Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * selfuncs.c
4 : * Selectivity functions and index cost estimation functions for
5 : * standard operators and index access methods.
6 : *
7 : * Selectivity routines are registered in the pg_operator catalog
8 : * in the "oprrest" and "oprjoin" attributes.
9 : *
10 : * Index cost functions are located via the index AM's API struct,
11 : * which is obtained from the handler function registered in pg_am.
12 : *
13 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
14 : * Portions Copyright (c) 1994, Regents of the University of California
15 : *
16 : *
17 : * IDENTIFICATION
18 : * src/backend/utils/adt/selfuncs.c
19 : *
20 : *-------------------------------------------------------------------------
21 : */
22 :
23 : /*----------
24 : * Operator selectivity estimation functions are called to estimate the
25 : * selectivity of WHERE clauses whose top-level operator is their operator.
26 : * We divide the problem into two cases:
27 : * Restriction clause estimation: the clause involves vars of just
28 : * one relation.
29 : * Join clause estimation: the clause involves vars of multiple rels.
30 : * Join selectivity estimation is far more difficult and usually less accurate
31 : * than restriction estimation.
32 : *
33 : * When dealing with the inner scan of a nestloop join, we consider the
34 : * join's joinclauses as restriction clauses for the inner relation, and
35 : * treat vars of the outer relation as parameters (a/k/a constants of unknown
36 : * values). So, restriction estimators need to be able to accept an argument
37 : * telling which relation is to be treated as the variable.
38 : *
39 : * The call convention for a restriction estimator (oprrest function) is
40 : *
41 : * Selectivity oprrest (PlannerInfo *root,
42 : * Oid operator,
43 : * List *args,
44 : * int varRelid);
45 : *
46 : * root: general information about the query (rtable and RelOptInfo lists
47 : * are particularly important for the estimator).
48 : * operator: OID of the specific operator in question.
49 : * args: argument list from the operator clause.
50 : * varRelid: if not zero, the relid (rtable index) of the relation to
51 : * be treated as the variable relation. May be zero if the args list
52 : * is known to contain vars of only one relation.
53 : *
54 : * This is represented at the SQL level (in pg_proc) as
55 : *
56 : * float8 oprrest (internal, oid, internal, int4);
57 : *
58 : * The result is a selectivity, that is, a fraction (0 to 1) of the rows
59 : * of the relation that are expected to produce a TRUE result for the
60 : * given operator.
61 : *
62 : * The call convention for a join estimator (oprjoin function) is similar
63 : * except that varRelid is not needed, and instead join information is
64 : * supplied:
65 : *
66 : * Selectivity oprjoin (PlannerInfo *root,
67 : * Oid operator,
68 : * List *args,
69 : * JoinType jointype,
70 : * SpecialJoinInfo *sjinfo);
71 : *
72 : * float8 oprjoin (internal, oid, internal, int2, internal);
73 : *
74 : * (Before Postgres 8.4, join estimators had only the first four of these
75 : * parameters. That signature is still allowed, but deprecated.) The
76 : * relationship between jointype and sjinfo is explained in the comments for
77 : * clause_selectivity() --- the short version is that jointype is usually
78 : * best ignored in favor of examining sjinfo.
79 : *
80 : * Join selectivity for regular inner and outer joins is defined as the
81 : * fraction (0 to 1) of the cross product of the relations that is expected
82 : * to produce a TRUE result for the given operator. For both semi and anti
83 : * joins, however, the selectivity is defined as the fraction of the left-hand
84 : * side relation's rows that are expected to have a match (ie, at least one
85 : * row with a TRUE result) in the right-hand side.
86 : *
87 : * For both oprrest and oprjoin functions, the operator's input collation OID
88 : * (if any) is passed using the standard fmgr mechanism, so that the estimator
89 : * function can fetch it with PG_GET_COLLATION(). Note, however, that all
90 : * statistics in pg_statistic are currently built using the relevant column's
91 : * collation.
92 : *----------
93 : */
94 :
95 : #include "postgres.h"
96 :
97 : #include <ctype.h>
98 : #include <math.h>
99 :
100 : #include "access/brin.h"
101 : #include "access/brin_page.h"
102 : #include "access/gin.h"
103 : #include "access/table.h"
104 : #include "access/tableam.h"
105 : #include "access/visibilitymap.h"
106 : #include "catalog/pg_am.h"
107 : #include "catalog/pg_collation.h"
108 : #include "catalog/pg_operator.h"
109 : #include "catalog/pg_statistic.h"
110 : #include "catalog/pg_statistic_ext.h"
111 : #include "executor/nodeAgg.h"
112 : #include "miscadmin.h"
113 : #include "nodes/makefuncs.h"
114 : #include "nodes/nodeFuncs.h"
115 : #include "optimizer/clauses.h"
116 : #include "optimizer/cost.h"
117 : #include "optimizer/optimizer.h"
118 : #include "optimizer/pathnode.h"
119 : #include "optimizer/paths.h"
120 : #include "optimizer/plancat.h"
121 : #include "parser/parse_clause.h"
122 : #include "parser/parse_relation.h"
123 : #include "parser/parsetree.h"
124 : #include "rewrite/rewriteManip.h"
125 : #include "statistics/statistics.h"
126 : #include "storage/bufmgr.h"
127 : #include "utils/acl.h"
128 : #include "utils/array.h"
129 : #include "utils/builtins.h"
130 : #include "utils/date.h"
131 : #include "utils/datum.h"
132 : #include "utils/fmgroids.h"
133 : #include "utils/index_selfuncs.h"
134 : #include "utils/lsyscache.h"
135 : #include "utils/memutils.h"
136 : #include "utils/pg_locale.h"
137 : #include "utils/rel.h"
138 : #include "utils/selfuncs.h"
139 : #include "utils/snapmgr.h"
140 : #include "utils/spccache.h"
141 : #include "utils/syscache.h"
142 : #include "utils/timestamp.h"
143 : #include "utils/typcache.h"
144 :
145 : #define DEFAULT_PAGE_CPU_MULTIPLIER 50.0
146 :
147 : /* Hooks for plugins to get control when we ask for stats */
148 : get_relation_stats_hook_type get_relation_stats_hook = NULL;
149 : get_index_stats_hook_type get_index_stats_hook = NULL;
150 :
151 : static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
152 : static double eqjoinsel_inner(Oid opfuncoid, Oid collation,
153 : VariableStatData *vardata1, VariableStatData *vardata2,
154 : double nd1, double nd2,
155 : bool isdefault1, bool isdefault2,
156 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
157 : Form_pg_statistic stats1, Form_pg_statistic stats2,
158 : bool have_mcvs1, bool have_mcvs2);
159 : static double eqjoinsel_semi(Oid opfuncoid, Oid collation,
160 : VariableStatData *vardata1, VariableStatData *vardata2,
161 : double nd1, double nd2,
162 : bool isdefault1, bool isdefault2,
163 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
164 : Form_pg_statistic stats1, Form_pg_statistic stats2,
165 : bool have_mcvs1, bool have_mcvs2,
166 : RelOptInfo *inner_rel);
167 : static bool estimate_multivariate_ndistinct(PlannerInfo *root,
168 : RelOptInfo *rel, List **varinfos, double *ndistinct);
169 : static bool convert_to_scalar(Datum value, Oid valuetypid, Oid collid,
170 : double *scaledvalue,
171 : Datum lobound, Datum hibound, Oid boundstypid,
172 : double *scaledlobound, double *scaledhibound);
173 : static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure);
174 : static void convert_string_to_scalar(char *value,
175 : double *scaledvalue,
176 : char *lobound,
177 : double *scaledlobound,
178 : char *hibound,
179 : double *scaledhibound);
180 : static void convert_bytea_to_scalar(Datum value,
181 : double *scaledvalue,
182 : Datum lobound,
183 : double *scaledlobound,
184 : Datum hibound,
185 : double *scaledhibound);
186 : static double convert_one_string_to_scalar(char *value,
187 : int rangelo, int rangehi);
188 : static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
189 : int rangelo, int rangehi);
190 : static char *convert_string_datum(Datum value, Oid typid, Oid collid,
191 : bool *failure);
192 : static double convert_timevalue_to_scalar(Datum value, Oid typid,
193 : bool *failure);
194 : static void examine_simple_variable(PlannerInfo *root, Var *var,
195 : VariableStatData *vardata);
196 : static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata,
197 : Oid sortop, Oid collation,
198 : Datum *min, Datum *max);
199 : static void get_stats_slot_range(AttStatsSlot *sslot,
200 : Oid opfuncoid, FmgrInfo *opproc,
201 : Oid collation, int16 typLen, bool typByVal,
202 : Datum *min, Datum *max, bool *p_have_data);
203 : static bool get_actual_variable_range(PlannerInfo *root,
204 : VariableStatData *vardata,
205 : Oid sortop, Oid collation,
206 : Datum *min, Datum *max);
207 : static bool get_actual_variable_endpoint(Relation heapRel,
208 : Relation indexRel,
209 : ScanDirection indexscandir,
210 : ScanKey scankeys,
211 : int16 typLen,
212 : bool typByVal,
213 : TupleTableSlot *tableslot,
214 : MemoryContext outercontext,
215 : Datum *endpointDatum);
216 : static RelOptInfo *find_join_input_rel(PlannerInfo *root, Relids relids);
217 :
218 :
219 : /*
220 : * eqsel - Selectivity of "=" for any data types.
221 : *
222 : * Note: this routine is also used to estimate selectivity for some
223 : * operators that are not "=" but have comparable selectivity behavior,
224 : * such as "~=" (geometric approximate-match). Even for "=", we must
225 : * keep in mind that the left and right datatypes may differ.
226 : */
227 : Datum
228 619978 : eqsel(PG_FUNCTION_ARGS)
229 : {
230 619978 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
231 : }
232 :
233 : /*
234 : * Common code for eqsel() and neqsel()
235 : */
236 : static double
237 670714 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
238 : {
239 670714 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
240 670714 : Oid operator = PG_GETARG_OID(1);
241 670714 : List *args = (List *) PG_GETARG_POINTER(2);
242 670714 : int varRelid = PG_GETARG_INT32(3);
243 670714 : Oid collation = PG_GET_COLLATION();
244 : VariableStatData vardata;
245 : Node *other;
246 : bool varonleft;
247 : double selec;
248 :
249 : /*
250 : * When asked about <>, we do the estimation using the corresponding =
251 : * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
252 : */
253 670714 : if (negate)
254 : {
255 50736 : operator = get_negator(operator);
256 50736 : if (!OidIsValid(operator))
257 : {
258 : /* Use default selectivity (should we raise an error instead?) */
259 0 : return 1.0 - DEFAULT_EQ_SEL;
260 : }
261 : }
262 :
263 : /*
264 : * If expression is not variable = something or something = variable, then
265 : * punt and return a default estimate.
266 : */
267 670714 : if (!get_restriction_variable(root, args, varRelid,
268 : &vardata, &other, &varonleft))
269 4908 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
270 :
271 : /*
272 : * We can do a lot better if the something is a constant. (Note: the
273 : * Const might result from estimation rather than being a simple constant
274 : * in the query.)
275 : */
276 665806 : if (IsA(other, Const))
277 279940 : selec = var_eq_const(&vardata, operator, collation,
278 279940 : ((Const *) other)->constvalue,
279 279940 : ((Const *) other)->constisnull,
280 : varonleft, negate);
281 : else
282 385866 : selec = var_eq_non_const(&vardata, operator, collation, other,
283 : varonleft, negate);
284 :
285 665806 : ReleaseVariableStats(vardata);
286 :
287 665806 : return selec;
288 : }
289 :
290 : /*
291 : * var_eq_const --- eqsel for var = const case
292 : *
293 : * This is exported so that some other estimation functions can use it.
294 : */
295 : double
296 326380 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
297 : Datum constval, bool constisnull,
298 : bool varonleft, bool negate)
299 : {
300 : double selec;
301 326380 : double nullfrac = 0.0;
302 : bool isdefault;
303 : Oid opfuncoid;
304 :
305 : /*
306 : * If the constant is NULL, assume operator is strict and return zero, ie,
307 : * operator will never return TRUE. (It's zero even for a negator op.)
308 : */
309 326380 : if (constisnull)
310 354 : return 0.0;
311 :
312 : /*
313 : * Grab the nullfrac for use below. Note we allow use of nullfrac
314 : * regardless of security check.
315 : */
316 326026 : if (HeapTupleIsValid(vardata->statsTuple))
317 : {
318 : Form_pg_statistic stats;
319 :
320 240382 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
321 240382 : nullfrac = stats->stanullfrac;
322 : }
323 :
324 : /*
325 : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
326 : * assume there is exactly one match regardless of anything else. (This
327 : * is slightly bogus, since the index or clause's equality operator might
328 : * be different from ours, but it's much more likely to be right than
329 : * ignoring the information.)
330 : */
331 326026 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
332 : {
333 82632 : selec = 1.0 / vardata->rel->tuples;
334 : }
335 417858 : else if (HeapTupleIsValid(vardata->statsTuple) &&
336 174464 : statistic_proc_security_check(vardata,
337 174464 : (opfuncoid = get_opcode(oproid))))
338 174464 : {
339 : AttStatsSlot sslot;
340 174464 : bool match = false;
341 : int i;
342 :
343 : /*
344 : * Is the constant "=" to any of the column's most common values?
345 : * (Although the given operator may not really be "=", we will assume
346 : * that seeing whether it returns TRUE is an appropriate test. If you
347 : * don't like this, maybe you shouldn't be using eqsel for your
348 : * operator...)
349 : */
350 174464 : if (get_attstatsslot(&sslot, vardata->statsTuple,
351 : STATISTIC_KIND_MCV, InvalidOid,
352 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
353 : {
354 159750 : LOCAL_FCINFO(fcinfo, 2);
355 : FmgrInfo eqproc;
356 :
357 159750 : fmgr_info(opfuncoid, &eqproc);
358 :
359 : /*
360 : * Save a few cycles by setting up the fcinfo struct just once.
361 : * Using FunctionCallInvoke directly also avoids failure if the
362 : * eqproc returns NULL, though really equality functions should
363 : * never do that.
364 : */
365 159750 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
366 : NULL, NULL);
367 159750 : fcinfo->args[0].isnull = false;
368 159750 : fcinfo->args[1].isnull = false;
369 : /* be careful to apply operator right way 'round */
370 159750 : if (varonleft)
371 159718 : fcinfo->args[1].value = constval;
372 : else
373 32 : fcinfo->args[0].value = constval;
374 :
375 2604744 : for (i = 0; i < sslot.nvalues; i++)
376 : {
377 : Datum fresult;
378 :
379 2522042 : if (varonleft)
380 2521986 : fcinfo->args[0].value = sslot.values[i];
381 : else
382 56 : fcinfo->args[1].value = sslot.values[i];
383 2522042 : fcinfo->isnull = false;
384 2522042 : fresult = FunctionCallInvoke(fcinfo);
385 2522042 : if (!fcinfo->isnull && DatumGetBool(fresult))
386 : {
387 77048 : match = true;
388 77048 : break;
389 : }
390 : }
391 : }
392 : else
393 : {
394 : /* no most-common-value info available */
395 14714 : i = 0; /* keep compiler quiet */
396 : }
397 :
398 174464 : if (match)
399 : {
400 : /*
401 : * Constant is "=" to this common value. We know selectivity
402 : * exactly (or as exactly as ANALYZE could calculate it, anyway).
403 : */
404 77048 : selec = sslot.numbers[i];
405 : }
406 : else
407 : {
408 : /*
409 : * Comparison is against a constant that is neither NULL nor any
410 : * of the common values. Its selectivity cannot be more than
411 : * this:
412 : */
413 97416 : double sumcommon = 0.0;
414 : double otherdistinct;
415 :
416 2267280 : for (i = 0; i < sslot.nnumbers; i++)
417 2169864 : sumcommon += sslot.numbers[i];
418 97416 : selec = 1.0 - sumcommon - nullfrac;
419 97416 : CLAMP_PROBABILITY(selec);
420 :
421 : /*
422 : * and in fact it's probably a good deal less. We approximate that
423 : * all the not-common values share this remaining fraction
424 : * equally, so we divide by the number of other distinct values.
425 : */
426 97416 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
427 97416 : sslot.nnumbers;
428 97416 : if (otherdistinct > 1)
429 42018 : selec /= otherdistinct;
430 :
431 : /*
432 : * Another cross-check: selectivity shouldn't be estimated as more
433 : * than the least common "most common value".
434 : */
435 97416 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
436 0 : selec = sslot.numbers[sslot.nnumbers - 1];
437 : }
438 :
439 174464 : free_attstatsslot(&sslot);
440 : }
441 : else
442 : {
443 : /*
444 : * No ANALYZE stats available, so make a guess using estimated number
445 : * of distinct values and assuming they are equally common. (The guess
446 : * is unlikely to be very good, but we do know a few special cases.)
447 : */
448 68930 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
449 : }
450 :
451 : /* now adjust if we wanted <> rather than = */
452 326026 : if (negate)
453 42218 : selec = 1.0 - selec - nullfrac;
454 :
455 : /* result should be in range, but make sure... */
456 326026 : CLAMP_PROBABILITY(selec);
457 :
458 326026 : return selec;
459 : }
460 :
461 : /*
462 : * var_eq_non_const --- eqsel for var = something-other-than-const case
463 : *
464 : * This is exported so that some other estimation functions can use it.
465 : */
466 : double
467 385866 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
468 : Node *other,
469 : bool varonleft, bool negate)
470 : {
471 : double selec;
472 385866 : double nullfrac = 0.0;
473 : bool isdefault;
474 :
475 : /*
476 : * Grab the nullfrac for use below.
477 : */
478 385866 : if (HeapTupleIsValid(vardata->statsTuple))
479 : {
480 : Form_pg_statistic stats;
481 :
482 275690 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
483 275690 : nullfrac = stats->stanullfrac;
484 : }
485 :
486 : /*
487 : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
488 : * assume there is exactly one match regardless of anything else. (This
489 : * is slightly bogus, since the index or clause's equality operator might
490 : * be different from ours, but it's much more likely to be right than
491 : * ignoring the information.)
492 : */
493 385866 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
494 : {
495 151008 : selec = 1.0 / vardata->rel->tuples;
496 : }
497 234858 : else if (HeapTupleIsValid(vardata->statsTuple))
498 : {
499 : double ndistinct;
500 : AttStatsSlot sslot;
501 :
502 : /*
503 : * Search is for a value that we do not know a priori, but we will
504 : * assume it is not NULL. Estimate the selectivity as non-null
505 : * fraction divided by number of distinct values, so that we get a
506 : * result averaged over all possible values whether common or
507 : * uncommon. (Essentially, we are assuming that the not-yet-known
508 : * comparison value is equally likely to be any of the possible
509 : * values, regardless of their frequency in the table. Is that a good
510 : * idea?)
511 : */
512 140156 : selec = 1.0 - nullfrac;
513 140156 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
514 140156 : if (ndistinct > 1)
515 135228 : selec /= ndistinct;
516 :
517 : /*
518 : * Cross-check: selectivity should never be estimated as more than the
519 : * most common value's.
520 : */
521 140156 : if (get_attstatsslot(&sslot, vardata->statsTuple,
522 : STATISTIC_KIND_MCV, InvalidOid,
523 : ATTSTATSSLOT_NUMBERS))
524 : {
525 120428 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
526 558 : selec = sslot.numbers[0];
527 120428 : free_attstatsslot(&sslot);
528 : }
529 : }
530 : else
531 : {
532 : /*
533 : * No ANALYZE stats available, so make a guess using estimated number
534 : * of distinct values and assuming they are equally common. (The guess
535 : * is unlikely to be very good, but we do know a few special cases.)
536 : */
537 94702 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
538 : }
539 :
540 : /* now adjust if we wanted <> rather than = */
541 385866 : if (negate)
542 6332 : selec = 1.0 - selec - nullfrac;
543 :
544 : /* result should be in range, but make sure... */
545 385866 : CLAMP_PROBABILITY(selec);
546 :
547 385866 : return selec;
548 : }
549 :
550 : /*
551 : * neqsel - Selectivity of "!=" for any data types.
552 : *
553 : * This routine is also used for some operators that are not "!="
554 : * but have comparable selectivity behavior. See above comments
555 : * for eqsel().
556 : */
557 : Datum
558 50736 : neqsel(PG_FUNCTION_ARGS)
559 : {
560 50736 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
561 : }
562 :
563 : /*
564 : * scalarineqsel - Selectivity of "<", "<=", ">", ">=" for scalars.
565 : *
566 : * This is the guts of scalarltsel/scalarlesel/scalargtsel/scalargesel.
567 : * The isgt and iseq flags distinguish which of the four cases apply.
568 : *
569 : * The caller has commuted the clause, if necessary, so that we can treat
570 : * the variable as being on the left. The caller must also make sure that
571 : * the other side of the clause is a non-null Const, and dissect that into
572 : * a value and datatype. (This definition simplifies some callers that
573 : * want to estimate against a computed value instead of a Const node.)
574 : *
575 : * This routine works for any datatype (or pair of datatypes) known to
576 : * convert_to_scalar(). If it is applied to some other datatype,
577 : * it will return an approximate estimate based on assuming that the constant
578 : * value falls in the middle of the bin identified by binary search.
579 : */
580 : static double
581 299046 : scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
582 : Oid collation,
583 : VariableStatData *vardata, Datum constval, Oid consttype)
584 : {
585 : Form_pg_statistic stats;
586 : FmgrInfo opproc;
587 : double mcv_selec,
588 : hist_selec,
589 : sumcommon;
590 : double selec;
591 :
592 299046 : if (!HeapTupleIsValid(vardata->statsTuple))
593 : {
594 : /*
595 : * No stats are available. Typically this means we have to fall back
596 : * on the default estimate; but if the variable is CTID then we can
597 : * make an estimate based on comparing the constant to the table size.
598 : */
599 21872 : if (vardata->var && IsA(vardata->var, Var) &&
600 17122 : ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
601 : {
602 : ItemPointer itemptr;
603 : double block;
604 : double density;
605 :
606 : /*
607 : * If the relation's empty, we're going to include all of it.
608 : * (This is mostly to avoid divide-by-zero below.)
609 : */
610 1948 : if (vardata->rel->pages == 0)
611 0 : return 1.0;
612 :
613 1948 : itemptr = (ItemPointer) DatumGetPointer(constval);
614 1948 : block = ItemPointerGetBlockNumberNoCheck(itemptr);
615 :
616 : /*
617 : * Determine the average number of tuples per page (density).
618 : *
619 : * Since the last page will, on average, be only half full, we can
620 : * estimate it to have half as many tuples as earlier pages. So
621 : * give it half the weight of a regular page.
622 : */
623 1948 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
624 :
625 : /* If target is the last page, use half the density. */
626 1948 : if (block >= vardata->rel->pages - 1)
627 30 : density *= 0.5;
628 :
629 : /*
630 : * Using the average tuples per page, calculate how far into the
631 : * page the itemptr is likely to be and adjust block accordingly,
632 : * by adding that fraction of a whole block (but never more than a
633 : * whole block, no matter how high the itemptr's offset is). Here
634 : * we are ignoring the possibility of dead-tuple line pointers,
635 : * which is fairly bogus, but we lack the info to do better.
636 : */
637 1948 : if (density > 0.0)
638 : {
639 1948 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
640 :
641 1948 : block += Min(offset / density, 1.0);
642 : }
643 :
644 : /*
645 : * Convert relative block number to selectivity. Again, the last
646 : * page has only half weight.
647 : */
648 1948 : selec = block / (vardata->rel->pages - 0.5);
649 :
650 : /*
651 : * The calculation so far gave us a selectivity for the "<=" case.
652 : * We'll have one fewer tuple for "<" and one additional tuple for
653 : * ">=", the latter of which we'll reverse the selectivity for
654 : * below, so we can simply subtract one tuple for both cases. The
655 : * cases that need this adjustment can be identified by iseq being
656 : * equal to isgt.
657 : */
658 1948 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
659 1836 : selec -= (1.0 / vardata->rel->tuples);
660 :
661 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
662 1948 : if (isgt)
663 1834 : selec = 1.0 - selec;
664 :
665 1948 : CLAMP_PROBABILITY(selec);
666 1948 : return selec;
667 : }
668 :
669 : /* no stats available, so default result */
670 19924 : return DEFAULT_INEQ_SEL;
671 : }
672 277174 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
673 :
674 277174 : fmgr_info(get_opcode(operator), &opproc);
675 :
676 : /*
677 : * If we have most-common-values info, add up the fractions of the MCV
678 : * entries that satisfy MCV OP CONST. These fractions contribute directly
679 : * to the result selectivity. Also add up the total fraction represented
680 : * by MCV entries.
681 : */
682 277174 : mcv_selec = mcv_selectivity(vardata, &opproc, collation, constval, true,
683 : &sumcommon);
684 :
685 : /*
686 : * If there is a histogram, determine which bin the constant falls in, and
687 : * compute the resulting contribution to selectivity.
688 : */
689 277174 : hist_selec = ineq_histogram_selectivity(root, vardata,
690 : operator, &opproc, isgt, iseq,
691 : collation,
692 : constval, consttype);
693 :
694 : /*
695 : * Now merge the results from the MCV and histogram calculations,
696 : * realizing that the histogram covers only the non-null values that are
697 : * not listed in MCV.
698 : */
699 277174 : selec = 1.0 - stats->stanullfrac - sumcommon;
700 :
701 277174 : if (hist_selec >= 0.0)
702 204082 : selec *= hist_selec;
703 : else
704 : {
705 : /*
706 : * If no histogram but there are values not accounted for by MCV,
707 : * arbitrarily assume half of them will match.
708 : */
709 73092 : selec *= 0.5;
710 : }
711 :
712 277174 : selec += mcv_selec;
713 :
714 : /* result should be in range, but make sure... */
715 277174 : CLAMP_PROBABILITY(selec);
716 :
717 277174 : return selec;
718 : }
719 :
720 : /*
721 : * mcv_selectivity - Examine the MCV list for selectivity estimates
722 : *
723 : * Determine the fraction of the variable's MCV population that satisfies
724 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft. Also
725 : * compute the fraction of the total column population represented by the MCV
726 : * list. This code will work for any boolean-returning predicate operator.
727 : *
728 : * The function result is the MCV selectivity, and the fraction of the
729 : * total population is returned into *sumcommonp. Zeroes are returned
730 : * if there is no MCV list.
731 : */
732 : double
733 283020 : mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation,
734 : Datum constval, bool varonleft,
735 : double *sumcommonp)
736 : {
737 : double mcv_selec,
738 : sumcommon;
739 : AttStatsSlot sslot;
740 : int i;
741 :
742 283020 : mcv_selec = 0.0;
743 283020 : sumcommon = 0.0;
744 :
745 563676 : if (HeapTupleIsValid(vardata->statsTuple) &&
746 561258 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
747 280602 : get_attstatsslot(&sslot, vardata->statsTuple,
748 : STATISTIC_KIND_MCV, InvalidOid,
749 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
750 : {
751 131804 : LOCAL_FCINFO(fcinfo, 2);
752 :
753 : /*
754 : * We invoke the opproc "by hand" so that we won't fail on NULL
755 : * results. Such cases won't arise for normal comparison functions,
756 : * but generic_restriction_selectivity could perhaps be used with
757 : * operators that can return NULL. A small side benefit is to not
758 : * need to re-initialize the fcinfo struct from scratch each time.
759 : */
760 131804 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
761 : NULL, NULL);
762 131804 : fcinfo->args[0].isnull = false;
763 131804 : fcinfo->args[1].isnull = false;
764 : /* be careful to apply operator right way 'round */
765 131804 : if (varonleft)
766 131804 : fcinfo->args[1].value = constval;
767 : else
768 0 : fcinfo->args[0].value = constval;
769 :
770 3985802 : for (i = 0; i < sslot.nvalues; i++)
771 : {
772 : Datum fresult;
773 :
774 3853998 : if (varonleft)
775 3853998 : fcinfo->args[0].value = sslot.values[i];
776 : else
777 0 : fcinfo->args[1].value = sslot.values[i];
778 3853998 : fcinfo->isnull = false;
779 3853998 : fresult = FunctionCallInvoke(fcinfo);
780 3853998 : if (!fcinfo->isnull && DatumGetBool(fresult))
781 1411596 : mcv_selec += sslot.numbers[i];
782 3853998 : sumcommon += sslot.numbers[i];
783 : }
784 131804 : free_attstatsslot(&sslot);
785 : }
786 :
787 283020 : *sumcommonp = sumcommon;
788 283020 : return mcv_selec;
789 : }
790 :
791 : /*
792 : * histogram_selectivity - Examine the histogram for selectivity estimates
793 : *
794 : * Determine the fraction of the variable's histogram entries that satisfy
795 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft.
796 : *
797 : * This code will work for any boolean-returning predicate operator, whether
798 : * or not it has anything to do with the histogram sort operator. We are
799 : * essentially using the histogram just as a representative sample. However,
800 : * small histograms are unlikely to be all that representative, so the caller
801 : * should be prepared to fall back on some other estimation approach when the
802 : * histogram is missing or very small. It may also be prudent to combine this
803 : * approach with another one when the histogram is small.
804 : *
805 : * If the actual histogram size is not at least min_hist_size, we won't bother
806 : * to do the calculation at all. Also, if the n_skip parameter is > 0, we
807 : * ignore the first and last n_skip histogram elements, on the grounds that
808 : * they are outliers and hence not very representative. Typical values for
809 : * these parameters are 10 and 1.
810 : *
811 : * The function result is the selectivity, or -1 if there is no histogram
812 : * or it's smaller than min_hist_size.
813 : *
814 : * The output parameter *hist_size receives the actual histogram size,
815 : * or zero if no histogram. Callers may use this number to decide how
816 : * much faith to put in the function result.
817 : *
818 : * Note that the result disregards both the most-common-values (if any) and
819 : * null entries. The caller is expected to combine this result with
820 : * statistics for those portions of the column population. It may also be
821 : * prudent to clamp the result range, ie, disbelieve exact 0 or 1 outputs.
822 : */
823 : double
824 5846 : histogram_selectivity(VariableStatData *vardata,
825 : FmgrInfo *opproc, Oid collation,
826 : Datum constval, bool varonleft,
827 : int min_hist_size, int n_skip,
828 : int *hist_size)
829 : {
830 : double result;
831 : AttStatsSlot sslot;
832 :
833 : /* check sanity of parameters */
834 : Assert(n_skip >= 0);
835 : Assert(min_hist_size > 2 * n_skip);
836 :
837 9328 : if (HeapTupleIsValid(vardata->statsTuple) &&
838 6958 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
839 3476 : get_attstatsslot(&sslot, vardata->statsTuple,
840 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
841 : ATTSTATSSLOT_VALUES))
842 : {
843 3382 : *hist_size = sslot.nvalues;
844 3382 : if (sslot.nvalues >= min_hist_size)
845 : {
846 1676 : LOCAL_FCINFO(fcinfo, 2);
847 1676 : int nmatch = 0;
848 : int i;
849 :
850 : /*
851 : * We invoke the opproc "by hand" so that we won't fail on NULL
852 : * results. Such cases won't arise for normal comparison
853 : * functions, but generic_restriction_selectivity could perhaps be
854 : * used with operators that can return NULL. A small side benefit
855 : * is to not need to re-initialize the fcinfo struct from scratch
856 : * each time.
857 : */
858 1676 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
859 : NULL, NULL);
860 1676 : fcinfo->args[0].isnull = false;
861 1676 : fcinfo->args[1].isnull = false;
862 : /* be careful to apply operator right way 'round */
863 1676 : if (varonleft)
864 1676 : fcinfo->args[1].value = constval;
865 : else
866 0 : fcinfo->args[0].value = constval;
867 :
868 139508 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
869 : {
870 : Datum fresult;
871 :
872 137832 : if (varonleft)
873 137832 : fcinfo->args[0].value = sslot.values[i];
874 : else
875 0 : fcinfo->args[1].value = sslot.values[i];
876 137832 : fcinfo->isnull = false;
877 137832 : fresult = FunctionCallInvoke(fcinfo);
878 137832 : if (!fcinfo->isnull && DatumGetBool(fresult))
879 6694 : nmatch++;
880 : }
881 1676 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
882 : }
883 : else
884 1706 : result = -1;
885 3382 : free_attstatsslot(&sslot);
886 : }
887 : else
888 : {
889 2464 : *hist_size = 0;
890 2464 : result = -1;
891 : }
892 :
893 5846 : return result;
894 : }
895 :
896 : /*
897 : * generic_restriction_selectivity - Selectivity for almost anything
898 : *
899 : * This function estimates selectivity for operators that we don't have any
900 : * special knowledge about, but are on data types that we collect standard
901 : * MCV and/or histogram statistics for. (Additional assumptions are that
902 : * the operator is strict and immutable, or at least stable.)
903 : *
904 : * If we have "VAR OP CONST" or "CONST OP VAR", selectivity is estimated by
905 : * applying the operator to each element of the column's MCV and/or histogram
906 : * stats, and merging the results using the assumption that the histogram is
907 : * a reasonable random sample of the column's non-MCV population. Note that
908 : * if the operator's semantics are related to the histogram ordering, this
909 : * might not be such a great assumption; other functions such as
910 : * scalarineqsel() are probably a better match in such cases.
911 : *
912 : * Otherwise, fall back to the default selectivity provided by the caller.
913 : */
914 : double
915 1106 : generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation,
916 : List *args, int varRelid,
917 : double default_selectivity)
918 : {
919 : double selec;
920 : VariableStatData vardata;
921 : Node *other;
922 : bool varonleft;
923 :
924 : /*
925 : * If expression is not variable OP something or something OP variable,
926 : * then punt and return the default estimate.
927 : */
928 1106 : if (!get_restriction_variable(root, args, varRelid,
929 : &vardata, &other, &varonleft))
930 0 : return default_selectivity;
931 :
932 : /*
933 : * If the something is a NULL constant, assume operator is strict and
934 : * return zero, ie, operator will never return TRUE.
935 : */
936 1106 : if (IsA(other, Const) &&
937 1106 : ((Const *) other)->constisnull)
938 : {
939 0 : ReleaseVariableStats(vardata);
940 0 : return 0.0;
941 : }
942 :
943 1106 : if (IsA(other, Const))
944 : {
945 : /* Variable is being compared to a known non-null constant */
946 1106 : Datum constval = ((Const *) other)->constvalue;
947 : FmgrInfo opproc;
948 : double mcvsum;
949 : double mcvsel;
950 : double nullfrac;
951 : int hist_size;
952 :
953 1106 : fmgr_info(get_opcode(oproid), &opproc);
954 :
955 : /*
956 : * Calculate the selectivity for the column's most common values.
957 : */
958 1106 : mcvsel = mcv_selectivity(&vardata, &opproc, collation,
959 : constval, varonleft,
960 : &mcvsum);
961 :
962 : /*
963 : * If the histogram is large enough, see what fraction of it matches
964 : * the query, and assume that's representative of the non-MCV
965 : * population. Otherwise use the default selectivity for the non-MCV
966 : * population.
967 : */
968 1106 : selec = histogram_selectivity(&vardata, &opproc, collation,
969 : constval, varonleft,
970 : 10, 1, &hist_size);
971 1106 : if (selec < 0)
972 : {
973 : /* Nope, fall back on default */
974 1106 : selec = default_selectivity;
975 : }
976 0 : else if (hist_size < 100)
977 : {
978 : /*
979 : * For histogram sizes from 10 to 100, we combine the histogram
980 : * and default selectivities, putting increasingly more trust in
981 : * the histogram for larger sizes.
982 : */
983 0 : double hist_weight = hist_size / 100.0;
984 :
985 0 : selec = selec * hist_weight +
986 0 : default_selectivity * (1.0 - hist_weight);
987 : }
988 :
989 : /* In any case, don't believe extremely small or large estimates. */
990 1106 : if (selec < 0.0001)
991 0 : selec = 0.0001;
992 1106 : else if (selec > 0.9999)
993 0 : selec = 0.9999;
994 :
995 : /* Don't forget to account for nulls. */
996 1106 : if (HeapTupleIsValid(vardata.statsTuple))
997 84 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
998 : else
999 1022 : nullfrac = 0.0;
1000 :
1001 : /*
1002 : * Now merge the results from the MCV and histogram calculations,
1003 : * realizing that the histogram covers only the non-null values that
1004 : * are not listed in MCV.
1005 : */
1006 1106 : selec *= 1.0 - nullfrac - mcvsum;
1007 1106 : selec += mcvsel;
1008 : }
1009 : else
1010 : {
1011 : /* Comparison value is not constant, so we can't do anything */
1012 0 : selec = default_selectivity;
1013 : }
1014 :
1015 1106 : ReleaseVariableStats(vardata);
1016 :
1017 : /* result should be in range, but make sure... */
1018 1106 : CLAMP_PROBABILITY(selec);
1019 :
1020 1106 : return selec;
1021 : }
1022 :
1023 : /*
1024 : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
1025 : *
1026 : * Determine the fraction of the variable's histogram population that
1027 : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
1028 : * The isgt and iseq flags distinguish which of the four cases apply.
1029 : *
1030 : * While opproc could be looked up from the operator OID, common callers
1031 : * also need to call it separately, so we make the caller pass both.
1032 : *
1033 : * Returns -1 if there is no histogram (valid results will always be >= 0).
1034 : *
1035 : * Note that the result disregards both the most-common-values (if any) and
1036 : * null entries. The caller is expected to combine this result with
1037 : * statistics for those portions of the column population.
1038 : *
1039 : * This is exported so that some other estimation functions can use it.
1040 : */
1041 : double
1042 281876 : ineq_histogram_selectivity(PlannerInfo *root,
1043 : VariableStatData *vardata,
1044 : Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq,
1045 : Oid collation,
1046 : Datum constval, Oid consttype)
1047 : {
1048 : double hist_selec;
1049 : AttStatsSlot sslot;
1050 :
1051 281876 : hist_selec = -1.0;
1052 :
1053 : /*
1054 : * Someday, ANALYZE might store more than one histogram per rel/att,
1055 : * corresponding to more than one possible sort ordering defined for the
1056 : * column type. Right now, we know there is only one, so just grab it and
1057 : * see if it matches the query.
1058 : *
1059 : * Note that we can't use opoid as search argument; the staop appearing in
1060 : * pg_statistic will be for the relevant '<' operator, but what we have
1061 : * might be some other inequality operator such as '>='. (Even if opoid
1062 : * is a '<' operator, it could be cross-type.) Hence we must use
1063 : * comparison_ops_are_compatible() to see if the operators match.
