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 704370 : eqsel(PG_FUNCTION_ARGS)
301 : {
302 704370 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
303 : }
304 :
305 : /*
306 : * Common code for eqsel() and neqsel()
307 : */
308 : static double
309 751084 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
310 : {
311 751084 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
312 751084 : Oid operator = PG_GETARG_OID(1);
313 751084 : List *args = (List *) PG_GETARG_POINTER(2);
314 751084 : int varRelid = PG_GETARG_INT32(3);
315 751084 : 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 751084 : if (negate)
326 : {
327 46714 : operator = get_negator(operator);
328 46714 : 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 751084 : if (!get_restriction_variable(root, args, varRelid,
340 : &vardata, &other, &varonleft))
341 5124 : 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 745954 : if (IsA(other, Const))
349 308924 : selec = var_eq_const(&vardata, operator, collation,
350 308924 : ((Const *) other)->constvalue,
351 308924 : ((Const *) other)->constisnull,
352 : varonleft, negate);
353 : else
354 437030 : selec = var_eq_non_const(&vardata, operator, collation, other,
355 : varonleft, negate);
356 :
357 745954 : ReleaseVariableStats(vardata);
358 :
359 745954 : 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 353842 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
369 : Datum constval, bool constisnull,
370 : bool varonleft, bool negate)
371 : {
372 : double selec;
373 353842 : 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 353842 : if (constisnull)
382 410 : 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 353432 : if (HeapTupleIsValid(vardata->statsTuple))
389 : {
390 : Form_pg_statistic stats;
391 :
392 266238 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
393 266238 : 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 353432 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
404 : {
405 84422 : selec = 1.0 / vardata->rel->tuples;
406 : }
407 468974 : else if (HeapTupleIsValid(vardata->statsTuple) &&
408 199964 : statistic_proc_security_check(vardata,
409 199964 : (opfuncoid = get_opcode(oproid))))
410 199964 : {
411 : AttStatsSlot sslot;
412 199964 : 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 199964 : if (get_attstatsslot(&sslot, vardata->statsTuple,
423 : STATISTIC_KIND_MCV, InvalidOid,
424 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
425 : {
426 178610 : LOCAL_FCINFO(fcinfo, 2);
427 : FmgrInfo eqproc;
428 :
429 178610 : 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 178610 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
438 : NULL, NULL);
439 178610 : fcinfo->args[0].isnull = false;
440 178610 : fcinfo->args[1].isnull = false;
441 : /* be careful to apply operator right way 'round */
442 178610 : if (varonleft)
443 178578 : fcinfo->args[1].value = constval;
444 : else
445 32 : fcinfo->args[0].value = constval;
446 :
447 3069216 : for (i = 0; i < sslot.nvalues; i++)
448 : {
449 : Datum fresult;
450 :
451 2986322 : if (varonleft)
452 2986266 : fcinfo->args[0].value = sslot.values[i];
453 : else
454 56 : fcinfo->args[1].value = sslot.values[i];
455 2986322 : fcinfo->isnull = false;
456 2986322 : fresult = FunctionCallInvoke(fcinfo);
457 2986322 : if (!fcinfo->isnull && DatumGetBool(fresult))
458 : {
459 95716 : match = true;
460 95716 : break;
461 : }
462 : }
463 : }
464 : else
465 : {
466 : /* no most-common-value info available */
467 21354 : i = 0; /* keep compiler quiet */
468 : }
469 :
470 199964 : 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 95716 : 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 104248 : double sumcommon = 0.0;
486 : double otherdistinct;
487 :
488 2606120 : for (i = 0; i < sslot.nnumbers; i++)
489 2501872 : sumcommon += sslot.numbers[i];
490 104248 : selec = 1.0 - sumcommon - nullfrac;
491 104248 : 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 104248 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
499 104248 : sslot.nnumbers;
500 104248 : if (otherdistinct > 1)
501 53960 : 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 104248 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
508 0 : selec = sslot.numbers[sslot.nnumbers - 1];
509 : }
510 :
511 199964 : 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 69046 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
521 : }
522 :
523 : /* now adjust if we wanted <> rather than = */
524 353432 : if (negate)
525 37730 : selec = 1.0 - selec - nullfrac;
526 :
527 : /* result should be in range, but make sure... */
528 353432 : CLAMP_PROBABILITY(selec);
529 :
530 353432 : 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 437030 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
540 : Node *other,
541 : bool varonleft, bool negate)
542 : {
543 : double selec;
544 437030 : double nullfrac = 0.0;
545 : bool isdefault;
546 :
547 : /*
548 : * Grab the nullfrac for use below.
549 : */
550 437030 : if (HeapTupleIsValid(vardata->statsTuple))
551 : {
552 : Form_pg_statistic stats;
553 :
554 300094 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
555 300094 : 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 437030 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
566 : {
567 164588 : selec = 1.0 / vardata->rel->tuples;
568 : }
569 272442 : 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 151796 : selec = 1.0 - nullfrac;
585 151796 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
586 151796 : if (ndistinct > 1)
587 147968 : selec /= ndistinct;
588 :
589 : /*
590 : * Cross-check: selectivity should never be estimated as more than the
591 : * most common value's.
592 : */
593 151796 : if (get_attstatsslot(&sslot, vardata->statsTuple,
594 : STATISTIC_KIND_MCV, InvalidOid,
595 : ATTSTATSSLOT_NUMBERS))
596 : {
597 132488 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
598 582 : selec = sslot.numbers[0];
599 132488 : 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 120646 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
610 : }
611 :
612 : /* now adjust if we wanted <> rather than = */
613 437030 : if (negate)
614 6590 : selec = 1.0 - selec - nullfrac;
615 :
616 : /* result should be in range, but make sure... */
617 437030 : CLAMP_PROBABILITY(selec);
618 :
619 437030 : 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 46714 : neqsel(PG_FUNCTION_ARGS)
631 : {
632 46714 : 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 379624 : 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 379624 : 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 28506 : if (vardata->var && IsA(vardata->var, Var) &&
672 23540 : ((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 2020 : if (vardata->rel->pages == 0)
683 0 : return 1.0;
684 :
685 2020 : itemptr = (ItemPointer) DatumGetPointer(constval);
686 2020 : 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 2020 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
696 :
697 : /* If target is the last page, use half the density. */
698 2020 : if (block >= vardata->rel->pages - 1)
699 30 : 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 2020 : if (density > 0.0)
710 : {
711 2020 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
712 :
713 2020 : 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 2020 : 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 2020 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
731 1880 : selec -= (1.0 / vardata->rel->tuples);
732 :
733 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
734 2020 : if (isgt)
735 1862 : selec = 1.0 - selec;
736 :
737 2020 : CLAMP_PROBABILITY(selec);
738 2020 : return selec;
739 : }
740 :
741 : /* no stats available, so default result */
742 26486 : return DEFAULT_INEQ_SEL;
743 : }
744 351118 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
745 :
746 351118 : 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 351118 : 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 351118 : 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 351118 : selec = 1.0 - stats->stanullfrac - sumcommon;
772 :
773 351118 : if (hist_selec >= 0.0)
774 219282 : 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 131836 : selec *= 0.5;
782 : }
783 :
784 351118 : selec += mcv_selec;
785 :
786 : /* result should be in range, but make sure... */
787 351118 : CLAMP_PROBABILITY(selec);
788 :
789 351118 : 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 357408 : 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 357408 : mcv_selec = 0.0;
815 357408 : sumcommon = 0.0;
816 :
817 712348 : if (HeapTupleIsValid(vardata->statsTuple) &&
818 709550 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
819 354610 : get_attstatsslot(&sslot, vardata->statsTuple,
820 : STATISTIC_KIND_MCV, InvalidOid,
821 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
822 : {
823 196240 : 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 196240 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
833 : NULL, NULL);
834 196240 : fcinfo->args[0].isnull = false;
835 196240 : fcinfo->args[1].isnull = false;
836 : /* be careful to apply operator right way 'round */
837 196240 : if (varonleft)
838 196240 : fcinfo->args[1].value = constval;
839 : else
840 0 : fcinfo->args[0].value = constval;
841 :
842 4683962 : for (i = 0; i < sslot.nvalues; i++)
843 : {
844 : Datum fresult;
845 :
846 4487722 : if (varonleft)
847 4487722 : fcinfo->args[0].value = sslot.values[i];
848 : else
849 0 : fcinfo->args[1].value = sslot.values[i];
850 4487722 : fcinfo->isnull = false;
851 4487722 : fresult = FunctionCallInvoke(fcinfo);
852 4487722 : if (!fcinfo->isnull && DatumGetBool(fresult))
853 1732724 : mcv_selec += sslot.numbers[i];
854 4487722 : sumcommon += sslot.numbers[i];
855 : }
856 196240 : free_attstatsslot(&sslot);
857 : }
858 :
859 357408 : *sumcommonp = sumcommon;
860 357408 : 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 6290 : 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 10112 : if (HeapTupleIsValid(vardata->statsTuple) &&
910 7638 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
911 3816 : get_attstatsslot(&sslot, vardata->statsTuple,
912 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
913 : ATTSTATSSLOT_VALUES))
914 : {
915 3722 : *hist_size = sslot.nvalues;
916 3722 : if (sslot.nvalues >= min_hist_size)
917 : {
918 1790 : LOCAL_FCINFO(fcinfo, 2);
919 1790 : 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 1790 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
931 : NULL, NULL);
932 1790 : fcinfo->args[0].isnull = false;
933 1790 : fcinfo->args[1].isnull = false;
934 : /* be careful to apply operator right way 'round */
935 1790 : if (varonleft)
936 1790 : fcinfo->args[1].value = constval;
937 : else
938 0 : fcinfo->args[0].value = constval;
939 :
940 146566 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
941 : {
942 : Datum fresult;
943 :
944 144776 : if (varonleft)
945 144776 : fcinfo->args[0].value = sslot.values[i];
946 : else
947 0 : fcinfo->args[1].value = sslot.values[i];
948 144776 : fcinfo->isnull = false;
949 144776 : fresult = FunctionCallInvoke(fcinfo);
950 144776 : if (!fcinfo->isnull && DatumGetBool(fresult))
951 9616 : nmatch++;
952 : }
953 1790 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
954 : }
955 : else
956 1932 : result = -1;
957 3722 : free_attstatsslot(&sslot);
958 : }
959 : else
960 : {
961 2568 : *hist_size = 0;
962 2568 : result = -1;
963 : }
964 :
965 6290 : 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 1130 : 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 1130 : 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 1130 : if (IsA(other, Const) &&
1009 1130 : ((Const *) other)->constisnull)
1010 : {
1011 0 : ReleaseVariableStats(vardata);
1012 0 : return 0.0;
1013 : }
1014 :
1015 1130 : if (IsA(other, Const))
1016 : {
1017 : /* Variable is being compared to a known non-null constant */
1018 1130 : Datum constval = ((Const *) other)->constvalue;
1019 : FmgrInfo opproc;
1020 : double mcvsum;
1021 : double mcvsel;
1022 : double nullfrac;
1023 : int hist_size;
1024 :
1025 1130 : fmgr_info(get_opcode(oproid), &opproc);
1026 :
1027 : /*
1028 : * Calculate the selectivity for the column's most common values.
1029 : */
1030 1130 : 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 1130 : selec = histogram_selectivity(&vardata, &opproc, collation,
1041 : constval, varonleft,
1042 : 10, 1, &hist_size);
1043 1130 : if (selec < 0)
1044 : {
1045 : /* Nope, fall back on default */
1046 1130 : 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 1130 : if (selec < 0.0001)
1063 0 : selec = 0.0001;
1064 1130 : else if (selec > 0.9999)
1065 0 : selec = 0.9999;
1066 :
1067 : /* Don't forget to account for nulls. */
1068 1130 : if (HeapTupleIsValid(vardata.statsTuple))
1069 84 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
1070 : else
1071 1046 : 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 1130 : selec *= 1.0 - nullfrac - mcvsum;
1079 1130 : 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 1130 : ReleaseVariableStats(vardata);
1088 :
1089 : /* result should be in range, but make sure... */
1090 1130 : CLAMP_PROBABILITY(selec);
1091 :
1092 1130 : 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 356530 : 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 356530 : 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 712358 : if (HeapTupleIsValid(vardata->statsTuple) &&
1138 711332 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1139 355504 : get_attstatsslot(&sslot, vardata->statsTuple,
1140 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1141 : ATTSTATSSLOT_VALUES))
1142 : {
1143 223990 : if (sslot.nvalues > 1 &&
1144 447904 : sslot.stacoll == collation &&
1145 223914 : comparison_ops_are_compatible(sslot.staop, opoid))
1146 223806 : {
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 223806 : int lobound = 0; /* first possible slot to search */
1167 223806 : int hibound = sslot.nvalues; /* last+1 slot to search */
1168 223806 : 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 223806 : if (sslot.nvalues == 2)
1177 2958 : have_end = get_actual_variable_range(root,
1178 : vardata,
1179 : sslot.staop,
1180 : collation,
1181 : &sslot.values[0],
1182 2958 : &sslot.values[1]);
1183 :
1184 1484066 : while (lobound < hibound)
1185 : {
1186 1260260 : 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 1260260 : if (probe == 0 && sslot.nvalues > 2)
1195 110346 : have_end = get_actual_variable_range(root,
1196 : vardata,
1197 : sslot.staop,
1198 : collation,
1199 : &sslot.values[0],
1200 : NULL);
1201 1149914 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1202 76252 : have_end = get_actual_variable_range(root,
1203 : vardata,
1204 : sslot.staop,
1205 : collation,
1206 : NULL,
1207 76252 : &sslot.values[probe]);
1208 :
1209 1260260 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1210 : collation,
1211 1260260 : sslot.values[probe],
1212 : constval));
1213 1260260 : if (isgt)
1214 69142 : ltcmp = !ltcmp;
1215 1260260 : if (ltcmp)
1216 475734 : lobound = probe + 1;
1217 : else
1218 784526 : hibound = probe;
1219 : }
1220 :
1221 223806 : 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 95610 : histfrac = 0.0;
1232 : }
1233 128196 : else if (lobound >= sslot.nvalues)
1234 : {
1235 : /*
1236 : * Inverse case: constant is above upper histogram boundary.
1237 : */
1238 38224 : histfrac = 1.0;
1239 : }
1240 : else
1241 : {
1242 : /* We have values[i-1] <= constant <= values[i]. */
1243 89972 : int i = lobound;
1244 89972 : 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 89972 : if (i == 1 || isgt == iseq)
1267 : {
1268 : double otherdistinct;
1269 : bool isdefault;
1270 : AttStatsSlot mcvslot;
1271 :
1272 : /* Get estimated number of distinct values */
1273 38862 : otherdistinct = get_variable_numdistinct(vardata,
1274 : &isdefault);
1275 :
1276 : /* Subtract off the number of known MCVs */
1277 38862 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1278 : STATISTIC_KIND_MCV, InvalidOid,
1279 : ATTSTATSSLOT_NUMBERS))
1280 : {
1281 4096 : otherdistinct -= mcvslot.nnumbers;
1282 4096 : free_attstatsslot(&mcvslot);
1283 : }
1284 :
1285 : /* If result doesn't seem sane, leave eq_selec at 0 */
1286 38862 : if (otherdistinct > 1)
1287 38820 : 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 89972 : if (convert_to_scalar(constval, consttype, collation,
1296 : &val,
1297 89972 : sslot.values[i - 1], sslot.values[i],
1298 : vardata->vartype,
1299 : &low, &high))
1300 : {
1301 89972 : if (high <= low)
1302 : {
1303 : /* cope if bin boundaries appear identical */
1304 0 : binfrac = 0.5;
1305 : }
1306 89972 : else if (val <= low)
1307 19872 : binfrac = 0.0;
1308 70100 : else if (val >= high)
1309 3200 : binfrac = 1.0;
1310 : else
1311 : {
1312 66900 : 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 66900 : if (isnan(binfrac) ||
1321 66900 : 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 89972 : histfrac = (double) (i - 1) + binfrac;
1344 89972 : 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 89972 : if (i == 1)
1379 16672 : 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 89972 : if (isgt == iseq)
1387 29502 : 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 223806 : 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 223806 : if (have_end)
1406 127314 : CLAMP_PROBABILITY(hist_selec);
1407 : else
1408 : {
1409 96492 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1410 :
1411 96492 : if (hist_selec < cutoff)
1412 34080 : hist_selec = cutoff;
1413 62412 : else if (hist_selec > 1.0 - cutoff)
1414 22674 : hist_selec = 1.0 - cutoff;
1415 : }
1416 : }
1417 184 : 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 184 : LOCAL_FCINFO(fcinfo, 2);
1427 184 : int nmatch = 0;
1428 :
1429 184 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1430 : NULL, NULL);
1431 184 : fcinfo->args[0].isnull = false;
1432 184 : fcinfo->args[1].isnull = false;
1433 184 : fcinfo->args[1].value = constval;
1434 962508 : for (int i = 0; i < sslot.nvalues; i++)
1435 : {
1436 : Datum fresult;
1437 :
1438 962324 : fcinfo->args[0].value = sslot.values[i];
1439 962324 : fcinfo->isnull = false;
1440 962324 : fresult = FunctionCallInvoke(fcinfo);
1441 962324 : if (!fcinfo->isnull && DatumGetBool(fresult))
1442 2228 : nmatch++;
1443 : }
1444 184 : 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 184 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1454 :
1455 184 : if (hist_selec < cutoff)
1456 12 : hist_selec = cutoff;
1457 172 : else if (hist_selec > 1.0 - cutoff)
1458 12 : hist_selec = 1.0 - cutoff;
1459 : }
1460 : }
1461 :
1462 223990 : free_attstatsslot(&sslot);
1463 : }
1464 :
1465 356530 : return hist_selec;
1466 : }
1467 :
1468 : /*
1469 : * Common wrapper function for the selectivity estimators that simply
1470 : * invoke scalarineqsel().
1471 : */
1472 : static Datum
1473 51096 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1474 : {
1475 51096 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1476 51096 : Oid operator = PG_GETARG_OID(1);
1477 51096 : List *args = (List *) PG_GETARG_POINTER(2);
1478 51096 : int varRelid = PG_GETARG_INT32(3);
1479 51096 : 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 51096 : if (!get_restriction_variable(root, args, varRelid,
1492 : &vardata, &other, &varonleft))
1493 644 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1494 :
1495 : /*
1496 : * Can't do anything useful if the something is not a constant, either.
1497 : */
1498 50452 : if (!IsA(other, Const))
1499 : {
1500 2842 : ReleaseVariableStats(vardata);
1501 2842 : 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 47610 : if (((Const *) other)->constisnull)
1509 : {
1510 66 : ReleaseVariableStats(vardata);
1511 66 : PG_RETURN_FLOAT8(0.0);
1512 : }
1513 47544 : constval = ((Const *) other)->constvalue;
1514 47544 : consttype = ((Const *) other)->consttype;
1515 :
1516 : /*
1517 : * Force the var to be on the left to simplify logic in scalarineqsel.
1518 : */
1519 47544 : if (!varonleft)
1520 : {
1521 384 : operator = get_commutator(operator);
1522 384 : 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 384 : isgt = !isgt;
1529 : }
1530 :
1531 : /* The rest of the work is done by scalarineqsel(). */
1532 47544 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1533 : &vardata, constval, consttype);
1534 :
1535 47544 : ReleaseVariableStats(vardata);
1536 :
1537 47544 : PG_RETURN_FLOAT8((float8) selec);
1538 : }
1539 :
1540 : /*
1541 : * scalarltsel - Selectivity of "<" for scalars.
1542 : */
1543 : Datum
1544 15206 : scalarltsel(PG_FUNCTION_ARGS)
1545 : {
1546 15206 : return scalarineqsel_wrapper(fcinfo, false, false);
1547 : }
1548 :
1549 : /*
1550 : * scalarlesel - Selectivity of "<=" for scalars.
1551 : */
1552 : Datum
1553 4636 : scalarlesel(PG_FUNCTION_ARGS)
1554 : {
1555 4636 : return scalarineqsel_wrapper(fcinfo, false, true);
1556 : }
1557 :
1558 : /*
1559 : * scalargtsel - Selectivity of ">" for scalars.
1560 : */
1561 : Datum
1562 15660 : scalargtsel(PG_FUNCTION_ARGS)
1563 : {
1564 15660 : return scalarineqsel_wrapper(fcinfo, true, false);
1565 : }
1566 :
1567 : /*
1568 : * scalargesel - Selectivity of ">=" for scalars.
1569 : */
1570 : Datum
1571 15594 : scalargesel(PG_FUNCTION_ARGS)
1572 : {
1573 15594 : 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 56898 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1586 : {
1587 : VariableStatData vardata;
1588 : double selec;
1589 :
1590 56898 : examine_variable(root, arg, varRelid, &vardata);
1591 56898 : 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 35900 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1598 : BoolGetDatum(true), false, true, false);
1599 : }
1600 20998 : 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 12460 : selec = 0.3333333;
1610 : }
1611 : else
1612 : {
1613 : /* Otherwise, the default estimate is 0.5 */
1614 8538 : selec = 0.5;
1615 : }
1616 56898 : ReleaseVariableStats(vardata);
1617 56898 : return selec;
1618 : }
1619 :
1620 : /*
1621 : * booltestsel - Selectivity of BooleanTest Node.
1622 : */
1623 : Selectivity
1624 902 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1625 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1626 : {
1627 : VariableStatData vardata;
1628 : double selec;
1629 :
1630 902 : examine_variable(root, arg, varRelid, &vardata);
1631 :
1632 902 : if (HeapTupleIsValid(vardata.statsTuple))
1633 : {
1634 : Form_pg_statistic stats;
1635 : double freq_null;
1636 : AttStatsSlot sslot;
1637 :
1638 12 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1639 12 : freq_null = stats->stanullfrac;
1640 :
1641 12 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1642 : STATISTIC_KIND_MCV, InvalidOid,
1643 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1644 12 : && sslot.nnumbers > 0)
1645 12 : {
1646 : double freq_true;
1647 : double freq_false;
1648 :
1649 : /*
1650 : * Get first MCV frequency and derive frequency for true.
1651 : */
1652 12 : if (DatumGetBool(sslot.values[0]))
1653 0 : freq_true = sslot.numbers[0];
1654 : else
1655 12 : 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 12 : freq_false = 1.0 - freq_true - freq_null;
1662 :
1663 12 : 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 12 : case IS_TRUE:
1674 : /* select only TRUE values */
1675 12 : selec = freq_true;
1676 12 : 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 12 : 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 0 : switch (booltesttype)
1706 : {
1707 0 : case IS_UNKNOWN:
1708 : /* select only NULL values */
1709 0 : selec = freq_null;
1710 0 : 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 890 : switch (booltesttype)
1743 : {
1744 48 : case IS_UNKNOWN:
1745 48 : selec = DEFAULT_UNK_SEL;
1746 48 : break;
1747 108 : case IS_NOT_UNKNOWN:
1748 108 : selec = DEFAULT_NOT_UNK_SEL;
1749 108 : break;
1750 252 : case IS_TRUE:
1751 : case IS_NOT_FALSE:
1752 252 : selec = (double) clause_selectivity(root, arg,
1753 : varRelid,
1754 : jointype, sjinfo);
1755 252 : break;
1756 482 : case IS_FALSE:
1757 : case IS_NOT_TRUE:
1758 482 : selec = 1.0 - (double) clause_selectivity(root, arg,
1759 : varRelid,
1760 : jointype, sjinfo);
1761 482 : 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 902 : ReleaseVariableStats(vardata);
1771 :
1772 : /* result should be in range, but make sure... */
1773 902 : CLAMP_PROBABILITY(selec);
1774 :
1775 902 : return (Selectivity) selec;
1776 : }
1777 :
1778 : /*
1779 : * nulltestsel - Selectivity of NullTest Node.
1780 : */
1781 : Selectivity
1782 17516 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1783 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1784 : {
1785 : VariableStatData vardata;
1786 : double selec;
1787 :
1788 17516 : examine_variable(root, arg, varRelid, &vardata);
1789 :
1790 17516 : if (HeapTupleIsValid(vardata.statsTuple))
1791 : {
1792 : Form_pg_statistic stats;
1793 : double freq_null;
1794 :
1795 9782 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1796 9782 : freq_null = stats->stanullfrac;
1797 :
1798 9782 : switch (nulltesttype)
1799 : {
1800 7260 : case IS_NULL:
1801 :
1802 : /*
1803 : * Use freq_null directly.
1804 : */
1805 7260 : selec = freq_null;
1806 7260 : break;
1807 2522 : case IS_NOT_NULL:
1808 :
1809 : /*
1810 : * Select not unknown (not null) values. Calculate from
1811 : * freq_null.
1812 : */
1813 2522 : selec = 1.0 - freq_null;
1814 2522 : break;
1815 0 : default:
1816 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1817 : (int) nulltesttype);
1818 : return (Selectivity) 0; /* keep compiler quiet */
1819 : }
1820 : }
1821 7734 : else if (vardata.var && IsA(vardata.var, Var) &&
1822 6974 : ((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 104 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1829 : }
1830 : else
1831 : {
1832 : /*
1833 : * No ANALYZE stats available, so make a guess
1834 : */
1835 7630 : switch (nulltesttype)
1836 : {
1837 2098 : case IS_NULL:
1838 2098 : selec = DEFAULT_UNK_SEL;
1839 2098 : break;
1840 5532 : case IS_NOT_NULL:
1841 5532 : selec = DEFAULT_NOT_UNK_SEL;
1842 5532 : 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 17516 : ReleaseVariableStats(vardata);
1851 :
1852 : /* result should be in range, but make sure... */
1853 17516 : CLAMP_PROBABILITY(selec);
1854 :
1855 17516 : 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 130888 : strip_array_coercion(Node *node)
1868 : {
1869 : for (;;)
1870 : {
1871 131000 : if (node && IsA(node, ArrayCoerceExpr))
1872 112 : {
1873 3022 : 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 3022 : if (IsA(acoerce->elemexpr, RelabelType) &&
1880 112 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1881 112 : node = (Node *) acoerce->arg;
1882 : else
1883 : break;
1884 : }
1885 127978 : 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 130888 : return node;
1894 : }
1895 :
1896 : /*
1897 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1898 : */
1899 : Selectivity
1900 22960 : scalararraysel(PlannerInfo *root,
1901 : ScalarArrayOpExpr *clause,
1902 : bool is_join_clause,
1903 : int varRelid,
1904 : JoinType jointype,
1905 : SpecialJoinInfo *sjinfo)
1906 : {
1907 22960 : Oid operator = clause->opno;
1908 22960 : bool useOr = clause->useOr;
1909 22960 : bool isEquality = false;
1910 22960 : 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 22960 : leftop = (Node *) linitial(clause->args);
1924 22960 : rightop = (Node *) lsecond(clause->args);
1925 :
1926 : /* aggressively reduce both sides to constants */
1927 22960 : leftop = estimate_expression_value(root, leftop);
1928 22960 : rightop = estimate_expression_value(root, rightop);
1929 :
1930 : /* get nominal (after relabeling) element type of rightop */
1931 22960 : nominal_element_type = get_base_element_type(exprType(rightop));
1932 22960 : if (!OidIsValid(nominal_element_type))
1933 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1934 : /* get nominal collation, too, for generating constants */
1935 22960 : nominal_element_collation = exprCollation(rightop);
1936 :
1937 : /* look through any binary-compatible relabeling of rightop */
1938 22960 : 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 22960 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1945 22960 : if (OidIsValid(typentry->eq_opr))
1946 : {
1947 22956 : if (operator == typentry->eq_opr)
1948 19590 : isEquality = true;
1949 3366 : else if (get_negator(operator) == typentry->eq_opr)
1950 2800 : 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 22960 : if ((isEquality || isInequality) && !is_join_clause)
1960 : {
1961 22390 : s1 = scalararraysel_containment(root, leftop, rightop,
1962 : nominal_element_type,
1963 : isEquality, useOr, varRelid);
1964 22390 : if (s1 >= 0.0)
1965 118 : return s1;
1966 : }
1967 :
1968 : /*
1969 : * Look up the underlying operator's selectivity estimator. Punt if it
1970 : * hasn't got one.
1971 : */
1972 22842 : if (is_join_clause)
1973 0 : oprsel = get_oprjoin(operator);
1974 : else
1975 22842 : oprsel = get_oprrest(operator);
1976 22842 : if (!oprsel)
1977 4 : return (Selectivity) 0.5;
1978 22838 : 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 22838 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1989 19668 : isEquality = true;
1990 3170 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1991 2690 : 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 22838 : if (rightop && IsA(rightop, Const))
2006 18496 : {
2007 18526 : Datum arraydatum = ((Const *) rightop)->constvalue;
2008 18526 : 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 18526 : if (arrayisnull) /* qual can't succeed if null array */
2019 30 : return (Selectivity) 0.0;
2020 18496 : arrayval = DatumGetArrayTypeP(arraydatum);
2021 18496 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
2022 : &elmlen, &elmbyval, &elmalign);
2023 18496 : deconstruct_array(arrayval,
2024 : ARR_ELEMTYPE(arrayval),
2025 : elmlen, elmbyval, elmalign,
2026 : &elem_values, &elem_nulls, &num_elems);
2027 :
2028 : /*
2029 : * For generic operators, we assume the probability of success is
2030 : * independent for each array element. But for "= ANY" or "<> ALL",
2031 : * if the array elements are distinct (which'd typically be the case)
2032 : * then the probabilities are disjoint, and we should just sum them.
2033 : *
2034 : * If we were being really tense we would try to confirm that the
2035 : * elements are all distinct, but that would be expensive and it
2036 : * doesn't seem to be worth the cycles; it would amount to penalizing
2037 : * well-written queries in favor of poorly-written ones. However, we
2038 : * do protect ourselves a little bit by checking whether the
2039 : * disjointness assumption leads to an impossible (out of range)
2040 : * probability; if so, we fall back to the normal calculation.
2041 : */
2042 18496 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2043 :
2044 78642 : for (i = 0; i < num_elems; i++)
2045 : {
2046 : List *args;
2047 : Selectivity s2;
2048 :
2049 60146 : args = list_make2(leftop,
2050 : makeConst(nominal_element_type,
2051 : -1,
2052 : nominal_element_collation,
2053 : elmlen,
2054 : elem_values[i],
2055 : elem_nulls[i],
2056 : elmbyval));
2057 60146 : if (is_join_clause)
2058 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2059 : clause->inputcollid,
2060 : PointerGetDatum(root),
2061 : ObjectIdGetDatum(operator),
2062 : PointerGetDatum(args),
2063 : Int16GetDatum(jointype),
2064 : PointerGetDatum(sjinfo)));
2065 : else
2066 60146 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2067 : clause->inputcollid,
2068 : PointerGetDatum(root),
2069 : ObjectIdGetDatum(operator),
2070 : PointerGetDatum(args),
2071 : Int32GetDatum(varRelid)));
2072 :
2073 60146 : if (useOr)
2074 : {
2075 51464 : s1 = s1 + s2 - s1 * s2;
2076 51464 : if (isEquality)
2077 50420 : s1disjoint += s2;
2078 : }
2079 : else
2080 : {
2081 8682 : s1 = s1 * s2;
2082 8682 : if (isInequality)
2083 8370 : s1disjoint += s2 - 1.0;
2084 : }
2085 : }
2086 :
2087 : /* accept disjoint-probability estimate if in range */
2088 18496 : if ((useOr ? isEquality : isInequality) &&
2089 17866 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2090 17836 : s1 = s1disjoint;
2091 : }
2092 4312 : else if (rightop && IsA(rightop, ArrayExpr) &&
2093 384 : !((ArrayExpr *) rightop)->multidims)
2094 384 : {
2095 384 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2096 : int16 elmlen;
2097 : bool elmbyval;
2098 : ListCell *l;
2099 :
2100 384 : get_typlenbyval(arrayexpr->element_typeid,
2101 : &elmlen, &elmbyval);
2102 :
2103 : /*
2104 : * We use the assumption of disjoint probabilities here too, although
2105 : * the odds of equal array elements are rather higher if the elements
2106 : * are not all constants (which they won't be, else constant folding
2107 : * would have reduced the ArrayExpr to a Const). In this path it's
2108 : * critical to have the sanity check on the s1disjoint estimate.
2109 : */
2110 384 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2111 :
2112 1420 : foreach(l, arrayexpr->elements)
2113 : {
2114 1036 : Node *elem = (Node *) lfirst(l);
2115 : List *args;
2116 : Selectivity s2;
2117 :
2118 : /*
2119 : * Theoretically, if elem isn't of nominal_element_type we should
2120 : * insert a RelabelType, but it seems unlikely that any operator
2121 : * estimation function would really care ...
2122 : */
2123 1036 : args = list_make2(leftop, elem);
2124 1036 : if (is_join_clause)
2125 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2126 : clause->inputcollid,
2127 : PointerGetDatum(root),
2128 : ObjectIdGetDatum(operator),
2129 : PointerGetDatum(args),
2130 : Int16GetDatum(jointype),
2131 : PointerGetDatum(sjinfo)));
2132 : else
2133 1036 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2134 : clause->inputcollid,
2135 : PointerGetDatum(root),
2136 : ObjectIdGetDatum(operator),
2137 : PointerGetDatum(args),
2138 : Int32GetDatum(varRelid)));
2139 :
2140 1036 : if (useOr)
2141 : {
2142 1036 : s1 = s1 + s2 - s1 * s2;
2143 1036 : if (isEquality)
2144 1036 : s1disjoint += s2;
2145 : }
2146 : else
2147 : {
2148 0 : s1 = s1 * s2;
2149 0 : if (isInequality)
2150 0 : s1disjoint += s2 - 1.0;
2151 : }
2152 : }
2153 :
2154 : /* accept disjoint-probability estimate if in range */
2155 384 : if ((useOr ? isEquality : isInequality) &&
2156 384 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2157 384 : s1 = s1disjoint;
2158 : }
2159 : else
2160 : {
2161 : CaseTestExpr *dummyexpr;
2162 : List *args;
2163 : Selectivity s2;
2164 : int i;
2165 :
2166 : /*
2167 : * We need a dummy rightop to pass to the operator selectivity
2168 : * routine. It can be pretty much anything that doesn't look like a
2169 : * constant; CaseTestExpr is a convenient choice.
2170 : */
2171 3928 : dummyexpr = makeNode(CaseTestExpr);
2172 3928 : dummyexpr->typeId = nominal_element_type;
2173 3928 : dummyexpr->typeMod = -1;
2174 3928 : dummyexpr->collation = clause->inputcollid;
2175 3928 : args = list_make2(leftop, dummyexpr);
2176 3928 : if (is_join_clause)
2177 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2178 : clause->inputcollid,
2179 : PointerGetDatum(root),
2180 : ObjectIdGetDatum(operator),
2181 : PointerGetDatum(args),
2182 : Int16GetDatum(jointype),
2183 : PointerGetDatum(sjinfo)));
2184 : else
2185 3928 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2186 : clause->inputcollid,
2187 : PointerGetDatum(root),
2188 : ObjectIdGetDatum(operator),
2189 : PointerGetDatum(args),
2190 : Int32GetDatum(varRelid)));
2191 3928 : s1 = useOr ? 0.0 : 1.0;
2192 :
2193 : /*
2194 : * Arbitrarily assume 10 elements in the eventual array value (see
2195 : * also estimate_array_length). We don't risk an assumption of
2196 : * disjoint probabilities here.
2197 : */
2198 43208 : for (i = 0; i < 10; i++)
2199 : {
2200 39280 : if (useOr)
2201 39280 : s1 = s1 + s2 - s1 * s2;
2202 : else
2203 0 : s1 = s1 * s2;
2204 : }
2205 : }
2206 :
2207 : /* result should be in range, but make sure... */
2208 22808 : CLAMP_PROBABILITY(s1);
2209 :
2210 22808 : return s1;
2211 : }
2212 :
2213 : /*
2214 : * Estimate number of elements in the array yielded by an expression.
2215 : *
2216 : * Note: the result is integral, but we use "double" to avoid overflow
2217 : * concerns. Most callers will use it in double-type expressions anyway.
2218 : *
2219 : * Note: in some code paths root can be passed as NULL, resulting in
2220 : * slightly worse estimates.
2221 : */
2222 : double
2223 107928 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2224 : {
2225 : /* look through any binary-compatible relabeling of arrayexpr */
2226 107928 : arrayexpr = strip_array_coercion(arrayexpr);
2227 :
2228 107928 : if (arrayexpr && IsA(arrayexpr, Const))
2229 : {
2230 48564 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2231 48564 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2232 : ArrayType *arrayval;
2233 :
2234 48564 : if (arrayisnull)
2235 90 : return 0;
2236 48474 : arrayval = DatumGetArrayTypeP(arraydatum);
2237 48474 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2238 : }
2239 59364 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2240 680 : !((ArrayExpr *) arrayexpr)->multidims)
2241 : {
2242 680 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2243 : }
2244 58684 : else if (arrayexpr && root)
2245 : {
2246 : /* See if we can find any statistics about it */
2247 : VariableStatData vardata;
2248 : AttStatsSlot sslot;
2249 58660 : double nelem = 0;
2250 :
2251 58660 : examine_variable(root, arrayexpr, 0, &vardata);
2252 58660 : if (HeapTupleIsValid(vardata.statsTuple))
2253 : {
2254 : /*
2255 : * Found stats, so use the average element count, which is stored
2256 : * in the last stanumbers element of the DECHIST statistics.
