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) 1996-2024, 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 "statistics/statistics.h"
125 : #include "storage/bufmgr.h"
126 : #include "utils/acl.h"
127 : #include "utils/array.h"
128 : #include "utils/builtins.h"
129 : #include "utils/date.h"
130 : #include "utils/datum.h"
131 : #include "utils/fmgroids.h"
132 : #include "utils/index_selfuncs.h"
133 : #include "utils/lsyscache.h"
134 : #include "utils/memutils.h"
135 : #include "utils/pg_locale.h"
136 : #include "utils/rel.h"
137 : #include "utils/selfuncs.h"
138 : #include "utils/snapmgr.h"
139 : #include "utils/spccache.h"
140 : #include "utils/syscache.h"
141 : #include "utils/timestamp.h"
142 : #include "utils/typcache.h"
143 :
144 : #define DEFAULT_PAGE_CPU_MULTIPLIER 50.0
145 :
146 : /* Hooks for plugins to get control when we ask for stats */
147 : get_relation_stats_hook_type get_relation_stats_hook = NULL;
148 : get_index_stats_hook_type get_index_stats_hook = NULL;
149 :
150 : static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
151 : static double eqjoinsel_inner(Oid opfuncoid, Oid collation,
152 : VariableStatData *vardata1, VariableStatData *vardata2,
153 : double nd1, double nd2,
154 : bool isdefault1, bool isdefault2,
155 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
156 : Form_pg_statistic stats1, Form_pg_statistic stats2,
157 : bool have_mcvs1, bool have_mcvs2);
158 : static double eqjoinsel_semi(Oid opfuncoid, Oid collation,
159 : VariableStatData *vardata1, VariableStatData *vardata2,
160 : double nd1, double nd2,
161 : bool isdefault1, bool isdefault2,
162 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
163 : Form_pg_statistic stats1, Form_pg_statistic stats2,
164 : bool have_mcvs1, bool have_mcvs2,
165 : RelOptInfo *inner_rel);
166 : static bool estimate_multivariate_ndistinct(PlannerInfo *root,
167 : RelOptInfo *rel, List **varinfos, double *ndistinct);
168 : static bool convert_to_scalar(Datum value, Oid valuetypid, Oid collid,
169 : double *scaledvalue,
170 : Datum lobound, Datum hibound, Oid boundstypid,
171 : double *scaledlobound, double *scaledhibound);
172 : static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure);
173 : static void convert_string_to_scalar(char *value,
174 : double *scaledvalue,
175 : char *lobound,
176 : double *scaledlobound,
177 : char *hibound,
178 : double *scaledhibound);
179 : static void convert_bytea_to_scalar(Datum value,
180 : double *scaledvalue,
181 : Datum lobound,
182 : double *scaledlobound,
183 : Datum hibound,
184 : double *scaledhibound);
185 : static double convert_one_string_to_scalar(char *value,
186 : int rangelo, int rangehi);
187 : static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
188 : int rangelo, int rangehi);
189 : static char *convert_string_datum(Datum value, Oid typid, Oid collid,
190 : bool *failure);
191 : static double convert_timevalue_to_scalar(Datum value, Oid typid,
192 : bool *failure);
193 : static void examine_simple_variable(PlannerInfo *root, Var *var,
194 : VariableStatData *vardata);
195 : static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata,
196 : Oid sortop, Oid collation,
197 : Datum *min, Datum *max);
198 : static void get_stats_slot_range(AttStatsSlot *sslot,
199 : Oid opfuncoid, FmgrInfo *opproc,
200 : Oid collation, int16 typLen, bool typByVal,
201 : Datum *min, Datum *max, bool *p_have_data);
202 : static bool get_actual_variable_range(PlannerInfo *root,
203 : VariableStatData *vardata,
204 : Oid sortop, Oid collation,
205 : Datum *min, Datum *max);
206 : static bool get_actual_variable_endpoint(Relation heapRel,
207 : Relation indexRel,
208 : ScanDirection indexscandir,
209 : ScanKey scankeys,
210 : int16 typLen,
211 : bool typByVal,
212 : TupleTableSlot *tableslot,
213 : MemoryContext outercontext,
214 : Datum *endpointDatum);
215 : static RelOptInfo *find_join_input_rel(PlannerInfo *root, Relids relids);
216 :
217 :
218 : /*
219 : * eqsel - Selectivity of "=" for any data types.
220 : *
221 : * Note: this routine is also used to estimate selectivity for some
222 : * operators that are not "=" but have comparable selectivity behavior,
223 : * such as "~=" (geometric approximate-match). Even for "=", we must
224 : * keep in mind that the left and right datatypes may differ.
225 : */
226 : Datum
227 521522 : eqsel(PG_FUNCTION_ARGS)
228 : {
229 521522 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
230 : }
231 :
232 : /*
233 : * Common code for eqsel() and neqsel()
234 : */
235 : static double
236 562470 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
237 : {
238 562470 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
239 562470 : Oid operator = PG_GETARG_OID(1);
240 562470 : List *args = (List *) PG_GETARG_POINTER(2);
241 562470 : int varRelid = PG_GETARG_INT32(3);
242 562470 : Oid collation = PG_GET_COLLATION();
243 : VariableStatData vardata;
244 : Node *other;
245 : bool varonleft;
246 : double selec;
247 :
248 : /*
249 : * When asked about <>, we do the estimation using the corresponding =
250 : * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
251 : */
252 562470 : if (negate)
253 : {
254 40948 : operator = get_negator(operator);
255 40948 : if (!OidIsValid(operator))
256 : {
257 : /* Use default selectivity (should we raise an error instead?) */
258 0 : return 1.0 - DEFAULT_EQ_SEL;
259 : }
260 : }
261 :
262 : /*
263 : * If expression is not variable = something or something = variable, then
264 : * punt and return a default estimate.
265 : */
266 562470 : if (!get_restriction_variable(root, args, varRelid,
267 : &vardata, &other, &varonleft))
268 3686 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
269 :
270 : /*
271 : * We can do a lot better if the something is a constant. (Note: the
272 : * Const might result from estimation rather than being a simple constant
273 : * in the query.)
274 : */
275 558784 : if (IsA(other, Const))
276 246424 : selec = var_eq_const(&vardata, operator, collation,
277 246424 : ((Const *) other)->constvalue,
278 246424 : ((Const *) other)->constisnull,
279 : varonleft, negate);
280 : else
281 312360 : selec = var_eq_non_const(&vardata, operator, collation, other,
282 : varonleft, negate);
283 :
284 558784 : ReleaseVariableStats(vardata);
285 :
286 558784 : return selec;
287 : }
288 :
289 : /*
290 : * var_eq_const --- eqsel for var = const case
291 : *
292 : * This is exported so that some other estimation functions can use it.
293 : */
294 : double
295 290430 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
296 : Datum constval, bool constisnull,
297 : bool varonleft, bool negate)
298 : {
299 : double selec;
300 290430 : double nullfrac = 0.0;
301 : bool isdefault;
302 : Oid opfuncoid;
303 :
304 : /*
305 : * If the constant is NULL, assume operator is strict and return zero, ie,
306 : * operator will never return TRUE. (It's zero even for a negator op.)
307 : */
308 290430 : if (constisnull)
309 320 : return 0.0;
310 :
311 : /*
312 : * Grab the nullfrac for use below. Note we allow use of nullfrac
313 : * regardless of security check.
314 : */
315 290110 : if (HeapTupleIsValid(vardata->statsTuple))
316 : {
317 : Form_pg_statistic stats;
318 :
319 213580 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
320 213580 : nullfrac = stats->stanullfrac;
321 : }
322 :
323 : /*
324 : * If we matched the var to a unique index or DISTINCT clause, assume
325 : * there is exactly one match regardless of anything else. (This is
326 : * slightly bogus, since the index or clause's equality operator might be
327 : * different from ours, but it's much more likely to be right than
328 : * ignoring the information.)
329 : */
330 290110 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
331 : {
332 76668 : selec = 1.0 / vardata->rel->tuples;
333 : }
334 367558 : else if (HeapTupleIsValid(vardata->statsTuple) &&
335 154116 : statistic_proc_security_check(vardata,
336 154116 : (opfuncoid = get_opcode(oproid))))
337 154116 : {
338 : AttStatsSlot sslot;
339 154116 : bool match = false;
340 : int i;
341 :
342 : /*
343 : * Is the constant "=" to any of the column's most common values?
344 : * (Although the given operator may not really be "=", we will assume
345 : * that seeing whether it returns TRUE is an appropriate test. If you
346 : * don't like this, maybe you shouldn't be using eqsel for your
347 : * operator...)
348 : */
349 154116 : if (get_attstatsslot(&sslot, vardata->statsTuple,
350 : STATISTIC_KIND_MCV, InvalidOid,
351 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
352 : {
353 137930 : LOCAL_FCINFO(fcinfo, 2);
354 : FmgrInfo eqproc;
355 :
356 137930 : fmgr_info(opfuncoid, &eqproc);
357 :
358 : /*
359 : * Save a few cycles by setting up the fcinfo struct just once.
360 : * Using FunctionCallInvoke directly also avoids failure if the
361 : * eqproc returns NULL, though really equality functions should
362 : * never do that.
363 : */
364 137930 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
365 : NULL, NULL);
366 137930 : fcinfo->args[0].isnull = false;
367 137930 : fcinfo->args[1].isnull = false;
368 : /* be careful to apply operator right way 'round */
369 137930 : if (varonleft)
370 137898 : fcinfo->args[1].value = constval;
371 : else
372 32 : fcinfo->args[0].value = constval;
373 :
374 1935032 : for (i = 0; i < sslot.nvalues; i++)
375 : {
376 : Datum fresult;
377 :
378 1865910 : if (varonleft)
379 1865854 : fcinfo->args[0].value = sslot.values[i];
380 : else
381 56 : fcinfo->args[1].value = sslot.values[i];
382 1865910 : fcinfo->isnull = false;
383 1865910 : fresult = FunctionCallInvoke(fcinfo);
384 1865910 : if (!fcinfo->isnull && DatumGetBool(fresult))
385 : {
386 68808 : match = true;
387 68808 : break;
388 : }
389 : }
390 : }
391 : else
392 : {
393 : /* no most-common-value info available */
394 16186 : i = 0; /* keep compiler quiet */
395 : }
396 :
397 154116 : if (match)
398 : {
399 : /*
400 : * Constant is "=" to this common value. We know selectivity
401 : * exactly (or as exactly as ANALYZE could calculate it, anyway).
402 : */
403 68808 : selec = sslot.numbers[i];
404 : }
405 : else
406 : {
407 : /*
408 : * Comparison is against a constant that is neither NULL nor any
409 : * of the common values. Its selectivity cannot be more than
410 : * this:
411 : */
412 85308 : double sumcommon = 0.0;
413 : double otherdistinct;
414 :
415 1731824 : for (i = 0; i < sslot.nnumbers; i++)
416 1646516 : sumcommon += sslot.numbers[i];
417 85308 : selec = 1.0 - sumcommon - nullfrac;
418 85308 : CLAMP_PROBABILITY(selec);
419 :
420 : /*
421 : * and in fact it's probably a good deal less. We approximate that
422 : * all the not-common values share this remaining fraction
423 : * equally, so we divide by the number of other distinct values.
424 : */
425 85308 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
426 85308 : sslot.nnumbers;
427 85308 : if (otherdistinct > 1)
428 37720 : selec /= otherdistinct;
429 :
430 : /*
431 : * Another cross-check: selectivity shouldn't be estimated as more
432 : * than the least common "most common value".
433 : */
434 85308 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
435 0 : selec = sslot.numbers[sslot.nnumbers - 1];
436 : }
437 :
438 154116 : free_attstatsslot(&sslot);
439 : }
440 : else
441 : {
442 : /*
443 : * No ANALYZE stats available, so make a guess using estimated number
444 : * of distinct values and assuming they are equally common. (The guess
445 : * is unlikely to be very good, but we do know a few special cases.)
446 : */
447 59326 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
448 : }
449 :
450 : /* now adjust if we wanted <> rather than = */
451 290110 : if (negate)
452 33508 : selec = 1.0 - selec - nullfrac;
453 :
454 : /* result should be in range, but make sure... */
455 290110 : CLAMP_PROBABILITY(selec);
456 :
457 290110 : return selec;
458 : }
459 :
460 : /*
461 : * var_eq_non_const --- eqsel for var = something-other-than-const case
462 : *
463 : * This is exported so that some other estimation functions can use it.
464 : */
465 : double
466 312360 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
467 : Node *other,
468 : bool varonleft, bool negate)
469 : {
470 : double selec;
471 312360 : double nullfrac = 0.0;
472 : bool isdefault;
473 :
474 : /*
475 : * Grab the nullfrac for use below.
476 : */
477 312360 : if (HeapTupleIsValid(vardata->statsTuple))
478 : {
479 : Form_pg_statistic stats;
480 :
481 220998 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
482 220998 : nullfrac = stats->stanullfrac;
483 : }
484 :
485 : /*
486 : * If we matched the var to a unique index or DISTINCT clause, assume
487 : * there is exactly one match regardless of anything else. (This is
488 : * slightly bogus, since the index or clause's equality operator might be
489 : * different from ours, but it's much more likely to be right than
490 : * ignoring the information.)
491 : */
492 312360 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
493 : {
494 124770 : selec = 1.0 / vardata->rel->tuples;
495 : }
496 187590 : else if (HeapTupleIsValid(vardata->statsTuple))
497 : {
498 : double ndistinct;
499 : AttStatsSlot sslot;
500 :
501 : /*
502 : * Search is for a value that we do not know a priori, but we will
503 : * assume it is not NULL. Estimate the selectivity as non-null
504 : * fraction divided by number of distinct values, so that we get a
505 : * result averaged over all possible values whether common or
506 : * uncommon. (Essentially, we are assuming that the not-yet-known
507 : * comparison value is equally likely to be any of the possible
508 : * values, regardless of their frequency in the table. Is that a good
509 : * idea?)
510 : */
511 110116 : selec = 1.0 - nullfrac;
512 110116 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
513 110116 : if (ndistinct > 1)
514 106018 : selec /= ndistinct;
515 :
516 : /*
517 : * Cross-check: selectivity should never be estimated as more than the
518 : * most common value's.
519 : */
520 110116 : if (get_attstatsslot(&sslot, vardata->statsTuple,
521 : STATISTIC_KIND_MCV, InvalidOid,
522 : ATTSTATSSLOT_NUMBERS))
523 : {
524 91958 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
525 588 : selec = sslot.numbers[0];
526 91958 : free_attstatsslot(&sslot);
527 : }
528 : }
529 : else
530 : {
531 : /*
532 : * No ANALYZE stats available, so make a guess using estimated number
533 : * of distinct values and assuming they are equally common. (The guess
534 : * is unlikely to be very good, but we do know a few special cases.)
535 : */
536 77474 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
537 : }
538 :
539 : /* now adjust if we wanted <> rather than = */
540 312360 : if (negate)
541 5590 : selec = 1.0 - selec - nullfrac;
542 :
543 : /* result should be in range, but make sure... */
544 312360 : CLAMP_PROBABILITY(selec);
545 :
546 312360 : return selec;
547 : }
548 :
549 : /*
550 : * neqsel - Selectivity of "!=" for any data types.
551 : *
552 : * This routine is also used for some operators that are not "!="
553 : * but have comparable selectivity behavior. See above comments
554 : * for eqsel().
555 : */
556 : Datum
557 40948 : neqsel(PG_FUNCTION_ARGS)
558 : {
559 40948 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
560 : }
561 :
562 : /*
563 : * scalarineqsel - Selectivity of "<", "<=", ">", ">=" for scalars.
564 : *
565 : * This is the guts of scalarltsel/scalarlesel/scalargtsel/scalargesel.
566 : * The isgt and iseq flags distinguish which of the four cases apply.
567 : *
568 : * The caller has commuted the clause, if necessary, so that we can treat
569 : * the variable as being on the left. The caller must also make sure that
570 : * the other side of the clause is a non-null Const, and dissect that into
571 : * a value and datatype. (This definition simplifies some callers that
572 : * want to estimate against a computed value instead of a Const node.)
573 : *
574 : * This routine works for any datatype (or pair of datatypes) known to
575 : * convert_to_scalar(). If it is applied to some other datatype,
576 : * it will return an approximate estimate based on assuming that the constant
577 : * value falls in the middle of the bin identified by binary search.
578 : */
579 : static double
580 266244 : scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
581 : Oid collation,
582 : VariableStatData *vardata, Datum constval, Oid consttype)
583 : {
584 : Form_pg_statistic stats;
585 : FmgrInfo opproc;
586 : double mcv_selec,
587 : hist_selec,
588 : sumcommon;
589 : double selec;
590 :
591 266244 : if (!HeapTupleIsValid(vardata->statsTuple))
592 : {
593 : /*
594 : * No stats are available. Typically this means we have to fall back
595 : * on the default estimate; but if the variable is CTID then we can
596 : * make an estimate based on comparing the constant to the table size.
597 : */
598 19230 : if (vardata->var && IsA(vardata->var, Var) &&
599 14986 : ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
600 : {
601 : ItemPointer itemptr;
602 : double block;
603 : double density;
604 :
605 : /*
606 : * If the relation's empty, we're going to include all of it.
607 : * (This is mostly to avoid divide-by-zero below.)
608 : */
609 214 : if (vardata->rel->pages == 0)
610 0 : return 1.0;
611 :
612 214 : itemptr = (ItemPointer) DatumGetPointer(constval);
613 214 : block = ItemPointerGetBlockNumberNoCheck(itemptr);
614 :
615 : /*
616 : * Determine the average number of tuples per page (density).
617 : *
618 : * Since the last page will, on average, be only half full, we can
619 : * estimate it to have half as many tuples as earlier pages. So
620 : * give it half the weight of a regular page.
621 : */
622 214 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
623 :
624 : /* If target is the last page, use half the density. */
625 214 : if (block >= vardata->rel->pages - 1)
626 30 : density *= 0.5;
627 :
628 : /*
629 : * Using the average tuples per page, calculate how far into the
630 : * page the itemptr is likely to be and adjust block accordingly,
631 : * by adding that fraction of a whole block (but never more than a
632 : * whole block, no matter how high the itemptr's offset is). Here
633 : * we are ignoring the possibility of dead-tuple line pointers,
634 : * which is fairly bogus, but we lack the info to do better.
635 : */
636 214 : if (density > 0.0)
637 : {
638 214 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
639 :
640 214 : block += Min(offset / density, 1.0);
641 : }
642 :
643 : /*
644 : * Convert relative block number to selectivity. Again, the last
645 : * page has only half weight.
646 : */
647 214 : selec = block / (vardata->rel->pages - 0.5);
648 :
649 : /*
650 : * The calculation so far gave us a selectivity for the "<=" case.
651 : * We'll have one fewer tuple for "<" and one additional tuple for
652 : * ">=", the latter of which we'll reverse the selectivity for
653 : * below, so we can simply subtract one tuple for both cases. The
654 : * cases that need this adjustment can be identified by iseq being
655 : * equal to isgt.
656 : */
657 214 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
658 102 : selec -= (1.0 / vardata->rel->tuples);
659 :
660 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
661 214 : if (isgt)
662 100 : selec = 1.0 - selec;
663 :
664 214 : CLAMP_PROBABILITY(selec);
665 214 : return selec;
666 : }
667 :
668 : /* no stats available, so default result */
669 19016 : return DEFAULT_INEQ_SEL;
670 : }
671 247014 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
672 :
673 247014 : fmgr_info(get_opcode(operator), &opproc);
674 :
675 : /*
676 : * If we have most-common-values info, add up the fractions of the MCV
677 : * entries that satisfy MCV OP CONST. These fractions contribute directly
678 : * to the result selectivity. Also add up the total fraction represented
679 : * by MCV entries.
680 : */
681 247014 : mcv_selec = mcv_selectivity(vardata, &opproc, collation, constval, true,
682 : &sumcommon);
683 :
684 : /*
685 : * If there is a histogram, determine which bin the constant falls in, and
686 : * compute the resulting contribution to selectivity.
687 : */
688 247014 : hist_selec = ineq_histogram_selectivity(root, vardata,
689 : operator, &opproc, isgt, iseq,
690 : collation,
691 : constval, consttype);
692 :
693 : /*
694 : * Now merge the results from the MCV and histogram calculations,
695 : * realizing that the histogram covers only the non-null values that are
696 : * not listed in MCV.
697 : */
698 247014 : selec = 1.0 - stats->stanullfrac - sumcommon;
699 :
700 247014 : if (hist_selec >= 0.0)
701 193134 : selec *= hist_selec;
702 : else
703 : {
704 : /*
705 : * If no histogram but there are values not accounted for by MCV,
706 : * arbitrarily assume half of them will match.
707 : */
708 53880 : selec *= 0.5;
709 : }
710 :
711 247014 : selec += mcv_selec;
712 :
713 : /* result should be in range, but make sure... */
714 247014 : CLAMP_PROBABILITY(selec);
715 :
716 247014 : return selec;
717 : }
718 :
719 : /*
720 : * mcv_selectivity - Examine the MCV list for selectivity estimates
721 : *
722 : * Determine the fraction of the variable's MCV population that satisfies
723 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft. Also
724 : * compute the fraction of the total column population represented by the MCV
725 : * list. This code will work for any boolean-returning predicate operator.
726 : *
727 : * The function result is the MCV selectivity, and the fraction of the
728 : * total population is returned into *sumcommonp. Zeroes are returned
729 : * if there is no MCV list.
730 : */
731 : double
732 252014 : mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation,
733 : Datum constval, bool varonleft,
734 : double *sumcommonp)
735 : {
736 : double mcv_selec,
737 : sumcommon;
738 : AttStatsSlot sslot;
739 : int i;
740 :
741 252014 : mcv_selec = 0.0;
742 252014 : sumcommon = 0.0;
743 :
744 501760 : if (HeapTupleIsValid(vardata->statsTuple) &&
745 499450 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
746 249704 : get_attstatsslot(&sslot, vardata->statsTuple,
747 : STATISTIC_KIND_MCV, InvalidOid,
748 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
749 : {
750 112198 : LOCAL_FCINFO(fcinfo, 2);
751 :
752 : /*
753 : * We invoke the opproc "by hand" so that we won't fail on NULL
754 : * results. Such cases won't arise for normal comparison functions,
755 : * but generic_restriction_selectivity could perhaps be used with
756 : * operators that can return NULL. A small side benefit is to not
757 : * need to re-initialize the fcinfo struct from scratch each time.
758 : */
759 112198 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
760 : NULL, NULL);
761 112198 : fcinfo->args[0].isnull = false;
762 112198 : fcinfo->args[1].isnull = false;
763 : /* be careful to apply operator right way 'round */
764 112198 : if (varonleft)
765 112198 : fcinfo->args[1].value = constval;
766 : else
767 0 : fcinfo->args[0].value = constval;
768 :
769 3044232 : for (i = 0; i < sslot.nvalues; i++)
770 : {
771 : Datum fresult;
772 :
773 2932034 : if (varonleft)
774 2932034 : fcinfo->args[0].value = sslot.values[i];
775 : else
776 0 : fcinfo->args[1].value = sslot.values[i];
777 2932034 : fcinfo->isnull = false;
778 2932034 : fresult = FunctionCallInvoke(fcinfo);
779 2932034 : if (!fcinfo->isnull && DatumGetBool(fresult))
780 1215120 : mcv_selec += sslot.numbers[i];
781 2932034 : sumcommon += sslot.numbers[i];
782 : }
783 112198 : free_attstatsslot(&sslot);
784 : }
785 :
786 252014 : *sumcommonp = sumcommon;
787 252014 : return mcv_selec;
788 : }
789 :
790 : /*
791 : * histogram_selectivity - Examine the histogram for selectivity estimates
792 : *
793 : * Determine the fraction of the variable's histogram entries that satisfy
794 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft.
795 : *
796 : * This code will work for any boolean-returning predicate operator, whether
797 : * or not it has anything to do with the histogram sort operator. We are
798 : * essentially using the histogram just as a representative sample. However,
799 : * small histograms are unlikely to be all that representative, so the caller
800 : * should be prepared to fall back on some other estimation approach when the
801 : * histogram is missing or very small. It may also be prudent to combine this
802 : * approach with another one when the histogram is small.
803 : *
804 : * If the actual histogram size is not at least min_hist_size, we won't bother
805 : * to do the calculation at all. Also, if the n_skip parameter is > 0, we
806 : * ignore the first and last n_skip histogram elements, on the grounds that
807 : * they are outliers and hence not very representative. Typical values for
808 : * these parameters are 10 and 1.
809 : *
810 : * The function result is the selectivity, or -1 if there is no histogram
811 : * or it's smaller than min_hist_size.
812 : *
813 : * The output parameter *hist_size receives the actual histogram size,
814 : * or zero if no histogram. Callers may use this number to decide how
815 : * much faith to put in the function result.
816 : *
817 : * Note that the result disregards both the most-common-values (if any) and
818 : * null entries. The caller is expected to combine this result with
819 : * statistics for those portions of the column population. It may also be
820 : * prudent to clamp the result range, ie, disbelieve exact 0 or 1 outputs.
821 : */
822 : double
823 5000 : histogram_selectivity(VariableStatData *vardata,
824 : FmgrInfo *opproc, Oid collation,
825 : Datum constval, bool varonleft,
826 : int min_hist_size, int n_skip,
827 : int *hist_size)
828 : {
829 : double result;
830 : AttStatsSlot sslot;
831 :
832 : /* check sanity of parameters */
833 : Assert(n_skip >= 0);
834 : Assert(min_hist_size > 2 * n_skip);
835 :
836 7732 : if (HeapTupleIsValid(vardata->statsTuple) &&
837 5458 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
838 2726 : get_attstatsslot(&sslot, vardata->statsTuple,
839 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
840 : ATTSTATSSLOT_VALUES))
841 : {
842 2632 : *hist_size = sslot.nvalues;
843 2632 : if (sslot.nvalues >= min_hist_size)
844 : {
845 1550 : LOCAL_FCINFO(fcinfo, 2);
846 1550 : int nmatch = 0;
847 : int i;
848 :
849 : /*
850 : * We invoke the opproc "by hand" so that we won't fail on NULL
851 : * results. Such cases won't arise for normal comparison
852 : * functions, but generic_restriction_selectivity could perhaps be
853 : * used with operators that can return NULL. A small side benefit
854 : * is to not need to re-initialize the fcinfo struct from scratch
855 : * each time.
856 : */
857 1550 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
858 : NULL, NULL);
859 1550 : fcinfo->args[0].isnull = false;
860 1550 : fcinfo->args[1].isnull = false;
861 : /* be careful to apply operator right way 'round */
862 1550 : if (varonleft)
863 1550 : fcinfo->args[1].value = constval;
864 : else
865 0 : fcinfo->args[0].value = constval;
866 :
867 134080 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
868 : {
869 : Datum fresult;
870 :
871 132530 : if (varonleft)
872 132530 : fcinfo->args[0].value = sslot.values[i];
873 : else
874 0 : fcinfo->args[1].value = sslot.values[i];
875 132530 : fcinfo->isnull = false;
876 132530 : fresult = FunctionCallInvoke(fcinfo);
877 132530 : if (!fcinfo->isnull && DatumGetBool(fresult))
878 5424 : nmatch++;
879 : }
880 1550 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
881 : }
882 : else
883 1082 : result = -1;
884 2632 : free_attstatsslot(&sslot);
885 : }
886 : else
887 : {
888 2368 : *hist_size = 0;
889 2368 : result = -1;
890 : }
891 :
892 5000 : return result;
893 : }
894 :
895 : /*
896 : * generic_restriction_selectivity - Selectivity for almost anything
897 : *
898 : * This function estimates selectivity for operators that we don't have any
899 : * special knowledge about, but are on data types that we collect standard
900 : * MCV and/or histogram statistics for. (Additional assumptions are that
901 : * the operator is strict and immutable, or at least stable.)
902 : *
903 : * If we have "VAR OP CONST" or "CONST OP VAR", selectivity is estimated by
904 : * applying the operator to each element of the column's MCV and/or histogram
905 : * stats, and merging the results using the assumption that the histogram is
906 : * a reasonable random sample of the column's non-MCV population. Note that
907 : * if the operator's semantics are related to the histogram ordering, this
908 : * might not be such a great assumption; other functions such as
909 : * scalarineqsel() are probably a better match in such cases.
910 : *
911 : * Otherwise, fall back to the default selectivity provided by the caller.
912 : */
913 : double
914 1106 : generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation,
915 : List *args, int varRelid,
916 : double default_selectivity)
917 : {
918 : double selec;
919 : VariableStatData vardata;
920 : Node *other;
921 : bool varonleft;
922 :
923 : /*
924 : * If expression is not variable OP something or something OP variable,
925 : * then punt and return the default estimate.
926 : */
927 1106 : if (!get_restriction_variable(root, args, varRelid,
928 : &vardata, &other, &varonleft))
929 0 : return default_selectivity;
930 :
931 : /*
932 : * If the something is a NULL constant, assume operator is strict and
933 : * return zero, ie, operator will never return TRUE.
934 : */
935 1106 : if (IsA(other, Const) &&
936 1106 : ((Const *) other)->constisnull)
937 : {
938 0 : ReleaseVariableStats(vardata);
939 0 : return 0.0;
940 : }
941 :
942 1106 : if (IsA(other, Const))
943 : {
944 : /* Variable is being compared to a known non-null constant */
945 1106 : Datum constval = ((Const *) other)->constvalue;
946 : FmgrInfo opproc;
947 : double mcvsum;
948 : double mcvsel;
949 : double nullfrac;
950 : int hist_size;
951 :
952 1106 : fmgr_info(get_opcode(oproid), &opproc);
953 :
954 : /*
955 : * Calculate the selectivity for the column's most common values.
956 : */
957 1106 : mcvsel = mcv_selectivity(&vardata, &opproc, collation,
958 : constval, varonleft,
959 : &mcvsum);
960 :
961 : /*
962 : * If the histogram is large enough, see what fraction of it matches
963 : * the query, and assume that's representative of the non-MCV
964 : * population. Otherwise use the default selectivity for the non-MCV
965 : * population.
966 : */
967 1106 : selec = histogram_selectivity(&vardata, &opproc, collation,
968 : constval, varonleft,
969 : 10, 1, &hist_size);
970 1106 : if (selec < 0)
971 : {
972 : /* Nope, fall back on default */
973 1106 : selec = default_selectivity;
974 : }
975 0 : else if (hist_size < 100)
976 : {
977 : /*
978 : * For histogram sizes from 10 to 100, we combine the histogram
979 : * and default selectivities, putting increasingly more trust in
980 : * the histogram for larger sizes.
981 : */
982 0 : double hist_weight = hist_size / 100.0;
983 :
984 0 : selec = selec * hist_weight +
985 0 : default_selectivity * (1.0 - hist_weight);
986 : }
987 :
988 : /* In any case, don't believe extremely small or large estimates. */
989 1106 : if (selec < 0.0001)
990 0 : selec = 0.0001;
991 1106 : else if (selec > 0.9999)
992 0 : selec = 0.9999;
993 :
994 : /* Don't forget to account for nulls. */
995 1106 : if (HeapTupleIsValid(vardata.statsTuple))
996 84 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
997 : else
998 1022 : nullfrac = 0.0;
999 :
1000 : /*
1001 : * Now merge the results from the MCV and histogram calculations,
1002 : * realizing that the histogram covers only the non-null values that
1003 : * are not listed in MCV.
1004 : */
1005 1106 : selec *= 1.0 - nullfrac - mcvsum;
1006 1106 : selec += mcvsel;
1007 : }
1008 : else
1009 : {
1010 : /* Comparison value is not constant, so we can't do anything */
1011 0 : selec = default_selectivity;
1012 : }
1013 :
1014 1106 : ReleaseVariableStats(vardata);
1015 :
1016 : /* result should be in range, but make sure... */
1017 1106 : CLAMP_PROBABILITY(selec);
1018 :
1019 1106 : return selec;
1020 : }
1021 :
1022 : /*
1023 : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
1024 : *
1025 : * Determine the fraction of the variable's histogram population that
1026 : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
1027 : * The isgt and iseq flags distinguish which of the four cases apply.
1028 : *
1029 : * While opproc could be looked up from the operator OID, common callers
1030 : * also need to call it separately, so we make the caller pass both.
1031 : *
1032 : * Returns -1 if there is no histogram (valid results will always be >= 0).
1033 : *
1034 : * Note that the result disregards both the most-common-values (if any) and
1035 : * null entries. The caller is expected to combine this result with
1036 : * statistics for those portions of the column population.
1037 : *
1038 : * This is exported so that some other estimation functions can use it.
1039 : */
1040 : double
1041 250182 : ineq_histogram_selectivity(PlannerInfo *root,
1042 : VariableStatData *vardata,
1043 : Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq,
1044 : Oid collation,
1045 : Datum constval, Oid consttype)
1046 : {
1047 : double hist_selec;
1048 : AttStatsSlot sslot;
1049 :
1050 250182 : hist_selec = -1.0;
1051 :
1052 : /*
1053 : * Someday, ANALYZE might store more than one histogram per rel/att,
1054 : * corresponding to more than one possible sort ordering defined for the
1055 : * column type. Right now, we know there is only one, so just grab it and
1056 : * see if it matches the query.
1057 : *
1058 : * Note that we can't use opoid as search argument; the staop appearing in
1059 : * pg_statistic will be for the relevant '<' operator, but what we have
1060 : * might be some other inequality operator such as '>='. (Even if opoid
1061 : * is a '<' operator, it could be cross-type.) Hence we must use
1062 : * comparison_ops_are_compatible() to see if the operators match.
1063 : */
1064 499802 : if (HeapTupleIsValid(vardata->statsTuple) &&
1065 499204 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1066 249584 : get_attstatsslot(&sslot, vardata->statsTuple,
1067 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1068 : ATTSTATSSLOT_VALUES))
1069 : {
1070 195738 : if (sslot.nvalues > 1 &&
1071 391400 : sslot.stacoll == collation &&
1072 195662 : comparison_ops_are_compatible(sslot.staop, opoid))
1073 195554 : {
1074 : /*
1075 : * Use binary search to find the desired location, namely the
1076 : * right end of the histogram bin containing the comparison value,
1077 : * which is the leftmost entry for which the comparison operator
1078 : * succeeds (if isgt) or fails (if !isgt).
1079 : *
1080 : * In this loop, we pay no attention to whether the operator iseq
1081 : * or not; that detail will be mopped up below. (We cannot tell,
1082 : * anyway, whether the operator thinks the values are equal.)
1083 : *
1084 : * If the binary search accesses the first or last histogram
1085 : * entry, we try to replace that endpoint with the true column min
1086 : * or max as found by get_actual_variable_range(). This
1087 : * ameliorates misestimates when the min or max is moving as a
1088 : * result of changes since the last ANALYZE. Note that this could
1089 : * result in effectively including MCVs into the histogram that
1090 : * weren't there before, but we don't try to correct for that.
1091 : */
1092 : double histfrac;
1093 195554 : int lobound = 0; /* first possible slot to search */
1094 195554 : int hibound = sslot.nvalues; /* last+1 slot to search */
1095 195554 : bool have_end = false;
1096 :
1097 : /*
1098 : * If there are only two histogram entries, we'll want up-to-date
1099 : * values for both. (If there are more than two, we need at most
1100 : * one of them to be updated, so we deal with that within the
1101 : * loop.)
1102 : */
1103 195554 : if (sslot.nvalues == 2)
1104 1754 : have_end = get_actual_variable_range(root,
1105 : vardata,
1106 : sslot.staop,
1107 : collation,
1108 : &sslot.values[0],
1109 1754 : &sslot.values[1]);
1110 :
1111 1294784 : while (lobound < hibound)
1112 : {
1113 1099230 : int probe = (lobound + hibound) / 2;
1114 : bool ltcmp;
1115 :
1116 : /*
1117 : * If we find ourselves about to compare to the first or last
1118 : * histogram entry, first try to replace it with the actual
1119 : * current min or max (unless we already did so above).
1120 : */
1121 1099230 : if (probe == 0 && sslot.nvalues > 2)
1122 96308 : have_end = get_actual_variable_range(root,
1123 : vardata,
1124 : sslot.staop,
1125 : collation,
1126 : &sslot.values[0],
1127 : NULL);
1128 1002922 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1129 65638 : have_end = get_actual_variable_range(root,
1130 : vardata,
1131 : sslot.staop,
1132 : collation,
1133 : NULL,
1134 65638 : &sslot.values[probe]);
1135 :
1136 1099230 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1137 : collation,
1138 1099230 : sslot.values[probe],
1139 : constval));
1140 1099230 : if (isgt)
1141 55618 : ltcmp = !ltcmp;
1142 1099230 : if (ltcmp)
1143 413704 : lobound = probe + 1;
1144 : else
1145 685526 : hibound = probe;
1146 : }
1147 :
1148 195554 : if (lobound <= 0)
1149 : {
1150 : /*
1151 : * Constant is below lower histogram boundary. More
1152 : * precisely, we have found that no entry in the histogram
1153 : * satisfies the inequality clause (if !isgt) or they all do
1154 : * (if isgt). We estimate that that's true of the entire
1155 : * table, so set histfrac to 0.0 (which we'll flip to 1.0
1156 : * below, if isgt).
