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 562198 : eqsel(PG_FUNCTION_ARGS)
301 : {
302 562198 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
303 : }
304 :
305 : /*
306 : * Common code for eqsel() and neqsel()
307 : */
308 : static double
309 594736 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
310 : {
311 594736 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
312 594736 : Oid operator = PG_GETARG_OID(1);
313 594736 : List *args = (List *) PG_GETARG_POINTER(2);
314 594736 : int varRelid = PG_GETARG_INT32(3);
315 594736 : 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 594736 : if (negate)
326 : {
327 32538 : operator = get_negator(operator);
328 32538 : 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 594736 : if (!get_restriction_variable(root, args, varRelid,
340 : &vardata, &other, &varonleft))
341 3618 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
342 :
343 : /*
344 : * We can do a lot better if the something is a constant. (Note: the
345 : * Const might result from estimation rather than being a simple constant
346 : * in the query.)
347 : */
348 591114 : if (IsA(other, Const))
349 223573 : selec = var_eq_const(&vardata, operator, collation,
350 223573 : ((Const *) other)->constvalue,
351 223573 : ((Const *) other)->constisnull,
352 : varonleft, negate);
353 : else
354 367541 : selec = var_eq_non_const(&vardata, operator, collation, other,
355 : varonleft, negate);
356 :
357 591114 : ReleaseVariableStats(vardata);
358 :
359 591114 : 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 253551 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
369 : Datum constval, bool constisnull,
370 : bool varonleft, bool negate)
371 : {
372 : double selec;
373 253551 : 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 253551 : if (constisnull)
382 272 : 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 253279 : if (HeapTupleIsValid(vardata->statsTuple))
389 : {
390 : Form_pg_statistic stats;
391 :
392 180480 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
393 180480 : 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 253279 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
404 : {
405 51111 : selec = 1.0 / vardata->rel->tuples;
406 : }
407 340620 : else if (HeapTupleIsValid(vardata->statsTuple) &&
408 138452 : statistic_proc_security_check(vardata,
409 138452 : (opfuncoid = get_opcode(oproid))))
410 138452 : {
411 : AttStatsSlot sslot;
412 138452 : 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 138452 : if (get_attstatsslot(&sslot, vardata->statsTuple,
423 : STATISTIC_KIND_MCV, InvalidOid,
424 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
425 : {
426 125289 : LOCAL_FCINFO(fcinfo, 2);
427 : FmgrInfo eqproc;
428 :
429 125289 : 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 125289 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
438 : NULL, NULL);
439 125289 : fcinfo->args[0].isnull = false;
440 125289 : fcinfo->args[1].isnull = false;
441 : /* be careful to apply operator right way 'round */
442 125289 : if (varonleft)
443 125263 : fcinfo->args[1].value = constval;
444 : else
445 26 : fcinfo->args[0].value = constval;
446 :
447 2090443 : for (i = 0; i < sslot.nvalues; i++)
448 : {
449 : Datum fresult;
450 :
451 2029550 : if (varonleft)
452 2029504 : fcinfo->args[0].value = sslot.values[i];
453 : else
454 46 : fcinfo->args[1].value = sslot.values[i];
455 2029550 : fcinfo->isnull = false;
456 2029550 : fresult = FunctionCallInvoke(fcinfo);
457 2029550 : if (!fcinfo->isnull && DatumGetBool(fresult))
458 : {
459 64396 : match = true;
460 64396 : break;
461 : }
462 : }
463 : }
464 : else
465 : {
466 : /* no most-common-value info available */
467 13163 : i = 0; /* keep compiler quiet */
468 : }
469 :
470 138452 : 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 64396 : 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 74056 : double sumcommon = 0.0;
486 : double otherdistinct;
487 :
488 1771936 : for (i = 0; i < sslot.nnumbers; i++)
489 1697880 : sumcommon += sslot.numbers[i];
490 74056 : selec = 1.0 - sumcommon - nullfrac;
491 74056 : 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 74056 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
499 74056 : sslot.nnumbers;
500 74056 : if (otherdistinct > 1)
501 34580 : 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 74056 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
508 0 : selec = sslot.numbers[sslot.nnumbers - 1];
509 : }
510 :
511 138452 : 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 63716 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
521 : }
522 :
523 : /* now adjust if we wanted <> rather than = */
524 253279 : if (negate)
525 25801 : selec = 1.0 - selec - nullfrac;
526 :
527 : /* result should be in range, but make sure... */
528 253279 : CLAMP_PROBABILITY(selec);
529 :
530 253279 : 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 367541 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
540 : Node *other,
541 : bool varonleft, bool negate)
542 : {
543 : double selec;
544 367541 : double nullfrac = 0.0;
545 : bool isdefault;
546 :
547 : /*
548 : * Grab the nullfrac for use below.
549 : */
550 367541 : if (HeapTupleIsValid(vardata->statsTuple))
551 : {
552 : Form_pg_statistic stats;
553 :
554 219688 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
555 219688 : 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 367541 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
566 : {
567 129638 : selec = 1.0 / vardata->rel->tuples;
568 : }
569 237903 : 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 110886 : selec = 1.0 - nullfrac;
585 110886 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
586 110886 : if (ndistinct > 1)
587 108307 : selec /= ndistinct;
588 :
589 : /*
590 : * Cross-check: selectivity should never be estimated as more than the
591 : * most common value's.
592 : */
593 110886 : if (get_attstatsslot(&sslot, vardata->statsTuple,
594 : STATISTIC_KIND_MCV, InvalidOid,
595 : ATTSTATSSLOT_NUMBERS))
596 : {
597 95042 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
598 483 : selec = sslot.numbers[0];
599 95042 : 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 127017 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
610 : }
611 :
612 : /* now adjust if we wanted <> rather than = */
613 367541 : if (negate)
614 5148 : selec = 1.0 - selec - nullfrac;
615 :
616 : /* result should be in range, but make sure... */
617 367541 : CLAMP_PROBABILITY(selec);
618 :
619 367541 : 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 32538 : neqsel(PG_FUNCTION_ARGS)
631 : {
632 32538 : 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 247709 : 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 247709 : 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 21423 : if (vardata->var && IsA(vardata->var, Var) &&
672 17376 : ((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 1679 : if (vardata->rel->pages == 0)
683 0 : return 1.0;
684 :
685 1679 : itemptr = (ItemPointer) DatumGetPointer(constval);
686 1679 : 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 1679 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
696 :
697 : /* If target is the last page, use half the density. */
698 1679 : if (block >= vardata->rel->pages - 1)
699 31 : 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 1679 : if (density > 0.0)
710 : {
711 1679 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
712 :
713 1679 : 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 1679 : 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 1679 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
731 1562 : selec -= (1.0 / vardata->rel->tuples);
732 :
733 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
734 1679 : if (isgt)
735 1547 : selec = 1.0 - selec;
736 :
737 1679 : CLAMP_PROBABILITY(selec);
738 1679 : return selec;
739 : }
740 :
741 : /* no stats available, so default result */
742 19744 : return DEFAULT_INEQ_SEL;
743 : }
744 226286 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
745 :
746 226286 : 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 226286 : 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 226286 : 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 226286 : selec = 1.0 - stats->stanullfrac - sumcommon;
772 :
773 226286 : if (hist_selec >= 0.0)
774 138170 : 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 88116 : selec *= 0.5;
782 : }
783 :
784 226286 : selec += mcv_selec;
785 :
786 : /* result should be in range, but make sure... */
787 226286 : CLAMP_PROBABILITY(selec);
788 :
789 226286 : 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 230693 : 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 230693 : mcv_selec = 0.0;
815 230693 : sumcommon = 0.0;
816 :
817 459584 : if (HeapTupleIsValid(vardata->statsTuple) &&
818 457507 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
819 228616 : get_attstatsslot(&sslot, vardata->statsTuple,
820 : STATISTIC_KIND_MCV, InvalidOid,
821 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
822 : {
823 128117 : 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 128117 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
833 : NULL, NULL);
834 128117 : fcinfo->args[0].isnull = false;
835 128117 : fcinfo->args[1].isnull = false;
836 : /* be careful to apply operator right way 'round */
837 128117 : if (varonleft)
838 128117 : fcinfo->args[1].value = constval;
839 : else
840 0 : fcinfo->args[0].value = constval;
841 :
842 3027354 : for (i = 0; i < sslot.nvalues; i++)
843 : {
844 : Datum fresult;
845 :
846 2899237 : if (varonleft)
847 2899237 : fcinfo->args[0].value = sslot.values[i];
848 : else
849 0 : fcinfo->args[1].value = sslot.values[i];
850 2899237 : fcinfo->isnull = false;
851 2899237 : fresult = FunctionCallInvoke(fcinfo);
852 2899237 : if (!fcinfo->isnull && DatumGetBool(fresult))
853 1185871 : mcv_selec += sslot.numbers[i];
854 2899237 : sumcommon += sslot.numbers[i];
855 : }
856 128117 : free_attstatsslot(&sslot);
857 : }
858 :
859 230693 : *sumcommonp = sumcommon;
860 230693 : 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 4407 : 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 7012 : if (HeapTupleIsValid(vardata->statsTuple) &&
910 5205 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
911 2600 : get_attstatsslot(&sslot, vardata->statsTuple,
912 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
913 : ATTSTATSSLOT_VALUES))
914 : {
915 2522 : *hist_size = sslot.nvalues;
916 2522 : if (sslot.nvalues >= min_hist_size)
917 : {
918 1376 : LOCAL_FCINFO(fcinfo, 2);
919 1376 : 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 1376 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
931 : NULL, NULL);
932 1376 : fcinfo->args[0].isnull = false;
933 1376 : fcinfo->args[1].isnull = false;
934 : /* be careful to apply operator right way 'round */
935 1376 : if (varonleft)
936 1376 : fcinfo->args[1].value = constval;
937 : else
938 0 : fcinfo->args[0].value = constval;
939 :
940 115625 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
941 : {
942 : Datum fresult;
943 :
944 114249 : if (varonleft)
945 114249 : fcinfo->args[0].value = sslot.values[i];
946 : else
947 0 : fcinfo->args[1].value = sslot.values[i];
948 114249 : fcinfo->isnull = false;
949 114249 : fresult = FunctionCallInvoke(fcinfo);
950 114249 : if (!fcinfo->isnull && DatumGetBool(fresult))
951 4397 : nmatch++;
952 : }
953 1376 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
954 : }
955 : else
956 1146 : result = -1;
957 2522 : free_attstatsslot(&sslot);
958 : }
959 : else
960 : {
961 1885 : *hist_size = 0;
962 1885 : result = -1;
963 : }
964 :
965 4407 : return result;
966 : }
967 :
968 : /*
969 : * generic_restriction_selectivity - Selectivity for almost anything
970 : *
971 : * This function estimates selectivity for operators that we don't have any
972 : * special knowledge about, but are on data types that we collect standard
973 : * MCV and/or histogram statistics for. (Additional assumptions are that
974 : * the operator is strict and immutable, or at least stable.)
975 : *
976 : * If we have "VAR OP CONST" or "CONST OP VAR", selectivity is estimated by
977 : * applying the operator to each element of the column's MCV and/or histogram
978 : * stats, and merging the results using the assumption that the histogram is
979 : * a reasonable random sample of the column's non-MCV population. Note that
980 : * if the operator's semantics are related to the histogram ordering, this
981 : * might not be such a great assumption; other functions such as
982 : * scalarineqsel() are probably a better match in such cases.
983 : *
984 : * Otherwise, fall back to the default selectivity provided by the caller.
985 : */
986 : double
987 845 : generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation,
988 : List *args, int varRelid,
989 : double default_selectivity)
990 : {
991 : double selec;
992 : VariableStatData vardata;
993 : Node *other;
994 : bool varonleft;
995 :
996 : /*
997 : * If expression is not variable OP something or something OP variable,
998 : * then punt and return the default estimate.
999 : */
1000 845 : if (!get_restriction_variable(root, args, varRelid,
1001 : &vardata, &other, &varonleft))
1002 0 : return default_selectivity;
1003 :
1004 : /*
1005 : * If the something is a NULL constant, assume operator is strict and
1006 : * return zero, ie, operator will never return TRUE.
1007 : */
1008 845 : if (IsA(other, Const) &&
1009 845 : ((Const *) other)->constisnull)
1010 : {
1011 0 : ReleaseVariableStats(vardata);
1012 0 : return 0.0;
1013 : }
1014 :
1015 845 : if (IsA(other, Const))
1016 : {
1017 : /* Variable is being compared to a known non-null constant */
1018 845 : Datum constval = ((Const *) other)->constvalue;
1019 : FmgrInfo opproc;
1020 : double mcvsum;
1021 : double mcvsel;
1022 : double nullfrac;
1023 : int hist_size;
1024 :
1025 845 : fmgr_info(get_opcode(oproid), &opproc);
1026 :
1027 : /*
1028 : * Calculate the selectivity for the column's most common values.
1029 : */
1030 845 : mcvsel = mcv_selectivity(&vardata, &opproc, collation,
1031 : constval, varonleft,
1032 : &mcvsum);
1033 :
1034 : /*
1035 : * If the histogram is large enough, see what fraction of it matches
1036 : * the query, and assume that's representative of the non-MCV
1037 : * population. Otherwise use the default selectivity for the non-MCV
1038 : * population.
1039 : */
1040 845 : selec = histogram_selectivity(&vardata, &opproc, collation,
1041 : constval, varonleft,
1042 : 10, 1, &hist_size);
1043 845 : if (selec < 0)
1044 : {
1045 : /* Nope, fall back on default */
1046 845 : selec = default_selectivity;
1047 : }
1048 0 : else if (hist_size < 100)
1049 : {
1050 : /*
1051 : * For histogram sizes from 10 to 100, we combine the histogram
1052 : * and default selectivities, putting increasingly more trust in
1053 : * the histogram for larger sizes.
1054 : */
1055 0 : double hist_weight = hist_size / 100.0;
1056 :
1057 0 : selec = selec * hist_weight +
1058 0 : default_selectivity * (1.0 - hist_weight);
1059 : }
1060 :
1061 : /* In any case, don't believe extremely small or large estimates. */
1062 845 : if (selec < 0.0001)
1063 0 : selec = 0.0001;
1064 845 : else if (selec > 0.9999)
1065 0 : selec = 0.9999;
1066 :
1067 : /* Don't forget to account for nulls. */
1068 845 : if (HeapTupleIsValid(vardata.statsTuple))
1069 70 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
1070 : else
1071 775 : nullfrac = 0.0;
1072 :
1073 : /*
1074 : * Now merge the results from the MCV and histogram calculations,
1075 : * realizing that the histogram covers only the non-null values that
1076 : * are not listed in MCV.
1077 : */
1078 845 : selec *= 1.0 - nullfrac - mcvsum;
1079 845 : selec += mcvsel;
1080 : }
1081 : else
1082 : {
1083 : /* Comparison value is not constant, so we can't do anything */
1084 0 : selec = default_selectivity;
1085 : }
1086 :
1087 845 : ReleaseVariableStats(vardata);
1088 :
1089 : /* result should be in range, but make sure... */
1090 845 : CLAMP_PROBABILITY(selec);
1091 :
1092 845 : return selec;
1093 : }
1094 :
1095 : /*
1096 : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
1097 : *
1098 : * Determine the fraction of the variable's histogram population that
1099 : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
1100 : * The isgt and iseq flags distinguish which of the four cases apply.
1101 : *
1102 : * While opproc could be looked up from the operator OID, common callers
1103 : * also need to call it separately, so we make the caller pass both.
1104 : *
1105 : * Returns -1 if there is no histogram (valid results will always be >= 0).
1106 : *
1107 : * Note that the result disregards both the most-common-values (if any) and
1108 : * null entries. The caller is expected to combine this result with
1109 : * statistics for those portions of the column population.
1110 : *
1111 : * This is exported so that some other estimation functions can use it.
1112 : */
1113 : double
1114 229572 : 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 229572 : 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 458634 : if (HeapTupleIsValid(vardata->statsTuple) &&
1138 457854 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1139 228792 : get_attstatsslot(&sslot, vardata->statsTuple,
1140 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1141 : ATTSTATSSLOT_VALUES))
1142 : {
1143 140944 : if (sslot.nvalues > 1 &&
1144 281826 : sslot.stacoll == collation &&
1145 140882 : comparison_ops_are_compatible(sslot.staop, opoid))
1146 140792 : {
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 140792 : int lobound = 0; /* first possible slot to search */
1167 140792 : int hibound = sslot.nvalues; /* last+1 slot to search */
1168 140792 : 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 140792 : if (sslot.nvalues == 2)
1177 2339 : have_end = get_actual_variable_range(root,
1178 : vardata,
1179 : sslot.staop,
1180 : collation,
1181 : &sslot.values[0],
1182 2339 : &sslot.values[1]);
1183 :
1184 921910 : while (lobound < hibound)
1185 : {
1186 781118 : 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 781118 : if (probe == 0 && sslot.nvalues > 2)
1195 68768 : have_end = get_actual_variable_range(root,
1196 : vardata,
1197 : sslot.staop,
1198 : collation,
1199 : &sslot.values[0],
1200 : NULL);
1201 712350 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1202 47987 : have_end = get_actual_variable_range(root,
1203 : vardata,
1204 : sslot.staop,
1205 : collation,
1206 : NULL,
1207 47987 : &sslot.values[probe]);
1208 :
1209 781118 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1210 : collation,
1211 781118 : sslot.values[probe],
1212 : constval));
1213 781118 : if (isgt)
1214 50561 : ltcmp = !ltcmp;
1215 781118 : if (ltcmp)
1216 292115 : lobound = probe + 1;
1217 : else
1218 489003 : hibound = probe;
1219 : }
1220 :
1221 140792 : 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 60264 : histfrac = 0.0;
1232 : }
1233 80528 : else if (lobound >= sslot.nvalues)
1234 : {
1235 : /*
1236 : * Inverse case: constant is above upper histogram boundary.
1237 : */
1238 24228 : histfrac = 1.0;
1239 : }
1240 : else
1241 : {
1242 : /* We have values[i-1] <= constant <= values[i]. */
1243 56300 : int i = lobound;
1244 56300 : 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 56300 : if (i == 1 || isgt == iseq)
1267 : {
1268 : double otherdistinct;
1269 : bool isdefault;
1270 : AttStatsSlot mcvslot;
1271 :
1272 : /* Get estimated number of distinct values */
1273 23825 : otherdistinct = get_variable_numdistinct(vardata,
1274 : &isdefault);
1275 :
1276 : /* Subtract off the number of known MCVs */
1277 23825 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1278 : STATISTIC_KIND_MCV, InvalidOid,
1279 : ATTSTATSSLOT_NUMBERS))
1280 : {
1281 2851 : otherdistinct -= mcvslot.nnumbers;
1282 2851 : free_attstatsslot(&mcvslot);
1283 : }
1284 :
1285 : /* If result doesn't seem sane, leave eq_selec at 0 */
1286 23825 : if (otherdistinct > 1)
1287 23804 : 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 56300 : if (convert_to_scalar(constval, consttype, collation,
1296 : &val,
1297 56300 : sslot.values[i - 1], sslot.values[i],
1298 : vardata->vartype,
1299 : &low, &high))
1300 : {
1301 56300 : if (high <= low)
1302 : {
1303 : /* cope if bin boundaries appear identical */
1304 0 : binfrac = 0.5;
1305 : }
1306 56300 : else if (val <= low)
1307 11603 : binfrac = 0.0;
1308 44697 : else if (val >= high)
1309 2130 : binfrac = 1.0;
1310 : else
1311 : {
1312 42567 : 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 42567 : if (isnan(binfrac) ||
1321 42567 : 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 56300 : histfrac = (double) (i - 1) + binfrac;
1344 56300 : 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 56300 : if (i == 1)
1379 10092 : 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 56300 : if (isgt == iseq)
1387 18726 : 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 140792 : 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 140792 : if (have_end)
1406 79442 : CLAMP_PROBABILITY(hist_selec);
1407 : else
1408 : {
1409 61350 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1410 :
1411 61350 : if (hist_selec < cutoff)
1412 20621 : hist_selec = cutoff;
1413 40729 : else if (hist_selec > 1.0 - cutoff)
1414 15144 : hist_selec = 1.0 - cutoff;
1415 : }
1416 : }
1417 152 : else if (sslot.nvalues > 1)
1418 : {
1419 : /*
1420 : * If we get here, we have a histogram but it's not sorted the way
1421 : * we want. Do a brute-force search to see how many of the
1422 : * entries satisfy the comparison condition, and take that
1423 : * fraction as our estimate. (This is identical to the inner loop
1424 : * of histogram_selectivity; maybe share code?)
1425 : */
1426 152 : LOCAL_FCINFO(fcinfo, 2);
1427 152 : int nmatch = 0;
1428 :
1429 152 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1430 : NULL, NULL);
1431 152 : fcinfo->args[0].isnull = false;
1432 152 : fcinfo->args[1].isnull = false;
1433 152 : fcinfo->args[1].value = constval;
1434 801750 : for (int i = 0; i < sslot.nvalues; i++)
1435 : {
1436 : Datum fresult;
1437 :
1438 801598 : fcinfo->args[0].value = sslot.values[i];
1439 801598 : fcinfo->isnull = false;
1440 801598 : fresult = FunctionCallInvoke(fcinfo);
1441 801598 : if (!fcinfo->isnull && DatumGetBool(fresult))
1442 1614 : nmatch++;
1443 : }
1444 152 : hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1445 :
1446 : /*
1447 : * As above, clamp to a hundredth of the histogram resolution.
1448 : * This case is surely even less trustworthy than the normal one,
1449 : * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1450 : * clamp should be more restrictive in this case?)
1451 : */
1452 : {
1453 152 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1454 :
1455 152 : if (hist_selec < cutoff)
1456 10 : hist_selec = cutoff;
1457 142 : else if (hist_selec > 1.0 - cutoff)
1458 10 : hist_selec = 1.0 - cutoff;
1459 : }
1460 : }
1461 :
1462 140944 : free_attstatsslot(&sslot);
1463 : }
1464 :
1465 229572 : return hist_selec;
1466 : }
1467 :
1468 : /*
1469 : * Common wrapper function for the selectivity estimators that simply
1470 : * invoke scalarineqsel().
1471 : */
1472 : static Datum
1473 39031 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1474 : {
1475 39031 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1476 39031 : Oid operator = PG_GETARG_OID(1);
1477 39031 : List *args = (List *) PG_GETARG_POINTER(2);
1478 39031 : int varRelid = PG_GETARG_INT32(3);
1479 39031 : 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 39031 : if (!get_restriction_variable(root, args, varRelid,
1492 : &vardata, &other, &varonleft))
1493 462 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1494 :
1495 : /*
1496 : * Can't do anything useful if the something is not a constant, either.
1497 : */
1498 38569 : if (!IsA(other, Const))
1499 : {
1500 2365 : ReleaseVariableStats(vardata);
1501 2365 : 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 36204 : if (((Const *) other)->constisnull)
1509 : {
1510 55 : ReleaseVariableStats(vardata);
1511 55 : PG_RETURN_FLOAT8(0.0);
1512 : }
1513 36149 : constval = ((Const *) other)->constvalue;
1514 36149 : consttype = ((Const *) other)->consttype;
1515 :
1516 : /*
1517 : * Force the var to be on the left to simplify logic in scalarineqsel.
1518 : */
1519 36149 : if (!varonleft)
1520 : {
1521 318 : operator = get_commutator(operator);
1522 318 : if (!operator)
1523 : {
1524 : /* Use default selectivity (should we raise an error instead?) */
1525 0 : ReleaseVariableStats(vardata);
1526 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1527 : }
1528 318 : isgt = !isgt;
1529 : }
1530 :
1531 : /* The rest of the work is done by scalarineqsel(). */
1532 36149 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1533 : &vardata, constval, consttype);
1534 :
1535 36149 : ReleaseVariableStats(vardata);
1536 :
1537 36149 : PG_RETURN_FLOAT8((float8) selec);
1538 : }
1539 :
1540 : /*
1541 : * scalarltsel - Selectivity of "<" for scalars.
1542 : */
1543 : Datum
1544 12221 : scalarltsel(PG_FUNCTION_ARGS)
1545 : {
1546 12221 : return scalarineqsel_wrapper(fcinfo, false, false);
1547 : }
1548 :
1549 : /*
1550 : * scalarlesel - Selectivity of "<=" for scalars.
1551 : */
1552 : Datum
1553 3812 : scalarlesel(PG_FUNCTION_ARGS)
1554 : {
1555 3812 : return scalarineqsel_wrapper(fcinfo, false, true);
1556 : }
1557 :
1558 : /*
1559 : * scalargtsel - Selectivity of ">" for scalars.
1560 : */
1561 : Datum
1562 12418 : scalargtsel(PG_FUNCTION_ARGS)
1563 : {
1564 12418 : return scalarineqsel_wrapper(fcinfo, true, false);
1565 : }
1566 :
1567 : /*
1568 : * scalargesel - Selectivity of ">=" for scalars.
1569 : */
1570 : Datum
1571 10580 : scalargesel(PG_FUNCTION_ARGS)
1572 : {
1573 10580 : 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 47673 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1586 : {
1587 : VariableStatData vardata;
1588 : double selec;
1589 :
1590 47673 : examine_variable(root, arg, varRelid, &vardata);
1591 47673 : 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 22901 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1598 : BoolGetDatum(true), false, true, false);
1599 : }
1600 24772 : 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 11456 : selec = 0.3333333;
1610 : }
1611 : else
1612 : {
1613 : /* Otherwise, the default estimate is 0.5 */
1614 13316 : selec = 0.5;
1615 : }
1616 47673 : ReleaseVariableStats(vardata);
1617 47673 : return selec;
1618 : }
1619 :
1620 : /*
1621 : * booltestsel - Selectivity of BooleanTest Node.
1622 : */
1623 : Selectivity
1624 791 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1625 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1626 : {
1627 : VariableStatData vardata;
1628 : double selec;
1629 :
1630 791 : examine_variable(root, arg, varRelid, &vardata);
1631 :
1632 791 : if (HeapTupleIsValid(vardata.statsTuple))
1633 : {
1634 : Form_pg_statistic stats;
1635 : double freq_null;
1636 : AttStatsSlot sslot;
1637 :
1638 20 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1639 20 : freq_null = stats->stanullfrac;
1640 :
1641 20 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1642 : STATISTIC_KIND_MCV, InvalidOid,
1643 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1644 10 : && sslot.nnumbers > 0)
1645 10 : {
1646 : double freq_true;
1647 : double freq_false;
1648 :
1649 : /*
1650 : * Get first MCV frequency and derive frequency for true.
1651 : */
1652 10 : if (DatumGetBool(sslot.values[0]))
1653 0 : freq_true = sslot.numbers[0];
1654 : else
1655 10 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1656 :
1657 : /*
1658 : * Next derive frequency for false. Then use these as appropriate
1659 : * to derive frequency for each case.
1660 : */
1661 10 : freq_false = 1.0 - freq_true - freq_null;
1662 :
1663 10 : switch (booltesttype)
1664 : {
1665 0 : case IS_UNKNOWN:
1666 : /* select only NULL values */
1667 0 : selec = freq_null;
1668 0 : break;
1669 0 : case IS_NOT_UNKNOWN:
1670 : /* select non-NULL values */
1671 0 : selec = 1.0 - freq_null;
1672 0 : break;
1673 10 : case IS_TRUE:
1674 : /* select only TRUE values */
1675 10 : selec = freq_true;
1676 10 : break;
1677 0 : case IS_NOT_TRUE:
1678 : /* select non-TRUE values */
1679 0 : selec = 1.0 - freq_true;
1680 0 : break;
1681 0 : case IS_FALSE:
1682 : /* select only FALSE values */
1683 0 : selec = freq_false;
1684 0 : break;
1685 0 : case IS_NOT_FALSE:
1686 : /* select non-FALSE values */
1687 0 : selec = 1.0 - freq_false;
1688 0 : break;
1689 0 : default:
1690 0 : elog(ERROR, "unrecognized booltesttype: %d",
1691 : (int) booltesttype);
1692 : selec = 0.0; /* Keep compiler quiet */
1693 : break;
1694 : }
1695 :
1696 10 : free_attstatsslot(&sslot);
1697 : }
1698 : else
1699 : {
1700 : /*
1701 : * No most-common-value info available. Still have null fraction
1702 : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1703 : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1704 : */
1705 10 : switch (booltesttype)
1706 : {
1707 10 : case IS_UNKNOWN:
1708 : /* select only NULL values */
1709 10 : selec = freq_null;
1710 10 : break;
1711 0 : case IS_NOT_UNKNOWN:
1712 : /* select non-NULL values */
1713 0 : selec = 1.0 - freq_null;
1714 0 : break;
1715 0 : case IS_TRUE:
1716 : case IS_FALSE:
1717 : /* Assume we select half of the non-NULL values */
1718 0 : selec = (1.0 - freq_null) / 2.0;
1719 0 : break;
1720 0 : case IS_NOT_TRUE:
1721 : case IS_NOT_FALSE:
1722 : /* Assume we select NULLs plus half of the non-NULLs */
1723 : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1724 0 : selec = (freq_null + 1.0) / 2.0;
1725 0 : break;
1726 0 : default:
1727 0 : elog(ERROR, "unrecognized booltesttype: %d",
1728 : (int) booltesttype);
1729 : selec = 0.0; /* Keep compiler quiet */
1730 : break;
1731 : }
1732 : }
1733 : }
1734 : else
1735 : {
1736 : /*
1737 : * If we can't get variable statistics for the argument, perhaps
1738 : * clause_selectivity can do something with it. We ignore the
1739 : * possibility of a NULL value when using clause_selectivity, and just
1740 : * assume the value is either TRUE or FALSE.
1741 : */
1742 771 : switch (booltesttype)
1743 : {
1744 40 : case IS_UNKNOWN:
1745 40 : selec = DEFAULT_UNK_SEL;
1746 40 : break;
1747 90 : case IS_NOT_UNKNOWN:
1748 90 : selec = DEFAULT_NOT_UNK_SEL;
1749 90 : break;
1750 220 : case IS_TRUE:
1751 : case IS_NOT_FALSE:
1752 220 : selec = (double) clause_selectivity(root, arg,
1753 : varRelid,
1754 : jointype, sjinfo);
1755 220 : break;
1756 421 : case IS_FALSE:
1757 : case IS_NOT_TRUE:
1758 421 : selec = 1.0 - (double) clause_selectivity(root, arg,
1759 : varRelid,
1760 : jointype, sjinfo);
1761 421 : break;
1762 0 : default:
1763 0 : elog(ERROR, "unrecognized booltesttype: %d",
1764 : (int) booltesttype);
1765 : selec = 0.0; /* Keep compiler quiet */
1766 : break;
1767 : }
1768 : }
1769 :
1770 791 : ReleaseVariableStats(vardata);
1771 :
1772 : /* result should be in range, but make sure... */
1773 791 : CLAMP_PROBABILITY(selec);
1774 :
1775 791 : return (Selectivity) selec;
1776 : }
1777 :
1778 : /*
1779 : * nulltestsel - Selectivity of NullTest Node.
1780 : */
1781 : Selectivity
1782 14260 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1783 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1784 : {
1785 : VariableStatData vardata;
1786 : double selec;
1787 :
1788 14260 : examine_variable(root, arg, varRelid, &vardata);
1789 :
1790 14260 : if (HeapTupleIsValid(vardata.statsTuple))
1791 : {
1792 : Form_pg_statistic stats;
1793 : double freq_null;
1794 :
1795 7764 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1796 7764 : freq_null = stats->stanullfrac;
1797 :
1798 7764 : switch (nulltesttype)
1799 : {
1800 5976 : case IS_NULL:
1801 :
1802 : /*
1803 : * Use freq_null directly.
1804 : */
1805 5976 : selec = freq_null;
1806 5976 : break;
1807 1788 : case IS_NOT_NULL:
1808 :
1809 : /*
1810 : * Select not unknown (not null) values. Calculate from
1811 : * freq_null.
1812 : */
1813 1788 : selec = 1.0 - freq_null;
1814 1788 : break;
1815 0 : default:
1816 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1817 : (int) nulltesttype);
1818 : return (Selectivity) 0; /* keep compiler quiet */
1819 : }
1820 : }
1821 6496 : else if (vardata.var && IsA(vardata.var, Var) &&
1822 5962 : ((Var *) vardata.var)->varattno < 0)
1823 : {
1824 : /*
1825 : * There are no stats for system columns, but we know they are never
1826 : * NULL.
1827 : */
1828 89 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1829 : }
1830 : else
1831 : {
1832 : /*
1833 : * No ANALYZE stats available, so make a guess
1834 : */
1835 6407 : switch (nulltesttype)
1836 : {
1837 1648 : case IS_NULL:
1838 1648 : selec = DEFAULT_UNK_SEL;
1839 1648 : break;
1840 4759 : case IS_NOT_NULL:
1841 4759 : selec = DEFAULT_NOT_UNK_SEL;
1842 4759 : 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 14260 : ReleaseVariableStats(vardata);
1851 :
1852 : /* result should be in range, but make sure... */
1853 14260 : CLAMP_PROBABILITY(selec);
1854 :
1855 14260 : 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 105660 : strip_array_coercion(Node *node)
1868 : {
1869 : for (;;)
1870 : {
1871 105748 : if (node && IsA(node, ArrayCoerceExpr))
1872 88 : {
1873 1974 : 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 1974 : if (IsA(acoerce->elemexpr, RelabelType) &&
1880 88 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1881 88 : node = (Node *) acoerce->arg;
1882 : else
1883 : break;
1884 : }
1885 103774 : 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 105660 : return node;
1894 : }
1895 :
1896 : /*
1897 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1898 : */
1899 : Selectivity
1900 18276 : scalararraysel(PlannerInfo *root,
1901 : ScalarArrayOpExpr *clause,
1902 : bool is_join_clause,
1903 : int varRelid,
1904 : JoinType jointype,
1905 : SpecialJoinInfo *sjinfo)
1906 : {
1907 18276 : Oid operator = clause->opno;
1908 18276 : bool useOr = clause->useOr;
1909 18276 : bool isEquality = false;
1910 18276 : 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 18276 : leftop = (Node *) linitial(clause->args);
1924 18276 : rightop = (Node *) lsecond(clause->args);
1925 :
1926 : /* aggressively reduce both sides to constants */
1927 18276 : leftop = estimate_expression_value(root, leftop);
1928 18276 : rightop = estimate_expression_value(root, rightop);
1929 :
1930 : /* get nominal (after relabeling) element type of rightop */
1931 18276 : nominal_element_type = get_base_element_type(exprType(rightop));
1932 18276 : if (!OidIsValid(nominal_element_type))
1933 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1934 : /* get nominal collation, too, for generating constants */
1935 18276 : nominal_element_collation = exprCollation(rightop);
1936 :
1937 : /* look through any binary-compatible relabeling of rightop */
1938 18276 : 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 18276 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1945 18276 : if (OidIsValid(typentry->eq_opr))
1946 : {
1947 18274 : if (operator == typentry->eq_opr)
1948 15905 : isEquality = true;
1949 2369 : else if (get_negator(operator) == typentry->eq_opr)
1950 1890 : 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 18276 : if ((isEquality || isInequality) && !is_join_clause)
1960 : {
1961 17794 : s1 = scalararraysel_containment(root, leftop, rightop,
1962 : nominal_element_type,
1963 : isEquality, useOr, varRelid);
1964 17794 : if (s1 >= 0.0)
1965 95 : return s1;
1966 : }
1967 :
1968 : /*
1969 : * Look up the underlying operator's selectivity estimator. Punt if it
1970 : * hasn't got one.
1971 : */
1972 18181 : if (is_join_clause)
1973 1 : oprsel = get_oprjoin(operator);
1974 : else
1975 18180 : oprsel = get_oprrest(operator);
1976 18181 : if (!oprsel)
1977 2 : return (Selectivity) 0.5;
1978 18179 : 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 18179 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1989 15960 : isEquality = true;
1990 2219 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1991 1819 : 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 18179 : if (rightop && IsA(rightop, Const))
2006 14654 : {
2007 14684 : Datum arraydatum = ((Const *) rightop)->constvalue;
2008 14684 : 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 14684 : if (arrayisnull) /* qual can't succeed if null array */
2019 30 : return (Selectivity) 0.0;
2020 14659 : arrayval = DatumGetArrayTypeP(arraydatum);
2021 :
2022 : /*
2023 : * When the array contains a NULL constant, same as var_eq_const, we
2024 : * assume the operator is strict and nothing will match, thus return
2025 : * 0.0.
2026 : */
2027 14659 : if (!useOr && array_contains_nulls(arrayval))
2028 5 : return (Selectivity) 0.0;
2029 :
2030 14654 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
2031 : &elmlen, &elmbyval, &elmalign);
2032 14654 : deconstruct_array(arrayval,
2033 : ARR_ELEMTYPE(arrayval),
2034 : elmlen, elmbyval, elmalign,
2035 : &elem_values, &elem_nulls, &num_elems);
2036 :
2037 : /*
2038 : * For generic operators, we assume the probability of success is
2039 : * independent for each array element. But for "= ANY" or "<> ALL",
2040 : * if the array elements are distinct (which'd typically be the case)
2041 : * then the probabilities are disjoint, and we should just sum them.
2042 : *
2043 : * If we were being really tense we would try to confirm that the
2044 : * elements are all distinct, but that would be expensive and it
2045 : * doesn't seem to be worth the cycles; it would amount to penalizing
2046 : * well-written queries in favor of poorly-written ones. However, we
2047 : * do protect ourselves a little bit by checking whether the
2048 : * disjointness assumption leads to an impossible (out of range)
2049 : * probability; if so, we fall back to the normal calculation.
