Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * selfuncs.c
4 : * Selectivity functions and index cost estimation functions for
5 : * standard operators and index access methods.
6 : *
7 : * Selectivity routines are registered in the pg_operator catalog
8 : * in the "oprrest" and "oprjoin" attributes.
9 : *
10 : * Index cost functions are located via the index AM's API struct,
11 : * which is obtained from the handler function registered in pg_am.
12 : *
13 : * Portions Copyright (c) 1996-2019, 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. Thus, in most cases where we are looking at statistics, we
92 : * should ignore the operator collation and use the stats entry's collation.
93 : * We expect that the error induced by doing this is usually not large enough
94 : * to justify complicating matters. In any case, doing otherwise would yield
95 : * entirely garbage results for ordered stats data such as histograms.
96 : *----------
97 : */
98 :
99 : #include "postgres.h"
100 :
101 : #include <ctype.h>
102 : #include <math.h>
103 :
104 : #include "access/brin.h"
105 : #include "access/gin.h"
106 : #include "access/table.h"
107 : #include "access/tableam.h"
108 : #include "access/visibilitymap.h"
109 : #include "catalog/pg_am.h"
110 : #include "catalog/pg_collation.h"
111 : #include "catalog/pg_operator.h"
112 : #include "catalog/pg_statistic.h"
113 : #include "catalog/pg_statistic_ext.h"
114 : #include "executor/nodeAgg.h"
115 : #include "miscadmin.h"
116 : #include "nodes/makefuncs.h"
117 : #include "nodes/nodeFuncs.h"
118 : #include "optimizer/clauses.h"
119 : #include "optimizer/cost.h"
120 : #include "optimizer/optimizer.h"
121 : #include "optimizer/pathnode.h"
122 : #include "optimizer/paths.h"
123 : #include "optimizer/plancat.h"
124 : #include "parser/parse_clause.h"
125 : #include "parser/parsetree.h"
126 : #include "statistics/statistics.h"
127 : #include "storage/bufmgr.h"
128 : #include "utils/builtins.h"
129 : #include "utils/date.h"
130 : #include "utils/datum.h"
131 : #include "utils/fmgroids.h"
132 : #include "utils/index_selfuncs.h"
133 : #include "utils/lsyscache.h"
134 : #include "utils/memutils.h"
135 : #include "utils/pg_locale.h"
136 : #include "utils/rel.h"
137 : #include "utils/selfuncs.h"
138 : #include "utils/snapmgr.h"
139 : #include "utils/spccache.h"
140 : #include "utils/syscache.h"
141 : #include "utils/timestamp.h"
142 : #include "utils/typcache.h"
143 :
144 :
145 : /* Hooks for plugins to get control when we ask for stats */
146 : get_relation_stats_hook_type get_relation_stats_hook = NULL;
147 : get_index_stats_hook_type get_index_stats_hook = NULL;
148 :
149 : static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
150 : static double eqjoinsel_inner(Oid opfuncoid,
151 : VariableStatData *vardata1, VariableStatData *vardata2,
152 : double nd1, double nd2,
153 : bool isdefault1, bool isdefault2,
154 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
155 : Form_pg_statistic stats1, Form_pg_statistic stats2,
156 : bool have_mcvs1, bool have_mcvs2);
157 : static double eqjoinsel_semi(Oid opfuncoid,
158 : VariableStatData *vardata1, VariableStatData *vardata2,
159 : double nd1, double nd2,
160 : bool isdefault1, bool isdefault2,
161 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
162 : Form_pg_statistic stats1, Form_pg_statistic stats2,
163 : bool have_mcvs1, bool have_mcvs2,
164 : RelOptInfo *inner_rel);
165 : static bool estimate_multivariate_ndistinct(PlannerInfo *root,
166 : RelOptInfo *rel, List **varinfos, double *ndistinct);
167 : static bool convert_to_scalar(Datum value, Oid valuetypid, Oid collid,
168 : double *scaledvalue,
169 : Datum lobound, Datum hibound, Oid boundstypid,
170 : double *scaledlobound, double *scaledhibound);
171 : static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure);
172 : static void convert_string_to_scalar(char *value,
173 : double *scaledvalue,
174 : char *lobound,
175 : double *scaledlobound,
176 : char *hibound,
177 : double *scaledhibound);
178 : static void convert_bytea_to_scalar(Datum value,
179 : double *scaledvalue,
180 : Datum lobound,
181 : double *scaledlobound,
182 : Datum hibound,
183 : double *scaledhibound);
184 : static double convert_one_string_to_scalar(char *value,
185 : int rangelo, int rangehi);
186 : static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
187 : int rangelo, int rangehi);
188 : static char *convert_string_datum(Datum value, Oid typid, Oid collid,
189 : bool *failure);
190 : static double convert_timevalue_to_scalar(Datum value, Oid typid,
191 : bool *failure);
192 : static void examine_simple_variable(PlannerInfo *root, Var *var,
193 : VariableStatData *vardata);
194 : static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata,
195 : Oid sortop, Datum *min, Datum *max);
196 : static bool get_actual_variable_range(PlannerInfo *root,
197 : VariableStatData *vardata,
198 : Oid sortop,
199 : Datum *min, Datum *max);
200 : static bool get_actual_variable_endpoint(Relation heapRel,
201 : Relation indexRel,
202 : ScanDirection indexscandir,
203 : ScanKey scankeys,
204 : int16 typLen,
205 : bool typByVal,
206 : TupleTableSlot *tableslot,
207 : MemoryContext outercontext,
208 : Datum *endpointDatum);
209 : static RelOptInfo *find_join_input_rel(PlannerInfo *root, Relids relids);
210 :
211 :
212 : /*
213 : * eqsel - Selectivity of "=" for any data types.
214 : *
215 : * Note: this routine is also used to estimate selectivity for some
216 : * operators that are not "=" but have comparable selectivity behavior,
217 : * such as "~=" (geometric approximate-match). Even for "=", we must
218 : * keep in mind that the left and right datatypes may differ.
219 : */
220 : Datum
221 295798 : eqsel(PG_FUNCTION_ARGS)
222 : {
223 295798 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
224 : }
225 :
226 : /*
227 : * Common code for eqsel() and neqsel()
228 : */
229 : static double
230 310662 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
231 : {
232 310662 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
233 310662 : Oid operator = PG_GETARG_OID(1);
234 310662 : List *args = (List *) PG_GETARG_POINTER(2);
235 310662 : int varRelid = PG_GETARG_INT32(3);
236 : VariableStatData vardata;
237 : Node *other;
238 : bool varonleft;
239 : double selec;
240 :
241 : /*
242 : * When asked about <>, we do the estimation using the corresponding =
243 : * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
244 : */
245 310662 : if (negate)
246 : {
247 14864 : operator = get_negator(operator);
248 14864 : if (!OidIsValid(operator))
249 : {
250 : /* Use default selectivity (should we raise an error instead?) */
251 0 : return 1.0 - DEFAULT_EQ_SEL;
252 : }
253 : }
254 :
255 : /*
256 : * If expression is not variable = something or something = variable, then
257 : * punt and return a default estimate.
258 : */
259 310662 : if (!get_restriction_variable(root, args, varRelid,
260 : &vardata, &other, &varonleft))
261 1532 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
262 :
263 : /*
264 : * We can do a lot better if the something is a constant. (Note: the
265 : * Const might result from estimation rather than being a simple constant
266 : * in the query.)
267 : */
268 309130 : if (IsA(other, Const))
269 418866 : selec = var_eq_const(&vardata, operator,
270 139622 : ((Const *) other)->constvalue,
271 139622 : ((Const *) other)->constisnull,
272 : varonleft, negate);
273 : else
274 169508 : selec = var_eq_non_const(&vardata, operator, other,
275 : varonleft, negate);
276 :
277 309130 : ReleaseVariableStats(vardata);
278 :
279 309130 : return selec;
280 : }
281 :
282 : /*
283 : * var_eq_const --- eqsel for var = const case
284 : *
285 : * This is exported so that some other estimation functions can use it.
286 : */
287 : double
288 153558 : var_eq_const(VariableStatData *vardata, Oid operator,
289 : Datum constval, bool constisnull,
290 : bool varonleft, bool negate)
291 : {
292 : double selec;
293 153558 : double nullfrac = 0.0;
294 : bool isdefault;
295 : Oid opfuncoid;
296 :
297 : /*
298 : * If the constant is NULL, assume operator is strict and return zero, ie,
299 : * operator will never return TRUE. (It's zero even for a negator op.)
300 : */
301 153558 : if (constisnull)
302 24 : return 0.0;
303 :
304 : /*
305 : * Grab the nullfrac for use below. Note we allow use of nullfrac
306 : * regardless of security check.
307 : */
308 153534 : if (HeapTupleIsValid(vardata->statsTuple))
309 : {
310 : Form_pg_statistic stats;
311 :
312 105044 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
313 105044 : nullfrac = stats->stanullfrac;
314 : }
315 :
316 : /*
317 : * If we matched the var to a unique index or DISTINCT clause, assume
318 : * there is exactly one match regardless of anything else. (This is
319 : * slightly bogus, since the index or clause's equality operator might be
320 : * different from ours, but it's much more likely to be right than
321 : * ignoring the information.)
322 : */
323 153534 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
324 : {
325 30062 : selec = 1.0 / vardata->rel->tuples;
326 : }
327 208558 : else if (HeapTupleIsValid(vardata->statsTuple) &&
328 85086 : statistic_proc_security_check(vardata,
329 : (opfuncoid = get_opcode(operator))))
330 85086 : {
331 : AttStatsSlot sslot;
332 85086 : bool match = false;
333 : int i;
334 :
335 : /*
336 : * Is the constant "=" to any of the column's most common values?
337 : * (Although the given operator may not really be "=", we will assume
338 : * that seeing whether it returns TRUE is an appropriate test. If you
339 : * don't like this, maybe you shouldn't be using eqsel for your
340 : * operator...)
341 : */
342 85086 : if (get_attstatsslot(&sslot, vardata->statsTuple,
343 : STATISTIC_KIND_MCV, InvalidOid,
344 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
345 : {
346 : FmgrInfo eqproc;
347 :
348 76106 : fmgr_info(opfuncoid, &eqproc);
349 :
350 1460838 : for (i = 0; i < sslot.nvalues; i++)
351 : {
352 : /* be careful to apply operator right way 'round */
353 1420630 : if (varonleft)
354 1420630 : match = DatumGetBool(FunctionCall2Coll(&eqproc,
355 : sslot.stacoll,
356 : sslot.values[i],
357 : constval));
358 : else
359 0 : match = DatumGetBool(FunctionCall2Coll(&eqproc,
360 : sslot.stacoll,
361 : constval,
362 : sslot.values[i]));
363 1420630 : if (match)
364 35898 : break;
365 : }
366 : }
367 : else
368 : {
369 : /* no most-common-value info available */
370 8980 : i = 0; /* keep compiler quiet */
371 : }
372 :
373 85086 : if (match)
374 : {
375 : /*
376 : * Constant is "=" to this common value. We know selectivity
377 : * exactly (or as exactly as ANALYZE could calculate it, anyway).
378 : */
379 35898 : selec = sslot.numbers[i];
380 : }
381 : else
382 : {
383 : /*
384 : * Comparison is against a constant that is neither NULL nor any
385 : * of the common values. Its selectivity cannot be more than
386 : * this:
387 : */
388 49188 : double sumcommon = 0.0;
389 : double otherdistinct;
390 :
391 1335480 : for (i = 0; i < sslot.nnumbers; i++)
392 1286292 : sumcommon += sslot.numbers[i];
393 49188 : selec = 1.0 - sumcommon - nullfrac;
394 49188 : CLAMP_PROBABILITY(selec);
395 :
396 : /*
397 : * and in fact it's probably a good deal less. We approximate that
398 : * all the not-common values share this remaining fraction
399 : * equally, so we divide by the number of other distinct values.
400 : */
401 98376 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
402 49188 : sslot.nnumbers;
403 49188 : if (otherdistinct > 1)
404 28868 : selec /= otherdistinct;
405 :
406 : /*
407 : * Another cross-check: selectivity shouldn't be estimated as more
408 : * than the least common "most common value".
409 : */
410 49188 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
411 0 : selec = sslot.numbers[sslot.nnumbers - 1];
412 : }
413 :
414 85086 : free_attstatsslot(&sslot);
415 : }
416 : else
417 : {
418 : /*
419 : * No ANALYZE stats available, so make a guess using estimated number
420 : * of distinct values and assuming they are equally common. (The guess
421 : * is unlikely to be very good, but we do know a few special cases.)
422 : */
423 38386 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
424 : }
425 :
426 : /* now adjust if we wanted <> rather than = */
427 153534 : if (negate)
428 11300 : selec = 1.0 - selec - nullfrac;
429 :
430 : /* result should be in range, but make sure... */
431 153534 : CLAMP_PROBABILITY(selec);
432 :
433 153534 : return selec;
434 : }
435 :
436 : /*
437 : * var_eq_non_const --- eqsel for var = something-other-than-const case
438 : *
439 : * This is exported so that some other estimation functions can use it.
440 : */
441 : double
442 169508 : var_eq_non_const(VariableStatData *vardata, Oid operator,
443 : Node *other,
444 : bool varonleft, bool negate)
445 : {
446 : double selec;
447 169508 : double nullfrac = 0.0;
448 : bool isdefault;
449 :
450 : /*
451 : * Grab the nullfrac for use below.
452 : */
453 169508 : if (HeapTupleIsValid(vardata->statsTuple))
454 : {
455 : Form_pg_statistic stats;
456 :
457 128002 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
458 128002 : nullfrac = stats->stanullfrac;
459 : }
460 :
461 : /*
462 : * If we matched the var to a unique index or DISTINCT clause, assume
463 : * there is exactly one match regardless of anything else. (This is
464 : * slightly bogus, since the index or clause's equality operator might be
465 : * different from ours, but it's much more likely to be right than
466 : * ignoring the information.)
467 : */
468 169508 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
469 : {
470 83386 : selec = 1.0 / vardata->rel->tuples;
471 : }
472 86122 : else if (HeapTupleIsValid(vardata->statsTuple))
473 : {
474 : double ndistinct;
475 : AttStatsSlot sslot;
476 :
477 : /*
478 : * Search is for a value that we do not know a priori, but we will
479 : * assume it is not NULL. Estimate the selectivity as non-null
480 : * fraction divided by number of distinct values, so that we get a
481 : * result averaged over all possible values whether common or
482 : * uncommon. (Essentially, we are assuming that the not-yet-known
483 : * comparison value is equally likely to be any of the possible
484 : * values, regardless of their frequency in the table. Is that a good
485 : * idea?)
