6066 : * statistic_proc_security_check().
6067 : */
6068 1158081 : vardata->acl_ok =
6069 1158081 : all_rows_selectable(root, var->varno,
6070 1158081 : bms_make_singleton(var->varattno - FirstLowInvalidHeapAttributeNumber));
6071 : }
6072 : else
6073 : {
6074 : /* suppress any possible leakproofness checks later */
6075 576721 : vardata->acl_ok = true;
6076 : }
6077 : }
6078 91036 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
6079 83483 : (rte->rtekind == RTE_CTE && !rte->self_reference))
6080 : {
6081 : /*
6082 : * Plain subquery (not one that was converted to an appendrel) or
6083 : * non-recursive CTE. In either case, we can try to find out what the
6084 : * Var refers to within the subquery. We skip this for appendrel and
6085 : * recursive-CTE cases because any column stats we did find would
6086 : * likely not be very relevant.
6087 : */
6088 : PlannerInfo *subroot;
6089 : Query *subquery;
6090 : List *subtlist;
6091 : TargetEntry *ste;
6092 :
6093 : /*
6094 : * Punt if it's a whole-row var rather than a plain column reference.
6095 : */
6096 15511 : if (var->varattno == InvalidAttrNumber)
6097 0 : return;
6098 :
6099 : /*
6100 : * Otherwise, find the subquery's planner subroot.
6101 : */
6102 15511 : if (rte->rtekind == RTE_SUBQUERY)
6103 : {
6104 : RelOptInfo *rel;
6105 :
6106 : /*
6107 : * Fetch RelOptInfo for subquery. Note that we don't change the
6108 : * rel returned in vardata, since caller expects it to be a rel of
6109 : * the caller's query level. Because we might already be
6110 : * recursing, we can't use that rel pointer either, but have to
6111 : * look up the Var's rel afresh.
6112 : */
6113 7553 : rel = find_base_rel(root, var->varno);
6114 :
6115 7553 : subroot = rel->subroot;
6116 : }
6117 : else
6118 : {
6119 : /* CTE case is more difficult */
6120 : PlannerInfo *cteroot;
6121 : Index levelsup;
6122 : int ndx;
6123 : int plan_id;
6124 : ListCell *lc;
6125 :
6126 : /*
6127 : * Find the referenced CTE, and locate the subroot previously made
6128 : * for it.
6129 : */
6130 7958 : levelsup = rte->ctelevelsup;
6131 7958 : cteroot = root;
6132 19022 : while (levelsup-- > 0)
6133 : {
6134 11064 : cteroot = cteroot->parent_root;
6135 11064 : if (!cteroot) /* shouldn't happen */
6136 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
6137 : }
6138 :
6139 : /*
6140 : * Note: cte_plan_ids can be shorter than cteList, if we are still
6141 : * working on planning the CTEs (ie, this is a side-reference from
6142 : * another CTE). So we mustn't use forboth here.
6143 : */
6144 7958 : ndx = 0;
6145 11008 : foreach(lc, cteroot->parse->cteList)
6146 : {
6147 11008 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
6148 :
6149 11008 : if (strcmp(cte->ctename, rte->ctename) == 0)
6150 7958 : break;
6151 3050 : ndx++;
6152 : }
6153 7958 : if (lc == NULL) /* shouldn't happen */
6154 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
6155 7958 : if (ndx >= list_length(cteroot->cte_plan_ids))
6156 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
6157 7958 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
6158 7958 : if (plan_id <= 0)
6159 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
6160 7958 : subroot = list_nth(root->glob->subroots, plan_id - 1);
6161 : }
6162 :
6163 : /* If the subquery hasn't been planned yet, we have to punt */
6164 15511 : if (subroot == NULL)
6165 0 : return;
6166 : Assert(IsA(subroot, PlannerInfo));
6167 :
6168 : /*
6169 : * We must use the subquery parsetree as mangled by the planner, not
6170 : * the raw version from the RTE, because we need a Var that will refer
6171 : * to the subroot's live RelOptInfos. For instance, if any subquery
6172 : * pullup happened during planning, Vars in the targetlist might have
6173 : * gotten replaced, and we need to see the replacement expressions.
6174 : */
6175 15511 : subquery = subroot->parse;
6176 : Assert(IsA(subquery, Query));
6177 :
6178 : /*
6179 : * Punt if subquery uses set operations or grouping sets, as these
6180 : * will mash underlying columns' stats beyond recognition. (Set ops
6181 : * are particularly nasty; if we forged ahead, we would return stats
6182 : * relevant to only the leftmost subselect...) DISTINCT is also
6183 : * problematic, but we check that later because there is a possibility
6184 : * of learning something even with it.
6185 : */
6186 15511 : if (subquery->setOperations ||
6187 14139 : subquery->groupingSets)
6188 1449 : return;
6189 :
6190 : /* Get the subquery output expression referenced by the upper Var */
6191 14062 : if (subquery->returningList)
6192 153 : subtlist = subquery->returningList;
6193 : else
6194 13909 : subtlist = subquery->targetList;
6195 14062 : ste = get_tle_by_resno(subtlist, var->varattno);
6196 14062 : if (ste == NULL || ste->resjunk)
6197 0 : elog(ERROR, "subquery %s does not have attribute %d",
6198 : rte->eref->aliasname, var->varattno);
6199 14062 : var = (Var *) ste->expr;
6200 :
6201 : /*
6202 : * If subquery uses DISTINCT, we can't make use of any stats for the
6203 : * variable ... but, if it's the only DISTINCT column, we are entitled
6204 : * to consider it unique. We do the test this way so that it works
6205 : * for cases involving DISTINCT ON.
6206 : */
6207 14062 : if (subquery->distinctClause)
6208 : {
6209 1132 : if (list_length(subquery->distinctClause) == 1 &&
6210 361 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
6211 200 : vardata->isunique = true;
6212 : /* cannot go further */
6213 771 : return;
6214 : }
6215 :
6216 : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
6217 13291 : if (subquery->groupClause)
6218 : {
6219 891 : if (list_length(subquery->groupClause) == 1 &&
6220 374 : targetIsInSortList(ste, InvalidOid, subquery->groupClause))
6221 294 : vardata->isunique = true;
6222 : /* cannot go further */
6223 517 : return;
6224 : }
6225 :
6226 : /*
6227 : * If the sub-query originated from a view with the security_barrier
6228 : * attribute, we must not look at the variable's statistics, though it
6229 : * seems all right to notice the existence of a DISTINCT clause. So
6230 : * stop here.
6231 : *
6232 : * This is probably a harsher restriction than necessary; it's
6233 : * certainly OK for the selectivity estimator (which is a C function,
6234 : * and therefore omnipotent anyway) to look at the statistics. But
6235 : * many selectivity estimators will happily *invoke the operator
6236 : * function* to try to work out a good estimate - and that's not OK.
6237 : * So for now, don't dig down for stats.
6238 : */
6239 12774 : if (rte->security_barrier)
6240 457 : return;
6241 :
6242 : /* Can only handle a simple Var of subquery's query level */
6243 12317 : if (var && IsA(var, Var) &&
6244 8078 : var->varlevelsup == 0)
6245 : {
6246 : /*
6247 : * OK, recurse into the subquery. Note that the original setting
6248 : * of vardata->isunique (which will surely be false) is left
6249 : * unchanged in this situation. That's what we want, since even
6250 : * if the underlying column is unique, the subquery may have
6251 : * joined to other tables in a way that creates duplicates.
6252 : */
6253 8078 : examine_simple_variable(subroot, var, vardata);
6254 : }
6255 : }
6256 : else
6257 : {
6258 : /*
6259 : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
6260 : * see RTE_JOIN here because join alias Vars have already been
6261 : * flattened.) There's not much we can do with function outputs, but
6262 : * maybe someday try to be smarter about VALUES.
6263 : */
6264 : }
6265 : }
6266 :
6267 : /*
6268 : * all_rows_selectable
6269 : * Test whether the user has permission to select all rows from a given
6270 : * relation.
6271 : *
6272 : * Inputs:
6273 : * root: the planner info
6274 : * varno: the index of the relation (assumed to be an RTE_RELATION)
6275 : * varattnos: the attributes for which permission is required, or NULL if
6276 : * whole-table access is required
6277 : *
6278 : * Returns true if the user has the required select permissions, and there are
6279 : * no securityQuals from security barrier views or RLS policies.
6280 : *
6281 : * Note that if the relation is an inheritance child relation, securityQuals
6282 : * and access permissions are checked against the inheritance root parent (the
6283 : * relation actually mentioned in the query) --- see the comments in
6284 : * expand_single_inheritance_child() for an explanation of why it has to be
6285 : * done this way.
6286 : *
6287 : * If varattnos is non-NULL, its attribute numbers should be offset by
6288 : * FirstLowInvalidHeapAttributeNumber so that system attributes can be
6289 : * checked. If varattnos is NULL, only table-level SELECT privileges are
6290 : * checked, not any column-level privileges.
6291 : *
6292 : * Note: if the relation is accessed via a view, this function actually tests
6293 : * whether the view owner has permission to select from the relation. To
6294 : * ensure that the current user has permission, it is also necessary to check
6295 : * that the current user has permission to select from the view, which we do
6296 : * at planner-startup --- see subquery_planner().
6297 : *
6298 : * This is exported so that other estimation functions can use it.
6299 : */
6300 : bool
6301 1162054 : all_rows_selectable(PlannerInfo *root, Index varno, Bitmapset *varattnos)
6302 : {
6303 1162054 : RelOptInfo *rel = find_base_rel_noerr(root, varno);
6304 1162054 : RangeTblEntry *rte = planner_rt_fetch(varno, root);
6305 : Oid userid;
6306 : int varattno;
6307 :
6308 : Assert(rte->rtekind == RTE_RELATION);
6309 :
6310 : /*
6311 : * Determine the user ID to use for privilege checks (either the current
6312 : * user or the view owner, if we're accessing the table via a view).
6313 : *
6314 : * Normally the relation will have an associated RelOptInfo from which we
6315 : * can find the userid, but it might not if it's a RETURNING Var for an
6316 : * INSERT target relation. In that case use the RTEPermissionInfo
6317 : * associated with the RTE.
6318 : *
6319 : * If we navigate up to a parent relation, we keep using the same userid,
6320 : * since it's the same in all relations of a given inheritance tree.
6321 : */
6322 1162054 : if (rel)
6323 1162021 : userid = rel->userid;
6324 : else
6325 : {
6326 : RTEPermissionInfo *perminfo;
6327 :
6328 33 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
6329 33 : userid = perminfo->checkAsUser;
6330 : }
6331 1162054 : if (!OidIsValid(userid))
6332 1032857 : userid = GetUserId();
6333 :
6334 : /*
6335 : * Permissions and securityQuals must be checked on the table actually
6336 : * mentioned in the query, so if this is an inheritance child, navigate up
6337 : * to the inheritance root parent. If the user can read the whole table
6338 : * or the required columns there, then they can read from the child table
6339 : * too. For per-column checks, we must find out which of the root
6340 : * parent's attributes the child relation's attributes correspond to.
6341 : */
6342 1162054 : if (root->append_rel_array != NULL)
6343 : {
6344 : AppendRelInfo *appinfo;
6345 :
6346 177829 : appinfo = root->append_rel_array[varno];
6347 :
6348 : /*
6349 : * Partitions are mapped to their immediate parent, not the root
6350 : * parent, so must be ready to walk up multiple AppendRelInfos. But
6351 : * stop if we hit a parent that is not RTE_RELATION --- that's a
6352 : * flattened UNION ALL subquery, not an inheritance parent.
6353 : */
6354 334211 : while (appinfo &&
6355 156767 : planner_rt_fetch(appinfo->parent_relid,
6356 156767 : root)->rtekind == RTE_RELATION)
6357 : {
6358 156382 : Bitmapset *parent_varattnos = NULL;
6359 :
6360 : /*
6361 : * For each child attribute, find the corresponding parent
6362 : * attribute. In rare cases, the attribute may be local to the
6363 : * child table, in which case, we've got to live with having no
6364 : * access to this column.
6365 : */
6366 156382 : varattno = -1;
6367 310497 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6368 : {
6369 : AttrNumber attno;
6370 : AttrNumber parent_attno;
6371 :
6372 154115 : attno = varattno + FirstLowInvalidHeapAttributeNumber;
6373 :
6374 154115 : if (attno == InvalidAttrNumber)
6375 : {
6376 : /*
6377 : * Whole-row reference, so must map each column of the
6378 : * child to the parent table.
6379 : */
6380 30 : for (attno = 1; attno <= appinfo->num_child_cols; attno++)
6381 : {
6382 20 : parent_attno = appinfo->parent_colnos[attno - 1];
6383 20 : if (parent_attno == 0)
6384 0 : return false; /* attr is local to child */
6385 : parent_varattnos =
6386 20 : bms_add_member(parent_varattnos,
6387 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6388 : }
6389 : }
6390 : else
6391 : {
6392 154105 : if (attno < 0)
6393 : {
6394 : /* System attnos are the same in all tables */
6395 0 : parent_attno = attno;
6396 : }
6397 : else
6398 : {
6399 154105 : if (attno > appinfo->num_child_cols)
6400 0 : return false; /* safety check */
6401 154105 : parent_attno = appinfo->parent_colnos[attno - 1];
6402 154105 : if (parent_attno == 0)
6403 0 : return false; /* attr is local to child */
6404 : }
6405 : parent_varattnos =
6406 154105 : bms_add_member(parent_varattnos,
6407 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6408 : }
6409 : }
6410 :
6411 : /* If the parent is itself a child, continue up */
6412 156382 : varno = appinfo->parent_relid;
6413 156382 : varattnos = parent_varattnos;
6414 156382 : appinfo = root->append_rel_array[varno];
6415 : }
6416 :
6417 : /* Perform the access check on this parent rel */
6418 177829 : rte = planner_rt_fetch(varno, root);
6419 : Assert(rte->rtekind == RTE_RELATION);
6420 : }
6421 :
6422 : /*
6423 : * For all rows to be accessible, there must be no securityQuals from
6424 : * security barrier views or RLS policies.
6425 : */
6426 1162054 : if (rte->securityQuals != NIL)
6427 690 : return false;
6428 :
6429 : /*
6430 : * Test for table-level SELECT privilege.
6431 : *
6432 : * If varattnos is non-NULL, this is sufficient to give access to all
6433 : * requested attributes, even for a child table, since we have verified
6434 : * that all required child columns have matching parent columns.
6435 : *
6436 : * If varattnos is NULL (whole-table access requested), this doesn't
6437 : * necessarily guarantee that the user can read all columns of a child
6438 : * table, but we allow it anyway (see comments in examine_variable()) and
6439 : * don't bother checking any column privileges.
6440 : */
6441 1161364 : if (pg_class_aclcheck(rte->relid, userid, ACL_SELECT) == ACLCHECK_OK)
6442 1161021 : return true;
6443 :
6444 343 : if (varattnos == NULL)
6445 10 : return false; /* whole-table access requested */
6446 :
6447 : /*
6448 : * Don't have table-level SELECT privilege, so check per-column
6449 : * privileges.