2257 : * Actually that is the average count of *distinct* elements;
2258 : * perhaps we should scale it up somewhat?
2259 : */
2260 13556 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2261 : STATISTIC_KIND_DECHIST, InvalidOid,
2262 : ATTSTATSSLOT_NUMBERS))
2263 : {
2264 13442 : if (sslot.nnumbers > 0)
2265 13442 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2266 13442 : free_attstatsslot(&sslot);
2267 : }
2268 : }
2269 58660 : ReleaseVariableStats(vardata);
2270 :
2271 58660 : if (nelem > 0)
2272 13442 : return nelem;
2273 : }
2274 :
2275 : /* Else use a default guess --- this should match scalararraysel */
2276 45242 : return 10;
2277 : }
2278 :
2279 : /*
2280 : * rowcomparesel - Selectivity of RowCompareExpr Node.
2281 : *
2282 : * We estimate RowCompare selectivity by considering just the first (high
2283 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2284 : * this estimate could be refined by considering additional columns, it
2285 : * seems unlikely that we could do a lot better without multi-column
2286 : * statistics.
2287 : */
2288 : Selectivity
2289 252 : rowcomparesel(PlannerInfo *root,
2290 : RowCompareExpr *clause,
2291 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2292 : {
2293 : Selectivity s1;
2294 252 : Oid opno = linitial_oid(clause->opnos);
2295 252 : Oid inputcollid = linitial_oid(clause->inputcollids);
2296 : List *opargs;
2297 : bool is_join_clause;
2298 :
2299 : /* Build equivalent arg list for single operator */
2300 252 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2301 :
2302 : /*
2303 : * Decide if it's a join clause. This should match clausesel.c's
2304 : * treat_as_join_clause(), except that we intentionally consider only the
2305 : * leading columns and not the rest of the clause.
2306 : */
2307 252 : if (varRelid != 0)
2308 : {
2309 : /*
2310 : * Caller is forcing restriction mode (eg, because we are examining an
2311 : * inner indexscan qual).
2312 : */
2313 54 : is_join_clause = false;
2314 : }
2315 198 : else if (sjinfo == NULL)
2316 : {
2317 : /*
2318 : * It must be a restriction clause, since it's being evaluated at a
2319 : * scan node.
2320 : */
2321 186 : is_join_clause = false;
2322 : }
2323 : else
2324 : {
2325 : /*
2326 : * Otherwise, it's a join if there's more than one base relation used.
2327 : */
2328 12 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2329 : }
2330 :
2331 252 : if (is_join_clause)
2332 : {
2333 : /* Estimate selectivity for a join clause. */
2334 12 : s1 = join_selectivity(root, opno,
2335 : opargs,
2336 : inputcollid,
2337 : jointype,
2338 : sjinfo);
2339 : }
2340 : else
2341 : {
2342 : /* Estimate selectivity for a restriction clause. */
2343 240 : s1 = restriction_selectivity(root, opno,
2344 : opargs,
2345 : inputcollid,
2346 : varRelid);
2347 : }
2348 :
2349 252 : return s1;
2350 : }
2351 :
2352 : /*
2353 : * eqjoinsel - Join selectivity of "="
2354 : */
2355 : Datum
2356 268514 : eqjoinsel(PG_FUNCTION_ARGS)
2357 : {
2358 268514 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2359 268514 : Oid operator = PG_GETARG_OID(1);
2360 268514 : List *args = (List *) PG_GETARG_POINTER(2);
2361 :
2362 : #ifdef NOT_USED
2363 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2364 : #endif
2365 268514 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2366 268514 : Oid collation = PG_GET_COLLATION();
2367 : double selec;
2368 : double selec_inner;
2369 : VariableStatData vardata1;
2370 : VariableStatData vardata2;
2371 : double nd1;
2372 : double nd2;
2373 : bool isdefault1;
2374 : bool isdefault2;
2375 : Oid opfuncoid;
2376 : FmgrInfo eqproc;
2377 268514 : Oid hashLeft = InvalidOid;
2378 268514 : Oid hashRight = InvalidOid;
2379 : AttStatsSlot sslot1;
2380 : AttStatsSlot sslot2;
2381 268514 : Form_pg_statistic stats1 = NULL;
2382 268514 : Form_pg_statistic stats2 = NULL;
2383 268514 : bool have_mcvs1 = false;
2384 268514 : bool have_mcvs2 = false;
2385 268514 : bool *hasmatch1 = NULL;
2386 268514 : bool *hasmatch2 = NULL;
2387 268514 : int nmatches = 0;
2388 : bool get_mcv_stats;
2389 : bool join_is_reversed;
2390 : RelOptInfo *inner_rel;
2391 :
2392 268514 : get_join_variables(root, args, sjinfo,
2393 : &vardata1, &vardata2, &join_is_reversed);
2394 :
2395 268514 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2396 268514 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2397 :
2398 268514 : opfuncoid = get_opcode(operator);
2399 :
2400 268514 : memset(&sslot1, 0, sizeof(sslot1));
2401 268514 : memset(&sslot2, 0, sizeof(sslot2));
2402 :
2403 : /*
2404 : * There is no use in fetching one side's MCVs if we lack MCVs for the
2405 : * other side, so do a quick check to verify that both stats exist.
2406 : */
2407 740870 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2408 363286 : HeapTupleIsValid(vardata2.statsTuple) &&
2409 159444 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2410 : STATISTIC_KIND_MCV, InvalidOid,
2411 472356 : 0) &&
2412 72834 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2413 : STATISTIC_KIND_MCV, InvalidOid,
2414 : 0));
2415 :
2416 268514 : if (HeapTupleIsValid(vardata1.statsTuple))
2417 : {
2418 : /* note we allow use of nullfrac regardless of security check */
2419 203842 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2420 232382 : if (get_mcv_stats &&
2421 28540 : statistic_proc_security_check(&vardata1, opfuncoid))
2422 28540 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2423 : STATISTIC_KIND_MCV, InvalidOid,
2424 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2425 : }
2426 :
2427 268514 : if (HeapTupleIsValid(vardata2.statsTuple))
2428 : {
2429 : /* note we allow use of nullfrac regardless of security check */
2430 180130 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2431 208670 : if (get_mcv_stats &&
2432 28540 : statistic_proc_security_check(&vardata2, opfuncoid))
2433 28540 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2434 : STATISTIC_KIND_MCV, InvalidOid,
2435 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2436 : }
2437 :
2438 : /* Prepare info usable by both eqjoinsel_inner and eqjoinsel_semi */
2439 268514 : if (have_mcvs1 && have_mcvs2)
2440 : {
2441 28540 : fmgr_info(opfuncoid, &eqproc);
2442 28540 : hasmatch1 = (bool *) palloc0(sslot1.nvalues * sizeof(bool));
2443 28540 : hasmatch2 = (bool *) palloc0(sslot2.nvalues * sizeof(bool));
2444 :
2445 : /*
2446 : * If the MCV lists are long enough to justify hashing, try to look up
2447 : * hash functions for the join operator.
2448 : */
2449 28540 : if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
2450 1668 : (void) get_op_hash_functions(operator, &hashLeft, &hashRight);
2451 : }
2452 : else
2453 239974 : memset(&eqproc, 0, sizeof(eqproc)); /* silence uninit-var warnings */
2454 :
2455 : /* We need to compute the inner-join selectivity in all cases */
2456 268514 : selec_inner = eqjoinsel_inner(&eqproc, collation,
2457 : hashLeft, hashRight,
2458 : &vardata1, &vardata2,
2459 : nd1, nd2,
2460 : isdefault1, isdefault2,
2461 : &sslot1, &sslot2,
2462 : stats1, stats2,
2463 : have_mcvs1, have_mcvs2,
2464 : hasmatch1, hasmatch2,
2465 : &nmatches);
2466 :
2467 268514 : switch (sjinfo->jointype)
2468 : {
2469 257630 : case JOIN_INNER:
2470 : case JOIN_LEFT:
2471 : case JOIN_FULL:
2472 257630 : selec = selec_inner;
2473 257630 : break;
2474 10884 : case JOIN_SEMI:
2475 : case JOIN_ANTI:
2476 :
2477 : /*
2478 : * Look up the join's inner relation. min_righthand is sufficient
2479 : * information because neither SEMI nor ANTI joins permit any
2480 : * reassociation into or out of their RHS, so the righthand will
2481 : * always be exactly that set of rels.
2482 : */
2483 10884 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2484 :
2485 10884 : if (!join_is_reversed)
2486 6734 : selec = eqjoinsel_semi(&eqproc, collation,
2487 : hashLeft, hashRight,
2488 : false,
2489 : &vardata1, &vardata2,
2490 : nd1, nd2,
2491 : isdefault1, isdefault2,
2492 : &sslot1, &sslot2,
2493 : stats1, stats2,
2494 : have_mcvs1, have_mcvs2,
2495 : hasmatch1, hasmatch2,
2496 : &nmatches,
2497 : inner_rel);
2498 : else
2499 4150 : selec = eqjoinsel_semi(&eqproc, collation,
2500 : hashLeft, hashRight,
2501 : true,
2502 : &vardata2, &vardata1,
2503 : nd2, nd1,
2504 : isdefault2, isdefault1,
2505 : &sslot2, &sslot1,
2506 : stats2, stats1,
2507 : have_mcvs2, have_mcvs1,
2508 : hasmatch2, hasmatch1,
2509 : &nmatches,
2510 : inner_rel);
2511 :
2512 : /*
2513 : * We should never estimate the output of a semijoin to be more
2514 : * rows than we estimate for an inner join with the same input
2515 : * rels and join condition; it's obviously impossible for that to
2516 : * happen. The former estimate is N1 * Ssemi while the latter is
2517 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2518 : * this is worthwhile because of the shakier estimation rules we
2519 : * use in eqjoinsel_semi, particularly in cases where it has to
2520 : * punt entirely.
2521 : */
2522 10884 : selec = Min(selec, inner_rel->rows * selec_inner);
2523 10884 : break;
2524 0 : default:
2525 : /* other values not expected here */
2526 0 : elog(ERROR, "unrecognized join type: %d",
2527 : (int) sjinfo->jointype);
2528 : selec = 0; /* keep compiler quiet */
2529 : break;
2530 : }
2531 :
2532 268514 : free_attstatsslot(&sslot1);
2533 268514 : free_attstatsslot(&sslot2);
2534 :
2535 268514 : ReleaseVariableStats(vardata1);
2536 268514 : ReleaseVariableStats(vardata2);
2537 :
2538 268514 : if (hasmatch1)
2539 28540 : pfree(hasmatch1);
2540 268514 : if (hasmatch2)
2541 28540 : pfree(hasmatch2);
2542 :
2543 268514 : CLAMP_PROBABILITY(selec);
2544 :
2545 268514 : PG_RETURN_FLOAT8((float8) selec);
2546 : }
2547 :
2548 : /*
2549 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2550 : *
2551 : * In addition to computing the selectivity estimate, this will fill
2552 : * hasmatch1[], hasmatch2[], and *p_nmatches (if have_mcvs1 && have_mcvs2).
2553 : * We may be able to re-use that data in eqjoinsel_semi.
2554 : *
2555 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2556 : * that it's worth trying to distinguish them here.
2557 : */
2558 : static double
2559 268514 : eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
2560 : Oid hashLeft, Oid hashRight,
2561 : VariableStatData *vardata1, VariableStatData *vardata2,
2562 : double nd1, double nd2,
2563 : bool isdefault1, bool isdefault2,
2564 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2565 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2566 : bool have_mcvs1, bool have_mcvs2,
2567 : bool *hasmatch1, bool *hasmatch2,
2568 : int *p_nmatches)
2569 : {
2570 : double selec;
2571 :
2572 268514 : if (have_mcvs1 && have_mcvs2)
2573 28540 : {
2574 : /*
2575 : * We have most-common-value lists for both relations. Run through
2576 : * the lists to see which MCVs actually join to each other with the
2577 : * given operator. This allows us to determine the exact join
2578 : * selectivity for the portion of the relations represented by the MCV
2579 : * lists. We still have to estimate for the remaining population, but
2580 : * in a skewed distribution this gives us a big leg up in accuracy.
2581 : * For motivation see the analysis in Y. Ioannidis and S.
2582 : * Christodoulakis, "On the propagation of errors in the size of join
2583 : * results", Technical Report 1018, Computer Science Dept., University
2584 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2585 : */
2586 28540 : double nullfrac1 = stats1->stanullfrac;
2587 28540 : double nullfrac2 = stats2->stanullfrac;
2588 : double matchprodfreq,
2589 : matchfreq1,
2590 : matchfreq2,
2591 : unmatchfreq1,
2592 : unmatchfreq2,
2593 : otherfreq1,
2594 : otherfreq2,
2595 : totalsel1,
2596 : totalsel2;
2597 : int i,
2598 : nmatches;
2599 :
2600 : /* Fill the match arrays */
2601 28540 : eqjoinsel_find_matches(eqproc, collation,
2602 : hashLeft, hashRight,
2603 : false,
2604 : sslot1, sslot2,
2605 : sslot1->nvalues, sslot2->nvalues,
2606 : hasmatch1, hasmatch2,
2607 : p_nmatches, &matchprodfreq);
2608 28540 : nmatches = *p_nmatches;
2609 28540 : CLAMP_PROBABILITY(matchprodfreq);
2610 :
2611 : /* Sum up frequencies of matched and unmatched MCVs */
2612 28540 : matchfreq1 = unmatchfreq1 = 0.0;
2613 706610 : for (i = 0; i < sslot1->nvalues; i++)
2614 : {
2615 678070 : if (hasmatch1[i])
2616 294418 : matchfreq1 += sslot1->numbers[i];
2617 : else
2618 383652 : unmatchfreq1 += sslot1->numbers[i];
2619 : }
2620 28540 : CLAMP_PROBABILITY(matchfreq1);
2621 28540 : CLAMP_PROBABILITY(unmatchfreq1);
2622 28540 : matchfreq2 = unmatchfreq2 = 0.0;
2623 534640 : for (i = 0; i < sslot2->nvalues; i++)
2624 : {
2625 506100 : if (hasmatch2[i])
2626 294418 : matchfreq2 += sslot2->numbers[i];
2627 : else
2628 211682 : unmatchfreq2 += sslot2->numbers[i];
2629 : }
2630 28540 : CLAMP_PROBABILITY(matchfreq2);
2631 28540 : CLAMP_PROBABILITY(unmatchfreq2);
2632 :
2633 : /*
2634 : * Compute total frequency of non-null values that are not in the MCV
2635 : * lists.
2636 : */
2637 28540 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2638 28540 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2639 28540 : CLAMP_PROBABILITY(otherfreq1);
2640 28540 : CLAMP_PROBABILITY(otherfreq2);
2641 :
2642 : /*
2643 : * We can estimate the total selectivity from the point of view of
2644 : * relation 1 as: the known selectivity for matched MCVs, plus
2645 : * unmatched MCVs that are assumed to match against random members of
2646 : * relation 2's non-MCV population, plus non-MCV values that are
2647 : * assumed to match against random members of relation 2's unmatched
2648 : * MCVs plus non-MCV values.
2649 : */
2650 28540 : totalsel1 = matchprodfreq;
2651 28540 : if (nd2 > sslot2->nvalues)
2652 6234 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2653 28540 : if (nd2 > nmatches)
2654 11176 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2655 11176 : (nd2 - nmatches);
2656 : /* Same estimate from the point of view of relation 2. */
2657 28540 : totalsel2 = matchprodfreq;
2658 28540 : if (nd1 > sslot1->nvalues)
2659 7032 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2660 28540 : if (nd1 > nmatches)
2661 9896 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2662 9896 : (nd1 - nmatches);
2663 :
2664 : /*
2665 : * Use the smaller of the two estimates. This can be justified in
2666 : * essentially the same terms as given below for the no-stats case: to
2667 : * a first approximation, we are estimating from the point of view of
2668 : * the relation with smaller nd.
2669 : */
2670 28540 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2671 : }
2672 : else
2673 : {
2674 : /*
2675 : * We do not have MCV lists for both sides. Estimate the join
2676 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2677 : * is plausible if we assume that the join operator is strict and the
2678 : * non-null values are about equally distributed: a given non-null
2679 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2680 : * of rel2, so total join rows are at most
2681 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2682 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2683 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2684 : * with MIN() is an upper bound. Using the MIN() means we estimate
2685 : * from the point of view of the relation with smaller nd (since the
2686 : * larger nd is determining the MIN). It is reasonable to assume that
2687 : * most tuples in this rel will have join partners, so the bound is
2688 : * probably reasonably tight and should be taken as-is.
2689 : *
2690 : * XXX Can we be smarter if we have an MCV list for just one side? It
2691 : * seems that if we assume equal distribution for the other side, we
2692 : * end up with the same answer anyway.
2693 : */
2694 239974 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2695 239974 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2696 :
2697 239974 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2698 239974 : if (nd1 > nd2)
2699 127206 : selec /= nd1;
2700 : else
2701 112768 : selec /= nd2;
2702 : }
2703 :
2704 268514 : return selec;
2705 : }
2706 :
2707 : /*
2708 : * eqjoinsel_semi --- eqjoinsel for semi join
2709 : *
2710 : * (Also used for anti join, which we are supposed to estimate the same way.)
2711 : * Caller has ensured that vardata1 is the LHS variable; however, eqproc
2712 : * is for the original join operator, which might now need to have the inputs
2713 : * swapped in order to apply correctly. Also, if have_mcvs1 && have_mcvs2
2714 : * then hasmatch1[], hasmatch2[], and *p_nmatches were filled by
2715 : * eqjoinsel_inner.
2716 : */
2717 : static double
2718 10884 : eqjoinsel_semi(FmgrInfo *eqproc, Oid collation,
2719 : Oid hashLeft, Oid hashRight,
2720 : bool op_is_reversed,
2721 : VariableStatData *vardata1, VariableStatData *vardata2,
2722 : double nd1, double nd2,
2723 : bool isdefault1, bool isdefault2,
2724 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2725 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2726 : bool have_mcvs1, bool have_mcvs2,
2727 : bool *hasmatch1, bool *hasmatch2,
2728 : int *p_nmatches,
2729 : RelOptInfo *inner_rel)
2730 : {
2731 : double selec;
2732 :
2733 : /*
2734 : * We clamp nd2 to be not more than what we estimate the inner relation's
2735 : * size to be. This is intuitively somewhat reasonable since obviously
2736 : * there can't be more than that many distinct values coming from the
2737 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2738 : * likewise) is that this is the only pathway by which restriction clauses
2739 : * applied to the inner rel will affect the join result size estimate,
2740 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2741 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2742 : * the selectivity of outer-rel restrictions.
2743 : *
2744 : * We can apply this clamping both with respect to the base relation from
2745 : * which the join variable comes (if there is just one), and to the
2746 : * immediate inner input relation of the current join.
2747 : *
2748 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2749 : * great, maybe, but it didn't come out of nowhere either. This is most
2750 : * helpful when the inner relation is empty and consequently has no stats.
2751 : */
2752 10884 : if (vardata2->rel)
2753 : {
2754 10878 : if (nd2 >= vardata2->rel->rows)
2755 : {
2756 8736 : nd2 = vardata2->rel->rows;
2757 8736 : isdefault2 = false;
2758 : }
2759 : }
2760 10884 : if (nd2 >= inner_rel->rows)
2761 : {
2762 8680 : nd2 = inner_rel->rows;
2763 8680 : isdefault2 = false;
2764 : }
2765 :
2766 10884 : if (have_mcvs1 && have_mcvs2)
2767 624 : {
2768 : /*
2769 : * We have most-common-value lists for both relations. Run through
2770 : * the lists to see which MCVs actually join to each other with the
2771 : * given operator. This allows us to determine the exact join
2772 : * selectivity for the portion of the relations represented by the MCV
2773 : * lists. We still have to estimate for the remaining population, but
2774 : * in a skewed distribution this gives us a big leg up in accuracy.
2775 : */
2776 624 : double nullfrac1 = stats1->stanullfrac;
2777 : double matchprodfreq,
2778 : matchfreq1,
2779 : uncertainfrac,
2780 : uncertain;
2781 : int i,
2782 : nmatches,
2783 : clamped_nvalues2;
2784 :
2785 : /*
2786 : * The clamping above could have resulted in nd2 being less than
2787 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2788 : * most common values in the relation will appear in the join input,
2789 : * and so compare to only the first nd2 members of the MCV list. Of
2790 : * course this is frequently wrong, but it's the best bet we can make.
2791 : */
2792 624 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2793 :
2794 : /*
2795 : * If we did not set clamped_nvalues2 to less than sslot2->nvalues,
2796 : * then the hasmatch1[] and hasmatch2[] match flags computed by
2797 : * eqjoinsel_inner are still perfectly applicable, so we need not
2798 : * re-do the matching work. Note that it does not matter if
2799 : * op_is_reversed: we'd get the same answers.
2800 : *
2801 : * If we did clamp, then a different set of sslot2 values is to be
2802 : * compared, so we have to re-do the matching.
2803 : */
2804 624 : if (clamped_nvalues2 != sslot2->nvalues)
2805 : {
2806 : /* Must re-zero the arrays */
2807 0 : memset(hasmatch1, 0, sslot1->nvalues * sizeof(bool));
2808 0 : memset(hasmatch2, 0, clamped_nvalues2 * sizeof(bool));
2809 : /* Re-fill the match arrays */
2810 0 : eqjoinsel_find_matches(eqproc, collation,
2811 : hashLeft, hashRight,
2812 : op_is_reversed,
2813 : sslot1, sslot2,
2814 : sslot1->nvalues, clamped_nvalues2,
2815 : hasmatch1, hasmatch2,
2816 : p_nmatches, &matchprodfreq);
2817 : }
2818 624 : nmatches = *p_nmatches;
2819 :
2820 : /* Sum up frequencies of matched MCVs */
2821 624 : matchfreq1 = 0.0;
2822 13738 : for (i = 0; i < sslot1->nvalues; i++)
2823 : {
2824 13114 : if (hasmatch1[i])
2825 11448 : matchfreq1 += sslot1->numbers[i];
2826 : }
2827 624 : CLAMP_PROBABILITY(matchfreq1);
2828 :
2829 : /*
2830 : * Now we need to estimate the fraction of relation 1 that has at
2831 : * least one join partner. We know for certain that the matched MCVs
2832 : * do, so that gives us a lower bound, but we're really in the dark
2833 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2834 : * assume all non-null rel1 rows have join partners, else assume for
2835 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2836 : * can discount the known-matched MCVs from the distinct-values counts
2837 : * before doing the division.
2838 : *
2839 : * Crude as the above is, it's completely useless if we don't have
2840 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2841 : * nd2 is default, punt and assume half of the uncertain rows have
2842 : * join partners.
2843 : */
2844 624 : if (!isdefault1 && !isdefault2)
2845 : {
2846 624 : nd1 -= nmatches;
2847 624 : nd2 -= nmatches;
2848 624 : if (nd1 <= nd2 || nd2 < 0)
2849 588 : uncertainfrac = 1.0;
2850 : else
2851 36 : uncertainfrac = nd2 / nd1;
2852 : }
2853 : else
2854 0 : uncertainfrac = 0.5;
2855 624 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2856 624 : CLAMP_PROBABILITY(uncertain);
2857 624 : selec = matchfreq1 + uncertainfrac * uncertain;
2858 : }
2859 : else
2860 : {
2861 : /*
2862 : * Without MCV lists for both sides, we can only use the heuristic
2863 : * about nd1 vs nd2.
2864 : */
2865 10260 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2866 :
2867 10260 : if (!isdefault1 && !isdefault2)
2868 : {
2869 7862 : if (nd1 <= nd2 || nd2 < 0)
2870 4918 : selec = 1.0 - nullfrac1;
2871 : else
2872 2944 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2873 : }
2874 : else
2875 2398 : selec = 0.5 * (1.0 - nullfrac1);
2876 : }
2877 :
2878 10884 : return selec;
2879 : }
2880 :
2881 : /*
2882 : * Identify matching MCVs for eqjoinsel_inner or eqjoinsel_semi.
2883 : *
2884 : * Inputs:
2885 : * eqproc: FmgrInfo for equality function to use (might be reversed)
2886 : * collation: OID of collation to use
2887 : * hashLeft, hashRight: OIDs of hash functions associated with equality op,
2888 : * or InvalidOid if we're not to use hashing
2889 : * op_is_reversed: indicates that eqproc compares right type to left type
2890 : * sslot1, sslot2: MCV values for the lefthand and righthand inputs
2891 : * nvalues1, nvalues2: number of values to be considered (can be less than
2892 : * sslotN->nvalues, but not more)
2893 : * Outputs:
2894 : * hasmatch1[], hasmatch2[]: pre-zeroed arrays of lengths nvalues1, nvalues2;
2895 : * entries are set to true if that MCV has a match on the other side
2896 : * *p_nmatches: receives number of MCV pairs that match
2897 : * *p_matchprodfreq: receives sum(sslot1->numbers[i] * sslot2->numbers[j])
2898 : * for matching MCVs
2899 : *
2900 : * Note that hashLeft is for the eqproc's left-hand input type, hashRight
2901 : * for its right, regardless of op_is_reversed.
2902 : *
2903 : * Note we assume that each MCV will match at most one member of the other
2904 : * MCV list. If the operator isn't really equality, there could be multiple
2905 : * matches --- but we don't look for them, both for speed and because the
2906 : * math wouldn't add up...
2907 : */
2908 : static void
2909 28540 : eqjoinsel_find_matches(FmgrInfo *eqproc, Oid collation,
2910 : Oid hashLeft, Oid hashRight,
2911 : bool op_is_reversed,
2912 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2913 : int nvalues1, int nvalues2,
2914 : bool *hasmatch1, bool *hasmatch2,
2915 : int *p_nmatches, double *p_matchprodfreq)
2916 : {
2917 28540 : LOCAL_FCINFO(fcinfo, 2);
2918 28540 : double matchprodfreq = 0.0;
2919 28540 : int nmatches = 0;
2920 :
2921 : /*
2922 : * Save a few cycles by setting up the fcinfo struct just once. Using
2923 : * FunctionCallInvoke directly also avoids failure if the eqproc returns
2924 : * NULL, though really equality functions should never do that.
2925 : */
2926 28540 : InitFunctionCallInfoData(*fcinfo, eqproc, 2, collation,
2927 : NULL, NULL);
2928 28540 : fcinfo->args[0].isnull = false;
2929 28540 : fcinfo->args[1].isnull = false;
2930 :
2931 28540 : if (OidIsValid(hashLeft) && OidIsValid(hashRight))
2932 1668 : {
2933 : /* Use a hash table to speed up the matching */
2934 1668 : LOCAL_FCINFO(hash_fcinfo, 1);
2935 : FmgrInfo hash_proc;
2936 : MCVHashContext hashContext;
2937 : MCVHashTable_hash *hashTable;
2938 : AttStatsSlot *statsProbe;
2939 : AttStatsSlot *statsHash;
2940 : bool *hasMatchProbe;
2941 : bool *hasMatchHash;
2942 : int nvaluesProbe;
2943 : int nvaluesHash;
2944 :
2945 : /* Make sure we build the hash table on the smaller array. */
2946 1668 : if (sslot1->nvalues >= sslot2->nvalues)
2947 : {
2948 1668 : statsProbe = sslot1;
2949 1668 : statsHash = sslot2;
2950 1668 : hasMatchProbe = hasmatch1;
2951 1668 : hasMatchHash = hasmatch2;
2952 1668 : nvaluesProbe = nvalues1;
2953 1668 : nvaluesHash = nvalues2;
2954 : }
2955 : else
2956 : {
2957 : /* We'll have to reverse the direction of use of the operator. */
2958 0 : op_is_reversed = !op_is_reversed;
2959 0 : statsProbe = sslot2;
2960 0 : statsHash = sslot1;
2961 0 : hasMatchProbe = hasmatch2;
2962 0 : hasMatchHash = hasmatch1;
2963 0 : nvaluesProbe = nvalues2;
2964 0 : nvaluesHash = nvalues1;
2965 : }
2966 :
2967 : /*
2968 : * Build the hash table on the smaller array, using the appropriate
2969 : * hash function for its data type.
2970 : */
2971 1668 : fmgr_info(op_is_reversed ? hashLeft : hashRight, &hash_proc);
2972 1668 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
2973 : NULL, NULL);
2974 1668 : hash_fcinfo->args[0].isnull = false;
2975 :
2976 1668 : hashContext.equal_fcinfo = fcinfo;
2977 1668 : hashContext.hash_fcinfo = hash_fcinfo;
2978 1668 : hashContext.op_is_reversed = op_is_reversed;
2979 1668 : hashContext.insert_mode = true;
2980 1668 : get_typlenbyval(statsHash->valuetype,
2981 : &hashContext.hash_typlen,
2982 : &hashContext.hash_typbyval);
2983 :
2984 1668 : hashTable = MCVHashTable_create(CurrentMemoryContext,
2985 : nvaluesHash,
2986 : &hashContext);
2987 :
2988 168468 : for (int i = 0; i < nvaluesHash; i++)
2989 : {
2990 166800 : bool found = false;
2991 166800 : MCVHashEntry *entry = MCVHashTable_insert(hashTable,
2992 166800 : statsHash->values[i],
2993 : &found);
2994 :
2995 : /*
2996 : * MCVHashTable_insert will only report "found" if the new value
2997 : * is equal to some previous one per datum_image_eq(). That
2998 : * probably shouldn't happen, since we're not expecting duplicates
2999 : * in the MCV list. If we do find a dup, just ignore it, leaving
3000 : * the hash entry's index pointing at the first occurrence. That
3001 : * matches the behavior that the non-hashed code path would have.
3002 : */
3003 166800 : if (likely(!found))
3004 166800 : entry->index = i;
3005 : }
3006 :
3007 : /*
3008 : * Prepare to probe the hash table. If the probe values are of a
3009 : * different data type, then we need to change hash functions. (This
3010 : * code relies on the assumption that since we defined SH_STORE_HASH,
3011 : * simplehash.h will never need to compute hash values for existing
3012 : * hash table entries.)
3013 : */
3014 1668 : hashContext.insert_mode = false;
3015 1668 : if (hashLeft != hashRight)
3016 : {
3017 0 : fmgr_info(op_is_reversed ? hashRight : hashLeft, &hash_proc);
3018 : /* Resetting hash_fcinfo is probably unnecessary, but be safe */
3019 0 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
3020 : NULL, NULL);
3021 0 : hash_fcinfo->args[0].isnull = false;
3022 : }
3023 :
3024 : /* Look up each probe value in turn. */
3025 168468 : for (int i = 0; i < nvaluesProbe; i++)
3026 : {
3027 166800 : MCVHashEntry *entry = MCVHashTable_lookup(hashTable,
3028 166800 : statsProbe->values[i]);
3029 :
3030 : /* As in the other code path, skip already-matched hash entries */
3031 166800 : if (entry != NULL && !hasMatchHash[entry->index])
3032 : {
3033 65542 : hasMatchHash[entry->index] = hasMatchProbe[i] = true;
3034 65542 : nmatches++;
3035 65542 : matchprodfreq += statsHash->numbers[entry->index] * statsProbe->numbers[i];
3036 : }
3037 : }
3038 :
3039 1668 : MCVHashTable_destroy(hashTable);
3040 : }
3041 : else
3042 : {
3043 : /* We're not to use hashing, so do it the O(N^2) way */
3044 : int index1,
3045 : index2;
3046 :
3047 : /* Set up to supply the values in the order the operator expects */
3048 26872 : if (op_is_reversed)
3049 : {
3050 0 : index1 = 1;
3051 0 : index2 = 0;
3052 : }
3053 : else
3054 : {
3055 26872 : index1 = 0;
3056 26872 : index2 = 1;
3057 : }
3058 :
3059 538142 : for (int i = 0; i < nvalues1; i++)
3060 : {
3061 511270 : fcinfo->args[index1].value = sslot1->values[i];
3062 :
3063 10535870 : for (int j = 0; j < nvalues2; j++)
3064 : {
3065 : Datum fresult;
3066 :
3067 10253476 : if (hasmatch2[j])
3068 3110328 : continue;
3069 7143148 : fcinfo->args[index2].value = sslot2->values[j];
3070 7143148 : fcinfo->isnull = false;
3071 7143148 : fresult = FunctionCallInvoke(fcinfo);
3072 7143148 : if (!fcinfo->isnull && DatumGetBool(fresult))
3073 : {
3074 228876 : hasmatch1[i] = hasmatch2[j] = true;
3075 228876 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
3076 228876 : nmatches++;
3077 228876 : break;
3078 : }
3079 : }
3080 : }
3081 : }
3082 :
3083 28540 : *p_nmatches = nmatches;
3084 28540 : *p_matchprodfreq = matchprodfreq;
3085 28540 : }
3086 :
3087 : /*
3088 : * Support functions for the hash tables used by eqjoinsel_find_matches
3089 : */
3090 : static uint32
3091 333600 : hash_mcv(MCVHashTable_hash *tab, Datum key)
3092 : {
3093 333600 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3094 333600 : FunctionCallInfo fcinfo = context->hash_fcinfo;
3095 : Datum fresult;
3096 :
3097 333600 : fcinfo->args[0].value = key;
3098 333600 : fcinfo->isnull = false;
3099 333600 : fresult = FunctionCallInvoke(fcinfo);
3100 : Assert(!fcinfo->isnull);
3101 333600 : return DatumGetUInt32(fresult);
3102 : }
3103 :
3104 : static bool
3105 65542 : mcvs_equal(MCVHashTable_hash *tab, Datum key0, Datum key1)
3106 : {
3107 65542 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3108 :
3109 65542 : if (context->insert_mode)
3110 : {
3111 : /*
3112 : * During the insertion step, any comparisons will be between two
3113 : * Datums of the hash table's data type, so if the given operator is
3114 : * cross-type it will be the wrong thing to use. Fortunately, we can
3115 : * use datum_image_eq instead. The MCV values should all be distinct
3116 : * anyway, so it's mostly pro-forma to compare them at all.
3117 : */
3118 0 : return datum_image_eq(key0, key1,
3119 0 : context->hash_typbyval, context->hash_typlen);
3120 : }
3121 : else
3122 : {
3123 65542 : FunctionCallInfo fcinfo = context->equal_fcinfo;
3124 : Datum fresult;
3125 :
3126 : /*
3127 : * Apply the operator the correct way around. Although simplehash.h
3128 : * doesn't document this explicitly, during lookups key0 is from the
3129 : * hash table while key1 is the probe value, so we should compare them
3130 : * in that order only if op_is_reversed.
3131 : */
3132 65542 : if (context->op_is_reversed)
3133 : {
3134 0 : fcinfo->args[0].value = key0;
3135 0 : fcinfo->args[1].value = key1;
3136 : }
3137 : else
3138 : {
3139 65542 : fcinfo->args[0].value = key1;
3140 65542 : fcinfo->args[1].value = key0;
3141 : }
3142 65542 : fcinfo->isnull = false;
3143 65542 : fresult = FunctionCallInvoke(fcinfo);
3144 65542 : return (!fcinfo->isnull && DatumGetBool(fresult));
3145 : }
3146 : }
3147 :
3148 : /*
3149 : * neqjoinsel - Join selectivity of "!="
3150 : */
3151 : Datum
3152 3770 : neqjoinsel(PG_FUNCTION_ARGS)
3153 : {
3154 3770 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3155 3770 : Oid operator = PG_GETARG_OID(1);
3156 3770 : List *args = (List *) PG_GETARG_POINTER(2);
3157 3770 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
3158 3770 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
3159 3770 : Oid collation = PG_GET_COLLATION();
3160 : float8 result;
3161 :
3162 3770 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
3163 1278 : {
3164 : /*
3165 : * For semi-joins, if there is more than one distinct value in the RHS
3166 : * relation then every non-null LHS row must find a row to join since
3167 : * it can only be equal to one of them. We'll assume that there is
3168 : * always more than one distinct RHS value for the sake of stability,
3169 : * though in theory we could have special cases for empty RHS
3170 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
3171 : * fraction of LHS that has the same value as the single RHS value).
3172 : *
3173 : * For anti-joins, if we use the same assumption that there is more
3174 : * than one distinct key in the RHS relation, then every non-null LHS
3175 : * row must be suppressed by the anti-join.
3176 : *
3177 : * So either way, the selectivity estimate should be 1 - nullfrac.
3178 : */
3179 : VariableStatData leftvar;
3180 : VariableStatData rightvar;
3181 : bool reversed;
3182 : HeapTuple statsTuple;
3183 : double nullfrac;
3184 :
3185 1278 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
3186 1278 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
3187 1278 : if (HeapTupleIsValid(statsTuple))
3188 1042 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
3189 : else
3190 236 : nullfrac = 0.0;
3191 1278 : ReleaseVariableStats(leftvar);
3192 1278 : ReleaseVariableStats(rightvar);
3193 :
3194 1278 : result = 1.0 - nullfrac;
3195 : }
3196 : else
3197 : {
3198 : /*
3199 : * We want 1 - eqjoinsel() where the equality operator is the one
3200 : * associated with this != operator, that is, its negator.
