PostgreSQL Source Code git master
All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Pages
selfuncs.c
Go to the documentation of this file.
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"
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"
123#include "parser/parsetree.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 */
149
150static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
151static 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,
157 bool have_mcvs1, bool have_mcvs2);
158static 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,
164 bool have_mcvs1, bool have_mcvs2,
165 RelOptInfo *inner_rel);
167 RelOptInfo *rel, List **varinfos, double *ndistinct);
168static 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);
172static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure);
173static void convert_string_to_scalar(char *value,
174 double *scaledvalue,
175 char *lobound,
176 double *scaledlobound,
177 char *hibound,
178 double *scaledhibound);
180 double *scaledvalue,
181 Datum lobound,
182 double *scaledlobound,
183 Datum hibound,
184 double *scaledhibound);
185static double convert_one_string_to_scalar(char *value,
186 int rangelo, int rangehi);
187static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
188 int rangelo, int rangehi);
189static char *convert_string_datum(Datum value, Oid typid, Oid collid,
190 bool *failure);
191static double convert_timevalue_to_scalar(Datum value, Oid typid,
192 bool *failure);
194 VariableStatData *vardata);
196 Oid sortop, Oid collation,
197 Datum *min, Datum *max);
198static 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);
203 VariableStatData *vardata,
204 Oid sortop, Oid collation,
205 Datum *min, Datum *max);
206static 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);
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 */
226Datum
228{
229 PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
230}
231
232/*
233 * Common code for eqsel() and neqsel()
234 */
235static double
237{
239 Oid operator = PG_GETARG_OID(1);
241 int varRelid = PG_GETARG_INT32(3);
242 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 if (negate)
253 {
254 operator = get_negator(operator);
255 if (!OidIsValid(operator))
256 {
257 /* Use default selectivity (should we raise an error instead?) */
258 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 if (!get_restriction_variable(root, args, varRelid,
267 &vardata, &other, &varonleft))
268 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 if (IsA(other, Const))
276 selec = var_eq_const(&vardata, operator, collation,
277 ((Const *) other)->constvalue,
278 ((Const *) other)->constisnull,
279 varonleft, negate);
280 else
281 selec = var_eq_non_const(&vardata, operator, collation, other,
282 varonleft, negate);
283
284 ReleaseVariableStats(vardata);
285
286 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 */
294double
295var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
296 Datum constval, bool constisnull,
297 bool varonleft, bool negate)
298{
299 double selec;
300 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 if (constisnull)
309 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 if (HeapTupleIsValid(vardata->statsTuple))
316 {
317 Form_pg_statistic stats;
318
319 stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
320 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 if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
331 {
332 selec = 1.0 / vardata->rel->tuples;
333 }
334 else if (HeapTupleIsValid(vardata->statsTuple) &&
336 (opfuncoid = get_opcode(oproid))))
337 {
338 AttStatsSlot sslot;
339 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 if (get_attstatsslot(&sslot, vardata->statsTuple,
350 STATISTIC_KIND_MCV, InvalidOid,
352 {
353 LOCAL_FCINFO(fcinfo, 2);
354 FmgrInfo eqproc;
355
356 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 InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
365 NULL, NULL);
366 fcinfo->args[0].isnull = false;
367 fcinfo->args[1].isnull = false;
368 /* be careful to apply operator right way 'round */
369 if (varonleft)
370 fcinfo->args[1].value = constval;
371 else
372 fcinfo->args[0].value = constval;
373
374 for (i = 0; i < sslot.nvalues; i++)
375 {
376 Datum fresult;
377
378 if (varonleft)
379 fcinfo->args[0].value = sslot.values[i];
380 else
381 fcinfo->args[1].value = sslot.values[i];
382 fcinfo->isnull = false;
383 fresult = FunctionCallInvoke(fcinfo);
384 if (!fcinfo->isnull && DatumGetBool(fresult))
385 {
386 match = true;
387 break;
388 }
389 }
390 }
391 else
392 {
393 /* no most-common-value info available */
394 i = 0; /* keep compiler quiet */
395 }
396
397 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 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 double sumcommon = 0.0;
413 double otherdistinct;
414
415 for (i = 0; i < sslot.nnumbers; i++)
416 sumcommon += sslot.numbers[i];
417 selec = 1.0 - sumcommon - nullfrac;
418 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 otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
426 sslot.nnumbers;
427 if (otherdistinct > 1)
428 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 if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
435 selec = sslot.numbers[sslot.nnumbers - 1];
436 }
437
438 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 selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
448 }
449
450 /* now adjust if we wanted <> rather than = */
451 if (negate)
452 selec = 1.0 - selec - nullfrac;
453
454 /* result should be in range, but make sure... */
455 CLAMP_PROBABILITY(selec);
456
457 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 */
465double
466var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
467 Node *other,
468 bool varonleft, bool negate)
469{
470 double selec;
471 double nullfrac = 0.0;
472 bool isdefault;
473
474 /*
475 * Grab the nullfrac for use below.
476 */
477 if (HeapTupleIsValid(vardata->statsTuple))
478 {
479 Form_pg_statistic stats;
480
481 stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
482 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 if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
493 {
494 selec = 1.0 / vardata->rel->tuples;
495 }
496 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 selec = 1.0 - nullfrac;
512 ndistinct = get_variable_numdistinct(vardata, &isdefault);
513 if (ndistinct > 1)
514 selec /= ndistinct;
515
516 /*
517 * Cross-check: selectivity should never be estimated as more than the
518 * most common value's.
519 */
520 if (get_attstatsslot(&sslot, vardata->statsTuple,
521 STATISTIC_KIND_MCV, InvalidOid,
523 {
524 if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
525 selec = sslot.numbers[0];
526 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 selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
537 }
538
539 /* now adjust if we wanted <> rather than = */
540 if (negate)
541 selec = 1.0 - selec - nullfrac;
542
543 /* result should be in range, but make sure... */
544 CLAMP_PROBABILITY(selec);
545
546 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 */
556Datum
558{
559 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 */
579static double
580scalarineqsel(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 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 if (vardata->var && IsA(vardata->var, Var) &&
599 ((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 if (vardata->rel->pages == 0)
610 return 1.0;
611
612 itemptr = (ItemPointer) DatumGetPointer(constval);
613 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 density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
623
624 /* If target is the last page, use half the density. */
625 if (block >= vardata->rel->pages - 1)
626 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 if (density > 0.0)
637 {
639
640 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 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 if (iseq == isgt && vardata->rel->tuples >= 1.0)
658 selec -= (1.0 / vardata->rel->tuples);
659
660 /* Finally, reverse the selectivity for the ">", ">=" cases. */
661 if (isgt)
662 selec = 1.0 - selec;
663
664 CLAMP_PROBABILITY(selec);
665 return selec;
666 }
667
668 /* no stats available, so default result */
669 return DEFAULT_INEQ_SEL;
670 }
671 stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
672
673 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 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 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 selec = 1.0 - stats->stanullfrac - sumcommon;
699
700 if (hist_selec >= 0.0)
701 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 selec *= 0.5;
709 }
710
711 selec += mcv_selec;
712
713 /* result should be in range, but make sure... */
714 CLAMP_PROBABILITY(selec);
715
716 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 */
731double
732mcv_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 mcv_selec = 0.0;
742 sumcommon = 0.0;
743
744 if (HeapTupleIsValid(vardata->statsTuple) &&
745 statistic_proc_security_check(vardata, opproc->fn_oid) &&
746 get_attstatsslot(&sslot, vardata->statsTuple,
747 STATISTIC_KIND_MCV, InvalidOid,
749 {
750 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 InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
760 NULL, NULL);
761 fcinfo->args[0].isnull = false;
762 fcinfo->args[1].isnull = false;
763 /* be careful to apply operator right way 'round */
764 if (varonleft)
765 fcinfo->args[1].value = constval;
766 else
767 fcinfo->args[0].value = constval;
768
769 for (i = 0; i < sslot.nvalues; i++)
770 {
771 Datum fresult;
772
773 if (varonleft)
774 fcinfo->args[0].value = sslot.values[i];
775 else
776 fcinfo->args[1].value = sslot.values[i];
777 fcinfo->isnull = false;
778 fresult = FunctionCallInvoke(fcinfo);
779 if (!fcinfo->isnull && DatumGetBool(fresult))
780 mcv_selec += sslot.numbers[i];
781 sumcommon += sslot.numbers[i];
782 }
783 free_attstatsslot(&sslot);
784 }
785
786 *sumcommonp = sumcommon;
787 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 */
822double
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 if (HeapTupleIsValid(vardata->statsTuple) &&
837 statistic_proc_security_check(vardata, opproc->fn_oid) &&
838 get_attstatsslot(&sslot, vardata->statsTuple,
839 STATISTIC_KIND_HISTOGRAM, InvalidOid,
841 {
842 *hist_size = sslot.nvalues;
843 if (sslot.nvalues >= min_hist_size)
844 {
845 LOCAL_FCINFO(fcinfo, 2);
846 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 InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
858 NULL, NULL);
859 fcinfo->args[0].isnull = false;
860 fcinfo->args[1].isnull = false;
861 /* be careful to apply operator right way 'round */
862 if (varonleft)
863 fcinfo->args[1].value = constval;
864 else
865 fcinfo->args[0].value = constval;
866
867 for (i = n_skip; i < sslot.nvalues - n_skip; i++)
868 {
869 Datum fresult;
870
871 if (varonleft)
872 fcinfo->args[0].value = sslot.values[i];
873 else
874 fcinfo->args[1].value = sslot.values[i];
875 fcinfo->isnull = false;
876 fresult = FunctionCallInvoke(fcinfo);
877 if (!fcinfo->isnull && DatumGetBool(fresult))
878 nmatch++;
879 }
880 result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
881 }
882 else
883 result = -1;
884 free_attstatsslot(&sslot);
885 }
886 else
887 {
888 *hist_size = 0;
889 result = -1;
890 }
891
892 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 */
913double
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 if (!get_restriction_variable(root, args, varRelid,
928 &vardata, &other, &varonleft))
929 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 if (IsA(other, Const) &&
936 ((Const *) other)->constisnull)
937 {
938 ReleaseVariableStats(vardata);
939 return 0.0;
940 }
941
942 if (IsA(other, Const))
943 {
944 /* Variable is being compared to a known non-null constant */
945 Datum constval = ((Const *) other)->constvalue;
946 FmgrInfo opproc;
947 double mcvsum;
948 double mcvsel;
949 double nullfrac;
950 int hist_size;
951
952 fmgr_info(get_opcode(oproid), &opproc);
953
954 /*
955 * Calculate the selectivity for the column's most common values.
956 */
957 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 selec = histogram_selectivity(&vardata, &opproc, collation,
968 constval, varonleft,
969 10, 1, &hist_size);
970 if (selec < 0)
971 {
972 /* Nope, fall back on default */
973 selec = default_selectivity;
974 }
975 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 double hist_weight = hist_size / 100.0;
983
984 selec = selec * hist_weight +
985 default_selectivity * (1.0 - hist_weight);
986 }
987
988 /* In any case, don't believe extremely small or large estimates. */
989 if (selec < 0.0001)
990 selec = 0.0001;
991 else if (selec > 0.9999)
992 selec = 0.9999;
993
994 /* Don't forget to account for nulls. */
995 if (HeapTupleIsValid(vardata.statsTuple))
996 nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
997 else
998 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 selec *= 1.0 - nullfrac - mcvsum;
1006 selec += mcvsel;
1007 }
1008 else
1009 {
1010 /* Comparison value is not constant, so we can't do anything */
1011 selec = default_selectivity;
1012 }
1013
1014 ReleaseVariableStats(vardata);
1015
1016 /* result should be in range, but make sure... */
1017 CLAMP_PROBABILITY(selec);
1018
1019 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 */
1040double
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 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 if (HeapTupleIsValid(vardata->statsTuple) &&
1065 statistic_proc_security_check(vardata, opproc->fn_oid) &&
1066 get_attstatsslot(&sslot, vardata->statsTuple,
1067 STATISTIC_KIND_HISTOGRAM, InvalidOid,
1069 {
1070 if (sslot.nvalues > 1 &&
1071 sslot.stacoll == collation &&
1073 {
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 int lobound = 0; /* first possible slot to search */
1094 int hibound = sslot.nvalues; /* last+1 slot to search */
1095 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 if (sslot.nvalues == 2)
1105 vardata,
1106 sslot.staop,
1107 collation,
1108 &sslot.values[0],
1109 &sslot.values[1]);
1110
1111 while (lobound < hibound)
1112 {
1113 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 if (probe == 0 && sslot.nvalues > 2)
1123 vardata,
1124 sslot.staop,
1125 collation,
1126 &sslot.values[0],
1127 NULL);
1128 else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1130 vardata,
1131 sslot.staop,
1132 collation,
1133 NULL,
1134 &sslot.values[probe]);
1135
1136 ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1137 collation,
1138 sslot.values[probe],
1139 constval));
1140 if (isgt)
1141 ltcmp = !ltcmp;
1142 if (ltcmp)
1143 lobound = probe + 1;
1144 else
1145 hibound = probe;
1146 }
1147
1148 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 histfrac = 0.0;
1159 }
1160 else if (lobound >= sslot.nvalues)
1161 {
1162 /*
1163 * Inverse case: constant is above upper histogram boundary.
1164 */
1165 histfrac = 1.0;
1166 }
1167 else
1168 {
1169 /* We have values[i-1] <= constant <= values[i]. */
1170 int i = lobound;
1171 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 if (i == 1 || isgt == iseq)
1194 {
1195 double otherdistinct;
1196 bool isdefault;
1197 AttStatsSlot mcvslot;
1198
1199 /* Get estimated number of distinct values */
1200 otherdistinct = get_variable_numdistinct(vardata,
1201 &isdefault);
1202
1203 /* Subtract off the number of known MCVs */
1204 if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1205 STATISTIC_KIND_MCV, InvalidOid,
1207 {
1208 otherdistinct -= mcvslot.nnumbers;
1209 free_attstatsslot(&mcvslot);
1210 }
1211
1212 /* If result doesn't seem sane, leave eq_selec at 0 */
1213 if (otherdistinct > 1)
1214 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 if (convert_to_scalar(constval, consttype, collation,
1223 &val,
1224 sslot.values[i - 1], sslot.values[i],
1225 vardata->vartype,
1226 &low, &high))
1227 {
1228 if (high <= low)
1229 {
1230 /* cope if bin boundaries appear identical */
1231 binfrac = 0.5;
1232 }
1233 else if (val <= low)
1234 binfrac = 0.0;
1235 else if (val >= high)
1236 binfrac = 1.0;
1237 else
1238 {
1239 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 if (isnan(binfrac) ||
1248 binfrac < 0.0 || binfrac > 1.0)
1249 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 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 histfrac = (double) (i - 1) + binfrac;
1271 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 if (i == 1)
1306 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 if (isgt == iseq)
1314 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 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 if (have_end)
1333 CLAMP_PROBABILITY(hist_selec);
1334 else
1335 {
1336 double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1337
1338 if (hist_selec < cutoff)
1339 hist_selec = cutoff;
1340 else if (hist_selec > 1.0 - cutoff)
1341 hist_selec = 1.0 - cutoff;
1342 }
1343 }
1344 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 LOCAL_FCINFO(fcinfo, 2);
1354 int nmatch = 0;
1355
1356 InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1357 NULL, NULL);
1358 fcinfo->args[0].isnull = false;
1359 fcinfo->args[1].isnull = false;
1360 fcinfo->args[1].value = constval;
1361 for (int i = 0; i < sslot.nvalues; i++)
1362 {
1363 Datum fresult;
1364
1365 fcinfo->args[0].value = sslot.values[i];
1366 fcinfo->isnull = false;
1367 fresult = FunctionCallInvoke(fcinfo);
1368 if (!fcinfo->isnull && DatumGetBool(fresult))
1369 nmatch++;
1370 }
1371 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 double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1381
1382 if (hist_selec < cutoff)
1383 hist_selec = cutoff;
1384 else if (hist_selec > 1.0 - cutoff)
1385 hist_selec = 1.0 - cutoff;
1386 }
1387 }
1388
1389 free_attstatsslot(&sslot);
1390 }
1391
1392 return hist_selec;
1393}
1394
1395/*
1396 * Common wrapper function for the selectivity estimators that simply
1397 * invoke scalarineqsel().
1398 */
1399static Datum
1401{
1403 Oid operator = PG_GETARG_OID(1);
1404 List *args = (List *) PG_GETARG_POINTER(2);
1405 int varRelid = PG_GETARG_INT32(3);
1406 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 if (!get_restriction_variable(root, args, varRelid,
1419 &vardata, &other, &varonleft))
1421
1422 /*
1423 * Can't do anything useful if the something is not a constant, either.
1424 */
1425 if (!IsA(other, Const))
1426 {
1427 ReleaseVariableStats(vardata);
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 if (((Const *) other)->constisnull)
1436 {
1437 ReleaseVariableStats(vardata);
1438 PG_RETURN_FLOAT8(0.0);
1439 }
1440 constval = ((Const *) other)->constvalue;
1441 consttype = ((Const *) other)->consttype;
1442
1443 /*
1444 * Force the var to be on the left to simplify logic in scalarineqsel.
1445 */
1446 if (!varonleft)
1447 {
1448 operator = get_commutator(operator);
1449 if (!operator)
1450 {
1451 /* Use default selectivity (should we raise an error instead?) */
1452 ReleaseVariableStats(vardata);
1454 }
1455 isgt = !isgt;
1456 }
1457
1458 /* The rest of the work is done by scalarineqsel(). */
1459 selec = scalarineqsel(root, operator, isgt, iseq, collation,
1460 &vardata, constval, consttype);
1461
1462 ReleaseVariableStats(vardata);
1463
1464 PG_RETURN_FLOAT8((float8) selec);
1465}
1466
1467/*
1468 * scalarltsel - Selectivity of "<" for scalars.
1469 */
1470Datum
1472{
1473 return scalarineqsel_wrapper(fcinfo, false, false);
1474}
1475
1476/*
1477 * scalarlesel - Selectivity of "<=" for scalars.
1478 */
1479Datum
1481{
1482 return scalarineqsel_wrapper(fcinfo, false, true);
1483}
1484
1485/*
1486 * scalargtsel - Selectivity of ">" for scalars.
1487 */
1488Datum
1490{
1491 return scalarineqsel_wrapper(fcinfo, true, false);
1492}
1493
1494/*
1495 * scalargesel - Selectivity of ">=" for scalars.
1496 */
1497Datum
1499{
1500 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 */
1513{
1514 VariableStatData vardata;
1515 double selec;
1516
1517 examine_variable(root, arg, varRelid, &vardata);
1518 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 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 selec = 0.5;
1531 }
1532 ReleaseVariableStats(vardata);
1533 return selec;
1534}
1535
1536/*
1537 * booltestsel - Selectivity of BooleanTest Node.
1538 */
1541 int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1542{
1543 VariableStatData vardata;
1544 double selec;
1545
1546 examine_variable(root, arg, varRelid, &vardata);
1547
1548 if (HeapTupleIsValid(vardata.statsTuple))
1549 {
1550 Form_pg_statistic stats;
1551 double freq_null;
1552 AttStatsSlot sslot;
1553
1554 stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1555 freq_null = stats->stanullfrac;
1556
1557 if (get_attstatsslot(&sslot, vardata.statsTuple,
1558 STATISTIC_KIND_MCV, InvalidOid,
1560 && sslot.nnumbers > 0)
1561 {
1562 double freq_true;
1563 double freq_false;
1564
1565 /*
1566 * Get first MCV frequency and derive frequency for true.
1567 */
1568 if (DatumGetBool(sslot.values[0]))
1569 freq_true = sslot.numbers[0];
1570 else
1571 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 freq_false = 1.0 - freq_true - freq_null;
1578
1579 switch (booltesttype)
1580 {
1581 case IS_UNKNOWN:
1582 /* select only NULL values */
1583 selec = freq_null;
1584 break;
1585 case IS_NOT_UNKNOWN:
1586 /* select non-NULL values */
1587 selec = 1.0 - freq_null;
1588 break;
1589 case IS_TRUE:
1590 /* select only TRUE values */
1591 selec = freq_true;
1592 break;
1593 case IS_NOT_TRUE:
1594 /* select non-TRUE values */
1595 selec = 1.0 - freq_true;
1596 break;
1597 case IS_FALSE:
1598 /* select only FALSE values */
1599 selec = freq_false;
1600 break;
1601 case IS_NOT_FALSE:
1602 /* select non-FALSE values */
1603 selec = 1.0 - freq_false;
1604 break;
1605 default:
1606 elog(ERROR, "unrecognized booltesttype: %d",
1607 (int) booltesttype);
1608 selec = 0.0; /* Keep compiler quiet */
1609 break;
1610 }
1611
1612 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 switch (booltesttype)
1622 {
1623 case IS_UNKNOWN:
1624 /* select only NULL values */
1625 selec = freq_null;
1626 break;
1627 case IS_NOT_UNKNOWN:
1628 /* select non-NULL values */
1629 selec = 1.0 - freq_null;
1630 break;
1631 case IS_TRUE:
1632 case IS_FALSE:
1633 /* Assume we select half of the non-NULL values */
1634 selec = (1.0 - freq_null) / 2.0;
1635 break;
1636 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 selec = (freq_null + 1.0) / 2.0;
1641 break;
1642 default:
1643 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 switch (booltesttype)
1659 {
1660 case IS_UNKNOWN:
1661 selec = DEFAULT_UNK_SEL;
1662 break;
1663 case IS_NOT_UNKNOWN:
1664 selec = DEFAULT_NOT_UNK_SEL;
1665 break;
1666 case IS_TRUE:
1667 case IS_NOT_FALSE:
1668 selec = (double) clause_selectivity(root, arg,
1669 varRelid,
1670 jointype, sjinfo);
1671 break;
1672 case IS_FALSE:
1673 case IS_NOT_TRUE:
1674 selec = 1.0 - (double) clause_selectivity(root, arg,
1675 varRelid,
1676 jointype, sjinfo);
1677 break;
1678 default:
1679 elog(ERROR, "unrecognized booltesttype: %d",
1680 (int) booltesttype);
1681 selec = 0.0; /* Keep compiler quiet */
1682 break;
1683 }
1684 }
1685
1686 ReleaseVariableStats(vardata);
1687
1688 /* result should be in range, but make sure... */
1689 CLAMP_PROBABILITY(selec);
1690
1691 return (Selectivity) selec;
1692}
1693
1694/*
1695 * nulltestsel - Selectivity of NullTest Node.
1696 */
1699 int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1700{
1701 VariableStatData vardata;
1702 double selec;
1703
1704 examine_variable(root, arg, varRelid, &vardata);
1705
1706 if (HeapTupleIsValid(vardata.statsTuple))
1707 {
1708 Form_pg_statistic stats;
1709 double freq_null;
1710
1711 stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1712 freq_null = stats->stanullfrac;
1713
1714 switch (nulltesttype)
1715 {
1716 case IS_NULL:
1717
1718 /*
1719 * Use freq_null directly.
1720 */
1721 selec = freq_null;
1722 break;
1723 case IS_NOT_NULL:
1724
1725 /*
1726 * Select not unknown (not null) values. Calculate from
1727 * freq_null.
1728 */
1729 selec = 1.0 - freq_null;
1730 break;
1731 default:
1732 elog(ERROR, "unrecognized nulltesttype: %d",
1733 (int) nulltesttype);
1734 return (Selectivity) 0; /* keep compiler quiet */
1735 }
1736 }
1737 else if (vardata.var && IsA(vardata.var, Var) &&
1738 ((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 selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1745 }
1746 else
1747 {
1748 /*
1749 * No ANALYZE stats available, so make a guess
1750 */
1751 switch (nulltesttype)
1752 {
1753 case IS_NULL:
1754 selec = DEFAULT_UNK_SEL;
1755 break;
1756 case IS_NOT_NULL:
1757 selec = DEFAULT_NOT_UNK_SEL;
1758 break;
1759 default:
1760 elog(ERROR, "unrecognized nulltesttype: %d",
1761 (int) nulltesttype);
1762 return (Selectivity) 0; /* keep compiler quiet */
1763 }
1764 }
1765
1766 ReleaseVariableStats(vardata);
1767
1768 /* result should be in range, but make sure... */
1769 CLAMP_PROBABILITY(selec);
1770
1771 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 */
1782static Node *
1784{
1785 for (;;)
1786 {
1787 if (node && IsA(node, ArrayCoerceExpr))
1788 {
1789 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 if (IsA(acoerce->elemexpr, RelabelType) &&
1796 IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1797 node = (Node *) acoerce->arg;
1798 else
1799 break;
1800 }
1801 else if (node && IsA(node, RelabelType))
1802 {
1803 /* We don't really expect this case, but may as well cope */
1804 node = (Node *) ((RelabelType *) node)->arg;
1805 }
1806 else
1807 break;
1808 }
1809 return node;
1810}
1811
1812/*
1813 * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1814 */
1817 ScalarArrayOpExpr *clause,
1818 bool is_join_clause,
1819 int varRelid,
1820 JoinType jointype,
1821 SpecialJoinInfo *sjinfo)
1822{
1823 Oid operator = clause->opno;
1824 bool useOr = clause->useOr;
1825 bool isEquality = false;
1826 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;
1835 Selectivity s1disjoint;
1836
1837 /* First, deconstruct the expression */
1838 Assert(list_length(clause->args) == 2);
1839 leftop = (Node *) linitial(clause->args);
1840 rightop = (Node *) lsecond(clause->args);
1841
1842 /* aggressively reduce both sides to constants */
1843 leftop = estimate_expression_value(root, leftop);
1844 rightop = estimate_expression_value(root, rightop);
1845
1846 /* get nominal (after relabeling) element type of rightop */
1847 nominal_element_type = get_base_element_type(exprType(rightop));
1848 if (!OidIsValid(nominal_element_type))
1849 return (Selectivity) 0.5; /* probably shouldn't happen */
1850 /* get nominal collation, too, for generating constants */
1851 nominal_element_collation = exprCollation(rightop);
1852
1853 /* look through any binary-compatible relabeling of rightop */
1854 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 typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1861 if (OidIsValid(typentry->eq_opr))
1862 {
1863 if (operator == typentry->eq_opr)
1864 isEquality = true;
1865 else if (get_negator(operator) == typentry->eq_opr)
1866 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 if ((isEquality || isInequality) && !is_join_clause)
1876 {
1877 s1 = scalararraysel_containment(root, leftop, rightop,
1878 nominal_element_type,
1879 isEquality, useOr, varRelid);
1880 if (s1 >= 0.0)
1881 return s1;
1882 }
1883
1884 /*
1885 * Look up the underlying operator's selectivity estimator. Punt if it
1886 * hasn't got one.
1887 */
1888 if (is_join_clause)
1889 oprsel = get_oprjoin(operator);
1890 else
1891 oprsel = get_oprrest(operator);
1892 if (!oprsel)
1893 return (Selectivity) 0.5;
1894 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 if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1905 isEquality = true;
1906 else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1907 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 if (rightop && IsA(rightop, Const))
1922 {
1923 Datum arraydatum = ((Const *) rightop)->constvalue;
1924 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 if (arrayisnull) /* qual can't succeed if null array */
1935 return (Selectivity) 0.0;
1936 arrayval = DatumGetArrayTypeP(arraydatum);
1938 &elmlen, &elmbyval, &elmalign);
1939 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 s1 = s1disjoint = (useOr ? 0.0 : 1.0);
1959
1960 for (i = 0; i < num_elems; i++)
1961 {
1962 List *args;
1964
1965 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 if (is_join_clause)
1974 s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1975 clause->inputcollid,
1977 ObjectIdGetDatum(operator),
1979 Int16GetDatum(jointype),
1980 PointerGetDatum(sjinfo)));
1981 else
1982 s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1983 clause->inputcollid,
1985 ObjectIdGetDatum(operator),
1987 Int32GetDatum(varRelid)));
1988
1989 if (useOr)
1990 {
1991 s1 = s1 + s2 - s1 * s2;
1992 if (isEquality)
1993 s1disjoint += s2;
1994 }
1995 else
1996 {
1997 s1 = s1 * s2;
1998 if (isInequality)
1999 s1disjoint += s2 - 1.0;
2000 }
2001 }
2002
2003 /* accept disjoint-probability estimate if in range */
2004 if ((useOr ? isEquality : isInequality) &&
2005 s1disjoint >= 0.0 && s1disjoint <= 1.0)
2006 s1 = s1disjoint;
2007 }
2008 else if (rightop && IsA(rightop, ArrayExpr) &&
2009 !((ArrayExpr *) rightop)->multidims)
2010 {
2011 ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2012 int16 elmlen;
2013 bool elmbyval;
2014 ListCell *l;
2015
2016 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 s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2027
2028 foreach(l, arrayexpr->elements)
2029 {
2030 Node *elem = (Node *) lfirst(l);
2031 List *args;
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 args = list_make2(leftop, elem);
2040 if (is_join_clause)
2041 s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2042 clause->inputcollid,
2044 ObjectIdGetDatum(operator),
2046 Int16GetDatum(jointype),
2047 PointerGetDatum(sjinfo)));
2048 else
2049 s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2050 clause->inputcollid,
2052 ObjectIdGetDatum(operator),
2054 Int32GetDatum(varRelid)));
2055
2056 if (useOr)
2057 {
2058 s1 = s1 + s2 - s1 * s2;
2059 if (isEquality)
2060 s1disjoint += s2;
2061 }
2062 else
2063 {
2064 s1 = s1 * s2;
2065 if (isInequality)
2066 s1disjoint += s2 - 1.0;
2067 }
2068 }
2069
2070 /* accept disjoint-probability estimate if in range */
2071 if ((useOr ? isEquality : isInequality) &&
2072 s1disjoint >= 0.0 && s1disjoint <= 1.0)
2073 s1 = s1disjoint;
2074 }
2075 else
2076 {
2077 CaseTestExpr *dummyexpr;
2078 List *args;
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 dummyexpr = makeNode(CaseTestExpr);
2088 dummyexpr->typeId = nominal_element_type;
2089 dummyexpr->typeMod = -1;
2090 dummyexpr->collation = clause->inputcollid;
2091 args = list_make2(leftop, dummyexpr);
2092 if (is_join_clause)
2093 s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2094 clause->inputcollid,
2096 ObjectIdGetDatum(operator),
2098 Int16GetDatum(jointype),
2099 PointerGetDatum(sjinfo)));
2100 else
2101 s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2102 clause->inputcollid,
2104 ObjectIdGetDatum(operator),
2106 Int32GetDatum(varRelid)));
2107 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 for (i = 0; i < 10; i++)
2115 {
2116 if (useOr)
2117 s1 = s1 + s2 - s1 * s2;
2118 else
2119 s1 = s1 * s2;
2120 }
2121 }
2122
2123 /* result should be in range, but make sure... */
2125
2126 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 * Note: in some code paths root can be passed as NULL, resulting in
2136 * slightly worse estimates.
2137 */
2138double
2140{
2141 /* look through any binary-compatible relabeling of arrayexpr */
2142 arrayexpr = strip_array_coercion(arrayexpr);
2143
2144 if (arrayexpr && IsA(arrayexpr, Const))
2145 {
2146 Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2147 bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2148 ArrayType *arrayval;
2149
2150 if (arrayisnull)
2151 return 0;
2152 arrayval = DatumGetArrayTypeP(arraydatum);
2153 return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2154 }
2155 else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2156 !((ArrayExpr *) arrayexpr)->multidims)
2157 {
2158 return list_length(((ArrayExpr *) arrayexpr)->elements);
2159 }
2160 else if (arrayexpr && root)
2161 {
2162 /* See if we can find any statistics about it */
2163 VariableStatData vardata;
2164 AttStatsSlot sslot;
2165 double nelem = 0;
2166
2167 examine_variable(root, arrayexpr, 0, &vardata);
2168 if (HeapTupleIsValid(vardata.statsTuple))
2169 {
2170 /*
2171 * Found stats, so use the average element count, which is stored
2172 * in the last stanumbers element of the DECHIST statistics.
2173 * Actually that is the average count of *distinct* elements;
2174 * perhaps we should scale it up somewhat?
2175 */
2176 if (get_attstatsslot(&sslot, vardata.statsTuple,
2177 STATISTIC_KIND_DECHIST, InvalidOid,
2179 {
2180 if (sslot.nnumbers > 0)
2181 nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2182 free_attstatsslot(&sslot);
2183 }
2184 }
2185 ReleaseVariableStats(vardata);
2186
2187 if (nelem > 0)
2188 return nelem;
2189 }
2190
2191 /* Else use a default guess --- this should match scalararraysel */
2192 return 10;
2193}
2194
2195/*
2196 * rowcomparesel - Selectivity of RowCompareExpr Node.
2197 *
2198 * We estimate RowCompare selectivity by considering just the first (high
2199 * order) columns, which makes it equivalent to an ordinary OpExpr. While
2200 * this estimate could be refined by considering additional columns, it
2201 * seems unlikely that we could do a lot better without multi-column
2202 * statistics.
2203 */
2206 RowCompareExpr *clause,
2207 int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2208{
2210 Oid opno = linitial_oid(clause->opnos);
2211 Oid inputcollid = linitial_oid(clause->inputcollids);
2212 List *opargs;
2213 bool is_join_clause;
2214
2215 /* Build equivalent arg list for single operator */
2216 opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2217
2218 /*
2219 * Decide if it's a join clause. This should match clausesel.c's
2220 * treat_as_join_clause(), except that we intentionally consider only the
2221 * leading columns and not the rest of the clause.