1157 : */
1158 86258 : histfrac = 0.0;
1159 : }
1160 109296 : else if (lobound >= sslot.nvalues)
1161 : {
1162 : /*
1163 : * Inverse case: constant is above upper histogram boundary.
1164 : */
1165 32244 : histfrac = 1.0;
1166 : }
1167 : else
1168 : {
1169 : /* We have values[i-1] <= constant <= values[i]. */
1170 77052 : int i = lobound;
1171 77052 : double eq_selec = 0;
1172 : double val,
1173 : high,
1174 : low;
1175 : double binfrac;
1176 :
1177 : /*
1178 : * In the cases where we'll need it below, obtain an estimate
1179 : * of the selectivity of "x = constval". We use a calculation
1180 : * similar to what var_eq_const() does for a non-MCV constant,
1181 : * ie, estimate that all distinct non-MCV values occur equally
1182 : * often. But multiplication by "1.0 - sumcommon - nullfrac"
1183 : * will be done by our caller, so we shouldn't do that here.
1184 : * Therefore we can't try to clamp the estimate by reference
1185 : * to the least common MCV; the result would be too small.
1186 : *
1187 : * Note: since this is effectively assuming that constval
1188 : * isn't an MCV, it's logically dubious if constval in fact is
1189 : * one. But we have to apply *some* correction for equality,
1190 : * and anyway we cannot tell if constval is an MCV, since we
1191 : * don't have a suitable equality operator at hand.
1192 : */
1193 77052 : if (i == 1 || isgt == iseq)
1194 : {
1195 : double otherdistinct;
1196 : bool isdefault;
1197 : AttStatsSlot mcvslot;
1198 :
1199 : /* Get estimated number of distinct values */
1200 25158 : otherdistinct = get_variable_numdistinct(vardata,
1201 : &isdefault);
1202 :
1203 : /* Subtract off the number of known MCVs */
1204 25158 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1205 : STATISTIC_KIND_MCV, InvalidOid,
1206 : ATTSTATSSLOT_NUMBERS))
1207 : {
1208 3496 : otherdistinct -= mcvslot.nnumbers;
1209 3496 : free_attstatsslot(&mcvslot);
1210 : }
1211 :
1212 : /* If result doesn't seem sane, leave eq_selec at 0 */
1213 25158 : if (otherdistinct > 1)
1214 25116 : eq_selec = 1.0 / otherdistinct;
1215 : }
1216 :
1217 : /*
1218 : * Convert the constant and the two nearest bin boundary
1219 : * values to a uniform comparison scale, and do a linear
1220 : * interpolation within this bin.
1221 : */
1222 77052 : if (convert_to_scalar(constval, consttype, collation,
1223 : &val,
1224 77052 : sslot.values[i - 1], sslot.values[i],
1225 : vardata->vartype,
1226 : &low, &high))
1227 : {
1228 77052 : if (high <= low)
1229 : {
1230 : /* cope if bin boundaries appear identical */
1231 0 : binfrac = 0.5;
1232 : }
1233 77052 : else if (val <= low)
1234 16312 : binfrac = 0.0;
1235 60740 : else if (val >= high)
1236 3572 : binfrac = 1.0;
1237 : else
1238 : {
1239 57168 : binfrac = (val - low) / (high - low);
1240 :
1241 : /*
1242 : * Watch out for the possibility that we got a NaN or
1243 : * Infinity from the division. This can happen
1244 : * despite the previous checks, if for example "low"
1245 : * is -Infinity.
1246 : */
1247 57168 : if (isnan(binfrac) ||
1248 57168 : binfrac < 0.0 || binfrac > 1.0)
1249 0 : binfrac = 0.5;
1250 : }
1251 : }
1252 : else
1253 : {
1254 : /*
1255 : * Ideally we'd produce an error here, on the grounds that
1256 : * the given operator shouldn't have scalarXXsel
1257 : * registered as its selectivity func unless we can deal
1258 : * with its operand types. But currently, all manner of
1259 : * stuff is invoking scalarXXsel, so give a default
1260 : * estimate until that can be fixed.
1261 : */
1262 0 : binfrac = 0.5;
1263 : }
1264 :
1265 : /*
1266 : * Now, compute the overall selectivity across the values
1267 : * represented by the histogram. We have i-1 full bins and
1268 : * binfrac partial bin below the constant.
1269 : */
1270 77052 : histfrac = (double) (i - 1) + binfrac;
1271 77052 : histfrac /= (double) (sslot.nvalues - 1);
1272 :
1273 : /*
1274 : * At this point, histfrac is an estimate of the fraction of
1275 : * the population represented by the histogram that satisfies
1276 : * "x <= constval". Somewhat remarkably, this statement is
1277 : * true regardless of which operator we were doing the probes
1278 : * with, so long as convert_to_scalar() delivers reasonable
1279 : * results. If the probe constant is equal to some histogram
1280 : * entry, we would have considered the bin to the left of that
1281 : * entry if probing with "<" or ">=", or the bin to the right
1282 : * if probing with "<=" or ">"; but binfrac would have come
1283 : * out as 1.0 in the first case and 0.0 in the second, leading
1284 : * to the same histfrac in either case. For probe constants
1285 : * between histogram entries, we find the same bin and get the
1286 : * same estimate with any operator.
1287 : *
1288 : * The fact that the estimate corresponds to "x <= constval"
1289 : * and not "x < constval" is because of the way that ANALYZE
1290 : * constructs the histogram: each entry is, effectively, the
1291 : * rightmost value in its sample bucket. So selectivity
1292 : * values that are exact multiples of 1/(histogram_size-1)
1293 : * should be understood as estimates including a histogram
1294 : * entry plus everything to its left.
1295 : *
1296 : * However, that breaks down for the first histogram entry,
1297 : * which necessarily is the leftmost value in its sample
1298 : * bucket. That means the first histogram bin is slightly
1299 : * narrower than the rest, by an amount equal to eq_selec.
1300 : * Another way to say that is that we want "x <= leftmost" to
1301 : * be estimated as eq_selec not zero. So, if we're dealing
1302 : * with the first bin (i==1), rescale to make that true while
1303 : * adjusting the rest of that bin linearly.
1304 : */
1305 77052 : if (i == 1)
1306 11332 : histfrac += eq_selec * (1.0 - binfrac);
1307 :
1308 : /*
1309 : * "x <= constval" is good if we want an estimate for "<=" or
1310 : * ">", but if we are estimating for "<" or ">=", we now need
1311 : * to decrease the estimate by eq_selec.
1312 : */
1313 77052 : if (isgt == iseq)
1314 20424 : histfrac -= eq_selec;
1315 : }
1316 :
1317 : /*
1318 : * Now the estimate is finished for "<" and "<=" cases. If we are
1319 : * estimating for ">" or ">=", flip it.
1320 : */
1321 195554 : hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1322 :
1323 : /*
1324 : * The histogram boundaries are only approximate to begin with,
1325 : * and may well be out of date anyway. Therefore, don't believe
1326 : * extremely small or large selectivity estimates --- unless we
1327 : * got actual current endpoint values from the table, in which
1328 : * case just do the usual sanity clamp. Somewhat arbitrarily, we
1329 : * set the cutoff for other cases at a hundredth of the histogram
1330 : * resolution.
1331 : */
1332 195554 : if (have_end)
1333 102342 : CLAMP_PROBABILITY(hist_selec);
1334 : else
1335 : {
1336 93212 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1337 :
1338 93212 : if (hist_selec < cutoff)
1339 31590 : hist_selec = cutoff;
1340 61622 : else if (hist_selec > 1.0 - cutoff)
1341 24974 : hist_selec = 1.0 - cutoff;
1342 : }
1343 : }
1344 184 : else if (sslot.nvalues > 1)
1345 : {
1346 : /*
1347 : * If we get here, we have a histogram but it's not sorted the way
1348 : * we want. Do a brute-force search to see how many of the
1349 : * entries satisfy the comparison condition, and take that
1350 : * fraction as our estimate. (This is identical to the inner loop
1351 : * of histogram_selectivity; maybe share code?)
1352 : */
1353 184 : LOCAL_FCINFO(fcinfo, 2);
1354 184 : int nmatch = 0;
1355 :
1356 184 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1357 : NULL, NULL);
1358 184 : fcinfo->args[0].isnull = false;
1359 184 : fcinfo->args[1].isnull = false;
1360 184 : fcinfo->args[1].value = constval;
1361 962084 : for (int i = 0; i < sslot.nvalues; i++)
1362 : {
1363 : Datum fresult;
1364 :
1365 961900 : fcinfo->args[0].value = sslot.values[i];
1366 961900 : fcinfo->isnull = false;
1367 961900 : fresult = FunctionCallInvoke(fcinfo);
1368 961900 : if (!fcinfo->isnull && DatumGetBool(fresult))
1369 1924 : nmatch++;
1370 : }
1371 184 : hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1372 :
1373 : /*
1374 : * As above, clamp to a hundredth of the histogram resolution.
1375 : * This case is surely even less trustworthy than the normal one,
1376 : * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1377 : * clamp should be more restrictive in this case?)
1378 : */
1379 : {
1380 184 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1381 :
1382 184 : if (hist_selec < cutoff)
1383 12 : hist_selec = cutoff;
1384 172 : else if (hist_selec > 1.0 - cutoff)
1385 12 : hist_selec = 1.0 - cutoff;
1386 : }
1387 : }
1388 :
1389 195738 : free_attstatsslot(&sslot);
1390 : }
1391 :
1392 250182 : return hist_selec;
1393 : }
1394 :
1395 : /*
1396 : * Common wrapper function for the selectivity estimators that simply
1397 : * invoke scalarineqsel().
1398 : */
1399 : static Datum
1400 38992 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1401 : {
1402 38992 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1403 38992 : Oid operator = PG_GETARG_OID(1);
1404 38992 : List *args = (List *) PG_GETARG_POINTER(2);
1405 38992 : int varRelid = PG_GETARG_INT32(3);
1406 38992 : Oid collation = PG_GET_COLLATION();
1407 : VariableStatData vardata;
1408 : Node *other;
1409 : bool varonleft;
1410 : Datum constval;
1411 : Oid consttype;
1412 : double selec;
1413 :
1414 : /*
1415 : * If expression is not variable op something or something op variable,
1416 : * then punt and return a default estimate.
1417 : */
1418 38992 : if (!get_restriction_variable(root, args, varRelid,
1419 : &vardata, &other, &varonleft))
1420 890 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1421 :
1422 : /*
1423 : * Can't do anything useful if the something is not a constant, either.
1424 : */
1425 38102 : if (!IsA(other, Const))
1426 : {
1427 2636 : ReleaseVariableStats(vardata);
1428 2636 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1429 : }
1430 :
1431 : /*
1432 : * If the constant is NULL, assume operator is strict and return zero, ie,
1433 : * operator will never return TRUE.
1434 : */
1435 35466 : if (((Const *) other)->constisnull)
1436 : {
1437 54 : ReleaseVariableStats(vardata);
1438 54 : PG_RETURN_FLOAT8(0.0);
1439 : }
1440 35412 : constval = ((Const *) other)->constvalue;
1441 35412 : consttype = ((Const *) other)->consttype;
1442 :
1443 : /*
1444 : * Force the var to be on the left to simplify logic in scalarineqsel.
1445 : */
1446 35412 : if (!varonleft)
1447 : {
1448 336 : operator = get_commutator(operator);
1449 336 : if (!operator)
1450 : {
1451 : /* Use default selectivity (should we raise an error instead?) */
1452 0 : ReleaseVariableStats(vardata);
1453 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1454 : }
1455 336 : isgt = !isgt;
1456 : }
1457 :
1458 : /* The rest of the work is done by scalarineqsel(). */
1459 35412 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1460 : &vardata, constval, consttype);
1461 :
1462 35412 : ReleaseVariableStats(vardata);
1463 :
1464 35412 : PG_RETURN_FLOAT8((float8) selec);
1465 : }
1466 :
1467 : /*
1468 : * scalarltsel - Selectivity of "<" for scalars.
1469 : */
1470 : Datum
1471 13536 : scalarltsel(PG_FUNCTION_ARGS)
1472 : {
1473 13536 : return scalarineqsel_wrapper(fcinfo, false, false);
1474 : }
1475 :
1476 : /*
1477 : * scalarlesel - Selectivity of "<=" for scalars.
1478 : */
1479 : Datum
1480 4290 : scalarlesel(PG_FUNCTION_ARGS)
1481 : {
1482 4290 : return scalarineqsel_wrapper(fcinfo, false, true);
1483 : }
1484 :
1485 : /*
1486 : * scalargtsel - Selectivity of ">" for scalars.
1487 : */
1488 : Datum
1489 13860 : scalargtsel(PG_FUNCTION_ARGS)
1490 : {
1491 13860 : return scalarineqsel_wrapper(fcinfo, true, false);
1492 : }
1493 :
1494 : /*
1495 : * scalargesel - Selectivity of ">=" for scalars.
1496 : */
1497 : Datum
1498 7306 : scalargesel(PG_FUNCTION_ARGS)
1499 : {
1500 7306 : return scalarineqsel_wrapper(fcinfo, true, true);
1501 : }
1502 :
1503 : /*
1504 : * boolvarsel - Selectivity of Boolean variable.
1505 : *
1506 : * This can actually be called on any boolean-valued expression. If it
1507 : * involves only Vars of the specified relation, and if there are statistics
1508 : * about the Var or expression (the latter is possible if it's indexed) then
1509 : * we'll produce a real estimate; otherwise it's just a default.
1510 : */
1511 : Selectivity
1512 43712 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1513 : {
1514 : VariableStatData vardata;
1515 : double selec;
1516 :
1517 43712 : examine_variable(root, arg, varRelid, &vardata);
1518 43712 : if (HeapTupleIsValid(vardata.statsTuple))
1519 : {
1520 : /*
1521 : * A boolean variable V is equivalent to the clause V = 't', so we
1522 : * compute the selectivity as if that is what we have.
1523 : */
1524 36662 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1525 : BoolGetDatum(true), false, true, false);
1526 : }
1527 : else
1528 : {
1529 : /* Otherwise, the default estimate is 0.5 */
1530 7050 : selec = 0.5;
1531 : }
1532 43712 : ReleaseVariableStats(vardata);
1533 43712 : return selec;
1534 : }
1535 :
1536 : /*
1537 : * booltestsel - Selectivity of BooleanTest Node.
1538 : */
1539 : Selectivity
1540 850 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1541 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1542 : {
1543 : VariableStatData vardata;
1544 : double selec;
1545 :
1546 850 : examine_variable(root, arg, varRelid, &vardata);
1547 :
1548 850 : if (HeapTupleIsValid(vardata.statsTuple))
1549 : {
1550 : Form_pg_statistic stats;
1551 : double freq_null;
1552 : AttStatsSlot sslot;
1553 :
1554 0 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1555 0 : freq_null = stats->stanullfrac;
1556 :
1557 0 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1558 : STATISTIC_KIND_MCV, InvalidOid,
1559 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1560 0 : && sslot.nnumbers > 0)
1561 0 : {
1562 : double freq_true;
1563 : double freq_false;
1564 :
1565 : /*
1566 : * Get first MCV frequency and derive frequency for true.
1567 : */
1568 0 : if (DatumGetBool(sslot.values[0]))
1569 0 : freq_true = sslot.numbers[0];
1570 : else
1571 0 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1572 :
1573 : /*
1574 : * Next derive frequency for false. Then use these as appropriate
1575 : * to derive frequency for each case.
1576 : */
1577 0 : freq_false = 1.0 - freq_true - freq_null;
1578 :
1579 0 : switch (booltesttype)
1580 : {
1581 0 : case IS_UNKNOWN:
1582 : /* select only NULL values */
1583 0 : selec = freq_null;
1584 0 : break;
1585 0 : case IS_NOT_UNKNOWN:
1586 : /* select non-NULL values */
1587 0 : selec = 1.0 - freq_null;
1588 0 : break;
1589 0 : case IS_TRUE:
1590 : /* select only TRUE values */
1591 0 : selec = freq_true;
1592 0 : break;
1593 0 : case IS_NOT_TRUE:
1594 : /* select non-TRUE values */
1595 0 : selec = 1.0 - freq_true;
1596 0 : break;
1597 0 : case IS_FALSE:
1598 : /* select only FALSE values */
1599 0 : selec = freq_false;
1600 0 : break;
1601 0 : case IS_NOT_FALSE:
1602 : /* select non-FALSE values */
1603 0 : selec = 1.0 - freq_false;
1604 0 : break;
1605 0 : default:
1606 0 : elog(ERROR, "unrecognized booltesttype: %d",
1607 : (int) booltesttype);
1608 : selec = 0.0; /* Keep compiler quiet */
1609 : break;
1610 : }
1611 :
1612 0 : free_attstatsslot(&sslot);
1613 : }
1614 : else
1615 : {
1616 : /*
1617 : * No most-common-value info available. Still have null fraction
1618 : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1619 : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1620 : */
1621 0 : switch (booltesttype)
1622 : {
1623 0 : case IS_UNKNOWN:
1624 : /* select only NULL values */
1625 0 : selec = freq_null;
1626 0 : break;
1627 0 : case IS_NOT_UNKNOWN:
1628 : /* select non-NULL values */
1629 0 : selec = 1.0 - freq_null;
1630 0 : break;
1631 0 : case IS_TRUE:
1632 : case IS_FALSE:
1633 : /* Assume we select half of the non-NULL values */
1634 0 : selec = (1.0 - freq_null) / 2.0;
1635 0 : break;
1636 0 : case IS_NOT_TRUE:
1637 : case IS_NOT_FALSE:
1638 : /* Assume we select NULLs plus half of the non-NULLs */
1639 : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1640 0 : selec = (freq_null + 1.0) / 2.0;
1641 0 : break;
1642 0 : default:
1643 0 : elog(ERROR, "unrecognized booltesttype: %d",
1644 : (int) booltesttype);
1645 : selec = 0.0; /* Keep compiler quiet */
1646 : break;
1647 : }
1648 : }
1649 : }
1650 : else
1651 : {
1652 : /*
1653 : * If we can't get variable statistics for the argument, perhaps
1654 : * clause_selectivity can do something with it. We ignore the
1655 : * possibility of a NULL value when using clause_selectivity, and just
1656 : * assume the value is either TRUE or FALSE.
1657 : */
1658 850 : switch (booltesttype)
1659 : {
1660 48 : case IS_UNKNOWN:
1661 48 : selec = DEFAULT_UNK_SEL;
1662 48 : break;
1663 108 : case IS_NOT_UNKNOWN:
1664 108 : selec = DEFAULT_NOT_UNK_SEL;
1665 108 : break;
1666 252 : case IS_TRUE:
1667 : case IS_NOT_FALSE:
1668 252 : selec = (double) clause_selectivity(root, arg,
1669 : varRelid,
1670 : jointype, sjinfo);
1671 252 : break;
1672 442 : case IS_FALSE:
1673 : case IS_NOT_TRUE:
1674 442 : selec = 1.0 - (double) clause_selectivity(root, arg,
1675 : varRelid,
1676 : jointype, sjinfo);
1677 442 : break;
1678 0 : default:
1679 0 : elog(ERROR, "unrecognized booltesttype: %d",
1680 : (int) booltesttype);
1681 : selec = 0.0; /* Keep compiler quiet */
1682 : break;
1683 : }
1684 : }
1685 :
1686 850 : ReleaseVariableStats(vardata);
1687 :
1688 : /* result should be in range, but make sure... */
1689 850 : CLAMP_PROBABILITY(selec);
1690 :
1691 850 : return (Selectivity) selec;
1692 : }
1693 :
1694 : /*
1695 : * nulltestsel - Selectivity of NullTest Node.
1696 : */
1697 : Selectivity
1698 16366 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1699 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1700 : {
1701 : VariableStatData vardata;
1702 : double selec;
1703 :
1704 16366 : examine_variable(root, arg, varRelid, &vardata);
1705 :
1706 16366 : if (HeapTupleIsValid(vardata.statsTuple))
1707 : {
1708 : Form_pg_statistic stats;
1709 : double freq_null;
1710 :
1711 9510 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1712 9510 : freq_null = stats->stanullfrac;
1713 :
1714 9510 : switch (nulltesttype)
1715 : {
1716 6682 : case IS_NULL:
1717 :
1718 : /*
1719 : * Use freq_null directly.
1720 : */
1721 6682 : selec = freq_null;
1722 6682 : break;
1723 2828 : case IS_NOT_NULL:
1724 :
1725 : /*
1726 : * Select not unknown (not null) values. Calculate from
1727 : * freq_null.
1728 : */
1729 2828 : selec = 1.0 - freq_null;
1730 2828 : break;
1731 0 : default:
1732 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1733 : (int) nulltesttype);
1734 : return (Selectivity) 0; /* keep compiler quiet */
1735 : }
1736 : }
1737 6856 : else if (vardata.var && IsA(vardata.var, Var) &&
1738 6190 : ((Var *) vardata.var)->varattno < 0)
1739 : {
1740 : /*
1741 : * There are no stats for system columns, but we know they are never
1742 : * NULL.
1743 : */
1744 60 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1745 : }
1746 : else
1747 : {
1748 : /*
1749 : * No ANALYZE stats available, so make a guess
1750 : */
1751 6796 : switch (nulltesttype)
1752 : {
1753 1992 : case IS_NULL:
1754 1992 : selec = DEFAULT_UNK_SEL;
1755 1992 : break;
1756 4804 : case IS_NOT_NULL:
1757 4804 : selec = DEFAULT_NOT_UNK_SEL;
1758 4804 : break;
1759 0 : default:
1760 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1761 : (int) nulltesttype);
1762 : return (Selectivity) 0; /* keep compiler quiet */
1763 : }
1764 : }
1765 :
1766 16366 : ReleaseVariableStats(vardata);
1767 :
1768 : /* result should be in range, but make sure... */
1769 16366 : CLAMP_PROBABILITY(selec);
1770 :
1771 16366 : return (Selectivity) selec;
1772 : }
1773 :
1774 : /*
1775 : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1776 : *
1777 : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1778 : * but it seems possible that RelabelType might show up. Also, the planner
1779 : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1780 : * so we need to be ready to deal with more than one level.
1781 : */
1782 : static Node *
1783 108626 : strip_array_coercion(Node *node)
1784 : {
1785 : for (;;)
1786 : {
1787 108626 : if (node && IsA(node, ArrayCoerceExpr))
1788 36 : {
1789 2702 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1790 :
1791 : /*
1792 : * If the per-element expression is just a RelabelType on top of
1793 : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1794 : */
1795 2702 : if (IsA(acoerce->elemexpr, RelabelType) &&
1796 36 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1797 36 : node = (Node *) acoerce->arg;
1798 : else
1799 : break;
1800 : }
1801 105924 : else if (node && IsA(node, RelabelType))
1802 : {
1803 : /* We don't really expect this case, but may as well cope */
1804 0 : node = (Node *) ((RelabelType *) node)->arg;
1805 : }
1806 : else
1807 : break;
1808 : }
1809 108590 : return node;
1810 : }
1811 :
1812 : /*
1813 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1814 : */
1815 : Selectivity
1816 16032 : scalararraysel(PlannerInfo *root,
1817 : ScalarArrayOpExpr *clause,
1818 : bool is_join_clause,
1819 : int varRelid,
1820 : JoinType jointype,
1821 : SpecialJoinInfo *sjinfo)
1822 : {
1823 16032 : Oid operator = clause->opno;
1824 16032 : bool useOr = clause->useOr;
1825 16032 : bool isEquality = false;
1826 16032 : bool isInequality = false;
1827 : Node *leftop;
1828 : Node *rightop;
1829 : Oid nominal_element_type;
1830 : Oid nominal_element_collation;
1831 : TypeCacheEntry *typentry;
1832 : RegProcedure oprsel;
1833 : FmgrInfo oprselproc;
1834 : Selectivity s1;
1835 : Selectivity s1disjoint;
1836 :
1837 : /* First, deconstruct the expression */
1838 : Assert(list_length(clause->args) == 2);
1839 16032 : leftop = (Node *) linitial(clause->args);
1840 16032 : rightop = (Node *) lsecond(clause->args);
1841 :
1842 : /* aggressively reduce both sides to constants */
1843 16032 : leftop = estimate_expression_value(root, leftop);
1844 16032 : rightop = estimate_expression_value(root, rightop);
1845 :
1846 : /* get nominal (after relabeling) element type of rightop */
1847 16032 : nominal_element_type = get_base_element_type(exprType(rightop));
1848 16032 : if (!OidIsValid(nominal_element_type))
1849 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1850 : /* get nominal collation, too, for generating constants */
1851 16032 : nominal_element_collation = exprCollation(rightop);
1852 :
1853 : /* look through any binary-compatible relabeling of rightop */
1854 16032 : rightop = strip_array_coercion(rightop);
1855 :
1856 : /*
1857 : * Detect whether the operator is the default equality or inequality
1858 : * operator of the array element type.
1859 : */
1860 16032 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1861 16032 : if (OidIsValid(typentry->eq_opr))
1862 : {
1863 16032 : if (operator == typentry->eq_opr)
1864 13620 : isEquality = true;
1865 2412 : else if (get_negator(operator) == typentry->eq_opr)
1866 1932 : isInequality = true;
1867 : }
1868 :
1869 : /*
1870 : * If it is equality or inequality, we might be able to estimate this as a
1871 : * form of array containment; for instance "const = ANY(column)" can be
1872 : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1873 : * that, and returns the selectivity estimate if successful, or -1 if not.
1874 : */
1875 16032 : if ((isEquality || isInequality) && !is_join_clause)
1876 : {
1877 15552 : s1 = scalararraysel_containment(root, leftop, rightop,
1878 : nominal_element_type,
1879 : isEquality, useOr, varRelid);
1880 15552 : if (s1 >= 0.0)
1881 118 : return s1;
1882 : }
1883 :
1884 : /*
1885 : * Look up the underlying operator's selectivity estimator. Punt if it
1886 : * hasn't got one.
1887 : */
1888 15914 : if (is_join_clause)
1889 0 : oprsel = get_oprjoin(operator);
1890 : else
1891 15914 : oprsel = get_oprrest(operator);
1892 15914 : if (!oprsel)
1893 0 : return (Selectivity) 0.5;
1894 15914 : fmgr_info(oprsel, &oprselproc);
1895 :
1896 : /*
1897 : * In the array-containment check above, we must only believe that an
1898 : * operator is equality or inequality if it is the default btree equality
1899 : * operator (or its negator) for the element type, since those are the
1900 : * operators that array containment will use. But in what follows, we can
1901 : * be a little laxer, and also believe that any operators using eqsel() or
1902 : * neqsel() as selectivity estimator act like equality or inequality.
1903 : */
1904 15914 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1905 13666 : isEquality = true;
1906 2248 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1907 1822 : isInequality = true;
1908 :
1909 : /*
1910 : * We consider three cases:
1911 : *
1912 : * 1. rightop is an Array constant: deconstruct the array, apply the
1913 : * operator's selectivity function for each array element, and merge the
1914 : * results in the same way that clausesel.c does for AND/OR combinations.
1915 : *
1916 : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
1917 : * function for each element of the ARRAY[] construct, and merge.
1918 : *
1919 : * 3. otherwise, make a guess ...
1920 : */
1921 15914 : if (rightop && IsA(rightop, Const))
1922 12452 : {
1923 12470 : Datum arraydatum = ((Const *) rightop)->constvalue;
1924 12470 : bool arrayisnull = ((Const *) rightop)->constisnull;
1925 : ArrayType *arrayval;
1926 : int16 elmlen;
1927 : bool elmbyval;
1928 : char elmalign;
1929 : int num_elems;
1930 : Datum *elem_values;
1931 : bool *elem_nulls;
1932 : int i;
1933 :
1934 12470 : if (arrayisnull) /* qual can't succeed if null array */
1935 18 : return (Selectivity) 0.0;
1936 12452 : arrayval = DatumGetArrayTypeP(arraydatum);
1937 12452 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
1938 : &elmlen, &elmbyval, &elmalign);
1939 12452 : deconstruct_array(arrayval,
1940 : ARR_ELEMTYPE(arrayval),
1941 : elmlen, elmbyval, elmalign,
1942 : &elem_values, &elem_nulls, &num_elems);
1943 :
1944 : /*
1945 : * For generic operators, we assume the probability of success is
1946 : * independent for each array element. But for "= ANY" or "<> ALL",
1947 : * if the array elements are distinct (which'd typically be the case)
1948 : * then the probabilities are disjoint, and we should just sum them.
1949 : *
1950 : * If we were being really tense we would try to confirm that the
1951 : * elements are all distinct, but that would be expensive and it
1952 : * doesn't seem to be worth the cycles; it would amount to penalizing
1953 : * well-written queries in favor of poorly-written ones. However, we
1954 : * do protect ourselves a little bit by checking whether the
1955 : * disjointness assumption leads to an impossible (out of range)
1956 : * probability; if so, we fall back to the normal calculation.
1957 : */
1958 12452 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
1959 :
1960 50820 : for (i = 0; i < num_elems; i++)
1961 : {
1962 : List *args;
1963 : Selectivity s2;
1964 :
1965 38368 : args = list_make2(leftop,
1966 : makeConst(nominal_element_type,
1967 : -1,
1968 : nominal_element_collation,
1969 : elmlen,
1970 : elem_values[i],
1971 : elem_nulls[i],
1972 : elmbyval));
1973 38368 : if (is_join_clause)
1974 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1975 : clause->inputcollid,
1976 : PointerGetDatum(root),
1977 : ObjectIdGetDatum(operator),
1978 : PointerGetDatum(args),
1979 : Int16GetDatum(jointype),
1980 : PointerGetDatum(sjinfo)));
1981 : else
1982 38368 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1983 : clause->inputcollid,
1984 : PointerGetDatum(root),
1985 : ObjectIdGetDatum(operator),
1986 : PointerGetDatum(args),
1987 : Int32GetDatum(varRelid)));
1988 :
1989 38368 : if (useOr)
1990 : {
1991 32476 : s1 = s1 + s2 - s1 * s2;
1992 32476 : if (isEquality)
1993 31612 : s1disjoint += s2;
1994 : }
1995 : else
1996 : {
1997 5892 : s1 = s1 * s2;
1998 5892 : if (isInequality)
1999 5580 : s1disjoint += s2 - 1.0;
2000 : }
2001 : }
2002 :
2003 : /* accept disjoint-probability estimate if in range */
2004 12452 : if ((useOr ? isEquality : isInequality) &&
2005 11876 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2006 11846 : s1 = s1disjoint;
2007 : }
2008 3444 : else if (rightop && IsA(rightop, ArrayExpr) &&
2009 122 : !((ArrayExpr *) rightop)->multidims)
2010 122 : {
2011 122 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2012 : int16 elmlen;
2013 : bool elmbyval;
2014 : ListCell *l;
2015 :
2016 122 : get_typlenbyval(arrayexpr->element_typeid,
2017 : &elmlen, &elmbyval);
2018 :
2019 : /*
2020 : * We use the assumption of disjoint probabilities here too, although
2021 : * the odds of equal array elements are rather higher if the elements
2022 : * are not all constants (which they won't be, else constant folding
2023 : * would have reduced the ArrayExpr to a Const). In this path it's
2024 : * critical to have the sanity check on the s1disjoint estimate.
2025 : */
2026 122 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2027 :
2028 440 : foreach(l, arrayexpr->elements)
2029 : {
2030 318 : Node *elem = (Node *) lfirst(l);
2031 : List *args;
2032 : Selectivity s2;
2033 :
2034 : /*
2035 : * Theoretically, if elem isn't of nominal_element_type we should
2036 : * insert a RelabelType, but it seems unlikely that any operator
2037 : * estimation function would really care ...
2038 : */
2039 318 : args = list_make2(leftop, elem);
2040 318 : if (is_join_clause)
2041 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2042 : clause->inputcollid,
2043 : PointerGetDatum(root),
2044 : ObjectIdGetDatum(operator),
2045 : PointerGetDatum(args),
2046 : Int16GetDatum(jointype),
2047 : PointerGetDatum(sjinfo)));
2048 : else
2049 318 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2050 : clause->inputcollid,
2051 : PointerGetDatum(root),
2052 : ObjectIdGetDatum(operator),
2053 : PointerGetDatum(args),
2054 : Int32GetDatum(varRelid)));
2055 :
2056 318 : if (useOr)
2057 : {
2058 318 : s1 = s1 + s2 - s1 * s2;
2059 318 : if (isEquality)
2060 318 : s1disjoint += s2;
2061 : }
2062 : else
2063 : {
2064 0 : s1 = s1 * s2;
2065 0 : if (isInequality)
2066 0 : s1disjoint += s2 - 1.0;
2067 : }
2068 : }
2069 :
2070 : /* accept disjoint-probability estimate if in range */
2071 122 : if ((useOr ? isEquality : isInequality) &&
2072 122 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2073 122 : s1 = s1disjoint;
2074 : }
2075 : else
2076 : {
2077 : CaseTestExpr *dummyexpr;
2078 : List *args;
2079 : Selectivity s2;
2080 : int i;
2081 :
2082 : /*
2083 : * We need a dummy rightop to pass to the operator selectivity
2084 : * routine. It can be pretty much anything that doesn't look like a
2085 : * constant; CaseTestExpr is a convenient choice.
2086 : */
2087 3322 : dummyexpr = makeNode(CaseTestExpr);
2088 3322 : dummyexpr->typeId = nominal_element_type;
2089 3322 : dummyexpr->typeMod = -1;
2090 3322 : dummyexpr->collation = clause->inputcollid;
2091 3322 : args = list_make2(leftop, dummyexpr);
2092 3322 : if (is_join_clause)
2093 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2094 : clause->inputcollid,
2095 : PointerGetDatum(root),
2096 : ObjectIdGetDatum(operator),
2097 : PointerGetDatum(args),
2098 : Int16GetDatum(jointype),
2099 : PointerGetDatum(sjinfo)));
2100 : else
2101 3322 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2102 : clause->inputcollid,
2103 : PointerGetDatum(root),
2104 : ObjectIdGetDatum(operator),
2105 : PointerGetDatum(args),
2106 : Int32GetDatum(varRelid)));
2107 3322 : s1 = useOr ? 0.0 : 1.0;
2108 :
2109 : /*
2110 : * Arbitrarily assume 10 elements in the eventual array value (see
2111 : * also estimate_array_length). We don't risk an assumption of
2112 : * disjoint probabilities here.
2113 : */
2114 36542 : for (i = 0; i < 10; i++)
2115 : {
2116 33220 : if (useOr)
2117 33220 : s1 = s1 + s2 - s1 * s2;
2118 : else
2119 0 : s1 = s1 * s2;
2120 : }
2121 : }
2122 :
2123 : /* result should be in range, but make sure... */
2124 15896 : CLAMP_PROBABILITY(s1);
2125 :
2126 15896 : return s1;
2127 : }
2128 :
2129 : /*
2130 : * Estimate number of elements in the array yielded by an expression.
2131 : *
2132 : * Note: the result is integral, but we use "double" to avoid overflow
2133 : * concerns. Most callers will use it in double-type expressions anyway.
2134 : */
2135 : double
2136 92558 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2137 : {
2138 : /* look through any binary-compatible relabeling of arrayexpr */
2139 92558 : arrayexpr = strip_array_coercion(arrayexpr);
2140 :
2141 92558 : if (arrayexpr && IsA(arrayexpr, Const))
2142 : {
2143 39734 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2144 39734 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2145 : ArrayType *arrayval;
2146 :
2147 39734 : if (arrayisnull)
2148 36 : return 0;
2149 39698 : arrayval = DatumGetArrayTypeP(arraydatum);
2150 39698 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2151 : }
2152 52824 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2153 490 : !((ArrayExpr *) arrayexpr)->multidims)
2154 : {
2155 490 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2156 : }
2157 52334 : else if (arrayexpr)
2158 : {
2159 : /* See if we can find any statistics about it */
2160 : VariableStatData vardata;
2161 : AttStatsSlot sslot;
2162 52334 : double nelem = 0;
2163 :
2164 52334 : examine_variable(root, arrayexpr, 0, &vardata);
2165 52334 : if (HeapTupleIsValid(vardata.statsTuple))
2166 : {
2167 : /*
2168 : * Found stats, so use the average element count, which is stored
2169 : * in the last stanumbers element of the DECHIST statistics.
2170 : * Actually that is the average count of *distinct* elements;
2171 : * perhaps we should scale it up somewhat?