3201 : */
3202 2492 : Oid eqop = get_negator(operator);
3203 :
3204 2492 : if (eqop)
3205 : {
3206 : result =
3207 2492 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
3208 : collation,
3209 : PointerGetDatum(root),
3210 : ObjectIdGetDatum(eqop),
3211 : PointerGetDatum(args),
3212 : Int16GetDatum(jointype),
3213 : PointerGetDatum(sjinfo)));
3214 : }
3215 : else
3216 : {
3217 : /* Use default selectivity (should we raise an error instead?) */
3218 0 : result = DEFAULT_EQ_SEL;
3219 : }
3220 2492 : result = 1.0 - result;
3221 : }
3222 :
3223 3770 : PG_RETURN_FLOAT8(result);
3224 : }
3225 :
3226 : /*
3227 : * scalarltjoinsel - Join selectivity of "<" for scalars
3228 : */
3229 : Datum
3230 324 : scalarltjoinsel(PG_FUNCTION_ARGS)
3231 : {
3232 324 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3233 : }
3234 :
3235 : /*
3236 : * scalarlejoinsel - Join selectivity of "<=" for scalars
3237 : */
3238 : Datum
3239 276 : scalarlejoinsel(PG_FUNCTION_ARGS)
3240 : {
3241 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3242 : }
3243 :
3244 : /*
3245 : * scalargtjoinsel - Join selectivity of ">" for scalars
3246 : */
3247 : Datum
3248 276 : scalargtjoinsel(PG_FUNCTION_ARGS)
3249 : {
3250 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3251 : }
3252 :
3253 : /*
3254 : * scalargejoinsel - Join selectivity of ">=" for scalars
3255 : */
3256 : Datum
3257 184 : scalargejoinsel(PG_FUNCTION_ARGS)
3258 : {
3259 184 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3260 : }
3261 :
3262 :
3263 : /*
3264 : * mergejoinscansel - Scan selectivity of merge join.
3265 : *
3266 : * A merge join will stop as soon as it exhausts either input stream.
3267 : * Therefore, if we can estimate the ranges of both input variables,
3268 : * we can estimate how much of the input will actually be read. This
3269 : * can have a considerable impact on the cost when using indexscans.
3270 : *
3271 : * Also, we can estimate how much of each input has to be read before the
3272 : * first join pair is found, which will affect the join's startup time.
3273 : *
3274 : * clause should be a clause already known to be mergejoinable. opfamily,
3275 : * cmptype, and nulls_first specify the sort ordering being used.
3276 : *
3277 : * The outputs are:
3278 : * *leftstart is set to the fraction of the left-hand variable expected
3279 : * to be scanned before the first join pair is found (0 to 1).
3280 : * *leftend is set to the fraction of the left-hand variable expected
3281 : * to be scanned before the join terminates (0 to 1).
3282 : * *rightstart, *rightend similarly for the right-hand variable.
3283 : */
3284 : void
3285 143264 : mergejoinscansel(PlannerInfo *root, Node *clause,
3286 : Oid opfamily, CompareType cmptype, bool nulls_first,
3287 : Selectivity *leftstart, Selectivity *leftend,
3288 : Selectivity *rightstart, Selectivity *rightend)
3289 : {
3290 : Node *left,
3291 : *right;
3292 : VariableStatData leftvar,
3293 : rightvar;
3294 : Oid opmethod;
3295 : int op_strategy;
3296 : Oid op_lefttype;
3297 : Oid op_righttype;
3298 : Oid opno,
3299 : collation,
3300 : lsortop,
3301 : rsortop,
3302 : lstatop,
3303 : rstatop,
3304 : ltop,
3305 : leop,
3306 : revltop,
3307 : revleop;
3308 : StrategyNumber ltstrat,
3309 : lestrat,
3310 : gtstrat,
3311 : gestrat;
3312 : bool isgt;
3313 : Datum leftmin,
3314 : leftmax,
3315 : rightmin,
3316 : rightmax;
3317 : double selec;
3318 :
3319 : /* Set default results if we can't figure anything out. */
3320 : /* XXX should default "start" fraction be a bit more than 0? */
3321 143264 : *leftstart = *rightstart = 0.0;
3322 143264 : *leftend = *rightend = 1.0;
3323 :
3324 : /* Deconstruct the merge clause */
3325 143264 : if (!is_opclause(clause))
3326 0 : return; /* shouldn't happen */
3327 143264 : opno = ((OpExpr *) clause)->opno;
3328 143264 : collation = ((OpExpr *) clause)->inputcollid;
3329 143264 : left = get_leftop((Expr *) clause);
3330 143264 : right = get_rightop((Expr *) clause);
3331 143264 : if (!right)
3332 0 : return; /* shouldn't happen */
3333 :
3334 : /* Look for stats for the inputs */
3335 143264 : examine_variable(root, left, 0, &leftvar);
3336 143264 : examine_variable(root, right, 0, &rightvar);
3337 :
3338 143264 : opmethod = get_opfamily_method(opfamily);
3339 :
3340 : /* Extract the operator's declared left/right datatypes */
3341 143264 : get_op_opfamily_properties(opno, opfamily, false,
3342 : &op_strategy,
3343 : &op_lefttype,
3344 : &op_righttype);
3345 : Assert(IndexAmTranslateStrategy(op_strategy, opmethod, opfamily, true) == COMPARE_EQ);
3346 :
3347 : /*
3348 : * Look up the various operators we need. If we don't find them all, it
3349 : * probably means the opfamily is broken, but we just fail silently.
3350 : *
3351 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3352 : * operator, regardless of which sort direction we are considering.
3353 : */
3354 143264 : switch (cmptype)
3355 : {
3356 143228 : case COMPARE_LT:
3357 143228 : isgt = false;
3358 143228 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3359 143228 : lestrat = IndexAmTranslateCompareType(COMPARE_LE, opmethod, opfamily, true);
3360 143228 : if (op_lefttype == op_righttype)
3361 : {
3362 : /* easy case */
3363 141454 : ltop = get_opfamily_member(opfamily,
3364 : op_lefttype, op_righttype,
3365 : ltstrat);
3366 141454 : leop = get_opfamily_member(opfamily,
3367 : op_lefttype, op_righttype,
3368 : lestrat);
3369 141454 : lsortop = ltop;
3370 141454 : rsortop = ltop;
3371 141454 : lstatop = lsortop;
3372 141454 : rstatop = rsortop;
3373 141454 : revltop = ltop;
3374 141454 : revleop = leop;
3375 : }
3376 : else
3377 : {
3378 1774 : ltop = get_opfamily_member(opfamily,
3379 : op_lefttype, op_righttype,
3380 : ltstrat);
3381 1774 : leop = get_opfamily_member(opfamily,
3382 : op_lefttype, op_righttype,
3383 : lestrat);
3384 1774 : lsortop = get_opfamily_member(opfamily,
3385 : op_lefttype, op_lefttype,
3386 : ltstrat);
3387 1774 : rsortop = get_opfamily_member(opfamily,
3388 : op_righttype, op_righttype,
3389 : ltstrat);
3390 1774 : lstatop = lsortop;
3391 1774 : rstatop = rsortop;
3392 1774 : revltop = get_opfamily_member(opfamily,
3393 : op_righttype, op_lefttype,
3394 : ltstrat);
3395 1774 : revleop = get_opfamily_member(opfamily,
3396 : op_righttype, op_lefttype,
3397 : lestrat);
3398 : }
3399 143228 : break;
3400 36 : case COMPARE_GT:
3401 : /* descending-order case */
3402 36 : isgt = true;
3403 36 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3404 36 : gtstrat = IndexAmTranslateCompareType(COMPARE_GT, opmethod, opfamily, true);
3405 36 : gestrat = IndexAmTranslateCompareType(COMPARE_GE, opmethod, opfamily, true);
3406 36 : if (op_lefttype == op_righttype)
3407 : {
3408 : /* easy case */
3409 36 : ltop = get_opfamily_member(opfamily,
3410 : op_lefttype, op_righttype,
3411 : gtstrat);
3412 36 : leop = get_opfamily_member(opfamily,
3413 : op_lefttype, op_righttype,
3414 : gestrat);
3415 36 : lsortop = ltop;
3416 36 : rsortop = ltop;
3417 36 : lstatop = get_opfamily_member(opfamily,
3418 : op_lefttype, op_lefttype,
3419 : ltstrat);
3420 36 : rstatop = lstatop;
3421 36 : revltop = ltop;
3422 36 : revleop = leop;
3423 : }
3424 : else
3425 : {
3426 0 : ltop = get_opfamily_member(opfamily,
3427 : op_lefttype, op_righttype,
3428 : gtstrat);
3429 0 : leop = get_opfamily_member(opfamily,
3430 : op_lefttype, op_righttype,
3431 : gestrat);
3432 0 : lsortop = get_opfamily_member(opfamily,
3433 : op_lefttype, op_lefttype,
3434 : gtstrat);
3435 0 : rsortop = get_opfamily_member(opfamily,
3436 : op_righttype, op_righttype,
3437 : gtstrat);
3438 0 : lstatop = get_opfamily_member(opfamily,
3439 : op_lefttype, op_lefttype,
3440 : ltstrat);
3441 0 : rstatop = get_opfamily_member(opfamily,
3442 : op_righttype, op_righttype,
3443 : ltstrat);
3444 0 : revltop = get_opfamily_member(opfamily,
3445 : op_righttype, op_lefttype,
3446 : gtstrat);
3447 0 : revleop = get_opfamily_member(opfamily,
3448 : op_righttype, op_lefttype,
3449 : gestrat);
3450 : }
3451 36 : break;
3452 0 : default:
3453 0 : goto fail; /* shouldn't get here */
3454 : }
3455 :
3456 143264 : if (!OidIsValid(lsortop) ||
3457 143264 : !OidIsValid(rsortop) ||
3458 143264 : !OidIsValid(lstatop) ||
3459 143264 : !OidIsValid(rstatop) ||
3460 143252 : !OidIsValid(ltop) ||
3461 143252 : !OidIsValid(leop) ||
3462 143252 : !OidIsValid(revltop) ||
3463 : !OidIsValid(revleop))
3464 12 : goto fail; /* insufficient info in catalogs */
3465 :
3466 : /* Try to get ranges of both inputs */
3467 143252 : if (!isgt)
3468 : {
3469 143216 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3470 : &leftmin, &leftmax))
3471 34760 : goto fail; /* no range available from stats */
3472 108456 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3473 : &rightmin, &rightmax))
3474 25442 : goto fail; /* no range available from stats */
3475 : }
3476 : else
3477 : {
3478 : /* need to swap the max and min */
3479 36 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3480 : &leftmax, &leftmin))
3481 30 : goto fail; /* no range available from stats */
3482 6 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3483 : &rightmax, &rightmin))
3484 0 : goto fail; /* no range available from stats */
3485 : }
3486 :
3487 : /*
3488 : * Now, the fraction of the left variable that will be scanned is the
3489 : * fraction that's <= the right-side maximum value. But only believe
3490 : * non-default estimates, else stick with our 1.0.
3491 : */
3492 83020 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3493 : rightmax, op_righttype);
3494 83020 : if (selec != DEFAULT_INEQ_SEL)
3495 83014 : *leftend = selec;
3496 :
3497 : /* And similarly for the right variable. */
3498 83020 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3499 : leftmax, op_lefttype);
3500 83020 : if (selec != DEFAULT_INEQ_SEL)
3501 83020 : *rightend = selec;
3502 :
3503 : /*
3504 : * Only one of the two "end" fractions can really be less than 1.0;
3505 : * believe the smaller estimate and reset the other one to exactly 1.0. If
3506 : * we get exactly equal estimates (as can easily happen with self-joins),
3507 : * believe neither.
3508 : */
3509 83020 : if (*leftend > *rightend)
3510 24754 : *leftend = 1.0;
3511 58266 : else if (*leftend < *rightend)
3512 33650 : *rightend = 1.0;
3513 : else
3514 24616 : *leftend = *rightend = 1.0;
3515 :
3516 : /*
3517 : * Also, the fraction of the left variable that will be scanned before the
3518 : * first join pair is found is the fraction that's < the right-side
3519 : * minimum value. But only believe non-default estimates, else stick with
3520 : * our own default.
3521 : */
3522 83020 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3523 : rightmin, op_righttype);
3524 83020 : if (selec != DEFAULT_INEQ_SEL)
3525 83020 : *leftstart = selec;
3526 :
3527 : /* And similarly for the right variable. */
3528 83020 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3529 : leftmin, op_lefttype);
3530 83020 : if (selec != DEFAULT_INEQ_SEL)
3531 83020 : *rightstart = selec;
3532 :
3533 : /*
3534 : * Only one of the two "start" fractions can really be more than zero;
3535 : * believe the larger estimate and reset the other one to exactly 0.0. If
3536 : * we get exactly equal estimates (as can easily happen with self-joins),
3537 : * believe neither.
3538 : */
3539 83020 : if (*leftstart < *rightstart)
3540 16984 : *leftstart = 0.0;
3541 66036 : else if (*leftstart > *rightstart)
3542 24436 : *rightstart = 0.0;
3543 : else
3544 41600 : *leftstart = *rightstart = 0.0;
3545 :
3546 : /*
3547 : * If the sort order is nulls-first, we're going to have to skip over any
3548 : * nulls too. These would not have been counted by scalarineqsel, and we
3549 : * can safely add in this fraction regardless of whether we believe
3550 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3551 : */
3552 83020 : if (nulls_first)
3553 : {
3554 : Form_pg_statistic stats;
3555 :
3556 6 : if (HeapTupleIsValid(leftvar.statsTuple))
3557 : {
3558 6 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3559 6 : *leftstart += stats->stanullfrac;
3560 6 : CLAMP_PROBABILITY(*leftstart);
3561 6 : *leftend += stats->stanullfrac;
3562 6 : CLAMP_PROBABILITY(*leftend);
3563 : }
3564 6 : if (HeapTupleIsValid(rightvar.statsTuple))
3565 : {
3566 6 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3567 6 : *rightstart += stats->stanullfrac;
3568 6 : CLAMP_PROBABILITY(*rightstart);
3569 6 : *rightend += stats->stanullfrac;
3570 6 : CLAMP_PROBABILITY(*rightend);
3571 : }
3572 : }
3573 :
3574 : /* Disbelieve start >= end, just in case that can happen */
3575 83020 : if (*leftstart >= *leftend)
3576 : {
3577 164 : *leftstart = 0.0;
3578 164 : *leftend = 1.0;
3579 : }
3580 83020 : if (*rightstart >= *rightend)
3581 : {
3582 1110 : *rightstart = 0.0;
3583 1110 : *rightend = 1.0;
3584 : }
3585 :
3586 81910 : fail:
3587 143264 : ReleaseVariableStats(leftvar);
3588 143264 : ReleaseVariableStats(rightvar);
3589 : }
3590 :
3591 :
3592 : /*
3593 : * matchingsel -- generic matching-operator selectivity support
3594 : *
3595 : * Use these for any operators that (a) are on data types for which we collect
3596 : * standard statistics, and (b) have behavior for which the default estimate
3597 : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3598 : * operators.
3599 : */
3600 :
3601 : Datum
3602 1130 : matchingsel(PG_FUNCTION_ARGS)
3603 : {
3604 1130 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3605 1130 : Oid operator = PG_GETARG_OID(1);
3606 1130 : List *args = (List *) PG_GETARG_POINTER(2);
3607 1130 : int varRelid = PG_GETARG_INT32(3);
3608 1130 : Oid collation = PG_GET_COLLATION();
3609 : double selec;
3610 :
3611 : /* Use generic restriction selectivity logic. */
3612 1130 : selec = generic_restriction_selectivity(root, operator, collation,
3613 : args, varRelid,
3614 : DEFAULT_MATCHING_SEL);
3615 :
3616 1130 : PG_RETURN_FLOAT8((float8) selec);
3617 : }
3618 :
3619 : Datum
3620 6 : matchingjoinsel(PG_FUNCTION_ARGS)
3621 : {
3622 : /* Just punt, for the moment. */
3623 6 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3624 : }
3625 :
3626 :
3627 : /*
3628 : * Helper routine for estimate_num_groups: add an item to a list of
3629 : * GroupVarInfos, but only if it's not known equal to any of the existing
3630 : * entries.
3631 : */
3632 : typedef struct
3633 : {
3634 : Node *var; /* might be an expression, not just a Var */
3635 : RelOptInfo *rel; /* relation it belongs to */
3636 : double ndistinct; /* # distinct values */
3637 : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3638 : } GroupVarInfo;
3639 :
3640 : static List *
3641 397288 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3642 : Node *var, VariableStatData *vardata)
3643 : {
3644 : GroupVarInfo *varinfo;
3645 : double ndistinct;
3646 : bool isdefault;
3647 : ListCell *lc;
3648 :
3649 397288 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3650 :
3651 : /*
3652 : * The nullingrels bits within the var could cause the same var to be
3653 : * counted multiple times if it's marked with different nullingrels. They
3654 : * could also prevent us from matching the var to the expressions in
3655 : * extended statistics (see estimate_multivariate_ndistinct). So strip
3656 : * them out first.
3657 : */
3658 397288 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3659 :
3660 480306 : foreach(lc, varinfos)
3661 : {
3662 84150 : varinfo = (GroupVarInfo *) lfirst(lc);
3663 :
3664 : /* Drop exact duplicates */
3665 84150 : if (equal(var, varinfo->var))
3666 1132 : return varinfos;
3667 :
3668 : /*
3669 : * Drop known-equal vars, but only if they belong to different
3670 : * relations (see comments for estimate_num_groups). We aren't too
3671 : * fussy about the semantics of "equal" here.
3672 : */
3673 90294 : if (vardata->rel != varinfo->rel &&
3674 7000 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3675 : {
3676 300 : if (varinfo->ndistinct <= ndistinct)
3677 : {
3678 : /* Keep older item, forget new one */
3679 276 : return varinfos;
3680 : }
3681 : else
3682 : {
3683 : /* Delete the older item */
3684 24 : varinfos = foreach_delete_current(varinfos, lc);
3685 : }
3686 : }
3687 : }
3688 :
3689 396156 : varinfo = palloc_object(GroupVarInfo);
3690 :
3691 396156 : varinfo->var = var;
3692 396156 : varinfo->rel = vardata->rel;
3693 396156 : varinfo->ndistinct = ndistinct;
3694 396156 : varinfo->isdefault = isdefault;
3695 396156 : varinfos = lappend(varinfos, varinfo);
3696 396156 : return varinfos;
3697 : }
3698 :
3699 : /*
3700 : * estimate_num_groups - Estimate number of groups in a grouped query
3701 : *
3702 : * Given a query having a GROUP BY clause, estimate how many groups there
3703 : * will be --- ie, the number of distinct combinations of the GROUP BY
3704 : * expressions.
3705 : *
3706 : * This routine is also used to estimate the number of rows emitted by
3707 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3708 : * actually, we only use it for DISTINCT when there's no grouping or
3709 : * aggregation ahead of the DISTINCT.)
3710 : *
3711 : * Inputs:
3712 : * root - the query
3713 : * groupExprs - list of expressions being grouped by
3714 : * input_rows - number of rows estimated to arrive at the group/unique
3715 : * filter step
3716 : * pgset - NULL, or a List** pointing to a grouping set to filter the
3717 : * groupExprs against
3718 : *
3719 : * Outputs:
3720 : * estinfo - When passed as non-NULL, the function will set bits in the
3721 : * "flags" field in order to provide callers with additional information
3722 : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3723 : * bit if we used any default values in the estimation.
3724 : *
3725 : * Given the lack of any cross-correlation statistics in the system, it's
3726 : * impossible to do anything really trustworthy with GROUP BY conditions
3727 : * involving multiple Vars. We should however avoid assuming the worst
3728 : * case (all possible cross-product terms actually appear as groups) since
3729 : * very often the grouped-by Vars are highly correlated. Our current approach
3730 : * is as follows:
3731 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3732 : * independently of their content, and are ignored in the subsequent
3733 : * steps. This is mainly because tests like "col IS NULL" break the
3734 : * heuristic used in step 2 especially badly.
3735 : * 2. Reduce the given expressions to a list of unique Vars used. For
3736 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3737 : * It is clearly correct not to count the same Var more than once.
3738 : * It is also reasonable to treat f(x) the same as x: f() cannot
3739 : * increase the number of distinct values (unless it is volatile,
3740 : * which we consider unlikely for grouping), but it probably won't
3741 : * reduce the number of distinct values much either.
3742 : * As a special case, if a GROUP BY expression can be matched to an
3743 : * expressional index for which we have statistics, then we treat the
3744 : * whole expression as though it were just a Var.
3745 : * 3. If the list contains Vars of different relations that are known equal
3746 : * due to equivalence classes, then drop all but one of the Vars from each
3747 : * known-equal set, keeping the one with smallest estimated # of values
3748 : * (since the extra values of the others can't appear in joined rows).
3749 : * Note the reason we only consider Vars of different relations is that
3750 : * if we considered ones of the same rel, we'd be double-counting the
3751 : * restriction selectivity of the equality in the next step.
3752 : * 4. For Vars within a single source rel, we multiply together the numbers
3753 : * of values, clamp to the number of rows in the rel (divided by 10 if
3754 : * more than one Var), and then multiply by a factor based on the
3755 : * selectivity of the restriction clauses for that rel. When there's
3756 : * more than one Var, the initial product is probably too high (it's the
3757 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3758 : * helpful heuristic for not letting the estimate get out of hand. (The
3759 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3760 : * we multiply by to adjust for the restriction selectivity assumes that
3761 : * the restriction clauses are independent of the grouping, which may not
3762 : * be a valid assumption, but it's hard to do better.
3763 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3764 : * rel, and multiply the results together.
3765 : * Note that rels not containing grouped Vars are ignored completely, as are
3766 : * join clauses. Such rels cannot increase the number of groups, and we
3767 : * assume such clauses do not reduce the number either (somewhat bogus,
3768 : * but we don't have the info to do better).
3769 : */
3770 : double
3771 345682 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3772 : List **pgset, EstimationInfo *estinfo)
3773 : {
3774 345682 : List *varinfos = NIL;
3775 345682 : double srf_multiplier = 1.0;
3776 : double numdistinct;
3777 : ListCell *l;
3778 : int i;
3779 :
3780 : /* Zero the estinfo output parameter, if non-NULL */
3781 345682 : if (estinfo != NULL)
3782 295294 : memset(estinfo, 0, sizeof(EstimationInfo));
3783 :
3784 : /*
3785 : * We don't ever want to return an estimate of zero groups, as that tends
3786 : * to lead to division-by-zero and other unpleasantness. The input_rows
3787 : * estimate is usually already at least 1, but clamp it just in case it
3788 : * isn't.
3789 : */
3790 345682 : input_rows = clamp_row_est(input_rows);
3791 :
3792 : /*
3793 : * If no grouping columns, there's exactly one group. (This can't happen
3794 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3795 : * corner cases with set operations.)
3796 : */
3797 345682 : if (groupExprs == NIL || (pgset && *pgset == NIL))
3798 1136 : return 1.0;
3799 :
3800 : /*
3801 : * Count groups derived from boolean grouping expressions. For other
3802 : * expressions, find the unique Vars used, treating an expression as a Var
3803 : * if we can find stats for it. For each one, record the statistical
3804 : * estimate of number of distinct values (total in its table, without
3805 : * regard for filtering).
3806 : */
3807 344546 : numdistinct = 1.0;
3808 :
3809 344546 : i = 0;
3810 740038 : foreach(l, groupExprs)
3811 : {
3812 395540 : Node *groupexpr = (Node *) lfirst(l);
3813 : double this_srf_multiplier;
3814 : VariableStatData vardata;
3815 : List *varshere;
3816 : ListCell *l2;
3817 :
3818 : /* is expression in this grouping set? */
3819 395540 : if (pgset && !list_member_int(*pgset, i++))
3820 326774 : continue;
3821 :
3822 : /*
3823 : * Set-returning functions in grouping columns are a bit problematic.
3824 : * The code below will effectively ignore their SRF nature and come up
3825 : * with a numdistinct estimate as though they were scalar functions.
3826 : * We compensate by scaling up the end result by the largest SRF
3827 : * rowcount estimate. (This will be an overestimate if the SRF
3828 : * produces multiple copies of any output value, but it seems best to
3829 : * assume the SRF's outputs are distinct. In any case, it's probably
3830 : * pointless to worry too much about this without much better
3831 : * estimates for SRF output rowcounts than we have today.)
3832 : */
3833 394728 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3834 394728 : if (srf_multiplier < this_srf_multiplier)
3835 144 : srf_multiplier = this_srf_multiplier;
3836 :
3837 : /* Short-circuit for expressions returning boolean */
3838 394728 : if (exprType(groupexpr) == BOOLOID)
3839 : {
3840 204 : numdistinct *= 2.0;
3841 204 : continue;
3842 : }
3843 :
3844 : /*
3845 : * If examine_variable is able to deduce anything about the GROUP BY
3846 : * expression, treat it as a single variable even if it's really more
3847 : * complicated.
3848 : *
3849 : * XXX This has the consequence that if there's a statistics object on
3850 : * the expression, we don't split it into individual Vars. This
3851 : * affects our selection of statistics in
3852 : * estimate_multivariate_ndistinct, because it's probably better to
3853 : * use more accurate estimate for each expression and treat them as
3854 : * independent, than to combine estimates for the extracted variables
3855 : * when we don't know how that relates to the expressions.
3856 : */
3857 394524 : examine_variable(root, groupexpr, 0, &vardata);
3858 394524 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3859 : {
3860 325074 : varinfos = add_unique_group_var(root, varinfos,
3861 : groupexpr, &vardata);
3862 325074 : ReleaseVariableStats(vardata);
3863 325074 : continue;
3864 : }
3865 69450 : ReleaseVariableStats(vardata);
3866 :
3867 : /*
3868 : * Else pull out the component Vars. Handle PlaceHolderVars by
3869 : * recursing into their arguments (effectively assuming that the
3870 : * PlaceHolderVar doesn't change the number of groups, which boils
3871 : * down to ignoring the possible addition of nulls to the result set).
3872 : */
3873 69450 : varshere = pull_var_clause(groupexpr,
3874 : PVC_RECURSE_AGGREGATES |
3875 : PVC_RECURSE_WINDOWFUNCS |
3876 : PVC_RECURSE_PLACEHOLDERS);
3877 :
3878 : /*
3879 : * If we find any variable-free GROUP BY item, then either it is a
3880 : * constant (and we can ignore it) or it contains a volatile function;
3881 : * in the latter case we punt and assume that each input row will
3882 : * yield a distinct group.
3883 : */
3884 69450 : if (varshere == NIL)
3885 : {
3886 732 : if (contain_volatile_functions(groupexpr))
3887 48 : return input_rows;
3888 684 : continue;
3889 : }
3890 :
3891 : /*
3892 : * Else add variables to varinfos list
3893 : */
3894 140932 : foreach(l2, varshere)
3895 : {
3896 72214 : Node *var = (Node *) lfirst(l2);
3897 :
3898 72214 : examine_variable(root, var, 0, &vardata);
3899 72214 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3900 72214 : ReleaseVariableStats(vardata);
3901 : }
3902 : }
3903 :
3904 : /*
3905 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3906 : * list.
3907 : */
3908 344498 : if (varinfos == NIL)
3909 : {
3910 : /* Apply SRF multiplier as we would do in the long path */
3911 400 : numdistinct *= srf_multiplier;
3912 : /* Round off */
3913 400 : numdistinct = ceil(numdistinct);
3914 : /* Guard against out-of-range answers */
3915 400 : if (numdistinct > input_rows)
3916 44 : numdistinct = input_rows;
3917 400 : if (numdistinct < 1.0)
3918 0 : numdistinct = 1.0;
3919 400 : return numdistinct;
3920 : }
3921 :
3922 : /*
3923 : * Group Vars by relation and estimate total numdistinct.
3924 : *
3925 : * For each iteration of the outer loop, we process the frontmost Var in
3926 : * varinfos, plus all other Vars in the same relation. We remove these
3927 : * Vars from the newvarinfos list for the next iteration. This is the
3928 : * easiest way to group Vars of same rel together.
3929 : */
3930 : do
3931 : {
3932 347012 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3933 347012 : RelOptInfo *rel = varinfo1->rel;
3934 347012 : double reldistinct = 1;
3935 347012 : double relmaxndistinct = reldistinct;
3936 347012 : int relvarcount = 0;
3937 347012 : List *newvarinfos = NIL;
3938 347012 : List *relvarinfos = NIL;
3939 :
3940 : /*
3941 : * Split the list of varinfos in two - one for the current rel, one
3942 : * for remaining Vars on other rels.
3943 : */
3944 347012 : relvarinfos = lappend(relvarinfos, varinfo1);
3945 401720 : for_each_from(l, varinfos, 1)
3946 : {
3947 54708 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3948 :
3949 54708 : if (varinfo2->rel == varinfo1->rel)
3950 : {
3951 : /* varinfos on current rel */
3952 49120 : relvarinfos = lappend(relvarinfos, varinfo2);
3953 : }
3954 : else
3955 : {
3956 : /* not time to process varinfo2 yet */
3957 5588 : newvarinfos = lappend(newvarinfos, varinfo2);
3958 : }
3959 : }
3960 :
3961 : /*
3962 : * Get the numdistinct estimate for the Vars of this rel. We
3963 : * iteratively search for multivariate n-distinct with maximum number
3964 : * of vars; assuming that each var group is independent of the others,
3965 : * we multiply them together. Any remaining relvarinfos after no more
3966 : * multivariate matches are found are assumed independent too, so
3967 : * their individual ndistinct estimates are multiplied also.
3968 : *
3969 : * While iterating, count how many separate numdistinct values we
3970 : * apply. We apply a fudge factor below, but only if we multiplied
3971 : * more than one such values.
3972 : */
3973 694150 : while (relvarinfos)
3974 : {
3975 : double mvndistinct;
3976 :
3977 347138 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3978 : &mvndistinct))
3979 : {
3980 414 : reldistinct *= mvndistinct;
3981 414 : if (relmaxndistinct < mvndistinct)
3982 402 : relmaxndistinct = mvndistinct;
3983 414 : relvarcount++;
3984 : }
3985 : else
3986 : {
3987 741980 : foreach(l, relvarinfos)
3988 : {
3989 395256 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3990 :
3991 395256 : reldistinct *= varinfo2->ndistinct;
3992 395256 : if (relmaxndistinct < varinfo2->ndistinct)
3993 348266 : relmaxndistinct = varinfo2->ndistinct;
3994 395256 : relvarcount++;
3995 :
3996 : /*
3997 : * When varinfo2's isdefault is set then we'd better set
3998 : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3999 : */
4000 395256 : if (estinfo != NULL && varinfo2->isdefault)
4001 19072 : estinfo->flags |= SELFLAG_USED_DEFAULT;
4002 : }
4003 :
4004 : /* we're done with this relation */
4005 346724 : relvarinfos = NIL;
4006 : }
4007 : }
4008 :
4009 : /*
4010 : * Sanity check --- don't divide by zero if empty relation.
4011 : */
4012 : Assert(IS_SIMPLE_REL(rel));
4013 347012 : if (rel->tuples > 0)
4014 : {
4015 : /*
4016 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
4017 : * fudge factor is because the Vars are probably correlated but we
4018 : * don't know by how much. We should never clamp to less than the
4019 : * largest ndistinct value for any of the Vars, though, since
4020 : * there will surely be at least that many groups.
4021 : */
4022 345968 : double clamp = rel->tuples;
4023 :
4024 345968 : if (relvarcount > 1)
4025 : {
4026 44330 : clamp *= 0.1;
4027 44330 : if (clamp < relmaxndistinct)
4028 : {
4029 41652 : clamp = relmaxndistinct;
4030 : /* for sanity in case some ndistinct is too large: */
4031 41652 : if (clamp > rel->tuples)
4032 78 : clamp = rel->tuples;
4033 : }
4034 : }
4035 345968 : if (reldistinct > clamp)
4036 36220 : reldistinct = clamp;
4037 :
4038 : /*
4039 : * Update the estimate based on the restriction selectivity,
4040 : * guarding against division by zero when reldistinct is zero.
4041 : * Also skip this if we know that we are returning all rows.
4042 : */
4043 345968 : if (reldistinct > 0 && rel->rows < rel->tuples)
4044 : {
4045 : /*
4046 : * Given a table containing N rows with n distinct values in a
4047 : * uniform distribution, if we select p rows at random then
4048 : * the expected number of distinct values selected is
4049 : *
4050 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
4051 : *
4052 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
4053 : *
4054 : * See "Approximating block accesses in database
4055 : * organizations", S. B. Yao, Communications of the ACM,
4056 : * Volume 20 Issue 4, April 1977 Pages 260-261.
4057 : *
4058 : * Alternatively, re-arranging the terms from the factorials,
4059 : * this may be written as
4060 : *
4061 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
4062 : *
4063 : * This form of the formula is more efficient to compute in
4064 : * the common case where p is larger than N/n. Additionally,
4065 : * as pointed out by Dell'Era, if i << N for all terms in the
4066 : * product, it can be approximated by
4067 : *
4068 : * n * (1 - ((N-p)/N)^(N/n))
4069 : *
4070 : * See "Expected distinct values when selecting from a bag
4071 : * without replacement", Alberto Dell'Era,
4072 : * http://www.adellera.it/investigations/distinct_balls/.
4073 : *
4074 : * The condition i << N is equivalent to n >> 1, so this is a
4075 : * good approximation when the number of distinct values in
4076 : * the table is large. It turns out that this formula also
4077 : * works well even when n is small.
4078 : */
4079 108980 : reldistinct *=
4080 108980 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
4081 108980 : rel->tuples / reldistinct));
4082 : }
4083 345968 : reldistinct = clamp_row_est(reldistinct);
4084 :
4085 : /*
4086 : * Update estimate of total distinct groups.
4087 : */
4088 345968 : numdistinct *= reldistinct;
4089 : }
4090 :
4091 347012 : varinfos = newvarinfos;
4092 347012 : } while (varinfos != NIL);
4093 :
4094 : /* Now we can account for the effects of any SRFs */
4095 344098 : numdistinct *= srf_multiplier;
4096 :
4097 : /* Round off */
4098 344098 : numdistinct = ceil(numdistinct);
4099 :
4100 : /* Guard against out-of-range answers */
4101 344098 : if (numdistinct > input_rows)
4102 71390 : numdistinct = input_rows;
4103 344098 : if (numdistinct < 1.0)
4104 0 : numdistinct = 1.0;
4105 :
4106 344098 : return numdistinct;
4107 : }
4108 :
4109 : /*
4110 : * Try to estimate the bucket size of the hash join inner side when the join
4111 : * condition contains two or more clauses by employing extended statistics.
4112 : *
4113 : * The main idea of this approach is that the distinct value generated by
4114 : * multivariate estimation on two or more columns would provide less bucket size
4115 : * than estimation on one separate column.
4116 : *
4117 : * IMPORTANT: It is crucial to synchronize the approach of combining different
4118 : * estimations with the caller's method.
4119 : *
4120 : * Return a list of clauses that didn't fetch any extended statistics.
4121 : */
4122 : List *
4123 454278 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
4124 : List *hashclauses,
4125 : Selectivity *innerbucketsize)
4126 : {
4127 : List *clauses;
4128 : List *otherclauses;
4129 : double ndistinct;
4130 :
4131 454278 : if (list_length(hashclauses) <= 1)
4132 : {
4133 : /*
4134 : * Nothing to do for a single clause. Could we employ univariate
4135 : * extended stat here?
4136 : */
4137 417536 : return hashclauses;
4138 : }
4139 :
4140 : /* "clauses" is the list of hashclauses we've not dealt with yet */
4141 36742 : clauses = list_copy(hashclauses);
4142 : /* "otherclauses" holds clauses we are going to return to caller */
4143 36742 : otherclauses = NIL;
4144 : /* current estimate of ndistinct */
4145 36742 : ndistinct = 1.0;
4146 73496 : while (clauses != NIL)
4147 : {
4148 : ListCell *lc;
4149 36754 : int relid = -1;
4150 36754 : List *varinfos = NIL;
4151 36754 : List *origin_rinfos = NIL;
4152 : double mvndistinct;
4153 : List *origin_varinfos;
4154 36754 : int group_relid = -1;
4155 36754 : RelOptInfo *group_rel = NULL;
4156 : ListCell *lc1,
4157 : *lc2;
4158 :
4159 : /*
4160 : * Find clauses, referencing the same single base relation and try to
4161 : * estimate such a group with extended statistics. Create varinfo for
4162 : * an approved clause, push it to otherclauses, if it can't be
4163 : * estimated here or ignore to process at the next iteration.
4164 : */
4165 110874 : foreach(lc, clauses)
4166 : {
4167 74120 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
4168 : Node *expr;
4169 : Relids relids;
4170 : GroupVarInfo *varinfo;
4171 :
4172 : /*
4173 : * Find the inner side of the join, which we need to estimate the
4174 : * number of buckets. Use outer_is_left because the
4175 : * clause_sides_match_join routine has called on hash clauses.