1064 : */
1065 563096 : if (HeapTupleIsValid(vardata->statsTuple) &&
1066 562392 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1067 281172 : get_attstatsslot(&sslot, vardata->statsTuple,
1068 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1069 : ATTSTATSSLOT_VALUES))
1070 : {
1071 208126 : if (sslot.nvalues > 1 &&
1072 416176 : sslot.stacoll == collation &&
1073 208050 : comparison_ops_are_compatible(sslot.staop, opoid))
1074 207942 : {
1075 : /*
1076 : * Use binary search to find the desired location, namely the
1077 : * right end of the histogram bin containing the comparison value,
1078 : * which is the leftmost entry for which the comparison operator
1079 : * succeeds (if isgt) or fails (if !isgt).
1080 : *
1081 : * In this loop, we pay no attention to whether the operator iseq
1082 : * or not; that detail will be mopped up below. (We cannot tell,
1083 : * anyway, whether the operator thinks the values are equal.)
1084 : *
1085 : * If the binary search accesses the first or last histogram
1086 : * entry, we try to replace that endpoint with the true column min
1087 : * or max as found by get_actual_variable_range(). This
1088 : * ameliorates misestimates when the min or max is moving as a
1089 : * result of changes since the last ANALYZE. Note that this could
1090 : * result in effectively including MCVs into the histogram that
1091 : * weren't there before, but we don't try to correct for that.
1092 : */
1093 : double histfrac;
1094 207942 : int lobound = 0; /* first possible slot to search */
1095 207942 : int hibound = sslot.nvalues; /* last+1 slot to search */
1096 207942 : bool have_end = false;
1097 :
1098 : /*
1099 : * If there are only two histogram entries, we'll want up-to-date
1100 : * values for both. (If there are more than two, we need at most
1101 : * one of them to be updated, so we deal with that within the
1102 : * loop.)
1103 : */
1104 207942 : if (sslot.nvalues == 2)
1105 2018 : have_end = get_actual_variable_range(root,
1106 : vardata,
1107 : sslot.staop,
1108 : collation,
1109 : &sslot.values[0],
1110 2018 : &sslot.values[1]);
1111 :
1112 1343882 : while (lobound < hibound)
1113 : {
1114 1135940 : int probe = (lobound + hibound) / 2;
1115 : bool ltcmp;
1116 :
1117 : /*
1118 : * If we find ourselves about to compare to the first or last
1119 : * histogram entry, first try to replace it with the actual
1120 : * current min or max (unless we already did so above).
1121 : */
1122 1135940 : if (probe == 0 && sslot.nvalues > 2)
1123 105116 : have_end = get_actual_variable_range(root,
1124 : vardata,
1125 : sslot.staop,
1126 : collation,
1127 : &sslot.values[0],
1128 : NULL);
1129 1030824 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1130 69406 : have_end = get_actual_variable_range(root,
1131 : vardata,
1132 : sslot.staop,
1133 : collation,
1134 : NULL,
1135 69406 : &sslot.values[probe]);
1136 :
1137 1135940 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1138 : collation,
1139 1135940 : sslot.values[probe],
1140 : constval));
1141 1135940 : if (isgt)
1142 64808 : ltcmp = !ltcmp;
1143 1135940 : if (ltcmp)
1144 428662 : lobound = probe + 1;
1145 : else
1146 707278 : hibound = probe;
1147 : }
1148 :
1149 207942 : if (lobound <= 0)
1150 : {
1151 : /*
1152 : * Constant is below lower histogram boundary. More
1153 : * precisely, we have found that no entry in the histogram
1154 : * satisfies the inequality clause (if !isgt) or they all do
1155 : * (if isgt). We estimate that that's true of the entire
1156 : * table, so set histfrac to 0.0 (which we'll flip to 1.0
1157 : * below, if isgt).
1158 : */
1159 90746 : histfrac = 0.0;
1160 : }
1161 117196 : else if (lobound >= sslot.nvalues)
1162 : {
1163 : /*
1164 : * Inverse case: constant is above upper histogram boundary.
1165 : */
1166 30544 : histfrac = 1.0;
1167 : }
1168 : else
1169 : {
1170 : /* We have values[i-1] <= constant <= values[i]. */
1171 86652 : int i = lobound;
1172 86652 : double eq_selec = 0;
1173 : double val,
1174 : high,
1175 : low;
1176 : double binfrac;
1177 :
1178 : /*
1179 : * In the cases where we'll need it below, obtain an estimate
1180 : * of the selectivity of "x = constval". We use a calculation
1181 : * similar to what var_eq_const() does for a non-MCV constant,
1182 : * ie, estimate that all distinct non-MCV values occur equally
1183 : * often. But multiplication by "1.0 - sumcommon - nullfrac"
1184 : * will be done by our caller, so we shouldn't do that here.
1185 : * Therefore we can't try to clamp the estimate by reference
1186 : * to the least common MCV; the result would be too small.
1187 : *
1188 : * Note: since this is effectively assuming that constval
1189 : * isn't an MCV, it's logically dubious if constval in fact is
1190 : * one. But we have to apply *some* correction for equality,
1191 : * and anyway we cannot tell if constval is an MCV, since we
1192 : * don't have a suitable equality operator at hand.
1193 : */
1194 86652 : if (i == 1 || isgt == iseq)
1195 : {
1196 : double otherdistinct;
1197 : bool isdefault;
1198 : AttStatsSlot mcvslot;
1199 :
1200 : /* Get estimated number of distinct values */
1201 34960 : otherdistinct = get_variable_numdistinct(vardata,
1202 : &isdefault);
1203 :
1204 : /* Subtract off the number of known MCVs */
1205 34960 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1206 : STATISTIC_KIND_MCV, InvalidOid,
1207 : ATTSTATSSLOT_NUMBERS))
1208 : {
1209 3350 : otherdistinct -= mcvslot.nnumbers;
1210 3350 : free_attstatsslot(&mcvslot);
1211 : }
1212 :
1213 : /* If result doesn't seem sane, leave eq_selec at 0 */
1214 34960 : if (otherdistinct > 1)
1215 34918 : eq_selec = 1.0 / otherdistinct;
1216 : }
1217 :
1218 : /*
1219 : * Convert the constant and the two nearest bin boundary
1220 : * values to a uniform comparison scale, and do a linear
1221 : * interpolation within this bin.
1222 : */
1223 86652 : if (convert_to_scalar(constval, consttype, collation,
1224 : &val,
1225 86652 : sslot.values[i - 1], sslot.values[i],
1226 : vardata->vartype,
1227 : &low, &high))
1228 : {
1229 86652 : if (high <= low)
1230 : {
1231 : /* cope if bin boundaries appear identical */
1232 0 : binfrac = 0.5;
1233 : }
1234 86652 : else if (val <= low)
1235 21010 : binfrac = 0.0;
1236 65642 : else if (val >= high)
1237 3864 : binfrac = 1.0;
1238 : else
1239 : {
1240 61778 : binfrac = (val - low) / (high - low);
1241 :
1242 : /*
1243 : * Watch out for the possibility that we got a NaN or
1244 : * Infinity from the division. This can happen
1245 : * despite the previous checks, if for example "low"
1246 : * is -Infinity.
1247 : */
1248 61778 : if (isnan(binfrac) ||
1249 61778 : binfrac < 0.0 || binfrac > 1.0)
1250 0 : binfrac = 0.5;
1251 : }
1252 : }
1253 : else
1254 : {
1255 : /*
1256 : * Ideally we'd produce an error here, on the grounds that
1257 : * the given operator shouldn't have scalarXXsel
1258 : * registered as its selectivity func unless we can deal
1259 : * with its operand types. But currently, all manner of
1260 : * stuff is invoking scalarXXsel, so give a default
1261 : * estimate until that can be fixed.
1262 : */
1263 0 : binfrac = 0.5;
1264 : }
1265 :
1266 : /*
1267 : * Now, compute the overall selectivity across the values
1268 : * represented by the histogram. We have i-1 full bins and
1269 : * binfrac partial bin below the constant.
1270 : */
1271 86652 : histfrac = (double) (i - 1) + binfrac;
1272 86652 : histfrac /= (double) (sslot.nvalues - 1);
1273 :
1274 : /*
1275 : * At this point, histfrac is an estimate of the fraction of
1276 : * the population represented by the histogram that satisfies
1277 : * "x <= constval". Somewhat remarkably, this statement is
1278 : * true regardless of which operator we were doing the probes
1279 : * with, so long as convert_to_scalar() delivers reasonable
1280 : * results. If the probe constant is equal to some histogram
1281 : * entry, we would have considered the bin to the left of that
1282 : * entry if probing with "<" or ">=", or the bin to the right
1283 : * if probing with "<=" or ">"; but binfrac would have come
1284 : * out as 1.0 in the first case and 0.0 in the second, leading
1285 : * to the same histfrac in either case. For probe constants
1286 : * between histogram entries, we find the same bin and get the
1287 : * same estimate with any operator.
1288 : *
1289 : * The fact that the estimate corresponds to "x <= constval"
1290 : * and not "x < constval" is because of the way that ANALYZE
1291 : * constructs the histogram: each entry is, effectively, the
1292 : * rightmost value in its sample bucket. So selectivity
1293 : * values that are exact multiples of 1/(histogram_size-1)
1294 : * should be understood as estimates including a histogram
1295 : * entry plus everything to its left.
1296 : *
1297 : * However, that breaks down for the first histogram entry,
1298 : * which necessarily is the leftmost value in its sample
1299 : * bucket. That means the first histogram bin is slightly
1300 : * narrower than the rest, by an amount equal to eq_selec.
1301 : * Another way to say that is that we want "x <= leftmost" to
1302 : * be estimated as eq_selec not zero. So, if we're dealing
1303 : * with the first bin (i==1), rescale to make that true while
1304 : * adjusting the rest of that bin linearly.
1305 : */
1306 86652 : if (i == 1)
1307 15706 : histfrac += eq_selec * (1.0 - binfrac);
1308 :
1309 : /*
1310 : * "x <= constval" is good if we want an estimate for "<=" or
1311 : * ">", but if we are estimating for "<" or ">=", we now need
1312 : * to decrease the estimate by eq_selec.
1313 : */
1314 86652 : if (isgt == iseq)
1315 25870 : histfrac -= eq_selec;
1316 : }
1317 :
1318 : /*
1319 : * Now the estimate is finished for "<" and "<=" cases. If we are
1320 : * estimating for ">" or ">=", flip it.
1321 : */
1322 207942 : hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1323 :
1324 : /*
1325 : * The histogram boundaries are only approximate to begin with,
1326 : * and may well be out of date anyway. Therefore, don't believe
1327 : * extremely small or large selectivity estimates --- unless we
1328 : * got actual current endpoint values from the table, in which
1329 : * case just do the usual sanity clamp. Somewhat arbitrarily, we
1330 : * set the cutoff for other cases at a hundredth of the histogram
1331 : * resolution.
1332 : */
1333 207942 : if (have_end)
1334 117962 : CLAMP_PROBABILITY(hist_selec);
1335 : else
1336 : {
1337 89980 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1338 :
1339 89980 : if (hist_selec < cutoff)
1340 32390 : hist_selec = cutoff;
1341 57590 : else if (hist_selec > 1.0 - cutoff)
1342 20996 : hist_selec = 1.0 - cutoff;
1343 : }
1344 : }
1345 184 : else if (sslot.nvalues > 1)
1346 : {
1347 : /*
1348 : * If we get here, we have a histogram but it's not sorted the way
1349 : * we want. Do a brute-force search to see how many of the
1350 : * entries satisfy the comparison condition, and take that
1351 : * fraction as our estimate. (This is identical to the inner loop
1352 : * of histogram_selectivity; maybe share code?)
1353 : */
1354 184 : LOCAL_FCINFO(fcinfo, 2);
1355 184 : int nmatch = 0;
1356 :
1357 184 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1358 : NULL, NULL);
1359 184 : fcinfo->args[0].isnull = false;
1360 184 : fcinfo->args[1].isnull = false;
1361 184 : fcinfo->args[1].value = constval;
1362 962108 : for (int i = 0; i < sslot.nvalues; i++)
1363 : {
1364 : Datum fresult;
1365 :
1366 961924 : fcinfo->args[0].value = sslot.values[i];
1367 961924 : fcinfo->isnull = false;
1368 961924 : fresult = FunctionCallInvoke(fcinfo);
1369 961924 : if (!fcinfo->isnull && DatumGetBool(fresult))
1370 1940 : nmatch++;
1371 : }
1372 184 : hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1373 :
1374 : /*
1375 : * As above, clamp to a hundredth of the histogram resolution.
1376 : * This case is surely even less trustworthy than the normal one,
1377 : * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1378 : * clamp should be more restrictive in this case?)
1379 : */
1380 : {
1381 184 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1382 :
1383 184 : if (hist_selec < cutoff)
1384 12 : hist_selec = cutoff;
1385 172 : else if (hist_selec > 1.0 - cutoff)
1386 12 : hist_selec = 1.0 - cutoff;
1387 : }
1388 : }
1389 :
1390 208126 : free_attstatsslot(&sslot);
1391 : }
1392 :
1393 281876 : return hist_selec;
1394 : }
1395 :
1396 : /*
1397 : * Common wrapper function for the selectivity estimators that simply
1398 : * invoke scalarineqsel().
1399 : */
1400 : static Datum
1401 43254 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1402 : {
1403 43254 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1404 43254 : Oid operator = PG_GETARG_OID(1);
1405 43254 : List *args = (List *) PG_GETARG_POINTER(2);
1406 43254 : int varRelid = PG_GETARG_INT32(3);
1407 43254 : Oid collation = PG_GET_COLLATION();
1408 : VariableStatData vardata;
1409 : Node *other;
1410 : bool varonleft;
1411 : Datum constval;
1412 : Oid consttype;
1413 : double selec;
1414 :
1415 : /*
1416 : * If expression is not variable op something or something op variable,
1417 : * then punt and return a default estimate.
1418 : */
1419 43254 : if (!get_restriction_variable(root, args, varRelid,
1420 : &vardata, &other, &varonleft))
1421 624 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1422 :
1423 : /*
1424 : * Can't do anything useful if the something is not a constant, either.
1425 : */
1426 42630 : if (!IsA(other, Const))
1427 : {
1428 2686 : ReleaseVariableStats(vardata);
1429 2686 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1430 : }
1431 :
1432 : /*
1433 : * If the constant is NULL, assume operator is strict and return zero, ie,
1434 : * operator will never return TRUE.
1435 : */
1436 39944 : if (((Const *) other)->constisnull)
1437 : {
1438 66 : ReleaseVariableStats(vardata);
1439 66 : PG_RETURN_FLOAT8(0.0);
1440 : }
1441 39878 : constval = ((Const *) other)->constvalue;
1442 39878 : consttype = ((Const *) other)->consttype;
1443 :
1444 : /*
1445 : * Force the var to be on the left to simplify logic in scalarineqsel.
1446 : */
1447 39878 : if (!varonleft)
1448 : {
1449 366 : operator = get_commutator(operator);
1450 366 : if (!operator)
1451 : {
1452 : /* Use default selectivity (should we raise an error instead?) */
1453 0 : ReleaseVariableStats(vardata);
1454 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1455 : }
1456 366 : isgt = !isgt;
1457 : }
1458 :
1459 : /* The rest of the work is done by scalarineqsel(). */
1460 39878 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1461 : &vardata, constval, consttype);
1462 :
1463 39878 : ReleaseVariableStats(vardata);
1464 :
1465 39878 : PG_RETURN_FLOAT8((float8) selec);
1466 : }
1467 :
1468 : /*
1469 : * scalarltsel - Selectivity of "<" for scalars.
1470 : */
1471 : Datum
1472 14228 : scalarltsel(PG_FUNCTION_ARGS)
1473 : {
1474 14228 : return scalarineqsel_wrapper(fcinfo, false, false);
1475 : }
1476 :
1477 : /*
1478 : * scalarlesel - Selectivity of "<=" for scalars.
1479 : */
1480 : Datum
1481 4434 : scalarlesel(PG_FUNCTION_ARGS)
1482 : {
1483 4434 : return scalarineqsel_wrapper(fcinfo, false, true);
1484 : }
1485 :
1486 : /*
1487 : * scalargtsel - Selectivity of ">" for scalars.
1488 : */
1489 : Datum
1490 14952 : scalargtsel(PG_FUNCTION_ARGS)
1491 : {
1492 14952 : return scalarineqsel_wrapper(fcinfo, true, false);
1493 : }
1494 :
1495 : /*
1496 : * scalargesel - Selectivity of ">=" for scalars.
1497 : */
1498 : Datum
1499 9640 : scalargesel(PG_FUNCTION_ARGS)
1500 : {
1501 9640 : return scalarineqsel_wrapper(fcinfo, true, true);
1502 : }
1503 :
1504 : /*
1505 : * boolvarsel - Selectivity of Boolean variable.
1506 : *
1507 : * This can actually be called on any boolean-valued expression. If it
1508 : * involves only Vars of the specified relation, and if there are statistics
1509 : * about the Var or expression (the latter is possible if it's indexed) then
1510 : * we'll produce a real estimate; otherwise it's just a default.
1511 : */
1512 : Selectivity
1513 46186 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1514 : {
1515 : VariableStatData vardata;
1516 : double selec;
1517 :
1518 46186 : examine_variable(root, arg, varRelid, &vardata);
1519 46186 : if (HeapTupleIsValid(vardata.statsTuple))
1520 : {
1521 : /*
1522 : * A boolean variable V is equivalent to the clause V = 't', so we
1523 : * compute the selectivity as if that is what we have.
1524 : */
1525 37926 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1526 : BoolGetDatum(true), false, true, false);
1527 : }
1528 : else
1529 : {
1530 : /* Otherwise, the default estimate is 0.5 */
1531 8260 : selec = 0.5;
1532 : }
1533 46186 : ReleaseVariableStats(vardata);
1534 46186 : return selec;
1535 : }
1536 :
1537 : /*
1538 : * booltestsel - Selectivity of BooleanTest Node.
1539 : */
1540 : Selectivity
1541 872 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1542 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1543 : {
1544 : VariableStatData vardata;
1545 : double selec;
1546 :
1547 872 : examine_variable(root, arg, varRelid, &vardata);
1548 :
1549 872 : if (HeapTupleIsValid(vardata.statsTuple))
1550 : {
1551 : Form_pg_statistic stats;
1552 : double freq_null;
1553 : AttStatsSlot sslot;
1554 :
1555 0 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1556 0 : freq_null = stats->stanullfrac;
1557 :
1558 0 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1559 : STATISTIC_KIND_MCV, InvalidOid,
1560 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1561 0 : && sslot.nnumbers > 0)
1562 0 : {
1563 : double freq_true;
1564 : double freq_false;
1565 :
1566 : /*
1567 : * Get first MCV frequency and derive frequency for true.
1568 : */
1569 0 : if (DatumGetBool(sslot.values[0]))
1570 0 : freq_true = sslot.numbers[0];
1571 : else
1572 0 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1573 :
1574 : /*
1575 : * Next derive frequency for false. Then use these as appropriate
1576 : * to derive frequency for each case.
1577 : */
1578 0 : freq_false = 1.0 - freq_true - freq_null;
1579 :
1580 0 : switch (booltesttype)
1581 : {
1582 0 : case IS_UNKNOWN:
1583 : /* select only NULL values */
1584 0 : selec = freq_null;
1585 0 : break;
1586 0 : case IS_NOT_UNKNOWN:
1587 : /* select non-NULL values */
1588 0 : selec = 1.0 - freq_null;
1589 0 : break;
1590 0 : case IS_TRUE:
1591 : /* select only TRUE values */
1592 0 : selec = freq_true;
1593 0 : break;
1594 0 : case IS_NOT_TRUE:
1595 : /* select non-TRUE values */
1596 0 : selec = 1.0 - freq_true;
1597 0 : break;
1598 0 : case IS_FALSE:
1599 : /* select only FALSE values */
1600 0 : selec = freq_false;
1601 0 : break;
1602 0 : case IS_NOT_FALSE:
1603 : /* select non-FALSE values */
1604 0 : selec = 1.0 - freq_false;
1605 0 : break;
1606 0 : default:
1607 0 : elog(ERROR, "unrecognized booltesttype: %d",
1608 : (int) booltesttype);
1609 : selec = 0.0; /* Keep compiler quiet */
1610 : break;
1611 : }
1612 :
1613 0 : free_attstatsslot(&sslot);
1614 : }
1615 : else
1616 : {
1617 : /*
1618 : * No most-common-value info available. Still have null fraction
1619 : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1620 : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1621 : */
1622 0 : switch (booltesttype)
1623 : {
1624 0 : case IS_UNKNOWN:
1625 : /* select only NULL values */
1626 0 : selec = freq_null;
1627 0 : break;
1628 0 : case IS_NOT_UNKNOWN:
1629 : /* select non-NULL values */
1630 0 : selec = 1.0 - freq_null;
1631 0 : break;
1632 0 : case IS_TRUE:
1633 : case IS_FALSE:
1634 : /* Assume we select half of the non-NULL values */
1635 0 : selec = (1.0 - freq_null) / 2.0;
1636 0 : break;
1637 0 : case IS_NOT_TRUE:
1638 : case IS_NOT_FALSE:
1639 : /* Assume we select NULLs plus half of the non-NULLs */
1640 : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1641 0 : selec = (freq_null + 1.0) / 2.0;
1642 0 : break;
1643 0 : default:
1644 0 : elog(ERROR, "unrecognized booltesttype: %d",
1645 : (int) booltesttype);
1646 : selec = 0.0; /* Keep compiler quiet */
1647 : break;
1648 : }
1649 : }
1650 : }
1651 : else
1652 : {
1653 : /*
1654 : * If we can't get variable statistics for the argument, perhaps
1655 : * clause_selectivity can do something with it. We ignore the
1656 : * possibility of a NULL value when using clause_selectivity, and just
1657 : * assume the value is either TRUE or FALSE.
1658 : */
1659 872 : switch (booltesttype)
1660 : {
1661 48 : case IS_UNKNOWN:
1662 48 : selec = DEFAULT_UNK_SEL;
1663 48 : break;
1664 108 : case IS_NOT_UNKNOWN:
1665 108 : selec = DEFAULT_NOT_UNK_SEL;
1666 108 : break;
1667 252 : case IS_TRUE:
1668 : case IS_NOT_FALSE:
1669 252 : selec = (double) clause_selectivity(root, arg,
1670 : varRelid,
1671 : jointype, sjinfo);
1672 252 : break;
1673 464 : case IS_FALSE:
1674 : case IS_NOT_TRUE:
1675 464 : selec = 1.0 - (double) clause_selectivity(root, arg,
1676 : varRelid,
1677 : jointype, sjinfo);
1678 464 : break;
1679 0 : default:
1680 0 : elog(ERROR, "unrecognized booltesttype: %d",
1681 : (int) booltesttype);
1682 : selec = 0.0; /* Keep compiler quiet */
1683 : break;
1684 : }
1685 : }
1686 :
1687 872 : ReleaseVariableStats(vardata);
1688 :
1689 : /* result should be in range, but make sure... */
1690 872 : CLAMP_PROBABILITY(selec);
1691 :
1692 872 : return (Selectivity) selec;
1693 : }
1694 :
1695 : /*
1696 : * nulltestsel - Selectivity of NullTest Node.
1697 : */
1698 : Selectivity
1699 16632 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1700 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1701 : {
1702 : VariableStatData vardata;
1703 : double selec;
1704 :
1705 16632 : examine_variable(root, arg, varRelid, &vardata);
1706 :
1707 16632 : if (HeapTupleIsValid(vardata.statsTuple))
1708 : {
1709 : Form_pg_statistic stats;
1710 : double freq_null;
1711 :
1712 9374 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1713 9374 : freq_null = stats->stanullfrac;
1714 :
1715 9374 : switch (nulltesttype)
1716 : {
1717 6862 : case IS_NULL:
1718 :
1719 : /*
1720 : * Use freq_null directly.
1721 : */
1722 6862 : selec = freq_null;
1723 6862 : break;
1724 2512 : case IS_NOT_NULL:
1725 :
1726 : /*
1727 : * Select not unknown (not null) values. Calculate from
1728 : * freq_null.
1729 : */
1730 2512 : selec = 1.0 - freq_null;
1731 2512 : break;
1732 0 : default:
1733 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1734 : (int) nulltesttype);
1735 : return (Selectivity) 0; /* keep compiler quiet */
1736 : }
1737 : }
1738 7258 : else if (vardata.var && IsA(vardata.var, Var) &&
1739 6556 : ((Var *) vardata.var)->varattno < 0)
1740 : {
1741 : /*
1742 : * There are no stats for system columns, but we know they are never
1743 : * NULL.
1744 : */
1745 60 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1746 : }
1747 : else
1748 : {
1749 : /*
1750 : * No ANALYZE stats available, so make a guess
1751 : */
1752 7198 : switch (nulltesttype)
1753 : {
1754 2116 : case IS_NULL:
1755 2116 : selec = DEFAULT_UNK_SEL;
1756 2116 : break;
1757 5082 : case IS_NOT_NULL:
1758 5082 : selec = DEFAULT_NOT_UNK_SEL;
1759 5082 : break;
1760 0 : default:
1761 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1762 : (int) nulltesttype);
1763 : return (Selectivity) 0; /* keep compiler quiet */
1764 : }
1765 : }
1766 :
1767 16632 : ReleaseVariableStats(vardata);
1768 :
1769 : /* result should be in range, but make sure... */
1770 16632 : CLAMP_PROBABILITY(selec);
1771 :
1772 16632 : return (Selectivity) selec;
1773 : }
1774 :
1775 : /*
1776 : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1777 : *
1778 : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1779 : * but it seems possible that RelabelType might show up. Also, the planner
1780 : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1781 : * so we need to be ready to deal with more than one level.
1782 : */
1783 : static Node *
1784 126236 : strip_array_coercion(Node *node)
1785 : {
1786 : for (;;)
1787 : {
1788 126236 : if (node && IsA(node, ArrayCoerceExpr))
1789 36 : {
1790 2866 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1791 :
1792 : /*
1793 : * If the per-element expression is just a RelabelType on top of
1794 : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1795 : */
1796 2866 : if (IsA(acoerce->elemexpr, RelabelType) &&
1797 36 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1798 36 : node = (Node *) acoerce->arg;
1799 : else
1800 : break;
1801 : }
1802 123370 : else if (node && IsA(node, RelabelType))
1803 : {
1804 : /* We don't really expect this case, but may as well cope */
1805 0 : node = (Node *) ((RelabelType *) node)->arg;
1806 : }
1807 : else
1808 : break;
1809 : }
1810 126200 : return node;
1811 : }
1812 :
1813 : /*
1814 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1815 : */
1816 : Selectivity
1817 22386 : scalararraysel(PlannerInfo *root,
1818 : ScalarArrayOpExpr *clause,
1819 : bool is_join_clause,
1820 : int varRelid,
1821 : JoinType jointype,
1822 : SpecialJoinInfo *sjinfo)
1823 : {
1824 22386 : Oid operator = clause->opno;
1825 22386 : bool useOr = clause->useOr;
1826 22386 : bool isEquality = false;
1827 22386 : bool isInequality = false;
1828 : Node *leftop;
1829 : Node *rightop;
1830 : Oid nominal_element_type;
1831 : Oid nominal_element_collation;
1832 : TypeCacheEntry *typentry;
1833 : RegProcedure oprsel;
1834 : FmgrInfo oprselproc;
1835 : Selectivity s1;
1836 : Selectivity s1disjoint;
1837 :
1838 : /* First, deconstruct the expression */
1839 : Assert(list_length(clause->args) == 2);
1840 22386 : leftop = (Node *) linitial(clause->args);
1841 22386 : rightop = (Node *) lsecond(clause->args);
1842 :
1843 : /* aggressively reduce both sides to constants */
1844 22386 : leftop = estimate_expression_value(root, leftop);
1845 22386 : rightop = estimate_expression_value(root, rightop);
1846 :
1847 : /* get nominal (after relabeling) element type of rightop */
1848 22386 : nominal_element_type = get_base_element_type(exprType(rightop));
1849 22386 : if (!OidIsValid(nominal_element_type))
1850 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1851 : /* get nominal collation, too, for generating constants */
1852 22386 : nominal_element_collation = exprCollation(rightop);
1853 :
1854 : /* look through any binary-compatible relabeling of rightop */
1855 22386 : rightop = strip_array_coercion(rightop);
1856 :
1857 : /*
1858 : * Detect whether the operator is the default equality or inequality
1859 : * operator of the array element type.
1860 : */
1861 22386 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1862 22386 : if (OidIsValid(typentry->eq_opr))
1863 : {
1864 22384 : if (operator == typentry->eq_opr)
1865 16656 : isEquality = true;
1866 5728 : else if (get_negator(operator) == typentry->eq_opr)
1867 5176 : isInequality = true;
1868 : }
1869 :
1870 : /*
1871 : * If it is equality or inequality, we might be able to estimate this as a
1872 : * form of array containment; for instance "const = ANY(column)" can be
1873 : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1874 : * that, and returns the selectivity estimate if successful, or -1 if not.
1875 : */
1876 22386 : if ((isEquality || isInequality) && !is_join_clause)
1877 : {
1878 21832 : s1 = scalararraysel_containment(root, leftop, rightop,
1879 : nominal_element_type,
1880 : isEquality, useOr, varRelid);
1881 21832 : if (s1 >= 0.0)
1882 118 : return s1;
1883 : }
1884 :
1885 : /*
1886 : * Look up the underlying operator's selectivity estimator. Punt if it
1887 : * hasn't got one.
1888 : */
1889 22268 : if (is_join_clause)
1890 0 : oprsel = get_oprjoin(operator);
1891 : else
1892 22268 : oprsel = get_oprrest(operator);
1893 22268 : if (!oprsel)
1894 2 : return (Selectivity) 0.5;
1895 22266 : fmgr_info(oprsel, &oprselproc);
1896 :
1897 : /*
1898 : * In the array-containment check above, we must only believe that an
1899 : * operator is equality or inequality if it is the default btree equality
1900 : * operator (or its negator) for the element type, since those are the
1901 : * operators that array containment will use. But in what follows, we can
1902 : * be a little laxer, and also believe that any operators using eqsel() or
1903 : * neqsel() as selectivity estimator act like equality or inequality.
1904 : */
1905 22266 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1906 16720 : isEquality = true;
1907 5546 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1908 5066 : isInequality = true;
1909 :
1910 : /*
1911 : * We consider three cases:
1912 : *
1913 : * 1. rightop is an Array constant: deconstruct the array, apply the
1914 : * operator's selectivity function for each array element, and merge the
1915 : * results in the same way that clausesel.c does for AND/OR combinations.
1916 : *
1917 : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
1918 : * function for each element of the ARRAY[] construct, and merge.
1919 : *
1920 : * 3. otherwise, make a guess ...
1921 : */
1922 22266 : if (rightop && IsA(rightop, Const))
1923 18188 : {
1924 18206 : Datum arraydatum = ((Const *) rightop)->constvalue;
1925 18206 : bool arrayisnull = ((Const *) rightop)->constisnull;
1926 : ArrayType *arrayval;
1927 : int16 elmlen;
1928 : bool elmbyval;
1929 : char elmalign;
1930 : int num_elems;
1931 : Datum *elem_values;
1932 : bool *elem_nulls;
1933 : int i;
1934 :
1935 18206 : if (arrayisnull) /* qual can't succeed if null array */
1936 18 : return (Selectivity) 0.0;
1937 18188 : arrayval = DatumGetArrayTypeP(arraydatum);
1938 18188 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
1939 : &elmlen, &elmbyval, &elmalign);
1940 18188 : deconstruct_array(arrayval,
1941 : ARR_ELEMTYPE(arrayval),
1942 : elmlen, elmbyval, elmalign,
1943 : &elem_values, &elem_nulls, &num_elems);
1944 :
1945 : /*
1946 : * For generic operators, we assume the probability of success is
1947 : * independent for each array element. But for "= ANY" or "<> ALL",
1948 : * if the array elements are distinct (which'd typically be the case)
1949 : * then the probabilities are disjoint, and we should just sum them.
1950 : *
1951 : * If we were being really tense we would try to confirm that the
1952 : * elements are all distinct, but that would be expensive and it
1953 : * doesn't seem to be worth the cycles; it would amount to penalizing
1954 : * well-written queries in favor of poorly-written ones. However, we
1955 : * do protect ourselves a little bit by checking whether the
1956 : * disjointness assumption leads to an impossible (out of range)
1957 : * probability; if so, we fall back to the normal calculation.
1958 : */
1959 18188 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
1960 :
1961 71288 : for (i = 0; i < num_elems; i++)
1962 : {
1963 : List *args;
1964 : Selectivity s2;
1965 :
1966 53100 : args = list_make2(leftop,
1967 : makeConst(nominal_element_type,
1968 : -1,
1969 : nominal_element_collation,
1970 : elmlen,
1971 : elem_values[i],
1972 : elem_nulls[i],
1973 : elmbyval));
1974 53100 : if (is_join_clause)
1975 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1976 : clause->inputcollid,
1977 : PointerGetDatum(root),
1978 : ObjectIdGetDatum(operator),
1979 : PointerGetDatum(args),
1980 : Int16GetDatum(jointype),
1981 : PointerGetDatum(sjinfo)));
1982 : else
1983 53100 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1984 : clause->inputcollid,
1985 : PointerGetDatum(root),
1986 : ObjectIdGetDatum(operator),
1987 : PointerGetDatum(args),
1988 : Int32GetDatum(varRelid)));
1989 :
1990 53100 : if (useOr)
1991 : {
1992 40010 : s1 = s1 + s2 - s1 * s2;
1993 40010 : if (isEquality)
1994 38966 : s1disjoint += s2;
1995 : }
1996 : else
1997 : {
1998 13090 : s1 = s1 * s2;
1999 13090 : if (isInequality)
2000 12778 : s1disjoint += s2 - 1.0;
2001 : }
2002 : }
2003 :
2004 : /* accept disjoint-probability estimate if in range */
2005 18188 : if ((useOr ? isEquality : isInequality) &&
2006 17534 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2007 17504 : s1 = s1disjoint;
2008 : }
2009 4060 : else if (rightop && IsA(rightop, ArrayExpr) &&
2010 292 : !((ArrayExpr *) rightop)->multidims)
2011 292 : {
2012 292 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2013 : int16 elmlen;
2014 : bool elmbyval;
2015 : ListCell *l;
2016 :
2017 292 : get_typlenbyval(arrayexpr->element_typeid,
2018 : &elmlen, &elmbyval);
2019 :
2020 : /*
2021 : * We use the assumption of disjoint probabilities here too, although
2022 : * the odds of equal array elements are rather higher if the elements
2023 : * are not all constants (which they won't be, else constant folding
2024 : * would have reduced the ArrayExpr to a Const). In this path it's
2025 : * critical to have the sanity check on the s1disjoint estimate.
2026 : */
2027 292 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2028 :
2029 1038 : foreach(l, arrayexpr->elements)
2030 : {
2031 746 : Node *elem = (Node *) lfirst(l);
2032 : List *args;
2033 : Selectivity s2;
2034 :
2035 : /*
2036 : * Theoretically, if elem isn't of nominal_element_type we should
2037 : * insert a RelabelType, but it seems unlikely that any operator
2038 : * estimation function would really care ...
2039 : */
2040 746 : args = list_make2(leftop, elem);
2041 746 : if (is_join_clause)
2042 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2043 : clause->inputcollid,
2044 : PointerGetDatum(root),
2045 : ObjectIdGetDatum(operator),
2046 : PointerGetDatum(args),
2047 : Int16GetDatum(jointype),
2048 : PointerGetDatum(sjinfo)));
2049 : else
2050 746 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2051 : clause->inputcollid,
2052 : PointerGetDatum(root),
2053 : ObjectIdGetDatum(operator),
2054 : PointerGetDatum(args),
2055 : Int32GetDatum(varRelid)));
2056 :
2057 746 : if (useOr)
2058 : {
2059 746 : s1 = s1 + s2 - s1 * s2;
2060 746 : if (isEquality)
2061 746 : s1disjoint += s2;
2062 : }
2063 : else
2064 : {
2065 0 : s1 = s1 * s2;
2066 0 : if (isInequality)
2067 0 : s1disjoint += s2 - 1.0;
2068 : }
2069 : }
2070 :
2071 : /* accept disjoint-probability estimate if in range */
2072 292 : if ((useOr ? isEquality : isInequality) &&
2073 292 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2074 292 : s1 = s1disjoint;
2075 : }
2076 : else
2077 : {
2078 : CaseTestExpr *dummyexpr;
2079 : List *args;
2080 : Selectivity s2;
2081 : int i;
2082 :
2083 : /*
2084 : * We need a dummy rightop to pass to the operator selectivity
2085 : * routine. It can be pretty much anything that doesn't look like a
2086 : * constant; CaseTestExpr is a convenient choice.
2087 : */
2088 3768 : dummyexpr = makeNode(CaseTestExpr);
2089 3768 : dummyexpr->typeId = nominal_element_type;
2090 3768 : dummyexpr->typeMod = -1;
2091 3768 : dummyexpr->collation = clause->inputcollid;
2092 3768 : args = list_make2(leftop, dummyexpr);
2093 3768 : if (is_join_clause)
2094 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2095 : clause->inputcollid,
2096 : PointerGetDatum(root),
2097 : ObjectIdGetDatum(operator),
2098 : PointerGetDatum(args),
2099 : Int16GetDatum(jointype),
2100 : PointerGetDatum(sjinfo)));
2101 : else
2102 3768 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2103 : clause->inputcollid,
2104 : PointerGetDatum(root),
2105 : ObjectIdGetDatum(operator),
2106 : PointerGetDatum(args),
2107 : Int32GetDatum(varRelid)));
2108 3768 : s1 = useOr ? 0.0 : 1.0;
2109 :
2110 : /*
2111 : * Arbitrarily assume 10 elements in the eventual array value (see
2112 : * also estimate_array_length). We don't risk an assumption of
2113 : * disjoint probabilities here.
2114 : */
2115 41448 : for (i = 0; i < 10; i++)
2116 : {
2117 37680 : if (useOr)
2118 37680 : s1 = s1 + s2 - s1 * s2;
2119 : else
2120 0 : s1 = s1 * s2;
2121 : }
2122 : }
2123 :
2124 : /* result should be in range, but make sure... */
2125 22248 : CLAMP_PROBABILITY(s1);
2126 :
2127 22248 : return s1;
2128 : }
2129 :
2130 : /*
2131 : * Estimate number of elements in the array yielded by an expression.
2132 : *
2133 : * Note: the result is integral, but we use "double" to avoid overflow
2134 : * concerns. Most callers will use it in double-type expressions anyway.