486 : */
487 59472 : selec = 1.0 - nullfrac;
488 59472 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
489 59472 : if (ndistinct > 1)
490 56918 : selec /= ndistinct;
491 :
492 : /*
493 : * Cross-check: selectivity should never be estimated as more than the
494 : * most common value's.
495 : */
496 59472 : if (get_attstatsslot(&sslot, vardata->statsTuple,
497 : STATISTIC_KIND_MCV, InvalidOid,
498 : ATTSTATSSLOT_NUMBERS))
499 : {
500 51670 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
501 140 : selec = sslot.numbers[0];
502 51670 : free_attstatsslot(&sslot);
503 : }
504 : }
505 : else
506 : {
507 : /*
508 : * No ANALYZE stats available, so make a guess using estimated number
509 : * of distinct values and assuming they are equally common. (The guess
510 : * is unlikely to be very good, but we do know a few special cases.)
511 : */
512 26650 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
513 : }
514 :
515 : /* now adjust if we wanted <> rather than = */
516 169508 : if (negate)
517 2628 : selec = 1.0 - selec - nullfrac;
518 :
519 : /* result should be in range, but make sure... */
520 169508 : CLAMP_PROBABILITY(selec);
521 :
522 169508 : return selec;
523 : }
524 :
525 : /*
526 : * neqsel - Selectivity of "!=" for any data types.
527 : *
528 : * This routine is also used for some operators that are not "!="
529 : * but have comparable selectivity behavior. See above comments
530 : * for eqsel().
531 : */
532 : Datum
533 14864 : neqsel(PG_FUNCTION_ARGS)
534 : {
535 14864 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
536 : }
537 :
538 : /*
539 : * scalarineqsel - Selectivity of "<", "<=", ">", ">=" for scalars.
540 : *
541 : * This is the guts of scalarltsel/scalarlesel/scalargtsel/scalargesel.
542 : * The isgt and iseq flags distinguish which of the four cases apply.
543 : *
544 : * The caller has commuted the clause, if necessary, so that we can treat
545 : * the variable as being on the left. The caller must also make sure that
546 : * the other side of the clause is a non-null Const, and dissect that into
547 : * a value and datatype. (This definition simplifies some callers that
548 : * want to estimate against a computed value instead of a Const node.)
549 : *
550 : * This routine works for any datatype (or pair of datatypes) known to
551 : * convert_to_scalar(). If it is applied to some other datatype,
552 : * it will return an approximate estimate based on assuming that the constant
553 : * value falls in the middle of the bin identified by binary search.
554 : */
555 : static double
556 161532 : scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
557 : VariableStatData *vardata, Datum constval, Oid consttype)
558 : {
559 : Form_pg_statistic stats;
560 : FmgrInfo opproc;
561 : double mcv_selec,
562 : hist_selec,
563 : sumcommon;
564 : double selec;
565 :
566 161532 : if (!HeapTupleIsValid(vardata->statsTuple))
567 : {
568 : /*
569 : * No stats are available. Typically this means we have to fall back
570 : * on the default estimate; but if the variable is CTID then we can
571 : * make an estimate based on comparing the constant to the table size.
572 : */
573 24460 : if (vardata->var && IsA(vardata->var, Var) &&
574 10862 : ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
575 : {
576 : ItemPointer itemptr;
577 : double block;
578 : double density;
579 :
580 : /*
581 : * If the relation's empty, we're going to include all of it.
582 : * (This is mostly to avoid divide-by-zero below.)
583 : */
584 4 : if (vardata->rel->pages == 0)
585 0 : return 1.0;
586 :
587 4 : itemptr = (ItemPointer) DatumGetPointer(constval);
588 4 : block = ItemPointerGetBlockNumberNoCheck(itemptr);
589 :
590 : /*
591 : * Determine the average number of tuples per page (density).
592 : *
593 : * Since the last page will, on average, be only half full, we can
594 : * estimate it to have half as many tuples as earlier pages. So
595 : * give it half the weight of a regular page.
596 : */
597 4 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
598 :
599 : /* If target is the last page, use half the density. */
600 4 : if (block >= vardata->rel->pages - 1)
601 0 : density *= 0.5;
602 :
603 : /*
604 : * Using the average tuples per page, calculate how far into the
605 : * page the itemptr is likely to be and adjust block accordingly,
606 : * by adding that fraction of a whole block (but never more than a
607 : * whole block, no matter how high the itemptr's offset is). Here
608 : * we are ignoring the possibility of dead-tuple line pointers,
609 : * which is fairly bogus, but we lack the info to do better.
610 : */
611 4 : if (density > 0.0)
612 : {
613 4 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
614 :
615 4 : block += Min(offset / density, 1.0);
616 : }
617 :
618 : /*
619 : * Convert relative block number to selectivity. Again, the last
620 : * page has only half weight.
621 : */
622 4 : selec = block / (vardata->rel->pages - 0.5);
623 :
624 : /*
625 : * The calculation so far gave us a selectivity for the "<=" case.
626 : * We'll have one less tuple for "<" and one additional tuple for
627 : * ">=", the latter of which we'll reverse the selectivity for
628 : * below, so we can simply subtract one tuple for both cases. The
629 : * cases that need this adjustment can be identified by iseq being
630 : * equal to isgt.
631 : */
632 4 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
633 0 : selec -= (1.0 / vardata->rel->tuples);
634 :
635 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
636 4 : if (isgt)
637 4 : selec = 1.0 - selec;
638 :
639 4 : CLAMP_PROBABILITY(selec);
640 4 : return selec;
641 : }
642 :
643 : /* no stats available, so default result */
644 13594 : return DEFAULT_INEQ_SEL;
645 : }
646 147934 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
647 :
648 147934 : fmgr_info(get_opcode(operator), &opproc);
649 :
650 : /*
651 : * If we have most-common-values info, add up the fractions of the MCV
652 : * entries that satisfy MCV OP CONST. These fractions contribute directly
653 : * to the result selectivity. Also add up the total fraction represented
654 : * by MCV entries.
655 : */
656 147934 : mcv_selec = mcv_selectivity(vardata, &opproc, constval, true,
657 : &sumcommon);
658 :
659 : /*
660 : * If there is a histogram, determine which bin the constant falls in, and
661 : * compute the resulting contribution to selectivity.
662 : */
663 147934 : hist_selec = ineq_histogram_selectivity(root, vardata,
664 : &opproc, isgt, iseq,
665 : constval, consttype);
666 :
667 : /*
668 : * Now merge the results from the MCV and histogram calculations,
669 : * realizing that the histogram covers only the non-null values that are
670 : * not listed in MCV.
671 : */
672 147934 : selec = 1.0 - stats->stanullfrac - sumcommon;
673 :
674 147934 : if (hist_selec >= 0.0)
675 126942 : selec *= hist_selec;
676 : else
677 : {
678 : /*
679 : * If no histogram but there are values not accounted for by MCV,
680 : * arbitrarily assume half of them will match.
681 : */
682 20992 : selec *= 0.5;
683 : }
684 :
685 147934 : selec += mcv_selec;
686 :
687 : /* result should be in range, but make sure... */
688 147934 : CLAMP_PROBABILITY(selec);
689 :
690 147934 : return selec;
691 : }
692 :
693 : /*
694 : * mcv_selectivity - Examine the MCV list for selectivity estimates
695 : *
696 : * Determine the fraction of the variable's MCV population that satisfies
697 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft. Also
698 : * compute the fraction of the total column population represented by the MCV
699 : * list. This code will work for any boolean-returning predicate operator.
700 : *
701 : * The function result is the MCV selectivity, and the fraction of the
702 : * total population is returned into *sumcommonp. Zeroes are returned
703 : * if there is no MCV list.
704 : */
705 : double
706 150280 : mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc,
707 : Datum constval, bool varonleft,
708 : double *sumcommonp)
709 : {
710 : double mcv_selec,
711 : sumcommon;
712 : AttStatsSlot sslot;
713 : int i;
714 :
715 150280 : mcv_selec = 0.0;
716 150280 : sumcommon = 0.0;
717 :
718 299316 : if (HeapTupleIsValid(vardata->statsTuple) &&
719 298020 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
720 148984 : get_attstatsslot(&sslot, vardata->statsTuple,
721 : STATISTIC_KIND_MCV, InvalidOid,
722 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
723 : {
724 2603046 : for (i = 0; i < sslot.nvalues; i++)
725 : {
726 5054016 : if (varonleft ?
727 2527008 : DatumGetBool(FunctionCall2Coll(opproc,
728 : sslot.stacoll,
729 : sslot.values[i],
730 : constval)) :
731 0 : DatumGetBool(FunctionCall2Coll(opproc,
732 : sslot.stacoll,
733 : constval,
734 : sslot.values[i])))
735 1333836 : mcv_selec += sslot.numbers[i];
736 2527008 : sumcommon += sslot.numbers[i];
737 : }
738 76038 : free_attstatsslot(&sslot);
739 : }
740 :
741 150280 : *sumcommonp = sumcommon;
742 150280 : return mcv_selec;
743 : }
744 :
745 : /*
746 : * histogram_selectivity - Examine the histogram for selectivity estimates
747 : *
748 : * Determine the fraction of the variable's histogram entries that satisfy
749 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft.
750 : *
751 : * This code will work for any boolean-returning predicate operator, whether
752 : * or not it has anything to do with the histogram sort operator. We are
753 : * essentially using the histogram just as a representative sample. However,
754 : * small histograms are unlikely to be all that representative, so the caller
755 : * should be prepared to fall back on some other estimation approach when the
756 : * histogram is missing or very small. It may also be prudent to combine this
757 : * approach with another one when the histogram is small.
758 : *
759 : * If the actual histogram size is not at least min_hist_size, we won't bother
760 : * to do the calculation at all. Also, if the n_skip parameter is > 0, we
761 : * ignore the first and last n_skip histogram elements, on the grounds that
762 : * they are outliers and hence not very representative. Typical values for
763 : * these parameters are 10 and 1.
764 : *
765 : * The function result is the selectivity, or -1 if there is no histogram
766 : * or it's smaller than min_hist_size.
767 : *
768 : * The output parameter *hist_size receives the actual histogram size,
769 : * or zero if no histogram. Callers may use this number to decide how
770 : * much faith to put in the function result.
771 : *
772 : * Note that the result disregards both the most-common-values (if any) and
773 : * null entries. The caller is expected to combine this result with
774 : * statistics for those portions of the column population. It may also be
775 : * prudent to clamp the result range, ie, disbelieve exact 0 or 1 outputs.
776 : */
777 : double
778 2346 : histogram_selectivity(VariableStatData *vardata, FmgrInfo *opproc,
779 : Datum constval, bool varonleft,
780 : int min_hist_size, int n_skip,
781 : int *hist_size)
782 : {
783 : double result;
784 : AttStatsSlot sslot;
785 :
786 : /* check sanity of parameters */
787 : Assert(n_skip >= 0);
788 : Assert(min_hist_size > 2 * n_skip);
789 :
790 3448 : if (HeapTupleIsValid(vardata->statsTuple) &&
791 2204 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
792 1102 : get_attstatsslot(&sslot, vardata->statsTuple,
793 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
794 : ATTSTATSSLOT_VALUES))
795 : {
796 1098 : *hist_size = sslot.nvalues;
797 1098 : if (sslot.nvalues >= min_hist_size)
798 : {
799 742 : int nmatch = 0;
800 : int i;
801 :
802 65388 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
803 : {
804 129292 : if (varonleft ?
805 64646 : DatumGetBool(FunctionCall2Coll(opproc,
806 : sslot.stacoll,
807 : sslot.values[i],
808 : constval)) :
809 0 : DatumGetBool(FunctionCall2Coll(opproc,
810 : sslot.stacoll,
811 : constval,
812 : sslot.values[i])))
813 2772 : nmatch++;
814 : }
815 742 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
816 : }
817 : else
818 356 : result = -1;
819 1098 : free_attstatsslot(&sslot);
820 : }
821 : else
822 : {
823 1248 : *hist_size = 0;
824 1248 : result = -1;
825 : }
826 :
827 2346 : return result;
828 : }
829 :
830 : /*
831 : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
832 : *
833 : * Determine the fraction of the variable's histogram population that
834 : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
835 : * The isgt and iseq flags distinguish which of the four cases apply.
836 : *
837 : * Returns -1 if there is no histogram (valid results will always be >= 0).
838 : *
839 : * Note that the result disregards both the most-common-values (if any) and
840 : * null entries. The caller is expected to combine this result with
841 : * statistics for those portions of the column population.
842 : *
843 : * This is exported so that some other estimation functions can use it.
844 : */
845 : double
846 149700 : ineq_histogram_selectivity(PlannerInfo *root,
847 : VariableStatData *vardata,
848 : FmgrInfo *opproc, bool isgt, bool iseq,
849 : Datum constval, Oid consttype)
850 : {
851 : double hist_selec;
852 : AttStatsSlot sslot;
853 :
854 149700 : hist_selec = -1.0;
855 :
856 : /*
857 : * Someday, ANALYZE might store more than one histogram per rel/att,
858 : * corresponding to more than one possible sort ordering defined for the
859 : * column type. However, to make that work we will need to figure out
860 : * which staop to search for --- it's not necessarily the one we have at
861 : * hand! (For example, we might have a '<=' operator rather than the '<'
862 : * operator that will appear in staop.) For now, assume that whatever
863 : * appears in pg_statistic is sorted the same way our operator sorts, or
864 : * the reverse way if isgt is true.
865 : */
866 298588 : if (HeapTupleIsValid(vardata->statsTuple) &&
867 297724 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
868 148836 : get_attstatsslot(&sslot, vardata->statsTuple,
869 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
870 : ATTSTATSSLOT_VALUES))
871 : {
872 127894 : if (sslot.nvalues > 1)
873 : {
874 : /*
875 : * Use binary search to find the desired location, namely the
876 : * right end of the histogram bin containing the comparison value,
877 : * which is the leftmost entry for which the comparison operator
878 : * succeeds (if isgt) or fails (if !isgt). (If the given operator
879 : * isn't actually sort-compatible with the histogram, you'll get
880 : * garbage results ... but probably not any more garbage-y than
881 : * you would have from the old linear search.)
882 : *
883 : * In this loop, we pay no attention to whether the operator iseq
884 : * or not; that detail will be mopped up below. (We cannot tell,
885 : * anyway, whether the operator thinks the values are equal.)
886 : *
887 : * If the binary search accesses the first or last histogram
888 : * entry, we try to replace that endpoint with the true column min
889 : * or max as found by get_actual_variable_range(). This
890 : * ameliorates misestimates when the min or max is moving as a
891 : * result of changes since the last ANALYZE. Note that this could
892 : * result in effectively including MCVs into the histogram that
893 : * weren't there before, but we don't try to correct for that.