2050 : */
2051 14654 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2052 :
2053 57056 : for (i = 0; i < num_elems; i++)
2054 : {
2055 : List *args;
2056 : Selectivity s2;
2057 :
2058 42402 : args = list_make2(leftop,
2059 : makeConst(nominal_element_type,
2060 : -1,
2061 : nominal_element_collation,
2062 : elmlen,
2063 : elem_values[i],
2064 : elem_nulls[i],
2065 : elmbyval));
2066 42402 : if (is_join_clause)
2067 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2068 : clause->inputcollid,
2069 : PointerGetDatum(root),
2070 : ObjectIdGetDatum(operator),
2071 : PointerGetDatum(args),
2072 : Int16GetDatum(jointype),
2073 : PointerGetDatum(sjinfo)));
2074 : else
2075 42402 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2076 : clause->inputcollid,
2077 : PointerGetDatum(root),
2078 : ObjectIdGetDatum(operator),
2079 : PointerGetDatum(args),
2080 : Int32GetDatum(varRelid)));
2081 :
2082 42402 : if (useOr)
2083 : {
2084 36568 : s1 = s1 + s2 - s1 * s2;
2085 36568 : if (isEquality)
2086 35698 : s1disjoint += s2;
2087 : }
2088 : else
2089 : {
2090 5834 : s1 = s1 * s2;
2091 5834 : if (isInequality)
2092 5574 : s1disjoint += s2 - 1.0;
2093 : }
2094 : }
2095 :
2096 : /* accept disjoint-probability estimate if in range */
2097 14654 : if ((useOr ? isEquality : isInequality) &&
2098 14149 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2099 14102 : s1 = s1disjoint;
2100 : }
2101 3495 : else if (rightop && IsA(rightop, ArrayExpr) &&
2102 296 : !((ArrayExpr *) rightop)->multidims)
2103 291 : {
2104 296 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2105 : int16 elmlen;
2106 : bool elmbyval;
2107 : ListCell *l;
2108 :
2109 296 : get_typlenbyval(arrayexpr->element_typeid,
2110 : &elmlen, &elmbyval);
2111 :
2112 : /*
2113 : * We use the assumption of disjoint probabilities here too, although
2114 : * the odds of equal array elements are rather higher if the elements
2115 : * are not all constants (which they won't be, else constant folding
2116 : * would have reduced the ArrayExpr to a Const). In this path it's
2117 : * critical to have the sanity check on the s1disjoint estimate.
2118 : */
2119 296 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2120 :
2121 1075 : foreach(l, arrayexpr->elements)
2122 : {
2123 784 : Node *elem = (Node *) lfirst(l);
2124 : List *args;
2125 : Selectivity s2;
2126 :
2127 : /*
2128 : * When the array contains a NULL constant, same as var_eq_const,
2129 : * we assume the operator is strict and nothing will match, thus
2130 : * return 0.0.
2131 : */
2132 784 : if (!useOr && IsA(elem, Const) && ((Const *) elem)->constisnull)
2133 5 : return (Selectivity) 0.0;
2134 :
2135 : /*
2136 : * Theoretically, if elem isn't of nominal_element_type we should
2137 : * insert a RelabelType, but it seems unlikely that any operator
2138 : * estimation function would really care ...
2139 : */
2140 779 : args = list_make2(leftop, elem);
2141 779 : if (is_join_clause)
2142 3 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2143 : clause->inputcollid,
2144 : PointerGetDatum(root),
2145 : ObjectIdGetDatum(operator),
2146 : PointerGetDatum(args),
2147 : Int16GetDatum(jointype),
2148 : PointerGetDatum(sjinfo)));
2149 : else
2150 776 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2151 : clause->inputcollid,
2152 : PointerGetDatum(root),
2153 : ObjectIdGetDatum(operator),
2154 : PointerGetDatum(args),
2155 : Int32GetDatum(varRelid)));
2156 :
2157 779 : if (useOr)
2158 : {
2159 759 : s1 = s1 + s2 - s1 * s2;
2160 759 : if (isEquality)
2161 759 : s1disjoint += s2;
2162 : }
2163 : else
2164 : {
2165 20 : s1 = s1 * s2;
2166 20 : if (isInequality)
2167 20 : s1disjoint += s2 - 1.0;
2168 : }
2169 : }
2170 :
2171 : /* accept disjoint-probability estimate if in range */
2172 291 : if ((useOr ? isEquality : isInequality) &&
2173 291 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2174 291 : s1 = s1disjoint;
2175 : }
2176 : else
2177 : {
2178 : CaseTestExpr *dummyexpr;
2179 : List *args;
2180 : Selectivity s2;
2181 : int i;
2182 :
2183 : /*
2184 : * We need a dummy rightop to pass to the operator selectivity
2185 : * routine. It can be pretty much anything that doesn't look like a
2186 : * constant; CaseTestExpr is a convenient choice.
2187 : */
2188 3199 : dummyexpr = makeNode(CaseTestExpr);
2189 3199 : dummyexpr->typeId = nominal_element_type;
2190 3199 : dummyexpr->typeMod = -1;
2191 3199 : dummyexpr->collation = clause->inputcollid;
2192 3199 : args = list_make2(leftop, dummyexpr);
2193 3199 : if (is_join_clause)
2194 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2195 : clause->inputcollid,
2196 : PointerGetDatum(root),
2197 : ObjectIdGetDatum(operator),
2198 : PointerGetDatum(args),
2199 : Int16GetDatum(jointype),
2200 : PointerGetDatum(sjinfo)));
2201 : else
2202 3199 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2203 : clause->inputcollid,
2204 : PointerGetDatum(root),
2205 : ObjectIdGetDatum(operator),
2206 : PointerGetDatum(args),
2207 : Int32GetDatum(varRelid)));
2208 3199 : s1 = useOr ? 0.0 : 1.0;
2209 :
2210 : /*
2211 : * Arbitrarily assume 10 elements in the eventual array value (see
2212 : * also estimate_array_length). We don't risk an assumption of
2213 : * disjoint probabilities here.
2214 : */
2215 35189 : for (i = 0; i < 10; i++)
2216 : {
2217 31990 : if (useOr)
2218 31990 : s1 = s1 + s2 - s1 * s2;
2219 : else
2220 0 : s1 = s1 * s2;
2221 : }
2222 : }
2223 :
2224 : /* result should be in range, but make sure... */
2225 18144 : CLAMP_PROBABILITY(s1);
2226 :
2227 18144 : return s1;
2228 : }
2229 :
2230 : /*
2231 : * Estimate number of elements in the array yielded by an expression.
2232 : *
2233 : * Note: the result is integral, but we use "double" to avoid overflow
2234 : * concerns. Most callers will use it in double-type expressions anyway.
2235 : *
2236 : * Note: in some code paths root can be passed as NULL, resulting in
2237 : * slightly worse estimates.
2238 : */
2239 : double
2240 87384 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2241 : {
2242 : /* look through any binary-compatible relabeling of arrayexpr */
2243 87384 : arrayexpr = strip_array_coercion(arrayexpr);
2244 :
2245 87384 : if (arrayexpr && IsA(arrayexpr, Const))
2246 : {
2247 38417 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2248 38417 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2249 : ArrayType *arrayval;
2250 :
2251 38417 : if (arrayisnull)
2252 75 : return 0;
2253 38342 : arrayval = DatumGetArrayTypeP(arraydatum);
2254 38342 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2255 : }
2256 48967 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2257 512 : !((ArrayExpr *) arrayexpr)->multidims)
2258 : {
2259 512 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2260 : }
2261 48455 : else if (arrayexpr && root)
2262 : {
2263 : /* See if we can find any statistics about it */
2264 : VariableStatData vardata;
2265 : AttStatsSlot sslot;
2266 48435 : double nelem = 0;
2267 :
2268 : /*
2269 : * Skip calling examine_variable for Var with varno 0, which has no
2270 : * valid relation entry and would error in find_base_rel. Such a Var
2271 : * can appear when a nested set operation's output type doesn't match
2272 : * the parent's expected type, because recurse_set_operations builds a
2273 : * projection target list using generate_setop_tlist with varno 0, and
2274 : * if the required type coercion involves an ArrayCoerceExpr, we can
2275 : * be called on that Var.
2276 : */
2277 48435 : if (IsA(arrayexpr, Var) && ((Var *) arrayexpr)->varno == 0)
2278 6421 : return 10; /* default guess, should match scalararraysel */
2279 :
2280 48430 : examine_variable(root, arrayexpr, 0, &vardata);
2281 48430 : if (HeapTupleIsValid(vardata.statsTuple))
2282 : {
2283 : /*
2284 : * Found stats, so use the average element count, which is stored
2285 : * in the last stanumbers element of the DECHIST statistics.
2286 : * Actually that is the average count of *distinct* elements;
2287 : * perhaps we should scale it up somewhat?
2288 : */
2289 6511 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2290 : STATISTIC_KIND_DECHIST, InvalidOid,
2291 : ATTSTATSSLOT_NUMBERS))
2292 : {
2293 6416 : if (sslot.nnumbers > 0)
2294 6416 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2295 6416 : free_attstatsslot(&sslot);
2296 : }
2297 : }
2298 48430 : ReleaseVariableStats(vardata);
2299 :
2300 48430 : if (nelem > 0)
2301 6416 : return nelem;
2302 : }
2303 :
2304 : /* Else use a default guess --- this should match scalararraysel */
2305 42034 : return 10;
2306 : }
2307 :
2308 : /*
2309 : * rowcomparesel - Selectivity of RowCompareExpr Node.
2310 : *
2311 : * We estimate RowCompare selectivity by considering just the first (high
2312 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2313 : * this estimate could be refined by considering additional columns, it
2314 : * seems unlikely that we could do a lot better without multi-column
2315 : * statistics.
2316 : */
2317 : Selectivity
2318 240 : rowcomparesel(PlannerInfo *root,
2319 : RowCompareExpr *clause,
2320 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2321 : {
2322 : Selectivity s1;
2323 240 : Oid opno = linitial_oid(clause->opnos);
2324 240 : Oid inputcollid = linitial_oid(clause->inputcollids);
2325 : List *opargs;
2326 : bool is_join_clause;
2327 :
2328 : /* Build equivalent arg list for single operator */
2329 240 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2330 :
2331 : /*
2332 : * Decide if it's a join clause. This should match clausesel.c's
2333 : * treat_as_join_clause(), except that we intentionally consider only the
2334 : * leading columns and not the rest of the clause.
2335 : */
2336 240 : if (varRelid != 0)
2337 : {
2338 : /*
2339 : * Caller is forcing restriction mode (eg, because we are examining an
2340 : * inner indexscan qual).
2341 : */
2342 45 : is_join_clause = false;
2343 : }
2344 195 : else if (sjinfo == NULL)
2345 : {
2346 : /*
2347 : * It must be a restriction clause, since it's being evaluated at a
2348 : * scan node.
2349 : */
2350 175 : is_join_clause = false;
2351 : }
2352 : else
2353 : {
2354 : /*
2355 : * Otherwise, it's a join if there's more than one base relation used.
2356 : */
2357 20 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2358 : }
2359 :
2360 240 : if (is_join_clause)
2361 : {
2362 : /* Estimate selectivity for a join clause. */
2363 20 : s1 = join_selectivity(root, opno,
2364 : opargs,
2365 : inputcollid,
2366 : jointype,
2367 : sjinfo);
2368 : }
2369 : else
2370 : {
2371 : /* Estimate selectivity for a restriction clause. */
2372 220 : s1 = restriction_selectivity(root, opno,
2373 : opargs,
2374 : inputcollid,
2375 : varRelid);
2376 : }
2377 :
2378 240 : return s1;
2379 : }
2380 :
2381 : /*
2382 : * eqjoinsel - Join selectivity of "="
2383 : */
2384 : Datum
2385 218790 : eqjoinsel(PG_FUNCTION_ARGS)
2386 : {
2387 218790 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2388 218790 : Oid operator = PG_GETARG_OID(1);
2389 218790 : List *args = (List *) PG_GETARG_POINTER(2);
2390 :
2391 : #ifdef NOT_USED
2392 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2393 : #endif
2394 218790 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2395 218790 : Oid collation = PG_GET_COLLATION();
2396 : double selec;
2397 : double selec_inner;
2398 : VariableStatData vardata1;
2399 : VariableStatData vardata2;
2400 : double nd1;
2401 : double nd2;
2402 : bool isdefault1;
2403 : bool isdefault2;
2404 : Oid opfuncoid;
2405 : FmgrInfo eqproc;
2406 218790 : Oid hashLeft = InvalidOid;
2407 218790 : Oid hashRight = InvalidOid;
2408 : AttStatsSlot sslot1;
2409 : AttStatsSlot sslot2;
2410 218790 : Form_pg_statistic stats1 = NULL;
2411 218790 : Form_pg_statistic stats2 = NULL;
2412 218790 : bool have_mcvs1 = false;
2413 218790 : bool have_mcvs2 = false;
2414 218790 : bool *hasmatch1 = NULL;
2415 218790 : bool *hasmatch2 = NULL;
2416 218790 : int nmatches = 0;
2417 : bool get_mcv_stats;
2418 : bool join_is_reversed;
2419 : RelOptInfo *inner_rel;
2420 :
2421 218790 : get_join_variables(root, args, sjinfo,
2422 : &vardata1, &vardata2, &join_is_reversed);
2423 :
2424 218790 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2425 218790 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2426 :
2427 218790 : opfuncoid = get_opcode(operator);
2428 :
2429 218790 : memset(&sslot1, 0, sizeof(sslot1));
2430 218790 : memset(&sslot2, 0, sizeof(sslot2));
2431 :
2432 : /*
2433 : * There is no use in fetching one side's MCVs if we lack MCVs for the
2434 : * other side, so do a quick check to verify that both stats exist.
2435 : */
2436 587865 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2437 262524 : HeapTupleIsValid(vardata2.statsTuple) &&
2438 112239 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2439 : STATISTIC_KIND_MCV, InvalidOid,
2440 369075 : 0) &&
2441 56257 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2442 : STATISTIC_KIND_MCV, InvalidOid,
2443 : 0));
2444 :
2445 218790 : if (HeapTupleIsValid(vardata1.statsTuple))
2446 : {
2447 : /* note we allow use of nullfrac regardless of security check */
2448 150285 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2449 172839 : if (get_mcv_stats &&
2450 22554 : statistic_proc_security_check(&vardata1, opfuncoid))
2451 22554 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2452 : STATISTIC_KIND_MCV, InvalidOid,
2453 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2454 : }
2455 :
2456 218790 : if (HeapTupleIsValid(vardata2.statsTuple))
2457 : {
2458 : /* note we allow use of nullfrac regardless of security check */
2459 129643 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2460 152197 : if (get_mcv_stats &&
2461 22554 : statistic_proc_security_check(&vardata2, opfuncoid))
2462 22554 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2463 : STATISTIC_KIND_MCV, InvalidOid,
2464 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2465 : }
2466 :
2467 : /* Prepare info usable by both eqjoinsel_inner and eqjoinsel_semi */
2468 218790 : if (have_mcvs1 && have_mcvs2)
2469 : {
2470 22554 : fmgr_info(opfuncoid, &eqproc);
2471 22554 : hasmatch1 = (bool *) palloc0(sslot1.nvalues * sizeof(bool));
2472 22554 : hasmatch2 = (bool *) palloc0(sslot2.nvalues * sizeof(bool));
2473 :
2474 : /*
2475 : * If the MCV lists are long enough to justify hashing, try to look up
2476 : * hash functions for the join operator.
2477 : */
2478 22554 : if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
2479 1288 : (void) get_op_hash_functions_ext(operator,
2480 1288 : exprType((Node *) linitial(args)),
2481 : &hashLeft, &hashRight);
2482 : }
2483 : else
2484 196236 : memset(&eqproc, 0, sizeof(eqproc)); /* silence uninit-var warnings */
2485 :
2486 : /* We need to compute the inner-join selectivity in all cases */
2487 218790 : selec_inner = eqjoinsel_inner(&eqproc, collation,
2488 : hashLeft, hashRight,
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 :
2498 218790 : switch (sjinfo->jointype)
2499 : {
2500 207382 : case JOIN_INNER:
2501 : case JOIN_LEFT:
2502 : case JOIN_FULL:
2503 207382 : selec = selec_inner;
2504 207382 : break;
2505 11408 : case JOIN_SEMI:
2506 : case JOIN_ANTI:
2507 :
2508 : /*
2509 : * Look up the join's inner relation. min_righthand is sufficient
2510 : * information because neither SEMI nor ANTI joins permit any
2511 : * reassociation into or out of their RHS, so the righthand will
2512 : * always be exactly that set of rels.
2513 : */
2514 11408 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2515 :
2516 11408 : if (!join_is_reversed)
2517 5419 : selec = eqjoinsel_semi(&eqproc, collation,
2518 : hashLeft, hashRight,
2519 : false,
2520 : &vardata1, &vardata2,
2521 : nd1, nd2,
2522 : isdefault1, isdefault2,
2523 : &sslot1, &sslot2,
2524 : stats1, stats2,
2525 : have_mcvs1, have_mcvs2,
2526 : hasmatch1, hasmatch2,
2527 : &nmatches,
2528 : inner_rel);
2529 : else
2530 5989 : selec = eqjoinsel_semi(&eqproc, collation,
2531 : hashLeft, hashRight,
2532 : true,
2533 : &vardata2, &vardata1,
2534 : nd2, nd1,
2535 : isdefault2, isdefault1,
2536 : &sslot2, &sslot1,
2537 : stats2, stats1,
2538 : have_mcvs2, have_mcvs1,
2539 : hasmatch2, hasmatch1,
2540 : &nmatches,
2541 : inner_rel);
2542 :
2543 : /*
2544 : * We should never estimate the output of a semijoin to be more
2545 : * rows than we estimate for an inner join with the same input
2546 : * rels and join condition; it's obviously impossible for that to
2547 : * happen. The former estimate is N1 * Ssemi while the latter is
2548 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2549 : * this is worthwhile because of the shakier estimation rules we
2550 : * use in eqjoinsel_semi, particularly in cases where it has to
2551 : * punt entirely.
2552 : */
2553 11408 : selec = Min(selec, inner_rel->rows * selec_inner);
2554 11408 : break;
2555 0 : default:
2556 : /* other values not expected here */
2557 0 : elog(ERROR, "unrecognized join type: %d",
2558 : (int) sjinfo->jointype);
2559 : selec = 0; /* keep compiler quiet */
2560 : break;
2561 : }
2562 :
2563 218790 : free_attstatsslot(&sslot1);
2564 218790 : free_attstatsslot(&sslot2);
2565 :
2566 218790 : ReleaseVariableStats(vardata1);
2567 218790 : ReleaseVariableStats(vardata2);
2568 :
2569 218790 : if (hasmatch1)
2570 22554 : pfree(hasmatch1);
2571 218790 : if (hasmatch2)
2572 22554 : pfree(hasmatch2);
2573 :
2574 218790 : CLAMP_PROBABILITY(selec);
2575 :
2576 218790 : PG_RETURN_FLOAT8((float8) selec);
2577 : }
2578 :
2579 : /*
2580 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2581 : *
2582 : * In addition to computing the selectivity estimate, this will fill
2583 : * hasmatch1[], hasmatch2[], and *p_nmatches (if have_mcvs1 && have_mcvs2).
2584 : * We may be able to re-use that data in eqjoinsel_semi.
2585 : *
2586 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2587 : * that it's worth trying to distinguish them here.
2588 : */
2589 : static double
2590 218790 : eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
2591 : Oid hashLeft, Oid hashRight,
2592 : VariableStatData *vardata1, VariableStatData *vardata2,
2593 : double nd1, double nd2,
2594 : bool isdefault1, bool isdefault2,
2595 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2596 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2597 : bool have_mcvs1, bool have_mcvs2,
2598 : bool *hasmatch1, bool *hasmatch2,
2599 : int *p_nmatches)
2600 : {
2601 : double selec;
2602 :
2603 218790 : if (have_mcvs1 && have_mcvs2)
2604 22554 : {
2605 : /*
2606 : * We have most-common-value lists for both relations. Run through
2607 : * the lists to see which MCVs actually join to each other with the
2608 : * given operator. This allows us to determine the exact join
2609 : * selectivity for the portion of the relations represented by the MCV
2610 : * lists. We still have to estimate for the remaining population, but
2611 : * in a skewed distribution this gives us a big leg up in accuracy.
2612 : * For motivation see the analysis in Y. Ioannidis and S.
2613 : * Christodoulakis, "On the propagation of errors in the size of join
2614 : * results", Technical Report 1018, Computer Science Dept., University
2615 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2616 : */
2617 22554 : double nullfrac1 = stats1->stanullfrac;
2618 22554 : double nullfrac2 = stats2->stanullfrac;
2619 : double matchprodfreq,
2620 : matchfreq1,
2621 : matchfreq2,
2622 : unmatchfreq1,
2623 : unmatchfreq2,
2624 : otherfreq1,
2625 : otherfreq2,
2626 : totalsel1,
2627 : totalsel2;
2628 : int i,
2629 : nmatches;
2630 :
2631 : /* Fill the match arrays */
2632 22554 : eqjoinsel_find_matches(eqproc, collation,
2633 : hashLeft, hashRight,
2634 : false,
2635 : sslot1, sslot2,
2636 : sslot1->nvalues, sslot2->nvalues,
2637 : hasmatch1, hasmatch2,
2638 : p_nmatches, &matchprodfreq);
2639 22554 : nmatches = *p_nmatches;
2640 22554 : CLAMP_PROBABILITY(matchprodfreq);
2641 :
2642 : /* Sum up frequencies of matched and unmatched MCVs */
2643 22554 : matchfreq1 = unmatchfreq1 = 0.0;
2644 526593 : for (i = 0; i < sslot1->nvalues; i++)
2645 : {
2646 504039 : if (hasmatch1[i])
2647 234753 : matchfreq1 += sslot1->numbers[i];
2648 : else
2649 269286 : unmatchfreq1 += sslot1->numbers[i];
2650 : }
2651 22554 : CLAMP_PROBABILITY(matchfreq1);
2652 22554 : CLAMP_PROBABILITY(unmatchfreq1);
2653 22554 : matchfreq2 = unmatchfreq2 = 0.0;
2654 411388 : for (i = 0; i < sslot2->nvalues; i++)
2655 : {
2656 388834 : if (hasmatch2[i])
2657 234753 : matchfreq2 += sslot2->numbers[i];
2658 : else
2659 154081 : unmatchfreq2 += sslot2->numbers[i];
2660 : }
2661 22554 : CLAMP_PROBABILITY(matchfreq2);
2662 22554 : CLAMP_PROBABILITY(unmatchfreq2);
2663 :
2664 : /*
2665 : * Compute total frequency of non-null values that are not in the MCV
2666 : * lists.
2667 : */
2668 22554 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2669 22554 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2670 22554 : CLAMP_PROBABILITY(otherfreq1);
2671 22554 : CLAMP_PROBABILITY(otherfreq2);
2672 :
2673 : /*
2674 : * We can estimate the total selectivity from the point of view of
2675 : * relation 1 as: the known selectivity for matched MCVs, plus
2676 : * unmatched MCVs that are assumed to match against random members of
2677 : * relation 2's non-MCV population, plus non-MCV values that are
2678 : * assumed to match against random members of relation 2's unmatched
2679 : * MCVs plus non-MCV values.
2680 : */
2681 22554 : totalsel1 = matchprodfreq;
2682 22554 : if (nd2 > sslot2->nvalues)
2683 4310 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2684 22554 : if (nd2 > nmatches)
2685 8235 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2686 8235 : (nd2 - nmatches);
2687 : /* Same estimate from the point of view of relation 2. */
2688 22554 : totalsel2 = matchprodfreq;
2689 22554 : if (nd1 > sslot1->nvalues)
2690 5018 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2691 22554 : if (nd1 > nmatches)
2692 7219 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2693 7219 : (nd1 - nmatches);
2694 :
2695 : /*
2696 : * Use the smaller of the two estimates. This can be justified in
2697 : * essentially the same terms as given below for the no-stats case: to
2698 : * a first approximation, we are estimating from the point of view of
2699 : * the relation with smaller nd.
2700 : */
2701 22554 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2702 : }
2703 : else
2704 : {
2705 : /*
2706 : * We do not have MCV lists for both sides. Estimate the join
2707 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2708 : * is plausible if we assume that the join operator is strict and the
2709 : * non-null values are about equally distributed: a given non-null
2710 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2711 : * of rel2, so total join rows are at most
2712 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2713 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2714 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2715 : * with MIN() is an upper bound. Using the MIN() means we estimate
2716 : * from the point of view of the relation with smaller nd (since the
2717 : * larger nd is determining the MIN). It is reasonable to assume that
2718 : * most tuples in this rel will have join partners, so the bound is
2719 : * probably reasonably tight and should be taken as-is.
2720 : *
2721 : * XXX Can we be smarter if we have an MCV list for just one side? It
2722 : * seems that if we assume equal distribution for the other side, we
2723 : * end up with the same answer anyway.
2724 : */
2725 196236 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2726 196236 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2727 :
2728 196236 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2729 196236 : if (nd1 > nd2)
2730 96311 : selec /= nd1;
2731 : else
2732 99925 : selec /= nd2;
2733 : }
2734 :
2735 218790 : return selec;
2736 : }
2737 :
2738 : /*
2739 : * eqjoinsel_semi --- eqjoinsel for semi join
2740 : *
2741 : * (Also used for anti join, which we are supposed to estimate the same way.)
2742 : * Caller has ensured that vardata1 is the LHS variable; however, eqproc
2743 : * is for the original join operator, which might now need to have the inputs
2744 : * swapped in order to apply correctly. Also, if have_mcvs1 && have_mcvs2
2745 : * then hasmatch1[], hasmatch2[], and *p_nmatches were filled by
2746 : * eqjoinsel_inner.
2747 : */
2748 : static double
2749 11408 : eqjoinsel_semi(FmgrInfo *eqproc, Oid collation,
2750 : Oid hashLeft, Oid hashRight,
2751 : bool op_is_reversed,
2752 : VariableStatData *vardata1, VariableStatData *vardata2,
2753 : double nd1, double nd2,
2754 : bool isdefault1, bool isdefault2,
2755 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2756 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2757 : bool have_mcvs1, bool have_mcvs2,
2758 : bool *hasmatch1, bool *hasmatch2,
2759 : int *p_nmatches,
2760 : RelOptInfo *inner_rel)
2761 : {
2762 : double selec;
2763 :
2764 : /*
2765 : * We clamp nd2 to be not more than what we estimate the inner relation's
2766 : * size to be. This is intuitively somewhat reasonable since obviously
2767 : * there can't be more than that many distinct values coming from the
2768 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2769 : * likewise) is that this is the only pathway by which restriction clauses
2770 : * applied to the inner rel will affect the join result size estimate,
2771 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2772 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2773 : * the selectivity of outer-rel restrictions.
2774 : *
2775 : * We can apply this clamping both with respect to the base relation from
2776 : * which the join variable comes (if there is just one), and to the
2777 : * immediate inner input relation of the current join.
2778 : *
2779 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2780 : * great, maybe, but it didn't come out of nowhere either. This is most
2781 : * helpful when the inner relation is empty and consequently has no stats.
2782 : */
2783 11408 : if (vardata2->rel)
2784 : {
2785 11403 : if (nd2 >= vardata2->rel->rows)
2786 : {
2787 9643 : nd2 = vardata2->rel->rows;
2788 9643 : isdefault2 = false;
2789 : }
2790 : }
2791 11408 : if (nd2 >= inner_rel->rows)
2792 : {
2793 9593 : nd2 = inner_rel->rows;
2794 9593 : isdefault2 = false;
2795 : }
2796 :
2797 11408 : if (have_mcvs1 && have_mcvs2)
2798 519 : {
2799 : /*
2800 : * We have most-common-value lists for both relations. Run through
2801 : * the lists to see which MCVs actually join to each other with the
2802 : * given operator. This allows us to determine the exact join
2803 : * selectivity for the portion of the relations represented by the MCV
2804 : * lists. We still have to estimate for the remaining population, but
2805 : * in a skewed distribution this gives us a big leg up in accuracy.
2806 : */
2807 519 : double nullfrac1 = stats1->stanullfrac;
2808 : double matchprodfreq,
2809 : matchfreq1,
2810 : uncertainfrac,
2811 : uncertain;
2812 : int i,
2813 : nmatches,
2814 : clamped_nvalues2;
2815 :
2816 : /*
2817 : * The clamping above could have resulted in nd2 being less than
2818 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2819 : * most common values in the relation will appear in the join input,
2820 : * and so compare to only the first nd2 members of the MCV list. Of
2821 : * course this is frequently wrong, but it's the best bet we can make.
2822 : */
2823 519 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2824 :
2825 : /*
2826 : * If we did not set clamped_nvalues2 to less than sslot2->nvalues,
2827 : * then the hasmatch1[] and hasmatch2[] match flags computed by
2828 : * eqjoinsel_inner are still perfectly applicable, so we need not
2829 : * re-do the matching work. Note that it does not matter if
2830 : * op_is_reversed: we'd get the same answers.
2831 : *
2832 : * If we did clamp, then a different set of sslot2 values is to be
2833 : * compared, so we have to re-do the matching.
2834 : */
2835 519 : if (clamped_nvalues2 != sslot2->nvalues)
2836 : {
2837 : /* Must re-zero the arrays */
2838 0 : memset(hasmatch1, 0, sslot1->nvalues * sizeof(bool));
2839 0 : memset(hasmatch2, 0, clamped_nvalues2 * sizeof(bool));
2840 : /* Re-fill the match arrays */
2841 0 : eqjoinsel_find_matches(eqproc, collation,
2842 : hashLeft, hashRight,
2843 : op_is_reversed,
2844 : sslot1, sslot2,
2845 : sslot1->nvalues, clamped_nvalues2,
2846 : hasmatch1, hasmatch2,
2847 : p_nmatches, &matchprodfreq);
2848 : }
2849 519 : nmatches = *p_nmatches;
2850 :
2851 : /* Sum up frequencies of matched MCVs */
2852 519 : matchfreq1 = 0.0;
2853 11531 : for (i = 0; i < sslot1->nvalues; i++)
2854 : {
2855 11012 : if (hasmatch1[i])
2856 9529 : matchfreq1 += sslot1->numbers[i];
2857 : }
2858 519 : CLAMP_PROBABILITY(matchfreq1);
2859 :
2860 : /*
2861 : * Now we need to estimate the fraction of relation 1 that has at
2862 : * least one join partner. We know for certain that the matched MCVs
2863 : * do, so that gives us a lower bound, but we're really in the dark
2864 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2865 : * assume all non-null rel1 rows have join partners, else assume for
2866 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2867 : * can discount the known-matched MCVs from the distinct-values counts
2868 : * before doing the division.
2869 : *
2870 : * Crude as the above is, it's completely useless if we don't have
2871 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2872 : * nd2 is default, punt and assume half of the uncertain rows have
2873 : * join partners.
2874 : */
2875 519 : if (!isdefault1 && !isdefault2)
2876 : {
2877 519 : nd1 -= nmatches;
2878 519 : nd2 -= nmatches;
2879 519 : if (nd1 <= nd2 || nd2 < 0)
2880 489 : uncertainfrac = 1.0;
2881 : else
2882 30 : uncertainfrac = nd2 / nd1;
2883 : }
2884 : else
2885 0 : uncertainfrac = 0.5;
2886 519 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2887 519 : CLAMP_PROBABILITY(uncertain);
2888 519 : selec = matchfreq1 + uncertainfrac * uncertain;
2889 : }
2890 : else
2891 : {
2892 : /*
2893 : * Without MCV lists for both sides, we can only use the heuristic
2894 : * about nd1 vs nd2.
2895 : */
2896 10889 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2897 :
2898 10889 : if (!isdefault1 && !isdefault2)
2899 : {
2900 8075 : if (nd1 <= nd2 || nd2 < 0)
2901 4221 : selec = 1.0 - nullfrac1;
2902 : else
2903 3854 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2904 : }
2905 : else
2906 2814 : selec = 0.5 * (1.0 - nullfrac1);
2907 : }
2908 :
2909 11408 : return selec;
2910 : }
2911 :
2912 : /*
2913 : * Identify matching MCVs for eqjoinsel_inner or eqjoinsel_semi.
2914 : *
2915 : * Inputs:
2916 : * eqproc: FmgrInfo for equality function to use (might be reversed)
2917 : * collation: OID of collation to use
2918 : * hashLeft, hashRight: OIDs of hash functions associated with equality op,
2919 : * or InvalidOid if we're not to use hashing
2920 : * op_is_reversed: indicates that eqproc compares right type to left type
2921 : * sslot1, sslot2: MCV values for the lefthand and righthand inputs
2922 : * nvalues1, nvalues2: number of values to be considered (can be less than
2923 : * sslotN->nvalues, but not more)
2924 : * Outputs:
2925 : * hasmatch1[], hasmatch2[]: pre-zeroed arrays of lengths nvalues1, nvalues2;
2926 : * entries are set to true if that MCV has a match on the other side
2927 : * *p_nmatches: receives number of MCV pairs that match
2928 : * *p_matchprodfreq: receives sum(sslot1->numbers[i] * sslot2->numbers[j])
2929 : * for matching MCVs
2930 : *
2931 : * Note that hashLeft is for the eqproc's left-hand input type, hashRight
2932 : * for its right, regardless of op_is_reversed.
2933 : *
2934 : * Note we assume that each MCV will match at most one member of the other
2935 : * MCV list. If the operator isn't really equality, there could be multiple
2936 : * matches --- but we don't look for them, both for speed and because the
2937 : * math wouldn't add up...
2938 : */
2939 : static void
2940 22554 : eqjoinsel_find_matches(FmgrInfo *eqproc, Oid collation,
2941 : Oid hashLeft, Oid hashRight,
2942 : bool op_is_reversed,
2943 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2944 : int nvalues1, int nvalues2,
2945 : bool *hasmatch1, bool *hasmatch2,
2946 : int *p_nmatches, double *p_matchprodfreq)
2947 : {
2948 22554 : LOCAL_FCINFO(fcinfo, 2);
2949 22554 : double matchprodfreq = 0.0;
2950 22554 : int nmatches = 0;
2951 :
2952 : /*
2953 : * Save a few cycles by setting up the fcinfo struct just once. Using
2954 : * FunctionCallInvoke directly also avoids failure if the eqproc returns
2955 : * NULL, though really equality functions should never do that.
2956 : */
2957 22554 : InitFunctionCallInfoData(*fcinfo, eqproc, 2, collation,
2958 : NULL, NULL);
2959 22554 : fcinfo->args[0].isnull = false;
2960 22554 : fcinfo->args[1].isnull = false;
2961 :
2962 22554 : if (OidIsValid(hashLeft) && OidIsValid(hashRight))
2963 1288 : {
2964 : /* Use a hash table to speed up the matching */
2965 1288 : LOCAL_FCINFO(hash_fcinfo, 1);
2966 : FmgrInfo hash_proc;
2967 : MCVHashContext hashContext;
2968 : MCVHashTable_hash *hashTable;
2969 : AttStatsSlot *statsProbe;
2970 : AttStatsSlot *statsHash;
2971 : bool *hasMatchProbe;
2972 : bool *hasMatchHash;
2973 : int nvaluesProbe;
2974 : int nvaluesHash;
2975 :
2976 : /* Make sure we build the hash table on the smaller array. */
2977 1288 : if (sslot1->nvalues >= sslot2->nvalues)
2978 : {
2979 1288 : statsProbe = sslot1;
2980 1288 : statsHash = sslot2;
2981 1288 : hasMatchProbe = hasmatch1;
2982 1288 : hasMatchHash = hasmatch2;
2983 1288 : nvaluesProbe = nvalues1;
2984 1288 : nvaluesHash = nvalues2;
2985 : }
2986 : else
2987 : {
2988 : /* We'll have to reverse the direction of use of the operator. */
2989 0 : op_is_reversed = !op_is_reversed;
2990 0 : statsProbe = sslot2;
2991 0 : statsHash = sslot1;
2992 0 : hasMatchProbe = hasmatch2;
2993 0 : hasMatchHash = hasmatch1;
2994 0 : nvaluesProbe = nvalues2;
2995 0 : nvaluesHash = nvalues1;
2996 : }
2997 :
2998 : /*
2999 : * Build the hash table on the smaller array, using the appropriate
3000 : * hash function for its data type.
3001 : */
3002 1288 : fmgr_info(op_is_reversed ? hashLeft : hashRight, &hash_proc);
3003 1288 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
3004 : NULL, NULL);
3005 1288 : hash_fcinfo->args[0].isnull = false;
3006 :
3007 1288 : hashContext.equal_fcinfo = fcinfo;
3008 1288 : hashContext.hash_fcinfo = hash_fcinfo;
3009 1288 : hashContext.op_is_reversed = op_is_reversed;
3010 1288 : hashContext.insert_mode = true;
3011 1288 : get_typlenbyval(statsHash->valuetype,
3012 : &hashContext.hash_typlen,
3013 : &hashContext.hash_typbyval);
3014 :
3015 1288 : hashTable = MCVHashTable_create(CurrentMemoryContext,
3016 : nvaluesHash,
3017 : &hashContext);
3018 :
3019 130088 : for (int i = 0; i < nvaluesHash; i++)
3020 : {
3021 128800 : bool found = false;
3022 128800 : MCVHashEntry *entry = MCVHashTable_insert(hashTable,
3023 128800 : statsHash->values[i],
3024 : &found);
3025 :
3026 : /*
3027 : * MCVHashTable_insert will only report "found" if the new value
3028 : * is equal to some previous one per datum_image_eq(). That
3029 : * probably shouldn't happen, since we're not expecting duplicates
3030 : * in the MCV list. If we do find a dup, just ignore it, leaving
3031 : * the hash entry's index pointing at the first occurrence. That
3032 : * matches the behavior that the non-hashed code path would have.
3033 : */
3034 128800 : if (likely(!found))
3035 128800 : entry->index = i;
3036 : }
3037 :
3038 : /*
3039 : * Prepare to probe the hash table. If the probe values are of a
3040 : * different data type, then we need to change hash functions. (This
3041 : * code relies on the assumption that since we defined SH_STORE_HASH,
3042 : * simplehash.h will never need to compute hash values for existing
3043 : * hash table entries.)