2135 : *
2136 : * Note: in some code paths root can be passed as NULL, resulting in
2137 : * slightly worse estimates.
2138 : */
2139 : double
2140 103814 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2141 : {
2142 : /* look through any binary-compatible relabeling of arrayexpr */
2143 103814 : arrayexpr = strip_array_coercion(arrayexpr);
2144 :
2145 103814 : if (arrayexpr && IsA(arrayexpr, Const))
2146 : {
2147 46974 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2148 46974 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2149 : ArrayType *arrayval;
2150 :
2151 46974 : if (arrayisnull)
2152 42 : return 0;
2153 46932 : arrayval = DatumGetArrayTypeP(arraydatum);
2154 46932 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2155 : }
2156 56840 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2157 494 : !((ArrayExpr *) arrayexpr)->multidims)
2158 : {
2159 494 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2160 : }
2161 56346 : else if (arrayexpr && root)
2162 : {
2163 : /* See if we can find any statistics about it */
2164 : VariableStatData vardata;
2165 : AttStatsSlot sslot;
2166 56346 : double nelem = 0;
2167 :
2168 56346 : examine_variable(root, arrayexpr, 0, &vardata);
2169 56346 : if (HeapTupleIsValid(vardata.statsTuple))
2170 : {
2171 : /*
2172 : * Found stats, so use the average element count, which is stored
2173 : * in the last stanumbers element of the DECHIST statistics.
2174 : * Actually that is the average count of *distinct* elements;
2175 : * perhaps we should scale it up somewhat?
2176 : */
2177 8152 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2178 : STATISTIC_KIND_DECHIST, InvalidOid,
2179 : ATTSTATSSLOT_NUMBERS))
2180 : {
2181 8038 : if (sslot.nnumbers > 0)
2182 8038 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2183 8038 : free_attstatsslot(&sslot);
2184 : }
2185 : }
2186 56346 : ReleaseVariableStats(vardata);
2187 :
2188 56346 : if (nelem > 0)
2189 8038 : return nelem;
2190 : }
2191 :
2192 : /* Else use a default guess --- this should match scalararraysel */
2193 48308 : return 10;
2194 : }
2195 :
2196 : /*
2197 : * rowcomparesel - Selectivity of RowCompareExpr Node.
2198 : *
2199 : * We estimate RowCompare selectivity by considering just the first (high
2200 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2201 : * this estimate could be refined by considering additional columns, it
2202 : * seems unlikely that we could do a lot better without multi-column
2203 : * statistics.
2204 : */
2205 : Selectivity
2206 216 : rowcomparesel(PlannerInfo *root,
2207 : RowCompareExpr *clause,
2208 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2209 : {
2210 : Selectivity s1;
2211 216 : Oid opno = linitial_oid(clause->opnos);
2212 216 : Oid inputcollid = linitial_oid(clause->inputcollids);
2213 : List *opargs;
2214 : bool is_join_clause;
2215 :
2216 : /* Build equivalent arg list for single operator */
2217 216 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2218 :
2219 : /*
2220 : * Decide if it's a join clause. This should match clausesel.c's
2221 : * treat_as_join_clause(), except that we intentionally consider only the
2222 : * leading columns and not the rest of the clause.
2223 : */
2224 216 : if (varRelid != 0)
2225 : {
2226 : /*
2227 : * Caller is forcing restriction mode (eg, because we are examining an
2228 : * inner indexscan qual).
2229 : */
2230 54 : is_join_clause = false;
2231 : }
2232 162 : else if (sjinfo == NULL)
2233 : {
2234 : /*
2235 : * It must be a restriction clause, since it's being evaluated at a
2236 : * scan node.
2237 : */
2238 150 : is_join_clause = false;
2239 : }
2240 : else
2241 : {
2242 : /*
2243 : * Otherwise, it's a join if there's more than one base relation used.
2244 : */
2245 12 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2246 : }
2247 :
2248 216 : if (is_join_clause)
2249 : {
2250 : /* Estimate selectivity for a join clause. */
2251 12 : s1 = join_selectivity(root, opno,
2252 : opargs,
2253 : inputcollid,
2254 : jointype,
2255 : sjinfo);
2256 : }
2257 : else
2258 : {
2259 : /* Estimate selectivity for a restriction clause. */
2260 204 : s1 = restriction_selectivity(root, opno,
2261 : opargs,
2262 : inputcollid,
2263 : varRelid);
2264 : }
2265 :
2266 216 : return s1;
2267 : }
2268 :
2269 : /*
2270 : * eqjoinsel - Join selectivity of "="
2271 : */
2272 : Datum
2273 218840 : eqjoinsel(PG_FUNCTION_ARGS)
2274 : {
2275 218840 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2276 218840 : Oid operator = PG_GETARG_OID(1);
2277 218840 : List *args = (List *) PG_GETARG_POINTER(2);
2278 :
2279 : #ifdef NOT_USED
2280 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2281 : #endif
2282 218840 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2283 218840 : Oid collation = PG_GET_COLLATION();
2284 : double selec;
2285 : double selec_inner;
2286 : VariableStatData vardata1;
2287 : VariableStatData vardata2;
2288 : double nd1;
2289 : double nd2;
2290 : bool isdefault1;
2291 : bool isdefault2;
2292 : Oid opfuncoid;
2293 : AttStatsSlot sslot1;
2294 : AttStatsSlot sslot2;
2295 218840 : Form_pg_statistic stats1 = NULL;
2296 218840 : Form_pg_statistic stats2 = NULL;
2297 218840 : bool have_mcvs1 = false;
2298 218840 : bool have_mcvs2 = false;
2299 : bool get_mcv_stats;
2300 : bool join_is_reversed;
2301 : RelOptInfo *inner_rel;
2302 :
2303 218840 : get_join_variables(root, args, sjinfo,
2304 : &vardata1, &vardata2, &join_is_reversed);
2305 :
2306 218840 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2307 218840 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2308 :
2309 218840 : opfuncoid = get_opcode(operator);
2310 :
2311 218840 : memset(&sslot1, 0, sizeof(sslot1));
2312 218840 : memset(&sslot2, 0, sizeof(sslot2));
2313 :
2314 : /*
2315 : * There is no use in fetching one side's MCVs if we lack MCVs for the
2316 : * other side, so do a quick check to verify that both stats exist.
2317 : */
2318 608720 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2319 305074 : HeapTupleIsValid(vardata2.statsTuple) &&
2320 134034 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2321 : STATISTIC_KIND_MCV, InvalidOid,
2322 389880 : 0) &&
2323 53948 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2324 : STATISTIC_KIND_MCV, InvalidOid,
2325 : 0));
2326 :
2327 218840 : if (HeapTupleIsValid(vardata1.statsTuple))
2328 : {
2329 : /* note we allow use of nullfrac regardless of security check */
2330 171040 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2331 184444 : if (get_mcv_stats &&
2332 13404 : statistic_proc_security_check(&vardata1, opfuncoid))
2333 13404 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2334 : STATISTIC_KIND_MCV, InvalidOid,
2335 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2336 : }
2337 :
2338 218840 : if (HeapTupleIsValid(vardata2.statsTuple))
2339 : {
2340 : /* note we allow use of nullfrac regardless of security check */
2341 150890 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2342 164294 : if (get_mcv_stats &&
2343 13404 : statistic_proc_security_check(&vardata2, opfuncoid))
2344 13404 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2345 : STATISTIC_KIND_MCV, InvalidOid,
2346 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2347 : }
2348 :
2349 : /* We need to compute the inner-join selectivity in all cases */
2350 218840 : selec_inner = eqjoinsel_inner(opfuncoid, collation,
2351 : &vardata1, &vardata2,
2352 : nd1, nd2,
2353 : isdefault1, isdefault2,
2354 : &sslot1, &sslot2,
2355 : stats1, stats2,
2356 : have_mcvs1, have_mcvs2);
2357 :
2358 218840 : switch (sjinfo->jointype)
2359 : {
2360 211118 : case JOIN_INNER:
2361 : case JOIN_LEFT:
2362 : case JOIN_FULL:
2363 211118 : selec = selec_inner;
2364 211118 : break;
2365 7722 : case JOIN_SEMI:
2366 : case JOIN_ANTI:
2367 :
2368 : /*
2369 : * Look up the join's inner relation. min_righthand is sufficient
2370 : * information because neither SEMI nor ANTI joins permit any
2371 : * reassociation into or out of their RHS, so the righthand will
2372 : * always be exactly that set of rels.
2373 : */
2374 7722 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2375 :
2376 7722 : if (!join_is_reversed)
2377 3788 : selec = eqjoinsel_semi(opfuncoid, collation,
2378 : &vardata1, &vardata2,
2379 : nd1, nd2,
2380 : isdefault1, isdefault2,
2381 : &sslot1, &sslot2,
2382 : stats1, stats2,
2383 : have_mcvs1, have_mcvs2,
2384 : inner_rel);
2385 : else
2386 : {
2387 3934 : Oid commop = get_commutator(operator);
2388 3934 : Oid commopfuncoid = OidIsValid(commop) ? get_opcode(commop) : InvalidOid;
2389 :
2390 3934 : selec = eqjoinsel_semi(commopfuncoid, collation,
2391 : &vardata2, &vardata1,
2392 : nd2, nd1,
2393 : isdefault2, isdefault1,
2394 : &sslot2, &sslot1,
2395 : stats2, stats1,
2396 : have_mcvs2, have_mcvs1,
2397 : inner_rel);
2398 : }
2399 :
2400 : /*
2401 : * We should never estimate the output of a semijoin to be more
2402 : * rows than we estimate for an inner join with the same input
2403 : * rels and join condition; it's obviously impossible for that to
2404 : * happen. The former estimate is N1 * Ssemi while the latter is
2405 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2406 : * this is worthwhile because of the shakier estimation rules we
2407 : * use in eqjoinsel_semi, particularly in cases where it has to
2408 : * punt entirely.
2409 : */
2410 7722 : selec = Min(selec, inner_rel->rows * selec_inner);
2411 7722 : break;
2412 0 : default:
2413 : /* other values not expected here */
2414 0 : elog(ERROR, "unrecognized join type: %d",
2415 : (int) sjinfo->jointype);
2416 : selec = 0; /* keep compiler quiet */
2417 : break;
2418 : }
2419 :
2420 218840 : free_attstatsslot(&sslot1);
2421 218840 : free_attstatsslot(&sslot2);
2422 :
2423 218840 : ReleaseVariableStats(vardata1);
2424 218840 : ReleaseVariableStats(vardata2);
2425 :
2426 218840 : CLAMP_PROBABILITY(selec);
2427 :
2428 218840 : PG_RETURN_FLOAT8((float8) selec);
2429 : }
2430 :
2431 : /*
2432 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2433 : *
2434 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2435 : * that it's worth trying to distinguish them here.
2436 : */
2437 : static double
2438 218840 : eqjoinsel_inner(Oid opfuncoid, Oid collation,
2439 : VariableStatData *vardata1, VariableStatData *vardata2,
2440 : double nd1, double nd2,
2441 : bool isdefault1, bool isdefault2,
2442 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2443 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2444 : bool have_mcvs1, bool have_mcvs2)
2445 : {
2446 : double selec;
2447 :
2448 218840 : if (have_mcvs1 && have_mcvs2)
2449 13404 : {
2450 : /*
2451 : * We have most-common-value lists for both relations. Run through
2452 : * the lists to see which MCVs actually join to each other with the
2453 : * given operator. This allows us to determine the exact join
2454 : * selectivity for the portion of the relations represented by the MCV
2455 : * lists. We still have to estimate for the remaining population, but
2456 : * in a skewed distribution this gives us a big leg up in accuracy.
2457 : * For motivation see the analysis in Y. Ioannidis and S.
2458 : * Christodoulakis, "On the propagation of errors in the size of join
2459 : * results", Technical Report 1018, Computer Science Dept., University
2460 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2461 : */
2462 13404 : LOCAL_FCINFO(fcinfo, 2);
2463 : FmgrInfo eqproc;
2464 : bool *hasmatch1;
2465 : bool *hasmatch2;
2466 13404 : double nullfrac1 = stats1->stanullfrac;
2467 13404 : double nullfrac2 = stats2->stanullfrac;
2468 : double matchprodfreq,
2469 : matchfreq1,
2470 : matchfreq2,
2471 : unmatchfreq1,
2472 : unmatchfreq2,
2473 : otherfreq1,
2474 : otherfreq2,
2475 : totalsel1,
2476 : totalsel2;
2477 : int i,
2478 : nmatches;
2479 :
2480 13404 : fmgr_info(opfuncoid, &eqproc);
2481 :
2482 : /*
2483 : * Save a few cycles by setting up the fcinfo struct just once. Using
2484 : * FunctionCallInvoke directly also avoids failure if the eqproc
2485 : * returns NULL, though really equality functions should never do
2486 : * that.
2487 : */
2488 13404 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2489 : NULL, NULL);
2490 13404 : fcinfo->args[0].isnull = false;
2491 13404 : fcinfo->args[1].isnull = false;
2492 :
2493 13404 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2494 13404 : hasmatch2 = (bool *) palloc0(sslot2->nvalues * sizeof(bool));
2495 :
2496 : /*
2497 : * Note we assume that each MCV will match at most one member of the
2498 : * other MCV list. If the operator isn't really equality, there could
2499 : * be multiple matches --- but we don't look for them, both for speed
2500 : * and because the math wouldn't add up...
2501 : */
2502 13404 : matchprodfreq = 0.0;
2503 13404 : nmatches = 0;
2504 568956 : for (i = 0; i < sslot1->nvalues; i++)
2505 : {
2506 : int j;
2507 :
2508 555552 : fcinfo->args[0].value = sslot1->values[i];
2509 :
2510 21877538 : for (j = 0; j < sslot2->nvalues; j++)
2511 : {
2512 : Datum fresult;
2513 :
2514 21514526 : if (hasmatch2[j])
2515 5779534 : continue;
2516 15734992 : fcinfo->args[1].value = sslot2->values[j];
2517 15734992 : fcinfo->isnull = false;
2518 15734992 : fresult = FunctionCallInvoke(fcinfo);
2519 15734992 : if (!fcinfo->isnull && DatumGetBool(fresult))
2520 : {
2521 192540 : hasmatch1[i] = hasmatch2[j] = true;
2522 192540 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
2523 192540 : nmatches++;
2524 192540 : break;
2525 : }
2526 : }
2527 : }
2528 13404 : CLAMP_PROBABILITY(matchprodfreq);
2529 : /* Sum up frequencies of matched and unmatched MCVs */
2530 13404 : matchfreq1 = unmatchfreq1 = 0.0;
2531 568956 : for (i = 0; i < sslot1->nvalues; i++)
2532 : {
2533 555552 : if (hasmatch1[i])
2534 192540 : matchfreq1 += sslot1->numbers[i];
2535 : else
2536 363012 : unmatchfreq1 += sslot1->numbers[i];
2537 : }
2538 13404 : CLAMP_PROBABILITY(matchfreq1);
2539 13404 : CLAMP_PROBABILITY(unmatchfreq1);
2540 13404 : matchfreq2 = unmatchfreq2 = 0.0;
2541 390392 : for (i = 0; i < sslot2->nvalues; i++)
2542 : {
2543 376988 : if (hasmatch2[i])
2544 192540 : matchfreq2 += sslot2->numbers[i];
2545 : else
2546 184448 : unmatchfreq2 += sslot2->numbers[i];
2547 : }
2548 13404 : CLAMP_PROBABILITY(matchfreq2);
2549 13404 : CLAMP_PROBABILITY(unmatchfreq2);
2550 13404 : pfree(hasmatch1);
2551 13404 : pfree(hasmatch2);
2552 :
2553 : /*
2554 : * Compute total frequency of non-null values that are not in the MCV
2555 : * lists.
2556 : */
2557 13404 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2558 13404 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2559 13404 : CLAMP_PROBABILITY(otherfreq1);
2560 13404 : CLAMP_PROBABILITY(otherfreq2);
2561 :
2562 : /*
2563 : * We can estimate the total selectivity from the point of view of
2564 : * relation 1 as: the known selectivity for matched MCVs, plus
2565 : * unmatched MCVs that are assumed to match against random members of
2566 : * relation 2's non-MCV population, plus non-MCV values that are
2567 : * assumed to match against random members of relation 2's unmatched
2568 : * MCVs plus non-MCV values.
2569 : */
2570 13404 : totalsel1 = matchprodfreq;
2571 13404 : if (nd2 > sslot2->nvalues)
2572 5646 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2573 13404 : if (nd2 > nmatches)
2574 10396 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2575 10396 : (nd2 - nmatches);
2576 : /* Same estimate from the point of view of relation 2. */
2577 13404 : totalsel2 = matchprodfreq;
2578 13404 : if (nd1 > sslot1->nvalues)
2579 6776 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2580 13404 : if (nd1 > nmatches)
2581 9164 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2582 9164 : (nd1 - nmatches);
2583 :
2584 : /*
2585 : * Use the smaller of the two estimates. This can be justified in
2586 : * essentially the same terms as given below for the no-stats case: to
2587 : * a first approximation, we are estimating from the point of view of
2588 : * the relation with smaller nd.
2589 : */
2590 13404 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2591 : }
2592 : else
2593 : {
2594 : /*
2595 : * We do not have MCV lists for both sides. Estimate the join
2596 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2597 : * is plausible if we assume that the join operator is strict and the
2598 : * non-null values are about equally distributed: a given non-null
2599 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2600 : * of rel2, so total join rows are at most
2601 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2602 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2603 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2604 : * with MIN() is an upper bound. Using the MIN() means we estimate
2605 : * from the point of view of the relation with smaller nd (since the
2606 : * larger nd is determining the MIN). It is reasonable to assume that
2607 : * most tuples in this rel will have join partners, so the bound is
2608 : * probably reasonably tight and should be taken as-is.
2609 : *
2610 : * XXX Can we be smarter if we have an MCV list for just one side? It
2611 : * seems that if we assume equal distribution for the other side, we
2612 : * end up with the same answer anyway.
2613 : */
2614 205436 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2615 205436 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2616 :
2617 205436 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2618 205436 : if (nd1 > nd2)
2619 102452 : selec /= nd1;
2620 : else
2621 102984 : selec /= nd2;
2622 : }
2623 :
2624 218840 : return selec;
2625 : }
2626 :
2627 : /*
2628 : * eqjoinsel_semi --- eqjoinsel for semi join
2629 : *
2630 : * (Also used for anti join, which we are supposed to estimate the same way.)
2631 : * Caller has ensured that vardata1 is the LHS variable.
2632 : * Unlike eqjoinsel_inner, we have to cope with opfuncoid being InvalidOid.
2633 : */
2634 : static double
2635 7722 : eqjoinsel_semi(Oid opfuncoid, Oid collation,
2636 : VariableStatData *vardata1, VariableStatData *vardata2,
2637 : double nd1, double nd2,
2638 : bool isdefault1, bool isdefault2,
2639 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2640 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2641 : bool have_mcvs1, bool have_mcvs2,
2642 : RelOptInfo *inner_rel)
2643 : {
2644 : double selec;
2645 :
2646 : /*
2647 : * We clamp nd2 to be not more than what we estimate the inner relation's
2648 : * size to be. This is intuitively somewhat reasonable since obviously
2649 : * there can't be more than that many distinct values coming from the
2650 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2651 : * likewise) is that this is the only pathway by which restriction clauses
2652 : * applied to the inner rel will affect the join result size estimate,
2653 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2654 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2655 : * the selectivity of outer-rel restrictions.
2656 : *
2657 : * We can apply this clamping both with respect to the base relation from
2658 : * which the join variable comes (if there is just one), and to the
2659 : * immediate inner input relation of the current join.
2660 : *
2661 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2662 : * great, maybe, but it didn't come out of nowhere either. This is most
2663 : * helpful when the inner relation is empty and consequently has no stats.
2664 : */
2665 7722 : if (vardata2->rel)
2666 : {
2667 7716 : if (nd2 >= vardata2->rel->rows)
2668 : {
2669 5870 : nd2 = vardata2->rel->rows;
2670 5870 : isdefault2 = false;
2671 : }
2672 : }
2673 7722 : if (nd2 >= inner_rel->rows)
2674 : {
2675 5812 : nd2 = inner_rel->rows;
2676 5812 : isdefault2 = false;
2677 : }
2678 :
2679 7722 : if (have_mcvs1 && have_mcvs2 && OidIsValid(opfuncoid))
2680 516 : {
2681 : /*
2682 : * We have most-common-value lists for both relations. Run through
2683 : * the lists to see which MCVs actually join to each other with the
2684 : * given operator. This allows us to determine the exact join
2685 : * selectivity for the portion of the relations represented by the MCV
2686 : * lists. We still have to estimate for the remaining population, but
2687 : * in a skewed distribution this gives us a big leg up in accuracy.
2688 : */
2689 516 : LOCAL_FCINFO(fcinfo, 2);
2690 : FmgrInfo eqproc;
2691 : bool *hasmatch1;
2692 : bool *hasmatch2;
2693 516 : double nullfrac1 = stats1->stanullfrac;
2694 : double matchfreq1,
2695 : uncertainfrac,
2696 : uncertain;
2697 : int i,
2698 : nmatches,
2699 : clamped_nvalues2;
2700 :
2701 : /*
2702 : * The clamping above could have resulted in nd2 being less than
2703 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2704 : * most common values in the relation will appear in the join input,
2705 : * and so compare to only the first nd2 members of the MCV list. Of
2706 : * course this is frequently wrong, but it's the best bet we can make.
2707 : */
2708 516 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2709 :
2710 516 : fmgr_info(opfuncoid, &eqproc);
2711 :
2712 : /*
2713 : * Save a few cycles by setting up the fcinfo struct just once. Using
2714 : * FunctionCallInvoke directly also avoids failure if the eqproc
2715 : * returns NULL, though really equality functions should never do
2716 : * that.
2717 : */
2718 516 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2719 : NULL, NULL);
2720 516 : fcinfo->args[0].isnull = false;
2721 516 : fcinfo->args[1].isnull = false;
2722 :
2723 516 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2724 516 : hasmatch2 = (bool *) palloc0(clamped_nvalues2 * sizeof(bool));
2725 :
2726 : /*
2727 : * Note we assume that each MCV will match at most one member of the
2728 : * other MCV list. If the operator isn't really equality, there could
2729 : * be multiple matches --- but we don't look for them, both for speed
2730 : * and because the math wouldn't add up...
2731 : */
2732 516 : nmatches = 0;
2733 12022 : for (i = 0; i < sslot1->nvalues; i++)
2734 : {
2735 : int j;
2736 :
2737 11506 : fcinfo->args[0].value = sslot1->values[i];
2738 :
2739 451462 : for (j = 0; j < clamped_nvalues2; j++)
2740 : {
2741 : Datum fresult;
2742 :
2743 450214 : if (hasmatch2[j])
2744 379588 : continue;
2745 70626 : fcinfo->args[1].value = sslot2->values[j];
2746 70626 : fcinfo->isnull = false;
2747 70626 : fresult = FunctionCallInvoke(fcinfo);
2748 70626 : if (!fcinfo->isnull && DatumGetBool(fresult))
2749 : {
2750 10258 : hasmatch1[i] = hasmatch2[j] = true;
2751 10258 : nmatches++;
2752 10258 : break;
2753 : }
2754 : }
2755 : }
2756 : /* Sum up frequencies of matched MCVs */
2757 516 : matchfreq1 = 0.0;
2758 12022 : for (i = 0; i < sslot1->nvalues; i++)
2759 : {
2760 11506 : if (hasmatch1[i])
2761 10258 : matchfreq1 += sslot1->numbers[i];
2762 : }
2763 516 : CLAMP_PROBABILITY(matchfreq1);
2764 516 : pfree(hasmatch1);
2765 516 : pfree(hasmatch2);
2766 :
2767 : /*
2768 : * Now we need to estimate the fraction of relation 1 that has at
2769 : * least one join partner. We know for certain that the matched MCVs
2770 : * do, so that gives us a lower bound, but we're really in the dark
2771 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2772 : * assume all non-null rel1 rows have join partners, else assume for
2773 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2774 : * can discount the known-matched MCVs from the distinct-values counts
2775 : * before doing the division.
2776 : *
2777 : * Crude as the above is, it's completely useless if we don't have
2778 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2779 : * nd2 is default, punt and assume half of the uncertain rows have
2780 : * join partners.
2781 : */
2782 516 : if (!isdefault1 && !isdefault2)
2783 : {
2784 516 : nd1 -= nmatches;
2785 516 : nd2 -= nmatches;
2786 516 : if (nd1 <= nd2 || nd2 < 0)
2787 486 : uncertainfrac = 1.0;
2788 : else
2789 30 : uncertainfrac = nd2 / nd1;
2790 : }
2791 : else
2792 0 : uncertainfrac = 0.5;
2793 516 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2794 516 : CLAMP_PROBABILITY(uncertain);
2795 516 : selec = matchfreq1 + uncertainfrac * uncertain;
2796 : }
2797 : else
2798 : {
2799 : /*
2800 : * Without MCV lists for both sides, we can only use the heuristic
2801 : * about nd1 vs nd2.
2802 : */
2803 7206 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2804 :
2805 7206 : if (!isdefault1 && !isdefault2)
2806 : {
2807 5050 : if (nd1 <= nd2 || nd2 < 0)
2808 2560 : selec = 1.0 - nullfrac1;
2809 : else
2810 2490 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2811 : }
2812 : else
2813 2156 : selec = 0.5 * (1.0 - nullfrac1);
2814 : }
2815 :
2816 7722 : return selec;
2817 : }
2818 :
2819 : /*
2820 : * neqjoinsel - Join selectivity of "!="
2821 : */
2822 : Datum
2823 3630 : neqjoinsel(PG_FUNCTION_ARGS)
2824 : {
2825 3630 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2826 3630 : Oid operator = PG_GETARG_OID(1);
2827 3630 : List *args = (List *) PG_GETARG_POINTER(2);
2828 3630 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2829 3630 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2830 3630 : Oid collation = PG_GET_COLLATION();
2831 : float8 result;
2832 :
2833 3630 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
2834 1224 : {
2835 : /*
2836 : * For semi-joins, if there is more than one distinct value in the RHS
2837 : * relation then every non-null LHS row must find a row to join since
2838 : * it can only be equal to one of them. We'll assume that there is
2839 : * always more than one distinct RHS value for the sake of stability,
2840 : * though in theory we could have special cases for empty RHS
2841 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
2842 : * fraction of LHS that has the same value as the single RHS value).
2843 : *
2844 : * For anti-joins, if we use the same assumption that there is more
2845 : * than one distinct key in the RHS relation, then every non-null LHS
2846 : * row must be suppressed by the anti-join.
2847 : *
2848 : * So either way, the selectivity estimate should be 1 - nullfrac.
2849 : */
2850 : VariableStatData leftvar;
2851 : VariableStatData rightvar;
2852 : bool reversed;
2853 : HeapTuple statsTuple;
2854 : double nullfrac;
2855 :
2856 1224 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
2857 1224 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
2858 1224 : if (HeapTupleIsValid(statsTuple))
2859 1002 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
2860 : else
2861 222 : nullfrac = 0.0;
2862 1224 : ReleaseVariableStats(leftvar);
2863 1224 : ReleaseVariableStats(rightvar);
2864 :
2865 1224 : result = 1.0 - nullfrac;
2866 : }
2867 : else
2868 : {
2869 : /*
2870 : * We want 1 - eqjoinsel() where the equality operator is the one
2871 : * associated with this != operator, that is, its negator.
2872 : */
2873 2406 : Oid eqop = get_negator(operator);
2874 :
2875 2406 : if (eqop)
2876 : {
2877 : result =
2878 2406 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
2879 : collation,
2880 : PointerGetDatum(root),
2881 : ObjectIdGetDatum(eqop),
2882 : PointerGetDatum(args),
2883 : Int16GetDatum(jointype),
2884 : PointerGetDatum(sjinfo)));
2885 : }
2886 : else
2887 : {
2888 : /* Use default selectivity (should we raise an error instead?) */
2889 0 : result = DEFAULT_EQ_SEL;
2890 : }
2891 2406 : result = 1.0 - result;
2892 : }
2893 :
2894 3630 : PG_RETURN_FLOAT8(result);
2895 : }
2896 :
2897 : /*
2898 : * scalarltjoinsel - Join selectivity of "<" for scalars
2899 : */
2900 : Datum
2901 318 : scalarltjoinsel(PG_FUNCTION_ARGS)
2902 : {
2903 318 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2904 : }
2905 :
2906 : /*
2907 : * scalarlejoinsel - Join selectivity of "<=" for scalars
2908 : */
2909 : Datum
2910 276 : scalarlejoinsel(PG_FUNCTION_ARGS)
2911 : {
2912 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2913 : }
2914 :
2915 : /*
2916 : * scalargtjoinsel - Join selectivity of ">" for scalars
2917 : */
2918 : Datum
2919 240 : scalargtjoinsel(PG_FUNCTION_ARGS)
2920 : {
2921 240 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2922 : }
2923 :
2924 : /*
2925 : * scalargejoinsel - Join selectivity of ">=" for scalars
2926 : */
2927 : Datum
2928 184 : scalargejoinsel(PG_FUNCTION_ARGS)
2929 : {
2930 184 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2931 : }
2932 :
2933 :
2934 : /*
2935 : * mergejoinscansel - Scan selectivity of merge join.
2936 : *
2937 : * A merge join will stop as soon as it exhausts either input stream.
2938 : * Therefore, if we can estimate the ranges of both input variables,
2939 : * we can estimate how much of the input will actually be read. This
2940 : * can have a considerable impact on the cost when using indexscans.
2941 : *
2942 : * Also, we can estimate how much of each input has to be read before the
2943 : * first join pair is found, which will affect the join's startup time.
2944 : *
2945 : * clause should be a clause already known to be mergejoinable. opfamily,
2946 : * strategy, and nulls_first specify the sort ordering being used.
2947 : *
2948 : * The outputs are:
2949 : * *leftstart is set to the fraction of the left-hand variable expected
2950 : * to be scanned before the first join pair is found (0 to 1).
2951 : * *leftend is set to the fraction of the left-hand variable expected
2952 : * to be scanned before the join terminates (0 to 1).
2953 : * *rightstart, *rightend similarly for the right-hand variable.
2954 : */
2955 : void
2956 110938 : mergejoinscansel(PlannerInfo *root, Node *clause,
2957 : Oid opfamily, int strategy, bool nulls_first,
2958 : Selectivity *leftstart, Selectivity *leftend,
2959 : Selectivity *rightstart, Selectivity *rightend)
2960 : {
2961 : Node *left,
2962 : *right;
2963 : VariableStatData leftvar,
2964 : rightvar;
2965 : int op_strategy;
2966 : Oid op_lefttype;
2967 : Oid op_righttype;
2968 : Oid opno,
2969 : collation,
2970 : lsortop,
2971 : rsortop,
2972 : lstatop,
2973 : rstatop,
2974 : ltop,
2975 : leop,
2976 : revltop,
2977 : revleop;
2978 : bool isgt;
2979 : Datum leftmin,
2980 : leftmax,
2981 : rightmin,
2982 : rightmax;
2983 : double selec;
2984 :
2985 : /* Set default results if we can't figure anything out. */
2986 : /* XXX should default "start" fraction be a bit more than 0? */
2987 110938 : *leftstart = *rightstart = 0.0;
2988 110938 : *leftend = *rightend = 1.0;
2989 :
2990 : /* Deconstruct the merge clause */
2991 110938 : if (!is_opclause(clause))
2992 0 : return; /* shouldn't happen */
2993 110938 : opno = ((OpExpr *) clause)->opno;
2994 110938 : collation = ((OpExpr *) clause)->inputcollid;
2995 110938 : left = get_leftop((Expr *) clause);
2996 110938 : right = get_rightop((Expr *) clause);
2997 110938 : if (!right)
2998 0 : return; /* shouldn't happen */
2999 :
3000 : /* Look for stats for the inputs */
3001 110938 : examine_variable(root, left, 0, &leftvar);
3002 110938 : examine_variable(root, right, 0, &rightvar);
3003 :
3004 : /* Extract the operator's declared left/right datatypes */
3005 110938 : get_op_opfamily_properties(opno, opfamily, false,
3006 : &op_strategy,
3007 : &op_lefttype,
3008 : &op_righttype);
3009 : Assert(op_strategy == BTEqualStrategyNumber);
3010 :
3011 : /*
3012 : * Look up the various operators we need. If we don't find them all, it
3013 : * probably means the opfamily is broken, but we just fail silently.
3014 : *
3015 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3016 : * operator, regardless of which sort direction we are considering.
3017 : */
3018 110938 : switch (strategy)
3019 : {
3020 110884 : case BTLessStrategyNumber:
3021 110884 : isgt = false;
3022 110884 : if (op_lefttype == op_righttype)
3023 : {
3024 : /* easy case */
3025 109512 : ltop = get_opfamily_member(opfamily,
3026 : op_lefttype, op_righttype,
3027 : BTLessStrategyNumber);
3028 109512 : leop = get_opfamily_member(opfamily,
3029 : op_lefttype, op_righttype,
3030 : BTLessEqualStrategyNumber);
3031 109512 : lsortop = ltop;
3032 109512 : rsortop = ltop;
3033 109512 : lstatop = lsortop;
3034 109512 : rstatop = rsortop;
3035 109512 : revltop = ltop;
3036 109512 : revleop = leop;
3037 : }
3038 : else
3039 : {
3040 1372 : ltop = get_opfamily_member(opfamily,
3041 : op_lefttype, op_righttype,
3042 : BTLessStrategyNumber);
3043 1372 : leop = get_opfamily_member(opfamily,
3044 : op_lefttype, op_righttype,
3045 : BTLessEqualStrategyNumber);
3046 1372 : lsortop = get_opfamily_member(opfamily,
3047 : op_lefttype, op_lefttype,
3048 : BTLessStrategyNumber);
3049 1372 : rsortop = get_opfamily_member(opfamily,
3050 : op_righttype, op_righttype,
3051 : BTLessStrategyNumber);
3052 1372 : lstatop = lsortop;
3053 1372 : rstatop = rsortop;
3054 1372 : revltop = get_opfamily_member(opfamily,
3055 : op_righttype, op_lefttype,
3056 : BTLessStrategyNumber);
3057 1372 : revleop = get_opfamily_member(opfamily,
3058 : op_righttype, op_lefttype,
3059 : BTLessEqualStrategyNumber);
3060 : }
3061 110884 : break;
3062 54 : case BTGreaterStrategyNumber:
3063 : /* descending-order case */
3064 54 : isgt = true;
3065 54 : if (op_lefttype == op_righttype)
3066 : {
3067 : /* easy case */
3068 54 : ltop = get_opfamily_member(opfamily,
3069 : op_lefttype, op_righttype,
3070 : BTGreaterStrategyNumber);
3071 54 : leop = get_opfamily_member(opfamily,
3072 : op_lefttype, op_righttype,
3073 : BTGreaterEqualStrategyNumber);
3074 54 : lsortop = ltop;
3075 54 : rsortop = ltop;
3076 54 : lstatop = get_opfamily_member(opfamily,
3077 : op_lefttype, op_lefttype,
3078 : BTLessStrategyNumber);
3079 54 : rstatop = lstatop;
3080 54 : revltop = ltop;
3081 54 : revleop = leop;
3082 : }
3083 : else
3084 : {
3085 0 : ltop = get_opfamily_member(opfamily,
3086 : op_lefttype, op_righttype,
3087 : BTGreaterStrategyNumber);
3088 0 : leop = get_opfamily_member(opfamily,
3089 : op_lefttype, op_righttype,
3090 : BTGreaterEqualStrategyNumber);
3091 0 : lsortop = get_opfamily_member(opfamily,
3092 : op_lefttype, op_lefttype,
3093 : BTGreaterStrategyNumber);
3094 0 : rsortop = get_opfamily_member(opfamily,
3095 : op_righttype, op_righttype,
3096 : BTGreaterStrategyNumber);
3097 0 : lstatop = get_opfamily_member(opfamily,
3098 : op_lefttype, op_lefttype,
3099 : BTLessStrategyNumber);
3100 0 : rstatop = get_opfamily_member(opfamily,
3101 : op_righttype, op_righttype,
3102 : BTLessStrategyNumber);
3103 0 : revltop = get_opfamily_member(opfamily,
3104 : op_righttype, op_lefttype,
3105 : BTGreaterStrategyNumber);
3106 0 : revleop = get_opfamily_member(opfamily,
3107 : op_righttype, op_lefttype,
3108 : BTGreaterEqualStrategyNumber);
3109 : }
3110 54 : break;
3111 0 : default:
3112 0 : goto fail; /* shouldn't get here */
3113 : }
3114 :
3115 110938 : if (!OidIsValid(lsortop) ||
3116 110938 : !OidIsValid(rsortop) ||
3117 110938 : !OidIsValid(lstatop) ||
3118 110938 : !OidIsValid(rstatop) ||
3119 110926 : !OidIsValid(ltop) ||
3120 110926 : !OidIsValid(leop) ||
3121 110926 : !OidIsValid(revltop) ||
3122 : !OidIsValid(revleop))
3123 12 : goto fail; /* insufficient info in catalogs */
3124 :
3125 : /* Try to get ranges of both inputs */
3126 110926 : if (!isgt)
3127 : {
3128 110872 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3129 : &leftmin, &leftmax))
3130 25610 : goto fail; /* no range available from stats */
3131 85262 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3132 : &rightmin, &rightmax))
3133 20494 : goto fail; /* no range available from stats */
3134 : }
3135 : else
3136 : {
3137 : /* need to swap the max and min */
3138 54 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3139 : &leftmax, &leftmin))
3140 30 : goto fail; /* no range available from stats */
3141 24 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3142 : &rightmax, &rightmin))
3143 0 : goto fail; /* no range available from stats */
3144 : }
3145 :
3146 : /*
3147 : * Now, the fraction of the left variable that will be scanned is the
3148 : * fraction that's <= the right-side maximum value. But only believe
3149 : * non-default estimates, else stick with our 1.0.
3150 : */
3151 64792 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3152 : rightmax, op_righttype);
3153 64792 : if (selec != DEFAULT_INEQ_SEL)
3154 64786 : *leftend = selec;
3155 :
3156 : /* And similarly for the right variable. */
3157 64792 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3158 : leftmax, op_lefttype);
3159 64792 : if (selec != DEFAULT_INEQ_SEL)
3160 64792 : *rightend = selec;
3161 :
3162 : /*
3163 : * Only one of the two "end" fractions can really be less than 1.0;
3164 : * believe the smaller estimate and reset the other one to exactly 1.0. If
3165 : * we get exactly equal estimates (as can easily happen with self-joins),
3166 : * believe neither.