894 : */
895 : double histfrac;
896 127894 : int lobound = 0; /* first possible slot to search */
897 127894 : int hibound = sslot.nvalues; /* last+1 slot to search */
898 127894 : bool have_end = false;
899 :
900 : /*
901 : * If there are only two histogram entries, we'll want up-to-date
902 : * values for both. (If there are more than two, we need at most
903 : * one of them to be updated, so we deal with that within the
904 : * loop.)
905 : */
906 127894 : if (sslot.nvalues == 2)
907 540 : have_end = get_actual_variable_range(root,
908 : vardata,
909 : sslot.staop,
910 : &sslot.values[0],
911 540 : &sslot.values[1]);
912 :
913 1004944 : while (lobound < hibound)
914 : {
915 749156 : int probe = (lobound + hibound) / 2;
916 : bool ltcmp;
917 :
918 : /*
919 : * If we find ourselves about to compare to the first or last
920 : * histogram entry, first try to replace it with the actual
921 : * current min or max (unless we already did so above).
922 : */
923 749156 : if (probe == 0 && sslot.nvalues > 2)
924 65220 : have_end = get_actual_variable_range(root,
925 : vardata,
926 : sslot.staop,
927 : &sslot.values[0],
928 : NULL);
929 683936 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
930 43750 : have_end = get_actual_variable_range(root,
931 : vardata,
932 : sslot.staop,
933 : NULL,
934 43750 : &sslot.values[probe]);
935 :
936 749156 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
937 : sslot.stacoll,
938 : sslot.values[probe],
939 : constval));
940 749156 : if (isgt)
941 56138 : ltcmp = !ltcmp;
942 749156 : if (ltcmp)
943 281674 : lobound = probe + 1;
944 : else
945 467482 : hibound = probe;
946 : }
947 :
948 127894 : if (lobound <= 0)
949 : {
950 : /*
951 : * Constant is below lower histogram boundary. More
952 : * precisely, we have found that no entry in the histogram
953 : * satisfies the inequality clause (if !isgt) or they all do
954 : * (if isgt). We estimate that that's true of the entire
955 : * table, so set histfrac to 0.0 (which we'll flip to 1.0
956 : * below, if isgt).
957 : */
958 61188 : histfrac = 0.0;
959 : }
960 66706 : else if (lobound >= sslot.nvalues)
961 : {
962 : /*
963 : * Inverse case: constant is above upper histogram boundary.
964 : */
965 21770 : histfrac = 1.0;
966 : }
967 : else
968 : {
969 : /* We have values[i-1] <= constant <= values[i]. */
970 44936 : int i = lobound;
971 44936 : double eq_selec = 0;
972 : double val,
973 : high,
974 : low;
975 : double binfrac;
976 :
977 : /*
978 : * In the cases where we'll need it below, obtain an estimate
979 : * of the selectivity of "x = constval". We use a calculation
980 : * similar to what var_eq_const() does for a non-MCV constant,
981 : * ie, estimate that all distinct non-MCV values occur equally
982 : * often. But multiplication by "1.0 - sumcommon - nullfrac"
983 : * will be done by our caller, so we shouldn't do that here.
984 : * Therefore we can't try to clamp the estimate by reference
985 : * to the least common MCV; the result would be too small.
986 : *
987 : * Note: since this is effectively assuming that constval
988 : * isn't an MCV, it's logically dubious if constval in fact is
989 : * one. But we have to apply *some* correction for equality,
990 : * and anyway we cannot tell if constval is an MCV, since we
991 : * don't have a suitable equality operator at hand.
992 : */
993 44936 : if (i == 1 || isgt == iseq)
994 : {
995 : double otherdistinct;
996 : bool isdefault;
997 : AttStatsSlot mcvslot;
998 :
999 : /* Get estimated number of distinct values */
1000 10650 : otherdistinct = get_variable_numdistinct(vardata,
1001 : &isdefault);
1002 :
1003 : /* Subtract off the number of known MCVs */
1004 10650 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1005 : STATISTIC_KIND_MCV, InvalidOid,
1006 : ATTSTATSSLOT_NUMBERS))
1007 : {
1008 2208 : otherdistinct -= mcvslot.nnumbers;
1009 2208 : free_attstatsslot(&mcvslot);
1010 : }
1011 :
1012 : /* If result doesn't seem sane, leave eq_selec at 0 */
1013 10650 : if (otherdistinct > 1)
1014 10650 : eq_selec = 1.0 / otherdistinct;
1015 : }
1016 :
1017 : /*
1018 : * Convert the constant and the two nearest bin boundary
1019 : * values to a uniform comparison scale, and do a linear
1020 : * interpolation within this bin.
1021 : */
1022 134808 : if (convert_to_scalar(constval, consttype, sslot.stacoll,
1023 : &val,
1024 89872 : sslot.values[i - 1], sslot.values[i],
1025 : vardata->vartype,
1026 : &low, &high))
1027 : {
1028 44932 : if (high <= low)
1029 : {
1030 : /* cope if bin boundaries appear identical */
1031 0 : binfrac = 0.5;
1032 : }
1033 44932 : else if (val <= low)
1034 5716 : binfrac = 0.0;
1035 39216 : else if (val >= high)
1036 982 : binfrac = 1.0;
1037 : else
1038 : {
1039 38234 : binfrac = (val - low) / (high - low);
1040 :
1041 : /*
1042 : * Watch out for the possibility that we got a NaN or
1043 : * Infinity from the division. This can happen
1044 : * despite the previous checks, if for example "low"
1045 : * is -Infinity.
1046 : */
1047 38234 : if (isnan(binfrac) ||
1048 38234 : binfrac < 0.0 || binfrac > 1.0)
1049 0 : binfrac = 0.5;
1050 : }
1051 : }
1052 : else
1053 : {
1054 : /*
1055 : * Ideally we'd produce an error here, on the grounds that
1056 : * the given operator shouldn't have scalarXXsel
1057 : * registered as its selectivity func unless we can deal
1058 : * with its operand types. But currently, all manner of
1059 : * stuff is invoking scalarXXsel, so give a default
1060 : * estimate until that can be fixed.
1061 : */
1062 4 : binfrac = 0.5;
1063 : }
1064 :
1065 : /*
1066 : * Now, compute the overall selectivity across the values
1067 : * represented by the histogram. We have i-1 full bins and
1068 : * binfrac partial bin below the constant.
1069 : */
1070 44936 : histfrac = (double) (i - 1) + binfrac;
1071 44936 : histfrac /= (double) (sslot.nvalues - 1);
1072 :
1073 : /*
1074 : * At this point, histfrac is an estimate of the fraction of
1075 : * the population represented by the histogram that satisfies
1076 : * "x <= constval". Somewhat remarkably, this statement is
1077 : * true regardless of which operator we were doing the probes
1078 : * with, so long as convert_to_scalar() delivers reasonable
1079 : * results. If the probe constant is equal to some histogram
1080 : * entry, we would have considered the bin to the left of that
1081 : * entry if probing with "<" or ">=", or the bin to the right
1082 : * if probing with "<=" or ">"; but binfrac would have come
1083 : * out as 1.0 in the first case and 0.0 in the second, leading
1084 : * to the same histfrac in either case. For probe constants
1085 : * between histogram entries, we find the same bin and get the
1086 : * same estimate with any operator.
1087 : *
1088 : * The fact that the estimate corresponds to "x <= constval"
1089 : * and not "x < constval" is because of the way that ANALYZE
1090 : * constructs the histogram: each entry is, effectively, the
1091 : * rightmost value in its sample bucket. So selectivity
1092 : * values that are exact multiples of 1/(histogram_size-1)
1093 : * should be understood as estimates including a histogram
1094 : * entry plus everything to its left.
1095 : *
1096 : * However, that breaks down for the first histogram entry,
1097 : * which necessarily is the leftmost value in its sample
1098 : * bucket. That means the first histogram bin is slightly
1099 : * narrower than the rest, by an amount equal to eq_selec.
1100 : * Another way to say that is that we want "x <= leftmost" to
1101 : * be estimated as eq_selec not zero. So, if we're dealing
1102 : * with the first bin (i==1), rescale to make that true while
1103 : * adjusting the rest of that bin linearly.
1104 : */
1105 44936 : if (i == 1)
1106 4370 : histfrac += eq_selec * (1.0 - binfrac);
1107 :
1108 : /*
1109 : * "x <= constval" is good if we want an estimate for "<=" or
1110 : * ">", but if we are estimating for "<" or ">=", we now need
1111 : * to decrease the estimate by eq_selec.
1112 : */
1113 44936 : if (isgt == iseq)
1114 9760 : histfrac -= eq_selec;
1115 : }
1116 :
1117 : /*
1118 : * Now the estimate is finished for "<" and "<=" cases. If we are
1119 : * estimating for ">" or ">=", flip it.
1120 : */
1121 127894 : hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1122 :
1123 : /*
1124 : * The histogram boundaries are only approximate to begin with,
1125 : * and may well be out of date anyway. Therefore, don't believe
1126 : * extremely small or large selectivity estimates --- unless we
1127 : * got actual current endpoint values from the table, in which
1128 : * case just do the usual sanity clamp. Somewhat arbitrarily, we
1129 : * set the cutoff for other cases at a hundredth of the histogram
1130 : * resolution.
1131 : */
1132 127894 : if (have_end)
1133 65532 : CLAMP_PROBABILITY(hist_selec);
1134 : else
1135 : {
1136 62362 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1137 :
1138 62362 : if (hist_selec < cutoff)
1139 19956 : hist_selec = cutoff;
1140 42406 : else if (hist_selec > 1.0 - cutoff)
1141 21362 : hist_selec = 1.0 - cutoff;
1142 : }
1143 : }
1144 :
1145 127894 : free_attstatsslot(&sslot);
1146 : }
1147 :
1148 149700 : return hist_selec;
1149 : }
1150 :
1151 : /*
1152 : * Common wrapper function for the selectivity estimators that simply
1153 : * invoke scalarineqsel().
1154 : */
1155 : static Datum
1156 29628 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1157 : {
1158 29628 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1159 29628 : Oid operator = PG_GETARG_OID(1);
1160 29628 : List *args = (List *) PG_GETARG_POINTER(2);
1161 29628 : int varRelid = PG_GETARG_INT32(3);
1162 : VariableStatData vardata;
1163 : Node *other;
1164 : bool varonleft;
1165 : Datum constval;
1166 : Oid consttype;
1167 : double selec;
1168 :
1169 : /*
1170 : * If expression is not variable op something or something op variable,
1171 : * then punt and return a default estimate.
1172 : */
1173 29628 : if (!get_restriction_variable(root, args, varRelid,
1174 : &vardata, &other, &varonleft))
1175 176 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1176 :
1177 : /*
1178 : * Can't do anything useful if the something is not a constant, either.
1179 : */
1180 29452 : if (!IsA(other, Const))
1181 : {
1182 1760 : ReleaseVariableStats(vardata);
1183 1760 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1184 : }
1185 :
1186 : /*
1187 : * If the constant is NULL, assume operator is strict and return zero, ie,
1188 : * operator will never return TRUE.
1189 : */
1190 27692 : if (((Const *) other)->constisnull)
1191 : {
1192 0 : ReleaseVariableStats(vardata);
1193 0 : PG_RETURN_FLOAT8(0.0);
1194 : }
1195 27692 : constval = ((Const *) other)->constvalue;
1196 27692 : consttype = ((Const *) other)->consttype;
1197 :
1198 : /*
1199 : * Force the var to be on the left to simplify logic in scalarineqsel.
1200 : */
1201 27692 : if (!varonleft)
1202 : {
1203 92 : operator = get_commutator(operator);
1204 92 : if (!operator)
1205 : {
1206 : /* Use default selectivity (should we raise an error instead?) */
1207 0 : ReleaseVariableStats(vardata);
1208 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1209 : }
1210 92 : isgt = !isgt;
1211 : }
1212 :
1213 : /* The rest of the work is done by scalarineqsel(). */
1214 27692 : selec = scalarineqsel(root, operator, isgt, iseq,
1215 : &vardata, constval, consttype);
1216 :
1217 27692 : ReleaseVariableStats(vardata);
1218 :
1219 27692 : PG_RETURN_FLOAT8((float8) selec);
1220 : }
1221 :
1222 : /*
1223 : * scalarltsel - Selectivity of "<" for scalars.
1224 : */
1225 : Datum
1226 7802 : scalarltsel(PG_FUNCTION_ARGS)
1227 : {
1228 7802 : return scalarineqsel_wrapper(fcinfo, false, false);
1229 : }
1230 :
1231 : /*
1232 : * scalarlesel - Selectivity of "<=" for scalars.
1233 : */
1234 : Datum
1235 3144 : scalarlesel(PG_FUNCTION_ARGS)
1236 : {
1237 3144 : return scalarineqsel_wrapper(fcinfo, false, true);
1238 : }
1239 :
1240 : /*
1241 : * scalargtsel - Selectivity of ">" for scalars.
1242 : */
1243 : Datum
1244 11746 : scalargtsel(PG_FUNCTION_ARGS)
1245 : {
1246 11746 : return scalarineqsel_wrapper(fcinfo, true, false);
1247 : }
1248 :
1249 : /*
1250 : * scalargesel - Selectivity of ">=" for scalars.
1251 : */
1252 : Datum
1253 6936 : scalargesel(PG_FUNCTION_ARGS)
1254 : {
1255 6936 : return scalarineqsel_wrapper(fcinfo, true, true);
1256 : }
1257 :
1258 : /*
1259 : * boolvarsel - Selectivity of Boolean variable.
1260 : *
1261 : * This can actually be called on any boolean-valued expression. If it
1262 : * involves only Vars of the specified relation, and if there are statistics
1263 : * about the Var or expression (the latter is possible if it's indexed) then
1264 : * we'll produce a real estimate; otherwise it's just a default.
1265 : */
1266 : Selectivity
1267 14140 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1268 : {
1269 : VariableStatData vardata;
1270 : double selec;
1271 :
1272 14140 : examine_variable(root, arg, varRelid, &vardata);
1273 14140 : if (HeapTupleIsValid(vardata.statsTuple))
1274 : {
1275 : /*
1276 : * A boolean variable V is equivalent to the clause V = 't', so we
1277 : * compute the selectivity as if that is what we have.
1278 : */
1279 11384 : selec = var_eq_const(&vardata, BooleanEqualOperator,
1280 : BoolGetDatum(true), false, true, false);
1281 : }
1282 : else
1283 : {
1284 : /* Otherwise, the default estimate is 0.5 */
1285 2756 : selec = 0.5;
1286 : }
1287 14140 : ReleaseVariableStats(vardata);
1288 14140 : return selec;
1289 : }
1290 :
1291 : /*
1292 : * booltestsel - Selectivity of BooleanTest Node.