2222 */
2223 if (varRelid != 0)
2224 {
2225 /*
2226 * Caller is forcing restriction mode (eg, because we are examining an
2227 * inner indexscan qual).
2228 */
2229 is_join_clause = false;
2230 }
2231 else if (sjinfo == NULL)
2232 {
2233 /*
2234 * It must be a restriction clause, since it's being evaluated at a
2235 * scan node.
2236 */
2237 is_join_clause = false;
2238 }
2239 else
2240 {
2241 /*
2242 * Otherwise, it's a join if there's more than one base relation used.
2243 */
2244 is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2245 }
2246
2247 if (is_join_clause)
2248 {
2249 /* Estimate selectivity for a join clause. */
2250 s1 = join_selectivity(root, opno,
2251 opargs,
2252 inputcollid,
2253 jointype,
2254 sjinfo);
2255 }
2256 else
2257 {
2258 /* Estimate selectivity for a restriction clause. */
2260 opargs,
2261 inputcollid,
2262 varRelid);
2263 }
2264
2265 return s1;
2266}
2267
2268/*
2269 * eqjoinsel - Join selectivity of "="
2270 */
2271Datum
2273{
2275 Oid operator = PG_GETARG_OID(1);
2276 List *args = (List *) PG_GETARG_POINTER(2);
2277
2278#ifdef NOT_USED
2279 JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2280#endif
2282 Oid collation = PG_GET_COLLATION();
2283 double selec;
2284 double selec_inner;
2285 VariableStatData vardata1;
2286 VariableStatData vardata2;
2287 double nd1;
2288 double nd2;
2289 bool isdefault1;
2290 bool isdefault2;
2291 Oid opfuncoid;
2292 AttStatsSlot sslot1;
2293 AttStatsSlot sslot2;
2294 Form_pg_statistic stats1 = NULL;
2295 Form_pg_statistic stats2 = NULL;
2296 bool have_mcvs1 = false;
2297 bool have_mcvs2 = false;
2298 bool get_mcv_stats;
2299 bool join_is_reversed;
2300 RelOptInfo *inner_rel;
2301
2302 get_join_variables(root, args, sjinfo,
2303 &vardata1, &vardata2, &join_is_reversed);
2304
2305 nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2306 nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2307
2308 opfuncoid = get_opcode(operator);
2309
2310 memset(&sslot1, 0, sizeof(sslot1));
2311 memset(&sslot2, 0, sizeof(sslot2));
2312
2313 /*
2314 * There is no use in fetching one side's MCVs if we lack MCVs for the
2315 * other side, so do a quick check to verify that both stats exist.
2316 */
2317 get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2318 HeapTupleIsValid(vardata2.statsTuple) &&
2319 get_attstatsslot(&sslot1, vardata1.statsTuple,
2320 STATISTIC_KIND_MCV, InvalidOid,
2321 0) &&
2322 get_attstatsslot(&sslot2, vardata2.statsTuple,
2323 STATISTIC_KIND_MCV, InvalidOid,
2324 0));
2325
2326 if (HeapTupleIsValid(vardata1.statsTuple))
2327 {
2328 /* note we allow use of nullfrac regardless of security check */
2329 stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2330 if (get_mcv_stats &&
2331 statistic_proc_security_check(&vardata1, opfuncoid))
2332 have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2333 STATISTIC_KIND_MCV, InvalidOid,
2335 }
2336
2337 if (HeapTupleIsValid(vardata2.statsTuple))
2338 {
2339 /* note we allow use of nullfrac regardless of security check */
2340 stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2341 if (get_mcv_stats &&
2342 statistic_proc_security_check(&vardata2, opfuncoid))
2343 have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2344 STATISTIC_KIND_MCV, InvalidOid,
2346 }
2347
2348 /* We need to compute the inner-join selectivity in all cases */
2349 selec_inner = eqjoinsel_inner(opfuncoid, collation,
2350 &vardata1, &vardata2,
2351 nd1, nd2,
2352 isdefault1, isdefault2,
2353 &sslot1, &sslot2,
2354 stats1, stats2,
2355 have_mcvs1, have_mcvs2);
2356
2357 switch (sjinfo->jointype)
2358 {
2359 case JOIN_INNER:
2360 case JOIN_LEFT:
2361 case JOIN_FULL:
2362 selec = selec_inner;
2363 break;
2364 case JOIN_SEMI:
2365 case JOIN_ANTI:
2366
2367 /*
2368 * Look up the join's inner relation. min_righthand is sufficient
2369 * information because neither SEMI nor ANTI joins permit any
2370 * reassociation into or out of their RHS, so the righthand will
2371 * always be exactly that set of rels.
2372 */
2373 inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2374
2375 if (!join_is_reversed)
2376 selec = eqjoinsel_semi(opfuncoid, collation,
2377 &vardata1, &vardata2,
2378 nd1, nd2,
2379 isdefault1, isdefault2,
2380 &sslot1, &sslot2,
2381 stats1, stats2,
2382 have_mcvs1, have_mcvs2,
2383 inner_rel);
2384 else
2385 {
2386 Oid commop = get_commutator(operator);
2387 Oid commopfuncoid = OidIsValid(commop) ? get_opcode(commop) : InvalidOid;
2388
2389 selec = eqjoinsel_semi(commopfuncoid, collation,
2390 &vardata2, &vardata1,
2391 nd2, nd1,
2392 isdefault2, isdefault1,
2393 &sslot2, &sslot1,
2394 stats2, stats1,
2395 have_mcvs2, have_mcvs1,
2396 inner_rel);
2397 }
2398
2399 /*
2400 * We should never estimate the output of a semijoin to be more
2401 * rows than we estimate for an inner join with the same input
2402 * rels and join condition; it's obviously impossible for that to
2403 * happen. The former estimate is N1 * Ssemi while the latter is
2404 * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2405 * this is worthwhile because of the shakier estimation rules we
2406 * use in eqjoinsel_semi, particularly in cases where it has to
2407 * punt entirely.
2408 */
2409 selec = Min(selec, inner_rel->rows * selec_inner);
2410 break;
2411 default:
2412 /* other values not expected here */
2413 elog(ERROR, "unrecognized join type: %d",
2414 (int) sjinfo->jointype);
2415 selec = 0; /* keep compiler quiet */
2416 break;
2417 }
2418
2419 free_attstatsslot(&sslot1);
2420 free_attstatsslot(&sslot2);
2421
2422 ReleaseVariableStats(vardata1);
2423 ReleaseVariableStats(vardata2);
2424
2425 CLAMP_PROBABILITY(selec);
2426
2427 PG_RETURN_FLOAT8((float8) selec);
2428}
2429
2430/*
2431 * eqjoinsel_inner --- eqjoinsel for normal inner join
2432 *
2433 * We also use this for LEFT/FULL outer joins; it's not presently clear
2434 * that it's worth trying to distinguish them here.
2435 */
2436static double
2437eqjoinsel_inner(Oid opfuncoid, Oid collation,
2438 VariableStatData *vardata1, VariableStatData *vardata2,
2439 double nd1, double nd2,
2440 bool isdefault1, bool isdefault2,
2441 AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2442 Form_pg_statistic stats1, Form_pg_statistic stats2,
2443 bool have_mcvs1, bool have_mcvs2)
2444{
2445 double selec;
2446
2447 if (have_mcvs1 && have_mcvs2)
2448 {
2449 /*
2450 * We have most-common-value lists for both relations. Run through
2451 * the lists to see which MCVs actually join to each other with the
2452 * given operator. This allows us to determine the exact join
2453 * selectivity for the portion of the relations represented by the MCV
2454 * lists. We still have to estimate for the remaining population, but
2455 * in a skewed distribution this gives us a big leg up in accuracy.
2456 * For motivation see the analysis in Y. Ioannidis and S.
2457 * Christodoulakis, "On the propagation of errors in the size of join
2458 * results", Technical Report 1018, Computer Science Dept., University
2459 * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2460 */
2461 LOCAL_FCINFO(fcinfo, 2);
2462 FmgrInfo eqproc;
2463 bool *hasmatch1;
2464 bool *hasmatch2;
2465 double nullfrac1 = stats1->stanullfrac;
2466 double nullfrac2 = stats2->stanullfrac;
2467 double matchprodfreq,
2468 matchfreq1,
2469 matchfreq2,
2470 unmatchfreq1,
2471 unmatchfreq2,
2472 otherfreq1,
2473 otherfreq2,
2474 totalsel1,
2475 totalsel2;
2476 int i,
2477 nmatches;
2478
2479 fmgr_info(opfuncoid, &eqproc);
2480
2481 /*
2482 * Save a few cycles by setting up the fcinfo struct just once. Using
2483 * FunctionCallInvoke directly also avoids failure if the eqproc
2484 * returns NULL, though really equality functions should never do
2485 * that.
2486 */
2487 InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2488 NULL, NULL);
2489 fcinfo->args[0].isnull = false;
2490 fcinfo->args[1].isnull = false;
2491
2492 hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2493 hasmatch2 = (bool *) palloc0(sslot2->nvalues * sizeof(bool));
2494
2495 /*
2496 * Note we assume that each MCV will match at most one member of the
2497 * other MCV list. If the operator isn't really equality, there could
2498 * be multiple matches --- but we don't look for them, both for speed
2499 * and because the math wouldn't add up...
2500 */
2501 matchprodfreq = 0.0;
2502 nmatches = 0;
2503 for (i = 0; i < sslot1->nvalues; i++)
2504 {
2505 int j;
2506
2507 fcinfo->args[0].value = sslot1->values[i];
2508
2509 for (j = 0; j < sslot2->nvalues; j++)
2510 {
2511 Datum fresult;
2512
2513 if (hasmatch2[j])
2514 continue;
2515 fcinfo->args[1].value = sslot2->values[j];
2516 fcinfo->isnull = false;
2517 fresult = FunctionCallInvoke(fcinfo);
2518 if (!fcinfo->isnull && DatumGetBool(fresult))
2519 {
2520 hasmatch1[i] = hasmatch2[j] = true;
2521 matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
2522 nmatches++;
2523 break;
2524 }
2525 }
2526 }
2527 CLAMP_PROBABILITY(matchprodfreq);
2528 /* Sum up frequencies of matched and unmatched MCVs */
2529 matchfreq1 = unmatchfreq1 = 0.0;
2530 for (i = 0; i < sslot1->nvalues; i++)
2531 {
2532 if (hasmatch1[i])
2533 matchfreq1 += sslot1->numbers[i];
2534 else
2535 unmatchfreq1 += sslot1->numbers[i];
2536 }
2537 CLAMP_PROBABILITY(matchfreq1);
2538 CLAMP_PROBABILITY(unmatchfreq1);
2539 matchfreq2 = unmatchfreq2 = 0.0;
2540 for (i = 0; i < sslot2->nvalues; i++)
2541 {
2542 if (hasmatch2[i])
2543 matchfreq2 += sslot2->numbers[i];
2544 else
2545 unmatchfreq2 += sslot2->numbers[i];
2546 }
2547 CLAMP_PROBABILITY(matchfreq2);
2548 CLAMP_PROBABILITY(unmatchfreq2);
2549 pfree(hasmatch1);
2550 pfree(hasmatch2);
2551
2552 /*
2553 * Compute total frequency of non-null values that are not in the MCV
2554 * lists.
2555 */
2556 otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2557 otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2558 CLAMP_PROBABILITY(otherfreq1);
2559 CLAMP_PROBABILITY(otherfreq2);
2560
2561 /*
2562 * We can estimate the total selectivity from the point of view of
2563 * relation 1 as: the known selectivity for matched MCVs, plus
2564 * unmatched MCVs that are assumed to match against random members of
2565 * relation 2's non-MCV population, plus non-MCV values that are
2566 * assumed to match against random members of relation 2's unmatched
2567 * MCVs plus non-MCV values.
2568 */
2569 totalsel1 = matchprodfreq;
2570 if (nd2 > sslot2->nvalues)
2571 totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2572 if (nd2 > nmatches)
2573 totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2574 (nd2 - nmatches);
2575 /* Same estimate from the point of view of relation 2. */
2576 totalsel2 = matchprodfreq;
2577 if (nd1 > sslot1->nvalues)
2578 totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2579 if (nd1 > nmatches)
2580 totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2581 (nd1 - nmatches);
2582
2583 /*
2584 * Use the smaller of the two estimates. This can be justified in
2585 * essentially the same terms as given below for the no-stats case: to
2586 * a first approximation, we are estimating from the point of view of
2587 * the relation with smaller nd.
2588 */
2589 selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2590 }
2591 else
2592 {
2593 /*
2594 * We do not have MCV lists for both sides. Estimate the join
2595 * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2596 * is plausible if we assume that the join operator is strict and the
2597 * non-null values are about equally distributed: a given non-null
2598 * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2599 * of rel2, so total join rows are at most
2600 * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2601 * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2602 * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2603 * with MIN() is an upper bound. Using the MIN() means we estimate
2604 * from the point of view of the relation with smaller nd (since the
2605 * larger nd is determining the MIN). It is reasonable to assume that
2606 * most tuples in this rel will have join partners, so the bound is
2607 * probably reasonably tight and should be taken as-is.
2608 *
2609 * XXX Can we be smarter if we have an MCV list for just one side? It
2610 * seems that if we assume equal distribution for the other side, we
2611 * end up with the same answer anyway.
2612 */
2613 double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2614 double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2615
2616 selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2617 if (nd1 > nd2)
2618 selec /= nd1;
2619 else
2620 selec /= nd2;
2621 }
2622
2623 return selec;
2624}
2625
2626/*
2627 * eqjoinsel_semi --- eqjoinsel for semi join
2628 *
2629 * (Also used for anti join, which we are supposed to estimate the same way.)
2630 * Caller has ensured that vardata1 is the LHS variable.
2631 * Unlike eqjoinsel_inner, we have to cope with opfuncoid being InvalidOid.
2632 */
2633static double
2634eqjoinsel_semi(Oid opfuncoid, Oid collation,
2635 VariableStatData *vardata1, VariableStatData *vardata2,
2636 double nd1, double nd2,
2637 bool isdefault1, bool isdefault2,
2638 AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2639 Form_pg_statistic stats1, Form_pg_statistic stats2,
2640 bool have_mcvs1, bool have_mcvs2,
2641 RelOptInfo *inner_rel)
2642{
2643 double selec;
2644
2645 /*
2646 * We clamp nd2 to be not more than what we estimate the inner relation's
2647 * size to be. This is intuitively somewhat reasonable since obviously
2648 * there can't be more than that many distinct values coming from the
2649 * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2650 * likewise) is that this is the only pathway by which restriction clauses
2651 * applied to the inner rel will affect the join result size estimate,
2652 * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2653 * only the outer rel's size. If we clamped nd1 we'd be double-counting
2654 * the selectivity of outer-rel restrictions.
2655 *
2656 * We can apply this clamping both with respect to the base relation from
2657 * which the join variable comes (if there is just one), and to the
2658 * immediate inner input relation of the current join.
2659 *
2660 * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2661 * great, maybe, but it didn't come out of nowhere either. This is most
2662 * helpful when the inner relation is empty and consequently has no stats.
2663 */
2664 if (vardata2->rel)
2665 {
2666 if (nd2 >= vardata2->rel->rows)
2667 {
2668 nd2 = vardata2->rel->rows;
2669 isdefault2 = false;
2670 }
2671 }
2672 if (nd2 >= inner_rel->rows)
2673 {
2674 nd2 = inner_rel->rows;
2675 isdefault2 = false;
2676 }
2677
2678 if (have_mcvs1 && have_mcvs2 && OidIsValid(opfuncoid))
2679 {
2680 /*
2681 * We have most-common-value lists for both relations. Run through
2682 * the lists to see which MCVs actually join to each other with the
2683 * given operator. This allows us to determine the exact join
2684 * selectivity for the portion of the relations represented by the MCV
2685 * lists. We still have to estimate for the remaining population, but
2686 * in a skewed distribution this gives us a big leg up in accuracy.
2687 */
2688 LOCAL_FCINFO(fcinfo, 2);
2689 FmgrInfo eqproc;
2690 bool *hasmatch1;
2691 bool *hasmatch2;
2692 double nullfrac1 = stats1->stanullfrac;
2693 double matchfreq1,
2694 uncertainfrac,
2695 uncertain;
2696 int i,
2697 nmatches,
2698 clamped_nvalues2;
2699
2700 /*
2701 * The clamping above could have resulted in nd2 being less than
2702 * sslot2->nvalues; in which case, we assume that precisely the nd2
2703 * most common values in the relation will appear in the join input,
2704 * and so compare to only the first nd2 members of the MCV list. Of
2705 * course this is frequently wrong, but it's the best bet we can make.
2706 */
2707 clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2708
2709 fmgr_info(opfuncoid, &eqproc);
2710
2711 /*
2712 * Save a few cycles by setting up the fcinfo struct just once. Using
2713 * FunctionCallInvoke directly also avoids failure if the eqproc
2714 * returns NULL, though really equality functions should never do
2715 * that.
2716 */
2717 InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2718 NULL, NULL);
2719 fcinfo->args[0].isnull = false;
2720 fcinfo->args[1].isnull = false;
2721
2722 hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2723 hasmatch2 = (bool *) palloc0(clamped_nvalues2 * sizeof(bool));
2724
2725 /*
2726 * Note we assume that each MCV will match at most one member of the
2727 * other MCV list. If the operator isn't really equality, there could
2728 * be multiple matches --- but we don't look for them, both for speed
2729 * and because the math wouldn't add up...
2730 */
2731 nmatches = 0;
2732 for (i = 0; i < sslot1->nvalues; i++)
2733 {
2734 int j;
2735
2736 fcinfo->args[0].value = sslot1->values[i];
2737
2738 for (j = 0; j < clamped_nvalues2; j++)
2739 {
2740 Datum fresult;
2741
2742 if (hasmatch2[j])
2743 continue;
2744 fcinfo->args[1].value = sslot2->values[j];
2745 fcinfo->isnull = false;
2746 fresult = FunctionCallInvoke(fcinfo);
2747 if (!fcinfo->isnull && DatumGetBool(fresult))
2748 {
2749 hasmatch1[i] = hasmatch2[j] = true;
2750 nmatches++;
2751 break;
2752 }
2753 }
2754 }
2755 /* Sum up frequencies of matched MCVs */
2756 matchfreq1 = 0.0;
2757 for (i = 0; i < sslot1->nvalues; i++)
2758 {
2759 if (hasmatch1[i])
2760 matchfreq1 += sslot1->numbers[i];
2761 }
2762 CLAMP_PROBABILITY(matchfreq1);
2763 pfree(hasmatch1);
2764 pfree(hasmatch2);
2765
2766 /*
2767 * Now we need to estimate the fraction of relation 1 that has at
2768 * least one join partner. We know for certain that the matched MCVs
2769 * do, so that gives us a lower bound, but we're really in the dark
2770 * about everything else. Our crude approach is: if nd1 <= nd2 then
2771 * assume all non-null rel1 rows have join partners, else assume for
2772 * the uncertain rows that a fraction nd2/nd1 have join partners. We
2773 * can discount the known-matched MCVs from the distinct-values counts
2774 * before doing the division.
2775 *
2776 * Crude as the above is, it's completely useless if we don't have
2777 * reliable ndistinct values for both sides. Hence, if either nd1 or
2778 * nd2 is default, punt and assume half of the uncertain rows have
2779 * join partners.
2780 */
2781 if (!isdefault1 && !isdefault2)
2782 {
2783 nd1 -= nmatches;
2784 nd2 -= nmatches;
2785 if (nd1 <= nd2 || nd2 < 0)
2786 uncertainfrac = 1.0;
2787 else
2788 uncertainfrac = nd2 / nd1;
2789 }
2790 else
2791 uncertainfrac = 0.5;
2792 uncertain = 1.0 - matchfreq1 - nullfrac1;
2793 CLAMP_PROBABILITY(uncertain);
2794 selec = matchfreq1 + uncertainfrac * uncertain;
2795 }
2796 else
2797 {
2798 /*
2799 * Without MCV lists for both sides, we can only use the heuristic
2800 * about nd1 vs nd2.
2801 */
2802 double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2803
2804 if (!isdefault1 && !isdefault2)
2805 {
2806 if (nd1 <= nd2 || nd2 < 0)
2807 selec = 1.0 - nullfrac1;
2808 else
2809 selec = (nd2 / nd1) * (1.0 - nullfrac1);
2810 }
2811 else
2812 selec = 0.5 * (1.0 - nullfrac1);
2813 }
2814
2815 return selec;
2816}
2817
2818/*
2819 * neqjoinsel - Join selectivity of "!="
2820 */
2821Datum
2823{
2825 Oid operator = PG_GETARG_OID(1);
2826 List *args = (List *) PG_GETARG_POINTER(2);
2827 JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2829 Oid collation = PG_GET_COLLATION();
2830 float8 result;
2831
2832 if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
2833 {
2834 /*
2835 * For semi-joins, if there is more than one distinct value in the RHS
2836 * relation then every non-null LHS row must find a row to join since
2837 * it can only be equal to one of them. We'll assume that there is
2838 * always more than one distinct RHS value for the sake of stability,
2839 * though in theory we could have special cases for empty RHS
2840 * (selectivity = 0) and single-distinct-value RHS (selectivity =
2841 * fraction of LHS that has the same value as the single RHS value).
2842 *
2843 * For anti-joins, if we use the same assumption that there is more
2844 * than one distinct key in the RHS relation, then every non-null LHS
2845 * row must be suppressed by the anti-join.
2846 *
2847 * So either way, the selectivity estimate should be 1 - nullfrac.
2848 */
2849 VariableStatData leftvar;
2850 VariableStatData rightvar;
2851 bool reversed;
2852 HeapTuple statsTuple;
2853 double nullfrac;
2854
2855 get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
2856 statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
2857 if (HeapTupleIsValid(statsTuple))
2858 nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
2859 else
2860 nullfrac = 0.0;
2861 ReleaseVariableStats(leftvar);
2862 ReleaseVariableStats(rightvar);
2863
2864 result = 1.0 - nullfrac;
2865 }
2866 else
2867 {
2868 /*
2869 * We want 1 - eqjoinsel() where the equality operator is the one
2870 * associated with this != operator, that is, its negator.
2871 */
2872 Oid eqop = get_negator(operator);
2873
2874 if (eqop)
2875 {
2876 result =
2878 collation,
2880 ObjectIdGetDatum(eqop),
2882 Int16GetDatum(jointype),
2883 PointerGetDatum(sjinfo)));
2884 }
2885 else
2886 {
2887 /* Use default selectivity (should we raise an error instead?) */
2888 result = DEFAULT_EQ_SEL;
2889 }
2890 result = 1.0 - result;
2891 }
2892
2893 PG_RETURN_FLOAT8(result);
2894}
2895
2896/*
2897 * scalarltjoinsel - Join selectivity of "<" for scalars
2898 */
2899Datum
2901{
2903}
2904
2905/*
2906 * scalarlejoinsel - Join selectivity of "<=" for scalars
2907 */
2908Datum
2910{
2912}
2913
2914/*
2915 * scalargtjoinsel - Join selectivity of ">" for scalars
2916 */
2917Datum
2919{
2921}
2922
2923/*
2924 * scalargejoinsel - Join selectivity of ">=" for scalars
2925 */
2926Datum
2928{
2930}
2931
2932
2933/*
2934 * mergejoinscansel - Scan selectivity of merge join.
2935 *
2936 * A merge join will stop as soon as it exhausts either input stream.
2937 * Therefore, if we can estimate the ranges of both input variables,
2938 * we can estimate how much of the input will actually be read. This
2939 * can have a considerable impact on the cost when using indexscans.
2940 *
2941 * Also, we can estimate how much of each input has to be read before the
2942 * first join pair is found, which will affect the join's startup time.
2943 *
2944 * clause should be a clause already known to be mergejoinable. opfamily,
2945 * strategy, and nulls_first specify the sort ordering being used.
2946 *
2947 * The outputs are:
2948 * *leftstart is set to the fraction of the left-hand variable expected
2949 * to be scanned before the first join pair is found (0 to 1).
2950 * *leftend is set to the fraction of the left-hand variable expected
2951 * to be scanned before the join terminates (0 to 1).
2952 * *rightstart, *rightend similarly for the right-hand variable.
2953 */
2954void
2956 Oid opfamily, int strategy, bool nulls_first,
2957 Selectivity *leftstart, Selectivity *leftend,
2958 Selectivity *rightstart, Selectivity *rightend)
2959{
2960 Node *left,
2961 *right;
2962 VariableStatData leftvar,
2963 rightvar;
2964 int op_strategy;
2965 Oid op_lefttype;
2966 Oid op_righttype;
2967 Oid opno,
2968 collation,
2969 lsortop,
2970 rsortop,
2971 lstatop,
2972 rstatop,
2973 ltop,
2974 leop,
2975 revltop,
2976 revleop;
2977 bool isgt;
2978 Datum leftmin,
2979 leftmax,
2980 rightmin,
2981 rightmax;
2982 double selec;
2983
2984 /* Set default results if we can't figure anything out. */
2985 /* XXX should default "start" fraction be a bit more than 0? */
2986 *leftstart = *rightstart = 0.0;
2987 *leftend = *rightend = 1.0;
2988
2989 /* Deconstruct the merge clause */
2990 if (!is_opclause(clause))
2991 return; /* shouldn't happen */
2992 opno = ((OpExpr *) clause)->opno;
2993 collation = ((OpExpr *) clause)->inputcollid;
2994 left = get_leftop((Expr *) clause);
2995 right = get_rightop((Expr *) clause);
2996 if (!right)
2997 return; /* shouldn't happen */
2998
2999 /* Look for stats for the inputs */
3000 examine_variable(root, left, 0, &leftvar);
3001 examine_variable(root, right, 0, &rightvar);
3002
3003 /* Extract the operator's declared left/right datatypes */
3004 get_op_opfamily_properties(opno, opfamily, false,
3005 &op_strategy,
3006 &op_lefttype,
3007 &op_righttype);
3008 Assert(op_strategy == BTEqualStrategyNumber);
3009
3010 /*
3011 * Look up the various operators we need. If we don't find them all, it
3012 * probably means the opfamily is broken, but we just fail silently.
3013 *
3014 * Note: we expect that pg_statistic histograms will be sorted by the '<'
3015 * operator, regardless of which sort direction we are considering.
3016 */
3017 switch (strategy)
3018 {
3020 isgt = false;
3021 if (op_lefttype == op_righttype)
3022 {
3023 /* easy case */
3024 ltop = get_opfamily_member(opfamily,
3025 op_lefttype, op_righttype,
3027 leop = get_opfamily_member(opfamily,
3028 op_lefttype, op_righttype,
3030 lsortop = ltop;
3031 rsortop = ltop;
3032 lstatop = lsortop;
3033 rstatop = rsortop;
3034 revltop = ltop;
3035 revleop = leop;
3036 }
3037 else
3038 {
3039 ltop = get_opfamily_member(opfamily,
3040 op_lefttype, op_righttype,
3042 leop = get_opfamily_member(opfamily,
3043 op_lefttype, op_righttype,
3045 lsortop = get_opfamily_member(opfamily,
3046 op_lefttype, op_lefttype,
3048 rsortop = get_opfamily_member(opfamily,
3049 op_righttype, op_righttype,
3051 lstatop = lsortop;
3052 rstatop = rsortop;
3053 revltop = get_opfamily_member(opfamily,
3054 op_righttype, op_lefttype,
3056 revleop = get_opfamily_member(opfamily,
3057 op_righttype, op_lefttype,
3059 }
3060 break;
3062 /* descending-order case */
3063 isgt = true;
3064 if (op_lefttype == op_righttype)
3065 {
3066 /* easy case */
3067 ltop = get_opfamily_member(opfamily,
3068 op_lefttype, op_righttype,
3070 leop = get_opfamily_member(opfamily,
3071 op_lefttype, op_righttype,
3073 lsortop = ltop;
3074 rsortop = ltop;
3075 lstatop = get_opfamily_member(opfamily,
3076 op_lefttype, op_lefttype,
3078 rstatop = lstatop;
3079 revltop = ltop;
3080 revleop = leop;
3081 }
3082 else
3083 {
3084 ltop = get_opfamily_member(opfamily,
3085 op_lefttype, op_righttype,
3087 leop = get_opfamily_member(opfamily,
3088 op_lefttype, op_righttype,
3090 lsortop = get_opfamily_member(opfamily,
3091 op_lefttype, op_lefttype,
3093 rsortop = get_opfamily_member(opfamily,
3094 op_righttype, op_righttype,
3096 lstatop = get_opfamily_member(opfamily,
3097 op_lefttype, op_lefttype,
3099 rstatop = get_opfamily_member(opfamily,
3100 op_righttype, op_righttype,
3102 revltop = get_opfamily_member(opfamily,
3103 op_righttype, op_lefttype,
3105 revleop = get_opfamily_member(opfamily,
3106 op_righttype, op_lefttype,
3108 }
3109 break;
3110 default:
3111 goto fail; /* shouldn't get here */
3112 }
3113
3114 if (!OidIsValid(lsortop) ||
3115 !OidIsValid(rsortop) ||
3116 !OidIsValid(lstatop) ||
3117 !OidIsValid(rstatop) ||
3118 !OidIsValid(ltop) ||
3119 !OidIsValid(leop) ||
3120 !OidIsValid(revltop) ||
3121 !OidIsValid(revleop))
3122 goto fail; /* insufficient info in catalogs */
3123
3124 /* Try to get ranges of both inputs */
3125 if (!isgt)
3126 {
3127 if (!get_variable_range(root, &leftvar, lstatop, collation,
3128 &leftmin, &leftmax))
3129 goto fail; /* no range available from stats */
3130 if (!get_variable_range(root, &rightvar, rstatop, collation,
3131 &rightmin, &rightmax))
3132 goto fail; /* no range available from stats */
3133 }
3134 else
3135 {
3136 /* need to swap the max and min */
3137 if (!get_variable_range(root, &leftvar, lstatop, collation,
3138 &leftmax, &leftmin))
3139 goto fail; /* no range available from stats */
3140 if (!get_variable_range(root, &rightvar, rstatop, collation,
3141 &rightmax, &rightmin))
3142 goto fail; /* no range available from stats */
3143 }
3144
3145 /*
3146 * Now, the fraction of the left variable that will be scanned is the
3147 * fraction that's <= the right-side maximum value. But only believe
3148 * non-default estimates, else stick with our 1.0.
3149 */
3150 selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3151 rightmax, op_righttype);
3152 if (selec != DEFAULT_INEQ_SEL)
3153 *leftend = selec;
3154
3155 /* And similarly for the right variable. */
3156 selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3157 leftmax, op_lefttype);
3158 if (selec != DEFAULT_INEQ_SEL)
3159 *rightend = selec;
3160
3161 /*
3162 * Only one of the two "end" fractions can really be less than 1.0;
3163 * believe the smaller estimate and reset the other one to exactly 1.0. If
3164 * we get exactly equal estimates (as can easily happen with self-joins),
3165 * believe neither.
3166 */
3167 if (*leftend > *rightend)
3168 *leftend = 1.0;
3169 else if (*leftend < *rightend)
3170 *rightend = 1.0;
3171 else
3172 *leftend = *rightend = 1.0;
3173
3174 /*
3175 * Also, the fraction of the left variable that will be scanned before the
3176 * first join pair is found is the fraction that's < the right-side
3177 * minimum value. But only believe non-default estimates, else stick with
3178 * our own default.
3179 */
3180 selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3181 rightmin, op_righttype);
3182 if (selec != DEFAULT_INEQ_SEL)
3183 *leftstart = selec;
3184
3185 /* And similarly for the right variable. */
3186 selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3187 leftmin, op_lefttype);
3188 if (selec != DEFAULT_INEQ_SEL)
3189 *rightstart = selec;
3190
3191 /*
3192 * Only one of the two "start" fractions can really be more than zero;
3193 * believe the larger estimate and reset the other one to exactly 0.0. If
3194 * we get exactly equal estimates (as can easily happen with self-joins),
3195 * believe neither.
3196 */
3197 if (*leftstart < *rightstart)
3198 *leftstart = 0.0;
3199 else if (*leftstart > *rightstart)
3200 *rightstart = 0.0;
3201 else
3202 *leftstart = *rightstart = 0.0;
3203
3204 /*
3205 * If the sort order is nulls-first, we're going to have to skip over any
3206 * nulls too. These would not have been counted by scalarineqsel, and we
3207 * can safely add in this fraction regardless of whether we believe
3208 * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3209 */
3210 if (nulls_first)
3211 {
3212 Form_pg_statistic stats;
3213
3214 if (HeapTupleIsValid(leftvar.statsTuple))
3215 {
3216 stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3217 *leftstart += stats->stanullfrac;
3218 CLAMP_PROBABILITY(*leftstart);
3219 *leftend += stats->stanullfrac;
3220 CLAMP_PROBABILITY(*leftend);
3221 }
3222 if (HeapTupleIsValid(rightvar.statsTuple))
3223 {
3224 stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3225 *rightstart += stats->stanullfrac;
3226 CLAMP_PROBABILITY(*rightstart);
3227 *rightend += stats->stanullfrac;
3228 CLAMP_PROBABILITY(*rightend);
3229 }
3230 }
3231
3232 /* Disbelieve start >= end, just in case that can happen */
3233 if (*leftstart >= *leftend)
3234 {
3235 *leftstart = 0.0;
3236 *leftend = 1.0;
3237 }
3238 if (*rightstart >= *rightend)
3239 {
3240 *rightstart = 0.0;
3241 *rightend = 1.0;
3242 }
3243
3244fail:
3245 ReleaseVariableStats(leftvar);
3246 ReleaseVariableStats(rightvar);
3247}
3248
3249
3250/*
3251 * matchingsel -- generic matching-operator selectivity support
3252 *
3253 * Use these for any operators that (a) are on data types for which we collect
3254 * standard statistics, and (b) have behavior for which the default estimate
3255 * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3256 * operators.