3167 : */
3168 64792 : if (*leftend > *rightend)
3169 24812 : *leftend = 1.0;
3170 39980 : else if (*leftend < *rightend)
3171 33434 : *rightend = 1.0;
3172 : else
3173 6546 : *leftend = *rightend = 1.0;
3174 :
3175 : /*
3176 : * Also, the fraction of the left variable that will be scanned before the
3177 : * first join pair is found is the fraction that's < the right-side
3178 : * minimum value. But only believe non-default estimates, else stick with
3179 : * our own default.
3180 : */
3181 64792 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3182 : rightmin, op_righttype);
3183 64792 : if (selec != DEFAULT_INEQ_SEL)
3184 64792 : *leftstart = selec;
3185 :
3186 : /* And similarly for the right variable. */
3187 64792 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3188 : leftmin, op_lefttype);
3189 64792 : if (selec != DEFAULT_INEQ_SEL)
3190 64792 : *rightstart = selec;
3191 :
3192 : /*
3193 : * Only one of the two "start" fractions can really be more than zero;
3194 : * believe the larger estimate and reset the other one to exactly 0.0. If
3195 : * we get exactly equal estimates (as can easily happen with self-joins),
3196 : * believe neither.
3197 : */
3198 64792 : if (*leftstart < *rightstart)
3199 19910 : *leftstart = 0.0;
3200 44882 : else if (*leftstart > *rightstart)
3201 25712 : *rightstart = 0.0;
3202 : else
3203 19170 : *leftstart = *rightstart = 0.0;
3204 :
3205 : /*
3206 : * If the sort order is nulls-first, we're going to have to skip over any
3207 : * nulls too. These would not have been counted by scalarineqsel, and we
3208 : * can safely add in this fraction regardless of whether we believe
3209 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3210 : */
3211 64792 : if (nulls_first)
3212 : {
3213 : Form_pg_statistic stats;
3214 :
3215 24 : if (HeapTupleIsValid(leftvar.statsTuple))
3216 : {
3217 24 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3218 24 : *leftstart += stats->stanullfrac;
3219 24 : CLAMP_PROBABILITY(*leftstart);
3220 24 : *leftend += stats->stanullfrac;
3221 24 : CLAMP_PROBABILITY(*leftend);
3222 : }
3223 24 : if (HeapTupleIsValid(rightvar.statsTuple))
3224 : {
3225 24 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3226 24 : *rightstart += stats->stanullfrac;
3227 24 : CLAMP_PROBABILITY(*rightstart);
3228 24 : *rightend += stats->stanullfrac;
3229 24 : CLAMP_PROBABILITY(*rightend);
3230 : }
3231 : }
3232 :
3233 : /* Disbelieve start >= end, just in case that can happen */
3234 64792 : if (*leftstart >= *leftend)
3235 : {
3236 202 : *leftstart = 0.0;
3237 202 : *leftend = 1.0;
3238 : }
3239 64792 : if (*rightstart >= *rightend)
3240 : {
3241 1016 : *rightstart = 0.0;
3242 1016 : *rightend = 1.0;
3243 : }
3244 :
3245 63776 : fail:
3246 110938 : ReleaseVariableStats(leftvar);
3247 110938 : ReleaseVariableStats(rightvar);
3248 : }
3249 :
3250 :
3251 : /*
3252 : * matchingsel -- generic matching-operator selectivity support
3253 : *
3254 : * Use these for any operators that (a) are on data types for which we collect
3255 : * standard statistics, and (b) have behavior for which the default estimate
3256 : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3257 : * operators.
3258 : */
3259 :
3260 : Datum
3261 1106 : matchingsel(PG_FUNCTION_ARGS)
3262 : {
3263 1106 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3264 1106 : Oid operator = PG_GETARG_OID(1);
3265 1106 : List *args = (List *) PG_GETARG_POINTER(2);
3266 1106 : int varRelid = PG_GETARG_INT32(3);
3267 1106 : Oid collation = PG_GET_COLLATION();
3268 : double selec;
3269 :
3270 : /* Use generic restriction selectivity logic. */
3271 1106 : selec = generic_restriction_selectivity(root, operator, collation,
3272 : args, varRelid,
3273 : DEFAULT_MATCHING_SEL);
3274 :
3275 1106 : PG_RETURN_FLOAT8((float8) selec);
3276 : }
3277 :
3278 : Datum
3279 6 : matchingjoinsel(PG_FUNCTION_ARGS)
3280 : {
3281 : /* Just punt, for the moment. */
3282 6 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3283 : }
3284 :
3285 :
3286 : /*
3287 : * Helper routine for estimate_num_groups: add an item to a list of
3288 : * GroupVarInfos, but only if it's not known equal to any of the existing
3289 : * entries.
3290 : */
3291 : typedef struct
3292 : {
3293 : Node *var; /* might be an expression, not just a Var */
3294 : RelOptInfo *rel; /* relation it belongs to */
3295 : double ndistinct; /* # distinct values */
3296 : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3297 : } GroupVarInfo;
3298 :
3299 : static List *
3300 342886 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3301 : Node *var, VariableStatData *vardata)
3302 : {
3303 : GroupVarInfo *varinfo;
3304 : double ndistinct;
3305 : bool isdefault;
3306 : ListCell *lc;
3307 :
3308 342886 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3309 :
3310 : /*
3311 : * The nullingrels bits within the var could cause the same var to be
3312 : * counted multiple times if it's marked with different nullingrels. They
3313 : * could also prevent us from matching the var to the expressions in
3314 : * extended statistics (see estimate_multivariate_ndistinct). So strip
3315 : * them out first.
3316 : */
3317 342886 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3318 :
3319 417732 : foreach(lc, varinfos)
3320 : {
3321 75666 : varinfo = (GroupVarInfo *) lfirst(lc);
3322 :
3323 : /* Drop exact duplicates */
3324 75666 : if (equal(var, varinfo->var))
3325 820 : return varinfos;
3326 :
3327 : /*
3328 : * Drop known-equal vars, but only if they belong to different
3329 : * relations (see comments for estimate_num_groups). We aren't too
3330 : * fussy about the semantics of "equal" here.
3331 : */
3332 79592 : if (vardata->rel != varinfo->rel &&
3333 4626 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3334 : {
3335 120 : if (varinfo->ndistinct <= ndistinct)
3336 : {
3337 : /* Keep older item, forget new one */
3338 120 : return varinfos;
3339 : }
3340 : else
3341 : {
3342 : /* Delete the older item */
3343 0 : varinfos = foreach_delete_current(varinfos, lc);
3344 : }
3345 : }
3346 : }
3347 :
3348 342066 : varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
3349 :
3350 342066 : varinfo->var = var;
3351 342066 : varinfo->rel = vardata->rel;
3352 342066 : varinfo->ndistinct = ndistinct;
3353 342066 : varinfo->isdefault = isdefault;
3354 342066 : varinfos = lappend(varinfos, varinfo);
3355 342066 : return varinfos;
3356 : }
3357 :
3358 : /*
3359 : * estimate_num_groups - Estimate number of groups in a grouped query
3360 : *
3361 : * Given a query having a GROUP BY clause, estimate how many groups there
3362 : * will be --- ie, the number of distinct combinations of the GROUP BY
3363 : * expressions.
3364 : *
3365 : * This routine is also used to estimate the number of rows emitted by
3366 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3367 : * actually, we only use it for DISTINCT when there's no grouping or
3368 : * aggregation ahead of the DISTINCT.)
3369 : *
3370 : * Inputs:
3371 : * root - the query
3372 : * groupExprs - list of expressions being grouped by
3373 : * input_rows - number of rows estimated to arrive at the group/unique
3374 : * filter step
3375 : * pgset - NULL, or a List** pointing to a grouping set to filter the
3376 : * groupExprs against
3377 : *
3378 : * Outputs:
3379 : * estinfo - When passed as non-NULL, the function will set bits in the
3380 : * "flags" field in order to provide callers with additional information
3381 : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3382 : * bit if we used any default values in the estimation.
3383 : *
3384 : * Given the lack of any cross-correlation statistics in the system, it's
3385 : * impossible to do anything really trustworthy with GROUP BY conditions
3386 : * involving multiple Vars. We should however avoid assuming the worst
3387 : * case (all possible cross-product terms actually appear as groups) since
3388 : * very often the grouped-by Vars are highly correlated. Our current approach
3389 : * is as follows:
3390 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3391 : * independently of their content, and are ignored in the subsequent
3392 : * steps. This is mainly because tests like "col IS NULL" break the
3393 : * heuristic used in step 2 especially badly.
3394 : * 2. Reduce the given expressions to a list of unique Vars used. For
3395 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3396 : * It is clearly correct not to count the same Var more than once.
3397 : * It is also reasonable to treat f(x) the same as x: f() cannot
3398 : * increase the number of distinct values (unless it is volatile,
3399 : * which we consider unlikely for grouping), but it probably won't
3400 : * reduce the number of distinct values much either.
3401 : * As a special case, if a GROUP BY expression can be matched to an
3402 : * expressional index for which we have statistics, then we treat the
3403 : * whole expression as though it were just a Var.
3404 : * 3. If the list contains Vars of different relations that are known equal
3405 : * due to equivalence classes, then drop all but one of the Vars from each
3406 : * known-equal set, keeping the one with smallest estimated # of values
3407 : * (since the extra values of the others can't appear in joined rows).
3408 : * Note the reason we only consider Vars of different relations is that
3409 : * if we considered ones of the same rel, we'd be double-counting the
3410 : * restriction selectivity of the equality in the next step.
3411 : * 4. For Vars within a single source rel, we multiply together the numbers
3412 : * of values, clamp to the number of rows in the rel (divided by 10 if
3413 : * more than one Var), and then multiply by a factor based on the
3414 : * selectivity of the restriction clauses for that rel. When there's
3415 : * more than one Var, the initial product is probably too high (it's the
3416 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3417 : * helpful heuristic for not letting the estimate get out of hand. (The
3418 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3419 : * we multiply by to adjust for the restriction selectivity assumes that
3420 : * the restriction clauses are independent of the grouping, which may not
3421 : * be a valid assumption, but it's hard to do better.
3422 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3423 : * rel, and multiply the results together.
3424 : * Note that rels not containing grouped Vars are ignored completely, as are
3425 : * join clauses. Such rels cannot increase the number of groups, and we
3426 : * assume such clauses do not reduce the number either (somewhat bogus,
3427 : * but we don't have the info to do better).
3428 : */
3429 : double
3430 298092 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3431 : List **pgset, EstimationInfo *estinfo)
3432 : {
3433 298092 : List *varinfos = NIL;
3434 298092 : double srf_multiplier = 1.0;
3435 : double numdistinct;
3436 : ListCell *l;
3437 : int i;
3438 :
3439 : /* Zero the estinfo output parameter, if non-NULL */
3440 298092 : if (estinfo != NULL)
3441 266634 : memset(estinfo, 0, sizeof(EstimationInfo));
3442 :
3443 : /*
3444 : * We don't ever want to return an estimate of zero groups, as that tends
3445 : * to lead to division-by-zero and other unpleasantness. The input_rows
3446 : * estimate is usually already at least 1, but clamp it just in case it
3447 : * isn't.
3448 : */
3449 298092 : input_rows = clamp_row_est(input_rows);
3450 :
3451 : /*
3452 : * If no grouping columns, there's exactly one group. (This can't happen
3453 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3454 : * corner cases with set operations.)
3455 : */
3456 298092 : if (groupExprs == NIL || (pgset && *pgset == NIL))
3457 1014 : return 1.0;
3458 :
3459 : /*
3460 : * Count groups derived from boolean grouping expressions. For other
3461 : * expressions, find the unique Vars used, treating an expression as a Var
3462 : * if we can find stats for it. For each one, record the statistical
3463 : * estimate of number of distinct values (total in its table, without
3464 : * regard for filtering).
3465 : */
3466 297078 : numdistinct = 1.0;
3467 :
3468 297078 : i = 0;
3469 638396 : foreach(l, groupExprs)
3470 : {
3471 341366 : Node *groupexpr = (Node *) lfirst(l);
3472 : double this_srf_multiplier;
3473 : VariableStatData vardata;
3474 : List *varshere;
3475 : ListCell *l2;
3476 :
3477 : /* is expression in this grouping set? */
3478 341366 : if (pgset && !list_member_int(*pgset, i++))
3479 287204 : continue;
3480 :
3481 : /*
3482 : * Set-returning functions in grouping columns are a bit problematic.
3483 : * The code below will effectively ignore their SRF nature and come up
3484 : * with a numdistinct estimate as though they were scalar functions.
3485 : * We compensate by scaling up the end result by the largest SRF
3486 : * rowcount estimate. (This will be an overestimate if the SRF
3487 : * produces multiple copies of any output value, but it seems best to
3488 : * assume the SRF's outputs are distinct. In any case, it's probably
3489 : * pointless to worry too much about this without much better
3490 : * estimates for SRF output rowcounts than we have today.)
3491 : */
3492 340578 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3493 340578 : if (srf_multiplier < this_srf_multiplier)
3494 132 : srf_multiplier = this_srf_multiplier;
3495 :
3496 : /* Short-circuit for expressions returning boolean */
3497 340578 : if (exprType(groupexpr) == BOOLOID)
3498 : {
3499 192 : numdistinct *= 2.0;
3500 192 : continue;
3501 : }
3502 :
3503 : /*
3504 : * If examine_variable is able to deduce anything about the GROUP BY
3505 : * expression, treat it as a single variable even if it's really more
3506 : * complicated.
3507 : *
3508 : * XXX This has the consequence that if there's a statistics object on
3509 : * the expression, we don't split it into individual Vars. This
3510 : * affects our selection of statistics in
3511 : * estimate_multivariate_ndistinct, because it's probably better to
3512 : * use more accurate estimate for each expression and treat them as
3513 : * independent, than to combine estimates for the extracted variables
3514 : * when we don't know how that relates to the expressions.
3515 : */
3516 340386 : examine_variable(root, groupexpr, 0, &vardata);
3517 340386 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3518 : {
3519 285636 : varinfos = add_unique_group_var(root, varinfos,
3520 : groupexpr, &vardata);
3521 285636 : ReleaseVariableStats(vardata);
3522 285636 : continue;
3523 : }
3524 54750 : ReleaseVariableStats(vardata);
3525 :
3526 : /*
3527 : * Else pull out the component Vars. Handle PlaceHolderVars by
3528 : * recursing into their arguments (effectively assuming that the
3529 : * PlaceHolderVar doesn't change the number of groups, which boils
3530 : * down to ignoring the possible addition of nulls to the result set).
3531 : */
3532 54750 : varshere = pull_var_clause(groupexpr,
3533 : PVC_RECURSE_AGGREGATES |
3534 : PVC_RECURSE_WINDOWFUNCS |
3535 : PVC_RECURSE_PLACEHOLDERS);
3536 :
3537 : /*
3538 : * If we find any variable-free GROUP BY item, then either it is a
3539 : * constant (and we can ignore it) or it contains a volatile function;
3540 : * in the latter case we punt and assume that each input row will
3541 : * yield a distinct group.
3542 : */
3543 54750 : if (varshere == NIL)
3544 : {
3545 636 : if (contain_volatile_functions(groupexpr))
3546 48 : return input_rows;
3547 588 : continue;
3548 : }
3549 :
3550 : /*
3551 : * Else add variables to varinfos list
3552 : */
3553 111364 : foreach(l2, varshere)
3554 : {
3555 57250 : Node *var = (Node *) lfirst(l2);
3556 :
3557 57250 : examine_variable(root, var, 0, &vardata);
3558 57250 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3559 57250 : ReleaseVariableStats(vardata);
3560 : }
3561 : }
3562 :
3563 : /*
3564 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3565 : * list.
3566 : */
3567 297030 : if (varinfos == NIL)
3568 : {
3569 : /* Apply SRF multiplier as we would do in the long path */
3570 346 : numdistinct *= srf_multiplier;
3571 : /* Round off */
3572 346 : numdistinct = ceil(numdistinct);
3573 : /* Guard against out-of-range answers */
3574 346 : if (numdistinct > input_rows)
3575 38 : numdistinct = input_rows;
3576 346 : if (numdistinct < 1.0)
3577 0 : numdistinct = 1.0;
3578 346 : return numdistinct;
3579 : }
3580 :
3581 : /*
3582 : * Group Vars by relation and estimate total numdistinct.
3583 : *
3584 : * For each iteration of the outer loop, we process the frontmost Var in
3585 : * varinfos, plus all other Vars in the same relation. We remove these
3586 : * Vars from the newvarinfos list for the next iteration. This is the
3587 : * easiest way to group Vars of same rel together.
3588 : */
3589 : do
3590 : {
3591 298612 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3592 298612 : RelOptInfo *rel = varinfo1->rel;
3593 298612 : double reldistinct = 1;
3594 298612 : double relmaxndistinct = reldistinct;
3595 298612 : int relvarcount = 0;
3596 298612 : List *newvarinfos = NIL;
3597 298612 : List *relvarinfos = NIL;
3598 :
3599 : /*
3600 : * Split the list of varinfos in two - one for the current rel, one
3601 : * for remaining Vars on other rels.
3602 : */
3603 298612 : relvarinfos = lappend(relvarinfos, varinfo1);
3604 345930 : for_each_from(l, varinfos, 1)
3605 : {
3606 47318 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3607 :
3608 47318 : if (varinfo2->rel == varinfo1->rel)
3609 : {
3610 : /* varinfos on current rel */
3611 43454 : relvarinfos = lappend(relvarinfos, varinfo2);
3612 : }
3613 : else
3614 : {
3615 : /* not time to process varinfo2 yet */
3616 3864 : newvarinfos = lappend(newvarinfos, varinfo2);
3617 : }
3618 : }
3619 :
3620 : /*
3621 : * Get the numdistinct estimate for the Vars of this rel. We
3622 : * iteratively search for multivariate n-distinct with maximum number
3623 : * of vars; assuming that each var group is independent of the others,
3624 : * we multiply them together. Any remaining relvarinfos after no more
3625 : * multivariate matches are found are assumed independent too, so
3626 : * their individual ndistinct estimates are multiplied also.
3627 : *
3628 : * While iterating, count how many separate numdistinct values we
3629 : * apply. We apply a fudge factor below, but only if we multiplied
3630 : * more than one such values.
3631 : */
3632 597350 : while (relvarinfos)
3633 : {
3634 : double mvndistinct;
3635 :
3636 298738 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3637 : &mvndistinct))
3638 : {
3639 402 : reldistinct *= mvndistinct;
3640 402 : if (relmaxndistinct < mvndistinct)
3641 390 : relmaxndistinct = mvndistinct;
3642 402 : relvarcount++;
3643 : }
3644 : else
3645 : {
3646 639550 : foreach(l, relvarinfos)
3647 : {
3648 341214 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3649 :
3650 341214 : reldistinct *= varinfo2->ndistinct;
3651 341214 : if (relmaxndistinct < varinfo2->ndistinct)
3652 298496 : relmaxndistinct = varinfo2->ndistinct;
3653 341214 : relvarcount++;
3654 :
3655 : /*
3656 : * When varinfo2's isdefault is set then we'd better set
3657 : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3658 : */
3659 341214 : if (estinfo != NULL && varinfo2->isdefault)
3660 20676 : estinfo->flags |= SELFLAG_USED_DEFAULT;
3661 : }
3662 :
3663 : /* we're done with this relation */
3664 298336 : relvarinfos = NIL;
3665 : }
3666 : }
3667 :
3668 : /*
3669 : * Sanity check --- don't divide by zero if empty relation.
3670 : */
3671 : Assert(IS_SIMPLE_REL(rel));
3672 298612 : if (rel->tuples > 0)
3673 : {
3674 : /*
3675 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
3676 : * fudge factor is because the Vars are probably correlated but we
3677 : * don't know by how much. We should never clamp to less than the
3678 : * largest ndistinct value for any of the Vars, though, since
3679 : * there will surely be at least that many groups.
3680 : */
3681 297600 : double clamp = rel->tuples;
3682 :
3683 297600 : if (relvarcount > 1)
3684 : {
3685 38952 : clamp *= 0.1;
3686 38952 : if (clamp < relmaxndistinct)
3687 : {
3688 37340 : clamp = relmaxndistinct;
3689 : /* for sanity in case some ndistinct is too large: */
3690 37340 : if (clamp > rel->tuples)
3691 78 : clamp = rel->tuples;
3692 : }
3693 : }
3694 297600 : if (reldistinct > clamp)
3695 37036 : reldistinct = clamp;
3696 :
3697 : /*
3698 : * Update the estimate based on the restriction selectivity,
3699 : * guarding against division by zero when reldistinct is zero.
3700 : * Also skip this if we know that we are returning all rows.
3701 : */
3702 297600 : if (reldistinct > 0 && rel->rows < rel->tuples)
3703 : {
3704 : /*
3705 : * Given a table containing N rows with n distinct values in a
3706 : * uniform distribution, if we select p rows at random then
3707 : * the expected number of distinct values selected is
3708 : *
3709 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
3710 : *
3711 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
3712 : *
3713 : * See "Approximating block accesses in database
3714 : * organizations", S. B. Yao, Communications of the ACM,
3715 : * Volume 20 Issue 4, April 1977 Pages 260-261.
3716 : *
3717 : * Alternatively, re-arranging the terms from the factorials,
3718 : * this may be written as
3719 : *
3720 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
3721 : *
3722 : * This form of the formula is more efficient to compute in
3723 : * the common case where p is larger than N/n. Additionally,
3724 : * as pointed out by Dell'Era, if i << N for all terms in the
3725 : * product, it can be approximated by
3726 : *
3727 : * n * (1 - ((N-p)/N)^(N/n))
3728 : *
3729 : * See "Expected distinct values when selecting from a bag
3730 : * without replacement", Alberto Dell'Era,
3731 : * http://www.adellera.it/investigations/distinct_balls/.
3732 : *
3733 : * The condition i << N is equivalent to n >> 1, so this is a
3734 : * good approximation when the number of distinct values in
3735 : * the table is large. It turns out that this formula also
3736 : * works well even when n is small.
3737 : */
3738 91224 : reldistinct *=
3739 91224 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
3740 91224 : rel->tuples / reldistinct));
3741 : }
3742 297600 : reldistinct = clamp_row_est(reldistinct);
3743 :
3744 : /*
3745 : * Update estimate of total distinct groups.
3746 : */
3747 297600 : numdistinct *= reldistinct;
3748 : }
3749 :
3750 298612 : varinfos = newvarinfos;
3751 298612 : } while (varinfos != NIL);
3752 :
3753 : /* Now we can account for the effects of any SRFs */
3754 296684 : numdistinct *= srf_multiplier;
3755 :
3756 : /* Round off */
3757 296684 : numdistinct = ceil(numdistinct);
3758 :
3759 : /* Guard against out-of-range answers */
3760 296684 : if (numdistinct > input_rows)
3761 59980 : numdistinct = input_rows;
3762 296684 : if (numdistinct < 1.0)
3763 0 : numdistinct = 1.0;
3764 :
3765 296684 : return numdistinct;
3766 : }
3767 :
3768 : /*
3769 : * Try to estimate the bucket size of the hash join inner side when the join
3770 : * condition contains two or more clauses by employing extended statistics.
3771 : *
3772 : * The main idea of this approach is that the distinct value generated by
3773 : * multivariate estimation on two or more columns would provide less bucket size
3774 : * than estimation on one separate column.
3775 : *
3776 : * IMPORTANT: It is crucial to synchronize the approach of combining different
3777 : * estimations with the caller's method.
3778 : *
3779 : * Return a list of clauses that didn't fetch any extended statistics.
3780 : */
3781 : List *
3782 263296 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
3783 : List *hashclauses,
3784 : Selectivity *innerbucketsize)
3785 : {
3786 263296 : List *clauses = list_copy(hashclauses);
3787 263296 : List *otherclauses = NIL;
3788 263296 : double ndistinct = 1.0;
3789 :
3790 263296 : if (list_length(hashclauses) <= 1)
3791 :
3792 : /*
3793 : * Nothing to do for a single clause. Could we employ univariate
3794 : * extended stat here?
3795 : */
3796 236218 : return hashclauses;
3797 :
3798 54156 : while (clauses != NIL)
3799 : {
3800 : ListCell *lc;
3801 27078 : int relid = -1;
3802 27078 : List *varinfos = NIL;
3803 27078 : List *origin_rinfos = NIL;
3804 : double mvndistinct;
3805 : List *origin_varinfos;
3806 27078 : int group_relid = -1;
3807 27078 : RelOptInfo *group_rel = NULL;
3808 : ListCell *lc1,
3809 : *lc2;
3810 :
3811 : /*
3812 : * Find clauses, referencing the same single base relation and try to
3813 : * estimate such a group with extended statistics. Create varinfo for
3814 : * an approved clause, push it to otherclauses, if it can't be
3815 : * estimated here or ignore to process at the next iteration.
3816 : */
3817 81846 : foreach(lc, clauses)
3818 : {
3819 54768 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
3820 : Node *expr;
3821 : Relids relids;
3822 : GroupVarInfo *varinfo;
3823 :
3824 : /*
3825 : * Find the inner side of the join, which we need to estimate the
3826 : * number of buckets. Use outer_is_left because the
3827 : * clause_sides_match_join routine has called on hash clauses.
3828 : */
3829 109536 : relids = rinfo->outer_is_left ?
3830 54768 : rinfo->right_relids : rinfo->left_relids;
3831 109536 : expr = rinfo->outer_is_left ?
3832 54768 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
3833 :
3834 54768 : if (bms_get_singleton_member(relids, &relid) &&
3835 54114 : root->simple_rel_array[relid]->statlist != NIL)
3836 : {
3837 : /*
3838 : * This inner-side expression references only one relation.
3839 : * Extended statistics on this clause can exist.
3840 : */
3841 18 : if (group_relid < 0)
3842 : {
3843 6 : RangeTblEntry *rte = root->simple_rte_array[relid];
3844 :
3845 6 : if (!rte || (rte->relkind != RELKIND_RELATION &&
3846 0 : rte->relkind != RELKIND_MATVIEW &&
3847 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
3848 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
3849 : {
3850 : /* Extended statistics can't exist in principle */
3851 0 : otherclauses = lappend(otherclauses, rinfo);
3852 0 : clauses = foreach_delete_current(clauses, lc);
3853 0 : continue;
3854 : }
3855 :
3856 6 : group_relid = relid;
3857 6 : group_rel = root->simple_rel_array[relid];
3858 : }
3859 12 : else if (group_relid != relid)
3860 :
3861 : /*
3862 : * Being in the group forming state we don't need other
3863 : * clauses.
3864 : */
3865 0 : continue;
3866 :
3867 18 : varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
3868 18 : varinfo->var = expr;
3869 18 : varinfo->rel = root->simple_rel_array[relid];
3870 18 : varinfo->ndistinct = 0.0;
3871 18 : varinfo->isdefault = false;
3872 18 : varinfos = lappend(varinfos, varinfo);
3873 :
3874 : /*
3875 : * Remember the link to RestrictInfo for the case the clause
3876 : * is failed to be estimated.
3877 : */
3878 18 : origin_rinfos = lappend(origin_rinfos, rinfo);
3879 : }
3880 : else
3881 : /* This clause can't be estimated with extended statistics */
3882 54750 : otherclauses = lappend(otherclauses, rinfo);
3883 :
3884 54768 : clauses = foreach_delete_current(clauses, lc);
3885 : }
3886 :
3887 27078 : if (list_length(varinfos) < 2)
3888 : {
3889 : /*
3890 : * Multivariate statistics doesn't apply to single columns except
3891 : * for expressions, but it has not been implemented yet.
3892 : */
3893 27072 : otherclauses = list_concat(otherclauses, origin_rinfos);
3894 27072 : list_free_deep(varinfos);
3895 27072 : list_free(origin_rinfos);
3896 27072 : continue;
3897 : }
3898 :
3899 : Assert(group_rel != NULL);
3900 :
3901 : /* Employ the extended statistics. */
3902 6 : origin_varinfos = varinfos;
3903 : for (;;)
3904 6 : {
3905 12 : bool estimated = estimate_multivariate_ndistinct(root,
3906 : group_rel,
3907 : &varinfos,
3908 : &mvndistinct);
3909 :
3910 12 : if (!estimated)
3911 6 : break;
3912 :
3913 : /*
3914 : * We've got an estimation. Use ndistinct value in a consistent
3915 : * way - according to the caller's logic (see
3916 : * final_cost_hashjoin).
3917 : */
3918 6 : if (ndistinct < mvndistinct)
3919 6 : ndistinct = mvndistinct;
3920 : Assert(ndistinct >= 1.0);
3921 : }
3922 :
3923 : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
3924 :
3925 : /* Collect unmatched clauses as otherclauses. */
3926 24 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
3927 : {
3928 18 : GroupVarInfo *vinfo = lfirst(lc1);
3929 :
3930 18 : if (!list_member_ptr(varinfos, vinfo))
3931 : /* Already estimated */
3932 18 : continue;
3933 :
3934 : /* Can't be estimated here - push to the returning list */
3935 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
3936 : }
3937 : }
3938 :
3939 27078 : *innerbucketsize = 1.0 / ndistinct;
3940 27078 : return otherclauses;
3941 : }
3942 :
3943 : /*
3944 : * Estimate hash bucket statistics when the specified expression is used
3945 : * as a hash key for the given number of buckets.
3946 : *
3947 : * This attempts to determine two values:
3948 : *
3949 : * 1. The frequency of the most common value of the expression (returns
3950 : * zero into *mcv_freq if we can't get that).
3951 : *
3952 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
3953 : * divided by total tuples in relation.
3954 : *
3955 : * XXX This is really pretty bogus since we're effectively assuming that the
3956 : * distribution of hash keys will be the same after applying restriction
3957 : * clauses as it was in the underlying relation. However, we are not nearly
3958 : * smart enough to figure out how the restrict clauses might change the
3959 : * distribution, so this will have to do for now.
3960 : *
3961 : * We are passed the number of buckets the executor will use for the given
3962 : * input relation. If the data were perfectly distributed, with the same
3963 : * number of tuples going into each available bucket, then the bucketsize
3964 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
3965 : * only if (a) there are at least nbuckets distinct data values, and (b)
3966 : * we have a not-too-skewed data distribution. Otherwise the buckets will
3967 : * be nonuniformly occupied. If the other relation in the join has a key
3968 : * distribution similar to this one's, then the most-loaded buckets are
3969 : * exactly those that will be probed most often. Therefore, the "average"
3970 : * bucket size for costing purposes should really be taken as something close
3971 : * to the "worst case" bucket size. We try to estimate this by adjusting the
3972 : * fraction if there are too few distinct data values, and then scaling up
3973 : * by the ratio of the most common value's frequency to the average frequency.
3974 : *
3975 : * If no statistics are available, use a default estimate of 0.1. This will
3976 : * discourage use of a hash rather strongly if the inner relation is large,
3977 : * which is what we want. We do not want to hash unless we know that the
3978 : * inner rel is well-dispersed (or the alternatives seem much worse).
3979 : *
3980 : * The caller should also check that the mcv_freq is not so large that the
3981 : * most common value would by itself require an impractically large bucket.
3982 : * In a hash join, the executor can split buckets if they get too big, but
3983 : * obviously that doesn't help for a bucket that contains many duplicates of
3984 : * the same value.
3985 : */
3986 : void
3987 156130 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
3988 : Selectivity *mcv_freq,
3989 : Selectivity *bucketsize_frac)
3990 : {
3991 : VariableStatData vardata;
3992 : double estfract,
3993 : ndistinct,
3994 : stanullfrac,
3995 : avgfreq;
3996 : bool isdefault;
3997 : AttStatsSlot sslot;
3998 :
3999 156130 : examine_variable(root, hashkey, 0, &vardata);
4000 :
4001 : /* Look up the frequency of the most common value, if available */
4002 156130 : *mcv_freq = 0.0;
4003 :
4004 156130 : if (HeapTupleIsValid(vardata.statsTuple))
4005 : {
4006 109358 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4007 : STATISTIC_KIND_MCV, InvalidOid,
4008 : ATTSTATSSLOT_NUMBERS))
4009 : {
4010 : /*
4011 : * The first MCV stat is for the most common value.
4012 : */
4013 52478 : if (sslot.nnumbers > 0)
4014 52478 : *mcv_freq = sslot.numbers[0];
4015 52478 : free_attstatsslot(&sslot);
4016 : }
4017 : }
4018 :
4019 : /* Get number of distinct values */
4020 156130 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4021 :
4022 : /*
4023 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4024 : * mcv_freq is known to be even higher than that, use it instead.
4025 : */
4026 156130 : if (isdefault)
4027 : {
4028 21946 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4029 21946 : ReleaseVariableStats(vardata);
4030 21946 : return;
4031 : }
4032 :
4033 : /* Get fraction that are null */
4034 134184 : if (HeapTupleIsValid(vardata.statsTuple))
4035 : {
4036 : Form_pg_statistic stats;
4037 :
4038 109340 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
4039 109340 : stanullfrac = stats->stanullfrac;
4040 : }
4041 : else
4042 24844 : stanullfrac = 0.0;
4043 :
4044 : /* Compute avg freq of all distinct data values in raw relation */
4045 134184 : avgfreq = (1.0 - stanullfrac) / ndistinct;
4046 :
4047 : /*
4048 : * Adjust ndistinct to account for restriction clauses. Observe we are
4049 : * assuming that the data distribution is affected uniformly by the
4050 : * restriction clauses!
4051 : *
4052 : * XXX Possibly better way, but much more expensive: multiply by
4053 : * selectivity of rel's restriction clauses that mention the target Var.
4054 : */
4055 134184 : if (vardata.rel && vardata.rel->tuples > 0)
4056 : {
4057 134170 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
4058 134170 : ndistinct = clamp_row_est(ndistinct);
4059 : }
4060 :
4061 : /*
4062 : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
4063 : * number of buckets is less than the expected number of distinct values;
4064 : * otherwise it is 1/ndistinct.
4065 : */
4066 134184 : if (ndistinct > nbuckets)
4067 92 : estfract = 1.0 / nbuckets;
4068 : else
4069 134092 : estfract = 1.0 / ndistinct;
4070 :
4071 : /*
4072 : * Adjust estimated bucketsize upward to account for skewed distribution.
4073 : */
4074 134184 : if (avgfreq > 0.0 && *mcv_freq > avgfreq)
4075 47394 : estfract *= *mcv_freq / avgfreq;
4076 :
4077 : /*
4078 : * Clamp bucketsize to sane range (the above adjustment could easily
4079 : * produce an out-of-range result). We set the lower bound a little above
4080 : * zero, since zero isn't a very sane result.
4081 : */
4082 134184 : if (estfract < 1.0e-6)
4083 0 : estfract = 1.0e-6;
4084 134184 : else if (estfract > 1.0)
4085 33290 : estfract = 1.0;
4086 :
4087 134184 : *bucketsize_frac = (Selectivity) estfract;
4088 :
4089 134184 : ReleaseVariableStats(vardata);
4090 : }
4091 :
4092 : /*
4093 : * estimate_hashagg_tablesize
4094 : * estimate the number of bytes that a hash aggregate hashtable will
4095 : * require based on the agg_costs, path width and number of groups.
4096 : *
4097 : * We return the result as "double" to forestall any possible overflow
4098 : * problem in the multiplication by dNumGroups.
4099 : *
4100 : * XXX this may be over-estimating the size now that hashagg knows to omit
4101 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4102 : * grouping columns not in the hashed set are counted here even though hashagg
4103 : * won't store them. Is this a problem?
4104 : */
4105 : double
4106 2338 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4107 : const AggClauseCosts *agg_costs, double dNumGroups)
4108 : {
4109 : Size hashentrysize;
4110 :
4111 2338 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4112 2338 : path->pathtarget->width,
4113 : agg_costs->transitionSpace);
4114 :
4115 : /*
4116 : * Note that this disregards the effect of fill-factor and growth policy
4117 : * of the hash table. That's probably ok, given that the default
4118 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4119 : * "double-in-size" growth policies here.
4120 : */
4121 2338 : return hashentrysize * dNumGroups;
4122 : }
4123 :
4124 :
4125 : /*-------------------------------------------------------------------------
4126 : *
4127 : * Support routines
4128 : *
4129 : *-------------------------------------------------------------------------
4130 : */
4131 :
4132 : /*
4133 : * Find applicable ndistinct statistics for the given list of VarInfos (which
4134 : * must all belong to the given rel), and update *ndistinct to the estimate of
4135 : * the MVNDistinctItem that best matches. If a match it found, *varinfos is
4136 : * updated to remove the list of matched varinfos.
4137 : *
4138 : * Varinfos that aren't for simple Vars are ignored.
4139 : *
4140 : * Return true if we're able to find a match, false otherwise.
4141 : */
4142 : static bool
4143 298750 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4144 : List **varinfos, double *ndistinct)
4145 : {
4146 : ListCell *lc;
4147 : int nmatches_vars;
4148 : int nmatches_exprs;
4149 298750 : Oid statOid = InvalidOid;
4150 : MVNDistinct *stats;
4151 298750 : StatisticExtInfo *matched_info = NULL;
4152 298750 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4153 :
4154 : /* bail out immediately if the table has no extended statistics */
4155 298750 : if (!rel->statlist)
4156 298210 : return false;
4157 :
4158 : /* look for the ndistinct statistics object matching the most vars */
4159 540 : nmatches_vars = 0; /* we require at least two matches */
4160 540 : nmatches_exprs = 0;
4161 2148 : foreach(lc, rel->statlist)
4162 : {
4163 : ListCell *lc2;
4164 1608 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
4165 1608 : int nshared_vars = 0;
4166 1608 : int nshared_exprs = 0;
4167 :
4168 : /* skip statistics of other kinds */
4169 1608 : if (info->kind != STATS_EXT_NDISTINCT)
4170 750 : continue;
4171 :
4172 : /* skip statistics with mismatching stxdinherit value */
4173 858 : if (info->inherit != rte->inh)
4174 24 : continue;
4175 :
4176 : /*
4177 : * Determine how many expressions (and variables in non-matched
4178 : * expressions) match. We'll then use these numbers to pick the
4179 : * statistics object that best matches the clauses.
4180 : */
4181 2658 : foreach(lc2, *varinfos)
4182 : {
4183 : ListCell *lc3;
4184 1824 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4185 : AttrNumber attnum;
4186 :
4187 : Assert(varinfo->rel == rel);
4188 :
4189 : /* simple Var, search in statistics keys directly */
4190 1824 : if (IsA(varinfo->var, Var))
4191 : {
4192 1458 : attnum = ((Var *) varinfo->var)->varattno;
4193 :
4194 : /*
4195 : * Ignore system attributes - we don't support statistics on
4196 : * them, so can't match them (and it'd fail as the values are
4197 : * negative).