1293 : */
1294 : Selectivity
1295 132 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1296 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1297 : {
1298 : VariableStatData vardata;
1299 : double selec;
1300 :
1301 132 : examine_variable(root, arg, varRelid, &vardata);
1302 :
1303 132 : if (HeapTupleIsValid(vardata.statsTuple))
1304 : {
1305 : Form_pg_statistic stats;
1306 : double freq_null;
1307 : AttStatsSlot sslot;
1308 :
1309 0 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1310 0 : freq_null = stats->stanullfrac;
1311 :
1312 0 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1313 : STATISTIC_KIND_MCV, InvalidOid,
1314 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1315 0 : && sslot.nnumbers > 0)
1316 0 : {
1317 : double freq_true;
1318 : double freq_false;
1319 :
1320 : /*
1321 : * Get first MCV frequency and derive frequency for true.
1322 : */
1323 0 : if (DatumGetBool(sslot.values[0]))
1324 0 : freq_true = sslot.numbers[0];
1325 : else
1326 0 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1327 :
1328 : /*
1329 : * Next derive frequency for false. Then use these as appropriate
1330 : * to derive frequency for each case.
1331 : */
1332 0 : freq_false = 1.0 - freq_true - freq_null;
1333 :
1334 0 : switch (booltesttype)
1335 : {
1336 : case IS_UNKNOWN:
1337 : /* select only NULL values */
1338 0 : selec = freq_null;
1339 0 : break;
1340 : case IS_NOT_UNKNOWN:
1341 : /* select non-NULL values */
1342 0 : selec = 1.0 - freq_null;
1343 0 : break;
1344 : case IS_TRUE:
1345 : /* select only TRUE values */
1346 0 : selec = freq_true;
1347 0 : break;
1348 : case IS_NOT_TRUE:
1349 : /* select non-TRUE values */
1350 0 : selec = 1.0 - freq_true;
1351 0 : break;
1352 : case IS_FALSE:
1353 : /* select only FALSE values */
1354 0 : selec = freq_false;
1355 0 : break;
1356 : case IS_NOT_FALSE:
1357 : /* select non-FALSE values */
1358 0 : selec = 1.0 - freq_false;
1359 0 : break;
1360 : default:
1361 0 : elog(ERROR, "unrecognized booltesttype: %d",
1362 : (int) booltesttype);
1363 : selec = 0.0; /* Keep compiler quiet */
1364 : break;
1365 : }
1366 :
1367 0 : free_attstatsslot(&sslot);
1368 : }
1369 : else
1370 : {
1371 : /*
1372 : * No most-common-value info available. Still have null fraction
1373 : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1374 : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1375 : */
1376 0 : switch (booltesttype)
1377 : {
1378 : case IS_UNKNOWN:
1379 : /* select only NULL values */
1380 0 : selec = freq_null;
1381 0 : break;
1382 : case IS_NOT_UNKNOWN:
1383 : /* select non-NULL values */
1384 0 : selec = 1.0 - freq_null;
1385 0 : break;
1386 : case IS_TRUE:
1387 : case IS_FALSE:
1388 : /* Assume we select half of the non-NULL values */
1389 0 : selec = (1.0 - freq_null) / 2.0;
1390 0 : break;
1391 : case IS_NOT_TRUE:
1392 : case IS_NOT_FALSE:
1393 : /* Assume we select NULLs plus half of the non-NULLs */
1394 : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1395 0 : selec = (freq_null + 1.0) / 2.0;
1396 0 : break;
1397 : default:
1398 0 : elog(ERROR, "unrecognized booltesttype: %d",
1399 : (int) booltesttype);
1400 : selec = 0.0; /* Keep compiler quiet */
1401 : break;
1402 : }
1403 : }
1404 : }
1405 : else
1406 : {
1407 : /*
1408 : * If we can't get variable statistics for the argument, perhaps
1409 : * clause_selectivity can do something with it. We ignore the
1410 : * possibility of a NULL value when using clause_selectivity, and just
1411 : * assume the value is either TRUE or FALSE.
1412 : */
1413 132 : switch (booltesttype)
1414 : {
1415 : case IS_UNKNOWN:
1416 12 : selec = DEFAULT_UNK_SEL;
1417 12 : break;
1418 : case IS_NOT_UNKNOWN:
1419 12 : selec = DEFAULT_NOT_UNK_SEL;
1420 12 : break;
1421 : case IS_TRUE:
1422 : case IS_NOT_FALSE:
1423 32 : selec = (double) clause_selectivity(root, arg,
1424 : varRelid,
1425 : jointype, sjinfo);
1426 32 : break;
1427 : case IS_FALSE:
1428 : case IS_NOT_TRUE:
1429 76 : selec = 1.0 - (double) clause_selectivity(root, arg,
1430 : varRelid,
1431 : jointype, sjinfo);
1432 76 : break;
1433 : default:
1434 0 : elog(ERROR, "unrecognized booltesttype: %d",
1435 : (int) booltesttype);
1436 : selec = 0.0; /* Keep compiler quiet */
1437 : break;
1438 : }
1439 : }
1440 :
1441 132 : ReleaseVariableStats(vardata);
1442 :
1443 : /* result should be in range, but make sure... */
1444 132 : CLAMP_PROBABILITY(selec);
1445 :
1446 132 : return (Selectivity) selec;
1447 : }
1448 :
1449 : /*
1450 : * nulltestsel - Selectivity of NullTest Node.
1451 : */
1452 : Selectivity
1453 12258 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1454 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1455 : {
1456 : VariableStatData vardata;
1457 : double selec;
1458 :
1459 12258 : examine_variable(root, arg, varRelid, &vardata);
1460 :
1461 12258 : if (HeapTupleIsValid(vardata.statsTuple))
1462 : {
1463 : Form_pg_statistic stats;
1464 : double freq_null;
1465 :
1466 3730 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1467 3730 : freq_null = stats->stanullfrac;
1468 :
1469 3730 : switch (nulltesttype)
1470 : {
1471 : case IS_NULL:
1472 :
1473 : /*
1474 : * Use freq_null directly.
1475 : */
1476 3246 : selec = freq_null;
1477 3246 : break;
1478 : case IS_NOT_NULL:
1479 :
1480 : /*
1481 : * Select not unknown (not null) values. Calculate from
1482 : * freq_null.
1483 : */
1484 484 : selec = 1.0 - freq_null;
1485 484 : break;
1486 : default:
1487 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1488 : (int) nulltesttype);
1489 : return (Selectivity) 0; /* keep compiler quiet */
1490 : }
1491 : }
1492 14448 : else if (vardata.var && IsA(vardata.var, Var) &&
1493 5920 : ((Var *) vardata.var)->varattno < 0)
1494 : {
1495 : /*
1496 : * There are no stats for system columns, but we know they are never
1497 : * NULL.
1498 : */
1499 0 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1500 : }
1501 : else
1502 : {
1503 : /*
1504 : * No ANALYZE stats available, so make a guess
1505 : */
1506 8528 : switch (nulltesttype)
1507 : {
1508 : case IS_NULL:
1509 972 : selec = DEFAULT_UNK_SEL;
1510 972 : break;
1511 : case IS_NOT_NULL:
1512 7556 : selec = DEFAULT_NOT_UNK_SEL;
1513 7556 : break;
1514 : default:
1515 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1516 : (int) nulltesttype);
1517 : return (Selectivity) 0; /* keep compiler quiet */
1518 : }
1519 : }
1520 :
1521 12258 : ReleaseVariableStats(vardata);
1522 :
1523 : /* result should be in range, but make sure... */
1524 12258 : CLAMP_PROBABILITY(selec);
1525 :
1526 12258 : return (Selectivity) selec;
1527 : }
1528 :
1529 : /*
1530 : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1531 : *
1532 : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1533 : * but it seems possible that RelabelType might show up. Also, the planner
1534 : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1535 : * so we need to be ready to deal with more than one level.
1536 : */
1537 : static Node *
1538 62934 : strip_array_coercion(Node *node)
1539 : {
1540 : for (;;)
1541 : {
1542 62946 : if (node && IsA(node, ArrayCoerceExpr))
1543 12 : {
1544 220 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1545 :
1546 : /*
1547 : * If the per-element expression is just a RelabelType on top of
1548 : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1549 : */
1550 232 : if (IsA(acoerce->elemexpr, RelabelType) &&
1551 12 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1552 12 : node = (Node *) acoerce->arg;
1553 : else
1554 : break;
1555 : }
1556 62714 : else if (node && IsA(node, RelabelType))
1557 : {
1558 : /* We don't really expect this case, but may as well cope */
1559 0 : node = (Node *) ((RelabelType *) node)->arg;
1560 : }
1561 : else
1562 : break;
1563 : }
1564 62922 : return node;
1565 : }
1566 :
1567 : /*
1568 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1569 : */
1570 : Selectivity
1571 13150 : scalararraysel(PlannerInfo *root,
1572 : ScalarArrayOpExpr *clause,
1573 : bool is_join_clause,
1574 : int varRelid,
1575 : JoinType jointype,
1576 : SpecialJoinInfo *sjinfo)
1577 : {
1578 13150 : Oid operator = clause->opno;
1579 13150 : bool useOr = clause->useOr;
1580 13150 : bool isEquality = false;
1581 13150 : bool isInequality = false;
1582 : Node *leftop;
1583 : Node *rightop;
1584 : Oid nominal_element_type;
1585 : Oid nominal_element_collation;
1586 : TypeCacheEntry *typentry;
1587 : RegProcedure oprsel;
1588 : FmgrInfo oprselproc;
1589 : Selectivity s1;
1590 : Selectivity s1disjoint;
1591 :
1592 : /* First, deconstruct the expression */
1593 : Assert(list_length(clause->args) == 2);
1594 13150 : leftop = (Node *) linitial(clause->args);
1595 13150 : rightop = (Node *) lsecond(clause->args);
1596 :
1597 : /* aggressively reduce both sides to constants */
1598 13150 : leftop = estimate_expression_value(root, leftop);
1599 13150 : rightop = estimate_expression_value(root, rightop);
1600 :
1601 : /* get nominal (after relabeling) element type of rightop */
1602 13150 : nominal_element_type = get_base_element_type(exprType(rightop));
1603 13150 : if (!OidIsValid(nominal_element_type))
1604 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1605 : /* get nominal collation, too, for generating constants */
1606 13150 : nominal_element_collation = exprCollation(rightop);
1607 :
1608 : /* look through any binary-compatible relabeling of rightop */
1609 13150 : rightop = strip_array_coercion(rightop);
1610 :
1611 : /*
1612 : * Detect whether the operator is the default equality or inequality
1613 : * operator of the array element type.
1614 : */
1615 13150 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1616 13150 : if (OidIsValid(typentry->eq_opr))
1617 : {
1618 13150 : if (operator == typentry->eq_opr)
1619 12302 : isEquality = true;
1620 848 : else if (get_negator(operator) == typentry->eq_opr)
1621 740 : isInequality = true;
1622 : }
1623 :
1624 : /*
1625 : * If it is equality or inequality, we might be able to estimate this as a
1626 : * form of array containment; for instance "const = ANY(column)" can be
1627 : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1628 : * that, and returns the selectivity estimate if successful, or -1 if not.
1629 : */
1630 13150 : if ((isEquality || isInequality) && !is_join_clause)
1631 : {
1632 13042 : s1 = scalararraysel_containment(root, leftop, rightop,
1633 : nominal_element_type,
1634 : isEquality, useOr, varRelid);
1635 13042 : if (s1 >= 0.0)
1636 38 : return s1;
1637 : }
1638 :
1639 : /*
1640 : * Look up the underlying operator's selectivity estimator. Punt if it
1641 : * hasn't got one.
1642 : */
1643 13112 : if (is_join_clause)
1644 0 : oprsel = get_oprjoin(operator);
1645 : else
1646 13112 : oprsel = get_oprrest(operator);
1647 13112 : if (!oprsel)
1648 0 : return (Selectivity) 0.5;
1649 13112 : fmgr_info(oprsel, &oprselproc);
1650 :
1651 : /*
1652 : * In the array-containment check above, we must only believe that an
1653 : * operator is equality or inequality if it is the default btree equality
1654 : * operator (or its negator) for the element type, since those are the
1655 : * operators that array containment will use. But in what follows, we can
1656 : * be a little laxer, and also believe that any operators using eqsel() or
1657 : * neqsel() as selectivity estimator act like equality or inequality.
1658 : */
1659 13112 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1660 12324 : isEquality = true;
1661 788 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1662 716 : isInequality = true;
1663 :
1664 : /*
1665 : * We consider three cases:
1666 : *
1667 : * 1. rightop is an Array constant: deconstruct the array, apply the
1668 : * operator's selectivity function for each array element, and merge the
1669 : * results in the same way that clausesel.c does for AND/OR combinations.
1670 : *
1671 : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
1672 : * function for each element of the ARRAY[] construct, and merge.
1673 : *
1674 : * 3. otherwise, make a guess ...
1675 : */
1676 13112 : if (rightop && IsA(rightop, Const))
1677 6640 : {
1678 6652 : Datum arraydatum = ((Const *) rightop)->constvalue;
1679 6652 : bool arrayisnull = ((Const *) rightop)->constisnull;
1680 : ArrayType *arrayval;
1681 : int16 elmlen;
1682 : bool elmbyval;
1683 : char elmalign;
1684 : int num_elems;
1685 : Datum *elem_values;
1686 : bool *elem_nulls;
1687 : int i;
1688 :
1689 6652 : if (arrayisnull) /* qual can't succeed if null array */
1690 12 : return (Selectivity) 0.0;
1691 6640 : arrayval = DatumGetArrayTypeP(arraydatum);
1692 6640 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
1693 : &elmlen, &elmbyval, &elmalign);
1694 6640 : deconstruct_array(arrayval,
1695 : ARR_ELEMTYPE(arrayval),
1696 : elmlen, elmbyval, elmalign,
1697 : &elem_values, &elem_nulls, &num_elems);
1698 :
1699 : /*
1700 : * For generic operators, we assume the probability of success is
1701 : * independent for each array element. But for "= ANY" or "<> ALL",
1702 : * if the array elements are distinct (which'd typically be the case)
1703 : * then the probabilities are disjoint, and we should just sum them.
1704 : *
1705 : * If we were being really tense we would try to confirm that the
1706 : * elements are all distinct, but that would be expensive and it
1707 : * doesn't seem to be worth the cycles; it would amount to penalizing
1708 : * well-written queries in favor of poorly-written ones. However, we
1709 : * do protect ourselves a little bit by checking whether the
1710 : * disjointness assumption leads to an impossible (out of range)
1711 : * probability; if so, we fall back to the normal calculation.