3044 : */
3045 1288 : hashContext.insert_mode = false;
3046 1288 : if (hashLeft != hashRight)
3047 : {
3048 0 : fmgr_info(op_is_reversed ? hashRight : hashLeft, &hash_proc);
3049 : /* Resetting hash_fcinfo is probably unnecessary, but be safe */
3050 0 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
3051 : NULL, NULL);
3052 0 : hash_fcinfo->args[0].isnull = false;
3053 : }
3054 :
3055 : /* Look up each probe value in turn. */
3056 130088 : for (int i = 0; i < nvaluesProbe; i++)
3057 : {
3058 128800 : MCVHashEntry *entry = MCVHashTable_lookup(hashTable,
3059 128800 : statsProbe->values[i]);
3060 :
3061 : /* As in the other code path, skip already-matched hash entries */
3062 128800 : if (entry != NULL && !hasMatchHash[entry->index])
3063 : {
3064 54186 : hasMatchHash[entry->index] = hasMatchProbe[i] = true;
3065 54186 : nmatches++;
3066 54186 : matchprodfreq += statsHash->numbers[entry->index] * statsProbe->numbers[i];
3067 : }
3068 : }
3069 :
3070 1288 : MCVHashTable_destroy(hashTable);
3071 : }
3072 : else
3073 : {
3074 : /* We're not to use hashing, so do it the O(N^2) way */
3075 : int index1,
3076 : index2;
3077 :
3078 : /* Set up to supply the values in the order the operator expects */
3079 21266 : if (op_is_reversed)
3080 : {
3081 0 : index1 = 1;
3082 0 : index2 = 0;
3083 : }
3084 : else
3085 : {
3086 21266 : index1 = 0;
3087 21266 : index2 = 1;
3088 : }
3089 :
3090 396505 : for (int i = 0; i < nvalues1; i++)
3091 : {
3092 375239 : fcinfo->args[index1].value = sslot1->values[i];
3093 :
3094 8000408 : for (int j = 0; j < nvalues2; j++)
3095 : {
3096 : Datum fresult;
3097 :
3098 7805736 : if (hasmatch2[j])
3099 2471443 : continue;
3100 5334293 : fcinfo->args[index2].value = sslot2->values[j];
3101 5334293 : fcinfo->isnull = false;
3102 5334293 : fresult = FunctionCallInvoke(fcinfo);
3103 5334293 : if (!fcinfo->isnull && DatumGetBool(fresult))
3104 : {
3105 180567 : hasmatch1[i] = hasmatch2[j] = true;
3106 180567 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
3107 180567 : nmatches++;
3108 180567 : break;
3109 : }
3110 : }
3111 : }
3112 : }
3113 :
3114 22554 : *p_nmatches = nmatches;
3115 22554 : *p_matchprodfreq = matchprodfreq;
3116 22554 : }
3117 :
3118 : /*
3119 : * Support functions for the hash tables used by eqjoinsel_find_matches
3120 : */
3121 : static uint32
3122 257600 : hash_mcv(MCVHashTable_hash *tab, Datum key)
3123 : {
3124 257600 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3125 257600 : FunctionCallInfo fcinfo = context->hash_fcinfo;
3126 : Datum fresult;
3127 :
3128 257600 : fcinfo->args[0].value = key;
3129 257600 : fcinfo->isnull = false;
3130 257600 : fresult = FunctionCallInvoke(fcinfo);
3131 : Assert(!fcinfo->isnull);
3132 257600 : return DatumGetUInt32(fresult);
3133 : }
3134 :
3135 : static bool
3136 54186 : mcvs_equal(MCVHashTable_hash *tab, Datum key0, Datum key1)
3137 : {
3138 54186 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3139 :
3140 54186 : if (context->insert_mode)
3141 : {
3142 : /*
3143 : * During the insertion step, any comparisons will be between two
3144 : * Datums of the hash table's data type, so if the given operator is
3145 : * cross-type it will be the wrong thing to use. Fortunately, we can
3146 : * use datum_image_eq instead. The MCV values should all be distinct
3147 : * anyway, so it's mostly pro-forma to compare them at all.
3148 : */
3149 0 : return datum_image_eq(key0, key1,
3150 0 : context->hash_typbyval, context->hash_typlen);
3151 : }
3152 : else
3153 : {
3154 54186 : FunctionCallInfo fcinfo = context->equal_fcinfo;
3155 : Datum fresult;
3156 :
3157 : /*
3158 : * Apply the operator the correct way around. Although simplehash.h
3159 : * doesn't document this explicitly, during lookups key0 is from the
3160 : * hash table while key1 is the probe value, so we should compare them
3161 : * in that order only if op_is_reversed.
3162 : */
3163 54186 : if (context->op_is_reversed)
3164 : {
3165 0 : fcinfo->args[0].value = key0;
3166 0 : fcinfo->args[1].value = key1;
3167 : }
3168 : else
3169 : {
3170 54186 : fcinfo->args[0].value = key1;
3171 54186 : fcinfo->args[1].value = key0;
3172 : }
3173 54186 : fcinfo->isnull = false;
3174 54186 : fresult = FunctionCallInvoke(fcinfo);
3175 54186 : return (!fcinfo->isnull && DatumGetBool(fresult));
3176 : }
3177 : }
3178 :
3179 : /*
3180 : * neqjoinsel - Join selectivity of "!="
3181 : */
3182 : Datum
3183 2938 : neqjoinsel(PG_FUNCTION_ARGS)
3184 : {
3185 2938 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3186 2938 : Oid operator = PG_GETARG_OID(1);
3187 2938 : List *args = (List *) PG_GETARG_POINTER(2);
3188 2938 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
3189 2938 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
3190 2938 : Oid collation = PG_GET_COLLATION();
3191 : float8 result;
3192 :
3193 2938 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
3194 1015 : {
3195 : /*
3196 : * For semi-joins, if there is more than one distinct value in the RHS
3197 : * relation then every non-null LHS row must find a row to join since
3198 : * it can only be equal to one of them. We'll assume that there is
3199 : * always more than one distinct RHS value for the sake of stability,
3200 : * though in theory we could have special cases for empty RHS
3201 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
3202 : * fraction of LHS that has the same value as the single RHS value).
3203 : *
3204 : * For anti-joins, if we use the same assumption that there is more
3205 : * than one distinct key in the RHS relation, then every non-null LHS
3206 : * row must be suppressed by the anti-join.
3207 : *
3208 : * So either way, the selectivity estimate should be 1 - nullfrac.
3209 : */
3210 : VariableStatData leftvar;
3211 : VariableStatData rightvar;
3212 : bool reversed;
3213 : HeapTuple statsTuple;
3214 : double nullfrac;
3215 :
3216 1015 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
3217 1015 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
3218 1015 : if (HeapTupleIsValid(statsTuple))
3219 838 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
3220 : else
3221 177 : nullfrac = 0.0;
3222 1015 : ReleaseVariableStats(leftvar);
3223 1015 : ReleaseVariableStats(rightvar);
3224 :
3225 1015 : result = 1.0 - nullfrac;
3226 : }
3227 : else
3228 : {
3229 : /*
3230 : * We want 1 - eqjoinsel() where the equality operator is the one
3231 : * associated with this != operator, that is, its negator.
3232 : */
3233 1923 : Oid eqop = get_negator(operator);
3234 :
3235 1923 : if (eqop)
3236 : {
3237 : result =
3238 1923 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
3239 : collation,
3240 : PointerGetDatum(root),
3241 : ObjectIdGetDatum(eqop),
3242 : PointerGetDatum(args),
3243 : Int16GetDatum(jointype),
3244 : PointerGetDatum(sjinfo)));
3245 : }
3246 : else
3247 : {
3248 : /* Use default selectivity (should we raise an error instead?) */
3249 0 : result = DEFAULT_EQ_SEL;
3250 : }
3251 1923 : result = 1.0 - result;
3252 : }
3253 :
3254 2938 : PG_RETURN_FLOAT8(result);
3255 : }
3256 :
3257 : /*
3258 : * scalarltjoinsel - Join selectivity of "<" for scalars
3259 : */
3260 : Datum
3261 270 : scalarltjoinsel(PG_FUNCTION_ARGS)
3262 : {
3263 270 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3264 : }
3265 :
3266 : /*
3267 : * scalarlejoinsel - Join selectivity of "<=" for scalars
3268 : */
3269 : Datum
3270 198 : scalarlejoinsel(PG_FUNCTION_ARGS)
3271 : {
3272 198 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3273 : }
3274 :
3275 : /*
3276 : * scalargtjoinsel - Join selectivity of ">" for scalars
3277 : */
3278 : Datum
3279 240 : scalargtjoinsel(PG_FUNCTION_ARGS)
3280 : {
3281 240 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3282 : }
3283 :
3284 : /*
3285 : * scalargejoinsel - Join selectivity of ">=" for scalars
3286 : */
3287 : Datum
3288 152 : scalargejoinsel(PG_FUNCTION_ARGS)
3289 : {
3290 152 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3291 : }
3292 :
3293 :
3294 : /*
3295 : * mergejoinscansel - Scan selectivity of merge join.
3296 : *
3297 : * A merge join will stop as soon as it exhausts either input stream.
3298 : * Therefore, if we can estimate the ranges of both input variables,
3299 : * we can estimate how much of the input will actually be read. This
3300 : * can have a considerable impact on the cost when using indexscans.
3301 : *
3302 : * Also, we can estimate how much of each input has to be read before the
3303 : * first join pair is found, which will affect the join's startup time.
3304 : *
3305 : * clause should be a clause already known to be mergejoinable. opfamily,
3306 : * cmptype, and nulls_first specify the sort ordering being used.
3307 : *
3308 : * The outputs are:
3309 : * *leftstart is set to the fraction of the left-hand variable expected
3310 : * to be scanned before the first join pair is found (0 to 1).
3311 : * *leftend is set to the fraction of the left-hand variable expected
3312 : * to be scanned before the join terminates (0 to 1).
3313 : * *rightstart, *rightend similarly for the right-hand variable.
3314 : */
3315 : void
3316 102851 : mergejoinscansel(PlannerInfo *root, Node *clause,
3317 : Oid opfamily, CompareType cmptype, bool nulls_first,
3318 : Selectivity *leftstart, Selectivity *leftend,
3319 : Selectivity *rightstart, Selectivity *rightend)
3320 : {
3321 : Node *left,
3322 : *right;
3323 : VariableStatData leftvar,
3324 : rightvar;
3325 : Oid opmethod;
3326 : int op_strategy;
3327 : Oid op_lefttype;
3328 : Oid op_righttype;
3329 : Oid opno,
3330 : collation,
3331 : lsortop,
3332 : rsortop,
3333 : lstatop,
3334 : rstatop,
3335 : ltop,
3336 : leop,
3337 : revltop,
3338 : revleop;
3339 : StrategyNumber ltstrat,
3340 : lestrat,
3341 : gtstrat,
3342 : gestrat;
3343 : bool isgt;
3344 : Datum leftmin,
3345 : leftmax,
3346 : rightmin,
3347 : rightmax;
3348 : double selec;
3349 :
3350 : /* Set default results if we can't figure anything out. */
3351 : /* XXX should default "start" fraction be a bit more than 0? */
3352 102851 : *leftstart = *rightstart = 0.0;
3353 102851 : *leftend = *rightend = 1.0;
3354 :
3355 : /* Deconstruct the merge clause */
3356 102851 : if (!is_opclause(clause))
3357 0 : return; /* shouldn't happen */
3358 102851 : opno = ((OpExpr *) clause)->opno;
3359 102851 : collation = ((OpExpr *) clause)->inputcollid;
3360 102851 : left = get_leftop((Expr *) clause);
3361 102851 : right = get_rightop((Expr *) clause);
3362 102851 : if (!right)
3363 0 : return; /* shouldn't happen */
3364 :
3365 : /* Look for stats for the inputs */
3366 102851 : examine_variable(root, left, 0, &leftvar);
3367 102851 : examine_variable(root, right, 0, &rightvar);
3368 :
3369 102851 : opmethod = get_opfamily_method(opfamily);
3370 :
3371 : /* Extract the operator's declared left/right datatypes */
3372 102851 : get_op_opfamily_properties(opno, opfamily, false,
3373 : &op_strategy,
3374 : &op_lefttype,
3375 : &op_righttype);
3376 : Assert(IndexAmTranslateStrategy(op_strategy, opmethod, opfamily, true) == COMPARE_EQ);
3377 :
3378 : /*
3379 : * Look up the various operators we need. If we don't find them all, it
3380 : * probably means the opfamily is broken, but we just fail silently.
3381 : *
3382 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3383 : * operator, regardless of which sort direction we are considering.
3384 : */
3385 102851 : switch (cmptype)
3386 : {
3387 102822 : case COMPARE_LT:
3388 102822 : isgt = false;
3389 102822 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3390 102822 : lestrat = IndexAmTranslateCompareType(COMPARE_LE, opmethod, opfamily, true);
3391 102822 : if (op_lefttype == op_righttype)
3392 : {
3393 : /* easy case */
3394 101652 : ltop = get_opfamily_member(opfamily,
3395 : op_lefttype, op_righttype,
3396 : ltstrat);
3397 101652 : leop = get_opfamily_member(opfamily,
3398 : op_lefttype, op_righttype,
3399 : lestrat);
3400 101652 : lsortop = ltop;
3401 101652 : rsortop = ltop;
3402 101652 : lstatop = lsortop;
3403 101652 : rstatop = rsortop;
3404 101652 : revltop = ltop;
3405 101652 : revleop = leop;
3406 : }
3407 : else
3408 : {
3409 1170 : ltop = get_opfamily_member(opfamily,
3410 : op_lefttype, op_righttype,
3411 : ltstrat);
3412 1170 : leop = get_opfamily_member(opfamily,
3413 : op_lefttype, op_righttype,
3414 : lestrat);
3415 1170 : lsortop = get_opfamily_member(opfamily,
3416 : op_lefttype, op_lefttype,
3417 : ltstrat);
3418 1170 : rsortop = get_opfamily_member(opfamily,
3419 : op_righttype, op_righttype,
3420 : ltstrat);
3421 1170 : lstatop = lsortop;
3422 1170 : rstatop = rsortop;
3423 1170 : revltop = get_opfamily_member(opfamily,
3424 : op_righttype, op_lefttype,
3425 : ltstrat);
3426 1170 : revleop = get_opfamily_member(opfamily,
3427 : op_righttype, op_lefttype,
3428 : lestrat);
3429 : }
3430 102822 : break;
3431 29 : case COMPARE_GT:
3432 : /* descending-order case */
3433 29 : isgt = true;
3434 29 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3435 29 : gtstrat = IndexAmTranslateCompareType(COMPARE_GT, opmethod, opfamily, true);
3436 29 : gestrat = IndexAmTranslateCompareType(COMPARE_GE, opmethod, opfamily, true);
3437 29 : if (op_lefttype == op_righttype)
3438 : {
3439 : /* easy case */
3440 29 : ltop = get_opfamily_member(opfamily,
3441 : op_lefttype, op_righttype,
3442 : gtstrat);
3443 29 : leop = get_opfamily_member(opfamily,
3444 : op_lefttype, op_righttype,
3445 : gestrat);
3446 29 : lsortop = ltop;
3447 29 : rsortop = ltop;
3448 29 : lstatop = get_opfamily_member(opfamily,
3449 : op_lefttype, op_lefttype,
3450 : ltstrat);
3451 29 : rstatop = lstatop;
3452 29 : revltop = ltop;
3453 29 : revleop = leop;
3454 : }
3455 : else
3456 : {
3457 0 : ltop = get_opfamily_member(opfamily,
3458 : op_lefttype, op_righttype,
3459 : gtstrat);
3460 0 : leop = get_opfamily_member(opfamily,
3461 : op_lefttype, op_righttype,
3462 : gestrat);
3463 0 : lsortop = get_opfamily_member(opfamily,
3464 : op_lefttype, op_lefttype,
3465 : gtstrat);
3466 0 : rsortop = get_opfamily_member(opfamily,
3467 : op_righttype, op_righttype,
3468 : gtstrat);
3469 0 : lstatop = get_opfamily_member(opfamily,
3470 : op_lefttype, op_lefttype,
3471 : ltstrat);
3472 0 : rstatop = get_opfamily_member(opfamily,
3473 : op_righttype, op_righttype,
3474 : ltstrat);
3475 0 : revltop = get_opfamily_member(opfamily,
3476 : op_righttype, op_lefttype,
3477 : gtstrat);
3478 0 : revleop = get_opfamily_member(opfamily,
3479 : op_righttype, op_lefttype,
3480 : gestrat);
3481 : }
3482 29 : break;
3483 0 : default:
3484 0 : goto fail; /* shouldn't get here */
3485 : }
3486 :
3487 102851 : if (!OidIsValid(lsortop) ||
3488 102851 : !OidIsValid(rsortop) ||
3489 102851 : !OidIsValid(lstatop) ||
3490 102851 : !OidIsValid(rstatop) ||
3491 102841 : !OidIsValid(ltop) ||
3492 102841 : !OidIsValid(leop) ||
3493 102841 : !OidIsValid(revltop) ||
3494 : !OidIsValid(revleop))
3495 10 : goto fail; /* insufficient info in catalogs */
3496 :
3497 : /* Try to get ranges of both inputs */
3498 102841 : if (!isgt)
3499 : {
3500 102812 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3501 : &leftmin, &leftmax))
3502 32240 : goto fail; /* no range available from stats */
3503 70572 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3504 : &rightmin, &rightmax))
3505 17687 : goto fail; /* no range available from stats */
3506 : }
3507 : else
3508 : {
3509 : /* need to swap the max and min */
3510 29 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3511 : &leftmax, &leftmin))
3512 24 : goto fail; /* no range available from stats */
3513 5 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3514 : &rightmax, &rightmin))
3515 0 : goto fail; /* no range available from stats */
3516 : }
3517 :
3518 : /*
3519 : * Now, the fraction of the left variable that will be scanned is the
3520 : * fraction that's <= the right-side maximum value. But only believe
3521 : * non-default estimates, else stick with our 1.0.
3522 : */
3523 52890 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3524 : rightmax, op_righttype);
3525 52890 : if (selec != DEFAULT_INEQ_SEL)
3526 52886 : *leftend = selec;
3527 :
3528 : /* And similarly for the right variable. */
3529 52890 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3530 : leftmax, op_lefttype);
3531 52890 : if (selec != DEFAULT_INEQ_SEL)
3532 52890 : *rightend = selec;
3533 :
3534 : /*
3535 : * Only one of the two "end" fractions can really be less than 1.0;
3536 : * believe the smaller estimate and reset the other one to exactly 1.0. If
3537 : * we get exactly equal estimates (as can easily happen with self-joins),
3538 : * believe neither.
3539 : */
3540 52890 : if (*leftend > *rightend)
3541 16396 : *leftend = 1.0;
3542 36494 : else if (*leftend < *rightend)
3543 20350 : *rightend = 1.0;
3544 : else
3545 16144 : *leftend = *rightend = 1.0;
3546 :
3547 : /*
3548 : * Also, the fraction of the left variable that will be scanned before the
3549 : * first join pair is found is the fraction that's < the right-side
3550 : * minimum value. But only believe non-default estimates, else stick with
3551 : * our own default.
3552 : */
3553 52890 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3554 : rightmin, op_righttype);
3555 52890 : if (selec != DEFAULT_INEQ_SEL)
3556 52890 : *leftstart = selec;
3557 :
3558 : /* And similarly for the right variable. */
3559 52890 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3560 : leftmin, op_lefttype);
3561 52890 : if (selec != DEFAULT_INEQ_SEL)
3562 52890 : *rightstart = selec;
3563 :
3564 : /*
3565 : * Only one of the two "start" fractions can really be more than zero;
3566 : * believe the larger estimate and reset the other one to exactly 0.0. If
3567 : * we get exactly equal estimates (as can easily happen with self-joins),
3568 : * believe neither.
3569 : */
3570 52890 : if (*leftstart < *rightstart)
3571 10669 : *leftstart = 0.0;
3572 42221 : else if (*leftstart > *rightstart)
3573 15362 : *rightstart = 0.0;
3574 : else
3575 26859 : *leftstart = *rightstart = 0.0;
3576 :
3577 : /*
3578 : * If the sort order is nulls-first, we're going to have to skip over any
3579 : * nulls too. These would not have been counted by scalarineqsel, and we
3580 : * can safely add in this fraction regardless of whether we believe
3581 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3582 : */
3583 52890 : if (nulls_first)
3584 : {
3585 : Form_pg_statistic stats;
3586 :
3587 5 : if (HeapTupleIsValid(leftvar.statsTuple))
3588 : {
3589 5 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3590 5 : *leftstart += stats->stanullfrac;
3591 5 : CLAMP_PROBABILITY(*leftstart);
3592 5 : *leftend += stats->stanullfrac;
3593 5 : CLAMP_PROBABILITY(*leftend);
3594 : }
3595 5 : if (HeapTupleIsValid(rightvar.statsTuple))
3596 : {
3597 5 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3598 5 : *rightstart += stats->stanullfrac;
3599 5 : CLAMP_PROBABILITY(*rightstart);
3600 5 : *rightend += stats->stanullfrac;
3601 5 : CLAMP_PROBABILITY(*rightend);
3602 : }
3603 : }
3604 :
3605 : /* Disbelieve start >= end, just in case that can happen */
3606 52890 : if (*leftstart >= *leftend)
3607 : {
3608 117 : *leftstart = 0.0;
3609 117 : *leftend = 1.0;
3610 : }
3611 52890 : if (*rightstart >= *rightend)
3612 : {
3613 784 : *rightstart = 0.0;
3614 784 : *rightend = 1.0;
3615 : }
3616 :
3617 52106 : fail:
3618 102851 : ReleaseVariableStats(leftvar);
3619 102851 : ReleaseVariableStats(rightvar);
3620 : }
3621 :
3622 :
3623 : /*
3624 : * matchingsel -- generic matching-operator selectivity support
3625 : *
3626 : * Use these for any operators that (a) are on data types for which we collect
3627 : * standard statistics, and (b) have behavior for which the default estimate
3628 : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3629 : * operators.
3630 : */
3631 :
3632 : Datum
3633 845 : matchingsel(PG_FUNCTION_ARGS)
3634 : {
3635 845 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3636 845 : Oid operator = PG_GETARG_OID(1);
3637 845 : List *args = (List *) PG_GETARG_POINTER(2);
3638 845 : int varRelid = PG_GETARG_INT32(3);
3639 845 : Oid collation = PG_GET_COLLATION();
3640 : double selec;
3641 :
3642 : /* Use generic restriction selectivity logic. */
3643 845 : selec = generic_restriction_selectivity(root, operator, collation,
3644 : args, varRelid,
3645 : DEFAULT_MATCHING_SEL);
3646 :
3647 845 : PG_RETURN_FLOAT8((float8) selec);
3648 : }
3649 :
3650 : Datum
3651 5 : matchingjoinsel(PG_FUNCTION_ARGS)
3652 : {
3653 : /* Just punt, for the moment. */
3654 5 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3655 : }
3656 :
3657 :
3658 : /*
3659 : * Helper routine for estimate_num_groups: add an item to a list of
3660 : * GroupVarInfos, but only if it's not known equal to any of the existing
3661 : * entries.
3662 : */
3663 : typedef struct
3664 : {
3665 : Node *var; /* might be an expression, not just a Var */
3666 : RelOptInfo *rel; /* relation it belongs to */
3667 : double ndistinct; /* # distinct values */
3668 : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3669 : } GroupVarInfo;
3670 :
3671 : static List *
3672 282490 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3673 : Node *var, VariableStatData *vardata)
3674 : {
3675 : GroupVarInfo *varinfo;
3676 : double ndistinct;
3677 : bool isdefault;
3678 : ListCell *lc;
3679 :
3680 282490 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3681 :
3682 : /*
3683 : * The nullingrels bits within the var could cause the same var to be
3684 : * counted multiple times if it's marked with different nullingrels. They
3685 : * could also prevent us from matching the var to the expressions in
3686 : * extended statistics (see estimate_multivariate_ndistinct). So strip
3687 : * them out first.
3688 : */
3689 282490 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3690 :
3691 348626 : foreach(lc, varinfos)
3692 : {
3693 66915 : varinfo = (GroupVarInfo *) lfirst(lc);
3694 :
3695 : /* Drop exact duplicates */
3696 66915 : if (equal(var, varinfo->var))
3697 779 : return varinfos;
3698 :
3699 : /*
3700 : * Drop known-equal vars, but only if they belong to different
3701 : * relations (see comments for estimate_num_groups). We aren't too
3702 : * fussy about the semantics of "equal" here.
3703 : */
3704 70520 : if (vardata->rel != varinfo->rel &&
3705 4190 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3706 : {
3707 210 : if (varinfo->ndistinct <= ndistinct)
3708 : {
3709 : /* Keep older item, forget new one */
3710 194 : return varinfos;
3711 : }
3712 : else
3713 : {
3714 : /* Delete the older item */
3715 16 : varinfos = foreach_delete_current(varinfos, lc);
3716 : }
3717 : }
3718 : }
3719 :
3720 281711 : varinfo = palloc_object(GroupVarInfo);
3721 :
3722 281711 : varinfo->var = var;
3723 281711 : varinfo->rel = vardata->rel;
3724 281711 : varinfo->ndistinct = ndistinct;
3725 281711 : varinfo->isdefault = isdefault;
3726 281711 : varinfos = lappend(varinfos, varinfo);
3727 281711 : return varinfos;
3728 : }
3729 :
3730 : /*
3731 : * estimate_num_groups - Estimate number of groups in a grouped query
3732 : *
3733 : * Given a query having a GROUP BY clause, estimate how many groups there
3734 : * will be --- ie, the number of distinct combinations of the GROUP BY
3735 : * expressions.
3736 : *
3737 : * This routine is also used to estimate the number of rows emitted by
3738 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3739 : * actually, we only use it for DISTINCT when there's no grouping or
3740 : * aggregation ahead of the DISTINCT.)
3741 : *
3742 : * Inputs:
3743 : * root - the query
3744 : * groupExprs - list of expressions being grouped by
3745 : * input_rows - number of rows estimated to arrive at the group/unique
3746 : * filter step
3747 : * pgset - NULL, or a List** pointing to a grouping set to filter the
3748 : * groupExprs against
3749 : *
3750 : * Outputs:
3751 : * estinfo - When passed as non-NULL, the function will set bits in the
3752 : * "flags" field in order to provide callers with additional information
3753 : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3754 : * bit if we used any default values in the estimation.
3755 : *
3756 : * Given the lack of any cross-correlation statistics in the system, it's
3757 : * impossible to do anything really trustworthy with GROUP BY conditions
3758 : * involving multiple Vars. We should however avoid assuming the worst
3759 : * case (all possible cross-product terms actually appear as groups) since
3760 : * very often the grouped-by Vars are highly correlated. Our current approach
3761 : * is as follows:
3762 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3763 : * independently of their content, and are ignored in the subsequent
3764 : * steps. This is mainly because tests like "col IS NULL" break the
3765 : * heuristic used in step 2 especially badly.
3766 : * 2. Reduce the given expressions to a list of unique Vars used. For
3767 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3768 : * It is clearly correct not to count the same Var more than once.
3769 : * It is also reasonable to treat f(x) the same as x: f() cannot
3770 : * increase the number of distinct values (unless it is volatile,
3771 : * which we consider unlikely for grouping), but it probably won't
3772 : * reduce the number of distinct values much either.
3773 : * As a special case, if a GROUP BY expression can be matched to an
3774 : * expressional index for which we have statistics, then we treat the
3775 : * whole expression as though it were just a Var.
3776 : * 3. If the list contains Vars of different relations that are known equal
3777 : * due to equivalence classes, then drop all but one of the Vars from each
3778 : * known-equal set, keeping the one with smallest estimated # of values
3779 : * (since the extra values of the others can't appear in joined rows).
3780 : * Note the reason we only consider Vars of different relations is that
3781 : * if we considered ones of the same rel, we'd be double-counting the
3782 : * restriction selectivity of the equality in the next step.
3783 : * 4. For Vars within a single source rel, we multiply together the numbers
3784 : * of values, clamp to the number of rows in the rel (divided by 10 if
3785 : * more than one Var), and then multiply by a factor based on the
3786 : * selectivity of the restriction clauses for that rel. When there's
3787 : * more than one Var, the initial product is probably too high (it's the
3788 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3789 : * helpful heuristic for not letting the estimate get out of hand. (The
3790 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3791 : * we multiply by to adjust for the restriction selectivity assumes that
3792 : * the restriction clauses are independent of the grouping, which may not
3793 : * be a valid assumption, but it's hard to do better.
3794 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3795 : * rel, and multiply the results together.
3796 : * Note that rels not containing grouped Vars are ignored completely, as are
3797 : * join clauses. Such rels cannot increase the number of groups, and we
3798 : * assume such clauses do not reduce the number either (somewhat bogus,
3799 : * but we don't have the info to do better).
3800 : */
3801 : double
3802 245375 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3803 : List **pgset, EstimationInfo *estinfo)
3804 : {
3805 245375 : List *varinfos = NIL;
3806 245375 : double srf_multiplier = 1.0;
3807 : double numdistinct;
3808 : ListCell *l;
3809 : int i;
3810 :
3811 : /* Zero the estinfo output parameter, if non-NULL */
3812 245375 : if (estinfo != NULL)
3813 204384 : memset(estinfo, 0, sizeof(EstimationInfo));
3814 :
3815 : /*
3816 : * We don't ever want to return an estimate of zero groups, as that tends
3817 : * to lead to division-by-zero and other unpleasantness. The input_rows
3818 : * estimate is usually already at least 1, but clamp it just in case it
3819 : * isn't.
3820 : */
3821 245375 : input_rows = clamp_row_est(input_rows);
3822 :
3823 : /*
3824 : * If no grouping columns, there's exactly one group. (This can't happen
3825 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3826 : * corner cases with set operations.)
3827 : */
3828 245375 : if (groupExprs == NIL || (pgset && *pgset == NIL))
3829 924 : return 1.0;
3830 :
3831 : /*
3832 : * Count groups derived from boolean grouping expressions. For other
3833 : * expressions, find the unique Vars used, treating an expression as a Var
3834 : * if we can find stats for it. For each one, record the statistical
3835 : * estimate of number of distinct values (total in its table, without
3836 : * regard for filtering).
3837 : */
3838 244451 : numdistinct = 1.0;
3839 :
3840 244451 : i = 0;
3841 526315 : foreach(l, groupExprs)
3842 : {
3843 281904 : Node *groupexpr = (Node *) lfirst(l);
3844 : double this_srf_multiplier;
3845 : VariableStatData vardata;
3846 : List *varshere;
3847 : ListCell *l2;
3848 :
3849 : /* is expression in this grouping set? */
3850 281904 : if (pgset && !list_member_int(*pgset, i++))
3851 222443 : continue;
3852 :
3853 : /*
3854 : * Set-returning functions in grouping columns are a bit problematic.
3855 : * The code below will effectively ignore their SRF nature and come up
3856 : * with a numdistinct estimate as though they were scalar functions.
3857 : * We compensate by scaling up the end result by the largest SRF
3858 : * rowcount estimate. (This will be an overestimate if the SRF
3859 : * produces multiple copies of any output value, but it seems best to
3860 : * assume the SRF's outputs are distinct. In any case, it's probably
3861 : * pointless to worry too much about this without much better
3862 : * estimates for SRF output rowcounts than we have today.)
3863 : */
3864 281236 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3865 281236 : if (srf_multiplier < this_srf_multiplier)
3866 130 : srf_multiplier = this_srf_multiplier;
3867 :
3868 : /* Short-circuit for expressions returning boolean */
3869 281236 : if (exprType(groupexpr) == BOOLOID)
3870 : {
3871 292 : numdistinct *= 2.0;
3872 292 : continue;
3873 : }
3874 :
3875 : /*
3876 : * If examine_variable is able to deduce anything about the GROUP BY
3877 : * expression, treat it as a single variable even if it's really more
3878 : * complicated.
3879 : *
3880 : * XXX This has the consequence that if there's a statistics object on
3881 : * the expression, we don't split it into individual Vars. This
3882 : * affects our selection of statistics in
3883 : * estimate_multivariate_ndistinct, because it's probably better to
3884 : * use more accurate estimate for each expression and treat them as
3885 : * independent, than to combine estimates for the extracted variables
3886 : * when we don't know how that relates to the expressions.
3887 : */
3888 280944 : examine_variable(root, groupexpr, 0, &vardata);
3889 280944 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3890 : {
3891 220985 : varinfos = add_unique_group_var(root, varinfos,
3892 : groupexpr, &vardata);
3893 220985 : ReleaseVariableStats(vardata);
3894 220985 : continue;
3895 : }
3896 59959 : ReleaseVariableStats(vardata);
3897 :
3898 : /*
3899 : * Else pull out the component Vars. Handle PlaceHolderVars by
3900 : * recursing into their arguments (effectively assuming that the
3901 : * PlaceHolderVar doesn't change the number of groups, which boils
3902 : * down to ignoring the possible addition of nulls to the result set).
3903 : */
3904 59959 : varshere = pull_var_clause(groupexpr,
3905 : PVC_RECURSE_AGGREGATES |
3906 : PVC_RECURSE_WINDOWFUNCS |
3907 : PVC_RECURSE_PLACEHOLDERS);
3908 :
3909 : /*
3910 : * If we find any variable-free GROUP BY item, then either it is a
3911 : * constant (and we can ignore it) or it contains a volatile function;
3912 : * in the latter case we punt and assume that each input row will
3913 : * yield a distinct group.
3914 : */
3915 59959 : if (varshere == NIL)
3916 : {
3917 538 : if (contain_volatile_functions(groupexpr))
3918 40 : return input_rows;
3919 498 : continue;
3920 : }
3921 :
3922 : /*
3923 : * Else add variables to varinfos list
3924 : */
3925 120926 : foreach(l2, varshere)
3926 : {
3927 61505 : Node *var = (Node *) lfirst(l2);
3928 :
3929 61505 : examine_variable(root, var, 0, &vardata);
3930 61505 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3931 61505 : ReleaseVariableStats(vardata);
3932 : }
3933 : }
3934 :
3935 : /*
3936 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3937 : * list.
3938 : */
3939 244411 : if (varinfos == NIL)
3940 : {
3941 : /* Apply SRF multiplier as we would do in the long path */
3942 385 : numdistinct *= srf_multiplier;
3943 : /* Round off */
3944 385 : numdistinct = ceil(numdistinct);
3945 : /* Guard against out-of-range answers */
3946 385 : if (numdistinct > input_rows)
3947 71 : numdistinct = input_rows;
3948 385 : if (numdistinct < 1.0)
3949 0 : numdistinct = 1.0;
3950 385 : return numdistinct;
3951 : }
3952 :
3953 : /*
3954 : * Group Vars by relation and estimate total numdistinct.
3955 : *
3956 : * For each iteration of the outer loop, we process the frontmost Var in
3957 : * varinfos, plus all other Vars in the same relation. We remove these
3958 : * Vars from the newvarinfos list for the next iteration. This is the
3959 : * easiest way to group Vars of same rel together.
3960 : */
3961 : do
3962 : {
3963 246194 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3964 246194 : RelOptInfo *rel = varinfo1->rel;
3965 246194 : double reldistinct = 1;
3966 246194 : double relmaxndistinct = reldistinct;
3967 246194 : int relvarcount = 0;
3968 246194 : List *newvarinfos = NIL;
3969 246194 : List *relvarinfos = NIL;
3970 :
3971 : /*
3972 : * Split the list of varinfos in two - one for the current rel, one
3973 : * for remaining Vars on other rels.
3974 : */
3975 246194 : relvarinfos = lappend(relvarinfos, varinfo1);
3976 284850 : for_each_from(l, varinfos, 1)
3977 : {
3978 38656 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3979 :
3980 38656 : if (varinfo2->rel == varinfo1->rel)
3981 : {
3982 : /* varinfos on current rel */
3983 35501 : relvarinfos = lappend(relvarinfos, varinfo2);
3984 : }
3985 : else
3986 : {
3987 : /* not time to process varinfo2 yet */
3988 3155 : newvarinfos = lappend(newvarinfos, varinfo2);
3989 : }
3990 : }
3991 :
3992 : /*
3993 : * Get the numdistinct estimate for the Vars of this rel. We
3994 : * iteratively search for multivariate n-distinct with maximum number
3995 : * of vars; assuming that each var group is independent of the others,
3996 : * we multiply them together. Any remaining relvarinfos after no more
3997 : * multivariate matches are found are assumed independent too, so
3998 : * their individual ndistinct estimates are multiplied also.
3999 : *
4000 : * While iterating, count how many separate numdistinct values we
4001 : * apply. We apply a fudge factor below, but only if we multiplied
4002 : * more than one such values.
4003 : */
4004 492493 : while (relvarinfos)
4005 : {
4006 : double mvndistinct;
4007 :
4008 246299 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
4009 : &mvndistinct))
4010 : {
4011 345 : reldistinct *= mvndistinct;
4012 345 : if (relmaxndistinct < mvndistinct)
4013 335 : relmaxndistinct = mvndistinct;
4014 345 : relvarcount++;
4015 : }
4016 : else
4017 : {
4018 526919 : foreach(l, relvarinfos)
4019 : {
4020 280965 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
4021 :
4022 280965 : reldistinct *= varinfo2->ndistinct;
4023 280965 : if (relmaxndistinct < varinfo2->ndistinct)
4024 249696 : relmaxndistinct = varinfo2->ndistinct;
4025 280965 : relvarcount++;
4026 :
4027 : /*
4028 : * When varinfo2's isdefault is set then we'd better set
4029 : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
4030 : */
4031 280965 : if (estinfo != NULL && varinfo2->isdefault)
4032 18073 : estinfo->flags |= SELFLAG_USED_DEFAULT;
4033 : }
4034 :
4035 : /* we're done with this relation */
4036 245954 : relvarinfos = NIL;
4037 : }
4038 : }
4039 :
4040 : /*
4041 : * Sanity check --- don't divide by zero if empty relation.
4042 : */
4043 : Assert(IS_SIMPLE_REL(rel));
4044 246194 : if (rel->tuples > 0)
4045 : {
4046 : /*
4047 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
4048 : * fudge factor is because the Vars are probably correlated but we
4049 : * don't know by how much. We should never clamp to less than the
4050 : * largest ndistinct value for any of the Vars, though, since
4051 : * there will surely be at least that many groups.
4052 : */
4053 245315 : double clamp = rel->tuples;
4054 :
4055 245315 : if (relvarcount > 1)
4056 : {
4057 30951 : clamp *= 0.1;
4058 30951 : if (clamp < relmaxndistinct)
4059 : {
4060 28806 : clamp = relmaxndistinct;
4061 : /* for sanity in case some ndistinct is too large: */
4062 28806 : if (clamp > rel->tuples)
4063 65 : clamp = rel->tuples;
4064 : }
4065 : }
4066 245315 : if (reldistinct > clamp)
4067 25266 : reldistinct = clamp;
4068 :
4069 : /*
4070 : * Update the estimate based on the restriction selectivity,
4071 : * guarding against division by zero when reldistinct is zero.