3257 */
3258
3259Datum
3261{
3263 Oid operator = PG_GETARG_OID(1);
3264 List *args = (List *) PG_GETARG_POINTER(2);
3265 int varRelid = PG_GETARG_INT32(3);
3266 Oid collation = PG_GET_COLLATION();
3267 double selec;
3268
3269 /* Use generic restriction selectivity logic. */
3270 selec = generic_restriction_selectivity(root, operator, collation,
3271 args, varRelid,
3273
3274 PG_RETURN_FLOAT8((float8) selec);
3275}
3276
3277Datum
3279{
3280 /* Just punt, for the moment. */
3282}
3283
3284
3285/*
3286 * Helper routine for estimate_num_groups: add an item to a list of
3287 * GroupVarInfos, but only if it's not known equal to any of the existing
3288 * entries.
3289 */
3290typedef struct
3291{
3292 Node *var; /* might be an expression, not just a Var */
3293 RelOptInfo *rel; /* relation it belongs to */
3294 double ndistinct; /* # distinct values */
3295 bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3296} GroupVarInfo;
3297
3298static List *
3300 Node *var, VariableStatData *vardata)
3301{
3302 GroupVarInfo *varinfo;
3303 double ndistinct;
3304 bool isdefault;
3305 ListCell *lc;
3306
3307 ndistinct = get_variable_numdistinct(vardata, &isdefault);
3308
3309 foreach(lc, varinfos)
3310 {
3311 varinfo = (GroupVarInfo *) lfirst(lc);
3312
3313 /* Drop exact duplicates */
3314 if (equal(var, varinfo->var))
3315 return varinfos;
3316
3317 /*
3318 * Drop known-equal vars, but only if they belong to different
3319 * relations (see comments for estimate_num_groups). We aren't too
3320 * fussy about the semantics of "equal" here.
3321 */
3322 if (vardata->rel != varinfo->rel &&
3323 exprs_known_equal(root, var, varinfo->var, InvalidOid))
3324 {
3325 if (varinfo->ndistinct <= ndistinct)
3326 {
3327 /* Keep older item, forget new one */
3328 return varinfos;
3329 }
3330 else
3331 {
3332 /* Delete the older item */
3333 varinfos = foreach_delete_current(varinfos, lc);
3334 }
3335 }
3336 }
3337
3338 varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
3339
3340 varinfo->var = var;
3341 varinfo->rel = vardata->rel;
3342 varinfo->ndistinct = ndistinct;
3343 varinfo->isdefault = isdefault;
3344 varinfos = lappend(varinfos, varinfo);
3345 return varinfos;
3346}
3347
3348/*
3349 * estimate_num_groups - Estimate number of groups in a grouped query
3350 *
3351 * Given a query having a GROUP BY clause, estimate how many groups there
3352 * will be --- ie, the number of distinct combinations of the GROUP BY
3353 * expressions.
3354 *
3355 * This routine is also used to estimate the number of rows emitted by
3356 * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3357 * actually, we only use it for DISTINCT when there's no grouping or
3358 * aggregation ahead of the DISTINCT.)
3359 *
3360 * Inputs:
3361 * root - the query
3362 * groupExprs - list of expressions being grouped by
3363 * input_rows - number of rows estimated to arrive at the group/unique
3364 * filter step
3365 * pgset - NULL, or a List** pointing to a grouping set to filter the
3366 * groupExprs against
3367 *
3368 * Outputs:
3369 * estinfo - When passed as non-NULL, the function will set bits in the
3370 * "flags" field in order to provide callers with additional information
3371 * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3372 * bit if we used any default values in the estimation.
3373 *
3374 * Given the lack of any cross-correlation statistics in the system, it's
3375 * impossible to do anything really trustworthy with GROUP BY conditions
3376 * involving multiple Vars. We should however avoid assuming the worst
3377 * case (all possible cross-product terms actually appear as groups) since
3378 * very often the grouped-by Vars are highly correlated. Our current approach
3379 * is as follows:
3380 * 1. Expressions yielding boolean are assumed to contribute two groups,
3381 * independently of their content, and are ignored in the subsequent
3382 * steps. This is mainly because tests like "col IS NULL" break the
3383 * heuristic used in step 2 especially badly.
3384 * 2. Reduce the given expressions to a list of unique Vars used. For
3385 * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3386 * It is clearly correct not to count the same Var more than once.
3387 * It is also reasonable to treat f(x) the same as x: f() cannot
3388 * increase the number of distinct values (unless it is volatile,
3389 * which we consider unlikely for grouping), but it probably won't
3390 * reduce the number of distinct values much either.
3391 * As a special case, if a GROUP BY expression can be matched to an
3392 * expressional index for which we have statistics, then we treat the
3393 * whole expression as though it were just a Var.
3394 * 3. If the list contains Vars of different relations that are known equal
3395 * due to equivalence classes, then drop all but one of the Vars from each
3396 * known-equal set, keeping the one with smallest estimated # of values
3397 * (since the extra values of the others can't appear in joined rows).
3398 * Note the reason we only consider Vars of different relations is that
3399 * if we considered ones of the same rel, we'd be double-counting the
3400 * restriction selectivity of the equality in the next step.
3401 * 4. For Vars within a single source rel, we multiply together the numbers
3402 * of values, clamp to the number of rows in the rel (divided by 10 if
3403 * more than one Var), and then multiply by a factor based on the
3404 * selectivity of the restriction clauses for that rel. When there's
3405 * more than one Var, the initial product is probably too high (it's the
3406 * worst case) but clamping to a fraction of the rel's rows seems to be a
3407 * helpful heuristic for not letting the estimate get out of hand. (The
3408 * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3409 * we multiply by to adjust for the restriction selectivity assumes that
3410 * the restriction clauses are independent of the grouping, which may not
3411 * be a valid assumption, but it's hard to do better.
3412 * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3413 * rel, and multiply the results together.
3414 * Note that rels not containing grouped Vars are ignored completely, as are
3415 * join clauses. Such rels cannot increase the number of groups, and we
3416 * assume such clauses do not reduce the number either (somewhat bogus,
3417 * but we don't have the info to do better).
3418 */
3419double
3420estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3421 List **pgset, EstimationInfo *estinfo)
3422{
3423 List *varinfos = NIL;
3424 double srf_multiplier = 1.0;
3425 double numdistinct;
3426 ListCell *l;
3427 int i;
3428
3429 /* Zero the estinfo output parameter, if non-NULL */
3430 if (estinfo != NULL)
3431 memset(estinfo, 0, sizeof(EstimationInfo));
3432
3433 /*
3434 * We don't ever want to return an estimate of zero groups, as that tends
3435 * to lead to division-by-zero and other unpleasantness. The input_rows
3436 * estimate is usually already at least 1, but clamp it just in case it
3437 * isn't.
3438 */
3439 input_rows = clamp_row_est(input_rows);
3440
3441 /*
3442 * If no grouping columns, there's exactly one group. (This can't happen
3443 * for normal cases with GROUP BY or DISTINCT, but it is possible for
3444 * corner cases with set operations.)
3445 */
3446 if (groupExprs == NIL || (pgset && *pgset == NIL))
3447 return 1.0;
3448
3449 /*
3450 * Count groups derived from boolean grouping expressions. For other
3451 * expressions, find the unique Vars used, treating an expression as a Var
3452 * if we can find stats for it. For each one, record the statistical
3453 * estimate of number of distinct values (total in its table, without
3454 * regard for filtering).
3455 */
3456 numdistinct = 1.0;
3457
3458 i = 0;
3459 foreach(l, groupExprs)
3460 {
3461 Node *groupexpr = (Node *) lfirst(l);
3462 double this_srf_multiplier;
3463 VariableStatData vardata;
3464 List *varshere;
3465 ListCell *l2;
3466
3467 /* is expression in this grouping set? */
3468 if (pgset && !list_member_int(*pgset, i++))
3469 continue;
3470
3471 /*
3472 * Set-returning functions in grouping columns are a bit problematic.
3473 * The code below will effectively ignore their SRF nature and come up
3474 * with a numdistinct estimate as though they were scalar functions.
3475 * We compensate by scaling up the end result by the largest SRF
3476 * rowcount estimate. (This will be an overestimate if the SRF
3477 * produces multiple copies of any output value, but it seems best to
3478 * assume the SRF's outputs are distinct. In any case, it's probably
3479 * pointless to worry too much about this without much better
3480 * estimates for SRF output rowcounts than we have today.)
3481 */
3482 this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3483 if (srf_multiplier < this_srf_multiplier)
3484 srf_multiplier = this_srf_multiplier;
3485
3486 /* Short-circuit for expressions returning boolean */
3487 if (exprType(groupexpr) == BOOLOID)
3488 {
3489 numdistinct *= 2.0;
3490 continue;
3491 }
3492
3493 /*
3494 * If examine_variable is able to deduce anything about the GROUP BY
3495 * expression, treat it as a single variable even if it's really more
3496 * complicated.
3497 *
3498 * XXX This has the consequence that if there's a statistics object on
3499 * the expression, we don't split it into individual Vars. This
3500 * affects our selection of statistics in
3501 * estimate_multivariate_ndistinct, because it's probably better to
3502 * use more accurate estimate for each expression and treat them as
3503 * independent, than to combine estimates for the extracted variables
3504 * when we don't know how that relates to the expressions.
3505 */
3506 examine_variable(root, groupexpr, 0, &vardata);
3507 if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3508 {
3509 varinfos = add_unique_group_var(root, varinfos,
3510 groupexpr, &vardata);
3511 ReleaseVariableStats(vardata);
3512 continue;
3513 }
3514 ReleaseVariableStats(vardata);
3515
3516 /*
3517 * Else pull out the component Vars. Handle PlaceHolderVars by
3518 * recursing into their arguments (effectively assuming that the
3519 * PlaceHolderVar doesn't change the number of groups, which boils
3520 * down to ignoring the possible addition of nulls to the result set).
3521 */
3522 varshere = pull_var_clause(groupexpr,
3526
3527 /*
3528 * If we find any variable-free GROUP BY item, then either it is a
3529 * constant (and we can ignore it) or it contains a volatile function;
3530 * in the latter case we punt and assume that each input row will
3531 * yield a distinct group.
3532 */
3533 if (varshere == NIL)
3534 {
3535 if (contain_volatile_functions(groupexpr))
3536 return input_rows;
3537 continue;
3538 }
3539
3540 /*
3541 * Else add variables to varinfos list
3542 */
3543 foreach(l2, varshere)
3544 {
3545 Node *var = (Node *) lfirst(l2);
3546
3547 examine_variable(root, var, 0, &vardata);
3548 varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3549 ReleaseVariableStats(vardata);
3550 }
3551 }
3552
3553 /*
3554 * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3555 * list.
3556 */
3557 if (varinfos == NIL)
3558 {
3559 /* Apply SRF multiplier as we would do in the long path */
3560 numdistinct *= srf_multiplier;
3561 /* Round off */
3562 numdistinct = ceil(numdistinct);
3563 /* Guard against out-of-range answers */
3564 if (numdistinct > input_rows)
3565 numdistinct = input_rows;
3566 if (numdistinct < 1.0)
3567 numdistinct = 1.0;
3568 return numdistinct;
3569 }
3570
3571 /*
3572 * Group Vars by relation and estimate total numdistinct.
3573 *
3574 * For each iteration of the outer loop, we process the frontmost Var in
3575 * varinfos, plus all other Vars in the same relation. We remove these
3576 * Vars from the newvarinfos list for the next iteration. This is the
3577 * easiest way to group Vars of same rel together.
3578 */
3579 do
3580 {
3581 GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3582 RelOptInfo *rel = varinfo1->rel;
3583 double reldistinct = 1;
3584 double relmaxndistinct = reldistinct;
3585 int relvarcount = 0;
3586 List *newvarinfos = NIL;
3587 List *relvarinfos = NIL;
3588
3589 /*
3590 * Split the list of varinfos in two - one for the current rel, one
3591 * for remaining Vars on other rels.
3592 */
3593 relvarinfos = lappend(relvarinfos, varinfo1);
3594 for_each_from(l, varinfos, 1)
3595 {
3596 GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3597
3598 if (varinfo2->rel == varinfo1->rel)
3599 {
3600 /* varinfos on current rel */
3601 relvarinfos = lappend(relvarinfos, varinfo2);
3602 }
3603 else
3604 {
3605 /* not time to process varinfo2 yet */
3606 newvarinfos = lappend(newvarinfos, varinfo2);
3607 }
3608 }
3609
3610 /*
3611 * Get the numdistinct estimate for the Vars of this rel. We
3612 * iteratively search for multivariate n-distinct with maximum number
3613 * of vars; assuming that each var group is independent of the others,
3614 * we multiply them together. Any remaining relvarinfos after no more
3615 * multivariate matches are found are assumed independent too, so
3616 * their individual ndistinct estimates are multiplied also.
3617 *
3618 * While iterating, count how many separate numdistinct values we
3619 * apply. We apply a fudge factor below, but only if we multiplied
3620 * more than one such values.
3621 */
3622 while (relvarinfos)
3623 {
3624 double mvndistinct;
3625
3626 if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3627 &mvndistinct))
3628 {
3629 reldistinct *= mvndistinct;
3630 if (relmaxndistinct < mvndistinct)
3631 relmaxndistinct = mvndistinct;
3632 relvarcount++;
3633 }
3634 else
3635 {
3636 foreach(l, relvarinfos)
3637 {
3638 GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3639
3640 reldistinct *= varinfo2->ndistinct;
3641 if (relmaxndistinct < varinfo2->ndistinct)
3642 relmaxndistinct = varinfo2->ndistinct;
3643 relvarcount++;
3644
3645 /*
3646 * When varinfo2's isdefault is set then we'd better set
3647 * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3648 */
3649 if (estinfo != NULL && varinfo2->isdefault)
3650 estinfo->flags |= SELFLAG_USED_DEFAULT;
3651 }
3652
3653 /* we're done with this relation */
3654 relvarinfos = NIL;
3655 }
3656 }
3657
3658 /*
3659 * Sanity check --- don't divide by zero if empty relation.
3660 */
3661 Assert(IS_SIMPLE_REL(rel));
3662 if (rel->tuples > 0)
3663 {
3664 /*
3665 * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
3666 * fudge factor is because the Vars are probably correlated but we
3667 * don't know by how much. We should never clamp to less than the
3668 * largest ndistinct value for any of the Vars, though, since
3669 * there will surely be at least that many groups.
3670 */
3671 double clamp = rel->tuples;
3672
3673 if (relvarcount > 1)
3674 {
3675 clamp *= 0.1;
3676 if (clamp < relmaxndistinct)
3677 {
3678 clamp = relmaxndistinct;
3679 /* for sanity in case some ndistinct is too large: */
3680 if (clamp > rel->tuples)
3681 clamp = rel->tuples;
3682 }
3683 }
3684 if (reldistinct > clamp)
3685 reldistinct = clamp;
3686
3687 /*
3688 * Update the estimate based on the restriction selectivity,
3689 * guarding against division by zero when reldistinct is zero.
3690 * Also skip this if we know that we are returning all rows.
3691 */
3692 if (reldistinct > 0 && rel->rows < rel->tuples)
3693 {
3694 /*
3695 * Given a table containing N rows with n distinct values in a
3696 * uniform distribution, if we select p rows at random then
3697 * the expected number of distinct values selected is
3698 *
3699 * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
3700 *
3701 * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
3702 *
3703 * See "Approximating block accesses in database
3704 * organizations", S. B. Yao, Communications of the ACM,
3705 * Volume 20 Issue 4, April 1977 Pages 260-261.
3706 *
3707 * Alternatively, re-arranging the terms from the factorials,
3708 * this may be written as
3709 *
3710 * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
3711 *
3712 * This form of the formula is more efficient to compute in
3713 * the common case where p is larger than N/n. Additionally,
3714 * as pointed out by Dell'Era, if i << N for all terms in the
3715 * product, it can be approximated by
3716 *
3717 * n * (1 - ((N-p)/N)^(N/n))
3718 *
3719 * See "Expected distinct values when selecting from a bag
3720 * without replacement", Alberto Dell'Era,
3721 * http://www.adellera.it/investigations/distinct_balls/.
3722 *
3723 * The condition i << N is equivalent to n >> 1, so this is a
3724 * good approximation when the number of distinct values in
3725 * the table is large. It turns out that this formula also
3726 * works well even when n is small.
3727 */
3728 reldistinct *=
3729 (1 - pow((rel->tuples - rel->rows) / rel->tuples,
3730 rel->tuples / reldistinct));
3731 }
3732 reldistinct = clamp_row_est(reldistinct);
3733
3734 /*
3735 * Update estimate of total distinct groups.
3736 */
3737 numdistinct *= reldistinct;
3738 }
3739
3740 varinfos = newvarinfos;
3741 } while (varinfos != NIL);
3742
3743 /* Now we can account for the effects of any SRFs */
3744 numdistinct *= srf_multiplier;
3745
3746 /* Round off */
3747 numdistinct = ceil(numdistinct);
3748
3749 /* Guard against out-of-range answers */
3750 if (numdistinct > input_rows)
3751 numdistinct = input_rows;
3752 if (numdistinct < 1.0)
3753 numdistinct = 1.0;
3754
3755 return numdistinct;
3756}
3757
3758/*
3759 * Estimate hash bucket statistics when the specified expression is used
3760 * as a hash key for the given number of buckets.
3761 *
3762 * This attempts to determine two values:
3763 *
3764 * 1. The frequency of the most common value of the expression (returns
3765 * zero into *mcv_freq if we can't get that).
3766 *
3767 * 2. The "bucketsize fraction", ie, average number of entries in a bucket
3768 * divided by total tuples in relation.
3769 *
3770 * XXX This is really pretty bogus since we're effectively assuming that the
3771 * distribution of hash keys will be the same after applying restriction
3772 * clauses as it was in the underlying relation. However, we are not nearly
3773 * smart enough to figure out how the restrict clauses might change the
3774 * distribution, so this will have to do for now.
3775 *
3776 * We are passed the number of buckets the executor will use for the given
3777 * input relation. If the data were perfectly distributed, with the same
3778 * number of tuples going into each available bucket, then the bucketsize
3779 * fraction would be 1/nbuckets. But this happy state of affairs will occur
3780 * only if (a) there are at least nbuckets distinct data values, and (b)
3781 * we have a not-too-skewed data distribution. Otherwise the buckets will
3782 * be nonuniformly occupied. If the other relation in the join has a key
3783 * distribution similar to this one's, then the most-loaded buckets are
3784 * exactly those that will be probed most often. Therefore, the "average"
3785 * bucket size for costing purposes should really be taken as something close
3786 * to the "worst case" bucket size. We try to estimate this by adjusting the
3787 * fraction if there are too few distinct data values, and then scaling up
3788 * by the ratio of the most common value's frequency to the average frequency.
3789 *
3790 * If no statistics are available, use a default estimate of 0.1. This will
3791 * discourage use of a hash rather strongly if the inner relation is large,
3792 * which is what we want. We do not want to hash unless we know that the
3793 * inner rel is well-dispersed (or the alternatives seem much worse).
3794 *
3795 * The caller should also check that the mcv_freq is not so large that the
3796 * most common value would by itself require an impractically large bucket.
3797 * In a hash join, the executor can split buckets if they get too big, but
3798 * obviously that doesn't help for a bucket that contains many duplicates of
3799 * the same value.
3800 */
3801void
3803 Selectivity *mcv_freq,
3804 Selectivity *bucketsize_frac)
3805{
3806 VariableStatData vardata;
3807 double estfract,
3808 ndistinct,
3809 stanullfrac,
3810 avgfreq;
3811 bool isdefault;
3812 AttStatsSlot sslot;
3813
3814 examine_variable(root, hashkey, 0, &vardata);
3815
3816 /* Look up the frequency of the most common value, if available */
3817 *mcv_freq = 0.0;
3818
3819 if (HeapTupleIsValid(vardata.statsTuple))
3820 {
3821 if (get_attstatsslot(&sslot, vardata.statsTuple,
3822 STATISTIC_KIND_MCV, InvalidOid,
3824 {
3825 /*
3826 * The first MCV stat is for the most common value.
3827 */
3828 if (sslot.nnumbers > 0)
3829 *mcv_freq = sslot.numbers[0];
3830 free_attstatsslot(&sslot);
3831 }
3832 }
3833
3834 /* Get number of distinct values */
3835 ndistinct = get_variable_numdistinct(&vardata, &isdefault);
3836
3837 /*
3838 * If ndistinct isn't real, punt. We normally return 0.1, but if the
3839 * mcv_freq is known to be even higher than that, use it instead.
3840 */
3841 if (isdefault)
3842 {
3843 *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
3844 ReleaseVariableStats(vardata);
3845 return;
3846 }
3847
3848 /* Get fraction that are null */
3849 if (HeapTupleIsValid(vardata.statsTuple))
3850 {
3851 Form_pg_statistic stats;
3852
3853 stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
3854 stanullfrac = stats->stanullfrac;
3855 }
3856 else
3857 stanullfrac = 0.0;
3858
3859 /* Compute avg freq of all distinct data values in raw relation */
3860 avgfreq = (1.0 - stanullfrac) / ndistinct;
3861
3862 /*
3863 * Adjust ndistinct to account for restriction clauses. Observe we are
3864 * assuming that the data distribution is affected uniformly by the
3865 * restriction clauses!
3866 *
3867 * XXX Possibly better way, but much more expensive: multiply by
3868 * selectivity of rel's restriction clauses that mention the target Var.
3869 */
3870 if (vardata.rel && vardata.rel->tuples > 0)
3871 {
3872 ndistinct *= vardata.rel->rows / vardata.rel->tuples;
3873 ndistinct = clamp_row_est(ndistinct);
3874 }
3875
3876 /*
3877 * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
3878 * number of buckets is less than the expected number of distinct values;
3879 * otherwise it is 1/ndistinct.
3880 */
3881 if (ndistinct > nbuckets)
3882 estfract = 1.0 / nbuckets;
3883 else
3884 estfract = 1.0 / ndistinct;
3885
3886 /*
3887 * Adjust estimated bucketsize upward to account for skewed distribution.
3888 */
3889 if (avgfreq > 0.0 && *mcv_freq > avgfreq)
3890 estfract *= *mcv_freq / avgfreq;
3891
3892 /*
3893 * Clamp bucketsize to sane range (the above adjustment could easily
3894 * produce an out-of-range result). We set the lower bound a little above
3895 * zero, since zero isn't a very sane result.
3896 */
3897 if (estfract < 1.0e-6)
3898 estfract = 1.0e-6;
3899 else if (estfract > 1.0)
3900 estfract = 1.0;
3901
3902 *bucketsize_frac = (Selectivity) estfract;
3903
3904 ReleaseVariableStats(vardata);
3905}
3906
3907/*
3908 * estimate_hashagg_tablesize
3909 * estimate the number of bytes that a hash aggregate hashtable will
3910 * require based on the agg_costs, path width and number of groups.
3911 *
3912 * We return the result as "double" to forestall any possible overflow
3913 * problem in the multiplication by dNumGroups.
3914 *
3915 * XXX this may be over-estimating the size now that hashagg knows to omit
3916 * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
3917 * grouping columns not in the hashed set are counted here even though hashagg
3918 * won't store them. Is this a problem?
3919 */
3920double
3922 const AggClauseCosts *agg_costs, double dNumGroups)
3923{
3924 Size hashentrysize;
3925
3926 hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
3927 path->pathtarget->width,
3928 agg_costs->transitionSpace);
3929
3930 /*
3931 * Note that this disregards the effect of fill-factor and growth policy
3932 * of the hash table. That's probably ok, given that the default
3933 * fill-factor is relatively high. It'd be hard to meaningfully factor in
3934 * "double-in-size" growth policies here.
3935 */
3936 return hashentrysize * dNumGroups;
3937}
3938
3939
3940/*-------------------------------------------------------------------------
3941 *
3942 * Support routines
3943 *
3944 *-------------------------------------------------------------------------
3945 */
3946
3947/*
3948 * Find applicable ndistinct statistics for the given list of VarInfos (which
3949 * must all belong to the given rel), and update *ndistinct to the estimate of
3950 * the MVNDistinctItem that best matches. If a match it found, *varinfos is
3951 * updated to remove the list of matched varinfos.
3952 *
3953 * Varinfos that aren't for simple Vars are ignored.
3954 *
3955 * Return true if we're able to find a match, false otherwise.
3956 */
3957static bool
3959 List **varinfos, double *ndistinct)
3960{
3961 ListCell *lc;
3962 int nmatches_vars;
3963 int nmatches_exprs;
3964 Oid statOid = InvalidOid;
3965 MVNDistinct *stats;
3966 StatisticExtInfo *matched_info = NULL;
3968
3969 /* bail out immediately if the table has no extended statistics */
3970 if (!rel->statlist)
3971 return false;
3972
3973 /* look for the ndistinct statistics object matching the most vars */
3974 nmatches_vars = 0; /* we require at least two matches */
3975 nmatches_exprs = 0;
3976 foreach(lc, rel->statlist)
3977 {
3978 ListCell *lc2;
3980 int nshared_vars = 0;
3981 int nshared_exprs = 0;
3982
3983 /* skip statistics of other kinds */
3984 if (info->kind != STATS_EXT_NDISTINCT)
3985 continue;
3986
3987 /* skip statistics with mismatching stxdinherit value */
3988 if (info->inherit != rte->inh)
3989 continue;
3990
3991 /*
3992 * Determine how many expressions (and variables in non-matched
3993 * expressions) match. We'll then use these numbers to pick the
3994 * statistics object that best matches the clauses.
3995 */
3996 foreach(lc2, *varinfos)
3997 {
3998 ListCell *lc3;
3999 GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4001
4002 Assert(varinfo->rel == rel);
4003
4004 /* simple Var, search in statistics keys directly */
4005 if (IsA(varinfo->var, Var))
4006 {
4007 attnum = ((Var *) varinfo->var)->varattno;
4008
4009 /*
4010 * Ignore system attributes - we don't support statistics on
4011 * them, so can't match them (and it'd fail as the values are
4012 * negative).
4013 */
4015 continue;
4016
4017 if (bms_is_member(attnum, info->keys))
4018 nshared_vars++;
4019
4020 continue;
4021 }
4022
4023 /* expression - see if it's in the statistics object */
4024 foreach(lc3, info->exprs)
4025 {
4026 Node *expr = (Node *) lfirst(lc3);
4027
4028 if (equal(varinfo->var, expr))
4029 {
4030 nshared_exprs++;
4031 break;
4032 }
4033 }
4034 }
4035
4036 if (nshared_vars + nshared_exprs < 2)
4037 continue;
4038
4039 /*
4040 * Does this statistics object match more columns than the currently
4041 * best object? If so, use this one instead.
4042 *
4043 * XXX This should break ties using name of the object, or something
4044 * like that, to make the outcome stable.
4045 */
4046 if ((nshared_exprs > nmatches_exprs) ||
4047 (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4048 {
4049 statOid = info->statOid;
4050 nmatches_vars = nshared_vars;
4051 nmatches_exprs = nshared_exprs;
4052 matched_info = info;
4053 }
4054 }
4055
4056 /* No match? */
4057 if (statOid == InvalidOid)
4058 return false;
4059
4060 Assert(nmatches_vars + nmatches_exprs > 1);
4061
4062 stats = statext_ndistinct_load(statOid, rte->inh);
4063
4064 /*
4065 * If we have a match, search it for the specific item that matches (there
4066 * must be one), and construct the output values.
4067 */
4068 if (stats)
4069 {
4070 int i;
4071 List *newlist = NIL;
4072 MVNDistinctItem *item = NULL;
4073 ListCell *lc2;
4074 Bitmapset *matched = NULL;
4075 AttrNumber attnum_offset;
4076
4077 /*
4078 * How much we need to offset the attnums? If there are no
4079 * expressions, no offset is needed. Otherwise offset enough to move
4080 * the lowest one (which is equal to number of expressions) to 1.
4081 */
4082 if (matched_info->exprs)
4083 attnum_offset = (list_length(matched_info->exprs) + 1);
4084 else
4085 attnum_offset = 0;
4086
4087 /* see what actually matched */
4088 foreach(lc2, *varinfos)
4089 {
4090 ListCell *lc3;
4091 int idx;
4092 bool found = false;
4093
4094 GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4095
4096 /*
4097 * Process a simple Var expression, by matching it to keys
4098 * directly. If there's a matching expression, we'll try matching
4099 * it later.
4100 */
4101 if (IsA(varinfo->var, Var))
4102 {
4103 AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4104
4105 /*
4106 * Ignore expressions on system attributes. Can't rely on the
4107 * bms check for negative values.
4108 */
4110 continue;
4111
4112 /* Is the variable covered by the statistics object? */
4113 if (!bms_is_member(attnum, matched_info->keys))
4114 continue;
4115
4116 attnum = attnum + attnum_offset;
4117
4118 /* ensure sufficient offset */
4120
4121 matched = bms_add_member(matched, attnum);
4122
4123 found = true;
4124 }
4125
4126 /*
4127 * XXX Maybe we should allow searching the expressions even if we
4128 * found an attribute matching the expression? That would handle
4129 * trivial expressions like "(a)" but it seems fairly useless.
4130 */
4131 if (found)
4132 continue;
4133
4134 /* expression - see if it's in the statistics object */
4135 idx = 0;
4136 foreach(lc3, matched_info->exprs)
4137 {
4138 Node *expr = (Node *) lfirst(lc3);
4139
4140 if (equal(varinfo->var, expr))
4141 {
4142 AttrNumber attnum = -(idx + 1);
4143
4144 attnum = attnum + attnum_offset;
4145
4146 /* ensure sufficient offset */
4148
4149 matched = bms_add_member(matched, attnum);
4150
4151 /* there should be just one matching expression */
4152 break;
4153 }
4154
4155 idx++;
4156 }
4157 }
4158
4159 /* Find the specific item that exactly matches the combination */
4160 for (i = 0; i < stats->nitems; i++)
4161 {
4162 int j;
4163 MVNDistinctItem *tmpitem = &stats->items[i];
4164
4165 if (tmpitem->nattributes != bms_num_members(matched))
4166 continue;
4167
4168 /* assume it's the right item */
4169 item = tmpitem;
4170
4171 /* check that all item attributes/expressions fit the match */
4172 for (j = 0; j < tmpitem->nattributes; j++)
4173 {
4174 AttrNumber attnum = tmpitem->attributes[j];
4175
4176 /*
4177 * Thanks to how we constructed the matched bitmap above, we
4178 * can just offset all attnums the same way.
4179 */
4180 attnum = attnum + attnum_offset;
4181
4182 if (!bms_is_member(attnum, matched))
4183 {
4184 /* nah, it's not this item */
4185 item = NULL;
4186 break;
4187 }
4188 }
4189
4190 /*
4191 * If the item has all the matched attributes, we know it's the
4192 * right one - there can't be a better one. matching more.
4193 */
4194 if (item)
4195 break;
4196 }
4197
4198 /*
4199 * Make sure we found an item. There has to be one, because ndistinct
4200 * statistics includes all combinations of attributes.
4201 */
4202 if (!item)
4203 elog(ERROR, "corrupt MVNDistinct entry");
4204
4205 /* Form the output varinfo list, keeping only unmatched ones */
4206 foreach(lc, *varinfos)
4207 {
4208 GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4209 ListCell *lc3;
4210 bool found = false;
4211
4212 /*
4213 * Let's look at plain variables first, because it's the most
4214 * common case and the check is quite cheap. We can simply get the
4215 * attnum and check (with an offset) matched bitmap.
4216 */
4217 if (IsA(varinfo->var, Var))
4218 {
4219 AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4220
4221 /*
4222 * If it's a system attribute, we're done. We don't support
4223 * extended statistics on system attributes, so it's clearly
4224 * not matched. Just keep the expression and continue.
4225 */
4227 {
4228 newlist = lappend(newlist, varinfo);
4229 continue;
4230 }
4231
4232 /* apply the same offset as above */
4233 attnum += attnum_offset;
4234
4235 /* if it's not matched, keep the varinfo */
4236 if (!bms_is_member(attnum, matched))
4237 newlist = lappend(newlist, varinfo);
4238
4239 /* The rest of the loop deals with complex expressions. */
4240 continue;
4241 }
4242
4243 /*
4244 * Process complex expressions, not just simple Vars.
4245 *
4246 * First, we search for an exact match of an expression. If we
4247 * find one, we can just discard the whole GroupVarInfo, with all
4248 * the variables we extracted from it.
4249 *
4250 * Otherwise we inspect the individual vars, and try matching it
4251 * to variables in the item.