1712 : */
1713 6640 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
1714 :
1715 27710 : for (i = 0; i < num_elems; i++)
1716 : {
1717 : List *args;
1718 : Selectivity s2;
1719 :
1720 21070 : args = list_make2(leftop,
1721 : makeConst(nominal_element_type,
1722 : -1,
1723 : nominal_element_collation,
1724 : elmlen,
1725 : elem_values[i],
1726 : elem_nulls[i],
1727 : elmbyval));
1728 21070 : if (is_join_clause)
1729 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1730 : clause->inputcollid,
1731 : PointerGetDatum(root),
1732 : ObjectIdGetDatum(operator),
1733 : PointerGetDatum(args),
1734 : Int16GetDatum(jointype),
1735 : PointerGetDatum(sjinfo)));
1736 : else
1737 21070 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1738 : clause->inputcollid,
1739 : PointerGetDatum(root),
1740 : ObjectIdGetDatum(operator),
1741 : PointerGetDatum(args),
1742 : Int32GetDatum(varRelid)));
1743 :
1744 21070 : if (useOr)
1745 : {
1746 18924 : s1 = s1 + s2 - s1 * s2;
1747 18924 : if (isEquality)
1748 18824 : s1disjoint += s2;
1749 : }
1750 : else
1751 : {
1752 2146 : s1 = s1 * s2;
1753 2146 : if (isInequality)
1754 2106 : s1disjoint += s2 - 1.0;
1755 : }
1756 : }
1757 :
1758 : /* accept disjoint-probability estimate if in range */
1759 6640 : if ((useOr ? isEquality : isInequality) &&
1760 6524 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
1761 6504 : s1 = s1disjoint;
1762 : }
1763 6514 : else if (rightop && IsA(rightop, ArrayExpr) &&
1764 54 : !((ArrayExpr *) rightop)->multidims)
1765 54 : {
1766 54 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
1767 : int16 elmlen;
1768 : bool elmbyval;
1769 : ListCell *l;
1770 :
1771 54 : get_typlenbyval(arrayexpr->element_typeid,
1772 : &elmlen, &elmbyval);
1773 :
1774 : /*
1775 : * We use the assumption of disjoint probabilities here too, although
1776 : * the odds of equal array elements are rather higher if the elements
1777 : * are not all constants (which they won't be, else constant folding
1778 : * would have reduced the ArrayExpr to a Const). In this path it's
1779 : * critical to have the sanity check on the s1disjoint estimate.
1780 : */
1781 54 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
1782 :
1783 200 : foreach(l, arrayexpr->elements)
1784 : {
1785 146 : Node *elem = (Node *) lfirst(l);
1786 : List *args;
1787 : Selectivity s2;
1788 :
1789 : /*
1790 : * Theoretically, if elem isn't of nominal_element_type we should
1791 : * insert a RelabelType, but it seems unlikely that any operator
1792 : * estimation function would really care ...
1793 : */
1794 146 : args = list_make2(leftop, elem);
1795 146 : if (is_join_clause)
1796 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1797 : clause->inputcollid,
1798 : PointerGetDatum(root),
1799 : ObjectIdGetDatum(operator),
1800 : PointerGetDatum(args),
1801 : Int16GetDatum(jointype),
1802 : PointerGetDatum(sjinfo)));
1803 : else
1804 146 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1805 : clause->inputcollid,
1806 : PointerGetDatum(root),
1807 : ObjectIdGetDatum(operator),
1808 : PointerGetDatum(args),
1809 : Int32GetDatum(varRelid)));
1810 :
1811 146 : if (useOr)
1812 : {
1813 146 : s1 = s1 + s2 - s1 * s2;
1814 146 : if (isEquality)
1815 146 : s1disjoint += s2;
1816 : }
1817 : else
1818 : {
1819 0 : s1 = s1 * s2;
1820 0 : if (isInequality)
1821 0 : s1disjoint += s2 - 1.0;
1822 : }
1823 : }
1824 :
1825 : /* accept disjoint-probability estimate if in range */
1826 54 : if ((useOr ? isEquality : isInequality) &&
1827 54 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
1828 54 : s1 = s1disjoint;
1829 : }
1830 : else
1831 : {
1832 : CaseTestExpr *dummyexpr;
1833 : List *args;
1834 : Selectivity s2;
1835 : int i;
1836 :
1837 : /*
1838 : * We need a dummy rightop to pass to the operator selectivity
1839 : * routine. It can be pretty much anything that doesn't look like a
1840 : * constant; CaseTestExpr is a convenient choice.
1841 : */
1842 6406 : dummyexpr = makeNode(CaseTestExpr);
1843 6406 : dummyexpr->typeId = nominal_element_type;
1844 6406 : dummyexpr->typeMod = -1;
1845 6406 : dummyexpr->collation = clause->inputcollid;
1846 6406 : args = list_make2(leftop, dummyexpr);
1847 6406 : if (is_join_clause)
1848 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1849 : clause->inputcollid,
1850 : PointerGetDatum(root),
1851 : ObjectIdGetDatum(operator),
1852 : PointerGetDatum(args),
1853 : Int16GetDatum(jointype),
1854 : PointerGetDatum(sjinfo)));
1855 : else
1856 6406 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1857 : clause->inputcollid,
1858 : PointerGetDatum(root),
1859 : ObjectIdGetDatum(operator),
1860 : PointerGetDatum(args),
1861 : Int32GetDatum(varRelid)));
1862 6406 : s1 = useOr ? 0.0 : 1.0;
1863 :
1864 : /*
1865 : * Arbitrarily assume 10 elements in the eventual array value (see
1866 : * also estimate_array_length). We don't risk an assumption of
1867 : * disjoint probabilities here.
1868 : */
1869 70466 : for (i = 0; i < 10; i++)
1870 : {
1871 64060 : if (useOr)
1872 64060 : s1 = s1 + s2 - s1 * s2;
1873 : else
1874 0 : s1 = s1 * s2;
1875 : }
1876 : }
1877 :
1878 : /* result should be in range, but make sure... */
1879 13100 : CLAMP_PROBABILITY(s1);
1880 :
1881 13100 : return s1;
1882 : }
1883 :
1884 : /*
1885 : * Estimate number of elements in the array yielded by an expression.
1886 : *
1887 : * It's important that this agree with scalararraysel.
1888 : */
1889 : int
1890 49772 : estimate_array_length(Node *arrayexpr)
1891 : {
1892 : /* look through any binary-compatible relabeling of arrayexpr */
1893 49772 : arrayexpr = strip_array_coercion(arrayexpr);
1894 :
1895 49772 : if (arrayexpr && IsA(arrayexpr, Const))
1896 : {
1897 9104 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
1898 9104 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
1899 : ArrayType *arrayval;
1900 :
1901 9104 : if (arrayisnull)
1902 12 : return 0;
1903 9092 : arrayval = DatumGetArrayTypeP(arraydatum);
1904 9092 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
1905 : }
1906 40764 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
1907 96 : !((ArrayExpr *) arrayexpr)->multidims)
1908 : {
1909 96 : return list_length(((ArrayExpr *) arrayexpr)->elements);
1910 : }
1911 : else
1912 : {
1913 : /* default guess --- see also scalararraysel */
1914 40572 : return 10;
1915 : }
1916 : }
1917 :
1918 : /*
1919 : * rowcomparesel - Selectivity of RowCompareExpr Node.
1920 : *
1921 : * We estimate RowCompare selectivity by considering just the first (high
1922 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
1923 : * this estimate could be refined by considering additional columns, it
1924 : * seems unlikely that we could do a lot better without multi-column
1925 : * statistics.
1926 : */
1927 : Selectivity
1928 104 : rowcomparesel(PlannerInfo *root,
1929 : RowCompareExpr *clause,
1930 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1931 : {
1932 : Selectivity s1;
1933 104 : Oid opno = linitial_oid(clause->opnos);
1934 104 : Oid inputcollid = linitial_oid(clause->inputcollids);
1935 : List *opargs;
1936 : bool is_join_clause;
1937 :
1938 : /* Build equivalent arg list for single operator */
1939 104 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
1940 :
1941 : /*
1942 : * Decide if it's a join clause. This should match clausesel.c's
1943 : * treat_as_join_clause(), except that we intentionally consider only the
1944 : * leading columns and not the rest of the clause.
1945 : */
1946 104 : if (varRelid != 0)
1947 : {
1948 : /*
1949 : * Caller is forcing restriction mode (eg, because we are examining an
1950 : * inner indexscan qual).
1951 : */
1952 36 : is_join_clause = false;
1953 : }
1954 68 : else if (sjinfo == NULL)
1955 : {
1956 : /*
1957 : * It must be a restriction clause, since it's being evaluated at a
1958 : * scan node.
1959 : */
1960 60 : is_join_clause = false;
1961 : }
1962 : else
1963 : {
1964 : /*
1965 : * Otherwise, it's a join if there's more than one relation used.
1966 : */
1967 8 : is_join_clause = (NumRelids((Node *) opargs) > 1);
1968 : }
1969 :
1970 104 : if (is_join_clause)
1971 : {
1972 : /* Estimate selectivity for a join clause. */
1973 8 : s1 = join_selectivity(root, opno,
1974 : opargs,
1975 : inputcollid,
1976 : jointype,
1977 : sjinfo);
1978 : }
1979 : else
1980 : {
1981 : /* Estimate selectivity for a restriction clause. */
1982 96 : s1 = restriction_selectivity(root, opno,
1983 : opargs,
1984 : inputcollid,
1985 : varRelid);
1986 : }
1987 :
1988 104 : return s1;
1989 : }
1990 :
1991 : /*
1992 : * eqjoinsel - Join selectivity of "="
1993 : */
1994 : Datum
1995 112960 : eqjoinsel(PG_FUNCTION_ARGS)
1996 : {
1997 112960 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1998 112960 : Oid operator = PG_GETARG_OID(1);
1999 112960 : List *args = (List *) PG_GETARG_POINTER(2);
2000 :
2001 : #ifdef NOT_USED
2002 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2003 : #endif
2004 112960 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2005 : double selec;
2006 : double selec_inner;
2007 : VariableStatData vardata1;
2008 : VariableStatData vardata2;
2009 : double nd1;
2010 : double nd2;
2011 : bool isdefault1;
2012 : bool isdefault2;
2013 : Oid opfuncoid;
2014 : AttStatsSlot sslot1;
2015 : AttStatsSlot sslot2;
2016 112960 : Form_pg_statistic stats1 = NULL;
2017 112960 : Form_pg_statistic stats2 = NULL;
2018 112960 : bool have_mcvs1 = false;
2019 112960 : bool have_mcvs2 = false;
2020 : bool join_is_reversed;
2021 : RelOptInfo *inner_rel;
2022 :
2023 112960 : get_join_variables(root, args, sjinfo,
2024 : &vardata1, &vardata2, &join_is_reversed);
2025 :
2026 112960 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2027 112960 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2028 :
2029 112960 : opfuncoid = get_opcode(operator);
2030 :
2031 112960 : memset(&sslot1, 0, sizeof(sslot1));
2032 112960 : memset(&sslot2, 0, sizeof(sslot2));
2033 :
2034 112960 : if (HeapTupleIsValid(vardata1.statsTuple))
2035 : {
2036 : /* note we allow use of nullfrac regardless of security check */
2037 72404 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2038 72404 : if (statistic_proc_security_check(&vardata1, opfuncoid))
2039 72404 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2040 : STATISTIC_KIND_MCV, InvalidOid,
2041 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2042 : }
2043 :
2044 112960 : if (HeapTupleIsValid(vardata2.statsTuple))
2045 : {
2046 : /* note we allow use of nullfrac regardless of security check */
2047 72994 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2048 72994 : if (statistic_proc_security_check(&vardata2, opfuncoid))
2049 72994 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2050 : STATISTIC_KIND_MCV, InvalidOid,
2051 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2052 : }
2053 :
2054 : /* We need to compute the inner-join selectivity in all cases */
2055 112960 : selec_inner = eqjoinsel_inner(opfuncoid,
2056 : &vardata1, &vardata2,
2057 : nd1, nd2,
2058 : isdefault1, isdefault2,
2059 : &sslot1, &sslot2,
2060 : stats1, stats2,
2061 : have_mcvs1, have_mcvs2);
2062 :
2063 112960 : switch (sjinfo->jointype)
2064 : {
2065 : case JOIN_INNER:
2066 : case JOIN_LEFT:
2067 : case JOIN_FULL:
2068 103832 : selec = selec_inner;
2069 103832 : break;
2070 : case JOIN_SEMI:
2071 : case JOIN_ANTI:
2072 :
2073 : /*
2074 : * Look up the join's inner relation. min_righthand is sufficient
2075 : * information because neither SEMI nor ANTI joins permit any
2076 : * reassociation into or out of their RHS, so the righthand will
2077 : * always be exactly that set of rels.
2078 : */
2079 9128 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2080 :
2081 9128 : if (!join_is_reversed)
2082 6710 : selec = eqjoinsel_semi(opfuncoid,
2083 : &vardata1, &vardata2,
2084 : nd1, nd2,
2085 : isdefault1, isdefault2,
2086 : &sslot1, &sslot2,
2087 : stats1, stats2,
2088 : have_mcvs1, have_mcvs2,
2089 : inner_rel);
2090 : else
2091 : {
2092 2418 : Oid commop = get_commutator(operator);
2093 2418 : Oid commopfuncoid = OidIsValid(commop) ? get_opcode(commop) : InvalidOid;
2094 :
2095 2418 : selec = eqjoinsel_semi(commopfuncoid,
2096 : &vardata2, &vardata1,
2097 : nd2, nd1,
2098 : isdefault2, isdefault1,
2099 : &sslot2, &sslot1,
2100 : stats2, stats1,
2101 : have_mcvs2, have_mcvs1,
2102 : inner_rel);
2103 : }
2104 :
2105 : /*
2106 : * We should never estimate the output of a semijoin to be more
2107 : * rows than we estimate for an inner join with the same input
2108 : * rels and join condition; it's obviously impossible for that to
2109 : * happen. The former estimate is N1 * Ssemi while the latter is
2110 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2111 : * this is worthwhile because of the shakier estimation rules we
2112 : * use in eqjoinsel_semi, particularly in cases where it has to
2113 : * punt entirely.
2114 : */
2115 9128 : selec = Min(selec, inner_rel->rows * selec_inner);
2116 9128 : break;
2117 : default:
2118 : /* other values not expected here */
2119 0 : elog(ERROR, "unrecognized join type: %d",
2120 : (int) sjinfo->jointype);
2121 : selec = 0; /* keep compiler quiet */
2122 : break;
2123 : }
2124 :
2125 112960 : free_attstatsslot(&sslot1);
2126 112960 : free_attstatsslot(&sslot2);
2127 :
2128 112960 : ReleaseVariableStats(vardata1);
2129 112960 : ReleaseVariableStats(vardata2);
2130 :
2131 112960 : CLAMP_PROBABILITY(selec);
2132 :
2133 112960 : PG_RETURN_FLOAT8((float8) selec);
2134 : }
2135 :
2136 : /*
2137 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2138 : *
2139 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2140 : * that it's worth trying to distinguish them here.