4072 : * Also skip this if we know that we are returning all rows.
4073 : */
4074 245315 : if (reldistinct > 0 && rel->rows < rel->tuples)
4075 : {
4076 : /*
4077 : * Given a table containing N rows with n distinct values in a
4078 : * uniform distribution, if we select p rows at random then
4079 : * the expected number of distinct values selected is
4080 : *
4081 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
4082 : *
4083 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
4084 : *
4085 : * See "Approximating block accesses in database
4086 : * organizations", S. B. Yao, Communications of the ACM,
4087 : * Volume 20 Issue 4, April 1977 Pages 260-261.
4088 : *
4089 : * Alternatively, re-arranging the terms from the factorials,
4090 : * this may be written as
4091 : *
4092 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
4093 : *
4094 : * This form of the formula is more efficient to compute in
4095 : * the common case where p is larger than N/n. Additionally,
4096 : * as pointed out by Dell'Era, if i << N for all terms in the
4097 : * product, it can be approximated by
4098 : *
4099 : * n * (1 - ((N-p)/N)^(N/n))
4100 : *
4101 : * See "Expected distinct values when selecting from a bag
4102 : * without replacement", Alberto Dell'Era,
4103 : * http://www.adellera.it/investigations/distinct_balls/.
4104 : *
4105 : * The condition i << N is equivalent to n >> 1, so this is a
4106 : * good approximation when the number of distinct values in
4107 : * the table is large. It turns out that this formula also
4108 : * works well even when n is small.
4109 : */
4110 72876 : reldistinct *=
4111 72876 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
4112 72876 : rel->tuples / reldistinct));
4113 : }
4114 245315 : reldistinct = clamp_row_est(reldistinct);
4115 :
4116 : /*
4117 : * Update estimate of total distinct groups.
4118 : */
4119 245315 : numdistinct *= reldistinct;
4120 : }
4121 :
4122 246194 : varinfos = newvarinfos;
4123 246194 : } while (varinfos != NIL);
4124 :
4125 : /* Now we can account for the effects of any SRFs */
4126 244026 : numdistinct *= srf_multiplier;
4127 :
4128 : /* Round off */
4129 244026 : numdistinct = ceil(numdistinct);
4130 :
4131 : /* Guard against out-of-range answers */
4132 244026 : if (numdistinct > input_rows)
4133 52690 : numdistinct = input_rows;
4134 244026 : if (numdistinct < 1.0)
4135 0 : numdistinct = 1.0;
4136 :
4137 244026 : return numdistinct;
4138 : }
4139 :
4140 : /*
4141 : * Try to estimate the bucket size of the hash join inner side when the join
4142 : * condition contains two or more clauses by employing extended statistics.
4143 : *
4144 : * The main idea of this approach is that the distinct value generated by
4145 : * multivariate estimation on two or more columns would provide less bucket size
4146 : * than estimation on one separate column.
4147 : *
4148 : * IMPORTANT: It is crucial to synchronize the approach of combining different
4149 : * estimations with the caller's method.
4150 : *
4151 : * Return a list of clauses that didn't fetch any extended statistics.
4152 : */
4153 : List *
4154 346043 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
4155 : List *hashclauses,
4156 : Selectivity *innerbucketsize)
4157 : {
4158 : List *clauses;
4159 : List *otherclauses;
4160 : double ndistinct;
4161 :
4162 346043 : if (list_length(hashclauses) <= 1)
4163 : {
4164 : /*
4165 : * Nothing to do for a single clause. Could we employ univariate
4166 : * extended stat here?
4167 : */
4168 316509 : return hashclauses;
4169 : }
4170 :
4171 : /* "clauses" is the list of hashclauses we've not dealt with yet */
4172 29534 : clauses = list_copy(hashclauses);
4173 : /* "otherclauses" holds clauses we are going to return to caller */
4174 29534 : otherclauses = NIL;
4175 : /* current estimate of ndistinct */
4176 29534 : ndistinct = 1.0;
4177 59078 : while (clauses != NIL)
4178 : {
4179 : ListCell *lc;
4180 29544 : int relid = -1;
4181 29544 : List *varinfos = NIL;
4182 29544 : List *origin_rinfos = NIL;
4183 : double mvndistinct;
4184 : List *origin_varinfos;
4185 29544 : int group_relid = -1;
4186 29544 : RelOptInfo *group_rel = NULL;
4187 : ListCell *lc1,
4188 : *lc2;
4189 :
4190 : /*
4191 : * Find clauses, referencing the same single base relation and try to
4192 : * estimate such a group with extended statistics. Create varinfo for
4193 : * an approved clause, push it to otherclauses, if it can't be
4194 : * estimated here or ignore to process at the next iteration.
4195 : */
4196 90671 : foreach(lc, clauses)
4197 : {
4198 61127 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
4199 : Node *expr;
4200 : Relids relids;
4201 : GroupVarInfo *varinfo;
4202 :
4203 : /*
4204 : * Find the inner side of the join, which we need to estimate the
4205 : * number of buckets. Use outer_is_left because the
4206 : * clause_sides_match_join routine has called on hash clauses.
4207 : */
4208 122254 : relids = rinfo->outer_is_left ?
4209 61127 : rinfo->right_relids : rinfo->left_relids;
4210 122254 : expr = rinfo->outer_is_left ?
4211 61127 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
4212 :
4213 61127 : if (bms_get_singleton_member(relids, &relid) &&
4214 59507 : root->simple_rel_array[relid]->statlist != NIL)
4215 40 : {
4216 50 : bool is_duplicate = false;
4217 :
4218 : /*
4219 : * This inner-side expression references only one relation.
4220 : * Extended statistics on this clause can exist.
4221 : */
4222 50 : if (group_relid < 0)
4223 : {
4224 25 : RangeTblEntry *rte = root->simple_rte_array[relid];
4225 :
4226 25 : if (!rte || (rte->relkind != RELKIND_RELATION &&
4227 0 : rte->relkind != RELKIND_MATVIEW &&
4228 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
4229 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
4230 : {
4231 : /* Extended statistics can't exist in principle */
4232 0 : otherclauses = lappend(otherclauses, rinfo);
4233 0 : clauses = foreach_delete_current(clauses, lc);
4234 0 : continue;
4235 : }
4236 :
4237 25 : group_relid = relid;
4238 25 : group_rel = root->simple_rel_array[relid];
4239 : }
4240 25 : else if (group_relid != relid)
4241 : {
4242 : /*
4243 : * Being in the group forming state we don't need other
4244 : * clauses.
4245 : */
4246 0 : continue;
4247 : }
4248 :
4249 : /*
4250 : * We're going to add the new clause to the varinfos list. We
4251 : * might re-use add_unique_group_var(), but we don't do so for
4252 : * two reasons.
4253 : *
4254 : * 1) We must keep the origin_rinfos list ordered exactly the
4255 : * same way as varinfos.
4256 : *
4257 : * 2) add_unique_group_var() is designed for
4258 : * estimate_num_groups(), where a larger number of groups is
4259 : * worse. While estimating the number of hash buckets, we
4260 : * have the opposite: a lesser number of groups is worse.
4261 : * Therefore, we don't have to remove "known equal" vars: the
4262 : * removed var may valuably contribute to the multivariate
4263 : * statistics to grow the number of groups.
4264 : */
4265 :
4266 : /*
4267 : * Clear nullingrels to correctly match hash keys. See
4268 : * add_unique_group_var()'s comment for details.
4269 : */
4270 50 : expr = remove_nulling_relids(expr, root->outer_join_rels, NULL);
4271 :
4272 : /*
4273 : * Detect and exclude exact duplicates from the list of hash
4274 : * keys (like add_unique_group_var does).
4275 : */
4276 70 : foreach(lc1, varinfos)
4277 : {
4278 30 : varinfo = (GroupVarInfo *) lfirst(lc1);
4279 :
4280 30 : if (!equal(expr, varinfo->var))
4281 20 : continue;
4282 :
4283 10 : is_duplicate = true;
4284 10 : break;
4285 : }
4286 :
4287 50 : if (is_duplicate)
4288 : {
4289 : /*
4290 : * Skip exact duplicates. Adding them to the otherclauses
4291 : * list also doesn't make sense.
4292 : */
4293 10 : continue;
4294 : }
4295 :
4296 : /*
4297 : * Initialize GroupVarInfo. We only use it to call
4298 : * estimate_multivariate_ndistinct(), which doesn't care about
4299 : * ndistinct and isdefault fields. Thus, skip these fields.
4300 : */
4301 40 : varinfo = palloc0_object(GroupVarInfo);
4302 40 : varinfo->var = expr;
4303 40 : varinfo->rel = root->simple_rel_array[relid];
4304 40 : varinfos = lappend(varinfos, varinfo);
4305 :
4306 : /*
4307 : * Remember the link to RestrictInfo for the case the clause
4308 : * is failed to be estimated.
4309 : */
4310 40 : origin_rinfos = lappend(origin_rinfos, rinfo);
4311 : }
4312 : else
4313 : {
4314 : /* This clause can't be estimated with extended statistics */
4315 61077 : otherclauses = lappend(otherclauses, rinfo);
4316 : }
4317 :
4318 61117 : clauses = foreach_delete_current(clauses, lc);
4319 : }
4320 :
4321 29544 : if (list_length(varinfos) < 2)
4322 : {
4323 : /*
4324 : * Multivariate statistics doesn't apply to single columns except
4325 : * for expressions, but it has not been implemented yet.
4326 : */
4327 29534 : otherclauses = list_concat(otherclauses, origin_rinfos);
4328 29534 : list_free_deep(varinfos);
4329 29534 : list_free(origin_rinfos);
4330 29534 : continue;
4331 : }
4332 :
4333 : Assert(group_rel != NULL);
4334 :
4335 : /* Employ the extended statistics. */
4336 10 : origin_varinfos = varinfos;
4337 : for (;;)
4338 10 : {
4339 20 : bool estimated = estimate_multivariate_ndistinct(root,
4340 : group_rel,
4341 : &varinfos,
4342 : &mvndistinct);
4343 :
4344 20 : if (!estimated)
4345 10 : break;
4346 :
4347 : /*
4348 : * We've got an estimation. Use ndistinct value in a consistent
4349 : * way - according to the caller's logic (see
4350 : * final_cost_hashjoin).
4351 : */
4352 10 : if (ndistinct < mvndistinct)
4353 10 : ndistinct = mvndistinct;
4354 : Assert(ndistinct >= 1.0);
4355 : }
4356 :
4357 : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
4358 :
4359 : /* Collect unmatched clauses as otherclauses. */
4360 35 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
4361 : {
4362 25 : GroupVarInfo *vinfo = lfirst(lc1);
4363 :
4364 25 : if (!list_member_ptr(varinfos, vinfo))
4365 : /* Already estimated */
4366 25 : continue;
4367 :
4368 : /* Can't be estimated here - push to the returning list */
4369 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
4370 : }
4371 : }
4372 :
4373 29534 : *innerbucketsize = 1.0 / ndistinct;
4374 29534 : return otherclauses;
4375 : }
4376 :
4377 : /*
4378 : * Estimate hash bucket statistics when the specified expression is used
4379 : * as a hash key for the given number of buckets.
4380 : *
4381 : * This attempts to determine two values:
4382 : *
4383 : * 1. The frequency of the most common value of the expression (returns
4384 : * zero into *mcv_freq if we can't get that). This will be frequency
4385 : * relative to the entire underlying table.
4386 : *
4387 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
4388 : * divided by total number of tuples to be hashed.
4389 : *
4390 : * XXX This is really pretty bogus since we're effectively assuming that the
4391 : * distribution of hash keys will be the same after applying restriction
4392 : * clauses as it was in the underlying relation. However, we are not nearly
4393 : * smart enough to figure out how the restrict clauses might change the
4394 : * distribution, so this will have to do for now.
4395 : *
4396 : * We are passed the number of buckets the executor will use for the given
4397 : * input relation. If the data were perfectly distributed, with the same
4398 : * number of tuples going into each available bucket, then the bucketsize
4399 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
4400 : * only if (a) there are at least nbuckets distinct data values, and (b)
4401 : * we have a not-too-skewed data distribution. Otherwise the buckets will
4402 : * be nonuniformly occupied. If the other relation in the join has a key
4403 : * distribution similar to this one's, then the most-loaded buckets are
4404 : * exactly those that will be probed most often. Therefore, the "average"
4405 : * bucket size for costing purposes should really be taken as something close
4406 : * to the "worst case" bucket size. We try to estimate this by adjusting the
4407 : * fraction if there are too few distinct data values, and then clamping to
4408 : * at least the bucket size implied by the most common value's frequency.
4409 : *
4410 : * If no statistics are available, use a default estimate of 0.1. This will
4411 : * discourage use of a hash rather strongly if the inner relation is large,
4412 : * which is what we want. We do not want to hash unless we know that the
4413 : * inner rel is well-dispersed (or the alternatives seem much worse).
4414 : *
4415 : * The caller should also check that the mcv_freq is not so large that the
4416 : * most common value would by itself require an impractically large bucket.
4417 : * In a hash join, the executor can split buckets if they get too big, but
4418 : * obviously that doesn't help for a bucket that contains many duplicates of
4419 : * the same value.
4420 : */
4421 : void
4422 154770 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
4423 : Selectivity *mcv_freq,
4424 : Selectivity *bucketsize_frac)
4425 : {
4426 : VariableStatData vardata;
4427 : double estfract,
4428 : ndistinct;
4429 : bool isdefault;
4430 : AttStatsSlot sslot;
4431 :
4432 154770 : examine_variable(root, hashkey, 0, &vardata);
4433 :
4434 : /* Initialize *mcv_freq to "unknown" */
4435 154770 : *mcv_freq = 0.0;
4436 :
4437 : /* Look up the frequency of the most common value, if available */
4438 154770 : if (HeapTupleIsValid(vardata.statsTuple))
4439 : {
4440 100793 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4441 : STATISTIC_KIND_MCV, InvalidOid,
4442 : ATTSTATSSLOT_NUMBERS))
4443 : {
4444 : /*
4445 : * The first MCV stat is for the most common value.
4446 : */
4447 59873 : if (sslot.nnumbers > 0)
4448 59873 : *mcv_freq = sslot.numbers[0];
4449 59873 : free_attstatsslot(&sslot);
4450 : }
4451 40920 : else if (get_attstatsslot(&sslot, vardata.statsTuple,
4452 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
4453 : 0))
4454 : {
4455 : /*
4456 : * If there are no recorded MCVs, but we do have a histogram, then
4457 : * assume that ANALYZE determined that the column is unique.
4458 : */
4459 39444 : if (vardata.rel && vardata.rel->tuples > 0)
4460 39429 : *mcv_freq = 1.0 / vardata.rel->tuples;
4461 : }
4462 : }
4463 :
4464 : /* Get number of distinct values */
4465 154770 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4466 :
4467 : /*
4468 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4469 : * mcv_freq is known to be even higher than that, use it instead.
4470 : */
4471 154770 : if (isdefault)
4472 : {
4473 24828 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4474 24828 : ReleaseVariableStats(vardata);
4475 24828 : return;
4476 : }
4477 :
4478 : /*
4479 : * Adjust ndistinct to account for restriction clauses. Observe we are
4480 : * assuming that the data distribution is affected uniformly by the
4481 : * restriction clauses!
4482 : *
4483 : * XXX Possibly better way, but much more expensive: multiply by
4484 : * selectivity of rel's restriction clauses that mention the target Var.
4485 : */
4486 129942 : if (vardata.rel && vardata.rel->tuples > 0)
4487 : {
4488 129915 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
4489 129915 : ndistinct = clamp_row_est(ndistinct);
4490 : }
4491 :
4492 : /*
4493 : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
4494 : * number of buckets is less than the expected number of distinct values;
4495 : * otherwise it is 1/ndistinct.
4496 : */
4497 129942 : if (ndistinct > nbuckets)
4498 72 : estfract = 1.0 / nbuckets;
4499 : else
4500 129870 : estfract = 1.0 / ndistinct;
4501 :
4502 : /*
4503 : * Clamp the bucketsize fraction to be not less than the MCV frequency,
4504 : * since whichever bucket the MCV values end up in will have at least that
4505 : * size. This has no effect if *mcv_freq is still zero.
4506 : */
4507 129942 : estfract = Max(estfract, *mcv_freq);
4508 :
4509 129942 : *bucketsize_frac = (Selectivity) estfract;
4510 :
4511 129942 : ReleaseVariableStats(vardata);
4512 : }
4513 :
4514 : /*
4515 : * estimate_hashagg_tablesize
4516 : * estimate the number of bytes that a hash aggregate hashtable will
4517 : * require based on the agg_costs, path width and number of groups.
4518 : *
4519 : * We return the result as "double" to forestall any possible overflow
4520 : * problem in the multiplication by dNumGroups.
4521 : *
4522 : * XXX this may be over-estimating the size now that hashagg knows to omit
4523 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4524 : * grouping columns not in the hashed set are counted here even though hashagg
4525 : * won't store them. Is this a problem?
4526 : */
4527 : double
4528 2424 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4529 : const AggClauseCosts *agg_costs, double dNumGroups)
4530 : {
4531 : Size hashentrysize;
4532 :
4533 2424 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4534 2424 : path->pathtarget->width,
4535 2424 : agg_costs->transitionSpace);
4536 :
4537 : /*
4538 : * Note that this disregards the effect of fill-factor and growth policy
4539 : * of the hash table. That's probably ok, given that the default
4540 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4541 : * "double-in-size" growth policies here.
4542 : */
4543 2424 : return hashentrysize * dNumGroups;
4544 : }
4545 :
4546 :
4547 : /*-------------------------------------------------------------------------
4548 : *
4549 : * Support routines
4550 : *
4551 : *-------------------------------------------------------------------------
4552 : */
4553 :
4554 : /*
4555 : * Find the best matching ndistinct extended statistics for the given list of
4556 : * GroupVarInfos.
4557 : *
4558 : * Callers must ensure that the given GroupVarInfos all belong to 'rel' and
4559 : * the GroupVarInfos list does not contain any duplicate Vars or expressions.
4560 : *
4561 : * When statistics are found that match > 1 of the given GroupVarInfo, the
4562 : * *ndistinct parameter is set according to the ndistinct estimate and a new
4563 : * list is built with the matching GroupVarInfos removed, which is output via
4564 : * the *varinfos parameter before returning true. When no matching stats are
4565 : * found, false is returned and the *varinfos and *ndistinct parameters are
4566 : * left untouched.
4567 : */
4568 : static bool
4569 246319 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4570 : List **varinfos, double *ndistinct)
4571 : {
4572 : ListCell *lc;
4573 : int nmatches_vars;
4574 : int nmatches_exprs;
4575 246319 : Oid statOid = InvalidOid;
4576 : MVNDistinct *stats;
4577 246319 : StatisticExtInfo *matched_info = NULL;
4578 246319 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4579 :
4580 : /* bail out immediately if the table has no extended statistics */
4581 246319 : if (!rel->statlist)
4582 245848 : return false;
4583 :
4584 : /* look for the ndistinct statistics object matching the most vars */
4585 471 : nmatches_vars = 0; /* we require at least two matches */
4586 471 : nmatches_exprs = 0;
4587 1875 : foreach(lc, rel->statlist)
4588 : {
4589 : ListCell *lc2;
4590 1404 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
4591 1404 : int nshared_vars = 0;
4592 1404 : int nshared_exprs = 0;
4593 :
4594 : /* skip statistics of other kinds */
4595 1404 : if (info->kind != STATS_EXT_NDISTINCT)
4596 663 : continue;
4597 :
4598 : /* skip statistics with mismatching stxdinherit value */
4599 741 : if (info->inherit != rte->inh)
4600 25 : continue;
4601 :
4602 : /*
4603 : * Determine how many expressions (and variables in non-matched
4604 : * expressions) match. We'll then use these numbers to pick the
4605 : * statistics object that best matches the clauses.
4606 : */
4607 2267 : foreach(lc2, *varinfos)
4608 : {
4609 : ListCell *lc3;
4610 1551 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4611 : AttrNumber attnum;
4612 :
4613 : Assert(varinfo->rel == rel);
4614 :
4615 : /* simple Var, search in statistics keys directly */
4616 1551 : if (IsA(varinfo->var, Var))
4617 : {
4618 1246 : attnum = ((Var *) varinfo->var)->varattno;
4619 :
4620 : /*
4621 : * Ignore system attributes - we don't support statistics on
4622 : * them, so can't match them (and it'd fail as the values are
4623 : * negative).
4624 : */
4625 1246 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4626 10 : continue;
4627 :
4628 1236 : if (bms_is_member(attnum, info->keys))
4629 730 : nshared_vars++;
4630 :
4631 1236 : continue;
4632 : }
4633 :
4634 : /* expression - see if it's in the statistics object */
4635 550 : foreach(lc3, info->exprs)
4636 : {
4637 440 : Node *expr = (Node *) lfirst(lc3);
4638 :
4639 440 : if (equal(varinfo->var, expr))
4640 : {
4641 195 : nshared_exprs++;
4642 195 : break;
4643 : }
4644 : }
4645 : }
4646 :
4647 : /*
4648 : * The ndistinct extended statistics contain estimates for a minimum
4649 : * of pairs of columns which the statistics are defined on and
4650 : * certainly not single columns. Here we skip unless we managed to
4651 : * match to at least two columns.
4652 : */
4653 716 : if (nshared_vars + nshared_exprs < 2)
4654 331 : continue;
4655 :
4656 : /*
4657 : * Check if these statistics are a better match than the previous best
4658 : * match and if so, take note of the StatisticExtInfo.
4659 : *
4660 : * The statslist is sorted by statOid, so the StatisticExtInfo we
4661 : * select as the best match is deterministic even when multiple sets
4662 : * of statistics match equally as well.
4663 : */
4664 385 : if ((nshared_exprs > nmatches_exprs) ||
4665 295 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4666 : {
4667 365 : statOid = info->statOid;
4668 365 : nmatches_vars = nshared_vars;
4669 365 : nmatches_exprs = nshared_exprs;
4670 365 : matched_info = info;
4671 : }
4672 : }
4673 :
4674 : /* No match? */
4675 471 : if (statOid == InvalidOid)
4676 116 : return false;
4677 :
4678 : Assert(nmatches_vars + nmatches_exprs > 1);
4679 :
4680 355 : stats = statext_ndistinct_load(statOid, rte->inh);
4681 :
4682 : /*
4683 : * If we have a match, search it for the specific item that matches (there
4684 : * must be one), and construct the output values.
4685 : */
4686 355 : if (stats)
4687 : {
4688 : int i;
4689 355 : List *newlist = NIL;
4690 355 : MVNDistinctItem *item = NULL;
4691 : ListCell *lc2;
4692 355 : Bitmapset *matched = NULL;
4693 : AttrNumber attnum_offset;
4694 :
4695 : /*
4696 : * How much we need to offset the attnums? If there are no
4697 : * expressions, no offset is needed. Otherwise offset enough to move
4698 : * the lowest one (which is equal to number of expressions) to 1.
4699 : */
4700 355 : if (matched_info->exprs)
4701 125 : attnum_offset = (list_length(matched_info->exprs) + 1);
4702 : else
4703 230 : attnum_offset = 0;
4704 :
4705 : /* see what actually matched */
4706 1240 : foreach(lc2, *varinfos)
4707 : {
4708 : ListCell *lc3;
4709 : int idx;
4710 885 : bool found = false;
4711 :
4712 885 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4713 :
4714 : /*
4715 : * Process a simple Var expression, by matching it to keys
4716 : * directly. If there's a matching expression, we'll try matching
4717 : * it later.
4718 : */
4719 885 : if (IsA(varinfo->var, Var))
4720 : {
4721 730 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4722 :
4723 : /*
4724 : * Ignore expressions on system attributes. Can't rely on the
4725 : * bms check for negative values.
4726 : */
4727 730 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4728 5 : continue;
4729 :
4730 : /* Is the variable covered by the statistics object? */
4731 725 : if (!bms_is_member(attnum, matched_info->keys))
4732 100 : continue;
4733 :
4734 625 : attnum = attnum + attnum_offset;
4735 :
4736 : /* ensure sufficient offset */
4737 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4738 :
4739 625 : matched = bms_add_member(matched, attnum);
4740 :
4741 625 : found = true;
4742 : }
4743 :
4744 : /*
4745 : * XXX Maybe we should allow searching the expressions even if we
4746 : * found an attribute matching the expression? That would handle
4747 : * trivial expressions like "(a)" but it seems fairly useless.
4748 : */
4749 780 : if (found)
4750 625 : continue;
4751 :
4752 : /* expression - see if it's in the statistics object */
4753 155 : idx = 0;
4754 255 : foreach(lc3, matched_info->exprs)
4755 : {
4756 230 : Node *expr = (Node *) lfirst(lc3);
4757 :
4758 230 : if (equal(varinfo->var, expr))
4759 : {
4760 130 : AttrNumber attnum = -(idx + 1);
4761 :
4762 130 : attnum = attnum + attnum_offset;
4763 :
4764 : /* ensure sufficient offset */
4765 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4766 :
4767 130 : matched = bms_add_member(matched, attnum);
4768 :
4769 : /* there should be just one matching expression */
4770 130 : break;
4771 : }
4772 :
4773 100 : idx++;
4774 : }
4775 : }
4776 :
4777 : /* Find the specific item that exactly matches the combination */
4778 720 : for (i = 0; i < stats->nitems; i++)
4779 : {
4780 : int j;
4781 720 : MVNDistinctItem *tmpitem = &stats->items[i];
4782 :
4783 720 : if (tmpitem->nattributes != bms_num_members(matched))
4784 135 : continue;
4785 :
4786 : /* assume it's the right item */
4787 585 : item = tmpitem;
4788 :
4789 : /* check that all item attributes/expressions fit the match */
4790 1410 : for (j = 0; j < tmpitem->nattributes; j++)
4791 : {
4792 1055 : AttrNumber attnum = tmpitem->attributes[j];
4793 :
4794 : /*
4795 : * Thanks to how we constructed the matched bitmap above, we
4796 : * can just offset all attnums the same way.
4797 : */
4798 1055 : attnum = attnum + attnum_offset;
4799 :
4800 1055 : if (!bms_is_member(attnum, matched))
4801 : {
4802 : /* nah, it's not this item */
4803 230 : item = NULL;
4804 230 : break;
4805 : }
4806 : }
4807 :
4808 : /*
4809 : * If the item has all the matched attributes, we know it's the
4810 : * right one - there can't be a better one. matching more.
4811 : */
4812 585 : if (item)
4813 355 : break;
4814 : }
4815 :
4816 : /*
4817 : * Make sure we found an item. There has to be one, because ndistinct
4818 : * statistics includes all combinations of attributes.
4819 : */
4820 355 : if (!item)
4821 0 : elog(ERROR, "corrupt MVNDistinct entry");
4822 :
4823 : /* Form the output varinfo list, keeping only unmatched ones */
4824 1240 : foreach(lc, *varinfos)
4825 : {
4826 885 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4827 : ListCell *lc3;
4828 885 : bool found = false;
4829 :
4830 : /*
4831 : * Let's look at plain variables first, because it's the most
4832 : * common case and the check is quite cheap. We can simply get the
4833 : * attnum and check (with an offset) matched bitmap.
4834 : */
4835 885 : if (IsA(varinfo->var, Var))
4836 725 : {
4837 730 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4838 :
4839 : /*
4840 : * If it's a system attribute, we're done. We don't support
4841 : * extended statistics on system attributes, so it's clearly
4842 : * not matched. Just keep the expression and continue.
4843 : */
4844 730 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4845 : {
4846 5 : newlist = lappend(newlist, varinfo);
4847 5 : continue;
4848 : }
4849 :
4850 : /* apply the same offset as above */
4851 725 : attnum += attnum_offset;
4852 :
4853 : /* if it's not matched, keep the varinfo */
4854 725 : if (!bms_is_member(attnum, matched))
4855 100 : newlist = lappend(newlist, varinfo);
4856 :
4857 : /* The rest of the loop deals with complex expressions. */
4858 725 : continue;
4859 : }
4860 :
4861 : /*
4862 : * Process complex expressions, not just simple Vars.
4863 : *
4864 : * First, we search for an exact match of an expression. If we
4865 : * find one, we can just discard the whole GroupVarInfo, with all
4866 : * the variables we extracted from it.
4867 : *
4868 : * Otherwise we inspect the individual vars, and try matching it
4869 : * to variables in the item.
4870 : */
4871 255 : foreach(lc3, matched_info->exprs)
4872 : {
4873 230 : Node *expr = (Node *) lfirst(lc3);
4874 :
4875 230 : if (equal(varinfo->var, expr))
4876 : {
4877 130 : found = true;
4878 130 : break;
4879 : }
4880 : }
4881 :
4882 : /* found exact match, skip */
4883 155 : if (found)
4884 130 : continue;
4885 :
4886 25 : newlist = lappend(newlist, varinfo);
4887 : }
4888 :
4889 355 : *varinfos = newlist;
4890 355 : *ndistinct = item->ndistinct;
4891 355 : return true;
4892 : }
4893 :
4894 0 : return false;
4895 : }
4896 :
4897 : /*
4898 : * convert_to_scalar
4899 : * Convert non-NULL values of the indicated types to the comparison
4900 : * scale needed by scalarineqsel().
4901 : * Returns "true" if successful.
4902 : *
4903 : * XXX this routine is a hack: ideally we should look up the conversion
4904 : * subroutines in pg_type.
4905 : *
4906 : * All numeric datatypes are simply converted to their equivalent
4907 : * "double" values. (NUMERIC values that are outside the range of "double"
4908 : * are clamped to +/- HUGE_VAL.)
4909 : *
4910 : * String datatypes are converted by convert_string_to_scalar(),
4911 : * which is explained below. The reason why this routine deals with
4912 : * three values at a time, not just one, is that we need it for strings.
4913 : *
4914 : * The bytea datatype is just enough different from strings that it has
4915 : * to be treated separately.
4916 : *
4917 : * The several datatypes representing absolute times are all converted
4918 : * to Timestamp, which is actually an int64, and then we promote that to
4919 : * a double. Note this will give correct results even for the "special"
4920 : * values of Timestamp, since those are chosen to compare correctly;
4921 : * see timestamp_cmp.
4922 : *
4923 : * The several datatypes representing relative times (intervals) are all
4924 : * converted to measurements expressed in seconds.
4925 : */
4926 : static bool
4927 56300 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4928 : Datum lobound, Datum hibound, Oid boundstypid,
4929 : double *scaledlobound, double *scaledhibound)
4930 : {
4931 56300 : bool failure = false;
4932 :
4933 : /*
4934 : * Both the valuetypid and the boundstypid should exactly match the
4935 : * declared input type(s) of the operator we are invoked for. However,
4936 : * extensions might try to use scalarineqsel as estimator for operators
4937 : * with input type(s) we don't handle here; in such cases, we want to
4938 : * return false, not fail. In any case, we mustn't assume that valuetypid
4939 : * and boundstypid are identical.
4940 : *
4941 : * XXX The histogram we are interpolating between points of could belong
4942 : * to a column that's only binary-compatible with the declared type. In
4943 : * essence we are assuming that the semantics of binary-compatible types
4944 : * are enough alike that we can use a histogram generated with one type's
4945 : * operators to estimate selectivity for the other's. This is outright
4946 : * wrong in some cases --- in particular signed versus unsigned
4947 : * interpretation could trip us up. But it's useful enough in the
4948 : * majority of cases that we do it anyway. Should think about more
4949 : * rigorous ways to do it.
4950 : */
4951 56300 : switch (valuetypid)
4952 : {
4953 : /*
4954 : * Built-in numeric types
4955 : */
4956 51413 : case BOOLOID:
4957 : case INT2OID:
4958 : case INT4OID:
4959 : case INT8OID:
4960 : case FLOAT4OID:
4961 : case FLOAT8OID:
4962 : case NUMERICOID:
4963 : case OIDOID:
4964 : case REGPROCOID:
4965 : case REGPROCEDUREOID:
4966 : case REGOPEROID:
4967 : case REGOPERATOROID:
4968 : case REGCLASSOID:
4969 : case REGTYPEOID:
4970 : case REGCOLLATIONOID:
4971 : case REGCONFIGOID:
4972 : case REGDICTIONARYOID:
4973 : case REGROLEOID:
4974 : case REGNAMESPACEOID:
4975 : case REGDATABASEOID:
4976 51413 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4977 : &failure);
4978 51413 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4979 : &failure);
4980 51413 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4981 : &failure);
4982 51413 : return !failure;
4983 :
4984 : /*
4985 : * Built-in string types
4986 : */
4987 4887 : case CHAROID:
4988 : case BPCHAROID:
4989 : case VARCHAROID:
4990 : case TEXTOID:
4991 : case NAMEOID:
4992 : {
4993 4887 : char *valstr = convert_string_datum(value, valuetypid,
4994 : collid, &failure);
4995 4887 : char *lostr = convert_string_datum(lobound, boundstypid,
4996 : collid, &failure);
4997 4887 : char *histr = convert_string_datum(hibound, boundstypid,
4998 : collid, &failure);
4999 :
5000 : /*
5001 : * Bail out if any of the values is not of string type. We
5002 : * might leak converted strings for the other value(s), but
5003 : * that's not worth troubling over.
5004 : */
5005 4887 : if (failure)
5006 0 : return false;
5007 :
5008 4887 : convert_string_to_scalar(valstr, scaledvalue,
5009 : lostr, scaledlobound,
5010 : histr, scaledhibound);
5011 4887 : pfree(valstr);
5012 4887 : pfree(lostr);
5013 4887 : pfree(histr);
5014 4887 : return true;
5015 : }
5016 :
5017 : /*
5018 : * Built-in bytea type
5019 : */
5020 0 : case BYTEAOID:
5021 : {
5022 : /* We only support bytea vs bytea comparison */
5023 0 : if (boundstypid != BYTEAOID)
5024 0 : return false;
5025 0 : convert_bytea_to_scalar(value, scaledvalue,
5026 : lobound, scaledlobound,
5027 : hibound, scaledhibound);
5028 0 : return true;
5029 : }
5030 :
5031 : /*
5032 : * Built-in time types
5033 : */
5034 0 : case TIMESTAMPOID:
5035 : case TIMESTAMPTZOID:
5036 : case DATEOID:
5037 : case INTERVALOID:
5038 : case TIMEOID:
5039 : case TIMETZOID:
5040 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
5041 : &failure);
5042 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
5043 : &failure);
5044 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
5045 : &failure);
5046 0 : return !failure;
5047 :
5048 : /*
5049 : * Built-in network types
5050 : */
5051 0 : case INETOID:
5052 : case CIDROID:
5053 : case MACADDROID:
5054 : case MACADDR8OID:
5055 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
5056 : &failure);
5057 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
5058 : &failure);
5059 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
5060 : &failure);
5061 0 : return !failure;
5062 : }
5063 : /* Don't know how to convert */
5064 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
5065 0 : return false;
5066 : }
5067 :
5068 : /*
5069 : * Do convert_to_scalar()'s work for any numeric data type.
5070 : *
5071 : * On failure (e.g., unsupported typid), set *failure to true;
5072 : * otherwise, that variable is not changed.
5073 : */
5074 : static double
5075 154239 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
5076 : {
5077 154239 : switch (typid)
5078 : {
5079 0 : case BOOLOID:
5080 0 : return (double) DatumGetBool(value);
5081 10 : case INT2OID:
5082 10 : return (double) DatumGetInt16(value);
5083 22169 : case INT4OID:
5084 22169 : return (double) DatumGetInt32(value);
5085 0 : case INT8OID:
5086 0 : return (double) DatumGetInt64(value);
5087 0 : case FLOAT4OID:
5088 0 : return (double) DatumGetFloat4(value);
5089 45 : case FLOAT8OID:
5090 45 : return (double) DatumGetFloat8(value);
5091 0 : case NUMERICOID:
5092 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
5093 0 : return (double)
5094 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
5095 : value));
5096 132015 : case OIDOID:
5097 : case REGPROCOID:
5098 : case REGPROCEDUREOID:
5099 : case REGOPEROID:
5100 : case REGOPERATOROID:
5101 : case REGCLASSOID:
5102 : case REGTYPEOID:
5103 : case REGCOLLATIONOID:
5104 : case REGCONFIGOID:
5105 : case REGDICTIONARYOID:
5106 : case REGROLEOID:
5107 : case REGNAMESPACEOID:
5108 : case REGDATABASEOID:
5109 : /* we can treat OIDs as integers... */
5110 132015 : return (double) DatumGetObjectId(value);
5111 : }
5112 :
5113 0 : *failure = true;
5114 0 : return 0;
5115 : }
5116 :
5117 : /*
5118 : * Do convert_to_scalar()'s work for any character-string data type.
5119 : *
5120 : * String datatypes are converted to a scale that ranges from 0 to 1,
5121 : * where we visualize the bytes of the string as fractional digits.
5122 : *
5123 : * We do not want the base to be 256, however, since that tends to
5124 : * generate inflated selectivity estimates; few databases will have
5125 : * occurrences of all 256 possible byte values at each position.
5126 : * Instead, use the smallest and largest byte values seen in the bounds
5127 : * as the estimated range for each byte, after some fudging to deal with
5128 : * the fact that we probably aren't going to see the full range that way.
5129 : *
5130 : * An additional refinement is that we discard any common prefix of the
5131 : * three strings before computing the scaled values. This allows us to
5132 : * "zoom in" when we encounter a narrow data range. An example is a phone
5133 : * number database where all the values begin with the same area code.
5134 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
5135 : * so this is more likely to happen than you might think.)