4252 */
4253 foreach(lc3, matched_info->exprs)
4254 {
4255 Node *expr = (Node *) lfirst(lc3);
4256
4257 if (equal(varinfo->var, expr))
4258 {
4259 found = true;
4260 break;
4261 }
4262 }
4263
4264 /* found exact match, skip */
4265 if (found)
4266 continue;
4267
4268 newlist = lappend(newlist, varinfo);
4269 }
4270
4271 *varinfos = newlist;
4272 *ndistinct = item->ndistinct;
4273 return true;
4274 }
4275
4276 return false;
4277}
4278
4279/*
4280 * convert_to_scalar
4281 * Convert non-NULL values of the indicated types to the comparison
4282 * scale needed by scalarineqsel().
4283 * Returns "true" if successful.
4284 *
4285 * XXX this routine is a hack: ideally we should look up the conversion
4286 * subroutines in pg_type.
4287 *
4288 * All numeric datatypes are simply converted to their equivalent
4289 * "double" values. (NUMERIC values that are outside the range of "double"
4290 * are clamped to +/- HUGE_VAL.)
4291 *
4292 * String datatypes are converted by convert_string_to_scalar(),
4293 * which is explained below. The reason why this routine deals with
4294 * three values at a time, not just one, is that we need it for strings.
4295 *
4296 * The bytea datatype is just enough different from strings that it has
4297 * to be treated separately.
4298 *
4299 * The several datatypes representing absolute times are all converted
4300 * to Timestamp, which is actually an int64, and then we promote that to
4301 * a double. Note this will give correct results even for the "special"
4302 * values of Timestamp, since those are chosen to compare correctly;
4303 * see timestamp_cmp.
4304 *
4305 * The several datatypes representing relative times (intervals) are all
4306 * converted to measurements expressed in seconds.
4307 */
4308static bool
4309convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4310 Datum lobound, Datum hibound, Oid boundstypid,
4311 double *scaledlobound, double *scaledhibound)
4312{
4313 bool failure = false;
4314
4315 /*
4316 * Both the valuetypid and the boundstypid should exactly match the
4317 * declared input type(s) of the operator we are invoked for. However,
4318 * extensions might try to use scalarineqsel as estimator for operators
4319 * with input type(s) we don't handle here; in such cases, we want to
4320 * return false, not fail. In any case, we mustn't assume that valuetypid
4321 * and boundstypid are identical.
4322 *
4323 * XXX The histogram we are interpolating between points of could belong
4324 * to a column that's only binary-compatible with the declared type. In
4325 * essence we are assuming that the semantics of binary-compatible types
4326 * are enough alike that we can use a histogram generated with one type's
4327 * operators to estimate selectivity for the other's. This is outright
4328 * wrong in some cases --- in particular signed versus unsigned
4329 * interpretation could trip us up. But it's useful enough in the
4330 * majority of cases that we do it anyway. Should think about more
4331 * rigorous ways to do it.
4332 */
4333 switch (valuetypid)
4334 {
4335 /*
4336 * Built-in numeric types
4337 */
4338 case BOOLOID:
4339 case INT2OID:
4340 case INT4OID:
4341 case INT8OID:
4342 case FLOAT4OID:
4343 case FLOAT8OID:
4344 case NUMERICOID:
4345 case OIDOID:
4346 case REGPROCOID:
4347 case REGPROCEDUREOID:
4348 case REGOPEROID:
4349 case REGOPERATOROID:
4350 case REGCLASSOID:
4351 case REGTYPEOID:
4352 case REGCOLLATIONOID:
4353 case REGCONFIGOID:
4354 case REGDICTIONARYOID:
4355 case REGROLEOID:
4356 case REGNAMESPACEOID:
4357 *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4358 &failure);
4359 *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4360 &failure);
4361 *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4362 &failure);
4363 return !failure;
4364
4365 /*
4366 * Built-in string types
4367 */
4368 case CHAROID:
4369 case BPCHAROID:
4370 case VARCHAROID:
4371 case TEXTOID:
4372 case NAMEOID:
4373 {
4374 char *valstr = convert_string_datum(value, valuetypid,
4375 collid, &failure);
4376 char *lostr = convert_string_datum(lobound, boundstypid,
4377 collid, &failure);
4378 char *histr = convert_string_datum(hibound, boundstypid,
4379 collid, &failure);
4380
4381 /*
4382 * Bail out if any of the values is not of string type. We
4383 * might leak converted strings for the other value(s), but
4384 * that's not worth troubling over.
4385 */
4386 if (failure)
4387 return false;
4388
4389 convert_string_to_scalar(valstr, scaledvalue,
4390 lostr, scaledlobound,
4391 histr, scaledhibound);
4392 pfree(valstr);
4393 pfree(lostr);
4394 pfree(histr);
4395 return true;
4396 }
4397
4398 /*
4399 * Built-in bytea type
4400 */
4401 case BYTEAOID:
4402 {
4403 /* We only support bytea vs bytea comparison */
4404 if (boundstypid != BYTEAOID)
4405 return false;
4406 convert_bytea_to_scalar(value, scaledvalue,
4407 lobound, scaledlobound,
4408 hibound, scaledhibound);
4409 return true;
4410 }
4411
4412 /*
4413 * Built-in time types
4414 */
4415 case TIMESTAMPOID:
4416 case TIMESTAMPTZOID:
4417 case DATEOID:
4418 case INTERVALOID:
4419 case TIMEOID:
4420 case TIMETZOID:
4421 *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
4422 &failure);
4423 *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
4424 &failure);
4425 *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
4426 &failure);
4427 return !failure;
4428
4429 /*
4430 * Built-in network types
4431 */
4432 case INETOID:
4433 case CIDROID:
4434 case MACADDROID:
4435 case MACADDR8OID:
4436 *scaledvalue = convert_network_to_scalar(value, valuetypid,
4437 &failure);
4438 *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
4439 &failure);
4440 *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
4441 &failure);
4442 return !failure;
4443 }
4444 /* Don't know how to convert */
4445 *scaledvalue = *scaledlobound = *scaledhibound = 0;
4446 return false;
4447}
4448
4449/*
4450 * Do convert_to_scalar()'s work for any numeric data type.
4451 *
4452 * On failure (e.g., unsupported typid), set *failure to true;
4453 * otherwise, that variable is not changed.
4454 */
4455static double
4457{
4458 switch (typid)
4459 {
4460 case BOOLOID:
4461 return (double) DatumGetBool(value);
4462 case INT2OID:
4463 return (double) DatumGetInt16(value);
4464 case INT4OID:
4465 return (double) DatumGetInt32(value);
4466 case INT8OID:
4467 return (double) DatumGetInt64(value);
4468 case FLOAT4OID:
4469 return (double) DatumGetFloat4(value);
4470 case FLOAT8OID:
4471 return (double) DatumGetFloat8(value);
4472 case NUMERICOID:
4473 /* Note: out-of-range values will be clamped to +-HUGE_VAL */
4474 return (double)
4476 value));
4477 case OIDOID:
4478 case REGPROCOID:
4479 case REGPROCEDUREOID:
4480 case REGOPEROID:
4481 case REGOPERATOROID:
4482 case REGCLASSOID:
4483 case REGTYPEOID:
4484 case REGCOLLATIONOID:
4485 case REGCONFIGOID:
4486 case REGDICTIONARYOID:
4487 case REGROLEOID:
4488 case REGNAMESPACEOID:
4489 /* we can treat OIDs as integers... */
4490 return (double) DatumGetObjectId(value);
4491 }
4492
4493 *failure = true;
4494 return 0;
4495}
4496
4497/*
4498 * Do convert_to_scalar()'s work for any character-string data type.
4499 *
4500 * String datatypes are converted to a scale that ranges from 0 to 1,
4501 * where we visualize the bytes of the string as fractional digits.
4502 *
4503 * We do not want the base to be 256, however, since that tends to
4504 * generate inflated selectivity estimates; few databases will have
4505 * occurrences of all 256 possible byte values at each position.
4506 * Instead, use the smallest and largest byte values seen in the bounds
4507 * as the estimated range for each byte, after some fudging to deal with
4508 * the fact that we probably aren't going to see the full range that way.
4509 *
4510 * An additional refinement is that we discard any common prefix of the
4511 * three strings before computing the scaled values. This allows us to
4512 * "zoom in" when we encounter a narrow data range. An example is a phone
4513 * number database where all the values begin with the same area code.
4514 * (Actually, the bounds will be adjacent histogram-bin-boundary values,
4515 * so this is more likely to happen than you might think.)
4516 */
4517static void
4519 double *scaledvalue,
4520 char *lobound,
4521 double *scaledlobound,
4522 char *hibound,
4523 double *scaledhibound)
4524{
4525 int rangelo,
4526 rangehi;
4527 char *sptr;
4528
4529 rangelo = rangehi = (unsigned char) hibound[0];
4530 for (sptr = lobound; *sptr; sptr++)
4531 {
4532 if (rangelo > (unsigned char) *sptr)
4533 rangelo = (unsigned char) *sptr;
4534 if (rangehi < (unsigned char) *sptr)
4535 rangehi = (unsigned char) *sptr;
4536 }
4537 for (sptr = hibound; *sptr; sptr++)
4538 {
4539 if (rangelo > (unsigned char) *sptr)
4540 rangelo = (unsigned char) *sptr;
4541 if (rangehi < (unsigned char) *sptr)
4542 rangehi = (unsigned char) *sptr;
4543 }
4544 /* If range includes any upper-case ASCII chars, make it include all */
4545 if (rangelo <= 'Z' && rangehi >= 'A')
4546 {
4547 if (rangelo > 'A')
4548 rangelo = 'A';
4549 if (rangehi < 'Z')
4550 rangehi = 'Z';
4551 }
4552 /* Ditto lower-case */
4553 if (rangelo <= 'z' && rangehi >= 'a')
4554 {
4555 if (rangelo > 'a')
4556 rangelo = 'a';
4557 if (rangehi < 'z')
4558 rangehi = 'z';
4559 }
4560 /* Ditto digits */
4561 if (rangelo <= '9' && rangehi >= '0')
4562 {
4563 if (rangelo > '0')
4564 rangelo = '0';
4565 if (rangehi < '9')
4566 rangehi = '9';
4567 }
4568
4569 /*
4570 * If range includes less than 10 chars, assume we have not got enough
4571 * data, and make it include regular ASCII set.
4572 */
4573 if (rangehi - rangelo < 9)
4574 {
4575 rangelo = ' ';
4576 rangehi = 127;
4577 }
4578
4579 /*
4580 * Now strip any common prefix of the three strings.
4581 */
4582 while (*lobound)
4583 {
4584 if (*lobound != *hibound || *lobound != *value)
4585 break;
4586 lobound++, hibound++, value++;
4587 }
4588
4589 /*
4590 * Now we can do the conversions.
4591 */
4592 *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
4593 *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
4594 *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
4595}
4596
4597static double
4598convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
4599{
4600 int slen = strlen(value);
4601 double num,
4602 denom,
4603 base;
4604
4605 if (slen <= 0)
4606 return 0.0; /* empty string has scalar value 0 */
4607
4608 /*
4609 * There seems little point in considering more than a dozen bytes from
4610 * the string. Since base is at least 10, that will give us nominal
4611 * resolution of at least 12 decimal digits, which is surely far more
4612 * precision than this estimation technique has got anyway (especially in
4613 * non-C locales). Also, even with the maximum possible base of 256, this
4614 * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
4615 * overflow on any known machine.
4616 */
4617 if (slen > 12)
4618 slen = 12;
4619
4620 /* Convert initial characters to fraction */
4621 base = rangehi - rangelo + 1;
4622 num = 0.0;
4623 denom = base;
4624 while (slen-- > 0)
4625 {
4626 int ch = (unsigned char) *value++;
4627
4628 if (ch < rangelo)
4629 ch = rangelo - 1;
4630 else if (ch > rangehi)
4631 ch = rangehi + 1;
4632 num += ((double) (ch - rangelo)) / denom;
4633 denom *= base;
4634 }
4635
4636 return num;
4637}
4638
4639/*
4640 * Convert a string-type Datum into a palloc'd, null-terminated string.
4641 *
4642 * On failure (e.g., unsupported typid), set *failure to true;
4643 * otherwise, that variable is not changed. (We'll return NULL on failure.)
4644 *
4645 * When using a non-C locale, we must pass the string through pg_strxfrm()
4646 * before continuing, so as to generate correct locale-specific results.
4647 */
4648static char *
4650{
4651 char *val;
4652 pg_locale_t mylocale;
4653
4654 switch (typid)
4655 {
4656 case CHAROID:
4657 val = (char *) palloc(2);
4658 val[0] = DatumGetChar(value);
4659 val[1] = '\0';
4660 break;
4661 case BPCHAROID:
4662 case VARCHAROID:
4663 case TEXTOID:
4665 break;
4666 case NAMEOID:
4667 {
4669
4670 val = pstrdup(NameStr(*nm));
4671 break;
4672 }
4673 default:
4674 *failure = true;
4675 return NULL;
4676 }
4677
4679
4680 if (!mylocale->collate_is_c)
4681 {
4682 char *xfrmstr;
4683 size_t xfrmlen;
4684 size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
4685
4686 /*
4687 * XXX: We could guess at a suitable output buffer size and only call
4688 * pg_strxfrm() twice if our guess is too small.
4689 *
4690 * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
4691 * bogus data or set an error. This is not really a problem unless it
4692 * crashes since it will only give an estimation error and nothing
4693 * fatal.
4694 *
4695 * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
4696 * some cases, libc strxfrm() may return the wrong results, but that
4697 * will only lead to an estimation error.
4698 */
4699 xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
4700#ifdef WIN32
4701
4702 /*
4703 * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
4704 * of trying to allocate this much memory (and fail), just return the
4705 * original string unmodified as if we were in the C locale.
4706 */
4707 if (xfrmlen == INT_MAX)
4708 return val;
4709#endif
4710 xfrmstr = (char *) palloc(xfrmlen + 1);
4711 xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
4712
4713 /*
4714 * Some systems (e.g., glibc) can return a smaller value from the
4715 * second call than the first; thus the Assert must be <= not ==.
4716 */
4717 Assert(xfrmlen2 <= xfrmlen);
4718 pfree(val);
4719 val = xfrmstr;
4720 }
4721
4722 return val;
4723}
4724
4725/*
4726 * Do convert_to_scalar()'s work for any bytea data type.
4727 *
4728 * Very similar to convert_string_to_scalar except we can't assume
4729 * null-termination and therefore pass explicit lengths around.
4730 *
4731 * Also, assumptions about likely "normal" ranges of characters have been
4732 * removed - a data range of 0..255 is always used, for now. (Perhaps
4733 * someday we will add information about actual byte data range to
4734 * pg_statistic.)
4735 */
4736static void
4738 double *scaledvalue,
4739 Datum lobound,
4740 double *scaledlobound,
4741 Datum hibound,
4742 double *scaledhibound)
4743{
4744 bytea *valuep = DatumGetByteaPP(value);
4745 bytea *loboundp = DatumGetByteaPP(lobound);
4746 bytea *hiboundp = DatumGetByteaPP(hibound);
4747 int rangelo,
4748 rangehi,
4749 valuelen = VARSIZE_ANY_EXHDR(valuep),
4750 loboundlen = VARSIZE_ANY_EXHDR(loboundp),
4751 hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
4752 i,
4753 minlen;
4754 unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
4755 unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
4756 unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
4757
4758 /*
4759 * Assume bytea data is uniformly distributed across all byte values.
4760 */
4761 rangelo = 0;
4762 rangehi = 255;
4763
4764 /*
4765 * Now strip any common prefix of the three strings.
4766 */
4767 minlen = Min(Min(valuelen, loboundlen), hiboundlen);
4768 for (i = 0; i < minlen; i++)
4769 {
4770 if (*lostr != *histr || *lostr != *valstr)
4771 break;
4772 lostr++, histr++, valstr++;
4773 loboundlen--, hiboundlen--, valuelen--;
4774 }
4775
4776 /*
4777 * Now we can do the conversions.
4778 */
4779 *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
4780 *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
4781 *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
4782}
4783
4784static double
4785convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
4786 int rangelo, int rangehi)
4787{
4788 double num,
4789 denom,
4790 base;
4791
4792 if (valuelen <= 0)
4793 return 0.0; /* empty string has scalar value 0 */
4794
4795 /*
4796 * Since base is 256, need not consider more than about 10 chars (even
4797 * this many seems like overkill)
4798 */
4799 if (valuelen > 10)
4800 valuelen = 10;
4801
4802 /* Convert initial characters to fraction */
4803 base = rangehi - rangelo + 1;
4804 num = 0.0;
4805 denom = base;
4806 while (valuelen-- > 0)
4807 {
4808 int ch = *value++;
4809
4810 if (ch < rangelo)
4811 ch = rangelo - 1;
4812 else if (ch > rangehi)
4813 ch = rangehi + 1;
4814 num += ((double) (ch - rangelo)) / denom;
4815 denom *= base;
4816 }
4817
4818 return num;
4819}
4820
4821/*
4822 * Do convert_to_scalar()'s work for any timevalue data type.
4823 *
4824 * On failure (e.g., unsupported typid), set *failure to true;
4825 * otherwise, that variable is not changed.
4826 */
4827static double
4829{
4830 switch (typid)
4831 {
4832 case TIMESTAMPOID:
4833 return DatumGetTimestamp(value);
4834 case TIMESTAMPTZOID:
4835 return DatumGetTimestampTz(value);
4836 case DATEOID:
4838 case INTERVALOID:
4839 {
4841
4842 /*
4843 * Convert the month part of Interval to days using assumed
4844 * average month length of 365.25/12.0 days. Not too
4845 * accurate, but plenty good enough for our purposes.
4846 *
4847 * This also works for infinite intervals, which just have all
4848 * fields set to INT_MIN/INT_MAX, and so will produce a result
4849 * smaller/larger than any finite interval.
4850 */
4851 return interval->time + interval->day * (double) USECS_PER_DAY +
4853 }
4854 case TIMEOID:
4855 return DatumGetTimeADT(value);
4856 case TIMETZOID:
4857 {
4859
4860 /* use GMT-equivalent time */
4861 return (double) (timetz->time + (timetz->zone * 1000000.0));
4862 }
4863 }
4864
4865 *failure = true;
4866 return 0;
4867}
4868
4869
4870/*
4871 * get_restriction_variable
4872 * Examine the args of a restriction clause to see if it's of the
4873 * form (variable op pseudoconstant) or (pseudoconstant op variable),
4874 * where "variable" could be either a Var or an expression in vars of a
4875 * single relation. If so, extract information about the variable,
4876 * and also indicate which side it was on and the other argument.
4877 *
4878 * Inputs:
4879 * root: the planner info
4880 * args: clause argument list
4881 * varRelid: see specs for restriction selectivity functions
4882 *
4883 * Outputs: (these are valid only if true is returned)
4884 * *vardata: gets information about variable (see examine_variable)
4885 * *other: gets other clause argument, aggressively reduced to a constant
4886 * *varonleft: set true if variable is on the left, false if on the right
4887 *
4888 * Returns true if a variable is identified, otherwise false.
4889 *
4890 * Note: if there are Vars on both sides of the clause, we must fail, because
4891 * callers are expecting that the other side will act like a pseudoconstant.
4892 */
4893bool
4895 VariableStatData *vardata, Node **other,
4896 bool *varonleft)
4897{
4898 Node *left,
4899 *right;
4900 VariableStatData rdata;
4901
4902 /* Fail if not a binary opclause (probably shouldn't happen) */
4903 if (list_length(args) != 2)
4904 return false;
4905
4906 left = (Node *) linitial(args);
4907 right = (Node *) lsecond(args);
4908
4909 /*
4910 * Examine both sides. Note that when varRelid is nonzero, Vars of other
4911 * relations will be treated as pseudoconstants.
4912 */
4913 examine_variable(root, left, varRelid, vardata);
4914 examine_variable(root, right, varRelid, &rdata);
4915
4916 /*
4917 * If one side is a variable and the other not, we win.
4918 */
4919 if (vardata->rel && rdata.rel == NULL)
4920 {
4921 *varonleft = true;
4922 *other = estimate_expression_value(root, rdata.var);
4923 /* Assume we need no ReleaseVariableStats(rdata) here */
4924 return true;
4925 }
4926
4927 if (vardata->rel == NULL && rdata.rel)
4928 {
4929 *varonleft = false;
4930 *other = estimate_expression_value(root, vardata->var);
4931 /* Assume we need no ReleaseVariableStats(*vardata) here */
4932 *vardata = rdata;
4933 return true;
4934 }
4935
4936 /* Oops, clause has wrong structure (probably var op var) */
4937 ReleaseVariableStats(*vardata);
4938 ReleaseVariableStats(rdata);
4939
4940 return false;
4941}
4942
4943/*
4944 * get_join_variables
4945 * Apply examine_variable() to each side of a join clause.
4946 * Also, attempt to identify whether the join clause has the same
4947 * or reversed sense compared to the SpecialJoinInfo.
4948 *
4949 * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
4950 * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
4951 * where we can't tell for sure, we default to assuming it's normal.
4952 */
4953void
4955 VariableStatData *vardata1, VariableStatData *vardata2,
4956 bool *join_is_reversed)
4957{
4958 Node *left,
4959 *right;
4960
4961 if (list_length(args) != 2)
4962 elog(ERROR, "join operator should take two arguments");
4963
4964 left = (Node *) linitial(args);
4965 right = (Node *) lsecond(args);
4966
4967 examine_variable(root, left, 0, vardata1);
4968 examine_variable(root, right, 0, vardata2);
4969
4970 if (vardata1->rel &&
4971 bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
4972 *join_is_reversed = true; /* var1 is on RHS */
4973 else if (vardata2->rel &&
4974 bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
4975 *join_is_reversed = true; /* var2 is on LHS */
4976 else
4977 *join_is_reversed = false;
4978}
4979
4980/* statext_expressions_load copies the tuple, so just pfree it. */
4981static void
4983{
4984 pfree(tuple);
4985}
4986
4987/*
4988 * examine_variable
4989 * Try to look up statistical data about an expression.
4990 * Fill in a VariableStatData struct to describe the expression.
4991 *
4992 * Inputs:
4993 * root: the planner info
4994 * node: the expression tree to examine
4995 * varRelid: see specs for restriction selectivity functions
4996 *
4997 * Outputs: *vardata is filled as follows:
4998 * var: the input expression (with any binary relabeling stripped, if
4999 * it is or contains a variable; but otherwise the type is preserved)
5000 * rel: RelOptInfo for relation containing variable; NULL if expression
5001 * contains no Vars (NOTE this could point to a RelOptInfo of a
5002 * subquery, not one in the current query).
5003 * statsTuple: the pg_statistic entry for the variable, if one exists;
5004 * otherwise NULL.
5005 * freefunc: pointer to a function to release statsTuple with.
5006 * vartype: exposed type of the expression; this should always match
5007 * the declared input type of the operator we are estimating for.
5008 * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5009 * commonly the same as the exposed type of the variable argument,
5010 * but can be different in binary-compatible-type cases.
5011 * isunique: true if we were able to match the var to a unique index or a
5012 * single-column DISTINCT clause, implying its values are unique for
5013 * this query. (Caution: this should be trusted for statistical
5014 * purposes only, since we do not check indimmediate nor verify that
5015 * the exact same definition of equality applies.)
5016 * acl_ok: true if current user has permission to read the column(s)
5017 * underlying the pg_statistic entry. This is consulted by
5018 * statistic_proc_security_check().
5019 *
5020 * Caller is responsible for doing ReleaseVariableStats() before exiting.
5021 */
5022void
5024 VariableStatData *vardata)
5025{
5026 Node *basenode;
5027 Relids varnos;
5028 RelOptInfo *onerel;
5029
5030 /* Make sure we don't return dangling pointers in vardata */
5031 MemSet(vardata, 0, sizeof(VariableStatData));
5032
5033 /* Save the exposed type of the expression */
5034 vardata->vartype = exprType(node);
5035
5036 /* Look inside any binary-compatible relabeling */
5037
5038 if (IsA(node, RelabelType))
5039 basenode = (Node *) ((RelabelType *) node)->arg;
5040 else
5041 basenode = node;
5042
5043 /* Fast path for a simple Var */
5044
5045 if (IsA(basenode, Var) &&
5046 (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5047 {
5048 Var *var = (Var *) basenode;
5049
5050 /* Set up result fields other than the stats tuple */
5051 vardata->var = basenode; /* return Var without relabeling */
5052 vardata->rel = find_base_rel(root, var->varno);
5053 vardata->atttype = var->vartype;
5054 vardata->atttypmod = var->vartypmod;
5055 vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5056
5057 /* Try to locate some stats */
5058 examine_simple_variable(root, var, vardata);
5059
5060 return;
5061 }
5062
5063 /*
5064 * Okay, it's a more complicated expression. Determine variable
5065 * membership. Note that when varRelid isn't zero, only vars of that
5066 * relation are considered "real" vars.
5067 */
5068 varnos = pull_varnos(root, basenode);
5069
5070 onerel = NULL;
5071
5072 if (bms_is_empty(varnos))
5073 {
5074 /* No Vars at all ... must be pseudo-constant clause */
5075 }
5076 else
5077 {
5078 int relid;
5079
5080 if (bms_get_singleton_member(varnos, &relid))
5081 {
5082 if (varRelid == 0 || varRelid == relid)
5083 {
5084 onerel = find_base_rel(root, relid);
5085 vardata->rel = onerel;
5086 node = basenode; /* strip any relabeling */
5087 }
5088 /* else treat it as a constant */
5089 }
5090 else
5091 {
5092 /* varnos has multiple relids */
5093 if (varRelid == 0)
5094 {
5095 /* treat it as a variable of a join relation */
5096 vardata->rel = find_join_rel(root, varnos);
5097 node = basenode; /* strip any relabeling */
5098 }
5099 else if (bms_is_member(varRelid, varnos))
5100 {
5101 /* ignore the vars belonging to other relations */
5102 vardata->rel = find_base_rel(root, varRelid);
5103 node = basenode; /* strip any relabeling */
5104 /* note: no point in expressional-index search here */
5105 }
5106 /* else treat it as a constant */
5107 }
5108 }
5109
5110 bms_free(varnos);
5111
5112 vardata->var = node;
5113 vardata->atttype = exprType(node);
5114 vardata->atttypmod = exprTypmod(node);
5115
5116 if (onerel)
5117 {
5118 /*
5119 * We have an expression in vars of a single relation. Try to match
5120 * it to expressional index columns, in hopes of finding some
5121 * statistics.
5122 *
5123 * Note that we consider all index columns including INCLUDE columns,
5124 * since there could be stats for such columns. But the test for
5125 * uniqueness needs to be warier.
5126 *
5127 * XXX it's conceivable that there are multiple matches with different
5128 * index opfamilies; if so, we need to pick one that matches the
5129 * operator we are estimating for. FIXME later.
5130 */
5131 ListCell *ilist;
5132 ListCell *slist;
5133 Oid userid;
5134
5135 /*
5136 * Determine the user ID to use for privilege checks: either
5137 * onerel->userid if it's set (e.g., in case we're accessing the table
5138 * via a view), or the current user otherwise.
5139 *
5140 * If we drill down to child relations, we keep using the same userid:
5141 * it's going to be the same anyway, due to how we set up the relation
5142 * tree (q.v. build_simple_rel).
5143 */
5144 userid = OidIsValid(onerel->userid) ? onerel->userid : GetUserId();
5145
5146 foreach(ilist, onerel->indexlist)
5147 {
5148 IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5149 ListCell *indexpr_item;
5150 int pos;
5151
5152 indexpr_item = list_head(index->indexprs);
5153 if (indexpr_item == NULL)
5154 continue; /* no expressions here... */
5155
5156 for (pos = 0; pos < index->ncolumns; pos++)
5157 {
5158 if (index->indexkeys[pos] == 0)
5159 {
5160 Node *indexkey;
5161
5162 if (indexpr_item == NULL)
5163 elog(ERROR, "too few entries in indexprs list");
5164 indexkey = (Node *) lfirst(indexpr_item);
5165 if (indexkey && IsA(indexkey, RelabelType))
5166 indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5167 if (equal(node, indexkey))
5168 {
5169 /*
5170 * Found a match ... is it a unique index? Tests here
5171 * should match has_unique_index().
5172 */
5173 if (index->unique &&
5174 index->nkeycolumns == 1 &&
5175 pos == 0 &&
5176 (index->indpred == NIL || index->predOK))
5177 vardata->isunique = true;
5178
5179 /*
5180 * Has it got stats? We only consider stats for
5181 * non-partial indexes, since partial indexes probably
5182 * don't reflect whole-relation statistics; the above
5183 * check for uniqueness is the only info we take from
5184 * a partial index.
5185 *
5186 * An index stats hook, however, must make its own
5187 * decisions about what to do with partial indexes.
5188 */
5190 (*get_index_stats_hook) (root, index->indexoid,
5191 pos + 1, vardata))
5192 {
5193 /*
5194 * The hook took control of acquiring a stats
5195 * tuple. If it did supply a tuple, it'd better
5196 * have supplied a freefunc.
5197 */
5198 if (HeapTupleIsValid(vardata->statsTuple) &&
5199 !vardata->freefunc)
5200 elog(ERROR, "no function provided to release variable stats with");
5201 }
5202 else if (index->indpred == NIL)
5203 {
5204 vardata->statsTuple =
5205 SearchSysCache3(STATRELATTINH,
5206 ObjectIdGetDatum(index->indexoid),
5207 Int16GetDatum(pos + 1),
5208 BoolGetDatum(false));
5209 vardata->freefunc = ReleaseSysCache;
5210
5211 if (HeapTupleIsValid(vardata->statsTuple))
5212 {
5213 /* Get index's table for permission check */
5214 RangeTblEntry *rte;
5215
5216 rte = planner_rt_fetch(index->rel->relid, root);
5217 Assert(rte->rtekind == RTE_RELATION);
5218
5219 /*
5220 * For simplicity, we insist on the whole
5221 * table being selectable, rather than trying
5222 * to identify which column(s) the index
5223 * depends on. Also require all rows to be
5224 * selectable --- there must be no
5225 * securityQuals from security barrier views
5226 * or RLS policies.
5227 */
5228 vardata->acl_ok =
5229 rte->securityQuals == NIL &&
5230 (pg_class_aclcheck(rte->relid, userid,
5232
5233 /*
5234 * If the user doesn't have permissions to
5235 * access an inheritance child relation, check
5236 * the permissions of the table actually
5237 * mentioned in the query, since most likely
5238 * the user does have that permission. Note
5239 * that whole-table select privilege on the
5240 * parent doesn't quite guarantee that the
5241 * user could read all columns of the child.
5242 * But in practice it's unlikely that any
5243 * interesting security violation could result
5244 * from allowing access to the expression
5245 * index's stats, so we allow it anyway. See
5246 * similar code in examine_simple_variable()
5247 * for additional comments.
5248 */
5249 if (!vardata->acl_ok &&
5250 root->append_rel_array != NULL)
5251 {
5252 AppendRelInfo *appinfo;
5253 Index varno = index->rel->relid;
5254
5255 appinfo = root->append_rel_array[varno];
5256 while (appinfo &&
5258 root)->rtekind == RTE_RELATION)
5259 {
5260 varno = appinfo->parent_relid;
5261 appinfo = root->append_rel_array[varno];
5262 }
5263 if (varno != index->rel->relid)
5264 {
5265 /* Repeat access check on this rel */
5266 rte = planner_rt_fetch(varno, root);
5267 Assert(rte->rtekind == RTE_RELATION);
5268
5269 vardata->acl_ok =
5270 rte->securityQuals == NIL &&
5272 userid,
5274 }
5275 }
5276 }
5277 else
5278 {
5279 /* suppress leakproofness checks later */
5280 vardata->acl_ok = true;
5281 }
5282 }
5283 if (vardata->statsTuple)
5284 break;
5285 }
5286 indexpr_item = lnext(index->indexprs, indexpr_item);
5287 }
5288 }
5289 if (vardata->statsTuple)
5290 break;
5291 }
5292
5293 /*
5294 * Search extended statistics for one with a matching expression.
5295 * There might be multiple ones, so just grab the first one. In the
5296 * future, we might consider the statistics target (and pick the most
5297 * accurate statistics) and maybe some other parameters.
5298 */
5299 foreach(slist, onerel->statlist)
5300 {
5301 StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5302 RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5303 ListCell *expr_item;
5304 int pos;
5305
5306 /*
5307 * Stop once we've found statistics for the expression (either
5308 * from extended stats, or for an index in the preceding loop).
5309 */
5310 if (vardata->statsTuple)
5311 break;
5312
5313 /* skip stats without per-expression stats */
5314 if (info->kind != STATS_EXT_EXPRESSIONS)
5315 continue;
5316
5317 /* skip stats with mismatching stxdinherit value */
5318 if (info->inherit != rte->inh)
5319 continue;
5320
5321 pos = 0;
5322 foreach(expr_item, info->exprs)
5323 {
5324 Node *expr = (Node *) lfirst(expr_item);
5325
5326 Assert(expr);
5327
5328 /* strip RelabelType before comparing it */
5329 if (expr && IsA(expr, RelabelType))
5330 expr = (Node *) ((RelabelType *) expr)->arg;
5331
5332 /* found a match, see if we can extract pg_statistic row */
5333 if (equal(node, expr))
5334 {
5335 /*
5336 * XXX Not sure if we should cache the tuple somewhere.