2172 : */
2173 2988 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2174 : STATISTIC_KIND_DECHIST, InvalidOid,
2175 : ATTSTATSSLOT_NUMBERS))
2176 : {
2177 2874 : if (sslot.nnumbers > 0)
2178 2874 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2179 2874 : free_attstatsslot(&sslot);
2180 : }
2181 : }
2182 52334 : ReleaseVariableStats(vardata);
2183 :
2184 52334 : if (nelem > 0)
2185 2874 : return nelem;
2186 : }
2187 :
2188 : /* Else use a default guess --- this should match scalararraysel */
2189 49460 : return 10;
2190 : }
2191 :
2192 : /*
2193 : * rowcomparesel - Selectivity of RowCompareExpr Node.
2194 : *
2195 : * We estimate RowCompare selectivity by considering just the first (high
2196 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2197 : * this estimate could be refined by considering additional columns, it
2198 : * seems unlikely that we could do a lot better without multi-column
2199 : * statistics.
2200 : */
2201 : Selectivity
2202 156 : rowcomparesel(PlannerInfo *root,
2203 : RowCompareExpr *clause,
2204 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2205 : {
2206 : Selectivity s1;
2207 156 : Oid opno = linitial_oid(clause->opnos);
2208 156 : Oid inputcollid = linitial_oid(clause->inputcollids);
2209 : List *opargs;
2210 : bool is_join_clause;
2211 :
2212 : /* Build equivalent arg list for single operator */
2213 156 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2214 :
2215 : /*
2216 : * Decide if it's a join clause. This should match clausesel.c's
2217 : * treat_as_join_clause(), except that we intentionally consider only the
2218 : * leading columns and not the rest of the clause.
2219 : */
2220 156 : if (varRelid != 0)
2221 : {
2222 : /*
2223 : * Caller is forcing restriction mode (eg, because we are examining an
2224 : * inner indexscan qual).
2225 : */
2226 54 : is_join_clause = false;
2227 : }
2228 102 : else if (sjinfo == NULL)
2229 : {
2230 : /*
2231 : * It must be a restriction clause, since it's being evaluated at a
2232 : * scan node.
2233 : */
2234 90 : is_join_clause = false;
2235 : }
2236 : else
2237 : {
2238 : /*
2239 : * Otherwise, it's a join if there's more than one base relation used.
2240 : */
2241 12 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2242 : }
2243 :
2244 156 : if (is_join_clause)
2245 : {
2246 : /* Estimate selectivity for a join clause. */
2247 12 : s1 = join_selectivity(root, opno,
2248 : opargs,
2249 : inputcollid,
2250 : jointype,
2251 : sjinfo);
2252 : }
2253 : else
2254 : {
2255 : /* Estimate selectivity for a restriction clause. */
2256 144 : s1 = restriction_selectivity(root, opno,
2257 : opargs,
2258 : inputcollid,
2259 : varRelid);
2260 : }
2261 :
2262 156 : return s1;
2263 : }
2264 :
2265 : /*
2266 : * eqjoinsel - Join selectivity of "="
2267 : */
2268 : Datum
2269 189276 : eqjoinsel(PG_FUNCTION_ARGS)
2270 : {
2271 189276 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2272 189276 : Oid operator = PG_GETARG_OID(1);
2273 189276 : List *args = (List *) PG_GETARG_POINTER(2);
2274 :
2275 : #ifdef NOT_USED
2276 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2277 : #endif
2278 189276 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2279 189276 : Oid collation = PG_GET_COLLATION();
2280 : double selec;
2281 : double selec_inner;
2282 : VariableStatData vardata1;
2283 : VariableStatData vardata2;
2284 : double nd1;
2285 : double nd2;
2286 : bool isdefault1;
2287 : bool isdefault2;
2288 : Oid opfuncoid;
2289 : AttStatsSlot sslot1;
2290 : AttStatsSlot sslot2;
2291 189276 : Form_pg_statistic stats1 = NULL;
2292 189276 : Form_pg_statistic stats2 = NULL;
2293 189276 : bool have_mcvs1 = false;
2294 189276 : bool have_mcvs2 = false;
2295 : bool get_mcv_stats;
2296 : bool join_is_reversed;
2297 : RelOptInfo *inner_rel;
2298 :
2299 189276 : get_join_variables(root, args, sjinfo,
2300 : &vardata1, &vardata2, &join_is_reversed);
2301 :
2302 189276 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2303 189276 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2304 :
2305 189276 : opfuncoid = get_opcode(operator);
2306 :
2307 189276 : memset(&sslot1, 0, sizeof(sslot1));
2308 189276 : memset(&sslot2, 0, sizeof(sslot2));
2309 :
2310 : /*
2311 : * There is no use in fetching one side's MCVs if we lack MCVs for the
2312 : * other side, so do a quick check to verify that both stats exist.
2313 : */
2314 531364 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2315 271040 : HeapTupleIsValid(vardata2.statsTuple) &&
2316 118228 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2317 : STATISTIC_KIND_MCV, InvalidOid,
2318 342088 : 0) &&
2319 54862 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2320 : STATISTIC_KIND_MCV, InvalidOid,
2321 : 0));
2322 :
2323 189276 : if (HeapTupleIsValid(vardata1.statsTuple))
2324 : {
2325 : /* note we allow use of nullfrac regardless of security check */
2326 152812 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2327 169450 : if (get_mcv_stats &&
2328 16638 : statistic_proc_security_check(&vardata1, opfuncoid))
2329 16638 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2330 : STATISTIC_KIND_MCV, InvalidOid,
2331 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2332 : }
2333 :
2334 189276 : if (HeapTupleIsValid(vardata2.statsTuple))
2335 : {
2336 : /* note we allow use of nullfrac regardless of security check */
2337 127004 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2338 143642 : if (get_mcv_stats &&
2339 16638 : statistic_proc_security_check(&vardata2, opfuncoid))
2340 16638 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2341 : STATISTIC_KIND_MCV, InvalidOid,
2342 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2343 : }
2344 :
2345 : /* We need to compute the inner-join selectivity in all cases */
2346 189276 : selec_inner = eqjoinsel_inner(opfuncoid, collation,
2347 : &vardata1, &vardata2,
2348 : nd1, nd2,
2349 : isdefault1, isdefault2,
2350 : &sslot1, &sslot2,
2351 : stats1, stats2,
2352 : have_mcvs1, have_mcvs2);
2353 :
2354 189276 : switch (sjinfo->jointype)
2355 : {
2356 182236 : case JOIN_INNER:
2357 : case JOIN_LEFT:
2358 : case JOIN_FULL:
2359 182236 : selec = selec_inner;
2360 182236 : break;
2361 7040 : case JOIN_SEMI:
2362 : case JOIN_ANTI:
2363 :
2364 : /*
2365 : * Look up the join's inner relation. min_righthand is sufficient
2366 : * information because neither SEMI nor ANTI joins permit any
2367 : * reassociation into or out of their RHS, so the righthand will
2368 : * always be exactly that set of rels.
2369 : */
2370 7040 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2371 :
2372 7040 : if (!join_is_reversed)
2373 3428 : selec = eqjoinsel_semi(opfuncoid, collation,
2374 : &vardata1, &vardata2,
2375 : nd1, nd2,
2376 : isdefault1, isdefault2,
2377 : &sslot1, &sslot2,
2378 : stats1, stats2,
2379 : have_mcvs1, have_mcvs2,
2380 : inner_rel);
2381 : else
2382 : {
2383 3612 : Oid commop = get_commutator(operator);
2384 3612 : Oid commopfuncoid = OidIsValid(commop) ? get_opcode(commop) : InvalidOid;
2385 :
2386 3612 : selec = eqjoinsel_semi(commopfuncoid, collation,
2387 : &vardata2, &vardata1,
2388 : nd2, nd1,
2389 : isdefault2, isdefault1,
2390 : &sslot2, &sslot1,
2391 : stats2, stats1,
2392 : have_mcvs2, have_mcvs1,
2393 : inner_rel);
2394 : }
2395 :
2396 : /*
2397 : * We should never estimate the output of a semijoin to be more
2398 : * rows than we estimate for an inner join with the same input
2399 : * rels and join condition; it's obviously impossible for that to
2400 : * happen. The former estimate is N1 * Ssemi while the latter is
2401 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2402 : * this is worthwhile because of the shakier estimation rules we
2403 : * use in eqjoinsel_semi, particularly in cases where it has to
2404 : * punt entirely.
2405 : */
2406 7040 : selec = Min(selec, inner_rel->rows * selec_inner);
2407 7040 : break;
2408 0 : default:
2409 : /* other values not expected here */
2410 0 : elog(ERROR, "unrecognized join type: %d",
2411 : (int) sjinfo->jointype);
2412 : selec = 0; /* keep compiler quiet */
2413 : break;
2414 : }
2415 :
2416 189276 : free_attstatsslot(&sslot1);
2417 189276 : free_attstatsslot(&sslot2);
2418 :
2419 189276 : ReleaseVariableStats(vardata1);
2420 189276 : ReleaseVariableStats(vardata2);
2421 :
2422 189276 : CLAMP_PROBABILITY(selec);
2423 :
2424 189276 : PG_RETURN_FLOAT8((float8) selec);
2425 : }
2426 :
2427 : /*
2428 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2429 : *
2430 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2431 : * that it's worth trying to distinguish them here.
2432 : */
2433 : static double
2434 189276 : eqjoinsel_inner(Oid opfuncoid, Oid collation,
2435 : VariableStatData *vardata1, VariableStatData *vardata2,
2436 : double nd1, double nd2,
2437 : bool isdefault1, bool isdefault2,
2438 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2439 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2440 : bool have_mcvs1, bool have_mcvs2)
2441 : {
2442 : double selec;
2443 :
2444 189276 : if (have_mcvs1 && have_mcvs2)
2445 16638 : {
2446 : /*
2447 : * We have most-common-value lists for both relations. Run through
2448 : * the lists to see which MCVs actually join to each other with the
2449 : * given operator. This allows us to determine the exact join
2450 : * selectivity for the portion of the relations represented by the MCV
2451 : * lists. We still have to estimate for the remaining population, but
2452 : * in a skewed distribution this gives us a big leg up in accuracy.
2453 : * For motivation see the analysis in Y. Ioannidis and S.
2454 : * Christodoulakis, "On the propagation of errors in the size of join
2455 : * results", Technical Report 1018, Computer Science Dept., University
2456 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2457 : */
2458 16638 : LOCAL_FCINFO(fcinfo, 2);
2459 : FmgrInfo eqproc;
2460 : bool *hasmatch1;
2461 : bool *hasmatch2;
2462 16638 : double nullfrac1 = stats1->stanullfrac;
2463 16638 : double nullfrac2 = stats2->stanullfrac;
2464 : double matchprodfreq,
2465 : matchfreq1,
2466 : matchfreq2,
2467 : unmatchfreq1,
2468 : unmatchfreq2,
2469 : otherfreq1,
2470 : otherfreq2,
2471 : totalsel1,
2472 : totalsel2;
2473 : int i,
2474 : nmatches;
2475 :
2476 16638 : fmgr_info(opfuncoid, &eqproc);
2477 :
2478 : /*
2479 : * Save a few cycles by setting up the fcinfo struct just once. Using
2480 : * FunctionCallInvoke directly also avoids failure if the eqproc
2481 : * returns NULL, though really equality functions should never do
2482 : * that.
2483 : */
2484 16638 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2485 : NULL, NULL);
2486 16638 : fcinfo->args[0].isnull = false;
2487 16638 : fcinfo->args[1].isnull = false;
2488 :
2489 16638 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2490 16638 : hasmatch2 = (bool *) palloc0(sslot2->nvalues * sizeof(bool));
2491 :
2492 : /*
2493 : * Note we assume that each MCV will match at most one member of the
2494 : * other MCV list. If the operator isn't really equality, there could
2495 : * be multiple matches --- but we don't look for them, both for speed
2496 : * and because the math wouldn't add up...
2497 : */
2498 16638 : matchprodfreq = 0.0;
2499 16638 : nmatches = 0;
2500 522708 : for (i = 0; i < sslot1->nvalues; i++)
2501 : {
2502 : int j;
2503 :
2504 506070 : fcinfo->args[0].value = sslot1->values[i];
2505 :
2506 18984574 : for (j = 0; j < sslot2->nvalues; j++)
2507 : {
2508 : Datum fresult;
2509 :
2510 18643528 : if (hasmatch2[j])
2511 5172842 : continue;
2512 13470686 : fcinfo->args[1].value = sslot2->values[j];
2513 13470686 : fcinfo->isnull = false;
2514 13470686 : fresult = FunctionCallInvoke(fcinfo);
2515 13470686 : if (!fcinfo->isnull && DatumGetBool(fresult))
2516 : {
2517 165024 : hasmatch1[i] = hasmatch2[j] = true;
2518 165024 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
2519 165024 : nmatches++;
2520 165024 : break;
2521 : }
2522 : }
2523 : }
2524 16638 : CLAMP_PROBABILITY(matchprodfreq);
2525 : /* Sum up frequencies of matched and unmatched MCVs */
2526 16638 : matchfreq1 = unmatchfreq1 = 0.0;
2527 522708 : for (i = 0; i < sslot1->nvalues; i++)
2528 : {
2529 506070 : if (hasmatch1[i])
2530 165024 : matchfreq1 += sslot1->numbers[i];
2531 : else
2532 341046 : unmatchfreq1 += sslot1->numbers[i];
2533 : }
2534 16638 : CLAMP_PROBABILITY(matchfreq1);
2535 16638 : CLAMP_PROBABILITY(unmatchfreq1);
2536 16638 : matchfreq2 = unmatchfreq2 = 0.0;
2537 544136 : for (i = 0; i < sslot2->nvalues; i++)
2538 : {
2539 527498 : if (hasmatch2[i])
2540 165024 : matchfreq2 += sslot2->numbers[i];
2541 : else
2542 362474 : unmatchfreq2 += sslot2->numbers[i];
2543 : }
2544 16638 : CLAMP_PROBABILITY(matchfreq2);
2545 16638 : CLAMP_PROBABILITY(unmatchfreq2);
2546 16638 : pfree(hasmatch1);
2547 16638 : pfree(hasmatch2);
2548 :
2549 : /*
2550 : * Compute total frequency of non-null values that are not in the MCV
2551 : * lists.
2552 : */
2553 16638 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2554 16638 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2555 16638 : CLAMP_PROBABILITY(otherfreq1);
2556 16638 : CLAMP_PROBABILITY(otherfreq2);
2557 :
2558 : /*
2559 : * We can estimate the total selectivity from the point of view of
2560 : * relation 1 as: the known selectivity for matched MCVs, plus
2561 : * unmatched MCVs that are assumed to match against random members of
2562 : * relation 2's non-MCV population, plus non-MCV values that are
2563 : * assumed to match against random members of relation 2's unmatched
2564 : * MCVs plus non-MCV values.
2565 : */
2566 16638 : totalsel1 = matchprodfreq;
2567 16638 : if (nd2 > sslot2->nvalues)
2568 10388 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2569 16638 : if (nd2 > nmatches)
2570 13898 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2571 13898 : (nd2 - nmatches);
2572 : /* Same estimate from the point of view of relation 2. */
2573 16638 : totalsel2 = matchprodfreq;
2574 16638 : if (nd1 > sslot1->nvalues)
2575 10474 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2576 16638 : if (nd1 > nmatches)
2577 12874 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2578 12874 : (nd1 - nmatches);
2579 :
2580 : /*
2581 : * Use the smaller of the two estimates. This can be justified in
2582 : * essentially the same terms as given below for the no-stats case: to
2583 : * a first approximation, we are estimating from the point of view of
2584 : * the relation with smaller nd.
2585 : */
2586 16638 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2587 : }
2588 : else
2589 : {
2590 : /*
2591 : * We do not have MCV lists for both sides. Estimate the join
2592 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2593 : * is plausible if we assume that the join operator is strict and the
2594 : * non-null values are about equally distributed: a given non-null
2595 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2596 : * of rel2, so total join rows are at most
2597 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2598 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2599 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2600 : * with MIN() is an upper bound. Using the MIN() means we estimate
2601 : * from the point of view of the relation with smaller nd (since the
2602 : * larger nd is determining the MIN). It is reasonable to assume that
2603 : * most tuples in this rel will have join partners, so the bound is
2604 : * probably reasonably tight and should be taken as-is.
2605 : *
2606 : * XXX Can we be smarter if we have an MCV list for just one side? It
2607 : * seems that if we assume equal distribution for the other side, we
2608 : * end up with the same answer anyway.
2609 : */
2610 172638 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2611 172638 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2612 :
2613 172638 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2614 172638 : if (nd1 > nd2)
2615 78620 : selec /= nd1;
2616 : else
2617 94018 : selec /= nd2;
2618 : }
2619 :
2620 189276 : return selec;
2621 : }
2622 :
2623 : /*
2624 : * eqjoinsel_semi --- eqjoinsel for semi join
2625 : *
2626 : * (Also used for anti join, which we are supposed to estimate the same way.)
2627 : * Caller has ensured that vardata1 is the LHS variable.
2628 : * Unlike eqjoinsel_inner, we have to cope with opfuncoid being InvalidOid.
2629 : */
2630 : static double
2631 7040 : eqjoinsel_semi(Oid opfuncoid, Oid collation,
2632 : VariableStatData *vardata1, VariableStatData *vardata2,
2633 : double nd1, double nd2,
2634 : bool isdefault1, bool isdefault2,
2635 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2636 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2637 : bool have_mcvs1, bool have_mcvs2,
2638 : RelOptInfo *inner_rel)
2639 : {
2640 : double selec;
2641 :
2642 : /*
2643 : * We clamp nd2 to be not more than what we estimate the inner relation's
2644 : * size to be. This is intuitively somewhat reasonable since obviously
2645 : * there can't be more than that many distinct values coming from the
2646 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2647 : * likewise) is that this is the only pathway by which restriction clauses
2648 : * applied to the inner rel will affect the join result size estimate,
2649 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2650 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2651 : * the selectivity of outer-rel restrictions.
2652 : *
2653 : * We can apply this clamping both with respect to the base relation from
2654 : * which the join variable comes (if there is just one), and to the
2655 : * immediate inner input relation of the current join.
2656 : *
2657 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2658 : * great, maybe, but it didn't come out of nowhere either. This is most
2659 : * helpful when the inner relation is empty and consequently has no stats.
2660 : */
2661 7040 : if (vardata2->rel)
2662 : {
2663 7034 : if (nd2 >= vardata2->rel->rows)
2664 : {
2665 5398 : nd2 = vardata2->rel->rows;
2666 5398 : isdefault2 = false;
2667 : }
2668 : }
2669 7040 : if (nd2 >= inner_rel->rows)
2670 : {
2671 5338 : nd2 = inner_rel->rows;
2672 5338 : isdefault2 = false;
2673 : }
2674 :
2675 7040 : if (have_mcvs1 && have_mcvs2 && OidIsValid(opfuncoid))
2676 510 : {
2677 : /*
2678 : * We have most-common-value lists for both relations. Run through
2679 : * the lists to see which MCVs actually join to each other with the
2680 : * given operator. This allows us to determine the exact join
2681 : * selectivity for the portion of the relations represented by the MCV
2682 : * lists. We still have to estimate for the remaining population, but
2683 : * in a skewed distribution this gives us a big leg up in accuracy.
2684 : */
2685 510 : LOCAL_FCINFO(fcinfo, 2);
2686 : FmgrInfo eqproc;
2687 : bool *hasmatch1;
2688 : bool *hasmatch2;
2689 510 : double nullfrac1 = stats1->stanullfrac;
2690 : double matchfreq1,
2691 : uncertainfrac,
2692 : uncertain;
2693 : int i,
2694 : nmatches,
2695 : clamped_nvalues2;
2696 :
2697 : /*
2698 : * The clamping above could have resulted in nd2 being less than
2699 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2700 : * most common values in the relation will appear in the join input,
2701 : * and so compare to only the first nd2 members of the MCV list. Of
2702 : * course this is frequently wrong, but it's the best bet we can make.
2703 : */
2704 510 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2705 :
2706 510 : fmgr_info(opfuncoid, &eqproc);
2707 :
2708 : /*
2709 : * Save a few cycles by setting up the fcinfo struct just once. Using
2710 : * FunctionCallInvoke directly also avoids failure if the eqproc
2711 : * returns NULL, though really equality functions should never do
2712 : * that.
2713 : */
2714 510 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2715 : NULL, NULL);
2716 510 : fcinfo->args[0].isnull = false;
2717 510 : fcinfo->args[1].isnull = false;
2718 :
2719 510 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2720 510 : hasmatch2 = (bool *) palloc0(clamped_nvalues2 * sizeof(bool));
2721 :
2722 : /*
2723 : * Note we assume that each MCV will match at most one member of the
2724 : * other MCV list. If the operator isn't really equality, there could
2725 : * be multiple matches --- but we don't look for them, both for speed
2726 : * and because the math wouldn't add up...
2727 : */
2728 510 : nmatches = 0;
2729 11724 : for (i = 0; i < sslot1->nvalues; i++)
2730 : {
2731 : int j;
2732 :
2733 11214 : fcinfo->args[0].value = sslot1->values[i];
2734 :
2735 424050 : for (j = 0; j < clamped_nvalues2; j++)
2736 : {
2737 : Datum fresult;
2738 :
2739 423042 : if (hasmatch2[j])
2740 374016 : continue;
2741 49026 : fcinfo->args[1].value = sslot2->values[j];
2742 49026 : fcinfo->isnull = false;
2743 49026 : fresult = FunctionCallInvoke(fcinfo);
2744 49026 : if (!fcinfo->isnull && DatumGetBool(fresult))
2745 : {
2746 10206 : hasmatch1[i] = hasmatch2[j] = true;
2747 10206 : nmatches++;
2748 10206 : break;
2749 : }
2750 : }
2751 : }
2752 : /* Sum up frequencies of matched MCVs */
2753 510 : matchfreq1 = 0.0;
2754 11724 : for (i = 0; i < sslot1->nvalues; i++)
2755 : {
2756 11214 : if (hasmatch1[i])
2757 10206 : matchfreq1 += sslot1->numbers[i];
2758 : }
2759 510 : CLAMP_PROBABILITY(matchfreq1);
2760 510 : pfree(hasmatch1);
2761 510 : pfree(hasmatch2);
2762 :
2763 : /*
2764 : * Now we need to estimate the fraction of relation 1 that has at
2765 : * least one join partner. We know for certain that the matched MCVs
2766 : * do, so that gives us a lower bound, but we're really in the dark
2767 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2768 : * assume all non-null rel1 rows have join partners, else assume for
2769 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2770 : * can discount the known-matched MCVs from the distinct-values counts
2771 : * before doing the division.
2772 : *
2773 : * Crude as the above is, it's completely useless if we don't have
2774 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2775 : * nd2 is default, punt and assume half of the uncertain rows have
2776 : * join partners.
2777 : */
2778 510 : if (!isdefault1 && !isdefault2)
2779 : {
2780 510 : nd1 -= nmatches;
2781 510 : nd2 -= nmatches;
2782 510 : if (nd1 <= nd2 || nd2 < 0)
2783 480 : uncertainfrac = 1.0;
2784 : else
2785 30 : uncertainfrac = nd2 / nd1;
2786 : }
2787 : else
2788 0 : uncertainfrac = 0.5;
2789 510 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2790 510 : CLAMP_PROBABILITY(uncertain);
2791 510 : selec = matchfreq1 + uncertainfrac * uncertain;
2792 : }
2793 : else
2794 : {
2795 : /*
2796 : * Without MCV lists for both sides, we can only use the heuristic
2797 : * about nd1 vs nd2.
2798 : */
2799 6530 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2800 :
2801 6530 : if (!isdefault1 && !isdefault2)
2802 : {
2803 4616 : if (nd1 <= nd2 || nd2 < 0)
2804 2500 : selec = 1.0 - nullfrac1;
2805 : else
2806 2116 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2807 : }
2808 : else
2809 1914 : selec = 0.5 * (1.0 - nullfrac1);
2810 : }
2811 :
2812 7040 : return selec;
2813 : }
2814 :
2815 : /*
2816 : * neqjoinsel - Join selectivity of "!="
2817 : */
2818 : Datum
2819 3292 : neqjoinsel(PG_FUNCTION_ARGS)
2820 : {
2821 3292 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2822 3292 : Oid operator = PG_GETARG_OID(1);
2823 3292 : List *args = (List *) PG_GETARG_POINTER(2);
2824 3292 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2825 3292 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2826 3292 : Oid collation = PG_GET_COLLATION();
2827 : float8 result;
2828 :
2829 3292 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
2830 1090 : {
2831 : /*
2832 : * For semi-joins, if there is more than one distinct value in the RHS
2833 : * relation then every non-null LHS row must find a row to join since
2834 : * it can only be equal to one of them. We'll assume that there is
2835 : * always more than one distinct RHS value for the sake of stability,
2836 : * though in theory we could have special cases for empty RHS
2837 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
2838 : * fraction of LHS that has the same value as the single RHS value).
2839 : *
2840 : * For anti-joins, if we use the same assumption that there is more
2841 : * than one distinct key in the RHS relation, then every non-null LHS
2842 : * row must be suppressed by the anti-join.
2843 : *
2844 : * So either way, the selectivity estimate should be 1 - nullfrac.
2845 : */
2846 : VariableStatData leftvar;
2847 : VariableStatData rightvar;
2848 : bool reversed;
2849 : HeapTuple statsTuple;
2850 : double nullfrac;
2851 :
2852 1090 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
2853 1090 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
2854 1090 : if (HeapTupleIsValid(statsTuple))
2855 896 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
2856 : else
2857 194 : nullfrac = 0.0;
2858 1090 : ReleaseVariableStats(leftvar);
2859 1090 : ReleaseVariableStats(rightvar);
2860 :
2861 1090 : result = 1.0 - nullfrac;
2862 : }
2863 : else
2864 : {
2865 : /*
2866 : * We want 1 - eqjoinsel() where the equality operator is the one
2867 : * associated with this != operator, that is, its negator.
2868 : */
2869 2202 : Oid eqop = get_negator(operator);
2870 :
2871 2202 : if (eqop)
2872 : {
2873 : result =
2874 2202 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
2875 : collation,
2876 : PointerGetDatum(root),
2877 : ObjectIdGetDatum(eqop),
2878 : PointerGetDatum(args),
2879 : Int16GetDatum(jointype),
2880 : PointerGetDatum(sjinfo)));
2881 : }
2882 : else
2883 : {
2884 : /* Use default selectivity (should we raise an error instead?) */
2885 0 : result = DEFAULT_EQ_SEL;
2886 : }
2887 2202 : result = 1.0 - result;
2888 : }
2889 :
2890 3292 : PG_RETURN_FLOAT8(result);
2891 : }
2892 :
2893 : /*
2894 : * scalarltjoinsel - Join selectivity of "<" for scalars
2895 : */
2896 : Datum
2897 318 : scalarltjoinsel(PG_FUNCTION_ARGS)
2898 : {
2899 318 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2900 : }
2901 :
2902 : /*
2903 : * scalarlejoinsel - Join selectivity of "<=" for scalars
2904 : */
2905 : Datum
2906 276 : scalarlejoinsel(PG_FUNCTION_ARGS)
2907 : {
2908 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2909 : }
2910 :
2911 : /*
2912 : * scalargtjoinsel - Join selectivity of ">" for scalars
2913 : */
2914 : Datum
2915 228 : scalargtjoinsel(PG_FUNCTION_ARGS)
2916 : {
2917 228 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2918 : }
2919 :
2920 : /*
2921 : * scalargejoinsel - Join selectivity of ">=" for scalars
2922 : */
2923 : Datum
2924 184 : scalargejoinsel(PG_FUNCTION_ARGS)
2925 : {
2926 184 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2927 : }
2928 :
2929 :
2930 : /*
2931 : * mergejoinscansel - Scan selectivity of merge join.
2932 : *
2933 : * A merge join will stop as soon as it exhausts either input stream.
2934 : * Therefore, if we can estimate the ranges of both input variables,
2935 : * we can estimate how much of the input will actually be read. This
2936 : * can have a considerable impact on the cost when using indexscans.
2937 : *
2938 : * Also, we can estimate how much of each input has to be read before the
2939 : * first join pair is found, which will affect the join's startup time.
2940 : *
2941 : * clause should be a clause already known to be mergejoinable. opfamily,
2942 : * strategy, and nulls_first specify the sort ordering being used.
2943 : *
2944 : * The outputs are:
2945 : * *leftstart is set to the fraction of the left-hand variable expected
2946 : * to be scanned before the first join pair is found (0 to 1).
2947 : * *leftend is set to the fraction of the left-hand variable expected
2948 : * to be scanned before the join terminates (0 to 1).
2949 : * *rightstart, *rightend similarly for the right-hand variable.
2950 : */
2951 : void
2952 95742 : mergejoinscansel(PlannerInfo *root, Node *clause,
2953 : Oid opfamily, int strategy, bool nulls_first,
2954 : Selectivity *leftstart, Selectivity *leftend,
2955 : Selectivity *rightstart, Selectivity *rightend)
2956 : {
2957 : Node *left,
2958 : *right;
2959 : VariableStatData leftvar,
2960 : rightvar;
2961 : int op_strategy;
2962 : Oid op_lefttype;
2963 : Oid op_righttype;
2964 : Oid opno,
2965 : collation,
2966 : lsortop,
2967 : rsortop,
2968 : lstatop,
2969 : rstatop,
2970 : ltop,
2971 : leop,
2972 : revltop,
2973 : revleop;
2974 : bool isgt;
2975 : Datum leftmin,
2976 : leftmax,
2977 : rightmin,
2978 : rightmax;
2979 : double selec;
2980 :
2981 : /* Set default results if we can't figure anything out. */
2982 : /* XXX should default "start" fraction be a bit more than 0? */
2983 95742 : *leftstart = *rightstart = 0.0;
2984 95742 : *leftend = *rightend = 1.0;
2985 :
2986 : /* Deconstruct the merge clause */
2987 95742 : if (!is_opclause(clause))
2988 0 : return; /* shouldn't happen */
2989 95742 : opno = ((OpExpr *) clause)->opno;
2990 95742 : collation = ((OpExpr *) clause)->inputcollid;
2991 95742 : left = get_leftop((Expr *) clause);
2992 95742 : right = get_rightop((Expr *) clause);
2993 95742 : if (!right)
2994 0 : return; /* shouldn't happen */
2995 :
2996 : /* Look for stats for the inputs */
2997 95742 : examine_variable(root, left, 0, &leftvar);
2998 95742 : examine_variable(root, right, 0, &rightvar);
2999 :
3000 : /* Extract the operator's declared left/right datatypes */
3001 95742 : get_op_opfamily_properties(opno, opfamily, false,
3002 : &op_strategy,
3003 : &op_lefttype,
3004 : &op_righttype);
3005 : Assert(op_strategy == BTEqualStrategyNumber);
3006 :
3007 : /*
3008 : * Look up the various operators we need. If we don't find them all, it
3009 : * probably means the opfamily is broken, but we just fail silently.
3010 : *
3011 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3012 : * operator, regardless of which sort direction we are considering.
3013 : */
3014 95742 : switch (strategy)
3015 : {
3016 95688 : case BTLessStrategyNumber:
3017 95688 : isgt = false;
3018 95688 : if (op_lefttype == op_righttype)
3019 : {
3020 : /* easy case */
3021 94438 : ltop = get_opfamily_member(opfamily,
3022 : op_lefttype, op_righttype,
3023 : BTLessStrategyNumber);
3024 94438 : leop = get_opfamily_member(opfamily,
3025 : op_lefttype, op_righttype,
3026 : BTLessEqualStrategyNumber);
3027 94438 : lsortop = ltop;
3028 94438 : rsortop = ltop;
3029 94438 : lstatop = lsortop;
3030 94438 : rstatop = rsortop;
3031 94438 : revltop = ltop;
3032 94438 : revleop = leop;
3033 : }
3034 : else
3035 : {
3036 1250 : ltop = get_opfamily_member(opfamily,
3037 : op_lefttype, op_righttype,
3038 : BTLessStrategyNumber);
3039 1250 : leop = get_opfamily_member(opfamily,
3040 : op_lefttype, op_righttype,
3041 : BTLessEqualStrategyNumber);
3042 1250 : lsortop = get_opfamily_member(opfamily,
3043 : op_lefttype, op_lefttype,
3044 : BTLessStrategyNumber);
3045 1250 : rsortop = get_opfamily_member(opfamily,
3046 : op_righttype, op_righttype,
3047 : BTLessStrategyNumber);
3048 1250 : lstatop = lsortop;
3049 1250 : rstatop = rsortop;
3050 1250 : revltop = get_opfamily_member(opfamily,
3051 : op_righttype, op_lefttype,
3052 : BTLessStrategyNumber);
3053 1250 : revleop = get_opfamily_member(opfamily,
3054 : op_righttype, op_lefttype,
3055 : BTLessEqualStrategyNumber);
3056 : }
3057 95688 : break;
3058 54 : case BTGreaterStrategyNumber:
3059 : /* descending-order case */
3060 54 : isgt = true;
3061 54 : if (op_lefttype == op_righttype)
3062 : {
3063 : /* easy case */
3064 54 : ltop = get_opfamily_member(opfamily,
3065 : op_lefttype, op_righttype,
3066 : BTGreaterStrategyNumber);
3067 54 : leop = get_opfamily_member(opfamily,
3068 : op_lefttype, op_righttype,
3069 : BTGreaterEqualStrategyNumber);
3070 54 : lsortop = ltop;
3071 54 : rsortop = ltop;
3072 54 : lstatop = get_opfamily_member(opfamily,
3073 : op_lefttype, op_lefttype,
3074 : BTLessStrategyNumber);
3075 54 : rstatop = lstatop;
3076 54 : revltop = ltop;
3077 54 : revleop = leop;
3078 : }
3079 : else
3080 : {
3081 0 : ltop = get_opfamily_member(opfamily,
3082 : op_lefttype, op_righttype,
3083 : BTGreaterStrategyNumber);
3084 0 : leop = get_opfamily_member(opfamily,
3085 : op_lefttype, op_righttype,
3086 : BTGreaterEqualStrategyNumber);
3087 0 : lsortop = get_opfamily_member(opfamily,
3088 : op_lefttype, op_lefttype,
3089 : BTGreaterStrategyNumber);
3090 0 : rsortop = get_opfamily_member(opfamily,
3091 : op_righttype, op_righttype,
3092 : BTGreaterStrategyNumber);
3093 0 : lstatop = get_opfamily_member(opfamily,
3094 : op_lefttype, op_lefttype,
3095 : BTLessStrategyNumber);
3096 0 : rstatop = get_opfamily_member(opfamily,
3097 : op_righttype, op_righttype,
3098 : BTLessStrategyNumber);
3099 0 : revltop = get_opfamily_member(opfamily,
3100 : op_righttype, op_lefttype,
3101 : BTGreaterStrategyNumber);
3102 0 : revleop = get_opfamily_member(opfamily,
3103 : op_righttype, op_lefttype,
3104 : BTGreaterEqualStrategyNumber);
3105 : }
3106 54 : break;
3107 0 : default:
3108 0 : goto fail; /* shouldn't get here */
3109 : }
3110 :
3111 95742 : if (!OidIsValid(lsortop) ||
3112 95742 : !OidIsValid(rsortop) ||
3113 95742 : !OidIsValid(lstatop) ||
3114 95742 : !OidIsValid(rstatop) ||
3115 95730 : !OidIsValid(ltop) ||
3116 95730 : !OidIsValid(leop) ||
3117 95730 : !OidIsValid(revltop) ||
3118 : !OidIsValid(revleop))
3119 12 : goto fail; /* insufficient info in catalogs */
3120 :
3121 : /* Try to get ranges of both inputs */
3122 95730 : if (!isgt)
3123 : {
3124 95676 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3125 : &leftmin, &leftmax))
3126 19382 : goto fail; /* no range available from stats */
3127 76294 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3128 : &rightmin, &rightmax))
3129 18610 : goto fail; /* no range available from stats */
3130 : }
3131 : else
3132 : {
3133 : /* need to swap the max and min */
3134 54 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3135 : &leftmax, &leftmin))
3136 30 : goto fail; /* no range available from stats */
3137 24 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3138 : &rightmax, &rightmin))
3139 0 : goto fail; /* no range available from stats */
3140 : }
3141 :
3142 : /*
3143 : * Now, the fraction of the left variable that will be scanned is the
3144 : * fraction that's <= the right-side maximum value. But only believe
3145 : * non-default estimates, else stick with our 1.0.