4176 : */
4177 148240 : relids = rinfo->outer_is_left ?
4178 74120 : rinfo->right_relids : rinfo->left_relids;
4179 148240 : expr = rinfo->outer_is_left ?
4180 74120 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
4181 :
4182 74120 : if (bms_get_singleton_member(relids, &relid) &&
4183 71506 : root->simple_rel_array[relid]->statlist != NIL)
4184 48 : {
4185 60 : bool is_duplicate = false;
4186 :
4187 : /*
4188 : * This inner-side expression references only one relation.
4189 : * Extended statistics on this clause can exist.
4190 : */
4191 60 : if (group_relid < 0)
4192 : {
4193 30 : RangeTblEntry *rte = root->simple_rte_array[relid];
4194 :
4195 30 : if (!rte || (rte->relkind != RELKIND_RELATION &&
4196 0 : rte->relkind != RELKIND_MATVIEW &&
4197 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
4198 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
4199 : {
4200 : /* Extended statistics can't exist in principle */
4201 0 : otherclauses = lappend(otherclauses, rinfo);
4202 0 : clauses = foreach_delete_current(clauses, lc);
4203 0 : continue;
4204 : }
4205 :
4206 30 : group_relid = relid;
4207 30 : group_rel = root->simple_rel_array[relid];
4208 : }
4209 30 : else if (group_relid != relid)
4210 : {
4211 : /*
4212 : * Being in the group forming state we don't need other
4213 : * clauses.
4214 : */
4215 0 : continue;
4216 : }
4217 :
4218 : /*
4219 : * We're going to add the new clause to the varinfos list. We
4220 : * might re-use add_unique_group_var(), but we don't do so for
4221 : * two reasons.
4222 : *
4223 : * 1) We must keep the origin_rinfos list ordered exactly the
4224 : * same way as varinfos.
4225 : *
4226 : * 2) add_unique_group_var() is designed for
4227 : * estimate_num_groups(), where a larger number of groups is
4228 : * worse. While estimating the number of hash buckets, we
4229 : * have the opposite: a lesser number of groups is worse.
4230 : * Therefore, we don't have to remove "known equal" vars: the
4231 : * removed var may valuably contribute to the multivariate
4232 : * statistics to grow the number of groups.
4233 : */
4234 :
4235 : /*
4236 : * Clear nullingrels to correctly match hash keys. See
4237 : * add_unique_group_var()'s comment for details.
4238 : */
4239 60 : expr = remove_nulling_relids(expr, root->outer_join_rels, NULL);
4240 :
4241 : /*
4242 : * Detect and exclude exact duplicates from the list of hash
4243 : * keys (like add_unique_group_var does).
4244 : */
4245 84 : foreach(lc1, varinfos)
4246 : {
4247 36 : varinfo = (GroupVarInfo *) lfirst(lc1);
4248 :
4249 36 : if (!equal(expr, varinfo->var))
4250 24 : continue;
4251 :
4252 12 : is_duplicate = true;
4253 12 : break;
4254 : }
4255 :
4256 60 : if (is_duplicate)
4257 : {
4258 : /*
4259 : * Skip exact duplicates. Adding them to the otherclauses
4260 : * list also doesn't make sense.
4261 : */
4262 12 : continue;
4263 : }
4264 :
4265 : /*
4266 : * Initialize GroupVarInfo. We only use it to call
4267 : * estimate_multivariate_ndistinct(), which doesn't care about
4268 : * ndistinct and isdefault fields. Thus, skip these fields.
4269 : */
4270 48 : varinfo = palloc0_object(GroupVarInfo);
4271 48 : varinfo->var = expr;
4272 48 : varinfo->rel = root->simple_rel_array[relid];
4273 48 : varinfos = lappend(varinfos, varinfo);
4274 :
4275 : /*
4276 : * Remember the link to RestrictInfo for the case the clause
4277 : * is failed to be estimated.
4278 : */
4279 48 : origin_rinfos = lappend(origin_rinfos, rinfo);
4280 : }
4281 : else
4282 : {
4283 : /* This clause can't be estimated with extended statistics */
4284 74060 : otherclauses = lappend(otherclauses, rinfo);
4285 : }
4286 :
4287 74108 : clauses = foreach_delete_current(clauses, lc);
4288 : }
4289 :
4290 36754 : if (list_length(varinfos) < 2)
4291 : {
4292 : /*
4293 : * Multivariate statistics doesn't apply to single columns except
4294 : * for expressions, but it has not been implemented yet.
4295 : */
4296 36742 : otherclauses = list_concat(otherclauses, origin_rinfos);
4297 36742 : list_free_deep(varinfos);
4298 36742 : list_free(origin_rinfos);
4299 36742 : continue;
4300 : }
4301 :
4302 : Assert(group_rel != NULL);
4303 :
4304 : /* Employ the extended statistics. */
4305 12 : origin_varinfos = varinfos;
4306 : for (;;)
4307 12 : {
4308 24 : bool estimated = estimate_multivariate_ndistinct(root,
4309 : group_rel,
4310 : &varinfos,
4311 : &mvndistinct);
4312 :
4313 24 : if (!estimated)
4314 12 : break;
4315 :
4316 : /*
4317 : * We've got an estimation. Use ndistinct value in a consistent
4318 : * way - according to the caller's logic (see
4319 : * final_cost_hashjoin).
4320 : */
4321 12 : if (ndistinct < mvndistinct)
4322 12 : ndistinct = mvndistinct;
4323 : Assert(ndistinct >= 1.0);
4324 : }
4325 :
4326 : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
4327 :
4328 : /* Collect unmatched clauses as otherclauses. */
4329 42 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
4330 : {
4331 30 : GroupVarInfo *vinfo = lfirst(lc1);
4332 :
4333 30 : if (!list_member_ptr(varinfos, vinfo))
4334 : /* Already estimated */
4335 30 : continue;
4336 :
4337 : /* Can't be estimated here - push to the returning list */
4338 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
4339 : }
4340 : }
4341 :
4342 36742 : *innerbucketsize = 1.0 / ndistinct;
4343 36742 : return otherclauses;
4344 : }
4345 :
4346 : /*
4347 : * Estimate hash bucket statistics when the specified expression is used
4348 : * as a hash key for the given number of buckets.
4349 : *
4350 : * This attempts to determine two values:
4351 : *
4352 : * 1. The frequency of the most common value of the expression (returns
4353 : * zero into *mcv_freq if we can't get that).
4354 : *
4355 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
4356 : * divided by total tuples in relation.
4357 : *
4358 : * XXX This is really pretty bogus since we're effectively assuming that the
4359 : * distribution of hash keys will be the same after applying restriction
4360 : * clauses as it was in the underlying relation. However, we are not nearly
4361 : * smart enough to figure out how the restrict clauses might change the
4362 : * distribution, so this will have to do for now.
4363 : *
4364 : * We are passed the number of buckets the executor will use for the given
4365 : * input relation. If the data were perfectly distributed, with the same
4366 : * number of tuples going into each available bucket, then the bucketsize
4367 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
4368 : * only if (a) there are at least nbuckets distinct data values, and (b)
4369 : * we have a not-too-skewed data distribution. Otherwise the buckets will
4370 : * be nonuniformly occupied. If the other relation in the join has a key
4371 : * distribution similar to this one's, then the most-loaded buckets are
4372 : * exactly those that will be probed most often. Therefore, the "average"
4373 : * bucket size for costing purposes should really be taken as something close
4374 : * to the "worst case" bucket size. We try to estimate this by adjusting the
4375 : * fraction if there are too few distinct data values, and then scaling up
4376 : * by the ratio of the most common value's frequency to the average frequency.
4377 : *
4378 : * If no statistics are available, use a default estimate of 0.1. This will
4379 : * discourage use of a hash rather strongly if the inner relation is large,
4380 : * which is what we want. We do not want to hash unless we know that the
4381 : * inner rel is well-dispersed (or the alternatives seem much worse).
4382 : *
4383 : * The caller should also check that the mcv_freq is not so large that the
4384 : * most common value would by itself require an impractically large bucket.
4385 : * In a hash join, the executor can split buckets if they get too big, but
4386 : * obviously that doesn't help for a bucket that contains many duplicates of
4387 : * the same value.
4388 : */
4389 : void
4390 205774 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
4391 : Selectivity *mcv_freq,
4392 : Selectivity *bucketsize_frac)
4393 : {
4394 : VariableStatData vardata;
4395 : double estfract,
4396 : ndistinct,
4397 : stanullfrac,
4398 : avgfreq;
4399 : bool isdefault;
4400 : AttStatsSlot sslot;
4401 :
4402 205774 : examine_variable(root, hashkey, 0, &vardata);
4403 :
4404 : /* Initialize *mcv_freq to "unknown" */
4405 205774 : *mcv_freq = 0.0;
4406 :
4407 : /* Look up the frequency of the most common value, if available */
4408 205774 : if (HeapTupleIsValid(vardata.statsTuple))
4409 : {
4410 148262 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4411 : STATISTIC_KIND_MCV, InvalidOid,
4412 : ATTSTATSSLOT_NUMBERS))
4413 : {
4414 : /*
4415 : * The first MCV stat is for the most common value.
4416 : */
4417 86660 : if (sslot.nnumbers > 0)
4418 86660 : *mcv_freq = sslot.numbers[0];
4419 86660 : free_attstatsslot(&sslot);
4420 : }
4421 61602 : else if (get_attstatsslot(&sslot, vardata.statsTuple,
4422 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
4423 : 0))
4424 : {
4425 : /*
4426 : * If there are no recorded MCVs, but we do have a histogram, then
4427 : * assume that ANALYZE determined that the column is unique.
4428 : */
4429 59296 : if (vardata.rel && vardata.rel->rows > 0)
4430 59278 : *mcv_freq = 1.0 / vardata.rel->rows;
4431 : }
4432 : }
4433 :
4434 : /* Get number of distinct values */
4435 205774 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4436 :
4437 : /*
4438 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4439 : * mcv_freq is known to be even higher than that, use it instead.
4440 : */
4441 205774 : if (isdefault)
4442 : {
4443 25724 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4444 25724 : ReleaseVariableStats(vardata);
4445 25724 : return;
4446 : }
4447 :
4448 : /* Get fraction that are null */
4449 180050 : if (HeapTupleIsValid(vardata.statsTuple))
4450 : {
4451 : Form_pg_statistic stats;
4452 :
4453 148244 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
4454 148244 : stanullfrac = stats->stanullfrac;
4455 : }
4456 : else
4457 31806 : stanullfrac = 0.0;
4458 :
4459 : /* Compute avg freq of all distinct data values in raw relation */
4460 180050 : avgfreq = (1.0 - stanullfrac) / ndistinct;
4461 :
4462 : /*
4463 : * Adjust ndistinct to account for restriction clauses. Observe we are
4464 : * assuming that the data distribution is affected uniformly by the
4465 : * restriction clauses!
4466 : *
4467 : * XXX Possibly better way, but much more expensive: multiply by
4468 : * selectivity of rel's restriction clauses that mention the target Var.
4469 : */
4470 180050 : if (vardata.rel && vardata.rel->tuples > 0)
4471 : {
4472 179994 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
4473 179994 : ndistinct = clamp_row_est(ndistinct);
4474 : }
4475 :
4476 : /*
4477 : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
4478 : * number of buckets is less than the expected number of distinct values;
4479 : * otherwise it is 1/ndistinct.
4480 : */
4481 180050 : if (ndistinct > nbuckets)
4482 84 : estfract = 1.0 / nbuckets;
4483 : else
4484 179966 : estfract = 1.0 / ndistinct;
4485 :
4486 : /*
4487 : * Adjust estimated bucketsize upward to account for skewed distribution.
4488 : */
4489 180050 : if (avgfreq > 0.0 && *mcv_freq > avgfreq)
4490 96028 : estfract *= *mcv_freq / avgfreq;
4491 :
4492 : /*
4493 : * Clamp bucketsize to sane range (the above adjustment could easily
4494 : * produce an out-of-range result). We set the lower bound a little above
4495 : * zero, since zero isn't a very sane result.
4496 : */
4497 180050 : if (estfract < 1.0e-6)
4498 0 : estfract = 1.0e-6;
4499 180050 : else if (estfract > 1.0)
4500 47776 : estfract = 1.0;
4501 :
4502 180050 : *bucketsize_frac = (Selectivity) estfract;
4503 :
4504 180050 : ReleaseVariableStats(vardata);
4505 : }
4506 :
4507 : /*
4508 : * estimate_hashagg_tablesize
4509 : * estimate the number of bytes that a hash aggregate hashtable will
4510 : * require based on the agg_costs, path width and number of groups.
4511 : *
4512 : * We return the result as "double" to forestall any possible overflow
4513 : * problem in the multiplication by dNumGroups.
4514 : *
4515 : * XXX this may be over-estimating the size now that hashagg knows to omit
4516 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4517 : * grouping columns not in the hashed set are counted here even though hashagg
4518 : * won't store them. Is this a problem?
4519 : */
4520 : double
4521 2614 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4522 : const AggClauseCosts *agg_costs, double dNumGroups)
4523 : {
4524 : Size hashentrysize;
4525 :
4526 2614 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4527 2614 : path->pathtarget->width,
4528 2614 : agg_costs->transitionSpace);
4529 :
4530 : /*
4531 : * Note that this disregards the effect of fill-factor and growth policy
4532 : * of the hash table. That's probably ok, given that the default
4533 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4534 : * "double-in-size" growth policies here.
4535 : */
4536 2614 : return hashentrysize * dNumGroups;
4537 : }
4538 :
4539 :
4540 : /*-------------------------------------------------------------------------
4541 : *
4542 : * Support routines
4543 : *
4544 : *-------------------------------------------------------------------------
4545 : */
4546 :
4547 : /*
4548 : * Find the best matching ndistinct extended statistics for the given list of
4549 : * GroupVarInfos.
4550 : *
4551 : * Callers must ensure that the given GroupVarInfos all belong to 'rel' and
4552 : * the GroupVarInfos list does not contain any duplicate Vars or expressions.
4553 : *
4554 : * When statistics are found that match > 1 of the given GroupVarInfo, the
4555 : * *ndistinct parameter is set according to the ndistinct estimate and a new
4556 : * list is built with the matching GroupVarInfos removed, which is output via
4557 : * the *varinfos parameter before returning true. When no matching stats are
4558 : * found, false is returned and the *varinfos and *ndistinct parameters are
4559 : * left untouched.
4560 : */
4561 : static bool
4562 347162 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4563 : List **varinfos, double *ndistinct)
4564 : {
4565 : ListCell *lc;
4566 : int nmatches_vars;
4567 : int nmatches_exprs;
4568 347162 : Oid statOid = InvalidOid;
4569 : MVNDistinct *stats;
4570 347162 : StatisticExtInfo *matched_info = NULL;
4571 347162 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4572 :
4573 : /* bail out immediately if the table has no extended statistics */
4574 347162 : if (!rel->statlist)
4575 346598 : return false;
4576 :
4577 : /* look for the ndistinct statistics object matching the most vars */
4578 564 : nmatches_vars = 0; /* we require at least two matches */
4579 564 : nmatches_exprs = 0;
4580 2244 : foreach(lc, rel->statlist)
4581 : {
4582 : ListCell *lc2;
4583 1680 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
4584 1680 : int nshared_vars = 0;
4585 1680 : int nshared_exprs = 0;
4586 :
4587 : /* skip statistics of other kinds */
4588 1680 : if (info->kind != STATS_EXT_NDISTINCT)
4589 792 : continue;
4590 :
4591 : /* skip statistics with mismatching stxdinherit value */
4592 888 : if (info->inherit != rte->inh)
4593 30 : continue;
4594 :
4595 : /*
4596 : * Determine how many expressions (and variables in non-matched
4597 : * expressions) match. We'll then use these numbers to pick the
4598 : * statistics object that best matches the clauses.
4599 : */
4600 2718 : foreach(lc2, *varinfos)
4601 : {
4602 : ListCell *lc3;
4603 1860 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4604 : AttrNumber attnum;
4605 :
4606 : Assert(varinfo->rel == rel);
4607 :
4608 : /* simple Var, search in statistics keys directly */
4609 1860 : if (IsA(varinfo->var, Var))
4610 : {
4611 1494 : attnum = ((Var *) varinfo->var)->varattno;
4612 :
4613 : /*
4614 : * Ignore system attributes - we don't support statistics on
4615 : * them, so can't match them (and it'd fail as the values are
4616 : * negative).
4617 : */
4618 1494 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4619 12 : continue;
4620 :
4621 1482 : if (bms_is_member(attnum, info->keys))
4622 876 : nshared_vars++;
4623 :
4624 1482 : continue;
4625 : }
4626 :
4627 : /* expression - see if it's in the statistics object */
4628 660 : foreach(lc3, info->exprs)
4629 : {
4630 528 : Node *expr = (Node *) lfirst(lc3);
4631 :
4632 528 : if (equal(varinfo->var, expr))
4633 : {
4634 234 : nshared_exprs++;
4635 234 : break;
4636 : }
4637 : }
4638 : }
4639 :
4640 : /*
4641 : * The ndistinct extended statistics contain estimates for a minimum
4642 : * of pairs of columns which the statistics are defined on and
4643 : * certainly not single columns. Here we skip unless we managed to
4644 : * match to at least two columns.
4645 : */
4646 858 : if (nshared_vars + nshared_exprs < 2)
4647 396 : continue;
4648 :
4649 : /*
4650 : * Check if these statistics are a better match than the previous best
4651 : * match and if so, take note of the StatisticExtInfo.
4652 : *
4653 : * The statslist is sorted by statOid, so the StatisticExtInfo we
4654 : * select as the best match is deterministic even when multiple sets
4655 : * of statistics match equally as well.
4656 : */
4657 462 : if ((nshared_exprs > nmatches_exprs) ||
4658 354 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4659 : {
4660 438 : statOid = info->statOid;
4661 438 : nmatches_vars = nshared_vars;
4662 438 : nmatches_exprs = nshared_exprs;
4663 438 : matched_info = info;
4664 : }
4665 : }
4666 :
4667 : /* No match? */
4668 564 : if (statOid == InvalidOid)
4669 138 : return false;
4670 :
4671 : Assert(nmatches_vars + nmatches_exprs > 1);
4672 :
4673 426 : stats = statext_ndistinct_load(statOid, rte->inh);
4674 :
4675 : /*
4676 : * If we have a match, search it for the specific item that matches (there
4677 : * must be one), and construct the output values.
4678 : */
4679 426 : if (stats)
4680 : {
4681 : int i;
4682 426 : List *newlist = NIL;
4683 426 : MVNDistinctItem *item = NULL;
4684 : ListCell *lc2;
4685 426 : Bitmapset *matched = NULL;
4686 : AttrNumber attnum_offset;
4687 :
4688 : /*
4689 : * How much we need to offset the attnums? If there are no
4690 : * expressions, no offset is needed. Otherwise offset enough to move
4691 : * the lowest one (which is equal to number of expressions) to 1.
4692 : */
4693 426 : if (matched_info->exprs)
4694 150 : attnum_offset = (list_length(matched_info->exprs) + 1);
4695 : else
4696 276 : attnum_offset = 0;
4697 :
4698 : /* see what actually matched */
4699 1488 : foreach(lc2, *varinfos)
4700 : {
4701 : ListCell *lc3;
4702 : int idx;
4703 1062 : bool found = false;
4704 :
4705 1062 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4706 :
4707 : /*
4708 : * Process a simple Var expression, by matching it to keys
4709 : * directly. If there's a matching expression, we'll try matching
4710 : * it later.
4711 : */
4712 1062 : if (IsA(varinfo->var, Var))
4713 : {
4714 876 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4715 :
4716 : /*
4717 : * Ignore expressions on system attributes. Can't rely on the
4718 : * bms check for negative values.
4719 : */
4720 876 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4721 6 : continue;
4722 :
4723 : /* Is the variable covered by the statistics object? */
4724 870 : if (!bms_is_member(attnum, matched_info->keys))
4725 120 : continue;
4726 :
4727 750 : attnum = attnum + attnum_offset;
4728 :
4729 : /* ensure sufficient offset */
4730 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4731 :
4732 750 : matched = bms_add_member(matched, attnum);
4733 :
4734 750 : found = true;
4735 : }
4736 :
4737 : /*
4738 : * XXX Maybe we should allow searching the expressions even if we
4739 : * found an attribute matching the expression? That would handle
4740 : * trivial expressions like "(a)" but it seems fairly useless.
4741 : */
4742 936 : if (found)
4743 750 : continue;
4744 :
4745 : /* expression - see if it's in the statistics object */
4746 186 : idx = 0;
4747 306 : foreach(lc3, matched_info->exprs)
4748 : {
4749 276 : Node *expr = (Node *) lfirst(lc3);
4750 :
4751 276 : if (equal(varinfo->var, expr))
4752 : {
4753 156 : AttrNumber attnum = -(idx + 1);
4754 :
4755 156 : attnum = attnum + attnum_offset;
4756 :
4757 : /* ensure sufficient offset */
4758 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4759 :
4760 156 : matched = bms_add_member(matched, attnum);
4761 :
4762 : /* there should be just one matching expression */
4763 156 : break;
4764 : }
4765 :
4766 120 : idx++;
4767 : }
4768 : }
4769 :
4770 : /* Find the specific item that exactly matches the combination */
4771 864 : for (i = 0; i < stats->nitems; i++)
4772 : {
4773 : int j;
4774 864 : MVNDistinctItem *tmpitem = &stats->items[i];
4775 :
4776 864 : if (tmpitem->nattributes != bms_num_members(matched))
4777 162 : continue;
4778 :
4779 : /* assume it's the right item */
4780 702 : item = tmpitem;
4781 :
4782 : /* check that all item attributes/expressions fit the match */
4783 1692 : for (j = 0; j < tmpitem->nattributes; j++)
4784 : {
4785 1266 : AttrNumber attnum = tmpitem->attributes[j];
4786 :
4787 : /*
4788 : * Thanks to how we constructed the matched bitmap above, we
4789 : * can just offset all attnums the same way.
4790 : */
4791 1266 : attnum = attnum + attnum_offset;
4792 :
4793 1266 : if (!bms_is_member(attnum, matched))
4794 : {
4795 : /* nah, it's not this item */
4796 276 : item = NULL;
4797 276 : break;
4798 : }
4799 : }
4800 :
4801 : /*
4802 : * If the item has all the matched attributes, we know it's the
4803 : * right one - there can't be a better one. matching more.
4804 : */
4805 702 : if (item)
4806 426 : break;
4807 : }
4808 :
4809 : /*
4810 : * Make sure we found an item. There has to be one, because ndistinct
4811 : * statistics includes all combinations of attributes.
4812 : */
4813 426 : if (!item)
4814 0 : elog(ERROR, "corrupt MVNDistinct entry");
4815 :
4816 : /* Form the output varinfo list, keeping only unmatched ones */
4817 1488 : foreach(lc, *varinfos)
4818 : {
4819 1062 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4820 : ListCell *lc3;
4821 1062 : bool found = false;
4822 :
4823 : /*
4824 : * Let's look at plain variables first, because it's the most
4825 : * common case and the check is quite cheap. We can simply get the
4826 : * attnum and check (with an offset) matched bitmap.
4827 : */
4828 1062 : if (IsA(varinfo->var, Var))
4829 870 : {
4830 876 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4831 :
4832 : /*
4833 : * If it's a system attribute, we're done. We don't support
4834 : * extended statistics on system attributes, so it's clearly
4835 : * not matched. Just keep the expression and continue.
4836 : */
4837 876 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4838 : {
4839 6 : newlist = lappend(newlist, varinfo);
4840 6 : continue;
4841 : }
4842 :
4843 : /* apply the same offset as above */
4844 870 : attnum += attnum_offset;
4845 :
4846 : /* if it's not matched, keep the varinfo */
4847 870 : if (!bms_is_member(attnum, matched))
4848 120 : newlist = lappend(newlist, varinfo);
4849 :
4850 : /* The rest of the loop deals with complex expressions. */
4851 870 : continue;
4852 : }
4853 :
4854 : /*
4855 : * Process complex expressions, not just simple Vars.
4856 : *
4857 : * First, we search for an exact match of an expression. If we
4858 : * find one, we can just discard the whole GroupVarInfo, with all
4859 : * the variables we extracted from it.
4860 : *
4861 : * Otherwise we inspect the individual vars, and try matching it
4862 : * to variables in the item.
4863 : */
4864 306 : foreach(lc3, matched_info->exprs)
4865 : {
4866 276 : Node *expr = (Node *) lfirst(lc3);
4867 :
4868 276 : if (equal(varinfo->var, expr))
4869 : {
4870 156 : found = true;
4871 156 : break;
4872 : }
4873 : }
4874 :
4875 : /* found exact match, skip */
4876 186 : if (found)
4877 156 : continue;
4878 :
4879 30 : newlist = lappend(newlist, varinfo);
4880 : }
4881 :
4882 426 : *varinfos = newlist;
4883 426 : *ndistinct = item->ndistinct;
4884 426 : return true;
4885 : }
4886 :
4887 0 : return false;
4888 : }
4889 :
4890 : /*
4891 : * convert_to_scalar
4892 : * Convert non-NULL values of the indicated types to the comparison
4893 : * scale needed by scalarineqsel().
4894 : * Returns "true" if successful.
4895 : *
4896 : * XXX this routine is a hack: ideally we should look up the conversion
4897 : * subroutines in pg_type.
4898 : *
4899 : * All numeric datatypes are simply converted to their equivalent
4900 : * "double" values. (NUMERIC values that are outside the range of "double"
4901 : * are clamped to +/- HUGE_VAL.)
4902 : *
4903 : * String datatypes are converted by convert_string_to_scalar(),
4904 : * which is explained below. The reason why this routine deals with
4905 : * three values at a time, not just one, is that we need it for strings.
4906 : *
4907 : * The bytea datatype is just enough different from strings that it has
4908 : * to be treated separately.
4909 : *
4910 : * The several datatypes representing absolute times are all converted
4911 : * to Timestamp, which is actually an int64, and then we promote that to
4912 : * a double. Note this will give correct results even for the "special"
4913 : * values of Timestamp, since those are chosen to compare correctly;
4914 : * see timestamp_cmp.
4915 : *
4916 : * The several datatypes representing relative times (intervals) are all
4917 : * converted to measurements expressed in seconds.
4918 : */
4919 : static bool
4920 89972 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4921 : Datum lobound, Datum hibound, Oid boundstypid,
4922 : double *scaledlobound, double *scaledhibound)
4923 : {
4924 89972 : bool failure = false;
4925 :
4926 : /*
4927 : * Both the valuetypid and the boundstypid should exactly match the
4928 : * declared input type(s) of the operator we are invoked for. However,
4929 : * extensions might try to use scalarineqsel as estimator for operators
4930 : * with input type(s) we don't handle here; in such cases, we want to
4931 : * return false, not fail. In any case, we mustn't assume that valuetypid
4932 : * and boundstypid are identical.
4933 : *
4934 : * XXX The histogram we are interpolating between points of could belong
4935 : * to a column that's only binary-compatible with the declared type. In
4936 : * essence we are assuming that the semantics of binary-compatible types
4937 : * are enough alike that we can use a histogram generated with one type's
4938 : * operators to estimate selectivity for the other's. This is outright
4939 : * wrong in some cases --- in particular signed versus unsigned
4940 : * interpretation could trip us up. But it's useful enough in the
4941 : * majority of cases that we do it anyway. Should think about more
4942 : * rigorous ways to do it.
4943 : */
4944 89972 : switch (valuetypid)
4945 : {
4946 : /*
4947 : * Built-in numeric types
4948 : */
4949 82814 : case BOOLOID:
4950 : case INT2OID:
4951 : case INT4OID:
4952 : case INT8OID:
4953 : case FLOAT4OID:
4954 : case FLOAT8OID:
4955 : case NUMERICOID:
4956 : case OIDOID:
4957 : case REGPROCOID:
4958 : case REGPROCEDUREOID:
4959 : case REGOPEROID:
4960 : case REGOPERATOROID:
4961 : case REGCLASSOID:
4962 : case REGTYPEOID:
4963 : case REGCOLLATIONOID:
4964 : case REGCONFIGOID:
4965 : case REGDICTIONARYOID:
4966 : case REGROLEOID:
4967 : case REGNAMESPACEOID:
4968 : case REGDATABASEOID:
4969 82814 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4970 : &failure);
4971 82814 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4972 : &failure);
4973 82814 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4974 : &failure);
4975 82814 : return !failure;
4976 :
4977 : /*
4978 : * Built-in string types
4979 : */
4980 7158 : case CHAROID:
4981 : case BPCHAROID:
4982 : case VARCHAROID:
4983 : case TEXTOID:
4984 : case NAMEOID:
4985 : {
4986 7158 : char *valstr = convert_string_datum(value, valuetypid,
4987 : collid, &failure);
4988 7158 : char *lostr = convert_string_datum(lobound, boundstypid,
4989 : collid, &failure);
4990 7158 : char *histr = convert_string_datum(hibound, boundstypid,
4991 : collid, &failure);
4992 :
4993 : /*
4994 : * Bail out if any of the values is not of string type. We
4995 : * might leak converted strings for the other value(s), but
4996 : * that's not worth troubling over.
4997 : */
4998 7158 : if (failure)
4999 0 : return false;
5000 :
5001 7158 : convert_string_to_scalar(valstr, scaledvalue,
5002 : lostr, scaledlobound,
5003 : histr, scaledhibound);
5004 7158 : pfree(valstr);
5005 7158 : pfree(lostr);
5006 7158 : pfree(histr);
5007 7158 : return true;
5008 : }
5009 :
5010 : /*
5011 : * Built-in bytea type
5012 : */
5013 0 : case BYTEAOID:
5014 : {
5015 : /* We only support bytea vs bytea comparison */
5016 0 : if (boundstypid != BYTEAOID)
5017 0 : return false;
5018 0 : convert_bytea_to_scalar(value, scaledvalue,
5019 : lobound, scaledlobound,
5020 : hibound, scaledhibound);
5021 0 : return true;
5022 : }
5023 :
5024 : /*
5025 : * Built-in time types
5026 : */
5027 0 : case TIMESTAMPOID:
5028 : case TIMESTAMPTZOID:
5029 : case DATEOID:
5030 : case INTERVALOID:
5031 : case TIMEOID:
5032 : case TIMETZOID:
5033 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
5034 : &failure);
5035 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
5036 : &failure);
5037 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
5038 : &failure);
5039 0 : return !failure;
5040 :
5041 : /*
5042 : * Built-in network types
5043 : */
5044 0 : case INETOID:
5045 : case CIDROID:
5046 : case MACADDROID:
5047 : case MACADDR8OID:
5048 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
5049 : &failure);
5050 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
5051 : &failure);
5052 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
5053 : &failure);
5054 0 : return !failure;
5055 : }
5056 : /* Don't know how to convert */
5057 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
5058 0 : return false;
5059 : }
5060 :
5061 : /*
5062 : * Do convert_to_scalar()'s work for any numeric data type.
5063 : *
5064 : * On failure (e.g., unsupported typid), set *failure to true;
5065 : * otherwise, that variable is not changed.
5066 : */
5067 : static double
5068 248442 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
5069 : {
5070 248442 : switch (typid)
5071 : {
5072 0 : case BOOLOID:
5073 0 : return (double) DatumGetBool(value);
5074 12 : case INT2OID:
5075 12 : return (double) DatumGetInt16(value);
5076 31842 : case INT4OID:
5077 31842 : return (double) DatumGetInt32(value);
5078 0 : case INT8OID:
5079 0 : return (double) DatumGetInt64(value);
5080 0 : case FLOAT4OID:
5081 0 : return (double) DatumGetFloat4(value);
5082 54 : case FLOAT8OID:
5083 54 : return (double) DatumGetFloat8(value);
5084 0 : case NUMERICOID:
5085 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
5086 0 : return (double)
5087 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
5088 : value));
5089 216534 : case OIDOID:
5090 : case REGPROCOID:
5091 : case REGPROCEDUREOID:
5092 : case REGOPEROID:
5093 : case REGOPERATOROID:
5094 : case REGCLASSOID:
5095 : case REGTYPEOID:
5096 : case REGCOLLATIONOID:
5097 : case REGCONFIGOID:
5098 : case REGDICTIONARYOID:
5099 : case REGROLEOID:
5100 : case REGNAMESPACEOID:
5101 : case REGDATABASEOID:
5102 : /* we can treat OIDs as integers... */
5103 216534 : return (double) DatumGetObjectId(value);
5104 : }
5105 :
5106 0 : *failure = true;
5107 0 : return 0;
5108 : }
5109 :
5110 : /*
5111 : * Do convert_to_scalar()'s work for any character-string data type.
5112 : *
5113 : * String datatypes are converted to a scale that ranges from 0 to 1,
5114 : * where we visualize the bytes of the string as fractional digits.
5115 : *
5116 : * We do not want the base to be 256, however, since that tends to
5117 : * generate inflated selectivity estimates; few databases will have
5118 : * occurrences of all 256 possible byte values at each position.
5119 : * Instead, use the smallest and largest byte values seen in the bounds
5120 : * as the estimated range for each byte, after some fudging to deal with
5121 : * the fact that we probably aren't going to see the full range that way.
5122 : *
5123 : * An additional refinement is that we discard any common prefix of the
5124 : * three strings before computing the scaled values. This allows us to
5125 : * "zoom in" when we encounter a narrow data range. An example is a phone
5126 : * number database where all the values begin with the same area code.
5127 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
5128 : * so this is more likely to happen than you might think.)
5129 : */
5130 : static void
5131 7158 : convert_string_to_scalar(char *value,
5132 : double *scaledvalue,
5133 : char *lobound,
5134 : double *scaledlobound,
5135 : char *hibound,
5136 : double *scaledhibound)
5137 : {
5138 : int rangelo,
5139 : rangehi;
5140 : char *sptr;
5141 :
5142 7158 : rangelo = rangehi = (unsigned char) hibound[0];
5143 89742 : for (sptr = lobound; *sptr; sptr++)
5144 : {
5145 82584 : if (rangelo > (unsigned char) *sptr)
5146 16798 : rangelo = (unsigned char) *sptr;
5147 82584 : if (rangehi < (unsigned char) *sptr)
5148 8912 : rangehi = (unsigned char) *sptr;
5149 : }
5150 83552 : for (sptr = hibound; *sptr; sptr++)
5151 : {
5152 76394 : if (rangelo > (unsigned char) *sptr)
5153 1248 : rangelo = (unsigned char) *sptr;
5154 76394 : if (rangehi < (unsigned char) *sptr)
5155 3372 : rangehi = (unsigned char) *sptr;
5156 : }
5157 : /* If range includes any upper-case ASCII chars, make it include all */
5158 7158 : if (rangelo <= 'Z' && rangehi >= 'A')
5159 : {
5160 1516 : if (rangelo > 'A')
5161 222 : rangelo = 'A';
5162 1516 : if (rangehi < 'Z')
5163 480 : rangehi = 'Z';
5164 : }
5165 : /* Ditto lower-case */
5166 7158 : if (rangelo <= 'z' && rangehi >= 'a')
5167 : {
5168 6656 : if (rangelo > 'a')
5169 102 : rangelo = 'a';
5170 6656 : if (rangehi < 'z')
5171 6566 : rangehi = 'z';
5172 : }
5173 : /* Ditto digits */
5174 7158 : if (rangelo <= '9' && rangehi >= '0')
5175 : {
5176 836 : if (rangelo > '0')
5177 732 : rangelo = '0';
5178 836 : if (rangehi < '9')
5179 14 : rangehi = '9';
5180 : }
5181 :
5182 : /*
5183 : * If range includes less than 10 chars, assume we have not got enough
5184 : * data, and make it include regular ASCII set.
5185 : */
5186 7158 : if (rangehi - rangelo < 9)
5187 : {
5188 0 : rangelo = ' ';
5189 0 : rangehi = 127;
5190 : }
5191 :
5192 : /*
5193 : * Now strip any common prefix of the three strings.
5194 : */
5195 15374 : while (*lobound)
5196 : {
5197 15354 : if (*lobound != *hibound || *lobound != *value)
5198 : break;
5199 8216 : lobound++, hibound++, value++;
5200 : }
5201 :
5202 : /*
5203 : * Now we can do the conversions.
5204 : */
5205 7158 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
5206 7158 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
5207 7158 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
5208 7158 : }
5209 :
5210 : static double
5211 21474 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
5212 : {
5213 21474 : int slen = strlen(value);
5214 : double num,
5215 : denom,
5216 : base;
5217 :
5218 21474 : if (slen <= 0)
5219 20 : return 0.0; /* empty string has scalar value 0 */
5220 :
5221 : /*
5222 : * There seems little point in considering more than a dozen bytes from
5223 : * the string. Since base is at least 10, that will give us nominal
5224 : * resolution of at least 12 decimal digits, which is surely far more
5225 : * precision than this estimation technique has got anyway (especially in
5226 : * non-C locales). Also, even with the maximum possible base of 256, this
5227 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
5228 : * overflow on any known machine.