5136 : */
5137 : static void
5138 4887 : convert_string_to_scalar(char *value,
5139 : double *scaledvalue,
5140 : char *lobound,
5141 : double *scaledlobound,
5142 : char *hibound,
5143 : double *scaledhibound)
5144 : {
5145 : int rangelo,
5146 : rangehi;
5147 : char *sptr;
5148 :
5149 4887 : rangelo = rangehi = (unsigned char) hibound[0];
5150 68846 : for (sptr = lobound; *sptr; sptr++)
5151 : {
5152 63959 : if (rangelo > (unsigned char) *sptr)
5153 12098 : rangelo = (unsigned char) *sptr;
5154 63959 : if (rangehi < (unsigned char) *sptr)
5155 6197 : rangehi = (unsigned char) *sptr;
5156 : }
5157 56313 : for (sptr = hibound; *sptr; sptr++)
5158 : {
5159 51426 : if (rangelo > (unsigned char) *sptr)
5160 545 : rangelo = (unsigned char) *sptr;
5161 51426 : if (rangehi < (unsigned char) *sptr)
5162 1862 : rangehi = (unsigned char) *sptr;
5163 : }
5164 : /* If range includes any upper-case ASCII chars, make it include all */
5165 4887 : if (rangelo <= 'Z' && rangehi >= 'A')
5166 : {
5167 1152 : if (rangelo > 'A')
5168 185 : rangelo = 'A';
5169 1152 : if (rangehi < 'Z')
5170 425 : rangehi = 'Z';
5171 : }
5172 : /* Ditto lower-case */
5173 4887 : if (rangelo <= 'z' && rangehi >= 'a')
5174 : {
5175 4447 : if (rangelo > 'a')
5176 5 : rangelo = 'a';
5177 4447 : if (rangehi < 'z')
5178 4399 : rangehi = 'z';
5179 : }
5180 : /* Ditto digits */
5181 4887 : if (rangelo <= '9' && rangehi >= '0')
5182 : {
5183 557 : if (rangelo > '0')
5184 484 : rangelo = '0';
5185 557 : if (rangehi < '9')
5186 9 : rangehi = '9';
5187 : }
5188 :
5189 : /*
5190 : * If range includes less than 10 chars, assume we have not got enough
5191 : * data, and make it include regular ASCII set.
5192 : */
5193 4887 : if (rangehi - rangelo < 9)
5194 : {
5195 0 : rangelo = ' ';
5196 0 : rangehi = 127;
5197 : }
5198 :
5199 : /*
5200 : * Now strip any common prefix of the three strings.
5201 : */
5202 9366 : while (*lobound)
5203 : {
5204 9366 : if (*lobound != *hibound || *lobound != *value)
5205 : break;
5206 4479 : lobound++, hibound++, value++;
5207 : }
5208 :
5209 : /*
5210 : * Now we can do the conversions.
5211 : */
5212 4887 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
5213 4887 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
5214 4887 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
5215 4887 : }
5216 :
5217 : static double
5218 14661 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
5219 : {
5220 14661 : int slen = strlen(value);
5221 : double num,
5222 : denom,
5223 : base;
5224 :
5225 14661 : if (slen <= 0)
5226 0 : return 0.0; /* empty string has scalar value 0 */
5227 :
5228 : /*
5229 : * There seems little point in considering more than a dozen bytes from
5230 : * the string. Since base is at least 10, that will give us nominal
5231 : * resolution of at least 12 decimal digits, which is surely far more
5232 : * precision than this estimation technique has got anyway (especially in
5233 : * non-C locales). Also, even with the maximum possible base of 256, this
5234 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
5235 : * overflow on any known machine.
5236 : */
5237 14661 : if (slen > 12)
5238 4008 : slen = 12;
5239 :
5240 : /* Convert initial characters to fraction */
5241 14661 : base = rangehi - rangelo + 1;
5242 14661 : num = 0.0;
5243 14661 : denom = base;
5244 124860 : while (slen-- > 0)
5245 : {
5246 110199 : int ch = (unsigned char) *value++;
5247 :
5248 110199 : if (ch < rangelo)
5249 180 : ch = rangelo - 1;
5250 110019 : else if (ch > rangehi)
5251 0 : ch = rangehi + 1;
5252 110199 : num += ((double) (ch - rangelo)) / denom;
5253 110199 : denom *= base;
5254 : }
5255 :
5256 14661 : return num;
5257 : }
5258 :
5259 : /*
5260 : * Convert a string-type Datum into a palloc'd, null-terminated string.
5261 : *
5262 : * On failure (e.g., unsupported typid), set *failure to true;
5263 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
5264 : *
5265 : * When using a non-C locale, we must pass the string through pg_strxfrm()
5266 : * before continuing, so as to generate correct locale-specific results.
5267 : */
5268 : static char *
5269 14661 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
5270 : {
5271 : char *val;
5272 : pg_locale_t mylocale;
5273 :
5274 14661 : switch (typid)
5275 : {
5276 0 : case CHAROID:
5277 0 : val = (char *) palloc(2);
5278 0 : val[0] = DatumGetChar(value);
5279 0 : val[1] = '\0';
5280 0 : break;
5281 4418 : case BPCHAROID:
5282 : case VARCHAROID:
5283 : case TEXTOID:
5284 4418 : val = TextDatumGetCString(value);
5285 4418 : break;
5286 10243 : case NAMEOID:
5287 : {
5288 10243 : NameData *nm = (NameData *) DatumGetPointer(value);
5289 :
5290 10243 : val = pstrdup(NameStr(*nm));
5291 10243 : break;
5292 : }
5293 0 : default:
5294 0 : *failure = true;
5295 0 : return NULL;
5296 : }
5297 :
5298 14661 : mylocale = pg_newlocale_from_collation(collid);
5299 :
5300 14661 : if (!mylocale->collate_is_c)
5301 : {
5302 : char *xfrmstr;
5303 : size_t xfrmlen;
5304 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
5305 :
5306 : /*
5307 : * XXX: We could guess at a suitable output buffer size and only call
5308 : * pg_strxfrm() twice if our guess is too small.
5309 : *
5310 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
5311 : * bogus data or set an error. This is not really a problem unless it
5312 : * crashes since it will only give an estimation error and nothing
5313 : * fatal.
5314 : *
5315 : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
5316 : * some cases, libc strxfrm() may return the wrong results, but that
5317 : * will only lead to an estimation error.
5318 : */
5319 66 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
5320 : #ifdef WIN32
5321 :
5322 : /*
5323 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
5324 : * of trying to allocate this much memory (and fail), just return the
5325 : * original string unmodified as if we were in the C locale.
5326 : */
5327 : if (xfrmlen == INT_MAX)
5328 : return val;
5329 : #endif
5330 66 : xfrmstr = (char *) palloc(xfrmlen + 1);
5331 66 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
5332 :
5333 : /*
5334 : * Some systems (e.g., glibc) can return a smaller value from the
5335 : * second call than the first; thus the Assert must be <= not ==.
5336 : */
5337 : Assert(xfrmlen2 <= xfrmlen);
5338 66 : pfree(val);
5339 66 : val = xfrmstr;
5340 : }
5341 :
5342 14661 : return val;
5343 : }
5344 :
5345 : /*
5346 : * Do convert_to_scalar()'s work for any bytea data type.
5347 : *
5348 : * Very similar to convert_string_to_scalar except we can't assume
5349 : * null-termination and therefore pass explicit lengths around.
5350 : *
5351 : * Also, assumptions about likely "normal" ranges of characters have been
5352 : * removed - a data range of 0..255 is always used, for now. (Perhaps
5353 : * someday we will add information about actual byte data range to
5354 : * pg_statistic.)
5355 : */
5356 : static void
5357 0 : convert_bytea_to_scalar(Datum value,
5358 : double *scaledvalue,
5359 : Datum lobound,
5360 : double *scaledlobound,
5361 : Datum hibound,
5362 : double *scaledhibound)
5363 : {
5364 0 : bytea *valuep = DatumGetByteaPP(value);
5365 0 : bytea *loboundp = DatumGetByteaPP(lobound);
5366 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
5367 : int rangelo,
5368 : rangehi,
5369 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
5370 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
5371 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
5372 : i,
5373 : minlen;
5374 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
5375 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
5376 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
5377 :
5378 : /*
5379 : * Assume bytea data is uniformly distributed across all byte values.
5380 : */
5381 0 : rangelo = 0;
5382 0 : rangehi = 255;
5383 :
5384 : /*
5385 : * Now strip any common prefix of the three strings.
5386 : */
5387 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
5388 0 : for (i = 0; i < minlen; i++)
5389 : {
5390 0 : if (*lostr != *histr || *lostr != *valstr)
5391 : break;
5392 0 : lostr++, histr++, valstr++;
5393 0 : loboundlen--, hiboundlen--, valuelen--;
5394 : }
5395 :
5396 : /*
5397 : * Now we can do the conversions.
5398 : */
5399 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
5400 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
5401 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
5402 0 : }
5403 :
5404 : static double
5405 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
5406 : int rangelo, int rangehi)
5407 : {
5408 : double num,
5409 : denom,
5410 : base;
5411 :
5412 0 : if (valuelen <= 0)
5413 0 : return 0.0; /* empty string has scalar value 0 */
5414 :
5415 : /*
5416 : * Since base is 256, need not consider more than about 10 chars (even
5417 : * this many seems like overkill)
5418 : */
5419 0 : if (valuelen > 10)
5420 0 : valuelen = 10;
5421 :
5422 : /* Convert initial characters to fraction */
5423 0 : base = rangehi - rangelo + 1;
5424 0 : num = 0.0;
5425 0 : denom = base;
5426 0 : while (valuelen-- > 0)
5427 : {
5428 0 : int ch = *value++;
5429 :
5430 0 : if (ch < rangelo)
5431 0 : ch = rangelo - 1;
5432 0 : else if (ch > rangehi)
5433 0 : ch = rangehi + 1;
5434 0 : num += ((double) (ch - rangelo)) / denom;
5435 0 : denom *= base;
5436 : }
5437 :
5438 0 : return num;
5439 : }
5440 :
5441 : /*
5442 : * Do convert_to_scalar()'s work for any timevalue data type.
5443 : *
5444 : * On failure (e.g., unsupported typid), set *failure to true;
5445 : * otherwise, that variable is not changed.
5446 : */
5447 : static double
5448 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5449 : {
5450 0 : switch (typid)
5451 : {
5452 0 : case TIMESTAMPOID:
5453 0 : return DatumGetTimestamp(value);
5454 0 : case TIMESTAMPTZOID:
5455 0 : return DatumGetTimestampTz(value);
5456 0 : case DATEOID:
5457 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
5458 0 : case INTERVALOID:
5459 : {
5460 0 : Interval *interval = DatumGetIntervalP(value);
5461 :
5462 : /*
5463 : * Convert the month part of Interval to days using assumed
5464 : * average month length of 365.25/12.0 days. Not too
5465 : * accurate, but plenty good enough for our purposes.
5466 : *
5467 : * This also works for infinite intervals, which just have all
5468 : * fields set to INT_MIN/INT_MAX, and so will produce a result
5469 : * smaller/larger than any finite interval.
5470 : */
5471 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5472 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5473 : }
5474 0 : case TIMEOID:
5475 0 : return DatumGetTimeADT(value);
5476 0 : case TIMETZOID:
5477 : {
5478 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5479 :
5480 : /* use GMT-equivalent time */
5481 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5482 : }
5483 : }
5484 :
5485 0 : *failure = true;
5486 0 : return 0;
5487 : }
5488 :
5489 :
5490 : /*
5491 : * get_restriction_variable
5492 : * Examine the args of a restriction clause to see if it's of the
5493 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5494 : * where "variable" could be either a Var or an expression in vars of a
5495 : * single relation. If so, extract information about the variable,
5496 : * and also indicate which side it was on and the other argument.
5497 : *
5498 : * Inputs:
5499 : * root: the planner info
5500 : * args: clause argument list
5501 : * varRelid: see specs for restriction selectivity functions
5502 : *
5503 : * Outputs: (these are valid only if true is returned)
5504 : * *vardata: gets information about variable (see examine_variable)
5505 : * *other: gets other clause argument, aggressively reduced to a constant
5506 : * *varonleft: set true if variable is on the left, false if on the right
5507 : *
5508 : * Returns true if a variable is identified, otherwise false.
5509 : *
5510 : * Note: if there are Vars on both sides of the clause, we must fail, because
5511 : * callers are expecting that the other side will act like a pseudoconstant.
5512 : */
5513 : bool
5514 650165 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5515 : VariableStatData *vardata, Node **other,
5516 : bool *varonleft)
5517 : {
5518 : Node *left,
5519 : *right;
5520 : VariableStatData rdata;
5521 :
5522 : /* Fail if not a binary opclause (probably shouldn't happen) */
5523 650165 : if (list_length(args) != 2)
5524 0 : return false;
5525 :
5526 650165 : left = (Node *) linitial(args);
5527 650165 : right = (Node *) lsecond(args);
5528 :
5529 : /*
5530 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5531 : * relations will be treated as pseudoconstants.
5532 : */
5533 650165 : examine_variable(root, left, varRelid, vardata);
5534 650165 : examine_variable(root, right, varRelid, &rdata);
5535 :
5536 : /*
5537 : * If one side is a variable and the other not, we win.
5538 : */
5539 650165 : if (vardata->rel && rdata.rel == NULL)
5540 : {
5541 578744 : *varonleft = true;
5542 578744 : *other = estimate_expression_value(root, rdata.var);
5543 : /* Assume we need no ReleaseVariableStats(rdata) here */
5544 578740 : return true;
5545 : }
5546 :
5547 71421 : if (vardata->rel == NULL && rdata.rel)
5548 : {
5549 67103 : *varonleft = false;
5550 67103 : *other = estimate_expression_value(root, vardata->var);
5551 : /* Assume we need no ReleaseVariableStats(*vardata) here */
5552 67103 : *vardata = rdata;
5553 67103 : return true;
5554 : }
5555 :
5556 : /* Oops, clause has wrong structure (probably var op var) */
5557 4318 : ReleaseVariableStats(*vardata);
5558 4318 : ReleaseVariableStats(rdata);
5559 :
5560 4318 : return false;
5561 : }
5562 :
5563 : /*
5564 : * get_join_variables
5565 : * Apply examine_variable() to each side of a join clause.
5566 : * Also, attempt to identify whether the join clause has the same
5567 : * or reversed sense compared to the SpecialJoinInfo.
5568 : *
5569 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5570 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5571 : * where we can't tell for sure, we default to assuming it's normal.
5572 : */
5573 : void
5574 219805 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5575 : VariableStatData *vardata1, VariableStatData *vardata2,
5576 : bool *join_is_reversed)
5577 : {
5578 : Node *left,
5579 : *right;
5580 :
5581 219805 : if (list_length(args) != 2)
5582 0 : elog(ERROR, "join operator should take two arguments");
5583 :
5584 219805 : left = (Node *) linitial(args);
5585 219805 : right = (Node *) lsecond(args);
5586 :
5587 219805 : examine_variable(root, left, 0, vardata1);
5588 219805 : examine_variable(root, right, 0, vardata2);
5589 :
5590 439399 : if (vardata1->rel &&
5591 219594 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
5592 71821 : *join_is_reversed = true; /* var1 is on RHS */
5593 295757 : else if (vardata2->rel &&
5594 147773 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
5595 304 : *join_is_reversed = true; /* var2 is on LHS */
5596 : else
5597 147680 : *join_is_reversed = false;
5598 219805 : }
5599 :
5600 : /* statext_expressions_load copies the tuple, so just pfree it. */
5601 : static void
5602 1415 : ReleaseDummy(HeapTuple tuple)
5603 : {
5604 1415 : pfree(tuple);
5605 1415 : }
5606 :
5607 : /*
5608 : * examine_variable
5609 : * Try to look up statistical data about an expression.
5610 : * Fill in a VariableStatData struct to describe the expression.
5611 : *
5612 : * Inputs:
5613 : * root: the planner info
5614 : * node: the expression tree to examine
5615 : * varRelid: see specs for restriction selectivity functions
5616 : *
5617 : * Outputs: *vardata is filled as follows:
5618 : * var: the input expression (with any phvs or binary relabeling stripped,
5619 : * if it is or contains a variable; but otherwise unchanged)
5620 : * rel: RelOptInfo for relation containing variable; NULL if expression
5621 : * contains no Vars (NOTE this could point to a RelOptInfo of a
5622 : * subquery, not one in the current query).
5623 : * statsTuple: the pg_statistic entry for the variable, if one exists;
5624 : * otherwise NULL.
5625 : * freefunc: pointer to a function to release statsTuple with.
5626 : * vartype: exposed type of the expression; this should always match
5627 : * the declared input type of the operator we are estimating for.
5628 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5629 : * commonly the same as the exposed type of the variable argument,
5630 : * but can be different in binary-compatible-type cases.
5631 : * isunique: true if we were able to match the var to a unique index, a
5632 : * single-column DISTINCT or GROUP-BY clause, implying its values are
5633 : * unique for this query. (Caution: this should be trusted for
5634 : * statistical purposes only, since we do not check indimmediate nor
5635 : * verify that the exact same definition of equality applies.)
5636 : * acl_ok: true if current user has permission to read all table rows from
5637 : * the column(s) underlying the pg_statistic entry. This is consulted by
5638 : * statistic_proc_security_check().
5639 : *
5640 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5641 : */
5642 : void
5643 2571809 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5644 : VariableStatData *vardata)
5645 : {
5646 : Node *basenode;
5647 : Relids varnos;
5648 : Relids basevarnos;
5649 : RelOptInfo *onerel;
5650 :
5651 : /* Make sure we don't return dangling pointers in vardata */
5652 18002663 : MemSet(vardata, 0, sizeof(VariableStatData));
5653 :
5654 : /* Save the exposed type of the expression */
5655 2571809 : vardata->vartype = exprType(node);
5656 :
5657 : /*
5658 : * PlaceHolderVars are transparent for the purpose of statistics lookup;
5659 : * they do not alter the value distribution of the underlying expression.
5660 : * However, they can obscure the structure, preventing us from recognizing
5661 : * matches to base columns, index expressions, or extended statistics. So
5662 : * strip them out first.
5663 : */
5664 2571809 : basenode = strip_all_phvs_deep(root, node);
5665 :
5666 : /*
5667 : * Look inside any binary-compatible relabeling. We need to handle nested
5668 : * RelabelType nodes here, because the prior stripping of PlaceHolderVars
5669 : * may have brought separate RelabelTypes into adjacency.
5670 : */
5671 2612268 : while (IsA(basenode, RelabelType))
5672 40459 : basenode = (Node *) ((RelabelType *) basenode)->arg;
5673 :
5674 : /* Fast path for a simple Var */
5675 2571809 : if (IsA(basenode, Var) &&
5676 675807 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5677 : {
5678 1831936 : Var *var = (Var *) basenode;
5679 :
5680 : /* Set up result fields other than the stats tuple */
5681 1831936 : vardata->var = basenode; /* return Var without phvs or relabeling */
5682 1831936 : vardata->rel = find_base_rel(root, var->varno);
5683 1831936 : vardata->atttype = var->vartype;
5684 1831936 : vardata->atttypmod = var->vartypmod;
5685 1831936 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5686 :
5687 : /* Try to locate some stats */
5688 1831936 : examine_simple_variable(root, var, vardata);
5689 :
5690 1831936 : return;
5691 : }
5692 :
5693 : /*
5694 : * Okay, it's a more complicated expression. Determine variable
5695 : * membership. Note that when varRelid isn't zero, only vars of that
5696 : * relation are considered "real" vars.
5697 : */
5698 739873 : varnos = pull_varnos(root, basenode);
5699 739873 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5700 :
5701 739873 : onerel = NULL;
5702 :
5703 739873 : if (bms_is_empty(basevarnos))
5704 : {
5705 : /* No Vars at all ... must be pseudo-constant clause */
5706 : }
5707 : else
5708 : {
5709 : int relid;
5710 :
5711 : /* Check if the expression is in vars of a single base relation */
5712 394823 : if (bms_get_singleton_member(basevarnos, &relid))
5713 : {
5714 391149 : if (varRelid == 0 || varRelid == relid)
5715 : {
5716 54231 : onerel = find_base_rel(root, relid);
5717 54231 : vardata->rel = onerel;
5718 54231 : node = basenode; /* strip any phvs or relabeling */
5719 : }
5720 : /* else treat it as a constant */
5721 : }
5722 : else
5723 : {
5724 : /* varnos has multiple relids */
5725 3674 : if (varRelid == 0)
5726 : {
5727 : /* treat it as a variable of a join relation */
5728 2688 : vardata->rel = find_join_rel(root, varnos);
5729 2688 : node = basenode; /* strip any phvs or relabeling */
5730 : }
5731 986 : else if (bms_is_member(varRelid, varnos))
5732 : {
5733 : /* ignore the vars belonging to other relations */
5734 891 : vardata->rel = find_base_rel(root, varRelid);
5735 891 : node = basenode; /* strip any phvs or relabeling */
5736 : /* note: no point in expressional-index search here */
5737 : }
5738 : /* else treat it as a constant */
5739 : }
5740 : }
5741 :
5742 739873 : bms_free(basevarnos);
5743 :
5744 739873 : vardata->var = node;
5745 739873 : vardata->atttype = exprType(node);
5746 739873 : vardata->atttypmod = exprTypmod(node);
5747 :
5748 739873 : if (onerel)
5749 : {
5750 : /*
5751 : * We have an expression in vars of a single relation. Try to match
5752 : * it to expressional index columns, in hopes of finding some
5753 : * statistics.
5754 : *
5755 : * Note that we consider all index columns including INCLUDE columns,
5756 : * since there could be stats for such columns. But the test for
5757 : * uniqueness needs to be warier.
5758 : *
5759 : * XXX it's conceivable that there are multiple matches with different
5760 : * index opfamilies; if so, we need to pick one that matches the
5761 : * operator we are estimating for. FIXME later.
5762 : */
5763 : ListCell *ilist;
5764 : ListCell *slist;
5765 :
5766 : /*
5767 : * The nullingrels bits within the expression could prevent us from
5768 : * matching it to expressional index columns or to the expressions in
5769 : * extended statistics. So strip them out first.
5770 : */
5771 54231 : if (bms_overlap(varnos, root->outer_join_rels))
5772 1577 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5773 :
5774 119484 : foreach(ilist, onerel->indexlist)
5775 : {
5776 67630 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5777 : ListCell *indexpr_item;
5778 : int pos;
5779 :
5780 67630 : indexpr_item = list_head(index->indexprs);
5781 67630 : if (indexpr_item == NULL)
5782 63678 : continue; /* no expressions here... */
5783 :
5784 5587 : for (pos = 0; pos < index->ncolumns; pos++)
5785 : {
5786 4012 : if (index->indexkeys[pos] == 0)
5787 : {
5788 : Node *indexkey;
5789 :
5790 3952 : if (indexpr_item == NULL)
5791 0 : elog(ERROR, "too few entries in indexprs list");
5792 3952 : indexkey = (Node *) lfirst(indexpr_item);
5793 3952 : if (indexkey && IsA(indexkey, RelabelType))
5794 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5795 3952 : if (equal(node, indexkey))
5796 : {
5797 : /*
5798 : * Found a match ... is it a unique index? Tests here
5799 : * should match has_unique_index().
5800 : */
5801 2907 : if (index->unique &&
5802 365 : index->nkeycolumns == 1 &&
5803 365 : pos == 0 &&
5804 365 : (index->indpred == NIL || index->predOK))
5805 365 : vardata->isunique = true;
5806 :
5807 : /*
5808 : * Has it got stats? We only consider stats for
5809 : * non-partial indexes, since partial indexes probably
5810 : * don't reflect whole-relation statistics; the above
5811 : * check for uniqueness is the only info we take from
5812 : * a partial index.
5813 : *
5814 : * An index stats hook, however, must make its own
5815 : * decisions about what to do with partial indexes.
5816 : */
5817 2907 : if (get_index_stats_hook &&
5818 0 : (*get_index_stats_hook) (root, index->indexoid,
5819 0 : pos + 1, vardata))
5820 : {
5821 : /*
5822 : * The hook took control of acquiring a stats
5823 : * tuple. If it did supply a tuple, it'd better
5824 : * have supplied a freefunc.
5825 : */
5826 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5827 0 : !vardata->freefunc)
5828 0 : elog(ERROR, "no function provided to release variable stats with");
5829 : }
5830 2907 : else if (index->indpred == NIL)
5831 : {
5832 2907 : vardata->statsTuple =
5833 5814 : SearchSysCache3(STATRELATTINH,
5834 : ObjectIdGetDatum(index->indexoid),
5835 2907 : Int16GetDatum(pos + 1),
5836 : BoolGetDatum(false));
5837 2907 : vardata->freefunc = ReleaseSysCache;
5838 :
5839 2907 : if (HeapTupleIsValid(vardata->statsTuple))
5840 : {
5841 : /*
5842 : * Test if user has permission to access all
5843 : * rows from the index's table.
5844 : *
5845 : * For simplicity, we insist on the whole
5846 : * table being selectable, rather than trying
5847 : * to identify which column(s) the index
5848 : * depends on.
5849 : *
5850 : * Note that for an inheritance child,
5851 : * permissions are checked on the inheritance
5852 : * root parent, and whole-table select
5853 : * privilege on the parent doesn't quite
5854 : * guarantee that the user could read all
5855 : * columns of the child. But in practice it's
5856 : * unlikely that any interesting security
5857 : * violation could result from allowing access
5858 : * to the expression index's stats, so we
5859 : * allow it anyway. See similar code in
5860 : * examine_simple_variable() for additional
5861 : * comments.
5862 : */
5863 2377 : vardata->acl_ok =
5864 2377 : all_rows_selectable(root,
5865 2377 : index->rel->relid,
5866 : NULL);
5867 : }
5868 : else
5869 : {
5870 : /* suppress leakproofness checks later */
5871 530 : vardata->acl_ok = true;
5872 : }
5873 : }
5874 2907 : if (vardata->statsTuple)
5875 2377 : break;
5876 : }
5877 1575 : indexpr_item = lnext(index->indexprs, indexpr_item);
5878 : }
5879 : }
5880 3952 : if (vardata->statsTuple)
5881 2377 : break;
5882 : }
5883 :
5884 : /*
5885 : * Search extended statistics for one with a matching expression.
5886 : * There might be multiple ones, so just grab the first one. In the
5887 : * future, we might consider the statistics target (and pick the most
5888 : * accurate statistics) and maybe some other parameters.
5889 : */
5890 57760 : foreach(slist, onerel->statlist)
5891 : {
5892 3774 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5893 3774 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5894 : ListCell *expr_item;
5895 : int pos;
5896 :
5897 : /*
5898 : * Stop once we've found statistics for the expression (either
5899 : * from extended stats, or for an index in the preceding loop).
5900 : */
5901 3774 : if (vardata->statsTuple)
5902 245 : break;
5903 :
5904 : /* skip stats without per-expression stats */
5905 3529 : if (info->kind != STATS_EXT_EXPRESSIONS)
5906 1808 : continue;
5907 :
5908 : /* skip stats with mismatching stxdinherit value */
5909 1721 : if (info->inherit != rte->inh)
5910 5 : continue;
5911 :
5912 1716 : pos = 0;
5913 2836 : foreach(expr_item, info->exprs)
5914 : {
5915 2536 : Node *expr = (Node *) lfirst(expr_item);
5916 :
5917 : Assert(expr);
5918 :
5919 : /* strip RelabelType before comparing it */
5920 2536 : if (expr && IsA(expr, RelabelType))
5921 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5922 :
5923 : /* found a match, see if we can extract pg_statistic row */
5924 2536 : if (equal(node, expr))
5925 : {
5926 : /*
5927 : * XXX Not sure if we should cache the tuple somewhere.
5928 : * Now we just create a new copy every time.
5929 : */
5930 1416 : vardata->statsTuple =
5931 1416 : statext_expressions_load(info->statOid, rte->inh, pos);
5932 :
5933 : /* Nothing to release if no data found */
5934 1416 : if (vardata->statsTuple != NULL)
5935 : {
5936 1415 : vardata->freefunc = ReleaseDummy;
5937 : }
5938 :
5939 : /*
5940 : * Test if user has permission to access all rows from the
5941 : * table.
5942 : *
5943 : * For simplicity, we insist on the whole table being
5944 : * selectable, rather than trying to identify which
5945 : * column(s) the statistics object depends on.
5946 : *
5947 : * Note that for an inheritance child, permissions are
5948 : * checked on the inheritance root parent, and whole-table
5949 : * select privilege on the parent doesn't quite guarantee
5950 : * that the user could read all columns of the child. But
5951 : * in practice it's unlikely that any interesting security
5952 : * violation could result from allowing access to the
5953 : * expression stats, so we allow it anyway. See similar
5954 : * code in examine_simple_variable() for additional
5955 : * comments.
5956 : */
5957 1416 : vardata->acl_ok = all_rows_selectable(root,
5958 : onerel->relid,
5959 : NULL);
5960 :
5961 1416 : break;
5962 : }
5963 :
5964 1120 : pos++;
5965 : }
5966 : }
5967 : }
5968 :
5969 739873 : bms_free(varnos);
5970 : }
5971 :
5972 : /*
5973 : * strip_all_phvs_deep
5974 : * Deeply strip all PlaceHolderVars in an expression.
5975 : *
5976 : * As a performance optimization, we first use a lightweight walker to check
5977 : * for the presence of any PlaceHolderVars. The expensive mutator is invoked
5978 : * only if a PlaceHolderVar is found, avoiding unnecessary memory allocation
5979 : * and tree copying in the common case where no PlaceHolderVars are present.
5980 : */
5981 : static Node *
5982 2571809 : strip_all_phvs_deep(PlannerInfo *root, Node *node)
5983 : {
5984 : /* If there are no PHVs anywhere, we needn't work hard */
5985 2571809 : if (root->glob->lastPHId == 0)
5986 2545786 : return node;
5987 :
5988 26023 : if (!contain_placeholder_walker(node, NULL))
5989 22345 : return node;
5990 3678 : return strip_all_phvs_mutator(node, NULL);
5991 : }
5992 :
5993 : /*
5994 : * contain_placeholder_walker
5995 : * Lightweight walker to check if an expression contains any
5996 : * PlaceHolderVars
5997 : */
5998 : static bool
5999 29349 : contain_placeholder_walker(Node *node, void *context)
6000 : {
6001 29349 : if (node == NULL)
6002 109 : return false;
6003 29240 : if (IsA(node, PlaceHolderVar))
6004 3678 : return true;
6005 :
6006 25562 : return expression_tree_walker(node, contain_placeholder_walker, context);
6007 : }
6008 :
6009 : /*
6010 : * strip_all_phvs_mutator
6011 : * Mutator to deeply strip all PlaceHolderVars
6012 : */
6013 : static Node *
6014 9809 : strip_all_phvs_mutator(Node *node, void *context)
6015 : {
6016 9809 : if (node == NULL)
6017 34 : return NULL;
6018 9775 : if (IsA(node, PlaceHolderVar))
6019 : {
6020 : /* Strip it and recurse into its contained expression */
6021 3798 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6022 :
6023 3798 : return strip_all_phvs_mutator((Node *) phv->phexpr, context);
6024 : }
6025 :
6026 5977 : return expression_tree_mutator(node, strip_all_phvs_mutator, context);
6027 : }
6028 :
6029 : /*
6030 : * examine_simple_variable
6031 : * Handle a simple Var for examine_variable
6032 : *
6033 : * This is split out as a subroutine so that we can recurse to deal with
6034 : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
6035 : *
6036 : * We already filled in all the fields of *vardata except for the stats tuple.
6037 : */
6038 : static void
6039 1840040 : examine_simple_variable(PlannerInfo *root, Var *var,
6040 : VariableStatData *vardata)
6041 : {
6042 1840040 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
6043 :
6044 : Assert(IsA(rte, RangeTblEntry));
6045 :
6046 1840040 : if (get_relation_stats_hook &&
6047 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
6048 : {
6049 : /*
6050 : * The hook took control of acquiring a stats tuple. If it did supply
6051 : * a tuple, it'd better have supplied a freefunc.
6052 : */
6053 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6054 0 : !vardata->freefunc)
6055 0 : elog(ERROR, "no function provided to release variable stats with");
6056 : }
6057 1840040 : else if (rte->rtekind == RTE_RELATION)
6058 : {
6059 : /*
6060 : * Plain table or parent of an inheritance appendrel, so look up the
6061 : * column in pg_statistic
6062 : */
6063 1748020 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6064 : ObjectIdGetDatum(rte->relid),
6065 1748020 : Int16GetDatum(var->varattno),
6066 1748020 : BoolGetDatum(rte->inh));
6067 1748020 : vardata->freefunc = ReleaseSysCache;
6068 :
6069 1748020 : if (HeapTupleIsValid(vardata->statsTuple))
6070 : {
6071 : /*
6072 : * Test if user has permission to read all rows from this column.
6073 : *
6074 : * This requires that the user has the appropriate SELECT
6075 : * privileges and that there are no securityQuals from security
6076 : * barrier views or RLS policies. If that's not the case, then we
6077 : * only permit leakproof functions to be passed pg_statistic data
6078 : * in vardata, otherwise the functions might reveal data that the
6079 : * user doesn't have permission to see --- see
6080 : * statistic_proc_security_check().
6081 : */
6082 1153241 : vardata->acl_ok =
6083 1153241 : all_rows_selectable(root, var->varno,
6084 1153241 : bms_make_singleton(var->varattno - FirstLowInvalidHeapAttributeNumber));
6085 : }
6086 : else
6087 : {
6088 : /* suppress any possible leakproofness checks later */
6089 594779 : vardata->acl_ok = true;
6090 : }
6091 : }
6092 92020 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
6093 84205 : (rte->rtekind == RTE_CTE && !rte->self_reference))
6094 : {
6095 : /*
6096 : * Plain subquery (not one that was converted to an appendrel) or
6097 : * non-recursive CTE. In either case, we can try to find out what the
6098 : * Var refers to within the subquery. We skip this for appendrel and
6099 : * recursive-CTE cases because any column stats we did find would
6100 : * likely not be very relevant.
6101 : */
6102 : PlannerInfo *subroot;
6103 : Query *subquery;
6104 : List *subtlist;
6105 : TargetEntry *ste;
6106 :
6107 : /*
6108 : * Punt if it's a whole-row var rather than a plain column reference.
6109 : */
6110 15969 : if (var->varattno == InvalidAttrNumber)
6111 0 : return;
6112 :
6113 : /*
6114 : * Otherwise, find the subquery's planner subroot.
6115 : */
6116 15969 : if (rte->rtekind == RTE_SUBQUERY)
6117 : {
6118 : RelOptInfo *rel;
6119 :
6120 : /*
6121 : * Fetch RelOptInfo for subquery. Note that we don't change the
6122 : * rel returned in vardata, since caller expects it to be a rel of
6123 : * the caller's query level. Because we might already be
6124 : * recursing, we can't use that rel pointer either, but have to
6125 : * look up the Var's rel afresh.
6126 : */
6127 7815 : rel = find_base_rel(root, var->varno);
6128 :
6129 7815 : subroot = rel->subroot;
6130 : }
6131 : else
6132 : {
6133 : /* CTE case is more difficult */
6134 : PlannerInfo *cteroot;
6135 : Index levelsup;
6136 : int ndx;
6137 : int plan_id;
6138 : ListCell *lc;
6139 :
6140 : /*
6141 : * Find the referenced CTE, and locate the subroot previously made
6142 : * for it.
6143 : */
6144 8154 : levelsup = rte->ctelevelsup;
6145 8154 : cteroot = root;
6146 19208 : while (levelsup-- > 0)
6147 : {
6148 11054 : cteroot = cteroot->parent_root;
6149 11054 : if (!cteroot) /* shouldn't happen */
6150 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
6151 : }
6152 :
6153 : /*
6154 : * Note: cte_plan_ids can be shorter than cteList, if we are still
6155 : * working on planning the CTEs (ie, this is a side-reference from
6156 : * another CTE). So we mustn't use forboth here.
6157 : */
6158 8154 : ndx = 0;
6159 11204 : foreach(lc, cteroot->parse->cteList)
6160 : {
6161 11204 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
6162 :
6163 11204 : if (strcmp(cte->ctename, rte->ctename) == 0)
6164 8154 : break;
6165 3050 : ndx++;
6166 : }
6167 8154 : if (lc == NULL) /* shouldn't happen */
6168 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
6169 8154 : if (ndx >= list_length(cteroot->cte_plan_ids))
6170 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
6171 8154 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
6172 8154 : if (plan_id <= 0)
6173 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
6174 8154 : subroot = list_nth(root->glob->subroots, plan_id - 1);
6175 : }
6176 :
6177 : /* If the subquery hasn't been planned yet, we have to punt */
6178 15969 : if (subroot == NULL)
6179 0 : return;
6180 : Assert(IsA(subroot, PlannerInfo));
6181 :
6182 : /*
6183 : * We must use the subquery parsetree as mangled by the planner, not
6184 : * the raw version from the RTE, because we need a Var that will refer
6185 : * to the subroot's live RelOptInfos. For instance, if any subquery
6186 : * pullup happened during planning, Vars in the targetlist might have
6187 : * gotten replaced, and we need to see the replacement expressions.
6188 : */
6189 15969 : subquery = subroot->parse;
6190 : Assert(IsA(subquery, Query));
6191 :
6192 : /*
6193 : * Punt if subquery uses set operations or grouping sets, as these
6194 : * will mash underlying columns' stats beyond recognition. (Set ops
6195 : * are particularly nasty; if we forged ahead, we would return stats
6196 : * relevant to only the leftmost subselect...) DISTINCT is also
6197 : * problematic, but we check that later because there is a possibility
6198 : * of learning something even with it.
6199 : */
6200 15969 : if (subquery->setOperations ||
6201 14391 : subquery->groupingSets)
6202 1655 : return;
6203 :
6204 : /* Get the subquery output expression referenced by the upper Var */
6205 14314 : if (subquery->returningList)
6206 179 : subtlist = subquery->returningList;
6207 : else
6208 14135 : subtlist = subquery->targetList;
6209 14314 : ste = get_tle_by_resno(subtlist, var->varattno);
6210 14314 : if (ste == NULL || ste->resjunk)
6211 0 : elog(ERROR, "subquery %s does not have attribute %d",
6212 : rte->eref->aliasname, var->varattno);
6213 14314 : var = (Var *) ste->expr;
6214 :
6215 : /*
6216 : * If subquery uses DISTINCT, we can't make use of any stats for the
6217 : * variable ... but, if it's the only DISTINCT column, we are entitled
6218 : * to consider it unique. We do the test this way so that it works
6219 : * for cases involving DISTINCT ON.