3146 : */
3147 57708 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3148 : rightmax, op_righttype);
3149 57708 : if (selec != DEFAULT_INEQ_SEL)
3150 57702 : *leftend = selec;
3151 :
3152 : /* And similarly for the right variable. */
3153 57708 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3154 : leftmax, op_lefttype);
3155 57708 : if (selec != DEFAULT_INEQ_SEL)
3156 57708 : *rightend = selec;
3157 :
3158 : /*
3159 : * Only one of the two "end" fractions can really be less than 1.0;
3160 : * believe the smaller estimate and reset the other one to exactly 1.0. If
3161 : * we get exactly equal estimates (as can easily happen with self-joins),
3162 : * believe neither.
3163 : */
3164 57708 : if (*leftend > *rightend)
3165 29028 : *leftend = 1.0;
3166 28680 : else if (*leftend < *rightend)
3167 22870 : *rightend = 1.0;
3168 : else
3169 5810 : *leftend = *rightend = 1.0;
3170 :
3171 : /*
3172 : * Also, the fraction of the left variable that will be scanned before the
3173 : * first join pair is found is the fraction that's < the right-side
3174 : * minimum value. But only believe non-default estimates, else stick with
3175 : * our own default.
3176 : */
3177 57708 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3178 : rightmin, op_righttype);
3179 57708 : if (selec != DEFAULT_INEQ_SEL)
3180 57708 : *leftstart = selec;
3181 :
3182 : /* And similarly for the right variable. */
3183 57708 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3184 : leftmin, op_lefttype);
3185 57708 : if (selec != DEFAULT_INEQ_SEL)
3186 57708 : *rightstart = selec;
3187 :
3188 : /*
3189 : * Only one of the two "start" fractions can really be more than zero;
3190 : * believe the larger estimate and reset the other one to exactly 0.0. If
3191 : * we get exactly equal estimates (as can easily happen with self-joins),
3192 : * believe neither.
3193 : */
3194 57708 : if (*leftstart < *rightstart)
3195 16306 : *leftstart = 0.0;
3196 41402 : else if (*leftstart > *rightstart)
3197 25684 : *rightstart = 0.0;
3198 : else
3199 15718 : *leftstart = *rightstart = 0.0;
3200 :
3201 : /*
3202 : * If the sort order is nulls-first, we're going to have to skip over any
3203 : * nulls too. These would not have been counted by scalarineqsel, and we
3204 : * can safely add in this fraction regardless of whether we believe
3205 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3206 : */
3207 57708 : if (nulls_first)
3208 : {
3209 : Form_pg_statistic stats;
3210 :
3211 24 : if (HeapTupleIsValid(leftvar.statsTuple))
3212 : {
3213 24 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3214 24 : *leftstart += stats->stanullfrac;
3215 24 : CLAMP_PROBABILITY(*leftstart);
3216 24 : *leftend += stats->stanullfrac;
3217 24 : CLAMP_PROBABILITY(*leftend);
3218 : }
3219 24 : if (HeapTupleIsValid(rightvar.statsTuple))
3220 : {
3221 24 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3222 24 : *rightstart += stats->stanullfrac;
3223 24 : CLAMP_PROBABILITY(*rightstart);
3224 24 : *rightend += stats->stanullfrac;
3225 24 : CLAMP_PROBABILITY(*rightend);
3226 : }
3227 : }
3228 :
3229 : /* Disbelieve start >= end, just in case that can happen */
3230 57708 : if (*leftstart >= *leftend)
3231 : {
3232 194 : *leftstart = 0.0;
3233 194 : *leftend = 1.0;
3234 : }
3235 57708 : if (*rightstart >= *rightend)
3236 : {
3237 914 : *rightstart = 0.0;
3238 914 : *rightend = 1.0;
3239 : }
3240 :
3241 56794 : fail:
3242 95742 : ReleaseVariableStats(leftvar);
3243 95742 : ReleaseVariableStats(rightvar);
3244 : }
3245 :
3246 :
3247 : /*
3248 : * matchingsel -- generic matching-operator selectivity support
3249 : *
3250 : * Use these for any operators that (a) are on data types for which we collect
3251 : * standard statistics, and (b) have behavior for which the default estimate
3252 : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3253 : * operators.
3254 : */
3255 :
3256 : Datum
3257 1106 : matchingsel(PG_FUNCTION_ARGS)
3258 : {
3259 1106 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3260 1106 : Oid operator = PG_GETARG_OID(1);
3261 1106 : List *args = (List *) PG_GETARG_POINTER(2);
3262 1106 : int varRelid = PG_GETARG_INT32(3);
3263 1106 : Oid collation = PG_GET_COLLATION();
3264 : double selec;
3265 :
3266 : /* Use generic restriction selectivity logic. */
3267 1106 : selec = generic_restriction_selectivity(root, operator, collation,
3268 : args, varRelid,
3269 : DEFAULT_MATCHING_SEL);
3270 :
3271 1106 : PG_RETURN_FLOAT8((float8) selec);
3272 : }
3273 :
3274 : Datum
3275 6 : matchingjoinsel(PG_FUNCTION_ARGS)
3276 : {
3277 : /* Just punt, for the moment. */
3278 6 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3279 : }
3280 :
3281 :
3282 : /*
3283 : * Helper routine for estimate_num_groups: add an item to a list of
3284 : * GroupVarInfos, but only if it's not known equal to any of the existing
3285 : * entries.
3286 : */
3287 : typedef struct
3288 : {
3289 : Node *var; /* might be an expression, not just a Var */
3290 : RelOptInfo *rel; /* relation it belongs to */
3291 : double ndistinct; /* # distinct values */
3292 : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3293 : } GroupVarInfo;
3294 :
3295 : static List *
3296 286028 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3297 : Node *var, VariableStatData *vardata)
3298 : {
3299 : GroupVarInfo *varinfo;
3300 : double ndistinct;
3301 : bool isdefault;
3302 : ListCell *lc;
3303 :
3304 286028 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3305 :
3306 329128 : foreach(lc, varinfos)
3307 : {
3308 43852 : varinfo = (GroupVarInfo *) lfirst(lc);
3309 :
3310 : /* Drop exact duplicates */
3311 43852 : if (equal(var, varinfo->var))
3312 752 : return varinfos;
3313 :
3314 : /*
3315 : * Drop known-equal vars, but only if they belong to different
3316 : * relations (see comments for estimate_num_groups)
3317 : */
3318 46430 : if (vardata->rel != varinfo->rel &&
3319 3210 : exprs_known_equal(root, var, varinfo->var))
3320 : {
3321 120 : if (varinfo->ndistinct <= ndistinct)
3322 : {
3323 : /* Keep older item, forget new one */
3324 120 : return varinfos;
3325 : }
3326 : else
3327 : {
3328 : /* Delete the older item */
3329 0 : varinfos = foreach_delete_current(varinfos, lc);
3330 : }
3331 : }
3332 : }
3333 :
3334 285276 : varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
3335 :
3336 285276 : varinfo->var = var;
3337 285276 : varinfo->rel = vardata->rel;
3338 285276 : varinfo->ndistinct = ndistinct;
3339 285276 : varinfo->isdefault = isdefault;
3340 285276 : varinfos = lappend(varinfos, varinfo);
3341 285276 : return varinfos;
3342 : }
3343 :
3344 : /*
3345 : * estimate_num_groups - Estimate number of groups in a grouped query
3346 : *
3347 : * Given a query having a GROUP BY clause, estimate how many groups there
3348 : * will be --- ie, the number of distinct combinations of the GROUP BY
3349 : * expressions.
3350 : *
3351 : * This routine is also used to estimate the number of rows emitted by
3352 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3353 : * actually, we only use it for DISTINCT when there's no grouping or
3354 : * aggregation ahead of the DISTINCT.)
3355 : *
3356 : * Inputs:
3357 : * root - the query
3358 : * groupExprs - list of expressions being grouped by
3359 : * input_rows - number of rows estimated to arrive at the group/unique
3360 : * filter step
3361 : * pgset - NULL, or a List** pointing to a grouping set to filter the
3362 : * groupExprs against
3363 : *
3364 : * Outputs:
3365 : * estinfo - When passed as non-NULL, the function will set bits in the
3366 : * "flags" field in order to provide callers with additional information
3367 : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3368 : * bit if we used any default values in the estimation.
3369 : *
3370 : * Given the lack of any cross-correlation statistics in the system, it's
3371 : * impossible to do anything really trustworthy with GROUP BY conditions
3372 : * involving multiple Vars. We should however avoid assuming the worst
3373 : * case (all possible cross-product terms actually appear as groups) since
3374 : * very often the grouped-by Vars are highly correlated. Our current approach
3375 : * is as follows:
3376 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3377 : * independently of their content, and are ignored in the subsequent
3378 : * steps. This is mainly because tests like "col IS NULL" break the
3379 : * heuristic used in step 2 especially badly.
3380 : * 2. Reduce the given expressions to a list of unique Vars used. For
3381 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3382 : * It is clearly correct not to count the same Var more than once.
3383 : * It is also reasonable to treat f(x) the same as x: f() cannot
3384 : * increase the number of distinct values (unless it is volatile,
3385 : * which we consider unlikely for grouping), but it probably won't
3386 : * reduce the number of distinct values much either.
3387 : * As a special case, if a GROUP BY expression can be matched to an
3388 : * expressional index for which we have statistics, then we treat the
3389 : * whole expression as though it were just a Var.
3390 : * 3. If the list contains Vars of different relations that are known equal
3391 : * due to equivalence classes, then drop all but one of the Vars from each
3392 : * known-equal set, keeping the one with smallest estimated # of values
3393 : * (since the extra values of the others can't appear in joined rows).
3394 : * Note the reason we only consider Vars of different relations is that
3395 : * if we considered ones of the same rel, we'd be double-counting the
3396 : * restriction selectivity of the equality in the next step.
3397 : * 4. For Vars within a single source rel, we multiply together the numbers
3398 : * of values, clamp to the number of rows in the rel (divided by 10 if
3399 : * more than one Var), and then multiply by a factor based on the
3400 : * selectivity of the restriction clauses for that rel. When there's
3401 : * more than one Var, the initial product is probably too high (it's the
3402 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3403 : * helpful heuristic for not letting the estimate get out of hand. (The
3404 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3405 : * we multiply by to adjust for the restriction selectivity assumes that
3406 : * the restriction clauses are independent of the grouping, which may not
3407 : * be a valid assumption, but it's hard to do better.
3408 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3409 : * rel, and multiply the results together.
3410 : * Note that rels not containing grouped Vars are ignored completely, as are
3411 : * join clauses. Such rels cannot increase the number of groups, and we
3412 : * assume such clauses do not reduce the number either (somewhat bogus,
3413 : * but we don't have the info to do better).
3414 : */
3415 : double
3416 252334 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3417 : List **pgset, EstimationInfo *estinfo)
3418 : {
3419 252334 : List *varinfos = NIL;
3420 252334 : double srf_multiplier = 1.0;
3421 : double numdistinct;
3422 : ListCell *l;
3423 : int i;
3424 :
3425 : /* Zero the estinfo output parameter, if non-NULL */
3426 252334 : if (estinfo != NULL)
3427 225630 : memset(estinfo, 0, sizeof(EstimationInfo));
3428 :
3429 : /*
3430 : * We don't ever want to return an estimate of zero groups, as that tends
3431 : * to lead to division-by-zero and other unpleasantness. The input_rows
3432 : * estimate is usually already at least 1, but clamp it just in case it
3433 : * isn't.
3434 : */
3435 252334 : input_rows = clamp_row_est(input_rows);
3436 :
3437 : /*
3438 : * If no grouping columns, there's exactly one group. (This can't happen
3439 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3440 : * corner cases with set operations.)
3441 : */
3442 252334 : if (groupExprs == NIL || (pgset && *pgset == NIL))
3443 956 : return 1.0;
3444 :
3445 : /*
3446 : * Count groups derived from boolean grouping expressions. For other
3447 : * expressions, find the unique Vars used, treating an expression as a Var
3448 : * if we can find stats for it. For each one, record the statistical
3449 : * estimate of number of distinct values (total in its table, without
3450 : * regard for filtering).
3451 : */
3452 251378 : numdistinct = 1.0;
3453 :
3454 251378 : i = 0;
3455 535660 : foreach(l, groupExprs)
3456 : {
3457 284330 : Node *groupexpr = (Node *) lfirst(l);
3458 : double this_srf_multiplier;
3459 : VariableStatData vardata;
3460 : List *varshere;
3461 : ListCell *l2;
3462 :
3463 : /* is expression in this grouping set? */
3464 284330 : if (pgset && !list_member_int(*pgset, i++))
3465 238492 : continue;
3466 :
3467 : /*
3468 : * Set-returning functions in grouping columns are a bit problematic.
3469 : * The code below will effectively ignore their SRF nature and come up
3470 : * with a numdistinct estimate as though they were scalar functions.
3471 : * We compensate by scaling up the end result by the largest SRF
3472 : * rowcount estimate. (This will be an overestimate if the SRF
3473 : * produces multiple copies of any output value, but it seems best to
3474 : * assume the SRF's outputs are distinct. In any case, it's probably
3475 : * pointless to worry too much about this without much better
3476 : * estimates for SRF output rowcounts than we have today.)
3477 : */
3478 283602 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3479 283602 : if (srf_multiplier < this_srf_multiplier)
3480 132 : srf_multiplier = this_srf_multiplier;
3481 :
3482 : /* Short-circuit for expressions returning boolean */
3483 283602 : if (exprType(groupexpr) == BOOLOID)
3484 : {
3485 50 : numdistinct *= 2.0;
3486 50 : continue;
3487 : }
3488 :
3489 : /*
3490 : * If examine_variable is able to deduce anything about the GROUP BY
3491 : * expression, treat it as a single variable even if it's really more
3492 : * complicated.
3493 : *
3494 : * XXX This has the consequence that if there's a statistics object on
3495 : * the expression, we don't split it into individual Vars. This
3496 : * affects our selection of statistics in
3497 : * estimate_multivariate_ndistinct, because it's probably better to
3498 : * use more accurate estimate for each expression and treat them as
3499 : * independent, than to combine estimates for the extracted variables
3500 : * when we don't know how that relates to the expressions.
3501 : */
3502 283552 : examine_variable(root, groupexpr, 0, &vardata);
3503 283552 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3504 : {
3505 237198 : varinfos = add_unique_group_var(root, varinfos,
3506 : groupexpr, &vardata);
3507 237198 : ReleaseVariableStats(vardata);
3508 237198 : continue;
3509 : }
3510 46354 : ReleaseVariableStats(vardata);
3511 :
3512 : /*
3513 : * Else pull out the component Vars. Handle PlaceHolderVars by
3514 : * recursing into their arguments (effectively assuming that the
3515 : * PlaceHolderVar doesn't change the number of groups, which boils
3516 : * down to ignoring the possible addition of nulls to the result set).
3517 : */
3518 46354 : varshere = pull_var_clause(groupexpr,
3519 : PVC_RECURSE_AGGREGATES |
3520 : PVC_RECURSE_WINDOWFUNCS |
3521 : PVC_RECURSE_PLACEHOLDERS);
3522 :
3523 : /*
3524 : * If we find any variable-free GROUP BY item, then either it is a
3525 : * constant (and we can ignore it) or it contains a volatile function;
3526 : * in the latter case we punt and assume that each input row will
3527 : * yield a distinct group.
3528 : */
3529 46354 : if (varshere == NIL)
3530 : {
3531 564 : if (contain_volatile_functions(groupexpr))
3532 48 : return input_rows;
3533 516 : continue;
3534 : }
3535 :
3536 : /*
3537 : * Else add variables to varinfos list
3538 : */
3539 94620 : foreach(l2, varshere)
3540 : {
3541 48830 : Node *var = (Node *) lfirst(l2);
3542 :
3543 48830 : examine_variable(root, var, 0, &vardata);
3544 48830 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3545 48830 : ReleaseVariableStats(vardata);
3546 : }
3547 : }
3548 :
3549 : /*
3550 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3551 : * list.
3552 : */
3553 251330 : if (varinfos == NIL)
3554 : {
3555 : /* Apply SRF multiplier as we would do in the long path */
3556 292 : numdistinct *= srf_multiplier;
3557 : /* Round off */
3558 292 : numdistinct = ceil(numdistinct);
3559 : /* Guard against out-of-range answers */
3560 292 : if (numdistinct > input_rows)
3561 14 : numdistinct = input_rows;
3562 292 : if (numdistinct < 1.0)
3563 0 : numdistinct = 1.0;
3564 292 : return numdistinct;
3565 : }
3566 :
3567 : /*
3568 : * Group Vars by relation and estimate total numdistinct.
3569 : *
3570 : * For each iteration of the outer loop, we process the frontmost Var in
3571 : * varinfos, plus all other Vars in the same relation. We remove these
3572 : * Vars from the newvarinfos list for the next iteration. This is the
3573 : * easiest way to group Vars of same rel together.
3574 : */
3575 : do
3576 : {
3577 252766 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3578 252766 : RelOptInfo *rel = varinfo1->rel;
3579 252766 : double reldistinct = 1;
3580 252766 : double relmaxndistinct = reldistinct;
3581 252766 : int relvarcount = 0;
3582 252766 : List *newvarinfos = NIL;
3583 252766 : List *relvarinfos = NIL;
3584 :
3585 : /*
3586 : * Split the list of varinfos in two - one for the current rel, one
3587 : * for remaining Vars on other rels.
3588 : */
3589 252766 : relvarinfos = lappend(relvarinfos, varinfo1);
3590 287652 : for_each_from(l, varinfos, 1)
3591 : {
3592 34886 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3593 :
3594 34886 : if (varinfo2->rel == varinfo1->rel)
3595 : {
3596 : /* varinfos on current rel */
3597 32510 : relvarinfos = lappend(relvarinfos, varinfo2);
3598 : }
3599 : else
3600 : {
3601 : /* not time to process varinfo2 yet */
3602 2376 : newvarinfos = lappend(newvarinfos, varinfo2);
3603 : }
3604 : }
3605 :
3606 : /*
3607 : * Get the numdistinct estimate for the Vars of this rel. We
3608 : * iteratively search for multivariate n-distinct with maximum number
3609 : * of vars; assuming that each var group is independent of the others,
3610 : * we multiply them together. Any remaining relvarinfos after no more
3611 : * multivariate matches are found are assumed independent too, so
3612 : * their individual ndistinct estimates are multiplied also.
3613 : *
3614 : * While iterating, count how many separate numdistinct values we
3615 : * apply. We apply a fudge factor below, but only if we multiplied
3616 : * more than one such values.
3617 : */
3618 505658 : while (relvarinfos)
3619 : {
3620 : double mvndistinct;
3621 :
3622 252892 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3623 : &mvndistinct))
3624 : {
3625 402 : reldistinct *= mvndistinct;
3626 402 : if (relmaxndistinct < mvndistinct)
3627 390 : relmaxndistinct = mvndistinct;
3628 402 : relvarcount++;
3629 : }
3630 : else
3631 : {
3632 536914 : foreach(l, relvarinfos)
3633 : {
3634 284424 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3635 :
3636 284424 : reldistinct *= varinfo2->ndistinct;
3637 284424 : if (relmaxndistinct < varinfo2->ndistinct)
3638 257260 : relmaxndistinct = varinfo2->ndistinct;
3639 284424 : relvarcount++;
3640 :
3641 : /*
3642 : * When varinfo2's isdefault is set then we'd better set
3643 : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3644 : */
3645 284424 : if (estinfo != NULL && varinfo2->isdefault)
3646 12170 : estinfo->flags |= SELFLAG_USED_DEFAULT;
3647 : }
3648 :
3649 : /* we're done with this relation */
3650 252490 : relvarinfos = NIL;
3651 : }
3652 : }
3653 :
3654 : /*
3655 : * Sanity check --- don't divide by zero if empty relation.
3656 : */
3657 : Assert(IS_SIMPLE_REL(rel));
3658 252766 : if (rel->tuples > 0)
3659 : {
3660 : /*
3661 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
3662 : * fudge factor is because the Vars are probably correlated but we
3663 : * don't know by how much. We should never clamp to less than the
3664 : * largest ndistinct value for any of the Vars, though, since
3665 : * there will surely be at least that many groups.
3666 : */
3667 251818 : double clamp = rel->tuples;
3668 :
3669 251818 : if (relvarcount > 1)
3670 : {
3671 29600 : clamp *= 0.1;
3672 29600 : if (clamp < relmaxndistinct)
3673 : {
3674 28126 : clamp = relmaxndistinct;
3675 : /* for sanity in case some ndistinct is too large: */
3676 28126 : if (clamp > rel->tuples)
3677 78 : clamp = rel->tuples;
3678 : }
3679 : }
3680 251818 : if (reldistinct > clamp)
3681 27840 : reldistinct = clamp;
3682 :
3683 : /*
3684 : * Update the estimate based on the restriction selectivity,
3685 : * guarding against division by zero when reldistinct is zero.
3686 : * Also skip this if we know that we are returning all rows.
3687 : */
3688 251818 : if (reldistinct > 0 && rel->rows < rel->tuples)
3689 : {
3690 : /*
3691 : * Given a table containing N rows with n distinct values in a
3692 : * uniform distribution, if we select p rows at random then
3693 : * the expected number of distinct values selected is
3694 : *
3695 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
3696 : *
3697 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
3698 : *
3699 : * See "Approximating block accesses in database
3700 : * organizations", S. B. Yao, Communications of the ACM,
3701 : * Volume 20 Issue 4, April 1977 Pages 260-261.
3702 : *
3703 : * Alternatively, re-arranging the terms from the factorials,
3704 : * this may be written as
3705 : *
3706 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
3707 : *
3708 : * This form of the formula is more efficient to compute in
3709 : * the common case where p is larger than N/n. Additionally,
3710 : * as pointed out by Dell'Era, if i << N for all terms in the
3711 : * product, it can be approximated by
3712 : *
3713 : * n * (1 - ((N-p)/N)^(N/n))
3714 : *
3715 : * See "Expected distinct values when selecting from a bag
3716 : * without replacement", Alberto Dell'Era,
3717 : * http://www.adellera.it/investigations/distinct_balls/.
3718 : *
3719 : * The condition i << N is equivalent to n >> 1, so this is a
3720 : * good approximation when the number of distinct values in
3721 : * the table is large. It turns out that this formula also
3722 : * works well even when n is small.
3723 : */
3724 70530 : reldistinct *=
3725 70530 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
3726 70530 : rel->tuples / reldistinct));
3727 : }
3728 251818 : reldistinct = clamp_row_est(reldistinct);
3729 :
3730 : /*
3731 : * Update estimate of total distinct groups.
3732 : */
3733 251818 : numdistinct *= reldistinct;
3734 : }
3735 :
3736 252766 : varinfos = newvarinfos;
3737 252766 : } while (varinfos != NIL);
3738 :
3739 : /* Now we can account for the effects of any SRFs */
3740 251038 : numdistinct *= srf_multiplier;
3741 :
3742 : /* Round off */
3743 251038 : numdistinct = ceil(numdistinct);
3744 :
3745 : /* Guard against out-of-range answers */
3746 251038 : if (numdistinct > input_rows)
3747 56172 : numdistinct = input_rows;
3748 251038 : if (numdistinct < 1.0)
3749 0 : numdistinct = 1.0;
3750 :
3751 251038 : return numdistinct;
3752 : }
3753 :
3754 : /*
3755 : * Estimate hash bucket statistics when the specified expression is used
3756 : * as a hash key for the given number of buckets.
3757 : *
3758 : * This attempts to determine two values:
3759 : *
3760 : * 1. The frequency of the most common value of the expression (returns
3761 : * zero into *mcv_freq if we can't get that).
3762 : *
3763 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
3764 : * divided by total tuples in relation.
3765 : *
3766 : * XXX This is really pretty bogus since we're effectively assuming that the
3767 : * distribution of hash keys will be the same after applying restriction
3768 : * clauses as it was in the underlying relation. However, we are not nearly
3769 : * smart enough to figure out how the restrict clauses might change the
3770 : * distribution, so this will have to do for now.
3771 : *
3772 : * We are passed the number of buckets the executor will use for the given
3773 : * input relation. If the data were perfectly distributed, with the same
3774 : * number of tuples going into each available bucket, then the bucketsize
3775 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
3776 : * only if (a) there are at least nbuckets distinct data values, and (b)
3777 : * we have a not-too-skewed data distribution. Otherwise the buckets will
3778 : * be nonuniformly occupied. If the other relation in the join has a key
3779 : * distribution similar to this one's, then the most-loaded buckets are
3780 : * exactly those that will be probed most often. Therefore, the "average"
3781 : * bucket size for costing purposes should really be taken as something close
3782 : * to the "worst case" bucket size. We try to estimate this by adjusting the
3783 : * fraction if there are too few distinct data values, and then scaling up
3784 : * by the ratio of the most common value's frequency to the average frequency.
3785 : *
3786 : * If no statistics are available, use a default estimate of 0.1. This will
3787 : * discourage use of a hash rather strongly if the inner relation is large,
3788 : * which is what we want. We do not want to hash unless we know that the
3789 : * inner rel is well-dispersed (or the alternatives seem much worse).
3790 : *
3791 : * The caller should also check that the mcv_freq is not so large that the
3792 : * most common value would by itself require an impractically large bucket.
3793 : * In a hash join, the executor can split buckets if they get too big, but
3794 : * obviously that doesn't help for a bucket that contains many duplicates of
3795 : * the same value.
3796 : */
3797 : void
3798 129264 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
3799 : Selectivity *mcv_freq,
3800 : Selectivity *bucketsize_frac)
3801 : {
3802 : VariableStatData vardata;
3803 : double estfract,
3804 : ndistinct,
3805 : stanullfrac,
3806 : avgfreq;
3807 : bool isdefault;
3808 : AttStatsSlot sslot;
3809 :
3810 129264 : examine_variable(root, hashkey, 0, &vardata);
3811 :
3812 : /* Look up the frequency of the most common value, if available */
3813 129264 : *mcv_freq = 0.0;
3814 :
3815 129264 : if (HeapTupleIsValid(vardata.statsTuple))
3816 : {
3817 88950 : if (get_attstatsslot(&sslot, vardata.statsTuple,
3818 : STATISTIC_KIND_MCV, InvalidOid,
3819 : ATTSTATSSLOT_NUMBERS))
3820 : {
3821 : /*
3822 : * The first MCV stat is for the most common value.
3823 : */
3824 40150 : if (sslot.nnumbers > 0)
3825 40150 : *mcv_freq = sslot.numbers[0];
3826 40150 : free_attstatsslot(&sslot);
3827 : }
3828 : }
3829 :
3830 : /* Get number of distinct values */
3831 129264 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
3832 :
3833 : /*
3834 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
3835 : * mcv_freq is known to be even higher than that, use it instead.
3836 : */
3837 129264 : if (isdefault)
3838 : {
3839 17452 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
3840 17452 : ReleaseVariableStats(vardata);
3841 17452 : return;
3842 : }
3843 :
3844 : /* Get fraction that are null */
3845 111812 : if (HeapTupleIsValid(vardata.statsTuple))
3846 : {
3847 : Form_pg_statistic stats;
3848 :
3849 88932 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
3850 88932 : stanullfrac = stats->stanullfrac;
3851 : }
3852 : else
3853 22880 : stanullfrac = 0.0;
3854 :
3855 : /* Compute avg freq of all distinct data values in raw relation */
3856 111812 : avgfreq = (1.0 - stanullfrac) / ndistinct;
3857 :
3858 : /*
3859 : * Adjust ndistinct to account for restriction clauses. Observe we are
3860 : * assuming that the data distribution is affected uniformly by the
3861 : * restriction clauses!
3862 : *
3863 : * XXX Possibly better way, but much more expensive: multiply by
3864 : * selectivity of rel's restriction clauses that mention the target Var.
3865 : */
3866 111812 : if (vardata.rel && vardata.rel->tuples > 0)
3867 : {
3868 111798 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
3869 111798 : ndistinct = clamp_row_est(ndistinct);
3870 : }
3871 :
3872 : /*
3873 : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
3874 : * number of buckets is less than the expected number of distinct values;
3875 : * otherwise it is 1/ndistinct.
3876 : */
3877 111812 : if (ndistinct > nbuckets)
3878 96 : estfract = 1.0 / nbuckets;
3879 : else
3880 111716 : estfract = 1.0 / ndistinct;
3881 :
3882 : /*
3883 : * Adjust estimated bucketsize upward to account for skewed distribution.
3884 : */
3885 111812 : if (avgfreq > 0.0 && *mcv_freq > avgfreq)
3886 35508 : estfract *= *mcv_freq / avgfreq;
3887 :
3888 : /*
3889 : * Clamp bucketsize to sane range (the above adjustment could easily
3890 : * produce an out-of-range result). We set the lower bound a little above
3891 : * zero, since zero isn't a very sane result.
3892 : */
3893 111812 : if (estfract < 1.0e-6)
3894 0 : estfract = 1.0e-6;
3895 111812 : else if (estfract > 1.0)
3896 24868 : estfract = 1.0;
3897 :
3898 111812 : *bucketsize_frac = (Selectivity) estfract;
3899 :
3900 111812 : ReleaseVariableStats(vardata);
3901 : }
3902 :
3903 : /*
3904 : * estimate_hashagg_tablesize
3905 : * estimate the number of bytes that a hash aggregate hashtable will
3906 : * require based on the agg_costs, path width and number of groups.
3907 : *
3908 : * We return the result as "double" to forestall any possible overflow
3909 : * problem in the multiplication by dNumGroups.
3910 : *
3911 : * XXX this may be over-estimating the size now that hashagg knows to omit
3912 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
3913 : * grouping columns not in the hashed set are counted here even though hashagg
3914 : * won't store them. Is this a problem?
3915 : */
3916 : double
3917 1978 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
3918 : const AggClauseCosts *agg_costs, double dNumGroups)
3919 : {
3920 : Size hashentrysize;
3921 :
3922 1978 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
3923 1978 : path->pathtarget->width,
3924 : agg_costs->transitionSpace);
3925 :
3926 : /*
3927 : * Note that this disregards the effect of fill-factor and growth policy
3928 : * of the hash table. That's probably ok, given that the default
3929 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
3930 : * "double-in-size" growth policies here.
3931 : */
3932 1978 : return hashentrysize * dNumGroups;
3933 : }
3934 :
3935 :
3936 : /*-------------------------------------------------------------------------
3937 : *
3938 : * Support routines
3939 : *
3940 : *-------------------------------------------------------------------------
3941 : */
3942 :
3943 : /*
3944 : * Find applicable ndistinct statistics for the given list of VarInfos (which
3945 : * must all belong to the given rel), and update *ndistinct to the estimate of
3946 : * the MVNDistinctItem that best matches. If a match it found, *varinfos is
3947 : * updated to remove the list of matched varinfos.
3948 : *
3949 : * Varinfos that aren't for simple Vars are ignored.
3950 : *
3951 : * Return true if we're able to find a match, false otherwise.
3952 : */
3953 : static bool
3954 252892 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
3955 : List **varinfos, double *ndistinct)
3956 : {
3957 : ListCell *lc;
3958 : int nmatches_vars;
3959 : int nmatches_exprs;
3960 252892 : Oid statOid = InvalidOid;
3961 : MVNDistinct *stats;
3962 252892 : StatisticExtInfo *matched_info = NULL;
3963 252892 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
3964 :
3965 : /* bail out immediately if the table has no extended statistics */
3966 252892 : if (!rel->statlist)
3967 252370 : return false;
3968 :
3969 : /* look for the ndistinct statistics object matching the most vars */
3970 522 : nmatches_vars = 0; /* we require at least two matches */
3971 522 : nmatches_exprs = 0;
3972 2112 : foreach(lc, rel->statlist)
3973 : {
3974 : ListCell *lc2;
3975 1590 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
3976 1590 : int nshared_vars = 0;
3977 1590 : int nshared_exprs = 0;
3978 :
3979 : /* skip statistics of other kinds */
3980 1590 : if (info->kind != STATS_EXT_NDISTINCT)
3981 750 : continue;
3982 :
3983 : /* skip statistics with mismatching stxdinherit value */
3984 840 : if (info->inherit != rte->inh)
3985 24 : continue;
3986 :
3987 : /*
3988 : * Determine how many expressions (and variables in non-matched
3989 : * expressions) match. We'll then use these numbers to pick the
3990 : * statistics object that best matches the clauses.
3991 : */
3992 2616 : foreach(lc2, *varinfos)
3993 : {
3994 : ListCell *lc3;
3995 1800 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
3996 : AttrNumber attnum;
3997 :
3998 : Assert(varinfo->rel == rel);
3999 :
4000 : /* simple Var, search in statistics keys directly */
4001 1800 : if (IsA(varinfo->var, Var))
4002 : {
4003 1434 : attnum = ((Var *) varinfo->var)->varattno;
4004 :
4005 : /*
4006 : * Ignore system attributes - we don't support statistics on
4007 : * them, so can't match them (and it'd fail as the values are
4008 : * negative).
4009 : */
4010 1434 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4011 12 : continue;
4012 :
4013 1422 : if (bms_is_member(attnum, info->keys))
4014 816 : nshared_vars++;
4015 :
4016 1422 : continue;
4017 : }
4018 :
4019 : /* expression - see if it's in the statistics object */
4020 660 : foreach(lc3, info->exprs)
4021 : {
4022 528 : Node *expr = (Node *) lfirst(lc3);
4023 :
4024 528 : if (equal(varinfo->var, expr))
4025 : {
4026 234 : nshared_exprs++;
4027 234 : break;
4028 : }
4029 : }
4030 : }
4031 :
4032 816 : if (nshared_vars + nshared_exprs < 2)
4033 378 : continue;
4034 :
4035 : /*
4036 : * Does this statistics object match more columns than the currently
4037 : * best object? If so, use this one instead.
4038 : *
4039 : * XXX This should break ties using name of the object, or something
4040 : * like that, to make the outcome stable.
4041 : */
4042 438 : if ((nshared_exprs > nmatches_exprs) ||
4043 330 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4044 : {
4045 414 : statOid = info->statOid;
4046 414 : nmatches_vars = nshared_vars;
4047 414 : nmatches_exprs = nshared_exprs;
4048 414 : matched_info = info;
4049 : }
4050 : }
4051 :
4052 : /* No match? */
4053 522 : if (statOid == InvalidOid)
4054 120 : return false;
4055 :
4056 : Assert(nmatches_vars + nmatches_exprs > 1);
4057 :
4058 402 : stats = statext_ndistinct_load(statOid, rte->inh);
4059 :
4060 : /*
4061 : * If we have a match, search it for the specific item that matches (there
4062 : * must be one), and construct the output values.
4063 : */
4064 402 : if (stats)
4065 : {
4066 : int i;
4067 402 : List *newlist = NIL;
4068 402 : MVNDistinctItem *item = NULL;
4069 : ListCell *lc2;
4070 402 : Bitmapset *matched = NULL;
4071 : AttrNumber attnum_offset;
4072 :
4073 : /*
4074 : * How much we need to offset the attnums? If there are no
4075 : * expressions, no offset is needed. Otherwise offset enough to move
4076 : * the lowest one (which is equal to number of expressions) to 1.
4077 : */
4078 402 : if (matched_info->exprs)
4079 144 : attnum_offset = (list_length(matched_info->exprs) + 1);
4080 : else
4081 258 : attnum_offset = 0;
4082 :
4083 : /* see what actually matched */
4084 1410 : foreach(lc2, *varinfos)
4085 : {
4086 : ListCell *lc3;
4087 : int idx;
4088 1008 : bool found = false;
4089 :
4090 1008 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4091 :
4092 : /*
4093 : * Process a simple Var expression, by matching it to keys
4094 : * directly. If there's a matching expression, we'll try matching
4095 : * it later.
4096 : */
4097 1008 : if (IsA(varinfo->var, Var))
4098 : {
4099 822 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4100 :
4101 : /*
4102 : * Ignore expressions on system attributes. Can't rely on the
4103 : * bms check for negative values.
4104 : */
4105 822 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4106 6 : continue;
4107 :
4108 : /* Is the variable covered by the statistics object? */
4109 816 : if (!bms_is_member(attnum, matched_info->keys))
4110 120 : continue;
4111 :
4112 696 : attnum = attnum + attnum_offset;
4113 :
4114 : /* ensure sufficient offset */
4115 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4116 :
4117 696 : matched = bms_add_member(matched, attnum);
4118 :
4119 696 : found = true;
4120 : }
4121 :
4122 : /*
4123 : * XXX Maybe we should allow searching the expressions even if we
4124 : * found an attribute matching the expression? That would handle
4125 : * trivial expressions like "(a)" but it seems fairly useless.