5229 : */
5230 21454 : if (slen > 12)
5231 5390 : slen = 12;
5232 :
5233 : /* Convert initial characters to fraction */
5234 21454 : base = rangehi - rangelo + 1;
5235 21454 : num = 0.0;
5236 21454 : denom = base;
5237 176840 : while (slen-- > 0)
5238 : {
5239 155386 : int ch = (unsigned char) *value++;
5240 :
5241 155386 : if (ch < rangelo)
5242 160 : ch = rangelo - 1;
5243 155226 : else if (ch > rangehi)
5244 0 : ch = rangehi + 1;
5245 155386 : num += ((double) (ch - rangelo)) / denom;
5246 155386 : denom *= base;
5247 : }
5248 :
5249 21454 : return num;
5250 : }
5251 :
5252 : /*
5253 : * Convert a string-type Datum into a palloc'd, null-terminated string.
5254 : *
5255 : * On failure (e.g., unsupported typid), set *failure to true;
5256 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
5257 : *
5258 : * When using a non-C locale, we must pass the string through pg_strxfrm()
5259 : * before continuing, so as to generate correct locale-specific results.
5260 : */
5261 : static char *
5262 21474 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
5263 : {
5264 : char *val;
5265 : pg_locale_t mylocale;
5266 :
5267 21474 : switch (typid)
5268 : {
5269 0 : case CHAROID:
5270 0 : val = (char *) palloc(2);
5271 0 : val[0] = DatumGetChar(value);
5272 0 : val[1] = '\0';
5273 0 : break;
5274 6622 : case BPCHAROID:
5275 : case VARCHAROID:
5276 : case TEXTOID:
5277 6622 : val = TextDatumGetCString(value);
5278 6622 : break;
5279 14852 : case NAMEOID:
5280 : {
5281 14852 : NameData *nm = (NameData *) DatumGetPointer(value);
5282 :
5283 14852 : val = pstrdup(NameStr(*nm));
5284 14852 : break;
5285 : }
5286 0 : default:
5287 0 : *failure = true;
5288 0 : return NULL;
5289 : }
5290 :
5291 21474 : mylocale = pg_newlocale_from_collation(collid);
5292 :
5293 21474 : if (!mylocale->collate_is_c)
5294 : {
5295 : char *xfrmstr;
5296 : size_t xfrmlen;
5297 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
5298 :
5299 : /*
5300 : * XXX: We could guess at a suitable output buffer size and only call
5301 : * pg_strxfrm() twice if our guess is too small.
5302 : *
5303 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
5304 : * bogus data or set an error. This is not really a problem unless it
5305 : * crashes since it will only give an estimation error and nothing
5306 : * fatal.
5307 : *
5308 : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
5309 : * some cases, libc strxfrm() may return the wrong results, but that
5310 : * will only lead to an estimation error.
5311 : */
5312 72 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
5313 : #ifdef WIN32
5314 :
5315 : /*
5316 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
5317 : * of trying to allocate this much memory (and fail), just return the
5318 : * original string unmodified as if we were in the C locale.
5319 : */
5320 : if (xfrmlen == INT_MAX)
5321 : return val;
5322 : #endif
5323 72 : xfrmstr = (char *) palloc(xfrmlen + 1);
5324 72 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
5325 :
5326 : /*
5327 : * Some systems (e.g., glibc) can return a smaller value from the
5328 : * second call than the first; thus the Assert must be <= not ==.
5329 : */
5330 : Assert(xfrmlen2 <= xfrmlen);
5331 72 : pfree(val);
5332 72 : val = xfrmstr;
5333 : }
5334 :
5335 21474 : return val;
5336 : }
5337 :
5338 : /*
5339 : * Do convert_to_scalar()'s work for any bytea data type.
5340 : *
5341 : * Very similar to convert_string_to_scalar except we can't assume
5342 : * null-termination and therefore pass explicit lengths around.
5343 : *
5344 : * Also, assumptions about likely "normal" ranges of characters have been
5345 : * removed - a data range of 0..255 is always used, for now. (Perhaps
5346 : * someday we will add information about actual byte data range to
5347 : * pg_statistic.)
5348 : */
5349 : static void
5350 0 : convert_bytea_to_scalar(Datum value,
5351 : double *scaledvalue,
5352 : Datum lobound,
5353 : double *scaledlobound,
5354 : Datum hibound,
5355 : double *scaledhibound)
5356 : {
5357 0 : bytea *valuep = DatumGetByteaPP(value);
5358 0 : bytea *loboundp = DatumGetByteaPP(lobound);
5359 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
5360 : int rangelo,
5361 : rangehi,
5362 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
5363 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
5364 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
5365 : i,
5366 : minlen;
5367 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
5368 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
5369 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
5370 :
5371 : /*
5372 : * Assume bytea data is uniformly distributed across all byte values.
5373 : */
5374 0 : rangelo = 0;
5375 0 : rangehi = 255;
5376 :
5377 : /*
5378 : * Now strip any common prefix of the three strings.
5379 : */
5380 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
5381 0 : for (i = 0; i < minlen; i++)
5382 : {
5383 0 : if (*lostr != *histr || *lostr != *valstr)
5384 : break;
5385 0 : lostr++, histr++, valstr++;
5386 0 : loboundlen--, hiboundlen--, valuelen--;
5387 : }
5388 :
5389 : /*
5390 : * Now we can do the conversions.
5391 : */
5392 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
5393 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
5394 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
5395 0 : }
5396 :
5397 : static double
5398 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
5399 : int rangelo, int rangehi)
5400 : {
5401 : double num,
5402 : denom,
5403 : base;
5404 :
5405 0 : if (valuelen <= 0)
5406 0 : return 0.0; /* empty string has scalar value 0 */
5407 :
5408 : /*
5409 : * Since base is 256, need not consider more than about 10 chars (even
5410 : * this many seems like overkill)
5411 : */
5412 0 : if (valuelen > 10)
5413 0 : valuelen = 10;
5414 :
5415 : /* Convert initial characters to fraction */
5416 0 : base = rangehi - rangelo + 1;
5417 0 : num = 0.0;
5418 0 : denom = base;
5419 0 : while (valuelen-- > 0)
5420 : {
5421 0 : int ch = *value++;
5422 :
5423 0 : if (ch < rangelo)
5424 0 : ch = rangelo - 1;
5425 0 : else if (ch > rangehi)
5426 0 : ch = rangehi + 1;
5427 0 : num += ((double) (ch - rangelo)) / denom;
5428 0 : denom *= base;
5429 : }
5430 :
5431 0 : return num;
5432 : }
5433 :
5434 : /*
5435 : * Do convert_to_scalar()'s work for any timevalue data type.
5436 : *
5437 : * On failure (e.g., unsupported typid), set *failure to true;
5438 : * otherwise, that variable is not changed.
5439 : */
5440 : static double
5441 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5442 : {
5443 0 : switch (typid)
5444 : {
5445 0 : case TIMESTAMPOID:
5446 0 : return DatumGetTimestamp(value);
5447 0 : case TIMESTAMPTZOID:
5448 0 : return DatumGetTimestampTz(value);
5449 0 : case DATEOID:
5450 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
5451 0 : case INTERVALOID:
5452 : {
5453 0 : Interval *interval = DatumGetIntervalP(value);
5454 :
5455 : /*
5456 : * Convert the month part of Interval to days using assumed
5457 : * average month length of 365.25/12.0 days. Not too
5458 : * accurate, but plenty good enough for our purposes.
5459 : *
5460 : * This also works for infinite intervals, which just have all
5461 : * fields set to INT_MIN/INT_MAX, and so will produce a result
5462 : * smaller/larger than any finite interval.
5463 : */
5464 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5465 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5466 : }
5467 0 : case TIMEOID:
5468 0 : return DatumGetTimeADT(value);
5469 0 : case TIMETZOID:
5470 : {
5471 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5472 :
5473 : /* use GMT-equivalent time */
5474 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5475 : }
5476 : }
5477 :
5478 0 : *failure = true;
5479 0 : return 0;
5480 : }
5481 :
5482 :
5483 : /*
5484 : * get_restriction_variable
5485 : * Examine the args of a restriction clause to see if it's of the
5486 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5487 : * where "variable" could be either a Var or an expression in vars of a
5488 : * single relation. If so, extract information about the variable,
5489 : * and also indicate which side it was on and the other argument.
5490 : *
5491 : * Inputs:
5492 : * root: the planner info
5493 : * args: clause argument list
5494 : * varRelid: see specs for restriction selectivity functions
5495 : *
5496 : * Outputs: (these are valid only if true is returned)
5497 : * *vardata: gets information about variable (see examine_variable)
5498 : * *other: gets other clause argument, aggressively reduced to a constant
5499 : * *varonleft: set true if variable is on the left, false if on the right
5500 : *
5501 : * Returns true if a variable is identified, otherwise false.
5502 : *
5503 : * Note: if there are Vars on both sides of the clause, we must fail, because
5504 : * callers are expecting that the other side will act like a pseudoconstant.
5505 : */
5506 : bool
5507 821598 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5508 : VariableStatData *vardata, Node **other,
5509 : bool *varonleft)
5510 : {
5511 : Node *left,
5512 : *right;
5513 : VariableStatData rdata;
5514 :
5515 : /* Fail if not a binary opclause (probably shouldn't happen) */
5516 821598 : if (list_length(args) != 2)
5517 0 : return false;
5518 :
5519 821598 : left = (Node *) linitial(args);
5520 821598 : right = (Node *) lsecond(args);
5521 :
5522 : /*
5523 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5524 : * relations will be treated as pseudoconstants.
5525 : */
5526 821598 : examine_variable(root, left, varRelid, vardata);
5527 821598 : examine_variable(root, right, varRelid, &rdata);
5528 :
5529 : /*
5530 : * If one side is a variable and the other not, we win.
5531 : */
5532 821598 : if (vardata->rel && rdata.rel == NULL)
5533 : {
5534 733046 : *varonleft = true;
5535 733046 : *other = estimate_expression_value(root, rdata.var);
5536 : /* Assume we need no ReleaseVariableStats(rdata) here */
5537 733040 : return true;
5538 : }
5539 :
5540 88552 : if (vardata->rel == NULL && rdata.rel)
5541 : {
5542 82340 : *varonleft = false;
5543 82340 : *other = estimate_expression_value(root, vardata->var);
5544 : /* Assume we need no ReleaseVariableStats(*vardata) here */
5545 82340 : *vardata = rdata;
5546 82340 : return true;
5547 : }
5548 :
5549 : /* Oops, clause has wrong structure (probably var op var) */
5550 6212 : ReleaseVariableStats(*vardata);
5551 6212 : ReleaseVariableStats(rdata);
5552 :
5553 6212 : return false;
5554 : }
5555 :
5556 : /*
5557 : * get_join_variables
5558 : * Apply examine_variable() to each side of a join clause.
5559 : * Also, attempt to identify whether the join clause has the same
5560 : * or reversed sense compared to the SpecialJoinInfo.
5561 : *
5562 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5563 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5564 : * where we can't tell for sure, we default to assuming it's normal.
5565 : */
5566 : void
5567 269792 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5568 : VariableStatData *vardata1, VariableStatData *vardata2,
5569 : bool *join_is_reversed)
5570 : {
5571 : Node *left,
5572 : *right;
5573 :
5574 269792 : if (list_length(args) != 2)
5575 0 : elog(ERROR, "join operator should take two arguments");
5576 :
5577 269792 : left = (Node *) linitial(args);
5578 269792 : right = (Node *) lsecond(args);
5579 :
5580 269792 : examine_variable(root, left, 0, vardata1);
5581 269792 : examine_variable(root, right, 0, vardata2);
5582 :
5583 539330 : if (vardata1->rel &&
5584 269538 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
5585 91058 : *join_is_reversed = true; /* var1 is on RHS */
5586 357222 : else if (vardata2->rel &&
5587 178488 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
5588 206 : *join_is_reversed = true; /* var2 is on LHS */
5589 : else
5590 178528 : *join_is_reversed = false;
5591 269792 : }
5592 :
5593 : /* statext_expressions_load copies the tuple, so just pfree it. */
5594 : static void
5595 1650 : ReleaseDummy(HeapTuple tuple)
5596 : {
5597 1650 : pfree(tuple);
5598 1650 : }
5599 :
5600 : /*
5601 : * examine_variable
5602 : * Try to look up statistical data about an expression.
5603 : * Fill in a VariableStatData struct to describe the expression.
5604 : *
5605 : * Inputs:
5606 : * root: the planner info
5607 : * node: the expression tree to examine
5608 : * varRelid: see specs for restriction selectivity functions
5609 : *
5610 : * Outputs: *vardata is filled as follows:
5611 : * var: the input expression (with any phvs or binary relabeling stripped,
5612 : * if it is or contains a variable; but otherwise unchanged)
5613 : * rel: RelOptInfo for relation containing variable; NULL if expression
5614 : * contains no Vars (NOTE this could point to a RelOptInfo of a
5615 : * subquery, not one in the current query).
5616 : * statsTuple: the pg_statistic entry for the variable, if one exists;
5617 : * otherwise NULL.
5618 : * freefunc: pointer to a function to release statsTuple with.
5619 : * vartype: exposed type of the expression; this should always match
5620 : * the declared input type of the operator we are estimating for.
5621 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5622 : * commonly the same as the exposed type of the variable argument,
5623 : * but can be different in binary-compatible-type cases.
5624 : * isunique: true if we were able to match the var to a unique index, a
5625 : * single-column DISTINCT or GROUP-BY clause, implying its values are
5626 : * unique for this query. (Caution: this should be trusted for
5627 : * statistical purposes only, since we do not check indimmediate nor
5628 : * verify that the exact same definition of equality applies.)
5629 : * acl_ok: true if current user has permission to read all table rows from
5630 : * the column(s) underlying the pg_statistic entry. This is consulted by
5631 : * statistic_proc_security_check().
5632 : *
5633 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5634 : */
5635 : void
5636 3298186 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5637 : VariableStatData *vardata)
5638 : {
5639 : Node *basenode;
5640 : Relids varnos;
5641 : Relids basevarnos;
5642 : RelOptInfo *onerel;
5643 :
5644 : /* Make sure we don't return dangling pointers in vardata */
5645 23087302 : MemSet(vardata, 0, sizeof(VariableStatData));
5646 :
5647 : /* Save the exposed type of the expression */
5648 3298186 : vardata->vartype = exprType(node);
5649 :
5650 : /*
5651 : * PlaceHolderVars are transparent for the purpose of statistics lookup;
5652 : * they do not alter the value distribution of the underlying expression.
5653 : * However, they can obscure the structure, preventing us from recognizing
5654 : * matches to base columns, index expressions, or extended statistics. So
5655 : * strip them out first.
5656 : */
5657 3298186 : basenode = strip_all_phvs_deep(root, node);
5658 :
5659 : /*
5660 : * Look inside any binary-compatible relabeling. We need to handle nested
5661 : * RelabelType nodes here, because the prior stripping of PlaceHolderVars
5662 : * may have brought separate RelabelTypes into adjacency.
5663 : */
5664 3346764 : while (IsA(basenode, RelabelType))
5665 48578 : basenode = (Node *) ((RelabelType *) basenode)->arg;
5666 :
5667 : /* Fast path for a simple Var */
5668 3298186 : if (IsA(basenode, Var) &&
5669 795670 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5670 : {
5671 2362614 : Var *var = (Var *) basenode;
5672 :
5673 : /* Set up result fields other than the stats tuple */
5674 2362614 : vardata->var = basenode; /* return Var without phvs or relabeling */
5675 2362614 : vardata->rel = find_base_rel(root, var->varno);
5676 2362614 : vardata->atttype = var->vartype;
5677 2362614 : vardata->atttypmod = var->vartypmod;
5678 2362614 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5679 :
5680 : /* Try to locate some stats */
5681 2362614 : examine_simple_variable(root, var, vardata);
5682 :
5683 2362614 : return;
5684 : }
5685 :
5686 : /*
5687 : * Okay, it's a more complicated expression. Determine variable
5688 : * membership. Note that when varRelid isn't zero, only vars of that
5689 : * relation are considered "real" vars.
5690 : */
5691 935572 : varnos = pull_varnos(root, basenode);
5692 935572 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5693 :
5694 935572 : onerel = NULL;
5695 :
5696 935572 : if (bms_is_empty(basevarnos))
5697 : {
5698 : /* No Vars at all ... must be pseudo-constant clause */
5699 : }
5700 : else
5701 : {
5702 : int relid;
5703 :
5704 : /* Check if the expression is in vars of a single base relation */
5705 473286 : if (bms_get_singleton_member(basevarnos, &relid))
5706 : {
5707 468458 : if (varRelid == 0 || varRelid == relid)
5708 : {
5709 69062 : onerel = find_base_rel(root, relid);
5710 69062 : vardata->rel = onerel;
5711 69062 : node = basenode; /* strip any phvs or relabeling */
5712 : }
5713 : /* else treat it as a constant */
5714 : }
5715 : else
5716 : {
5717 : /* varnos has multiple relids */
5718 4828 : if (varRelid == 0)
5719 : {
5720 : /* treat it as a variable of a join relation */
5721 3516 : vardata->rel = find_join_rel(root, varnos);
5722 3516 : node = basenode; /* strip any phvs or relabeling */
5723 : }
5724 1312 : else if (bms_is_member(varRelid, varnos))
5725 : {
5726 : /* ignore the vars belonging to other relations */
5727 1198 : vardata->rel = find_base_rel(root, varRelid);
5728 1198 : node = basenode; /* strip any phvs or relabeling */
5729 : /* note: no point in expressional-index search here */
5730 : }
5731 : /* else treat it as a constant */
5732 : }
5733 : }
5734 :
5735 935572 : bms_free(basevarnos);
5736 :
5737 935572 : vardata->var = node;
5738 935572 : vardata->atttype = exprType(node);
5739 935572 : vardata->atttypmod = exprTypmod(node);
5740 :
5741 935572 : if (onerel)
5742 : {
5743 : /*
5744 : * We have an expression in vars of a single relation. Try to match
5745 : * it to expressional index columns, in hopes of finding some
5746 : * statistics.
5747 : *
5748 : * Note that we consider all index columns including INCLUDE columns,
5749 : * since there could be stats for such columns. But the test for
5750 : * uniqueness needs to be warier.
5751 : *
5752 : * XXX it's conceivable that there are multiple matches with different
5753 : * index opfamilies; if so, we need to pick one that matches the
5754 : * operator we are estimating for. FIXME later.
5755 : */
5756 : ListCell *ilist;
5757 : ListCell *slist;
5758 :
5759 : /*
5760 : * The nullingrels bits within the expression could prevent us from
5761 : * matching it to expressional index columns or to the expressions in
5762 : * extended statistics. So strip them out first.
5763 : */
5764 69062 : if (bms_overlap(varnos, root->outer_join_rels))
5765 2522 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5766 :
5767 150262 : foreach(ilist, onerel->indexlist)
5768 : {
5769 84182 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5770 : ListCell *indexpr_item;
5771 : int pos;
5772 :
5773 84182 : indexpr_item = list_head(index->indexprs);
5774 84182 : if (indexpr_item == NULL)
5775 79310 : continue; /* no expressions here... */
5776 :
5777 6834 : for (pos = 0; pos < index->ncolumns; pos++)
5778 : {
5779 4944 : if (index->indexkeys[pos] == 0)
5780 : {
5781 : Node *indexkey;
5782 :
5783 4872 : if (indexpr_item == NULL)
5784 0 : elog(ERROR, "too few entries in indexprs list");
5785 4872 : indexkey = (Node *) lfirst(indexpr_item);
5786 4872 : if (indexkey && IsA(indexkey, RelabelType))
5787 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5788 4872 : if (equal(node, indexkey))
5789 : {
5790 : /*
5791 : * Found a match ... is it a unique index? Tests here
5792 : * should match has_unique_index().
5793 : */
5794 3618 : if (index->unique &&
5795 438 : index->nkeycolumns == 1 &&
5796 438 : pos == 0 &&
5797 438 : (index->indpred == NIL || index->predOK))
5798 438 : vardata->isunique = true;
5799 :
5800 : /*
5801 : * Has it got stats? We only consider stats for
5802 : * non-partial indexes, since partial indexes probably
5803 : * don't reflect whole-relation statistics; the above
5804 : * check for uniqueness is the only info we take from
5805 : * a partial index.
5806 : *
5807 : * An index stats hook, however, must make its own
5808 : * decisions about what to do with partial indexes.
5809 : */
5810 3618 : if (get_index_stats_hook &&
5811 0 : (*get_index_stats_hook) (root, index->indexoid,
5812 0 : pos + 1, vardata))
5813 : {
5814 : /*
5815 : * The hook took control of acquiring a stats
5816 : * tuple. If it did supply a tuple, it'd better
5817 : * have supplied a freefunc.
5818 : */
5819 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5820 0 : !vardata->freefunc)
5821 0 : elog(ERROR, "no function provided to release variable stats with");
5822 : }
5823 3618 : else if (index->indpred == NIL)
5824 : {
5825 3618 : vardata->statsTuple =
5826 7236 : SearchSysCache3(STATRELATTINH,
5827 : ObjectIdGetDatum(index->indexoid),
5828 3618 : Int16GetDatum(pos + 1),
5829 : BoolGetDatum(false));
5830 3618 : vardata->freefunc = ReleaseSysCache;
5831 :
5832 3618 : if (HeapTupleIsValid(vardata->statsTuple))
5833 : {
5834 : /*
5835 : * Test if user has permission to access all
5836 : * rows from the index's table.
5837 : *
5838 : * For simplicity, we insist on the whole
5839 : * table being selectable, rather than trying
5840 : * to identify which column(s) the index
5841 : * depends on.
5842 : *
5843 : * Note that for an inheritance child,
5844 : * permissions are checked on the inheritance
5845 : * root parent, and whole-table select
5846 : * privilege on the parent doesn't quite
5847 : * guarantee that the user could read all
5848 : * columns of the child. But in practice it's
5849 : * unlikely that any interesting security
5850 : * violation could result from allowing access
5851 : * to the expression index's stats, so we
5852 : * allow it anyway. See similar code in
5853 : * examine_simple_variable() for additional
5854 : * comments.
5855 : */
5856 2982 : vardata->acl_ok =
5857 2982 : all_rows_selectable(root,
5858 2982 : index->rel->relid,
5859 : NULL);
5860 : }
5861 : else
5862 : {
5863 : /* suppress leakproofness checks later */
5864 636 : vardata->acl_ok = true;
5865 : }
5866 : }
5867 3618 : if (vardata->statsTuple)
5868 2982 : break;
5869 : }
5870 1890 : indexpr_item = lnext(index->indexprs, indexpr_item);
5871 : }
5872 : }
5873 4872 : if (vardata->statsTuple)
5874 2982 : break;
5875 : }
5876 :
5877 : /*
5878 : * Search extended statistics for one with a matching expression.
5879 : * There might be multiple ones, so just grab the first one. In the
5880 : * future, we might consider the statistics target (and pick the most
5881 : * accurate statistics) and maybe some other parameters.
5882 : */
5883 73178 : foreach(slist, onerel->statlist)
5884 : {
5885 4404 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5886 4404 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5887 : ListCell *expr_item;
5888 : int pos;
5889 :
5890 : /*
5891 : * Stop once we've found statistics for the expression (either
5892 : * from extended stats, or for an index in the preceding loop).
5893 : */
5894 4404 : if (vardata->statsTuple)
5895 288 : break;
5896 :
5897 : /* skip stats without per-expression stats */
5898 4116 : if (info->kind != STATS_EXT_EXPRESSIONS)
5899 2106 : continue;
5900 :
5901 : /* skip stats with mismatching stxdinherit value */
5902 2010 : if (info->inherit != rte->inh)
5903 6 : continue;
5904 :
5905 2004 : pos = 0;
5906 3306 : foreach(expr_item, info->exprs)
5907 : {
5908 2952 : Node *expr = (Node *) lfirst(expr_item);
5909 :
5910 : Assert(expr);
5911 :
5912 : /* strip RelabelType before comparing it */
5913 2952 : if (expr && IsA(expr, RelabelType))
5914 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5915 :
5916 : /* found a match, see if we can extract pg_statistic row */
5917 2952 : if (equal(node, expr))
5918 : {
5919 : /*
5920 : * XXX Not sure if we should cache the tuple somewhere.
5921 : * Now we just create a new copy every time.
5922 : */
5923 1650 : vardata->statsTuple =
5924 1650 : statext_expressions_load(info->statOid, rte->inh, pos);
5925 :
5926 1650 : vardata->freefunc = ReleaseDummy;
5927 :
5928 : /*
5929 : * Test if user has permission to access all rows from the
5930 : * table.
5931 : *
5932 : * For simplicity, we insist on the whole table being
5933 : * selectable, rather than trying to identify which
5934 : * column(s) the statistics object depends on.
5935 : *
5936 : * Note that for an inheritance child, permissions are
5937 : * checked on the inheritance root parent, and whole-table
5938 : * select privilege on the parent doesn't quite guarantee
5939 : * that the user could read all columns of the child. But
5940 : * in practice it's unlikely that any interesting security
5941 : * violation could result from allowing access to the
5942 : * expression stats, so we allow it anyway. See similar
5943 : * code in examine_simple_variable() for additional
5944 : * comments.
5945 : */
5946 1650 : vardata->acl_ok = all_rows_selectable(root,
5947 : onerel->relid,
5948 : NULL);
5949 :
5950 1650 : break;
5951 : }
5952 :
5953 1302 : pos++;
5954 : }
5955 : }
5956 : }
5957 :
5958 935572 : bms_free(varnos);
5959 : }
5960 :
5961 : /*
5962 : * strip_all_phvs_deep
5963 : * Deeply strip all PlaceHolderVars in an expression.
5964 :
5965 : * As a performance optimization, we first use a lightweight walker to check
5966 : * for the presence of any PlaceHolderVars. The expensive mutator is invoked
5967 : * only if a PlaceHolderVar is found, avoiding unnecessary memory allocation
5968 : * and tree copying in the common case where no PlaceHolderVars are present.
5969 : */
5970 : static Node *
5971 3298186 : strip_all_phvs_deep(PlannerInfo *root, Node *node)
5972 : {
5973 : /* If there are no PHVs anywhere, we needn't work hard */
5974 3298186 : if (root->glob->lastPHId == 0)
5975 3263592 : return node;
5976 :
5977 34594 : if (!contain_placeholder_walker(node, NULL))
5978 30260 : return node;
5979 4334 : return strip_all_phvs_mutator(node, NULL);
5980 : }
5981 :
5982 : /*
5983 : * contain_placeholder_walker
5984 : * Lightweight walker to check if an expression contains any
5985 : * PlaceHolderVars
5986 : */
5987 : static bool
5988 39192 : contain_placeholder_walker(Node *node, void *context)
5989 : {
5990 39192 : if (node == NULL)
5991 210 : return false;
5992 38982 : if (IsA(node, PlaceHolderVar))
5993 4334 : return true;
5994 :
5995 34648 : return expression_tree_walker(node, contain_placeholder_walker, context);
5996 : }
5997 :
5998 : /*
5999 : * strip_all_phvs_mutator
6000 : * Mutator to deeply strip all PlaceHolderVars
6001 : */
6002 : static Node *
6003 11686 : strip_all_phvs_mutator(Node *node, void *context)
6004 : {
6005 11686 : if (node == NULL)
6006 48 : return NULL;
6007 11638 : if (IsA(node, PlaceHolderVar))
6008 : {
6009 : /* Strip it and recurse into its contained expression */
6010 4484 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6011 :
6012 4484 : return strip_all_phvs_mutator((Node *) phv->phexpr, context);
6013 : }
6014 :
6015 7154 : return expression_tree_mutator(node, strip_all_phvs_mutator, context);
6016 : }
6017 :
6018 : /*
6019 : * examine_simple_variable
6020 : * Handle a simple Var for examine_variable
6021 : *
6022 : * This is split out as a subroutine so that we can recurse to deal with
6023 : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
6024 : *
6025 : * We already filled in all the fields of *vardata except for the stats tuple.
6026 : */
6027 : static void
6028 2368988 : examine_simple_variable(PlannerInfo *root, Var *var,
6029 : VariableStatData *vardata)
6030 : {
6031 2368988 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
6032 :
6033 : Assert(IsA(rte, RangeTblEntry));
6034 :
6035 2368988 : if (get_relation_stats_hook &&
6036 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
6037 : {
6038 : /*
6039 : * The hook took control of acquiring a stats tuple. If it did supply
6040 : * a tuple, it'd better have supplied a freefunc.
6041 : */
6042 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6043 0 : !vardata->freefunc)
6044 0 : elog(ERROR, "no function provided to release variable stats with");
6045 : }
6046 2368988 : else if (rte->rtekind == RTE_RELATION)
6047 : {
6048 : /*
6049 : * Plain table or parent of an inheritance appendrel, so look up the
6050 : * column in pg_statistic
6051 : */
6052 2245734 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6053 : ObjectIdGetDatum(rte->relid),
6054 2245734 : Int16GetDatum(var->varattno),
6055 2245734 : BoolGetDatum(rte->inh));
6056 2245734 : vardata->freefunc = ReleaseSysCache;
6057 :
6058 2245734 : if (HeapTupleIsValid(vardata->statsTuple))
6059 : {
6060 : /*
6061 : * Test if user has permission to read all rows from this column.
6062 : *
6063 : * This requires that the user has the appropriate SELECT
6064 : * privileges and that there are no securityQuals from security
6065 : * barrier views or RLS policies. If that's not the case, then we
6066 : * only permit leakproof functions to be passed pg_statistic data
6067 : * in vardata, otherwise the functions might reveal data that the
6068 : * user doesn't have permission to see --- see
6069 : * statistic_proc_security_check().
6070 : */
6071 1664084 : vardata->acl_ok =
6072 1664084 : all_rows_selectable(root, var->varno,
6073 1664084 : bms_make_singleton(var->varattno - FirstLowInvalidHeapAttributeNumber));
6074 : }
6075 : else
6076 : {
6077 : /* suppress any possible leakproofness checks later */
6078 581650 : vardata->acl_ok = true;
6079 : }
6080 : }
6081 123254 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
6082 112898 : (rte->rtekind == RTE_CTE && !rte->self_reference))
6083 : {
6084 : /*
6085 : * Plain subquery (not one that was converted to an appendrel) or
6086 : * non-recursive CTE. In either case, we can try to find out what the
6087 : * Var refers to within the subquery. We skip this for appendrel and
6088 : * recursive-CTE cases because any column stats we did find would
6089 : * likely not be very relevant.
6090 : */
6091 : PlannerInfo *subroot;
6092 : Query *subquery;
6093 : List *subtlist;
6094 : TargetEntry *ste;
6095 :
6096 : /*
6097 : * Punt if it's a whole-row var rather than a plain column reference.
6098 : */
6099 17604 : if (var->varattno == InvalidAttrNumber)
6100 0 : return;
6101 :
6102 : /*
6103 : * Otherwise, find the subquery's planner subroot.
6104 : */
6105 17604 : if (rte->rtekind == RTE_SUBQUERY)
6106 : {
6107 : RelOptInfo *rel;
6108 :
6109 : /*
6110 : * Fetch RelOptInfo for subquery. Note that we don't change the
6111 : * rel returned in vardata, since caller expects it to be a rel of
6112 : * the caller's query level. Because we might already be
6113 : * recursing, we can't use that rel pointer either, but have to
6114 : * look up the Var's rel afresh.
6115 : */
6116 10356 : rel = find_base_rel(root, var->varno);
6117 :
6118 10356 : subroot = rel->subroot;
6119 : }
6120 : else
6121 : {
6122 : /* CTE case is more difficult */
6123 : PlannerInfo *cteroot;
6124 : Index levelsup;
6125 : int ndx;
6126 : int plan_id;
6127 : ListCell *lc;
6128 :
6129 : /*
6130 : * Find the referenced CTE, and locate the subroot previously made
6131 : * for it.
6132 : */
6133 7248 : levelsup = rte->ctelevelsup;
6134 7248 : cteroot = root;
6135 13634 : while (levelsup-- > 0)
6136 : {
6137 6386 : cteroot = cteroot->parent_root;
6138 6386 : if (!cteroot) /* shouldn't happen */
6139 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
6140 : }
6141 :
6142 : /*
6143 : * Note: cte_plan_ids can be shorter than cteList, if we are still
6144 : * working on planning the CTEs (ie, this is a side-reference from
6145 : * another CTE). So we mustn't use forboth here.
6146 : */
6147 7248 : ndx = 0;
6148 9470 : foreach(lc, cteroot->parse->cteList)
6149 : {
6150 9470 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
6151 :
6152 9470 : if (strcmp(cte->ctename, rte->ctename) == 0)
6153 7248 : break;
6154 2222 : ndx++;
6155 : }
6156 7248 : if (lc == NULL) /* shouldn't happen */
6157 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
6158 7248 : if (ndx >= list_length(cteroot->cte_plan_ids))
6159 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
6160 7248 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
6161 7248 : if (plan_id <= 0)
6162 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
6163 7248 : subroot = list_nth(root->glob->subroots, plan_id - 1);
6164 : }
6165 :
6166 : /* If the subquery hasn't been planned yet, we have to punt */
6167 17604 : if (subroot == NULL)
6168 0 : return;
6169 : Assert(IsA(subroot, PlannerInfo));
6170 :
6171 : /*
6172 : * We must use the subquery parsetree as mangled by the planner, not
6173 : * the raw version from the RTE, because we need a Var that will refer
6174 : * to the subroot's live RelOptInfos. For instance, if any subquery
6175 : * pullup happened during planning, Vars in the targetlist might have
6176 : * gotten replaced, and we need to see the replacement expressions.
6177 : */
6178 17604 : subquery = subroot->parse;
6179 : Assert(IsA(subquery, Query));
6180 :
6181 : /*
6182 : * Punt if subquery uses set operations or grouping sets, as these
6183 : * will mash underlying columns' stats beyond recognition. (Set ops
6184 : * are particularly nasty; if we forged ahead, we would return stats
6185 : * relevant to only the leftmost subselect...) DISTINCT is also
6186 : * problematic, but we check that later because there is a possibility
6187 : * of learning something even with it.
6188 : */
6189 17604 : if (subquery->setOperations ||
6190 15318 : subquery->groupingSets)
6191 2382 : return;
6192 :
6193 : /* Get the subquery output expression referenced by the upper Var */
6194 15222 : if (subquery->returningList)
6195 206 : subtlist = subquery->returningList;
6196 : else
6197 15016 : subtlist = subquery->targetList;
6198 15222 : ste = get_tle_by_resno(subtlist, var->varattno);
6199 15222 : if (ste == NULL || ste->resjunk)
6200 0 : elog(ERROR, "subquery %s does not have attribute %d",
6201 : rte->eref->aliasname, var->varattno);
6202 15222 : var = (Var *) ste->expr;
6203 :
6204 : /*
6205 : * If subquery uses DISTINCT, we can't make use of any stats for the
6206 : * variable ... but, if it's the only DISTINCT column, we are entitled
6207 : * to consider it unique. We do the test this way so that it works
6208 : * for cases involving DISTINCT ON.
6209 : */
6210 15222 : if (subquery->distinctClause)
6211 : {
6212 1838 : if (list_length(subquery->distinctClause) == 1 &&
6213 616 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
6214 308 : vardata->isunique = true;
6215 : /* cannot go further */
6216 1222 : return;
6217 : }
6218 :
6219 : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
6220 14000 : if (subquery->groupClause)
6221 : {
6222 1080 : if (list_length(subquery->groupClause) == 1 &&
6223 450 : targetIsInSortList(ste, InvalidOid, subquery->groupClause))
6224 338 : vardata->isunique = true;
6225 : /* cannot go further */
6226 630 : return;
6227 : }
6228 :
6229 : /*
6230 : * If the sub-query originated from a view with the security_barrier
6231 : * attribute, we must not look at the variable's statistics, though it
6232 : * seems all right to notice the existence of a DISTINCT clause. So
6233 : * stop here.
6234 : *
6235 : * This is probably a harsher restriction than necessary; it's
6236 : * certainly OK for the selectivity estimator (which is a C function,
6237 : * and therefore omnipotent anyway) to look at the statistics. But
6238 : * many selectivity estimators will happily *invoke the operator
6239 : * function* to try to work out a good estimate - and that's not OK.
6240 : * So for now, don't dig down for stats.
6241 : */
6242 13370 : if (rte->security_barrier)
6243 1374 : return;
6244 :
6245 : /* Can only handle a simple Var of subquery's query level */
6246 11996 : if (var && IsA(var, Var) &&
6247 6374 : var->varlevelsup == 0)
6248 : {
6249 : /*
6250 : * OK, recurse into the subquery. Note that the original setting
6251 : * of vardata->isunique (which will surely be false) is left
6252 : * unchanged in this situation. That's what we want, since even
6253 : * if the underlying column is unique, the subquery may have
6254 : * joined to other tables in a way that creates duplicates.
6255 : */
6256 6374 : examine_simple_variable(subroot, var, vardata);
6257 : }
6258 : }
6259 : else
6260 : {
6261 : /*
6262 : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
6263 : * see RTE_JOIN here because join alias Vars have already been
6264 : * flattened.) There's not much we can do with function outputs, but
6265 : * maybe someday try to be smarter about VALUES.