5337 * Now we just create a new copy every time.
5338 */
5339 vardata->statsTuple =
5340 statext_expressions_load(info->statOid, rte->inh, pos);
5341
5342 vardata->freefunc = ReleaseDummy;
5343
5344 /*
5345 * For simplicity, we insist on the whole table being
5346 * selectable, rather than trying to identify which
5347 * column(s) the statistics object depends on. Also
5348 * require all rows to be selectable --- there must be no
5349 * securityQuals from security barrier views or RLS
5350 * policies.
5351 */
5352 vardata->acl_ok =
5353 rte->securityQuals == NIL &&
5354 (pg_class_aclcheck(rte->relid, userid,
5356
5357 /*
5358 * If the user doesn't have permissions to access an
5359 * inheritance child relation, check the permissions of
5360 * the table actually mentioned in the query, since most
5361 * likely the user does have that permission. Note that
5362 * whole-table select privilege on the parent doesn't
5363 * quite guarantee that the user could read all columns of
5364 * the child. But in practice it's unlikely that any
5365 * interesting security violation could result from
5366 * allowing access to the expression stats, so we allow it
5367 * anyway. See similar code in examine_simple_variable()
5368 * for additional comments.
5369 */
5370 if (!vardata->acl_ok &&
5371 root->append_rel_array != NULL)
5372 {
5373 AppendRelInfo *appinfo;
5374 Index varno = onerel->relid;
5375
5376 appinfo = root->append_rel_array[varno];
5377 while (appinfo &&
5379 root)->rtekind == RTE_RELATION)
5380 {
5381 varno = appinfo->parent_relid;
5382 appinfo = root->append_rel_array[varno];
5383 }
5384 if (varno != onerel->relid)
5385 {
5386 /* Repeat access check on this rel */
5387 rte = planner_rt_fetch(varno, root);
5388 Assert(rte->rtekind == RTE_RELATION);
5389
5390 vardata->acl_ok =
5391 rte->securityQuals == NIL &&
5393 userid,
5395 }
5396 }
5397
5398 break;
5399 }
5400
5401 pos++;
5402 }
5403 }
5404 }
5405}
5406
5407/*
5408 * examine_simple_variable
5409 * Handle a simple Var for examine_variable
5410 *
5411 * This is split out as a subroutine so that we can recurse to deal with
5412 * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
5413 *
5414 * We already filled in all the fields of *vardata except for the stats tuple.
5415 */
5416static void
5418 VariableStatData *vardata)
5419{
5420 RangeTblEntry *rte = root->simple_rte_array[var->varno];
5421
5422 Assert(IsA(rte, RangeTblEntry));
5423
5425 (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
5426 {
5427 /*
5428 * The hook took control of acquiring a stats tuple. If it did supply
5429 * a tuple, it'd better have supplied a freefunc.
5430 */
5431 if (HeapTupleIsValid(vardata->statsTuple) &&
5432 !vardata->freefunc)
5433 elog(ERROR, "no function provided to release variable stats with");
5434 }
5435 else if (rte->rtekind == RTE_RELATION)
5436 {
5437 /*
5438 * Plain table or parent of an inheritance appendrel, so look up the
5439 * column in pg_statistic
5440 */
5441 vardata->statsTuple = SearchSysCache3(STATRELATTINH,
5442 ObjectIdGetDatum(rte->relid),
5443 Int16GetDatum(var->varattno),
5444 BoolGetDatum(rte->inh));
5445 vardata->freefunc = ReleaseSysCache;
5446
5447 if (HeapTupleIsValid(vardata->statsTuple))
5448 {
5449 RelOptInfo *onerel = find_base_rel_noerr(root, var->varno);
5450 Oid userid;
5451
5452 /*
5453 * Check if user has permission to read this column. We require
5454 * all rows to be accessible, so there must be no securityQuals
5455 * from security barrier views or RLS policies.
5456 *
5457 * Normally the Var will have an associated RelOptInfo from which
5458 * we can find out which userid to do the check as; but it might
5459 * not if it's a RETURNING Var for an INSERT target relation. In
5460 * that case use the RTEPermissionInfo associated with the RTE.
5461 */
5462 if (onerel)
5463 userid = onerel->userid;
5464 else
5465 {
5466 RTEPermissionInfo *perminfo;
5467
5468 perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
5469 userid = perminfo->checkAsUser;
5470 }
5471 if (!OidIsValid(userid))
5472 userid = GetUserId();
5473
5474 vardata->acl_ok =
5475 rte->securityQuals == NIL &&
5476 ((pg_class_aclcheck(rte->relid, userid,
5477 ACL_SELECT) == ACLCHECK_OK) ||
5478 (pg_attribute_aclcheck(rte->relid, var->varattno, userid,
5479 ACL_SELECT) == ACLCHECK_OK));
5480
5481 /*
5482 * If the user doesn't have permissions to access an inheritance
5483 * child relation or specifically this attribute, check the
5484 * permissions of the table/column actually mentioned in the
5485 * query, since most likely the user does have that permission
5486 * (else the query will fail at runtime), and if the user can read
5487 * the column there then he can get the values of the child table
5488 * too. To do that, we must find out which of the root parent's
5489 * attributes the child relation's attribute corresponds to.
5490 */
5491 if (!vardata->acl_ok && var->varattno > 0 &&
5492 root->append_rel_array != NULL)
5493 {
5494 AppendRelInfo *appinfo;
5495 Index varno = var->varno;
5496 int varattno = var->varattno;
5497 bool found = false;
5498
5499 appinfo = root->append_rel_array[varno];
5500
5501 /*
5502 * Partitions are mapped to their immediate parent, not the
5503 * root parent, so must be ready to walk up multiple
5504 * AppendRelInfos. But stop if we hit a parent that is not
5505 * RTE_RELATION --- that's a flattened UNION ALL subquery, not
5506 * an inheritance parent.
5507 */
5508 while (appinfo &&
5510 root)->rtekind == RTE_RELATION)
5511 {
5512 int parent_varattno;
5513
5514 found = false;
5515 if (varattno <= 0 || varattno > appinfo->num_child_cols)
5516 break; /* safety check */
5517 parent_varattno = appinfo->parent_colnos[varattno - 1];
5518 if (parent_varattno == 0)
5519 break; /* Var is local to child */
5520
5521 varno = appinfo->parent_relid;
5522 varattno = parent_varattno;
5523 found = true;
5524
5525 /* If the parent is itself a child, continue up. */
5526 appinfo = root->append_rel_array[varno];
5527 }
5528
5529 /*
5530 * In rare cases, the Var may be local to the child table, in
5531 * which case, we've got to live with having no access to this
5532 * column's stats.
5533 */
5534 if (!found)
5535 return;
5536
5537 /* Repeat the access check on this parent rel & column */
5538 rte = planner_rt_fetch(varno, root);
5539 Assert(rte->rtekind == RTE_RELATION);
5540
5541 /*
5542 * Fine to use the same userid as it's the same in all
5543 * relations of a given inheritance tree.
5544 */
5545 vardata->acl_ok =
5546 rte->securityQuals == NIL &&
5547 ((pg_class_aclcheck(rte->relid, userid,
5548 ACL_SELECT) == ACLCHECK_OK) ||
5549 (pg_attribute_aclcheck(rte->relid, varattno, userid,
5550 ACL_SELECT) == ACLCHECK_OK));
5551 }
5552 }
5553 else
5554 {
5555 /* suppress any possible leakproofness checks later */
5556 vardata->acl_ok = true;
5557 }
5558 }
5559 else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
5560 (rte->rtekind == RTE_CTE && !rte->self_reference))
5561 {
5562 /*
5563 * Plain subquery (not one that was converted to an appendrel) or
5564 * non-recursive CTE. In either case, we can try to find out what the
5565 * Var refers to within the subquery. We skip this for appendrel and
5566 * recursive-CTE cases because any column stats we did find would
5567 * likely not be very relevant.
5568 */
5569 PlannerInfo *subroot;
5570 Query *subquery;
5571 List *subtlist;
5572 TargetEntry *ste;
5573
5574 /*
5575 * Punt if it's a whole-row var rather than a plain column reference.
5576 */
5577 if (var->varattno == InvalidAttrNumber)
5578 return;
5579
5580 /*
5581 * Otherwise, find the subquery's planner subroot.
5582 */
5583 if (rte->rtekind == RTE_SUBQUERY)
5584 {
5585 RelOptInfo *rel;
5586
5587 /*
5588 * Fetch RelOptInfo for subquery. Note that we don't change the
5589 * rel returned in vardata, since caller expects it to be a rel of
5590 * the caller's query level. Because we might already be
5591 * recursing, we can't use that rel pointer either, but have to
5592 * look up the Var's rel afresh.
5593 */
5594 rel = find_base_rel(root, var->varno);
5595
5596 subroot = rel->subroot;
5597 }
5598 else
5599 {
5600 /* CTE case is more difficult */
5601 PlannerInfo *cteroot;
5602 Index levelsup;
5603 int ndx;
5604 int plan_id;
5605 ListCell *lc;
5606
5607 /*
5608 * Find the referenced CTE, and locate the subroot previously made
5609 * for it.
5610 */
5611 levelsup = rte->ctelevelsup;
5612 cteroot = root;
5613 while (levelsup-- > 0)
5614 {
5615 cteroot = cteroot->parent_root;
5616 if (!cteroot) /* shouldn't happen */
5617 elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
5618 }
5619
5620 /*
5621 * Note: cte_plan_ids can be shorter than cteList, if we are still
5622 * working on planning the CTEs (ie, this is a side-reference from
5623 * another CTE). So we mustn't use forboth here.
5624 */
5625 ndx = 0;
5626 foreach(lc, cteroot->parse->cteList)
5627 {
5628 CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
5629
5630 if (strcmp(cte->ctename, rte->ctename) == 0)
5631 break;
5632 ndx++;
5633 }
5634 if (lc == NULL) /* shouldn't happen */
5635 elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
5636 if (ndx >= list_length(cteroot->cte_plan_ids))
5637 elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
5638 plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
5639 if (plan_id <= 0)
5640 elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
5641 subroot = list_nth(root->glob->subroots, plan_id - 1);
5642 }
5643
5644 /* If the subquery hasn't been planned yet, we have to punt */
5645 if (subroot == NULL)
5646 return;
5647 Assert(IsA(subroot, PlannerInfo));
5648
5649 /*
5650 * We must use the subquery parsetree as mangled by the planner, not
5651 * the raw version from the RTE, because we need a Var that will refer
5652 * to the subroot's live RelOptInfos. For instance, if any subquery
5653 * pullup happened during planning, Vars in the targetlist might have
5654 * gotten replaced, and we need to see the replacement expressions.
5655 */
5656 subquery = subroot->parse;
5657 Assert(IsA(subquery, Query));
5658
5659 /*
5660 * Punt if subquery uses set operations or GROUP BY, as these will
5661 * mash underlying columns' stats beyond recognition. (Set ops are
5662 * particularly nasty; if we forged ahead, we would return stats
5663 * relevant to only the leftmost subselect...) DISTINCT is also
5664 * problematic, but we check that later because there is a possibility
5665 * of learning something even with it.
5666 */
5667 if (subquery->setOperations ||
5668 subquery->groupClause ||
5669 subquery->groupingSets)
5670 return;
5671
5672 /* Get the subquery output expression referenced by the upper Var */
5673 if (subquery->returningList)
5674 subtlist = subquery->returningList;
5675 else
5676 subtlist = subquery->targetList;
5677 ste = get_tle_by_resno(subtlist, var->varattno);
5678 if (ste == NULL || ste->resjunk)
5679 elog(ERROR, "subquery %s does not have attribute %d",
5680 rte->eref->aliasname, var->varattno);
5681 var = (Var *) ste->expr;
5682
5683 /*
5684 * If subquery uses DISTINCT, we can't make use of any stats for the
5685 * variable ... but, if it's the only DISTINCT column, we are entitled
5686 * to consider it unique. We do the test this way so that it works
5687 * for cases involving DISTINCT ON.
5688 */
5689 if (subquery->distinctClause)
5690 {
5691 if (list_length(subquery->distinctClause) == 1 &&
5693 vardata->isunique = true;
5694 /* cannot go further */
5695 return;
5696 }
5697
5698 /*
5699 * If the sub-query originated from a view with the security_barrier
5700 * attribute, we must not look at the variable's statistics, though it
5701 * seems all right to notice the existence of a DISTINCT clause. So
5702 * stop here.
5703 *
5704 * This is probably a harsher restriction than necessary; it's
5705 * certainly OK for the selectivity estimator (which is a C function,
5706 * and therefore omnipotent anyway) to look at the statistics. But
5707 * many selectivity estimators will happily *invoke the operator
5708 * function* to try to work out a good estimate - and that's not OK.
5709 * So for now, don't dig down for stats.
5710 */
5711 if (rte->security_barrier)
5712 return;
5713
5714 /* Can only handle a simple Var of subquery's query level */
5715 if (var && IsA(var, Var) &&
5716 var->varlevelsup == 0)
5717 {
5718 /*
5719 * OK, recurse into the subquery. Note that the original setting
5720 * of vardata->isunique (which will surely be false) is left
5721 * unchanged in this situation. That's what we want, since even
5722 * if the underlying column is unique, the subquery may have
5723 * joined to other tables in a way that creates duplicates.
5724 */
5725 examine_simple_variable(subroot, var, vardata);
5726 }
5727 }
5728 else
5729 {
5730 /*
5731 * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
5732 * see RTE_JOIN here because join alias Vars have already been
5733 * flattened.) There's not much we can do with function outputs, but
5734 * maybe someday try to be smarter about VALUES.
5735 */
5736 }
5737}
5738
5739/*
5740 * Check whether it is permitted to call func_oid passing some of the
5741 * pg_statistic data in vardata. We allow this either if the user has SELECT
5742 * privileges on the table or column underlying the pg_statistic data or if
5743 * the function is marked leak-proof.
5744 */
5745bool
5747{
5748 if (vardata->acl_ok)
5749 return true;
5750
5751 if (!OidIsValid(func_oid))
5752 return false;
5753
5754 if (get_func_leakproof(func_oid))
5755 return true;
5756
5758 (errmsg_internal("not using statistics because function \"%s\" is not leak-proof",
5759 get_func_name(func_oid))));
5760 return false;
5761}
5762
5763/*
5764 * get_variable_numdistinct
5765 * Estimate the number of distinct values of a variable.
5766 *
5767 * vardata: results of examine_variable
5768 * *isdefault: set to true if the result is a default rather than based on
5769 * anything meaningful.
5770 *
5771 * NB: be careful to produce a positive integral result, since callers may
5772 * compare the result to exact integer counts, or might divide by it.
5773 */
5774double
5776{
5777 double stadistinct;
5778 double stanullfrac = 0.0;
5779 double ntuples;
5780
5781 *isdefault = false;
5782
5783 /*
5784 * Determine the stadistinct value to use. There are cases where we can
5785 * get an estimate even without a pg_statistic entry, or can get a better
5786 * value than is in pg_statistic. Grab stanullfrac too if we can find it
5787 * (otherwise, assume no nulls, for lack of any better idea).
5788 */
5789 if (HeapTupleIsValid(vardata->statsTuple))
5790 {
5791 /* Use the pg_statistic entry */
5792 Form_pg_statistic stats;
5793
5794 stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
5795 stadistinct = stats->stadistinct;
5796 stanullfrac = stats->stanullfrac;
5797 }
5798 else if (vardata->vartype == BOOLOID)
5799 {
5800 /*
5801 * Special-case boolean columns: presumably, two distinct values.
5802 *
5803 * Are there any other datatypes we should wire in special estimates
5804 * for?
5805 */
5806 stadistinct = 2.0;
5807 }
5808 else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
5809 {
5810 /*
5811 * If the Var represents a column of a VALUES RTE, assume it's unique.
5812 * This could of course be very wrong, but it should tend to be true
5813 * in well-written queries. We could consider examining the VALUES'
5814 * contents to get some real statistics; but that only works if the
5815 * entries are all constants, and it would be pretty expensive anyway.
5816 */
5817 stadistinct = -1.0; /* unique (and all non null) */
5818 }
5819 else
5820 {
5821 /*
5822 * We don't keep statistics for system columns, but in some cases we
5823 * can infer distinctness anyway.
5824 */
5825 if (vardata->var && IsA(vardata->var, Var))
5826 {
5827 switch (((Var *) vardata->var)->varattno)
5828 {
5830 stadistinct = -1.0; /* unique (and all non null) */
5831 break;
5833 stadistinct = 1.0; /* only 1 value */
5834 break;
5835 default:
5836 stadistinct = 0.0; /* means "unknown" */
5837 break;
5838 }
5839 }
5840 else
5841 stadistinct = 0.0; /* means "unknown" */
5842
5843 /*
5844 * XXX consider using estimate_num_groups on expressions?
5845 */
5846 }
5847
5848 /*
5849 * If there is a unique index or DISTINCT clause for the variable, assume
5850 * it is unique no matter what pg_statistic says; the statistics could be
5851 * out of date, or we might have found a partial unique index that proves
5852 * the var is unique for this query. However, we'd better still believe
5853 * the null-fraction statistic.
5854 */
5855 if (vardata->isunique)
5856 stadistinct = -1.0 * (1.0 - stanullfrac);
5857
5858 /*
5859 * If we had an absolute estimate, use that.
5860 */
5861 if (stadistinct > 0.0)
5862 return clamp_row_est(stadistinct);
5863
5864 /*
5865 * Otherwise we need to get the relation size; punt if not available.
5866 */
5867 if (vardata->rel == NULL)
5868 {
5869 *isdefault = true;
5870 return DEFAULT_NUM_DISTINCT;
5871 }
5872 ntuples = vardata->rel->tuples;
5873 if (ntuples <= 0.0)
5874 {
5875 *isdefault = true;
5876 return DEFAULT_NUM_DISTINCT;
5877 }
5878
5879 /*
5880 * If we had a relative estimate, use that.
5881 */
5882 if (stadistinct < 0.0)
5883 return clamp_row_est(-stadistinct * ntuples);
5884
5885 /*
5886 * With no data, estimate ndistinct = ntuples if the table is small, else
5887 * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
5888 * that the behavior isn't discontinuous.
5889 */
5890 if (ntuples < DEFAULT_NUM_DISTINCT)
5891 return clamp_row_est(ntuples);
5892
5893 *isdefault = true;
5894 return DEFAULT_NUM_DISTINCT;
5895}
5896
5897/*
5898 * get_variable_range
5899 * Estimate the minimum and maximum value of the specified variable.
5900 * If successful, store values in *min and *max, and return true.
5901 * If no data available, return false.
5902 *
5903 * sortop is the "<" comparison operator to use. This should generally
5904 * be "<" not ">", as only the former is likely to be found in pg_statistic.
5905 * The collation must be specified too.
5906 */
5907static bool
5909 Oid sortop, Oid collation,
5910 Datum *min, Datum *max)
5911{
5912 Datum tmin = 0;
5913 Datum tmax = 0;
5914 bool have_data = false;
5915 int16 typLen;
5916 bool typByVal;
5917 Oid opfuncoid;
5918 FmgrInfo opproc;
5919 AttStatsSlot sslot;
5920
5921 /*
5922 * XXX It's very tempting to try to use the actual column min and max, if
5923 * we can get them relatively-cheaply with an index probe. However, since
5924 * this function is called many times during join planning, that could
5925 * have unpleasant effects on planning speed. Need more investigation
5926 * before enabling this.
5927 */
5928#ifdef NOT_USED
5929 if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
5930 return true;
5931#endif
5932
5933 if (!HeapTupleIsValid(vardata->statsTuple))
5934 {
5935 /* no stats available, so default result */
5936 return false;
5937 }
5938
5939 /*
5940 * If we can't apply the sortop to the stats data, just fail. In
5941 * principle, if there's a histogram and no MCVs, we could return the
5942 * histogram endpoints without ever applying the sortop ... but it's
5943 * probably not worth trying, because whatever the caller wants to do with
5944 * the endpoints would likely fail the security check too.
5945 */
5946 if (!statistic_proc_security_check(vardata,
5947 (opfuncoid = get_opcode(sortop))))
5948 return false;
5949
5950 opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
5951
5952 get_typlenbyval(vardata->atttype, &typLen, &typByVal);
5953
5954 /*
5955 * If there is a histogram with the ordering we want, grab the first and
5956 * last values.
5957 */
5958 if (get_attstatsslot(&sslot, vardata->statsTuple,
5959 STATISTIC_KIND_HISTOGRAM, sortop,
5961 {
5962 if (sslot.stacoll == collation && sslot.nvalues > 0)
5963 {
5964 tmin = datumCopy(sslot.values[0], typByVal, typLen);
5965 tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
5966 have_data = true;
5967 }
5968 free_attstatsslot(&sslot);
5969 }
5970
5971 /*
5972 * Otherwise, if there is a histogram with some other ordering, scan it
5973 * and get the min and max values according to the ordering we want. This
5974 * of course may not find values that are really extremal according to our
5975 * ordering, but it beats ignoring available data.
5976 */
5977 if (!have_data &&
5978 get_attstatsslot(&sslot, vardata->statsTuple,
5979 STATISTIC_KIND_HISTOGRAM, InvalidOid,
5981 {
5982 get_stats_slot_range(&sslot, opfuncoid, &opproc,
5983 collation, typLen, typByVal,
5984 &tmin, &tmax, &have_data);
5985 free_attstatsslot(&sslot);
5986 }
5987
5988 /*
5989 * If we have most-common-values info, look for extreme MCVs. This is
5990 * needed even if we also have a histogram, since the histogram excludes
5991 * the MCVs. However, if we *only* have MCVs and no histogram, we should
5992 * be pretty wary of deciding that that is a full representation of the
5993 * data. Proceed only if the MCVs represent the whole table (to within
5994 * roundoff error).
5995 */
5996 if (get_attstatsslot(&sslot, vardata->statsTuple,
5997 STATISTIC_KIND_MCV, InvalidOid,
5998 have_data ? ATTSTATSSLOT_VALUES :
6000 {
6001 bool use_mcvs = have_data;
6002
6003 if (!have_data)
6004 {
6005 double sumcommon = 0.0;
6006 double nullfrac;
6007 int i;
6008
6009 for (i = 0; i < sslot.nnumbers; i++)
6010 sumcommon += sslot.numbers[i];
6011 nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6012 if (sumcommon + nullfrac > 0.99999)
6013 use_mcvs = true;
6014 }
6015
6016 if (use_mcvs)
6017 get_stats_slot_range(&sslot, opfuncoid, &opproc,
6018 collation, typLen, typByVal,
6019 &tmin, &tmax, &have_data);
6020 free_attstatsslot(&sslot);
6021 }
6022
6023 *min = tmin;
6024 *max = tmax;
6025 return have_data;
6026}
6027
6028/*
6029 * get_stats_slot_range: scan sslot for min/max values
6030 *
6031 * Subroutine for get_variable_range: update min/max/have_data according
6032 * to what we find in the statistics array.
6033 */
6034static void
6036 Oid collation, int16 typLen, bool typByVal,
6037 Datum *min, Datum *max, bool *p_have_data)
6038{
6039 Datum tmin = *min;
6040 Datum tmax = *max;
6041 bool have_data = *p_have_data;
6042 bool found_tmin = false;
6043 bool found_tmax = false;
6044
6045 /* Look up the comparison function, if we didn't already do so */
6046 if (opproc->fn_oid != opfuncoid)
6047 fmgr_info(opfuncoid, opproc);
6048
6049 /* Scan all the slot's values */
6050 for (int i = 0; i < sslot->nvalues; i++)
6051 {
6052 if (!have_data)
6053 {
6054 tmin = tmax = sslot->values[i];
6055 found_tmin = found_tmax = true;
6056 *p_have_data = have_data = true;
6057 continue;
6058 }
6060 collation,
6061 sslot->values[i], tmin)))
6062 {
6063 tmin = sslot->values[i];
6064 found_tmin = true;
6065 }
6067 collation,
6068 tmax, sslot->values[i])))
6069 {
6070 tmax = sslot->values[i];
6071 found_tmax = true;
6072 }
6073 }
6074
6075 /*
6076 * Copy the slot's values, if we found new extreme values.
6077 */
6078 if (found_tmin)
6079 *min = datumCopy(tmin, typByVal, typLen);
6080 if (found_tmax)
6081 *max = datumCopy(tmax, typByVal, typLen);
6082}
6083
6084
6085/*
6086 * get_actual_variable_range
6087 * Attempt to identify the current *actual* minimum and/or maximum
6088 * of the specified variable, by looking for a suitable btree index
6089 * and fetching its low and/or high values.
6090 * If successful, store values in *min and *max, and return true.
6091 * (Either pointer can be NULL if that endpoint isn't needed.)
6092 * If unsuccessful, return false.
6093 *
6094 * sortop is the "<" comparison operator to use.
6095 * collation is the required collation.
6096 */
6097static bool
6099 Oid sortop, Oid collation,
6100 Datum *min, Datum *max)
6101{
6102 bool have_data = false;
6103 RelOptInfo *rel = vardata->rel;
6104 RangeTblEntry *rte;
6105 ListCell *lc;
6106
6107 /* No hope if no relation or it doesn't have indexes */
6108 if (rel == NULL || rel->indexlist == NIL)
6109 return false;
6110 /* If it has indexes it must be a plain relation */
6111 rte = root->simple_rte_array[rel->relid];
6112 Assert(rte->rtekind == RTE_RELATION);
6113
6114 /* ignore partitioned tables. Any indexes here are not real indexes */
6115 if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6116 return false;
6117
6118 /* Search through the indexes to see if any match our problem */
6119 foreach(lc, rel->indexlist)
6120 {
6122 ScanDirection indexscandir;
6123
6124 /* Ignore non-btree indexes */
6125 if (index->relam != BTREE_AM_OID)
6126 continue;
6127
6128 /*
6129 * Ignore partial indexes --- we only want stats that cover the entire
6130 * relation.
6131 */
6132 if (index->indpred != NIL)
6133 continue;
6134
6135 /*
6136 * The index list might include hypothetical indexes inserted by a
6137 * get_relation_info hook --- don't try to access them.
6138 */
6139 if (index->hypothetical)
6140 continue;
6141
6142 /*
6143 * The first index column must match the desired variable, sortop, and
6144 * collation --- but we can use a descending-order index.
6145 */
6146 if (collation != index->indexcollations[0])
6147 continue; /* test first 'cause it's cheapest */
6148 if (!match_index_to_operand(vardata->var, 0, index))
6149 continue;
6150 switch (get_op_opfamily_strategy(sortop, index->sortopfamily[0]))
6151 {
6153 if (index->reverse_sort[0])
6154 indexscandir = BackwardScanDirection;
6155 else
6156 indexscandir = ForwardScanDirection;
6157 break;
6159 if (index->reverse_sort[0])
6160 indexscandir = ForwardScanDirection;
6161 else
6162 indexscandir = BackwardScanDirection;
6163 break;
6164 default:
6165 /* index doesn't match the sortop */
6166 continue;
6167 }
6168
6169 /*
6170 * Found a suitable index to extract data from. Set up some data that
6171 * can be used by both invocations of get_actual_variable_endpoint.
6172 */
6173 {
6174 MemoryContext tmpcontext;
6175 MemoryContext oldcontext;
6176 Relation heapRel;
6177 Relation indexRel;
6178 TupleTableSlot *slot;
6179 int16 typLen;
6180 bool typByVal;
6181 ScanKeyData scankeys[1];
6182
6183 /* Make sure any cruft gets recycled when we're done */
6185 "get_actual_variable_range workspace",
6187 oldcontext = MemoryContextSwitchTo(tmpcontext);
6188
6189 /*
6190 * Open the table and index so we can read from them. We should
6191 * already have some type of lock on each.
6192 */
6193 heapRel = table_open(rte->relid, NoLock);
6194 indexRel = index_open(index->indexoid, NoLock);
6195
6196 /* build some stuff needed for indexscan execution */
6197 slot = table_slot_create(heapRel, NULL);
6198 get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6199
6200 /* set up an IS NOT NULL scan key so that we ignore nulls */
6201 ScanKeyEntryInitialize(&scankeys[0],
6203 1, /* index col to scan */
6204 InvalidStrategy, /* no strategy */
6205 InvalidOid, /* no strategy subtype */
6206 InvalidOid, /* no collation */
6207 InvalidOid, /* no reg proc for this */
6208 (Datum) 0); /* constant */
6209
6210 /* If min is requested ... */
6211 if (min)
6212 {
6213 have_data = get_actual_variable_endpoint(heapRel,
6214 indexRel,
6215 indexscandir,
6216 scankeys,
6217 typLen,
6218 typByVal,
6219 slot,
6220 oldcontext,
6221 min);
6222 }
6223 else
6224 {
6225 /* If min not requested, still want to fetch max */
6226 have_data = true;
6227 }
6228
6229 /* If max is requested, and we didn't already fail ... */
6230 if (max && have_data)
6231 {
6232 /* scan in the opposite direction; all else is the same */
6233 have_data = get_actual_variable_endpoint(heapRel,
6234 indexRel,
6235 -indexscandir,
6236 scankeys,
6237 typLen,
6238 typByVal,
6239 slot,
6240 oldcontext,
6241 max);
6242 }
6243
6244 /* Clean everything up */
6246
6247 index_close(indexRel, NoLock);
6248 table_close(heapRel, NoLock);
6249
6250 MemoryContextSwitchTo(oldcontext);
6251 MemoryContextDelete(tmpcontext);
6252
6253 /* And we're done */
6254 break;
6255 }
6256 }
6257
6258 return have_data;
6259}
6260
6261/*
6262 * Get one endpoint datum (min or max depending on indexscandir) from the
6263 * specified index. Return true if successful, false if not.
6264 * On success, endpoint value is stored to *endpointDatum (and copied into
6265 * outercontext).
6266 *
6267 * scankeys is a 1-element scankey array set up to reject nulls.
6268 * typLen/typByVal describe the datatype of the index's first column.
6269 * tableslot is a slot suitable to hold table tuples, in case we need
6270 * to probe the heap.
6271 * (We could compute these values locally, but that would mean computing them
6272 * twice when get_actual_variable_range needs both the min and the max.)
6273 *
6274 * Failure occurs either when the index is empty, or we decide that it's
6275 * taking too long to find a suitable tuple.
6276 */
6277static bool
6279 Relation indexRel,
6280 ScanDirection indexscandir,
6281 ScanKey scankeys,
6282 int16 typLen,
6283 bool typByVal,
6284 TupleTableSlot *tableslot,
6285 MemoryContext outercontext,
6286 Datum *endpointDatum)
6287{
6288 bool have_data = false;
6289 SnapshotData SnapshotNonVacuumable;
6290 IndexScanDesc index_scan;
6291 Buffer vmbuffer = InvalidBuffer;
6292 BlockNumber last_heap_block = InvalidBlockNumber;
6293 int n_visited_heap_pages = 0;
6294 ItemPointer tid;
6296 bool isnull[INDEX_MAX_KEYS];
6297 MemoryContext oldcontext;
6298
6299 /*
6300 * We use the index-only-scan machinery for this. With mostly-static
6301 * tables that's a win because it avoids a heap visit. It's also a win
6302 * for dynamic data, but the reason is less obvious; read on for details.
6303 *
6304 * In principle, we should scan the index with our current active
6305 * snapshot, which is the best approximation we've got to what the query
6306 * will see when executed. But that won't be exact if a new snap is taken
6307 * before running the query, and it can be very expensive if a lot of
6308 * recently-dead or uncommitted rows exist at the beginning or end of the
6309 * index (because we'll laboriously fetch each one and reject it).
6310 * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
6311 * and uncommitted rows as well as normal visible rows. On the other
6312 * hand, it will reject known-dead rows, and thus not give a bogus answer
6313 * when the extreme value has been deleted (unless the deletion was quite
6314 * recent); that case motivates not using SnapshotAny here.
6315 *
6316 * A crucial point here is that SnapshotNonVacuumable, with
6317 * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
6318 * condition that the indexscan will use to decide that index entries are
6319 * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
6320 * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
6321 * have to continue scanning past it, we know that the indexscan will mark
6322 * that index entry killed. That means that the next
6323 * get_actual_variable_endpoint() call will not have to re-consider that
6324 * index entry. In this way we avoid repetitive work when this function
6325 * is used a lot during planning.
6326 *
6327 * But using SnapshotNonVacuumable creates a hazard of its own. In a
6328 * recently-created index, some index entries may point at "broken" HOT
6329 * chains in which not all the tuple versions contain data matching the
6330 * index entry. The live tuple version(s) certainly do match the index,
6331 * but SnapshotNonVacuumable can accept recently-dead tuple versions that
6332 * don't match. Hence, if we took data from the selected heap tuple, we
6333 * might get a bogus answer that's not close to the index extremal value,
6334 * or could even be NULL. We avoid this hazard because we take the data
6335 * from the index entry not the heap.
6336 *
6337 * Despite all this care, there are situations where we might find many
6338 * non-visible tuples near the end of the index. We don't want to expend
6339 * a huge amount of time here, so we give up once we've read too many heap
6340 * pages. When we fail for that reason, the caller will end up using
6341 * whatever extremal value is recorded in pg_statistic.