4126 : */
4127 882 : if (found)
4128 696 : continue;
4129 :
4130 : /* expression - see if it's in the statistics object */
4131 186 : idx = 0;
4132 306 : foreach(lc3, matched_info->exprs)
4133 : {
4134 276 : Node *expr = (Node *) lfirst(lc3);
4135 :
4136 276 : if (equal(varinfo->var, expr))
4137 : {
4138 156 : AttrNumber attnum = -(idx + 1);
4139 :
4140 156 : attnum = attnum + attnum_offset;
4141 :
4142 : /* ensure sufficient offset */
4143 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4144 :
4145 156 : matched = bms_add_member(matched, attnum);
4146 :
4147 : /* there should be just one matching expression */
4148 156 : break;
4149 : }
4150 :
4151 120 : idx++;
4152 : }
4153 : }
4154 :
4155 : /* Find the specific item that exactly matches the combination */
4156 822 : for (i = 0; i < stats->nitems; i++)
4157 : {
4158 : int j;
4159 822 : MVNDistinctItem *tmpitem = &stats->items[i];
4160 :
4161 822 : if (tmpitem->nattributes != bms_num_members(matched))
4162 144 : continue;
4163 :
4164 : /* assume it's the right item */
4165 678 : item = tmpitem;
4166 :
4167 : /* check that all item attributes/expressions fit the match */
4168 1614 : for (j = 0; j < tmpitem->nattributes; j++)
4169 : {
4170 1212 : AttrNumber attnum = tmpitem->attributes[j];
4171 :
4172 : /*
4173 : * Thanks to how we constructed the matched bitmap above, we
4174 : * can just offset all attnums the same way.
4175 : */
4176 1212 : attnum = attnum + attnum_offset;
4177 :
4178 1212 : if (!bms_is_member(attnum, matched))
4179 : {
4180 : /* nah, it's not this item */
4181 276 : item = NULL;
4182 276 : break;
4183 : }
4184 : }
4185 :
4186 : /*
4187 : * If the item has all the matched attributes, we know it's the
4188 : * right one - there can't be a better one. matching more.
4189 : */
4190 678 : if (item)
4191 402 : break;
4192 : }
4193 :
4194 : /*
4195 : * Make sure we found an item. There has to be one, because ndistinct
4196 : * statistics includes all combinations of attributes.
4197 : */
4198 402 : if (!item)
4199 0 : elog(ERROR, "corrupt MVNDistinct entry");
4200 :
4201 : /* Form the output varinfo list, keeping only unmatched ones */
4202 1410 : foreach(lc, *varinfos)
4203 : {
4204 1008 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4205 : ListCell *lc3;
4206 1008 : bool found = false;
4207 :
4208 : /*
4209 : * Let's look at plain variables first, because it's the most
4210 : * common case and the check is quite cheap. We can simply get the
4211 : * attnum and check (with an offset) matched bitmap.
4212 : */
4213 1008 : if (IsA(varinfo->var, Var))
4214 : {
4215 822 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4216 :
4217 : /*
4218 : * If it's a system attribute, we're done. We don't support
4219 : * extended statistics on system attributes, so it's clearly
4220 : * not matched. Just keep the expression and continue.
4221 : */
4222 822 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4223 : {
4224 6 : newlist = lappend(newlist, varinfo);
4225 6 : continue;
4226 : }
4227 :
4228 : /* apply the same offset as above */
4229 816 : attnum += attnum_offset;
4230 :
4231 : /* if it's not matched, keep the varinfo */
4232 816 : if (!bms_is_member(attnum, matched))
4233 120 : newlist = lappend(newlist, varinfo);
4234 :
4235 : /* The rest of the loop deals with complex expressions. */
4236 816 : continue;
4237 : }
4238 :
4239 : /*
4240 : * Process complex expressions, not just simple Vars.
4241 : *
4242 : * First, we search for an exact match of an expression. If we
4243 : * find one, we can just discard the whole GroupVarInfo, with all
4244 : * the variables we extracted from it.
4245 : *
4246 : * Otherwise we inspect the individual vars, and try matching it
4247 : * to variables in the item.
4248 : */
4249 306 : foreach(lc3, matched_info->exprs)
4250 : {
4251 276 : Node *expr = (Node *) lfirst(lc3);
4252 :
4253 276 : if (equal(varinfo->var, expr))
4254 : {
4255 156 : found = true;
4256 156 : break;
4257 : }
4258 : }
4259 :
4260 : /* found exact match, skip */
4261 186 : if (found)
4262 156 : continue;
4263 :
4264 30 : newlist = lappend(newlist, varinfo);
4265 : }
4266 :
4267 402 : *varinfos = newlist;
4268 402 : *ndistinct = item->ndistinct;
4269 402 : return true;
4270 : }
4271 :
4272 0 : return false;
4273 : }
4274 :
4275 : /*
4276 : * convert_to_scalar
4277 : * Convert non-NULL values of the indicated types to the comparison
4278 : * scale needed by scalarineqsel().
4279 : * Returns "true" if successful.
4280 : *
4281 : * XXX this routine is a hack: ideally we should look up the conversion
4282 : * subroutines in pg_type.
4283 : *
4284 : * All numeric datatypes are simply converted to their equivalent
4285 : * "double" values. (NUMERIC values that are outside the range of "double"
4286 : * are clamped to +/- HUGE_VAL.)
4287 : *
4288 : * String datatypes are converted by convert_string_to_scalar(),
4289 : * which is explained below. The reason why this routine deals with
4290 : * three values at a time, not just one, is that we need it for strings.
4291 : *
4292 : * The bytea datatype is just enough different from strings that it has
4293 : * to be treated separately.
4294 : *
4295 : * The several datatypes representing absolute times are all converted
4296 : * to Timestamp, which is actually an int64, and then we promote that to
4297 : * a double. Note this will give correct results even for the "special"
4298 : * values of Timestamp, since those are chosen to compare correctly;
4299 : * see timestamp_cmp.
4300 : *
4301 : * The several datatypes representing relative times (intervals) are all
4302 : * converted to measurements expressed in seconds.
4303 : */
4304 : static bool
4305 77052 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4306 : Datum lobound, Datum hibound, Oid boundstypid,
4307 : double *scaledlobound, double *scaledhibound)
4308 : {
4309 77052 : bool failure = false;
4310 :
4311 : /*
4312 : * Both the valuetypid and the boundstypid should exactly match the
4313 : * declared input type(s) of the operator we are invoked for. However,
4314 : * extensions might try to use scalarineqsel as estimator for operators
4315 : * with input type(s) we don't handle here; in such cases, we want to
4316 : * return false, not fail. In any case, we mustn't assume that valuetypid
4317 : * and boundstypid are identical.
4318 : *
4319 : * XXX The histogram we are interpolating between points of could belong
4320 : * to a column that's only binary-compatible with the declared type. In
4321 : * essence we are assuming that the semantics of binary-compatible types
4322 : * are enough alike that we can use a histogram generated with one type's
4323 : * operators to estimate selectivity for the other's. This is outright
4324 : * wrong in some cases --- in particular signed versus unsigned
4325 : * interpretation could trip us up. But it's useful enough in the
4326 : * majority of cases that we do it anyway. Should think about more
4327 : * rigorous ways to do it.
4328 : */
4329 77052 : switch (valuetypid)
4330 : {
4331 : /*
4332 : * Built-in numeric types
4333 : */
4334 72154 : case BOOLOID:
4335 : case INT2OID:
4336 : case INT4OID:
4337 : case INT8OID:
4338 : case FLOAT4OID:
4339 : case FLOAT8OID:
4340 : case NUMERICOID:
4341 : case OIDOID:
4342 : case REGPROCOID:
4343 : case REGPROCEDUREOID:
4344 : case REGOPEROID:
4345 : case REGOPERATOROID:
4346 : case REGCLASSOID:
4347 : case REGTYPEOID:
4348 : case REGCOLLATIONOID:
4349 : case REGCONFIGOID:
4350 : case REGDICTIONARYOID:
4351 : case REGROLEOID:
4352 : case REGNAMESPACEOID:
4353 72154 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4354 : &failure);
4355 72154 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4356 : &failure);
4357 72154 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4358 : &failure);
4359 72154 : return !failure;
4360 :
4361 : /*
4362 : * Built-in string types
4363 : */
4364 4898 : case CHAROID:
4365 : case BPCHAROID:
4366 : case VARCHAROID:
4367 : case TEXTOID:
4368 : case NAMEOID:
4369 : {
4370 4898 : char *valstr = convert_string_datum(value, valuetypid,
4371 : collid, &failure);
4372 4898 : char *lostr = convert_string_datum(lobound, boundstypid,
4373 : collid, &failure);
4374 4898 : char *histr = convert_string_datum(hibound, boundstypid,
4375 : collid, &failure);
4376 :
4377 : /*
4378 : * Bail out if any of the values is not of string type. We
4379 : * might leak converted strings for the other value(s), but
4380 : * that's not worth troubling over.
4381 : */
4382 4898 : if (failure)
4383 0 : return false;
4384 :
4385 4898 : convert_string_to_scalar(valstr, scaledvalue,
4386 : lostr, scaledlobound,
4387 : histr, scaledhibound);
4388 4898 : pfree(valstr);
4389 4898 : pfree(lostr);
4390 4898 : pfree(histr);
4391 4898 : return true;
4392 : }
4393 :
4394 : /*
4395 : * Built-in bytea type
4396 : */
4397 0 : case BYTEAOID:
4398 : {
4399 : /* We only support bytea vs bytea comparison */
4400 0 : if (boundstypid != BYTEAOID)
4401 0 : return false;
4402 0 : convert_bytea_to_scalar(value, scaledvalue,
4403 : lobound, scaledlobound,
4404 : hibound, scaledhibound);
4405 0 : return true;
4406 : }
4407 :
4408 : /*
4409 : * Built-in time types
4410 : */
4411 0 : case TIMESTAMPOID:
4412 : case TIMESTAMPTZOID:
4413 : case DATEOID:
4414 : case INTERVALOID:
4415 : case TIMEOID:
4416 : case TIMETZOID:
4417 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
4418 : &failure);
4419 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
4420 : &failure);
4421 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
4422 : &failure);
4423 0 : return !failure;
4424 :
4425 : /*
4426 : * Built-in network types
4427 : */
4428 0 : case INETOID:
4429 : case CIDROID:
4430 : case MACADDROID:
4431 : case MACADDR8OID:
4432 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
4433 : &failure);
4434 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
4435 : &failure);
4436 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
4437 : &failure);
4438 0 : return !failure;
4439 : }
4440 : /* Don't know how to convert */
4441 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
4442 0 : return false;
4443 : }
4444 :
4445 : /*
4446 : * Do convert_to_scalar()'s work for any numeric data type.
4447 : *
4448 : * On failure (e.g., unsupported typid), set *failure to true;
4449 : * otherwise, that variable is not changed.
4450 : */
4451 : static double
4452 216462 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
4453 : {
4454 216462 : switch (typid)
4455 : {
4456 0 : case BOOLOID:
4457 0 : return (double) DatumGetBool(value);
4458 12 : case INT2OID:
4459 12 : return (double) DatumGetInt16(value);
4460 28986 : case INT4OID:
4461 28986 : return (double) DatumGetInt32(value);
4462 0 : case INT8OID:
4463 0 : return (double) DatumGetInt64(value);
4464 0 : case FLOAT4OID:
4465 0 : return (double) DatumGetFloat4(value);
4466 36 : case FLOAT8OID:
4467 36 : return (double) DatumGetFloat8(value);
4468 0 : case NUMERICOID:
4469 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
4470 0 : return (double)
4471 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
4472 : value));
4473 187428 : case OIDOID:
4474 : case REGPROCOID:
4475 : case REGPROCEDUREOID:
4476 : case REGOPEROID:
4477 : case REGOPERATOROID:
4478 : case REGCLASSOID:
4479 : case REGTYPEOID:
4480 : case REGCOLLATIONOID:
4481 : case REGCONFIGOID:
4482 : case REGDICTIONARYOID:
4483 : case REGROLEOID:
4484 : case REGNAMESPACEOID:
4485 : /* we can treat OIDs as integers... */
4486 187428 : return (double) DatumGetObjectId(value);
4487 : }
4488 :
4489 0 : *failure = true;
4490 0 : return 0;
4491 : }
4492 :
4493 : /*
4494 : * Do convert_to_scalar()'s work for any character-string data type.
4495 : *
4496 : * String datatypes are converted to a scale that ranges from 0 to 1,
4497 : * where we visualize the bytes of the string as fractional digits.
4498 : *
4499 : * We do not want the base to be 256, however, since that tends to
4500 : * generate inflated selectivity estimates; few databases will have
4501 : * occurrences of all 256 possible byte values at each position.
4502 : * Instead, use the smallest and largest byte values seen in the bounds
4503 : * as the estimated range for each byte, after some fudging to deal with
4504 : * the fact that we probably aren't going to see the full range that way.
4505 : *
4506 : * An additional refinement is that we discard any common prefix of the
4507 : * three strings before computing the scaled values. This allows us to
4508 : * "zoom in" when we encounter a narrow data range. An example is a phone
4509 : * number database where all the values begin with the same area code.
4510 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
4511 : * so this is more likely to happen than you might think.)
4512 : */
4513 : static void
4514 4898 : convert_string_to_scalar(char *value,
4515 : double *scaledvalue,
4516 : char *lobound,
4517 : double *scaledlobound,
4518 : char *hibound,
4519 : double *scaledhibound)
4520 : {
4521 : int rangelo,
4522 : rangehi;
4523 : char *sptr;
4524 :
4525 4898 : rangelo = rangehi = (unsigned char) hibound[0];
4526 64350 : for (sptr = lobound; *sptr; sptr++)
4527 : {
4528 59452 : if (rangelo > (unsigned char) *sptr)
4529 12138 : rangelo = (unsigned char) *sptr;
4530 59452 : if (rangehi < (unsigned char) *sptr)
4531 6410 : rangehi = (unsigned char) *sptr;
4532 : }
4533 68246 : for (sptr = hibound; *sptr; sptr++)
4534 : {
4535 63348 : if (rangelo > (unsigned char) *sptr)
4536 900 : rangelo = (unsigned char) *sptr;
4537 63348 : if (rangehi < (unsigned char) *sptr)
4538 2444 : rangehi = (unsigned char) *sptr;
4539 : }
4540 : /* If range includes any upper-case ASCII chars, make it include all */
4541 4898 : if (rangelo <= 'Z' && rangehi >= 'A')
4542 : {
4543 838 : if (rangelo > 'A')
4544 90 : rangelo = 'A';
4545 838 : if (rangehi < 'Z')
4546 480 : rangehi = 'Z';
4547 : }
4548 : /* Ditto lower-case */
4549 4898 : if (rangelo <= 'z' && rangehi >= 'a')
4550 : {
4551 4396 : if (rangelo > 'a')
4552 18 : rangelo = 'a';
4553 4396 : if (rangehi < 'z')
4554 4346 : rangehi = 'z';
4555 : }
4556 : /* Ditto digits */
4557 4898 : if (rangelo <= '9' && rangehi >= '0')
4558 : {
4559 290 : if (rangelo > '0')
4560 196 : rangelo = '0';
4561 290 : if (rangehi < '9')
4562 14 : rangehi = '9';
4563 : }
4564 :
4565 : /*
4566 : * If range includes less than 10 chars, assume we have not got enough
4567 : * data, and make it include regular ASCII set.
4568 : */
4569 4898 : if (rangehi - rangelo < 9)
4570 : {
4571 0 : rangelo = ' ';
4572 0 : rangehi = 127;
4573 : }
4574 :
4575 : /*
4576 : * Now strip any common prefix of the three strings.
4577 : */
4578 8872 : while (*lobound)
4579 : {
4580 8872 : if (*lobound != *hibound || *lobound != *value)
4581 : break;
4582 3974 : lobound++, hibound++, value++;
4583 : }
4584 :
4585 : /*
4586 : * Now we can do the conversions.
4587 : */
4588 4898 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
4589 4898 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
4590 4898 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
4591 4898 : }
4592 :
4593 : static double
4594 14694 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
4595 : {
4596 14694 : int slen = strlen(value);
4597 : double num,
4598 : denom,
4599 : base;
4600 :
4601 14694 : if (slen <= 0)
4602 0 : return 0.0; /* empty string has scalar value 0 */
4603 :
4604 : /*
4605 : * There seems little point in considering more than a dozen bytes from
4606 : * the string. Since base is at least 10, that will give us nominal
4607 : * resolution of at least 12 decimal digits, which is surely far more
4608 : * precision than this estimation technique has got anyway (especially in
4609 : * non-C locales). Also, even with the maximum possible base of 256, this
4610 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
4611 : * overflow on any known machine.
4612 : */
4613 14694 : if (slen > 12)
4614 4370 : slen = 12;
4615 :
4616 : /* Convert initial characters to fraction */
4617 14694 : base = rangehi - rangelo + 1;
4618 14694 : num = 0.0;
4619 14694 : denom = base;
4620 125546 : while (slen-- > 0)
4621 : {
4622 110852 : int ch = (unsigned char) *value++;
4623 :
4624 110852 : if (ch < rangelo)
4625 232 : ch = rangelo - 1;
4626 110620 : else if (ch > rangehi)
4627 0 : ch = rangehi + 1;
4628 110852 : num += ((double) (ch - rangelo)) / denom;
4629 110852 : denom *= base;
4630 : }
4631 :
4632 14694 : return num;
4633 : }
4634 :
4635 : /*
4636 : * Convert a string-type Datum into a palloc'd, null-terminated string.
4637 : *
4638 : * On failure (e.g., unsupported typid), set *failure to true;
4639 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
4640 : *
4641 : * When using a non-C locale, we must pass the string through strxfrm()
4642 : * before continuing, so as to generate correct locale-specific results.
4643 : */
4644 : static char *
4645 14694 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
4646 : {
4647 : char *val;
4648 :
4649 14694 : switch (typid)
4650 : {
4651 0 : case CHAROID:
4652 0 : val = (char *) palloc(2);
4653 0 : val[0] = DatumGetChar(value);
4654 0 : val[1] = '\0';
4655 0 : break;
4656 4494 : case BPCHAROID:
4657 : case VARCHAROID:
4658 : case TEXTOID:
4659 4494 : val = TextDatumGetCString(value);
4660 4494 : break;
4661 10200 : case NAMEOID:
4662 : {
4663 10200 : NameData *nm = (NameData *) DatumGetPointer(value);
4664 :
4665 10200 : val = pstrdup(NameStr(*nm));
4666 10200 : break;
4667 : }
4668 0 : default:
4669 0 : *failure = true;
4670 0 : return NULL;
4671 : }
4672 :
4673 14694 : if (!lc_collate_is_c(collid))
4674 : {
4675 : char *xfrmstr;
4676 : size_t xfrmlen;
4677 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
4678 :
4679 : /*
4680 : * XXX: We could guess at a suitable output buffer size and only call
4681 : * strxfrm twice if our guess is too small.
4682 : *
4683 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
4684 : * bogus data or set an error. This is not really a problem unless it
4685 : * crashes since it will only give an estimation error and nothing
4686 : * fatal.
4687 : */
4688 72 : xfrmlen = strxfrm(NULL, val, 0);
4689 : #ifdef WIN32
4690 :
4691 : /*
4692 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
4693 : * of trying to allocate this much memory (and fail), just return the
4694 : * original string unmodified as if we were in the C locale.
4695 : */
4696 : if (xfrmlen == INT_MAX)
4697 : return val;
4698 : #endif
4699 72 : xfrmstr = (char *) palloc(xfrmlen + 1);
4700 72 : xfrmlen2 = strxfrm(xfrmstr, val, xfrmlen + 1);
4701 :
4702 : /*
4703 : * Some systems (e.g., glibc) can return a smaller value from the
4704 : * second call than the first; thus the Assert must be <= not ==.
4705 : */
4706 : Assert(xfrmlen2 <= xfrmlen);
4707 72 : pfree(val);
4708 72 : val = xfrmstr;
4709 : }
4710 :
4711 14694 : return val;
4712 : }
4713 :
4714 : /*
4715 : * Do convert_to_scalar()'s work for any bytea data type.
4716 : *
4717 : * Very similar to convert_string_to_scalar except we can't assume
4718 : * null-termination and therefore pass explicit lengths around.
4719 : *
4720 : * Also, assumptions about likely "normal" ranges of characters have been
4721 : * removed - a data range of 0..255 is always used, for now. (Perhaps
4722 : * someday we will add information about actual byte data range to
4723 : * pg_statistic.)
4724 : */
4725 : static void
4726 0 : convert_bytea_to_scalar(Datum value,
4727 : double *scaledvalue,
4728 : Datum lobound,
4729 : double *scaledlobound,
4730 : Datum hibound,
4731 : double *scaledhibound)
4732 : {
4733 0 : bytea *valuep = DatumGetByteaPP(value);
4734 0 : bytea *loboundp = DatumGetByteaPP(lobound);
4735 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
4736 : int rangelo,
4737 : rangehi,
4738 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
4739 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
4740 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
4741 : i,
4742 : minlen;
4743 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
4744 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
4745 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
4746 :
4747 : /*
4748 : * Assume bytea data is uniformly distributed across all byte values.
4749 : */
4750 0 : rangelo = 0;
4751 0 : rangehi = 255;
4752 :
4753 : /*
4754 : * Now strip any common prefix of the three strings.
4755 : */
4756 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
4757 0 : for (i = 0; i < minlen; i++)
4758 : {
4759 0 : if (*lostr != *histr || *lostr != *valstr)
4760 : break;
4761 0 : lostr++, histr++, valstr++;
4762 0 : loboundlen--, hiboundlen--, valuelen--;
4763 : }
4764 :
4765 : /*
4766 : * Now we can do the conversions.
4767 : */
4768 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
4769 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
4770 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
4771 0 : }
4772 :
4773 : static double
4774 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
4775 : int rangelo, int rangehi)
4776 : {
4777 : double num,
4778 : denom,
4779 : base;
4780 :
4781 0 : if (valuelen <= 0)
4782 0 : return 0.0; /* empty string has scalar value 0 */
4783 :
4784 : /*
4785 : * Since base is 256, need not consider more than about 10 chars (even
4786 : * this many seems like overkill)
4787 : */
4788 0 : if (valuelen > 10)
4789 0 : valuelen = 10;
4790 :
4791 : /* Convert initial characters to fraction */
4792 0 : base = rangehi - rangelo + 1;
4793 0 : num = 0.0;
4794 0 : denom = base;
4795 0 : while (valuelen-- > 0)
4796 : {
4797 0 : int ch = *value++;
4798 :
4799 0 : if (ch < rangelo)
4800 0 : ch = rangelo - 1;
4801 0 : else if (ch > rangehi)
4802 0 : ch = rangehi + 1;
4803 0 : num += ((double) (ch - rangelo)) / denom;
4804 0 : denom *= base;
4805 : }
4806 :
4807 0 : return num;
4808 : }
4809 :
4810 : /*
4811 : * Do convert_to_scalar()'s work for any timevalue data type.
4812 : *
4813 : * On failure (e.g., unsupported typid), set *failure to true;
4814 : * otherwise, that variable is not changed.
4815 : */
4816 : static double
4817 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
4818 : {
4819 0 : switch (typid)
4820 : {
4821 0 : case TIMESTAMPOID:
4822 0 : return DatumGetTimestamp(value);
4823 0 : case TIMESTAMPTZOID:
4824 0 : return DatumGetTimestampTz(value);
4825 0 : case DATEOID:
4826 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
4827 0 : case INTERVALOID:
4828 : {
4829 0 : Interval *interval = DatumGetIntervalP(value);
4830 :
4831 : /*
4832 : * Convert the month part of Interval to days using assumed
4833 : * average month length of 365.25/12.0 days. Not too
4834 : * accurate, but plenty good enough for our purposes.
4835 : *
4836 : * This also works for infinite intervals, which just have all
4837 : * fields set to INT_MIN/INT_MAX, and so will produce a result
4838 : * smaller/larger than any finite interval.
4839 : */
4840 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
4841 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
4842 : }
4843 0 : case TIMEOID:
4844 0 : return DatumGetTimeADT(value);
4845 0 : case TIMETZOID:
4846 : {
4847 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
4848 :
4849 : /* use GMT-equivalent time */
4850 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
4851 : }
4852 : }
4853 :
4854 0 : *failure = true;
4855 0 : return 0;
4856 : }
4857 :
4858 :
4859 : /*
4860 : * get_restriction_variable
4861 : * Examine the args of a restriction clause to see if it's of the
4862 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
4863 : * where "variable" could be either a Var or an expression in vars of a
4864 : * single relation. If so, extract information about the variable,
4865 : * and also indicate which side it was on and the other argument.
4866 : *
4867 : * Inputs:
4868 : * root: the planner info
4869 : * args: clause argument list
4870 : * varRelid: see specs for restriction selectivity functions
4871 : *
4872 : * Outputs: (these are valid only if true is returned)
4873 : * *vardata: gets information about variable (see examine_variable)
4874 : * *other: gets other clause argument, aggressively reduced to a constant
4875 : * *varonleft: set true if variable is on the left, false if on the right
4876 : *
4877 : * Returns true if a variable is identified, otherwise false.
4878 : *
4879 : * Note: if there are Vars on both sides of the clause, we must fail, because
4880 : * callers are expecting that the other side will act like a pseudoconstant.
4881 : */
4882 : bool
4883 618642 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
4884 : VariableStatData *vardata, Node **other,
4885 : bool *varonleft)
4886 : {
4887 : Node *left,
4888 : *right;
4889 : VariableStatData rdata;
4890 :
4891 : /* Fail if not a binary opclause (probably shouldn't happen) */
4892 618642 : if (list_length(args) != 2)
4893 0 : return false;
4894 :
4895 618642 : left = (Node *) linitial(args);
4896 618642 : right = (Node *) lsecond(args);
4897 :
4898 : /*
4899 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
4900 : * relations will be treated as pseudoconstants.
4901 : */
4902 618642 : examine_variable(root, left, varRelid, vardata);
4903 618642 : examine_variable(root, right, varRelid, &rdata);
4904 :
4905 : /*
4906 : * If one side is a variable and the other not, we win.
4907 : */
4908 618642 : if (vardata->rel && rdata.rel == NULL)
4909 : {
4910 551256 : *varonleft = true;
4911 551256 : *other = estimate_expression_value(root, rdata.var);
4912 : /* Assume we need no ReleaseVariableStats(rdata) here */
4913 551256 : return true;
4914 : }
4915 :
4916 67386 : if (vardata->rel == NULL && rdata.rel)
4917 : {
4918 62554 : *varonleft = false;
4919 62554 : *other = estimate_expression_value(root, vardata->var);
4920 : /* Assume we need no ReleaseVariableStats(*vardata) here */
4921 62554 : *vardata = rdata;
4922 62554 : return true;
4923 : }
4924 :
4925 : /* Oops, clause has wrong structure (probably var op var) */
4926 4832 : ReleaseVariableStats(*vardata);
4927 4832 : ReleaseVariableStats(rdata);
4928 :
4929 4832 : return false;
4930 : }
4931 :
4932 : /*
4933 : * get_join_variables
4934 : * Apply examine_variable() to each side of a join clause.
4935 : * Also, attempt to identify whether the join clause has the same
4936 : * or reversed sense compared to the SpecialJoinInfo.
4937 : *
4938 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
4939 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
4940 : * where we can't tell for sure, we default to assuming it's normal.
4941 : */
4942 : void
4943 190366 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
4944 : VariableStatData *vardata1, VariableStatData *vardata2,
4945 : bool *join_is_reversed)
4946 : {
4947 : Node *left,
4948 : *right;
4949 :
4950 190366 : if (list_length(args) != 2)
4951 0 : elog(ERROR, "join operator should take two arguments");
4952 :
4953 190366 : left = (Node *) linitial(args);
4954 190366 : right = (Node *) lsecond(args);
4955 :
4956 190366 : examine_variable(root, left, 0, vardata1);
4957 190366 : examine_variable(root, right, 0, vardata2);
4958 :
4959 380562 : if (vardata1->rel &&
4960 190196 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
4961 61122 : *join_is_reversed = true; /* var1 is on RHS */
4962 258318 : else if (vardata2->rel &&
4963 129074 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
4964 140 : *join_is_reversed = true; /* var2 is on LHS */
4965 : else
4966 129104 : *join_is_reversed = false;
4967 190366 : }
4968 :
4969 : /* statext_expressions_load copies the tuple, so just pfree it. */
4970 : static void
4971 1644 : ReleaseDummy(HeapTuple tuple)
4972 : {
4973 1644 : pfree(tuple);
4974 1644 : }
4975 :
4976 : /*
4977 : * examine_variable
4978 : * Try to look up statistical data about an expression.
4979 : * Fill in a VariableStatData struct to describe the expression.
4980 : *
4981 : * Inputs:
4982 : * root: the planner info
4983 : * node: the expression tree to examine
4984 : * varRelid: see specs for restriction selectivity functions
4985 : *
4986 : * Outputs: *vardata is filled as follows:
4987 : * var: the input expression (with any binary relabeling stripped, if
4988 : * it is or contains a variable; but otherwise the type is preserved)
4989 : * rel: RelOptInfo for relation containing variable; NULL if expression
4990 : * contains no Vars (NOTE this could point to a RelOptInfo of a
4991 : * subquery, not one in the current query).
4992 : * statsTuple: the pg_statistic entry for the variable, if one exists;
4993 : * otherwise NULL.
4994 : * freefunc: pointer to a function to release statsTuple with.
4995 : * vartype: exposed type of the expression; this should always match
4996 : * the declared input type of the operator we are estimating for.
4997 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
4998 : * commonly the same as the exposed type of the variable argument,
4999 : * but can be different in binary-compatible-type cases.
5000 : * isunique: true if we were able to match the var to a unique index or a
5001 : * single-column DISTINCT clause, implying its values are unique for
5002 : * this query. (Caution: this should be trusted for statistical
5003 : * purposes only, since we do not check indimmediate nor verify that
5004 : * the exact same definition of equality applies.)
5005 : * acl_ok: true if current user has permission to read the column(s)
5006 : * underlying the pg_statistic entry. This is consulted by
5007 : * statistic_proc_security_check().
5008 : *
5009 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5010 : */
5011 : void
5012 2399960 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5013 : VariableStatData *vardata)
5014 : {
5015 : Node *basenode;
5016 : Relids varnos;
5017 : RelOptInfo *onerel;
5018 :
5019 : /* Make sure we don't return dangling pointers in vardata */
5020 16799720 : MemSet(vardata, 0, sizeof(VariableStatData));
5021 :
5022 : /* Save the exposed type of the expression */
5023 2399960 : vardata->vartype = exprType(node);
5024 :
5025 : /* Look inside any binary-compatible relabeling */
5026 :
5027 2399960 : if (IsA(node, RelabelType))
5028 22694 : basenode = (Node *) ((RelabelType *) node)->arg;
5029 : else
5030 2377266 : basenode = node;
5031 :
5032 : /* Fast path for a simple Var */
5033 :
5034 2399960 : if (IsA(basenode, Var) &&
5035 558360 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5036 : {
5037 1693836 : Var *var = (Var *) basenode;
5038 :
5039 : /* Set up result fields other than the stats tuple */
5040 1693836 : vardata->var = basenode; /* return Var without relabeling */
5041 1693836 : vardata->rel = find_base_rel(root, var->varno);
5042 1693836 : vardata->atttype = var->vartype;
5043 1693836 : vardata->atttypmod = var->vartypmod;
5044 1693836 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5045 :
5046 : /* Try to locate some stats */
5047 1693836 : examine_simple_variable(root, var, vardata);
5048 :
5049 1693836 : return;
5050 : }
5051 :
5052 : /*
5053 : * Okay, it's a more complicated expression. Determine variable
5054 : * membership. Note that when varRelid isn't zero, only vars of that
5055 : * relation are considered "real" vars.
5056 : */
5057 706124 : varnos = pull_varnos(root, basenode);
5058 :
5059 706124 : onerel = NULL;
5060 :
5061 706124 : if (bms_is_empty(varnos))
5062 : {
5063 : /* No Vars at all ... must be pseudo-constant clause */
5064 : }
5065 : else
5066 : {
5067 : int relid;
5068 :
5069 334452 : if (bms_get_singleton_member(varnos, &relid))
5070 : {
5071 328964 : if (varRelid == 0 || varRelid == relid)
5072 : {
5073 51664 : onerel = find_base_rel(root, relid);
5074 51664 : vardata->rel = onerel;
5075 51664 : node = basenode; /* strip any relabeling */
5076 : }
5077 : /* else treat it as a constant */
5078 : }
5079 : else
5080 : {
5081 : /* varnos has multiple relids */
5082 5488 : if (varRelid == 0)
5083 : {
5084 : /* treat it as a variable of a join relation */
5085 4986 : vardata->rel = find_join_rel(root, varnos);
5086 4986 : node = basenode; /* strip any relabeling */
5087 : }
5088 502 : else if (bms_is_member(varRelid, varnos))
5089 : {
5090 : /* ignore the vars belonging to other relations */
5091 92 : vardata->rel = find_base_rel(root, varRelid);
5092 92 : node = basenode; /* strip any relabeling */
5093 : /* note: no point in expressional-index search here */
5094 : }
5095 : /* else treat it as a constant */
5096 : }
5097 : }
5098 :
5099 706124 : bms_free(varnos);
5100 :
5101 706124 : vardata->var = node;
5102 706124 : vardata->atttype = exprType(node);
5103 706124 : vardata->atttypmod = exprTypmod(node);
5104 :
5105 706124 : if (onerel)
5106 : {
5107 : /*
5108 : * We have an expression in vars of a single relation. Try to match
5109 : * it to expressional index columns, in hopes of finding some
5110 : * statistics.
5111 : *
5112 : * Note that we consider all index columns including INCLUDE columns,
5113 : * since there could be stats for such columns. But the test for
5114 : * uniqueness needs to be warier.
5115 : *
5116 : * XXX it's conceivable that there are multiple matches with different
5117 : * index opfamilies; if so, we need to pick one that matches the
5118 : * operator we are estimating for. FIXME later.
5119 : */
5120 : ListCell *ilist;
5121 : ListCell *slist;
5122 : Oid userid;
5123 :
5124 : /*
5125 : * Determine the user ID to use for privilege checks: either
5126 : * onerel->userid if it's set (e.g., in case we're accessing the table
5127 : * via a view), or the current user otherwise.
5128 : *
5129 : * If we drill down to child relations, we keep using the same userid:
5130 : * it's going to be the same anyway, due to how we set up the relation
5131 : * tree (q.v. build_simple_rel).
5132 : */
5133 51664 : userid = OidIsValid(onerel->userid) ? onerel->userid : GetUserId();
5134 :
5135 101154 : foreach(ilist, onerel->indexlist)
5136 : {
5137 52268 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5138 : ListCell *indexpr_item;
5139 : int pos;
5140 :
5141 52268 : indexpr_item = list_head(index->indexprs);
5142 52268 : if (indexpr_item == NULL)
5143 47522 : continue; /* no expressions here... */
5144 :
5145 6786 : for (pos = 0; pos < index->ncolumns; pos++)
5146 : {
5147 4818 : if (index->indexkeys[pos] == 0)
5148 : {
5149 : Node *indexkey;
5150 :
5151 4746 : if (indexpr_item == NULL)
5152 0 : elog(ERROR, "too few entries in indexprs list");
5153 4746 : indexkey = (Node *) lfirst(indexpr_item);
5154 4746 : if (indexkey && IsA(indexkey, RelabelType))
5155 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5156 4746 : if (equal(node, indexkey))
5157 : {
5158 : /*
5159 : * Found a match ... is it a unique index? Tests here
5160 : * should match has_unique_index().
5161 : */
5162 3486 : if (index->unique &&
5163 510 : index->nkeycolumns == 1 &&
5164 510 : pos == 0 &&
5165 510 : (index->indpred == NIL || index->predOK))
5166 510 : vardata->isunique = true;
5167 :
5168 : /*
5169 : * Has it got stats? We only consider stats for
5170 : * non-partial indexes, since partial indexes probably
5171 : * don't reflect whole-relation statistics; the above
5172 : * check for uniqueness is the only info we take from
5173 : * a partial index.
5174 : *
5175 : * An index stats hook, however, must make its own
5176 : * decisions about what to do with partial indexes.
5177 : */
5178 3486 : if (get_index_stats_hook &&
5179 0 : (*get_index_stats_hook) (root, index->indexoid,
5180 0 : pos + 1, vardata))
5181 : {
5182 : /*
5183 : * The hook took control of acquiring a stats
5184 : * tuple. If it did supply a tuple, it'd better
5185 : * have supplied a freefunc.
5186 : */
5187 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5188 0 : !vardata->freefunc)
5189 0 : elog(ERROR, "no function provided to release variable stats with");
5190 : }
5191 3486 : else if (index->indpred == NIL)
5192 : {
5193 3486 : vardata->statsTuple =
5194 6972 : SearchSysCache3(STATRELATTINH,
5195 : ObjectIdGetDatum(index->indexoid),
5196 3486 : Int16GetDatum(pos + 1),
5197 : BoolGetDatum(false));
5198 3486 : vardata->freefunc = ReleaseSysCache;
5199 :
5200 3486 : if (HeapTupleIsValid(vardata->statsTuple))
5201 : {
5202 : /* Get index's table for permission check */
5203 : RangeTblEntry *rte;
5204 :
5205 2778 : rte = planner_rt_fetch(index->rel->relid, root);
5206 : Assert(rte->rtekind == RTE_RELATION);
5207 :
5208 : /*
5209 : * For simplicity, we insist on the whole
5210 : * table being selectable, rather than trying
5211 : * to identify which column(s) the index
5212 : * depends on. Also require all rows to be
5213 : * selectable --- there must be no
5214 : * securityQuals from security barrier views
5215 : * or RLS policies.