6266 : */
6267 : }
6268 : }
6269 :
6270 : /*
6271 : * all_rows_selectable
6272 : * Test whether the user has permission to select all rows from a given
6273 : * relation.
6274 : *
6275 : * Inputs:
6276 : * root: the planner info
6277 : * varno: the index of the relation (assumed to be an RTE_RELATION)
6278 : * varattnos: the attributes for which permission is required, or NULL if
6279 : * whole-table access is required
6280 : *
6281 : * Returns true if the user has the required select permissions, and there are
6282 : * no securityQuals from security barrier views or RLS policies.
6283 : *
6284 : * Note that if the relation is an inheritance child relation, securityQuals
6285 : * and access permissions are checked against the inheritance root parent (the
6286 : * relation actually mentioned in the query) --- see the comments in
6287 : * expand_single_inheritance_child() for an explanation of why it has to be
6288 : * done this way.
6289 : *
6290 : * If varattnos is non-NULL, its attribute numbers should be offset by
6291 : * FirstLowInvalidHeapAttributeNumber so that system attributes can be
6292 : * checked. If varattnos is NULL, only table-level SELECT privileges are
6293 : * checked, not any column-level privileges.
6294 : *
6295 : * Note: if the relation is accessed via a view, this function actually tests
6296 : * whether the view owner has permission to select from the relation. To
6297 : * ensure that the current user has permission, it is also necessary to check
6298 : * that the current user has permission to select from the view, which we do
6299 : * at planner-startup --- see subquery_planner().
6300 : *
6301 : * This is exported so that other estimation functions can use it.
6302 : */
6303 : bool
6304 1668968 : all_rows_selectable(PlannerInfo *root, Index varno, Bitmapset *varattnos)
6305 : {
6306 1668968 : RelOptInfo *rel = find_base_rel_noerr(root, varno);
6307 1668968 : RangeTblEntry *rte = planner_rt_fetch(varno, root);
6308 : Oid userid;
6309 : int varattno;
6310 :
6311 : Assert(rte->rtekind == RTE_RELATION);
6312 :
6313 : /*
6314 : * Determine the user ID to use for privilege checks (either the current
6315 : * user or the view owner, if we're accessing the table via a view).
6316 : *
6317 : * Normally the relation will have an associated RelOptInfo from which we
6318 : * can find the userid, but it might not if it's a RETURNING Var for an
6319 : * INSERT target relation. In that case use the RTEPermissionInfo
6320 : * associated with the RTE.
6321 : *
6322 : * If we navigate up to a parent relation, we keep using the same userid,
6323 : * since it's the same in all relations of a given inheritance tree.
6324 : */
6325 1668968 : if (rel)
6326 1668926 : userid = rel->userid;
6327 : else
6328 : {
6329 : RTEPermissionInfo *perminfo;
6330 :
6331 42 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
6332 42 : userid = perminfo->checkAsUser;
6333 : }
6334 1668968 : if (!OidIsValid(userid))
6335 1491400 : userid = GetUserId();
6336 :
6337 : /*
6338 : * Permissions and securityQuals must be checked on the table actually
6339 : * mentioned in the query, so if this is an inheritance child, navigate up
6340 : * to the inheritance root parent. If the user can read the whole table
6341 : * or the required columns there, then they can read from the child table
6342 : * too. For per-column checks, we must find out which of the root
6343 : * parent's attributes the child relation's attributes correspond to.
6344 : */
6345 1668968 : if (root->append_rel_array != NULL)
6346 : {
6347 : AppendRelInfo *appinfo;
6348 :
6349 234406 : appinfo = root->append_rel_array[varno];
6350 :
6351 : /*
6352 : * Partitions are mapped to their immediate parent, not the root
6353 : * parent, so must be ready to walk up multiple AppendRelInfos. But
6354 : * stop if we hit a parent that is not RTE_RELATION --- that's a
6355 : * flattened UNION ALL subquery, not an inheritance parent.
6356 : */
6357 436174 : while (appinfo &&
6358 202140 : planner_rt_fetch(appinfo->parent_relid,
6359 202140 : root)->rtekind == RTE_RELATION)
6360 : {
6361 201768 : Bitmapset *parent_varattnos = NULL;
6362 :
6363 : /*
6364 : * For each child attribute, find the corresponding parent
6365 : * attribute. In rare cases, the attribute may be local to the
6366 : * child table, in which case, we've got to live with having no
6367 : * access to this column.
6368 : */
6369 201768 : varattno = -1;
6370 400686 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6371 : {
6372 : AttrNumber attno;
6373 : AttrNumber parent_attno;
6374 :
6375 198918 : attno = varattno + FirstLowInvalidHeapAttributeNumber;
6376 :
6377 198918 : if (attno == InvalidAttrNumber)
6378 : {
6379 : /*
6380 : * Whole-row reference, so must map each column of the
6381 : * child to the parent table.
6382 : */
6383 36 : for (attno = 1; attno <= appinfo->num_child_cols; attno++)
6384 : {
6385 24 : parent_attno = appinfo->parent_colnos[attno - 1];
6386 24 : if (parent_attno == 0)
6387 0 : return false; /* attr is local to child */
6388 : parent_varattnos =
6389 24 : bms_add_member(parent_varattnos,
6390 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6391 : }
6392 : }
6393 : else
6394 : {
6395 198906 : if (attno < 0)
6396 : {
6397 : /* System attnos are the same in all tables */
6398 0 : parent_attno = attno;
6399 : }
6400 : else
6401 : {
6402 198906 : if (attno > appinfo->num_child_cols)
6403 0 : return false; /* safety check */
6404 198906 : parent_attno = appinfo->parent_colnos[attno - 1];
6405 198906 : if (parent_attno == 0)
6406 0 : return false; /* attr is local to child */
6407 : }
6408 : parent_varattnos =
6409 198906 : bms_add_member(parent_varattnos,
6410 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6411 : }
6412 : }
6413 :
6414 : /* If the parent is itself a child, continue up */
6415 201768 : varno = appinfo->parent_relid;
6416 201768 : varattnos = parent_varattnos;
6417 201768 : appinfo = root->append_rel_array[varno];
6418 : }
6419 :
6420 : /* Perform the access check on this parent rel */
6421 234406 : rte = planner_rt_fetch(varno, root);
6422 : Assert(rte->rtekind == RTE_RELATION);
6423 : }
6424 :
6425 : /*
6426 : * For all rows to be accessible, there must be no securityQuals from
6427 : * security barrier views or RLS policies.
6428 : */
6429 1668968 : if (rte->securityQuals != NIL)
6430 828 : return false;
6431 :
6432 : /*
6433 : * Test for table-level SELECT privilege.
6434 : *
6435 : * If varattnos is non-NULL, this is sufficient to give access to all
6436 : * requested attributes, even for a child table, since we have verified
6437 : * that all required child columns have matching parent columns.
6438 : *
6439 : * If varattnos is NULL (whole-table access requested), this doesn't
6440 : * necessarily guarantee that the user can read all columns of a child
6441 : * table, but we allow it anyway (see comments in examine_variable()) and
6442 : * don't bother checking any column privileges.
6443 : */
6444 1668140 : if (pg_class_aclcheck(rte->relid, userid, ACL_SELECT) == ACLCHECK_OK)
6445 1667688 : return true;
6446 :
6447 452 : if (varattnos == NULL)
6448 12 : return false; /* whole-table access requested */
6449 :
6450 : /*
6451 : * Don't have table-level SELECT privilege, so check per-column
6452 : * privileges.
6453 : */
6454 440 : varattno = -1;
6455 646 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6456 : {
6457 440 : AttrNumber attno = varattno + FirstLowInvalidHeapAttributeNumber;
6458 :
6459 440 : if (attno == InvalidAttrNumber)
6460 : {
6461 : /* Whole-row reference, so must have access to all columns */
6462 6 : if (pg_attribute_aclcheck_all(rte->relid, userid, ACL_SELECT,
6463 : ACLMASK_ALL) != ACLCHECK_OK)
6464 6 : return false;
6465 : }
6466 : else
6467 : {
6468 434 : if (pg_attribute_aclcheck(rte->relid, attno, userid,
6469 : ACL_SELECT) != ACLCHECK_OK)
6470 228 : return false;
6471 : }
6472 : }
6473 :
6474 : /* If we reach here, have all required column privileges */
6475 206 : return true;
6476 : }
6477 :
6478 : /*
6479 : * examine_indexcol_variable
6480 : * Try to look up statistical data about an index column/expression.
6481 : * Fill in a VariableStatData struct to describe the column.
6482 : *
6483 : * Inputs:
6484 : * root: the planner info
6485 : * index: the index whose column we're interested in
6486 : * indexcol: 0-based index column number (subscripts index->indexkeys[])
6487 : *
6488 : * Outputs: *vardata is filled as follows:
6489 : * var: the input expression (with any binary relabeling stripped, if
6490 : * it is or contains a variable; but otherwise the type is preserved)
6491 : * rel: RelOptInfo for table relation containing variable.
6492 : * statsTuple: the pg_statistic entry for the variable, if one exists;
6493 : * otherwise NULL.
6494 : * freefunc: pointer to a function to release statsTuple with.
6495 : *
6496 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
6497 : */
6498 : static void
6499 806606 : examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
6500 : int indexcol, VariableStatData *vardata)
6501 : {
6502 : AttrNumber colnum;
6503 : Oid relid;
6504 :
6505 806606 : if (index->indexkeys[indexcol] != 0)
6506 : {
6507 : /* Simple variable --- look to stats for the underlying table */
6508 804392 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6509 :
6510 : Assert(rte->rtekind == RTE_RELATION);
6511 804392 : relid = rte->relid;
6512 : Assert(relid != InvalidOid);
6513 804392 : colnum = index->indexkeys[indexcol];
6514 804392 : vardata->rel = index->rel;
6515 :
6516 804392 : if (get_relation_stats_hook &&
6517 0 : (*get_relation_stats_hook) (root, rte, colnum, vardata))
6518 : {
6519 : /*
6520 : * The hook took control of acquiring a stats tuple. If it did
6521 : * supply a tuple, it'd better have supplied a freefunc.
6522 : */
6523 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6524 0 : !vardata->freefunc)
6525 0 : elog(ERROR, "no function provided to release variable stats with");
6526 : }
6527 : else
6528 : {
6529 804392 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6530 : ObjectIdGetDatum(relid),
6531 : Int16GetDatum(colnum),
6532 804392 : BoolGetDatum(rte->inh));
6533 804392 : vardata->freefunc = ReleaseSysCache;
6534 : }
6535 : }
6536 : else
6537 : {
6538 : /* Expression --- maybe there are stats for the index itself */
6539 2214 : relid = index->indexoid;
6540 2214 : colnum = indexcol + 1;
6541 :
6542 2214 : if (get_index_stats_hook &&
6543 0 : (*get_index_stats_hook) (root, relid, colnum, vardata))
6544 : {
6545 : /*
6546 : * The hook took control of acquiring a stats tuple. If it did
6547 : * supply a tuple, it'd better have supplied a freefunc.
6548 : */
6549 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6550 0 : !vardata->freefunc)
6551 0 : elog(ERROR, "no function provided to release variable stats with");
6552 : }
6553 : else
6554 : {
6555 2214 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6556 : ObjectIdGetDatum(relid),
6557 : Int16GetDatum(colnum),
6558 : BoolGetDatum(false));
6559 2214 : vardata->freefunc = ReleaseSysCache;
6560 : }
6561 : }
6562 806606 : }
6563 :
6564 : /*
6565 : * Check whether it is permitted to call func_oid passing some of the
6566 : * pg_statistic data in vardata. We allow this if either of the following
6567 : * conditions is met: (1) the user has SELECT privileges on the table or
6568 : * column underlying the pg_statistic data and there are no securityQuals from
6569 : * security barrier views or RLS policies, or (2) the function is marked
6570 : * leakproof.
6571 : */
6572 : bool
6573 1171630 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
6574 : {
6575 1171630 : if (vardata->acl_ok)
6576 1169780 : return true; /* have SELECT privs and no securityQuals */
6577 :
6578 1850 : if (!OidIsValid(func_oid))
6579 0 : return false;
6580 :
6581 1850 : if (get_func_leakproof(func_oid))
6582 916 : return true;
6583 :
6584 934 : ereport(DEBUG2,
6585 : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
6586 : get_func_name(func_oid))));
6587 934 : return false;
6588 : }
6589 :
6590 : /*
6591 : * get_variable_numdistinct
6592 : * Estimate the number of distinct values of a variable.
6593 : *
6594 : * vardata: results of examine_variable
6595 : * *isdefault: set to true if the result is a default rather than based on
6596 : * anything meaningful.
6597 : *
6598 : * NB: be careful to produce a positive integral result, since callers may
6599 : * compare the result to exact integer counts, or might divide by it.
6600 : */
6601 : double
6602 1665030 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
6603 : {
6604 : double stadistinct;
6605 1665030 : double stanullfrac = 0.0;
6606 : double ntuples;
6607 :
6608 1665030 : *isdefault = false;
6609 :
6610 : /*
6611 : * Determine the stadistinct value to use. There are cases where we can
6612 : * get an estimate even without a pg_statistic entry, or can get a better
6613 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
6614 : * (otherwise, assume no nulls, for lack of any better idea).
6615 : */
6616 1665030 : if (HeapTupleIsValid(vardata->statsTuple))
6617 : {
6618 : /* Use the pg_statistic entry */
6619 : Form_pg_statistic stats;
6620 :
6621 1167104 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6622 1167104 : stadistinct = stats->stadistinct;
6623 1167104 : stanullfrac = stats->stanullfrac;
6624 : }
6625 497926 : else if (vardata->vartype == BOOLOID)
6626 : {
6627 : /*
6628 : * Special-case boolean columns: presumably, two distinct values.
6629 : *
6630 : * Are there any other datatypes we should wire in special estimates
6631 : * for?
6632 : */
6633 602 : stadistinct = 2.0;
6634 : }
6635 497324 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6636 : {
6637 : /*
6638 : * If the Var represents a column of a VALUES RTE, assume it's unique.
6639 : * This could of course be very wrong, but it should tend to be true
6640 : * in well-written queries. We could consider examining the VALUES'
6641 : * contents to get some real statistics; but that only works if the
6642 : * entries are all constants, and it would be pretty expensive anyway.
6643 : */
6644 3588 : stadistinct = -1.0; /* unique (and all non null) */
6645 : }
6646 : else
6647 : {
6648 : /*
6649 : * We don't keep statistics for system columns, but in some cases we
6650 : * can infer distinctness anyway.
6651 : */
6652 493736 : if (vardata->var && IsA(vardata->var, Var))
6653 : {
6654 455982 : switch (((Var *) vardata->var)->varattno)
6655 : {
6656 1224 : case SelfItemPointerAttributeNumber:
6657 1224 : stadistinct = -1.0; /* unique (and all non null) */
6658 1224 : break;
6659 26398 : case TableOidAttributeNumber:
6660 26398 : stadistinct = 1.0; /* only 1 value */
6661 26398 : break;
6662 428360 : default:
6663 428360 : stadistinct = 0.0; /* means "unknown" */
6664 428360 : break;
6665 : }
6666 : }
6667 : else
6668 37754 : stadistinct = 0.0; /* means "unknown" */
6669 :
6670 : /*
6671 : * XXX consider using estimate_num_groups on expressions?
6672 : */
6673 : }
6674 :
6675 : /*
6676 : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6677 : * variable, assume it is unique no matter what pg_statistic says; the
6678 : * statistics could be out of date, or we might have found a partial
6679 : * unique index that proves the var is unique for this query. However,
6680 : * we'd better still believe the null-fraction statistic.
6681 : */
6682 1665030 : if (vardata->isunique)
6683 413248 : stadistinct = -1.0 * (1.0 - stanullfrac);
6684 :
6685 : /*
6686 : * If we had an absolute estimate, use that.
6687 : */
6688 1665030 : if (stadistinct > 0.0)
6689 430674 : return clamp_row_est(stadistinct);
6690 :
6691 : /*
6692 : * Otherwise we need to get the relation size; punt if not available.
6693 : */
6694 1234356 : if (vardata->rel == NULL)
6695 : {
6696 704 : *isdefault = true;
6697 704 : return DEFAULT_NUM_DISTINCT;
6698 : }
6699 1233652 : ntuples = vardata->rel->tuples;
6700 1233652 : if (ntuples <= 0.0)
6701 : {
6702 118366 : *isdefault = true;
6703 118366 : return DEFAULT_NUM_DISTINCT;
6704 : }
6705 :
6706 : /*
6707 : * If we had a relative estimate, use that.
6708 : */
6709 1115286 : if (stadistinct < 0.0)
6710 809346 : return clamp_row_est(-stadistinct * ntuples);
6711 :
6712 : /*
6713 : * With no data, estimate ndistinct = ntuples if the table is small, else
6714 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6715 : * that the behavior isn't discontinuous.
6716 : */
6717 305940 : if (ntuples < DEFAULT_NUM_DISTINCT)
6718 143472 : return clamp_row_est(ntuples);
6719 :
6720 162468 : *isdefault = true;
6721 162468 : return DEFAULT_NUM_DISTINCT;
6722 : }
6723 :
6724 : /*
6725 : * get_variable_range
6726 : * Estimate the minimum and maximum value of the specified variable.
6727 : * If successful, store values in *min and *max, and return true.
6728 : * If no data available, return false.
6729 : *
6730 : * sortop is the "<" comparison operator to use. This should generally
6731 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6732 : * The collation must be specified too.
6733 : */
6734 : static bool
6735 251714 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6736 : Oid sortop, Oid collation,
6737 : Datum *min, Datum *max)
6738 : {
6739 251714 : Datum tmin = 0;
6740 251714 : Datum tmax = 0;
6741 251714 : bool have_data = false;
6742 : int16 typLen;
6743 : bool typByVal;
6744 : Oid opfuncoid;
6745 : FmgrInfo opproc;
6746 : AttStatsSlot sslot;
6747 :
6748 : /*
6749 : * XXX It's very tempting to try to use the actual column min and max, if
6750 : * we can get them relatively-cheaply with an index probe. However, since
6751 : * this function is called many times during join planning, that could
6752 : * have unpleasant effects on planning speed. Need more investigation
6753 : * before enabling this.
6754 : */
6755 : #ifdef NOT_USED
6756 : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6757 : return true;
6758 : #endif
6759 :
6760 251714 : if (!HeapTupleIsValid(vardata->statsTuple))
6761 : {
6762 : /* no stats available, so default result */
6763 56590 : return false;
6764 : }
6765 :
6766 : /*
6767 : * If we can't apply the sortop to the stats data, just fail. In
6768 : * principle, if there's a histogram and no MCVs, we could return the
6769 : * histogram endpoints without ever applying the sortop ... but it's
6770 : * probably not worth trying, because whatever the caller wants to do with
6771 : * the endpoints would likely fail the security check too.
6772 : */
6773 195124 : if (!statistic_proc_security_check(vardata,
6774 195124 : (opfuncoid = get_opcode(sortop))))
6775 0 : return false;
6776 :
6777 195124 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6778 :
6779 195124 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6780 :
6781 : /*
6782 : * If there is a histogram with the ordering we want, grab the first and
6783 : * last values.
6784 : */
6785 195124 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6786 : STATISTIC_KIND_HISTOGRAM, sortop,
6787 : ATTSTATSSLOT_VALUES))
6788 : {
6789 123448 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6790 : {
6791 123448 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6792 123448 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6793 123448 : have_data = true;
6794 : }
6795 123448 : free_attstatsslot(&sslot);
6796 : }
6797 :
6798 : /*
6799 : * Otherwise, if there is a histogram with some other ordering, scan it
6800 : * and get the min and max values according to the ordering we want. This
6801 : * of course may not find values that are really extremal according to our
6802 : * ordering, but it beats ignoring available data.
6803 : */
6804 266800 : if (!have_data &&
6805 71676 : get_attstatsslot(&sslot, vardata->statsTuple,
6806 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6807 : ATTSTATSSLOT_VALUES))
6808 : {
6809 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6810 : collation, typLen, typByVal,
6811 : &tmin, &tmax, &have_data);
6812 0 : free_attstatsslot(&sslot);
6813 : }
6814 :
6815 : /*
6816 : * If we have most-common-values info, look for extreme MCVs. This is
6817 : * needed even if we also have a histogram, since the histogram excludes
6818 : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6819 : * be pretty wary of deciding that that is a full representation of the
6820 : * data. Proceed only if the MCVs represent the whole table (to within
6821 : * roundoff error).
6822 : */
6823 195124 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6824 : STATISTIC_KIND_MCV, InvalidOid,
6825 195124 : have_data ? ATTSTATSSLOT_VALUES :
6826 : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6827 : {
6828 109608 : bool use_mcvs = have_data;
6829 :
6830 109608 : if (!have_data)
6831 : {
6832 70242 : double sumcommon = 0.0;
6833 : double nullfrac;
6834 : int i;
6835 :
6836 529628 : for (i = 0; i < sslot.nnumbers; i++)
6837 459386 : sumcommon += sslot.numbers[i];
6838 70242 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6839 70242 : if (sumcommon + nullfrac > 0.99999)
6840 68034 : use_mcvs = true;
6841 : }
6842 :
6843 109608 : if (use_mcvs)
6844 107400 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6845 : collation, typLen, typByVal,
6846 : &tmin, &tmax, &have_data);
6847 109608 : free_attstatsslot(&sslot);
6848 : }
6849 :
6850 195124 : *min = tmin;
6851 195124 : *max = tmax;
6852 195124 : return have_data;
6853 : }
6854 :
6855 : /*
6856 : * get_stats_slot_range: scan sslot for min/max values
6857 : *
6858 : * Subroutine for get_variable_range: update min/max/have_data according
6859 : * to what we find in the statistics array.
6860 : */
6861 : static void
6862 107400 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6863 : Oid collation, int16 typLen, bool typByVal,
6864 : Datum *min, Datum *max, bool *p_have_data)
6865 : {
6866 107400 : Datum tmin = *min;
6867 107400 : Datum tmax = *max;
6868 107400 : bool have_data = *p_have_data;
6869 107400 : bool found_tmin = false;
6870 107400 : bool found_tmax = false;
6871 :
6872 : /* Look up the comparison function, if we didn't already do so */
6873 107400 : if (opproc->fn_oid != opfuncoid)
6874 107400 : fmgr_info(opfuncoid, opproc);
6875 :
6876 : /* Scan all the slot's values */
6877 2628482 : for (int i = 0; i < sslot->nvalues; i++)
6878 : {
6879 2521082 : if (!have_data)
6880 : {
6881 68034 : tmin = tmax = sslot->values[i];
6882 68034 : found_tmin = found_tmax = true;
6883 68034 : *p_have_data = have_data = true;
6884 68034 : continue;
6885 : }
6886 2453048 : if (DatumGetBool(FunctionCall2Coll(opproc,
6887 : collation,
6888 2453048 : sslot->values[i], tmin)))
6889 : {
6890 61000 : tmin = sslot->values[i];
6891 61000 : found_tmin = true;
6892 : }
6893 2453048 : if (DatumGetBool(FunctionCall2Coll(opproc,
6894 : collation,
6895 2453048 : tmax, sslot->values[i])))
6896 : {
6897 264752 : tmax = sslot->values[i];
6898 264752 : found_tmax = true;
6899 : }
6900 : }
6901 :
6902 : /*
6903 : * Copy the slot's values, if we found new extreme values.
6904 : */
6905 107400 : if (found_tmin)
6906 91898 : *min = datumCopy(tmin, typByVal, typLen);
6907 107400 : if (found_tmax)
6908 72942 : *max = datumCopy(tmax, typByVal, typLen);
6909 107400 : }
6910 :
6911 :
6912 : /*
6913 : * get_actual_variable_range
6914 : * Attempt to identify the current *actual* minimum and/or maximum
6915 : * of the specified variable, by looking for a suitable btree index
6916 : * and fetching its low and/or high values.
6917 : * If successful, store values in *min and *max, and return true.
6918 : * (Either pointer can be NULL if that endpoint isn't needed.)
6919 : * If unsuccessful, return false.
6920 : *
6921 : * sortop is the "<" comparison operator to use.
6922 : * collation is the required collation.
6923 : */
6924 : static bool
6925 189556 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6926 : Oid sortop, Oid collation,
6927 : Datum *min, Datum *max)
6928 : {
6929 189556 : bool have_data = false;
6930 189556 : RelOptInfo *rel = vardata->rel;
6931 : RangeTblEntry *rte;
6932 : ListCell *lc;
6933 :
6934 : /* No hope if no relation or it doesn't have indexes */
6935 189556 : if (rel == NULL || rel->indexlist == NIL)
6936 13840 : return false;
6937 : /* If it has indexes it must be a plain relation */
6938 175716 : rte = root->simple_rte_array[rel->relid];
6939 : Assert(rte->rtekind == RTE_RELATION);
6940 :
6941 : /* ignore partitioned tables. Any indexes here are not real indexes */
6942 175716 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6943 756 : return false;
6944 :
6945 : /* Search through the indexes to see if any match our problem */
6946 340738 : foreach(lc, rel->indexlist)
6947 : {
6948 293092 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6949 : ScanDirection indexscandir;
6950 : StrategyNumber strategy;
6951 :
6952 : /* Ignore non-ordering indexes */
6953 293092 : if (index->sortopfamily == NULL)
6954 0 : continue;
6955 :
6956 : /*
6957 : * Ignore partial indexes --- we only want stats that cover the entire
6958 : * relation.
6959 : */
6960 293092 : if (index->indpred != NIL)
6961 288 : continue;
6962 :
6963 : /*
6964 : * The index list might include hypothetical indexes inserted by a
6965 : * get_relation_info hook --- don't try to access them.
6966 : */
6967 292804 : if (index->hypothetical)
6968 0 : continue;
6969 :
6970 : /*
6971 : * get_actual_variable_endpoint uses the index-only-scan machinery, so
6972 : * ignore indexes that can't use it on their first column.
6973 : */
6974 292804 : if (!index->canreturn[0])
6975 0 : continue;
6976 :
6977 : /*
6978 : * The first index column must match the desired variable, sortop, and
6979 : * collation --- but we can use a descending-order index.
6980 : */
6981 292804 : if (collation != index->indexcollations[0])
6982 38792 : continue; /* test first 'cause it's cheapest */
6983 254012 : if (!match_index_to_operand(vardata->var, 0, index))
6984 126698 : continue;
6985 127314 : strategy = get_op_opfamily_strategy(sortop, index->sortopfamily[0]);
6986 127314 : switch (IndexAmTranslateStrategy(strategy, index->relam, index->sortopfamily[0], true))
6987 : {
6988 127314 : case COMPARE_LT:
6989 127314 : if (index->reverse_sort[0])
6990 0 : indexscandir = BackwardScanDirection;
6991 : else
6992 127314 : indexscandir = ForwardScanDirection;
6993 127314 : break;
6994 0 : case COMPARE_GT:
6995 0 : if (index->reverse_sort[0])
6996 0 : indexscandir = ForwardScanDirection;
6997 : else
6998 0 : indexscandir = BackwardScanDirection;
6999 0 : break;
7000 0 : default:
7001 : /* index doesn't match the sortop */
7002 0 : continue;
7003 : }
7004 :
7005 : /*
7006 : * Found a suitable index to extract data from. Set up some data that
7007 : * can be used by both invocations of get_actual_variable_endpoint.
7008 : */
7009 : {
7010 : MemoryContext tmpcontext;
7011 : MemoryContext oldcontext;
7012 : Relation heapRel;
7013 : Relation indexRel;
7014 : TupleTableSlot *slot;
7015 : int16 typLen;
7016 : bool typByVal;
7017 : ScanKeyData scankeys[1];
7018 :
7019 : /* Make sure any cruft gets recycled when we're done */
7020 127314 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
7021 : "get_actual_variable_range workspace",
7022 : ALLOCSET_DEFAULT_SIZES);
7023 127314 : oldcontext = MemoryContextSwitchTo(tmpcontext);
7024 :
7025 : /*
7026 : * Open the table and index so we can read from them. We should
7027 : * already have some type of lock on each.
7028 : */
7029 127314 : heapRel = table_open(rte->relid, NoLock);
7030 127314 : indexRel = index_open(index->indexoid, NoLock);
7031 :
7032 : /* build some stuff needed for indexscan execution */
7033 127314 : slot = table_slot_create(heapRel, NULL);
7034 127314 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
7035 :
7036 : /* set up an IS NOT NULL scan key so that we ignore nulls */
7037 127314 : ScanKeyEntryInitialize(&scankeys[0],
7038 : SK_ISNULL | SK_SEARCHNOTNULL,
7039 : 1, /* index col to scan */
7040 : InvalidStrategy, /* no strategy */
7041 : InvalidOid, /* no strategy subtype */
7042 : InvalidOid, /* no collation */
7043 : InvalidOid, /* no reg proc for this */
7044 : (Datum) 0); /* constant */
7045 :
7046 : /* If min is requested ... */
7047 127314 : if (min)
7048 : {
7049 71838 : have_data = get_actual_variable_endpoint(heapRel,
7050 : indexRel,
7051 : indexscandir,
7052 : scankeys,
7053 : typLen,
7054 : typByVal,
7055 : slot,
7056 : oldcontext,
7057 : min);
7058 : }
7059 : else
7060 : {
7061 : /* If min not requested, still want to fetch max */
7062 55476 : have_data = true;
7063 : }
7064 :
7065 : /* If max is requested, and we didn't already fail ... */
7066 127314 : if (max && have_data)
7067 : {
7068 : /* scan in the opposite direction; all else is the same */
7069 57228 : have_data = get_actual_variable_endpoint(heapRel,
7070 : indexRel,
7071 57228 : -indexscandir,
7072 : scankeys,
7073 : typLen,
7074 : typByVal,
7075 : slot,
7076 : oldcontext,
7077 : max);
7078 : }
7079 :
7080 : /* Clean everything up */
7081 127314 : ExecDropSingleTupleTableSlot(slot);
7082 :
7083 127314 : index_close(indexRel, NoLock);
7084 127314 : table_close(heapRel, NoLock);
7085 :
7086 127314 : MemoryContextSwitchTo(oldcontext);
7087 127314 : MemoryContextDelete(tmpcontext);
7088 :
7089 : /* And we're done */
7090 127314 : break;
7091 : }
7092 : }
7093 :
7094 174960 : return have_data;
7095 : }
7096 :
7097 : /*
7098 : * Get one endpoint datum (min or max depending on indexscandir) from the
7099 : * specified index. Return true if successful, false if not.
7100 : * On success, endpoint value is stored to *endpointDatum (and copied into
7101 : * outercontext).
7102 : *
7103 : * scankeys is a 1-element scankey array set up to reject nulls.
7104 : * typLen/typByVal describe the datatype of the index's first column.
7105 : * tableslot is a slot suitable to hold table tuples, in case we need
7106 : * to probe the heap.
7107 : * (We could compute these values locally, but that would mean computing them
7108 : * twice when get_actual_variable_range needs both the min and the max.)
7109 : *
7110 : * Failure occurs either when the index is empty, or we decide that it's
7111 : * taking too long to find a suitable tuple.
7112 : */
7113 : static bool
7114 129066 : get_actual_variable_endpoint(Relation heapRel,
7115 : Relation indexRel,
7116 : ScanDirection indexscandir,
7117 : ScanKey scankeys,
7118 : int16 typLen,
7119 : bool typByVal,
7120 : TupleTableSlot *tableslot,
7121 : MemoryContext outercontext,
7122 : Datum *endpointDatum)
7123 : {
7124 129066 : bool have_data = false;
7125 : SnapshotData SnapshotNonVacuumable;
7126 : IndexScanDesc index_scan;
7127 129066 : Buffer vmbuffer = InvalidBuffer;
7128 129066 : BlockNumber last_heap_block = InvalidBlockNumber;
7129 129066 : int n_visited_heap_pages = 0;
7130 : ItemPointer tid;
7131 : Datum values[INDEX_MAX_KEYS];
7132 : bool isnull[INDEX_MAX_KEYS];
7133 : MemoryContext oldcontext;
7134 :
7135 : /*
7136 : * We use the index-only-scan machinery for this. With mostly-static
7137 : * tables that's a win because it avoids a heap visit. It's also a win
7138 : * for dynamic data, but the reason is less obvious; read on for details.
7139 : *
7140 : * In principle, we should scan the index with our current active
7141 : * snapshot, which is the best approximation we've got to what the query
7142 : * will see when executed. But that won't be exact if a new snap is taken
7143 : * before running the query, and it can be very expensive if a lot of
7144 : * recently-dead or uncommitted rows exist at the beginning or end of the
7145 : * index (because we'll laboriously fetch each one and reject it).
7146 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
7147 : * and uncommitted rows as well as normal visible rows. On the other
7148 : * hand, it will reject known-dead rows, and thus not give a bogus answer
7149 : * when the extreme value has been deleted (unless the deletion was quite
7150 : * recent); that case motivates not using SnapshotAny here.
7151 : *
7152 : * A crucial point here is that SnapshotNonVacuumable, with
7153 : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
7154 : * condition that the indexscan will use to decide that index entries are
7155 : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
7156 : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
7157 : * have to continue scanning past it, we know that the indexscan will mark
7158 : * that index entry killed. That means that the next
7159 : * get_actual_variable_endpoint() call will not have to re-consider that
7160 : * index entry. In this way we avoid repetitive work when this function
7161 : * is used a lot during planning.
7162 : *
7163 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
7164 : * recently-created index, some index entries may point at "broken" HOT
7165 : * chains in which not all the tuple versions contain data matching the
7166 : * index entry. The live tuple version(s) certainly do match the index,
7167 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
7168 : * don't match. Hence, if we took data from the selected heap tuple, we
7169 : * might get a bogus answer that's not close to the index extremal value,
7170 : * or could even be NULL. We avoid this hazard because we take the data
7171 : * from the index entry not the heap.
7172 : *
7173 : * Despite all this care, there are situations where we might find many
7174 : * non-visible tuples near the end of the index. We don't want to expend
7175 : * a huge amount of time here, so we give up once we've read too many heap
7176 : * pages. When we fail for that reason, the caller will end up using
7177 : * whatever extremal value is recorded in pg_statistic.
7178 : */
7179 129066 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
7180 : GlobalVisTestFor(heapRel));
7181 :
7182 129066 : index_scan = index_beginscan(heapRel, indexRel,
7183 : &SnapshotNonVacuumable, NULL,
7184 : 1, 0);
7185 : /* Set it up for index-only scan */
7186 129066 : index_scan->xs_want_itup = true;
7187 129066 : index_rescan(index_scan, scankeys, 1, NULL, 0);
7188 :
7189 : /* Fetch first/next tuple in specified direction */
7190 151292 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
7191 : {
7192 151292 : BlockNumber block = ItemPointerGetBlockNumber(tid);
7193 :
7194 151292 : if (!VM_ALL_VISIBLE(heapRel,
7195 : block,
7196 : &vmbuffer))
7197 : {
7198 : /* Rats, we have to visit the heap to check visibility */
7199 105244 : if (!index_fetch_heap(index_scan, tableslot))
7200 : {
7201 : /*
7202 : * No visible tuple for this index entry, so we need to
7203 : * advance to the next entry. Before doing so, count heap
7204 : * page fetches and give up if we've done too many.
7205 : *
7206 : * We don't charge a page fetch if this is the same heap page
7207 : * as the previous tuple. This is on the conservative side,
7208 : * since other recently-accessed pages are probably still in
7209 : * buffers too; but it's good enough for this heuristic.
7210 : */
7211 : #define VISITED_PAGES_LIMIT 100
7212 :
7213 22226 : if (block != last_heap_block)
7214 : {
7215 2826 : last_heap_block = block;
7216 2826 : n_visited_heap_pages++;
7217 2826 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
7218 0 : break;
7219 : }
7220 :
7221 22226 : continue; /* no visible tuple, try next index entry */
7222 : }
7223 :
7224 : /* We don't actually need the heap tuple for anything */
7225 83018 : ExecClearTuple(tableslot);
7226 :
7227 : /*
7228 : * We don't care whether there's more than one visible tuple in
7229 : * the HOT chain; if any are visible, that's good enough.
7230 : */
7231 : }
7232 :
7233 : /*
7234 : * We expect that the index will return data in IndexTuple not
7235 : * HeapTuple format.
7236 : */
7237 129066 : if (!index_scan->xs_itup)
7238 0 : elog(ERROR, "no data returned for index-only scan");
7239 :
7240 : /*
7241 : * We do not yet support recheck here.
7242 : */
7243 129066 : if (index_scan->xs_recheck)
7244 0 : break;
7245 :
7246 : /* OK to deconstruct the index tuple */
7247 129066 : index_deform_tuple(index_scan->xs_itup,
7248 : index_scan->xs_itupdesc,
7249 : values, isnull);
7250 :
7251 : /* Shouldn't have got a null, but be careful */
7252 129066 : if (isnull[0])
7253 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
7254 : RelationGetRelationName(indexRel));
7255 :
7256 : /* Copy the index column value out to caller's context */
7257 129066 : oldcontext = MemoryContextSwitchTo(outercontext);
7258 129066 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
7259 129066 : MemoryContextSwitchTo(oldcontext);
7260 129066 : have_data = true;
7261 129066 : break;
7262 : }
7263 :
7264 129066 : if (vmbuffer != InvalidBuffer)
7265 116320 : ReleaseBuffer(vmbuffer);
7266 129066 : index_endscan(index_scan);
7267 :
7268 129066 : return have_data;
7269 : }
7270 :
7271 : /*
7272 : * find_join_input_rel
7273 : * Look up the input relation for a join.