6450 : */
6451 333 : varattno = -1;
6452 471 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6453 : {
6454 333 : AttrNumber attno = varattno + FirstLowInvalidHeapAttributeNumber;
6455 :
6456 333 : if (attno == InvalidAttrNumber)
6457 : {
6458 : /* Whole-row reference, so must have access to all columns */
6459 5 : if (pg_attribute_aclcheck_all(rte->relid, userid, ACL_SELECT,
6460 : ACLMASK_ALL) != ACLCHECK_OK)
6461 5 : return false;
6462 : }
6463 : else
6464 : {
6465 328 : if (pg_attribute_aclcheck(rte->relid, attno, userid,
6466 : ACL_SELECT) != ACLCHECK_OK)
6467 190 : return false;
6468 : }
6469 : }
6470 :
6471 : /* If we reach here, have all required column privileges */
6472 138 : return true;
6473 : }
6474 :
6475 : /*
6476 : * examine_indexcol_variable
6477 : * Try to look up statistical data about an index column/expression.
6478 : * Fill in a VariableStatData struct to describe the column.
6479 : *
6480 : * Inputs:
6481 : * root: the planner info
6482 : * index: the index whose column we're interested in
6483 : * indexcol: 0-based index column number (subscripts index->indexkeys[])
6484 : *
6485 : * Outputs: *vardata is filled as follows:
6486 : * var: the input expression (with any binary relabeling stripped, if
6487 : * it is or contains a variable; but otherwise the type is preserved)
6488 : * rel: RelOptInfo for table relation containing variable.
6489 : * statsTuple: the pg_statistic entry for the variable, if one exists;
6490 : * otherwise NULL.
6491 : * freefunc: pointer to a function to release statsTuple with.
6492 : *
6493 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
6494 : */
6495 : static void
6496 651809 : examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
6497 : int indexcol, VariableStatData *vardata)
6498 : {
6499 : AttrNumber colnum;
6500 : Oid relid;
6501 :
6502 651809 : if (index->indexkeys[indexcol] != 0)
6503 : {
6504 : /* Simple variable --- look to stats for the underlying table */
6505 649976 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6506 :
6507 : Assert(rte->rtekind == RTE_RELATION);
6508 649976 : relid = rte->relid;
6509 : Assert(relid != InvalidOid);
6510 649976 : colnum = index->indexkeys[indexcol];
6511 649976 : vardata->rel = index->rel;
6512 :
6513 649976 : if (get_relation_stats_hook &&
6514 0 : (*get_relation_stats_hook) (root, rte, colnum, vardata))
6515 : {
6516 : /*
6517 : * The hook took control of acquiring a stats tuple. If it did
6518 : * supply a tuple, it'd better have supplied a freefunc.
6519 : */
6520 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6521 0 : !vardata->freefunc)
6522 0 : elog(ERROR, "no function provided to release variable stats with");
6523 : }
6524 : else
6525 : {
6526 649976 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6527 : ObjectIdGetDatum(relid),
6528 : Int16GetDatum(colnum),
6529 649976 : BoolGetDatum(rte->inh));
6530 649976 : vardata->freefunc = ReleaseSysCache;
6531 : }
6532 : }
6533 : else
6534 : {
6535 : /* Expression --- maybe there are stats for the index itself */
6536 1833 : relid = index->indexoid;
6537 1833 : colnum = indexcol + 1;
6538 :
6539 1833 : if (get_index_stats_hook &&
6540 0 : (*get_index_stats_hook) (root, relid, colnum, vardata))
6541 : {
6542 : /*
6543 : * The hook took control of acquiring a stats tuple. If it did
6544 : * supply a tuple, it'd better have supplied a freefunc.
6545 : */
6546 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6547 0 : !vardata->freefunc)
6548 0 : elog(ERROR, "no function provided to release variable stats with");
6549 : }
6550 : else
6551 : {
6552 1833 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6553 : ObjectIdGetDatum(relid),
6554 : Int16GetDatum(colnum),
6555 : BoolGetDatum(false));
6556 1833 : vardata->freefunc = ReleaseSysCache;
6557 : }
6558 : }
6559 651809 : }
6560 :
6561 : /*
6562 : * Check whether it is permitted to call func_oid passing some of the
6563 : * pg_statistic data in vardata. We allow this if either of the following
6564 : * conditions is met: (1) the user has SELECT privileges on the table or
6565 : * column underlying the pg_statistic data and there are no securityQuals from
6566 : * security barrier views or RLS policies, or (2) the function is marked
6567 : * leakproof.
6568 : */
6569 : bool
6570 774603 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
6571 : {
6572 774603 : if (vardata->acl_ok)
6573 773092 : return true; /* have SELECT privs and no securityQuals */
6574 :
6575 1511 : if (!OidIsValid(func_oid))
6576 0 : return false;
6577 :
6578 1511 : if (get_func_leakproof(func_oid))
6579 748 : return true;
6580 :
6581 763 : ereport(DEBUG2,
6582 : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
6583 : get_func_name(func_oid))));
6584 763 : return false;
6585 : }
6586 :
6587 : /*
6588 : * get_variable_numdistinct
6589 : * Estimate the number of distinct values of a variable.
6590 : *
6591 : * vardata: results of examine_variable
6592 : * *isdefault: set to true if the result is a default rather than based on
6593 : * anything meaningful.
6594 : *
6595 : * NB: be careful to produce a positive integral result, since callers may
6596 : * compare the result to exact integer counts, or might divide by it.
6597 : */
6598 : double
6599 1293879 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
6600 : {
6601 : double stadistinct;
6602 1293879 : double stanullfrac = 0.0;
6603 : double ntuples;
6604 :
6605 1293879 : *isdefault = false;
6606 :
6607 : /*
6608 : * Determine the stadistinct value to use. There are cases where we can
6609 : * get an estimate even without a pg_statistic entry, or can get a better
6610 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
6611 : * (otherwise, assume no nulls, for lack of any better idea).
6612 : */
6613 1293879 : if (HeapTupleIsValid(vardata->statsTuple))
6614 : {
6615 : /* Use the pg_statistic entry */
6616 : Form_pg_statistic stats;
6617 :
6618 818437 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6619 818437 : stadistinct = stats->stadistinct;
6620 818437 : stanullfrac = stats->stanullfrac;
6621 : }
6622 475442 : else if (vardata->vartype == BOOLOID)
6623 : {
6624 : /*
6625 : * Special-case boolean columns: presumably, two distinct values.
6626 : *
6627 : * Are there any other datatypes we should wire in special estimates
6628 : * for?
6629 : */
6630 669 : stadistinct = 2.0;
6631 : }
6632 474773 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6633 : {
6634 : /*
6635 : * If the Var represents a column of a VALUES RTE, assume it's unique.
6636 : * This could of course be very wrong, but it should tend to be true
6637 : * in well-written queries. We could consider examining the VALUES'
6638 : * contents to get some real statistics; but that only works if the
6639 : * entries are all constants, and it would be pretty expensive anyway.
6640 : */
6641 2954 : stadistinct = -1.0; /* unique (and all non null) */
6642 : }
6643 : else
6644 : {
6645 : /*
6646 : * We don't keep statistics for system columns, but in some cases we
6647 : * can infer distinctness anyway.
6648 : */
6649 471819 : if (vardata->var && IsA(vardata->var, Var))
6650 : {
6651 441149 : switch (((Var *) vardata->var)->varattno)
6652 : {
6653 857 : case SelfItemPointerAttributeNumber:
6654 857 : stadistinct = -1.0; /* unique (and all non null) */
6655 857 : break;
6656 18252 : case TableOidAttributeNumber:
6657 18252 : stadistinct = 1.0; /* only 1 value */
6658 18252 : break;
6659 422040 : default:
6660 422040 : stadistinct = 0.0; /* means "unknown" */
6661 422040 : break;
6662 : }
6663 : }
6664 : else
6665 30670 : stadistinct = 0.0; /* means "unknown" */
6666 :
6667 : /*
6668 : * XXX consider using estimate_num_groups on expressions?
6669 : */
6670 : }
6671 :
6672 : /*
6673 : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6674 : * variable, assume it is unique no matter what pg_statistic says; the
6675 : * statistics could be out of date, or we might have found a partial
6676 : * unique index that proves the var is unique for this query. However,
6677 : * we'd better still believe the null-fraction statistic.
6678 : */
6679 1293879 : if (vardata->isunique)
6680 308557 : stadistinct = -1.0 * (1.0 - stanullfrac);
6681 :
6682 : /*
6683 : * If we had an absolute estimate, use that.
6684 : */
6685 1293879 : if (stadistinct > 0.0)
6686 315294 : return clamp_row_est(stadistinct);
6687 :
6688 : /*
6689 : * Otherwise we need to get the relation size; punt if not available.
6690 : */
6691 978585 : if (vardata->rel == NULL)
6692 : {
6693 533 : *isdefault = true;
6694 533 : return DEFAULT_NUM_DISTINCT;
6695 : }
6696 978052 : ntuples = vardata->rel->tuples;
6697 978052 : if (ntuples <= 0.0)
6698 : {
6699 109871 : *isdefault = true;
6700 109871 : return DEFAULT_NUM_DISTINCT;
6701 : }
6702 :
6703 : /*
6704 : * If we had a relative estimate, use that.
6705 : */
6706 868181 : if (stadistinct < 0.0)
6707 573233 : return clamp_row_est(-stadistinct * ntuples);
6708 :
6709 : /*
6710 : * With no data, estimate ndistinct = ntuples if the table is small, else
6711 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6712 : * that the behavior isn't discontinuous.
6713 : */
6714 294948 : if (ntuples < DEFAULT_NUM_DISTINCT)
6715 127345 : return clamp_row_est(ntuples);
6716 :
6717 167603 : *isdefault = true;
6718 167603 : return DEFAULT_NUM_DISTINCT;
6719 : }
6720 :
6721 : /*
6722 : * get_variable_range
6723 : * Estimate the minimum and maximum value of the specified variable.
6724 : * If successful, store values in *min and *max, and return true.
6725 : * If no data available, return false.
6726 : *
6727 : * sortop is the "<" comparison operator to use. This should generally
6728 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6729 : * The collation must be specified too.
6730 : */
6731 : static bool
6732 172424 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6733 : Oid sortop, Oid collation,
6734 : Datum *min, Datum *max)
6735 : {
6736 172424 : Datum tmin = 0;
6737 172424 : Datum tmax = 0;
6738 172424 : bool have_data = false;
6739 : int16 typLen;
6740 : bool typByVal;
6741 : Oid opfuncoid;
6742 : FmgrInfo opproc;
6743 : AttStatsSlot sslot;
6744 :
6745 : /*
6746 : * XXX It's very tempting to try to use the actual column min and max, if
6747 : * we can get them relatively-cheaply with an index probe. However, since
6748 : * this function is called many times during join planning, that could
6749 : * have unpleasant effects on planning speed. Need more investigation
6750 : * before enabling this.
6751 : */
6752 : #ifdef NOT_USED
6753 : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6754 : return true;
6755 : #endif
6756 :
6757 172424 : if (!HeapTupleIsValid(vardata->statsTuple))
6758 : {
6759 : /* no stats available, so default result */
6760 47231 : return false;
6761 : }
6762 :
6763 : /*
6764 : * If we can't apply the sortop to the stats data, just fail. In
6765 : * principle, if there's a histogram and no MCVs, we could return the
6766 : * histogram endpoints without ever applying the sortop ... but it's
6767 : * probably not worth trying, because whatever the caller wants to do with
6768 : * the endpoints would likely fail the security check too.
6769 : */
6770 125193 : if (!statistic_proc_security_check(vardata,
6771 125193 : (opfuncoid = get_opcode(sortop))))
6772 0 : return false;
6773 :
6774 125193 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6775 :
6776 125193 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6777 :
6778 : /*
6779 : * If there is a histogram with the ordering we want, grab the first and
6780 : * last values.
6781 : */
6782 125193 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6783 : STATISTIC_KIND_HISTOGRAM, sortop,
6784 : ATTSTATSSLOT_VALUES))
6785 : {
6786 77792 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6787 : {
6788 77792 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6789 77792 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6790 77792 : have_data = true;
6791 : }
6792 77792 : free_attstatsslot(&sslot);
6793 : }
6794 :
6795 : /*
6796 : * Otherwise, if there is a histogram with some other ordering, scan it
6797 : * and get the min and max values according to the ordering we want. This
6798 : * of course may not find values that are really extremal according to our
6799 : * ordering, but it beats ignoring available data.
6800 : */
6801 172594 : if (!have_data &&
6802 47401 : get_attstatsslot(&sslot, vardata->statsTuple,
6803 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6804 : ATTSTATSSLOT_VALUES))
6805 : {
6806 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6807 : collation, typLen, typByVal,
6808 : &tmin, &tmax, &have_data);
6809 0 : free_attstatsslot(&sslot);
6810 : }
6811 :
6812 : /*
6813 : * If we have most-common-values info, look for extreme MCVs. This is
6814 : * needed even if we also have a histogram, since the histogram excludes
6815 : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6816 : * be pretty wary of deciding that that is a full representation of the
6817 : * data. Proceed only if the MCVs represent the whole table (to within
6818 : * roundoff error).
6819 : */
6820 125193 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6821 : STATISTIC_KIND_MCV, InvalidOid,
6822 125193 : have_data ? ATTSTATSSLOT_VALUES :
6823 : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6824 : {
6825 70967 : bool use_mcvs = have_data;
6826 :
6827 70967 : if (!have_data)
6828 : {
6829 46500 : double sumcommon = 0.0;
6830 : double nullfrac;
6831 : int i;
6832 :
6833 372655 : for (i = 0; i < sslot.nnumbers; i++)
6834 326155 : sumcommon += sslot.numbers[i];
6835 46500 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6836 46500 : if (sumcommon + nullfrac > 0.99999)
6837 45501 : use_mcvs = true;
6838 : }
6839 :
6840 70967 : if (use_mcvs)
6841 69968 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6842 : collation, typLen, typByVal,
6843 : &tmin, &tmax, &have_data);
6844 70967 : free_attstatsslot(&sslot);
6845 : }
6846 :
6847 125193 : *min = tmin;
6848 125193 : *max = tmax;
6849 125193 : return have_data;
6850 : }
6851 :
6852 : /*
6853 : * get_stats_slot_range: scan sslot for min/max values
6854 : *
6855 : * Subroutine for get_variable_range: update min/max/have_data according
6856 : * to what we find in the statistics array.
6857 : */
6858 : static void
6859 69968 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6860 : Oid collation, int16 typLen, bool typByVal,
6861 : Datum *min, Datum *max, bool *p_have_data)
6862 : {
6863 69968 : Datum tmin = *min;
6864 69968 : Datum tmax = *max;
6865 69968 : bool have_data = *p_have_data;
6866 69968 : bool found_tmin = false;
6867 69968 : bool found_tmax = false;
6868 :
6869 : /* Look up the comparison function, if we didn't already do so */
6870 69968 : if (opproc->fn_oid != opfuncoid)
6871 69968 : fmgr_info(opfuncoid, opproc);
6872 :
6873 : /* Scan all the slot's values */
6874 1692275 : for (int i = 0; i < sslot->nvalues; i++)
6875 : {
6876 1622307 : if (!have_data)
6877 : {
6878 45501 : tmin = tmax = sslot->values[i];
6879 45501 : found_tmin = found_tmax = true;
6880 45501 : *p_have_data = have_data = true;
6881 45501 : continue;
6882 : }
6883 1576806 : if (DatumGetBool(FunctionCall2Coll(opproc,
6884 : collation,
6885 1576806 : sslot->values[i], tmin)))
6886 : {
6887 40095 : tmin = sslot->values[i];
6888 40095 : found_tmin = true;
6889 : }
6890 1576806 : if (DatumGetBool(FunctionCall2Coll(opproc,
6891 : collation,
6892 1576806 : tmax, sslot->values[i])))
6893 : {
6894 192009 : tmax = sslot->values[i];
6895 192009 : found_tmax = true;
6896 : }
6897 : }
6898 :
6899 : /*
6900 : * Copy the slot's values, if we found new extreme values.