5216 : */
5217 2778 : vardata->acl_ok =
5218 5556 : rte->securityQuals == NIL &&
5219 2778 : (pg_class_aclcheck(rte->relid, userid,
5220 : ACL_SELECT) == ACLCHECK_OK);
5221 :
5222 : /*
5223 : * If the user doesn't have permissions to
5224 : * access an inheritance child relation, check
5225 : * the permissions of the table actually
5226 : * mentioned in the query, since most likely
5227 : * the user does have that permission. Note
5228 : * that whole-table select privilege on the
5229 : * parent doesn't quite guarantee that the
5230 : * user could read all columns of the child.
5231 : * But in practice it's unlikely that any
5232 : * interesting security violation could result
5233 : * from allowing access to the expression
5234 : * index's stats, so we allow it anyway. See
5235 : * similar code in examine_simple_variable()
5236 : * for additional comments.
5237 : */
5238 2778 : if (!vardata->acl_ok &&
5239 18 : root->append_rel_array != NULL)
5240 : {
5241 : AppendRelInfo *appinfo;
5242 12 : Index varno = index->rel->relid;
5243 :
5244 12 : appinfo = root->append_rel_array[varno];
5245 36 : while (appinfo &&
5246 24 : planner_rt_fetch(appinfo->parent_relid,
5247 24 : root)->rtekind == RTE_RELATION)
5248 : {
5249 24 : varno = appinfo->parent_relid;
5250 24 : appinfo = root->append_rel_array[varno];
5251 : }
5252 12 : if (varno != index->rel->relid)
5253 : {
5254 : /* Repeat access check on this rel */
5255 12 : rte = planner_rt_fetch(varno, root);
5256 : Assert(rte->rtekind == RTE_RELATION);
5257 :
5258 12 : vardata->acl_ok =
5259 24 : rte->securityQuals == NIL &&
5260 12 : (pg_class_aclcheck(rte->relid,
5261 : userid,
5262 : ACL_SELECT) == ACLCHECK_OK);
5263 : }
5264 : }
5265 : }
5266 : else
5267 : {
5268 : /* suppress leakproofness checks later */
5269 708 : vardata->acl_ok = true;
5270 : }
5271 : }
5272 3486 : if (vardata->statsTuple)
5273 2778 : break;
5274 : }
5275 1968 : indexpr_item = lnext(index->indexprs, indexpr_item);
5276 : }
5277 : }
5278 4746 : if (vardata->statsTuple)
5279 2778 : break;
5280 : }
5281 :
5282 : /*
5283 : * Search extended statistics for one with a matching expression.
5284 : * There might be multiple ones, so just grab the first one. In the
5285 : * future, we might consider the statistics target (and pick the most
5286 : * accurate statistics) and maybe some other parameters.
5287 : */
5288 55684 : foreach(slist, onerel->statlist)
5289 : {
5290 4308 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5291 4308 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5292 : ListCell *expr_item;
5293 : int pos;
5294 :
5295 : /*
5296 : * Stop once we've found statistics for the expression (either
5297 : * from extended stats, or for an index in the preceding loop).
5298 : */
5299 4308 : if (vardata->statsTuple)
5300 288 : break;
5301 :
5302 : /* skip stats without per-expression stats */
5303 4020 : if (info->kind != STATS_EXT_EXPRESSIONS)
5304 2016 : continue;
5305 :
5306 : /* skip stats with mismatching stxdinherit value */
5307 2004 : if (info->inherit != rte->inh)
5308 6 : continue;
5309 :
5310 1998 : pos = 0;
5311 3300 : foreach(expr_item, info->exprs)
5312 : {
5313 2946 : Node *expr = (Node *) lfirst(expr_item);
5314 :
5315 : Assert(expr);
5316 :
5317 : /* strip RelabelType before comparing it */
5318 2946 : if (expr && IsA(expr, RelabelType))
5319 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5320 :
5321 : /* found a match, see if we can extract pg_statistic row */
5322 2946 : if (equal(node, expr))
5323 : {
5324 : /*
5325 : * XXX Not sure if we should cache the tuple somewhere.
5326 : * Now we just create a new copy every time.
5327 : */
5328 1644 : vardata->statsTuple =
5329 1644 : statext_expressions_load(info->statOid, rte->inh, pos);
5330 :
5331 1644 : vardata->freefunc = ReleaseDummy;
5332 :
5333 : /*
5334 : * For simplicity, we insist on the whole table being
5335 : * selectable, rather than trying to identify which
5336 : * column(s) the statistics object depends on. Also
5337 : * require all rows to be selectable --- there must be no
5338 : * securityQuals from security barrier views or RLS
5339 : * policies.
5340 : */
5341 1644 : vardata->acl_ok =
5342 3288 : rte->securityQuals == NIL &&
5343 1644 : (pg_class_aclcheck(rte->relid, userid,
5344 : ACL_SELECT) == ACLCHECK_OK);
5345 :
5346 : /*
5347 : * If the user doesn't have permissions to access an
5348 : * inheritance child relation, check the permissions of
5349 : * the table actually mentioned in the query, since most
5350 : * likely the user does have that permission. Note that
5351 : * whole-table select privilege on the parent doesn't
5352 : * quite guarantee that the user could read all columns of
5353 : * the child. But in practice it's unlikely that any
5354 : * interesting security violation could result from
5355 : * allowing access to the expression stats, so we allow it
5356 : * anyway. See similar code in examine_simple_variable()
5357 : * for additional comments.
5358 : */
5359 1644 : if (!vardata->acl_ok &&
5360 0 : root->append_rel_array != NULL)
5361 : {
5362 : AppendRelInfo *appinfo;
5363 0 : Index varno = onerel->relid;
5364 :
5365 0 : appinfo = root->append_rel_array[varno];
5366 0 : while (appinfo &&
5367 0 : planner_rt_fetch(appinfo->parent_relid,
5368 0 : root)->rtekind == RTE_RELATION)
5369 : {
5370 0 : varno = appinfo->parent_relid;
5371 0 : appinfo = root->append_rel_array[varno];
5372 : }
5373 0 : if (varno != onerel->relid)
5374 : {
5375 : /* Repeat access check on this rel */
5376 0 : rte = planner_rt_fetch(varno, root);
5377 : Assert(rte->rtekind == RTE_RELATION);
5378 :
5379 0 : vardata->acl_ok =
5380 0 : rte->securityQuals == NIL &&
5381 0 : (pg_class_aclcheck(rte->relid,
5382 : userid,
5383 : ACL_SELECT) == ACLCHECK_OK);
5384 : }
5385 : }
5386 :
5387 1644 : break;
5388 : }
5389 :
5390 1302 : pos++;
5391 : }
5392 : }
5393 : }
5394 : }
5395 :
5396 : /*
5397 : * examine_simple_variable
5398 : * Handle a simple Var for examine_variable
5399 : *
5400 : * This is split out as a subroutine so that we can recurse to deal with
5401 : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
5402 : *
5403 : * We already filled in all the fields of *vardata except for the stats tuple.
5404 : */
5405 : static void
5406 1697988 : examine_simple_variable(PlannerInfo *root, Var *var,
5407 : VariableStatData *vardata)
5408 : {
5409 1697988 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
5410 :
5411 : Assert(IsA(rte, RangeTblEntry));
5412 :
5413 1697988 : if (get_relation_stats_hook &&
5414 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
5415 : {
5416 : /*
5417 : * The hook took control of acquiring a stats tuple. If it did supply
5418 : * a tuple, it'd better have supplied a freefunc.
5419 : */
5420 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5421 0 : !vardata->freefunc)
5422 0 : elog(ERROR, "no function provided to release variable stats with");
5423 : }
5424 1697988 : else if (rte->rtekind == RTE_RELATION)
5425 : {
5426 : /*
5427 : * Plain table or parent of an inheritance appendrel, so look up the
5428 : * column in pg_statistic
5429 : */
5430 1624314 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
5431 : ObjectIdGetDatum(rte->relid),
5432 1624314 : Int16GetDatum(var->varattno),
5433 1624314 : BoolGetDatum(rte->inh));
5434 1624314 : vardata->freefunc = ReleaseSysCache;
5435 :
5436 1624314 : if (HeapTupleIsValid(vardata->statsTuple))
5437 : {
5438 1205288 : RelOptInfo *onerel = find_base_rel_noerr(root, var->varno);
5439 : Oid userid;
5440 :
5441 : /*
5442 : * Check if user has permission to read this column. We require
5443 : * all rows to be accessible, so there must be no securityQuals
5444 : * from security barrier views or RLS policies.
5445 : *
5446 : * Normally the Var will have an associated RelOptInfo from which
5447 : * we can find out which userid to do the check as; but it might
5448 : * not if it's a RETURNING Var for an INSERT target relation. In
5449 : * that case use the RTEPermissionInfo associated with the RTE.
5450 : */
5451 1205288 : if (onerel)
5452 1205246 : userid = onerel->userid;
5453 : else
5454 : {
5455 : RTEPermissionInfo *perminfo;
5456 :
5457 42 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
5458 42 : userid = perminfo->checkAsUser;
5459 : }
5460 1205288 : if (!OidIsValid(userid))
5461 1136126 : userid = GetUserId();
5462 :
5463 1205288 : vardata->acl_ok =
5464 2410824 : rte->securityQuals == NIL &&
5465 1205288 : ((pg_class_aclcheck(rte->relid, userid,
5466 248 : ACL_SELECT) == ACLCHECK_OK) ||
5467 248 : (pg_attribute_aclcheck(rte->relid, var->varattno, userid,
5468 : ACL_SELECT) == ACLCHECK_OK));
5469 :
5470 : /*
5471 : * If the user doesn't have permissions to access an inheritance
5472 : * child relation or specifically this attribute, check the
5473 : * permissions of the table/column actually mentioned in the
5474 : * query, since most likely the user does have that permission
5475 : * (else the query will fail at runtime), and if the user can read
5476 : * the column there then he can get the values of the child table
5477 : * too. To do that, we must find out which of the root parent's
5478 : * attributes the child relation's attribute corresponds to.
5479 : */
5480 1205288 : if (!vardata->acl_ok && var->varattno > 0 &&
5481 48 : root->append_rel_array != NULL)
5482 : {
5483 : AppendRelInfo *appinfo;
5484 12 : Index varno = var->varno;
5485 12 : int varattno = var->varattno;
5486 12 : bool found = false;
5487 :
5488 12 : appinfo = root->append_rel_array[varno];
5489 :
5490 : /*
5491 : * Partitions are mapped to their immediate parent, not the
5492 : * root parent, so must be ready to walk up multiple
5493 : * AppendRelInfos. But stop if we hit a parent that is not
5494 : * RTE_RELATION --- that's a flattened UNION ALL subquery, not
5495 : * an inheritance parent.
5496 : */
5497 36 : while (appinfo &&
5498 24 : planner_rt_fetch(appinfo->parent_relid,
5499 24 : root)->rtekind == RTE_RELATION)
5500 : {
5501 : int parent_varattno;
5502 :
5503 24 : found = false;
5504 24 : if (varattno <= 0 || varattno > appinfo->num_child_cols)
5505 : break; /* safety check */
5506 24 : parent_varattno = appinfo->parent_colnos[varattno - 1];
5507 24 : if (parent_varattno == 0)
5508 0 : break; /* Var is local to child */
5509 :
5510 24 : varno = appinfo->parent_relid;
5511 24 : varattno = parent_varattno;
5512 24 : found = true;
5513 :
5514 : /* If the parent is itself a child, continue up. */
5515 24 : appinfo = root->append_rel_array[varno];
5516 : }
5517 :
5518 : /*
5519 : * In rare cases, the Var may be local to the child table, in
5520 : * which case, we've got to live with having no access to this
5521 : * column's stats.
5522 : */
5523 12 : if (!found)
5524 0 : return;
5525 :
5526 : /* Repeat the access check on this parent rel & column */
5527 12 : rte = planner_rt_fetch(varno, root);
5528 : Assert(rte->rtekind == RTE_RELATION);
5529 :
5530 : /*
5531 : * Fine to use the same userid as it's the same in all
5532 : * relations of a given inheritance tree.
5533 : */
5534 12 : vardata->acl_ok =
5535 30 : rte->securityQuals == NIL &&
5536 12 : ((pg_class_aclcheck(rte->relid, userid,
5537 6 : ACL_SELECT) == ACLCHECK_OK) ||
5538 6 : (pg_attribute_aclcheck(rte->relid, varattno, userid,
5539 : ACL_SELECT) == ACLCHECK_OK));
5540 : }
5541 : }
5542 : else
5543 : {
5544 : /* suppress any possible leakproofness checks later */
5545 419026 : vardata->acl_ok = true;
5546 : }
5547 : }
5548 73674 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
5549 66466 : (rte->rtekind == RTE_CTE && !rte->self_reference))
5550 : {
5551 : /*
5552 : * Plain subquery (not one that was converted to an appendrel) or
5553 : * non-recursive CTE. In either case, we can try to find out what the
5554 : * Var refers to within the subquery. We skip this for appendrel and
5555 : * recursive-CTE cases because any column stats we did find would
5556 : * likely not be very relevant.
5557 : */
5558 : PlannerInfo *subroot;
5559 : Query *subquery;
5560 : List *subtlist;
5561 : TargetEntry *ste;
5562 :
5563 : /*
5564 : * Punt if it's a whole-row var rather than a plain column reference.
5565 : */
5566 12146 : if (var->varattno == InvalidAttrNumber)
5567 0 : return;
5568 :
5569 : /*
5570 : * Otherwise, find the subquery's planner subroot.
5571 : */
5572 12146 : if (rte->rtekind == RTE_SUBQUERY)
5573 : {
5574 : RelOptInfo *rel;
5575 :
5576 : /*
5577 : * Fetch RelOptInfo for subquery. Note that we don't change the
5578 : * rel returned in vardata, since caller expects it to be a rel of
5579 : * the caller's query level. Because we might already be
5580 : * recursing, we can't use that rel pointer either, but have to
5581 : * look up the Var's rel afresh.
5582 : */
5583 7208 : rel = find_base_rel(root, var->varno);
5584 :
5585 7208 : subroot = rel->subroot;
5586 : }
5587 : else
5588 : {
5589 : /* CTE case is more difficult */
5590 : PlannerInfo *cteroot;
5591 : Index levelsup;
5592 : int ndx;
5593 : int plan_id;
5594 : ListCell *lc;
5595 :
5596 : /*
5597 : * Find the referenced CTE, and locate the subroot previously made
5598 : * for it.
5599 : */
5600 4938 : levelsup = rte->ctelevelsup;
5601 4938 : cteroot = root;
5602 9590 : while (levelsup-- > 0)
5603 : {
5604 4652 : cteroot = cteroot->parent_root;
5605 4652 : if (!cteroot) /* shouldn't happen */
5606 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
5607 : }
5608 :
5609 : /*
5610 : * Note: cte_plan_ids can be shorter than cteList, if we are still
5611 : * working on planning the CTEs (ie, this is a side-reference from
5612 : * another CTE). So we mustn't use forboth here.
5613 : */
5614 4938 : ndx = 0;
5615 7192 : foreach(lc, cteroot->parse->cteList)
5616 : {
5617 7192 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
5618 :
5619 7192 : if (strcmp(cte->ctename, rte->ctename) == 0)
5620 4938 : break;
5621 2254 : ndx++;
5622 : }
5623 4938 : if (lc == NULL) /* shouldn't happen */
5624 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
5625 4938 : if (ndx >= list_length(cteroot->cte_plan_ids))
5626 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
5627 4938 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
5628 4938 : if (plan_id <= 0)
5629 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
5630 4938 : subroot = list_nth(root->glob->subroots, plan_id - 1);
5631 : }
5632 :
5633 : /* If the subquery hasn't been planned yet, we have to punt */
5634 12146 : if (subroot == NULL)
5635 0 : return;
5636 : Assert(IsA(subroot, PlannerInfo));
5637 :
5638 : /*
5639 : * We must use the subquery parsetree as mangled by the planner, not
5640 : * the raw version from the RTE, because we need a Var that will refer
5641 : * to the subroot's live RelOptInfos. For instance, if any subquery
5642 : * pullup happened during planning, Vars in the targetlist might have
5643 : * gotten replaced, and we need to see the replacement expressions.
5644 : */
5645 12146 : subquery = subroot->parse;
5646 : Assert(IsA(subquery, Query));
5647 :
5648 : /*
5649 : * Punt if subquery uses set operations or GROUP BY, as these will
5650 : * mash underlying columns' stats beyond recognition. (Set ops are
5651 : * particularly nasty; if we forged ahead, we would return stats
5652 : * relevant to only the leftmost subselect...) DISTINCT is also
5653 : * problematic, but we check that later because there is a possibility
5654 : * of learning something even with it.
5655 : */
5656 12146 : if (subquery->setOperations ||
5657 10870 : subquery->groupClause ||
5658 10214 : subquery->groupingSets)
5659 1932 : return;
5660 :
5661 : /* Get the subquery output expression referenced by the upper Var */
5662 10214 : if (subquery->returningList)
5663 206 : subtlist = subquery->returningList;
5664 : else
5665 10008 : subtlist = subquery->targetList;
5666 10214 : ste = get_tle_by_resno(subtlist, var->varattno);
5667 10214 : if (ste == NULL || ste->resjunk)
5668 0 : elog(ERROR, "subquery %s does not have attribute %d",
5669 : rte->eref->aliasname, var->varattno);
5670 10214 : var = (Var *) ste->expr;
5671 :
5672 : /*
5673 : * If subquery uses DISTINCT, we can't make use of any stats for the
5674 : * variable ... but, if it's the only DISTINCT column, we are entitled
5675 : * to consider it unique. We do the test this way so that it works
5676 : * for cases involving DISTINCT ON.
5677 : */
5678 10214 : if (subquery->distinctClause)
5679 : {
5680 1794 : if (list_length(subquery->distinctClause) == 1 &&
5681 606 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
5682 280 : vardata->isunique = true;
5683 : /* cannot go further */
5684 1188 : return;
5685 : }
5686 :
5687 : /*
5688 : * If the sub-query originated from a view with the security_barrier
5689 : * attribute, we must not look at the variable's statistics, though it
5690 : * seems all right to notice the existence of a DISTINCT clause. So
5691 : * stop here.
5692 : *
5693 : * This is probably a harsher restriction than necessary; it's
5694 : * certainly OK for the selectivity estimator (which is a C function,
5695 : * and therefore omnipotent anyway) to look at the statistics. But
5696 : * many selectivity estimators will happily *invoke the operator
5697 : * function* to try to work out a good estimate - and that's not OK.
5698 : * So for now, don't dig down for stats.
5699 : */
5700 9026 : if (rte->security_barrier)
5701 228 : return;
5702 :
5703 : /* Can only handle a simple Var of subquery's query level */
5704 8798 : if (var && IsA(var, Var) &&
5705 4152 : var->varlevelsup == 0)
5706 : {
5707 : /*
5708 : * OK, recurse into the subquery. Note that the original setting
5709 : * of vardata->isunique (which will surely be false) is left
5710 : * unchanged in this situation. That's what we want, since even
5711 : * if the underlying column is unique, the subquery may have
5712 : * joined to other tables in a way that creates duplicates.
5713 : */
5714 4152 : examine_simple_variable(subroot, var, vardata);
5715 : }
5716 : }
5717 : else
5718 : {
5719 : /*
5720 : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
5721 : * see RTE_JOIN here because join alias Vars have already been
5722 : * flattened.) There's not much we can do with function outputs, but
5723 : * maybe someday try to be smarter about VALUES.
5724 : */
5725 : }
5726 : }
5727 :
5728 : /*
5729 : * Check whether it is permitted to call func_oid passing some of the
5730 : * pg_statistic data in vardata. We allow this either if the user has SELECT
5731 : * privileges on the table or column underlying the pg_statistic data or if
5732 : * the function is marked leak-proof.
5733 : */
5734 : bool
5735 829618 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
5736 : {
5737 829618 : if (vardata->acl_ok)
5738 828692 : return true;
5739 :
5740 926 : if (!OidIsValid(func_oid))
5741 0 : return false;
5742 :
5743 926 : if (get_func_leakproof(func_oid))
5744 424 : return true;
5745 :
5746 502 : ereport(DEBUG2,
5747 : (errmsg_internal("not using statistics because function \"%s\" is not leak-proof",
5748 : get_func_name(func_oid))));
5749 502 : return false;
5750 : }
5751 :
5752 : /*
5753 : * get_variable_numdistinct
5754 : * Estimate the number of distinct values of a variable.
5755 : *
5756 : * vardata: results of examine_variable
5757 : * *isdefault: set to true if the result is a default rather than based on
5758 : * anything meaningful.
5759 : *
5760 : * NB: be careful to produce a positive integral result, since callers may
5761 : * compare the result to exact integer counts, or might divide by it.
5762 : */
5763 : double
5764 1151226 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
5765 : {
5766 : double stadistinct;
5767 1151226 : double stanullfrac = 0.0;
5768 : double ntuples;
5769 :
5770 1151226 : *isdefault = false;
5771 :
5772 : /*
5773 : * Determine the stadistinct value to use. There are cases where we can
5774 : * get an estimate even without a pg_statistic entry, or can get a better
5775 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
5776 : * (otherwise, assume no nulls, for lack of any better idea).
5777 : */
5778 1151226 : if (HeapTupleIsValid(vardata->statsTuple))
5779 : {
5780 : /* Use the pg_statistic entry */
5781 : Form_pg_statistic stats;
5782 :
5783 820012 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
5784 820012 : stadistinct = stats->stadistinct;
5785 820012 : stanullfrac = stats->stanullfrac;
5786 : }
5787 331214 : else if (vardata->vartype == BOOLOID)
5788 : {
5789 : /*
5790 : * Special-case boolean columns: presumably, two distinct values.
5791 : *
5792 : * Are there any other datatypes we should wire in special estimates
5793 : * for?
5794 : */
5795 324 : stadistinct = 2.0;
5796 : }
5797 330890 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
5798 : {
5799 : /*
5800 : * If the Var represents a column of a VALUES RTE, assume it's unique.
5801 : * This could of course be very wrong, but it should tend to be true
5802 : * in well-written queries. We could consider examining the VALUES'
5803 : * contents to get some real statistics; but that only works if the
5804 : * entries are all constants, and it would be pretty expensive anyway.
5805 : */
5806 2806 : stadistinct = -1.0; /* unique (and all non null) */
5807 : }
5808 : else
5809 : {
5810 : /*
5811 : * We don't keep statistics for system columns, but in some cases we
5812 : * can infer distinctness anyway.
5813 : */
5814 328084 : if (vardata->var && IsA(vardata->var, Var))
5815 : {
5816 307206 : switch (((Var *) vardata->var)->varattno)
5817 : {
5818 990 : case SelfItemPointerAttributeNumber:
5819 990 : stadistinct = -1.0; /* unique (and all non null) */
5820 990 : break;
5821 2518 : case TableOidAttributeNumber:
5822 2518 : stadistinct = 1.0; /* only 1 value */
5823 2518 : break;
5824 303698 : default:
5825 303698 : stadistinct = 0.0; /* means "unknown" */
5826 303698 : break;
5827 : }
5828 : }
5829 : else
5830 20878 : stadistinct = 0.0; /* means "unknown" */
5831 :
5832 : /*
5833 : * XXX consider using estimate_num_groups on expressions?
5834 : */
5835 : }
5836 :
5837 : /*
5838 : * If there is a unique index or DISTINCT clause for the variable, assume
5839 : * it is unique no matter what pg_statistic says; the statistics could be
5840 : * out of date, or we might have found a partial unique index that proves
5841 : * the var is unique for this query. However, we'd better still believe
5842 : * the null-fraction statistic.
5843 : */
5844 1151226 : if (vardata->isunique)
5845 320084 : stadistinct = -1.0 * (1.0 - stanullfrac);
5846 :
5847 : /*
5848 : * If we had an absolute estimate, use that.
5849 : */
5850 1151226 : if (stadistinct > 0.0)
5851 264406 : return clamp_row_est(stadistinct);
5852 :
5853 : /*
5854 : * Otherwise we need to get the relation size; punt if not available.
5855 : */
5856 886820 : if (vardata->rel == NULL)
5857 : {
5858 400 : *isdefault = true;
5859 400 : return DEFAULT_NUM_DISTINCT;
5860 : }
5861 886420 : ntuples = vardata->rel->tuples;
5862 886420 : if (ntuples <= 0.0)
5863 : {
5864 43276 : *isdefault = true;
5865 43276 : return DEFAULT_NUM_DISTINCT;
5866 : }
5867 :
5868 : /*
5869 : * If we had a relative estimate, use that.
5870 : */
5871 843144 : if (stadistinct < 0.0)
5872 594932 : return clamp_row_est(-stadistinct * ntuples);
5873 :
5874 : /*
5875 : * With no data, estimate ndistinct = ntuples if the table is small, else
5876 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
5877 : * that the behavior isn't discontinuous.
5878 : */
5879 248212 : if (ntuples < DEFAULT_NUM_DISTINCT)
5880 122904 : return clamp_row_est(ntuples);
5881 :
5882 125308 : *isdefault = true;
5883 125308 : return DEFAULT_NUM_DISTINCT;
5884 : }
5885 :
5886 : /*
5887 : * get_variable_range
5888 : * Estimate the minimum and maximum value of the specified variable.
5889 : * If successful, store values in *min and *max, and return true.
5890 : * If no data available, return false.
5891 : *
5892 : * sortop is the "<" comparison operator to use. This should generally
5893 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
5894 : * The collation must be specified too.
5895 : */
5896 : static bool
5897 172048 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
5898 : Oid sortop, Oid collation,
5899 : Datum *min, Datum *max)
5900 : {
5901 172048 : Datum tmin = 0;
5902 172048 : Datum tmax = 0;
5903 172048 : bool have_data = false;
5904 : int16 typLen;
5905 : bool typByVal;
5906 : Oid opfuncoid;
5907 : FmgrInfo opproc;
5908 : AttStatsSlot sslot;
5909 :
5910 : /*
5911 : * XXX It's very tempting to try to use the actual column min and max, if
5912 : * we can get them relatively-cheaply with an index probe. However, since
5913 : * this function is called many times during join planning, that could
5914 : * have unpleasant effects on planning speed. Need more investigation
5915 : * before enabling this.
5916 : */
5917 : #ifdef NOT_USED
5918 : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
5919 : return true;
5920 : #endif
5921 :
5922 172048 : if (!HeapTupleIsValid(vardata->statsTuple))
5923 : {
5924 : /* no stats available, so default result */
5925 36720 : return false;
5926 : }
5927 :
5928 : /*
5929 : * If we can't apply the sortop to the stats data, just fail. In
5930 : * principle, if there's a histogram and no MCVs, we could return the
5931 : * histogram endpoints without ever applying the sortop ... but it's
5932 : * probably not worth trying, because whatever the caller wants to do with
5933 : * the endpoints would likely fail the security check too.
5934 : */
5935 135328 : if (!statistic_proc_security_check(vardata,
5936 135328 : (opfuncoid = get_opcode(sortop))))
5937 0 : return false;
5938 :
5939 135328 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
5940 :
5941 135328 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
5942 :
5943 : /*
5944 : * If there is a histogram with the ordering we want, grab the first and
5945 : * last values.
5946 : */
5947 135328 : if (get_attstatsslot(&sslot, vardata->statsTuple,
5948 : STATISTIC_KIND_HISTOGRAM, sortop,
5949 : ATTSTATSSLOT_VALUES))
5950 : {
5951 107464 : if (sslot.stacoll == collation && sslot.nvalues > 0)
5952 : {
5953 107464 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
5954 107464 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
5955 107464 : have_data = true;
5956 : }
5957 107464 : free_attstatsslot(&sslot);
5958 : }
5959 :
5960 : /*
5961 : * Otherwise, if there is a histogram with some other ordering, scan it
5962 : * and get the min and max values according to the ordering we want. This
5963 : * of course may not find values that are really extremal according to our
5964 : * ordering, but it beats ignoring available data.
5965 : */
5966 163192 : if (!have_data &&
5967 27864 : get_attstatsslot(&sslot, vardata->statsTuple,
5968 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
5969 : ATTSTATSSLOT_VALUES))
5970 : {
5971 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
5972 : collation, typLen, typByVal,
5973 : &tmin, &tmax, &have_data);
5974 0 : free_attstatsslot(&sslot);
5975 : }
5976 :
5977 : /*
5978 : * If we have most-common-values info, look for extreme MCVs. This is
5979 : * needed even if we also have a histogram, since the histogram excludes
5980 : * the MCVs. However, if we *only* have MCVs and no histogram, we should
5981 : * be pretty wary of deciding that that is a full representation of the
5982 : * data. Proceed only if the MCVs represent the whole table (to within
5983 : * roundoff error).
5984 : */
5985 135328 : if (get_attstatsslot(&sslot, vardata->statsTuple,
5986 : STATISTIC_KIND_MCV, InvalidOid,
5987 135328 : have_data ? ATTSTATSSLOT_VALUES :
5988 : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
5989 : {
5990 61074 : bool use_mcvs = have_data;
5991 :
5992 61074 : if (!have_data)
5993 : {
5994 26752 : double sumcommon = 0.0;
5995 : double nullfrac;
5996 : int i;
5997 :
5998 130398 : for (i = 0; i < sslot.nnumbers; i++)
5999 103646 : sumcommon += sslot.numbers[i];
6000 26752 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6001 26752 : if (sumcommon + nullfrac > 0.99999)
6002 26562 : use_mcvs = true;
6003 : }
6004 :
6005 61074 : if (use_mcvs)
6006 60884 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6007 : collation, typLen, typByVal,
6008 : &tmin, &tmax, &have_data);
6009 61074 : free_attstatsslot(&sslot);
6010 : }
6011 :
6012 135328 : *min = tmin;
6013 135328 : *max = tmax;
6014 135328 : return have_data;
6015 : }
6016 :
6017 : /*
6018 : * get_stats_slot_range: scan sslot for min/max values
6019 : *
6020 : * Subroutine for get_variable_range: update min/max/have_data according
6021 : * to what we find in the statistics array.
6022 : */
6023 : static void
6024 60884 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6025 : Oid collation, int16 typLen, bool typByVal,
6026 : Datum *min, Datum *max, bool *p_have_data)
6027 : {
6028 60884 : Datum tmin = *min;
6029 60884 : Datum tmax = *max;
6030 60884 : bool have_data = *p_have_data;
6031 60884 : bool found_tmin = false;
6032 60884 : bool found_tmax = false;
6033 :
6034 : /* Look up the comparison function, if we didn't already do so */
6035 60884 : if (opproc->fn_oid != opfuncoid)
6036 60884 : fmgr_info(opfuncoid, opproc);
6037 :
6038 : /* Scan all the slot's values */
6039 1688296 : for (int i = 0; i < sslot->nvalues; i++)
6040 : {
6041 1627412 : if (!have_data)
6042 : {
6043 26562 : tmin = tmax = sslot->values[i];
6044 26562 : found_tmin = found_tmax = true;
6045 26562 : *p_have_data = have_data = true;
6046 26562 : continue;
6047 : }
6048 1600850 : if (DatumGetBool(FunctionCall2Coll(opproc,
6049 : collation,
6050 1600850 : sslot->values[i], tmin)))
6051 : {
6052 51266 : tmin = sslot->values[i];
6053 51266 : found_tmin = true;
6054 : }
6055 1600850 : if (DatumGetBool(FunctionCall2Coll(opproc,
6056 : collation,
6057 1600850 : tmax, sslot->values[i])))
6058 : {
6059 70154 : tmax = sslot->values[i];
6060 70154 : found_tmax = true;
6061 : }
6062 : }
6063 :
6064 : /*
6065 : * Copy the slot's values, if we found new extreme values.
6066 : */
6067 60884 : if (found_tmin)
6068 51760 : *min = datumCopy(tmin, typByVal, typLen);
6069 60884 : if (found_tmax)
6070 29714 : *max = datumCopy(tmax, typByVal, typLen);
6071 60884 : }
6072 :
6073 :
6074 : /*
6075 : * get_actual_variable_range
6076 : * Attempt to identify the current *actual* minimum and/or maximum
6077 : * of the specified variable, by looking for a suitable btree index
6078 : * and fetching its low and/or high values.
6079 : * If successful, store values in *min and *max, and return true.
6080 : * (Either pointer can be NULL if that endpoint isn't needed.)
6081 : * If unsuccessful, return false.
6082 : *
6083 : * sortop is the "<" comparison operator to use.
6084 : * collation is the required collation.
6085 : */
6086 : static bool
6087 163700 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6088 : Oid sortop, Oid collation,
6089 : Datum *min, Datum *max)
6090 : {
6091 163700 : bool have_data = false;
6092 163700 : RelOptInfo *rel = vardata->rel;
6093 : RangeTblEntry *rte;
6094 : ListCell *lc;
6095 :
6096 : /* No hope if no relation or it doesn't have indexes */
6097 163700 : if (rel == NULL || rel->indexlist == NIL)
6098 12748 : return false;
6099 : /* If it has indexes it must be a plain relation */
6100 150952 : rte = root->simple_rte_array[rel->relid];
6101 : Assert(rte->rtekind == RTE_RELATION);
6102 :
6103 : /* ignore partitioned tables. Any indexes here are not real indexes */
6104 150952 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6105 828 : return false;
6106 :
6107 : /* Search through the indexes to see if any match our problem */
6108 321040 : foreach(lc, rel->indexlist)
6109 : {
6110 273258 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6111 : ScanDirection indexscandir;
6112 :
6113 : /* Ignore non-btree indexes */
6114 273258 : if (index->relam != BTREE_AM_OID)
6115 0 : continue;
6116 :
6117 : /*
6118 : * Ignore partial indexes --- we only want stats that cover the entire
6119 : * relation.
6120 : */
6121 273258 : if (index->indpred != NIL)
6122 288 : continue;
6123 :
6124 : /*
6125 : * The index list might include hypothetical indexes inserted by a
6126 : * get_relation_info hook --- don't try to access them.
6127 : */
6128 272970 : if (index->hypothetical)
6129 0 : continue;
6130 :
6131 : /*
6132 : * The first index column must match the desired variable, sortop, and
6133 : * collation --- but we can use a descending-order index.
6134 : */
6135 272970 : if (collation != index->indexcollations[0])
6136 42006 : continue; /* test first 'cause it's cheapest */
6137 230964 : if (!match_index_to_operand(vardata->var, 0, index))
6138 128622 : continue;
6139 102342 : switch (get_op_opfamily_strategy(sortop, index->sortopfamily[0]))
6140 : {
6141 102342 : case BTLessStrategyNumber:
6142 102342 : if (index->reverse_sort[0])
6143 0 : indexscandir = BackwardScanDirection;
6144 : else
6145 102342 : indexscandir = ForwardScanDirection;
6146 102342 : break;
6147 0 : case BTGreaterStrategyNumber:
6148 0 : if (index->reverse_sort[0])
6149 0 : indexscandir = ForwardScanDirection;
6150 : else
6151 0 : indexscandir = BackwardScanDirection;
6152 0 : break;
6153 0 : default:
6154 : /* index doesn't match the sortop */
6155 0 : continue;
6156 : }
6157 :
6158 : /*
6159 : * Found a suitable index to extract data from. Set up some data that
6160 : * can be used by both invocations of get_actual_variable_endpoint.
6161 : */
6162 : {
6163 : MemoryContext tmpcontext;
6164 : MemoryContext oldcontext;
6165 : Relation heapRel;
6166 : Relation indexRel;
6167 : TupleTableSlot *slot;
6168 : int16 typLen;
6169 : bool typByVal;
6170 : ScanKeyData scankeys[1];
6171 :
6172 : /* Make sure any cruft gets recycled when we're done */
6173 102342 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
6174 : "get_actual_variable_range workspace",
6175 : ALLOCSET_DEFAULT_SIZES);
6176 102342 : oldcontext = MemoryContextSwitchTo(tmpcontext);
6177 :
6178 : /*
6179 : * Open the table and index so we can read from them. We should
6180 : * already have some type of lock on each.