7274 : *
7275 : * We assume that the input relation's RelOptInfo must have been constructed
7276 : * already.
7277 : */
7278 : static RelOptInfo *
7279 10884 : find_join_input_rel(PlannerInfo *root, Relids relids)
7280 : {
7281 10884 : RelOptInfo *rel = NULL;
7282 :
7283 10884 : if (!bms_is_empty(relids))
7284 : {
7285 : int relid;
7286 :
7287 10884 : if (bms_get_singleton_member(relids, &relid))
7288 10568 : rel = find_base_rel(root, relid);
7289 : else
7290 316 : rel = find_join_rel(root, relids);
7291 : }
7292 :
7293 10884 : if (rel == NULL)
7294 0 : elog(ERROR, "could not find RelOptInfo for given relids");
7295 :
7296 10884 : return rel;
7297 : }
7298 :
7299 :
7300 : /*-------------------------------------------------------------------------
7301 : *
7302 : * Index cost estimation functions
7303 : *
7304 : *-------------------------------------------------------------------------
7305 : */
7306 :
7307 : /*
7308 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
7309 : */
7310 : List *
7311 823854 : get_quals_from_indexclauses(List *indexclauses)
7312 : {
7313 823854 : List *result = NIL;
7314 : ListCell *lc;
7315 :
7316 1450382 : foreach(lc, indexclauses)
7317 : {
7318 626528 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7319 : ListCell *lc2;
7320 :
7321 1255978 : foreach(lc2, iclause->indexquals)
7322 : {
7323 629450 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7324 :
7325 629450 : result = lappend(result, rinfo);
7326 : }
7327 : }
7328 823854 : return result;
7329 : }
7330 :
7331 : /*
7332 : * Compute the total evaluation cost of the comparison operands in a list
7333 : * of index qual expressions. Since we know these will be evaluated just
7334 : * once per scan, there's no need to distinguish startup from per-row cost.
7335 : *
7336 : * This can be used either on the result of get_quals_from_indexclauses(),
7337 : * or directly on an indexorderbys list. In both cases, we expect that the
7338 : * index key expression is on the left side of binary clauses.
7339 : */
7340 : Cost
7341 1634702 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
7342 : {
7343 1634702 : Cost qual_arg_cost = 0;
7344 : ListCell *lc;
7345 :
7346 2264614 : foreach(lc, indexquals)
7347 : {
7348 629912 : Expr *clause = (Expr *) lfirst(lc);
7349 : Node *other_operand;
7350 : QualCost index_qual_cost;
7351 :
7352 : /*
7353 : * Index quals will have RestrictInfos, indexorderbys won't. Look
7354 : * through RestrictInfo if present.
7355 : */
7356 629912 : if (IsA(clause, RestrictInfo))
7357 629438 : clause = ((RestrictInfo *) clause)->clause;
7358 :
7359 629912 : if (IsA(clause, OpExpr))
7360 : {
7361 614860 : OpExpr *op = (OpExpr *) clause;
7362 :
7363 614860 : other_operand = (Node *) lsecond(op->args);
7364 : }
7365 15052 : else if (IsA(clause, RowCompareExpr))
7366 : {
7367 396 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7368 :
7369 396 : other_operand = (Node *) rc->rargs;
7370 : }
7371 14656 : else if (IsA(clause, ScalarArrayOpExpr))
7372 : {
7373 11730 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7374 :
7375 11730 : other_operand = (Node *) lsecond(saop->args);
7376 : }
7377 2926 : else if (IsA(clause, NullTest))
7378 : {
7379 2926 : other_operand = NULL;
7380 : }
7381 : else
7382 : {
7383 0 : elog(ERROR, "unsupported indexqual type: %d",
7384 : (int) nodeTag(clause));
7385 : other_operand = NULL; /* keep compiler quiet */
7386 : }
7387 :
7388 629912 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
7389 629912 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
7390 : }
7391 1634702 : return qual_arg_cost;
7392 : }
7393 :
7394 : void
7395 810860 : genericcostestimate(PlannerInfo *root,
7396 : IndexPath *path,
7397 : double loop_count,
7398 : GenericCosts *costs)
7399 : {
7400 810860 : IndexOptInfo *index = path->indexinfo;
7401 810860 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7402 810860 : List *indexOrderBys = path->indexorderbys;
7403 : Cost indexStartupCost;
7404 : Cost indexTotalCost;
7405 : Selectivity indexSelectivity;
7406 : double indexCorrelation;
7407 : double numIndexPages;
7408 : double numIndexTuples;
7409 : double spc_random_page_cost;
7410 : double num_sa_scans;
7411 : double num_outer_scans;
7412 : double num_scans;
7413 : double qual_op_cost;
7414 : double qual_arg_cost;
7415 : List *selectivityQuals;
7416 : ListCell *l;
7417 :
7418 : /*
7419 : * If the index is partial, AND the index predicate with the explicitly
7420 : * given indexquals to produce a more accurate idea of the index
7421 : * selectivity.
7422 : */
7423 810860 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7424 :
7425 : /*
7426 : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
7427 : * just assume that the number of index descents is the number of distinct
7428 : * combinations of array elements from all of the scan's SAOP clauses.
7429 : */
7430 810860 : num_sa_scans = costs->num_sa_scans;
7431 810860 : if (num_sa_scans < 1)
7432 : {
7433 7904 : num_sa_scans = 1;
7434 16592 : foreach(l, indexQuals)
7435 : {
7436 8688 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
7437 :
7438 8688 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
7439 : {
7440 26 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
7441 26 : double alength = estimate_array_length(root, lsecond(saop->args));
7442 :
7443 26 : if (alength > 1)
7444 26 : num_sa_scans *= alength;
7445 : }
7446 : }
7447 : }
7448 :
7449 : /* Estimate the fraction of main-table tuples that will be visited */
7450 810860 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7451 810860 : index->rel->relid,
7452 : JOIN_INNER,
7453 : NULL);
7454 :
7455 : /*
7456 : * If caller didn't give us an estimate, estimate the number of index
7457 : * tuples that will be visited. We do it in this rather peculiar-looking
7458 : * way in order to get the right answer for partial indexes.
7459 : */
7460 810860 : numIndexTuples = costs->numIndexTuples;
7461 810860 : if (numIndexTuples <= 0.0)
7462 : {
7463 92682 : numIndexTuples = indexSelectivity * index->rel->tuples;
7464 :
7465 : /*
7466 : * The above calculation counts all the tuples visited across all
7467 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
7468 : * average per-indexscan number, so adjust. This is a handy place to
7469 : * round to integer, too. (If caller supplied tuple estimate, it's
7470 : * responsible for handling these considerations.)
7471 : */
7472 92682 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7473 : }
7474 :
7475 : /*
7476 : * We can bound the number of tuples by the index size in any case. Also,
7477 : * always estimate at least one tuple is touched, even when
7478 : * indexSelectivity estimate is tiny.
7479 : */
7480 810860 : if (numIndexTuples > index->tuples)
7481 6602 : numIndexTuples = index->tuples;
7482 810860 : if (numIndexTuples < 1.0)
7483 93430 : numIndexTuples = 1.0;
7484 :
7485 : /*
7486 : * Estimate the number of index pages that will be retrieved.
7487 : *
7488 : * We use the simplistic method of taking a pro-rata fraction of the total
7489 : * number of index pages. In effect, this counts only leaf pages and not
7490 : * any overhead such as index metapage or upper tree levels.
7491 : *
7492 : * In practice access to upper index levels is often nearly free because
7493 : * those tend to stay in cache under load; moreover, the cost involved is
7494 : * highly dependent on index type. We therefore ignore such costs here
7495 : * and leave it to the caller to add a suitable charge if needed.
7496 : */
7497 810860 : if (index->pages > 1 && index->tuples > 1)
7498 746158 : numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
7499 : else
7500 64702 : numIndexPages = 1.0;
7501 :
7502 : /* fetch estimated page cost for tablespace containing index */
7503 810860 : get_tablespace_page_costs(index->reltablespace,
7504 : &spc_random_page_cost,
7505 : NULL);
7506 :
7507 : /*
7508 : * Now compute the disk access costs.
7509 : *
7510 : * The above calculations are all per-index-scan. However, if we are in a
7511 : * nestloop inner scan, we can expect the scan to be repeated (with
7512 : * different search keys) for each row of the outer relation. Likewise,
7513 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
7514 : * the potential for cache effects to reduce the number of disk page
7515 : * fetches needed. We want to estimate the average per-scan I/O cost in
7516 : * the presence of caching.
7517 : *
7518 : * We use the Mackert-Lohman formula (see costsize.c for details) to
7519 : * estimate the total number of page fetches that occur. While this
7520 : * wasn't what it was designed for, it seems a reasonable model anyway.
7521 : * Note that we are counting pages not tuples anymore, so we take N = T =
7522 : * index size, as if there were one "tuple" per page.
7523 : */
7524 810860 : num_outer_scans = loop_count;
7525 810860 : num_scans = num_sa_scans * num_outer_scans;
7526 :
7527 810860 : if (num_scans > 1)
7528 : {
7529 : double pages_fetched;
7530 :
7531 : /* total page fetches ignoring cache effects */
7532 96016 : pages_fetched = numIndexPages * num_scans;
7533 :
7534 : /* use Mackert and Lohman formula to adjust for cache effects */
7535 96016 : pages_fetched = index_pages_fetched(pages_fetched,
7536 : index->pages,
7537 96016 : (double) index->pages,
7538 : root);
7539 :
7540 : /*
7541 : * Now compute the total disk access cost, and then report a pro-rated
7542 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
7543 : * since that's internal to the indexscan.)
7544 : */
7545 96016 : indexTotalCost = (pages_fetched * spc_random_page_cost)
7546 : / num_outer_scans;
7547 : }
7548 : else
7549 : {
7550 : /*
7551 : * For a single index scan, we just charge spc_random_page_cost per
7552 : * page touched.
7553 : */
7554 714844 : indexTotalCost = numIndexPages * spc_random_page_cost;
7555 : }
7556 :
7557 : /*
7558 : * CPU cost: any complex expressions in the indexquals will need to be
7559 : * evaluated once at the start of the scan to reduce them to runtime keys
7560 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
7561 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
7562 : * indexqual operator. Because we have numIndexTuples as a per-scan
7563 : * number, we have to multiply by num_sa_scans to get the correct result
7564 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
7565 : * ORDER BY expressions.
7566 : *
7567 : * Note: this neglects the possible costs of rechecking lossy operators.
7568 : * Detecting that that might be needed seems more expensive than it's
7569 : * worth, though, considering all the other inaccuracies here ...
7570 : */
7571 810860 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
7572 810860 : index_other_operands_eval_cost(root, indexOrderBys);
7573 810860 : qual_op_cost = cpu_operator_cost *
7574 810860 : (list_length(indexQuals) + list_length(indexOrderBys));
7575 :
7576 810860 : indexStartupCost = qual_arg_cost;
7577 810860 : indexTotalCost += qual_arg_cost;
7578 810860 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
7579 :
7580 : /*
7581 : * Generic assumption about index correlation: there isn't any.
7582 : */
7583 810860 : indexCorrelation = 0.0;
7584 :
7585 : /*
7586 : * Return everything to caller.
7587 : */
7588 810860 : costs->indexStartupCost = indexStartupCost;
7589 810860 : costs->indexTotalCost = indexTotalCost;
7590 810860 : costs->indexSelectivity = indexSelectivity;
7591 810860 : costs->indexCorrelation = indexCorrelation;
7592 810860 : costs->numIndexPages = numIndexPages;
7593 810860 : costs->numIndexTuples = numIndexTuples;
7594 810860 : costs->spc_random_page_cost = spc_random_page_cost;
7595 810860 : costs->num_sa_scans = num_sa_scans;
7596 810860 : }
7597 :
7598 : /*
7599 : * If the index is partial, add its predicate to the given qual list.
7600 : *
7601 : * ANDing the index predicate with the explicitly given indexquals produces
7602 : * a more accurate idea of the index's selectivity. However, we need to be
7603 : * careful not to insert redundant clauses, because clauselist_selectivity()
7604 : * is easily fooled into computing a too-low selectivity estimate. Our
7605 : * approach is to add only the predicate clause(s) that cannot be proven to
7606 : * be implied by the given indexquals. This successfully handles cases such
7607 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
7608 : * There are many other cases where we won't detect redundancy, leading to a
7609 : * too-low selectivity estimate, which will bias the system in favor of using
7610 : * partial indexes where possible. That is not necessarily bad though.
7611 : *
7612 : * Note that indexQuals contains RestrictInfo nodes while the indpred
7613 : * does not, so the output list will be mixed. This is OK for both
7614 : * predicate_implied_by() and clauselist_selectivity(), but might be
7615 : * problematic if the result were passed to other things.
7616 : */
7617 : List *
7618 1369768 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
7619 : {
7620 1369768 : List *predExtraQuals = NIL;
7621 : ListCell *lc;
7622 :
7623 1369768 : if (index->indpred == NIL)
7624 1367750 : return indexQuals;
7625 :
7626 4048 : foreach(lc, index->indpred)
7627 : {
7628 2030 : Node *predQual = (Node *) lfirst(lc);
7629 2030 : List *oneQual = list_make1(predQual);
7630 :
7631 2030 : if (!predicate_implied_by(oneQual, indexQuals, false))
7632 1812 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7633 : }
7634 2018 : return list_concat(predExtraQuals, indexQuals);
7635 : }
7636 :
7637 : /*
7638 : * Estimate correlation of btree index's first column.
7639 : *
7640 : * If we can get an estimate of the first column's ordering correlation C
7641 : * from pg_statistic, estimate the index correlation as C for a single-column
7642 : * index, or C * 0.75 for multiple columns. The idea here is that multiple
7643 : * columns dilute the importance of the first column's ordering, but don't
7644 : * negate it entirely.
7645 : *
7646 : * We already filled in the stats tuple for *vardata when called.
7647 : */
7648 : static double
7649 600654 : btcost_correlation(IndexOptInfo *index, VariableStatData *vardata)
7650 : {
7651 : Oid sortop;
7652 : AttStatsSlot sslot;
7653 600654 : double indexCorrelation = 0;
7654 :
7655 : Assert(HeapTupleIsValid(vardata->statsTuple));
7656 :
7657 600654 : sortop = get_opfamily_member(index->opfamily[0],
7658 600654 : index->opcintype[0],
7659 600654 : index->opcintype[0],
7660 : BTLessStrategyNumber);
7661 1201308 : if (OidIsValid(sortop) &&
7662 600654 : get_attstatsslot(&sslot, vardata->statsTuple,
7663 : STATISTIC_KIND_CORRELATION, sortop,
7664 : ATTSTATSSLOT_NUMBERS))
7665 : {
7666 : double varCorrelation;
7667 :
7668 : Assert(sslot.nnumbers == 1);
7669 592424 : varCorrelation = sslot.numbers[0];
7670 :
7671 592424 : if (index->reverse_sort[0])
7672 0 : varCorrelation = -varCorrelation;
7673 :
7674 592424 : if (index->nkeycolumns > 1)
7675 207362 : indexCorrelation = varCorrelation * 0.75;
7676 : else
7677 385062 : indexCorrelation = varCorrelation;
7678 :
7679 592424 : free_attstatsslot(&sslot);
7680 : }
7681 :
7682 600654 : return indexCorrelation;
7683 : }
7684 :
7685 : void
7686 802956 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7687 : Cost *indexStartupCost, Cost *indexTotalCost,
7688 : Selectivity *indexSelectivity, double *indexCorrelation,
7689 : double *indexPages)
7690 : {
7691 802956 : IndexOptInfo *index = path->indexinfo;
7692 802956 : GenericCosts costs = {0};
7693 802956 : VariableStatData vardata = {0};
7694 : double numIndexTuples;
7695 : Cost descentCost;
7696 : List *indexBoundQuals;
7697 : List *indexSkipQuals;
7698 : int indexcol;
7699 : bool eqQualHere;
7700 : bool found_row_compare;
7701 : bool found_array;
7702 : bool found_is_null_op;
7703 802956 : bool have_correlation = false;
7704 : double num_sa_scans;
7705 802956 : double correlation = 0.0;
7706 : ListCell *lc;
7707 :
7708 : /*
7709 : * For a btree scan, only leading '=' quals plus inequality quals for the
7710 : * immediately next attribute contribute to index selectivity (these are
7711 : * the "boundary quals" that determine the starting and stopping points of
7712 : * the index scan). Additional quals can suppress visits to the heap, so
7713 : * it's OK to count them in indexSelectivity, but they should not count
7714 : * for estimating numIndexTuples. So we must examine the given indexquals
7715 : * to find out which ones count as boundary quals. We rely on the
7716 : * knowledge that they are given in index column order. Note that nbtree
7717 : * preprocessing can add skip arrays that act as leading '=' quals in the
7718 : * absence of ordinary input '=' quals, so in practice _most_ input quals
7719 : * are able to act as index bound quals (which we take into account here).
7720 : *
7721 : * For a RowCompareExpr, we consider only the first column, just as
7722 : * rowcomparesel() does.
7723 : *
7724 : * If there's a SAOP or skip array in the quals, we'll actually perform up
7725 : * to N index descents (not just one), but the underlying array key's
7726 : * operator can be considered to act the same as it normally does.
7727 : */
7728 802956 : indexBoundQuals = NIL;
7729 802956 : indexSkipQuals = NIL;
7730 802956 : indexcol = 0;
7731 802956 : eqQualHere = false;
7732 802956 : found_row_compare = false;
7733 802956 : found_array = false;
7734 802956 : found_is_null_op = false;
7735 802956 : num_sa_scans = 1;
7736 1367134 : foreach(lc, path->indexclauses)
7737 : {
7738 601554 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7739 : ListCell *lc2;
7740 :
7741 601554 : if (indexcol < iclause->indexcol)
7742 : {
7743 117044 : double num_sa_scans_prev_cols = num_sa_scans;
7744 :
7745 : /*
7746 : * Beginning of a new column's quals.
7747 : *
7748 : * Skip scans use skip arrays, which are ScalarArrayOp style
7749 : * arrays that generate their elements procedurally and on demand.
7750 : * Given a multi-column index on "(a, b)", and an SQL WHERE clause
7751 : * "WHERE b = 42", a skip scan will effectively use an indexqual
7752 : * "WHERE a = ANY('{every col a value}') AND b = 42". (Obviously,
7753 : * the array on "a" must also return "IS NULL" matches, since our
7754 : * WHERE clause used no strict operator on "a").
7755 : *
7756 : * Here we consider how nbtree will backfill skip arrays for any
7757 : * index columns that lacked an '=' qual. This maintains our
7758 : * num_sa_scans estimate, and determines if this new column (the
7759 : * "iclause->indexcol" column, not the prior "indexcol" column)
7760 : * can have its RestrictInfos/quals added to indexBoundQuals.
7761 : *
7762 : * We'll need to handle columns that have inequality quals, where
7763 : * the skip array generates values from a range constrained by the
7764 : * quals (not every possible value). We've been maintaining
7765 : * indexSkipQuals to help with this; it will now contain all of
7766 : * the prior column's quals (that is, indexcol's quals) when they
7767 : * might be used for this.
7768 : */
7769 117044 : if (found_row_compare)
7770 : {
7771 : /*
7772 : * Skip arrays can't be added after a RowCompare input qual
7773 : * due to limitations in nbtree
7774 : */
7775 24 : break;
7776 : }
7777 117020 : if (eqQualHere)
7778 : {
7779 : /*
7780 : * Don't need to add a skip array for an indexcol that already
7781 : * has an '=' qual/equality constraint
7782 : */
7783 80316 : indexcol++;
7784 80316 : indexSkipQuals = NIL;
7785 : }
7786 117020 : eqQualHere = false;
7787 :
7788 120010 : while (indexcol < iclause->indexcol)
7789 : {
7790 : double ndistinct;
7791 40342 : bool isdefault = true;
7792 :
7793 40342 : found_array = true;
7794 :
7795 : /*
7796 : * A skipped attribute's ndistinct forms the basis of our
7797 : * estimate of the total number of "array elements" used by
7798 : * its skip array at runtime. Look that up first.
7799 : */
7800 40342 : examine_indexcol_variable(root, index, indexcol, &vardata);
7801 40342 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
7802 :
7803 40342 : if (indexcol == 0)
7804 : {
7805 : /*
7806 : * Get an estimate of the leading column's correlation in
7807 : * passing (avoids rereading variable stats below)
7808 : */
7809 36692 : if (HeapTupleIsValid(vardata.statsTuple))
7810 23846 : correlation = btcost_correlation(index, &vardata);
7811 36692 : have_correlation = true;
7812 : }
7813 :
7814 40342 : ReleaseVariableStats(vardata);
7815 :
7816 : /*
7817 : * If ndistinct is a default estimate, conservatively assume
7818 : * that no skipping will happen at runtime
7819 : */
7820 40342 : if (isdefault)
7821 : {
7822 11544 : num_sa_scans = num_sa_scans_prev_cols;
7823 37352 : break; /* done building indexBoundQuals */
7824 : }
7825 :
7826 : /*
7827 : * Apply indexcol's indexSkipQuals selectivity to ndistinct
7828 : */
7829 28798 : if (indexSkipQuals != NIL)
7830 : {
7831 : List *partialSkipQuals;
7832 : Selectivity ndistinctfrac;
7833 :
7834 : /*
7835 : * If the index is partial, AND the index predicate with
7836 : * the index-bound quals to produce a more accurate idea
7837 : * of the number of distinct values for prior indexcol
7838 : */
7839 664 : partialSkipQuals = add_predicate_to_index_quals(index,
7840 : indexSkipQuals);
7841 :
7842 664 : ndistinctfrac = clauselist_selectivity(root, partialSkipQuals,
7843 664 : index->rel->relid,
7844 : JOIN_INNER,
7845 : NULL);
7846 :
7847 : /*
7848 : * If ndistinctfrac is selective (on its own), the scan is
7849 : * unlikely to benefit from repositioning itself using
7850 : * later quals. Do not allow iclause->indexcol's quals to
7851 : * be added to indexBoundQuals (it would increase descent
7852 : * costs, without lowering numIndexTuples costs by much).
7853 : */
7854 664 : if (ndistinctfrac < DEFAULT_RANGE_INEQ_SEL)
7855 : {
7856 374 : num_sa_scans = num_sa_scans_prev_cols;
7857 374 : break; /* done building indexBoundQuals */
7858 : }
7859 :
7860 : /* Adjust ndistinct downward */
7861 290 : ndistinct = rint(ndistinct * ndistinctfrac);
7862 290 : ndistinct = Max(ndistinct, 1);
7863 : }
7864 :
7865 : /*
7866 : * When there's no inequality quals, account for the need to
7867 : * find an initial value by counting -inf/+inf as a value.
7868 : *
7869 : * We don't charge anything extra for possible next/prior key
7870 : * index probes, which are sometimes used to find the next
7871 : * valid skip array element (ahead of using the located
7872 : * element value to relocate the scan to the next position
7873 : * that might contain matching tuples). It seems hard to do
7874 : * better here. Use of the skip support infrastructure often
7875 : * avoids most next/prior key probes. But even when it can't,
7876 : * there's a decent chance that most individual next/prior key
7877 : * probes will locate a leaf page whose key space overlaps all
7878 : * of the scan's keys (even the lower-order keys) -- which
7879 : * also avoids the need for a separate, extra index descent.
7880 : * Note also that these probes are much cheaper than non-probe
7881 : * primitive index scans: they're reliably very selective.
7882 : */
7883 28424 : if (indexSkipQuals == NIL)
7884 28134 : ndistinct += 1;
7885 :
7886 : /*
7887 : * Update num_sa_scans estimate by multiplying by ndistinct.
7888 : *
7889 : * We make the pessimistic assumption that there is no
7890 : * naturally occurring cross-column correlation. This is
7891 : * often wrong, but it seems best to err on the side of not
7892 : * expecting skipping to be helpful...
7893 : */
7894 28424 : num_sa_scans *= ndistinct;
7895 :
7896 : /*
7897 : * ...but back out of adding this latest group of 1 or more
7898 : * skip arrays when num_sa_scans exceeds the total number of
7899 : * index pages (revert to num_sa_scans from before indexcol).
7900 : * This causes a sharp discontinuity in cost (as a function of
7901 : * the indexcol's ndistinct), but that is representative of
7902 : * actual runtime costs.
7903 : *
7904 : * Note that skipping is helpful when each primitive index
7905 : * scan only manages to skip over 1 or 2 irrelevant leaf pages
7906 : * on average. Skip arrays bring savings in CPU costs due to
7907 : * the scan not needing to evaluate indexquals against every
7908 : * tuple, which can greatly exceed any savings in I/O costs.
7909 : * This test is a test of whether num_sa_scans implies that
7910 : * we're past the point where the ability to skip ceases to
7911 : * lower the scan's costs (even qual evaluation CPU costs).
7912 : */
7913 28424 : if (index->pages < num_sa_scans)
7914 : {
7915 25434 : num_sa_scans = num_sa_scans_prev_cols;
7916 25434 : break; /* done building indexBoundQuals */
7917 : }
7918 :
7919 2990 : indexcol++;
7920 2990 : indexSkipQuals = NIL;
7921 : }
7922 :
7923 : /*
7924 : * Finished considering the need to add skip arrays to bridge an
7925 : * initial eqQualHere gap between the old and new index columns
7926 : * (or there was no initial eqQualHere gap in the first place).
7927 : *
7928 : * If an initial gap could not be bridged, then new column's quals
7929 : * (i.e. iclause->indexcol's quals) won't go into indexBoundQuals,
7930 : * and so won't affect our final numIndexTuples estimate.
7931 : */
7932 117020 : if (indexcol != iclause->indexcol)
7933 37352 : break; /* done building indexBoundQuals */
7934 : }
7935 :
7936 : Assert(indexcol == iclause->indexcol);
7937 :
7938 : /* Examine each indexqual associated with this index clause */
7939 1131098 : foreach(lc2, iclause->indexquals)
7940 : {
7941 566920 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7942 566920 : Expr *clause = rinfo->clause;
7943 566920 : Oid clause_op = InvalidOid;
7944 : int op_strategy;
7945 :
7946 566920 : if (IsA(clause, OpExpr))
7947 : {
7948 552938 : OpExpr *op = (OpExpr *) clause;
7949 :
7950 552938 : clause_op = op->opno;
7951 : }
7952 13982 : else if (IsA(clause, RowCompareExpr))
7953 : {
7954 396 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7955 :
7956 396 : clause_op = linitial_oid(rc->opnos);
7957 396 : found_row_compare = true;
7958 : }
7959 13586 : else if (IsA(clause, ScalarArrayOpExpr))
7960 : {
7961 11302 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7962 11302 : Node *other_operand = (Node *) lsecond(saop->args);
7963 11302 : double alength = estimate_array_length(root, other_operand);
7964 :
7965 11302 : clause_op = saop->opno;
7966 11302 : found_array = true;
7967 : /* estimate SA descents by indexBoundQuals only */
7968 11302 : if (alength > 1)
7969 10994 : num_sa_scans *= alength;
7970 : }
7971 2284 : else if (IsA(clause, NullTest))
7972 : {
7973 2284 : NullTest *nt = (NullTest *) clause;
7974 :
7975 2284 : if (nt->nulltesttype == IS_NULL)
7976 : {
7977 240 : found_is_null_op = true;
7978 : /* IS NULL is like = for selectivity/skip scan purposes */
7979 240 : eqQualHere = true;
7980 : }
7981 : }
7982 : else
7983 0 : elog(ERROR, "unsupported indexqual type: %d",
7984 : (int) nodeTag(clause));
7985 :
7986 : /* check for equality operator */
7987 566920 : if (OidIsValid(clause_op))
7988 : {
7989 564636 : op_strategy = get_op_opfamily_strategy(clause_op,
7990 564636 : index->opfamily[indexcol]);
7991 : Assert(op_strategy != 0); /* not a member of opfamily?? */
7992 564636 : if (op_strategy == BTEqualStrategyNumber)
7993 532338 : eqQualHere = true;
7994 : }
7995 :
7996 566920 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
7997 :
7998 : /*
7999 : * We apply inequality selectivities to estimate index descent
8000 : * costs with scans that use skip arrays. Save this indexcol's
8001 : * RestrictInfos if it looks like they'll be needed for that.
8002 : */
8003 566920 : if (!eqQualHere && !found_row_compare &&
8004 33244 : indexcol < index->nkeycolumns - 1)
8005 5704 : indexSkipQuals = lappend(indexSkipQuals, rinfo);
8006 : }
8007 : }
8008 :
8009 : /*
8010 : * If index is unique and we found an '=' clause for each column, we can
8011 : * just assume numIndexTuples = 1 and skip the expensive
8012 : * clauselist_selectivity calculations. However, an array or NullTest
8013 : * always invalidates that theory (even when eqQualHere has been set).
8014 : */
8015 802956 : if (index->unique &&
8016 655270 : indexcol == index->nkeycolumns - 1 &&
8017 253298 : eqQualHere &&
8018 253298 : !found_array &&
8019 247024 : !found_is_null_op)
8020 246976 : numIndexTuples = 1.0;
8021 : else
8022 : {
8023 : List *selectivityQuals;
8024 : Selectivity btreeSelectivity;
8025 :
8026 : /*
8027 : * If the index is partial, AND the index predicate with the
8028 : * index-bound quals to produce a more accurate idea of the number of
8029 : * rows covered by the bound conditions.
8030 : */
8031 555980 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
8032 :
8033 555980 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
8034 555980 : index->rel->relid,
8035 : JOIN_INNER,
8036 : NULL);
8037 555980 : numIndexTuples = btreeSelectivity * index->rel->tuples;
8038 :
8039 : /*
8040 : * btree automatically combines individual array element primitive
8041 : * index scans whenever the tuples covered by the next set of array
8042 : * keys are close to tuples covered by the current set. That puts a
8043 : * natural ceiling on the worst case number of descents -- there
8044 : * cannot possibly be more than one descent per leaf page scanned.
8045 : *
8046 : * Clamp the number of descents to at most 1/3 the number of index
8047 : * pages. This avoids implausibly high estimates with low selectivity
8048 : * paths, where scans usually require only one or two descents. This
8049 : * is most likely to help when there are several SAOP clauses, where
8050 : * naively accepting the total number of distinct combinations of
8051 : * array elements as the number of descents would frequently lead to
8052 : * wild overestimates.
8053 : *
8054 : * We somewhat arbitrarily don't just make the cutoff the total number
8055 : * of leaf pages (we make it 1/3 the total number of pages instead) to
8056 : * give the btree code credit for its ability to continue on the leaf
8057 : * level with low selectivity scans.
8058 : *
8059 : * Note: num_sa_scans includes both ScalarArrayOp array elements and
8060 : * skip array elements whose qual affects our numIndexTuples estimate.
8061 : */
8062 555980 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
8063 555980 : num_sa_scans = Max(num_sa_scans, 1);
8064 :
8065 : /*
8066 : * As in genericcostestimate(), we have to adjust for any array quals
8067 : * included in indexBoundQuals, and then round to integer.
8068 : *
8069 : * It is tempting to make genericcostestimate behave as if array
8070 : * clauses work in almost the same way as scalar operators during
8071 : * btree scans, making the top-level scan look like a continuous scan
8072 : * (as opposed to num_sa_scans-many primitive index scans). After
8073 : * all, btree scans mostly work like that at runtime. However, such a
8074 : * scheme would badly bias genericcostestimate's simplistic approach
8075 : * to calculating numIndexPages through prorating.
8076 : *
8077 : * Stick with the approach taken by non-native SAOP scans for now.
8078 : * genericcostestimate will use the Mackert-Lohman formula to
8079 : * compensate for repeat page fetches, even though that definitely
8080 : * won't happen during btree scans (not for leaf pages, at least).
8081 : * We're usually very pessimistic about the number of primitive index
8082 : * scans that will be required, but it's not clear how to do better.
8083 : */
8084 555980 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
8085 : }
8086 :
8087 : /*
8088 : * Now do generic index cost estimation.
8089 : */
8090 802956 : costs.numIndexTuples = numIndexTuples;
8091 802956 : costs.num_sa_scans = num_sa_scans;
8092 :
8093 802956 : genericcostestimate(root, path, loop_count, &costs);
8094 :
8095 : /*
8096 : * Add a CPU-cost component to represent the costs of initial btree
8097 : * descent. We don't charge any I/O cost for touching upper btree levels,
8098 : * since they tend to stay in cache, but we still have to do about log2(N)
8099 : * comparisons to descend a btree of N leaf tuples. We charge one
8100 : * cpu_operator_cost per comparison.
8101 : *
8102 : * If there are SAOP or skip array keys, charge this once per estimated
8103 : * index descent. The ones after the first one are not startup cost so
8104 : * far as the overall plan goes, so just add them to "total" cost.
8105 : */
8106 802956 : if (index->tuples > 1) /* avoid computing log(0) */
8107 : {
8108 747118 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
8109 747118 : costs.indexStartupCost += descentCost;
8110 747118 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8111 : }
8112 :
8113 : /*
8114 : * Even though we're not charging I/O cost for touching upper btree pages,
8115 : * it's still reasonable to charge some CPU cost per page descended
8116 : * through. Moreover, if we had no such charge at all, bloated indexes
8117 : * would appear to have the same search cost as unbloated ones, at least
8118 : * in cases where only a single leaf page is expected to be visited. This
8119 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
8120 : * touched. The number of such pages is btree tree height plus one (ie,
8121 : * we charge for the leaf page too). As above, charge once per estimated
8122 : * SAOP/skip array descent.
8123 : */
8124 802956 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8125 802956 : costs.indexStartupCost += descentCost;
8126 802956 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8127 :
8128 802956 : if (!have_correlation)
8129 : {
8130 766264 : examine_indexcol_variable(root, index, 0, &vardata);
8131 766264 : if (HeapTupleIsValid(vardata.statsTuple))
8132 576808 : costs.indexCorrelation = btcost_correlation(index, &vardata);
8133 766264 : ReleaseVariableStats(vardata);
8134 : }
8135 : else
8136 : {
8137 : /* btcost_correlation already called earlier on */
8138 36692 : costs.indexCorrelation = correlation;
8139 : }
8140 :
8141 802956 : *indexStartupCost = costs.indexStartupCost;
8142 802956 : *indexTotalCost = costs.indexTotalCost;
8143 802956 : *indexSelectivity = costs.indexSelectivity;
8144 802956 : *indexCorrelation = costs.indexCorrelation;
8145 802956 : *indexPages = costs.numIndexPages;
8146 802956 : }
8147 :
8148 : void
8149 430 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8150 : Cost *indexStartupCost, Cost *indexTotalCost,
8151 : Selectivity *indexSelectivity, double *indexCorrelation,
8152 : double *indexPages)
8153 : {
8154 430 : GenericCosts costs = {0};
8155 :
8156 430 : genericcostestimate(root, path, loop_count, &costs);
8157 :
8158 : /*
8159 : * A hash index has no descent costs as such, since the index AM can go
8160 : * directly to the target bucket after computing the hash value. There
8161 : * are a couple of other hash-specific costs that we could conceivably add
8162 : * here, though:
8163 : *
8164 : * Ideally we'd charge spc_random_page_cost for each page in the target
8165 : * bucket, not just the numIndexPages pages that genericcostestimate
8166 : * thought we'd visit. However in most cases we don't know which bucket
8167 : * that will be. There's no point in considering the average bucket size
8168 : * because the hash AM makes sure that's always one page.
8169 : *
8170 : * Likewise, we could consider charging some CPU for each index tuple in
8171 : * the bucket, if we knew how many there were. But the per-tuple cost is
8172 : * just a hash value comparison, not a general datatype-dependent
8173 : * comparison, so any such charge ought to be quite a bit less than
8174 : * cpu_operator_cost; which makes it probably not worth worrying about.
8175 : *
8176 : * A bigger issue is that chance hash-value collisions will result in
8177 : * wasted probes into the heap. We don't currently attempt to model this
8178 : * cost on the grounds that it's rare, but maybe it's not rare enough.
8179 : * (Any fix for this ought to consider the generic lossy-operator problem,
8180 : * though; it's not entirely hash-specific.)
8181 : */
8182 :
8183 430 : *indexStartupCost = costs.indexStartupCost;
8184 430 : *indexTotalCost = costs.indexTotalCost;
8185 430 : *indexSelectivity = costs.indexSelectivity;
8186 430 : *indexCorrelation = costs.indexCorrelation;
8187 430 : *indexPages = costs.numIndexPages;
8188 430 : }
8189 :
8190 : void
8191 4878 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8192 : Cost *indexStartupCost, Cost *indexTotalCost,
8193 : Selectivity *indexSelectivity, double *indexCorrelation,
8194 : double *indexPages)
8195 : {
8196 4878 : IndexOptInfo *index = path->indexinfo;
8197 4878 : GenericCosts costs = {0};
8198 : Cost descentCost;
8199 :
8200 4878 : genericcostestimate(root, path, loop_count, &costs);
8201 :
8202 : /*
8203 : * We model index descent costs similarly to those for btree, but to do
8204 : * that we first need an idea of the tree height. We somewhat arbitrarily
8205 : * assume that the fanout is 100, meaning the tree height is at most
8206 : * log100(index->pages).