6220 : */
6221 14314 : if (subquery->distinctClause)
6222 : {
6223 1420 : if (list_length(subquery->distinctClause) == 1 &&
6224 475 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
6225 314 : vardata->isunique = true;
6226 : /* cannot go further */
6227 945 : return;
6228 : }
6229 :
6230 : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
6231 13369 : if (subquery->groupClause)
6232 : {
6233 947 : if (list_length(subquery->groupClause) == 1 &&
6234 402 : targetIsInSortList(ste, InvalidOid, subquery->groupClause))
6235 322 : vardata->isunique = true;
6236 : /* cannot go further */
6237 545 : return;
6238 : }
6239 :
6240 : /*
6241 : * If the sub-query originated from a view with the security_barrier
6242 : * attribute, we must not look at the variable's statistics, though it
6243 : * seems all right to notice the existence of a DISTINCT clause. So
6244 : * stop here.
6245 : *
6246 : * This is probably a harsher restriction than necessary; it's
6247 : * certainly OK for the selectivity estimator (which is a C function,
6248 : * and therefore omnipotent anyway) to look at the statistics. But
6249 : * many selectivity estimators will happily *invoke the operator
6250 : * function* to try to work out a good estimate - and that's not OK.
6251 : * So for now, don't dig down for stats.
6252 : */
6253 12824 : if (rte->security_barrier)
6254 473 : return;
6255 :
6256 : /* Can only handle a simple Var of subquery's query level */
6257 12351 : if (var && IsA(var, Var) &&
6258 8104 : var->varlevelsup == 0)
6259 : {
6260 : /*
6261 : * OK, recurse into the subquery. Note that the original setting
6262 : * of vardata->isunique (which will surely be false) is left
6263 : * unchanged in this situation. That's what we want, since even
6264 : * if the underlying column is unique, the subquery may have
6265 : * joined to other tables in a way that creates duplicates.
6266 : */
6267 8104 : examine_simple_variable(subroot, var, vardata);
6268 : }
6269 : }
6270 : else
6271 : {
6272 : /*
6273 : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
6274 : * see RTE_JOIN here because join alias Vars have already been
6275 : * flattened.) There's not much we can do with function outputs, but
6276 : * maybe someday try to be smarter about VALUES.
6277 : */
6278 : }
6279 : }
6280 :
6281 : /*
6282 : * all_rows_selectable
6283 : * Test whether the user has permission to select all rows from a given
6284 : * relation.
6285 : *
6286 : * Inputs:
6287 : * root: the planner info
6288 : * varno: the index of the relation (assumed to be an RTE_RELATION)
6289 : * varattnos: the attributes for which permission is required, or NULL if
6290 : * whole-table access is required
6291 : *
6292 : * Returns true if the user has the required select permissions, and there are
6293 : * no securityQuals from security barrier views or RLS policies.
6294 : *
6295 : * Note that if the relation is an inheritance child relation, securityQuals
6296 : * and access permissions are checked against the inheritance root parent (the
6297 : * relation actually mentioned in the query) --- see the comments in
6298 : * expand_single_inheritance_child() for an explanation of why it has to be
6299 : * done this way.
6300 : *
6301 : * If varattnos is non-NULL, its attribute numbers should be offset by
6302 : * FirstLowInvalidHeapAttributeNumber so that system attributes can be
6303 : * checked. If varattnos is NULL, only table-level SELECT privileges are
6304 : * checked, not any column-level privileges.
6305 : *
6306 : * Note: if the relation is accessed via a view, this function actually tests
6307 : * whether the view owner has permission to select from the relation. To
6308 : * ensure that the current user has permission, it is also necessary to check
6309 : * that the current user has permission to select from the view, which we do
6310 : * at planner-startup --- see subquery_planner().
6311 : *
6312 : * This is exported so that other estimation functions can use it.
6313 : */
6314 : bool
6315 1157244 : all_rows_selectable(PlannerInfo *root, Index varno, Bitmapset *varattnos)
6316 : {
6317 1157244 : RelOptInfo *rel = find_base_rel_noerr(root, varno);
6318 1157244 : RangeTblEntry *rte = planner_rt_fetch(varno, root);
6319 : Oid userid;
6320 : int varattno;
6321 :
6322 : Assert(rte->rtekind == RTE_RELATION);
6323 :
6324 : /*
6325 : * Determine the user ID to use for privilege checks (either the current
6326 : * user or the view owner, if we're accessing the table via a view).
6327 : *
6328 : * Normally the relation will have an associated RelOptInfo from which we
6329 : * can find the userid, but it might not if it's a RETURNING Var for an
6330 : * INSERT target relation. In that case use the RTEPermissionInfo
6331 : * associated with the RTE.
6332 : *
6333 : * If we navigate up to a parent relation, we keep using the same userid,
6334 : * since it's the same in all relations of a given inheritance tree.
6335 : */
6336 1157244 : if (rel)
6337 1157211 : userid = rel->userid;
6338 : else
6339 : {
6340 : RTEPermissionInfo *perminfo;
6341 :
6342 33 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
6343 33 : userid = perminfo->checkAsUser;
6344 : }
6345 1157244 : if (!OidIsValid(userid))
6346 1026518 : userid = GetUserId();
6347 :
6348 : /*
6349 : * Permissions and securityQuals must be checked on the table actually
6350 : * mentioned in the query, so if this is an inheritance child, navigate up
6351 : * to the inheritance root parent. If the user can read the whole table
6352 : * or the required columns there, then they can read from the child table
6353 : * too. For per-column checks, we must find out which of the root
6354 : * parent's attributes the child relation's attributes correspond to.
6355 : */
6356 1157244 : if (root->append_rel_array != NULL)
6357 : {
6358 : AppendRelInfo *appinfo;
6359 :
6360 177728 : appinfo = root->append_rel_array[varno];
6361 :
6362 : /*
6363 : * Partitions are mapped to their immediate parent, not the root
6364 : * parent, so must be ready to walk up multiple AppendRelInfos. But
6365 : * stop if we hit a parent that is not RTE_RELATION --- that's a
6366 : * flattened UNION ALL subquery, not an inheritance parent.
6367 : */
6368 334070 : while (appinfo &&
6369 156727 : planner_rt_fetch(appinfo->parent_relid,
6370 156727 : root)->rtekind == RTE_RELATION)
6371 : {
6372 156342 : Bitmapset *parent_varattnos = NULL;
6373 :
6374 : /*
6375 : * For each child attribute, find the corresponding parent
6376 : * attribute. In rare cases, the attribute may be local to the
6377 : * child table, in which case, we've got to live with having no
6378 : * access to this column.
6379 : */
6380 156342 : varattno = -1;
6381 310417 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6382 : {
6383 : AttrNumber attno;
6384 : AttrNumber parent_attno;
6385 :
6386 154075 : attno = varattno + FirstLowInvalidHeapAttributeNumber;
6387 :
6388 154075 : if (attno == InvalidAttrNumber)
6389 : {
6390 : /*
6391 : * Whole-row reference, so must map each column of the
6392 : * child to the parent table.
6393 : */
6394 30 : for (attno = 1; attno <= appinfo->num_child_cols; attno++)
6395 : {
6396 20 : parent_attno = appinfo->parent_colnos[attno - 1];
6397 20 : if (parent_attno == 0)
6398 0 : return false; /* attr is local to child */
6399 : parent_varattnos =
6400 20 : bms_add_member(parent_varattnos,
6401 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6402 : }
6403 : }
6404 : else
6405 : {
6406 154065 : if (attno < 0)
6407 : {
6408 : /* System attnos are the same in all tables */
6409 0 : parent_attno = attno;
6410 : }
6411 : else
6412 : {
6413 154065 : if (attno > appinfo->num_child_cols)
6414 0 : return false; /* safety check */
6415 154065 : parent_attno = appinfo->parent_colnos[attno - 1];
6416 154065 : if (parent_attno == 0)
6417 0 : return false; /* attr is local to child */
6418 : }
6419 : parent_varattnos =
6420 154065 : bms_add_member(parent_varattnos,
6421 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6422 : }
6423 : }
6424 :
6425 : /* If the parent is itself a child, continue up */
6426 156342 : varno = appinfo->parent_relid;
6427 156342 : varattnos = parent_varattnos;
6428 156342 : appinfo = root->append_rel_array[varno];
6429 : }
6430 :
6431 : /* Perform the access check on this parent rel */
6432 177728 : rte = planner_rt_fetch(varno, root);
6433 : Assert(rte->rtekind == RTE_RELATION);
6434 : }
6435 :
6436 : /*
6437 : * For all rows to be accessible, there must be no securityQuals from
6438 : * security barrier views or RLS policies.
6439 : */
6440 1157244 : if (rte->securityQuals != NIL)
6441 690 : return false;
6442 :
6443 : /*
6444 : * Test for table-level SELECT privilege.
6445 : *
6446 : * If varattnos is non-NULL, this is sufficient to give access to all
6447 : * requested attributes, even for a child table, since we have verified
6448 : * that all required child columns have matching parent columns.
6449 : *
6450 : * If varattnos is NULL (whole-table access requested), this doesn't
6451 : * necessarily guarantee that the user can read all columns of a child
6452 : * table, but we allow it anyway (see comments in examine_variable()) and
6453 : * don't bother checking any column privileges.
6454 : */
6455 1156554 : if (pg_class_aclcheck(rte->relid, userid, ACL_SELECT) == ACLCHECK_OK)
6456 1156211 : return true;
6457 :
6458 343 : if (varattnos == NULL)
6459 10 : return false; /* whole-table access requested */
6460 :
6461 : /*
6462 : * Don't have table-level SELECT privilege, so check per-column
6463 : * privileges.
6464 : */
6465 333 : varattno = -1;
6466 471 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6467 : {
6468 333 : AttrNumber attno = varattno + FirstLowInvalidHeapAttributeNumber;
6469 :
6470 333 : if (attno == InvalidAttrNumber)
6471 : {
6472 : /* Whole-row reference, so must have access to all columns */
6473 5 : if (pg_attribute_aclcheck_all(rte->relid, userid, ACL_SELECT,
6474 : ACLMASK_ALL) != ACLCHECK_OK)
6475 5 : return false;
6476 : }
6477 : else
6478 : {
6479 328 : if (pg_attribute_aclcheck(rte->relid, attno, userid,
6480 : ACL_SELECT) != ACLCHECK_OK)
6481 190 : return false;
6482 : }
6483 : }
6484 :
6485 : /* If we reach here, have all required column privileges */
6486 138 : return true;
6487 : }
6488 :
6489 : /*
6490 : * examine_indexcol_variable
6491 : * Try to look up statistical data about an index column/expression.
6492 : * Fill in a VariableStatData struct to describe the column.
6493 : *
6494 : * Inputs:
6495 : * root: the planner info
6496 : * index: the index whose column we're interested in
6497 : * indexcol: 0-based index column number (subscripts index->indexkeys[])
6498 : *
6499 : * Outputs: *vardata is filled as follows:
6500 : * var: the input expression (with any binary relabeling stripped, if
6501 : * it is or contains a variable; but otherwise the type is preserved)
6502 : * rel: RelOptInfo for table relation containing variable.
6503 : * statsTuple: the pg_statistic entry for the variable, if one exists;
6504 : * otherwise NULL.
6505 : * freefunc: pointer to a function to release statsTuple with.
6506 : *
6507 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
6508 : */
6509 : static void
6510 653238 : examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
6511 : int indexcol, VariableStatData *vardata)
6512 : {
6513 : AttrNumber colnum;
6514 : Oid relid;
6515 :
6516 653238 : if (index->indexkeys[indexcol] != 0)
6517 : {
6518 : /* Simple variable --- look to stats for the underlying table */
6519 651405 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6520 :
6521 : Assert(rte->rtekind == RTE_RELATION);
6522 651405 : relid = rte->relid;
6523 : Assert(relid != InvalidOid);
6524 651405 : colnum = index->indexkeys[indexcol];
6525 651405 : vardata->rel = index->rel;
6526 :
6527 651405 : if (get_relation_stats_hook &&
6528 0 : (*get_relation_stats_hook) (root, rte, colnum, vardata))
6529 : {
6530 : /*
6531 : * The hook took control of acquiring a stats tuple. If it did
6532 : * supply a tuple, it'd better have supplied a freefunc.
6533 : */
6534 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6535 0 : !vardata->freefunc)
6536 0 : elog(ERROR, "no function provided to release variable stats with");
6537 : }
6538 : else
6539 : {
6540 651405 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6541 : ObjectIdGetDatum(relid),
6542 : Int16GetDatum(colnum),
6543 651405 : BoolGetDatum(rte->inh));
6544 651405 : vardata->freefunc = ReleaseSysCache;
6545 : }
6546 : }
6547 : else
6548 : {
6549 : /* Expression --- maybe there are stats for the index itself */
6550 1833 : relid = index->indexoid;
6551 1833 : colnum = indexcol + 1;
6552 :
6553 1833 : if (get_index_stats_hook &&
6554 0 : (*get_index_stats_hook) (root, relid, colnum, vardata))
6555 : {
6556 : /*
6557 : * The hook took control of acquiring a stats tuple. If it did
6558 : * supply a tuple, it'd better have supplied a freefunc.
6559 : */
6560 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6561 0 : !vardata->freefunc)
6562 0 : elog(ERROR, "no function provided to release variable stats with");
6563 : }
6564 : else
6565 : {
6566 1833 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6567 : ObjectIdGetDatum(relid),
6568 : Int16GetDatum(colnum),
6569 : BoolGetDatum(false));
6570 1833 : vardata->freefunc = ReleaseSysCache;
6571 : }
6572 : }
6573 653238 : }
6574 :
6575 : /*
6576 : * Check whether it is permitted to call func_oid passing some of the
6577 : * pg_statistic data in vardata. We allow this if either of the following
6578 : * conditions is met: (1) the user has SELECT privileges on the table or
6579 : * column underlying the pg_statistic data and there are no securityQuals from
6580 : * security barrier views or RLS policies, or (2) the function is marked
6581 : * leakproof.
6582 : */
6583 : bool
6584 775401 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
6585 : {
6586 775401 : if (vardata->acl_ok)
6587 773890 : return true; /* have SELECT privs and no securityQuals */
6588 :
6589 1511 : if (!OidIsValid(func_oid))
6590 0 : return false;
6591 :
6592 1511 : if (get_func_leakproof(func_oid))
6593 748 : return true;
6594 :
6595 763 : ereport(DEBUG2,
6596 : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
6597 : get_func_name(func_oid))));
6598 763 : return false;
6599 : }
6600 :
6601 : /*
6602 : * get_variable_numdistinct
6603 : * Estimate the number of distinct values of a variable.
6604 : *
6605 : * vardata: results of examine_variable
6606 : * *isdefault: set to true if the result is a default rather than based on
6607 : * anything meaningful.
6608 : *
6609 : * NB: be careful to produce a positive integral result, since callers may
6610 : * compare the result to exact integer counts, or might divide by it.
6611 : */
6612 : double
6613 1303737 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
6614 : {
6615 : double stadistinct;
6616 1303737 : double stanullfrac = 0.0;
6617 : double ntuples;
6618 :
6619 1303737 : *isdefault = false;
6620 :
6621 : /*
6622 : * Determine the stadistinct value to use. There are cases where we can
6623 : * get an estimate even without a pg_statistic entry, or can get a better
6624 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
6625 : * (otherwise, assume no nulls, for lack of any better idea).
6626 : */
6627 1303737 : if (HeapTupleIsValid(vardata->statsTuple))
6628 : {
6629 : /* Use the pg_statistic entry */
6630 : Form_pg_statistic stats;
6631 :
6632 814179 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6633 814179 : stadistinct = stats->stadistinct;
6634 814179 : stanullfrac = stats->stanullfrac;
6635 : }
6636 489558 : else if (vardata->vartype == BOOLOID)
6637 : {
6638 : /*
6639 : * Special-case boolean columns: presumably, two distinct values.
6640 : *
6641 : * Are there any other datatypes we should wire in special estimates
6642 : * for?
6643 : */
6644 684 : stadistinct = 2.0;
6645 : }
6646 488874 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6647 : {
6648 : /*
6649 : * If the Var represents a column of a VALUES RTE, assume it's unique.
6650 : * This could of course be very wrong, but it should tend to be true
6651 : * in well-written queries. We could consider examining the VALUES'
6652 : * contents to get some real statistics; but that only works if the
6653 : * entries are all constants, and it would be pretty expensive anyway.
6654 : */
6655 2999 : stadistinct = -1.0; /* unique (and all non null) */
6656 : }
6657 : else
6658 : {
6659 : /*
6660 : * We don't keep statistics for system columns, but in some cases we
6661 : * can infer distinctness anyway.
6662 : */
6663 485875 : if (vardata->var && IsA(vardata->var, Var))
6664 : {
6665 454793 : switch (((Var *) vardata->var)->varattno)
6666 : {
6667 857 : case SelfItemPointerAttributeNumber:
6668 857 : stadistinct = -1.0; /* unique (and all non null) */
6669 857 : break;
6670 18252 : case TableOidAttributeNumber:
6671 18252 : stadistinct = 1.0; /* only 1 value */
6672 18252 : break;
6673 435684 : default:
6674 435684 : stadistinct = 0.0; /* means "unknown" */
6675 435684 : break;
6676 : }
6677 : }
6678 : else
6679 31082 : stadistinct = 0.0; /* means "unknown" */
6680 :
6681 : /*
6682 : * XXX consider using estimate_num_groups on expressions?
6683 : */
6684 : }
6685 :
6686 : /*
6687 : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6688 : * variable, assume it is unique no matter what pg_statistic says; the
6689 : * statistics could be out of date, or we might have found a partial
6690 : * unique index that proves the var is unique for this query. However,
6691 : * we'd better still believe the null-fraction statistic.
6692 : */
6693 1303737 : if (vardata->isunique)
6694 311802 : stadistinct = -1.0 * (1.0 - stanullfrac);
6695 :
6696 : /*
6697 : * If we had an absolute estimate, use that.
6698 : */
6699 1303737 : if (stadistinct > 0.0)
6700 316633 : return clamp_row_est(stadistinct);
6701 :
6702 : /*
6703 : * Otherwise we need to get the relation size; punt if not available.
6704 : */
6705 987104 : if (vardata->rel == NULL)
6706 : {
6707 591 : *isdefault = true;
6708 591 : return DEFAULT_NUM_DISTINCT;
6709 : }
6710 986513 : ntuples = vardata->rel->tuples;
6711 986513 : if (ntuples <= 0.0)
6712 : {
6713 111045 : *isdefault = true;
6714 111045 : return DEFAULT_NUM_DISTINCT;
6715 : }
6716 :
6717 : /*
6718 : * If we had a relative estimate, use that.
6719 : */
6720 875468 : if (stadistinct < 0.0)
6721 570345 : return clamp_row_est(-stadistinct * ntuples);
6722 :
6723 : /*
6724 : * With no data, estimate ndistinct = ntuples if the table is small, else
6725 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6726 : * that the behavior isn't discontinuous.
6727 : */
6728 305123 : if (ntuples < DEFAULT_NUM_DISTINCT)
6729 156012 : return clamp_row_est(ntuples);
6730 :
6731 149111 : *isdefault = true;
6732 149111 : return DEFAULT_NUM_DISTINCT;
6733 : }
6734 :
6735 : /*
6736 : * get_variable_range
6737 : * Estimate the minimum and maximum value of the specified variable.
6738 : * If successful, store values in *min and *max, and return true.
6739 : * If no data available, return false.
6740 : *
6741 : * sortop is the "<" comparison operator to use. This should generally
6742 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6743 : * The collation must be specified too.
6744 : */
6745 : static bool
6746 173418 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6747 : Oid sortop, Oid collation,
6748 : Datum *min, Datum *max)
6749 : {
6750 173418 : Datum tmin = 0;
6751 173418 : Datum tmax = 0;
6752 173418 : bool have_data = false;
6753 : int16 typLen;
6754 : bool typByVal;
6755 : Oid opfuncoid;
6756 : FmgrInfo opproc;
6757 : AttStatsSlot sslot;
6758 :
6759 : /*
6760 : * XXX It's very tempting to try to use the actual column min and max, if
6761 : * we can get them relatively-cheaply with an index probe. However, since
6762 : * this function is called many times during join planning, that could
6763 : * have unpleasant effects on planning speed. Need more investigation
6764 : * before enabling this.
6765 : */
6766 : #ifdef NOT_USED
6767 : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6768 : return true;
6769 : #endif
6770 :
6771 173418 : if (!HeapTupleIsValid(vardata->statsTuple))
6772 : {
6773 : /* no stats available, so default result */
6774 48199 : return false;
6775 : }
6776 :
6777 : /*
6778 : * If we can't apply the sortop to the stats data, just fail. In
6779 : * principle, if there's a histogram and no MCVs, we could return the
6780 : * histogram endpoints without ever applying the sortop ... but it's
6781 : * probably not worth trying, because whatever the caller wants to do with
6782 : * the endpoints would likely fail the security check too.
6783 : */
6784 125219 : if (!statistic_proc_security_check(vardata,
6785 125219 : (opfuncoid = get_opcode(sortop))))
6786 0 : return false;
6787 :
6788 125219 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6789 :
6790 125219 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6791 :
6792 : /*
6793 : * If there is a histogram with the ordering we want, grab the first and
6794 : * last values.
6795 : */
6796 125219 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6797 : STATISTIC_KIND_HISTOGRAM, sortop,
6798 : ATTSTATSSLOT_VALUES))
6799 : {
6800 77741 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6801 : {
6802 77741 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6803 77741 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6804 77741 : have_data = true;
6805 : }
6806 77741 : free_attstatsslot(&sslot);
6807 : }
6808 :
6809 : /*
6810 : * Otherwise, if there is a histogram with some other ordering, scan it
6811 : * and get the min and max values according to the ordering we want. This
6812 : * of course may not find values that are really extremal according to our
6813 : * ordering, but it beats ignoring available data.
6814 : */
6815 172697 : if (!have_data &&
6816 47478 : get_attstatsslot(&sslot, vardata->statsTuple,
6817 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6818 : ATTSTATSSLOT_VALUES))
6819 : {
6820 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6821 : collation, typLen, typByVal,
6822 : &tmin, &tmax, &have_data);
6823 0 : free_attstatsslot(&sslot);
6824 : }
6825 :
6826 : /*
6827 : * If we have most-common-values info, look for extreme MCVs. This is
6828 : * needed even if we also have a histogram, since the histogram excludes
6829 : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6830 : * be pretty wary of deciding that that is a full representation of the
6831 : * data. Proceed only if the MCVs represent the whole table (to within
6832 : * roundoff error).
6833 : */
6834 125219 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6835 : STATISTIC_KIND_MCV, InvalidOid,
6836 125219 : have_data ? ATTSTATSSLOT_VALUES :
6837 : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6838 : {
6839 70730 : bool use_mcvs = have_data;
6840 :
6841 70730 : if (!have_data)
6842 : {
6843 46552 : double sumcommon = 0.0;
6844 : double nullfrac;
6845 : int i;
6846 :
6847 375526 : for (i = 0; i < sslot.nnumbers; i++)
6848 328974 : sumcommon += sslot.numbers[i];
6849 46552 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6850 46552 : if (sumcommon + nullfrac > 0.99999)
6851 45726 : use_mcvs = true;
6852 : }
6853 :
6854 70730 : if (use_mcvs)
6855 69904 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6856 : collation, typLen, typByVal,
6857 : &tmin, &tmax, &have_data);
6858 70730 : free_attstatsslot(&sslot);
6859 : }
6860 :
6861 125219 : *min = tmin;
6862 125219 : *max = tmax;
6863 125219 : return have_data;
6864 : }
6865 :
6866 : /*
6867 : * get_stats_slot_range: scan sslot for min/max values
6868 : *
6869 : * Subroutine for get_variable_range: update min/max/have_data according
6870 : * to what we find in the statistics array.
6871 : */
6872 : static void
6873 69904 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6874 : Oid collation, int16 typLen, bool typByVal,
6875 : Datum *min, Datum *max, bool *p_have_data)
6876 : {
6877 69904 : Datum tmin = *min;
6878 69904 : Datum tmax = *max;
6879 69904 : bool have_data = *p_have_data;
6880 69904 : bool found_tmin = false;
6881 69904 : bool found_tmax = false;
6882 :
6883 : /* Look up the comparison function, if we didn't already do so */
6884 69904 : if (opproc->fn_oid != opfuncoid)
6885 69904 : fmgr_info(opfuncoid, opproc);
6886 :
6887 : /* Scan all the slot's values */
6888 1682602 : for (int i = 0; i < sslot->nvalues; i++)
6889 : {
6890 1612698 : if (!have_data)
6891 : {
6892 45726 : tmin = tmax = sslot->values[i];
6893 45726 : found_tmin = found_tmax = true;
6894 45726 : *p_have_data = have_data = true;
6895 45726 : continue;
6896 : }
6897 1566972 : if (DatumGetBool(FunctionCall2Coll(opproc,
6898 : collation,
6899 1566972 : sslot->values[i], tmin)))
6900 : {
6901 39539 : tmin = sslot->values[i];
6902 39539 : found_tmin = true;
6903 : }
6904 1566972 : if (DatumGetBool(FunctionCall2Coll(opproc,
6905 : collation,
6906 1566972 : tmax, sslot->values[i])))
6907 : {
6908 190928 : tmax = sslot->values[i];
6909 190928 : found_tmax = true;
6910 : }
6911 : }
6912 :
6913 : /*
6914 : * Copy the slot's values, if we found new extreme values.
6915 : */
6916 69904 : if (found_tmin)
6917 60510 : *min = datumCopy(tmin, typByVal, typLen);
6918 69904 : if (found_tmax)
6919 48256 : *max = datumCopy(tmax, typByVal, typLen);
6920 69904 : }
6921 :
6922 :
6923 : /*
6924 : * get_actual_variable_range
6925 : * Attempt to identify the current *actual* minimum and/or maximum
6926 : * of the specified variable, by looking for a suitable btree index
6927 : * and fetching its low and/or high values.
6928 : * If successful, store values in *min and *max, and return true.
6929 : * (Either pointer can be NULL if that endpoint isn't needed.)
6930 : * If unsuccessful, return false.
6931 : *
6932 : * sortop is the "<" comparison operator to use.
6933 : * collation is the required collation.
6934 : */
6935 : static bool
6936 119094 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6937 : Oid sortop, Oid collation,
6938 : Datum *min, Datum *max)
6939 : {
6940 119094 : bool have_data = false;
6941 119094 : RelOptInfo *rel = vardata->rel;
6942 : RangeTblEntry *rte;
6943 : ListCell *lc;
6944 :
6945 : /* No hope if no relation or it doesn't have indexes */
6946 119094 : if (rel == NULL || rel->indexlist == NIL)
6947 8991 : return false;
6948 : /* If it has indexes it must be a plain relation */
6949 110103 : rte = root->simple_rte_array[rel->relid];
6950 : Assert(rte->rtekind == RTE_RELATION);
6951 :
6952 : /* ignore partitioned tables. Any indexes here are not real indexes */
6953 110103 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6954 560 : return false;
6955 :
6956 : /* Search through the indexes to see if any match our problem */
6957 214202 : foreach(lc, rel->indexlist)
6958 : {
6959 184101 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6960 : ScanDirection indexscandir;
6961 : StrategyNumber strategy;
6962 :
6963 : /* Ignore non-ordering indexes */
6964 184101 : if (index->sortopfamily == NULL)
6965 3 : continue;
6966 :
6967 : /*
6968 : * Ignore partial indexes --- we only want stats that cover the entire
6969 : * relation.
6970 : */
6971 184098 : if (index->indpred != NIL)
6972 240 : continue;
6973 :
6974 : /*
6975 : * The index list might include hypothetical indexes inserted by a
6976 : * get_relation_info hook --- don't try to access them.
6977 : */
6978 183858 : if (index->hypothetical)
6979 0 : continue;
6980 :
6981 : /*
6982 : * get_actual_variable_endpoint uses the index-only-scan machinery, so
6983 : * ignore indexes that can't use it on their first column.
6984 : */
6985 183858 : if (!index->canreturn[0])
6986 0 : continue;
6987 :
6988 : /*
6989 : * The first index column must match the desired variable, sortop, and
6990 : * collation --- but we can use a descending-order index.
6991 : */
6992 183858 : if (collation != index->indexcollations[0])
6993 24192 : continue; /* test first 'cause it's cheapest */
6994 159666 : if (!match_index_to_operand(vardata->var, 0, index))
6995 80224 : continue;
6996 79442 : strategy = get_op_opfamily_strategy(sortop, index->sortopfamily[0]);
6997 79442 : switch (IndexAmTranslateStrategy(strategy, index->relam, index->sortopfamily[0], true))
6998 : {
6999 79442 : case COMPARE_LT:
7000 79442 : if (index->reverse_sort[0])
7001 0 : indexscandir = BackwardScanDirection;
7002 : else
7003 79442 : indexscandir = ForwardScanDirection;
7004 79442 : break;
7005 0 : case COMPARE_GT:
7006 0 : if (index->reverse_sort[0])
7007 0 : indexscandir = ForwardScanDirection;
7008 : else
7009 0 : indexscandir = BackwardScanDirection;
7010 0 : break;
7011 0 : default:
7012 : /* index doesn't match the sortop */
7013 0 : continue;
7014 : }
7015 :
7016 : /*
7017 : * Found a suitable index to extract data from. Set up some data that
7018 : * can be used by both invocations of get_actual_variable_endpoint.
7019 : */
7020 : {
7021 : MemoryContext tmpcontext;
7022 : MemoryContext oldcontext;
7023 : Relation heapRel;
7024 : Relation indexRel;
7025 : TupleTableSlot *slot;
7026 : int16 typLen;
7027 : bool typByVal;
7028 : ScanKeyData scankeys[1];
7029 :
7030 : /* Make sure any cruft gets recycled when we're done */
7031 79442 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
7032 : "get_actual_variable_range workspace",
7033 : ALLOCSET_DEFAULT_SIZES);
7034 79442 : oldcontext = MemoryContextSwitchTo(tmpcontext);
7035 :
7036 : /*
7037 : * Open the table and index so we can read from them. We should
7038 : * already have some type of lock on each.
7039 : */
7040 79442 : heapRel = table_open(rte->relid, NoLock);
7041 79442 : indexRel = index_open(index->indexoid, NoLock);
7042 :
7043 : /* build some stuff needed for indexscan execution */
7044 79442 : slot = table_slot_create(heapRel, NULL);
7045 79442 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
7046 :
7047 : /* set up an IS NOT NULL scan key so that we ignore nulls */
7048 79442 : ScanKeyEntryInitialize(&scankeys[0],
7049 : SK_ISNULL | SK_SEARCHNOTNULL,
7050 : 1, /* index col to scan */
7051 : InvalidStrategy, /* no strategy */
7052 : InvalidOid, /* no strategy subtype */
7053 : InvalidOid, /* no collation */
7054 : InvalidOid, /* no reg proc for this */
7055 : (Datum) 0); /* constant */
7056 :
7057 : /* If min is requested ... */
7058 79442 : if (min)
7059 : {
7060 44573 : have_data = get_actual_variable_endpoint(heapRel,
7061 : indexRel,
7062 : indexscandir,
7063 : scankeys,
7064 : typLen,
7065 : typByVal,
7066 : slot,
7067 : oldcontext,
7068 : min);
7069 : }
7070 : else
7071 : {
7072 : /* If min not requested, still want to fetch max */
7073 34869 : have_data = true;
7074 : }
7075 :
7076 : /* If max is requested, and we didn't already fail ... */
7077 79442 : if (max && have_data)
7078 : {
7079 : /* scan in the opposite direction; all else is the same */
7080 36039 : have_data = get_actual_variable_endpoint(heapRel,
7081 : indexRel,
7082 36039 : -indexscandir,
7083 : scankeys,
7084 : typLen,
7085 : typByVal,
7086 : slot,
7087 : oldcontext,
7088 : max);
7089 : }
7090 :
7091 : /* Clean everything up */
7092 79442 : ExecDropSingleTupleTableSlot(slot);
7093 :
7094 79442 : index_close(indexRel, NoLock);
7095 79442 : table_close(heapRel, NoLock);
7096 :
7097 79442 : MemoryContextSwitchTo(oldcontext);
7098 79442 : MemoryContextDelete(tmpcontext);
7099 :
7100 : /* And we're done */
7101 79442 : break;
7102 : }
7103 : }
7104 :
7105 109543 : return have_data;
7106 : }
7107 :
7108 : /*
7109 : * Get one endpoint datum (min or max depending on indexscandir) from the
7110 : * specified index. Return true if successful, false if not.
7111 : * On success, endpoint value is stored to *endpointDatum (and copied into
7112 : * outercontext).
7113 : *
7114 : * scankeys is a 1-element scankey array set up to reject nulls.
7115 : * typLen/typByVal describe the datatype of the index's first column.
7116 : * tableslot is a slot suitable to hold table tuples, in case we need
7117 : * to probe the heap.
7118 : * (We could compute these values locally, but that would mean computing them
7119 : * twice when get_actual_variable_range needs both the min and the max.)
7120 : *
7121 : * Failure occurs either when the index is empty, or we decide that it's
7122 : * taking too long to find a suitable tuple.
7123 : */
7124 : static bool
7125 80612 : get_actual_variable_endpoint(Relation heapRel,
7126 : Relation indexRel,
7127 : ScanDirection indexscandir,
7128 : ScanKey scankeys,
7129 : int16 typLen,
7130 : bool typByVal,
7131 : TupleTableSlot *tableslot,
7132 : MemoryContext outercontext,
7133 : Datum *endpointDatum)
7134 : {
7135 80612 : bool have_data = false;
7136 : SnapshotData SnapshotNonVacuumable;
7137 : IndexScanDesc index_scan;
7138 80612 : Buffer vmbuffer = InvalidBuffer;
7139 80612 : BlockNumber last_heap_block = InvalidBlockNumber;
7140 80612 : int n_visited_heap_pages = 0;
7141 : ItemPointer tid;
7142 : Datum values[INDEX_MAX_KEYS];
7143 : bool isnull[INDEX_MAX_KEYS];
7144 : MemoryContext oldcontext;
7145 :
7146 : /*
7147 : * We use the index-only-scan machinery for this. With mostly-static
7148 : * tables that's a win because it avoids a heap visit. It's also a win
7149 : * for dynamic data, but the reason is less obvious; read on for details.
7150 : *
7151 : * In principle, we should scan the index with our current active
7152 : * snapshot, which is the best approximation we've got to what the query
7153 : * will see when executed. But that won't be exact if a new snap is taken
7154 : * before running the query, and it can be very expensive if a lot of
7155 : * recently-dead or uncommitted rows exist at the beginning or end of the
7156 : * index (because we'll laboriously fetch each one and reject it).
7157 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
7158 : * and uncommitted rows as well as normal visible rows. On the other
7159 : * hand, it will reject known-dead rows, and thus not give a bogus answer
7160 : * when the extreme value has been deleted (unless the deletion was quite
7161 : * recent); that case motivates not using SnapshotAny here.
7162 : *
7163 : * A crucial point here is that SnapshotNonVacuumable, with
7164 : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
7165 : * condition that the indexscan will use to decide that index entries are
7166 : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
7167 : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
7168 : * have to continue scanning past it, we know that the indexscan will mark
7169 : * that index entry killed. That means that the next
7170 : * get_actual_variable_endpoint() call will not have to re-consider that
7171 : * index entry. In this way we avoid repetitive work when this function
7172 : * is used a lot during planning.
7173 : *
7174 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
7175 : * recently-created index, some index entries may point at "broken" HOT
7176 : * chains in which not all the tuple versions contain data matching the
7177 : * index entry. The live tuple version(s) certainly do match the index,
7178 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
7179 : * don't match. Hence, if we took data from the selected heap tuple, we
7180 : * might get a bogus answer that's not close to the index extremal value,
7181 : * or could even be NULL. We avoid this hazard because we take the data
7182 : * from the index entry not the heap.
7183 : *
7184 : * Despite all this care, there are situations where we might find many
7185 : * non-visible tuples near the end of the index. We don't want to expend
7186 : * a huge amount of time here, so we give up once we've read too many heap
7187 : * pages. When we fail for that reason, the caller will end up using
7188 : * whatever extremal value is recorded in pg_statistic.
7189 : */
7190 80612 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
7191 : GlobalVisTestFor(heapRel));
7192 :
7193 80612 : index_scan = index_beginscan(heapRel, indexRel,
7194 : &SnapshotNonVacuumable, NULL,
7195 : 1, 0,
7196 : SO_NONE);
7197 : /* Set it up for index-only scan */
7198 80612 : index_scan->xs_want_itup = true;
7199 80612 : index_rescan(index_scan, scankeys, 1, NULL, 0);
7200 :
7201 : /* Fetch first/next tuple in specified direction */
7202 94440 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
7203 : {
7204 94440 : BlockNumber block = ItemPointerGetBlockNumber(tid);
7205 :
7206 94440 : if (!VM_ALL_VISIBLE(heapRel,
7207 : block,
7208 : &vmbuffer))
7209 : {
7210 : /* Rats, we have to visit the heap to check visibility */
7211 63201 : if (!index_fetch_heap(index_scan, tableslot))
7212 : {
7213 : /*
7214 : * No visible tuple for this index entry, so we need to
7215 : * advance to the next entry. Before doing so, count heap
7216 : * page fetches and give up if we've done too many.
7217 : *
7218 : * We don't charge a page fetch if this is the same heap page
7219 : * as the previous tuple. This is on the conservative side,
7220 : * since other recently-accessed pages are probably still in
7221 : * buffers too; but it's good enough for this heuristic.
7222 : */
7223 : #define VISITED_PAGES_LIMIT 100
7224 :
7225 13828 : if (block != last_heap_block)
7226 : {
7227 1891 : last_heap_block = block;
7228 1891 : n_visited_heap_pages++;
7229 1891 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
7230 0 : break;
7231 : }
7232 :
7233 13828 : continue; /* no visible tuple, try next index entry */
7234 : }
7235 :
7236 : /* We don't actually need the heap tuple for anything */
7237 49373 : ExecClearTuple(tableslot);
7238 :
7239 : /*
7240 : * We don't care whether there's more than one visible tuple in
7241 : * the HOT chain; if any are visible, that's good enough.