6901 : */
6902 69968 : if (found_tmin)
6903 60551 : *min = datumCopy(tmin, typByVal, typLen);
6904 69968 : if (found_tmax)
6905 48257 : *max = datumCopy(tmax, typByVal, typLen);
6906 69968 : }
6907 :
6908 :
6909 : /*
6910 : * get_actual_variable_range
6911 : * Attempt to identify the current *actual* minimum and/or maximum
6912 : * of the specified variable, by looking for a suitable btree index
6913 : * and fetching its low and/or high values.
6914 : * If successful, store values in *min and *max, and return true.
6915 : * (Either pointer can be NULL if that endpoint isn't needed.)
6916 : * If unsuccessful, return false.
6917 : *
6918 : * sortop is the "<" comparison operator to use.
6919 : * collation is the required collation.
6920 : */
6921 : static bool
6922 119748 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6923 : Oid sortop, Oid collation,
6924 : Datum *min, Datum *max)
6925 : {
6926 119748 : bool have_data = false;
6927 119748 : RelOptInfo *rel = vardata->rel;
6928 : RangeTblEntry *rte;
6929 : ListCell *lc;
6930 :
6931 : /* No hope if no relation or it doesn't have indexes */
6932 119748 : if (rel == NULL || rel->indexlist == NIL)
6933 8938 : return false;
6934 : /* If it has indexes it must be a plain relation */
6935 110810 : rte = root->simple_rte_array[rel->relid];
6936 : Assert(rte->rtekind == RTE_RELATION);
6937 :
6938 : /* ignore partitioned tables. Any indexes here are not real indexes */
6939 110810 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6940 560 : return false;
6941 :
6942 : /* Search through the indexes to see if any match our problem */
6943 216287 : foreach(lc, rel->indexlist)
6944 : {
6945 185865 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6946 : ScanDirection indexscandir;
6947 : StrategyNumber strategy;
6948 :
6949 : /* Ignore non-ordering indexes */
6950 185865 : if (index->sortopfamily == NULL)
6951 3 : continue;
6952 :
6953 : /*
6954 : * Ignore partial indexes --- we only want stats that cover the entire
6955 : * relation.
6956 : */
6957 185862 : if (index->indpred != NIL)
6958 240 : continue;
6959 :
6960 : /*
6961 : * The index list might include hypothetical indexes inserted by a
6962 : * get_relation_info hook --- don't try to access them.
6963 : */
6964 185622 : if (index->hypothetical)
6965 0 : continue;
6966 :
6967 : /*
6968 : * get_actual_variable_endpoint uses the index-only-scan machinery, so
6969 : * ignore indexes that can't use it on their first column.
6970 : */
6971 185622 : if (!index->canreturn[0])
6972 0 : continue;
6973 :
6974 : /*
6975 : * The first index column must match the desired variable, sortop, and
6976 : * collation --- but we can use a descending-order index.
6977 : */
6978 185622 : if (collation != index->indexcollations[0])
6979 24651 : continue; /* test first 'cause it's cheapest */
6980 160971 : if (!match_index_to_operand(vardata->var, 0, index))
6981 81143 : continue;
6982 79828 : strategy = get_op_opfamily_strategy(sortop, index->sortopfamily[0]);
6983 79828 : switch (IndexAmTranslateStrategy(strategy, index->relam, index->sortopfamily[0], true))
6984 : {
6985 79828 : case COMPARE_LT:
6986 79828 : if (index->reverse_sort[0])
6987 0 : indexscandir = BackwardScanDirection;
6988 : else
6989 79828 : indexscandir = ForwardScanDirection;
6990 79828 : break;
6991 0 : case COMPARE_GT:
6992 0 : if (index->reverse_sort[0])
6993 0 : indexscandir = ForwardScanDirection;
6994 : else
6995 0 : indexscandir = BackwardScanDirection;
6996 0 : break;
6997 0 : default:
6998 : /* index doesn't match the sortop */
6999 0 : continue;
7000 : }
7001 :
7002 : /*
7003 : * Found a suitable index to extract data from. Set up some data that
7004 : * can be used by both invocations of get_actual_variable_endpoint.
7005 : */
7006 : {
7007 : MemoryContext tmpcontext;
7008 : MemoryContext oldcontext;
7009 : Relation heapRel;
7010 : Relation indexRel;
7011 : TupleTableSlot *slot;
7012 : int16 typLen;
7013 : bool typByVal;
7014 : ScanKeyData scankeys[1];
7015 :
7016 : /* Make sure any cruft gets recycled when we're done */
7017 79828 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
7018 : "get_actual_variable_range workspace",
7019 : ALLOCSET_DEFAULT_SIZES);
7020 79828 : oldcontext = MemoryContextSwitchTo(tmpcontext);
7021 :
7022 : /*
7023 : * Open the table and index so we can read from them. We should
7024 : * already have some type of lock on each.
7025 : */
7026 79828 : heapRel = table_open(rte->relid, NoLock);
7027 79828 : indexRel = index_open(index->indexoid, NoLock);
7028 :
7029 : /* build some stuff needed for indexscan execution */
7030 79828 : slot = table_slot_create(heapRel, NULL);
7031 79828 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
7032 :
7033 : /* set up an IS NOT NULL scan key so that we ignore nulls */
7034 79828 : ScanKeyEntryInitialize(&scankeys[0],
7035 : SK_ISNULL | SK_SEARCHNOTNULL,
7036 : 1, /* index col to scan */
7037 : InvalidStrategy, /* no strategy */
7038 : InvalidOid, /* no strategy subtype */
7039 : InvalidOid, /* no collation */
7040 : InvalidOid, /* no reg proc for this */
7041 : (Datum) 0); /* constant */
7042 :
7043 : /* If min is requested ... */
7044 79828 : if (min)
7045 : {
7046 44517 : have_data = get_actual_variable_endpoint(heapRel,
7047 : indexRel,
7048 : indexscandir,
7049 : scankeys,
7050 : typLen,
7051 : typByVal,
7052 : slot,
7053 : oldcontext,
7054 : min);
7055 : }
7056 : else
7057 : {
7058 : /* If min not requested, still want to fetch max */
7059 35311 : have_data = true;
7060 : }
7061 :
7062 : /* If max is requested, and we didn't already fail ... */
7063 79828 : if (max && have_data)
7064 : {
7065 : /* scan in the opposite direction; all else is the same */
7066 36469 : have_data = get_actual_variable_endpoint(heapRel,
7067 : indexRel,
7068 36469 : -indexscandir,
7069 : scankeys,
7070 : typLen,
7071 : typByVal,
7072 : slot,
7073 : oldcontext,
7074 : max);
7075 : }
7076 :
7077 : /* Clean everything up */
7078 79828 : ExecDropSingleTupleTableSlot(slot);
7079 :
7080 79828 : index_close(indexRel, NoLock);
7081 79828 : table_close(heapRel, NoLock);
7082 :
7083 79828 : MemoryContextSwitchTo(oldcontext);
7084 79828 : MemoryContextDelete(tmpcontext);
7085 :
7086 : /* And we're done */
7087 79828 : break;
7088 : }
7089 : }
7090 :
7091 110250 : return have_data;
7092 : }
7093 :
7094 : /*
7095 : * Get one endpoint datum (min or max depending on indexscandir) from the
7096 : * specified index. Return true if successful, false if not.
7097 : * On success, endpoint value is stored to *endpointDatum (and copied into
7098 : * outercontext).
7099 : *
7100 : * scankeys is a 1-element scankey array set up to reject nulls.
7101 : * typLen/typByVal describe the datatype of the index's first column.
7102 : * tableslot is a slot suitable to hold table tuples, in case we need
7103 : * to probe the heap.
7104 : * (We could compute these values locally, but that would mean computing them
7105 : * twice when get_actual_variable_range needs both the min and the max.)
7106 : *
7107 : * Failure occurs either when the index is empty, or we decide that it's
7108 : * taking too long to find a suitable tuple.
7109 : */
7110 : static bool
7111 80986 : get_actual_variable_endpoint(Relation heapRel,
7112 : Relation indexRel,
7113 : ScanDirection indexscandir,
7114 : ScanKey scankeys,
7115 : int16 typLen,
7116 : bool typByVal,
7117 : TupleTableSlot *tableslot,
7118 : MemoryContext outercontext,
7119 : Datum *endpointDatum)
7120 : {
7121 80986 : bool have_data = false;
7122 : SnapshotData SnapshotNonVacuumable;
7123 : IndexScanDesc index_scan;
7124 80986 : Buffer vmbuffer = InvalidBuffer;
7125 80986 : BlockNumber last_heap_block = InvalidBlockNumber;
7126 80986 : int n_visited_heap_pages = 0;
7127 : ItemPointer tid;
7128 : Datum values[INDEX_MAX_KEYS];
7129 : bool isnull[INDEX_MAX_KEYS];
7130 : MemoryContext oldcontext;
7131 :
7132 : /*
7133 : * We use the index-only-scan machinery for this. With mostly-static
7134 : * tables that's a win because it avoids a heap visit. It's also a win
7135 : * for dynamic data, but the reason is less obvious; read on for details.
7136 : *
7137 : * In principle, we should scan the index with our current active
7138 : * snapshot, which is the best approximation we've got to what the query
7139 : * will see when executed. But that won't be exact if a new snap is taken
7140 : * before running the query, and it can be very expensive if a lot of
7141 : * recently-dead or uncommitted rows exist at the beginning or end of the
7142 : * index (because we'll laboriously fetch each one and reject it).
7143 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
7144 : * and uncommitted rows as well as normal visible rows. On the other
7145 : * hand, it will reject known-dead rows, and thus not give a bogus answer
7146 : * when the extreme value has been deleted (unless the deletion was quite
7147 : * recent); that case motivates not using SnapshotAny here.
7148 : *
7149 : * A crucial point here is that SnapshotNonVacuumable, with
7150 : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
7151 : * condition that the indexscan will use to decide that index entries are
7152 : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
7153 : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
7154 : * have to continue scanning past it, we know that the indexscan will mark
7155 : * that index entry killed. That means that the next
7156 : * get_actual_variable_endpoint() call will not have to re-consider that
7157 : * index entry. In this way we avoid repetitive work when this function
7158 : * is used a lot during planning.
7159 : *
7160 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
7161 : * recently-created index, some index entries may point at "broken" HOT
7162 : * chains in which not all the tuple versions contain data matching the
7163 : * index entry. The live tuple version(s) certainly do match the index,
7164 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
7165 : * don't match. Hence, if we took data from the selected heap tuple, we
7166 : * might get a bogus answer that's not close to the index extremal value,
7167 : * or could even be NULL. We avoid this hazard because we take the data
7168 : * from the index entry not the heap.
7169 : *
7170 : * Despite all this care, there are situations where we might find many
7171 : * non-visible tuples near the end of the index. We don't want to expend
7172 : * a huge amount of time here, so we give up once we've read too many heap
7173 : * pages. When we fail for that reason, the caller will end up using
7174 : * whatever extremal value is recorded in pg_statistic.
7175 : */
7176 80986 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
7177 : GlobalVisTestFor(heapRel));
7178 :
7179 80986 : index_scan = index_beginscan(heapRel, indexRel,
7180 : &SnapshotNonVacuumable, NULL,
7181 : 1, 0);
7182 : /* Set it up for index-only scan */
7183 80986 : index_scan->xs_want_itup = true;
7184 80986 : index_rescan(index_scan, scankeys, 1, NULL, 0);
7185 :
7186 : /* Fetch first/next tuple in specified direction */
7187 99969 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
7188 : {
7189 99969 : BlockNumber block = ItemPointerGetBlockNumber(tid);
7190 :
7191 99969 : if (!VM_ALL_VISIBLE(heapRel,
7192 : block,
7193 : &vmbuffer))
7194 : {
7195 : /* Rats, we have to visit the heap to check visibility */
7196 68945 : if (!index_fetch_heap(index_scan, tableslot))
7197 : {
7198 : /*
7199 : * No visible tuple for this index entry, so we need to
7200 : * advance to the next entry. Before doing so, count heap
7201 : * page fetches and give up if we've done too many.
7202 : *
7203 : * We don't charge a page fetch if this is the same heap page
7204 : * as the previous tuple. This is on the conservative side,
7205 : * since other recently-accessed pages are probably still in
7206 : * buffers too; but it's good enough for this heuristic.
7207 : */
7208 : #define VISITED_PAGES_LIMIT 100
7209 :
7210 18983 : if (block != last_heap_block)
7211 : {
7212 1714 : last_heap_block = block;
7213 1714 : n_visited_heap_pages++;
7214 1714 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
7215 0 : break;
7216 : }
7217 :
7218 18983 : continue; /* no visible tuple, try next index entry */
7219 : }
7220 :
7221 : /* We don't actually need the heap tuple for anything */
7222 49962 : ExecClearTuple(tableslot);
7223 :
7224 : /*
7225 : * We don't care whether there's more than one visible tuple in
7226 : * the HOT chain; if any are visible, that's good enough.
7227 : */
7228 : }
7229 :
7230 : /*
7231 : * We expect that the index will return data in IndexTuple not
7232 : * HeapTuple format.
7233 : */
7234 80986 : if (!index_scan->xs_itup)
7235 0 : elog(ERROR, "no data returned for index-only scan");
7236 :
7237 : /*
7238 : * We do not yet support recheck here.
7239 : */
7240 80986 : if (index_scan->xs_recheck)
7241 0 : break;
7242 :
7243 : /* OK to deconstruct the index tuple */
7244 80986 : index_deform_tuple(index_scan->xs_itup,
7245 : index_scan->xs_itupdesc,
7246 : values, isnull);
7247 :
7248 : /* Shouldn't have got a null, but be careful */
7249 80986 : if (isnull[0])
7250 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
7251 : RelationGetRelationName(indexRel));
7252 :
7253 : /* Copy the index column value out to caller's context */
7254 80986 : oldcontext = MemoryContextSwitchTo(outercontext);
7255 80986 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
7256 80986 : MemoryContextSwitchTo(oldcontext);
7257 80986 : have_data = true;
7258 80986 : break;
7259 : }
7260 :
7261 80986 : if (vmbuffer != InvalidBuffer)
7262 73470 : ReleaseBuffer(vmbuffer);
7263 80986 : index_endscan(index_scan);
7264 :
7265 80986 : return have_data;
7266 : }
7267 :
7268 : /*
7269 : * find_join_input_rel
7270 : * Look up the input relation for a join.