2141 : */
2142 : static double
2143 112960 : eqjoinsel_inner(Oid opfuncoid,
2144 : VariableStatData *vardata1, VariableStatData *vardata2,
2145 : double nd1, double nd2,
2146 : bool isdefault1, bool isdefault2,
2147 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2148 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2149 : bool have_mcvs1, bool have_mcvs2)
2150 : {
2151 : double selec;
2152 :
2153 112960 : if (have_mcvs1 && have_mcvs2)
2154 8704 : {
2155 : /*
2156 : * We have most-common-value lists for both relations. Run through
2157 : * the lists to see which MCVs actually join to each other with the
2158 : * given operator. This allows us to determine the exact join
2159 : * selectivity for the portion of the relations represented by the MCV
2160 : * lists. We still have to estimate for the remaining population, but
2161 : * in a skewed distribution this gives us a big leg up in accuracy.
2162 : * For motivation see the analysis in Y. Ioannidis and S.
2163 : * Christodoulakis, "On the propagation of errors in the size of join
2164 : * results", Technical Report 1018, Computer Science Dept., University
2165 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2166 : */
2167 : FmgrInfo eqproc;
2168 : bool *hasmatch1;
2169 : bool *hasmatch2;
2170 8704 : double nullfrac1 = stats1->stanullfrac;
2171 8704 : double nullfrac2 = stats2->stanullfrac;
2172 : double matchprodfreq,
2173 : matchfreq1,
2174 : matchfreq2,
2175 : unmatchfreq1,
2176 : unmatchfreq2,
2177 : otherfreq1,
2178 : otherfreq2,
2179 : totalsel1,
2180 : totalsel2;
2181 : int i,
2182 : nmatches;
2183 :
2184 8704 : fmgr_info(opfuncoid, &eqproc);
2185 8704 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2186 8704 : hasmatch2 = (bool *) palloc0(sslot2->nvalues * sizeof(bool));
2187 :
2188 : /*
2189 : * Note we assume that each MCV will match at most one member of the
2190 : * other MCV list. If the operator isn't really equality, there could
2191 : * be multiple matches --- but we don't look for them, both for speed
2192 : * and because the math wouldn't add up...
2193 : */
2194 8704 : matchprodfreq = 0.0;
2195 8704 : nmatches = 0;
2196 225100 : for (i = 0; i < sslot1->nvalues; i++)
2197 : {
2198 : int j;
2199 :
2200 7562314 : for (j = 0; j < sslot2->nvalues; j++)
2201 : {
2202 7390884 : if (hasmatch2[j])
2203 1598898 : continue;
2204 5791986 : if (DatumGetBool(FunctionCall2Coll(&eqproc,
2205 : sslot1->stacoll,
2206 : sslot1->values[i],
2207 : sslot2->values[j])))
2208 : {
2209 44966 : hasmatch1[i] = hasmatch2[j] = true;
2210 44966 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
2211 44966 : nmatches++;
2212 44966 : break;
2213 : }
2214 : }
2215 : }
2216 8704 : CLAMP_PROBABILITY(matchprodfreq);
2217 : /* Sum up frequencies of matched and unmatched MCVs */
2218 8704 : matchfreq1 = unmatchfreq1 = 0.0;
2219 225100 : for (i = 0; i < sslot1->nvalues; i++)
2220 : {
2221 216396 : if (hasmatch1[i])
2222 44966 : matchfreq1 += sslot1->numbers[i];
2223 : else
2224 171430 : unmatchfreq1 += sslot1->numbers[i];
2225 : }
2226 8704 : CLAMP_PROBABILITY(matchfreq1);
2227 8704 : CLAMP_PROBABILITY(unmatchfreq1);
2228 8704 : matchfreq2 = unmatchfreq2 = 0.0;
2229 424200 : for (i = 0; i < sslot2->nvalues; i++)
2230 : {
2231 415496 : if (hasmatch2[i])
2232 44966 : matchfreq2 += sslot2->numbers[i];
2233 : else
2234 370530 : unmatchfreq2 += sslot2->numbers[i];
2235 : }
2236 8704 : CLAMP_PROBABILITY(matchfreq2);
2237 8704 : CLAMP_PROBABILITY(unmatchfreq2);
2238 8704 : pfree(hasmatch1);
2239 8704 : pfree(hasmatch2);
2240 :
2241 : /*
2242 : * Compute total frequency of non-null values that are not in the MCV
2243 : * lists.
2244 : */
2245 8704 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2246 8704 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2247 8704 : CLAMP_PROBABILITY(otherfreq1);
2248 8704 : CLAMP_PROBABILITY(otherfreq2);
2249 :
2250 : /*
2251 : * We can estimate the total selectivity from the point of view of
2252 : * relation 1 as: the known selectivity for matched MCVs, plus
2253 : * unmatched MCVs that are assumed to match against random members of
2254 : * relation 2's non-MCV population, plus non-MCV values that are
2255 : * assumed to match against random members of relation 2's unmatched
2256 : * MCVs plus non-MCV values.
2257 : */
2258 8704 : totalsel1 = matchprodfreq;
2259 8704 : if (nd2 > sslot2->nvalues)
2260 7364 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2261 8704 : if (nd2 > nmatches)
2262 16600 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2263 8300 : (nd2 - nmatches);
2264 : /* Same estimate from the point of view of relation 2. */
2265 8704 : totalsel2 = matchprodfreq;
2266 8704 : if (nd1 > sslot1->nvalues)
2267 7268 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2268 8704 : if (nd1 > nmatches)
2269 15168 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2270 7584 : (nd1 - nmatches);
2271 :
2272 : /*
2273 : * Use the smaller of the two estimates. This can be justified in
2274 : * essentially the same terms as given below for the no-stats case: to
2275 : * a first approximation, we are estimating from the point of view of
2276 : * the relation with smaller nd.
2277 : */
2278 8704 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2279 : }
2280 : else
2281 : {
2282 : /*
2283 : * We do not have MCV lists for both sides. Estimate the join
2284 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2285 : * is plausible if we assume that the join operator is strict and the
2286 : * non-null values are about equally distributed: a given non-null
2287 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2288 : * of rel2, so total join rows are at most
2289 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2290 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2291 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2292 : * with MIN() is an upper bound. Using the MIN() means we estimate
2293 : * from the point of view of the relation with smaller nd (since the
2294 : * larger nd is determining the MIN). It is reasonable to assume that
2295 : * most tuples in this rel will have join partners, so the bound is
2296 : * probably reasonably tight and should be taken as-is.
2297 : *
2298 : * XXX Can we be smarter if we have an MCV list for just one side? It
2299 : * seems that if we assume equal distribution for the other side, we
2300 : * end up with the same answer anyway.
2301 : */
2302 104256 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2303 104256 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2304 :
2305 104256 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2306 104256 : if (nd1 > nd2)
2307 34606 : selec /= nd1;
2308 : else
2309 69650 : selec /= nd2;
2310 : }
2311 :
2312 112960 : return selec;
2313 : }
2314 :
2315 : /*
2316 : * eqjoinsel_semi --- eqjoinsel for semi join
2317 : *
2318 : * (Also used for anti join, which we are supposed to estimate the same way.)
2319 : * Caller has ensured that vardata1 is the LHS variable.
2320 : * Unlike eqjoinsel_inner, we have to cope with opfuncoid being InvalidOid.
2321 : */
2322 : static double
2323 9128 : eqjoinsel_semi(Oid opfuncoid,
2324 : VariableStatData *vardata1, VariableStatData *vardata2,
2325 : double nd1, double nd2,
2326 : bool isdefault1, bool isdefault2,
2327 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2328 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2329 : bool have_mcvs1, bool have_mcvs2,
2330 : RelOptInfo *inner_rel)
2331 : {
2332 : double selec;
2333 :
2334 : /*
2335 : * We clamp nd2 to be not more than what we estimate the inner relation's
2336 : * size to be. This is intuitively somewhat reasonable since obviously
2337 : * there can't be more than that many distinct values coming from the
2338 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2339 : * likewise) is that this is the only pathway by which restriction clauses
2340 : * applied to the inner rel will affect the join result size estimate,
2341 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2342 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2343 : * the selectivity of outer-rel restrictions.
2344 : *
2345 : * We can apply this clamping both with respect to the base relation from
2346 : * which the join variable comes (if there is just one), and to the
2347 : * immediate inner input relation of the current join.
2348 : *
2349 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2350 : * great, maybe, but it didn't come out of nowhere either. This is most
2351 : * helpful when the inner relation is empty and consequently has no stats.
2352 : */
2353 9128 : if (vardata2->rel)
2354 : {
2355 9128 : if (nd2 >= vardata2->rel->rows)
2356 : {
2357 8536 : nd2 = vardata2->rel->rows;
2358 8536 : isdefault2 = false;
2359 : }
2360 : }
2361 9128 : if (nd2 >= inner_rel->rows)
2362 : {
2363 8504 : nd2 = inner_rel->rows;
2364 8504 : isdefault2 = false;
2365 : }
2366 :
2367 9128 : if (have_mcvs1 && have_mcvs2 && OidIsValid(opfuncoid))
2368 44 : {
2369 : /*
2370 : * We have most-common-value lists for both relations. Run through
2371 : * the lists to see which MCVs actually join to each other with the
2372 : * given operator. This allows us to determine the exact join
2373 : * selectivity for the portion of the relations represented by the MCV
2374 : * lists. We still have to estimate for the remaining population, but
2375 : * in a skewed distribution this gives us a big leg up in accuracy.
2376 : */
2377 : FmgrInfo eqproc;
2378 : bool *hasmatch1;
2379 : bool *hasmatch2;
2380 44 : double nullfrac1 = stats1->stanullfrac;
2381 : double matchfreq1,
2382 : uncertainfrac,
2383 : uncertain;
2384 : int i,
2385 : nmatches,
2386 : clamped_nvalues2;
2387 :
2388 : /*
2389 : * The clamping above could have resulted in nd2 being less than
2390 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2391 : * most common values in the relation will appear in the join input,
2392 : * and so compare to only the first nd2 members of the MCV list. Of
2393 : * course this is frequently wrong, but it's the best bet we can make.
2394 : */
2395 44 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2396 :
2397 44 : fmgr_info(opfuncoid, &eqproc);
2398 44 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2399 44 : hasmatch2 = (bool *) palloc0(clamped_nvalues2 * sizeof(bool));
2400 :
2401 : /*
2402 : * Note we assume that each MCV will match at most one member of the
2403 : * other MCV list. If the operator isn't really equality, there could
2404 : * be multiple matches --- but we don't look for them, both for speed
2405 : * and because the math wouldn't add up...
2406 : */
2407 44 : nmatches = 0;
2408 2724 : for (i = 0; i < sslot1->nvalues; i++)
2409 : {
2410 : int j;
2411 :
2412 103172 : for (j = 0; j < clamped_nvalues2; j++)
2413 : {
2414 102808 : if (hasmatch2[j])
2415 93204 : continue;
2416 9604 : if (DatumGetBool(FunctionCall2Coll(&eqproc,
2417 : sslot1->stacoll,
2418 : sslot1->values[i],
2419 : sslot2->values[j])))
2420 : {
2421 2316 : hasmatch1[i] = hasmatch2[j] = true;
2422 2316 : nmatches++;
2423 2316 : break;
2424 : }
2425 : }
2426 : }
2427 : /* Sum up frequencies of matched MCVs */
2428 44 : matchfreq1 = 0.0;
2429 2724 : for (i = 0; i < sslot1->nvalues; i++)
2430 : {
2431 2680 : if (hasmatch1[i])
2432 2316 : matchfreq1 += sslot1->numbers[i];
2433 : }
2434 44 : CLAMP_PROBABILITY(matchfreq1);
2435 44 : pfree(hasmatch1);
2436 44 : pfree(hasmatch2);
2437 :
2438 : /*
2439 : * Now we need to estimate the fraction of relation 1 that has at
2440 : * least one join partner. We know for certain that the matched MCVs
2441 : * do, so that gives us a lower bound, but we're really in the dark
2442 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2443 : * assume all non-null rel1 rows have join partners, else assume for
2444 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2445 : * can discount the known-matched MCVs from the distinct-values counts
2446 : * before doing the division.
2447 : *
2448 : * Crude as the above is, it's completely useless if we don't have
2449 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2450 : * nd2 is default, punt and assume half of the uncertain rows have
2451 : * join partners.
2452 : */
2453 44 : if (!isdefault1 && !isdefault2)
2454 : {
2455 44 : nd1 -= nmatches;
2456 44 : nd2 -= nmatches;
2457 88 : if (nd1 <= nd2 || nd2 < 0)
2458 32 : uncertainfrac = 1.0;
2459 : else
2460 12 : uncertainfrac = nd2 / nd1;
2461 : }
2462 : else
2463 0 : uncertainfrac = 0.5;
2464 44 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2465 44 : CLAMP_PROBABILITY(uncertain);
2466 44 : selec = matchfreq1 + uncertainfrac * uncertain;
2467 : }
2468 : else
2469 : {
2470 : /*
2471 : * Without MCV lists for both sides, we can only use the heuristic
2472 : * about nd1 vs nd2.
2473 : */
2474 9084 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2475 :
2476 9084 : if (!isdefault1 && !isdefault2)
2477 : {
2478 16044 : if (nd1 <= nd2 || nd2 < 0)
2479 6808 : selec = 1.0 - nullfrac1;
2480 : else
2481 1214 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2482 : }
2483 : else
2484 1062 : selec = 0.5 * (1.0 - nullfrac1);
2485 : }
2486 :
2487 9128 : return selec;
2488 : }
2489 :
2490 : /*
2491 : * neqjoinsel - Join selectivity of "!="
2492 : */
2493 : Datum
2494 1606 : neqjoinsel(PG_FUNCTION_ARGS)
2495 : {
2496 1606 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2497 1606 : Oid operator = PG_GETARG_OID(1);
2498 1606 : List *args = (List *) PG_GETARG_POINTER(2);
2499 1606 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2500 1606 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2501 : float8 result;
2502 :
2503 1606 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
2504 586 : {
2505 : /*
2506 : * For semi-joins, if there is more than one distinct value in the RHS
2507 : * relation then every non-null LHS row must find a row to join since
2508 : * it can only be equal to one of them. We'll assume that there is
2509 : * always more than one distinct RHS value for the sake of stability,
2510 : * though in theory we could have special cases for empty RHS
2511 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
2512 : * fraction of LHS that has the same value as the single RHS value).