8207 : *
8208 : * Although this computation isn't really expensive enough to require
8209 : * caching, we might as well use index->tree_height to cache it.
8210 : */
8211 4878 : if (index->tree_height < 0) /* unknown? */
8212 : {
8213 4864 : if (index->pages > 1) /* avoid computing log(0) */
8214 2720 : index->tree_height = (int) (log(index->pages) / log(100.0));
8215 : else
8216 2144 : index->tree_height = 0;
8217 : }
8218 :
8219 : /*
8220 : * Add a CPU-cost component to represent the costs of initial descent. We
8221 : * just use log(N) here not log2(N) since the branching factor isn't
8222 : * necessarily two anyway. As for btree, charge once per SA scan.
8223 : */
8224 4878 : if (index->tuples > 1) /* avoid computing log(0) */
8225 : {
8226 4878 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8227 4878 : costs.indexStartupCost += descentCost;
8228 4878 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8229 : }
8230 :
8231 : /*
8232 : * Likewise add a per-page charge, calculated the same as for btrees.
8233 : */
8234 4878 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8235 4878 : costs.indexStartupCost += descentCost;
8236 4878 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8237 :
8238 4878 : *indexStartupCost = costs.indexStartupCost;
8239 4878 : *indexTotalCost = costs.indexTotalCost;
8240 4878 : *indexSelectivity = costs.indexSelectivity;
8241 4878 : *indexCorrelation = costs.indexCorrelation;
8242 4878 : *indexPages = costs.numIndexPages;
8243 4878 : }
8244 :
8245 : void
8246 1784 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8247 : Cost *indexStartupCost, Cost *indexTotalCost,
8248 : Selectivity *indexSelectivity, double *indexCorrelation,
8249 : double *indexPages)
8250 : {
8251 1784 : IndexOptInfo *index = path->indexinfo;
8252 1784 : GenericCosts costs = {0};
8253 : Cost descentCost;
8254 :
8255 1784 : genericcostestimate(root, path, loop_count, &costs);
8256 :
8257 : /*
8258 : * We model index descent costs similarly to those for btree, but to do
8259 : * that we first need an idea of the tree height. We somewhat arbitrarily
8260 : * assume that the fanout is 100, meaning the tree height is at most
8261 : * log100(index->pages).
8262 : *
8263 : * Although this computation isn't really expensive enough to require
8264 : * caching, we might as well use index->tree_height to cache it.
8265 : */
8266 1784 : if (index->tree_height < 0) /* unknown? */
8267 : {
8268 1778 : if (index->pages > 1) /* avoid computing log(0) */
8269 1778 : index->tree_height = (int) (log(index->pages) / log(100.0));
8270 : else
8271 0 : index->tree_height = 0;
8272 : }
8273 :
8274 : /*
8275 : * Add a CPU-cost component to represent the costs of initial descent. We
8276 : * just use log(N) here not log2(N) since the branching factor isn't
8277 : * necessarily two anyway. As for btree, charge once per SA scan.
8278 : */
8279 1784 : if (index->tuples > 1) /* avoid computing log(0) */
8280 : {
8281 1784 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8282 1784 : costs.indexStartupCost += descentCost;
8283 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8284 : }
8285 :
8286 : /*
8287 : * Likewise add a per-page charge, calculated the same as for btrees.
8288 : */
8289 1784 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8290 1784 : costs.indexStartupCost += descentCost;
8291 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8292 :
8293 1784 : *indexStartupCost = costs.indexStartupCost;
8294 1784 : *indexTotalCost = costs.indexTotalCost;
8295 1784 : *indexSelectivity = costs.indexSelectivity;
8296 1784 : *indexCorrelation = costs.indexCorrelation;
8297 1784 : *indexPages = costs.numIndexPages;
8298 1784 : }
8299 :
8300 :
8301 : /*
8302 : * Support routines for gincostestimate
8303 : */
8304 :
8305 : typedef struct
8306 : {
8307 : bool attHasFullScan[INDEX_MAX_KEYS];
8308 : bool attHasNormalScan[INDEX_MAX_KEYS];
8309 : double partialEntries;
8310 : double exactEntries;
8311 : double searchEntries;
8312 : double arrayScans;
8313 : } GinQualCounts;
8314 :
8315 : /*
8316 : * Estimate the number of index terms that need to be searched for while
8317 : * testing the given GIN query, and increment the counts in *counts
8318 : * appropriately. If the query is unsatisfiable, return false.
8319 : */
8320 : static bool
8321 2480 : gincost_pattern(IndexOptInfo *index, int indexcol,
8322 : Oid clause_op, Datum query,
8323 : GinQualCounts *counts)
8324 : {
8325 : FmgrInfo flinfo;
8326 : Oid extractProcOid;
8327 : Oid collation;
8328 : int strategy_op;
8329 : Oid lefttype,
8330 : righttype;
8331 2480 : int32 nentries = 0;
8332 2480 : bool *partial_matches = NULL;
8333 2480 : Pointer *extra_data = NULL;
8334 2480 : bool *nullFlags = NULL;
8335 2480 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
8336 : int32 i;
8337 :
8338 : Assert(indexcol < index->nkeycolumns);
8339 :
8340 : /*
8341 : * Get the operator's strategy number and declared input data types within
8342 : * the index opfamily. (We don't need the latter, but we use
8343 : * get_op_opfamily_properties because it will throw error if it fails to
8344 : * find a matching pg_amop entry.)
8345 : */
8346 2480 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
8347 : &strategy_op, &lefttype, &righttype);
8348 :
8349 : /*
8350 : * GIN always uses the "default" support functions, which are those with
8351 : * lefttype == righttype == the opclass' opcintype (see
8352 : * IndexSupportInitialize in relcache.c).
8353 : */
8354 2480 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
8355 2480 : index->opcintype[indexcol],
8356 2480 : index->opcintype[indexcol],
8357 : GIN_EXTRACTQUERY_PROC);
8358 :
8359 2480 : if (!OidIsValid(extractProcOid))
8360 : {
8361 : /* should not happen; throw same error as index_getprocinfo */
8362 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
8363 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
8364 : get_rel_name(index->indexoid));
8365 : }
8366 :
8367 : /*
8368 : * Choose collation to pass to extractProc (should match initGinState).
8369 : */
8370 2480 : if (OidIsValid(index->indexcollations[indexcol]))
8371 414 : collation = index->indexcollations[indexcol];
8372 : else
8373 2066 : collation = DEFAULT_COLLATION_OID;
8374 :
8375 2480 : fmgr_info(extractProcOid, &flinfo);
8376 :
8377 2480 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
8378 :
8379 2480 : FunctionCall7Coll(&flinfo,
8380 : collation,
8381 : query,
8382 : PointerGetDatum(&nentries),
8383 : UInt16GetDatum(strategy_op),
8384 : PointerGetDatum(&partial_matches),
8385 : PointerGetDatum(&extra_data),
8386 : PointerGetDatum(&nullFlags),
8387 : PointerGetDatum(&searchMode));
8388 :
8389 2480 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
8390 : {
8391 : /* No match is possible */
8392 12 : return false;
8393 : }
8394 :
8395 9676 : for (i = 0; i < nentries; i++)
8396 : {
8397 : /*
8398 : * For partial match we haven't any information to estimate number of
8399 : * matched entries in index, so, we just estimate it as 100
8400 : */
8401 7208 : if (partial_matches && partial_matches[i])
8402 694 : counts->partialEntries += 100;
8403 : else
8404 6514 : counts->exactEntries++;
8405 :
8406 7208 : counts->searchEntries++;
8407 : }
8408 :
8409 2468 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
8410 : {
8411 1984 : counts->attHasNormalScan[indexcol] = true;
8412 : }
8413 484 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
8414 : {
8415 : /* Treat "include empty" like an exact-match item */
8416 44 : counts->attHasNormalScan[indexcol] = true;
8417 44 : counts->exactEntries++;
8418 44 : counts->searchEntries++;
8419 : }
8420 : else
8421 : {
8422 : /* It's GIN_SEARCH_MODE_ALL */
8423 440 : counts->attHasFullScan[indexcol] = true;
8424 : }
8425 :
8426 2468 : return true;
8427 : }
8428 :
8429 : /*
8430 : * Estimate the number of index terms that need to be searched for while
8431 : * testing the given GIN index clause, and increment the counts in *counts
8432 : * appropriately. If the query is unsatisfiable, return false.
8433 : */
8434 : static bool
8435 2468 : gincost_opexpr(PlannerInfo *root,
8436 : IndexOptInfo *index,
8437 : int indexcol,
8438 : OpExpr *clause,
8439 : GinQualCounts *counts)
8440 : {
8441 2468 : Oid clause_op = clause->opno;
8442 2468 : Node *operand = (Node *) lsecond(clause->args);
8443 :
8444 : /* aggressively reduce to a constant, and look through relabeling */
8445 2468 : operand = estimate_expression_value(root, operand);
8446 :
8447 2468 : if (IsA(operand, RelabelType))
8448 0 : operand = (Node *) ((RelabelType *) operand)->arg;
8449 :
8450 : /*
8451 : * It's impossible to call extractQuery method for unknown operand. So
8452 : * unless operand is a Const we can't do much; just assume there will be
8453 : * one ordinary search entry from the operand at runtime.
8454 : */
8455 2468 : if (!IsA(operand, Const))
8456 : {
8457 0 : counts->exactEntries++;
8458 0 : counts->searchEntries++;
8459 0 : return true;
8460 : }
8461 :
8462 : /* If Const is null, there can be no matches */
8463 2468 : if (((Const *) operand)->constisnull)
8464 0 : return false;
8465 :
8466 : /* Otherwise, apply extractQuery and get the actual term counts */
8467 2468 : return gincost_pattern(index, indexcol, clause_op,
8468 : ((Const *) operand)->constvalue,
8469 : counts);
8470 : }
8471 :
8472 : /*
8473 : * Estimate the number of index terms that need to be searched for while
8474 : * testing the given GIN index clause, and increment the counts in *counts
8475 : * appropriately. If the query is unsatisfiable, return false.
8476 : *
8477 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
8478 : * each of which involves one value from the RHS array, plus all the
8479 : * non-array quals (if any). To model this, we average the counts across
8480 : * the RHS elements, and add the averages to the counts in *counts (which
8481 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
8482 : * by N, causing gincostestimate to scale up its estimates accordingly.
8483 : */
8484 : static bool
8485 6 : gincost_scalararrayopexpr(PlannerInfo *root,
8486 : IndexOptInfo *index,
8487 : int indexcol,
8488 : ScalarArrayOpExpr *clause,
8489 : double numIndexEntries,
8490 : GinQualCounts *counts)
8491 : {
8492 6 : Oid clause_op = clause->opno;
8493 6 : Node *rightop = (Node *) lsecond(clause->args);
8494 : ArrayType *arrayval;
8495 : int16 elmlen;
8496 : bool elmbyval;
8497 : char elmalign;
8498 : int numElems;
8499 : Datum *elemValues;
8500 : bool *elemNulls;
8501 : GinQualCounts arraycounts;
8502 6 : int numPossible = 0;
8503 : int i;
8504 :
8505 : Assert(clause->useOr);
8506 :
8507 : /* aggressively reduce to a constant, and look through relabeling */
8508 6 : rightop = estimate_expression_value(root, rightop);
8509 :
8510 6 : if (IsA(rightop, RelabelType))
8511 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
8512 :
8513 : /*
8514 : * It's impossible to call extractQuery method for unknown operand. So
8515 : * unless operand is a Const we can't do much; just assume there will be
8516 : * one ordinary search entry from each array entry at runtime, and fall
8517 : * back on a probably-bad estimate of the number of array entries.
8518 : */
8519 6 : if (!IsA(rightop, Const))
8520 : {
8521 0 : counts->exactEntries++;
8522 0 : counts->searchEntries++;
8523 0 : counts->arrayScans *= estimate_array_length(root, rightop);
8524 0 : return true;
8525 : }
8526 :
8527 : /* If Const is null, there can be no matches */
8528 6 : if (((Const *) rightop)->constisnull)
8529 0 : return false;
8530 :
8531 : /* Otherwise, extract the array elements and iterate over them */
8532 6 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
8533 6 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
8534 : &elmlen, &elmbyval, &elmalign);
8535 6 : deconstruct_array(arrayval,
8536 : ARR_ELEMTYPE(arrayval),
8537 : elmlen, elmbyval, elmalign,
8538 : &elemValues, &elemNulls, &numElems);
8539 :
8540 6 : memset(&arraycounts, 0, sizeof(arraycounts));
8541 :
8542 18 : for (i = 0; i < numElems; i++)
8543 : {
8544 : GinQualCounts elemcounts;
8545 :
8546 : /* NULL can't match anything, so ignore, as the executor will */
8547 12 : if (elemNulls[i])
8548 0 : continue;
8549 :
8550 : /* Otherwise, apply extractQuery and get the actual term counts */
8551 12 : memset(&elemcounts, 0, sizeof(elemcounts));
8552 :
8553 12 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
8554 : &elemcounts))
8555 : {
8556 : /* We ignore array elements that are unsatisfiable patterns */
8557 12 : numPossible++;
8558 :
8559 12 : if (elemcounts.attHasFullScan[indexcol] &&
8560 0 : !elemcounts.attHasNormalScan[indexcol])
8561 : {
8562 : /*
8563 : * Full index scan will be required. We treat this as if
8564 : * every key in the index had been listed in the query; is
8565 : * that reasonable?
8566 : */
8567 0 : elemcounts.partialEntries = 0;
8568 0 : elemcounts.exactEntries = numIndexEntries;
8569 0 : elemcounts.searchEntries = numIndexEntries;
8570 : }
8571 12 : arraycounts.partialEntries += elemcounts.partialEntries;
8572 12 : arraycounts.exactEntries += elemcounts.exactEntries;
8573 12 : arraycounts.searchEntries += elemcounts.searchEntries;
8574 : }
8575 : }
8576 :
8577 6 : if (numPossible == 0)
8578 : {
8579 : /* No satisfiable patterns in the array */
8580 0 : return false;
8581 : }
8582 :
8583 : /*
8584 : * Now add the averages to the global counts. This will give us an
8585 : * estimate of the average number of terms searched for in each indexscan,
8586 : * including contributions from both array and non-array quals.
8587 : */
8588 6 : counts->partialEntries += arraycounts.partialEntries / numPossible;
8589 6 : counts->exactEntries += arraycounts.exactEntries / numPossible;
8590 6 : counts->searchEntries += arraycounts.searchEntries / numPossible;
8591 :
8592 6 : counts->arrayScans *= numPossible;
8593 :
8594 6 : return true;
8595 : }
8596 :
8597 : /*
8598 : * GIN has search behavior completely different from other index types
8599 : */
8600 : void
8601 2264 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8602 : Cost *indexStartupCost, Cost *indexTotalCost,
8603 : Selectivity *indexSelectivity, double *indexCorrelation,
8604 : double *indexPages)
8605 : {
8606 2264 : IndexOptInfo *index = path->indexinfo;
8607 2264 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8608 : List *selectivityQuals;
8609 2264 : double numPages = index->pages,
8610 2264 : numTuples = index->tuples;
8611 : double numEntryPages,
8612 : numDataPages,
8613 : numPendingPages,
8614 : numEntries;
8615 : GinQualCounts counts;
8616 : bool matchPossible;
8617 : bool fullIndexScan;
8618 : double partialScale;
8619 : double entryPagesFetched,
8620 : dataPagesFetched,
8621 : dataPagesFetchedBySel;
8622 : double qual_op_cost,
8623 : qual_arg_cost,
8624 : spc_random_page_cost,
8625 : outer_scans;
8626 : Cost descentCost;
8627 : Relation indexRel;
8628 : GinStatsData ginStats;
8629 : ListCell *lc;
8630 : int i;
8631 :
8632 : /*
8633 : * Obtain statistical information from the meta page, if possible. Else
8634 : * set ginStats to zeroes, and we'll cope below.
8635 : */
8636 2264 : if (!index->hypothetical)
8637 : {
8638 : /* Lock should have already been obtained in plancat.c */
8639 2264 : indexRel = index_open(index->indexoid, NoLock);
8640 2264 : ginGetStats(indexRel, &ginStats);
8641 2264 : index_close(indexRel, NoLock);
8642 : }
8643 : else
8644 : {
8645 0 : memset(&ginStats, 0, sizeof(ginStats));
8646 : }
8647 :
8648 : /*
8649 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
8650 : * trusted, but the other fields are data as of the last VACUUM. We can
8651 : * scale them up to account for growth since then, but that method only
8652 : * goes so far; in the worst case, the stats might be for a completely
8653 : * empty index, and scaling them will produce pretty bogus numbers.
8654 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
8655 : * it's grown more than that, fall back to estimating things only from the
8656 : * assumed-accurate index size. But we'll trust nPendingPages in any case
8657 : * so long as it's not clearly insane, ie, more than the index size.
8658 : */
8659 2264 : if (ginStats.nPendingPages < numPages)
8660 2264 : numPendingPages = ginStats.nPendingPages;
8661 : else
8662 0 : numPendingPages = 0;
8663 :
8664 2264 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
8665 2264 : ginStats.nTotalPages > numPages / 4 &&
8666 2212 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
8667 1948 : {
8668 : /*
8669 : * OK, the stats seem close enough to sane to be trusted. But we
8670 : * still need to scale them by the ratio numPages / nTotalPages to
8671 : * account for growth since the last VACUUM.
8672 : */
8673 1948 : double scale = numPages / ginStats.nTotalPages;
8674 :
8675 1948 : numEntryPages = ceil(ginStats.nEntryPages * scale);
8676 1948 : numDataPages = ceil(ginStats.nDataPages * scale);
8677 1948 : numEntries = ceil(ginStats.nEntries * scale);
8678 : /* ensure we didn't round up too much */
8679 1948 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
8680 1948 : numDataPages = Min(numDataPages,
8681 : numPages - numPendingPages - numEntryPages);
8682 : }
8683 : else
8684 : {
8685 : /*
8686 : * We might get here because it's a hypothetical index, or an index
8687 : * created pre-9.1 and never vacuumed since upgrading (in which case
8688 : * its stats would read as zeroes), or just because it's grown too
8689 : * much since the last VACUUM for us to put our faith in scaling.
8690 : *
8691 : * Invent some plausible internal statistics based on the index page
8692 : * count (and clamp that to at least 10 pages, just in case). We
8693 : * estimate that 90% of the index is entry pages, and the rest is data
8694 : * pages. Estimate 100 entries per entry page; this is rather bogus
8695 : * since it'll depend on the size of the keys, but it's more robust
8696 : * than trying to predict the number of entries per heap tuple.
8697 : */
8698 316 : numPages = Max(numPages, 10);
8699 316 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
8700 316 : numDataPages = numPages - numPendingPages - numEntryPages;
8701 316 : numEntries = floor(numEntryPages * 100);
8702 : }
8703 :
8704 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
8705 2264 : if (numEntries < 1)
8706 0 : numEntries = 1;
8707 :
8708 : /*
8709 : * If the index is partial, AND the index predicate with the index-bound
8710 : * quals to produce a more accurate idea of the number of rows covered by
8711 : * the bound conditions.
8712 : */
8713 2264 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
8714 :
8715 : /* Estimate the fraction of main-table tuples that will be visited */
8716 4528 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
8717 2264 : index->rel->relid,
8718 : JOIN_INNER,
8719 : NULL);
8720 :
8721 : /* fetch estimated page cost for tablespace containing index */
8722 2264 : get_tablespace_page_costs(index->reltablespace,
8723 : &spc_random_page_cost,
8724 : NULL);
8725 :
8726 : /*
8727 : * Generic assumption about index correlation: there isn't any.
8728 : */
8729 2264 : *indexCorrelation = 0.0;
8730 :
8731 : /*
8732 : * Examine quals to estimate number of search entries & partial matches
8733 : */
8734 2264 : memset(&counts, 0, sizeof(counts));
8735 2264 : counts.arrayScans = 1;
8736 2264 : matchPossible = true;
8737 :
8738 4738 : foreach(lc, path->indexclauses)
8739 : {
8740 2474 : IndexClause *iclause = lfirst_node(IndexClause, lc);
8741 : ListCell *lc2;
8742 :
8743 4936 : foreach(lc2, iclause->indexquals)
8744 : {
8745 2474 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8746 2474 : Expr *clause = rinfo->clause;
8747 :
8748 2474 : if (IsA(clause, OpExpr))
8749 : {
8750 2468 : matchPossible = gincost_opexpr(root,
8751 : index,
8752 2468 : iclause->indexcol,
8753 : (OpExpr *) clause,
8754 : &counts);
8755 2468 : if (!matchPossible)
8756 12 : break;
8757 : }
8758 6 : else if (IsA(clause, ScalarArrayOpExpr))
8759 : {
8760 6 : matchPossible = gincost_scalararrayopexpr(root,
8761 : index,
8762 6 : iclause->indexcol,
8763 : (ScalarArrayOpExpr *) clause,
8764 : numEntries,
8765 : &counts);
8766 6 : if (!matchPossible)
8767 0 : break;
8768 : }
8769 : else
8770 : {
8771 : /* shouldn't be anything else for a GIN index */
8772 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
8773 : (int) nodeTag(clause));
8774 : }
8775 : }
8776 : }
8777 :
8778 : /* Fall out if there were any provably-unsatisfiable quals */
8779 2264 : if (!matchPossible)
8780 : {
8781 12 : *indexStartupCost = 0;
8782 12 : *indexTotalCost = 0;
8783 12 : *indexSelectivity = 0;
8784 12 : return;
8785 : }
8786 :
8787 : /*
8788 : * If attribute has a full scan and at the same time doesn't have normal
8789 : * scan, then we'll have to scan all non-null entries of that attribute.
8790 : * Currently, we don't have per-attribute statistics for GIN. Thus, we
8791 : * must assume the whole GIN index has to be scanned in this case.
8792 : */
8793 2252 : fullIndexScan = false;
8794 4394 : for (i = 0; i < index->nkeycolumns; i++)
8795 : {
8796 2480 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
8797 : {
8798 338 : fullIndexScan = true;
8799 338 : break;
8800 : }
8801 : }
8802 :
8803 2252 : if (fullIndexScan || indexQuals == NIL)
8804 : {
8805 : /*
8806 : * Full index scan will be required. We treat this as if every key in
8807 : * the index had been listed in the query; is that reasonable?
8808 : */
8809 338 : counts.partialEntries = 0;
8810 338 : counts.exactEntries = numEntries;
8811 338 : counts.searchEntries = numEntries;
8812 : }
8813 :
8814 : /* Will we have more than one iteration of a nestloop scan? */
8815 2252 : outer_scans = loop_count;
8816 :
8817 : /*
8818 : * Compute cost to begin scan, first of all, pay attention to pending
8819 : * list.
8820 : */
8821 2252 : entryPagesFetched = numPendingPages;
8822 :
8823 : /*
8824 : * Estimate number of entry pages read. We need to do
8825 : * counts.searchEntries searches. Use a power function as it should be,
8826 : * but tuples on leaf pages usually is much greater. Here we include all
8827 : * searches in entry tree, including search of first entry in partial
8828 : * match algorithm
8829 : */
8830 2252 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8831 :
8832 : /*
8833 : * Add an estimate of entry pages read by partial match algorithm. It's a
8834 : * scan over leaf pages in entry tree. We haven't any useful stats here,
8835 : * so estimate it as proportion. Because counts.partialEntries is really
8836 : * pretty bogus (see code above), it's possible that it is more than
8837 : * numEntries; clamp the proportion to ensure sanity.
8838 : */
8839 2252 : partialScale = counts.partialEntries / numEntries;
8840 2252 : partialScale = Min(partialScale, 1.0);
8841 :
8842 2252 : entryPagesFetched += ceil(numEntryPages * partialScale);
8843 :
8844 : /*
8845 : * Partial match algorithm reads all data pages before doing actual scan,
8846 : * so it's a startup cost. Again, we haven't any useful stats here, so
8847 : * estimate it as proportion.
8848 : */
8849 2252 : dataPagesFetched = ceil(numDataPages * partialScale);
8850 :
8851 2252 : *indexStartupCost = 0;
8852 2252 : *indexTotalCost = 0;
8853 :
8854 : /*
8855 : * Add a CPU-cost component to represent the costs of initial entry btree
8856 : * descent. We don't charge any I/O cost for touching upper btree levels,
8857 : * since they tend to stay in cache, but we still have to do about log2(N)
8858 : * comparisons to descend a btree of N leaf tuples. We charge one
8859 : * cpu_operator_cost per comparison.
8860 : *
8861 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
8862 : * ones after the first one are not startup cost so far as the overall
8863 : * plan is concerned, so add them only to "total" cost.
8864 : */
8865 2252 : if (numEntries > 1) /* avoid computing log(0) */
8866 : {
8867 2252 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
8868 2252 : *indexStartupCost += descentCost * counts.searchEntries;
8869 2252 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
8870 : }
8871 :
8872 : /*
8873 : * Add a cpu cost per entry-page fetched. This is not amortized over a
8874 : * loop.
8875 : */
8876 2252 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8877 2252 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8878 :
8879 : /*
8880 : * Add a cpu cost per data-page fetched. This is also not amortized over a
8881 : * loop. Since those are the data pages from the partial match algorithm,
8882 : * charge them as startup cost.
8883 : */
8884 2252 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
8885 :
8886 : /*
8887 : * Since we add the startup cost to the total cost later on, remove the
8888 : * initial arrayscan from the total.
8889 : */
8890 2252 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8891 :
8892 : /*
8893 : * Calculate cache effects if more than one scan due to nestloops or array
8894 : * quals. The result is pro-rated per nestloop scan, but the array qual
8895 : * factor shouldn't be pro-rated (compare genericcostestimate).
8896 : */
8897 2252 : if (outer_scans > 1 || counts.arrayScans > 1)
8898 : {
8899 6 : entryPagesFetched *= outer_scans * counts.arrayScans;
8900 6 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
8901 : (BlockNumber) numEntryPages,
8902 : numEntryPages, root);
8903 6 : entryPagesFetched /= outer_scans;
8904 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8905 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8906 : (BlockNumber) numDataPages,
8907 : numDataPages, root);
8908 6 : dataPagesFetched /= outer_scans;
8909 : }
8910 :
8911 : /*
8912 : * Here we use random page cost because logically-close pages could be far
8913 : * apart on disk.
8914 : */
8915 2252 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
8916 :
8917 : /*
8918 : * Now compute the number of data pages fetched during the scan.
8919 : *
8920 : * We assume every entry to have the same number of items, and that there
8921 : * is no overlap between them. (XXX: tsvector and array opclasses collect
8922 : * statistics on the frequency of individual keys; it would be nice to use
8923 : * those here.)
8924 : */
8925 2252 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
8926 :
8927 : /*
8928 : * If there is a lot of overlap among the entries, in particular if one of
8929 : * the entries is very frequent, the above calculation can grossly
8930 : * under-estimate. As a simple cross-check, calculate a lower bound based
8931 : * on the overall selectivity of the quals. At a minimum, we must read
8932 : * one item pointer for each matching entry.
8933 : *
8934 : * The width of each item pointer varies, based on the level of
8935 : * compression. We don't have statistics on that, but an average of
8936 : * around 3 bytes per item is fairly typical.
8937 : */
8938 2252 : dataPagesFetchedBySel = ceil(*indexSelectivity *
8939 2252 : (numTuples / (BLCKSZ / 3)));
8940 2252 : if (dataPagesFetchedBySel > dataPagesFetched)
8941 1866 : dataPagesFetched = dataPagesFetchedBySel;
8942 :
8943 : /* Add one page cpu-cost to the startup cost */
8944 2252 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
8945 :
8946 : /*
8947 : * Add once again a CPU-cost for those data pages, before amortizing for
8948 : * cache.
8949 : */
8950 2252 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8951 :
8952 : /* Account for cache effects, the same as above */
8953 2252 : if (outer_scans > 1 || counts.arrayScans > 1)
8954 : {
8955 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8956 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8957 : (BlockNumber) numDataPages,
8958 : numDataPages, root);
8959 6 : dataPagesFetched /= outer_scans;
8960 : }
8961 :
8962 : /* And apply random_page_cost as the cost per page */
8963 2252 : *indexTotalCost += *indexStartupCost +
8964 2252 : dataPagesFetched * spc_random_page_cost;
8965 :
8966 : /*
8967 : * Add on index qual eval costs, much as in genericcostestimate. We charge
8968 : * cpu but we can disregard indexorderbys, since GIN doesn't support
8969 : * those.
8970 : */
8971 2252 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8972 2252 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
8973 :
8974 2252 : *indexStartupCost += qual_arg_cost;
8975 2252 : *indexTotalCost += qual_arg_cost;
8976 :
8977 : /*
8978 : * Add a cpu cost per search entry, corresponding to the actual visited
8979 : * entries.
8980 : */
8981 2252 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
8982 : /* Now add a cpu cost per tuple in the posting lists / trees */
8983 2252 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
8984 2252 : *indexPages = dataPagesFetched;
8985 : }
8986 :
8987 : /*
8988 : * BRIN has search behavior completely different from other index types
8989 : */
8990 : void
8991 10730 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8992 : Cost *indexStartupCost, Cost *indexTotalCost,
8993 : Selectivity *indexSelectivity, double *indexCorrelation,
8994 : double *indexPages)
8995 : {
8996 10730 : IndexOptInfo *index = path->indexinfo;
8997 10730 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8998 10730 : double numPages = index->pages;
8999 10730 : RelOptInfo *baserel = index->rel;
9000 10730 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
9001 : Cost spc_seq_page_cost;
9002 : Cost spc_random_page_cost;
9003 : double qual_arg_cost;
9004 : double qualSelectivity;
9005 : BrinStatsData statsData;
9006 : double indexRanges;
9007 : double minimalRanges;
9008 : double estimatedRanges;
9009 : double selec;
9010 : Relation indexRel;
9011 : ListCell *l;
9012 : VariableStatData vardata;
9013 :
9014 : Assert(rte->rtekind == RTE_RELATION);
9015 :
9016 : /* fetch estimated page cost for the tablespace containing the index */
9017 10730 : get_tablespace_page_costs(index->reltablespace,
9018 : &spc_random_page_cost,
9019 : &spc_seq_page_cost);
9020 :
9021 : /*
9022 : * Obtain some data from the index itself, if possible. Otherwise invent
9023 : * some plausible internal statistics based on the relation page count.
9024 : */
9025 10730 : if (!index->hypothetical)
9026 : {
9027 : /*
9028 : * A lock should have already been obtained on the index in plancat.c.
9029 : */
9030 10730 : indexRel = index_open(index->indexoid, NoLock);
9031 10730 : brinGetStats(indexRel, &statsData);
9032 10730 : index_close(indexRel, NoLock);
9033 :
9034 : /* work out the actual number of ranges in the index */
9035 10730 : indexRanges = Max(ceil((double) baserel->pages /
9036 : statsData.pagesPerRange), 1.0);
9037 : }
9038 : else
9039 : {
9040 : /*
9041 : * Assume default number of pages per range, and estimate the number
9042 : * of ranges based on that.
9043 : */
9044 0 : indexRanges = Max(ceil((double) baserel->pages /
9045 : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
9046 :
9047 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
9048 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
9049 : }
9050 :
9051 : /*
9052 : * Compute index correlation
9053 : *
9054 : * Because we can use all index quals equally when scanning, we can use
9055 : * the largest correlation (in absolute value) among columns used by the
9056 : * query. Start at zero, the worst possible case. If we cannot find any
9057 : * correlation statistics, we will keep it as 0.
9058 : */
9059 10730 : *indexCorrelation = 0;
9060 :
9061 21462 : foreach(l, path->indexclauses)
9062 : {
9063 10732 : IndexClause *iclause = lfirst_node(IndexClause, l);
9064 10732 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
9065 :
9066 : /* attempt to lookup stats in relation for this index column */
9067 10732 : if (attnum != 0)
9068 : {
9069 : /* Simple variable -- look to stats for the underlying table */
9070 10732 : if (get_relation_stats_hook &&
9071 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
9072 : {
9073 : /*
9074 : * The hook took control of acquiring a stats tuple. If it
9075 : * did supply a tuple, it'd better have supplied a freefunc.
9076 : */
9077 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
9078 0 : elog(ERROR,
9079 : "no function provided to release variable stats with");
9080 : }
9081 : else
9082 : {
9083 10732 : vardata.statsTuple =
9084 10732 : SearchSysCache3(STATRELATTINH,
9085 : ObjectIdGetDatum(rte->relid),
9086 : Int16GetDatum(attnum),
9087 : BoolGetDatum(false));
9088 10732 : vardata.freefunc = ReleaseSysCache;
9089 : }
9090 : }
9091 : else
9092 : {
9093 : /*
9094 : * Looks like we've found an expression column in the index. Let's
9095 : * see if there's any stats for it.
9096 : */
9097 :
9098 : /* get the attnum from the 0-based index. */
9099 0 : attnum = iclause->indexcol + 1;
9100 :
9101 0 : if (get_index_stats_hook &&
9102 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
9103 : {
9104 : /*
9105 : * The hook took control of acquiring a stats tuple. If it
9106 : * did supply a tuple, it'd better have supplied a freefunc.
9107 : */
9108 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
9109 0 : !vardata.freefunc)
9110 0 : elog(ERROR, "no function provided to release variable stats with");
9111 : }
9112 : else
9113 : {
9114 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
9115 : ObjectIdGetDatum(index->indexoid),
9116 : Int16GetDatum(attnum),
9117 : BoolGetDatum(false));
9118 0 : vardata.freefunc = ReleaseSysCache;
9119 : }
9120 : }
9121 :
9122 10732 : if (HeapTupleIsValid(vardata.statsTuple))
9123 : {
9124 : AttStatsSlot sslot;
9125 :
9126 36 : if (get_attstatsslot(&sslot, vardata.statsTuple,
9127 : STATISTIC_KIND_CORRELATION, InvalidOid,
9128 : ATTSTATSSLOT_NUMBERS))
9129 : {
9130 36 : double varCorrelation = 0.0;
9131 :
9132 36 : if (sslot.nnumbers > 0)
9133 36 : varCorrelation = fabs(sslot.numbers[0]);
9134 :
9135 36 : if (varCorrelation > *indexCorrelation)
9136 36 : *indexCorrelation = varCorrelation;
9137 :
9138 36 : free_attstatsslot(&sslot);
9139 : }
9140 : }
9141 :
9142 10732 : ReleaseVariableStats(vardata);
9143 : }
9144 :
9145 10730 : qualSelectivity = clauselist_selectivity(root, indexQuals,
9146 10730 : baserel->relid,
9147 : JOIN_INNER, NULL);
9148 :
9149 : /*
9150 : * Now calculate the minimum possible ranges we could match with if all of
9151 : * the rows were in the perfect order in the table's heap.
9152 : */
9153 10730 : minimalRanges = ceil(indexRanges * qualSelectivity);
9154 :
9155 : /*
9156 : * Now estimate the number of ranges that we'll touch by using the
9157 : * indexCorrelation from the stats. Careful not to divide by zero (note
9158 : * we're using the absolute value of the correlation).
9159 : */
9160 10730 : if (*indexCorrelation < 1.0e-10)
9161 10694 : estimatedRanges = indexRanges;
9162 : else
9163 36 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
9164 :
9165 : /* we expect to visit this portion of the table */
9166 10730 : selec = estimatedRanges / indexRanges;
9167 :
9168 10730 : CLAMP_PROBABILITY(selec);
9169 :
9170 10730 : *indexSelectivity = selec;
9171 :
9172 : /*
9173 : * Compute the index qual costs, much as in genericcostestimate, to add to
9174 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
9175 : * support those.
9176 : */
9177 10730 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
9178 :
9179 : /*
9180 : * Compute the startup cost as the cost to read the whole revmap
9181 : * sequentially, including the cost to execute the index quals.
9182 : */
9183 10730 : *indexStartupCost =
9184 10730 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
9185 10730 : *indexStartupCost += qual_arg_cost;
9186 :
9187 : /*
9188 : * To read a BRIN index there might be a bit of back and forth over
9189 : * regular pages, as revmap might point to them out of sequential order;
9190 : * calculate the total cost as reading the whole index in random order.
9191 : */
9192 10730 : *indexTotalCost = *indexStartupCost +
9193 10730 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
9194 :
9195 : /*
9196 : * Charge a small amount per range tuple which we expect to match to. This
9197 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
9198 : * will set a bit for each page in the range when we find a matching
9199 : * range, so we must multiply the charge by the number of pages in the
9200 : * range.
9201 : */
9202 10730 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
9203 10730 : statsData.pagesPerRange;
9204 :
9205 10730 : *indexPages = index->pages;
9206 10730 : }
|