6181 : */
6182 102342 : heapRel = table_open(rte->relid, NoLock);
6183 102342 : indexRel = index_open(index->indexoid, NoLock);
6184 :
6185 : /* build some stuff needed for indexscan execution */
6186 102342 : slot = table_slot_create(heapRel, NULL);
6187 102342 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6188 :
6189 : /* set up an IS NOT NULL scan key so that we ignore nulls */
6190 102342 : ScanKeyEntryInitialize(&scankeys[0],
6191 : SK_ISNULL | SK_SEARCHNOTNULL,
6192 : 1, /* index col to scan */
6193 : InvalidStrategy, /* no strategy */
6194 : InvalidOid, /* no strategy subtype */
6195 : InvalidOid, /* no collation */
6196 : InvalidOid, /* no reg proc for this */
6197 : (Datum) 0); /* constant */
6198 :
6199 : /* If min is requested ... */
6200 102342 : if (min)
6201 : {
6202 60790 : have_data = get_actual_variable_endpoint(heapRel,
6203 : indexRel,
6204 : indexscandir,
6205 : scankeys,
6206 : typLen,
6207 : typByVal,
6208 : slot,
6209 : oldcontext,
6210 : min);
6211 : }
6212 : else
6213 : {
6214 : /* If min not requested, still want to fetch max */
6215 41552 : have_data = true;
6216 : }
6217 :
6218 : /* If max is requested, and we didn't already fail ... */
6219 102342 : if (max && have_data)
6220 : {
6221 : /* scan in the opposite direction; all else is the same */
6222 42728 : have_data = get_actual_variable_endpoint(heapRel,
6223 : indexRel,
6224 42728 : -indexscandir,
6225 : scankeys,
6226 : typLen,
6227 : typByVal,
6228 : slot,
6229 : oldcontext,
6230 : max);
6231 : }
6232 :
6233 : /* Clean everything up */
6234 102342 : ExecDropSingleTupleTableSlot(slot);
6235 :
6236 102342 : index_close(indexRel, NoLock);
6237 102342 : table_close(heapRel, NoLock);
6238 :
6239 102342 : MemoryContextSwitchTo(oldcontext);
6240 102342 : MemoryContextDelete(tmpcontext);
6241 :
6242 : /* And we're done */
6243 102342 : break;
6244 : }
6245 : }
6246 :
6247 150124 : return have_data;
6248 : }
6249 :
6250 : /*
6251 : * Get one endpoint datum (min or max depending on indexscandir) from the
6252 : * specified index. Return true if successful, false if not.
6253 : * On success, endpoint value is stored to *endpointDatum (and copied into
6254 : * outercontext).
6255 : *
6256 : * scankeys is a 1-element scankey array set up to reject nulls.
6257 : * typLen/typByVal describe the datatype of the index's first column.
6258 : * tableslot is a slot suitable to hold table tuples, in case we need
6259 : * to probe the heap.
6260 : * (We could compute these values locally, but that would mean computing them
6261 : * twice when get_actual_variable_range needs both the min and the max.)
6262 : *
6263 : * Failure occurs either when the index is empty, or we decide that it's
6264 : * taking too long to find a suitable tuple.
6265 : */
6266 : static bool
6267 103518 : get_actual_variable_endpoint(Relation heapRel,
6268 : Relation indexRel,
6269 : ScanDirection indexscandir,
6270 : ScanKey scankeys,
6271 : int16 typLen,
6272 : bool typByVal,
6273 : TupleTableSlot *tableslot,
6274 : MemoryContext outercontext,
6275 : Datum *endpointDatum)
6276 : {
6277 103518 : bool have_data = false;
6278 : SnapshotData SnapshotNonVacuumable;
6279 : IndexScanDesc index_scan;
6280 103518 : Buffer vmbuffer = InvalidBuffer;
6281 103518 : BlockNumber last_heap_block = InvalidBlockNumber;
6282 103518 : int n_visited_heap_pages = 0;
6283 : ItemPointer tid;
6284 : Datum values[INDEX_MAX_KEYS];
6285 : bool isnull[INDEX_MAX_KEYS];
6286 : MemoryContext oldcontext;
6287 :
6288 : /*
6289 : * We use the index-only-scan machinery for this. With mostly-static
6290 : * tables that's a win because it avoids a heap visit. It's also a win
6291 : * for dynamic data, but the reason is less obvious; read on for details.
6292 : *
6293 : * In principle, we should scan the index with our current active
6294 : * snapshot, which is the best approximation we've got to what the query
6295 : * will see when executed. But that won't be exact if a new snap is taken
6296 : * before running the query, and it can be very expensive if a lot of
6297 : * recently-dead or uncommitted rows exist at the beginning or end of the
6298 : * index (because we'll laboriously fetch each one and reject it).
6299 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
6300 : * and uncommitted rows as well as normal visible rows. On the other
6301 : * hand, it will reject known-dead rows, and thus not give a bogus answer
6302 : * when the extreme value has been deleted (unless the deletion was quite
6303 : * recent); that case motivates not using SnapshotAny here.
6304 : *
6305 : * A crucial point here is that SnapshotNonVacuumable, with
6306 : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
6307 : * condition that the indexscan will use to decide that index entries are
6308 : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
6309 : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
6310 : * have to continue scanning past it, we know that the indexscan will mark
6311 : * that index entry killed. That means that the next
6312 : * get_actual_variable_endpoint() call will not have to re-consider that
6313 : * index entry. In this way we avoid repetitive work when this function
6314 : * is used a lot during planning.
6315 : *
6316 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
6317 : * recently-created index, some index entries may point at "broken" HOT
6318 : * chains in which not all the tuple versions contain data matching the
6319 : * index entry. The live tuple version(s) certainly do match the index,
6320 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
6321 : * don't match. Hence, if we took data from the selected heap tuple, we
6322 : * might get a bogus answer that's not close to the index extremal value,
6323 : * or could even be NULL. We avoid this hazard because we take the data
6324 : * from the index entry not the heap.
6325 : *
6326 : * Despite all this care, there are situations where we might find many
6327 : * non-visible tuples near the end of the index. We don't want to expend
6328 : * a huge amount of time here, so we give up once we've read too many heap
6329 : * pages. When we fail for that reason, the caller will end up using
6330 : * whatever extremal value is recorded in pg_statistic.
6331 : */
6332 103518 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
6333 : GlobalVisTestFor(heapRel));
6334 :
6335 103518 : index_scan = index_beginscan(heapRel, indexRel,
6336 : &SnapshotNonVacuumable,
6337 : 1, 0);
6338 : /* Set it up for index-only scan */
6339 103518 : index_scan->xs_want_itup = true;
6340 103518 : index_rescan(index_scan, scankeys, 1, NULL, 0);
6341 :
6342 : /* Fetch first/next tuple in specified direction */
6343 118668 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
6344 : {
6345 118668 : BlockNumber block = ItemPointerGetBlockNumber(tid);
6346 :
6347 118668 : if (!VM_ALL_VISIBLE(heapRel,
6348 : block,
6349 : &vmbuffer))
6350 : {
6351 : /* Rats, we have to visit the heap to check visibility */
6352 79244 : if (!index_fetch_heap(index_scan, tableslot))
6353 : {
6354 : /*
6355 : * No visible tuple for this index entry, so we need to
6356 : * advance to the next entry. Before doing so, count heap
6357 : * page fetches and give up if we've done too many.
6358 : *
6359 : * We don't charge a page fetch if this is the same heap page
6360 : * as the previous tuple. This is on the conservative side,
6361 : * since other recently-accessed pages are probably still in
6362 : * buffers too; but it's good enough for this heuristic.
6363 : */
6364 : #define VISITED_PAGES_LIMIT 100
6365 :
6366 15150 : if (block != last_heap_block)
6367 : {
6368 1472 : last_heap_block = block;
6369 1472 : n_visited_heap_pages++;
6370 1472 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
6371 0 : break;
6372 : }
6373 :
6374 15150 : continue; /* no visible tuple, try next index entry */
6375 : }
6376 :
6377 : /* We don't actually need the heap tuple for anything */
6378 64094 : ExecClearTuple(tableslot);
6379 :
6380 : /*
6381 : * We don't care whether there's more than one visible tuple in
6382 : * the HOT chain; if any are visible, that's good enough.
6383 : */
6384 : }
6385 :
6386 : /*
6387 : * We expect that btree will return data in IndexTuple not HeapTuple
6388 : * format. It's not lossy either.
6389 : */
6390 103518 : if (!index_scan->xs_itup)
6391 0 : elog(ERROR, "no data returned for index-only scan");
6392 103518 : if (index_scan->xs_recheck)
6393 0 : elog(ERROR, "unexpected recheck indication from btree");
6394 :
6395 : /* OK to deconstruct the index tuple */
6396 103518 : index_deform_tuple(index_scan->xs_itup,
6397 : index_scan->xs_itupdesc,
6398 : values, isnull);
6399 :
6400 : /* Shouldn't have got a null, but be careful */
6401 103518 : if (isnull[0])
6402 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
6403 : RelationGetRelationName(indexRel));
6404 :
6405 : /* Copy the index column value out to caller's context */
6406 103518 : oldcontext = MemoryContextSwitchTo(outercontext);
6407 103518 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
6408 103518 : MemoryContextSwitchTo(oldcontext);
6409 103518 : have_data = true;
6410 103518 : break;
6411 : }
6412 :
6413 103518 : if (vmbuffer != InvalidBuffer)
6414 92138 : ReleaseBuffer(vmbuffer);
6415 103518 : index_endscan(index_scan);
6416 :
6417 103518 : return have_data;
6418 : }
6419 :
6420 : /*
6421 : * find_join_input_rel
6422 : * Look up the input relation for a join.
6423 : *
6424 : * We assume that the input relation's RelOptInfo must have been constructed
6425 : * already.
6426 : */
6427 : static RelOptInfo *
6428 7040 : find_join_input_rel(PlannerInfo *root, Relids relids)
6429 : {
6430 7040 : RelOptInfo *rel = NULL;
6431 :
6432 7040 : if (!bms_is_empty(relids))
6433 : {
6434 : int relid;
6435 :
6436 7040 : if (bms_get_singleton_member(relids, &relid))
6437 6736 : rel = find_base_rel(root, relid);
6438 : else
6439 304 : rel = find_join_rel(root, relids);
6440 : }
6441 :
6442 7040 : if (rel == NULL)
6443 0 : elog(ERROR, "could not find RelOptInfo for given relids");
6444 :
6445 7040 : return rel;
6446 : }
6447 :
6448 :
6449 : /*-------------------------------------------------------------------------
6450 : *
6451 : * Index cost estimation functions
6452 : *
6453 : *-------------------------------------------------------------------------
6454 : */
6455 :
6456 : /*
6457 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
6458 : */
6459 : List *
6460 645082 : get_quals_from_indexclauses(List *indexclauses)
6461 : {
6462 645082 : List *result = NIL;
6463 : ListCell *lc;
6464 :
6465 1135368 : foreach(lc, indexclauses)
6466 : {
6467 490286 : IndexClause *iclause = lfirst_node(IndexClause, lc);
6468 : ListCell *lc2;
6469 :
6470 983392 : foreach(lc2, iclause->indexquals)
6471 : {
6472 493106 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6473 :
6474 493106 : result = lappend(result, rinfo);
6475 : }
6476 : }
6477 645082 : return result;
6478 : }
6479 :
6480 : /*
6481 : * Compute the total evaluation cost of the comparison operands in a list
6482 : * of index qual expressions. Since we know these will be evaluated just
6483 : * once per scan, there's no need to distinguish startup from per-row cost.
6484 : *
6485 : * This can be used either on the result of get_quals_from_indexclauses(),
6486 : * or directly on an indexorderbys list. In both cases, we expect that the
6487 : * index key expression is on the left side of binary clauses.
6488 : */
6489 : Cost
6490 1277564 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
6491 : {
6492 1277564 : Cost qual_arg_cost = 0;
6493 : ListCell *lc;
6494 :
6495 1771132 : foreach(lc, indexquals)
6496 : {
6497 493568 : Expr *clause = (Expr *) lfirst(lc);
6498 : Node *other_operand;
6499 : QualCost index_qual_cost;
6500 :
6501 : /*
6502 : * Index quals will have RestrictInfos, indexorderbys won't. Look
6503 : * through RestrictInfo if present.
6504 : */
6505 493568 : if (IsA(clause, RestrictInfo))
6506 493094 : clause = ((RestrictInfo *) clause)->clause;
6507 :
6508 493568 : if (IsA(clause, OpExpr))
6509 : {
6510 482956 : OpExpr *op = (OpExpr *) clause;
6511 :
6512 482956 : other_operand = (Node *) lsecond(op->args);
6513 : }
6514 10612 : else if (IsA(clause, RowCompareExpr))
6515 : {
6516 132 : RowCompareExpr *rc = (RowCompareExpr *) clause;
6517 :
6518 132 : other_operand = (Node *) rc->rargs;
6519 : }
6520 10480 : else if (IsA(clause, ScalarArrayOpExpr))
6521 : {
6522 7686 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6523 :
6524 7686 : other_operand = (Node *) lsecond(saop->args);
6525 : }
6526 2794 : else if (IsA(clause, NullTest))
6527 : {
6528 2794 : other_operand = NULL;
6529 : }
6530 : else
6531 : {
6532 0 : elog(ERROR, "unsupported indexqual type: %d",
6533 : (int) nodeTag(clause));
6534 : other_operand = NULL; /* keep compiler quiet */
6535 : }
6536 :
6537 493568 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
6538 493568 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
6539 : }
6540 1277564 : return qual_arg_cost;
6541 : }
6542 :
6543 : void
6544 632494 : genericcostestimate(PlannerInfo *root,
6545 : IndexPath *path,
6546 : double loop_count,
6547 : GenericCosts *costs)
6548 : {
6549 632494 : IndexOptInfo *index = path->indexinfo;
6550 632494 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
6551 632494 : List *indexOrderBys = path->indexorderbys;
6552 : Cost indexStartupCost;
6553 : Cost indexTotalCost;
6554 : Selectivity indexSelectivity;
6555 : double indexCorrelation;
6556 : double numIndexPages;
6557 : double numIndexTuples;
6558 : double spc_random_page_cost;
6559 : double num_sa_scans;
6560 : double num_outer_scans;
6561 : double num_scans;
6562 : double qual_op_cost;
6563 : double qual_arg_cost;
6564 : List *selectivityQuals;
6565 : ListCell *l;
6566 :
6567 : /*
6568 : * If the index is partial, AND the index predicate with the explicitly
6569 : * given indexquals to produce a more accurate idea of the index
6570 : * selectivity.
6571 : */
6572 632494 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
6573 :
6574 : /*
6575 : * Check for ScalarArrayOpExpr index quals, and estimate the number of
6576 : * index scans that will be performed.
6577 : */
6578 632494 : num_sa_scans = 1;
6579 1112812 : foreach(l, indexQuals)
6580 : {
6581 480318 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
6582 :
6583 480318 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
6584 : {
6585 7680 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
6586 7680 : double alength = estimate_array_length(root, lsecond(saop->args));
6587 :
6588 7680 : if (alength > 1)
6589 7584 : num_sa_scans *= alength;
6590 : }
6591 : }
6592 :
6593 : /* Estimate the fraction of main-table tuples that will be visited */
6594 632494 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
6595 632494 : index->rel->relid,
6596 : JOIN_INNER,
6597 : NULL);
6598 :
6599 : /*
6600 : * If caller didn't give us an estimate, estimate the number of index
6601 : * tuples that will be visited. We do it in this rather peculiar-looking
6602 : * way in order to get the right answer for partial indexes.
6603 : */
6604 632494 : numIndexTuples = costs->numIndexTuples;
6605 632494 : if (numIndexTuples <= 0.0)
6606 : {
6607 56534 : numIndexTuples = indexSelectivity * index->rel->tuples;
6608 :
6609 : /*
6610 : * The above calculation counts all the tuples visited across all
6611 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
6612 : * average per-indexscan number, so adjust. This is a handy place to
6613 : * round to integer, too. (If caller supplied tuple estimate, it's
6614 : * responsible for handling these considerations.)
6615 : */
6616 56534 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
6617 : }
6618 :
6619 : /*
6620 : * We can bound the number of tuples by the index size in any case. Also,
6621 : * always estimate at least one tuple is touched, even when
6622 : * indexSelectivity estimate is tiny.
6623 : */
6624 632494 : if (numIndexTuples > index->tuples)
6625 4908 : numIndexTuples = index->tuples;
6626 632494 : if (numIndexTuples < 1.0)
6627 55784 : numIndexTuples = 1.0;
6628 :
6629 : /*
6630 : * Estimate the number of index pages that will be retrieved.
6631 : *
6632 : * We use the simplistic method of taking a pro-rata fraction of the total
6633 : * number of index pages. In effect, this counts only leaf pages and not
6634 : * any overhead such as index metapage or upper tree levels.
6635 : *
6636 : * In practice access to upper index levels is often nearly free because
6637 : * those tend to stay in cache under load; moreover, the cost involved is
6638 : * highly dependent on index type. We therefore ignore such costs here
6639 : * and leave it to the caller to add a suitable charge if needed.
6640 : */
6641 632494 : if (index->pages > 1 && index->tuples > 1)
6642 591124 : numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
6643 : else
6644 41370 : numIndexPages = 1.0;
6645 :
6646 : /* fetch estimated page cost for tablespace containing index */
6647 632494 : get_tablespace_page_costs(index->reltablespace,
6648 : &spc_random_page_cost,
6649 : NULL);
6650 :
6651 : /*
6652 : * Now compute the disk access costs.
6653 : *
6654 : * The above calculations are all per-index-scan. However, if we are in a
6655 : * nestloop inner scan, we can expect the scan to be repeated (with
6656 : * different search keys) for each row of the outer relation. Likewise,
6657 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
6658 : * the potential for cache effects to reduce the number of disk page
6659 : * fetches needed. We want to estimate the average per-scan I/O cost in
6660 : * the presence of caching.
6661 : *
6662 : * We use the Mackert-Lohman formula (see costsize.c for details) to
6663 : * estimate the total number of page fetches that occur. While this
6664 : * wasn't what it was designed for, it seems a reasonable model anyway.
6665 : * Note that we are counting pages not tuples anymore, so we take N = T =
6666 : * index size, as if there were one "tuple" per page.
6667 : */
6668 632494 : num_outer_scans = loop_count;
6669 632494 : num_scans = num_sa_scans * num_outer_scans;
6670 :
6671 632494 : if (num_scans > 1)
6672 : {
6673 : double pages_fetched;
6674 :
6675 : /* total page fetches ignoring cache effects */
6676 70488 : pages_fetched = numIndexPages * num_scans;
6677 :
6678 : /* use Mackert and Lohman formula to adjust for cache effects */
6679 70488 : pages_fetched = index_pages_fetched(pages_fetched,
6680 : index->pages,
6681 70488 : (double) index->pages,
6682 : root);
6683 :
6684 : /*
6685 : * Now compute the total disk access cost, and then report a pro-rated
6686 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
6687 : * since that's internal to the indexscan.)
6688 : */
6689 70488 : indexTotalCost = (pages_fetched * spc_random_page_cost)
6690 : / num_outer_scans;
6691 : }
6692 : else
6693 : {
6694 : /*
6695 : * For a single index scan, we just charge spc_random_page_cost per
6696 : * page touched.
6697 : */
6698 562006 : indexTotalCost = numIndexPages * spc_random_page_cost;
6699 : }
6700 :
6701 : /*
6702 : * CPU cost: any complex expressions in the indexquals will need to be
6703 : * evaluated once at the start of the scan to reduce them to runtime keys
6704 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
6705 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
6706 : * indexqual operator. Because we have numIndexTuples as a per-scan
6707 : * number, we have to multiply by num_sa_scans to get the correct result
6708 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
6709 : * ORDER BY expressions.
6710 : *
6711 : * Note: this neglects the possible costs of rechecking lossy operators.
6712 : * Detecting that that might be needed seems more expensive than it's
6713 : * worth, though, considering all the other inaccuracies here ...
6714 : */
6715 632494 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
6716 632494 : index_other_operands_eval_cost(root, indexOrderBys);
6717 632494 : qual_op_cost = cpu_operator_cost *
6718 632494 : (list_length(indexQuals) + list_length(indexOrderBys));
6719 :
6720 632494 : indexStartupCost = qual_arg_cost;
6721 632494 : indexTotalCost += qual_arg_cost;
6722 632494 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
6723 :
6724 : /*
6725 : * Generic assumption about index correlation: there isn't any.
6726 : */
6727 632494 : indexCorrelation = 0.0;
6728 :
6729 : /*
6730 : * Return everything to caller.
6731 : */
6732 632494 : costs->indexStartupCost = indexStartupCost;
6733 632494 : costs->indexTotalCost = indexTotalCost;
6734 632494 : costs->indexSelectivity = indexSelectivity;
6735 632494 : costs->indexCorrelation = indexCorrelation;
6736 632494 : costs->numIndexPages = numIndexPages;
6737 632494 : costs->numIndexTuples = numIndexTuples;
6738 632494 : costs->spc_random_page_cost = spc_random_page_cost;
6739 632494 : costs->num_sa_scans = num_sa_scans;
6740 632494 : }
6741 :
6742 : /*
6743 : * If the index is partial, add its predicate to the given qual list.
6744 : *
6745 : * ANDing the index predicate with the explicitly given indexquals produces
6746 : * a more accurate idea of the index's selectivity. However, we need to be
6747 : * careful not to insert redundant clauses, because clauselist_selectivity()
6748 : * is easily fooled into computing a too-low selectivity estimate. Our
6749 : * approach is to add only the predicate clause(s) that cannot be proven to
6750 : * be implied by the given indexquals. This successfully handles cases such
6751 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
6752 : * There are many other cases where we won't detect redundancy, leading to a
6753 : * too-low selectivity estimate, which will bias the system in favor of using
6754 : * partial indexes where possible. That is not necessarily bad though.
6755 : *
6756 : * Note that indexQuals contains RestrictInfo nodes while the indpred
6757 : * does not, so the output list will be mixed. This is OK for both
6758 : * predicate_implied_by() and clauselist_selectivity(), but might be
6759 : * problematic if the result were passed to other things.
6760 : */
6761 : List *
6762 1044448 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
6763 : {
6764 1044448 : List *predExtraQuals = NIL;
6765 : ListCell *lc;
6766 :
6767 1044448 : if (index->indpred == NIL)
6768 1042446 : return indexQuals;
6769 :
6770 4016 : foreach(lc, index->indpred)
6771 : {
6772 2014 : Node *predQual = (Node *) lfirst(lc);
6773 2014 : List *oneQual = list_make1(predQual);
6774 :
6775 2014 : if (!predicate_implied_by(oneQual, indexQuals, false))
6776 1812 : predExtraQuals = list_concat(predExtraQuals, oneQual);
6777 : }
6778 2002 : return list_concat(predExtraQuals, indexQuals);
6779 : }
6780 :
6781 :
6782 : void
6783 625014 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
6784 : Cost *indexStartupCost, Cost *indexTotalCost,
6785 : Selectivity *indexSelectivity, double *indexCorrelation,
6786 : double *indexPages)
6787 : {
6788 625014 : IndexOptInfo *index = path->indexinfo;
6789 625014 : GenericCosts costs = {0};
6790 : Oid relid;
6791 : AttrNumber colnum;
6792 625014 : VariableStatData vardata = {0};
6793 : double numIndexTuples;
6794 : Cost descentCost;
6795 : List *indexBoundQuals;
6796 : int indexcol;
6797 : bool eqQualHere;
6798 : bool found_saop;
6799 : bool found_is_null_op;
6800 : double num_sa_scans;
6801 : ListCell *lc;
6802 :
6803 : /*
6804 : * For a btree scan, only leading '=' quals plus inequality quals for the
6805 : * immediately next attribute contribute to index selectivity (these are
6806 : * the "boundary quals" that determine the starting and stopping points of
6807 : * the index scan). Additional quals can suppress visits to the heap, so
6808 : * it's OK to count them in indexSelectivity, but they should not count
6809 : * for estimating numIndexTuples. So we must examine the given indexquals
6810 : * to find out which ones count as boundary quals. We rely on the
6811 : * knowledge that they are given in index column order.
6812 : *
6813 : * For a RowCompareExpr, we consider only the first column, just as
6814 : * rowcomparesel() does.
6815 : *
6816 : * If there's a ScalarArrayOpExpr in the quals, we'll actually perform N
6817 : * index scans not one, but the ScalarArrayOpExpr's operator can be
6818 : * considered to act the same as it normally does.
6819 : */
6820 625014 : indexBoundQuals = NIL;
6821 625014 : indexcol = 0;
6822 625014 : eqQualHere = false;
6823 625014 : found_saop = false;
6824 625014 : found_is_null_op = false;
6825 625014 : num_sa_scans = 1;
6826 1070632 : foreach(lc, path->indexclauses)
6827 : {
6828 467368 : IndexClause *iclause = lfirst_node(IndexClause, lc);
6829 : ListCell *lc2;
6830 :
6831 467368 : if (indexcol != iclause->indexcol)
6832 : {
6833 : /* Beginning of a new column's quals */
6834 76852 : if (!eqQualHere)
6835 20534 : break; /* done if no '=' qual for indexcol */
6836 56318 : eqQualHere = false;
6837 56318 : indexcol++;
6838 56318 : if (indexcol != iclause->indexcol)
6839 1216 : break; /* no quals at all for indexcol */
6840 : }
6841 :
6842 : /* Examine each indexqual associated with this index clause */
6843 893896 : foreach(lc2, iclause->indexquals)
6844 : {
6845 448278 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6846 448278 : Expr *clause = rinfo->clause;
6847 448278 : Oid clause_op = InvalidOid;
6848 : int op_strategy;
6849 :
6850 448278 : if (IsA(clause, OpExpr))
6851 : {
6852 438466 : OpExpr *op = (OpExpr *) clause;
6853 :
6854 438466 : clause_op = op->opno;
6855 : }
6856 9812 : else if (IsA(clause, RowCompareExpr))
6857 : {
6858 132 : RowCompareExpr *rc = (RowCompareExpr *) clause;
6859 :
6860 132 : clause_op = linitial_oid(rc->opnos);
6861 : }
6862 9680 : else if (IsA(clause, ScalarArrayOpExpr))
6863 : {
6864 7480 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6865 7480 : Node *other_operand = (Node *) lsecond(saop->args);
6866 7480 : double alength = estimate_array_length(root, other_operand);
6867 :
6868 7480 : clause_op = saop->opno;
6869 7480 : found_saop = true;
6870 : /* count number of SA scans induced by indexBoundQuals only */
6871 7480 : if (alength > 1)
6872 7384 : num_sa_scans *= alength;
6873 : }
6874 2200 : else if (IsA(clause, NullTest))
6875 : {
6876 2200 : NullTest *nt = (NullTest *) clause;
6877 :
6878 2200 : if (nt->nulltesttype == IS_NULL)
6879 : {
6880 192 : found_is_null_op = true;
6881 : /* IS NULL is like = for selectivity purposes */
6882 192 : eqQualHere = true;
6883 : }
6884 : }
6885 : else
6886 0 : elog(ERROR, "unsupported indexqual type: %d",
6887 : (int) nodeTag(clause));
6888 :
6889 : /* check for equality operator */
6890 448278 : if (OidIsValid(clause_op))
6891 : {
6892 446078 : op_strategy = get_op_opfamily_strategy(clause_op,
6893 446078 : index->opfamily[indexcol]);
6894 : Assert(op_strategy != 0); /* not a member of opfamily?? */
6895 446078 : if (op_strategy == BTEqualStrategyNumber)
6896 422814 : eqQualHere = true;
6897 : }
6898 :
6899 448278 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
6900 : }
6901 : }
6902 :
6903 : /*
6904 : * If index is unique and we found an '=' clause for each column, we can
6905 : * just assume numIndexTuples = 1 and skip the expensive
6906 : * clauselist_selectivity calculations. However, a ScalarArrayOp or
6907 : * NullTest invalidates that theory, even though it sets eqQualHere.
6908 : */
6909 625014 : if (index->unique &&
6910 518712 : indexcol == index->nkeycolumns - 1 &&
6911 219250 : eqQualHere &&
6912 219250 : !found_saop &&
6913 214984 : !found_is_null_op)
6914 214918 : numIndexTuples = 1.0;
6915 : else
6916 : {
6917 : List *selectivityQuals;
6918 : Selectivity btreeSelectivity;
6919 :
6920 : /*
6921 : * If the index is partial, AND the index predicate with the
6922 : * index-bound quals to produce a more accurate idea of the number of
6923 : * rows covered by the bound conditions.
6924 : */
6925 410096 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
6926 :
6927 410096 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
6928 410096 : index->rel->relid,
6929 : JOIN_INNER,
6930 : NULL);
6931 410096 : numIndexTuples = btreeSelectivity * index->rel->tuples;
6932 :
6933 : /*
6934 : * As in genericcostestimate(), we have to adjust for any
6935 : * ScalarArrayOpExpr quals included in indexBoundQuals, and then round
6936 : * to integer.
6937 : */
6938 410096 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
6939 : }
6940 :
6941 : /*
6942 : * Now do generic index cost estimation.
6943 : */
6944 625014 : costs.numIndexTuples = numIndexTuples;
6945 :
6946 625014 : genericcostestimate(root, path, loop_count, &costs);
6947 :
6948 : /*
6949 : * Add a CPU-cost component to represent the costs of initial btree
6950 : * descent. We don't charge any I/O cost for touching upper btree levels,
6951 : * since they tend to stay in cache, but we still have to do about log2(N)
6952 : * comparisons to descend a btree of N leaf tuples. We charge one
6953 : * cpu_operator_cost per comparison.
6954 : *
6955 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
6956 : * ones after the first one are not startup cost so far as the overall
6957 : * plan is concerned, so add them only to "total" cost.
6958 : */
6959 625014 : if (index->tuples > 1) /* avoid computing log(0) */
6960 : {
6961 591650 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
6962 591650 : costs.indexStartupCost += descentCost;
6963 591650 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
6964 : }
6965 :
6966 : /*
6967 : * Even though we're not charging I/O cost for touching upper btree pages,
6968 : * it's still reasonable to charge some CPU cost per page descended
6969 : * through. Moreover, if we had no such charge at all, bloated indexes
6970 : * would appear to have the same search cost as unbloated ones, at least
6971 : * in cases where only a single leaf page is expected to be visited. This
6972 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
6973 : * touched. The number of such pages is btree tree height plus one (ie,
6974 : * we charge for the leaf page too). As above, charge once per SA scan.
6975 : */
6976 625014 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
6977 625014 : costs.indexStartupCost += descentCost;
6978 625014 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
6979 :
6980 : /*
6981 : * If we can get an estimate of the first column's ordering correlation C
6982 : * from pg_statistic, estimate the index correlation as C for a
6983 : * single-column index, or C * 0.75 for multiple columns. (The idea here
6984 : * is that multiple columns dilute the importance of the first column's
6985 : * ordering, but don't negate it entirely. Before 8.0 we divided the
6986 : * correlation by the number of columns, but that seems too strong.)
6987 : */
6988 625014 : if (index->indexkeys[0] != 0)
6989 : {
6990 : /* Simple variable --- look to stats for the underlying table */
6991 622848 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6992 :
6993 : Assert(rte->rtekind == RTE_RELATION);
6994 622848 : relid = rte->relid;
6995 : Assert(relid != InvalidOid);
6996 622848 : colnum = index->indexkeys[0];
6997 :
6998 622848 : if (get_relation_stats_hook &&
6999 0 : (*get_relation_stats_hook) (root, rte, colnum, &vardata))
7000 : {
7001 : /*
7002 : * The hook took control of acquiring a stats tuple. If it did
7003 : * supply a tuple, it'd better have supplied a freefunc.
7004 : */
7005 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
7006 0 : !vardata.freefunc)
7007 0 : elog(ERROR, "no function provided to release variable stats with");
7008 : }
7009 : else
7010 : {
7011 622848 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
7012 : ObjectIdGetDatum(relid),
7013 : Int16GetDatum(colnum),
7014 622848 : BoolGetDatum(rte->inh));
7015 622848 : vardata.freefunc = ReleaseSysCache;
7016 : }
7017 : }
7018 : else
7019 : {
7020 : /* Expression --- maybe there are stats for the index itself */
7021 2166 : relid = index->indexoid;
7022 2166 : colnum = 1;
7023 :
7024 2166 : if (get_index_stats_hook &&
7025 0 : (*get_index_stats_hook) (root, relid, colnum, &vardata))
7026 : {
7027 : /*
7028 : * The hook took control of acquiring a stats tuple. If it did
7029 : * supply a tuple, it'd better have supplied a freefunc.
7030 : */
7031 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
7032 0 : !vardata.freefunc)
7033 0 : elog(ERROR, "no function provided to release variable stats with");
7034 : }
7035 : else
7036 : {
7037 2166 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
7038 : ObjectIdGetDatum(relid),
7039 : Int16GetDatum(colnum),
7040 : BoolGetDatum(false));
7041 2166 : vardata.freefunc = ReleaseSysCache;
7042 : }
7043 : }
7044 :
7045 625014 : if (HeapTupleIsValid(vardata.statsTuple))
7046 : {
7047 : Oid sortop;
7048 : AttStatsSlot sslot;
7049 :
7050 456782 : sortop = get_opfamily_member(index->opfamily[0],
7051 456782 : index->opcintype[0],
7052 456782 : index->opcintype[0],
7053 : BTLessStrategyNumber);
7054 913564 : if (OidIsValid(sortop) &&
7055 456782 : get_attstatsslot(&sslot, vardata.statsTuple,
7056 : STATISTIC_KIND_CORRELATION, sortop,
7057 : ATTSTATSSLOT_NUMBERS))
7058 : {
7059 : double varCorrelation;
7060 :
7061 : Assert(sslot.nnumbers == 1);
7062 451952 : varCorrelation = sslot.numbers[0];
7063 :
7064 451952 : if (index->reverse_sort[0])
7065 0 : varCorrelation = -varCorrelation;
7066 :
7067 451952 : if (index->nkeycolumns > 1)
7068 141266 : costs.indexCorrelation = varCorrelation * 0.75;
7069 : else
7070 310686 : costs.indexCorrelation = varCorrelation;
7071 :
7072 451952 : free_attstatsslot(&sslot);
7073 : }
7074 : }
7075 :
7076 625014 : ReleaseVariableStats(vardata);
7077 :
7078 625014 : *indexStartupCost = costs.indexStartupCost;
7079 625014 : *indexTotalCost = costs.indexTotalCost;
7080 625014 : *indexSelectivity = costs.indexSelectivity;
7081 625014 : *indexCorrelation = costs.indexCorrelation;
7082 625014 : *indexPages = costs.numIndexPages;
7083 625014 : }
7084 :
7085 : void
7086 398 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7087 : Cost *indexStartupCost, Cost *indexTotalCost,
7088 : Selectivity *indexSelectivity, double *indexCorrelation,
7089 : double *indexPages)
7090 : {
7091 398 : GenericCosts costs = {0};
7092 :
7093 398 : genericcostestimate(root, path, loop_count, &costs);
7094 :
7095 : /*
7096 : * A hash index has no descent costs as such, since the index AM can go
7097 : * directly to the target bucket after computing the hash value. There
7098 : * are a couple of other hash-specific costs that we could conceivably add
7099 : * here, though:
7100 : *
7101 : * Ideally we'd charge spc_random_page_cost for each page in the target
7102 : * bucket, not just the numIndexPages pages that genericcostestimate
7103 : * thought we'd visit. However in most cases we don't know which bucket
7104 : * that will be. There's no point in considering the average bucket size
7105 : * because the hash AM makes sure that's always one page.
7106 : *
7107 : * Likewise, we could consider charging some CPU for each index tuple in
7108 : * the bucket, if we knew how many there were. But the per-tuple cost is
7109 : * just a hash value comparison, not a general datatype-dependent
7110 : * comparison, so any such charge ought to be quite a bit less than
7111 : * cpu_operator_cost; which makes it probably not worth worrying about.
7112 : *
7113 : * A bigger issue is that chance hash-value collisions will result in
7114 : * wasted probes into the heap. We don't currently attempt to model this
7115 : * cost on the grounds that it's rare, but maybe it's not rare enough.
7116 : * (Any fix for this ought to consider the generic lossy-operator problem,
7117 : * though; it's not entirely hash-specific.)
7118 : */
7119 :
7120 398 : *indexStartupCost = costs.indexStartupCost;
7121 398 : *indexTotalCost = costs.indexTotalCost;
7122 398 : *indexSelectivity = costs.indexSelectivity;
7123 398 : *indexCorrelation = costs.indexCorrelation;
7124 398 : *indexPages = costs.numIndexPages;
7125 398 : }
7126 :
7127 : void
7128 4486 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7129 : Cost *indexStartupCost, Cost *indexTotalCost,
7130 : Selectivity *indexSelectivity, double *indexCorrelation,
7131 : double *indexPages)
7132 : {
7133 4486 : IndexOptInfo *index = path->indexinfo;
7134 4486 : GenericCosts costs = {0};
7135 : Cost descentCost;
7136 :
7137 4486 : genericcostestimate(root, path, loop_count, &costs);
7138 :
7139 : /*
7140 : * We model index descent costs similarly to those for btree, but to do
7141 : * that we first need an idea of the tree height. We somewhat arbitrarily
7142 : * assume that the fanout is 100, meaning the tree height is at most
7143 : * log100(index->pages).