2513 : *
2514 : * For anti-joins, if we use the same assumption that there is more
2515 : * than one distinct key in the RHS relation, then every non-null LHS
2516 : * row must be suppressed by the anti-join.
2517 : *
2518 : * So either way, the selectivity estimate should be 1 - nullfrac.
2519 : */
2520 : VariableStatData leftvar;
2521 : VariableStatData rightvar;
2522 : bool reversed;
2523 : HeapTuple statsTuple;
2524 : double nullfrac;
2525 :
2526 586 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
2527 586 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
2528 586 : if (HeapTupleIsValid(statsTuple))
2529 492 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
2530 : else
2531 94 : nullfrac = 0.0;
2532 586 : ReleaseVariableStats(leftvar);
2533 586 : ReleaseVariableStats(rightvar);
2534 :
2535 586 : result = 1.0 - nullfrac;
2536 : }
2537 : else
2538 : {
2539 : /*
2540 : * We want 1 - eqjoinsel() where the equality operator is the one
2541 : * associated with this != operator, that is, its negator.
2542 : */
2543 1020 : Oid eqop = get_negator(operator);
2544 :
2545 1020 : if (eqop)
2546 : {
2547 1020 : result = DatumGetFloat8(DirectFunctionCall5(eqjoinsel,
2548 : PointerGetDatum(root),
2549 : ObjectIdGetDatum(eqop),
2550 : PointerGetDatum(args),
2551 : Int16GetDatum(jointype),
2552 : PointerGetDatum(sjinfo)));
2553 : }
2554 : else
2555 : {
2556 : /* Use default selectivity (should we raise an error instead?) */
2557 0 : result = DEFAULT_EQ_SEL;
2558 : }
2559 1020 : result = 1.0 - result;
2560 : }
2561 :
2562 1606 : PG_RETURN_FLOAT8(result);
2563 : }
2564 :
2565 : /*
2566 : * scalarltjoinsel - Join selectivity of "<" for scalars
2567 : */
2568 : Datum
2569 192 : scalarltjoinsel(PG_FUNCTION_ARGS)
2570 : {
2571 192 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2572 : }
2573 :
2574 : /*
2575 : * scalarlejoinsel - Join selectivity of "<=" for scalars
2576 : */
2577 : Datum
2578 34 : scalarlejoinsel(PG_FUNCTION_ARGS)
2579 : {
2580 34 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2581 : }
2582 :
2583 : /*
2584 : * scalargtjoinsel - Join selectivity of ">" for scalars
2585 : */
2586 : Datum
2587 140 : scalargtjoinsel(PG_FUNCTION_ARGS)
2588 : {
2589 140 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2590 : }
2591 :
2592 : /*
2593 : * scalargejoinsel - Join selectivity of ">=" for scalars
2594 : */
2595 : Datum
2596 12 : scalargejoinsel(PG_FUNCTION_ARGS)
2597 : {
2598 12 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2599 : }
2600 :
2601 :
2602 : /*
2603 : * mergejoinscansel - Scan selectivity of merge join.
2604 : *
2605 : * A merge join will stop as soon as it exhausts either input stream.
2606 : * Therefore, if we can estimate the ranges of both input variables,
2607 : * we can estimate how much of the input will actually be read. This
2608 : * can have a considerable impact on the cost when using indexscans.
2609 : *
2610 : * Also, we can estimate how much of each input has to be read before the
2611 : * first join pair is found, which will affect the join's startup time.
2612 : *
2613 : * clause should be a clause already known to be mergejoinable. opfamily,
2614 : * strategy, and nulls_first specify the sort ordering being used.
2615 : *
2616 : * The outputs are:
2617 : * *leftstart is set to the fraction of the left-hand variable expected
2618 : * to be scanned before the first join pair is found (0 to 1).
2619 : * *leftend is set to the fraction of the left-hand variable expected
2620 : * to be scanned before the join terminates (0 to 1).
2621 : * *rightstart, *rightend similarly for the right-hand variable.
2622 : */
2623 : void
2624 52178 : mergejoinscansel(PlannerInfo *root, Node *clause,
2625 : Oid opfamily, int strategy, bool nulls_first,
2626 : Selectivity *leftstart, Selectivity *leftend,
2627 : Selectivity *rightstart, Selectivity *rightend)
2628 : {
2629 : Node *left,
2630 : *right;
2631 : VariableStatData leftvar,
2632 : rightvar;
2633 : int op_strategy;
2634 : Oid op_lefttype;
2635 : Oid op_righttype;
2636 : Oid opno,
2637 : lsortop,
2638 : rsortop,
2639 : lstatop,
2640 : rstatop,
2641 : ltop,
2642 : leop,
2643 : revltop,
2644 : revleop;
2645 : bool isgt;
2646 : Datum leftmin,
2647 : leftmax,
2648 : rightmin,
2649 : rightmax;
2650 : double selec;
2651 :
2652 : /* Set default results if we can't figure anything out. */
2653 : /* XXX should default "start" fraction be a bit more than 0? */
2654 52178 : *leftstart = *rightstart = 0.0;
2655 52178 : *leftend = *rightend = 1.0;
2656 :
2657 : /* Deconstruct the merge clause */
2658 52178 : if (!is_opclause(clause))
2659 0 : return; /* shouldn't happen */
2660 52178 : opno = ((OpExpr *) clause)->opno;
2661 52178 : left = get_leftop((Expr *) clause);
2662 52178 : right = get_rightop((Expr *) clause);
2663 52178 : if (!right)
2664 0 : return; /* shouldn't happen */
2665 :
2666 : /* Look for stats for the inputs */
2667 52178 : examine_variable(root, left, 0, &leftvar);
2668 52178 : examine_variable(root, right, 0, &rightvar);
2669 :
2670 : /* Extract the operator's declared left/right datatypes */
2671 52178 : get_op_opfamily_properties(opno, opfamily, false,
2672 : &op_strategy,
2673 : &op_lefttype,
2674 : &op_righttype);
2675 : Assert(op_strategy == BTEqualStrategyNumber);
2676 :
2677 : /*
2678 : * Look up the various operators we need. If we don't find them all, it
2679 : * probably means the opfamily is broken, but we just fail silently.
2680 : *
2681 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
2682 : * operator, regardless of which sort direction we are considering.
2683 : */
2684 52178 : switch (strategy)
2685 : {
2686 : case BTLessStrategyNumber:
2687 52174 : isgt = false;
2688 52174 : if (op_lefttype == op_righttype)
2689 : {
2690 : /* easy case */
2691 51056 : ltop = get_opfamily_member(opfamily,
2692 : op_lefttype, op_righttype,
2693 : BTLessStrategyNumber);
2694 51056 : leop = get_opfamily_member(opfamily,
2695 : op_lefttype, op_righttype,
2696 : BTLessEqualStrategyNumber);
2697 51056 : lsortop = ltop;
2698 51056 : rsortop = ltop;
2699 51056 : lstatop = lsortop;
2700 51056 : rstatop = rsortop;
2701 51056 : revltop = ltop;
2702 51056 : revleop = leop;
2703 : }
2704 : else
2705 : {
2706 1118 : ltop = get_opfamily_member(opfamily,
2707 : op_lefttype, op_righttype,
2708 : BTLessStrategyNumber);
2709 1118 : leop = get_opfamily_member(opfamily,
2710 : op_lefttype, op_righttype,
2711 : BTLessEqualStrategyNumber);
2712 1118 : lsortop = get_opfamily_member(opfamily,
2713 : op_lefttype, op_lefttype,
2714 : BTLessStrategyNumber);
2715 1118 : rsortop = get_opfamily_member(opfamily,
2716 : op_righttype, op_righttype,
2717 : BTLessStrategyNumber);
2718 1118 : lstatop = lsortop;
2719 1118 : rstatop = rsortop;
2720 1118 : revltop = get_opfamily_member(opfamily,
2721 : op_righttype, op_lefttype,
2722 : BTLessStrategyNumber);
2723 1118 : revleop = get_opfamily_member(opfamily,
2724 : op_righttype, op_lefttype,
2725 : BTLessEqualStrategyNumber);
2726 : }
2727 52174 : break;
2728 : case BTGreaterStrategyNumber:
2729 : /* descending-order case */
2730 4 : isgt = true;
2731 4 : if (op_lefttype == op_righttype)
2732 : {
2733 : /* easy case */
2734 4 : ltop = get_opfamily_member(opfamily,
2735 : op_lefttype, op_righttype,
2736 : BTGreaterStrategyNumber);
2737 4 : leop = get_opfamily_member(opfamily,
2738 : op_lefttype, op_righttype,
2739 : BTGreaterEqualStrategyNumber);
2740 4 : lsortop = ltop;
2741 4 : rsortop = ltop;
2742 4 : lstatop = get_opfamily_member(opfamily,
2743 : op_lefttype, op_lefttype,
2744 : BTLessStrategyNumber);
2745 4 : rstatop = lstatop;
2746 4 : revltop = ltop;
2747 4 : revleop = leop;
2748 : }
2749 : else
2750 : {
2751 0 : ltop = get_opfamily_member(opfamily,
2752 : op_lefttype, op_righttype,
2753 : BTGreaterStrategyNumber);
2754 0 : leop = get_opfamily_member(opfamily,
2755 : op_lefttype, op_righttype,
2756 : BTGreaterEqualStrategyNumber);
2757 0 : lsortop = get_opfamily_member(opfamily,
2758 : op_lefttype, op_lefttype,
2759 : BTGreaterStrategyNumber);
2760 0 : rsortop = get_opfamily_member(opfamily,
2761 : op_righttype, op_righttype,
2762 : BTGreaterStrategyNumber);
2763 0 : lstatop = get_opfamily_member(opfamily,
2764 : op_lefttype, op_lefttype,
2765 : BTLessStrategyNumber);
2766 0 : rstatop = get_opfamily_member(opfamily,
2767 : op_righttype, op_righttype,
2768 : BTLessStrategyNumber);
2769 0 : revltop = get_opfamily_member(opfamily,
2770 : op_righttype, op_lefttype,
2771 : BTGreaterStrategyNumber);
2772 0 : revleop = get_opfamily_member(opfamily,
2773 : op_righttype, op_lefttype,
2774 : BTGreaterEqualStrategyNumber);
2775 : }
2776 4 : break;
2777 : default:
2778 0 : goto fail; /* shouldn't get here */
2779 : }
2780 :
2781 52178 : if (!OidIsValid(lsortop) ||
2782 52178 : !OidIsValid(rsortop) ||
2783 52178 : !OidIsValid(lstatop) ||
2784 52178 : !OidIsValid(rstatop) ||
2785 52170 : !OidIsValid(ltop) ||
2786 52170 : !OidIsValid(leop) ||
2787 52170 : !OidIsValid(revltop) ||
2788 : !OidIsValid(revleop))
2789 : goto fail; /* insufficient info in catalogs */
2790 :
2791 : /* Try to get ranges of both inputs */
2792 52170 : if (!isgt)
2793 : {
2794 52166 : if (!get_variable_range(root, &leftvar, lstatop,
2795 : &leftmin, &leftmax))
2796 15358 : goto fail; /* no range available from stats */
2797 36808 : if (!get_variable_range(root, &rightvar, rstatop,
2798 : &rightmin, &rightmax))
2799 3348 : goto fail; /* no range available from stats */
2800 : }
2801 : else
2802 : {
2803 : /* need to swap the max and min */
2804 4 : if (!get_variable_range(root, &leftvar, lstatop,
2805 : &leftmax, &leftmin))
2806 4 : goto fail; /* no range available from stats */
2807 0 : if (!get_variable_range(root, &rightvar, rstatop,
2808 : &rightmax, &rightmin))
2809 0 : goto fail; /* no range available from stats */
2810 : }
2811 :
2812 : /*
2813 : * Now, the fraction of the left variable that will be scanned is the
2814 : * fraction that's <= the right-side maximum value. But only believe
2815 : * non-default estimates, else stick with our 1.0.
2816 : */
2817 33460 : selec = scalarineqsel(root, leop, isgt, true, &leftvar,
2818 : rightmax, op_righttype);
2819 33460 : if (selec != DEFAULT_INEQ_SEL)
2820 33456 : *leftend = selec;
2821 :
2822 : /* And similarly for the right variable. */
2823 33460 : selec = scalarineqsel(root, revleop, isgt, true, &rightvar,
2824 : leftmax, op_lefttype);
2825 33460 : if (selec != DEFAULT_INEQ_SEL)
2826 33460 : *rightend = selec;
2827 :
2828 : /*
2829 : * Only one of the two "end" fractions can really be less than 1.0;
2830 : * believe the smaller estimate and reset the other one to exactly 1.0. If
2831 : * we get exactly equal estimates (as can easily happen with self-joins),
2832 : * believe neither.
2833 : */
2834 33460 : if (*leftend > *rightend)
2835 19900 : *leftend = 1.0;
2836 13560 : else if (*leftend < *rightend)
2837 11830 : *rightend = 1.0;
2838 : else
2839 1730 : *leftend = *rightend = 1.0;
2840 :
2841 : /*
2842 : * Also, the fraction of the left variable that will be scanned before the
2843 : * first join pair is found is the fraction that's < the right-side
2844 : * minimum value. But only believe non-default estimates, else stick with
2845 : * our own default.
2846 : */
2847 33460 : selec = scalarineqsel(root, ltop, isgt, false, &leftvar,
2848 : rightmin, op_righttype);
2849 33460 : if (selec != DEFAULT_INEQ_SEL)
2850 33460 : *leftstart = selec;
2851 :
2852 : /* And similarly for the right variable. */
2853 33460 : selec = scalarineqsel(root, revltop, isgt, false, &rightvar,
2854 : leftmin, op_lefttype);
2855 33460 : if (selec != DEFAULT_INEQ_SEL)
2856 33460 : *rightstart = selec;
2857 :
2858 : /*
2859 : * Only one of the two "start" fractions can really be more than zero;
2860 : * believe the larger estimate and reset the other one to exactly 0.0. If
2861 : * we get exactly equal estimates (as can easily happen with self-joins),
2862 : * believe neither.