4198 : */
4199 1458 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4200 12 : continue;
4201 :
4202 1446 : if (bms_is_member(attnum, info->keys))
4203 840 : nshared_vars++;
4204 :
4205 1446 : continue;
4206 : }
4207 :
4208 : /* expression - see if it's in the statistics object */
4209 660 : foreach(lc3, info->exprs)
4210 : {
4211 528 : Node *expr = (Node *) lfirst(lc3);
4212 :
4213 528 : if (equal(varinfo->var, expr))
4214 : {
4215 234 : nshared_exprs++;
4216 234 : break;
4217 : }
4218 : }
4219 : }
4220 :
4221 834 : if (nshared_vars + nshared_exprs < 2)
4222 390 : continue;
4223 :
4224 : /*
4225 : * Does this statistics object match more columns than the currently
4226 : * best object? If so, use this one instead.
4227 : *
4228 : * XXX This should break ties using name of the object, or something
4229 : * like that, to make the outcome stable.
4230 : */
4231 444 : if ((nshared_exprs > nmatches_exprs) ||
4232 336 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4233 : {
4234 420 : statOid = info->statOid;
4235 420 : nmatches_vars = nshared_vars;
4236 420 : nmatches_exprs = nshared_exprs;
4237 420 : matched_info = info;
4238 : }
4239 : }
4240 :
4241 : /* No match? */
4242 540 : if (statOid == InvalidOid)
4243 132 : return false;
4244 :
4245 : Assert(nmatches_vars + nmatches_exprs > 1);
4246 :
4247 408 : stats = statext_ndistinct_load(statOid, rte->inh);
4248 :
4249 : /*
4250 : * If we have a match, search it for the specific item that matches (there
4251 : * must be one), and construct the output values.
4252 : */
4253 408 : if (stats)
4254 : {
4255 : int i;
4256 408 : List *newlist = NIL;
4257 408 : MVNDistinctItem *item = NULL;
4258 : ListCell *lc2;
4259 408 : Bitmapset *matched = NULL;
4260 : AttrNumber attnum_offset;
4261 :
4262 : /*
4263 : * How much we need to offset the attnums? If there are no
4264 : * expressions, no offset is needed. Otherwise offset enough to move
4265 : * the lowest one (which is equal to number of expressions) to 1.
4266 : */
4267 408 : if (matched_info->exprs)
4268 144 : attnum_offset = (list_length(matched_info->exprs) + 1);
4269 : else
4270 264 : attnum_offset = 0;
4271 :
4272 : /* see what actually matched */
4273 1434 : foreach(lc2, *varinfos)
4274 : {
4275 : ListCell *lc3;
4276 : int idx;
4277 1026 : bool found = false;
4278 :
4279 1026 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4280 :
4281 : /*
4282 : * Process a simple Var expression, by matching it to keys
4283 : * directly. If there's a matching expression, we'll try matching
4284 : * it later.
4285 : */
4286 1026 : if (IsA(varinfo->var, Var))
4287 : {
4288 840 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4289 :
4290 : /*
4291 : * Ignore expressions on system attributes. Can't rely on the
4292 : * bms check for negative values.
4293 : */
4294 840 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4295 6 : continue;
4296 :
4297 : /* Is the variable covered by the statistics object? */
4298 834 : if (!bms_is_member(attnum, matched_info->keys))
4299 120 : continue;
4300 :
4301 714 : attnum = attnum + attnum_offset;
4302 :
4303 : /* ensure sufficient offset */
4304 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4305 :
4306 714 : matched = bms_add_member(matched, attnum);
4307 :
4308 714 : found = true;
4309 : }
4310 :
4311 : /*
4312 : * XXX Maybe we should allow searching the expressions even if we
4313 : * found an attribute matching the expression? That would handle
4314 : * trivial expressions like "(a)" but it seems fairly useless.
4315 : */
4316 900 : if (found)
4317 714 : continue;
4318 :
4319 : /* expression - see if it's in the statistics object */
4320 186 : idx = 0;
4321 306 : foreach(lc3, matched_info->exprs)
4322 : {
4323 276 : Node *expr = (Node *) lfirst(lc3);
4324 :
4325 276 : if (equal(varinfo->var, expr))
4326 : {
4327 156 : AttrNumber attnum = -(idx + 1);
4328 :
4329 156 : attnum = attnum + attnum_offset;
4330 :
4331 : /* ensure sufficient offset */
4332 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4333 :
4334 156 : matched = bms_add_member(matched, attnum);
4335 :
4336 : /* there should be just one matching expression */
4337 156 : break;
4338 : }
4339 :
4340 120 : idx++;
4341 : }
4342 : }
4343 :
4344 : /* Find the specific item that exactly matches the combination */
4345 846 : for (i = 0; i < stats->nitems; i++)
4346 : {
4347 : int j;
4348 846 : MVNDistinctItem *tmpitem = &stats->items[i];
4349 :
4350 846 : if (tmpitem->nattributes != bms_num_members(matched))
4351 162 : continue;
4352 :
4353 : /* assume it's the right item */
4354 684 : item = tmpitem;
4355 :
4356 : /* check that all item attributes/expressions fit the match */
4357 1638 : for (j = 0; j < tmpitem->nattributes; j++)
4358 : {
4359 1230 : AttrNumber attnum = tmpitem->attributes[j];
4360 :
4361 : /*
4362 : * Thanks to how we constructed the matched bitmap above, we
4363 : * can just offset all attnums the same way.
4364 : */
4365 1230 : attnum = attnum + attnum_offset;
4366 :
4367 1230 : if (!bms_is_member(attnum, matched))
4368 : {
4369 : /* nah, it's not this item */
4370 276 : item = NULL;
4371 276 : break;
4372 : }
4373 : }
4374 :
4375 : /*
4376 : * If the item has all the matched attributes, we know it's the
4377 : * right one - there can't be a better one. matching more.
4378 : */
4379 684 : if (item)
4380 408 : break;
4381 : }
4382 :
4383 : /*
4384 : * Make sure we found an item. There has to be one, because ndistinct
4385 : * statistics includes all combinations of attributes.
4386 : */
4387 408 : if (!item)
4388 0 : elog(ERROR, "corrupt MVNDistinct entry");
4389 :
4390 : /* Form the output varinfo list, keeping only unmatched ones */
4391 1434 : foreach(lc, *varinfos)
4392 : {
4393 1026 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4394 : ListCell *lc3;
4395 1026 : bool found = false;
4396 :
4397 : /*
4398 : * Let's look at plain variables first, because it's the most
4399 : * common case and the check is quite cheap. We can simply get the
4400 : * attnum and check (with an offset) matched bitmap.
4401 : */
4402 1026 : if (IsA(varinfo->var, Var))
4403 : {
4404 840 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4405 :
4406 : /*
4407 : * If it's a system attribute, we're done. We don't support
4408 : * extended statistics on system attributes, so it's clearly
4409 : * not matched. Just keep the expression and continue.
4410 : */
4411 840 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4412 : {
4413 6 : newlist = lappend(newlist, varinfo);
4414 6 : continue;
4415 : }
4416 :
4417 : /* apply the same offset as above */
4418 834 : attnum += attnum_offset;
4419 :
4420 : /* if it's not matched, keep the varinfo */
4421 834 : if (!bms_is_member(attnum, matched))
4422 120 : newlist = lappend(newlist, varinfo);
4423 :
4424 : /* The rest of the loop deals with complex expressions. */
4425 834 : continue;
4426 : }
4427 :
4428 : /*
4429 : * Process complex expressions, not just simple Vars.
4430 : *
4431 : * First, we search for an exact match of an expression. If we
4432 : * find one, we can just discard the whole GroupVarInfo, with all
4433 : * the variables we extracted from it.
4434 : *
4435 : * Otherwise we inspect the individual vars, and try matching it
4436 : * to variables in the item.
4437 : */
4438 306 : foreach(lc3, matched_info->exprs)
4439 : {
4440 276 : Node *expr = (Node *) lfirst(lc3);
4441 :
4442 276 : if (equal(varinfo->var, expr))
4443 : {
4444 156 : found = true;
4445 156 : break;
4446 : }
4447 : }
4448 :
4449 : /* found exact match, skip */
4450 186 : if (found)
4451 156 : continue;
4452 :
4453 30 : newlist = lappend(newlist, varinfo);
4454 : }
4455 :
4456 408 : *varinfos = newlist;
4457 408 : *ndistinct = item->ndistinct;
4458 408 : return true;
4459 : }
4460 :
4461 0 : return false;
4462 : }
4463 :
4464 : /*
4465 : * convert_to_scalar
4466 : * Convert non-NULL values of the indicated types to the comparison
4467 : * scale needed by scalarineqsel().
4468 : * Returns "true" if successful.
4469 : *
4470 : * XXX this routine is a hack: ideally we should look up the conversion
4471 : * subroutines in pg_type.
4472 : *
4473 : * All numeric datatypes are simply converted to their equivalent
4474 : * "double" values. (NUMERIC values that are outside the range of "double"
4475 : * are clamped to +/- HUGE_VAL.)
4476 : *
4477 : * String datatypes are converted by convert_string_to_scalar(),
4478 : * which is explained below. The reason why this routine deals with
4479 : * three values at a time, not just one, is that we need it for strings.
4480 : *
4481 : * The bytea datatype is just enough different from strings that it has
4482 : * to be treated separately.
4483 : *
4484 : * The several datatypes representing absolute times are all converted
4485 : * to Timestamp, which is actually an int64, and then we promote that to
4486 : * a double. Note this will give correct results even for the "special"
4487 : * values of Timestamp, since those are chosen to compare correctly;
4488 : * see timestamp_cmp.
4489 : *
4490 : * The several datatypes representing relative times (intervals) are all
4491 : * converted to measurements expressed in seconds.
4492 : */
4493 : static bool
4494 86652 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4495 : Datum lobound, Datum hibound, Oid boundstypid,
4496 : double *scaledlobound, double *scaledhibound)
4497 : {
4498 86652 : bool failure = false;
4499 :
4500 : /*
4501 : * Both the valuetypid and the boundstypid should exactly match the
4502 : * declared input type(s) of the operator we are invoked for. However,
4503 : * extensions might try to use scalarineqsel as estimator for operators
4504 : * with input type(s) we don't handle here; in such cases, we want to
4505 : * return false, not fail. In any case, we mustn't assume that valuetypid
4506 : * and boundstypid are identical.
4507 : *
4508 : * XXX The histogram we are interpolating between points of could belong
4509 : * to a column that's only binary-compatible with the declared type. In
4510 : * essence we are assuming that the semantics of binary-compatible types
4511 : * are enough alike that we can use a histogram generated with one type's
4512 : * operators to estimate selectivity for the other's. This is outright
4513 : * wrong in some cases --- in particular signed versus unsigned
4514 : * interpretation could trip us up. But it's useful enough in the
4515 : * majority of cases that we do it anyway. Should think about more
4516 : * rigorous ways to do it.
4517 : */
4518 86652 : switch (valuetypid)
4519 : {
4520 : /*
4521 : * Built-in numeric types
4522 : */
4523 80280 : case BOOLOID:
4524 : case INT2OID:
4525 : case INT4OID:
4526 : case INT8OID:
4527 : case FLOAT4OID:
4528 : case FLOAT8OID:
4529 : case NUMERICOID:
4530 : case OIDOID:
4531 : case REGPROCOID:
4532 : case REGPROCEDUREOID:
4533 : case REGOPEROID:
4534 : case REGOPERATOROID:
4535 : case REGCLASSOID:
4536 : case REGTYPEOID:
4537 : case REGCOLLATIONOID:
4538 : case REGCONFIGOID:
4539 : case REGDICTIONARYOID:
4540 : case REGROLEOID:
4541 : case REGNAMESPACEOID:
4542 80280 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4543 : &failure);
4544 80280 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4545 : &failure);
4546 80280 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4547 : &failure);
4548 80280 : return !failure;
4549 :
4550 : /*
4551 : * Built-in string types
4552 : */
4553 6372 : case CHAROID:
4554 : case BPCHAROID:
4555 : case VARCHAROID:
4556 : case TEXTOID:
4557 : case NAMEOID:
4558 : {
4559 6372 : char *valstr = convert_string_datum(value, valuetypid,
4560 : collid, &failure);
4561 6372 : char *lostr = convert_string_datum(lobound, boundstypid,
4562 : collid, &failure);
4563 6372 : char *histr = convert_string_datum(hibound, boundstypid,
4564 : collid, &failure);
4565 :
4566 : /*
4567 : * Bail out if any of the values is not of string type. We
4568 : * might leak converted strings for the other value(s), but
4569 : * that's not worth troubling over.
4570 : */
4571 6372 : if (failure)
4572 0 : return false;
4573 :
4574 6372 : convert_string_to_scalar(valstr, scaledvalue,
4575 : lostr, scaledlobound,
4576 : histr, scaledhibound);
4577 6372 : pfree(valstr);
4578 6372 : pfree(lostr);
4579 6372 : pfree(histr);
4580 6372 : return true;
4581 : }
4582 :
4583 : /*
4584 : * Built-in bytea type
4585 : */
4586 0 : case BYTEAOID:
4587 : {
4588 : /* We only support bytea vs bytea comparison */
4589 0 : if (boundstypid != BYTEAOID)
4590 0 : return false;
4591 0 : convert_bytea_to_scalar(value, scaledvalue,
4592 : lobound, scaledlobound,
4593 : hibound, scaledhibound);
4594 0 : return true;
4595 : }
4596 :
4597 : /*
4598 : * Built-in time types
4599 : */
4600 0 : case TIMESTAMPOID:
4601 : case TIMESTAMPTZOID:
4602 : case DATEOID:
4603 : case INTERVALOID:
4604 : case TIMEOID:
4605 : case TIMETZOID:
4606 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
4607 : &failure);
4608 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
4609 : &failure);
4610 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
4611 : &failure);
4612 0 : return !failure;
4613 :
4614 : /*
4615 : * Built-in network types
4616 : */
4617 0 : case INETOID:
4618 : case CIDROID:
4619 : case MACADDROID:
4620 : case MACADDR8OID:
4621 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
4622 : &failure);
4623 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
4624 : &failure);
4625 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
4626 : &failure);
4627 0 : return !failure;
4628 : }
4629 : /* Don't know how to convert */
4630 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
4631 0 : return false;
4632 : }
4633 :
4634 : /*
4635 : * Do convert_to_scalar()'s work for any numeric data type.
4636 : *
4637 : * On failure (e.g., unsupported typid), set *failure to true;
4638 : * otherwise, that variable is not changed.
4639 : */
4640 : static double
4641 240840 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
4642 : {
4643 240840 : switch (typid)
4644 : {
4645 0 : case BOOLOID:
4646 0 : return (double) DatumGetBool(value);
4647 12 : case INT2OID:
4648 12 : return (double) DatumGetInt16(value);
4649 31146 : case INT4OID:
4650 31146 : return (double) DatumGetInt32(value);
4651 0 : case INT8OID:
4652 0 : return (double) DatumGetInt64(value);
4653 0 : case FLOAT4OID:
4654 0 : return (double) DatumGetFloat4(value);
4655 54 : case FLOAT8OID:
4656 54 : return (double) DatumGetFloat8(value);
4657 0 : case NUMERICOID:
4658 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
4659 0 : return (double)
4660 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
4661 : value));
4662 209628 : case OIDOID:
4663 : case REGPROCOID:
4664 : case REGPROCEDUREOID:
4665 : case REGOPEROID:
4666 : case REGOPERATOROID:
4667 : case REGCLASSOID:
4668 : case REGTYPEOID:
4669 : case REGCOLLATIONOID:
4670 : case REGCONFIGOID:
4671 : case REGDICTIONARYOID:
4672 : case REGROLEOID:
4673 : case REGNAMESPACEOID:
4674 : /* we can treat OIDs as integers... */
4675 209628 : return (double) DatumGetObjectId(value);
4676 : }
4677 :
4678 0 : *failure = true;
4679 0 : return 0;
4680 : }
4681 :
4682 : /*
4683 : * Do convert_to_scalar()'s work for any character-string data type.
4684 : *
4685 : * String datatypes are converted to a scale that ranges from 0 to 1,
4686 : * where we visualize the bytes of the string as fractional digits.
4687 : *
4688 : * We do not want the base to be 256, however, since that tends to
4689 : * generate inflated selectivity estimates; few databases will have
4690 : * occurrences of all 256 possible byte values at each position.
4691 : * Instead, use the smallest and largest byte values seen in the bounds
4692 : * as the estimated range for each byte, after some fudging to deal with
4693 : * the fact that we probably aren't going to see the full range that way.
4694 : *
4695 : * An additional refinement is that we discard any common prefix of the
4696 : * three strings before computing the scaled values. This allows us to
4697 : * "zoom in" when we encounter a narrow data range. An example is a phone
4698 : * number database where all the values begin with the same area code.
4699 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
4700 : * so this is more likely to happen than you might think.)
4701 : */
4702 : static void
4703 6372 : convert_string_to_scalar(char *value,
4704 : double *scaledvalue,
4705 : char *lobound,
4706 : double *scaledlobound,
4707 : char *hibound,
4708 : double *scaledhibound)
4709 : {
4710 : int rangelo,
4711 : rangehi;
4712 : char *sptr;
4713 :
4714 6372 : rangelo = rangehi = (unsigned char) hibound[0];
4715 78612 : for (sptr = lobound; *sptr; sptr++)
4716 : {
4717 72240 : if (rangelo > (unsigned char) *sptr)
4718 15088 : rangelo = (unsigned char) *sptr;
4719 72240 : if (rangehi < (unsigned char) *sptr)
4720 7920 : rangehi = (unsigned char) *sptr;
4721 : }
4722 79358 : for (sptr = hibound; *sptr; sptr++)
4723 : {
4724 72986 : if (rangelo > (unsigned char) *sptr)
4725 1184 : rangelo = (unsigned char) *sptr;
4726 72986 : if (rangehi < (unsigned char) *sptr)
4727 3138 : rangehi = (unsigned char) *sptr;
4728 : }
4729 : /* If range includes any upper-case ASCII chars, make it include all */
4730 6372 : if (rangelo <= 'Z' && rangehi >= 'A')
4731 : {
4732 1066 : if (rangelo > 'A')
4733 138 : rangelo = 'A';
4734 1066 : if (rangehi < 'Z')
4735 480 : rangehi = 'Z';
4736 : }
4737 : /* Ditto lower-case */
4738 6372 : if (rangelo <= 'z' && rangehi >= 'a')
4739 : {
4740 5870 : if (rangelo > 'a')
4741 22 : rangelo = 'a';
4742 5870 : if (rangehi < 'z')
4743 5804 : rangehi = 'z';
4744 : }
4745 : /* Ditto digits */
4746 6372 : if (rangelo <= '9' && rangehi >= '0')
4747 : {
4748 470 : if (rangelo > '0')
4749 376 : rangelo = '0';
4750 470 : if (rangehi < '9')
4751 14 : rangehi = '9';
4752 : }
4753 :
4754 : /*
4755 : * If range includes less than 10 chars, assume we have not got enough
4756 : * data, and make it include regular ASCII set.
4757 : */
4758 6372 : if (rangehi - rangelo < 9)
4759 : {
4760 0 : rangelo = ' ';
4761 0 : rangehi = 127;
4762 : }
4763 :
4764 : /*
4765 : * Now strip any common prefix of the three strings.
4766 : */
4767 12940 : while (*lobound)
4768 : {
4769 12940 : if (*lobound != *hibound || *lobound != *value)
4770 : break;
4771 6568 : lobound++, hibound++, value++;
4772 : }
4773 :
4774 : /*
4775 : * Now we can do the conversions.
4776 : */
4777 6372 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
4778 6372 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
4779 6372 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
4780 6372 : }
4781 :
4782 : static double
4783 19116 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
4784 : {
4785 19116 : int slen = strlen(value);
4786 : double num,
4787 : denom,
4788 : base;
4789 :
4790 19116 : if (slen <= 0)
4791 0 : return 0.0; /* empty string has scalar value 0 */
4792 :
4793 : /*
4794 : * There seems little point in considering more than a dozen bytes from
4795 : * the string. Since base is at least 10, that will give us nominal
4796 : * resolution of at least 12 decimal digits, which is surely far more
4797 : * precision than this estimation technique has got anyway (especially in
4798 : * non-C locales). Also, even with the maximum possible base of 256, this
4799 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
4800 : * overflow on any known machine.
4801 : */
4802 19116 : if (slen > 12)
4803 4978 : slen = 12;
4804 :
4805 : /* Convert initial characters to fraction */
4806 19116 : base = rangehi - rangelo + 1;
4807 19116 : num = 0.0;
4808 19116 : denom = base;
4809 159222 : while (slen-- > 0)
4810 : {
4811 140106 : int ch = (unsigned char) *value++;
4812 :
4813 140106 : if (ch < rangelo)
4814 208 : ch = rangelo - 1;
4815 139898 : else if (ch > rangehi)
4816 0 : ch = rangehi + 1;
4817 140106 : num += ((double) (ch - rangelo)) / denom;
4818 140106 : denom *= base;
4819 : }
4820 :
4821 19116 : return num;
4822 : }
4823 :
4824 : /*
4825 : * Convert a string-type Datum into a palloc'd, null-terminated string.
4826 : *
4827 : * On failure (e.g., unsupported typid), set *failure to true;
4828 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
4829 : *
4830 : * When using a non-C locale, we must pass the string through pg_strxfrm()
4831 : * before continuing, so as to generate correct locale-specific results.
4832 : */
4833 : static char *
4834 19116 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
4835 : {
4836 : char *val;
4837 : pg_locale_t mylocale;
4838 :
4839 19116 : switch (typid)
4840 : {
4841 0 : case CHAROID:
4842 0 : val = (char *) palloc(2);
4843 0 : val[0] = DatumGetChar(value);
4844 0 : val[1] = '\0';
4845 0 : break;
4846 5920 : case BPCHAROID:
4847 : case VARCHAROID:
4848 : case TEXTOID:
4849 5920 : val = TextDatumGetCString(value);
4850 5920 : break;
4851 13196 : case NAMEOID:
4852 : {
4853 13196 : NameData *nm = (NameData *) DatumGetPointer(value);
4854 :
4855 13196 : val = pstrdup(NameStr(*nm));
4856 13196 : break;
4857 : }
4858 0 : default:
4859 0 : *failure = true;
4860 0 : return NULL;
4861 : }
4862 :
4863 19116 : mylocale = pg_newlocale_from_collation(collid);
4864 :
4865 19116 : if (!mylocale->collate_is_c)
4866 : {
4867 : char *xfrmstr;
4868 : size_t xfrmlen;
4869 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
4870 :
4871 : /*
4872 : * XXX: We could guess at a suitable output buffer size and only call
4873 : * pg_strxfrm() twice if our guess is too small.
4874 : *
4875 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
4876 : * bogus data or set an error. This is not really a problem unless it
4877 : * crashes since it will only give an estimation error and nothing
4878 : * fatal.
4879 : *
4880 : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
4881 : * some cases, libc strxfrm() may return the wrong results, but that
4882 : * will only lead to an estimation error.
4883 : */
4884 72 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
4885 : #ifdef WIN32
4886 :
4887 : /*
4888 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
4889 : * of trying to allocate this much memory (and fail), just return the
4890 : * original string unmodified as if we were in the C locale.
4891 : */
4892 : if (xfrmlen == INT_MAX)
4893 : return val;
4894 : #endif
4895 72 : xfrmstr = (char *) palloc(xfrmlen + 1);
4896 72 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
4897 :
4898 : /*
4899 : * Some systems (e.g., glibc) can return a smaller value from the
4900 : * second call than the first; thus the Assert must be <= not ==.
4901 : */
4902 : Assert(xfrmlen2 <= xfrmlen);
4903 72 : pfree(val);
4904 72 : val = xfrmstr;
4905 : }
4906 :
4907 19116 : return val;
4908 : }
4909 :
4910 : /*
4911 : * Do convert_to_scalar()'s work for any bytea data type.
4912 : *
4913 : * Very similar to convert_string_to_scalar except we can't assume
4914 : * null-termination and therefore pass explicit lengths around.
4915 : *
4916 : * Also, assumptions about likely "normal" ranges of characters have been
4917 : * removed - a data range of 0..255 is always used, for now. (Perhaps
4918 : * someday we will add information about actual byte data range to
4919 : * pg_statistic.)
4920 : */
4921 : static void
4922 0 : convert_bytea_to_scalar(Datum value,
4923 : double *scaledvalue,
4924 : Datum lobound,
4925 : double *scaledlobound,
4926 : Datum hibound,
4927 : double *scaledhibound)
4928 : {
4929 0 : bytea *valuep = DatumGetByteaPP(value);
4930 0 : bytea *loboundp = DatumGetByteaPP(lobound);
4931 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
4932 : int rangelo,
4933 : rangehi,
4934 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
4935 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
4936 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
4937 : i,
4938 : minlen;
4939 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
4940 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
4941 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
4942 :
4943 : /*
4944 : * Assume bytea data is uniformly distributed across all byte values.
4945 : */
4946 0 : rangelo = 0;
4947 0 : rangehi = 255;
4948 :
4949 : /*
4950 : * Now strip any common prefix of the three strings.
4951 : */
4952 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
4953 0 : for (i = 0; i < minlen; i++)
4954 : {
4955 0 : if (*lostr != *histr || *lostr != *valstr)
4956 : break;
4957 0 : lostr++, histr++, valstr++;
4958 0 : loboundlen--, hiboundlen--, valuelen--;
4959 : }
4960 :
4961 : /*
4962 : * Now we can do the conversions.
4963 : */
4964 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
4965 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
4966 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
4967 0 : }
4968 :
4969 : static double
4970 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
4971 : int rangelo, int rangehi)
4972 : {
4973 : double num,
4974 : denom,
4975 : base;
4976 :
4977 0 : if (valuelen <= 0)
4978 0 : return 0.0; /* empty string has scalar value 0 */
4979 :
4980 : /*
4981 : * Since base is 256, need not consider more than about 10 chars (even
4982 : * this many seems like overkill)
4983 : */
4984 0 : if (valuelen > 10)
4985 0 : valuelen = 10;
4986 :
4987 : /* Convert initial characters to fraction */
4988 0 : base = rangehi - rangelo + 1;
4989 0 : num = 0.0;
4990 0 : denom = base;
4991 0 : while (valuelen-- > 0)
4992 : {
4993 0 : int ch = *value++;
4994 :
4995 0 : if (ch < rangelo)
4996 0 : ch = rangelo - 1;
4997 0 : else if (ch > rangehi)
4998 0 : ch = rangehi + 1;
4999 0 : num += ((double) (ch - rangelo)) / denom;
5000 0 : denom *= base;
5001 : }
5002 :
5003 0 : return num;
5004 : }
5005 :
5006 : /*
5007 : * Do convert_to_scalar()'s work for any timevalue data type.
5008 : *
5009 : * On failure (e.g., unsupported typid), set *failure to true;
5010 : * otherwise, that variable is not changed.
5011 : */
5012 : static double
5013 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5014 : {
5015 0 : switch (typid)
5016 : {
5017 0 : case TIMESTAMPOID:
5018 0 : return DatumGetTimestamp(value);
5019 0 : case TIMESTAMPTZOID:
5020 0 : return DatumGetTimestampTz(value);
5021 0 : case DATEOID:
5022 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
5023 0 : case INTERVALOID:
5024 : {
5025 0 : Interval *interval = DatumGetIntervalP(value);
5026 :
5027 : /*
5028 : * Convert the month part of Interval to days using assumed
5029 : * average month length of 365.25/12.0 days. Not too
5030 : * accurate, but plenty good enough for our purposes.
5031 : *
5032 : * This also works for infinite intervals, which just have all
5033 : * fields set to INT_MIN/INT_MAX, and so will produce a result
5034 : * smaller/larger than any finite interval.
5035 : */
5036 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5037 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5038 : }
5039 0 : case TIMEOID:
5040 0 : return DatumGetTimeADT(value);
5041 0 : case TIMETZOID:
5042 : {
5043 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5044 :
5045 : /* use GMT-equivalent time */
5046 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5047 : }
5048 : }
5049 :
5050 0 : *failure = true;
5051 0 : return 0;
5052 : }
5053 :
5054 :
5055 : /*
5056 : * get_restriction_variable
5057 : * Examine the args of a restriction clause to see if it's of the
5058 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5059 : * where "variable" could be either a Var or an expression in vars of a
5060 : * single relation. If so, extract information about the variable,
5061 : * and also indicate which side it was on and the other argument.
5062 : *
5063 : * Inputs:
5064 : * root: the planner info
5065 : * args: clause argument list
5066 : * varRelid: see specs for restriction selectivity functions
5067 : *
5068 : * Outputs: (these are valid only if true is returned)
5069 : * *vardata: gets information about variable (see examine_variable)
5070 : * *other: gets other clause argument, aggressively reduced to a constant
5071 : * *varonleft: set true if variable is on the left, false if on the right
5072 : *
5073 : * Returns true if a variable is identified, otherwise false.
5074 : *
5075 : * Note: if there are Vars on both sides of the clause, we must fail, because
5076 : * callers are expecting that the other side will act like a pseudoconstant.
5077 : */
5078 : bool
5079 732560 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5080 : VariableStatData *vardata, Node **other,
5081 : bool *varonleft)
5082 : {
5083 : Node *left,
5084 : *right;
5085 : VariableStatData rdata;
5086 :
5087 : /* Fail if not a binary opclause (probably shouldn't happen) */
5088 732560 : if (list_length(args) != 2)
5089 0 : return false;
5090 :
5091 732560 : left = (Node *) linitial(args);
5092 732560 : right = (Node *) lsecond(args);
5093 :
5094 : /*
5095 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5096 : * relations will be treated as pseudoconstants.
5097 : */
5098 732560 : examine_variable(root, left, varRelid, vardata);
5099 732560 : examine_variable(root, right, varRelid, &rdata);
5100 :
5101 : /*
5102 : * If one side is a variable and the other not, we win.
5103 : */
5104 732560 : if (vardata->rel && rdata.rel == NULL)
5105 : {
5106 655958 : *varonleft = true;
5107 655958 : *other = estimate_expression_value(root, rdata.var);
5108 : /* Assume we need no ReleaseVariableStats(rdata) here */
5109 655958 : return true;
5110 : }
5111 :
5112 76602 : if (vardata->rel == NULL && rdata.rel)
5113 : {
5114 70702 : *varonleft = false;
5115 70702 : *other = estimate_expression_value(root, vardata->var);
5116 : /* Assume we need no ReleaseVariableStats(*vardata) here */
5117 70702 : *vardata = rdata;
5118 70702 : return true;
5119 : }
5120 :
5121 : /* Oops, clause has wrong structure (probably var op var) */
5122 5900 : ReleaseVariableStats(*vardata);
5123 5900 : ReleaseVariableStats(rdata);
5124 :
5125 5900 : return false;
5126 : }
5127 :
5128 : /*
5129 : * get_join_variables
5130 : * Apply examine_variable() to each side of a join clause.
5131 : * Also, attempt to identify whether the join clause has the same
5132 : * or reversed sense compared to the SpecialJoinInfo.
5133 : *
5134 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5135 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5136 : * where we can't tell for sure, we default to assuming it's normal.
5137 : */
5138 : void
5139 220064 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5140 : VariableStatData *vardata1, VariableStatData *vardata2,
5141 : bool *join_is_reversed)
5142 : {
5143 : Node *left,
5144 : *right;
5145 :
5146 220064 : if (list_length(args) != 2)
5147 0 : elog(ERROR, "join operator should take two arguments");
5148 :
5149 220064 : left = (Node *) linitial(args);
5150 220064 : right = (Node *) lsecond(args);
5151 :
5152 220064 : examine_variable(root, left, 0, vardata1);
5153 220064 : examine_variable(root, right, 0, vardata2);
5154 :
5155 439952 : if (vardata1->rel &&
5156 219888 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
5157 78504 : *join_is_reversed = true; /* var1 is on RHS */
5158 282980 : else if (vardata2->rel &&
5159 141420 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
5160 126 : *join_is_reversed = true; /* var2 is on LHS */
5161 : else
5162 141434 : *join_is_reversed = false;
5163 220064 : }
5164 :
5165 : /* statext_expressions_load copies the tuple, so just pfree it. */
5166 : static void
5167 1644 : ReleaseDummy(HeapTuple tuple)
5168 : {
5169 1644 : pfree(tuple);
5170 1644 : }
5171 :
5172 : /*
5173 : * examine_variable
5174 : * Try to look up statistical data about an expression.
5175 : * Fill in a VariableStatData struct to describe the expression.
5176 : *
5177 : * Inputs:
5178 : * root: the planner info
5179 : * node: the expression tree to examine
5180 : * varRelid: see specs for restriction selectivity functions
5181 : *
5182 : * Outputs: *vardata is filled as follows:
5183 : * var: the input expression (with any binary relabeling stripped, if
5184 : * it is or contains a variable; but otherwise the type is preserved)
5185 : * rel: RelOptInfo for relation containing variable; NULL if expression
5186 : * contains no Vars (NOTE this could point to a RelOptInfo of a
5187 : * subquery, not one in the current query).
5188 : * statsTuple: the pg_statistic entry for the variable, if one exists;
5189 : * otherwise NULL.
5190 : * freefunc: pointer to a function to release statsTuple with.
5191 : * vartype: exposed type of the expression; this should always match
5192 : * the declared input type of the operator we are estimating for.
5193 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5194 : * commonly the same as the exposed type of the variable argument,
5195 : * but can be different in binary-compatible-type cases.
5196 : * isunique: true if we were able to match the var to a unique index, a
5197 : * single-column DISTINCT or GROUP-BY clause, implying its values are
5198 : * unique for this query. (Caution: this should be trusted for
5199 : * statistical purposes only, since we do not check indimmediate nor
5200 : * verify that the exact same definition of equality applies.)
5201 : * acl_ok: true if current user has permission to read the column(s)
5202 : * underlying the pg_statistic entry. This is consulted by
5203 : * statistic_proc_security_check().
5204 : *
5205 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5206 : */
5207 : void
5208 2822758 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5209 : VariableStatData *vardata)
5210 : {
5211 : Node *basenode;
5212 : Relids varnos;
5213 : Relids basevarnos;
5214 : RelOptInfo *onerel;
5215 :
5216 : /* Make sure we don't return dangling pointers in vardata */
5217 19759306 : MemSet(vardata, 0, sizeof(VariableStatData));
5218 :
5219 : /* Save the exposed type of the expression */
5220 2822758 : vardata->vartype = exprType(node);
5221 :
5222 : /* Look inside any binary-compatible relabeling */
5223 :
5224 2822758 : if (IsA(node, RelabelType))
5225 24936 : basenode = (Node *) ((RelabelType *) node)->arg;
5226 : else
5227 2797822 : basenode = node;
5228 :
5229 : /* Fast path for a simple Var */
5230 :
5231 2822758 : if (IsA(basenode, Var) &&
5232 687558 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5233 : {
5234 1992172 : Var *var = (Var *) basenode;
5235 :
5236 : /* Set up result fields other than the stats tuple */
5237 1992172 : vardata->var = basenode; /* return Var without relabeling */
5238 1992172 : vardata->rel = find_base_rel(root, var->varno);
5239 1992172 : vardata->atttype = var->vartype;
5240 1992172 : vardata->atttypmod = var->vartypmod;
5241 1992172 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5242 :
5243 : /* Try to locate some stats */
5244 1992172 : examine_simple_variable(root, var, vardata);
5245 :
5246 1992172 : return;
5247 : }
5248 :
5249 : /*
5250 : * Okay, it's a more complicated expression. Determine variable
5251 : * membership. Note that when varRelid isn't zero, only vars of that
5252 : * relation are considered "real" vars.
5253 : */
5254 830586 : varnos = pull_varnos(root, basenode);
5255 830586 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5256 :
5257 830586 : onerel = NULL;
5258 :
5259 830586 : if (bms_is_empty(basevarnos))
5260 : {
5261 : /* No Vars at all ... must be pseudo-constant clause */
5262 : }
5263 : else
5264 : {
5265 : int relid;
5266 :
5267 : /* Check if the expression is in vars of a single base relation */
5268 406986 : if (bms_get_singleton_member(basevarnos, &relid))
5269 : {
5270 402998 : if (varRelid == 0 || varRelid == relid)
5271 : {
5272 57986 : onerel = find_base_rel(root, relid);
5273 57986 : vardata->rel = onerel;
5274 57986 : node = basenode; /* strip any relabeling */
5275 : }
5276 : /* else treat it as a constant */
5277 : }
5278 : else
5279 : {
5280 : /* varnos has multiple relids */
5281 3988 : if (varRelid == 0)
5282 : {
5283 : /* treat it as a variable of a join relation */
5284 3670 : vardata->rel = find_join_rel(root, varnos);
5285 3670 : node = basenode; /* strip any relabeling */
5286 : }
5287 318 : else if (bms_is_member(varRelid, varnos))
5288 : {
5289 : /* ignore the vars belonging to other relations */
5290 156 : vardata->rel = find_base_rel(root, varRelid);
5291 156 : node = basenode; /* strip any relabeling */
5292 : /* note: no point in expressional-index search here */
5293 : }
5294 : /* else treat it as a constant */
5295 : }
5296 : }
5297 :
5298 830586 : bms_free(basevarnos);
5299 :
5300 830586 : vardata->var = node;
5301 830586 : vardata->atttype = exprType(node);
5302 830586 : vardata->atttypmod = exprTypmod(node);
5303 :
5304 830586 : if (onerel)
5305 : {
5306 : /*
5307 : * We have an expression in vars of a single relation. Try to match
5308 : * it to expressional index columns, in hopes of finding some
5309 : * statistics.
5310 : *
5311 : * Note that we consider all index columns including INCLUDE columns,
5312 : * since there could be stats for such columns. But the test for
5313 : * uniqueness needs to be warier.
5314 : *
5315 : * XXX it's conceivable that there are multiple matches with different
5316 : * index opfamilies; if so, we need to pick one that matches the
5317 : * operator we are estimating for. FIXME later.
5318 : */
5319 : ListCell *ilist;
5320 : ListCell *slist;
5321 : Oid userid;
5322 :
5323 : /*
5324 : * The nullingrels bits within the expression could prevent us from
5325 : * matching it to expressional index columns or to the expressions in
5326 : * extended statistics. So strip them out first.