7271 : *
7272 : * We assume that the input relation's RelOptInfo must have been constructed
7273 : * already.
7274 : */
7275 : static RelOptInfo *
7276 11208 : find_join_input_rel(PlannerInfo *root, Relids relids)
7277 : {
7278 11208 : RelOptInfo *rel = NULL;
7279 :
7280 11208 : if (!bms_is_empty(relids))
7281 : {
7282 : int relid;
7283 :
7284 11208 : if (bms_get_singleton_member(relids, &relid))
7285 10933 : rel = find_base_rel(root, relid);
7286 : else
7287 275 : rel = find_join_rel(root, relids);
7288 : }
7289 :
7290 11208 : if (rel == NULL)
7291 0 : elog(ERROR, "could not find RelOptInfo for given relids");
7292 :
7293 11208 : return rel;
7294 : }
7295 :
7296 :
7297 : /*-------------------------------------------------------------------------
7298 : *
7299 : * Index cost estimation functions
7300 : *
7301 : *-------------------------------------------------------------------------
7302 : */
7303 :
7304 : /*
7305 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
7306 : */
7307 : List *
7308 665927 : get_quals_from_indexclauses(List *indexclauses)
7309 : {
7310 665927 : List *result = NIL;
7311 : ListCell *lc;
7312 :
7313 1157513 : foreach(lc, indexclauses)
7314 : {
7315 491586 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7316 : ListCell *lc2;
7317 :
7318 985537 : foreach(lc2, iclause->indexquals)
7319 : {
7320 493951 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7321 :
7322 493951 : result = lappend(result, rinfo);
7323 : }
7324 : }
7325 665927 : return result;
7326 : }
7327 :
7328 : /*
7329 : * Compute the total evaluation cost of the comparison operands in a list
7330 : * of index qual expressions. Since we know these will be evaluated just
7331 : * once per scan, there's no need to distinguish startup from per-row cost.
7332 : *
7333 : * This can be used either on the result of get_quals_from_indexclauses(),
7334 : * or directly on an indexorderbys list. In both cases, we expect that the
7335 : * index key expression is on the left side of binary clauses.
7336 : */
7337 : Cost
7338 1321354 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
7339 : {
7340 1321354 : Cost qual_arg_cost = 0;
7341 : ListCell *lc;
7342 :
7343 1815654 : foreach(lc, indexquals)
7344 : {
7345 494300 : Expr *clause = (Expr *) lfirst(lc);
7346 : Node *other_operand;
7347 : QualCost index_qual_cost;
7348 :
7349 : /*
7350 : * Index quals will have RestrictInfos, indexorderbys won't. Look
7351 : * through RestrictInfo if present.
7352 : */
7353 494300 : if (IsA(clause, RestrictInfo))
7354 493941 : clause = ((RestrictInfo *) clause)->clause;
7355 :
7356 494300 : if (IsA(clause, OpExpr))
7357 : {
7358 478802 : OpExpr *op = (OpExpr *) clause;
7359 :
7360 478802 : other_operand = (Node *) lsecond(op->args);
7361 : }
7362 15498 : else if (IsA(clause, RowCompareExpr))
7363 : {
7364 330 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7365 :
7366 330 : other_operand = (Node *) rc->rargs;
7367 : }
7368 15168 : else if (IsA(clause, ScalarArrayOpExpr))
7369 : {
7370 12741 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7371 :
7372 12741 : other_operand = (Node *) lsecond(saop->args);
7373 : }
7374 2427 : else if (IsA(clause, NullTest))
7375 : {
7376 2427 : other_operand = NULL;
7377 : }
7378 : else
7379 : {
7380 0 : elog(ERROR, "unsupported indexqual type: %d",
7381 : (int) nodeTag(clause));
7382 : other_operand = NULL; /* keep compiler quiet */
7383 : }
7384 :
7385 494300 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
7386 494300 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
7387 : }
7388 1321354 : return qual_arg_cost;
7389 : }
7390 :
7391 : /*
7392 : * Compute generic index access cost estimates.
7393 : *
7394 : * See struct GenericCosts in selfuncs.h for more info.
7395 : */
7396 : void
7397 655437 : genericcostestimate(PlannerInfo *root,
7398 : IndexPath *path,
7399 : double loop_count,
7400 : GenericCosts *costs)
7401 : {
7402 655437 : IndexOptInfo *index = path->indexinfo;
7403 655437 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7404 655437 : List *indexOrderBys = path->indexorderbys;
7405 : Cost indexStartupCost;
7406 : Cost indexTotalCost;
7407 : Selectivity indexSelectivity;
7408 : double indexCorrelation;
7409 : double numIndexPages;
7410 : double numIndexTuples;
7411 : double spc_random_page_cost;
7412 : double num_sa_scans;
7413 : double num_outer_scans;
7414 : double num_scans;
7415 : double qual_op_cost;
7416 : double qual_arg_cost;
7417 : List *selectivityQuals;
7418 : ListCell *l;
7419 :
7420 : /*
7421 : * If the index is partial, AND the index predicate with the explicitly
7422 : * given indexquals to produce a more accurate idea of the index
7423 : * selectivity.
7424 : */
7425 655437 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7426 :
7427 : /*
7428 : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
7429 : * just assume that the number of index descents is the number of distinct
7430 : * combinations of array elements from all of the scan's SAOP clauses.
7431 : */
7432 655437 : num_sa_scans = costs->num_sa_scans;
7433 655437 : if (num_sa_scans < 1)
7434 : {
7435 5845 : num_sa_scans = 1;
7436 12282 : foreach(l, indexQuals)
7437 : {
7438 6437 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
7439 :
7440 6437 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
7441 : {
7442 21 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
7443 21 : double alength = estimate_array_length(root, lsecond(saop->args));
7444 :
7445 21 : if (alength > 1)
7446 21 : num_sa_scans *= alength;
7447 : }
7448 : }
7449 : }
7450 :
7451 : /* Estimate the fraction of main-table tuples that will be visited */
7452 655437 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7453 655437 : index->rel->relid,
7454 : JOIN_INNER,
7455 : NULL);
7456 :
7457 : /*
7458 : * If caller didn't give us an estimate, estimate the number of index
7459 : * tuples that will be visited. We do it in this rather peculiar-looking
7460 : * way in order to get the right answer for partial indexes.
7461 : */
7462 655437 : numIndexTuples = costs->numIndexTuples;
7463 655437 : if (numIndexTuples <= 0.0)
7464 : {
7465 83745 : numIndexTuples = indexSelectivity * index->rel->tuples;
7466 :
7467 : /*
7468 : * The above calculation counts all the tuples visited across all
7469 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
7470 : * average per-indexscan number, so adjust. This is a handy place to
7471 : * round to integer, too. (If caller supplied tuple estimate, it's
7472 : * responsible for handling these considerations.)
7473 : */
7474 83745 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7475 : }
7476 :
7477 : /*
7478 : * We can bound the number of tuples by the index size in any case. Also,
7479 : * always estimate at least one tuple is touched, even when
7480 : * indexSelectivity estimate is tiny.
7481 : */
7482 655437 : if (numIndexTuples > index->tuples)
7483 8379 : numIndexTuples = index->tuples;
7484 655437 : if (numIndexTuples < 1.0)
7485 87680 : numIndexTuples = 1.0;
7486 :
7487 : /*
7488 : * Estimate the number of index pages that will be retrieved.
7489 : *
7490 : * We use the simplistic method of taking a pro-rata fraction of the total
7491 : * number of index leaf pages. We disregard any overhead such as index
7492 : * metapages or upper tree levels.
7493 : *
7494 : * In practice access to upper index levels is often nearly free because
7495 : * those tend to stay in cache under load; moreover, the cost involved is
7496 : * highly dependent on index type. We therefore ignore such costs here
7497 : * and leave it to the caller to add a suitable charge if needed.
7498 : */
7499 655437 : if (index->pages > costs->numNonLeafPages && index->tuples > 1)
7500 589485 : numIndexPages =
7501 589485 : ceil(numIndexTuples * (index->pages - costs->numNonLeafPages)
7502 589485 : / index->tuples);
7503 : else
7504 65952 : numIndexPages = 1.0;
7505 :
7506 : /* fetch estimated page cost for tablespace containing index */
7507 655437 : get_tablespace_page_costs(index->reltablespace,
7508 : &spc_random_page_cost,
7509 : NULL);
7510 :
7511 : /*
7512 : * Now compute the disk access costs.
7513 : *
7514 : * The above calculations are all per-index-scan. However, if we are in a
7515 : * nestloop inner scan, we can expect the scan to be repeated (with
7516 : * different search keys) for each row of the outer relation. Likewise,
7517 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
7518 : * the potential for cache effects to reduce the number of disk page
7519 : * fetches needed. We want to estimate the average per-scan I/O cost in
7520 : * the presence of caching.
7521 : *
7522 : * We use the Mackert-Lohman formula (see costsize.c for details) to
7523 : * estimate the total number of page fetches that occur. While this
7524 : * wasn't what it was designed for, it seems a reasonable model anyway.
7525 : * Note that we are counting pages not tuples anymore, so we take N = T =
7526 : * index size, as if there were one "tuple" per page.
7527 : */
7528 655437 : num_outer_scans = loop_count;
7529 655437 : num_scans = num_sa_scans * num_outer_scans;
7530 :
7531 655437 : if (num_scans > 1)
7532 : {
7533 : double pages_fetched;
7534 :
7535 : /* total page fetches ignoring cache effects */
7536 79485 : pages_fetched = numIndexPages * num_scans;
7537 :
7538 : /* use Mackert and Lohman formula to adjust for cache effects */
7539 79485 : pages_fetched = index_pages_fetched(pages_fetched,
7540 : index->pages,
7541 79485 : (double) index->pages,
7542 : root);
7543 :
7544 : /*
7545 : * Now compute the total disk access cost, and then report a pro-rated
7546 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
7547 : * since that's internal to the indexscan.)
7548 : */
7549 79485 : indexTotalCost = (pages_fetched * spc_random_page_cost)
7550 : / num_outer_scans;
7551 : }
7552 : else
7553 : {
7554 : /*
7555 : * For a single index scan, we just charge spc_random_page_cost per
7556 : * page touched.
7557 : */
7558 575952 : indexTotalCost = numIndexPages * spc_random_page_cost;
7559 : }
7560 :
7561 : /*
7562 : * CPU cost: any complex expressions in the indexquals will need to be
7563 : * evaluated once at the start of the scan to reduce them to runtime keys
7564 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
7565 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
7566 : * indexqual operator. Because we have numIndexTuples as a per-scan
7567 : * number, we have to multiply by num_sa_scans to get the correct result
7568 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
7569 : * ORDER BY expressions.
7570 : *
7571 : * Note: this neglects the possible costs of rechecking lossy operators.
7572 : * Detecting that that might be needed seems more expensive than it's
7573 : * worth, though, considering all the other inaccuracies here ...
7574 : */
7575 655437 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
7576 655437 : index_other_operands_eval_cost(root, indexOrderBys);
7577 655437 : qual_op_cost = cpu_operator_cost *
7578 655437 : (list_length(indexQuals) + list_length(indexOrderBys));
7579 :
7580 655437 : indexStartupCost = qual_arg_cost;
7581 655437 : indexTotalCost += qual_arg_cost;
7582 655437 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
7583 :
7584 : /*
7585 : * Generic assumption about index correlation: there isn't any.
7586 : */
7587 655437 : indexCorrelation = 0.0;
7588 :
7589 : /*
7590 : * Return everything to caller.
7591 : */
7592 655437 : costs->indexStartupCost = indexStartupCost;
7593 655437 : costs->indexTotalCost = indexTotalCost;
7594 655437 : costs->indexSelectivity = indexSelectivity;
7595 655437 : costs->indexCorrelation = indexCorrelation;
7596 655437 : costs->numIndexPages = numIndexPages;
7597 655437 : costs->numIndexTuples = numIndexTuples;
7598 655437 : costs->spc_random_page_cost = spc_random_page_cost;
7599 655437 : costs->num_sa_scans = num_sa_scans;
7600 655437 : }
7601 :
7602 : /*
7603 : * If the index is partial, add its predicate to the given qual list.
7604 : *
7605 : * ANDing the index predicate with the explicitly given indexquals produces
7606 : * a more accurate idea of the index's selectivity. However, we need to be
7607 : * careful not to insert redundant clauses, because clauselist_selectivity()
7608 : * is easily fooled into computing a too-low selectivity estimate. Our
7609 : * approach is to add only the predicate clause(s) that cannot be proven to
7610 : * be implied by the given indexquals. This successfully handles cases such
7611 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
7612 : * There are many other cases where we won't detect redundancy, leading to a
7613 : * too-low selectivity estimate, which will bias the system in favor of using
7614 : * partial indexes where possible. That is not necessarily bad though.
7615 : *
7616 : * Note that indexQuals contains RestrictInfo nodes while the indpred
7617 : * does not, so the output list will be mixed. This is OK for both
7618 : * predicate_implied_by() and clauselist_selectivity(), but might be
7619 : * problematic if the result were passed to other things.
7620 : */
7621 : List *
7622 1124519 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
7623 : {
7624 1124519 : List *predExtraQuals = NIL;
7625 : ListCell *lc;
7626 :
7627 1124519 : if (index->indpred == NIL)
7628 1122983 : return indexQuals;
7629 :
7630 3082 : foreach(lc, index->indpred)
7631 : {
7632 1546 : Node *predQual = (Node *) lfirst(lc);
7633 1546 : List *oneQual = list_make1(predQual);
7634 :
7635 1546 : if (!predicate_implied_by(oneQual, indexQuals, false))
7636 1370 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7637 : }
7638 1536 : return list_concat(predExtraQuals, indexQuals);
7639 : }
7640 :
7641 : /*
7642 : * Estimate correlation of btree index's first column.
7643 : *
7644 : * If we can get an estimate of the first column's ordering correlation C
7645 : * from pg_statistic, estimate the index correlation as C for a single-column
7646 : * index, or C * 0.75 for multiple columns. The idea here is that multiple
7647 : * columns dilute the importance of the first column's ordering, but don't
7648 : * negate it entirely.
7649 : *
7650 : * We already filled in the stats tuple for *vardata when called.