6342 */
6343 InitNonVacuumableSnapshot(SnapshotNonVacuumable,
6344 GlobalVisTestFor(heapRel));
6345
6346 index_scan = index_beginscan(heapRel, indexRel,
6347 &SnapshotNonVacuumable,
6348 1, 0);
6349 /* Set it up for index-only scan */
6350 index_scan->xs_want_itup = true;
6351 index_rescan(index_scan, scankeys, 1, NULL, 0);
6352
6353 /* Fetch first/next tuple in specified direction */
6354 while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
6355 {
6357
6358 if (!VM_ALL_VISIBLE(heapRel,
6359 block,
6360 &vmbuffer))
6361 {
6362 /* Rats, we have to visit the heap to check visibility */
6363 if (!index_fetch_heap(index_scan, tableslot))
6364 {
6365 /*
6366 * No visible tuple for this index entry, so we need to
6367 * advance to the next entry. Before doing so, count heap
6368 * page fetches and give up if we've done too many.
6369 *
6370 * We don't charge a page fetch if this is the same heap page
6371 * as the previous tuple. This is on the conservative side,
6372 * since other recently-accessed pages are probably still in
6373 * buffers too; but it's good enough for this heuristic.
6374 */
6375#define VISITED_PAGES_LIMIT 100
6376
6377 if (block != last_heap_block)
6378 {
6379 last_heap_block = block;
6380 n_visited_heap_pages++;
6381 if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
6382 break;
6383 }
6384
6385 continue; /* no visible tuple, try next index entry */
6386 }
6387
6388 /* We don't actually need the heap tuple for anything */
6389 ExecClearTuple(tableslot);
6390
6391 /*
6392 * We don't care whether there's more than one visible tuple in
6393 * the HOT chain; if any are visible, that's good enough.
6394 */
6395 }
6396
6397 /*
6398 * We expect that btree will return data in IndexTuple not HeapTuple
6399 * format. It's not lossy either.
6400 */
6401 if (!index_scan->xs_itup)
6402 elog(ERROR, "no data returned for index-only scan");
6403 if (index_scan->xs_recheck)
6404 elog(ERROR, "unexpected recheck indication from btree");
6405
6406 /* OK to deconstruct the index tuple */
6407 index_deform_tuple(index_scan->xs_itup,
6408 index_scan->xs_itupdesc,
6409 values, isnull);
6410
6411 /* Shouldn't have got a null, but be careful */
6412 if (isnull[0])
6413 elog(ERROR, "found unexpected null value in index \"%s\"",
6414 RelationGetRelationName(indexRel));
6415
6416 /* Copy the index column value out to caller's context */
6417 oldcontext = MemoryContextSwitchTo(outercontext);
6418 *endpointDatum = datumCopy(values[0], typByVal, typLen);
6419 MemoryContextSwitchTo(oldcontext);
6420 have_data = true;
6421 break;
6422 }
6423
6424 if (vmbuffer != InvalidBuffer)
6425 ReleaseBuffer(vmbuffer);
6426 index_endscan(index_scan);
6427
6428 return have_data;
6429}
6430
6431/*
6432 * find_join_input_rel
6433 * Look up the input relation for a join.
6434 *
6435 * We assume that the input relation's RelOptInfo must have been constructed
6436 * already.
6437 */
6438static RelOptInfo *
6440{
6441 RelOptInfo *rel = NULL;
6442
6443 if (!bms_is_empty(relids))
6444 {
6445 int relid;
6446
6447 if (bms_get_singleton_member(relids, &relid))
6448 rel = find_base_rel(root, relid);
6449 else
6450 rel = find_join_rel(root, relids);
6451 }
6452
6453 if (rel == NULL)
6454 elog(ERROR, "could not find RelOptInfo for given relids");
6455
6456 return rel;
6457}
6458
6459
6460/*-------------------------------------------------------------------------
6461 *
6462 * Index cost estimation functions
6463 *
6464 *-------------------------------------------------------------------------
6465 */
6466
6467/*
6468 * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
6469 */
6470List *
6472{
6473 List *result = NIL;
6474 ListCell *lc;
6475
6476 foreach(lc, indexclauses)
6477 {
6478 IndexClause *iclause = lfirst_node(IndexClause, lc);
6479 ListCell *lc2;
6480
6481 foreach(lc2, iclause->indexquals)
6482 {
6483 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6484
6485 result = lappend(result, rinfo);
6486 }
6487 }
6488 return result;
6489}
6490
6491/*
6492 * Compute the total evaluation cost of the comparison operands in a list
6493 * of index qual expressions. Since we know these will be evaluated just
6494 * once per scan, there's no need to distinguish startup from per-row cost.
6495 *
6496 * This can be used either on the result of get_quals_from_indexclauses(),
6497 * or directly on an indexorderbys list. In both cases, we expect that the
6498 * index key expression is on the left side of binary clauses.
6499 */
6500Cost
6502{
6503 Cost qual_arg_cost = 0;
6504 ListCell *lc;
6505
6506 foreach(lc, indexquals)
6507 {
6508 Expr *clause = (Expr *) lfirst(lc);
6509 Node *other_operand;
6510 QualCost index_qual_cost;
6511
6512 /*
6513 * Index quals will have RestrictInfos, indexorderbys won't. Look
6514 * through RestrictInfo if present.
6515 */
6516 if (IsA(clause, RestrictInfo))
6517 clause = ((RestrictInfo *) clause)->clause;
6518
6519 if (IsA(clause, OpExpr))
6520 {
6521 OpExpr *op = (OpExpr *) clause;
6522
6523 other_operand = (Node *) lsecond(op->args);
6524 }
6525 else if (IsA(clause, RowCompareExpr))
6526 {
6527 RowCompareExpr *rc = (RowCompareExpr *) clause;
6528
6529 other_operand = (Node *) rc->rargs;
6530 }
6531 else if (IsA(clause, ScalarArrayOpExpr))
6532 {
6533 ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6534
6535 other_operand = (Node *) lsecond(saop->args);
6536 }
6537 else if (IsA(clause, NullTest))
6538 {
6539 other_operand = NULL;
6540 }
6541 else
6542 {
6543 elog(ERROR, "unsupported indexqual type: %d",
6544 (int) nodeTag(clause));
6545 other_operand = NULL; /* keep compiler quiet */
6546 }
6547
6548 cost_qual_eval_node(&index_qual_cost, other_operand, root);
6549 qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
6550 }
6551 return qual_arg_cost;
6552}
6553
6554void
6556 IndexPath *path,
6557 double loop_count,
6558 GenericCosts *costs)
6559{
6560 IndexOptInfo *index = path->indexinfo;
6561 List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
6562 List *indexOrderBys = path->indexorderbys;
6563 Cost indexStartupCost;
6564 Cost indexTotalCost;
6565 Selectivity indexSelectivity;
6566 double indexCorrelation;
6567 double numIndexPages;
6568 double numIndexTuples;
6569 double spc_random_page_cost;
6570 double num_sa_scans;
6571 double num_outer_scans;
6572 double num_scans;
6573 double qual_op_cost;
6574 double qual_arg_cost;
6575 List *selectivityQuals;
6576 ListCell *l;
6577
6578 /*
6579 * If the index is partial, AND the index predicate with the explicitly
6580 * given indexquals to produce a more accurate idea of the index
6581 * selectivity.
6582 */
6583 selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
6584
6585 /*
6586 * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
6587 * just assume that the number of index descents is the number of distinct
6588 * combinations of array elements from all of the scan's SAOP clauses.
6589 */
6590 num_sa_scans = costs->num_sa_scans;
6591 if (num_sa_scans < 1)
6592 {
6593 num_sa_scans = 1;
6594 foreach(l, indexQuals)
6595 {
6596 RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
6597
6598 if (IsA(rinfo->clause, ScalarArrayOpExpr))
6599 {
6600 ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
6601 double alength = estimate_array_length(root, lsecond(saop->args));
6602
6603 if (alength > 1)
6604 num_sa_scans *= alength;
6605 }
6606 }
6607 }
6608
6609 /* Estimate the fraction of main-table tuples that will be visited */
6610 indexSelectivity = clauselist_selectivity(root, selectivityQuals,
6611 index->rel->relid,
6612 JOIN_INNER,
6613 NULL);
6614
6615 /*
6616 * If caller didn't give us an estimate, estimate the number of index
6617 * tuples that will be visited. We do it in this rather peculiar-looking
6618 * way in order to get the right answer for partial indexes.
6619 */
6620 numIndexTuples = costs->numIndexTuples;
6621 if (numIndexTuples <= 0.0)
6622 {
6623 numIndexTuples = indexSelectivity * index->rel->tuples;
6624
6625 /*
6626 * The above calculation counts all the tuples visited across all
6627 * scans induced by ScalarArrayOpExpr nodes. We want to consider the
6628 * average per-indexscan number, so adjust. This is a handy place to
6629 * round to integer, too. (If caller supplied tuple estimate, it's
6630 * responsible for handling these considerations.)
6631 */
6632 numIndexTuples = rint(numIndexTuples / num_sa_scans);
6633 }
6634
6635 /*
6636 * We can bound the number of tuples by the index size in any case. Also,
6637 * always estimate at least one tuple is touched, even when
6638 * indexSelectivity estimate is tiny.
6639 */
6640 if (numIndexTuples > index->tuples)
6641 numIndexTuples = index->tuples;
6642 if (numIndexTuples < 1.0)
6643 numIndexTuples = 1.0;
6644
6645 /*
6646 * Estimate the number of index pages that will be retrieved.
6647 *
6648 * We use the simplistic method of taking a pro-rata fraction of the total
6649 * number of index pages. In effect, this counts only leaf pages and not
6650 * any overhead such as index metapage or upper tree levels.
6651 *
6652 * In practice access to upper index levels is often nearly free because
6653 * those tend to stay in cache under load; moreover, the cost involved is
6654 * highly dependent on index type. We therefore ignore such costs here
6655 * and leave it to the caller to add a suitable charge if needed.
6656 */
6657 if (index->pages > 1 && index->tuples > 1)
6658 numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
6659 else
6660 numIndexPages = 1.0;
6661
6662 /* fetch estimated page cost for tablespace containing index */
6663 get_tablespace_page_costs(index->reltablespace,
6664 &spc_random_page_cost,
6665 NULL);
6666
6667 /*
6668 * Now compute the disk access costs.
6669 *
6670 * The above calculations are all per-index-scan. However, if we are in a
6671 * nestloop inner scan, we can expect the scan to be repeated (with
6672 * different search keys) for each row of the outer relation. Likewise,
6673 * ScalarArrayOpExpr quals result in multiple index scans. This creates
6674 * the potential for cache effects to reduce the number of disk page
6675 * fetches needed. We want to estimate the average per-scan I/O cost in
6676 * the presence of caching.
6677 *
6678 * We use the Mackert-Lohman formula (see costsize.c for details) to
6679 * estimate the total number of page fetches that occur. While this
6680 * wasn't what it was designed for, it seems a reasonable model anyway.
6681 * Note that we are counting pages not tuples anymore, so we take N = T =
6682 * index size, as if there were one "tuple" per page.
6683 */
6684 num_outer_scans = loop_count;
6685 num_scans = num_sa_scans * num_outer_scans;
6686
6687 if (num_scans > 1)
6688 {
6689 double pages_fetched;
6690
6691 /* total page fetches ignoring cache effects */
6692 pages_fetched = numIndexPages * num_scans;
6693
6694 /* use Mackert and Lohman formula to adjust for cache effects */
6695 pages_fetched = index_pages_fetched(pages_fetched,
6696 index->pages,
6697 (double) index->pages,
6698 root);
6699
6700 /*
6701 * Now compute the total disk access cost, and then report a pro-rated
6702 * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
6703 * since that's internal to the indexscan.)
6704 */
6705 indexTotalCost = (pages_fetched * spc_random_page_cost)
6706 / num_outer_scans;
6707 }
6708 else
6709 {
6710 /*
6711 * For a single index scan, we just charge spc_random_page_cost per
6712 * page touched.
6713 */
6714 indexTotalCost = numIndexPages * spc_random_page_cost;
6715 }
6716
6717 /*
6718 * CPU cost: any complex expressions in the indexquals will need to be
6719 * evaluated once at the start of the scan to reduce them to runtime keys
6720 * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
6721 * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
6722 * indexqual operator. Because we have numIndexTuples as a per-scan
6723 * number, we have to multiply by num_sa_scans to get the correct result
6724 * for ScalarArrayOpExpr cases. Similarly add in costs for any index
6725 * ORDER BY expressions.
6726 *
6727 * Note: this neglects the possible costs of rechecking lossy operators.
6728 * Detecting that that might be needed seems more expensive than it's
6729 * worth, though, considering all the other inaccuracies here ...
6730 */
6731 qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
6732 index_other_operands_eval_cost(root, indexOrderBys);
6733 qual_op_cost = cpu_operator_cost *
6734 (list_length(indexQuals) + list_length(indexOrderBys));
6735
6736 indexStartupCost = qual_arg_cost;
6737 indexTotalCost += qual_arg_cost;
6738 indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
6739
6740 /*
6741 * Generic assumption about index correlation: there isn't any.
6742 */
6743 indexCorrelation = 0.0;
6744
6745 /*
6746 * Return everything to caller.
6747 */
6748 costs->indexStartupCost = indexStartupCost;
6749 costs->indexTotalCost = indexTotalCost;
6750 costs->indexSelectivity = indexSelectivity;
6751 costs->indexCorrelation = indexCorrelation;
6752 costs->numIndexPages = numIndexPages;
6753 costs->numIndexTuples = numIndexTuples;
6754 costs->spc_random_page_cost = spc_random_page_cost;
6755 costs->num_sa_scans = num_sa_scans;
6756}
6757
6758/*
6759 * If the index is partial, add its predicate to the given qual list.
6760 *
6761 * ANDing the index predicate with the explicitly given indexquals produces
6762 * a more accurate idea of the index's selectivity. However, we need to be
6763 * careful not to insert redundant clauses, because clauselist_selectivity()
6764 * is easily fooled into computing a too-low selectivity estimate. Our
6765 * approach is to add only the predicate clause(s) that cannot be proven to
6766 * be implied by the given indexquals. This successfully handles cases such
6767 * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
6768 * There are many other cases where we won't detect redundancy, leading to a
6769 * too-low selectivity estimate, which will bias the system in favor of using
6770 * partial indexes where possible. That is not necessarily bad though.
6771 *
6772 * Note that indexQuals contains RestrictInfo nodes while the indpred
6773 * does not, so the output list will be mixed. This is OK for both
6774 * predicate_implied_by() and clauselist_selectivity(), but might be
6775 * problematic if the result were passed to other things.
6776 */
6777List *
6779{
6780 List *predExtraQuals = NIL;
6781 ListCell *lc;
6782
6783 if (index->indpred == NIL)
6784 return indexQuals;
6785
6786 foreach(lc, index->indpred)
6787 {
6788 Node *predQual = (Node *) lfirst(lc);
6789 List *oneQual = list_make1(predQual);
6790
6791 if (!predicate_implied_by(oneQual, indexQuals, false))
6792 predExtraQuals = list_concat(predExtraQuals, oneQual);
6793 }
6794 return list_concat(predExtraQuals, indexQuals);
6795}
6796
6797
6798void
6799btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
6800 Cost *indexStartupCost, Cost *indexTotalCost,
6801 Selectivity *indexSelectivity, double *indexCorrelation,
6802 double *indexPages)
6803{
6804 IndexOptInfo *index = path->indexinfo;
6805 GenericCosts costs = {0};
6806 Oid relid;
6807 AttrNumber colnum;
6808 VariableStatData vardata = {0};
6809 double numIndexTuples;
6810 Cost descentCost;
6811 List *indexBoundQuals;
6812 int indexcol;
6813 bool eqQualHere;
6814 bool found_saop;
6815 bool found_is_null_op;
6816 double num_sa_scans;
6817 ListCell *lc;
6818
6819 /*
6820 * For a btree scan, only leading '=' quals plus inequality quals for the
6821 * immediately next attribute contribute to index selectivity (these are
6822 * the "boundary quals" that determine the starting and stopping points of
6823 * the index scan). Additional quals can suppress visits to the heap, so
6824 * it's OK to count them in indexSelectivity, but they should not count
6825 * for estimating numIndexTuples. So we must examine the given indexquals
6826 * to find out which ones count as boundary quals. We rely on the
6827 * knowledge that they are given in index column order.
6828 *
6829 * For a RowCompareExpr, we consider only the first column, just as
6830 * rowcomparesel() does.
6831 *
6832 * If there's a ScalarArrayOpExpr in the quals, we'll actually perform up
6833 * to N index descents (not just one), but the ScalarArrayOpExpr's
6834 * operator can be considered to act the same as it normally does.
6835 */
6836 indexBoundQuals = NIL;
6837 indexcol = 0;
6838 eqQualHere = false;
6839 found_saop = false;
6840 found_is_null_op = false;
6841 num_sa_scans = 1;
6842 foreach(lc, path->indexclauses)
6843 {
6844 IndexClause *iclause = lfirst_node(IndexClause, lc);
6845 ListCell *lc2;
6846
6847 if (indexcol != iclause->indexcol)
6848 {
6849 /* Beginning of a new column's quals */
6850 if (!eqQualHere)
6851 break; /* done if no '=' qual for indexcol */
6852 eqQualHere = false;
6853 indexcol++;
6854 if (indexcol != iclause->indexcol)
6855 break; /* no quals at all for indexcol */
6856 }
6857
6858 /* Examine each indexqual associated with this index clause */
6859 foreach(lc2, iclause->indexquals)
6860 {
6861 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6862 Expr *clause = rinfo->clause;
6863 Oid clause_op = InvalidOid;
6864 int op_strategy;
6865
6866 if (IsA(clause, OpExpr))
6867 {
6868 OpExpr *op = (OpExpr *) clause;
6869
6870 clause_op = op->opno;
6871 }
6872 else if (IsA(clause, RowCompareExpr))
6873 {
6874 RowCompareExpr *rc = (RowCompareExpr *) clause;
6875
6876 clause_op = linitial_oid(rc->opnos);
6877 }
6878 else if (IsA(clause, ScalarArrayOpExpr))
6879 {
6880 ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6881 Node *other_operand = (Node *) lsecond(saop->args);
6882 double alength = estimate_array_length(root, other_operand);
6883
6884 clause_op = saop->opno;
6885 found_saop = true;
6886 /* estimate SA descents by indexBoundQuals only */
6887 if (alength > 1)
6888 num_sa_scans *= alength;
6889 }
6890 else if (IsA(clause, NullTest))
6891 {
6892 NullTest *nt = (NullTest *) clause;
6893
6894 if (nt->nulltesttype == IS_NULL)
6895 {
6896 found_is_null_op = true;
6897 /* IS NULL is like = for selectivity purposes */
6898 eqQualHere = true;
6899 }
6900 }
6901 else
6902 elog(ERROR, "unsupported indexqual type: %d",
6903 (int) nodeTag(clause));
6904
6905 /* check for equality operator */
6906 if (OidIsValid(clause_op))
6907 {
6908 op_strategy = get_op_opfamily_strategy(clause_op,
6909 index->opfamily[indexcol]);
6910 Assert(op_strategy != 0); /* not a member of opfamily?? */
6911 if (op_strategy == BTEqualStrategyNumber)
6912 eqQualHere = true;
6913 }
6914
6915 indexBoundQuals = lappend(indexBoundQuals, rinfo);
6916 }
6917 }
6918
6919 /*
6920 * If index is unique and we found an '=' clause for each column, we can
6921 * just assume numIndexTuples = 1 and skip the expensive
6922 * clauselist_selectivity calculations. However, a ScalarArrayOp or
6923 * NullTest invalidates that theory, even though it sets eqQualHere.
6924 */
6925 if (index->unique &&
6926 indexcol == index->nkeycolumns - 1 &&
6927 eqQualHere &&
6928 !found_saop &&
6929 !found_is_null_op)
6930 numIndexTuples = 1.0;
6931 else
6932 {
6933 List *selectivityQuals;
6934 Selectivity btreeSelectivity;
6935
6936 /*
6937 * If the index is partial, AND the index predicate with the
6938 * index-bound quals to produce a more accurate idea of the number of
6939 * rows covered by the bound conditions.
6940 */
6941 selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
6942
6943 btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
6944 index->rel->relid,
6945 JOIN_INNER,
6946 NULL);
6947 numIndexTuples = btreeSelectivity * index->rel->tuples;
6948
6949 /*
6950 * btree automatically combines individual ScalarArrayOpExpr primitive
6951 * index scans whenever the tuples covered by the next set of array
6952 * keys are close to tuples covered by the current set. That puts a
6953 * natural ceiling on the worst case number of descents -- there
6954 * cannot possibly be more than one descent per leaf page scanned.
6955 *
6956 * Clamp the number of descents to at most 1/3 the number of index
6957 * pages. This avoids implausibly high estimates with low selectivity
6958 * paths, where scans usually require only one or two descents. This
6959 * is most likely to help when there are several SAOP clauses, where
6960 * naively accepting the total number of distinct combinations of
6961 * array elements as the number of descents would frequently lead to
6962 * wild overestimates.
6963 *
6964 * We somewhat arbitrarily don't just make the cutoff the total number
6965 * of leaf pages (we make it 1/3 the total number of pages instead) to
6966 * give the btree code credit for its ability to continue on the leaf
6967 * level with low selectivity scans.
6968 */
6969 num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
6970 num_sa_scans = Max(num_sa_scans, 1);
6971
6972 /*
6973 * As in genericcostestimate(), we have to adjust for any
6974 * ScalarArrayOpExpr quals included in indexBoundQuals, and then round
6975 * to integer.
6976 *
6977 * It is tempting to make genericcostestimate behave as if SAOP
6978 * clauses work in almost the same way as scalar operators during
6979 * btree scans, making the top-level scan look like a continuous scan
6980 * (as opposed to num_sa_scans-many primitive index scans). After
6981 * all, btree scans mostly work like that at runtime. However, such a
6982 * scheme would badly bias genericcostestimate's simplistic approach
6983 * to calculating numIndexPages through prorating.
6984 *
6985 * Stick with the approach taken by non-native SAOP scans for now.
6986 * genericcostestimate will use the Mackert-Lohman formula to
6987 * compensate for repeat page fetches, even though that definitely
6988 * won't happen during btree scans (not for leaf pages, at least).
6989 * We're usually very pessimistic about the number of primitive index
6990 * scans that will be required, but it's not clear how to do better.
6991 */
6992 numIndexTuples = rint(numIndexTuples / num_sa_scans);
6993 }
6994
6995 /*
6996 * Now do generic index cost estimation.
6997 */
6998 costs.numIndexTuples = numIndexTuples;
6999 costs.num_sa_scans = num_sa_scans;
7000
7001 genericcostestimate(root, path, loop_count, &costs);
7002
7003 /*
7004 * Add a CPU-cost component to represent the costs of initial btree
7005 * descent. We don't charge any I/O cost for touching upper btree levels,
7006 * since they tend to stay in cache, but we still have to do about log2(N)
7007 * comparisons to descend a btree of N leaf tuples. We charge one
7008 * cpu_operator_cost per comparison.
7009 *
7010 * If there are ScalarArrayOpExprs, charge this once per estimated SA
7011 * index descent. The ones after the first one are not startup cost so
7012 * far as the overall plan goes, so just add them to "total" cost.
7013 */
7014 if (index->tuples > 1) /* avoid computing log(0) */
7015 {
7016 descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
7017 costs.indexStartupCost += descentCost;
7018 costs.indexTotalCost += costs.num_sa_scans * descentCost;
7019 }
7020
7021 /*
7022 * Even though we're not charging I/O cost for touching upper btree pages,
7023 * it's still reasonable to charge some CPU cost per page descended
7024 * through. Moreover, if we had no such charge at all, bloated indexes
7025 * would appear to have the same search cost as unbloated ones, at least
7026 * in cases where only a single leaf page is expected to be visited. This
7027 * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
7028 * touched. The number of such pages is btree tree height plus one (ie,
7029 * we charge for the leaf page too). As above, charge once per estimated
7030 * SA index descent.
7031 */
7032 descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7033 costs.indexStartupCost += descentCost;
7034 costs.indexTotalCost += costs.num_sa_scans * descentCost;
7035
7036 /*
7037 * If we can get an estimate of the first column's ordering correlation C
7038 * from pg_statistic, estimate the index correlation as C for a
7039 * single-column index, or C * 0.75 for multiple columns. (The idea here
7040 * is that multiple columns dilute the importance of the first column's
7041 * ordering, but don't negate it entirely. Before 8.0 we divided the
7042 * correlation by the number of columns, but that seems too strong.)
7043 */
7044 if (index->indexkeys[0] != 0)
7045 {
7046 /* Simple variable --- look to stats for the underlying table */
7047 RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
7048
7049 Assert(rte->rtekind == RTE_RELATION);
7050 relid = rte->relid;
7051 Assert(relid != InvalidOid);
7052 colnum = index->indexkeys[0];
7053
7055 (*get_relation_stats_hook) (root, rte, colnum, &vardata))
7056 {
7057 /*
7058 * The hook took control of acquiring a stats tuple. If it did
7059 * supply a tuple, it'd better have supplied a freefunc.
7060 */
7061 if (HeapTupleIsValid(vardata.statsTuple) &&
7062 !vardata.freefunc)
7063 elog(ERROR, "no function provided to release variable stats with");
7064 }
7065 else
7066 {
7067 vardata.statsTuple = SearchSysCache3(STATRELATTINH,
7068 ObjectIdGetDatum(relid),
7069 Int16GetDatum(colnum),
7070 BoolGetDatum(rte->inh));
7071 vardata.freefunc = ReleaseSysCache;
7072 }
7073 }
7074 else
7075 {
7076 /* Expression --- maybe there are stats for the index itself */
7077 relid = index->indexoid;
7078 colnum = 1;
7079
7081 (*get_index_stats_hook) (root, relid, colnum, &vardata))
7082 {
7083 /*
7084 * The hook took control of acquiring a stats tuple. If it did
7085 * supply a tuple, it'd better have supplied a freefunc.
7086 */
7087 if (HeapTupleIsValid(vardata.statsTuple) &&
7088 !vardata.freefunc)
7089 elog(ERROR, "no function provided to release variable stats with");
7090 }
7091 else
7092 {
7093 vardata.statsTuple = SearchSysCache3(STATRELATTINH,
7094 ObjectIdGetDatum(relid),
7095 Int16GetDatum(colnum),
7096 BoolGetDatum(false));
7097 vardata.freefunc = ReleaseSysCache;
7098 }
7099 }
7100
7101 if (HeapTupleIsValid(vardata.statsTuple))
7102 {
7103 Oid sortop;
7104 AttStatsSlot sslot;
7105
7106 sortop = get_opfamily_member(index->opfamily[0],
7107 index->opcintype[0],
7108 index->opcintype[0],
7110 if (OidIsValid(sortop) &&
7111 get_attstatsslot(&sslot, vardata.statsTuple,
7112 STATISTIC_KIND_CORRELATION, sortop,
7114 {
7115 double varCorrelation;
7116
7117 Assert(sslot.nnumbers == 1);
7118 varCorrelation = sslot.numbers[0];
7119
7120 if (index->reverse_sort[0])
7121 varCorrelation = -varCorrelation;
7122
7123 if (index->nkeycolumns > 1)
7124 costs.indexCorrelation = varCorrelation * 0.75;
7125 else
7126 costs.indexCorrelation = varCorrelation;
7127
7128 free_attstatsslot(&sslot);
7129 }
7130 }
7131
7132 ReleaseVariableStats(vardata);
7133
7134 *indexStartupCost = costs.indexStartupCost;
7135 *indexTotalCost = costs.indexTotalCost;
7136 *indexSelectivity = costs.indexSelectivity;
7137 *indexCorrelation = costs.indexCorrelation;
7138 *indexPages = costs.numIndexPages;
7139}
7140
7141void
7142hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7143 Cost *indexStartupCost, Cost *indexTotalCost,
7144 Selectivity *indexSelectivity, double *indexCorrelation,
7145 double *indexPages)
7146{
7147 GenericCosts costs = {0};
7148
7149 genericcostestimate(root, path, loop_count, &costs);
7150
7151 /*
7152 * A hash index has no descent costs as such, since the index AM can go
7153 * directly to the target bucket after computing the hash value. There
7154 * are a couple of other hash-specific costs that we could conceivably add
7155 * here, though:
7156 *
7157 * Ideally we'd charge spc_random_page_cost for each page in the target
7158 * bucket, not just the numIndexPages pages that genericcostestimate
7159 * thought we'd visit. However in most cases we don't know which bucket
7160 * that will be. There's no point in considering the average bucket size
7161 * because the hash AM makes sure that's always one page.
7162 *
7163 * Likewise, we could consider charging some CPU for each index tuple in
7164 * the bucket, if we knew how many there were. But the per-tuple cost is
7165 * just a hash value comparison, not a general datatype-dependent
7166 * comparison, so any such charge ought to be quite a bit less than
7167 * cpu_operator_cost; which makes it probably not worth worrying about.
7168 *
7169 * A bigger issue is that chance hash-value collisions will result in
7170 * wasted probes into the heap. We don't currently attempt to model this
7171 * cost on the grounds that it's rare, but maybe it's not rare enough.
7172 * (Any fix for this ought to consider the generic lossy-operator problem,
7173 * though; it's not entirely hash-specific.)
7174 */
7175
7176 *indexStartupCost = costs.indexStartupCost;
7177 *indexTotalCost = costs.indexTotalCost;
7178 *indexSelectivity = costs.indexSelectivity;
7179 *indexCorrelation = costs.indexCorrelation;
7180 *indexPages = costs.numIndexPages;
7181}
7182
7183void
7184gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7185 Cost *indexStartupCost, Cost *indexTotalCost,
7186 Selectivity *indexSelectivity, double *indexCorrelation,
7187 double *indexPages)
7188{
7189 IndexOptInfo *index = path->indexinfo;
7190 GenericCosts costs = {0};
7191 Cost descentCost;
7192
7193 genericcostestimate(root, path, loop_count, &costs);
7194
7195 /*
7196 * We model index descent costs similarly to those for btree, but to do
7197 * that we first need an idea of the tree height. We somewhat arbitrarily
7198 * assume that the fanout is 100, meaning the tree height is at most
7199 * log100(index->pages).
7200 *
7201 * Although this computation isn't really expensive enough to require
7202 * caching, we might as well use index->tree_height to cache it.
7203 */
7204 if (index->tree_height < 0) /* unknown? */
7205 {
7206 if (index->pages > 1) /* avoid computing log(0) */
7207 index->tree_height = (int) (log(index->pages) / log(100.0));
7208 else
7209 index->tree_height = 0;
7210 }
7211
7212 /*
7213 * Add a CPU-cost component to represent the costs of initial descent. We
7214 * just use log(N) here not log2(N) since the branching factor isn't
7215 * necessarily two anyway. As for btree, charge once per SA scan.
7216 */
7217 if (index->tuples > 1) /* avoid computing log(0) */
7218 {
7219 descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7220 costs.indexStartupCost += descentCost;
7221 costs.indexTotalCost += costs.num_sa_scans * descentCost;
7222 }
7223
7224 /*
7225 * Likewise add a per-page charge, calculated the same as for btrees.
7226 */
7227 descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7228 costs.indexStartupCost += descentCost;
7229 costs.indexTotalCost += costs.num_sa_scans * descentCost;
7230
7231 *indexStartupCost = costs.indexStartupCost;
7232 *indexTotalCost = costs.indexTotalCost;
7233 *indexSelectivity = costs.indexSelectivity;
7234 *indexCorrelation = costs.indexCorrelation;
7235 *indexPages = costs.numIndexPages;
7236}
7237
7238void
7239spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7240 Cost *indexStartupCost, Cost *indexTotalCost,
7241 Selectivity *indexSelectivity, double *indexCorrelation,
7242 double *indexPages)
7243{
7244 IndexOptInfo *index = path->indexinfo;
7245 GenericCosts costs = {0};
7246 Cost descentCost;
7247
7248 genericcostestimate(root, path, loop_count, &costs);
7249
7250 /*
7251 * We model index descent costs similarly to those for btree, but to do
7252 * that we first need an idea of the tree height. We somewhat arbitrarily
7253 * assume that the fanout is 100, meaning the tree height is at most
7254 * log100(index->pages).
7255 *
7256 * Although this computation isn't really expensive enough to require
7257 * caching, we might as well use index->tree_height to cache it.
7258 */
7259 if (index->tree_height < 0) /* unknown? */
7260 {
7261 if (index->pages > 1) /* avoid computing log(0) */
7262 index->tree_height = (int) (log(index->pages) / log(100.0));
7263 else
7264 index->tree_height = 0;
7265 }
7266
7267 /*
7268 * Add a CPU-cost component to represent the costs of initial descent. We
7269 * just use log(N) here not log2(N) since the branching factor isn't
7270 * necessarily two anyway. As for btree, charge once per SA scan.
7271 */
7272 if (index->tuples > 1) /* avoid computing log(0) */
7273 {
7274 descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7275 costs.indexStartupCost += descentCost;
7276 costs.indexTotalCost += costs.num_sa_scans * descentCost;
7277 }
7278
7279 /*
7280 * Likewise add a per-page charge, calculated the same as for btrees.