7144 : *
7145 : * Although this computation isn't really expensive enough to require
7146 : * caching, we might as well use index->tree_height to cache it.
7147 : */
7148 4486 : if (index->tree_height < 0) /* unknown? */
7149 : {
7150 4472 : if (index->pages > 1) /* avoid computing log(0) */
7151 2704 : index->tree_height = (int) (log(index->pages) / log(100.0));
7152 : else
7153 1768 : index->tree_height = 0;
7154 : }
7155 :
7156 : /*
7157 : * Add a CPU-cost component to represent the costs of initial descent. We
7158 : * just use log(N) here not log2(N) since the branching factor isn't
7159 : * necessarily two anyway. As for btree, charge once per SA scan.
7160 : */
7161 4486 : if (index->tuples > 1) /* avoid computing log(0) */
7162 : {
7163 4486 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7164 4486 : costs.indexStartupCost += descentCost;
7165 4486 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7166 : }
7167 :
7168 : /*
7169 : * Likewise add a per-page charge, calculated the same as for btrees.
7170 : */
7171 4486 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7172 4486 : costs.indexStartupCost += descentCost;
7173 4486 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7174 :
7175 4486 : *indexStartupCost = costs.indexStartupCost;
7176 4486 : *indexTotalCost = costs.indexTotalCost;
7177 4486 : *indexSelectivity = costs.indexSelectivity;
7178 4486 : *indexCorrelation = costs.indexCorrelation;
7179 4486 : *indexPages = costs.numIndexPages;
7180 4486 : }
7181 :
7182 : void
7183 1784 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7184 : Cost *indexStartupCost, Cost *indexTotalCost,
7185 : Selectivity *indexSelectivity, double *indexCorrelation,
7186 : double *indexPages)
7187 : {
7188 1784 : IndexOptInfo *index = path->indexinfo;
7189 1784 : GenericCosts costs = {0};
7190 : Cost descentCost;
7191 :
7192 1784 : genericcostestimate(root, path, loop_count, &costs);
7193 :
7194 : /*
7195 : * We model index descent costs similarly to those for btree, but to do
7196 : * that we first need an idea of the tree height. We somewhat arbitrarily
7197 : * assume that the fanout is 100, meaning the tree height is at most
7198 : * log100(index->pages).
7199 : *
7200 : * Although this computation isn't really expensive enough to require
7201 : * caching, we might as well use index->tree_height to cache it.
7202 : */
7203 1784 : if (index->tree_height < 0) /* unknown? */
7204 : {
7205 1778 : if (index->pages > 1) /* avoid computing log(0) */
7206 1778 : index->tree_height = (int) (log(index->pages) / log(100.0));
7207 : else
7208 0 : index->tree_height = 0;
7209 : }
7210 :
7211 : /*
7212 : * Add a CPU-cost component to represent the costs of initial descent. We
7213 : * just use log(N) here not log2(N) since the branching factor isn't
7214 : * necessarily two anyway. As for btree, charge once per SA scan.
7215 : */
7216 1784 : if (index->tuples > 1) /* avoid computing log(0) */
7217 : {
7218 1784 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7219 1784 : costs.indexStartupCost += descentCost;
7220 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7221 : }
7222 :
7223 : /*
7224 : * Likewise add a per-page charge, calculated the same as for btrees.
7225 : */
7226 1784 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7227 1784 : costs.indexStartupCost += descentCost;
7228 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7229 :
7230 1784 : *indexStartupCost = costs.indexStartupCost;
7231 1784 : *indexTotalCost = costs.indexTotalCost;
7232 1784 : *indexSelectivity = costs.indexSelectivity;
7233 1784 : *indexCorrelation = costs.indexCorrelation;
7234 1784 : *indexPages = costs.numIndexPages;
7235 1784 : }
7236 :
7237 :
7238 : /*
7239 : * Support routines for gincostestimate
7240 : */
7241 :
7242 : typedef struct
7243 : {
7244 : bool attHasFullScan[INDEX_MAX_KEYS];
7245 : bool attHasNormalScan[INDEX_MAX_KEYS];
7246 : double partialEntries;
7247 : double exactEntries;
7248 : double searchEntries;
7249 : double arrayScans;
7250 : } GinQualCounts;
7251 :
7252 : /*
7253 : * Estimate the number of index terms that need to be searched for while
7254 : * testing the given GIN query, and increment the counts in *counts
7255 : * appropriately. If the query is unsatisfiable, return false.
7256 : */
7257 : static bool
7258 2062 : gincost_pattern(IndexOptInfo *index, int indexcol,
7259 : Oid clause_op, Datum query,
7260 : GinQualCounts *counts)
7261 : {
7262 : FmgrInfo flinfo;
7263 : Oid extractProcOid;
7264 : Oid collation;
7265 : int strategy_op;
7266 : Oid lefttype,
7267 : righttype;
7268 2062 : int32 nentries = 0;
7269 2062 : bool *partial_matches = NULL;
7270 2062 : Pointer *extra_data = NULL;
7271 2062 : bool *nullFlags = NULL;
7272 2062 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
7273 : int32 i;
7274 :
7275 : Assert(indexcol < index->nkeycolumns);
7276 :
7277 : /*
7278 : * Get the operator's strategy number and declared input data types within
7279 : * the index opfamily. (We don't need the latter, but we use
7280 : * get_op_opfamily_properties because it will throw error if it fails to
7281 : * find a matching pg_amop entry.)
7282 : */
7283 2062 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
7284 : &strategy_op, &lefttype, &righttype);
7285 :
7286 : /*
7287 : * GIN always uses the "default" support functions, which are those with
7288 : * lefttype == righttype == the opclass' opcintype (see
7289 : * IndexSupportInitialize in relcache.c).
7290 : */
7291 2062 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
7292 2062 : index->opcintype[indexcol],
7293 2062 : index->opcintype[indexcol],
7294 : GIN_EXTRACTQUERY_PROC);
7295 :
7296 2062 : if (!OidIsValid(extractProcOid))
7297 : {
7298 : /* should not happen; throw same error as index_getprocinfo */
7299 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
7300 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
7301 : get_rel_name(index->indexoid));
7302 : }
7303 :
7304 : /*
7305 : * Choose collation to pass to extractProc (should match initGinState).
7306 : */
7307 2062 : if (OidIsValid(index->indexcollations[indexcol]))
7308 364 : collation = index->indexcollations[indexcol];
7309 : else
7310 1698 : collation = DEFAULT_COLLATION_OID;
7311 :
7312 2062 : fmgr_info(extractProcOid, &flinfo);
7313 :
7314 2062 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
7315 :
7316 2062 : FunctionCall7Coll(&flinfo,
7317 : collation,
7318 : query,
7319 : PointerGetDatum(&nentries),
7320 : UInt16GetDatum(strategy_op),
7321 : PointerGetDatum(&partial_matches),
7322 : PointerGetDatum(&extra_data),
7323 : PointerGetDatum(&nullFlags),
7324 : PointerGetDatum(&searchMode));
7325 :
7326 2062 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
7327 : {
7328 : /* No match is possible */
7329 12 : return false;
7330 : }
7331 :
7332 5786 : for (i = 0; i < nentries; i++)
7333 : {
7334 : /*
7335 : * For partial match we haven't any information to estimate number of
7336 : * matched entries in index, so, we just estimate it as 100
7337 : */
7338 3736 : if (partial_matches && partial_matches[i])
7339 316 : counts->partialEntries += 100;
7340 : else
7341 3420 : counts->exactEntries++;
7342 :
7343 3736 : counts->searchEntries++;
7344 : }
7345 :
7346 2050 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
7347 : {
7348 1578 : counts->attHasNormalScan[indexcol] = true;
7349 : }
7350 472 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
7351 : {
7352 : /* Treat "include empty" like an exact-match item */
7353 44 : counts->attHasNormalScan[indexcol] = true;
7354 44 : counts->exactEntries++;
7355 44 : counts->searchEntries++;
7356 : }
7357 : else
7358 : {
7359 : /* It's GIN_SEARCH_MODE_ALL */
7360 428 : counts->attHasFullScan[indexcol] = true;
7361 : }
7362 :
7363 2050 : return true;
7364 : }
7365 :
7366 : /*
7367 : * Estimate the number of index terms that need to be searched for while
7368 : * testing the given GIN index clause, and increment the counts in *counts
7369 : * appropriately. If the query is unsatisfiable, return false.
7370 : */
7371 : static bool
7372 2050 : gincost_opexpr(PlannerInfo *root,
7373 : IndexOptInfo *index,
7374 : int indexcol,
7375 : OpExpr *clause,
7376 : GinQualCounts *counts)
7377 : {
7378 2050 : Oid clause_op = clause->opno;
7379 2050 : Node *operand = (Node *) lsecond(clause->args);
7380 :
7381 : /* aggressively reduce to a constant, and look through relabeling */
7382 2050 : operand = estimate_expression_value(root, operand);
7383 :
7384 2050 : if (IsA(operand, RelabelType))
7385 0 : operand = (Node *) ((RelabelType *) operand)->arg;
7386 :
7387 : /*
7388 : * It's impossible to call extractQuery method for unknown operand. So
7389 : * unless operand is a Const we can't do much; just assume there will be
7390 : * one ordinary search entry from the operand at runtime.
7391 : */
7392 2050 : if (!IsA(operand, Const))
7393 : {
7394 0 : counts->exactEntries++;
7395 0 : counts->searchEntries++;
7396 0 : return true;
7397 : }
7398 :
7399 : /* If Const is null, there can be no matches */
7400 2050 : if (((Const *) operand)->constisnull)
7401 0 : return false;
7402 :
7403 : /* Otherwise, apply extractQuery and get the actual term counts */
7404 2050 : return gincost_pattern(index, indexcol, clause_op,
7405 : ((Const *) operand)->constvalue,
7406 : counts);
7407 : }
7408 :
7409 : /*
7410 : * Estimate the number of index terms that need to be searched for while
7411 : * testing the given GIN index clause, and increment the counts in *counts
7412 : * appropriately. If the query is unsatisfiable, return false.
7413 : *
7414 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
7415 : * each of which involves one value from the RHS array, plus all the
7416 : * non-array quals (if any). To model this, we average the counts across
7417 : * the RHS elements, and add the averages to the counts in *counts (which
7418 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
7419 : * by N, causing gincostestimate to scale up its estimates accordingly.
7420 : */
7421 : static bool
7422 6 : gincost_scalararrayopexpr(PlannerInfo *root,
7423 : IndexOptInfo *index,
7424 : int indexcol,
7425 : ScalarArrayOpExpr *clause,
7426 : double numIndexEntries,
7427 : GinQualCounts *counts)
7428 : {
7429 6 : Oid clause_op = clause->opno;
7430 6 : Node *rightop = (Node *) lsecond(clause->args);
7431 : ArrayType *arrayval;
7432 : int16 elmlen;
7433 : bool elmbyval;
7434 : char elmalign;
7435 : int numElems;
7436 : Datum *elemValues;
7437 : bool *elemNulls;
7438 : GinQualCounts arraycounts;
7439 6 : int numPossible = 0;
7440 : int i;
7441 :
7442 : Assert(clause->useOr);
7443 :
7444 : /* aggressively reduce to a constant, and look through relabeling */
7445 6 : rightop = estimate_expression_value(root, rightop);
7446 :
7447 6 : if (IsA(rightop, RelabelType))
7448 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
7449 :
7450 : /*
7451 : * It's impossible to call extractQuery method for unknown operand. So
7452 : * unless operand is a Const we can't do much; just assume there will be
7453 : * one ordinary search entry from each array entry at runtime, and fall
7454 : * back on a probably-bad estimate of the number of array entries.
7455 : */
7456 6 : if (!IsA(rightop, Const))
7457 : {
7458 0 : counts->exactEntries++;
7459 0 : counts->searchEntries++;
7460 0 : counts->arrayScans *= estimate_array_length(root, rightop);
7461 0 : return true;
7462 : }
7463 :
7464 : /* If Const is null, there can be no matches */
7465 6 : if (((Const *) rightop)->constisnull)
7466 0 : return false;
7467 :
7468 : /* Otherwise, extract the array elements and iterate over them */
7469 6 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
7470 6 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
7471 : &elmlen, &elmbyval, &elmalign);
7472 6 : deconstruct_array(arrayval,
7473 : ARR_ELEMTYPE(arrayval),
7474 : elmlen, elmbyval, elmalign,
7475 : &elemValues, &elemNulls, &numElems);
7476 :
7477 6 : memset(&arraycounts, 0, sizeof(arraycounts));
7478 :
7479 18 : for (i = 0; i < numElems; i++)
7480 : {
7481 : GinQualCounts elemcounts;
7482 :
7483 : /* NULL can't match anything, so ignore, as the executor will */
7484 12 : if (elemNulls[i])
7485 0 : continue;
7486 :
7487 : /* Otherwise, apply extractQuery and get the actual term counts */
7488 12 : memset(&elemcounts, 0, sizeof(elemcounts));
7489 :
7490 12 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
7491 : &elemcounts))
7492 : {
7493 : /* We ignore array elements that are unsatisfiable patterns */
7494 12 : numPossible++;
7495 :
7496 12 : if (elemcounts.attHasFullScan[indexcol] &&
7497 0 : !elemcounts.attHasNormalScan[indexcol])
7498 : {
7499 : /*
7500 : * Full index scan will be required. We treat this as if
7501 : * every key in the index had been listed in the query; is
7502 : * that reasonable?
7503 : */
7504 0 : elemcounts.partialEntries = 0;
7505 0 : elemcounts.exactEntries = numIndexEntries;
7506 0 : elemcounts.searchEntries = numIndexEntries;
7507 : }
7508 12 : arraycounts.partialEntries += elemcounts.partialEntries;
7509 12 : arraycounts.exactEntries += elemcounts.exactEntries;
7510 12 : arraycounts.searchEntries += elemcounts.searchEntries;
7511 : }
7512 : }
7513 :
7514 6 : if (numPossible == 0)
7515 : {
7516 : /* No satisfiable patterns in the array */
7517 0 : return false;
7518 : }
7519 :
7520 : /*
7521 : * Now add the averages to the global counts. This will give us an
7522 : * estimate of the average number of terms searched for in each indexscan,
7523 : * including contributions from both array and non-array quals.
7524 : */
7525 6 : counts->partialEntries += arraycounts.partialEntries / numPossible;
7526 6 : counts->exactEntries += arraycounts.exactEntries / numPossible;
7527 6 : counts->searchEntries += arraycounts.searchEntries / numPossible;
7528 :
7529 6 : counts->arrayScans *= numPossible;
7530 :
7531 6 : return true;
7532 : }
7533 :
7534 : /*
7535 : * GIN has search behavior completely different from other index types
7536 : */
7537 : void
7538 1858 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7539 : Cost *indexStartupCost, Cost *indexTotalCost,
7540 : Selectivity *indexSelectivity, double *indexCorrelation,
7541 : double *indexPages)
7542 : {
7543 1858 : IndexOptInfo *index = path->indexinfo;
7544 1858 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7545 : List *selectivityQuals;
7546 1858 : double numPages = index->pages,
7547 1858 : numTuples = index->tuples;
7548 : double numEntryPages,
7549 : numDataPages,
7550 : numPendingPages,
7551 : numEntries;
7552 : GinQualCounts counts;
7553 : bool matchPossible;
7554 : bool fullIndexScan;
7555 : double partialScale;
7556 : double entryPagesFetched,
7557 : dataPagesFetched,
7558 : dataPagesFetchedBySel;
7559 : double qual_op_cost,
7560 : qual_arg_cost,
7561 : spc_random_page_cost,
7562 : outer_scans;
7563 : Cost descentCost;
7564 : Relation indexRel;
7565 : GinStatsData ginStats;
7566 : ListCell *lc;
7567 : int i;
7568 :
7569 : /*
7570 : * Obtain statistical information from the meta page, if possible. Else
7571 : * set ginStats to zeroes, and we'll cope below.
7572 : */
7573 1858 : if (!index->hypothetical)
7574 : {
7575 : /* Lock should have already been obtained in plancat.c */
7576 1858 : indexRel = index_open(index->indexoid, NoLock);
7577 1858 : ginGetStats(indexRel, &ginStats);
7578 1858 : index_close(indexRel, NoLock);
7579 : }
7580 : else
7581 : {
7582 0 : memset(&ginStats, 0, sizeof(ginStats));
7583 : }
7584 :
7585 : /*
7586 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
7587 : * trusted, but the other fields are data as of the last VACUUM. We can
7588 : * scale them up to account for growth since then, but that method only
7589 : * goes so far; in the worst case, the stats might be for a completely
7590 : * empty index, and scaling them will produce pretty bogus numbers.
7591 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
7592 : * it's grown more than that, fall back to estimating things only from the
7593 : * assumed-accurate index size. But we'll trust nPendingPages in any case
7594 : * so long as it's not clearly insane, ie, more than the index size.
7595 : */
7596 1858 : if (ginStats.nPendingPages < numPages)
7597 1858 : numPendingPages = ginStats.nPendingPages;
7598 : else
7599 0 : numPendingPages = 0;
7600 :
7601 1858 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
7602 1858 : ginStats.nTotalPages > numPages / 4 &&
7603 1816 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
7604 1558 : {
7605 : /*
7606 : * OK, the stats seem close enough to sane to be trusted. But we
7607 : * still need to scale them by the ratio numPages / nTotalPages to
7608 : * account for growth since the last VACUUM.
7609 : */
7610 1558 : double scale = numPages / ginStats.nTotalPages;
7611 :
7612 1558 : numEntryPages = ceil(ginStats.nEntryPages * scale);
7613 1558 : numDataPages = ceil(ginStats.nDataPages * scale);
7614 1558 : numEntries = ceil(ginStats.nEntries * scale);
7615 : /* ensure we didn't round up too much */
7616 1558 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
7617 1558 : numDataPages = Min(numDataPages,
7618 : numPages - numPendingPages - numEntryPages);
7619 : }
7620 : else
7621 : {
7622 : /*
7623 : * We might get here because it's a hypothetical index, or an index
7624 : * created pre-9.1 and never vacuumed since upgrading (in which case
7625 : * its stats would read as zeroes), or just because it's grown too
7626 : * much since the last VACUUM for us to put our faith in scaling.
7627 : *
7628 : * Invent some plausible internal statistics based on the index page
7629 : * count (and clamp that to at least 10 pages, just in case). We
7630 : * estimate that 90% of the index is entry pages, and the rest is data
7631 : * pages. Estimate 100 entries per entry page; this is rather bogus
7632 : * since it'll depend on the size of the keys, but it's more robust
7633 : * than trying to predict the number of entries per heap tuple.
7634 : */
7635 300 : numPages = Max(numPages, 10);
7636 300 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
7637 300 : numDataPages = numPages - numPendingPages - numEntryPages;
7638 300 : numEntries = floor(numEntryPages * 100);
7639 : }
7640 :
7641 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
7642 1858 : if (numEntries < 1)
7643 0 : numEntries = 1;
7644 :
7645 : /*
7646 : * If the index is partial, AND the index predicate with the index-bound
7647 : * quals to produce a more accurate idea of the number of rows covered by
7648 : * the bound conditions.
7649 : */
7650 1858 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7651 :
7652 : /* Estimate the fraction of main-table tuples that will be visited */
7653 3716 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7654 1858 : index->rel->relid,
7655 : JOIN_INNER,
7656 : NULL);
7657 :
7658 : /* fetch estimated page cost for tablespace containing index */
7659 1858 : get_tablespace_page_costs(index->reltablespace,
7660 : &spc_random_page_cost,
7661 : NULL);
7662 :
7663 : /*
7664 : * Generic assumption about index correlation: there isn't any.
7665 : */
7666 1858 : *indexCorrelation = 0.0;
7667 :
7668 : /*
7669 : * Examine quals to estimate number of search entries & partial matches
7670 : */
7671 1858 : memset(&counts, 0, sizeof(counts));
7672 1858 : counts.arrayScans = 1;
7673 1858 : matchPossible = true;
7674 :
7675 3914 : foreach(lc, path->indexclauses)
7676 : {
7677 2056 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7678 : ListCell *lc2;
7679 :
7680 4100 : foreach(lc2, iclause->indexquals)
7681 : {
7682 2056 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7683 2056 : Expr *clause = rinfo->clause;
7684 :
7685 2056 : if (IsA(clause, OpExpr))
7686 : {
7687 2050 : matchPossible = gincost_opexpr(root,
7688 : index,
7689 2050 : iclause->indexcol,
7690 : (OpExpr *) clause,
7691 : &counts);
7692 2050 : if (!matchPossible)
7693 12 : break;
7694 : }
7695 6 : else if (IsA(clause, ScalarArrayOpExpr))
7696 : {
7697 6 : matchPossible = gincost_scalararrayopexpr(root,
7698 : index,
7699 6 : iclause->indexcol,
7700 : (ScalarArrayOpExpr *) clause,
7701 : numEntries,
7702 : &counts);
7703 6 : if (!matchPossible)
7704 0 : break;
7705 : }
7706 : else
7707 : {
7708 : /* shouldn't be anything else for a GIN index */
7709 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
7710 : (int) nodeTag(clause));
7711 : }
7712 : }
7713 : }
7714 :
7715 : /* Fall out if there were any provably-unsatisfiable quals */
7716 1858 : if (!matchPossible)
7717 : {
7718 12 : *indexStartupCost = 0;
7719 12 : *indexTotalCost = 0;
7720 12 : *indexSelectivity = 0;
7721 12 : return;
7722 : }
7723 :
7724 : /*
7725 : * If attribute has a full scan and at the same time doesn't have normal
7726 : * scan, then we'll have to scan all non-null entries of that attribute.
7727 : * Currently, we don't have per-attribute statistics for GIN. Thus, we
7728 : * must assume the whole GIN index has to be scanned in this case.
7729 : */
7730 1846 : fullIndexScan = false;
7731 3582 : for (i = 0; i < index->nkeycolumns; i++)
7732 : {
7733 2074 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
7734 : {
7735 338 : fullIndexScan = true;
7736 338 : break;
7737 : }
7738 : }
7739 :
7740 1846 : if (fullIndexScan || indexQuals == NIL)
7741 : {
7742 : /*
7743 : * Full index scan will be required. We treat this as if every key in
7744 : * the index had been listed in the query; is that reasonable?
7745 : */
7746 338 : counts.partialEntries = 0;
7747 338 : counts.exactEntries = numEntries;
7748 338 : counts.searchEntries = numEntries;
7749 : }
7750 :
7751 : /* Will we have more than one iteration of a nestloop scan? */
7752 1846 : outer_scans = loop_count;
7753 :
7754 : /*
7755 : * Compute cost to begin scan, first of all, pay attention to pending
7756 : * list.
7757 : */
7758 1846 : entryPagesFetched = numPendingPages;
7759 :
7760 : /*
7761 : * Estimate number of entry pages read. We need to do
7762 : * counts.searchEntries searches. Use a power function as it should be,
7763 : * but tuples on leaf pages usually is much greater. Here we include all
7764 : * searches in entry tree, including search of first entry in partial
7765 : * match algorithm
7766 : */
7767 1846 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
7768 :
7769 : /*
7770 : * Add an estimate of entry pages read by partial match algorithm. It's a
7771 : * scan over leaf pages in entry tree. We haven't any useful stats here,
7772 : * so estimate it as proportion. Because counts.partialEntries is really
7773 : * pretty bogus (see code above), it's possible that it is more than
7774 : * numEntries; clamp the proportion to ensure sanity.
7775 : */
7776 1846 : partialScale = counts.partialEntries / numEntries;
7777 1846 : partialScale = Min(partialScale, 1.0);
7778 :
7779 1846 : entryPagesFetched += ceil(numEntryPages * partialScale);
7780 :
7781 : /*
7782 : * Partial match algorithm reads all data pages before doing actual scan,
7783 : * so it's a startup cost. Again, we haven't any useful stats here, so
7784 : * estimate it as proportion.
7785 : */
7786 1846 : dataPagesFetched = ceil(numDataPages * partialScale);
7787 :
7788 1846 : *indexStartupCost = 0;
7789 1846 : *indexTotalCost = 0;
7790 :
7791 : /*
7792 : * Add a CPU-cost component to represent the costs of initial entry btree
7793 : * descent. We don't charge any I/O cost for touching upper btree levels,
7794 : * since they tend to stay in cache, but we still have to do about log2(N)
7795 : * comparisons to descend a btree of N leaf tuples. We charge one
7796 : * cpu_operator_cost per comparison.
7797 : *
7798 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
7799 : * ones after the first one are not startup cost so far as the overall
7800 : * plan is concerned, so add them only to "total" cost.
7801 : */
7802 1846 : if (numEntries > 1) /* avoid computing log(0) */
7803 : {
7804 1846 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
7805 1846 : *indexStartupCost += descentCost * counts.searchEntries;
7806 1846 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
7807 : }
7808 :
7809 : /*
7810 : * Add a cpu cost per entry-page fetched. This is not amortized over a
7811 : * loop.
7812 : */
7813 1846 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7814 1846 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7815 :
7816 : /*
7817 : * Add a cpu cost per data-page fetched. This is also not amortized over a
7818 : * loop. Since those are the data pages from the partial match algorithm,
7819 : * charge them as startup cost.
7820 : */
7821 1846 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
7822 :
7823 : /*
7824 : * Since we add the startup cost to the total cost later on, remove the
7825 : * initial arrayscan from the total.
7826 : */
7827 1846 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7828 :
7829 : /*
7830 : * Calculate cache effects if more than one scan due to nestloops or array
7831 : * quals. The result is pro-rated per nestloop scan, but the array qual
7832 : * factor shouldn't be pro-rated (compare genericcostestimate).
7833 : */
7834 1846 : if (outer_scans > 1 || counts.arrayScans > 1)
7835 : {
7836 6 : entryPagesFetched *= outer_scans * counts.arrayScans;
7837 6 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
7838 : (BlockNumber) numEntryPages,
7839 : numEntryPages, root);
7840 6 : entryPagesFetched /= outer_scans;
7841 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
7842 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
7843 : (BlockNumber) numDataPages,
7844 : numDataPages, root);
7845 6 : dataPagesFetched /= outer_scans;
7846 : }
7847 :
7848 : /*
7849 : * Here we use random page cost because logically-close pages could be far
7850 : * apart on disk.
7851 : */
7852 1846 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
7853 :
7854 : /*
7855 : * Now compute the number of data pages fetched during the scan.
7856 : *
7857 : * We assume every entry to have the same number of items, and that there
7858 : * is no overlap between them. (XXX: tsvector and array opclasses collect
7859 : * statistics on the frequency of individual keys; it would be nice to use
7860 : * those here.)
7861 : */
7862 1846 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
7863 :
7864 : /*
7865 : * If there is a lot of overlap among the entries, in particular if one of
7866 : * the entries is very frequent, the above calculation can grossly
7867 : * under-estimate. As a simple cross-check, calculate a lower bound based
7868 : * on the overall selectivity of the quals. At a minimum, we must read
7869 : * one item pointer for each matching entry.
7870 : *
7871 : * The width of each item pointer varies, based on the level of
7872 : * compression. We don't have statistics on that, but an average of
7873 : * around 3 bytes per item is fairly typical.
7874 : */
7875 1846 : dataPagesFetchedBySel = ceil(*indexSelectivity *
7876 1846 : (numTuples / (BLCKSZ / 3)));
7877 1846 : if (dataPagesFetchedBySel > dataPagesFetched)
7878 1472 : dataPagesFetched = dataPagesFetchedBySel;
7879 :
7880 : /* Add one page cpu-cost to the startup cost */
7881 1846 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
7882 :
7883 : /*
7884 : * Add once again a CPU-cost for those data pages, before amortizing for
7885 : * cache.
7886 : */
7887 1846 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7888 :
7889 : /* Account for cache effects, the same as above */
7890 1846 : if (outer_scans > 1 || counts.arrayScans > 1)
7891 : {
7892 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
7893 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
7894 : (BlockNumber) numDataPages,
7895 : numDataPages, root);
7896 6 : dataPagesFetched /= outer_scans;
7897 : }
7898 :
7899 : /* And apply random_page_cost as the cost per page */
7900 1846 : *indexTotalCost += *indexStartupCost +
7901 1846 : dataPagesFetched * spc_random_page_cost;
7902 :
7903 : /*
7904 : * Add on index qual eval costs, much as in genericcostestimate. We charge
7905 : * cpu but we can disregard indexorderbys, since GIN doesn't support
7906 : * those.
7907 : */
7908 1846 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
7909 1846 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
7910 :
7911 1846 : *indexStartupCost += qual_arg_cost;
7912 1846 : *indexTotalCost += qual_arg_cost;
7913 :
7914 : /*
7915 : * Add a cpu cost per search entry, corresponding to the actual visited
7916 : * entries.
7917 : */
7918 1846 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
7919 : /* Now add a cpu cost per tuple in the posting lists / trees */
7920 1846 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
7921 1846 : *indexPages = dataPagesFetched;
7922 : }
7923 :
7924 : /*
7925 : * BRIN has search behavior completely different from other index types
7926 : */
7927 : void
7928 10730 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7929 : Cost *indexStartupCost, Cost *indexTotalCost,
7930 : Selectivity *indexSelectivity, double *indexCorrelation,
7931 : double *indexPages)
7932 : {
7933 10730 : IndexOptInfo *index = path->indexinfo;
7934 10730 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7935 10730 : double numPages = index->pages;
7936 10730 : RelOptInfo *baserel = index->rel;
7937 10730 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
7938 : Cost spc_seq_page_cost;
7939 : Cost spc_random_page_cost;
7940 : double qual_arg_cost;
7941 : double qualSelectivity;
7942 : BrinStatsData statsData;
7943 : double indexRanges;
7944 : double minimalRanges;
7945 : double estimatedRanges;
7946 : double selec;
7947 : Relation indexRel;
7948 : ListCell *l;
7949 : VariableStatData vardata;
7950 :
7951 : Assert(rte->rtekind == RTE_RELATION);
7952 :
7953 : /* fetch estimated page cost for the tablespace containing the index */
7954 10730 : get_tablespace_page_costs(index->reltablespace,
7955 : &spc_random_page_cost,
7956 : &spc_seq_page_cost);
7957 :
7958 : /*
7959 : * Obtain some data from the index itself, if possible. Otherwise invent
7960 : * some plausible internal statistics based on the relation page count.
7961 : */
7962 10730 : if (!index->hypothetical)
7963 : {
7964 : /*
7965 : * A lock should have already been obtained on the index in plancat.c.
7966 : */
7967 10730 : indexRel = index_open(index->indexoid, NoLock);
7968 10730 : brinGetStats(indexRel, &statsData);
7969 10730 : index_close(indexRel, NoLock);
7970 :
7971 : /* work out the actual number of ranges in the index */
7972 10730 : indexRanges = Max(ceil((double) baserel->pages /
7973 : statsData.pagesPerRange), 1.0);
7974 : }
7975 : else
7976 : {
7977 : /*
7978 : * Assume default number of pages per range, and estimate the number
7979 : * of ranges based on that.
7980 : */
7981 0 : indexRanges = Max(ceil((double) baserel->pages /
7982 : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
7983 :
7984 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
7985 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
7986 : }
7987 :
7988 : /*
7989 : * Compute index correlation
7990 : *
7991 : * Because we can use all index quals equally when scanning, we can use
7992 : * the largest correlation (in absolute value) among columns used by the
7993 : * query. Start at zero, the worst possible case. If we cannot find any
7994 : * correlation statistics, we will keep it as 0.
7995 : */
7996 10730 : *indexCorrelation = 0;
7997 :
7998 21462 : foreach(l, path->indexclauses)
7999 : {
8000 10732 : IndexClause *iclause = lfirst_node(IndexClause, l);
8001 10732 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
8002 :
8003 : /* attempt to lookup stats in relation for this index column */
8004 10732 : if (attnum != 0)
8005 : {
8006 : /* Simple variable -- look to stats for the underlying table */
8007 10732 : if (get_relation_stats_hook &&
8008 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
8009 : {
8010 : /*
8011 : * The hook took control of acquiring a stats tuple. If it
8012 : * did supply a tuple, it'd better have supplied a freefunc.
8013 : */
8014 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
8015 0 : elog(ERROR,
8016 : "no function provided to release variable stats with");
8017 : }
8018 : else
8019 : {
8020 10732 : vardata.statsTuple =
8021 10732 : SearchSysCache3(STATRELATTINH,
8022 : ObjectIdGetDatum(rte->relid),
8023 : Int16GetDatum(attnum),
8024 : BoolGetDatum(false));
8025 10732 : vardata.freefunc = ReleaseSysCache;
8026 : }
8027 : }
8028 : else
8029 : {
8030 : /*
8031 : * Looks like we've found an expression column in the index. Let's
8032 : * see if there's any stats for it.
8033 : */
8034 :
8035 : /* get the attnum from the 0-based index. */
8036 0 : attnum = iclause->indexcol + 1;
8037 :
8038 0 : if (get_index_stats_hook &&
8039 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
8040 : {
8041 : /*
8042 : * The hook took control of acquiring a stats tuple. If it
8043 : * did supply a tuple, it'd better have supplied a freefunc.
8044 : */
8045 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
8046 0 : !vardata.freefunc)
8047 0 : elog(ERROR, "no function provided to release variable stats with");
8048 : }
8049 : else
8050 : {
8051 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
8052 : ObjectIdGetDatum(index->indexoid),
8053 : Int16GetDatum(attnum),
8054 : BoolGetDatum(false));
8055 0 : vardata.freefunc = ReleaseSysCache;
8056 : }
8057 : }
8058 :
8059 10732 : if (HeapTupleIsValid(vardata.statsTuple))
8060 : {
8061 : AttStatsSlot sslot;
8062 :
8063 36 : if (get_attstatsslot(&sslot, vardata.statsTuple,
8064 : STATISTIC_KIND_CORRELATION, InvalidOid,
8065 : ATTSTATSSLOT_NUMBERS))
8066 : {
8067 36 : double varCorrelation = 0.0;
8068 :
8069 36 : if (sslot.nnumbers > 0)
8070 36 : varCorrelation = fabs(sslot.numbers[0]);
8071 :
8072 36 : if (varCorrelation > *indexCorrelation)
8073 36 : *indexCorrelation = varCorrelation;
8074 :
8075 36 : free_attstatsslot(&sslot);
8076 : }
8077 : }
8078 :
8079 10732 : ReleaseVariableStats(vardata);
8080 : }
8081 :
8082 10730 : qualSelectivity = clauselist_selectivity(root, indexQuals,
8083 10730 : baserel->relid,
8084 : JOIN_INNER, NULL);
8085 :
8086 : /*
8087 : * Now calculate the minimum possible ranges we could match with if all of
8088 : * the rows were in the perfect order in the table's heap.
8089 : */
8090 10730 : minimalRanges = ceil(indexRanges * qualSelectivity);
8091 :
8092 : /*
8093 : * Now estimate the number of ranges that we'll touch by using the
8094 : * indexCorrelation from the stats. Careful not to divide by zero (note
8095 : * we're using the absolute value of the correlation).
8096 : */
8097 10730 : if (*indexCorrelation < 1.0e-10)
8098 10694 : estimatedRanges = indexRanges;
8099 : else
8100 36 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
8101 :
8102 : /* we expect to visit this portion of the table */
8103 10730 : selec = estimatedRanges / indexRanges;
8104 :
8105 10730 : CLAMP_PROBABILITY(selec);
8106 :
8107 10730 : *indexSelectivity = selec;
8108 :
8109 : /*
8110 : * Compute the index qual costs, much as in genericcostestimate, to add to
8111 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
8112 : * support those.
8113 : */
8114 10730 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8115 :
8116 : /*
8117 : * Compute the startup cost as the cost to read the whole revmap
8118 : * sequentially, including the cost to execute the index quals.
8119 : */
8120 10730 : *indexStartupCost =
8121 10730 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
8122 10730 : *indexStartupCost += qual_arg_cost;
8123 :
8124 : /*
8125 : * To read a BRIN index there might be a bit of back and forth over
8126 : * regular pages, as revmap might point to them out of sequential order;
8127 : * calculate the total cost as reading the whole index in random order.
8128 : */
8129 10730 : *indexTotalCost = *indexStartupCost +
8130 10730 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
8131 :
8132 : /*
8133 : * Charge a small amount per range tuple which we expect to match to. This
8134 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
8135 : * will set a bit for each page in the range when we find a matching
8136 : * range, so we must multiply the charge by the number of pages in the
8137 : * range.
8138 : */
8139 10730 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
8140 10730 : statsData.pagesPerRange;
8141 :
8142 10730 : *indexPages = index->pages;
8143 10730 : }
|