5327 : */
5328 57986 : if (bms_overlap(varnos, root->outer_join_rels))
5329 3036 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5330 :
5331 : /*
5332 : * Determine the user ID to use for privilege checks: either
5333 : * onerel->userid if it's set (e.g., in case we're accessing the table
5334 : * via a view), or the current user otherwise.
5335 : *
5336 : * If we drill down to child relations, we keep using the same userid:
5337 : * it's going to be the same anyway, due to how we set up the relation
5338 : * tree (q.v. build_simple_rel).
5339 : */
5340 57986 : userid = OidIsValid(onerel->userid) ? onerel->userid : GetUserId();
5341 :
5342 115852 : foreach(ilist, onerel->indexlist)
5343 : {
5344 60680 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5345 : ListCell *indexpr_item;
5346 : int pos;
5347 :
5348 60680 : indexpr_item = list_head(index->indexprs);
5349 60680 : if (indexpr_item == NULL)
5350 55934 : continue; /* no expressions here... */
5351 :
5352 6750 : for (pos = 0; pos < index->ncolumns; pos++)
5353 : {
5354 4818 : if (index->indexkeys[pos] == 0)
5355 : {
5356 : Node *indexkey;
5357 :
5358 4746 : if (indexpr_item == NULL)
5359 0 : elog(ERROR, "too few entries in indexprs list");
5360 4746 : indexkey = (Node *) lfirst(indexpr_item);
5361 4746 : if (indexkey && IsA(indexkey, RelabelType))
5362 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5363 4746 : if (equal(node, indexkey))
5364 : {
5365 : /*
5366 : * Found a match ... is it a unique index? Tests here
5367 : * should match has_unique_index().
5368 : */
5369 3450 : if (index->unique &&
5370 438 : index->nkeycolumns == 1 &&
5371 438 : pos == 0 &&
5372 438 : (index->indpred == NIL || index->predOK))
5373 438 : vardata->isunique = true;
5374 :
5375 : /*
5376 : * Has it got stats? We only consider stats for
5377 : * non-partial indexes, since partial indexes probably
5378 : * don't reflect whole-relation statistics; the above
5379 : * check for uniqueness is the only info we take from
5380 : * a partial index.
5381 : *
5382 : * An index stats hook, however, must make its own
5383 : * decisions about what to do with partial indexes.
5384 : */
5385 3450 : if (get_index_stats_hook &&
5386 0 : (*get_index_stats_hook) (root, index->indexoid,
5387 0 : pos + 1, vardata))
5388 : {
5389 : /*
5390 : * The hook took control of acquiring a stats
5391 : * tuple. If it did supply a tuple, it'd better
5392 : * have supplied a freefunc.
5393 : */
5394 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5395 0 : !vardata->freefunc)
5396 0 : elog(ERROR, "no function provided to release variable stats with");
5397 : }
5398 3450 : else if (index->indpred == NIL)
5399 : {
5400 3450 : vardata->statsTuple =
5401 6900 : SearchSysCache3(STATRELATTINH,
5402 : ObjectIdGetDatum(index->indexoid),
5403 3450 : Int16GetDatum(pos + 1),
5404 : BoolGetDatum(false));
5405 3450 : vardata->freefunc = ReleaseSysCache;
5406 :
5407 3450 : if (HeapTupleIsValid(vardata->statsTuple))
5408 : {
5409 : /* Get index's table for permission check */
5410 : RangeTblEntry *rte;
5411 :
5412 2814 : rte = planner_rt_fetch(index->rel->relid, root);
5413 : Assert(rte->rtekind == RTE_RELATION);
5414 :
5415 : /*
5416 : * For simplicity, we insist on the whole
5417 : * table being selectable, rather than trying
5418 : * to identify which column(s) the index
5419 : * depends on. Also require all rows to be
5420 : * selectable --- there must be no
5421 : * securityQuals from security barrier views
5422 : * or RLS policies.
5423 : */
5424 2814 : vardata->acl_ok =
5425 5628 : rte->securityQuals == NIL &&
5426 2814 : (pg_class_aclcheck(rte->relid, userid,
5427 : ACL_SELECT) == ACLCHECK_OK);
5428 :
5429 : /*
5430 : * If the user doesn't have permissions to
5431 : * access an inheritance child relation, check
5432 : * the permissions of the table actually
5433 : * mentioned in the query, since most likely
5434 : * the user does have that permission. Note
5435 : * that whole-table select privilege on the
5436 : * parent doesn't quite guarantee that the
5437 : * user could read all columns of the child.
5438 : * But in practice it's unlikely that any
5439 : * interesting security violation could result
5440 : * from allowing access to the expression
5441 : * index's stats, so we allow it anyway. See
5442 : * similar code in examine_simple_variable()
5443 : * for additional comments.
5444 : */
5445 2814 : if (!vardata->acl_ok &&
5446 18 : root->append_rel_array != NULL)
5447 : {
5448 : AppendRelInfo *appinfo;
5449 12 : Index varno = index->rel->relid;
5450 :
5451 12 : appinfo = root->append_rel_array[varno];
5452 36 : while (appinfo &&
5453 24 : planner_rt_fetch(appinfo->parent_relid,
5454 24 : root)->rtekind == RTE_RELATION)
5455 : {
5456 24 : varno = appinfo->parent_relid;
5457 24 : appinfo = root->append_rel_array[varno];
5458 : }
5459 12 : if (varno != index->rel->relid)
5460 : {
5461 : /* Repeat access check on this rel */
5462 12 : rte = planner_rt_fetch(varno, root);
5463 : Assert(rte->rtekind == RTE_RELATION);
5464 :
5465 12 : vardata->acl_ok =
5466 24 : rte->securityQuals == NIL &&
5467 12 : (pg_class_aclcheck(rte->relid,
5468 : userid,
5469 : ACL_SELECT) == ACLCHECK_OK);
5470 : }
5471 : }
5472 : }
5473 : else
5474 : {
5475 : /* suppress leakproofness checks later */
5476 636 : vardata->acl_ok = true;
5477 : }
5478 : }
5479 3450 : if (vardata->statsTuple)
5480 2814 : break;
5481 : }
5482 1932 : indexpr_item = lnext(index->indexprs, indexpr_item);
5483 : }
5484 : }
5485 4746 : if (vardata->statsTuple)
5486 2814 : break;
5487 : }
5488 :
5489 : /*
5490 : * Search extended statistics for one with a matching expression.
5491 : * There might be multiple ones, so just grab the first one. In the
5492 : * future, we might consider the statistics target (and pick the most
5493 : * accurate statistics) and maybe some other parameters.
5494 : */
5495 62012 : foreach(slist, onerel->statlist)
5496 : {
5497 4314 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5498 4314 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5499 : ListCell *expr_item;
5500 : int pos;
5501 :
5502 : /*
5503 : * Stop once we've found statistics for the expression (either
5504 : * from extended stats, or for an index in the preceding loop).
5505 : */
5506 4314 : if (vardata->statsTuple)
5507 288 : break;
5508 :
5509 : /* skip stats without per-expression stats */
5510 4026 : if (info->kind != STATS_EXT_EXPRESSIONS)
5511 2022 : continue;
5512 :
5513 : /* skip stats with mismatching stxdinherit value */
5514 2004 : if (info->inherit != rte->inh)
5515 6 : continue;
5516 :
5517 1998 : pos = 0;
5518 3300 : foreach(expr_item, info->exprs)
5519 : {
5520 2946 : Node *expr = (Node *) lfirst(expr_item);
5521 :
5522 : Assert(expr);
5523 :
5524 : /* strip RelabelType before comparing it */
5525 2946 : if (expr && IsA(expr, RelabelType))
5526 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5527 :
5528 : /* found a match, see if we can extract pg_statistic row */
5529 2946 : if (equal(node, expr))
5530 : {
5531 : /*
5532 : * XXX Not sure if we should cache the tuple somewhere.
5533 : * Now we just create a new copy every time.
5534 : */
5535 1644 : vardata->statsTuple =
5536 1644 : statext_expressions_load(info->statOid, rte->inh, pos);
5537 :
5538 1644 : vardata->freefunc = ReleaseDummy;
5539 :
5540 : /*
5541 : * For simplicity, we insist on the whole table being
5542 : * selectable, rather than trying to identify which
5543 : * column(s) the statistics object depends on. Also
5544 : * require all rows to be selectable --- there must be no
5545 : * securityQuals from security barrier views or RLS
5546 : * policies.
5547 : */
5548 1644 : vardata->acl_ok =
5549 3288 : rte->securityQuals == NIL &&
5550 1644 : (pg_class_aclcheck(rte->relid, userid,
5551 : ACL_SELECT) == ACLCHECK_OK);
5552 :
5553 : /*
5554 : * If the user doesn't have permissions to access an
5555 : * inheritance child relation, check the permissions of
5556 : * the table actually mentioned in the query, since most
5557 : * likely the user does have that permission. Note that
5558 : * whole-table select privilege on the parent doesn't
5559 : * quite guarantee that the user could read all columns of
5560 : * the child. But in practice it's unlikely that any
5561 : * interesting security violation could result from
5562 : * allowing access to the expression stats, so we allow it
5563 : * anyway. See similar code in examine_simple_variable()
5564 : * for additional comments.
5565 : */
5566 1644 : if (!vardata->acl_ok &&
5567 0 : root->append_rel_array != NULL)
5568 : {
5569 : AppendRelInfo *appinfo;
5570 0 : Index varno = onerel->relid;
5571 :
5572 0 : appinfo = root->append_rel_array[varno];
5573 0 : while (appinfo &&
5574 0 : planner_rt_fetch(appinfo->parent_relid,
5575 0 : root)->rtekind == RTE_RELATION)
5576 : {
5577 0 : varno = appinfo->parent_relid;
5578 0 : appinfo = root->append_rel_array[varno];
5579 : }
5580 0 : if (varno != onerel->relid)
5581 : {
5582 : /* Repeat access check on this rel */
5583 0 : rte = planner_rt_fetch(varno, root);
5584 : Assert(rte->rtekind == RTE_RELATION);
5585 :
5586 0 : vardata->acl_ok =
5587 0 : rte->securityQuals == NIL &&
5588 0 : (pg_class_aclcheck(rte->relid,
5589 : userid,
5590 : ACL_SELECT) == ACLCHECK_OK);
5591 : }
5592 : }
5593 :
5594 1644 : break;
5595 : }
5596 :
5597 1302 : pos++;
5598 : }
5599 : }
5600 : }
5601 :
5602 830586 : bms_free(varnos);
5603 : }
5604 :
5605 : /*
5606 : * examine_simple_variable
5607 : * Handle a simple Var for examine_variable
5608 : *
5609 : * This is split out as a subroutine so that we can recurse to deal with
5610 : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
5611 : *
5612 : * We already filled in all the fields of *vardata except for the stats tuple.
5613 : */
5614 : static void
5615 1997692 : examine_simple_variable(PlannerInfo *root, Var *var,
5616 : VariableStatData *vardata)
5617 : {
5618 1997692 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
5619 :
5620 : Assert(IsA(rte, RangeTblEntry));
5621 :
5622 1997692 : if (get_relation_stats_hook &&
5623 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
5624 : {
5625 : /*
5626 : * The hook took control of acquiring a stats tuple. If it did supply
5627 : * a tuple, it'd better have supplied a freefunc.
5628 : */
5629 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5630 0 : !vardata->freefunc)
5631 0 : elog(ERROR, "no function provided to release variable stats with");
5632 : }
5633 1997692 : else if (rte->rtekind == RTE_RELATION)
5634 : {
5635 : /*
5636 : * Plain table or parent of an inheritance appendrel, so look up the
5637 : * column in pg_statistic
5638 : */
5639 1920558 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
5640 : ObjectIdGetDatum(rte->relid),
5641 1920558 : Int16GetDatum(var->varattno),
5642 1920558 : BoolGetDatum(rte->inh));
5643 1920558 : vardata->freefunc = ReleaseSysCache;
5644 :
5645 1920558 : if (HeapTupleIsValid(vardata->statsTuple))
5646 : {
5647 1423258 : RelOptInfo *onerel = find_base_rel_noerr(root, var->varno);
5648 : Oid userid;
5649 :
5650 : /*
5651 : * Check if user has permission to read this column. We require
5652 : * all rows to be accessible, so there must be no securityQuals
5653 : * from security barrier views or RLS policies.
5654 : *
5655 : * Normally the Var will have an associated RelOptInfo from which
5656 : * we can find out which userid to do the check as; but it might
5657 : * not if it's a RETURNING Var for an INSERT target relation. In
5658 : * that case use the RTEPermissionInfo associated with the RTE.
5659 : */
5660 1423258 : if (onerel)
5661 1423216 : userid = onerel->userid;
5662 : else
5663 : {
5664 : RTEPermissionInfo *perminfo;
5665 :
5666 42 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
5667 42 : userid = perminfo->checkAsUser;
5668 : }
5669 1423258 : if (!OidIsValid(userid))
5670 1278254 : userid = GetUserId();
5671 :
5672 1423258 : vardata->acl_ok =
5673 2846776 : rte->securityQuals == NIL &&
5674 1423258 : ((pg_class_aclcheck(rte->relid, userid,
5675 260 : ACL_SELECT) == ACLCHECK_OK) ||
5676 260 : (pg_attribute_aclcheck(rte->relid, var->varattno, userid,
5677 : ACL_SELECT) == ACLCHECK_OK));
5678 :
5679 : /*
5680 : * If the user doesn't have permissions to access an inheritance
5681 : * child relation or specifically this attribute, check the
5682 : * permissions of the table/column actually mentioned in the
5683 : * query, since most likely the user does have that permission
5684 : * (else the query will fail at runtime), and if the user can read
5685 : * the column there then he can get the values of the child table
5686 : * too. To do that, we must find out which of the root parent's
5687 : * attributes the child relation's attribute corresponds to.
5688 : */
5689 1423258 : if (!vardata->acl_ok && var->varattno > 0 &&
5690 60 : root->append_rel_array != NULL)
5691 : {
5692 : AppendRelInfo *appinfo;
5693 12 : Index varno = var->varno;
5694 12 : int varattno = var->varattno;
5695 12 : bool found = false;
5696 :
5697 12 : appinfo = root->append_rel_array[varno];
5698 :
5699 : /*
5700 : * Partitions are mapped to their immediate parent, not the
5701 : * root parent, so must be ready to walk up multiple
5702 : * AppendRelInfos. But stop if we hit a parent that is not
5703 : * RTE_RELATION --- that's a flattened UNION ALL subquery, not
5704 : * an inheritance parent.
5705 : */
5706 36 : while (appinfo &&
5707 24 : planner_rt_fetch(appinfo->parent_relid,
5708 24 : root)->rtekind == RTE_RELATION)
5709 : {
5710 : int parent_varattno;
5711 :
5712 24 : found = false;
5713 24 : if (varattno <= 0 || varattno > appinfo->num_child_cols)
5714 : break; /* safety check */
5715 24 : parent_varattno = appinfo->parent_colnos[varattno - 1];
5716 24 : if (parent_varattno == 0)
5717 0 : break; /* Var is local to child */
5718 :
5719 24 : varno = appinfo->parent_relid;
5720 24 : varattno = parent_varattno;
5721 24 : found = true;
5722 :
5723 : /* If the parent is itself a child, continue up. */
5724 24 : appinfo = root->append_rel_array[varno];
5725 : }
5726 :
5727 : /*
5728 : * In rare cases, the Var may be local to the child table, in
5729 : * which case, we've got to live with having no access to this
5730 : * column's stats.
5731 : */
5732 12 : if (!found)
5733 0 : return;
5734 :
5735 : /* Repeat the access check on this parent rel & column */
5736 12 : rte = planner_rt_fetch(varno, root);
5737 : Assert(rte->rtekind == RTE_RELATION);
5738 :
5739 : /*
5740 : * Fine to use the same userid as it's the same in all
5741 : * relations of a given inheritance tree.
5742 : */
5743 12 : vardata->acl_ok =
5744 30 : rte->securityQuals == NIL &&
5745 12 : ((pg_class_aclcheck(rte->relid, userid,
5746 6 : ACL_SELECT) == ACLCHECK_OK) ||
5747 6 : (pg_attribute_aclcheck(rte->relid, varattno, userid,
5748 : ACL_SELECT) == ACLCHECK_OK));
5749 : }
5750 : }
5751 : else
5752 : {
5753 : /* suppress any possible leakproofness checks later */
5754 497300 : vardata->acl_ok = true;
5755 : }
5756 : }
5757 77134 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
5758 69102 : (rte->rtekind == RTE_CTE && !rte->self_reference))
5759 : {
5760 : /*
5761 : * Plain subquery (not one that was converted to an appendrel) or
5762 : * non-recursive CTE. In either case, we can try to find out what the
5763 : * Var refers to within the subquery. We skip this for appendrel and
5764 : * recursive-CTE cases because any column stats we did find would
5765 : * likely not be very relevant.
5766 : */
5767 : PlannerInfo *subroot;
5768 : Query *subquery;
5769 : List *subtlist;
5770 : TargetEntry *ste;
5771 :
5772 : /*
5773 : * Punt if it's a whole-row var rather than a plain column reference.
5774 : */
5775 14426 : if (var->varattno == InvalidAttrNumber)
5776 0 : return;
5777 :
5778 : /*
5779 : * Otherwise, find the subquery's planner subroot.
5780 : */
5781 14426 : if (rte->rtekind == RTE_SUBQUERY)
5782 : {
5783 : RelOptInfo *rel;
5784 :
5785 : /*
5786 : * Fetch RelOptInfo for subquery. Note that we don't change the
5787 : * rel returned in vardata, since caller expects it to be a rel of
5788 : * the caller's query level. Because we might already be
5789 : * recursing, we can't use that rel pointer either, but have to
5790 : * look up the Var's rel afresh.
5791 : */
5792 8032 : rel = find_base_rel(root, var->varno);
5793 :
5794 8032 : subroot = rel->subroot;
5795 : }
5796 : else
5797 : {
5798 : /* CTE case is more difficult */
5799 : PlannerInfo *cteroot;
5800 : Index levelsup;
5801 : int ndx;
5802 : int plan_id;
5803 : ListCell *lc;
5804 :
5805 : /*
5806 : * Find the referenced CTE, and locate the subroot previously made
5807 : * for it.
5808 : */
5809 6394 : levelsup = rte->ctelevelsup;
5810 6394 : cteroot = root;
5811 12050 : while (levelsup-- > 0)
5812 : {
5813 5656 : cteroot = cteroot->parent_root;
5814 5656 : if (!cteroot) /* shouldn't happen */
5815 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
5816 : }
5817 :
5818 : /*
5819 : * Note: cte_plan_ids can be shorter than cteList, if we are still
5820 : * working on planning the CTEs (ie, this is a side-reference from
5821 : * another CTE). So we mustn't use forboth here.
5822 : */
5823 6394 : ndx = 0;
5824 8624 : foreach(lc, cteroot->parse->cteList)
5825 : {
5826 8624 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
5827 :
5828 8624 : if (strcmp(cte->ctename, rte->ctename) == 0)
5829 6394 : break;
5830 2230 : ndx++;
5831 : }
5832 6394 : if (lc == NULL) /* shouldn't happen */
5833 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
5834 6394 : if (ndx >= list_length(cteroot->cte_plan_ids))
5835 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
5836 6394 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
5837 6394 : if (plan_id <= 0)
5838 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
5839 6394 : subroot = list_nth(root->glob->subroots, plan_id - 1);
5840 : }
5841 :
5842 : /* If the subquery hasn't been planned yet, we have to punt */
5843 14426 : if (subroot == NULL)
5844 0 : return;
5845 : Assert(IsA(subroot, PlannerInfo));
5846 :
5847 : /*
5848 : * We must use the subquery parsetree as mangled by the planner, not
5849 : * the raw version from the RTE, because we need a Var that will refer
5850 : * to the subroot's live RelOptInfos. For instance, if any subquery
5851 : * pullup happened during planning, Vars in the targetlist might have
5852 : * gotten replaced, and we need to see the replacement expressions.
5853 : */
5854 14426 : subquery = subroot->parse;
5855 : Assert(IsA(subquery, Query));
5856 :
5857 : /*
5858 : * Punt if subquery uses set operations or grouping sets, as these
5859 : * will mash underlying columns' stats beyond recognition. (Set ops
5860 : * are particularly nasty; if we forged ahead, we would return stats
5861 : * relevant to only the leftmost subselect...) DISTINCT is also
5862 : * problematic, but we check that later because there is a possibility
5863 : * of learning something even with it.
5864 : */
5865 14426 : if (subquery->setOperations ||
5866 12732 : subquery->groupingSets)
5867 1718 : return;
5868 :
5869 : /* Get the subquery output expression referenced by the upper Var */
5870 12708 : if (subquery->returningList)
5871 206 : subtlist = subquery->returningList;
5872 : else
5873 12502 : subtlist = subquery->targetList;
5874 12708 : ste = get_tle_by_resno(subtlist, var->varattno);
5875 12708 : if (ste == NULL || ste->resjunk)
5876 0 : elog(ERROR, "subquery %s does not have attribute %d",
5877 : rte->eref->aliasname, var->varattno);
5878 12708 : var = (Var *) ste->expr;
5879 :
5880 : /*
5881 : * If subquery uses DISTINCT, we can't make use of any stats for the
5882 : * variable ... but, if it's the only DISTINCT column, we are entitled
5883 : * to consider it unique. We do the test this way so that it works
5884 : * for cases involving DISTINCT ON.
5885 : */
5886 12708 : if (subquery->distinctClause)
5887 : {
5888 1810 : if (list_length(subquery->distinctClause) == 1 &&
5889 602 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
5890 280 : vardata->isunique = true;
5891 : /* cannot go further */
5892 1208 : return;
5893 : }
5894 :
5895 : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
5896 11500 : if (subquery->groupClause)
5897 : {
5898 1040 : if (list_length(subquery->groupClause) == 1 &&
5899 430 : targetIsInSortList(ste, InvalidOid, subquery->groupClause))
5900 334 : vardata->isunique = true;
5901 : /* cannot go further */
5902 610 : return;
5903 : }
5904 :
5905 : /*
5906 : * If the sub-query originated from a view with the security_barrier
5907 : * attribute, we must not look at the variable's statistics, though it
5908 : * seems all right to notice the existence of a DISTINCT clause. So
5909 : * stop here.
5910 : *
5911 : * This is probably a harsher restriction than necessary; it's
5912 : * certainly OK for the selectivity estimator (which is a C function,
5913 : * and therefore omnipotent anyway) to look at the statistics. But
5914 : * many selectivity estimators will happily *invoke the operator
5915 : * function* to try to work out a good estimate - and that's not OK.
5916 : * So for now, don't dig down for stats.
5917 : */
5918 10890 : if (rte->security_barrier)
5919 234 : return;
5920 :
5921 : /* Can only handle a simple Var of subquery's query level */
5922 10656 : if (var && IsA(var, Var) &&
5923 5520 : var->varlevelsup == 0)
5924 : {
5925 : /*
5926 : * OK, recurse into the subquery. Note that the original setting
5927 : * of vardata->isunique (which will surely be false) is left
5928 : * unchanged in this situation. That's what we want, since even
5929 : * if the underlying column is unique, the subquery may have
5930 : * joined to other tables in a way that creates duplicates.
5931 : */
5932 5520 : examine_simple_variable(subroot, var, vardata);
5933 : }
5934 : }
5935 : else
5936 : {
5937 : /*
5938 : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
5939 : * see RTE_JOIN here because join alias Vars have already been
5940 : * flattened.) There's not much we can do with function outputs, but
5941 : * maybe someday try to be smarter about VALUES.
5942 : */
5943 : }
5944 : }
5945 :
5946 : /*
5947 : * Check whether it is permitted to call func_oid passing some of the
5948 : * pg_statistic data in vardata. We allow this either if the user has SELECT
5949 : * privileges on the table or column underlying the pg_statistic data or if
5950 : * the function is marked leakproof.
5951 : */
5952 : bool
5953 923694 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
5954 : {
5955 923694 : if (vardata->acl_ok)
5956 923104 : return true;
5957 :
5958 590 : if (!OidIsValid(func_oid))
5959 0 : return false;
5960 :
5961 590 : if (get_func_leakproof(func_oid))
5962 244 : return true;
5963 :
5964 346 : ereport(DEBUG2,
5965 : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
5966 : get_func_name(func_oid))));
5967 346 : return false;
5968 : }
5969 :
5970 : /*
5971 : * get_variable_numdistinct
5972 : * Estimate the number of distinct values of a variable.
5973 : *
5974 : * vardata: results of examine_variable
5975 : * *isdefault: set to true if the result is a default rather than based on
5976 : * anything meaningful.
5977 : *
5978 : * NB: be careful to produce a positive integral result, since callers may
5979 : * compare the result to exact integer counts, or might divide by it.
5980 : */
5981 : double
5982 1372860 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
5983 : {
5984 : double stadistinct;
5985 1372860 : double stanullfrac = 0.0;
5986 : double ntuples;
5987 :
5988 1372860 : *isdefault = false;
5989 :
5990 : /*
5991 : * Determine the stadistinct value to use. There are cases where we can
5992 : * get an estimate even without a pg_statistic entry, or can get a better
5993 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
5994 : * (otherwise, assume no nulls, for lack of any better idea).
5995 : */
5996 1372860 : if (HeapTupleIsValid(vardata->statsTuple))
5997 : {
5998 : /* Use the pg_statistic entry */
5999 : Form_pg_statistic stats;
6000 :
6001 978874 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6002 978874 : stadistinct = stats->stadistinct;
6003 978874 : stanullfrac = stats->stanullfrac;
6004 : }
6005 393986 : else if (vardata->vartype == BOOLOID)
6006 : {
6007 : /*
6008 : * Special-case boolean columns: presumably, two distinct values.
6009 : *
6010 : * Are there any other datatypes we should wire in special estimates
6011 : * for?
6012 : */
6013 576 : stadistinct = 2.0;
6014 : }
6015 393410 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6016 : {
6017 : /*
6018 : * If the Var represents a column of a VALUES RTE, assume it's unique.
6019 : * This could of course be very wrong, but it should tend to be true
6020 : * in well-written queries. We could consider examining the VALUES'
6021 : * contents to get some real statistics; but that only works if the
6022 : * entries are all constants, and it would be pretty expensive anyway.
6023 : */
6024 3284 : stadistinct = -1.0; /* unique (and all non null) */
6025 : }
6026 : else
6027 : {
6028 : /*
6029 : * We don't keep statistics for system columns, but in some cases we
6030 : * can infer distinctness anyway.
6031 : */
6032 390126 : if (vardata->var && IsA(vardata->var, Var))
6033 : {
6034 366578 : switch (((Var *) vardata->var)->varattno)
6035 : {
6036 1146 : case SelfItemPointerAttributeNumber:
6037 1146 : stadistinct = -1.0; /* unique (and all non null) */
6038 1146 : break;
6039 2518 : case TableOidAttributeNumber:
6040 2518 : stadistinct = 1.0; /* only 1 value */
6041 2518 : break;
6042 362914 : default:
6043 362914 : stadistinct = 0.0; /* means "unknown" */
6044 362914 : break;
6045 : }
6046 : }
6047 : else
6048 23548 : stadistinct = 0.0; /* means "unknown" */
6049 :
6050 : /*
6051 : * XXX consider using estimate_num_groups on expressions?
6052 : */
6053 : }
6054 :
6055 : /*
6056 : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6057 : * variable, assume it is unique no matter what pg_statistic says; the
6058 : * statistics could be out of date, or we might have found a partial
6059 : * unique index that proves the var is unique for this query. However,
6060 : * we'd better still believe the null-fraction statistic.
6061 : */
6062 1372860 : if (vardata->isunique)
6063 367924 : stadistinct = -1.0 * (1.0 - stanullfrac);
6064 :
6065 : /*
6066 : * If we had an absolute estimate, use that.
6067 : */
6068 1372860 : if (stadistinct > 0.0)
6069 320380 : return clamp_row_est(stadistinct);
6070 :
6071 : /*
6072 : * Otherwise we need to get the relation size; punt if not available.
6073 : */
6074 1052480 : if (vardata->rel == NULL)
6075 : {
6076 352 : *isdefault = true;
6077 352 : return DEFAULT_NUM_DISTINCT;
6078 : }
6079 1052128 : ntuples = vardata->rel->tuples;
6080 1052128 : if (ntuples <= 0.0)
6081 : {
6082 47476 : *isdefault = true;
6083 47476 : return DEFAULT_NUM_DISTINCT;
6084 : }
6085 :
6086 : /*
6087 : * If we had a relative estimate, use that.
6088 : */
6089 1004652 : if (stadistinct < 0.0)
6090 703636 : return clamp_row_est(-stadistinct * ntuples);
6091 :
6092 : /*
6093 : * With no data, estimate ndistinct = ntuples if the table is small, else
6094 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6095 : * that the behavior isn't discontinuous.
6096 : */
6097 301016 : if (ntuples < DEFAULT_NUM_DISTINCT)
6098 132766 : return clamp_row_est(ntuples);
6099 :
6100 168250 : *isdefault = true;
6101 168250 : return DEFAULT_NUM_DISTINCT;
6102 : }
6103 :
6104 : /*
6105 : * get_variable_range
6106 : * Estimate the minimum and maximum value of the specified variable.
6107 : * If successful, store values in *min and *max, and return true.
6108 : * If no data available, return false.
6109 : *
6110 : * sortop is the "<" comparison operator to use. This should generally
6111 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6112 : * The collation must be specified too.
6113 : */
6114 : static bool
6115 196212 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6116 : Oid sortop, Oid collation,
6117 : Datum *min, Datum *max)
6118 : {
6119 196212 : Datum tmin = 0;
6120 196212 : Datum tmax = 0;
6121 196212 : bool have_data = false;
6122 : int16 typLen;
6123 : bool typByVal;
6124 : Oid opfuncoid;
6125 : FmgrInfo opproc;
6126 : AttStatsSlot sslot;
6127 :
6128 : /*
6129 : * XXX It's very tempting to try to use the actual column min and max, if
6130 : * we can get them relatively-cheaply with an index probe. However, since
6131 : * this function is called many times during join planning, that could
6132 : * have unpleasant effects on planning speed. Need more investigation
6133 : * before enabling this.
6134 : */
6135 : #ifdef NOT_USED
6136 : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6137 : return true;
6138 : #endif
6139 :
6140 196212 : if (!HeapTupleIsValid(vardata->statsTuple))
6141 : {
6142 : /* no stats available, so default result */
6143 43780 : return false;
6144 : }
6145 :
6146 : /*
6147 : * If we can't apply the sortop to the stats data, just fail. In
6148 : * principle, if there's a histogram and no MCVs, we could return the
6149 : * histogram endpoints without ever applying the sortop ... but it's
6150 : * probably not worth trying, because whatever the caller wants to do with
6151 : * the endpoints would likely fail the security check too.
6152 : */
6153 152432 : if (!statistic_proc_security_check(vardata,
6154 152432 : (opfuncoid = get_opcode(sortop))))
6155 0 : return false;
6156 :
6157 152432 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6158 :
6159 152432 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6160 :
6161 : /*
6162 : * If there is a histogram with the ordering we want, grab the first and
6163 : * last values.
6164 : */
6165 152432 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6166 : STATISTIC_KIND_HISTOGRAM, sortop,
6167 : ATTSTATSSLOT_VALUES))
6168 : {
6169 113316 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6170 : {
6171 113316 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6172 113316 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6173 113316 : have_data = true;
6174 : }
6175 113316 : free_attstatsslot(&sslot);
6176 : }
6177 :
6178 : /*
6179 : * Otherwise, if there is a histogram with some other ordering, scan it
6180 : * and get the min and max values according to the ordering we want. This
6181 : * of course may not find values that are really extremal according to our
6182 : * ordering, but it beats ignoring available data.
6183 : */
6184 191548 : if (!have_data &&
6185 39116 : get_attstatsslot(&sslot, vardata->statsTuple,
6186 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6187 : ATTSTATSSLOT_VALUES))
6188 : {
6189 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6190 : collation, typLen, typByVal,
6191 : &tmin, &tmax, &have_data);
6192 0 : free_attstatsslot(&sslot);
6193 : }
6194 :
6195 : /*
6196 : * If we have most-common-values info, look for extreme MCVs. This is
6197 : * needed even if we also have a histogram, since the histogram excludes
6198 : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6199 : * be pretty wary of deciding that that is a full representation of the
6200 : * data. Proceed only if the MCVs represent the whole table (to within
6201 : * roundoff error).
6202 : */
6203 152432 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6204 : STATISTIC_KIND_MCV, InvalidOid,
6205 152432 : have_data ? ATTSTATSSLOT_VALUES :
6206 : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6207 : {
6208 72120 : bool use_mcvs = have_data;
6209 :
6210 72120 : if (!have_data)
6211 : {
6212 37742 : double sumcommon = 0.0;
6213 : double nullfrac;
6214 : int i;
6215 :
6216 274536 : for (i = 0; i < sslot.nnumbers; i++)
6217 236794 : sumcommon += sslot.numbers[i];
6218 37742 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6219 37742 : if (sumcommon + nullfrac > 0.99999)
6220 36762 : use_mcvs = true;
6221 : }
6222 :
6223 72120 : if (use_mcvs)
6224 71140 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6225 : collation, typLen, typByVal,
6226 : &tmin, &tmax, &have_data);
6227 72120 : free_attstatsslot(&sslot);
6228 : }
6229 :
6230 152432 : *min = tmin;
6231 152432 : *max = tmax;
6232 152432 : return have_data;
6233 : }
6234 :
6235 : /*
6236 : * get_stats_slot_range: scan sslot for min/max values
6237 : *
6238 : * Subroutine for get_variable_range: update min/max/have_data according
6239 : * to what we find in the statistics array.
6240 : */
6241 : static void
6242 71140 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6243 : Oid collation, int16 typLen, bool typByVal,
6244 : Datum *min, Datum *max, bool *p_have_data)
6245 : {
6246 71140 : Datum tmin = *min;
6247 71140 : Datum tmax = *max;
6248 71140 : bool have_data = *p_have_data;
6249 71140 : bool found_tmin = false;
6250 71140 : bool found_tmax = false;
6251 :
6252 : /* Look up the comparison function, if we didn't already do so */
6253 71140 : if (opproc->fn_oid != opfuncoid)
6254 71140 : fmgr_info(opfuncoid, opproc);
6255 :
6256 : /* Scan all the slot's values */
6257 2205762 : for (int i = 0; i < sslot->nvalues; i++)
6258 : {
6259 2134622 : if (!have_data)
6260 : {
6261 36762 : tmin = tmax = sslot->values[i];
6262 36762 : found_tmin = found_tmax = true;
6263 36762 : *p_have_data = have_data = true;
6264 36762 : continue;
6265 : }
6266 2097860 : if (DatumGetBool(FunctionCall2Coll(opproc,
6267 : collation,
6268 2097860 : sslot->values[i], tmin)))
6269 : {
6270 46356 : tmin = sslot->values[i];
6271 46356 : found_tmin = true;
6272 : }
6273 2097860 : if (DatumGetBool(FunctionCall2Coll(opproc,
6274 : collation,
6275 2097860 : tmax, sslot->values[i])))
6276 : {
6277 92264 : tmax = sslot->values[i];
6278 92264 : found_tmax = true;
6279 : }
6280 : }
6281 :
6282 : /*
6283 : * Copy the slot's values, if we found new extreme values.
6284 : */
6285 71140 : if (found_tmin)
6286 56246 : *min = datumCopy(tmin, typByVal, typLen);
6287 71140 : if (found_tmax)
6288 38264 : *max = datumCopy(tmax, typByVal, typLen);
6289 71140 : }
6290 :
6291 :
6292 : /*
6293 : * get_actual_variable_range
6294 : * Attempt to identify the current *actual* minimum and/or maximum
6295 : * of the specified variable, by looking for a suitable btree index
6296 : * and fetching its low and/or high values.
6297 : * If successful, store values in *min and *max, and return true.
6298 : * (Either pointer can be NULL if that endpoint isn't needed.)
6299 : * If unsuccessful, return false.
6300 : *
6301 : * sortop is the "<" comparison operator to use.
6302 : * collation is the required collation.
6303 : */
6304 : static bool
6305 176540 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6306 : Oid sortop, Oid collation,
6307 : Datum *min, Datum *max)
6308 : {
6309 176540 : bool have_data = false;
6310 176540 : RelOptInfo *rel = vardata->rel;
6311 : RangeTblEntry *rte;
6312 : ListCell *lc;
6313 :
6314 : /* No hope if no relation or it doesn't have indexes */
6315 176540 : if (rel == NULL || rel->indexlist == NIL)
6316 13356 : return false;
6317 : /* If it has indexes it must be a plain relation */
6318 163184 : rte = root->simple_rte_array[rel->relid];
6319 : Assert(rte->rtekind == RTE_RELATION);
6320 :
6321 : /* ignore partitioned tables. Any indexes here are not real indexes */
6322 163184 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6323 828 : return false;
6324 :
6325 : /* Search through the indexes to see if any match our problem */
6326 310074 : foreach(lc, rel->indexlist)
6327 : {
6328 265680 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6329 : ScanDirection indexscandir;
6330 :
6331 : /* Ignore non-btree indexes */
6332 265680 : if (index->relam != BTREE_AM_OID)
6333 0 : continue;
6334 :
6335 : /*
6336 : * Ignore partial indexes --- we only want stats that cover the entire
6337 : * relation.
6338 : */
6339 265680 : if (index->indpred != NIL)
6340 288 : continue;
6341 :
6342 : /*
6343 : * The index list might include hypothetical indexes inserted by a
6344 : * get_relation_info hook --- don't try to access them.
6345 : */
6346 265392 : if (index->hypothetical)
6347 0 : continue;
6348 :
6349 : /*
6350 : * The first index column must match the desired variable, sortop, and
6351 : * collation --- but we can use a descending-order index.
6352 : */
6353 265392 : if (collation != index->indexcollations[0])
6354 33160 : continue; /* test first 'cause it's cheapest */
6355 232232 : if (!match_index_to_operand(vardata->var, 0, index))
6356 114270 : continue;
6357 117962 : switch (get_op_opfamily_strategy(sortop, index->sortopfamily[0]))
6358 : {
6359 117962 : case BTLessStrategyNumber:
6360 117962 : if (index->reverse_sort[0])
6361 0 : indexscandir = BackwardScanDirection;
6362 : else
6363 117962 : indexscandir = ForwardScanDirection;
6364 117962 : break;
6365 0 : case BTGreaterStrategyNumber:
6366 0 : if (index->reverse_sort[0])
6367 0 : indexscandir = ForwardScanDirection;
6368 : else
6369 0 : indexscandir = BackwardScanDirection;
6370 0 : break;
6371 0 : default:
6372 : /* index doesn't match the sortop */
6373 0 : continue;
6374 : }
6375 :
6376 : /*
6377 : * Found a suitable index to extract data from. Set up some data that
6378 : * can be used by both invocations of get_actual_variable_endpoint.