7651 : */
7652 : static double
7653 430993 : btcost_correlation(IndexOptInfo *index, VariableStatData *vardata)
7654 : {
7655 : Oid sortop;
7656 : AttStatsSlot sslot;
7657 430993 : double indexCorrelation = 0;
7658 :
7659 : Assert(HeapTupleIsValid(vardata->statsTuple));
7660 :
7661 430993 : sortop = get_opfamily_member(index->opfamily[0],
7662 430993 : index->opcintype[0],
7663 430993 : index->opcintype[0],
7664 : BTLessStrategyNumber);
7665 861986 : if (OidIsValid(sortop) &&
7666 430993 : get_attstatsslot(&sslot, vardata->statsTuple,
7667 : STATISTIC_KIND_CORRELATION, sortop,
7668 : ATTSTATSSLOT_NUMBERS))
7669 : {
7670 : double varCorrelation;
7671 :
7672 : Assert(sslot.nnumbers == 1);
7673 425519 : varCorrelation = sslot.numbers[0];
7674 :
7675 425519 : if (index->reverse_sort[0])
7676 0 : varCorrelation = -varCorrelation;
7677 :
7678 425519 : if (index->nkeycolumns > 1)
7679 148393 : indexCorrelation = varCorrelation * 0.75;
7680 : else
7681 277126 : indexCorrelation = varCorrelation;
7682 :
7683 425519 : free_attstatsslot(&sslot);
7684 : }
7685 :
7686 430993 : return indexCorrelation;
7687 : }
7688 :
7689 : void
7690 649592 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7691 : Cost *indexStartupCost, Cost *indexTotalCost,
7692 : Selectivity *indexSelectivity, double *indexCorrelation,
7693 : double *indexPages)
7694 : {
7695 649592 : IndexOptInfo *index = path->indexinfo;
7696 649592 : GenericCosts costs = {0};
7697 649592 : VariableStatData vardata = {0};
7698 : double numIndexTuples;
7699 : Cost descentCost;
7700 : List *indexBoundQuals;
7701 : List *indexSkipQuals;
7702 : int indexcol;
7703 : bool eqQualHere;
7704 : bool found_row_compare;
7705 : bool found_array;
7706 : bool found_is_null_op;
7707 649592 : bool have_correlation = false;
7708 : double num_sa_scans;
7709 649592 : double correlation = 0.0;
7710 : ListCell *lc;
7711 :
7712 : /*
7713 : * For a btree scan, only leading '=' quals plus inequality quals for the
7714 : * immediately next attribute contribute to index selectivity (these are
7715 : * the "boundary quals" that determine the starting and stopping points of
7716 : * the index scan). Additional quals can suppress visits to the heap, so
7717 : * it's OK to count them in indexSelectivity, but they should not count
7718 : * for estimating numIndexTuples. So we must examine the given indexquals
7719 : * to find out which ones count as boundary quals. We rely on the
7720 : * knowledge that they are given in index column order. Note that nbtree
7721 : * preprocessing can add skip arrays that act as leading '=' quals in the
7722 : * absence of ordinary input '=' quals, so in practice _most_ input quals
7723 : * are able to act as index bound quals (which we take into account here).
7724 : *
7725 : * For a RowCompareExpr, we consider only the first column, just as
7726 : * rowcomparesel() does.
7727 : *
7728 : * If there's a SAOP or skip array in the quals, we'll actually perform up
7729 : * to N index descents (not just one), but the underlying array key's
7730 : * operator can be considered to act the same as it normally does.
7731 : */
7732 649592 : indexBoundQuals = NIL;
7733 649592 : indexSkipQuals = NIL;
7734 649592 : indexcol = 0;
7735 649592 : eqQualHere = false;
7736 649592 : found_row_compare = false;
7737 649592 : found_array = false;
7738 649592 : found_is_null_op = false;
7739 649592 : num_sa_scans = 1;
7740 1094189 : foreach(lc, path->indexclauses)
7741 : {
7742 472299 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7743 : ListCell *lc2;
7744 :
7745 472299 : if (indexcol < iclause->indexcol)
7746 : {
7747 93315 : double num_sa_scans_prev_cols = num_sa_scans;
7748 :
7749 : /*
7750 : * Beginning of a new column's quals.
7751 : *
7752 : * Skip scans use skip arrays, which are ScalarArrayOp style
7753 : * arrays that generate their elements procedurally and on demand.
7754 : * Given a multi-column index on "(a, b)", and an SQL WHERE clause
7755 : * "WHERE b = 42", a skip scan will effectively use an indexqual
7756 : * "WHERE a = ANY('{every col a value}') AND b = 42". (Obviously,
7757 : * the array on "a" must also return "IS NULL" matches, since our
7758 : * WHERE clause used no strict operator on "a").
7759 : *
7760 : * Here we consider how nbtree will backfill skip arrays for any
7761 : * index columns that lacked an '=' qual. This maintains our
7762 : * num_sa_scans estimate, and determines if this new column (the
7763 : * "iclause->indexcol" column, not the prior "indexcol" column)
7764 : * can have its RestrictInfos/quals added to indexBoundQuals.
7765 : *
7766 : * We'll need to handle columns that have inequality quals, where
7767 : * the skip array generates values from a range constrained by the
7768 : * quals (not every possible value). We've been maintaining
7769 : * indexSkipQuals to help with this; it will now contain all of
7770 : * the prior column's quals (that is, indexcol's quals) when they
7771 : * might be used for this.
7772 : */
7773 93315 : if (found_row_compare)
7774 : {
7775 : /*
7776 : * Skip arrays can't be added after a RowCompare input qual
7777 : * due to limitations in nbtree
7778 : */
7779 20 : break;
7780 : }
7781 93295 : if (eqQualHere)
7782 : {
7783 : /*
7784 : * Don't need to add a skip array for an indexcol that already
7785 : * has an '=' qual/equality constraint
7786 : */
7787 66062 : indexcol++;
7788 66062 : indexSkipQuals = NIL;
7789 : }
7790 93295 : eqQualHere = false;
7791 :
7792 95053 : while (indexcol < iclause->indexcol)
7793 : {
7794 : double ndistinct;
7795 29440 : bool isdefault = true;
7796 :
7797 29440 : found_array = true;
7798 :
7799 : /*
7800 : * A skipped attribute's ndistinct forms the basis of our
7801 : * estimate of the total number of "array elements" used by
7802 : * its skip array at runtime. Look that up first.
7803 : */
7804 29440 : examine_indexcol_variable(root, index, indexcol, &vardata);
7805 29440 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
7806 :
7807 29440 : if (indexcol == 0)
7808 : {
7809 : /*
7810 : * Get an estimate of the leading column's correlation in
7811 : * passing (avoids rereading variable stats below)
7812 : */
7813 27223 : if (HeapTupleIsValid(vardata.statsTuple))
7814 16129 : correlation = btcost_correlation(index, &vardata);
7815 27223 : have_correlation = true;
7816 : }
7817 :
7818 29440 : ReleaseVariableStats(vardata);
7819 :
7820 : /*
7821 : * If ndistinct is a default estimate, conservatively assume
7822 : * that no skipping will happen at runtime
7823 : */
7824 29440 : if (isdefault)
7825 : {
7826 9217 : num_sa_scans = num_sa_scans_prev_cols;
7827 27682 : break; /* done building indexBoundQuals */
7828 : }
7829 :
7830 : /*
7831 : * Apply indexcol's indexSkipQuals selectivity to ndistinct
7832 : */
7833 20223 : if (indexSkipQuals != NIL)
7834 : {
7835 : List *partialSkipQuals;
7836 : Selectivity ndistinctfrac;
7837 :
7838 : /*
7839 : * If the index is partial, AND the index predicate with
7840 : * the index-bound quals to produce a more accurate idea
7841 : * of the number of distinct values for prior indexcol
7842 : */
7843 552 : partialSkipQuals = add_predicate_to_index_quals(index,
7844 : indexSkipQuals);
7845 :
7846 552 : ndistinctfrac = clauselist_selectivity(root, partialSkipQuals,
7847 552 : index->rel->relid,
7848 : JOIN_INNER,
7849 : NULL);
7850 :
7851 : /*
7852 : * If ndistinctfrac is selective (on its own), the scan is
7853 : * unlikely to benefit from repositioning itself using
7854 : * later quals. Do not allow iclause->indexcol's quals to
7855 : * be added to indexBoundQuals (it would increase descent
7856 : * costs, without lowering numIndexTuples costs by much).
7857 : */
7858 552 : if (ndistinctfrac < DEFAULT_RANGE_INEQ_SEL)
7859 : {
7860 311 : num_sa_scans = num_sa_scans_prev_cols;
7861 311 : break; /* done building indexBoundQuals */
7862 : }
7863 :
7864 : /* Adjust ndistinct downward */
7865 241 : ndistinct = rint(ndistinct * ndistinctfrac);
7866 241 : ndistinct = Max(ndistinct, 1);
7867 : }
7868 :
7869 : /*
7870 : * When there's no inequality quals, account for the need to
7871 : * find an initial value by counting -inf/+inf as a value.
7872 : *
7873 : * We don't charge anything extra for possible next/prior key
7874 : * index probes, which are sometimes used to find the next
7875 : * valid skip array element (ahead of using the located
7876 : * element value to relocate the scan to the next position
7877 : * that might contain matching tuples). It seems hard to do
7878 : * better here. Use of the skip support infrastructure often
7879 : * avoids most next/prior key probes. But even when it can't,
7880 : * there's a decent chance that most individual next/prior key
7881 : * probes will locate a leaf page whose key space overlaps all
7882 : * of the scan's keys (even the lower-order keys) -- which
7883 : * also avoids the need for a separate, extra index descent.
7884 : * Note also that these probes are much cheaper than non-probe
7885 : * primitive index scans: they're reliably very selective.
7886 : */
7887 19912 : if (indexSkipQuals == NIL)
7888 19671 : ndistinct += 1;
7889 :
7890 : /*
7891 : * Update num_sa_scans estimate by multiplying by ndistinct.
7892 : *
7893 : * We make the pessimistic assumption that there is no
7894 : * naturally occurring cross-column correlation. This is
7895 : * often wrong, but it seems best to err on the side of not
7896 : * expecting skipping to be helpful...
7897 : */
7898 19912 : num_sa_scans *= ndistinct;
7899 :
7900 : /*
7901 : * ...but back out of adding this latest group of 1 or more
7902 : * skip arrays when num_sa_scans exceeds the total number of
7903 : * index pages (revert to num_sa_scans from before indexcol).
7904 : * This causes a sharp discontinuity in cost (as a function of
7905 : * the indexcol's ndistinct), but that is representative of
7906 : * actual runtime costs.
7907 : *
7908 : * Note that skipping is helpful when each primitive index
7909 : * scan only manages to skip over 1 or 2 irrelevant leaf pages
7910 : * on average. Skip arrays bring savings in CPU costs due to
7911 : * the scan not needing to evaluate indexquals against every
7912 : * tuple, which can greatly exceed any savings in I/O costs.
7913 : * This test is a test of whether num_sa_scans implies that
7914 : * we're past the point where the ability to skip ceases to
7915 : * lower the scan's costs (even qual evaluation CPU costs).
7916 : */
7917 19912 : if (index->pages < num_sa_scans)
7918 : {
7919 18154 : num_sa_scans = num_sa_scans_prev_cols;
7920 18154 : break; /* done building indexBoundQuals */
7921 : }
7922 :
7923 1758 : indexcol++;
7924 1758 : indexSkipQuals = NIL;
7925 : }
7926 :
7927 : /*
7928 : * Finished considering the need to add skip arrays to bridge an
7929 : * initial eqQualHere gap between the old and new index columns
7930 : * (or there was no initial eqQualHere gap in the first place).
7931 : *
7932 : * If an initial gap could not be bridged, then new column's quals
7933 : * (i.e. iclause->indexcol's quals) won't go into indexBoundQuals,
7934 : * and so won't affect our final numIndexTuples estimate.
7935 : */
7936 93295 : if (indexcol != iclause->indexcol)
7937 27682 : break; /* done building indexBoundQuals */
7938 : }
7939 :
7940 : Assert(indexcol == iclause->indexcol);
7941 :
7942 : /* Examine each indexqual associated with this index clause */
7943 891437 : foreach(lc2, iclause->indexquals)
7944 : {
7945 446840 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7946 446840 : Expr *clause = rinfo->clause;
7947 446840 : Oid clause_op = InvalidOid;
7948 : int op_strategy;
7949 :
7950 446840 : if (IsA(clause, OpExpr))
7951 : {
7952 432239 : OpExpr *op = (OpExpr *) clause;
7953 :
7954 432239 : clause_op = op->opno;
7955 : }
7956 14601 : else if (IsA(clause, RowCompareExpr))
7957 : {
7958 330 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7959 :
7960 330 : clause_op = linitial_oid(rc->opnos);
7961 330 : found_row_compare = true;
7962 : }
7963 14271 : else if (IsA(clause, ScalarArrayOpExpr))
7964 : {
7965 12385 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7966 12385 : Node *other_operand = (Node *) lsecond(saop->args);
7967 12385 : double alength = estimate_array_length(root, other_operand);
7968 :
7969 12385 : clause_op = saop->opno;
7970 12385 : found_array = true;
7971 : /* estimate SA descents by indexBoundQuals only */
7972 12385 : if (alength > 1)
7973 12165 : num_sa_scans *= alength;
7974 : }
7975 1886 : else if (IsA(clause, NullTest))
7976 : {
7977 1886 : NullTest *nt = (NullTest *) clause;
7978 :
7979 1886 : if (nt->nulltesttype == IS_NULL)
7980 : {
7981 200 : found_is_null_op = true;
7982 : /* IS NULL is like = for selectivity/skip scan purposes */
7983 200 : eqQualHere = true;
7984 : }
7985 : }
7986 : else
7987 0 : elog(ERROR, "unsupported indexqual type: %d",
7988 : (int) nodeTag(clause));
7989 :
7990 : /* check for equality operator */
7991 446840 : if (OidIsValid(clause_op))
7992 : {
7993 444954 : op_strategy = get_op_opfamily_strategy(clause_op,
7994 444954 : index->opfamily[indexcol]);
7995 : Assert(op_strategy != 0); /* not a member of opfamily?? */
7996 444954 : if (op_strategy == BTEqualStrategyNumber)
7997 421328 : eqQualHere = true;
7998 : }
7999 :
8000 446840 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
8001 :
8002 : /*
8003 : * We apply inequality selectivities to estimate index descent
8004 : * costs with scans that use skip arrays. Save this indexcol's
8005 : * RestrictInfos if it looks like they'll be needed for that.
8006 : */
8007 446840 : if (!eqQualHere && !found_row_compare &&
8008 24415 : indexcol < index->nkeycolumns - 1)
8009 4652 : indexSkipQuals = lappend(indexSkipQuals, rinfo);
8010 : }
8011 : }
8012 :
8013 : /*
8014 : * If index is unique and we found an '=' clause for each column, we can
8015 : * just assume numIndexTuples = 1 and skip the expensive
8016 : * clauselist_selectivity calculations. However, an array or NullTest
8017 : * always invalidates that theory (even when eqQualHere has been set).
8018 : */
8019 649592 : if (index->unique &&
8020 519539 : indexcol == index->nkeycolumns - 1 &&
8021 188238 : eqQualHere &&
8022 188238 : !found_array &&
8023 182648 : !found_is_null_op)
8024 182608 : numIndexTuples = 1.0;
8025 : else
8026 : {
8027 : List *selectivityQuals;
8028 : Selectivity btreeSelectivity;
8029 :
8030 : /*
8031 : * If the index is partial, AND the index predicate with the
8032 : * index-bound quals to produce a more accurate idea of the number of
8033 : * rows covered by the bound conditions.
8034 : */
8035 466984 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
8036 :
8037 466984 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
8038 466984 : index->rel->relid,
8039 : JOIN_INNER,
8040 : NULL);
8041 466984 : numIndexTuples = btreeSelectivity * index->rel->tuples;
8042 :
8043 : /*
8044 : * btree automatically combines individual array element primitive
8045 : * index scans whenever the tuples covered by the next set of array
8046 : * keys are close to tuples covered by the current set. That puts a
8047 : * natural ceiling on the worst case number of descents -- there
8048 : * cannot possibly be more than one descent per leaf page scanned.