7242 : */
7243 : }
7244 :
7245 : /*
7246 : * We expect that the index will return data in IndexTuple not
7247 : * HeapTuple format.
7248 : */
7249 80612 : if (!index_scan->xs_itup)
7250 0 : elog(ERROR, "no data returned for index-only scan");
7251 :
7252 : /*
7253 : * We do not yet support recheck here.
7254 : */
7255 80612 : if (index_scan->xs_recheck)
7256 0 : break;
7257 :
7258 : /* OK to deconstruct the index tuple */
7259 80612 : index_deform_tuple(index_scan->xs_itup,
7260 : index_scan->xs_itupdesc,
7261 : values, isnull);
7262 :
7263 : /* Shouldn't have got a null, but be careful */
7264 80612 : if (isnull[0])
7265 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
7266 : RelationGetRelationName(indexRel));
7267 :
7268 : /* Copy the index column value out to caller's context */
7269 80612 : oldcontext = MemoryContextSwitchTo(outercontext);
7270 80612 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
7271 80612 : MemoryContextSwitchTo(oldcontext);
7272 80612 : have_data = true;
7273 80612 : break;
7274 : }
7275 :
7276 80612 : if (vmbuffer != InvalidBuffer)
7277 73679 : ReleaseBuffer(vmbuffer);
7278 80612 : index_endscan(index_scan);
7279 :
7280 80612 : return have_data;
7281 : }
7282 :
7283 : /*
7284 : * find_join_input_rel
7285 : * Look up the input relation for a join.
7286 : *
7287 : * We assume that the input relation's RelOptInfo must have been constructed
7288 : * already.
7289 : */
7290 : static RelOptInfo *
7291 11408 : find_join_input_rel(PlannerInfo *root, Relids relids)
7292 : {
7293 11408 : RelOptInfo *rel = NULL;
7294 :
7295 11408 : if (!bms_is_empty(relids))
7296 : {
7297 : int relid;
7298 :
7299 11408 : if (bms_get_singleton_member(relids, &relid))
7300 11132 : rel = find_base_rel(root, relid);
7301 : else
7302 276 : rel = find_join_rel(root, relids);
7303 : }
7304 :
7305 11408 : if (rel == NULL)
7306 0 : elog(ERROR, "could not find RelOptInfo for given relids");
7307 :
7308 11408 : return rel;
7309 : }
7310 :
7311 :
7312 : /*-------------------------------------------------------------------------
7313 : *
7314 : * Index cost estimation functions
7315 : *
7316 : *-------------------------------------------------------------------------
7317 : */
7318 :
7319 : /*
7320 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
7321 : */
7322 : List *
7323 668612 : get_quals_from_indexclauses(List *indexclauses)
7324 : {
7325 668612 : List *result = NIL;
7326 : ListCell *lc;
7327 :
7328 1161057 : foreach(lc, indexclauses)
7329 : {
7330 492445 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7331 : ListCell *lc2;
7332 :
7333 987321 : foreach(lc2, iclause->indexquals)
7334 : {
7335 494876 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7336 :
7337 494876 : result = lappend(result, rinfo);
7338 : }
7339 : }
7340 668612 : return result;
7341 : }
7342 :
7343 : /*
7344 : * Compute the total evaluation cost of the comparison operands in a list
7345 : * of index qual expressions. Since we know these will be evaluated just
7346 : * once per scan, there's no need to distinguish startup from per-row cost.
7347 : *
7348 : * This can be used either on the result of get_quals_from_indexclauses(),
7349 : * or directly on an indexorderbys list. In both cases, we expect that the
7350 : * index key expression is on the left side of binary clauses.
7351 : */
7352 : Cost
7353 1326724 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
7354 : {
7355 1326724 : Cost qual_arg_cost = 0;
7356 : ListCell *lc;
7357 :
7358 1821949 : foreach(lc, indexquals)
7359 : {
7360 495225 : Expr *clause = (Expr *) lfirst(lc);
7361 : Node *other_operand;
7362 : QualCost index_qual_cost;
7363 :
7364 : /*
7365 : * Index quals will have RestrictInfos, indexorderbys won't. Look
7366 : * through RestrictInfo if present.
7367 : */
7368 495225 : if (IsA(clause, RestrictInfo))
7369 494866 : clause = ((RestrictInfo *) clause)->clause;
7370 :
7371 495225 : if (IsA(clause, OpExpr))
7372 : {
7373 479816 : OpExpr *op = (OpExpr *) clause;
7374 :
7375 479816 : other_operand = (Node *) lsecond(op->args);
7376 : }
7377 15409 : else if (IsA(clause, RowCompareExpr))
7378 : {
7379 330 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7380 :
7381 330 : other_operand = (Node *) rc->rargs;
7382 : }
7383 15079 : else if (IsA(clause, ScalarArrayOpExpr))
7384 : {
7385 12640 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7386 :
7387 12640 : other_operand = (Node *) lsecond(saop->args);
7388 : }
7389 2439 : else if (IsA(clause, NullTest))
7390 : {
7391 2439 : other_operand = NULL;
7392 : }
7393 : else
7394 : {
7395 0 : elog(ERROR, "unsupported indexqual type: %d",
7396 : (int) nodeTag(clause));
7397 : other_operand = NULL; /* keep compiler quiet */
7398 : }
7399 :
7400 495225 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
7401 495225 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
7402 : }
7403 1326724 : return qual_arg_cost;
7404 : }
7405 :
7406 : /*
7407 : * Compute generic index access cost estimates.
7408 : *
7409 : * See struct GenericCosts in selfuncs.h for more info.
7410 : */
7411 : void
7412 658122 : genericcostestimate(PlannerInfo *root,
7413 : IndexPath *path,
7414 : double loop_count,
7415 : GenericCosts *costs)
7416 : {
7417 658122 : IndexOptInfo *index = path->indexinfo;
7418 658122 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7419 658122 : List *indexOrderBys = path->indexorderbys;
7420 : Cost indexStartupCost;
7421 : Cost indexTotalCost;
7422 : Selectivity indexSelectivity;
7423 : double indexCorrelation;
7424 : double numIndexPages;
7425 : double numIndexTuples;
7426 : double spc_random_page_cost;
7427 : double num_sa_scans;
7428 : double num_outer_scans;
7429 : double num_scans;
7430 : double qual_op_cost;
7431 : double qual_arg_cost;
7432 : List *selectivityQuals;
7433 : ListCell *l;
7434 :
7435 : /*
7436 : * If the index is partial, AND the index predicate with the explicitly
7437 : * given indexquals to produce a more accurate idea of the index
7438 : * selectivity.
7439 : */
7440 658122 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7441 :
7442 : /*
7443 : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
7444 : * just assume that the number of index descents is the number of distinct
7445 : * combinations of array elements from all of the scan's SAOP clauses.
7446 : */
7447 658122 : num_sa_scans = costs->num_sa_scans;
7448 658122 : if (num_sa_scans < 1)
7449 : {
7450 6968 : num_sa_scans = 1;
7451 15246 : foreach(l, indexQuals)
7452 : {
7453 8278 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
7454 :
7455 8278 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
7456 : {
7457 46 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
7458 46 : double alength = estimate_array_length(root, lsecond(saop->args));
7459 :
7460 46 : if (alength > 1)
7461 46 : num_sa_scans *= alength;
7462 : }
7463 : }
7464 : }
7465 :
7466 : /* Estimate the fraction of main-table tuples that will be visited */
7467 658122 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7468 658122 : index->rel->relid,
7469 : JOIN_INNER,
7470 : NULL);
7471 :
7472 : /*
7473 : * If caller didn't give us an estimate, estimate the number of index
7474 : * tuples that will be visited. We do it in this rather peculiar-looking
7475 : * way in order to get the right answer for partial indexes.
7476 : */
7477 658122 : numIndexTuples = costs->numIndexTuples;
7478 658122 : if (numIndexTuples <= 0.0)
7479 : {
7480 84875 : numIndexTuples = indexSelectivity * index->rel->tuples;
7481 :
7482 : /*
7483 : * The above calculation counts all the tuples visited across all
7484 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
7485 : * average per-indexscan number, so adjust. This is a handy place to
7486 : * round to integer, too. (If caller supplied tuple estimate, it's
7487 : * responsible for handling these considerations.)
7488 : */
7489 84875 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7490 : }
7491 :
7492 : /*
7493 : * We can bound the number of tuples by the index size in any case. Also,
7494 : * always estimate at least one tuple is touched, even when
7495 : * indexSelectivity estimate is tiny.
7496 : */
7497 658122 : if (numIndexTuples > index->tuples)
7498 8525 : numIndexTuples = index->tuples;
7499 658122 : if (numIndexTuples < 1.0)
7500 88574 : numIndexTuples = 1.0;
7501 :
7502 : /*
7503 : * Estimate the number of index pages that will be retrieved.
7504 : *
7505 : * We use the simplistic method of taking a pro-rata fraction of the total
7506 : * number of index leaf pages. We disregard any overhead such as index
7507 : * metapages or upper tree levels.
7508 : *
7509 : * In practice access to upper index levels is often nearly free because
7510 : * those tend to stay in cache under load; moreover, the cost involved is
7511 : * highly dependent on index type. We therefore ignore such costs here
7512 : * and leave it to the caller to add a suitable charge if needed.
7513 : */
7514 658122 : if (index->pages > costs->numNonLeafPages && index->tuples > 1)
7515 591029 : numIndexPages =
7516 591029 : ceil(numIndexTuples * (index->pages - costs->numNonLeafPages)
7517 591029 : / index->tuples);
7518 : else
7519 67093 : numIndexPages = 1.0;
7520 :
7521 : /* fetch estimated page cost for tablespace containing index */
7522 658122 : get_tablespace_page_costs(index->reltablespace,
7523 : &spc_random_page_cost,
7524 : NULL);
7525 :
7526 : /*
7527 : * Now compute the disk access costs.
7528 : *
7529 : * The above calculations are all per-index-scan. However, if we are in a
7530 : * nestloop inner scan, we can expect the scan to be repeated (with
7531 : * different search keys) for each row of the outer relation. Likewise,
7532 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
7533 : * the potential for cache effects to reduce the number of disk page
7534 : * fetches needed. We want to estimate the average per-scan I/O cost in
7535 : * the presence of caching.
7536 : *
7537 : * We use the Mackert-Lohman formula (see costsize.c for details) to
7538 : * estimate the total number of page fetches that occur. While this
7539 : * wasn't what it was designed for, it seems a reasonable model anyway.
7540 : * Note that we are counting pages not tuples anymore, so we take N = T =
7541 : * index size, as if there were one "tuple" per page.
7542 : */
7543 658122 : num_outer_scans = loop_count;
7544 658122 : num_scans = num_sa_scans * num_outer_scans;
7545 :
7546 658122 : if (num_scans > 1)
7547 : {
7548 : double pages_fetched;
7549 :
7550 : /* total page fetches ignoring cache effects */
7551 80658 : pages_fetched = numIndexPages * num_scans;
7552 :
7553 : /* use Mackert and Lohman formula to adjust for cache effects */
7554 80658 : pages_fetched = index_pages_fetched(pages_fetched,
7555 : index->pages,
7556 80658 : (double) index->pages,
7557 : root);
7558 :
7559 : /*
7560 : * Now compute the total disk access cost, and then report a pro-rated
7561 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
7562 : * since that's internal to the indexscan.)
7563 : */
7564 80658 : indexTotalCost = (pages_fetched * spc_random_page_cost)
7565 : / num_outer_scans;
7566 : }
7567 : else
7568 : {
7569 : /*
7570 : * For a single index scan, we just charge spc_random_page_cost per
7571 : * page touched.
7572 : */
7573 577464 : indexTotalCost = numIndexPages * spc_random_page_cost;
7574 : }
7575 :
7576 : /*
7577 : * CPU cost: any complex expressions in the indexquals will need to be
7578 : * evaluated once at the start of the scan to reduce them to runtime keys
7579 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
7580 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
7581 : * indexqual operator. Because we have numIndexTuples as a per-scan
7582 : * number, we have to multiply by num_sa_scans to get the correct result
7583 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
7584 : * ORDER BY expressions.
7585 : *
7586 : * Note: this neglects the possible costs of rechecking lossy operators.
7587 : * Detecting that that might be needed seems more expensive than it's
7588 : * worth, though, considering all the other inaccuracies here ...
7589 : */
7590 658122 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
7591 658122 : index_other_operands_eval_cost(root, indexOrderBys);
7592 658122 : qual_op_cost = cpu_operator_cost *
7593 658122 : (list_length(indexQuals) + list_length(indexOrderBys));
7594 :
7595 658122 : indexStartupCost = qual_arg_cost;
7596 658122 : indexTotalCost += qual_arg_cost;
7597 658122 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
7598 :
7599 : /*
7600 : * Generic assumption about index correlation: there isn't any.
7601 : */
7602 658122 : indexCorrelation = 0.0;
7603 :
7604 : /*
7605 : * Return everything to caller.
7606 : */
7607 658122 : costs->indexStartupCost = indexStartupCost;
7608 658122 : costs->indexTotalCost = indexTotalCost;
7609 658122 : costs->indexSelectivity = indexSelectivity;
7610 658122 : costs->indexCorrelation = indexCorrelation;
7611 658122 : costs->numIndexPages = numIndexPages;
7612 658122 : costs->numIndexTuples = numIndexTuples;
7613 658122 : costs->spc_random_page_cost = spc_random_page_cost;
7614 658122 : costs->num_sa_scans = num_sa_scans;
7615 658122 : }
7616 :
7617 : /*
7618 : * If the index is partial, add its predicate to the given qual list.
7619 : *
7620 : * ANDing the index predicate with the explicitly given indexquals produces
7621 : * a more accurate idea of the index's selectivity. However, we need to be
7622 : * careful not to insert redundant clauses, because clauselist_selectivity()
7623 : * is easily fooled into computing a too-low selectivity estimate. Our
7624 : * approach is to add only the predicate clause(s) that cannot be proven to
7625 : * be implied by the given indexquals. This successfully handles cases such
7626 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
7627 : * There are many other cases where we won't detect redundancy, leading to a
7628 : * too-low selectivity estimate, which will bias the system in favor of using
7629 : * partial indexes where possible. That is not necessarily bad though.
7630 : *
7631 : * Note that indexQuals contains RestrictInfo nodes while the indpred
7632 : * does not, so the output list will be mixed. This is OK for both
7633 : * predicate_implied_by() and clauselist_selectivity(), but might be
7634 : * problematic if the result were passed to other things.
7635 : */
7636 : List *
7637 1130623 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
7638 : {
7639 1130623 : List *predExtraQuals = NIL;
7640 : ListCell *lc;
7641 :
7642 1130623 : if (index->indpred == NIL)
7643 1129087 : return indexQuals;
7644 :
7645 3082 : foreach(lc, index->indpred)
7646 : {
7647 1546 : Node *predQual = (Node *) lfirst(lc);
7648 1546 : List *oneQual = list_make1(predQual);
7649 :
7650 1546 : if (!predicate_implied_by(oneQual, indexQuals, false))
7651 1370 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7652 : }
7653 1536 : return list_concat(predExtraQuals, indexQuals);
7654 : }
7655 :
7656 : /*
7657 : * Estimate correlation of btree index's first column.
7658 : *
7659 : * If we can get an estimate of the first column's ordering correlation C
7660 : * from pg_statistic, estimate the index correlation as C for a single-column
7661 : * index, or C * 0.75 for multiple columns. The idea here is that multiple
7662 : * columns dilute the importance of the first column's ordering, but don't
7663 : * negate it entirely.
7664 : *
7665 : * We already filled in the stats tuple for *vardata when called.
7666 : */
7667 : static double
7668 426477 : btcost_correlation(IndexOptInfo *index, VariableStatData *vardata)
7669 : {
7670 : Oid sortop;
7671 : AttStatsSlot sslot;
7672 426477 : double indexCorrelation = 0;
7673 :
7674 : Assert(HeapTupleIsValid(vardata->statsTuple));
7675 :
7676 426477 : sortop = get_opfamily_member(index->opfamily[0],
7677 426477 : index->opcintype[0],
7678 426477 : index->opcintype[0],
7679 : BTLessStrategyNumber);
7680 852954 : if (OidIsValid(sortop) &&
7681 426477 : get_attstatsslot(&sslot, vardata->statsTuple,
7682 : STATISTIC_KIND_CORRELATION, sortop,
7683 : ATTSTATSSLOT_NUMBERS))
7684 : {
7685 : double varCorrelation;
7686 :
7687 : Assert(sslot.nnumbers == 1);
7688 420951 : varCorrelation = sslot.numbers[0];
7689 :
7690 420951 : if (index->reverse_sort[0])
7691 0 : varCorrelation = -varCorrelation;
7692 :
7693 420951 : if (index->nkeycolumns > 1)
7694 146471 : indexCorrelation = varCorrelation * 0.75;
7695 : else
7696 274480 : indexCorrelation = varCorrelation;
7697 :
7698 420951 : free_attstatsslot(&sslot);
7699 : }
7700 :
7701 426477 : return indexCorrelation;
7702 : }
7703 :
7704 : void
7705 651154 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7706 : Cost *indexStartupCost, Cost *indexTotalCost,
7707 : Selectivity *indexSelectivity, double *indexCorrelation,
7708 : double *indexPages)
7709 : {
7710 651154 : IndexOptInfo *index = path->indexinfo;
7711 651154 : GenericCosts costs = {0};
7712 651154 : VariableStatData vardata = {0};
7713 : double numIndexTuples;
7714 : Cost descentCost;
7715 : List *indexBoundQuals;
7716 : List *indexSkipQuals;
7717 : int indexcol;
7718 : bool eqQualHere;
7719 : bool found_row_compare;
7720 : bool found_array;
7721 : bool found_is_null_op;
7722 651154 : bool have_correlation = false;
7723 : double num_sa_scans;
7724 651154 : double correlation = 0.0;
7725 : ListCell *lc;
7726 :
7727 : /*
7728 : * For a btree scan, only leading '=' quals plus inequality quals for the
7729 : * immediately next attribute contribute to index selectivity (these are
7730 : * the "boundary quals" that determine the starting and stopping points of
7731 : * the index scan). Additional quals can suppress visits to the heap, so
7732 : * it's OK to count them in indexSelectivity, but they should not count
7733 : * for estimating numIndexTuples. So we must examine the given indexquals
7734 : * to find out which ones count as boundary quals. We rely on the
7735 : * knowledge that they are given in index column order. Note that nbtree
7736 : * preprocessing can add skip arrays that act as leading '=' quals in the
7737 : * absence of ordinary input '=' quals, so in practice _most_ input quals
7738 : * are able to act as index bound quals (which we take into account here).
7739 : *
7740 : * For a RowCompareExpr, we consider only the first column, just as
7741 : * rowcomparesel() does.
7742 : *
7743 : * If there's a SAOP or skip array in the quals, we'll actually perform up
7744 : * to N index descents (not just one), but the underlying array key's
7745 : * operator can be considered to act the same as it normally does.
7746 : */
7747 651154 : indexBoundQuals = NIL;
7748 651154 : indexSkipQuals = NIL;
7749 651154 : indexcol = 0;
7750 651154 : eqQualHere = false;
7751 651154 : found_row_compare = false;
7752 651154 : found_array = false;
7753 651154 : found_is_null_op = false;
7754 651154 : num_sa_scans = 1;
7755 1094650 : foreach(lc, path->indexclauses)
7756 : {
7757 471312 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7758 : ListCell *lc2;
7759 :
7760 471312 : if (indexcol < iclause->indexcol)
7761 : {
7762 93471 : double num_sa_scans_prev_cols = num_sa_scans;
7763 :
7764 : /*
7765 : * Beginning of a new column's quals.
7766 : *
7767 : * Skip scans use skip arrays, which are ScalarArrayOp style
7768 : * arrays that generate their elements procedurally and on demand.
7769 : * Given a multi-column index on "(a, b)", and an SQL WHERE clause
7770 : * "WHERE b = 42", a skip scan will effectively use an indexqual
7771 : * "WHERE a = ANY('{every col a value}') AND b = 42". (Obviously,
7772 : * the array on "a" must also return "IS NULL" matches, since our
7773 : * WHERE clause used no strict operator on "a").
7774 : *
7775 : * Here we consider how nbtree will backfill skip arrays for any
7776 : * index columns that lacked an '=' qual. This maintains our
7777 : * num_sa_scans estimate, and determines if this new column (the
7778 : * "iclause->indexcol" column, not the prior "indexcol" column)
7779 : * can have its RestrictInfos/quals added to indexBoundQuals.
7780 : *
7781 : * We'll need to handle columns that have inequality quals, where
7782 : * the skip array generates values from a range constrained by the
7783 : * quals (not every possible value). We've been maintaining
7784 : * indexSkipQuals to help with this; it will now contain all of
7785 : * the prior column's quals (that is, indexcol's quals) when they
7786 : * might be used for this.
7787 : */
7788 93471 : if (found_row_compare)
7789 : {
7790 : /*
7791 : * Skip arrays can't be added after a RowCompare input qual
7792 : * due to limitations in nbtree
7793 : */
7794 20 : break;
7795 : }
7796 93451 : if (eqQualHere)
7797 : {
7798 : /*
7799 : * Don't need to add a skip array for an indexcol that already
7800 : * has an '=' qual/equality constraint
7801 : */
7802 66128 : indexcol++;
7803 66128 : indexSkipQuals = NIL;
7804 : }
7805 93451 : eqQualHere = false;
7806 :
7807 95052 : while (indexcol < iclause->indexcol)
7808 : {
7809 : double ndistinct;
7810 29397 : bool isdefault = true;
7811 :
7812 29397 : found_array = true;
7813 :
7814 : /*
7815 : * A skipped attribute's ndistinct forms the basis of our
7816 : * estimate of the total number of "array elements" used by
7817 : * its skip array at runtime. Look that up first.
7818 : */
7819 29397 : examine_indexcol_variable(root, index, indexcol, &vardata);
7820 29397 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
7821 :
7822 29397 : if (indexcol == 0)
7823 : {
7824 : /*
7825 : * Get an estimate of the leading column's correlation in
7826 : * passing (avoids rereading variable stats below)
7827 : */
7828 27313 : if (HeapTupleIsValid(vardata.statsTuple))
7829 16014 : correlation = btcost_correlation(index, &vardata);
7830 27313 : have_correlation = true;
7831 : }
7832 :
7833 29397 : ReleaseVariableStats(vardata);
7834 :
7835 : /*
7836 : * If ndistinct is a default estimate, conservatively assume
7837 : * that no skipping will happen at runtime
7838 : */
7839 29397 : if (isdefault)
7840 : {
7841 8973 : num_sa_scans = num_sa_scans_prev_cols;
7842 27796 : break; /* done building indexBoundQuals */
7843 : }
7844 :
7845 : /*
7846 : * Apply indexcol's indexSkipQuals selectivity to ndistinct
7847 : */
7848 20424 : if (indexSkipQuals != NIL)
7849 : {
7850 : List *partialSkipQuals;
7851 : Selectivity ndistinctfrac;
7852 :
7853 : /*
7854 : * If the index is partial, AND the index predicate with
7855 : * the index-bound quals to produce a more accurate idea
7856 : * of the number of distinct values for prior indexcol
7857 : */
7858 552 : partialSkipQuals = add_predicate_to_index_quals(index,
7859 : indexSkipQuals);
7860 :
7861 552 : ndistinctfrac = clauselist_selectivity(root, partialSkipQuals,
7862 552 : index->rel->relid,
7863 : JOIN_INNER,
7864 : NULL);
7865 :
7866 : /*
7867 : * If ndistinctfrac is selective (on its own), the scan is
7868 : * unlikely to benefit from repositioning itself using
7869 : * later quals. Do not allow iclause->indexcol's quals to
7870 : * be added to indexBoundQuals (it would increase descent
7871 : * costs, without lowering numIndexTuples costs by much).
7872 : */
7873 552 : if (ndistinctfrac < DEFAULT_RANGE_INEQ_SEL)
7874 : {
7875 311 : num_sa_scans = num_sa_scans_prev_cols;
7876 311 : break; /* done building indexBoundQuals */
7877 : }
7878 :
7879 : /* Adjust ndistinct downward */
7880 241 : ndistinct = rint(ndistinct * ndistinctfrac);
7881 241 : ndistinct = Max(ndistinct, 1);
7882 : }
7883 :
7884 : /*
7885 : * When there's no inequality quals, account for the need to
7886 : * find an initial value by counting -inf/+inf as a value.
7887 : *
7888 : * We don't charge anything extra for possible next/prior key
7889 : * index probes, which are sometimes used to find the next
7890 : * valid skip array element (ahead of using the located
7891 : * element value to relocate the scan to the next position
7892 : * that might contain matching tuples). It seems hard to do
7893 : * better here. Use of the skip support infrastructure often
7894 : * avoids most next/prior key probes. But even when it can't,
7895 : * there's a decent chance that most individual next/prior key
7896 : * probes will locate a leaf page whose key space overlaps all
7897 : * of the scan's keys (even the lower-order keys) -- which
7898 : * also avoids the need for a separate, extra index descent.
7899 : * Note also that these probes are much cheaper than non-probe
7900 : * primitive index scans: they're reliably very selective.
7901 : */
7902 20113 : if (indexSkipQuals == NIL)
7903 19872 : ndistinct += 1;
7904 :
7905 : /*
7906 : * Update num_sa_scans estimate by multiplying by ndistinct.
7907 : *
7908 : * We make the pessimistic assumption that there is no
7909 : * naturally occurring cross-column correlation. This is
7910 : * often wrong, but it seems best to err on the side of not
7911 : * expecting skipping to be helpful...
7912 : */
7913 20113 : num_sa_scans *= ndistinct;
7914 :
7915 : /*
7916 : * ...but back out of adding this latest group of 1 or more
7917 : * skip arrays when num_sa_scans exceeds the total number of
7918 : * index pages (revert to num_sa_scans from before indexcol).
7919 : * This causes a sharp discontinuity in cost (as a function of
7920 : * the indexcol's ndistinct), but that is representative of
7921 : * actual runtime costs.
7922 : *
7923 : * Note that skipping is helpful when each primitive index
7924 : * scan only manages to skip over 1 or 2 irrelevant leaf pages
7925 : * on average. Skip arrays bring savings in CPU costs due to
7926 : * the scan not needing to evaluate indexquals against every
7927 : * tuple, which can greatly exceed any savings in I/O costs.
7928 : * This test is a test of whether num_sa_scans implies that
7929 : * we're past the point where the ability to skip ceases to
7930 : * lower the scan's costs (even qual evaluation CPU costs).
7931 : */
7932 20113 : if (index->pages < num_sa_scans)
7933 : {
7934 18512 : num_sa_scans = num_sa_scans_prev_cols;
7935 18512 : break; /* done building indexBoundQuals */
7936 : }
7937 :
7938 1601 : indexcol++;
7939 1601 : indexSkipQuals = NIL;
7940 : }
7941 :
7942 : /*
7943 : * Finished considering the need to add skip arrays to bridge an
7944 : * initial eqQualHere gap between the old and new index columns
7945 : * (or there was no initial eqQualHere gap in the first place).
7946 : *
7947 : * If an initial gap could not be bridged, then new column's quals
7948 : * (i.e. iclause->indexcol's quals) won't go into indexBoundQuals,
7949 : * and so won't affect our final numIndexTuples estimate.
7950 : */
7951 93451 : if (indexcol != iclause->indexcol)
7952 27796 : break; /* done building indexBoundQuals */
7953 : }
7954 :
7955 : Assert(indexcol == iclause->indexcol);
7956 :
7957 : /* Examine each indexqual associated with this index clause */
7958 889297 : foreach(lc2, iclause->indexquals)
7959 : {
7960 445801 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7961 445801 : Expr *clause = rinfo->clause;
7962 445801 : Oid clause_op = InvalidOid;
7963 : int op_strategy;
7964 :
7965 445801 : if (IsA(clause, OpExpr))
7966 : {
7967 431345 : OpExpr *op = (OpExpr *) clause;
7968 :
7969 431345 : clause_op = op->opno;
7970 : }
7971 14456 : else if (IsA(clause, RowCompareExpr))
7972 : {
7973 330 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7974 :
7975 330 : clause_op = linitial_oid(rc->opnos);
7976 330 : found_row_compare = true;
7977 : }
7978 14126 : else if (IsA(clause, ScalarArrayOpExpr))
7979 : {
7980 12228 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7981 12228 : Node *other_operand = (Node *) lsecond(saop->args);
7982 12228 : double alength = estimate_array_length(root, other_operand);
7983 :
7984 12228 : clause_op = saop->opno;
7985 12228 : found_array = true;
7986 : /* estimate SA descents by indexBoundQuals only */
7987 12228 : if (alength > 1)
7988 12004 : num_sa_scans *= alength;
7989 : }
7990 1898 : else if (IsA(clause, NullTest))
7991 : {
7992 1898 : NullTest *nt = (NullTest *) clause;
7993 :
7994 1898 : if (nt->nulltesttype == IS_NULL)
7995 : {
7996 200 : found_is_null_op = true;
7997 : /* IS NULL is like = for selectivity/skip scan purposes */
7998 200 : eqQualHere = true;
7999 : }
8000 : }
8001 : else
8002 0 : elog(ERROR, "unsupported indexqual type: %d",
8003 : (int) nodeTag(clause));
8004 :
8005 : /* check for equality operator */
8006 445801 : if (OidIsValid(clause_op))
8007 : {
8008 443903 : op_strategy = get_op_opfamily_strategy(clause_op,
8009 443903 : index->opfamily[indexcol]);
8010 : Assert(op_strategy != 0); /* not a member of opfamily?? */
8011 443903 : if (op_strategy == BTEqualStrategyNumber)
8012 420057 : eqQualHere = true;
8013 : }
8014 :
8015 445801 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
8016 :
8017 : /*
8018 : * We apply inequality selectivities to estimate index descent
8019 : * costs with scans that use skip arrays. Save this indexcol's
8020 : * RestrictInfos if it looks like they'll be needed for that.
8021 : */
8022 445801 : if (!eqQualHere && !found_row_compare &&
8023 24647 : indexcol < index->nkeycolumns - 1)
8024 4772 : indexSkipQuals = lappend(indexSkipQuals, rinfo);
8025 : }
8026 : }
8027 :
8028 : /*
8029 : * If index is unique and we found an '=' clause for each column, we can
8030 : * just assume numIndexTuples = 1 and skip the expensive
8031 : * clauselist_selectivity calculations. However, an array or NullTest
8032 : * always invalidates that theory (even when eqQualHere has been set).
8033 : */
8034 651154 : if (index->unique &&
8035 520558 : indexcol == index->nkeycolumns - 1 &&
8036 186236 : eqQualHere &&
8037 186236 : !found_array &&
8038 180791 : !found_is_null_op)
8039 180751 : numIndexTuples = 1.0;
8040 : else
8041 : {
8042 : List *selectivityQuals;
8043 : Selectivity btreeSelectivity;
8044 :
8045 : /*
8046 : * If the index is partial, AND the index predicate with the
8047 : * index-bound quals to produce a more accurate idea of the number of
8048 : * rows covered by the bound conditions.
8049 : */
8050 470403 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
8051 :
8052 470403 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
8053 470403 : index->rel->relid,
8054 : JOIN_INNER,
8055 : NULL);
8056 470403 : numIndexTuples = btreeSelectivity * index->rel->tuples;
8057 :
8058 : /*
8059 : * btree automatically combines individual array element primitive
8060 : * index scans whenever the tuples covered by the next set of array
8061 : * keys are close to tuples covered by the current set. That puts a
8062 : * natural ceiling on the worst case number of descents -- there
8063 : * cannot possibly be more than one descent per leaf page scanned.
8064 : *
8065 : * Clamp the number of descents to at most 1/3 the number of index
8066 : * pages. This avoids implausibly high estimates with low selectivity
8067 : * paths, where scans usually require only one or two descents. This
8068 : * is most likely to help when there are several SAOP clauses, where
8069 : * naively accepting the total number of distinct combinations of
8070 : * array elements as the number of descents would frequently lead to
8071 : * wild overestimates.
8072 : *
8073 : * We somewhat arbitrarily don't just make the cutoff the total number
8074 : * of leaf pages (we make it 1/3 the total number of pages instead) to
8075 : * give the btree code credit for its ability to continue on the leaf
8076 : * level with low selectivity scans.
8077 : *
8078 : * Note: num_sa_scans includes both ScalarArrayOp array elements and
8079 : * skip array elements whose qual affects our numIndexTuples estimate.
8080 : */
8081 470403 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
8082 470403 : num_sa_scans = Max(num_sa_scans, 1);
8083 :
8084 : /*
8085 : * As in genericcostestimate(), we have to adjust for any array quals
8086 : * included in indexBoundQuals, and then round to integer.
8087 : *
8088 : * It is tempting to make genericcostestimate behave as if array
8089 : * clauses work in almost the same way as scalar operators during
8090 : * btree scans, making the top-level scan look like a continuous scan
8091 : * (as opposed to num_sa_scans-many primitive index scans). After
8092 : * all, btree scans mostly work like that at runtime. However, such a
8093 : * scheme would badly bias genericcostestimate's simplistic approach
8094 : * to calculating numIndexPages through prorating.
8095 : *
8096 : * Stick with the approach taken by non-native SAOP scans for now.
8097 : * genericcostestimate will use the Mackert-Lohman formula to
8098 : * compensate for repeat page fetches, even though that definitely
8099 : * won't happen during btree scans (not for leaf pages, at least).
8100 : * We're usually very pessimistic about the number of primitive index
8101 : * scans that will be required, but it's not clear how to do better.
8102 : */
8103 470403 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
8104 : }
8105 :
8106 : /*
8107 : * Now do generic index cost estimation.
8108 : *
8109 : * While we expended effort to make realistic estimates of numIndexTuples
8110 : * and num_sa_scans, we are content to count only the btree metapage as
8111 : * non-leaf. btree fanout is typically high enough that upper pages are
8112 : * few relative to leaf pages, so accounting for them would move the
8113 : * estimates at most a percent or two. Given the uncertainty in just how
8114 : * many upper pages exist in a particular index, we'll skip trying to
8115 : * handle that.
8116 : */
8117 651154 : costs.numIndexTuples = numIndexTuples;
8118 651154 : costs.num_sa_scans = num_sa_scans;
8119 651154 : costs.numNonLeafPages = 1;
8120 :
8121 651154 : genericcostestimate(root, path, loop_count, &costs);
8122 :
8123 : /*
8124 : * Add a CPU-cost component to represent the costs of initial btree
8125 : * descent. We don't charge any I/O cost for touching upper btree levels,
8126 : * since they tend to stay in cache, but we still have to do about log2(N)
8127 : * comparisons to descend a btree of N leaf tuples. We charge one
8128 : * cpu_operator_cost per comparison.
8129 : *
8130 : * If there are SAOP or skip array keys, charge this once per estimated
8131 : * index descent. The ones after the first one are not startup cost so
8132 : * far as the overall plan goes, so just add them to "total" cost.
8133 : */
8134 651154 : if (index->tuples > 1) /* avoid computing log(0) */
8135 : {
8136 589676 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
8137 589676 : costs.indexStartupCost += descentCost;
8138 589676 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8139 : }
8140 :
8141 : /*
8142 : * Even though we're not charging I/O cost for touching upper btree pages,
8143 : * it's still reasonable to charge some CPU cost per page descended
8144 : * through. Moreover, if we had no such charge at all, bloated indexes
8145 : * would appear to have the same search cost as unbloated ones, at least
8146 : * in cases where only a single leaf page is expected to be visited. This
8147 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
8148 : * touched. The number of such pages is btree tree height plus one (ie,
8149 : * we charge for the leaf page too). As above, charge once per estimated
8150 : * SAOP/skip array descent.
8151 : */
8152 651154 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8153 651154 : costs.indexStartupCost += descentCost;
8154 651154 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8155 :
8156 651154 : if (!have_correlation)
8157 : {
8158 623841 : examine_indexcol_variable(root, index, 0, &vardata);
8159 623841 : if (HeapTupleIsValid(vardata.statsTuple))
8160 410463 : costs.indexCorrelation = btcost_correlation(index, &vardata);
8161 623841 : ReleaseVariableStats(vardata);
8162 : }
8163 : else
8164 : {
8165 : /* btcost_correlation already called earlier on */
8166 27313 : costs.indexCorrelation = correlation;
8167 : }
8168 :
8169 651154 : *indexStartupCost = costs.indexStartupCost;
8170 651154 : *indexTotalCost = costs.indexTotalCost;
8171 651154 : *indexSelectivity = costs.indexSelectivity;
8172 651154 : *indexCorrelation = costs.indexCorrelation;
8173 651154 : *indexPages = costs.numIndexPages;
8174 651154 : }
8175 :
8176 : void
8177 308 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8178 : Cost *indexStartupCost, Cost *indexTotalCost,
8179 : Selectivity *indexSelectivity, double *indexCorrelation,
8180 : double *indexPages)
8181 : {
8182 308 : GenericCosts costs = {0};
8183 :
8184 : /* As in btcostestimate, count only the metapage as non-leaf */
8185 308 : costs.numNonLeafPages = 1;
8186 :
8187 308 : genericcostestimate(root, path, loop_count, &costs);
8188 :
8189 : /*
8190 : * A hash index has no descent costs as such, since the index AM can go
8191 : * directly to the target bucket after computing the hash value. There
8192 : * are a couple of other hash-specific costs that we could conceivably add
8193 : * here, though:
8194 : *
8195 : * Ideally we'd charge spc_random_page_cost for each page in the target
8196 : * bucket, not just the numIndexPages pages that genericcostestimate
8197 : * thought we'd visit. However in most cases we don't know which bucket
8198 : * that will be. There's no point in considering the average bucket size
8199 : * because the hash AM makes sure that's always one page.
8200 : *
8201 : * Likewise, we could consider charging some CPU for each index tuple in
8202 : * the bucket, if we knew how many there were. But the per-tuple cost is
8203 : * just a hash value comparison, not a general datatype-dependent
8204 : * comparison, so any such charge ought to be quite a bit less than
8205 : * cpu_operator_cost; which makes it probably not worth worrying about.
8206 : *
8207 : * A bigger issue is that chance hash-value collisions will result in
8208 : * wasted probes into the heap. We don't currently attempt to model this
8209 : * cost on the grounds that it's rare, but maybe it's not rare enough.