7281 */
7282 descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7283 costs.indexStartupCost += descentCost;
7284 costs.indexTotalCost += costs.num_sa_scans * descentCost;
7285
7286 *indexStartupCost = costs.indexStartupCost;
7287 *indexTotalCost = costs.indexTotalCost;
7288 *indexSelectivity = costs.indexSelectivity;
7289 *indexCorrelation = costs.indexCorrelation;
7290 *indexPages = costs.numIndexPages;
7291}
7292
7293
7294/*
7295 * Support routines for gincostestimate
7296 */
7297
7298typedef struct
7299{
7300 bool attHasFullScan[INDEX_MAX_KEYS];
7301 bool attHasNormalScan[INDEX_MAX_KEYS];
7307
7308/*
7309 * Estimate the number of index terms that need to be searched for while
7310 * testing the given GIN query, and increment the counts in *counts
7311 * appropriately. If the query is unsatisfiable, return false.
7312 */
7313static bool
7315 Oid clause_op, Datum query,
7316 GinQualCounts *counts)
7317{
7318 FmgrInfo flinfo;
7319 Oid extractProcOid;
7320 Oid collation;
7321 int strategy_op;
7322 Oid lefttype,
7323 righttype;
7324 int32 nentries = 0;
7325 bool *partial_matches = NULL;
7326 Pointer *extra_data = NULL;
7327 bool *nullFlags = NULL;
7328 int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
7329 int32 i;
7330
7331 Assert(indexcol < index->nkeycolumns);
7332
7333 /*
7334 * Get the operator's strategy number and declared input data types within
7335 * the index opfamily. (We don't need the latter, but we use
7336 * get_op_opfamily_properties because it will throw error if it fails to
7337 * find a matching pg_amop entry.)
7338 */
7339 get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
7340 &strategy_op, &lefttype, &righttype);
7341
7342 /*
7343 * GIN always uses the "default" support functions, which are those with
7344 * lefttype == righttype == the opclass' opcintype (see
7345 * IndexSupportInitialize in relcache.c).
7346 */
7347 extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
7348 index->opcintype[indexcol],
7349 index->opcintype[indexcol],
7351
7352 if (!OidIsValid(extractProcOid))
7353 {
7354 /* should not happen; throw same error as index_getprocinfo */
7355 elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
7356 GIN_EXTRACTQUERY_PROC, indexcol + 1,
7357 get_rel_name(index->indexoid));
7358 }
7359
7360 /*
7361 * Choose collation to pass to extractProc (should match initGinState).
7362 */
7363 if (OidIsValid(index->indexcollations[indexcol]))
7364 collation = index->indexcollations[indexcol];
7365 else
7366 collation = DEFAULT_COLLATION_OID;
7367
7368 fmgr_info(extractProcOid, &flinfo);
7369
7370 set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
7371
7372 FunctionCall7Coll(&flinfo,
7373 collation,
7374 query,
7375 PointerGetDatum(&nentries),
7376 UInt16GetDatum(strategy_op),
7377 PointerGetDatum(&partial_matches),
7378 PointerGetDatum(&extra_data),
7379 PointerGetDatum(&nullFlags),
7380 PointerGetDatum(&searchMode));
7381
7382 if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
7383 {
7384 /* No match is possible */
7385 return false;
7386 }
7387
7388 for (i = 0; i < nentries; i++)
7389 {
7390 /*
7391 * For partial match we haven't any information to estimate number of
7392 * matched entries in index, so, we just estimate it as 100
7393 */
7394 if (partial_matches && partial_matches[i])
7395 counts->partialEntries += 100;
7396 else
7397 counts->exactEntries++;
7398
7399 counts->searchEntries++;
7400 }
7401
7402 if (searchMode == GIN_SEARCH_MODE_DEFAULT)
7403 {
7404 counts->attHasNormalScan[indexcol] = true;
7405 }
7406 else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
7407 {
7408 /* Treat "include empty" like an exact-match item */
7409 counts->attHasNormalScan[indexcol] = true;
7410 counts->exactEntries++;
7411 counts->searchEntries++;
7412 }
7413 else
7414 {
7415 /* It's GIN_SEARCH_MODE_ALL */
7416 counts->attHasFullScan[indexcol] = true;
7417 }
7418
7419 return true;
7420}
7421
7422/*
7423 * Estimate the number of index terms that need to be searched for while
7424 * testing the given GIN index clause, and increment the counts in *counts
7425 * appropriately. If the query is unsatisfiable, return false.
7426 */
7427static bool
7430 int indexcol,
7431 OpExpr *clause,
7432 GinQualCounts *counts)
7433{
7434 Oid clause_op = clause->opno;
7435 Node *operand = (Node *) lsecond(clause->args);
7436
7437 /* aggressively reduce to a constant, and look through relabeling */
7438 operand = estimate_expression_value(root, operand);
7439
7440 if (IsA(operand, RelabelType))
7441 operand = (Node *) ((RelabelType *) operand)->arg;
7442
7443 /*
7444 * It's impossible to call extractQuery method for unknown operand. So
7445 * unless operand is a Const we can't do much; just assume there will be
7446 * one ordinary search entry from the operand at runtime.
7447 */
7448 if (!IsA(operand, Const))
7449 {
7450 counts->exactEntries++;
7451 counts->searchEntries++;
7452 return true;
7453 }
7454
7455 /* If Const is null, there can be no matches */
7456 if (((Const *) operand)->constisnull)
7457 return false;
7458
7459 /* Otherwise, apply extractQuery and get the actual term counts */
7460 return gincost_pattern(index, indexcol, clause_op,
7461 ((Const *) operand)->constvalue,
7462 counts);
7463}
7464
7465/*
7466 * Estimate the number of index terms that need to be searched for while
7467 * testing the given GIN index clause, and increment the counts in *counts
7468 * appropriately. If the query is unsatisfiable, return false.
7469 *
7470 * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
7471 * each of which involves one value from the RHS array, plus all the
7472 * non-array quals (if any). To model this, we average the counts across
7473 * the RHS elements, and add the averages to the counts in *counts (which
7474 * correspond to per-indexscan costs). We also multiply counts->arrayScans
7475 * by N, causing gincostestimate to scale up its estimates accordingly.
7476 */
7477static bool
7480 int indexcol,
7481 ScalarArrayOpExpr *clause,
7482 double numIndexEntries,
7483 GinQualCounts *counts)
7484{
7485 Oid clause_op = clause->opno;
7486 Node *rightop = (Node *) lsecond(clause->args);
7487 ArrayType *arrayval;
7488 int16 elmlen;
7489 bool elmbyval;
7490 char elmalign;
7491 int numElems;
7492 Datum *elemValues;
7493 bool *elemNulls;
7494 GinQualCounts arraycounts;
7495 int numPossible = 0;
7496 int i;
7497
7498 Assert(clause->useOr);
7499
7500 /* aggressively reduce to a constant, and look through relabeling */
7501 rightop = estimate_expression_value(root, rightop);
7502
7503 if (IsA(rightop, RelabelType))
7504 rightop = (Node *) ((RelabelType *) rightop)->arg;
7505
7506 /*
7507 * It's impossible to call extractQuery method for unknown operand. So
7508 * unless operand is a Const we can't do much; just assume there will be
7509 * one ordinary search entry from each array entry at runtime, and fall
7510 * back on a probably-bad estimate of the number of array entries.
7511 */
7512 if (!IsA(rightop, Const))
7513 {
7514 counts->exactEntries++;
7515 counts->searchEntries++;
7516 counts->arrayScans *= estimate_array_length(root, rightop);
7517 return true;
7518 }
7519
7520 /* If Const is null, there can be no matches */
7521 if (((Const *) rightop)->constisnull)
7522 return false;
7523
7524 /* Otherwise, extract the array elements and iterate over them */
7525 arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
7527 &elmlen, &elmbyval, &elmalign);
7528 deconstruct_array(arrayval,
7529 ARR_ELEMTYPE(arrayval),
7530 elmlen, elmbyval, elmalign,
7531 &elemValues, &elemNulls, &numElems);
7532
7533 memset(&arraycounts, 0, sizeof(arraycounts));
7534
7535 for (i = 0; i < numElems; i++)
7536 {
7537 GinQualCounts elemcounts;
7538
7539 /* NULL can't match anything, so ignore, as the executor will */
7540 if (elemNulls[i])
7541 continue;
7542
7543 /* Otherwise, apply extractQuery and get the actual term counts */
7544 memset(&elemcounts, 0, sizeof(elemcounts));
7545
7546 if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
7547 &elemcounts))
7548 {
7549 /* We ignore array elements that are unsatisfiable patterns */
7550 numPossible++;
7551
7552 if (elemcounts.attHasFullScan[indexcol] &&
7553 !elemcounts.attHasNormalScan[indexcol])
7554 {
7555 /*
7556 * Full index scan will be required. We treat this as if
7557 * every key in the index had been listed in the query; is
7558 * that reasonable?
7559 */
7560 elemcounts.partialEntries = 0;
7561 elemcounts.exactEntries = numIndexEntries;
7562 elemcounts.searchEntries = numIndexEntries;
7563 }
7564 arraycounts.partialEntries += elemcounts.partialEntries;
7565 arraycounts.exactEntries += elemcounts.exactEntries;
7566 arraycounts.searchEntries += elemcounts.searchEntries;
7567 }
7568 }
7569
7570 if (numPossible == 0)
7571 {
7572 /* No satisfiable patterns in the array */
7573 return false;
7574 }
7575
7576 /*
7577 * Now add the averages to the global counts. This will give us an
7578 * estimate of the average number of terms searched for in each indexscan,
7579 * including contributions from both array and non-array quals.
7580 */
7581 counts->partialEntries += arraycounts.partialEntries / numPossible;
7582 counts->exactEntries += arraycounts.exactEntries / numPossible;
7583 counts->searchEntries += arraycounts.searchEntries / numPossible;
7584
7585 counts->arrayScans *= numPossible;
7586
7587 return true;
7588}
7589
7590/*
7591 * GIN has search behavior completely different from other index types
7592 */
7593void
7594gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7595 Cost *indexStartupCost, Cost *indexTotalCost,
7596 Selectivity *indexSelectivity, double *indexCorrelation,
7597 double *indexPages)
7598{
7599 IndexOptInfo *index = path->indexinfo;
7600 List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7601 List *selectivityQuals;
7602 double numPages = index->pages,
7603 numTuples = index->tuples;
7604 double numEntryPages,
7605 numDataPages,
7606 numPendingPages,
7607 numEntries;
7608 GinQualCounts counts;
7609 bool matchPossible;
7610 bool fullIndexScan;
7611 double partialScale;
7612 double entryPagesFetched,
7613 dataPagesFetched,
7614 dataPagesFetchedBySel;
7615 double qual_op_cost,
7616 qual_arg_cost,
7617 spc_random_page_cost,
7618 outer_scans;
7619 Cost descentCost;
7620 Relation indexRel;
7621 GinStatsData ginStats;
7622 ListCell *lc;
7623 int i;
7624
7625 /*
7626 * Obtain statistical information from the meta page, if possible. Else
7627 * set ginStats to zeroes, and we'll cope below.
7628 */
7629 if (!index->hypothetical)
7630 {
7631 /* Lock should have already been obtained in plancat.c */
7632 indexRel = index_open(index->indexoid, NoLock);
7633 ginGetStats(indexRel, &ginStats);
7634 index_close(indexRel, NoLock);
7635 }
7636 else
7637 {
7638 memset(&ginStats, 0, sizeof(ginStats));
7639 }
7640
7641 /*
7642 * Assuming we got valid (nonzero) stats at all, nPendingPages can be
7643 * trusted, but the other fields are data as of the last VACUUM. We can
7644 * scale them up to account for growth since then, but that method only
7645 * goes so far; in the worst case, the stats might be for a completely
7646 * empty index, and scaling them will produce pretty bogus numbers.
7647 * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
7648 * it's grown more than that, fall back to estimating things only from the
7649 * assumed-accurate index size. But we'll trust nPendingPages in any case
7650 * so long as it's not clearly insane, ie, more than the index size.
7651 */
7652 if (ginStats.nPendingPages < numPages)
7653 numPendingPages = ginStats.nPendingPages;
7654 else
7655 numPendingPages = 0;
7656
7657 if (numPages > 0 && ginStats.nTotalPages <= numPages &&
7658 ginStats.nTotalPages > numPages / 4 &&
7659 ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
7660 {
7661 /*
7662 * OK, the stats seem close enough to sane to be trusted. But we
7663 * still need to scale them by the ratio numPages / nTotalPages to
7664 * account for growth since the last VACUUM.
7665 */
7666 double scale = numPages / ginStats.nTotalPages;
7667
7668 numEntryPages = ceil(ginStats.nEntryPages * scale);
7669 numDataPages = ceil(ginStats.nDataPages * scale);
7670 numEntries = ceil(ginStats.nEntries * scale);
7671 /* ensure we didn't round up too much */
7672 numEntryPages = Min(numEntryPages, numPages - numPendingPages);
7673 numDataPages = Min(numDataPages,
7674 numPages - numPendingPages - numEntryPages);
7675 }
7676 else
7677 {
7678 /*
7679 * We might get here because it's a hypothetical index, or an index
7680 * created pre-9.1 and never vacuumed since upgrading (in which case
7681 * its stats would read as zeroes), or just because it's grown too
7682 * much since the last VACUUM for us to put our faith in scaling.
7683 *
7684 * Invent some plausible internal statistics based on the index page
7685 * count (and clamp that to at least 10 pages, just in case). We
7686 * estimate that 90% of the index is entry pages, and the rest is data
7687 * pages. Estimate 100 entries per entry page; this is rather bogus
7688 * since it'll depend on the size of the keys, but it's more robust
7689 * than trying to predict the number of entries per heap tuple.
7690 */
7691 numPages = Max(numPages, 10);
7692 numEntryPages = floor((numPages - numPendingPages) * 0.90);
7693 numDataPages = numPages - numPendingPages - numEntryPages;
7694 numEntries = floor(numEntryPages * 100);
7695 }
7696
7697 /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
7698 if (numEntries < 1)
7699 numEntries = 1;
7700
7701 /*
7702 * If the index is partial, AND the index predicate with the index-bound
7703 * quals to produce a more accurate idea of the number of rows covered by
7704 * the bound conditions.
7705 */
7706 selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7707
7708 /* Estimate the fraction of main-table tuples that will be visited */
7709 *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7710 index->rel->relid,
7711 JOIN_INNER,
7712 NULL);
7713
7714 /* fetch estimated page cost for tablespace containing index */
7715 get_tablespace_page_costs(index->reltablespace,
7716 &spc_random_page_cost,
7717 NULL);
7718
7719 /*
7720 * Generic assumption about index correlation: there isn't any.
7721 */
7722 *indexCorrelation = 0.0;
7723
7724 /*
7725 * Examine quals to estimate number of search entries & partial matches
7726 */
7727 memset(&counts, 0, sizeof(counts));
7728 counts.arrayScans = 1;
7729 matchPossible = true;
7730
7731 foreach(lc, path->indexclauses)
7732 {
7733 IndexClause *iclause = lfirst_node(IndexClause, lc);
7734 ListCell *lc2;
7735
7736 foreach(lc2, iclause->indexquals)
7737 {
7738 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7739 Expr *clause = rinfo->clause;
7740
7741 if (IsA(clause, OpExpr))
7742 {
7743 matchPossible = gincost_opexpr(root,
7744 index,
7745 iclause->indexcol,
7746 (OpExpr *) clause,
7747 &counts);
7748 if (!matchPossible)
7749 break;
7750 }
7751 else if (IsA(clause, ScalarArrayOpExpr))
7752 {
7753 matchPossible = gincost_scalararrayopexpr(root,
7754 index,
7755 iclause->indexcol,
7756 (ScalarArrayOpExpr *) clause,
7757 numEntries,
7758 &counts);
7759 if (!matchPossible)
7760 break;
7761 }
7762 else
7763 {
7764 /* shouldn't be anything else for a GIN index */
7765 elog(ERROR, "unsupported GIN indexqual type: %d",
7766 (int) nodeTag(clause));
7767 }
7768 }
7769 }
7770
7771 /* Fall out if there were any provably-unsatisfiable quals */
7772 if (!matchPossible)
7773 {
7774 *indexStartupCost = 0;
7775 *indexTotalCost = 0;
7776 *indexSelectivity = 0;
7777 return;
7778 }
7779
7780 /*
7781 * If attribute has a full scan and at the same time doesn't have normal
7782 * scan, then we'll have to scan all non-null entries of that attribute.
7783 * Currently, we don't have per-attribute statistics for GIN. Thus, we
7784 * must assume the whole GIN index has to be scanned in this case.
7785 */
7786 fullIndexScan = false;
7787 for (i = 0; i < index->nkeycolumns; i++)
7788 {
7789 if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
7790 {
7791 fullIndexScan = true;
7792 break;
7793 }
7794 }
7795
7796 if (fullIndexScan || indexQuals == NIL)
7797 {
7798 /*
7799 * Full index scan will be required. We treat this as if every key in
7800 * the index had been listed in the query; is that reasonable?
7801 */
7802 counts.partialEntries = 0;
7803 counts.exactEntries = numEntries;
7804 counts.searchEntries = numEntries;
7805 }
7806
7807 /* Will we have more than one iteration of a nestloop scan? */
7808 outer_scans = loop_count;
7809
7810 /*
7811 * Compute cost to begin scan, first of all, pay attention to pending
7812 * list.
7813 */
7814 entryPagesFetched = numPendingPages;
7815
7816 /*
7817 * Estimate number of entry pages read. We need to do
7818 * counts.searchEntries searches. Use a power function as it should be,
7819 * but tuples on leaf pages usually is much greater. Here we include all
7820 * searches in entry tree, including search of first entry in partial
7821 * match algorithm
7822 */
7823 entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
7824
7825 /*
7826 * Add an estimate of entry pages read by partial match algorithm. It's a
7827 * scan over leaf pages in entry tree. We haven't any useful stats here,
7828 * so estimate it as proportion. Because counts.partialEntries is really
7829 * pretty bogus (see code above), it's possible that it is more than
7830 * numEntries; clamp the proportion to ensure sanity.
7831 */
7832 partialScale = counts.partialEntries / numEntries;
7833 partialScale = Min(partialScale, 1.0);
7834
7835 entryPagesFetched += ceil(numEntryPages * partialScale);
7836
7837 /*
7838 * Partial match algorithm reads all data pages before doing actual scan,
7839 * so it's a startup cost. Again, we haven't any useful stats here, so
7840 * estimate it as proportion.
7841 */
7842 dataPagesFetched = ceil(numDataPages * partialScale);
7843
7844 *indexStartupCost = 0;
7845 *indexTotalCost = 0;
7846
7847 /*
7848 * Add a CPU-cost component to represent the costs of initial entry btree
7849 * descent. We don't charge any I/O cost for touching upper btree levels,
7850 * since they tend to stay in cache, but we still have to do about log2(N)
7851 * comparisons to descend a btree of N leaf tuples. We charge one
7852 * cpu_operator_cost per comparison.
7853 *
7854 * If there are ScalarArrayOpExprs, charge this once per SA scan. The
7855 * ones after the first one are not startup cost so far as the overall
7856 * plan is concerned, so add them only to "total" cost.
7857 */
7858 if (numEntries > 1) /* avoid computing log(0) */
7859 {
7860 descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
7861 *indexStartupCost += descentCost * counts.searchEntries;
7862 *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
7863 }
7864
7865 /*
7866 * Add a cpu cost per entry-page fetched. This is not amortized over a
7867 * loop.
7868 */
7869 *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7870 *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7871
7872 /*
7873 * Add a cpu cost per data-page fetched. This is also not amortized over a
7874 * loop. Since those are the data pages from the partial match algorithm,
7875 * charge them as startup cost.
7876 */
7877 *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
7878
7879 /*
7880 * Since we add the startup cost to the total cost later on, remove the
7881 * initial arrayscan from the total.
7882 */
7883 *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7884
7885 /*
7886 * Calculate cache effects if more than one scan due to nestloops or array
7887 * quals. The result is pro-rated per nestloop scan, but the array qual
7888 * factor shouldn't be pro-rated (compare genericcostestimate).
7889 */
7890 if (outer_scans > 1 || counts.arrayScans > 1)
7891 {
7892 entryPagesFetched *= outer_scans * counts.arrayScans;
7893 entryPagesFetched = index_pages_fetched(entryPagesFetched,
7894 (BlockNumber) numEntryPages,
7895 numEntryPages, root);
7896 entryPagesFetched /= outer_scans;
7897 dataPagesFetched *= outer_scans * counts.arrayScans;
7898 dataPagesFetched = index_pages_fetched(dataPagesFetched,
7899 (BlockNumber) numDataPages,
7900 numDataPages, root);
7901 dataPagesFetched /= outer_scans;
7902 }
7903
7904 /*
7905 * Here we use random page cost because logically-close pages could be far
7906 * apart on disk.
7907 */
7908 *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
7909
7910 /*
7911 * Now compute the number of data pages fetched during the scan.
7912 *
7913 * We assume every entry to have the same number of items, and that there
7914 * is no overlap between them. (XXX: tsvector and array opclasses collect
7915 * statistics on the frequency of individual keys; it would be nice to use
7916 * those here.)
7917 */
7918 dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
7919
7920 /*
7921 * If there is a lot of overlap among the entries, in particular if one of
7922 * the entries is very frequent, the above calculation can grossly
7923 * under-estimate. As a simple cross-check, calculate a lower bound based
7924 * on the overall selectivity of the quals. At a minimum, we must read
7925 * one item pointer for each matching entry.
7926 *
7927 * The width of each item pointer varies, based on the level of
7928 * compression. We don't have statistics on that, but an average of
7929 * around 3 bytes per item is fairly typical.
7930 */
7931 dataPagesFetchedBySel = ceil(*indexSelectivity *
7932 (numTuples / (BLCKSZ / 3)));
7933 if (dataPagesFetchedBySel > dataPagesFetched)
7934 dataPagesFetched = dataPagesFetchedBySel;
7935
7936 /* Add one page cpu-cost to the startup cost */
7937 *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
7938
7939 /*
7940 * Add once again a CPU-cost for those data pages, before amortizing for
7941 * cache.
7942 */
7943 *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7944
7945 /* Account for cache effects, the same as above */
7946 if (outer_scans > 1 || counts.arrayScans > 1)
7947 {
7948 dataPagesFetched *= outer_scans * counts.arrayScans;
7949 dataPagesFetched = index_pages_fetched(dataPagesFetched,
7950 (BlockNumber) numDataPages,
7951 numDataPages, root);
7952 dataPagesFetched /= outer_scans;
7953 }
7954
7955 /* And apply random_page_cost as the cost per page */
7956 *indexTotalCost += *indexStartupCost +
7957 dataPagesFetched * spc_random_page_cost;
7958
7959 /*
7960 * Add on index qual eval costs, much as in genericcostestimate. We charge
7961 * cpu but we can disregard indexorderbys, since GIN doesn't support
7962 * those.
7963 */
7964 qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
7965 qual_op_cost = cpu_operator_cost * list_length(indexQuals);
7966
7967 *indexStartupCost += qual_arg_cost;
7968 *indexTotalCost += qual_arg_cost;
7969
7970 /*
7971 * Add a cpu cost per search entry, corresponding to the actual visited
7972 * entries.
7973 */
7974 *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
7975 /* Now add a cpu cost per tuple in the posting lists / trees */
7976 *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
7977 *indexPages = dataPagesFetched;
7978}
7979
7980/*
7981 * BRIN has search behavior completely different from other index types
7982 */
7983void
7984brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7985 Cost *indexStartupCost, Cost *indexTotalCost,
7986 Selectivity *indexSelectivity, double *indexCorrelation,
7987 double *indexPages)
7988{
7989 IndexOptInfo *index = path->indexinfo;
7990 List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7991 double numPages = index->pages;
7992 RelOptInfo *baserel = index->rel;
7993 RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
7994 Cost spc_seq_page_cost;
7995 Cost spc_random_page_cost;
7996 double qual_arg_cost;
7997 double qualSelectivity;
7998 BrinStatsData statsData;
7999 double indexRanges;
8000 double minimalRanges;
8001 double estimatedRanges;
8002 double selec;
8003 Relation indexRel;
8004 ListCell *l;
8005 VariableStatData vardata;
8006
8007 Assert(rte->rtekind == RTE_RELATION);
8008
8009 /* fetch estimated page cost for the tablespace containing the index */
8010 get_tablespace_page_costs(index->reltablespace,
8011 &spc_random_page_cost,
8012 &spc_seq_page_cost);
8013
8014 /*
8015 * Obtain some data from the index itself, if possible. Otherwise invent
8016 * some plausible internal statistics based on the relation page count.
8017 */
8018 if (!index->hypothetical)
8019 {
8020 /*
8021 * A lock should have already been obtained on the index in plancat.c.
8022 */
8023 indexRel = index_open(index->indexoid, NoLock);
8024 brinGetStats(indexRel, &statsData);
8025 index_close(indexRel, NoLock);
8026
8027 /* work out the actual number of ranges in the index */
8028 indexRanges = Max(ceil((double) baserel->pages /
8029 statsData.pagesPerRange), 1.0);
8030 }
8031 else
8032 {
8033 /*
8034 * Assume default number of pages per range, and estimate the number
8035 * of ranges based on that.
8036 */
8037 indexRanges = Max(ceil((double) baserel->pages /
8039
8041 statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
8042 }
8043
8044 /*
8045 * Compute index correlation
8046 *
8047 * Because we can use all index quals equally when scanning, we can use
8048 * the largest correlation (in absolute value) among columns used by the
8049 * query. Start at zero, the worst possible case. If we cannot find any
8050 * correlation statistics, we will keep it as 0.
8051 */
8052 *indexCorrelation = 0;
8053
8054 foreach(l, path->indexclauses)
8055 {
8056 IndexClause *iclause = lfirst_node(IndexClause, l);
8057 AttrNumber attnum = index->indexkeys[iclause->indexcol];
8058
8059 /* attempt to lookup stats in relation for this index column */
8060 if (attnum != 0)
8061 {
8062 /* Simple variable -- look to stats for the underlying table */
8064 (*get_relation_stats_hook) (root, rte, attnum, &vardata))
8065 {
8066 /*
8067 * The hook took control of acquiring a stats tuple. If it
8068 * did supply a tuple, it'd better have supplied a freefunc.
8069 */
8070 if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
8071 elog(ERROR,
8072 "no function provided to release variable stats with");
8073 }
8074 else
8075 {
8076 vardata.statsTuple =
8077 SearchSysCache3(STATRELATTINH,
8078 ObjectIdGetDatum(rte->relid),
8080 BoolGetDatum(false));
8081 vardata.freefunc = ReleaseSysCache;
8082 }
8083 }
8084 else
8085 {
8086 /*
8087 * Looks like we've found an expression column in the index. Let's
8088 * see if there's any stats for it.
8089 */
8090
8091 /* get the attnum from the 0-based index. */
8092 attnum = iclause->indexcol + 1;
8093
8095 (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
8096 {
8097 /*
8098 * The hook took control of acquiring a stats tuple. If it
8099 * did supply a tuple, it'd better have supplied a freefunc.
8100 */
8101 if (HeapTupleIsValid(vardata.statsTuple) &&
8102 !vardata.freefunc)
8103 elog(ERROR, "no function provided to release variable stats with");
8104 }
8105 else
8106 {
8107 vardata.statsTuple = SearchSysCache3(STATRELATTINH,
8108 ObjectIdGetDatum(index->indexoid),
8110 BoolGetDatum(false));
8111 vardata.freefunc = ReleaseSysCache;
8112 }
8113 }
8114
8115 if (HeapTupleIsValid(vardata.statsTuple))
8116 {
8117 AttStatsSlot sslot;
8118
8119 if (get_attstatsslot(&sslot, vardata.statsTuple,
8120 STATISTIC_KIND_CORRELATION, InvalidOid,
8122 {
8123 double varCorrelation = 0.0;
8124
8125 if (sslot.nnumbers > 0)
8126 varCorrelation = fabs(sslot.numbers[0]);
8127
8128 if (varCorrelation > *indexCorrelation)
8129 *indexCorrelation = varCorrelation;
8130
8131 free_attstatsslot(&sslot);
8132 }
8133 }
8134
8135 ReleaseVariableStats(vardata);
8136 }
8137
8138 qualSelectivity = clauselist_selectivity(root, indexQuals,
8139 baserel->relid,
8140 JOIN_INNER, NULL);
8141
8142 /*
8143 * Now calculate the minimum possible ranges we could match with if all of
8144 * the rows were in the perfect order in the table's heap.
8145 */
8146 minimalRanges = ceil(indexRanges * qualSelectivity);
8147
8148 /*
8149 * Now estimate the number of ranges that we'll touch by using the
8150 * indexCorrelation from the stats. Careful not to divide by zero (note
8151 * we're using the absolute value of the correlation).
8152 */
8153 if (*indexCorrelation < 1.0e-10)
8154 estimatedRanges = indexRanges;
8155 else
8156 estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
8157
8158 /* we expect to visit this portion of the table */
8159 selec = estimatedRanges / indexRanges;
8160
8161 CLAMP_PROBABILITY(selec);
8162
8163 *indexSelectivity = selec;
8164
8165 /*
8166 * Compute the index qual costs, much as in genericcostestimate, to add to
8167 * the index costs. We can disregard indexorderbys, since BRIN doesn't
8168 * support those.
8169 */
8170 qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8171
8172 /*
8173 * Compute the startup cost as the cost to read the whole revmap
8174 * sequentially, including the cost to execute the index quals.
8175 */
8176 *indexStartupCost =
8177 spc_seq_page_cost * statsData.revmapNumPages * loop_count;
8178 *indexStartupCost += qual_arg_cost;
8179
8180 /*
8181 * To read a BRIN index there might be a bit of back and forth over
8182 * regular pages, as revmap might point to them out of sequential order;
8183 * calculate the total cost as reading the whole index in random order.
8184 */
8185 *indexTotalCost = *indexStartupCost +
8186 spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
8187
8188 /*
8189 * Charge a small amount per range tuple which we expect to match to. This
8190 * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
8191 * will set a bit for each page in the range when we find a matching
8192 * range, so we must multiply the charge by the number of pages in the
8193 * range.
8194 */
8195 *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
8196 statsData.pagesPerRange;
8197
8198 *indexPages = index->pages;
8199}
Datum idx(PG_FUNCTION_ARGS)
Definition: _int_op.c:259
@ ACLCHECK_OK
Definition: acl.h:183
AclResult pg_attribute_aclcheck(Oid table_oid, AttrNumber attnum, Oid roleid, AclMode mode)
Definition: aclchk.c:3842
AclResult pg_class_aclcheck(Oid table_oid, Oid roleid, AclMode mode)
Definition: aclchk.c:4013
#define ARR_NDIM(a)
Definition: array.h:290
#define DatumGetArrayTypeP(X)
Definition: array.h:261
#define ARR_ELEMTYPE(a)
Definition: array.h:292
#define ARR_DIMS(a)
Definition: array.h:294
Selectivity scalararraysel_containment(PlannerInfo *root, Node *leftop, Node *rightop, Oid elemtype, bool isEquality, bool useOr, int varRelid)
void deconstruct_array(ArrayType *array, Oid elmtype, int elmlen, bool elmbyval, char elmalign, Datum **elemsp, bool **nullsp, int *nelemsp)
Definition: arrayfuncs.c:3631
int ArrayGetNItems(int ndim, const int *dims)
Definition: arrayutils.c:57
int16 AttrNumber
Definition: attnum.h:21
#define AttrNumberIsForUserDefinedAttr(attributeNumber)
Definition: attnum.h:41
#define InvalidAttrNumber
Definition: attnum.h:23
Datum numeric_float8_no_overflow(PG_FUNCTION_ARGS)
Definition: numeric.c:4779
bool bms_is_subset(const Bitmapset *a, const Bitmapset *b)
Definition: bitmapset.c:412
void bms_free(Bitmapset *a)
Definition: bitmapset.c:239
int bms_num_members(const Bitmapset *a)
Definition: bitmapset.c:751
bool bms_is_member(int x, const Bitmapset *a)
Definition: bitmapset.c:510
Bitmapset * bms_add_member(Bitmapset *a, int x)
Definition: bitmapset.c:815
bool bms_get_singleton_member(const Bitmapset *a, int *member)
Definition: bitmapset.c:715
#define bms_is_empty(a)
Definition: bitmapset.h:118
uint32 BlockNumber
Definition: block.h:31
#define InvalidBlockNumber
Definition: block.h:33
static Datum values[MAXATTR]
Definition: bootstrap.c:151
void brinGetStats(Relation index, BrinStatsData *stats)
Definition: brin.c:1640
#define BRIN_DEFAULT_PAGES_PER_RANGE
Definition: brin.h:39
#define REVMAP_PAGE_MAXITEMS
Definition: brin_page.h:93
int Buffer
Definition: buf.h:23
#define InvalidBuffer
Definition: buf.h:25
void ReleaseBuffer(Buffer buffer)
Definition: bufmgr.c:4924
#define TextDatumGetCString(d)
Definition: builtins.h:98
#define NameStr(name)
Definition: c.h:700
#define Min(x, y)
Definition: c.h:958
#define PG_USED_FOR_ASSERTS_ONLY
Definition: c.h:201
#define Max(x, y)
Definition: c.h:952
char * Pointer
Definition: c.h:476
#define Assert(condition)
Definition: c.h:812
double float8
Definition: c.h:584
int16_t int16
Definition: c.h:480
regproc RegProcedure
Definition: c.h:604
int32_t int32
Definition: c.h:481
unsigned int Index
Definition: c.h:568
#define MemSet(start, val, len)
Definition: c.h:974
#define OidIsValid(objectId)
Definition: c.h:729
size_t Size
Definition: c.h:559
int NumRelids(PlannerInfo *root, Node *clause)
Definition: clauses.c:2129
Node * estimate_expression_value(PlannerInfo *root, Node *node)
Definition: clauses.c:2394
bool contain_volatile_functions(Node *clause)
Definition: clauses.c:537
double expression_returns_set_rows(PlannerInfo *root, Node *clause)
Definition: clauses.c:288
Selectivity clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:100
Selectivity clause_selectivity(PlannerInfo *root, Node *clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: clausesel.c:667
Oid collid
double cpu_operator_cost
Definition: costsize.c:134
double index_pages_fetched(double tuples_fetched, BlockNumber pages, double index_pages, PlannerInfo *root)
Definition: costsize.c:908
void cost_qual_eval_node(QualCost *cost, Node *qual, PlannerInfo *root)
Definition: costsize.c:4758
double clamp_row_est(double nrows)
Definition: costsize.c:213
double cpu_index_tuple_cost
Definition: costsize.c:133
#define MONTHS_PER_YEAR
Definition: timestamp.h:108
#define USECS_PER_DAY
Definition: timestamp.h:131
#define DAYS_PER_YEAR
Definition: timestamp.h:107
double date2timestamp_no_overflow(DateADT dateVal)
Definition: date.c:739
static TimeTzADT * DatumGetTimeTzADTP(Datum X)
Definition: date.h:66
static DateADT DatumGetDateADT(Datum X)
Definition: date.h:54
static TimeADT DatumGetTimeADT(Datum X)
Definition: date.h:60
Datum datumCopy(Datum value, bool typByVal, int typLen)
Definition: datum.c:132
int errmsg_internal(const char *fmt,...)