8049 : *
8050 : * Clamp the number of descents to at most 1/3 the number of index
8051 : * pages. This avoids implausibly high estimates with low selectivity
8052 : * paths, where scans usually require only one or two descents. This
8053 : * is most likely to help when there are several SAOP clauses, where
8054 : * naively accepting the total number of distinct combinations of
8055 : * array elements as the number of descents would frequently lead to
8056 : * wild overestimates.
8057 : *
8058 : * We somewhat arbitrarily don't just make the cutoff the total number
8059 : * of leaf pages (we make it 1/3 the total number of pages instead) to
8060 : * give the btree code credit for its ability to continue on the leaf
8061 : * level with low selectivity scans.
8062 : *
8063 : * Note: num_sa_scans includes both ScalarArrayOp array elements and
8064 : * skip array elements whose qual affects our numIndexTuples estimate.
8065 : */
8066 466984 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
8067 466984 : num_sa_scans = Max(num_sa_scans, 1);
8068 :
8069 : /*
8070 : * As in genericcostestimate(), we have to adjust for any array quals
8071 : * included in indexBoundQuals, and then round to integer.
8072 : *
8073 : * It is tempting to make genericcostestimate behave as if array
8074 : * clauses work in almost the same way as scalar operators during
8075 : * btree scans, making the top-level scan look like a continuous scan
8076 : * (as opposed to num_sa_scans-many primitive index scans). After
8077 : * all, btree scans mostly work like that at runtime. However, such a
8078 : * scheme would badly bias genericcostestimate's simplistic approach
8079 : * to calculating numIndexPages through prorating.
8080 : *
8081 : * Stick with the approach taken by non-native SAOP scans for now.
8082 : * genericcostestimate will use the Mackert-Lohman formula to
8083 : * compensate for repeat page fetches, even though that definitely
8084 : * won't happen during btree scans (not for leaf pages, at least).
8085 : * We're usually very pessimistic about the number of primitive index
8086 : * scans that will be required, but it's not clear how to do better.
8087 : */
8088 466984 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
8089 : }
8090 :
8091 : /*
8092 : * Now do generic index cost estimation.
8093 : *
8094 : * While we expended effort to make realistic estimates of numIndexTuples
8095 : * and num_sa_scans, we are content to count only the btree metapage as
8096 : * non-leaf. btree fanout is typically high enough that upper pages are
8097 : * few relative to leaf pages, so accounting for them would move the
8098 : * estimates at most a percent or two. Given the uncertainty in just how
8099 : * many upper pages exist in a particular index, we'll skip trying to
8100 : * handle that.
8101 : */
8102 649592 : costs.numIndexTuples = numIndexTuples;
8103 649592 : costs.num_sa_scans = num_sa_scans;
8104 649592 : costs.numNonLeafPages = 1;
8105 :
8106 649592 : genericcostestimate(root, path, loop_count, &costs);
8107 :
8108 : /*
8109 : * Add a CPU-cost component to represent the costs of initial btree
8110 : * descent. We don't charge any I/O cost for touching upper btree levels,
8111 : * since they tend to stay in cache, but we still have to do about log2(N)
8112 : * comparisons to descend a btree of N leaf tuples. We charge one
8113 : * cpu_operator_cost per comparison.
8114 : *
8115 : * If there are SAOP or skip array keys, charge this once per estimated
8116 : * index descent. The ones after the first one are not startup cost so
8117 : * far as the overall plan goes, so just add them to "total" cost.
8118 : */
8119 649592 : if (index->tuples > 1) /* avoid computing log(0) */
8120 : {
8121 589022 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
8122 589022 : costs.indexStartupCost += descentCost;
8123 589022 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8124 : }
8125 :
8126 : /*
8127 : * Even though we're not charging I/O cost for touching upper btree pages,
8128 : * it's still reasonable to charge some CPU cost per page descended
8129 : * through. Moreover, if we had no such charge at all, bloated indexes
8130 : * would appear to have the same search cost as unbloated ones, at least
8131 : * in cases where only a single leaf page is expected to be visited. This
8132 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
8133 : * touched. The number of such pages is btree tree height plus one (ie,
8134 : * we charge for the leaf page too). As above, charge once per estimated
8135 : * SAOP/skip array descent.
8136 : */
8137 649592 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8138 649592 : costs.indexStartupCost += descentCost;
8139 649592 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8140 :
8141 649592 : if (!have_correlation)
8142 : {
8143 622369 : examine_indexcol_variable(root, index, 0, &vardata);
8144 622369 : if (HeapTupleIsValid(vardata.statsTuple))
8145 414864 : costs.indexCorrelation = btcost_correlation(index, &vardata);
8146 622369 : ReleaseVariableStats(vardata);
8147 : }
8148 : else
8149 : {
8150 : /* btcost_correlation already called earlier on */
8151 27223 : costs.indexCorrelation = correlation;
8152 : }
8153 :
8154 649592 : *indexStartupCost = costs.indexStartupCost;
8155 649592 : *indexTotalCost = costs.indexTotalCost;
8156 649592 : *indexSelectivity = costs.indexSelectivity;
8157 649592 : *indexCorrelation = costs.indexCorrelation;
8158 649592 : *indexPages = costs.numIndexPages;
8159 649592 : }
8160 :
8161 : void
8162 308 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8163 : Cost *indexStartupCost, Cost *indexTotalCost,
8164 : Selectivity *indexSelectivity, double *indexCorrelation,
8165 : double *indexPages)
8166 : {
8167 308 : GenericCosts costs = {0};
8168 :
8169 : /* As in btcostestimate, count only the metapage as non-leaf */
8170 308 : costs.numNonLeafPages = 1;
8171 :
8172 308 : genericcostestimate(root, path, loop_count, &costs);
8173 :
8174 : /*
8175 : * A hash index has no descent costs as such, since the index AM can go
8176 : * directly to the target bucket after computing the hash value. There
8177 : * are a couple of other hash-specific costs that we could conceivably add
8178 : * here, though:
8179 : *
8180 : * Ideally we'd charge spc_random_page_cost for each page in the target
8181 : * bucket, not just the numIndexPages pages that genericcostestimate
8182 : * thought we'd visit. However in most cases we don't know which bucket
8183 : * that will be. There's no point in considering the average bucket size
8184 : * because the hash AM makes sure that's always one page.
8185 : *
8186 : * Likewise, we could consider charging some CPU for each index tuple in
8187 : * the bucket, if we knew how many there were. But the per-tuple cost is
8188 : * just a hash value comparison, not a general datatype-dependent
8189 : * comparison, so any such charge ought to be quite a bit less than
8190 : * cpu_operator_cost; which makes it probably not worth worrying about.
8191 : *
8192 : * A bigger issue is that chance hash-value collisions will result in
8193 : * wasted probes into the heap. We don't currently attempt to model this
8194 : * cost on the grounds that it's rare, but maybe it's not rare enough.
8195 : * (Any fix for this ought to consider the generic lossy-operator problem,
8196 : * though; it's not entirely hash-specific.)
8197 : */
8198 :
8199 308 : *indexStartupCost = costs.indexStartupCost;
8200 308 : *indexTotalCost = costs.indexTotalCost;
8201 308 : *indexSelectivity = costs.indexSelectivity;
8202 308 : *indexCorrelation = costs.indexCorrelation;
8203 308 : *indexPages = costs.numIndexPages;
8204 308 : }
8205 :
8206 : void
8207 3649 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8208 : Cost *indexStartupCost, Cost *indexTotalCost,
8209 : Selectivity *indexSelectivity, double *indexCorrelation,
8210 : double *indexPages)
8211 : {
8212 3649 : IndexOptInfo *index = path->indexinfo;
8213 3649 : GenericCosts costs = {0};
8214 : Cost descentCost;
8215 :
8216 : /* GiST has no metapage, so we treat all pages as leaf pages */
8217 :
8218 3649 : genericcostestimate(root, path, loop_count, &costs);
8219 :
8220 : /*
8221 : * We model index descent costs similarly to those for btree, but to do
8222 : * that we first need an idea of the tree height. We somewhat arbitrarily
8223 : * assume that the fanout is 100, meaning the tree height is at most
8224 : * log100(index->pages).
8225 : *
8226 : * Although this computation isn't really expensive enough to require
8227 : * caching, we might as well use index->tree_height to cache it.
8228 : */
8229 3649 : if (index->tree_height < 0) /* unknown? */
8230 : {
8231 3642 : if (index->pages > 1) /* avoid computing log(0) */
8232 1956 : index->tree_height = (int) (log(index->pages) / log(100.0));
8233 : else
8234 1686 : index->tree_height = 0;
8235 : }
8236 :
8237 : /*
8238 : * Add a CPU-cost component to represent the costs of initial descent. We
8239 : * just use log(N) here not log2(N) since the branching factor isn't
8240 : * necessarily two anyway. As for btree, charge once per SA scan.
8241 : */
8242 3649 : if (index->tuples > 1) /* avoid computing log(0) */
8243 : {
8244 3649 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8245 3649 : costs.indexStartupCost += descentCost;
8246 3649 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8247 : }
8248 :
8249 : /*
8250 : * Likewise add a per-page charge, calculated the same as for btrees.
8251 : */
8252 3649 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8253 3649 : costs.indexStartupCost += descentCost;
8254 3649 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8255 :
8256 3649 : *indexStartupCost = costs.indexStartupCost;
8257 3649 : *indexTotalCost = costs.indexTotalCost;
8258 3649 : *indexSelectivity = costs.indexSelectivity;
8259 3649 : *indexCorrelation = costs.indexCorrelation;
8260 3649 : *indexPages = costs.numIndexPages;
8261 3649 : }
8262 :
8263 : void
8264 1482 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8265 : Cost *indexStartupCost, Cost *indexTotalCost,
8266 : Selectivity *indexSelectivity, double *indexCorrelation,
8267 : double *indexPages)
8268 : {
8269 1482 : IndexOptInfo *index = path->indexinfo;
8270 1482 : GenericCosts costs = {0};
8271 : Cost descentCost;
8272 :
8273 : /* As in btcostestimate, count only the metapage as non-leaf */
8274 1482 : costs.numNonLeafPages = 1;
8275 :
8276 1482 : genericcostestimate(root, path, loop_count, &costs);
8277 :
8278 : /*
8279 : * We model index descent costs similarly to those for btree, but to do
8280 : * that we first need an idea of the tree height. We somewhat arbitrarily
8281 : * assume that the fanout is 100, meaning the tree height is at most
8282 : * log100(index->pages).
8283 : *
8284 : * Although this computation isn't really expensive enough to require
8285 : * caching, we might as well use index->tree_height to cache it.
8286 : */
8287 1482 : if (index->tree_height < 0) /* unknown? */
8288 : {
8289 1477 : if (index->pages > 1) /* avoid computing log(0) */
8290 1477 : index->tree_height = (int) (log(index->pages) / log(100.0));
8291 : else
8292 0 : index->tree_height = 0;
8293 : }
8294 :
8295 : /*
8296 : * Add a CPU-cost component to represent the costs of initial descent. We
8297 : * just use log(N) here not log2(N) since the branching factor isn't
8298 : * necessarily two anyway. As for btree, charge once per SA scan.
8299 : */
8300 1482 : if (index->tuples > 1) /* avoid computing log(0) */
8301 : {
8302 1482 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8303 1482 : costs.indexStartupCost += descentCost;
8304 1482 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8305 : }
8306 :
8307 : /*
8308 : * Likewise add a per-page charge, calculated the same as for btrees.
8309 : */
8310 1482 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8311 1482 : costs.indexStartupCost += descentCost;
8312 1482 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8313 :
8314 1482 : *indexStartupCost = costs.indexStartupCost;
8315 1482 : *indexTotalCost = costs.indexTotalCost;
8316 1482 : *indexSelectivity = costs.indexSelectivity;
8317 1482 : *indexCorrelation = costs.indexCorrelation;
8318 1482 : *indexPages = costs.numIndexPages;
8319 1482 : }
8320 :
8321 :
8322 : /*
8323 : * Support routines for gincostestimate
8324 : */
8325 :
8326 : typedef struct
8327 : {
8328 : bool attHasFullScan[INDEX_MAX_KEYS];
8329 : bool attHasNormalScan[INDEX_MAX_KEYS];
8330 : double partialEntries;
8331 : double exactEntries;
8332 : double searchEntries;
8333 : double arrayScans;
8334 : } GinQualCounts;
8335 :
8336 : /*
8337 : * Estimate the number of index terms that need to be searched for while
8338 : * testing the given GIN query, and increment the counts in *counts
8339 : * appropriately. If the query is unsatisfiable, return false.
8340 : */
8341 : static bool
8342 1714 : gincost_pattern(IndexOptInfo *index, int indexcol,
8343 : Oid clause_op, Datum query,
8344 : GinQualCounts *counts)
8345 : {
8346 : FmgrInfo flinfo;
8347 : Oid extractProcOid;
8348 : Oid collation;
8349 : int strategy_op;
8350 : Oid lefttype,
8351 : righttype;
8352 1714 : int32 nentries = 0;
8353 1714 : bool *partial_matches = NULL;
8354 1714 : Pointer *extra_data = NULL;
8355 1714 : bool *nullFlags = NULL;
8356 1714 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
8357 : int32 i;
8358 :
8359 : Assert(indexcol < index->nkeycolumns);
8360 :
8361 : /*
8362 : * Get the operator's strategy number and declared input data types within
8363 : * the index opfamily. (We don't need the latter, but we use
8364 : * get_op_opfamily_properties because it will throw error if it fails to
8365 : * find a matching pg_amop entry.)
8366 : */
8367 1714 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
8368 : &strategy_op, &lefttype, &righttype);
8369 :
8370 : /*
8371 : * GIN always uses the "default" support functions, which are those with
8372 : * lefttype == righttype == the opclass' opcintype (see
8373 : * IndexSupportInitialize in relcache.c).
8374 : */
8375 1714 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
8376 1714 : index->opcintype[indexcol],
8377 1714 : index->opcintype[indexcol],
8378 : GIN_EXTRACTQUERY_PROC);
8379 :
8380 1714 : if (!OidIsValid(extractProcOid))
8381 : {
8382 : /* should not happen; throw same error as index_getprocinfo */
8383 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
8384 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
8385 : get_rel_name(index->indexoid));
8386 : }
8387 :
8388 : /*
8389 : * Choose collation to pass to extractProc (should match initGinState).