8210 : * (Any fix for this ought to consider the generic lossy-operator problem,
8211 : * though; it's not entirely hash-specific.)
8212 : */
8213 :
8214 308 : *indexStartupCost = costs.indexStartupCost;
8215 308 : *indexTotalCost = costs.indexTotalCost;
8216 308 : *indexSelectivity = costs.indexSelectivity;
8217 308 : *indexCorrelation = costs.indexCorrelation;
8218 308 : *indexPages = costs.numIndexPages;
8219 308 : }
8220 :
8221 : void
8222 4772 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8223 : Cost *indexStartupCost, Cost *indexTotalCost,
8224 : Selectivity *indexSelectivity, double *indexCorrelation,
8225 : double *indexPages)
8226 : {
8227 4772 : IndexOptInfo *index = path->indexinfo;
8228 4772 : GenericCosts costs = {0};
8229 : Cost descentCost;
8230 :
8231 : /* GiST has no metapage, so we treat all pages as leaf pages */
8232 :
8233 4772 : genericcostestimate(root, path, loop_count, &costs);
8234 :
8235 : /*
8236 : * We model index descent costs similarly to those for btree, but to do
8237 : * that we first need an idea of the tree height. We somewhat arbitrarily
8238 : * assume that the fanout is 100, meaning the tree height is at most
8239 : * log100(index->pages).
8240 : *
8241 : * Although this computation isn't really expensive enough to require
8242 : * caching, we might as well use index->tree_height to cache it.
8243 : */
8244 4772 : if (index->tree_height < 0) /* unknown? */
8245 : {
8246 4745 : if (index->pages > 1) /* avoid computing log(0) */
8247 1956 : index->tree_height = (int) (log(index->pages) / log(100.0));
8248 : else
8249 2789 : index->tree_height = 0;
8250 : }
8251 :
8252 : /*
8253 : * Add a CPU-cost component to represent the costs of initial descent. We
8254 : * just use log(N) here not log2(N) since the branching factor isn't
8255 : * necessarily two anyway. As for btree, charge once per SA scan.
8256 : */
8257 4772 : if (index->tuples > 1) /* avoid computing log(0) */
8258 : {
8259 4772 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8260 4772 : costs.indexStartupCost += descentCost;
8261 4772 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8262 : }
8263 :
8264 : /*
8265 : * Likewise add a per-page charge, calculated the same as for btrees.
8266 : */
8267 4772 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8268 4772 : costs.indexStartupCost += descentCost;
8269 4772 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8270 :
8271 4772 : *indexStartupCost = costs.indexStartupCost;
8272 4772 : *indexTotalCost = costs.indexTotalCost;
8273 4772 : *indexSelectivity = costs.indexSelectivity;
8274 4772 : *indexCorrelation = costs.indexCorrelation;
8275 4772 : *indexPages = costs.numIndexPages;
8276 4772 : }
8277 :
8278 : void
8279 1482 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8280 : Cost *indexStartupCost, Cost *indexTotalCost,
8281 : Selectivity *indexSelectivity, double *indexCorrelation,
8282 : double *indexPages)
8283 : {
8284 1482 : IndexOptInfo *index = path->indexinfo;
8285 1482 : GenericCosts costs = {0};
8286 : Cost descentCost;
8287 :
8288 : /* As in btcostestimate, count only the metapage as non-leaf */
8289 1482 : costs.numNonLeafPages = 1;
8290 :
8291 1482 : genericcostestimate(root, path, loop_count, &costs);
8292 :
8293 : /*
8294 : * We model index descent costs similarly to those for btree, but to do
8295 : * that we first need an idea of the tree height. We somewhat arbitrarily
8296 : * assume that the fanout is 100, meaning the tree height is at most
8297 : * log100(index->pages).
8298 : *
8299 : * Although this computation isn't really expensive enough to require
8300 : * caching, we might as well use index->tree_height to cache it.
8301 : */
8302 1482 : if (index->tree_height < 0) /* unknown? */
8303 : {
8304 1477 : if (index->pages > 1) /* avoid computing log(0) */
8305 1477 : index->tree_height = (int) (log(index->pages) / log(100.0));
8306 : else
8307 0 : index->tree_height = 0;
8308 : }
8309 :
8310 : /*
8311 : * Add a CPU-cost component to represent the costs of initial descent. We
8312 : * just use log(N) here not log2(N) since the branching factor isn't
8313 : * necessarily two anyway. As for btree, charge once per SA scan.
8314 : */
8315 1482 : if (index->tuples > 1) /* avoid computing log(0) */
8316 : {
8317 1482 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8318 1482 : costs.indexStartupCost += descentCost;
8319 1482 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8320 : }
8321 :
8322 : /*
8323 : * Likewise add a per-page charge, calculated the same as for btrees.
8324 : */
8325 1482 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8326 1482 : costs.indexStartupCost += descentCost;
8327 1482 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8328 :
8329 1482 : *indexStartupCost = costs.indexStartupCost;
8330 1482 : *indexTotalCost = costs.indexTotalCost;
8331 1482 : *indexSelectivity = costs.indexSelectivity;
8332 1482 : *indexCorrelation = costs.indexCorrelation;
8333 1482 : *indexPages = costs.numIndexPages;
8334 1482 : }
8335 :
8336 :
8337 : /*
8338 : * Support routines for gincostestimate
8339 : */
8340 :
8341 : typedef struct
8342 : {
8343 : bool attHasFullScan[INDEX_MAX_KEYS];
8344 : bool attHasNormalScan[INDEX_MAX_KEYS];
8345 : double partialEntries;
8346 : double exactEntries;
8347 : double searchEntries;
8348 : double arrayScans;
8349 : } GinQualCounts;
8350 :
8351 : /*
8352 : * Estimate the number of index terms that need to be searched for while
8353 : * testing the given GIN query, and increment the counts in *counts
8354 : * appropriately. If the query is unsatisfiable, return false.
8355 : */
8356 : static bool
8357 1714 : gincost_pattern(IndexOptInfo *index, int indexcol,
8358 : Oid clause_op, Datum query,
8359 : GinQualCounts *counts)
8360 : {
8361 : FmgrInfo flinfo;
8362 : Oid extractProcOid;
8363 : Oid collation;
8364 : int strategy_op;
8365 : Oid lefttype,
8366 : righttype;
8367 1714 : int32 nentries = 0;
8368 1714 : bool *partial_matches = NULL;
8369 1714 : Pointer *extra_data = NULL;
8370 1714 : bool *nullFlags = NULL;
8371 1714 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
8372 : int32 i;
8373 :
8374 : Assert(indexcol < index->nkeycolumns);
8375 :
8376 : /*
8377 : * Get the operator's strategy number and declared input data types within
8378 : * the index opfamily. (We don't need the latter, but we use
8379 : * get_op_opfamily_properties because it will throw error if it fails to
8380 : * find a matching pg_amop entry.)
8381 : */
8382 1714 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
8383 : &strategy_op, &lefttype, &righttype);
8384 :
8385 : /*
8386 : * GIN always uses the "default" support functions, which are those with
8387 : * lefttype == righttype == the opclass' opcintype (see
8388 : * IndexSupportInitialize in relcache.c).
8389 : */
8390 1714 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
8391 1714 : index->opcintype[indexcol],
8392 1714 : index->opcintype[indexcol],
8393 : GIN_EXTRACTQUERY_PROC);
8394 :
8395 1714 : if (!OidIsValid(extractProcOid))
8396 : {
8397 : /* should not happen; throw same error as index_getprocinfo */
8398 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
8399 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
8400 : get_rel_name(index->indexoid));
8401 : }
8402 :
8403 : /*
8404 : * Choose collation to pass to extractProc (should match initGinState).
8405 : */
8406 1714 : if (OidIsValid(index->indexcollations[indexcol]))
8407 243 : collation = index->indexcollations[indexcol];
8408 : else
8409 1471 : collation = DEFAULT_COLLATION_OID;
8410 :
8411 1714 : fmgr_info(extractProcOid, &flinfo);
8412 :
8413 1714 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
8414 :
8415 1714 : FunctionCall7Coll(&flinfo,
8416 : collation,
8417 : query,
8418 : PointerGetDatum(&nentries),
8419 : UInt16GetDatum(strategy_op),
8420 : PointerGetDatum(&partial_matches),
8421 : PointerGetDatum(&extra_data),
8422 : PointerGetDatum(&nullFlags),
8423 : PointerGetDatum(&searchMode));
8424 :
8425 1714 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
8426 : {
8427 : /* No match is possible */
8428 10 : return false;
8429 : }
8430 :
8431 5854 : for (i = 0; i < nentries; i++)
8432 : {
8433 : /*
8434 : * For partial match we haven't any information to estimate number of
8435 : * matched entries in index, so, we just estimate it as 100
8436 : */
8437 4150 : if (partial_matches && partial_matches[i])
8438 361 : counts->partialEntries += 100;
8439 : else
8440 3789 : counts->exactEntries++;
8441 :
8442 4150 : counts->searchEntries++;
8443 : }
8444 :
8445 1704 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
8446 : {
8447 1322 : counts->attHasNormalScan[indexcol] = true;
8448 : }
8449 382 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
8450 : {
8451 : /* Treat "include empty" like an exact-match item */
8452 36 : counts->attHasNormalScan[indexcol] = true;
8453 36 : counts->exactEntries++;
8454 36 : counts->searchEntries++;
8455 : }
8456 : else
8457 : {
8458 : /* It's GIN_SEARCH_MODE_ALL */
8459 346 : counts->attHasFullScan[indexcol] = true;
8460 : }
8461 :
8462 1704 : return true;
8463 : }
8464 :
8465 : /*
8466 : * Estimate the number of index terms that need to be searched for while
8467 : * testing the given GIN index clause, and increment the counts in *counts
8468 : * appropriately. If the query is unsatisfiable, return false.
8469 : */
8470 : static bool
8471 1704 : gincost_opexpr(PlannerInfo *root,
8472 : IndexOptInfo *index,
8473 : int indexcol,
8474 : OpExpr *clause,
8475 : GinQualCounts *counts)
8476 : {
8477 1704 : Oid clause_op = clause->opno;
8478 1704 : Node *operand = (Node *) lsecond(clause->args);
8479 :
8480 : /* aggressively reduce to a constant, and look through relabeling */
8481 1704 : operand = estimate_expression_value(root, operand);
8482 :
8483 1704 : if (IsA(operand, RelabelType))
8484 0 : operand = (Node *) ((RelabelType *) operand)->arg;
8485 :
8486 : /*
8487 : * It's impossible to call extractQuery method for unknown operand. So
8488 : * unless operand is a Const we can't do much; just assume there will be
8489 : * one ordinary search entry from the operand at runtime.
8490 : */
8491 1704 : if (!IsA(operand, Const))
8492 : {
8493 0 : counts->exactEntries++;
8494 0 : counts->searchEntries++;
8495 0 : return true;
8496 : }
8497 :
8498 : /* If Const is null, there can be no matches */
8499 1704 : if (((Const *) operand)->constisnull)
8500 0 : return false;
8501 :
8502 : /* Otherwise, apply extractQuery and get the actual term counts */
8503 1704 : return gincost_pattern(index, indexcol, clause_op,
8504 : ((Const *) operand)->constvalue,
8505 : counts);
8506 : }
8507 :
8508 : /*
8509 : * Estimate the number of index terms that need to be searched for while
8510 : * testing the given GIN index clause, and increment the counts in *counts
8511 : * appropriately. If the query is unsatisfiable, return false.
8512 : *
8513 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
8514 : * each of which involves one value from the RHS array, plus all the
8515 : * non-array quals (if any). To model this, we average the counts across
8516 : * the RHS elements, and add the averages to the counts in *counts (which
8517 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
8518 : * by N, causing gincostestimate to scale up its estimates accordingly.
8519 : */
8520 : static bool
8521 5 : gincost_scalararrayopexpr(PlannerInfo *root,
8522 : IndexOptInfo *index,
8523 : int indexcol,
8524 : ScalarArrayOpExpr *clause,
8525 : double numIndexEntries,
8526 : GinQualCounts *counts)
8527 : {
8528 5 : Oid clause_op = clause->opno;
8529 5 : Node *rightop = (Node *) lsecond(clause->args);
8530 : ArrayType *arrayval;
8531 : int16 elmlen;
8532 : bool elmbyval;
8533 : char elmalign;
8534 : int numElems;
8535 : Datum *elemValues;
8536 : bool *elemNulls;
8537 : GinQualCounts arraycounts;
8538 5 : int numPossible = 0;
8539 : int i;
8540 :
8541 : Assert(clause->useOr);
8542 :
8543 : /* aggressively reduce to a constant, and look through relabeling */
8544 5 : rightop = estimate_expression_value(root, rightop);
8545 :
8546 5 : if (IsA(rightop, RelabelType))
8547 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
8548 :
8549 : /*
8550 : * It's impossible to call extractQuery method for unknown operand. So
8551 : * unless operand is a Const we can't do much; just assume there will be
8552 : * one ordinary search entry from each array entry at runtime, and fall
8553 : * back on a probably-bad estimate of the number of array entries.
8554 : */
8555 5 : if (!IsA(rightop, Const))
8556 : {
8557 0 : counts->exactEntries++;
8558 0 : counts->searchEntries++;
8559 0 : counts->arrayScans *= estimate_array_length(root, rightop);
8560 0 : return true;
8561 : }
8562 :
8563 : /* If Const is null, there can be no matches */
8564 5 : if (((Const *) rightop)->constisnull)
8565 0 : return false;
8566 :
8567 : /* Otherwise, extract the array elements and iterate over them */
8568 5 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
8569 5 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
8570 : &elmlen, &elmbyval, &elmalign);
8571 5 : deconstruct_array(arrayval,
8572 : ARR_ELEMTYPE(arrayval),
8573 : elmlen, elmbyval, elmalign,
8574 : &elemValues, &elemNulls, &numElems);
8575 :
8576 5 : memset(&arraycounts, 0, sizeof(arraycounts));
8577 :
8578 15 : for (i = 0; i < numElems; i++)
8579 : {
8580 : GinQualCounts elemcounts;
8581 :
8582 : /* NULL can't match anything, so ignore, as the executor will */
8583 10 : if (elemNulls[i])
8584 0 : continue;
8585 :
8586 : /* Otherwise, apply extractQuery and get the actual term counts */
8587 10 : memset(&elemcounts, 0, sizeof(elemcounts));
8588 :
8589 10 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
8590 : &elemcounts))
8591 : {
8592 : /* We ignore array elements that are unsatisfiable patterns */
8593 10 : numPossible++;
8594 :
8595 10 : if (elemcounts.attHasFullScan[indexcol] &&
8596 0 : !elemcounts.attHasNormalScan[indexcol])
8597 : {
8598 : /*
8599 : * Full index scan will be required. We treat this as if
8600 : * every key in the index had been listed in the query; is
8601 : * that reasonable?
8602 : */
8603 0 : elemcounts.partialEntries = 0;
8604 0 : elemcounts.exactEntries = numIndexEntries;
8605 0 : elemcounts.searchEntries = numIndexEntries;
8606 : }
8607 10 : arraycounts.partialEntries += elemcounts.partialEntries;
8608 10 : arraycounts.exactEntries += elemcounts.exactEntries;
8609 10 : arraycounts.searchEntries += elemcounts.searchEntries;
8610 : }
8611 : }
8612 :
8613 5 : if (numPossible == 0)
8614 : {
8615 : /* No satisfiable patterns in the array */
8616 0 : return false;
8617 : }
8618 :
8619 : /*
8620 : * Now add the averages to the global counts. This will give us an
8621 : * estimate of the average number of terms searched for in each indexscan,
8622 : * including contributions from both array and non-array quals.
8623 : */
8624 5 : counts->partialEntries += arraycounts.partialEntries / numPossible;
8625 5 : counts->exactEntries += arraycounts.exactEntries / numPossible;
8626 5 : counts->searchEntries += arraycounts.searchEntries / numPossible;
8627 :
8628 5 : counts->arrayScans *= numPossible;
8629 :
8630 5 : return true;
8631 : }
8632 :
8633 : /*
8634 : * GIN has search behavior completely different from other index types
8635 : */
8636 : void
8637 1546 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8638 : Cost *indexStartupCost, Cost *indexTotalCost,
8639 : Selectivity *indexSelectivity, double *indexCorrelation,
8640 : double *indexPages)
8641 : {
8642 1546 : IndexOptInfo *index = path->indexinfo;
8643 1546 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8644 : List *selectivityQuals;
8645 1546 : double numPages = index->pages,
8646 1546 : numTuples = index->tuples;
8647 : double numEntryPages,
8648 : numDataPages,
8649 : numPendingPages,
8650 : numEntries;
8651 : GinQualCounts counts;
8652 : bool matchPossible;
8653 : bool fullIndexScan;
8654 : double partialScale;
8655 : double entryPagesFetched,
8656 : dataPagesFetched,
8657 : dataPagesFetchedBySel;
8658 : double qual_op_cost,
8659 : qual_arg_cost,
8660 : spc_random_page_cost,
8661 : outer_scans;
8662 : Cost descentCost;
8663 : Relation indexRel;
8664 : GinStatsData ginStats;
8665 : ListCell *lc;
8666 : int i;
8667 :
8668 : /*
8669 : * Obtain statistical information from the meta page, if possible. Else
8670 : * set ginStats to zeroes, and we'll cope below.
8671 : */
8672 1546 : if (!index->hypothetical)
8673 : {
8674 : /* Lock should have already been obtained in plancat.c */
8675 1546 : indexRel = index_open(index->indexoid, NoLock);
8676 1546 : ginGetStats(indexRel, &ginStats);
8677 1546 : index_close(indexRel, NoLock);
8678 : }
8679 : else
8680 : {
8681 0 : memset(&ginStats, 0, sizeof(ginStats));
8682 : }
8683 :
8684 : /*
8685 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
8686 : * trusted, but the other fields are data as of the last VACUUM. We can
8687 : * scale them up to account for growth since then, but that method only
8688 : * goes so far; in the worst case, the stats might be for a completely
8689 : * empty index, and scaling them will produce pretty bogus numbers.
8690 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
8691 : * it's grown more than that, fall back to estimating things only from the
8692 : * assumed-accurate index size. But we'll trust nPendingPages in any case
8693 : * so long as it's not clearly insane, ie, more than the index size.
8694 : */
8695 1546 : if (ginStats.nPendingPages < numPages)
8696 1546 : numPendingPages = ginStats.nPendingPages;
8697 : else
8698 0 : numPendingPages = 0;
8699 :
8700 1546 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
8701 1546 : ginStats.nTotalPages > numPages / 4 &&
8702 1506 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
8703 1292 : {
8704 : /*
8705 : * OK, the stats seem close enough to sane to be trusted. But we
8706 : * still need to scale them by the ratio numPages / nTotalPages to
8707 : * account for growth since the last VACUUM.
8708 : */
8709 1292 : double scale = numPages / ginStats.nTotalPages;
8710 :
8711 1292 : numEntryPages = ceil(ginStats.nEntryPages * scale);
8712 1292 : numDataPages = ceil(ginStats.nDataPages * scale);
8713 1292 : numEntries = ceil(ginStats.nEntries * scale);
8714 : /* ensure we didn't round up too much */
8715 1292 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
8716 1292 : numDataPages = Min(numDataPages,
8717 : numPages - numPendingPages - numEntryPages);
8718 : }
8719 : else
8720 : {
8721 : /*
8722 : * We might get here because it's a hypothetical index, or an index
8723 : * created pre-9.1 and never vacuumed since upgrading (in which case
8724 : * its stats would read as zeroes), or just because it's grown too
8725 : * much since the last VACUUM for us to put our faith in scaling.
8726 : *
8727 : * Invent some plausible internal statistics based on the index page
8728 : * count (and clamp that to at least 10 pages, just in case). We
8729 : * estimate that 90% of the index is entry pages, and the rest is data
8730 : * pages. Estimate 100 entries per entry page; this is rather bogus
8731 : * since it'll depend on the size of the keys, but it's more robust
8732 : * than trying to predict the number of entries per heap tuple.
8733 : */
8734 254 : numPages = Max(numPages, 10);
8735 254 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
8736 254 : numDataPages = numPages - numPendingPages - numEntryPages;
8737 254 : numEntries = floor(numEntryPages * 100);
8738 : }
8739 :
8740 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
8741 1546 : if (numEntries < 1)
8742 0 : numEntries = 1;
8743 :
8744 : /*
8745 : * If the index is partial, AND the index predicate with the index-bound
8746 : * quals to produce a more accurate idea of the number of rows covered by
8747 : * the bound conditions.
8748 : */
8749 1546 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
8750 :
8751 : /* Estimate the fraction of main-table tuples that will be visited */
8752 3092 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
8753 1546 : index->rel->relid,
8754 : JOIN_INNER,
8755 : NULL);
8756 :
8757 : /* fetch estimated page cost for tablespace containing index */
8758 1546 : get_tablespace_page_costs(index->reltablespace,
8759 : &spc_random_page_cost,
8760 : NULL);
8761 :
8762 : /*
8763 : * Generic assumption about index correlation: there isn't any.
8764 : */
8765 1546 : *indexCorrelation = 0.0;
8766 :
8767 : /*
8768 : * Examine quals to estimate number of search entries & partial matches
8769 : */
8770 1546 : memset(&counts, 0, sizeof(counts));
8771 1546 : counts.arrayScans = 1;
8772 1546 : matchPossible = true;
8773 :
8774 3255 : foreach(lc, path->indexclauses)
8775 : {
8776 1709 : IndexClause *iclause = lfirst_node(IndexClause, lc);
8777 : ListCell *lc2;
8778 :
8779 3408 : foreach(lc2, iclause->indexquals)
8780 : {
8781 1709 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8782 1709 : Expr *clause = rinfo->clause;
8783 :
8784 1709 : if (IsA(clause, OpExpr))
8785 : {
8786 1704 : matchPossible = gincost_opexpr(root,
8787 : index,
8788 1704 : iclause->indexcol,
8789 : (OpExpr *) clause,
8790 : &counts);
8791 1704 : if (!matchPossible)
8792 10 : break;
8793 : }
8794 5 : else if (IsA(clause, ScalarArrayOpExpr))
8795 : {
8796 5 : matchPossible = gincost_scalararrayopexpr(root,
8797 : index,
8798 5 : iclause->indexcol,
8799 : (ScalarArrayOpExpr *) clause,
8800 : numEntries,
8801 : &counts);
8802 5 : if (!matchPossible)
8803 0 : break;
8804 : }
8805 : else
8806 : {
8807 : /* shouldn't be anything else for a GIN index */
8808 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
8809 : (int) nodeTag(clause));
8810 : }
8811 : }
8812 : }
8813 :
8814 : /* Fall out if there were any provably-unsatisfiable quals */
8815 1546 : if (!matchPossible)
8816 : {
8817 10 : *indexStartupCost = 0;
8818 10 : *indexTotalCost = 0;
8819 10 : *indexSelectivity = 0;
8820 10 : return;
8821 : }
8822 :
8823 : /*
8824 : * If attribute has a full scan and at the same time doesn't have normal
8825 : * scan, then we'll have to scan all non-null entries of that attribute.
8826 : * Currently, we don't have per-attribute statistics for GIN. Thus, we
8827 : * must assume the whole GIN index has to be scanned in this case.
8828 : */
8829 1536 : fullIndexScan = false;
8830 2989 : for (i = 0; i < index->nkeycolumns; i++)
8831 : {
8832 1726 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
8833 : {
8834 273 : fullIndexScan = true;
8835 273 : break;
8836 : }
8837 : }
8838 :
8839 1536 : if (fullIndexScan || indexQuals == NIL)
8840 : {
8841 : /*
8842 : * Full index scan will be required. We treat this as if every key in
8843 : * the index had been listed in the query; is that reasonable?
8844 : */
8845 273 : counts.partialEntries = 0;
8846 273 : counts.exactEntries = numEntries;
8847 273 : counts.searchEntries = numEntries;
8848 : }
8849 :
8850 : /* Will we have more than one iteration of a nestloop scan? */
8851 1536 : outer_scans = loop_count;
8852 :
8853 : /*
8854 : * Compute cost to begin scan, first of all, pay attention to pending
8855 : * list.
8856 : */
8857 1536 : entryPagesFetched = numPendingPages;
8858 :
8859 : /*
8860 : * Estimate number of entry pages read. We need to do
8861 : * counts.searchEntries searches. Use a power function as it should be,
8862 : * but tuples on leaf pages usually is much greater. Here we include all
8863 : * searches in entry tree, including search of first entry in partial
8864 : * match algorithm
8865 : */
8866 1536 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8867 :
8868 : /*
8869 : * Add an estimate of entry pages read by partial match algorithm. It's a
8870 : * scan over leaf pages in entry tree. We haven't any useful stats here,
8871 : * so estimate it as proportion. Because counts.partialEntries is really
8872 : * pretty bogus (see code above), it's possible that it is more than
8873 : * numEntries; clamp the proportion to ensure sanity.
8874 : */
8875 1536 : partialScale = counts.partialEntries / numEntries;
8876 1536 : partialScale = Min(partialScale, 1.0);
8877 :
8878 1536 : entryPagesFetched += ceil(numEntryPages * partialScale);
8879 :
8880 : /*
8881 : * Partial match algorithm reads all data pages before doing actual scan,
8882 : * so it's a startup cost. Again, we haven't any useful stats here, so
8883 : * estimate it as proportion.
8884 : */
8885 1536 : dataPagesFetched = ceil(numDataPages * partialScale);
8886 :
8887 1536 : *indexStartupCost = 0;
8888 1536 : *indexTotalCost = 0;
8889 :
8890 : /*
8891 : * Add a CPU-cost component to represent the costs of initial entry btree
8892 : * descent. We don't charge any I/O cost for touching upper btree levels,
8893 : * since they tend to stay in cache, but we still have to do about log2(N)
8894 : * comparisons to descend a btree of N leaf tuples. We charge one
8895 : * cpu_operator_cost per comparison.
8896 : *
8897 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
8898 : * ones after the first one are not startup cost so far as the overall
8899 : * plan is concerned, so add them only to "total" cost.
8900 : */
8901 1536 : if (numEntries > 1) /* avoid computing log(0) */
8902 : {
8903 1536 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
8904 1536 : *indexStartupCost += descentCost * counts.searchEntries;
8905 1536 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
8906 : }
8907 :
8908 : /*
8909 : * Add a cpu cost per entry-page fetched. This is not amortized over a
8910 : * loop.
8911 : */
8912 1536 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8913 1536 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8914 :
8915 : /*
8916 : * Add a cpu cost per data-page fetched. This is also not amortized over a
8917 : * loop. Since those are the data pages from the partial match algorithm,
8918 : * charge them as startup cost.
8919 : */
8920 1536 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
8921 :
8922 : /*
8923 : * Since we add the startup cost to the total cost later on, remove the
8924 : * initial arrayscan from the total.
8925 : */
8926 1536 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8927 :
8928 : /*
8929 : * Calculate cache effects if more than one scan due to nestloops or array
8930 : * quals. The result is pro-rated per nestloop scan, but the array qual
8931 : * factor shouldn't be pro-rated (compare genericcostestimate).
8932 : */
8933 1536 : if (outer_scans > 1 || counts.arrayScans > 1)
8934 : {
8935 5 : entryPagesFetched *= outer_scans * counts.arrayScans;
8936 5 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
8937 : (BlockNumber) numEntryPages,
8938 : numEntryPages, root);
8939 5 : entryPagesFetched /= outer_scans;
8940 5 : dataPagesFetched *= outer_scans * counts.arrayScans;
8941 5 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8942 : (BlockNumber) numDataPages,
8943 : numDataPages, root);
8944 5 : dataPagesFetched /= outer_scans;
8945 : }
8946 :
8947 : /*
8948 : * Here we use random page cost because logically-close pages could be far
8949 : * apart on disk.
8950 : */
8951 1536 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
8952 :
8953 : /*
8954 : * Now compute the number of data pages fetched during the scan.
8955 : *
8956 : * We assume every entry to have the same number of items, and that there
8957 : * is no overlap between them. (XXX: tsvector and array opclasses collect
8958 : * statistics on the frequency of individual keys; it would be nice to use
8959 : * those here.)
8960 : */
8961 1536 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
8962 :
8963 : /*
8964 : * If there is a lot of overlap among the entries, in particular if one of
8965 : * the entries is very frequent, the above calculation can grossly
8966 : * under-estimate. As a simple cross-check, calculate a lower bound based
8967 : * on the overall selectivity of the quals. At a minimum, we must read
8968 : * one item pointer for each matching entry.
8969 : *
8970 : * The width of each item pointer varies, based on the level of
8971 : * compression. We don't have statistics on that, but an average of
8972 : * around 3 bytes per item is fairly typical.
8973 : */
8974 1536 : dataPagesFetchedBySel = ceil(*indexSelectivity *
8975 1536 : (numTuples / (BLCKSZ / 3)));
8976 1536 : if (dataPagesFetchedBySel > dataPagesFetched)
8977 1243 : dataPagesFetched = dataPagesFetchedBySel;
8978 :
8979 : /* Add one page cpu-cost to the startup cost */
8980 1536 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
8981 :
8982 : /*
8983 : * Add once again a CPU-cost for those data pages, before amortizing for
8984 : * cache.
8985 : */
8986 1536 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8987 :
8988 : /* Account for cache effects, the same as above */
8989 1536 : if (outer_scans > 1 || counts.arrayScans > 1)
8990 : {
8991 5 : dataPagesFetched *= outer_scans * counts.arrayScans;
8992 5 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8993 : (BlockNumber) numDataPages,
8994 : numDataPages, root);
8995 5 : dataPagesFetched /= outer_scans;
8996 : }
8997 :
8998 : /* And apply random_page_cost as the cost per page */
8999 1536 : *indexTotalCost += *indexStartupCost +
9000 1536 : dataPagesFetched * spc_random_page_cost;
9001 :
9002 : /*
9003 : * Add on index qual eval costs, much as in genericcostestimate. We charge
9004 : * cpu but we can disregard indexorderbys, since GIN doesn't support
9005 : * those.
9006 : */
9007 1536 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
9008 1536 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
9009 :
9010 1536 : *indexStartupCost += qual_arg_cost;
9011 1536 : *indexTotalCost += qual_arg_cost;
9012 :
9013 : /*
9014 : * Add a cpu cost per search entry, corresponding to the actual visited
9015 : * entries.
9016 : */
9017 1536 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
9018 : /* Now add a cpu cost per tuple in the posting lists / trees */
9019 1536 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
9020 1536 : *indexPages = dataPagesFetched;
9021 : }
9022 :
9023 : /*
9024 : * BRIN has search behavior completely different from other index types
9025 : */
9026 : void
9027 8944 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
9028 : Cost *indexStartupCost, Cost *indexTotalCost,
9029 : Selectivity *indexSelectivity, double *indexCorrelation,
9030 : double *indexPages)
9031 : {
9032 8944 : IndexOptInfo *index = path->indexinfo;
9033 8944 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
9034 8944 : double numPages = index->pages;
9035 8944 : RelOptInfo *baserel = index->rel;
9036 8944 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
9037 : Cost spc_seq_page_cost;
9038 : Cost spc_random_page_cost;
9039 : double qual_arg_cost;
9040 : double qualSelectivity;
9041 : BrinStatsData statsData;
9042 : double indexRanges;
9043 : double minimalRanges;
9044 : double estimatedRanges;
9045 : double selec;
9046 : Relation indexRel;
9047 : ListCell *l;
9048 : VariableStatData vardata;
9049 :
9050 : Assert(rte->rtekind == RTE_RELATION);
9051 :
9052 : /* fetch estimated page cost for the tablespace containing the index */
9053 8944 : get_tablespace_page_costs(index->reltablespace,
9054 : &spc_random_page_cost,
9055 : &spc_seq_page_cost);
9056 :
9057 : /*
9058 : * Obtain some data from the index itself, if possible. Otherwise invent
9059 : * some plausible internal statistics based on the relation page count.
9060 : */
9061 8944 : if (!index->hypothetical)
9062 : {
9063 : /*
9064 : * A lock should have already been obtained on the index in plancat.c.
9065 : */
9066 8944 : indexRel = index_open(index->indexoid, NoLock);
9067 8944 : brinGetStats(indexRel, &statsData);
9068 8944 : index_close(indexRel, NoLock);
9069 :
9070 : /* work out the actual number of ranges in the index */
9071 8944 : indexRanges = Max(ceil((double) baserel->pages /
9072 : statsData.pagesPerRange), 1.0);
9073 : }
9074 : else
9075 : {
9076 : /*
9077 : * Assume default number of pages per range, and estimate the number
9078 : * of ranges based on that.
9079 : */
9080 0 : indexRanges = Max(ceil((double) baserel->pages /
9081 : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
9082 :
9083 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
9084 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
9085 : }
9086 :
9087 : /*
9088 : * Compute index correlation
9089 : *
9090 : * Because we can use all index quals equally when scanning, we can use
9091 : * the largest correlation (in absolute value) among columns used by the
9092 : * query. Start at zero, the worst possible case. If we cannot find any
9093 : * correlation statistics, we will keep it as 0.
9094 : */
9095 8944 : *indexCorrelation = 0;
9096 :
9097 17889 : foreach(l, path->indexclauses)
9098 : {
9099 8945 : IndexClause *iclause = lfirst_node(IndexClause, l);
9100 8945 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
9101 :
9102 : /* attempt to lookup stats in relation for this index column */
9103 8945 : if (attnum != 0)
9104 : {
9105 : /* Simple variable -- look to stats for the underlying table */
9106 8945 : if (get_relation_stats_hook &&
9107 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
9108 : {
9109 : /*
9110 : * The hook took control of acquiring a stats tuple. If it
9111 : * did supply a tuple, it'd better have supplied a freefunc.
9112 : */
9113 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
9114 0 : elog(ERROR,
9115 : "no function provided to release variable stats with");
9116 : }
9117 : else
9118 : {
9119 8945 : vardata.statsTuple =
9120 8945 : SearchSysCache3(STATRELATTINH,
9121 : ObjectIdGetDatum(rte->relid),
9122 : Int16GetDatum(attnum),
9123 : BoolGetDatum(false));
9124 8945 : vardata.freefunc = ReleaseSysCache;
9125 : }
9126 : }
9127 : else
9128 : {
9129 : /*
9130 : * Looks like we've found an expression column in the index. Let's
9131 : * see if there's any stats for it.
9132 : */
9133 :
9134 : /* get the attnum from the 0-based index. */
9135 0 : attnum = iclause->indexcol + 1;
9136 :
9137 0 : if (get_index_stats_hook &&
9138 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
9139 : {
9140 : /*
9141 : * The hook took control of acquiring a stats tuple. If it
9142 : * did supply a tuple, it'd better have supplied a freefunc.
9143 : */
9144 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
9145 0 : !vardata.freefunc)
9146 0 : elog(ERROR, "no function provided to release variable stats with");
9147 : }
9148 : else
9149 : {
9150 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
9151 : ObjectIdGetDatum(index->indexoid),
9152 : Int16GetDatum(attnum),
9153 : BoolGetDatum(false));
9154 0 : vardata.freefunc = ReleaseSysCache;
9155 : }
9156 : }
9157 :
9158 8945 : if (HeapTupleIsValid(vardata.statsTuple))
9159 : {
9160 : AttStatsSlot sslot;
9161 :
9162 33 : if (get_attstatsslot(&sslot, vardata.statsTuple,
9163 : STATISTIC_KIND_CORRELATION, InvalidOid,
9164 : ATTSTATSSLOT_NUMBERS))
9165 : {
9166 33 : double varCorrelation = 0.0;
9167 :
9168 33 : if (sslot.nnumbers > 0)
9169 33 : varCorrelation = fabs(sslot.numbers[0]);
9170 :
9171 33 : if (varCorrelation > *indexCorrelation)
9172 33 : *indexCorrelation = varCorrelation;
9173 :
9174 33 : free_attstatsslot(&sslot);
9175 : }
9176 : }
9177 :
9178 8945 : ReleaseVariableStats(vardata);
9179 : }
9180 :
9181 8944 : qualSelectivity = clauselist_selectivity(root, indexQuals,
9182 8944 : baserel->relid,
9183 : JOIN_INNER, NULL);
9184 :
9185 : /*
9186 : * Now calculate the minimum possible ranges we could match with if all of
9187 : * the rows were in the perfect order in the table's heap.
9188 : */
9189 8944 : minimalRanges = ceil(indexRanges * qualSelectivity);
9190 :
9191 : /*
9192 : * Now estimate the number of ranges that we'll touch by using the
9193 : * indexCorrelation from the stats. Careful not to divide by zero (note
9194 : * we're using the absolute value of the correlation).
9195 : */
9196 8944 : if (*indexCorrelation < 1.0e-10)
9197 8911 : estimatedRanges = indexRanges;
9198 : else
9199 33 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
9200 :
9201 : /* we expect to visit this portion of the table */
9202 8944 : selec = estimatedRanges / indexRanges;
9203 :
9204 8944 : CLAMP_PROBABILITY(selec);
9205 :
9206 8944 : *indexSelectivity = selec;
9207 :
9208 : /*
9209 : * Compute the index qual costs, much as in genericcostestimate, to add to
9210 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
9211 : * support those.
9212 : */
9213 8944 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
9214 :
9215 : /*
9216 : * Compute the startup cost as the cost to read the whole revmap
9217 : * sequentially, including the cost to execute the index quals.
9218 : */
9219 8944 : *indexStartupCost =
9220 8944 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
9221 8944 : *indexStartupCost += qual_arg_cost;
9222 :
9223 : /*
9224 : * To read a BRIN index there might be a bit of back and forth over
9225 : * regular pages, as revmap might point to them out of sequential order;
9226 : * calculate the total cost as reading the whole index in random order.
9227 : */
9228 8944 : *indexTotalCost = *indexStartupCost +
9229 8944 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
9230 :
9231 : /*
9232 : * Charge a small amount per range tuple which we expect to match to. This
9233 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
9234 : * will set a bit for each page in the range when we find a matching
9235 : * range, so we must multiply the charge by the number of pages in the
9236 : * range.
9237 : */
9238 8944 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
9239 8944 : statsData.pagesPerRange;
9240 :
9241 8944 : *indexPages = index->pages;
9242 8944 : }
|