Definition: elog.c:1157
#define DEBUG2
Definition: elog.h:29
#define ERROR
Definition: elog.h:39
#define elog(elevel,...)
Definition: elog.h:225
#define ereport(elevel,...)
Definition: elog.h:149
bool equal(const void *a, const void *b)
Definition: equalfuncs.c:223
bool exprs_known_equal(PlannerInfo *root, Node *item1, Node *item2, Oid opfamily)
Definition: equivclass.c:2498
void ExecDropSingleTupleTableSlot(TupleTableSlot *slot)
Definition: execTuples.c:1441
HeapTuple statext_expressions_load(Oid stxoid, bool inh, int idx)
Datum FunctionCall4Coll(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2, Datum arg3, Datum arg4)
Definition: fmgr.c:1196
void set_fn_opclass_options(FmgrInfo *flinfo, bytea *options)
Definition: fmgr.c:2070
Datum FunctionCall2Coll(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2)
Definition: fmgr.c:1149
void fmgr_info(Oid functionId, FmgrInfo *finfo)
Definition: fmgr.c:127
Datum FunctionCall5Coll(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2, Datum arg3, Datum arg4, Datum arg5)
Definition: fmgr.c:1223
Datum DirectFunctionCall5Coll(PGFunction func, Oid collation, Datum arg1, Datum arg2, Datum arg3, Datum arg4, Datum arg5)
Definition: fmgr.c:886
Datum FunctionCall7Coll(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2, Datum arg3, Datum arg4, Datum arg5, Datum arg6, Datum arg7)
Definition: fmgr.c:1284
#define PG_GETARG_OID(n)
Definition: fmgr.h:275
#define DatumGetByteaPP(X)
Definition: fmgr.h:291
#define PG_RETURN_FLOAT8(x)
Definition: fmgr.h:367
#define PG_GETARG_POINTER(n)
Definition: fmgr.h:276
#define InitFunctionCallInfoData(Fcinfo, Flinfo, Nargs, Collation, Context, Resultinfo)
Definition: fmgr.h:150
#define DirectFunctionCall1(func, arg1)
Definition: fmgr.h:641
#define LOCAL_FCINFO(name, nargs)
Definition: fmgr.h:110
#define FunctionCallInvoke(fcinfo)
Definition: fmgr.h:172
#define PG_GETARG_INT32(n)
Definition: fmgr.h:269
#define PG_GET_COLLATION()
Definition: fmgr.h:198
#define PG_FUNCTION_ARGS
Definition: fmgr.h:193
#define PG_GETARG_INT16(n)
Definition: fmgr.h:271
#define GIN_EXTRACTQUERY_PROC
Definition: gin.h:24
#define GIN_SEARCH_MODE_DEFAULT
Definition: gin.h:34
#define GIN_SEARCH_MODE_INCLUDE_EMPTY
Definition: gin.h:35
void ginGetStats(Relation index, GinStatsData *stats)
Definition: ginutil.c:624
#define HeapTupleIsValid(tuple)
Definition: htup.h:78
#define GETSTRUCT(TUP)
Definition: htup_details.h:653
void index_close(Relation relation, LOCKMODE lockmode)
Definition: indexam.c:177
ItemPointer index_getnext_tid(IndexScanDesc scan, ScanDirection direction)
Definition: indexam.c:576
IndexScanDesc index_beginscan(Relation heapRelation, Relation indexRelation, Snapshot snapshot, int nkeys, int norderbys)
Definition: indexam.c:256
bool index_fetch_heap(IndexScanDesc scan, TupleTableSlot *slot)
Definition: indexam.c:634
void index_endscan(IndexScanDesc scan)
Definition: indexam.c:378
Relation index_open(Oid relationId, LOCKMODE lockmode)
Definition: indexam.c:133
void index_rescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int norderbys)
Definition: indexam.c:352
void index_deform_tuple(IndexTuple tup, TupleDesc tupleDescriptor, Datum *values, bool *isnull)
Definition: indextuple.c:456
bool match_index_to_operand(Node *operand, int indexcol, IndexOptInfo *index)
Definition: indxpath.c:4374
static struct @161 value
long val
Definition: informix.c:689
int j
Definition: isn.c:73
int i
Definition: isn.c:72
if(TABLE==NULL||TABLE_index==NULL)
Definition: isn.c:76
static OffsetNumber ItemPointerGetOffsetNumberNoCheck(const ItemPointerData *pointer)
Definition: itemptr.h:114
static BlockNumber ItemPointerGetBlockNumber(const ItemPointerData *pointer)
Definition: itemptr.h:103
static BlockNumber ItemPointerGetBlockNumberNoCheck(const ItemPointerData *pointer)
Definition: itemptr.h:93
ItemPointerData * ItemPointer
Definition: itemptr.h:49
List * lappend(List *list, void *datum)
Definition: list.c:339
List * list_concat(List *list1, const List *list2)
Definition: list.c:561
bool list_member_int(const List *list, int datum)
Definition: list.c:702
#define NoLock
Definition: lockdefs.h:34
char * get_rel_name(Oid relid)
Definition: lsyscache.c:1928
void get_op_opfamily_properties(Oid opno, Oid opfamily, bool ordering_op, int *strategy, Oid *lefttype, Oid *righttype)
Definition: lsyscache.c:136
RegProcedure get_oprrest(Oid opno)
Definition: lsyscache.c:1557
void free_attstatsslot(AttStatsSlot *sslot)
Definition: lsyscache.c:3344
bool comparison_ops_are_compatible(Oid opno1, Oid opno2)
Definition: lsyscache.c:749
void get_typlenbyvalalign(Oid typid, int16 *typlen, bool *typbyval, char *typalign)
Definition: lsyscache.c:2271
Oid get_opfamily_proc(Oid opfamily, Oid lefttype, Oid righttype, int16 procnum)
Definition: lsyscache.c:796
RegProcedure get_oprjoin(Oid opno)
Definition: lsyscache.c:1581
void get_typlenbyval(Oid typid, int16 *typlen, bool *typbyval)
Definition: lsyscache.c:2251
RegProcedure get_opcode(Oid opno)
Definition: lsyscache.c:1285
int get_op_opfamily_strategy(Oid opno, Oid opfamily)
Definition: lsyscache.c:83
Oid get_opfamily_member(Oid opfamily, Oid lefttype, Oid righttype, int16 strategy)
Definition: lsyscache.c:166
bool get_func_leakproof(Oid funcid)
Definition: lsyscache.c:1837
char * get_func_name(Oid funcid)
Definition: lsyscache.c:1608
Oid get_base_element_type(Oid typid)
Definition: lsyscache.c:2832
bool get_attstatsslot(AttStatsSlot *sslot, HeapTuple statstuple, int reqkind, Oid reqop, int flags)
Definition: lsyscache.c:3234
Oid get_negator(Oid opno)
Definition: lsyscache.c:1533
Oid get_commutator(Oid opno)
Definition: lsyscache.c:1509
#define ATTSTATSSLOT_NUMBERS
Definition: lsyscache.h:43
#define ATTSTATSSLOT_VALUES
Definition: lsyscache.h:42
Const * makeConst(Oid consttype, int32 consttypmod, Oid constcollid, int constlen, Datum constvalue, bool constisnull, bool constbyval)
Definition: makefuncs.c:301
char * pstrdup(const char *in)
Definition: mcxt.c:1696
void pfree(void *pointer)
Definition: mcxt.c:1521
void * palloc0(Size size)
Definition: mcxt.c:1347
void * palloc(Size size)
Definition: mcxt.c:1317
MemoryContext CurrentMemoryContext
Definition: mcxt.c:143
void MemoryContextDelete(MemoryContext context)
Definition: mcxt.c:454
#define AllocSetContextCreate
Definition: memutils.h:129
#define ALLOCSET_DEFAULT_SIZES
Definition: memutils.h:160
Oid GetUserId(void)
Definition: miscinit.c:517
MVNDistinct * statext_ndistinct_load(Oid mvoid, bool inh)
Definition: mvdistinct.c:148
double convert_network_to_scalar(Datum value, Oid typid, bool *failure)
Definition: network.c:1496
Size hash_agg_entry_size(int numTrans, Size tupleWidth, Size transitionSpace)
Definition: nodeAgg.c:1694
Oid exprType(const Node *expr)
Definition: nodeFuncs.c:42
int32 exprTypmod(const Node *expr)
Definition: nodeFuncs.c:298
Oid exprCollation(const Node *expr)
Definition: nodeFuncs.c:816
static Node * get_rightop(const void *clause)
Definition: nodeFuncs.h:95
static bool is_opclause(const void *clause)
Definition: nodeFuncs.h:76
static Node * get_leftop(const void *clause)
Definition: nodeFuncs.h:83
#define IsA(nodeptr, _type_)
Definition: nodes.h:158
double Cost
Definition: nodes.h:251
#define nodeTag(nodeptr)
Definition: nodes.h:133
double Selectivity
Definition: nodes.h:250
#define makeNode(_type_)
Definition: nodes.h:155
JoinType
Definition: nodes.h:288
@ JOIN_SEMI
Definition: nodes.h:307
@ JOIN_FULL
Definition: nodes.h:295
@ JOIN_INNER
Definition: nodes.h:293
@ JOIN_LEFT
Definition: nodes.h:294
@ JOIN_ANTI
Definition: nodes.h:308
uint16 OffsetNumber
Definition: off.h:24
#define PVC_RECURSE_AGGREGATES
Definition: optimizer.h:188
#define PVC_RECURSE_PLACEHOLDERS
Definition: optimizer.h:192
#define PVC_RECURSE_WINDOWFUNCS
Definition: optimizer.h:190
bool targetIsInSortList(TargetEntry *tle, Oid sortop, List *sortList)
RTEPermissionInfo * getRTEPermissionInfo(List *rteperminfos, RangeTblEntry *rte)
TargetEntry * get_tle_by_resno(List *tlist, AttrNumber resno)
@ RTE_CTE
Definition: parsenodes.h:1023
@ RTE_VALUES
Definition: parsenodes.h:1022
@ RTE_SUBQUERY
Definition: parsenodes.h:1018
@ RTE_RELATION
Definition: parsenodes.h:1017
#define ACL_SELECT
Definition: parsenodes.h:77
#define IS_SIMPLE_REL(rel)
Definition: pathnodes.h:839
#define planner_rt_fetch(rti, root)
Definition: pathnodes.h:570
int16 attnum
Definition: pg_attribute.h:74
void * arg
#define INDEX_MAX_KEYS
#define lfirst(lc)
Definition: pg_list.h:172
#define lfirst_node(type, lc)
Definition: pg_list.h:176
static int list_length(const List *l)
Definition: pg_list.h:152
#define NIL
Definition: pg_list.h:68
#define foreach_delete_current(lst, var_or_cell)
Definition: pg_list.h:391
#define list_make1(x1)
Definition: pg_list.h:212
#define for_each_from(cell, lst, N)
Definition: pg_list.h:414
static void * list_nth(const List *list, int n)
Definition: pg_list.h:299
#define linitial(l)
Definition: pg_list.h:178
#define lsecond(l)
Definition: pg_list.h:183
static ListCell * list_head(const List *l)
Definition: pg_list.h:128
static ListCell * lnext(const List *l, const ListCell *c)
Definition: pg_list.h:343
#define linitial_oid(l)
Definition: pg_list.h:180
#define list_make2(x1, x2)
Definition: pg_list.h:214
static int list_nth_int(const List *list, int n)
Definition: pg_list.h:310
pg_locale_t pg_newlocale_from_collation(Oid collid)
Definition: pg_locale.c:1341
size_t pg_strxfrm(char *dest, const char *src, size_t destsize, pg_locale_t locale)
Definition: pg_locale.c:1643
FormData_pg_statistic * Form_pg_statistic
Definition: pg_statistic.h:135
static int scale
Definition: pgbench.c:181
Selectivity restriction_selectivity(PlannerInfo *root, Oid operatorid, List *args, Oid inputcollid, int varRelid)
Definition: plancat.c:1956
bool has_unique_index(RelOptInfo *rel, AttrNumber attno)
Definition: plancat.c:2217
Selectivity join_selectivity(PlannerInfo *root, Oid operatorid, List *args, Oid inputcollid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: plancat.c:1995
static bool DatumGetBool(Datum X)
Definition: postgres.h:90
static int64 DatumGetInt64(Datum X)
Definition: postgres.h:385
static Datum PointerGetDatum(const void *X)
Definition: postgres.h:322
uintptr_t Datum
Definition: postgres.h:64
static float4 DatumGetFloat4(Datum X)
Definition: postgres.h:458
static Oid DatumGetObjectId(Datum X)
Definition: postgres.h:242
static Datum Int16GetDatum(int16 X)
Definition: postgres.h:172
static Datum UInt16GetDatum(uint16 X)
Definition: postgres.h:192
static Datum BoolGetDatum(bool X)
Definition: postgres.h:102
static float8 DatumGetFloat8(Datum X)
Definition: postgres.h:494
static Datum ObjectIdGetDatum(Oid X)
Definition: postgres.h:252
static Pointer DatumGetPointer(Datum X)
Definition: postgres.h:312
static char DatumGetChar(Datum X)
Definition: postgres.h:112
static Datum Int32GetDatum(int32 X)
Definition: postgres.h:212
static int16 DatumGetInt16(Datum X)
Definition: postgres.h:162
static int32 DatumGetInt32(Datum X)
Definition: postgres.h:202
#define InvalidOid
Definition: postgres_ext.h:36
unsigned int Oid
Definition: postgres_ext.h:31
bool predicate_implied_by(List *predicate_list, List *clause_list, bool weak)
Definition: predtest.c:152
char * s1
char * s2
BoolTestType
Definition: primnodes.h:1975
@ IS_NOT_TRUE
Definition: primnodes.h:1976
@ IS_NOT_FALSE
Definition: primnodes.h:1976
@ IS_NOT_UNKNOWN
Definition: primnodes.h:1976
@ IS_TRUE
Definition: primnodes.h:1976
@ IS_UNKNOWN
Definition: primnodes.h:1976
@ IS_FALSE
Definition: primnodes.h:1976
NullTestType
Definition: primnodes.h:1951
@ IS_NULL
Definition: primnodes.h:1952
@ IS_NOT_NULL
Definition: primnodes.h:1952
GlobalVisState * GlobalVisTestFor(Relation rel)
Definition: procarray.c:4107
MemoryContextSwitchTo(old_ctx)
tree ctl root
Definition: radixtree.h:1888
#define RelationGetRelationName(relation)
Definition: rel.h:539
RelOptInfo * find_base_rel(PlannerInfo *root, int relid)
Definition: relnode.c:414
RelOptInfo * find_base_rel_noerr(PlannerInfo *root, int relid)
Definition: relnode.c:436
RelOptInfo * find_join_rel(PlannerInfo *root, Relids relids)
Definition: relnode.c:527
void ScanKeyEntryInitialize(ScanKey entry, int flags, AttrNumber attributeNumber, StrategyNumber strategy, Oid subtype, Oid collation, RegProcedure procedure, Datum argument)
Definition: scankey.c:32
ScanDirection
Definition: sdir.h:25
@ BackwardScanDirection
Definition: sdir.h:26
@ ForwardScanDirection
Definition: sdir.h:28
static bool get_actual_variable_endpoint(Relation heapRel, Relation indexRel, ScanDirection indexscandir, ScanKey scankeys, int16 typLen, bool typByVal, TupleTableSlot *tableslot, MemoryContext outercontext, Datum *endpointDatum)
Definition: selfuncs.c:6278
bool get_restriction_variable(PlannerInfo *root, List *args, int varRelid, VariableStatData *vardata, Node **other, bool *varonleft)
Definition: selfuncs.c:4894
Datum neqsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:557
static RelOptInfo * find_join_input_rel(PlannerInfo *root, Relids relids)
Definition: selfuncs.c:6439
static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata, Oid sortop, Oid collation, Datum *min, Datum *max)
Definition: selfuncs.c:5908
void btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
Definition: selfuncs.c:6799
List * get_quals_from_indexclauses(List *indexclauses)
Definition: selfuncs.c:6471
static void convert_string_to_scalar(char *value, double *scaledvalue, char *lobound, double *scaledlobound, char *hibound, double *scaledhibound)
Definition: selfuncs.c:4518
double var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation, Datum constval, bool constisnull, bool varonleft, bool negate)
Definition: selfuncs.c:295
List * add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
Definition: selfuncs.c:6778
double generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation, List *args, int varRelid, double default_selectivity)
Definition: selfuncs.c:914
#define VISITED_PAGES_LIMIT
void spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
Definition: selfuncs.c:7239
Datum scalargtsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:1489
#define DEFAULT_PAGE_CPU_MULTIPLIER
Definition: selfuncs.c:144
static bool estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel, List **varinfos, double *ndistinct)
Definition: selfuncs.c:3958
Selectivity booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: selfuncs.c:1540
Datum eqjoinsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:2272
double estimate_array_length(PlannerInfo *root, Node *arrayexpr)
Definition: selfuncs.c:2139
double mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation, Datum constval, bool varonleft, double *sumcommonp)
Definition: selfuncs.c:732
Selectivity nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: selfuncs.c:1698
static void examine_simple_variable(PlannerInfo *root, Var *var, VariableStatData *vardata)
Definition: selfuncs.c:5417
static List * add_unique_group_var(PlannerInfo *root, List *varinfos, Node *var, VariableStatData *vardata)
Definition: selfuncs.c:3299
Datum matchingsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:3260
static double eqjoinsel_inner(Oid opfuncoid, Oid collation, VariableStatData *vardata1, VariableStatData *vardata2, double nd1, double nd2, bool isdefault1, bool isdefault2, AttStatsSlot *sslot1, AttStatsSlot *sslot2, Form_pg_statistic stats1, Form_pg_statistic stats2, bool have_mcvs1, bool have_mcvs2)
Definition: selfuncs.c:2437
Datum eqsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:227
void examine_variable(PlannerInfo *root, Node *node, int varRelid, VariableStatData *vardata)
Definition: selfuncs.c:5023
Datum scalargtjoinsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:2918
static double convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
Definition: selfuncs.c:4598
void mergejoinscansel(PlannerInfo *root, Node *clause, Oid opfamily, int strategy, bool nulls_first, Selectivity *leftstart, Selectivity *leftend, Selectivity *rightstart, Selectivity *rightend)
Definition: selfuncs.c:2955
static Datum scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
Definition: selfuncs.c:1400
static double eqjoinsel_semi(Oid opfuncoid, Oid collation, VariableStatData *vardata1, VariableStatData *vardata2, double nd1, double nd2, bool isdefault1, bool isdefault2, AttStatsSlot *sslot1, AttStatsSlot *sslot2, Form_pg_statistic stats1, Form_pg_statistic stats2, bool have_mcvs1, bool have_mcvs2, RelOptInfo *inner_rel)
Definition: selfuncs.c:2634
void gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
Definition: selfuncs.c:7594
static double convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
Definition: selfuncs.c:4828
static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
Definition: selfuncs.c:4456
static Node * strip_array_coercion(Node *node)
Definition: selfuncs.c:1783
double estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows, List **pgset, EstimationInfo *estinfo)
Definition: selfuncs.c:3420
static bool convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue, Datum lobound, Datum hibound, Oid boundstypid, double *scaledlobound, double *scaledhibound)
Definition: selfuncs.c:4309
double ineq_histogram_selectivity(PlannerInfo *root, VariableStatData *vardata, Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq, Oid collation, Datum constval, Oid consttype)
Definition: selfuncs.c:1041
void genericcostestimate(PlannerInfo *root, IndexPath *path, double loop_count, GenericCosts *costs)
Definition: selfuncs.c:6555
Datum scalarltjoinsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:2900
static bool gincost_pattern(IndexOptInfo *index, int indexcol, Oid clause_op, Datum query, GinQualCounts *counts)
Definition: selfuncs.c:7314
void brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
Definition: selfuncs.c:7984
void gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
Definition: selfuncs.c:7184
Datum scalargejoinsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:2927
get_index_stats_hook_type get_index_stats_hook
Definition: selfuncs.c:148
Datum matchingjoinsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:3278
static bool gincost_scalararrayopexpr(PlannerInfo *root, IndexOptInfo *index, int indexcol, ScalarArrayOpExpr *clause, double numIndexEntries, GinQualCounts *counts)
Definition: selfuncs.c:7478
double histogram_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation, Datum constval, bool varonleft, int min_hist_size, int n_skip, int *hist_size)
Definition: selfuncs.c:823
Selectivity boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
Definition: selfuncs.c:1512
Datum scalarlesel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:1480
Datum scalargesel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:1498
static double scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq, Oid collation, VariableStatData *vardata, Datum constval, Oid consttype)
Definition: selfuncs.c:580
static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen, int rangelo, int rangehi)
Definition: selfuncs.c:4785
Selectivity scalararraysel(PlannerInfo *root, ScalarArrayOpExpr *clause, bool is_join_clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: selfuncs.c:1816
Datum scalarltsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:1471
double var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation, Node *other, bool varonleft, bool negate)
Definition: selfuncs.c:466
static bool get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata, Oid sortop, Oid collation, Datum *min, Datum *max)
Definition: selfuncs.c:6098
Datum scalarlejoinsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:2909
double get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
Definition: selfuncs.c:5775
bool statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
Definition: selfuncs.c:5746
void hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count, Cost *indexStartupCost, Cost *indexTotalCost, Selectivity *indexSelectivity, double *indexCorrelation, double *indexPages)
Definition: selfuncs.c:7142
Datum neqjoinsel(PG_FUNCTION_ARGS)
Definition: selfuncs.c:2822
double estimate_hashagg_tablesize(PlannerInfo *root, Path *path, const AggClauseCosts *agg_costs, double dNumGroups)
Definition: selfuncs.c:3921
void estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets, Selectivity *mcv_freq, Selectivity *bucketsize_frac)
Definition: selfuncs.c:3802
static void convert_bytea_to_scalar(Datum value, double *scaledvalue, Datum lobound, double *scaledlobound, Datum hibound, double *scaledhibound)
Definition: selfuncs.c:4737
Cost index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
Definition: selfuncs.c:6501
get_relation_stats_hook_type get_relation_stats_hook
Definition: selfuncs.c:147
Selectivity rowcomparesel(PlannerInfo *root, RowCompareExpr *clause, int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
Definition: selfuncs.c:2205
static bool gincost_opexpr(PlannerInfo *root, IndexOptInfo *index, int indexcol, OpExpr *clause, GinQualCounts *counts)
Definition: selfuncs.c:7428
static void ReleaseDummy(HeapTuple tuple)
Definition: selfuncs.c:4982
static char * convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
Definition: selfuncs.c:4649
static double eqsel_internal(PG_FUNCTION_ARGS, bool negate)
Definition: selfuncs.c:236
static void get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc, Oid collation, int16 typLen, bool typByVal, Datum *min, Datum *max, bool *p_have_data)
Definition: selfuncs.c:6035
void get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo, VariableStatData *vardata1, VariableStatData *vardata2, bool *join_is_reversed)
Definition: selfuncs.c:4954
#define DEFAULT_NOT_UNK_SEL
Definition: selfuncs.h:56
#define ReleaseVariableStats(vardata)
Definition: selfuncs.h:99
#define CLAMP_PROBABILITY(p)
Definition: selfuncs.h:63
bool(* get_relation_stats_hook_type)(PlannerInfo *root, RangeTblEntry *rte, AttrNumber attnum, VariableStatData *vardata)
Definition: selfuncs.h:138
#define DEFAULT_UNK_SEL
Definition: selfuncs.h:55
bool(* get_index_stats_hook_type)(PlannerInfo *root, Oid indexOid, AttrNumber indexattnum, VariableStatData *vardata)
Definition: selfuncs.h:143
#define DEFAULT_EQ_SEL
Definition: selfuncs.h:34
#define DEFAULT_MATCHING_SEL
Definition: selfuncs.h:49
#define DEFAULT_INEQ_SEL
Definition: selfuncs.h:37
#define DEFAULT_NUM_DISTINCT
Definition: selfuncs.h:52
#define SELFLAG_USED_DEFAULT
Definition: selfuncs.h:76
#define SK_SEARCHNOTNULL
Definition: skey.h:122
#define SK_ISNULL
Definition: skey.h:115
#define InitNonVacuumableSnapshot(snapshotdata, vistestp)
Definition: snapmgr.h:50
void get_tablespace_page_costs(Oid spcid, double *spc_random_page_cost, double *spc_seq_page_cost)
Definition: spccache.c:182
#define BTGreaterStrategyNumber
Definition: stratnum.h:33
#define InvalidStrategy
Definition: stratnum.h:24
#define BTLessStrategyNumber
Definition: stratnum.h:29
#define BTEqualStrategyNumber
Definition: stratnum.h:31
#define BTLessEqualStrategyNumber
Definition: stratnum.h:30
#define BTGreaterEqualStrategyNumber
Definition: stratnum.h:32
Size transitionSpace
Definition: pathnodes.h:62
Index parent_relid
Definition: pathnodes.h:2980
int num_child_cols
Definition: pathnodes.h:3016
List * elements
Definition: primnodes.h:1380
Datum * values
Definition: lsyscache.h:53
float4 * numbers
Definition: lsyscache.h:56
int nnumbers
Definition: lsyscache.h:57
BlockNumber revmapNumPages
Definition: brin.h:35
BlockNumber pagesPerRange
Definition: brin.h:34
uint32 flags
Definition: selfuncs.h:80
Definition: fmgr.h:57
Oid fn_oid
Definition: fmgr.h:59
Selectivity indexSelectivity
Definition: selfuncs.h:127
Cost indexStartupCost
Definition: selfuncs.h:125
double indexCorrelation
Definition: selfuncs.h:128
double spc_random_page_cost
Definition: selfuncs.h:133
double num_sa_scans
Definition: selfuncs.h:134
Cost indexTotalCost
Definition: selfuncs.h:126
double numIndexPages
Definition: selfuncs.h:131
double numIndexTuples
Definition: selfuncs.h:132
bool attHasNormalScan[INDEX_MAX_KEYS]
Definition: selfuncs.c:7301
double exactEntries
Definition: selfuncs.c:7303
double arrayScans
Definition: selfuncs.c:7305
double partialEntries
Definition: selfuncs.c:7302
bool attHasFullScan[INDEX_MAX_KEYS]
Definition: selfuncs.c:7300
double searchEntries
Definition: selfuncs.c:7304
BlockNumber nDataPages
Definition: gin.h:47
BlockNumber nPendingPages
Definition: gin.h:44
BlockNumber nEntryPages
Definition: gin.h:46
int64 nEntries
Definition: gin.h:48
BlockNumber nTotalPages
Definition: gin.h:45
RelOptInfo * rel
Definition: selfuncs.c:3293
double ndistinct
Definition: selfuncs.c:3294
bool isdefault
Definition: selfuncs.c:3295
Node * var
Definition: selfuncs.c:3292
AttrNumber indexcol
Definition: pathnodes.h:1773
List * indexquals
Definition: pathnodes.h:1771
List * indexclauses
Definition: pathnodes.h:1723
List * indexorderbys
Definition: pathnodes.h:1724
IndexOptInfo * indexinfo
Definition: pathnodes.h:1722
IndexTuple xs_itup
Definition: relscan.h:159
struct TupleDescData * xs_itupdesc
Definition: relscan.h:160
Definition: pg_list.h:54
double ndistinct
Definition: statistics.h:28
AttrNumber * attributes
Definition: statistics.h:30
uint32 nitems
Definition: statistics.h:38
MVNDistinctItem items[FLEXIBLE_ARRAY_MEMBER]
Definition: statistics.h:39
Definition: nodes.h:129
NullTestType nulltesttype
Definition: primnodes.h:1959
Oid opno
Definition: primnodes.h:818
List * args
Definition: primnodes.h:836
List * cte_plan_ids
Definition: pathnodes.h:305
Query * parse
Definition: pathnodes.h:202
Cost per_tuple
Definition: pathnodes.h:48
Cost startup
Definition: pathnodes.h:47
List * returningList
Definition: parsenodes.h:200
Node * setOperations
Definition: parsenodes.h:221
List * cteList
Definition: parsenodes.h:168
List * groupClause
Definition: parsenodes.h:202
List * targetList
Definition: parsenodes.h:193
List * groupingSets
Definition: parsenodes.h:205
List * distinctClause
Definition: parsenodes.h:211
char * ctename
Definition: parsenodes.h:1196
Index ctelevelsup
Definition: parsenodes.h:1198
RTEKind rtekind
Definition: parsenodes.h:1047
Relids relids
Definition: pathnodes.h:871
Index relid
Definition: pathnodes.h:918
List * statlist
Definition: pathnodes.h:946
Cardinality tuples
Definition: pathnodes.h:949
BlockNumber pages
Definition: pathnodes.h:948
List * indexlist
Definition: pathnodes.h:944
Oid userid
Definition: pathnodes.h:966
PlannerInfo * subroot
Definition: pathnodes.h:953
Cardinality rows
Definition: pathnodes.h:877
RTEKind rtekind
Definition: pathnodes.h:922
Expr * clause
Definition: pathnodes.h:2575
Relids syn_lefthand
Definition: pathnodes.h:2907
Relids min_righthand
Definition: pathnodes.h:2906
JoinType jointype
Definition: pathnodes.h:2909
Relids syn_righthand
Definition: pathnodes.h:2908
Bitmapset * keys
Definition: pathnodes.h:1294
Expr * expr
Definition: primnodes.h:2190
Definition: date.h:28
TimeADT time
Definition: date.h:29
int32 zone
Definition: date.h:30
Definition: primnodes.h:248
AttrNumber varattno
Definition: primnodes.h:260
int varno
Definition: primnodes.h:255
Index varlevelsup
Definition: primnodes.h:280
HeapTuple statsTuple
Definition: selfuncs.h:89
int32 atttypmod
Definition: selfuncs.h:94
RelOptInfo * rel
Definition: selfuncs.h:88
void(* freefunc)(HeapTuple tuple)
Definition: selfuncs.h:91
Definition: type.h:96
Definition: c.h:695
Definition: c.h:641
#define TableOidAttributeNumber
Definition: sysattr.h:26
#define SelfItemPointerAttributeNumber
Definition: sysattr.h:21
void ReleaseSysCache(HeapTuple tuple)
Definition: syscache.c:269
HeapTuple SearchSysCache3(int cacheId, Datum key1, Datum key2, Datum key3)
Definition: syscache.c:243
void table_close(Relation relation, LOCKMODE lockmode)
Definition: table.c:126
Relation table_open(Oid relationId, LOCKMODE lockmode)
Definition: table.c:40
TupleTableSlot * table_slot_create(Relation relation, List **reglist)
Definition: tableam.c:91
static TupleTableSlot * ExecClearTuple(TupleTableSlot *slot)
Definition: tuptable.h:454
TypeCacheEntry * lookup_type_cache(Oid type_id, int flags)
Definition: typcache.c:386
#define TYPECACHE_EQ_OPR
Definition: typcache.h:137
static Interval * DatumGetIntervalP(Datum X)
Definition: timestamp.h:40
static Timestamp DatumGetTimestamp(Datum X)
Definition: timestamp.h:28
static TimestampTz DatumGetTimestampTz(Datum X)
Definition: timestamp.h:34
Relids pull_varnos(PlannerInfo *root, Node *node)
Definition: var.c:113
List * pull_var_clause(Node *node, int flags)
Definition: var.c:609
#define VARDATA_ANY(PTR)
Definition: varatt.h:324
#define VARSIZE_ANY_EXHDR(PTR)
Definition: varatt.h:317
#define VM_ALL_VISIBLE(r, b, v)
Definition: visibilitymap.h:24