8390 : */
8391 1714 : if (OidIsValid(index->indexcollations[indexcol]))
8392 243 : collation = index->indexcollations[indexcol];
8393 : else
8394 1471 : collation = DEFAULT_COLLATION_OID;
8395 :
8396 1714 : fmgr_info(extractProcOid, &flinfo);
8397 :
8398 1714 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
8399 :
8400 1714 : FunctionCall7Coll(&flinfo,
8401 : collation,
8402 : query,
8403 : PointerGetDatum(&nentries),
8404 : UInt16GetDatum(strategy_op),
8405 : PointerGetDatum(&partial_matches),
8406 : PointerGetDatum(&extra_data),
8407 : PointerGetDatum(&nullFlags),
8408 : PointerGetDatum(&searchMode));
8409 :
8410 1714 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
8411 : {
8412 : /* No match is possible */
8413 10 : return false;
8414 : }
8415 :
8416 5854 : for (i = 0; i < nentries; i++)
8417 : {
8418 : /*
8419 : * For partial match we haven't any information to estimate number of
8420 : * matched entries in index, so, we just estimate it as 100
8421 : */
8422 4150 : if (partial_matches && partial_matches[i])
8423 361 : counts->partialEntries += 100;
8424 : else
8425 3789 : counts->exactEntries++;
8426 :
8427 4150 : counts->searchEntries++;
8428 : }
8429 :
8430 1704 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
8431 : {
8432 1322 : counts->attHasNormalScan[indexcol] = true;
8433 : }
8434 382 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
8435 : {
8436 : /* Treat "include empty" like an exact-match item */
8437 36 : counts->attHasNormalScan[indexcol] = true;
8438 36 : counts->exactEntries++;
8439 36 : counts->searchEntries++;
8440 : }
8441 : else
8442 : {
8443 : /* It's GIN_SEARCH_MODE_ALL */
8444 346 : counts->attHasFullScan[indexcol] = true;
8445 : }
8446 :
8447 1704 : return true;
8448 : }
8449 :
8450 : /*
8451 : * Estimate the number of index terms that need to be searched for while
8452 : * testing the given GIN index clause, and increment the counts in *counts
8453 : * appropriately. If the query is unsatisfiable, return false.
8454 : */
8455 : static bool
8456 1704 : gincost_opexpr(PlannerInfo *root,
8457 : IndexOptInfo *index,
8458 : int indexcol,
8459 : OpExpr *clause,
8460 : GinQualCounts *counts)
8461 : {
8462 1704 : Oid clause_op = clause->opno;
8463 1704 : Node *operand = (Node *) lsecond(clause->args);
8464 :
8465 : /* aggressively reduce to a constant, and look through relabeling */
8466 1704 : operand = estimate_expression_value(root, operand);
8467 :
8468 1704 : if (IsA(operand, RelabelType))
8469 0 : operand = (Node *) ((RelabelType *) operand)->arg;
8470 :
8471 : /*
8472 : * It's impossible to call extractQuery method for unknown operand. So
8473 : * unless operand is a Const we can't do much; just assume there will be
8474 : * one ordinary search entry from the operand at runtime.
8475 : */
8476 1704 : if (!IsA(operand, Const))
8477 : {
8478 0 : counts->exactEntries++;
8479 0 : counts->searchEntries++;
8480 0 : return true;
8481 : }
8482 :
8483 : /* If Const is null, there can be no matches */
8484 1704 : if (((Const *) operand)->constisnull)
8485 0 : return false;
8486 :
8487 : /* Otherwise, apply extractQuery and get the actual term counts */
8488 1704 : return gincost_pattern(index, indexcol, clause_op,
8489 : ((Const *) operand)->constvalue,
8490 : counts);
8491 : }
8492 :
8493 : /*
8494 : * Estimate the number of index terms that need to be searched for while
8495 : * testing the given GIN index clause, and increment the counts in *counts
8496 : * appropriately. If the query is unsatisfiable, return false.
8497 : *
8498 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
8499 : * each of which involves one value from the RHS array, plus all the
8500 : * non-array quals (if any). To model this, we average the counts across
8501 : * the RHS elements, and add the averages to the counts in *counts (which
8502 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
8503 : * by N, causing gincostestimate to scale up its estimates accordingly.
8504 : */
8505 : static bool
8506 5 : gincost_scalararrayopexpr(PlannerInfo *root,
8507 : IndexOptInfo *index,
8508 : int indexcol,
8509 : ScalarArrayOpExpr *clause,
8510 : double numIndexEntries,
8511 : GinQualCounts *counts)
8512 : {
8513 5 : Oid clause_op = clause->opno;
8514 5 : Node *rightop = (Node *) lsecond(clause->args);
8515 : ArrayType *arrayval;
8516 : int16 elmlen;
8517 : bool elmbyval;
8518 : char elmalign;
8519 : int numElems;
8520 : Datum *elemValues;
8521 : bool *elemNulls;
8522 : GinQualCounts arraycounts;
8523 5 : int numPossible = 0;
8524 : int i;
8525 :
8526 : Assert(clause->useOr);
8527 :
8528 : /* aggressively reduce to a constant, and look through relabeling */
8529 5 : rightop = estimate_expression_value(root, rightop);
8530 :
8531 5 : if (IsA(rightop, RelabelType))
8532 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
8533 :
8534 : /*
8535 : * It's impossible to call extractQuery method for unknown operand. So
8536 : * unless operand is a Const we can't do much; just assume there will be
8537 : * one ordinary search entry from each array entry at runtime, and fall
8538 : * back on a probably-bad estimate of the number of array entries.
8539 : */
8540 5 : if (!IsA(rightop, Const))
8541 : {
8542 0 : counts->exactEntries++;
8543 0 : counts->searchEntries++;
8544 0 : counts->arrayScans *= estimate_array_length(root, rightop);
8545 0 : return true;
8546 : }
8547 :
8548 : /* If Const is null, there can be no matches */
8549 5 : if (((Const *) rightop)->constisnull)
8550 0 : return false;
8551 :
8552 : /* Otherwise, extract the array elements and iterate over them */
8553 5 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
8554 5 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
8555 : &elmlen, &elmbyval, &elmalign);
8556 5 : deconstruct_array(arrayval,
8557 : ARR_ELEMTYPE(arrayval),
8558 : elmlen, elmbyval, elmalign,
8559 : &elemValues, &elemNulls, &numElems);
8560 :
8561 5 : memset(&arraycounts, 0, sizeof(arraycounts));
8562 :
8563 15 : for (i = 0; i < numElems; i++)
8564 : {
8565 : GinQualCounts elemcounts;
8566 :
8567 : /* NULL can't match anything, so ignore, as the executor will */
8568 10 : if (elemNulls[i])
8569 0 : continue;
8570 :
8571 : /* Otherwise, apply extractQuery and get the actual term counts */
8572 10 : memset(&elemcounts, 0, sizeof(elemcounts));
8573 :
8574 10 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
8575 : &elemcounts))
8576 : {
8577 : /* We ignore array elements that are unsatisfiable patterns */
8578 10 : numPossible++;
8579 :
8580 10 : if (elemcounts.attHasFullScan[indexcol] &&
8581 0 : !elemcounts.attHasNormalScan[indexcol])
8582 : {
8583 : /*
8584 : * Full index scan will be required. We treat this as if
8585 : * every key in the index had been listed in the query; is
8586 : * that reasonable?
8587 : */
8588 0 : elemcounts.partialEntries = 0;
8589 0 : elemcounts.exactEntries = numIndexEntries;
8590 0 : elemcounts.searchEntries = numIndexEntries;
8591 : }
8592 10 : arraycounts.partialEntries += elemcounts.partialEntries;
8593 10 : arraycounts.exactEntries += elemcounts.exactEntries;
8594 10 : arraycounts.searchEntries += elemcounts.searchEntries;
8595 : }
8596 : }
8597 :
8598 5 : if (numPossible == 0)
8599 : {
8600 : /* No satisfiable patterns in the array */
8601 0 : return false;
8602 : }
8603 :
8604 : /*
8605 : * Now add the averages to the global counts. This will give us an
8606 : * estimate of the average number of terms searched for in each indexscan,
8607 : * including contributions from both array and non-array quals.
8608 : */
8609 5 : counts->partialEntries += arraycounts.partialEntries / numPossible;
8610 5 : counts->exactEntries += arraycounts.exactEntries / numPossible;
8611 5 : counts->searchEntries += arraycounts.searchEntries / numPossible;
8612 :
8613 5 : counts->arrayScans *= numPossible;
8614 :
8615 5 : return true;
8616 : }
8617 :
8618 : /*
8619 : * GIN has search behavior completely different from other index types
8620 : */
8621 : void
8622 1546 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8623 : Cost *indexStartupCost, Cost *indexTotalCost,
8624 : Selectivity *indexSelectivity, double *indexCorrelation,
8625 : double *indexPages)
8626 : {
8627 1546 : IndexOptInfo *index = path->indexinfo;
8628 1546 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8629 : List *selectivityQuals;
8630 1546 : double numPages = index->pages,
8631 1546 : numTuples = index->tuples;
8632 : double numEntryPages,
8633 : numDataPages,
8634 : numPendingPages,
8635 : numEntries;
8636 : GinQualCounts counts;
8637 : bool matchPossible;
8638 : bool fullIndexScan;
8639 : double partialScale;
8640 : double entryPagesFetched,
8641 : dataPagesFetched,
8642 : dataPagesFetchedBySel;
8643 : double qual_op_cost,
8644 : qual_arg_cost,
8645 : spc_random_page_cost,
8646 : outer_scans;
8647 : Cost descentCost;
8648 : Relation indexRel;
8649 : GinStatsData ginStats;
8650 : ListCell *lc;
8651 : int i;
8652 :
8653 : /*
8654 : * Obtain statistical information from the meta page, if possible. Else
8655 : * set ginStats to zeroes, and we'll cope below.
8656 : */
8657 1546 : if (!index->hypothetical)
8658 : {
8659 : /* Lock should have already been obtained in plancat.c */
8660 1546 : indexRel = index_open(index->indexoid, NoLock);
8661 1546 : ginGetStats(indexRel, &ginStats);
8662 1546 : index_close(indexRel, NoLock);
8663 : }
8664 : else
8665 : {
8666 0 : memset(&ginStats, 0, sizeof(ginStats));
8667 : }
8668 :
8669 : /*
8670 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
8671 : * trusted, but the other fields are data as of the last VACUUM. We can
8672 : * scale them up to account for growth since then, but that method only
8673 : * goes so far; in the worst case, the stats might be for a completely
8674 : * empty index, and scaling them will produce pretty bogus numbers.
8675 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
8676 : * it's grown more than that, fall back to estimating things only from the
8677 : * assumed-accurate index size. But we'll trust nPendingPages in any case
8678 : * so long as it's not clearly insane, ie, more than the index size.
8679 : */
8680 1546 : if (ginStats.nPendingPages < numPages)
8681 1546 : numPendingPages = ginStats.nPendingPages;
8682 : else
8683 0 : numPendingPages = 0;
8684 :
8685 1546 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
8686 1546 : ginStats.nTotalPages > numPages / 4 &&
8687 1506 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
8688 1292 : {
8689 : /*
8690 : * OK, the stats seem close enough to sane to be trusted. But we
8691 : * still need to scale them by the ratio numPages / nTotalPages to
8692 : * account for growth since the last VACUUM.
8693 : */
8694 1292 : double scale = numPages / ginStats.nTotalPages;
8695 :
8696 1292 : numEntryPages = ceil(ginStats.nEntryPages * scale);
8697 1292 : numDataPages = ceil(ginStats.nDataPages * scale);
8698 1292 : numEntries = ceil(ginStats.nEntries * scale);
8699 : /* ensure we didn't round up too much */
8700 1292 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
8701 1292 : numDataPages = Min(numDataPages,
8702 : numPages - numPendingPages - numEntryPages);
8703 : }
8704 : else
8705 : {
8706 : /*
8707 : * We might get here because it's a hypothetical index, or an index
8708 : * created pre-9.1 and never vacuumed since upgrading (in which case
8709 : * its stats would read as zeroes), or just because it's grown too
8710 : * much since the last VACUUM for us to put our faith in scaling.
8711 : *
8712 : * Invent some plausible internal statistics based on the index page
8713 : * count (and clamp that to at least 10 pages, just in case). We
8714 : * estimate that 90% of the index is entry pages, and the rest is data
8715 : * pages. Estimate 100 entries per entry page; this is rather bogus
8716 : * since it'll depend on the size of the keys, but it's more robust
8717 : * than trying to predict the number of entries per heap tuple.
8718 : */
8719 254 : numPages = Max(numPages, 10);
8720 254 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
8721 254 : numDataPages = numPages - numPendingPages - numEntryPages;
8722 254 : numEntries = floor(numEntryPages * 100);
8723 : }
8724 :
8725 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
8726 1546 : if (numEntries < 1)
8727 0 : numEntries = 1;
8728 :
8729 : /*
8730 : * If the index is partial, AND the index predicate with the index-bound
8731 : * quals to produce a more accurate idea of the number of rows covered by
8732 : * the bound conditions.
8733 : */
8734 1546 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
8735 :
8736 : /* Estimate the fraction of main-table tuples that will be visited */
8737 3092 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
8738 1546 : index->rel->relid,
8739 : JOIN_INNER,
8740 : NULL);
8741 :
8742 : /* fetch estimated page cost for tablespace containing index */
8743 1546 : get_tablespace_page_costs(index->reltablespace,
8744 : &spc_random_page_cost,
8745 : NULL);
8746 :
8747 : /*
8748 : * Generic assumption about index correlation: there isn't any.
8749 : */
8750 1546 : *indexCorrelation = 0.0;
8751 :
8752 : /*
8753 : * Examine quals to estimate number of search entries & partial matches
8754 : */
8755 1546 : memset(&counts, 0, sizeof(counts));
8756 1546 : counts.arrayScans = 1;
8757 1546 : matchPossible = true;
8758 :
8759 3255 : foreach(lc, path->indexclauses)
8760 : {
8761 1709 : IndexClause *iclause = lfirst_node(IndexClause, lc);
8762 : ListCell *lc2;
8763 :
8764 3408 : foreach(lc2, iclause->indexquals)
8765 : {
8766 1709 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8767 1709 : Expr *clause = rinfo->clause;
8768 :
8769 1709 : if (IsA(clause, OpExpr))
8770 : {
8771 1704 : matchPossible = gincost_opexpr(root,
8772 : index,
8773 1704 : iclause->indexcol,
8774 : (OpExpr *) clause,
8775 : &counts);
8776 1704 : if (!matchPossible)
8777 10 : break;
8778 : }
8779 5 : else if (IsA(clause, ScalarArrayOpExpr))
8780 : {
8781 5 : matchPossible = gincost_scalararrayopexpr(root,
8782 : index,
8783 5 : iclause->indexcol,
8784 : (ScalarArrayOpExpr *) clause,
8785 : numEntries,
8786 : &counts);
8787 5 : if (!matchPossible)
8788 0 : break;
8789 : }
8790 : else
8791 : {
8792 : /* shouldn't be anything else for a GIN index */
8793 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
8794 : (int) nodeTag(clause));
8795 : }
8796 : }
8797 : }
8798 :
8799 : /* Fall out if there were any provably-unsatisfiable quals */
8800 1546 : if (!matchPossible)
8801 : {
8802 10 : *indexStartupCost = 0;
8803 10 : *indexTotalCost = 0;
8804 10 : *indexSelectivity = 0;
8805 10 : return;
8806 : }
8807 :
8808 : /*
8809 : * If attribute has a full scan and at the same time doesn't have normal
8810 : * scan, then we'll have to scan all non-null entries of that attribute.
8811 : * Currently, we don't have per-attribute statistics for GIN. Thus, we
8812 : * must assume the whole GIN index has to be scanned in this case.