6379 : */
6380 : {
6381 : MemoryContext tmpcontext;
6382 : MemoryContext oldcontext;
6383 : Relation heapRel;
6384 : Relation indexRel;
6385 : TupleTableSlot *slot;
6386 : int16 typLen;
6387 : bool typByVal;
6388 : ScanKeyData scankeys[1];
6389 :
6390 : /* Make sure any cruft gets recycled when we're done */
6391 117962 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
6392 : "get_actual_variable_range workspace",
6393 : ALLOCSET_DEFAULT_SIZES);
6394 117962 : oldcontext = MemoryContextSwitchTo(tmpcontext);
6395 :
6396 : /*
6397 : * Open the table and index so we can read from them. We should
6398 : * already have some type of lock on each.
6399 : */
6400 117962 : heapRel = table_open(rte->relid, NoLock);
6401 117962 : indexRel = index_open(index->indexoid, NoLock);
6402 :
6403 : /* build some stuff needed for indexscan execution */
6404 117962 : slot = table_slot_create(heapRel, NULL);
6405 117962 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6406 :
6407 : /* set up an IS NOT NULL scan key so that we ignore nulls */
6408 117962 : ScanKeyEntryInitialize(&scankeys[0],
6409 : SK_ISNULL | SK_SEARCHNOTNULL,
6410 : 1, /* index col to scan */
6411 : InvalidStrategy, /* no strategy */
6412 : InvalidOid, /* no strategy subtype */
6413 : InvalidOid, /* no collation */
6414 : InvalidOid, /* no reg proc for this */
6415 : (Datum) 0); /* constant */
6416 :
6417 : /* If min is requested ... */
6418 117962 : if (min)
6419 : {
6420 68096 : have_data = get_actual_variable_endpoint(heapRel,
6421 : indexRel,
6422 : indexscandir,
6423 : scankeys,
6424 : typLen,
6425 : typByVal,
6426 : slot,
6427 : oldcontext,
6428 : min);
6429 : }
6430 : else
6431 : {
6432 : /* If min not requested, still want to fetch max */
6433 49866 : have_data = true;
6434 : }
6435 :
6436 : /* If max is requested, and we didn't already fail ... */
6437 117962 : if (max && have_data)
6438 : {
6439 : /* scan in the opposite direction; all else is the same */
6440 51442 : have_data = get_actual_variable_endpoint(heapRel,
6441 : indexRel,
6442 51442 : -indexscandir,
6443 : scankeys,
6444 : typLen,
6445 : typByVal,
6446 : slot,
6447 : oldcontext,
6448 : max);
6449 : }
6450 :
6451 : /* Clean everything up */
6452 117962 : ExecDropSingleTupleTableSlot(slot);
6453 :
6454 117962 : index_close(indexRel, NoLock);
6455 117962 : table_close(heapRel, NoLock);
6456 :
6457 117962 : MemoryContextSwitchTo(oldcontext);
6458 117962 : MemoryContextDelete(tmpcontext);
6459 :
6460 : /* And we're done */
6461 117962 : break;
6462 : }
6463 : }
6464 :
6465 162356 : return have_data;
6466 : }
6467 :
6468 : /*
6469 : * Get one endpoint datum (min or max depending on indexscandir) from the
6470 : * specified index. Return true if successful, false if not.
6471 : * On success, endpoint value is stored to *endpointDatum (and copied into
6472 : * outercontext).
6473 : *
6474 : * scankeys is a 1-element scankey array set up to reject nulls.
6475 : * typLen/typByVal describe the datatype of the index's first column.
6476 : * tableslot is a slot suitable to hold table tuples, in case we need
6477 : * to probe the heap.
6478 : * (We could compute these values locally, but that would mean computing them
6479 : * twice when get_actual_variable_range needs both the min and the max.)
6480 : *
6481 : * Failure occurs either when the index is empty, or we decide that it's
6482 : * taking too long to find a suitable tuple.
6483 : */
6484 : static bool
6485 119538 : get_actual_variable_endpoint(Relation heapRel,
6486 : Relation indexRel,
6487 : ScanDirection indexscandir,
6488 : ScanKey scankeys,
6489 : int16 typLen,
6490 : bool typByVal,
6491 : TupleTableSlot *tableslot,
6492 : MemoryContext outercontext,
6493 : Datum *endpointDatum)
6494 : {
6495 119538 : bool have_data = false;
6496 : SnapshotData SnapshotNonVacuumable;
6497 : IndexScanDesc index_scan;
6498 119538 : Buffer vmbuffer = InvalidBuffer;
6499 119538 : BlockNumber last_heap_block = InvalidBlockNumber;
6500 119538 : int n_visited_heap_pages = 0;
6501 : ItemPointer tid;
6502 : Datum values[INDEX_MAX_KEYS];
6503 : bool isnull[INDEX_MAX_KEYS];
6504 : MemoryContext oldcontext;
6505 :
6506 : /*
6507 : * We use the index-only-scan machinery for this. With mostly-static
6508 : * tables that's a win because it avoids a heap visit. It's also a win
6509 : * for dynamic data, but the reason is less obvious; read on for details.
6510 : *
6511 : * In principle, we should scan the index with our current active
6512 : * snapshot, which is the best approximation we've got to what the query
6513 : * will see when executed. But that won't be exact if a new snap is taken
6514 : * before running the query, and it can be very expensive if a lot of
6515 : * recently-dead or uncommitted rows exist at the beginning or end of the
6516 : * index (because we'll laboriously fetch each one and reject it).
6517 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
6518 : * and uncommitted rows as well as normal visible rows. On the other
6519 : * hand, it will reject known-dead rows, and thus not give a bogus answer
6520 : * when the extreme value has been deleted (unless the deletion was quite
6521 : * recent); that case motivates not using SnapshotAny here.
6522 : *
6523 : * A crucial point here is that SnapshotNonVacuumable, with
6524 : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
6525 : * condition that the indexscan will use to decide that index entries are
6526 : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
6527 : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
6528 : * have to continue scanning past it, we know that the indexscan will mark
6529 : * that index entry killed. That means that the next
6530 : * get_actual_variable_endpoint() call will not have to re-consider that
6531 : * index entry. In this way we avoid repetitive work when this function
6532 : * is used a lot during planning.
6533 : *
6534 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
6535 : * recently-created index, some index entries may point at "broken" HOT
6536 : * chains in which not all the tuple versions contain data matching the
6537 : * index entry. The live tuple version(s) certainly do match the index,
6538 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
6539 : * don't match. Hence, if we took data from the selected heap tuple, we
6540 : * might get a bogus answer that's not close to the index extremal value,
6541 : * or could even be NULL. We avoid this hazard because we take the data
6542 : * from the index entry not the heap.
6543 : *
6544 : * Despite all this care, there are situations where we might find many
6545 : * non-visible tuples near the end of the index. We don't want to expend
6546 : * a huge amount of time here, so we give up once we've read too many heap
6547 : * pages. When we fail for that reason, the caller will end up using
6548 : * whatever extremal value is recorded in pg_statistic.
6549 : */
6550 119538 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
6551 : GlobalVisTestFor(heapRel));
6552 :
6553 119538 : index_scan = index_beginscan(heapRel, indexRel,
6554 : &SnapshotNonVacuumable, NULL,
6555 : 1, 0);
6556 : /* Set it up for index-only scan */
6557 119538 : index_scan->xs_want_itup = true;
6558 119538 : index_rescan(index_scan, scankeys, 1, NULL, 0);
6559 :
6560 : /* Fetch first/next tuple in specified direction */
6561 152994 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
6562 : {
6563 152994 : BlockNumber block = ItemPointerGetBlockNumber(tid);
6564 :
6565 152994 : if (!VM_ALL_VISIBLE(heapRel,
6566 : block,
6567 : &vmbuffer))
6568 : {
6569 : /* Rats, we have to visit the heap to check visibility */
6570 112166 : if (!index_fetch_heap(index_scan, tableslot))
6571 : {
6572 : /*
6573 : * No visible tuple for this index entry, so we need to
6574 : * advance to the next entry. Before doing so, count heap
6575 : * page fetches and give up if we've done too many.
6576 : *
6577 : * We don't charge a page fetch if this is the same heap page
6578 : * as the previous tuple. This is on the conservative side,
6579 : * since other recently-accessed pages are probably still in
6580 : * buffers too; but it's good enough for this heuristic.
6581 : */
6582 : #define VISITED_PAGES_LIMIT 100
6583 :
6584 33456 : if (block != last_heap_block)
6585 : {
6586 2956 : last_heap_block = block;
6587 2956 : n_visited_heap_pages++;
6588 2956 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
6589 0 : break;
6590 : }
6591 :
6592 33456 : continue; /* no visible tuple, try next index entry */
6593 : }
6594 :
6595 : /* We don't actually need the heap tuple for anything */
6596 78710 : ExecClearTuple(tableslot);
6597 :
6598 : /*
6599 : * We don't care whether there's more than one visible tuple in
6600 : * the HOT chain; if any are visible, that's good enough.
6601 : */
6602 : }
6603 :
6604 : /*
6605 : * We expect that btree will return data in IndexTuple not HeapTuple
6606 : * format. It's not lossy either.
6607 : */
6608 119538 : if (!index_scan->xs_itup)
6609 0 : elog(ERROR, "no data returned for index-only scan");
6610 119538 : if (index_scan->xs_recheck)
6611 0 : elog(ERROR, "unexpected recheck indication from btree");
6612 :
6613 : /* OK to deconstruct the index tuple */
6614 119538 : index_deform_tuple(index_scan->xs_itup,
6615 : index_scan->xs_itupdesc,
6616 : values, isnull);
6617 :
6618 : /* Shouldn't have got a null, but be careful */
6619 119538 : if (isnull[0])
6620 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
6621 : RelationGetRelationName(indexRel));
6622 :
6623 : /* Copy the index column value out to caller's context */
6624 119538 : oldcontext = MemoryContextSwitchTo(outercontext);
6625 119538 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
6626 119538 : MemoryContextSwitchTo(oldcontext);
6627 119538 : have_data = true;
6628 119538 : break;
6629 : }
6630 :
6631 119538 : if (vmbuffer != InvalidBuffer)
6632 106840 : ReleaseBuffer(vmbuffer);
6633 119538 : index_endscan(index_scan);
6634 :
6635 119538 : return have_data;
6636 : }
6637 :
6638 : /*
6639 : * find_join_input_rel
6640 : * Look up the input relation for a join.
6641 : *
6642 : * We assume that the input relation's RelOptInfo must have been constructed
6643 : * already.
6644 : */
6645 : static RelOptInfo *
6646 7722 : find_join_input_rel(PlannerInfo *root, Relids relids)
6647 : {
6648 7722 : RelOptInfo *rel = NULL;
6649 :
6650 7722 : if (!bms_is_empty(relids))
6651 : {
6652 : int relid;
6653 :
6654 7722 : if (bms_get_singleton_member(relids, &relid))
6655 7418 : rel = find_base_rel(root, relid);
6656 : else
6657 304 : rel = find_join_rel(root, relids);
6658 : }
6659 :
6660 7722 : if (rel == NULL)
6661 0 : elog(ERROR, "could not find RelOptInfo for given relids");
6662 :
6663 7722 : return rel;
6664 : }
6665 :
6666 :
6667 : /*-------------------------------------------------------------------------
6668 : *
6669 : * Index cost estimation functions
6670 : *
6671 : *-------------------------------------------------------------------------
6672 : */
6673 :
6674 : /*
6675 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
6676 : */
6677 : List *
6678 741284 : get_quals_from_indexclauses(List *indexclauses)
6679 : {
6680 741284 : List *result = NIL;
6681 : ListCell *lc;
6682 :
6683 1312598 : foreach(lc, indexclauses)
6684 : {
6685 571314 : IndexClause *iclause = lfirst_node(IndexClause, lc);
6686 : ListCell *lc2;
6687 :
6688 1145520 : foreach(lc2, iclause->indexquals)
6689 : {
6690 574206 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6691 :
6692 574206 : result = lappend(result, rinfo);
6693 : }
6694 : }
6695 741284 : return result;
6696 : }
6697 :
6698 : /*
6699 : * Compute the total evaluation cost of the comparison operands in a list
6700 : * of index qual expressions. Since we know these will be evaluated just
6701 : * once per scan, there's no need to distinguish startup from per-row cost.
6702 : *
6703 : * This can be used either on the result of get_quals_from_indexclauses(),
6704 : * or directly on an indexorderbys list. In both cases, we expect that the
6705 : * index key expression is on the left side of binary clauses.
6706 : */
6707 : Cost
6708 1469958 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
6709 : {
6710 1469958 : Cost qual_arg_cost = 0;
6711 : ListCell *lc;
6712 :
6713 2044626 : foreach(lc, indexquals)
6714 : {
6715 574668 : Expr *clause = (Expr *) lfirst(lc);
6716 : Node *other_operand;
6717 : QualCost index_qual_cost;
6718 :
6719 : /*
6720 : * Index quals will have RestrictInfos, indexorderbys won't. Look
6721 : * through RestrictInfo if present.
6722 : */
6723 574668 : if (IsA(clause, RestrictInfo))
6724 574194 : clause = ((RestrictInfo *) clause)->clause;
6725 :
6726 574668 : if (IsA(clause, OpExpr))
6727 : {
6728 561474 : OpExpr *op = (OpExpr *) clause;
6729 :
6730 561474 : other_operand = (Node *) lsecond(op->args);
6731 : }
6732 13194 : else if (IsA(clause, RowCompareExpr))
6733 : {
6734 300 : RowCompareExpr *rc = (RowCompareExpr *) clause;
6735 :
6736 300 : other_operand = (Node *) rc->rargs;
6737 : }
6738 12894 : else if (IsA(clause, ScalarArrayOpExpr))
6739 : {
6740 9998 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6741 :
6742 9998 : other_operand = (Node *) lsecond(saop->args);
6743 : }
6744 2896 : else if (IsA(clause, NullTest))
6745 : {
6746 2896 : other_operand = NULL;
6747 : }
6748 : else
6749 : {
6750 0 : elog(ERROR, "unsupported indexqual type: %d",
6751 : (int) nodeTag(clause));
6752 : other_operand = NULL; /* keep compiler quiet */
6753 : }
6754 :
6755 574668 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
6756 574668 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
6757 : }
6758 1469958 : return qual_arg_cost;
6759 : }
6760 :
6761 : void
6762 728686 : genericcostestimate(PlannerInfo *root,
6763 : IndexPath *path,
6764 : double loop_count,
6765 : GenericCosts *costs)
6766 : {
6767 728686 : IndexOptInfo *index = path->indexinfo;
6768 728686 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
6769 728686 : List *indexOrderBys = path->indexorderbys;
6770 : Cost indexStartupCost;
6771 : Cost indexTotalCost;
6772 : Selectivity indexSelectivity;
6773 : double indexCorrelation;
6774 : double numIndexPages;
6775 : double numIndexTuples;
6776 : double spc_random_page_cost;
6777 : double num_sa_scans;
6778 : double num_outer_scans;
6779 : double num_scans;
6780 : double qual_op_cost;
6781 : double qual_arg_cost;
6782 : List *selectivityQuals;
6783 : ListCell *l;
6784 :
6785 : /*
6786 : * If the index is partial, AND the index predicate with the explicitly
6787 : * given indexquals to produce a more accurate idea of the index
6788 : * selectivity.
6789 : */
6790 728686 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
6791 :
6792 : /*
6793 : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
6794 : * just assume that the number of index descents is the number of distinct
6795 : * combinations of array elements from all of the scan's SAOP clauses.
6796 : */
6797 728686 : num_sa_scans = costs->num_sa_scans;
6798 728686 : if (num_sa_scans < 1)
6799 : {
6800 7784 : num_sa_scans = 1;
6801 16292 : foreach(l, indexQuals)
6802 : {
6803 8508 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
6804 :
6805 8508 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
6806 : {
6807 26 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
6808 26 : double alength = estimate_array_length(root, lsecond(saop->args));
6809 :
6810 26 : if (alength > 1)
6811 26 : num_sa_scans *= alength;
6812 : }
6813 : }
6814 : }
6815 :
6816 : /* Estimate the fraction of main-table tuples that will be visited */
6817 728686 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
6818 728686 : index->rel->relid,
6819 : JOIN_INNER,
6820 : NULL);
6821 :
6822 : /*
6823 : * If caller didn't give us an estimate, estimate the number of index
6824 : * tuples that will be visited. We do it in this rather peculiar-looking
6825 : * way in order to get the right answer for partial indexes.
6826 : */
6827 728686 : numIndexTuples = costs->numIndexTuples;
6828 728686 : if (numIndexTuples <= 0.0)
6829 : {
6830 76008 : numIndexTuples = indexSelectivity * index->rel->tuples;
6831 :
6832 : /*
6833 : * The above calculation counts all the tuples visited across all
6834 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
6835 : * average per-indexscan number, so adjust. This is a handy place to
6836 : * round to integer, too. (If caller supplied tuple estimate, it's
6837 : * responsible for handling these considerations.)
6838 : */
6839 76008 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
6840 : }
6841 :
6842 : /*
6843 : * We can bound the number of tuples by the index size in any case. Also,
6844 : * always estimate at least one tuple is touched, even when
6845 : * indexSelectivity estimate is tiny.
6846 : */
6847 728686 : if (numIndexTuples > index->tuples)
6848 5438 : numIndexTuples = index->tuples;
6849 728686 : if (numIndexTuples < 1.0)
6850 75660 : numIndexTuples = 1.0;
6851 :
6852 : /*
6853 : * Estimate the number of index pages that will be retrieved.
6854 : *
6855 : * We use the simplistic method of taking a pro-rata fraction of the total
6856 : * number of index pages. In effect, this counts only leaf pages and not
6857 : * any overhead such as index metapage or upper tree levels.
6858 : *
6859 : * In practice access to upper index levels is often nearly free because
6860 : * those tend to stay in cache under load; moreover, the cost involved is
6861 : * highly dependent on index type. We therefore ignore such costs here
6862 : * and leave it to the caller to add a suitable charge if needed.
6863 : */
6864 728686 : if (index->pages > 1 && index->tuples > 1)
6865 683892 : numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
6866 : else
6867 44794 : numIndexPages = 1.0;
6868 :
6869 : /* fetch estimated page cost for tablespace containing index */
6870 728686 : get_tablespace_page_costs(index->reltablespace,
6871 : &spc_random_page_cost,
6872 : NULL);
6873 :
6874 : /*
6875 : * Now compute the disk access costs.
6876 : *
6877 : * The above calculations are all per-index-scan. However, if we are in a
6878 : * nestloop inner scan, we can expect the scan to be repeated (with
6879 : * different search keys) for each row of the outer relation. Likewise,
6880 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
6881 : * the potential for cache effects to reduce the number of disk page
6882 : * fetches needed. We want to estimate the average per-scan I/O cost in
6883 : * the presence of caching.
6884 : *
6885 : * We use the Mackert-Lohman formula (see costsize.c for details) to
6886 : * estimate the total number of page fetches that occur. While this
6887 : * wasn't what it was designed for, it seems a reasonable model anyway.
6888 : * Note that we are counting pages not tuples anymore, so we take N = T =
6889 : * index size, as if there were one "tuple" per page.
6890 : */
6891 728686 : num_outer_scans = loop_count;
6892 728686 : num_scans = num_sa_scans * num_outer_scans;
6893 :
6894 728686 : if (num_scans > 1)
6895 : {
6896 : double pages_fetched;
6897 :
6898 : /* total page fetches ignoring cache effects */
6899 83576 : pages_fetched = numIndexPages * num_scans;
6900 :
6901 : /* use Mackert and Lohman formula to adjust for cache effects */
6902 83576 : pages_fetched = index_pages_fetched(pages_fetched,
6903 : index->pages,
6904 83576 : (double) index->pages,
6905 : root);
6906 :
6907 : /*
6908 : * Now compute the total disk access cost, and then report a pro-rated
6909 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
6910 : * since that's internal to the indexscan.)
6911 : */
6912 83576 : indexTotalCost = (pages_fetched * spc_random_page_cost)
6913 : / num_outer_scans;
6914 : }
6915 : else
6916 : {
6917 : /*
6918 : * For a single index scan, we just charge spc_random_page_cost per
6919 : * page touched.
6920 : */
6921 645110 : indexTotalCost = numIndexPages * spc_random_page_cost;
6922 : }
6923 :
6924 : /*
6925 : * CPU cost: any complex expressions in the indexquals will need to be
6926 : * evaluated once at the start of the scan to reduce them to runtime keys
6927 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
6928 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
6929 : * indexqual operator. Because we have numIndexTuples as a per-scan
6930 : * number, we have to multiply by num_sa_scans to get the correct result
6931 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
6932 : * ORDER BY expressions.
6933 : *
6934 : * Note: this neglects the possible costs of rechecking lossy operators.
6935 : * Detecting that that might be needed seems more expensive than it's
6936 : * worth, though, considering all the other inaccuracies here ...
6937 : */
6938 728686 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
6939 728686 : index_other_operands_eval_cost(root, indexOrderBys);
6940 728686 : qual_op_cost = cpu_operator_cost *
6941 728686 : (list_length(indexQuals) + list_length(indexOrderBys));
6942 :
6943 728686 : indexStartupCost = qual_arg_cost;
6944 728686 : indexTotalCost += qual_arg_cost;
6945 728686 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
6946 :
6947 : /*
6948 : * Generic assumption about index correlation: there isn't any.
6949 : */
6950 728686 : indexCorrelation = 0.0;
6951 :
6952 : /*
6953 : * Return everything to caller.
6954 : */
6955 728686 : costs->indexStartupCost = indexStartupCost;
6956 728686 : costs->indexTotalCost = indexTotalCost;
6957 728686 : costs->indexSelectivity = indexSelectivity;
6958 728686 : costs->indexCorrelation = indexCorrelation;
6959 728686 : costs->numIndexPages = numIndexPages;
6960 728686 : costs->numIndexTuples = numIndexTuples;
6961 728686 : costs->spc_random_page_cost = spc_random_page_cost;
6962 728686 : costs->num_sa_scans = num_sa_scans;
6963 728686 : }
6964 :
6965 : /*
6966 : * If the index is partial, add its predicate to the given qual list.
6967 : *
6968 : * ANDing the index predicate with the explicitly given indexquals produces
6969 : * a more accurate idea of the index's selectivity. However, we need to be
6970 : * careful not to insert redundant clauses, because clauselist_selectivity()
6971 : * is easily fooled into computing a too-low selectivity estimate. Our
6972 : * approach is to add only the predicate clause(s) that cannot be proven to
6973 : * be implied by the given indexquals. This successfully handles cases such
6974 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
6975 : * There are many other cases where we won't detect redundancy, leading to a
6976 : * too-low selectivity estimate, which will bias the system in favor of using
6977 : * partial indexes where possible. That is not necessarily bad though.
6978 : *
6979 : * Note that indexQuals contains RestrictInfo nodes while the indpred
6980 : * does not, so the output list will be mixed. This is OK for both
6981 : * predicate_implied_by() and clauselist_selectivity(), but might be
6982 : * problematic if the result were passed to other things.
6983 : */
6984 : List *
6985 1211280 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
6986 : {
6987 1211280 : List *predExtraQuals = NIL;
6988 : ListCell *lc;
6989 :
6990 1211280 : if (index->indpred == NIL)
6991 1209254 : return indexQuals;
6992 :
6993 4064 : foreach(lc, index->indpred)
6994 : {
6995 2038 : Node *predQual = (Node *) lfirst(lc);
6996 2038 : List *oneQual = list_make1(predQual);
6997 :
6998 2038 : if (!predicate_implied_by(oneQual, indexQuals, false))
6999 1836 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7000 : }
7001 2026 : return list_concat(predExtraQuals, indexQuals);
7002 : }
7003 :
7004 :
7005 : void
7006 720902 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7007 : Cost *indexStartupCost, Cost *indexTotalCost,
7008 : Selectivity *indexSelectivity, double *indexCorrelation,
7009 : double *indexPages)
7010 : {
7011 720902 : IndexOptInfo *index = path->indexinfo;
7012 720902 : GenericCosts costs = {0};
7013 : Oid relid;
7014 : AttrNumber colnum;
7015 720902 : VariableStatData vardata = {0};
7016 : double numIndexTuples;
7017 : Cost descentCost;
7018 : List *indexBoundQuals;
7019 : int indexcol;
7020 : bool eqQualHere;
7021 : bool found_saop;
7022 : bool found_is_null_op;
7023 : double num_sa_scans;
7024 : ListCell *lc;
7025 :
7026 : /*
7027 : * For a btree scan, only leading '=' quals plus inequality quals for the
7028 : * immediately next attribute contribute to index selectivity (these are
7029 : * the "boundary quals" that determine the starting and stopping points of
7030 : * the index scan). Additional quals can suppress visits to the heap, so
7031 : * it's OK to count them in indexSelectivity, but they should not count
7032 : * for estimating numIndexTuples. So we must examine the given indexquals
7033 : * to find out which ones count as boundary quals. We rely on the
7034 : * knowledge that they are given in index column order.
7035 : *
7036 : * For a RowCompareExpr, we consider only the first column, just as
7037 : * rowcomparesel() does.
7038 : *
7039 : * If there's a ScalarArrayOpExpr in the quals, we'll actually perform up
7040 : * to N index descents (not just one), but the ScalarArrayOpExpr's
7041 : * operator can be considered to act the same as it normally does.
7042 : */
7043 720902 : indexBoundQuals = NIL;
7044 720902 : indexcol = 0;
7045 720902 : eqQualHere = false;
7046 720902 : found_saop = false;
7047 720902 : found_is_null_op = false;
7048 720902 : num_sa_scans = 1;
7049 1239592 : foreach(lc, path->indexclauses)
7050 : {
7051 547796 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7052 : ListCell *lc2;
7053 :
7054 547796 : if (indexcol != iclause->indexcol)
7055 : {
7056 : /* Beginning of a new column's quals */
7057 101136 : if (!eqQualHere)
7058 27592 : break; /* done if no '=' qual for indexcol */
7059 73544 : eqQualHere = false;
7060 73544 : indexcol++;
7061 73544 : if (indexcol != iclause->indexcol)
7062 1514 : break; /* no quals at all for indexcol */
7063 : }
7064 :
7065 : /* Examine each indexqual associated with this index clause */
7066 1040094 : foreach(lc2, iclause->indexquals)
7067 : {
7068 521404 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7069 521404 : Expr *clause = rinfo->clause;
7070 521404 : Oid clause_op = InvalidOid;
7071 : int op_strategy;
7072 :
7073 521404 : if (IsA(clause, OpExpr))
7074 : {
7075 509208 : OpExpr *op = (OpExpr *) clause;
7076 :
7077 509208 : clause_op = op->opno;
7078 : }
7079 12196 : else if (IsA(clause, RowCompareExpr))
7080 : {
7081 300 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7082 :
7083 300 : clause_op = linitial_oid(rc->opnos);
7084 : }
7085 11896 : else if (IsA(clause, ScalarArrayOpExpr))
7086 : {
7087 9642 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7088 9642 : Node *other_operand = (Node *) lsecond(saop->args);
7089 9642 : double alength = estimate_array_length(root, other_operand);
7090 :
7091 9642 : clause_op = saop->opno;
7092 9642 : found_saop = true;
7093 : /* estimate SA descents by indexBoundQuals only */
7094 9642 : if (alength > 1)
7095 9546 : num_sa_scans *= alength;
7096 : }
7097 2254 : else if (IsA(clause, NullTest))
7098 : {
7099 2254 : NullTest *nt = (NullTest *) clause;
7100 :
7101 2254 : if (nt->nulltesttype == IS_NULL)
7102 : {
7103 210 : found_is_null_op = true;
7104 : /* IS NULL is like = for selectivity purposes */
7105 210 : eqQualHere = true;
7106 : }
7107 : }
7108 : else
7109 0 : elog(ERROR, "unsupported indexqual type: %d",
7110 : (int) nodeTag(clause));
7111 :
7112 : /* check for equality operator */
7113 521404 : if (OidIsValid(clause_op))
7114 : {
7115 519150 : op_strategy = get_op_opfamily_strategy(clause_op,
7116 519150 : index->opfamily[indexcol]);
7117 : Assert(op_strategy != 0); /* not a member of opfamily?? */
7118 519150 : if (op_strategy == BTEqualStrategyNumber)
7119 493770 : eqQualHere = true;
7120 : }
7121 :
7122 521404 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
7123 : }
7124 : }
7125 :
7126 : /*
7127 : * If index is unique and we found an '=' clause for each column, we can
7128 : * just assume numIndexTuples = 1 and skip the expensive
7129 : * clauselist_selectivity calculations. However, a ScalarArrayOp or
7130 : * NullTest invalidates that theory, even though it sets eqQualHere.
7131 : */
7132 720902 : if (index->unique &&
7133 598114 : indexcol == index->nkeycolumns - 1 &&
7134 245572 : eqQualHere &&
7135 245572 : !found_saop &&
7136 240242 : !found_is_null_op)
7137 240176 : numIndexTuples = 1.0;
7138 : else
7139 : {
7140 : List *selectivityQuals;
7141 : Selectivity btreeSelectivity;
7142 :
7143 : /*
7144 : * If the index is partial, AND the index predicate with the
7145 : * index-bound quals to produce a more accurate idea of the number of
7146 : * rows covered by the bound conditions.
7147 : */
7148 480726 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
7149 :
7150 480726 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
7151 480726 : index->rel->relid,
7152 : JOIN_INNER,
7153 : NULL);
7154 480726 : numIndexTuples = btreeSelectivity * index->rel->tuples;
7155 :
7156 : /*
7157 : * btree automatically combines individual ScalarArrayOpExpr primitive
7158 : * index scans whenever the tuples covered by the next set of array
7159 : * keys are close to tuples covered by the current set. That puts a
7160 : * natural ceiling on the worst case number of descents -- there
7161 : * cannot possibly be more than one descent per leaf page scanned.
7162 : *
7163 : * Clamp the number of descents to at most 1/3 the number of index
7164 : * pages. This avoids implausibly high estimates with low selectivity
7165 : * paths, where scans usually require only one or two descents. This
7166 : * is most likely to help when there are several SAOP clauses, where
7167 : * naively accepting the total number of distinct combinations of
7168 : * array elements as the number of descents would frequently lead to
7169 : * wild overestimates.
7170 : *
7171 : * We somewhat arbitrarily don't just make the cutoff the total number
7172 : * of leaf pages (we make it 1/3 the total number of pages instead) to
7173 : * give the btree code credit for its ability to continue on the leaf
7174 : * level with low selectivity scans.
7175 : */
7176 480726 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
7177 480726 : num_sa_scans = Max(num_sa_scans, 1);
7178 :
7179 : /*
7180 : * As in genericcostestimate(), we have to adjust for any
7181 : * ScalarArrayOpExpr quals included in indexBoundQuals, and then round
7182 : * to integer.
7183 : *
7184 : * It is tempting to make genericcostestimate behave as if SAOP
7185 : * clauses work in almost the same way as scalar operators during
7186 : * btree scans, making the top-level scan look like a continuous scan
7187 : * (as opposed to num_sa_scans-many primitive index scans). After
7188 : * all, btree scans mostly work like that at runtime. However, such a
7189 : * scheme would badly bias genericcostestimate's simplistic approach
7190 : * to calculating numIndexPages through prorating.
7191 : *
7192 : * Stick with the approach taken by non-native SAOP scans for now.
7193 : * genericcostestimate will use the Mackert-Lohman formula to
7194 : * compensate for repeat page fetches, even though that definitely
7195 : * won't happen during btree scans (not for leaf pages, at least).
7196 : * We're usually very pessimistic about the number of primitive index
7197 : * scans that will be required, but it's not clear how to do better.
7198 : */
7199 480726 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7200 : }
7201 :
7202 : /*
7203 : * Now do generic index cost estimation.
7204 : */
7205 720902 : costs.numIndexTuples = numIndexTuples;
7206 720902 : costs.num_sa_scans = num_sa_scans;
7207 :
7208 720902 : genericcostestimate(root, path, loop_count, &costs);
7209 :
7210 : /*
7211 : * Add a CPU-cost component to represent the costs of initial btree
7212 : * descent. We don't charge any I/O cost for touching upper btree levels,
7213 : * since they tend to stay in cache, but we still have to do about log2(N)
7214 : * comparisons to descend a btree of N leaf tuples. We charge one
7215 : * cpu_operator_cost per comparison.
7216 : *
7217 : * If there are ScalarArrayOpExprs, charge this once per estimated SA
7218 : * index descent. The ones after the first one are not startup cost so
7219 : * far as the overall plan goes, so just add them to "total" cost.
7220 : */
7221 720902 : if (index->tuples > 1) /* avoid computing log(0) */
7222 : {
7223 684646 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
7224 684646 : costs.indexStartupCost += descentCost;
7225 684646 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7226 : }
7227 :
7228 : /*
7229 : * Even though we're not charging I/O cost for touching upper btree pages,
7230 : * it's still reasonable to charge some CPU cost per page descended
7231 : * through. Moreover, if we had no such charge at all, bloated indexes
7232 : * would appear to have the same search cost as unbloated ones, at least
7233 : * in cases where only a single leaf page is expected to be visited. This
7234 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
7235 : * touched. The number of such pages is btree tree height plus one (ie,
7236 : * we charge for the leaf page too). As above, charge once per estimated
7237 : * SA index descent.
7238 : */
7239 720902 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7240 720902 : costs.indexStartupCost += descentCost;
7241 720902 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7242 :
7243 : /*
7244 : * If we can get an estimate of the first column's ordering correlation C
7245 : * from pg_statistic, estimate the index correlation as C for a
7246 : * single-column index, or C * 0.75 for multiple columns. (The idea here
7247 : * is that multiple columns dilute the importance of the first column's
7248 : * ordering, but don't negate it entirely. Before 8.0 we divided the
7249 : * correlation by the number of columns, but that seems too strong.)
7250 : */
7251 720902 : if (index->indexkeys[0] != 0)
7252 : {
7253 : /* Simple variable --- look to stats for the underlying table */
7254 718712 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
7255 :
7256 : Assert(rte->rtekind == RTE_RELATION);
7257 718712 : relid = rte->relid;
7258 : Assert(relid != InvalidOid);
7259 718712 : colnum = index->indexkeys[0];
7260 :
7261 718712 : if (get_relation_stats_hook &&
7262 0 : (*get_relation_stats_hook) (root, rte, colnum, &vardata))
7263 : {
7264 : /*
7265 : * The hook took control of acquiring a stats tuple. If it did
7266 : * supply a tuple, it'd better have supplied a freefunc.
7267 : */
7268 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
7269 0 : !vardata.freefunc)
7270 0 : elog(ERROR, "no function provided to release variable stats with");
7271 : }
7272 : else
7273 : {
7274 718712 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
7275 : ObjectIdGetDatum(relid),
7276 : Int16GetDatum(colnum),
7277 718712 : BoolGetDatum(rte->inh));
7278 718712 : vardata.freefunc = ReleaseSysCache;
7279 : }
7280 : }
7281 : else
7282 : {
7283 : /* Expression --- maybe there are stats for the index itself */
7284 2190 : relid = index->indexoid;
7285 2190 : colnum = 1;
7286 :
7287 2190 : if (get_index_stats_hook &&
7288 0 : (*get_index_stats_hook) (root, relid, colnum, &vardata))
7289 : {
7290 : /*
7291 : * The hook took control of acquiring a stats tuple. If it did
7292 : * supply a tuple, it'd better have supplied a freefunc.
7293 : */
7294 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
7295 0 : !vardata.freefunc)
7296 0 : elog(ERROR, "no function provided to release variable stats with");
7297 : }
7298 : else
7299 : {
7300 2190 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
7301 : ObjectIdGetDatum(relid),
7302 : Int16GetDatum(colnum),
7303 : BoolGetDatum(false));
7304 2190 : vardata.freefunc = ReleaseSysCache;
7305 : }
7306 : }
7307 :
7308 720902 : if (HeapTupleIsValid(vardata.statsTuple))
7309 : {
7310 : Oid sortop;
7311 : AttStatsSlot sslot;
7312 :
7313 536280 : sortop = get_opfamily_member(index->opfamily[0],
7314 536280 : index->opcintype[0],
7315 536280 : index->opcintype[0],
7316 : BTLessStrategyNumber);
7317 1072560 : if (OidIsValid(sortop) &&
7318 536280 : get_attstatsslot(&sslot, vardata.statsTuple,
7319 : STATISTIC_KIND_CORRELATION, sortop,
7320 : ATTSTATSSLOT_NUMBERS))
7321 : {
7322 : double varCorrelation;
7323 :
7324 : Assert(sslot.nnumbers == 1);
7325 530120 : varCorrelation = sslot.numbers[0];
7326 :
7327 530120 : if (index->reverse_sort[0])
7328 0 : varCorrelation = -varCorrelation;
7329 :
7330 530120 : if (index->nkeycolumns > 1)
7331 181522 : costs.indexCorrelation = varCorrelation * 0.75;
7332 : else
7333 348598 : costs.indexCorrelation = varCorrelation;
7334 :
7335 530120 : free_attstatsslot(&sslot);
7336 : }
7337 : }
7338 :
7339 720902 : ReleaseVariableStats(vardata);
7340 :
7341 720902 : *indexStartupCost = costs.indexStartupCost;
7342 720902 : *indexTotalCost = costs.indexTotalCost;
7343 720902 : *indexSelectivity = costs.indexSelectivity;
7344 720902 : *indexCorrelation = costs.indexCorrelation;
7345 720902 : *indexPages = costs.numIndexPages;
7346 720902 : }
7347 :
7348 : void
7349 410 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7350 : Cost *indexStartupCost, Cost *indexTotalCost,
7351 : Selectivity *indexSelectivity, double *indexCorrelation,
7352 : double *indexPages)
7353 : {
7354 410 : GenericCosts costs = {0};
7355 :
7356 410 : genericcostestimate(root, path, loop_count, &costs);
7357 :
7358 : /*
7359 : * A hash index has no descent costs as such, since the index AM can go
7360 : * directly to the target bucket after computing the hash value. There
7361 : * are a couple of other hash-specific costs that we could conceivably add
7362 : * here, though:
7363 : *
7364 : * Ideally we'd charge spc_random_page_cost for each page in the target
7365 : * bucket, not just the numIndexPages pages that genericcostestimate
7366 : * thought we'd visit. However in most cases we don't know which bucket
7367 : * that will be. There's no point in considering the average bucket size
7368 : * because the hash AM makes sure that's always one page.
