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