8813 : */
8814 1536 : fullIndexScan = false;
8815 2989 : for (i = 0; i < index->nkeycolumns; i++)
8816 : {
8817 1726 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
8818 : {
8819 273 : fullIndexScan = true;
8820 273 : break;
8821 : }
8822 : }
8823 :
8824 1536 : if (fullIndexScan || indexQuals == NIL)
8825 : {
8826 : /*
8827 : * Full index scan will be required. We treat this as if every key in
8828 : * the index had been listed in the query; is that reasonable?
8829 : */
8830 273 : counts.partialEntries = 0;
8831 273 : counts.exactEntries = numEntries;
8832 273 : counts.searchEntries = numEntries;
8833 : }
8834 :
8835 : /* Will we have more than one iteration of a nestloop scan? */
8836 1536 : outer_scans = loop_count;
8837 :
8838 : /*
8839 : * Compute cost to begin scan, first of all, pay attention to pending
8840 : * list.
8841 : */
8842 1536 : entryPagesFetched = numPendingPages;
8843 :
8844 : /*
8845 : * Estimate number of entry pages read. We need to do
8846 : * counts.searchEntries searches. Use a power function as it should be,
8847 : * but tuples on leaf pages usually is much greater. Here we include all
8848 : * searches in entry tree, including search of first entry in partial
8849 : * match algorithm
8850 : */
8851 1536 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8852 :
8853 : /*
8854 : * Add an estimate of entry pages read by partial match algorithm. It's a
8855 : * scan over leaf pages in entry tree. We haven't any useful stats here,
8856 : * so estimate it as proportion. Because counts.partialEntries is really
8857 : * pretty bogus (see code above), it's possible that it is more than
8858 : * numEntries; clamp the proportion to ensure sanity.
8859 : */
8860 1536 : partialScale = counts.partialEntries / numEntries;
8861 1536 : partialScale = Min(partialScale, 1.0);
8862 :
8863 1536 : entryPagesFetched += ceil(numEntryPages * partialScale);
8864 :
8865 : /*
8866 : * Partial match algorithm reads all data pages before doing actual scan,
8867 : * so it's a startup cost. Again, we haven't any useful stats here, so
8868 : * estimate it as proportion.
8869 : */
8870 1536 : dataPagesFetched = ceil(numDataPages * partialScale);
8871 :
8872 1536 : *indexStartupCost = 0;
8873 1536 : *indexTotalCost = 0;
8874 :
8875 : /*
8876 : * Add a CPU-cost component to represent the costs of initial entry btree
8877 : * descent. We don't charge any I/O cost for touching upper btree levels,
8878 : * since they tend to stay in cache, but we still have to do about log2(N)
8879 : * comparisons to descend a btree of N leaf tuples. We charge one
8880 : * cpu_operator_cost per comparison.
8881 : *
8882 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
8883 : * ones after the first one are not startup cost so far as the overall
8884 : * plan is concerned, so add them only to "total" cost.
8885 : */
8886 1536 : if (numEntries > 1) /* avoid computing log(0) */
8887 : {
8888 1536 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
8889 1536 : *indexStartupCost += descentCost * counts.searchEntries;
8890 1536 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
8891 : }
8892 :
8893 : /*
8894 : * Add a cpu cost per entry-page fetched. This is not amortized over a
8895 : * loop.
8896 : */
8897 1536 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8898 1536 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8899 :
8900 : /*
8901 : * Add a cpu cost per data-page fetched. This is also not amortized over a
8902 : * loop. Since those are the data pages from the partial match algorithm,
8903 : * charge them as startup cost.
8904 : */
8905 1536 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
8906 :
8907 : /*
8908 : * Since we add the startup cost to the total cost later on, remove the
8909 : * initial arrayscan from the total.
8910 : */
8911 1536 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8912 :
8913 : /*
8914 : * Calculate cache effects if more than one scan due to nestloops or array
8915 : * quals. The result is pro-rated per nestloop scan, but the array qual
8916 : * factor shouldn't be pro-rated (compare genericcostestimate).
8917 : */
8918 1536 : if (outer_scans > 1 || counts.arrayScans > 1)
8919 : {
8920 5 : entryPagesFetched *= outer_scans * counts.arrayScans;
8921 5 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
8922 : (BlockNumber) numEntryPages,
8923 : numEntryPages, root);
8924 5 : entryPagesFetched /= outer_scans;
8925 5 : dataPagesFetched *= outer_scans * counts.arrayScans;
8926 5 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8927 : (BlockNumber) numDataPages,
8928 : numDataPages, root);
8929 5 : dataPagesFetched /= outer_scans;
8930 : }
8931 :
8932 : /*
8933 : * Here we use random page cost because logically-close pages could be far
8934 : * apart on disk.
8935 : */
8936 1536 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
8937 :
8938 : /*
8939 : * Now compute the number of data pages fetched during the scan.
8940 : *
8941 : * We assume every entry to have the same number of items, and that there
8942 : * is no overlap between them. (XXX: tsvector and array opclasses collect
8943 : * statistics on the frequency of individual keys; it would be nice to use
8944 : * those here.)
8945 : */
8946 1536 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
8947 :
8948 : /*
8949 : * If there is a lot of overlap among the entries, in particular if one of
8950 : * the entries is very frequent, the above calculation can grossly
8951 : * under-estimate. As a simple cross-check, calculate a lower bound based
8952 : * on the overall selectivity of the quals. At a minimum, we must read
8953 : * one item pointer for each matching entry.
8954 : *
8955 : * The width of each item pointer varies, based on the level of
8956 : * compression. We don't have statistics on that, but an average of
8957 : * around 3 bytes per item is fairly typical.
8958 : */
8959 1536 : dataPagesFetchedBySel = ceil(*indexSelectivity *
8960 1536 : (numTuples / (BLCKSZ / 3)));
8961 1536 : if (dataPagesFetchedBySel > dataPagesFetched)
8962 1243 : dataPagesFetched = dataPagesFetchedBySel;
8963 :
8964 : /* Add one page cpu-cost to the startup cost */
8965 1536 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
8966 :
8967 : /*
8968 : * Add once again a CPU-cost for those data pages, before amortizing for
8969 : * cache.
8970 : */
8971 1536 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8972 :
8973 : /* Account for cache effects, the same as above */
8974 1536 : if (outer_scans > 1 || counts.arrayScans > 1)
8975 : {
8976 5 : dataPagesFetched *= outer_scans * counts.arrayScans;
8977 5 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8978 : (BlockNumber) numDataPages,
8979 : numDataPages, root);
8980 5 : dataPagesFetched /= outer_scans;
8981 : }
8982 :
8983 : /* And apply random_page_cost as the cost per page */
8984 1536 : *indexTotalCost += *indexStartupCost +
8985 1536 : dataPagesFetched * spc_random_page_cost;
8986 :
8987 : /*
8988 : * Add on index qual eval costs, much as in genericcostestimate. We charge
8989 : * cpu but we can disregard indexorderbys, since GIN doesn't support
8990 : * those.
8991 : */
8992 1536 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8993 1536 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
8994 :
8995 1536 : *indexStartupCost += qual_arg_cost;
8996 1536 : *indexTotalCost += qual_arg_cost;
8997 :
8998 : /*
8999 : * Add a cpu cost per search entry, corresponding to the actual visited
9000 : * entries.
9001 : */
9002 1536 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
9003 : /* Now add a cpu cost per tuple in the posting lists / trees */
9004 1536 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
9005 1536 : *indexPages = dataPagesFetched;
9006 : }
9007 :
9008 : /*
9009 : * BRIN has search behavior completely different from other index types
9010 : */
9011 : void
9012 8944 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
9013 : Cost *indexStartupCost, Cost *indexTotalCost,
9014 : Selectivity *indexSelectivity, double *indexCorrelation,
9015 : double *indexPages)
9016 : {
9017 8944 : IndexOptInfo *index = path->indexinfo;
9018 8944 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
9019 8944 : double numPages = index->pages;
9020 8944 : RelOptInfo *baserel = index->rel;
9021 8944 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
9022 : Cost spc_seq_page_cost;
9023 : Cost spc_random_page_cost;
9024 : double qual_arg_cost;
9025 : double qualSelectivity;
9026 : BrinStatsData statsData;
9027 : double indexRanges;
9028 : double minimalRanges;
9029 : double estimatedRanges;
9030 : double selec;
9031 : Relation indexRel;
9032 : ListCell *l;
9033 : VariableStatData vardata;
9034 :
9035 : Assert(rte->rtekind == RTE_RELATION);
9036 :
9037 : /* fetch estimated page cost for the tablespace containing the index */
9038 8944 : get_tablespace_page_costs(index->reltablespace,
9039 : &spc_random_page_cost,
9040 : &spc_seq_page_cost);
9041 :
9042 : /*
9043 : * Obtain some data from the index itself, if possible. Otherwise invent
9044 : * some plausible internal statistics based on the relation page count.
9045 : */
9046 8944 : if (!index->hypothetical)
9047 : {
9048 : /*
9049 : * A lock should have already been obtained on the index in plancat.c.
9050 : */
9051 8944 : indexRel = index_open(index->indexoid, NoLock);
9052 8944 : brinGetStats(indexRel, &statsData);
9053 8944 : index_close(indexRel, NoLock);
9054 :
9055 : /* work out the actual number of ranges in the index */
9056 8944 : indexRanges = Max(ceil((double) baserel->pages /
9057 : statsData.pagesPerRange), 1.0);
9058 : }
9059 : else
9060 : {
9061 : /*
9062 : * Assume default number of pages per range, and estimate the number
9063 : * of ranges based on that.
9064 : */
9065 0 : indexRanges = Max(ceil((double) baserel->pages /
9066 : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
9067 :
9068 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
9069 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
9070 : }
9071 :
9072 : /*
9073 : * Compute index correlation
9074 : *
9075 : * Because we can use all index quals equally when scanning, we can use
9076 : * the largest correlation (in absolute value) among columns used by the
9077 : * query. Start at zero, the worst possible case. If we cannot find any
9078 : * correlation statistics, we will keep it as 0.
9079 : */
9080 8944 : *indexCorrelation = 0;
9081 :
9082 17889 : foreach(l, path->indexclauses)
9083 : {
9084 8945 : IndexClause *iclause = lfirst_node(IndexClause, l);
9085 8945 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
9086 :
9087 : /* attempt to lookup stats in relation for this index column */
9088 8945 : if (attnum != 0)
9089 : {
9090 : /* Simple variable -- look to stats for the underlying table */
9091 8945 : if (get_relation_stats_hook &&
9092 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
9093 : {
9094 : /*
9095 : * The hook took control of acquiring a stats tuple. If it
9096 : * did supply a tuple, it'd better have supplied a freefunc.
9097 : */
9098 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
9099 0 : elog(ERROR,
9100 : "no function provided to release variable stats with");
9101 : }
9102 : else
9103 : {
9104 8945 : vardata.statsTuple =
9105 8945 : SearchSysCache3(STATRELATTINH,
9106 : ObjectIdGetDatum(rte->relid),
9107 : Int16GetDatum(attnum),
9108 : BoolGetDatum(false));
9109 8945 : vardata.freefunc = ReleaseSysCache;
9110 : }
9111 : }
9112 : else
9113 : {
9114 : /*
9115 : * Looks like we've found an expression column in the index. Let's
9116 : * see if there's any stats for it.
9117 : */
9118 :
9119 : /* get the attnum from the 0-based index. */
9120 0 : attnum = iclause->indexcol + 1;
9121 :
9122 0 : if (get_index_stats_hook &&
9123 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
9124 : {
9125 : /*
9126 : * The hook took control of acquiring a stats tuple. If it
9127 : * did supply a tuple, it'd better have supplied a freefunc.
9128 : */
9129 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
9130 0 : !vardata.freefunc)
9131 0 : elog(ERROR, "no function provided to release variable stats with");
9132 : }
9133 : else
9134 : {
9135 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
9136 : ObjectIdGetDatum(index->indexoid),
9137 : Int16GetDatum(attnum),
9138 : BoolGetDatum(false));
9139 0 : vardata.freefunc = ReleaseSysCache;
9140 : }
9141 : }
9142 :
9143 8945 : if (HeapTupleIsValid(vardata.statsTuple))
9144 : {
9145 : AttStatsSlot sslot;
9146 :
9147 33 : if (get_attstatsslot(&sslot, vardata.statsTuple,
9148 : STATISTIC_KIND_CORRELATION, InvalidOid,
9149 : ATTSTATSSLOT_NUMBERS))
9150 : {
9151 33 : double varCorrelation = 0.0;
9152 :
9153 33 : if (sslot.nnumbers > 0)
9154 33 : varCorrelation = fabs(sslot.numbers[0]);
9155 :
9156 33 : if (varCorrelation > *indexCorrelation)
9157 33 : *indexCorrelation = varCorrelation;
9158 :
9159 33 : free_attstatsslot(&sslot);
9160 : }
9161 : }
9162 :
9163 8945 : ReleaseVariableStats(vardata);
9164 : }
9165 :
9166 8944 : qualSelectivity = clauselist_selectivity(root, indexQuals,
9167 8944 : baserel->relid,
9168 : JOIN_INNER, NULL);
9169 :
9170 : /*
9171 : * Now calculate the minimum possible ranges we could match with if all of
9172 : * the rows were in the perfect order in the table's heap.
9173 : */
9174 8944 : minimalRanges = ceil(indexRanges * qualSelectivity);
9175 :
9176 : /*
9177 : * Now estimate the number of ranges that we'll touch by using the
9178 : * indexCorrelation from the stats. Careful not to divide by zero (note
9179 : * we're using the absolute value of the correlation).
9180 : */
9181 8944 : if (*indexCorrelation < 1.0e-10)
9182 8911 : estimatedRanges = indexRanges;
9183 : else
9184 33 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
9185 :
9186 : /* we expect to visit this portion of the table */
9187 8944 : selec = estimatedRanges / indexRanges;
9188 :
9189 8944 : CLAMP_PROBABILITY(selec);
9190 :
9191 8944 : *indexSelectivity = selec;
9192 :
9193 : /*
9194 : * Compute the index qual costs, much as in genericcostestimate, to add to
9195 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
9196 : * support those.
9197 : */
9198 8944 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
9199 :
9200 : /*
9201 : * Compute the startup cost as the cost to read the whole revmap
9202 : * sequentially, including the cost to execute the index quals.
9203 : */
9204 8944 : *indexStartupCost =
9205 8944 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
9206 8944 : *indexStartupCost += qual_arg_cost;
9207 :
9208 : /*
9209 : * To read a BRIN index there might be a bit of back and forth over
9210 : * regular pages, as revmap might point to them out of sequential order;
9211 : * calculate the total cost as reading the whole index in random order.
9212 : */
9213 8944 : *indexTotalCost = *indexStartupCost +
9214 8944 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
9215 :
9216 : /*
9217 : * Charge a small amount per range tuple which we expect to match to. This
9218 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
9219 : * will set a bit for each page in the range when we find a matching
9220 : * range, so we must multiply the charge by the number of pages in the
9221 : * range.
9222 : */
9223 8944 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
9224 8944 : statsData.pagesPerRange;
9225 :
9226 8944 : *indexPages = index->pages;
9227 8944 : }
|