7369 : *
7370 : * Likewise, we could consider charging some CPU for each index tuple in
7371 : * the bucket, if we knew how many there were. But the per-tuple cost is
7372 : * just a hash value comparison, not a general datatype-dependent
7373 : * comparison, so any such charge ought to be quite a bit less than
7374 : * cpu_operator_cost; which makes it probably not worth worrying about.
7375 : *
7376 : * A bigger issue is that chance hash-value collisions will result in
7377 : * wasted probes into the heap. We don't currently attempt to model this
7378 : * cost on the grounds that it's rare, but maybe it's not rare enough.
7379 : * (Any fix for this ought to consider the generic lossy-operator problem,
7380 : * though; it's not entirely hash-specific.)
7381 : */
7382 :
7383 410 : *indexStartupCost = costs.indexStartupCost;
7384 410 : *indexTotalCost = costs.indexTotalCost;
7385 410 : *indexSelectivity = costs.indexSelectivity;
7386 410 : *indexCorrelation = costs.indexCorrelation;
7387 410 : *indexPages = costs.numIndexPages;
7388 410 : }
7389 :
7390 : void
7391 4778 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7392 : Cost *indexStartupCost, Cost *indexTotalCost,
7393 : Selectivity *indexSelectivity, double *indexCorrelation,
7394 : double *indexPages)
7395 : {
7396 4778 : IndexOptInfo *index = path->indexinfo;
7397 4778 : GenericCosts costs = {0};
7398 : Cost descentCost;
7399 :
7400 4778 : genericcostestimate(root, path, loop_count, &costs);
7401 :
7402 : /*
7403 : * We model index descent costs similarly to those for btree, but to do
7404 : * that we first need an idea of the tree height. We somewhat arbitrarily
7405 : * assume that the fanout is 100, meaning the tree height is at most
7406 : * log100(index->pages).
7407 : *
7408 : * Although this computation isn't really expensive enough to require
7409 : * caching, we might as well use index->tree_height to cache it.
7410 : */
7411 4778 : if (index->tree_height < 0) /* unknown? */
7412 : {
7413 4764 : if (index->pages > 1) /* avoid computing log(0) */
7414 2704 : index->tree_height = (int) (log(index->pages) / log(100.0));
7415 : else
7416 2060 : index->tree_height = 0;
7417 : }
7418 :
7419 : /*
7420 : * Add a CPU-cost component to represent the costs of initial descent. We
7421 : * just use log(N) here not log2(N) since the branching factor isn't
7422 : * necessarily two anyway. As for btree, charge once per SA scan.
7423 : */
7424 4778 : if (index->tuples > 1) /* avoid computing log(0) */
7425 : {
7426 4778 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7427 4778 : costs.indexStartupCost += descentCost;
7428 4778 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7429 : }
7430 :
7431 : /*
7432 : * Likewise add a per-page charge, calculated the same as for btrees.
7433 : */
7434 4778 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7435 4778 : costs.indexStartupCost += descentCost;
7436 4778 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7437 :
7438 4778 : *indexStartupCost = costs.indexStartupCost;
7439 4778 : *indexTotalCost = costs.indexTotalCost;
7440 4778 : *indexSelectivity = costs.indexSelectivity;
7441 4778 : *indexCorrelation = costs.indexCorrelation;
7442 4778 : *indexPages = costs.numIndexPages;
7443 4778 : }
7444 :
7445 : void
7446 1784 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7447 : Cost *indexStartupCost, Cost *indexTotalCost,
7448 : Selectivity *indexSelectivity, double *indexCorrelation,
7449 : double *indexPages)
7450 : {
7451 1784 : IndexOptInfo *index = path->indexinfo;
7452 1784 : GenericCosts costs = {0};
7453 : Cost descentCost;
7454 :
7455 1784 : genericcostestimate(root, path, loop_count, &costs);
7456 :
7457 : /*
7458 : * We model index descent costs similarly to those for btree, but to do
7459 : * that we first need an idea of the tree height. We somewhat arbitrarily
7460 : * assume that the fanout is 100, meaning the tree height is at most
7461 : * log100(index->pages).
7462 : *
7463 : * Although this computation isn't really expensive enough to require
7464 : * caching, we might as well use index->tree_height to cache it.
7465 : */
7466 1784 : if (index->tree_height < 0) /* unknown? */
7467 : {
7468 1778 : if (index->pages > 1) /* avoid computing log(0) */
7469 1778 : index->tree_height = (int) (log(index->pages) / log(100.0));
7470 : else
7471 0 : index->tree_height = 0;
7472 : }
7473 :
7474 : /*
7475 : * Add a CPU-cost component to represent the costs of initial descent. We
7476 : * just use log(N) here not log2(N) since the branching factor isn't
7477 : * necessarily two anyway. As for btree, charge once per SA scan.
7478 : */
7479 1784 : if (index->tuples > 1) /* avoid computing log(0) */
7480 : {
7481 1784 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7482 1784 : costs.indexStartupCost += descentCost;
7483 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7484 : }
7485 :
7486 : /*
7487 : * Likewise add a per-page charge, calculated the same as for btrees.
7488 : */
7489 1784 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7490 1784 : costs.indexStartupCost += descentCost;
7491 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7492 :
7493 1784 : *indexStartupCost = costs.indexStartupCost;
7494 1784 : *indexTotalCost = costs.indexTotalCost;
7495 1784 : *indexSelectivity = costs.indexSelectivity;
7496 1784 : *indexCorrelation = costs.indexCorrelation;
7497 1784 : *indexPages = costs.numIndexPages;
7498 1784 : }
7499 :
7500 :
7501 : /*
7502 : * Support routines for gincostestimate
7503 : */
7504 :
7505 : typedef struct
7506 : {
7507 : bool attHasFullScan[INDEX_MAX_KEYS];
7508 : bool attHasNormalScan[INDEX_MAX_KEYS];
7509 : double partialEntries;
7510 : double exactEntries;
7511 : double searchEntries;
7512 : double arrayScans;
7513 : } GinQualCounts;
7514 :
7515 : /*
7516 : * Estimate the number of index terms that need to be searched for while
7517 : * testing the given GIN query, and increment the counts in *counts
7518 : * appropriately. If the query is unsatisfiable, return false.
7519 : */
7520 : static bool
7521 2072 : gincost_pattern(IndexOptInfo *index, int indexcol,
7522 : Oid clause_op, Datum query,
7523 : GinQualCounts *counts)
7524 : {
7525 : FmgrInfo flinfo;
7526 : Oid extractProcOid;
7527 : Oid collation;
7528 : int strategy_op;
7529 : Oid lefttype,
7530 : righttype;
7531 2072 : int32 nentries = 0;
7532 2072 : bool *partial_matches = NULL;
7533 2072 : Pointer *extra_data = NULL;
7534 2072 : bool *nullFlags = NULL;
7535 2072 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
7536 : int32 i;
7537 :
7538 : Assert(indexcol < index->nkeycolumns);
7539 :
7540 : /*
7541 : * Get the operator's strategy number and declared input data types within
7542 : * the index opfamily. (We don't need the latter, but we use
7543 : * get_op_opfamily_properties because it will throw error if it fails to
7544 : * find a matching pg_amop entry.)
7545 : */
7546 2072 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
7547 : &strategy_op, &lefttype, &righttype);
7548 :
7549 : /*
7550 : * GIN always uses the "default" support functions, which are those with
7551 : * lefttype == righttype == the opclass' opcintype (see
7552 : * IndexSupportInitialize in relcache.c).
7553 : */
7554 2072 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
7555 2072 : index->opcintype[indexcol],
7556 2072 : index->opcintype[indexcol],
7557 : GIN_EXTRACTQUERY_PROC);
7558 :
7559 2072 : if (!OidIsValid(extractProcOid))
7560 : {
7561 : /* should not happen; throw same error as index_getprocinfo */
7562 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
7563 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
7564 : get_rel_name(index->indexoid));
7565 : }
7566 :
7567 : /*
7568 : * Choose collation to pass to extractProc (should match initGinState).
7569 : */
7570 2072 : if (OidIsValid(index->indexcollations[indexcol]))
7571 364 : collation = index->indexcollations[indexcol];
7572 : else
7573 1708 : collation = DEFAULT_COLLATION_OID;
7574 :
7575 2072 : fmgr_info(extractProcOid, &flinfo);
7576 :
7577 2072 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
7578 :
7579 2072 : FunctionCall7Coll(&flinfo,
7580 : collation,
7581 : query,
7582 : PointerGetDatum(&nentries),
7583 : UInt16GetDatum(strategy_op),
7584 : PointerGetDatum(&partial_matches),
7585 : PointerGetDatum(&extra_data),
7586 : PointerGetDatum(&nullFlags),
7587 : PointerGetDatum(&searchMode));
7588 :
7589 2072 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
7590 : {
7591 : /* No match is possible */
7592 12 : return false;
7593 : }
7594 :
7595 8796 : for (i = 0; i < nentries; i++)
7596 : {
7597 : /*
7598 : * For partial match we haven't any information to estimate number of
7599 : * matched entries in index, so, we just estimate it as 100
7600 : */
7601 6736 : if (partial_matches && partial_matches[i])
7602 316 : counts->partialEntries += 100;
7603 : else
7604 6420 : counts->exactEntries++;
7605 :
7606 6736 : counts->searchEntries++;
7607 : }
7608 :
7609 2060 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
7610 : {
7611 1588 : counts->attHasNormalScan[indexcol] = true;
7612 : }
7613 472 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
7614 : {
7615 : /* Treat "include empty" like an exact-match item */
7616 44 : counts->attHasNormalScan[indexcol] = true;
7617 44 : counts->exactEntries++;
7618 44 : counts->searchEntries++;
7619 : }
7620 : else
7621 : {
7622 : /* It's GIN_SEARCH_MODE_ALL */
7623 428 : counts->attHasFullScan[indexcol] = true;
7624 : }
7625 :
7626 2060 : return true;
7627 : }
7628 :
7629 : /*
7630 : * Estimate the number of index terms that need to be searched for while
7631 : * testing the given GIN index clause, and increment the counts in *counts
7632 : * appropriately. If the query is unsatisfiable, return false.
7633 : */
7634 : static bool
7635 2060 : gincost_opexpr(PlannerInfo *root,
7636 : IndexOptInfo *index,
7637 : int indexcol,
7638 : OpExpr *clause,
7639 : GinQualCounts *counts)
7640 : {
7641 2060 : Oid clause_op = clause->opno;
7642 2060 : Node *operand = (Node *) lsecond(clause->args);
7643 :
7644 : /* aggressively reduce to a constant, and look through relabeling */
7645 2060 : operand = estimate_expression_value(root, operand);
7646 :
7647 2060 : if (IsA(operand, RelabelType))
7648 0 : operand = (Node *) ((RelabelType *) operand)->arg;
7649 :
7650 : /*
7651 : * It's impossible to call extractQuery method for unknown operand. So
7652 : * unless operand is a Const we can't do much; just assume there will be
7653 : * one ordinary search entry from the operand at runtime.
7654 : */
7655 2060 : if (!IsA(operand, Const))
7656 : {
7657 0 : counts->exactEntries++;
7658 0 : counts->searchEntries++;
7659 0 : return true;
7660 : }
7661 :
7662 : /* If Const is null, there can be no matches */
7663 2060 : if (((Const *) operand)->constisnull)
7664 0 : return false;
7665 :
7666 : /* Otherwise, apply extractQuery and get the actual term counts */
7667 2060 : return gincost_pattern(index, indexcol, clause_op,
7668 : ((Const *) operand)->constvalue,
7669 : counts);
7670 : }
7671 :
7672 : /*
7673 : * Estimate the number of index terms that need to be searched for while
7674 : * testing the given GIN index clause, and increment the counts in *counts
7675 : * appropriately. If the query is unsatisfiable, return false.
7676 : *
7677 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
7678 : * each of which involves one value from the RHS array, plus all the
7679 : * non-array quals (if any). To model this, we average the counts across
7680 : * the RHS elements, and add the averages to the counts in *counts (which
7681 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
7682 : * by N, causing gincostestimate to scale up its estimates accordingly.
7683 : */
7684 : static bool
7685 6 : gincost_scalararrayopexpr(PlannerInfo *root,
7686 : IndexOptInfo *index,
7687 : int indexcol,
7688 : ScalarArrayOpExpr *clause,
7689 : double numIndexEntries,
7690 : GinQualCounts *counts)
7691 : {
7692 6 : Oid clause_op = clause->opno;
7693 6 : Node *rightop = (Node *) lsecond(clause->args);
7694 : ArrayType *arrayval;
7695 : int16 elmlen;
7696 : bool elmbyval;
7697 : char elmalign;
7698 : int numElems;
7699 : Datum *elemValues;
7700 : bool *elemNulls;
7701 : GinQualCounts arraycounts;
7702 6 : int numPossible = 0;
7703 : int i;
7704 :
7705 : Assert(clause->useOr);
7706 :
7707 : /* aggressively reduce to a constant, and look through relabeling */
7708 6 : rightop = estimate_expression_value(root, rightop);
7709 :
7710 6 : if (IsA(rightop, RelabelType))
7711 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
7712 :
7713 : /*
7714 : * It's impossible to call extractQuery method for unknown operand. So
7715 : * unless operand is a Const we can't do much; just assume there will be
7716 : * one ordinary search entry from each array entry at runtime, and fall
7717 : * back on a probably-bad estimate of the number of array entries.
7718 : */
7719 6 : if (!IsA(rightop, Const))
7720 : {
7721 0 : counts->exactEntries++;
7722 0 : counts->searchEntries++;
7723 0 : counts->arrayScans *= estimate_array_length(root, rightop);
7724 0 : return true;
7725 : }
7726 :
7727 : /* If Const is null, there can be no matches */
7728 6 : if (((Const *) rightop)->constisnull)
7729 0 : return false;
7730 :
7731 : /* Otherwise, extract the array elements and iterate over them */
7732 6 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
7733 6 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
7734 : &elmlen, &elmbyval, &elmalign);
7735 6 : deconstruct_array(arrayval,
7736 : ARR_ELEMTYPE(arrayval),
7737 : elmlen, elmbyval, elmalign,
7738 : &elemValues, &elemNulls, &numElems);
7739 :
7740 6 : memset(&arraycounts, 0, sizeof(arraycounts));
7741 :
7742 18 : for (i = 0; i < numElems; i++)
7743 : {
7744 : GinQualCounts elemcounts;
7745 :
7746 : /* NULL can't match anything, so ignore, as the executor will */
7747 12 : if (elemNulls[i])
7748 0 : continue;
7749 :
7750 : /* Otherwise, apply extractQuery and get the actual term counts */
7751 12 : memset(&elemcounts, 0, sizeof(elemcounts));
7752 :
7753 12 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
7754 : &elemcounts))
7755 : {
7756 : /* We ignore array elements that are unsatisfiable patterns */
7757 12 : numPossible++;
7758 :
7759 12 : if (elemcounts.attHasFullScan[indexcol] &&
7760 0 : !elemcounts.attHasNormalScan[indexcol])
7761 : {
7762 : /*
7763 : * Full index scan will be required. We treat this as if
7764 : * every key in the index had been listed in the query; is
7765 : * that reasonable?
7766 : */
7767 0 : elemcounts.partialEntries = 0;
7768 0 : elemcounts.exactEntries = numIndexEntries;
7769 0 : elemcounts.searchEntries = numIndexEntries;
7770 : }
7771 12 : arraycounts.partialEntries += elemcounts.partialEntries;
7772 12 : arraycounts.exactEntries += elemcounts.exactEntries;
7773 12 : arraycounts.searchEntries += elemcounts.searchEntries;
7774 : }
7775 : }
7776 :
7777 6 : if (numPossible == 0)
7778 : {
7779 : /* No satisfiable patterns in the array */
7780 0 : return false;
7781 : }
7782 :
7783 : /*
7784 : * Now add the averages to the global counts. This will give us an
7785 : * estimate of the average number of terms searched for in each indexscan,
7786 : * including contributions from both array and non-array quals.
7787 : */
7788 6 : counts->partialEntries += arraycounts.partialEntries / numPossible;
7789 6 : counts->exactEntries += arraycounts.exactEntries / numPossible;
7790 6 : counts->searchEntries += arraycounts.searchEntries / numPossible;
7791 :
7792 6 : counts->arrayScans *= numPossible;
7793 :
7794 6 : return true;
7795 : }
7796 :
7797 : /*
7798 : * GIN has search behavior completely different from other index types
7799 : */
7800 : void
7801 1868 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7802 : Cost *indexStartupCost, Cost *indexTotalCost,
7803 : Selectivity *indexSelectivity, double *indexCorrelation,
7804 : double *indexPages)
7805 : {
7806 1868 : IndexOptInfo *index = path->indexinfo;
7807 1868 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7808 : List *selectivityQuals;
7809 1868 : double numPages = index->pages,
7810 1868 : numTuples = index->tuples;
7811 : double numEntryPages,
7812 : numDataPages,
7813 : numPendingPages,
7814 : numEntries;
7815 : GinQualCounts counts;
7816 : bool matchPossible;
7817 : bool fullIndexScan;
7818 : double partialScale;
7819 : double entryPagesFetched,
7820 : dataPagesFetched,
7821 : dataPagesFetchedBySel;
7822 : double qual_op_cost,
7823 : qual_arg_cost,
7824 : spc_random_page_cost,
7825 : outer_scans;
7826 : Cost descentCost;
7827 : Relation indexRel;
7828 : GinStatsData ginStats;
7829 : ListCell *lc;
7830 : int i;
7831 :
7832 : /*
7833 : * Obtain statistical information from the meta page, if possible. Else
7834 : * set ginStats to zeroes, and we'll cope below.
7835 : */
7836 1868 : if (!index->hypothetical)
7837 : {
7838 : /* Lock should have already been obtained in plancat.c */
7839 1868 : indexRel = index_open(index->indexoid, NoLock);
7840 1868 : ginGetStats(indexRel, &ginStats);
7841 1868 : index_close(indexRel, NoLock);
7842 : }
7843 : else
7844 : {
7845 0 : memset(&ginStats, 0, sizeof(ginStats));
7846 : }
7847 :
7848 : /*
7849 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
7850 : * trusted, but the other fields are data as of the last VACUUM. We can
7851 : * scale them up to account for growth since then, but that method only
7852 : * goes so far; in the worst case, the stats might be for a completely
7853 : * empty index, and scaling them will produce pretty bogus numbers.
7854 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
7855 : * it's grown more than that, fall back to estimating things only from the
7856 : * assumed-accurate index size. But we'll trust nPendingPages in any case
7857 : * so long as it's not clearly insane, ie, more than the index size.
7858 : */
7859 1868 : if (ginStats.nPendingPages < numPages)
7860 1868 : numPendingPages = ginStats.nPendingPages;
7861 : else
7862 0 : numPendingPages = 0;
7863 :
7864 1868 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
7865 1868 : ginStats.nTotalPages > numPages / 4 &&
7866 1816 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
7867 1558 : {
7868 : /*
7869 : * OK, the stats seem close enough to sane to be trusted. But we
7870 : * still need to scale them by the ratio numPages / nTotalPages to
7871 : * account for growth since the last VACUUM.
7872 : */
7873 1558 : double scale = numPages / ginStats.nTotalPages;
7874 :
7875 1558 : numEntryPages = ceil(ginStats.nEntryPages * scale);
7876 1558 : numDataPages = ceil(ginStats.nDataPages * scale);
7877 1558 : numEntries = ceil(ginStats.nEntries * scale);
7878 : /* ensure we didn't round up too much */
7879 1558 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
7880 1558 : numDataPages = Min(numDataPages,
7881 : numPages - numPendingPages - numEntryPages);
7882 : }
7883 : else
7884 : {
7885 : /*
7886 : * We might get here because it's a hypothetical index, or an index
7887 : * created pre-9.1 and never vacuumed since upgrading (in which case
7888 : * its stats would read as zeroes), or just because it's grown too
7889 : * much since the last VACUUM for us to put our faith in scaling.
7890 : *
7891 : * Invent some plausible internal statistics based on the index page
7892 : * count (and clamp that to at least 10 pages, just in case). We
7893 : * estimate that 90% of the index is entry pages, and the rest is data
7894 : * pages. Estimate 100 entries per entry page; this is rather bogus
7895 : * since it'll depend on the size of the keys, but it's more robust
7896 : * than trying to predict the number of entries per heap tuple.
7897 : */
7898 310 : numPages = Max(numPages, 10);
7899 310 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
7900 310 : numDataPages = numPages - numPendingPages - numEntryPages;
7901 310 : numEntries = floor(numEntryPages * 100);
7902 : }
7903 :
7904 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
7905 1868 : if (numEntries < 1)
7906 0 : numEntries = 1;
7907 :
7908 : /*
7909 : * If the index is partial, AND the index predicate with the index-bound
7910 : * quals to produce a more accurate idea of the number of rows covered by
7911 : * the bound conditions.
7912 : */
7913 1868 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7914 :
7915 : /* Estimate the fraction of main-table tuples that will be visited */
7916 3736 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7917 1868 : index->rel->relid,
7918 : JOIN_INNER,
7919 : NULL);
7920 :
7921 : /* fetch estimated page cost for tablespace containing index */
7922 1868 : get_tablespace_page_costs(index->reltablespace,
7923 : &spc_random_page_cost,
7924 : NULL);
7925 :
7926 : /*
7927 : * Generic assumption about index correlation: there isn't any.
7928 : */
7929 1868 : *indexCorrelation = 0.0;
7930 :
7931 : /*
7932 : * Examine quals to estimate number of search entries & partial matches
7933 : */
7934 1868 : memset(&counts, 0, sizeof(counts));
7935 1868 : counts.arrayScans = 1;
7936 1868 : matchPossible = true;
7937 :
7938 3934 : foreach(lc, path->indexclauses)
7939 : {
7940 2066 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7941 : ListCell *lc2;
7942 :
7943 4120 : foreach(lc2, iclause->indexquals)
7944 : {
7945 2066 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7946 2066 : Expr *clause = rinfo->clause;
7947 :
7948 2066 : if (IsA(clause, OpExpr))
7949 : {
7950 2060 : matchPossible = gincost_opexpr(root,
7951 : index,
7952 2060 : iclause->indexcol,
7953 : (OpExpr *) clause,
7954 : &counts);
7955 2060 : if (!matchPossible)
7956 12 : break;
7957 : }
7958 6 : else if (IsA(clause, ScalarArrayOpExpr))
7959 : {
7960 6 : matchPossible = gincost_scalararrayopexpr(root,
7961 : index,
7962 6 : iclause->indexcol,
7963 : (ScalarArrayOpExpr *) clause,
7964 : numEntries,
7965 : &counts);
7966 6 : if (!matchPossible)
7967 0 : break;
7968 : }
7969 : else
7970 : {
7971 : /* shouldn't be anything else for a GIN index */
7972 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
7973 : (int) nodeTag(clause));
7974 : }
7975 : }
7976 : }
7977 :
7978 : /* Fall out if there were any provably-unsatisfiable quals */
7979 1868 : if (!matchPossible)
7980 : {
7981 12 : *indexStartupCost = 0;
7982 12 : *indexTotalCost = 0;
7983 12 : *indexSelectivity = 0;
7984 12 : return;
7985 : }
7986 :
7987 : /*
7988 : * If attribute has a full scan and at the same time doesn't have normal
7989 : * scan, then we'll have to scan all non-null entries of that attribute.
7990 : * Currently, we don't have per-attribute statistics for GIN. Thus, we
7991 : * must assume the whole GIN index has to be scanned in this case.
7992 : */
7993 1856 : fullIndexScan = false;
7994 3602 : for (i = 0; i < index->nkeycolumns; i++)
7995 : {
7996 2084 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
7997 : {
7998 338 : fullIndexScan = true;
7999 338 : break;
8000 : }
8001 : }
8002 :
8003 1856 : if (fullIndexScan || indexQuals == NIL)
8004 : {
8005 : /*
8006 : * Full index scan will be required. We treat this as if every key in
8007 : * the index had been listed in the query; is that reasonable?
8008 : */
8009 338 : counts.partialEntries = 0;
8010 338 : counts.exactEntries = numEntries;
8011 338 : counts.searchEntries = numEntries;
8012 : }
8013 :
8014 : /* Will we have more than one iteration of a nestloop scan? */
8015 1856 : outer_scans = loop_count;
8016 :
8017 : /*
8018 : * Compute cost to begin scan, first of all, pay attention to pending
8019 : * list.
8020 : */
8021 1856 : entryPagesFetched = numPendingPages;
8022 :
8023 : /*
8024 : * Estimate number of entry pages read. We need to do
8025 : * counts.searchEntries searches. Use a power function as it should be,
8026 : * but tuples on leaf pages usually is much greater. Here we include all
8027 : * searches in entry tree, including search of first entry in partial
8028 : * match algorithm
8029 : */
8030 1856 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8031 :
8032 : /*
8033 : * Add an estimate of entry pages read by partial match algorithm. It's a
8034 : * scan over leaf pages in entry tree. We haven't any useful stats here,
8035 : * so estimate it as proportion. Because counts.partialEntries is really
8036 : * pretty bogus (see code above), it's possible that it is more than
8037 : * numEntries; clamp the proportion to ensure sanity.
8038 : */
8039 1856 : partialScale = counts.partialEntries / numEntries;
8040 1856 : partialScale = Min(partialScale, 1.0);
8041 :
8042 1856 : entryPagesFetched += ceil(numEntryPages * partialScale);
8043 :
8044 : /*
8045 : * Partial match algorithm reads all data pages before doing actual scan,
8046 : * so it's a startup cost. Again, we haven't any useful stats here, so
8047 : * estimate it as proportion.
8048 : */
8049 1856 : dataPagesFetched = ceil(numDataPages * partialScale);
8050 :
8051 1856 : *indexStartupCost = 0;
8052 1856 : *indexTotalCost = 0;
8053 :
8054 : /*
8055 : * Add a CPU-cost component to represent the costs of initial entry btree
8056 : * descent. We don't charge any I/O cost for touching upper btree levels,
8057 : * since they tend to stay in cache, but we still have to do about log2(N)
8058 : * comparisons to descend a btree of N leaf tuples. We charge one
8059 : * cpu_operator_cost per comparison.
8060 : *
8061 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
8062 : * ones after the first one are not startup cost so far as the overall
8063 : * plan is concerned, so add them only to "total" cost.
8064 : */
8065 1856 : if (numEntries > 1) /* avoid computing log(0) */
8066 : {
8067 1856 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
8068 1856 : *indexStartupCost += descentCost * counts.searchEntries;
8069 1856 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
8070 : }
8071 :
8072 : /*
8073 : * Add a cpu cost per entry-page fetched. This is not amortized over a
8074 : * loop.
8075 : */
8076 1856 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8077 1856 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8078 :
8079 : /*
8080 : * Add a cpu cost per data-page fetched. This is also not amortized over a
8081 : * loop. Since those are the data pages from the partial match algorithm,
8082 : * charge them as startup cost.
8083 : */
8084 1856 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
8085 :
8086 : /*
8087 : * Since we add the startup cost to the total cost later on, remove the
8088 : * initial arrayscan from the total.
8089 : */
8090 1856 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8091 :
8092 : /*
8093 : * Calculate cache effects if more than one scan due to nestloops or array
8094 : * quals. The result is pro-rated per nestloop scan, but the array qual
8095 : * factor shouldn't be pro-rated (compare genericcostestimate).
8096 : */
8097 1856 : if (outer_scans > 1 || counts.arrayScans > 1)
8098 : {
8099 6 : entryPagesFetched *= outer_scans * counts.arrayScans;
8100 6 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
8101 : (BlockNumber) numEntryPages,
8102 : numEntryPages, root);
8103 6 : entryPagesFetched /= outer_scans;
8104 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8105 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8106 : (BlockNumber) numDataPages,
8107 : numDataPages, root);
8108 6 : dataPagesFetched /= outer_scans;
8109 : }
8110 :
8111 : /*
8112 : * Here we use random page cost because logically-close pages could be far
8113 : * apart on disk.
8114 : */
8115 1856 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
8116 :
8117 : /*
8118 : * Now compute the number of data pages fetched during the scan.
8119 : *
8120 : * We assume every entry to have the same number of items, and that there
8121 : * is no overlap between them. (XXX: tsvector and array opclasses collect
8122 : * statistics on the frequency of individual keys; it would be nice to use
8123 : * those here.)
8124 : */
8125 1856 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
8126 :
8127 : /*
8128 : * If there is a lot of overlap among the entries, in particular if one of
8129 : * the entries is very frequent, the above calculation can grossly
8130 : * under-estimate. As a simple cross-check, calculate a lower bound based
8131 : * on the overall selectivity of the quals. At a minimum, we must read
8132 : * one item pointer for each matching entry.
8133 : *
8134 : * The width of each item pointer varies, based on the level of
8135 : * compression. We don't have statistics on that, but an average of
8136 : * around 3 bytes per item is fairly typical.
8137 : */
8138 1856 : dataPagesFetchedBySel = ceil(*indexSelectivity *
8139 1856 : (numTuples / (BLCKSZ / 3)));
8140 1856 : if (dataPagesFetchedBySel > dataPagesFetched)
8141 1476 : dataPagesFetched = dataPagesFetchedBySel;
8142 :
8143 : /* Add one page cpu-cost to the startup cost */
8144 1856 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
8145 :
8146 : /*
8147 : * Add once again a CPU-cost for those data pages, before amortizing for
8148 : * cache.
8149 : */
8150 1856 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8151 :
8152 : /* Account for cache effects, the same as above */
8153 1856 : if (outer_scans > 1 || counts.arrayScans > 1)
8154 : {
8155 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8156 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8157 : (BlockNumber) numDataPages,
8158 : numDataPages, root);
8159 6 : dataPagesFetched /= outer_scans;
8160 : }
8161 :
8162 : /* And apply random_page_cost as the cost per page */
8163 1856 : *indexTotalCost += *indexStartupCost +
8164 1856 : dataPagesFetched * spc_random_page_cost;
8165 :
8166 : /*
8167 : * Add on index qual eval costs, much as in genericcostestimate. We charge
8168 : * cpu but we can disregard indexorderbys, since GIN doesn't support
8169 : * those.
8170 : */
8171 1856 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8172 1856 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
8173 :
8174 1856 : *indexStartupCost += qual_arg_cost;
8175 1856 : *indexTotalCost += qual_arg_cost;
8176 :
8177 : /*
8178 : * Add a cpu cost per search entry, corresponding to the actual visited
8179 : * entries.
8180 : */
8181 1856 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
8182 : /* Now add a cpu cost per tuple in the posting lists / trees */
8183 1856 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
8184 1856 : *indexPages = dataPagesFetched;
8185 : }
8186 :
8187 : /*
8188 : * BRIN has search behavior completely different from other index types
8189 : */
8190 : void
8191 10730 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8192 : Cost *indexStartupCost, Cost *indexTotalCost,
8193 : Selectivity *indexSelectivity, double *indexCorrelation,
8194 : double *indexPages)
8195 : {
8196 10730 : IndexOptInfo *index = path->indexinfo;
8197 10730 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8198 10730 : double numPages = index->pages;
8199 10730 : RelOptInfo *baserel = index->rel;
8200 10730 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
8201 : Cost spc_seq_page_cost;
8202 : Cost spc_random_page_cost;
8203 : double qual_arg_cost;
8204 : double qualSelectivity;
8205 : BrinStatsData statsData;
8206 : double indexRanges;
8207 : double minimalRanges;
8208 : double estimatedRanges;
8209 : double selec;
8210 : Relation indexRel;
8211 : ListCell *l;
8212 : VariableStatData vardata;
8213 :
8214 : Assert(rte->rtekind == RTE_RELATION);
8215 :
8216 : /* fetch estimated page cost for the tablespace containing the index */
8217 10730 : get_tablespace_page_costs(index->reltablespace,
8218 : &spc_random_page_cost,
8219 : &spc_seq_page_cost);
8220 :
8221 : /*
8222 : * Obtain some data from the index itself, if possible. Otherwise invent
8223 : * some plausible internal statistics based on the relation page count.
8224 : */
8225 10730 : if (!index->hypothetical)
8226 : {
8227 : /*
8228 : * A lock should have already been obtained on the index in plancat.c.
8229 : */
8230 10730 : indexRel = index_open(index->indexoid, NoLock);
8231 10730 : brinGetStats(indexRel, &statsData);
8232 10730 : index_close(indexRel, NoLock);
8233 :
8234 : /* work out the actual number of ranges in the index */
8235 10730 : indexRanges = Max(ceil((double) baserel->pages /
8236 : statsData.pagesPerRange), 1.0);
8237 : }
8238 : else
8239 : {
8240 : /*
8241 : * Assume default number of pages per range, and estimate the number
8242 : * of ranges based on that.
8243 : */
8244 0 : indexRanges = Max(ceil((double) baserel->pages /
8245 : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
8246 :
8247 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
8248 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
8249 : }
8250 :
8251 : /*
8252 : * Compute index correlation
8253 : *
8254 : * Because we can use all index quals equally when scanning, we can use
8255 : * the largest correlation (in absolute value) among columns used by the
8256 : * query. Start at zero, the worst possible case. If we cannot find any
8257 : * correlation statistics, we will keep it as 0.
8258 : */
8259 10730 : *indexCorrelation = 0;
8260 :
8261 21462 : foreach(l, path->indexclauses)
8262 : {
8263 10732 : IndexClause *iclause = lfirst_node(IndexClause, l);
8264 10732 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
8265 :
8266 : /* attempt to lookup stats in relation for this index column */
8267 10732 : if (attnum != 0)
8268 : {
8269 : /* Simple variable -- look to stats for the underlying table */
8270 10732 : if (get_relation_stats_hook &&
8271 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
8272 : {
8273 : /*
8274 : * The hook took control of acquiring a stats tuple. If it
8275 : * did supply a tuple, it'd better have supplied a freefunc.
8276 : */
8277 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
8278 0 : elog(ERROR,
8279 : "no function provided to release variable stats with");
8280 : }
8281 : else
8282 : {
8283 10732 : vardata.statsTuple =
8284 10732 : SearchSysCache3(STATRELATTINH,
8285 : ObjectIdGetDatum(rte->relid),
8286 : Int16GetDatum(attnum),
8287 : BoolGetDatum(false));
8288 10732 : vardata.freefunc = ReleaseSysCache;
8289 : }
8290 : }
8291 : else
8292 : {
8293 : /*
8294 : * Looks like we've found an expression column in the index. Let's
8295 : * see if there's any stats for it.
8296 : */
8297 :
8298 : /* get the attnum from the 0-based index. */
8299 0 : attnum = iclause->indexcol + 1;
8300 :
8301 0 : if (get_index_stats_hook &&
8302 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
8303 : {
8304 : /*
8305 : * The hook took control of acquiring a stats tuple. If it
8306 : * did supply a tuple, it'd better have supplied a freefunc.
8307 : */
8308 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
8309 0 : !vardata.freefunc)
8310 0 : elog(ERROR, "no function provided to release variable stats with");
8311 : }
8312 : else
8313 : {
8314 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
8315 : ObjectIdGetDatum(index->indexoid),
8316 : Int16GetDatum(attnum),
8317 : BoolGetDatum(false));
8318 0 : vardata.freefunc = ReleaseSysCache;
8319 : }
8320 : }
8321 :
8322 10732 : if (HeapTupleIsValid(vardata.statsTuple))
8323 : {
8324 : AttStatsSlot sslot;
8325 :
8326 36 : if (get_attstatsslot(&sslot, vardata.statsTuple,
8327 : STATISTIC_KIND_CORRELATION, InvalidOid,
8328 : ATTSTATSSLOT_NUMBERS))
8329 : {
8330 36 : double varCorrelation = 0.0;
8331 :
8332 36 : if (sslot.nnumbers > 0)
8333 36 : varCorrelation = fabs(sslot.numbers[0]);
8334 :
8335 36 : if (varCorrelation > *indexCorrelation)
8336 36 : *indexCorrelation = varCorrelation;
8337 :
8338 36 : free_attstatsslot(&sslot);
8339 : }
8340 : }
8341 :
8342 10732 : ReleaseVariableStats(vardata);
8343 : }
8344 :
8345 10730 : qualSelectivity = clauselist_selectivity(root, indexQuals,
8346 10730 : baserel->relid,
8347 : JOIN_INNER, NULL);
8348 :
8349 : /*
8350 : * Now calculate the minimum possible ranges we could match with if all of
8351 : * the rows were in the perfect order in the table's heap.
8352 : */
8353 10730 : minimalRanges = ceil(indexRanges * qualSelectivity);
8354 :
8355 : /*
8356 : * Now estimate the number of ranges that we'll touch by using the
8357 : * indexCorrelation from the stats. Careful not to divide by zero (note
8358 : * we're using the absolute value of the correlation).
8359 : */
8360 10730 : if (*indexCorrelation < 1.0e-10)
8361 10694 : estimatedRanges = indexRanges;
8362 : else
8363 36 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
8364 :
8365 : /* we expect to visit this portion of the table */
8366 10730 : selec = estimatedRanges / indexRanges;
8367 :
8368 10730 : CLAMP_PROBABILITY(selec);
8369 :
8370 10730 : *indexSelectivity = selec;
8371 :
8372 : /*
8373 : * Compute the index qual costs, much as in genericcostestimate, to add to
8374 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
8375 : * support those.
8376 : */
8377 10730 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8378 :
8379 : /*
8380 : * Compute the startup cost as the cost to read the whole revmap
8381 : * sequentially, including the cost to execute the index quals.
8382 : */
8383 10730 : *indexStartupCost =
8384 10730 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
8385 10730 : *indexStartupCost += qual_arg_cost;
8386 :
8387 : /*
8388 : * To read a BRIN index there might be a bit of back and forth over
8389 : * regular pages, as revmap might point to them out of sequential order;
8390 : * calculate the total cost as reading the whole index in random order.
8391 : */
8392 10730 : *indexTotalCost = *indexStartupCost +
8393 10730 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
8394 :
8395 : /*
8396 : * Charge a small amount per range tuple which we expect to match to. This
8397 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
8398 : * will set a bit for each page in the range when we find a matching
8399 : * range, so we must multiply the charge by the number of pages in the
8400 : * range.
8401 : */
8402 10730 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
8403 10730 : statsData.pagesPerRange;
8404 :
8405 10730 : *indexPages = index->pages;
8406 10730 : }
|