2863 : */
2864 33460 : if (*leftstart < *rightstart)
2865 8852 : *leftstart = 0.0;
2866 24608 : else if (*leftstart > *rightstart)
2867 18570 : *rightstart = 0.0;
2868 : else
2869 6038 : *leftstart = *rightstart = 0.0;
2870 :
2871 : /*
2872 : * If the sort order is nulls-first, we're going to have to skip over any
2873 : * nulls too. These would not have been counted by scalarineqsel, and we
2874 : * can safely add in this fraction regardless of whether we believe
2875 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
2876 : */
2877 33460 : if (nulls_first)
2878 : {
2879 : Form_pg_statistic stats;
2880 :
2881 0 : if (HeapTupleIsValid(leftvar.statsTuple))
2882 : {
2883 0 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
2884 0 : *leftstart += stats->stanullfrac;
2885 0 : CLAMP_PROBABILITY(*leftstart);
2886 0 : *leftend += stats->stanullfrac;
2887 0 : CLAMP_PROBABILITY(*leftend);
2888 : }
2889 0 : if (HeapTupleIsValid(rightvar.statsTuple))
2890 : {
2891 0 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
2892 0 : *rightstart += stats->stanullfrac;
2893 0 : CLAMP_PROBABILITY(*rightstart);
2894 0 : *rightend += stats->stanullfrac;
2895 0 : CLAMP_PROBABILITY(*rightend);
2896 : }
2897 : }
2898 :
2899 : /* Disbelieve start >= end, just in case that can happen */
2900 33460 : if (*leftstart >= *leftend)
2901 : {
2902 988 : *leftstart = 0.0;
2903 988 : *leftend = 1.0;
2904 : }
2905 33460 : if (*rightstart >= *rightend)
2906 : {
2907 1168 : *rightstart = 0.0;
2908 1168 : *rightend = 1.0;
2909 : }
2910 :
2911 : fail:
2912 52178 : ReleaseVariableStats(leftvar);
2913 52178 : ReleaseVariableStats(rightvar);
2914 : }
2915 :
2916 :
2917 : /*
2918 : * Helper routine for estimate_num_groups: add an item to a list of
2919 : * GroupVarInfos, but only if it's not known equal to any of the existing
2920 : * entries.
2921 : */
2922 : typedef struct
2923 : {
2924 : Node *var; /* might be an expression, not just a Var */
2925 : RelOptInfo *rel; /* relation it belongs to */
2926 : double ndistinct; /* # distinct values */
2927 : } GroupVarInfo;
2928 :
2929 : static List *
2930 8818 : add_unique_group_var(PlannerInfo *root, List *varinfos,
2931 : Node *var, VariableStatData *vardata)
2932 : {
2933 : GroupVarInfo *varinfo;
2934 : double ndistinct;
2935 : bool isdefault;
2936 : ListCell *lc;
2937 :
2938 8818 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
2939 :
2940 11430 : foreach(lc, varinfos)
2941 : {
2942 2804 : varinfo = (GroupVarInfo *) lfirst(lc);
2943 :
2944 : /* Drop exact duplicates */
2945 2804 : if (equal(var, varinfo->var))
2946 332 : return varinfos;
2947 :
2948 : /*
2949 : * Drop known-equal vars, but only if they belong to different
2950 : * relations (see comments for estimate_num_groups)
2951 : */
2952 3028 : if (vardata->rel != varinfo->rel &&
2953 364 : exprs_known_equal(root, var, varinfo->var))
2954 : {
2955 56 : if (varinfo->ndistinct <= ndistinct)
2956 : {
2957 : /* Keep older item, forget new one */
2958 52 : return varinfos;
2959 : }
2960 : else
2961 : {
2962 : /* Delete the older item */
2963 4 : varinfos = foreach_delete_current(varinfos, lc);
2964 : }
2965 : }
2966 : }
2967 :
2968 8626 : varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
2969 :
2970 8626 : varinfo->var = var;
2971 8626 : varinfo->rel = vardata->rel;
2972 8626 : varinfo->ndistinct = ndistinct;
2973 8626 : varinfos = lappend(varinfos, varinfo);
2974 8626 : return varinfos;
2975 : }
2976 :
2977 : /*
2978 : * estimate_num_groups - Estimate number of groups in a grouped query
2979 : *
2980 : * Given a query having a GROUP BY clause, estimate how many groups there
2981 : * will be --- ie, the number of distinct combinations of the GROUP BY
2982 : * expressions.
2983 : *
2984 : * This routine is also used to estimate the number of rows emitted by
2985 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
2986 : * actually, we only use it for DISTINCT when there's no grouping or
2987 : * aggregation ahead of the DISTINCT.)
2988 : *
2989 : * Inputs:
2990 : * root - the query
2991 : * groupExprs - list of expressions being grouped by
2992 : * input_rows - number of rows estimated to arrive at the group/unique
2993 : * filter step
2994 : * pgset - NULL, or a List** pointing to a grouping set to filter the
2995 : * groupExprs against
2996 : *
2997 : * Given the lack of any cross-correlation statistics in the system, it's
2998 : * impossible to do anything really trustworthy with GROUP BY conditions
2999 : * involving multiple Vars. We should however avoid assuming the worst
3000 : * case (all possible cross-product terms actually appear as groups) since
3001 : * very often the grouped-by Vars are highly correlated. Our current approach
3002 : * is as follows:
3003 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3004 : * independently of their content, and are ignored in the subsequent
3005 : * steps. This is mainly because tests like "col IS NULL" break the
3006 : * heuristic used in step 2 especially badly.
3007 : * 2. Reduce the given expressions to a list of unique Vars used. For
3008 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3009 : * It is clearly correct not to count the same Var more than once.
3010 : * It is also reasonable to treat f(x) the same as x: f() cannot
3011 : * increase the number of distinct values (unless it is volatile,
3012 : * which we consider unlikely for grouping), but it probably won't
3013 : * reduce the number of distinct values much either.
3014 : * As a special case, if a GROUP BY expression can be matched to an
3015 : * expressional index for which we have statistics, then we treat the
3016 : * whole expression as though it were just a Var.
3017 : * 3. If the list contains Vars of different relations that are known equal
3018 : * due to equivalence classes, then drop all but one of the Vars from each
3019 : * known-equal set, keeping the one with smallest estimated # of values
3020 : * (since the extra values of the others can't appear in joined rows).
3021 : * Note the reason we only consider Vars of different relations is that
3022 : * if we considered ones of the same rel, we'd be double-counting the
3023 : * restriction selectivity of the equality in the next step.
3024 : * 4. For Vars within a single source rel, we multiply together the numbers
3025 : * of values, clamp to the number of rows in the rel (divided by 10 if
3026 : * more than one Var), and then multiply by a factor based on the
3027 : * selectivity of the restriction clauses for that rel. When there's
3028 : * more than one Var, the initial product is probably too high (it's the
3029 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3030 : * helpful heuristic for not letting the estimate get out of hand. (The
3031 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3032 : * we multiply by to adjust for the restriction selectivity assumes that
3033 : * the restriction clauses are independent of the grouping, which may not
3034 : * be a valid assumption, but it's hard to do better.
3035 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3036 : * rel, and multiply the results together.
3037 : * Note that rels not containing grouped Vars are ignored completely, as are
3038 : * join clauses. Such rels cannot increase the number of groups, and we
3039 : * assume such clauses do not reduce the number either (somewhat bogus,
3040 : * but we don't have the info to do better).
3041 : */
3042 : double
3043 7288 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3044 : List **pgset)
3045 : {
3046 7288 : List *varinfos = NIL;
3047 7288 : double srf_multiplier = 1.0;
3048 : double numdistinct;
3049 : ListCell *l;
3050 : int i;
3051 :
3052 : /*
3053 : * We don't ever want to return an estimate of zero groups, as that tends
3054 : * to lead to division-by-zero and other unpleasantness. The input_rows
3055 : * estimate is usually already at least 1, but clamp it just in case it
3056 : * isn't.
3057 : */
3058 7288 : input_rows = clamp_row_est(input_rows);
3059 :
3060 : /*
3061 : * If no grouping columns, there's exactly one group. (This can't happen
3062 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3063 : * corner cases with set operations.)
3064 : */
3065 7288 : if (groupExprs == NIL || (pgset && list_length(*pgset) < 1))
3066 432 : return 1.0;
3067 :
3068 : /*
3069 : * Count groups derived from boolean grouping expressions. For other
3070 : * expressions, find the unique Vars used, treating an expression as a Var
3071 : * if we can find stats for it. For each one, record the statistical
3072 : * estimate of number of distinct values (total in its table, without
3073 : * regard for filtering).
3074 : */
3075 6856 : numdistinct = 1.0;
3076 :
3077 6856 : i = 0;
3078 16164 : foreach(l, groupExprs)
3079 : {
3080 9312 : Node *groupexpr = (Node *) lfirst(l);
3081 : double this_srf_multiplier;
3082 : VariableStatData vardata;
3083 : List *varshere;
3084 : ListCell *l2;
3085 :
3086 : /* is expression in this grouping set? */
3087 9312 : if (pgset && !list_member_int(*pgset, i++))
3088 5402 : continue;
3089 :
3090 : /*
3091 : * Set-returning functions in grouping columns are a bit problematic.
3092 : * The code below will effectively ignore their SRF nature and come up
3093 : * with a numdistinct estimate as though they were scalar functions.
3094 : * We compensate by scaling up the end result by the largest SRF
3095 : * rowcount estimate. (This will be an overestimate if the SRF
3096 : * produces multiple copies of any output value, but it seems best to
3097 : * assume the SRF's outputs are distinct. In any case, it's probably
3098 : * pointless to worry too much about this without much better
3099 : * estimates for SRF output rowcounts than we have today.)
3100 : */
3101 9036 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3102 9036 : if (srf_multiplier < this_srf_multiplier)
3103 72 : srf_multiplier = this_srf_multiplier;
3104 :
3105 : /* Short-circuit for expressions returning boolean */
3106 9036 : if (exprType(groupexpr) == BOOLOID)
3107 : {
3108 24 : numdistinct *= 2.0;
3109 24 : continue;
3110 : }
3111 :
3112 : /*
3113 : * If examine_variable is able to deduce anything about the GROUP BY
3114 : * expression, treat it as a single variable even if it's really more
3115 : * complicated.
3116 : */
3117 9012 : examine_variable(root, groupexpr, 0, &vardata);
3118 9012 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3119 : {
3120 4540 : varinfos = add_unique_group_var(root, varinfos,
3121 : groupexpr, &vardata);
3122 4540 : ReleaseVariableStats(vardata);
3123 4540 : continue;
3124 : }
3125 4472 : ReleaseVariableStats(vardata);
3126 :
3127 : /*
3128 : * Else pull out the component Vars. Handle PlaceHolderVars by
3129 : * recursing into their arguments (effectively assuming that the
3130 : * PlaceHolderVar doesn't change the number of groups, which boils
3131 : * down to ignoring the possible addition of nulls to the result set).
3132 : */
3133 4472 : varshere = pull_var_clause(groupexpr,
3134 : PVC_RECURSE_AGGREGATES |
3135 : PVC_RECURSE_WINDOWFUNCS |
3136 : PVC_RECURSE_PLACEHOLDERS);
3137 :
3138 : /*
3139 : * If we find any variable-free GROUP BY item, then either it is a
3140 : * constant (and we can ignore it) or it contains a volatile function;
3141 : * in the latter case we punt and assume that each input row will
3142 : * yield a distinct group.
3143 : */
3144 4472 : if (varshere == NIL)
3145 : {
3146 290 : if (contain_volatile_functions(groupexpr))
3147 4 : return input_rows;
3148 286 : continue;
3149 : }
3150 :
3151 : /*
3152 : * Else add variables to varinfos list
3153 : */
3154 8460 : foreach(l2, varshere)
3155 : {
3156 4278 : Node *var = (Node *) lfirst(l2);
3157 :
3158 4278 : examine_variable(root, var, 0, &vardata);
3159 4278 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3160 4278 : ReleaseVariableStats(vardata);
3161 : }
3162 : }
3163 :
3164 : /*
3165 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3166 : * list.
3167 : */
3168 6852 : if (varinfos == NIL)
3169 : {
3170 : /* Apply SRF multiplier as we would do in the long path */
3171 148 : numdistinct *= srf_multiplier;
3172 : /* Round off */
3173 148 : numdistinct = ceil(numdistinct);
3174 : /* Guard against out-of-range answers */
3175 148 : if (numdistinct > input_rows)
3176 0 : numdistinct = input_rows;
3177 148 : if (numdistinct < 1.0)
3178 0 : numdistinct = 1.0;
3179 148 : return numdistinct;
3180 : }
3181 :
3182 : /*
3183 : * Group Vars by relation and estimate total numdistinct.
3184 : *
3185 : * For each iteration of the outer loop, we process the frontmost Var in
3186 : * varinfos, plus all other Vars in the same relation. We remove these
3187 : * Vars from the newvarinfos list for the next iteration. This is the
3188 : * easiest way to group Vars of same rel together.
3189 : */
3190 : do
3191 : {
3192 6924 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3193 6924 : RelOptInfo *rel = varinfo1->rel;
3194 6924 : double reldistinct = 1;
3195 6924 : double relmaxndistinct = reldistinct;
3196 6924 : int relvarcount = 0;
3197 6924 : List *newvarinfos = NIL;
3198 6924 : List *relvarinfos = NIL;
3199 :
3200 : /*
3201 : * Split the list of varinfos in two - one for the current rel, one
3202 : * for remaining Vars on other rels.
3203 : */
3204 6924 : relvarinfos = lappend(relvarinfos, varinfo1);
3205 8870 : for_each_cell(l, varinfos, list_second_cell(varinfos))
3206 : {
3207 1946 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3208 :
3209 1946 : if (varinfo2->rel == varinfo1->rel)
3210 : {
3211 : /* varinfos on current rel */
3212 1698 : relvarinfos = lappend(relvarinfos, varinfo2);
3213 : }
3214 : else
3215 : {
3216 : /* not time to process varinfo2 yet */
3217 248 : newvarinfos = lappend(newvarinfos, varinfo2);
3218 : }
3219 : }
3220 :
3221 : /*
3222 : * Get the numdistinct estimate for the Vars of this rel. We
3223 : * iteratively search for multivariate n-distinct with maximum number
3224 : * of vars; assuming that each var group is independent of the others,
3225 : * we multiply them together. Any remaining relvarinfos after no more
3226 : * multivariate matches are found are assumed independent too, so
3227 : * their individual ndistinct estimates are multiplied also.
3228 : *
3229 : * While iterating, count how many separate numdistinct values we
3230 : * apply. We apply a fudge factor below, but only if we multiplied
3231 : * more than one such values.
3232 : */
3233 20788 : while (relvarinfos)
3234 : {
3235 : double mvndistinct;
3236 :
3237 6940 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3238 : &mvndistinct))
3239 : {
3240 36 : reldistinct *= mvndistinct;
3241 36 : if (relmaxndistinct < mvndistinct)
3242 36 : relmaxndistinct = mvndistinct;
3243 36 : relvarcount++;
3244 : }
3245 : else
3246 : {
3247 15438 : foreach(l, relvarinfos)
3248 : {
3249 8534 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3250 :
3251 8534 : reldistinct *= varinfo2->ndistinct;
3252 8534 : if (relmaxndistinct < varinfo2->ndistinct)
3253 7102 : relmaxndistinct = varinfo2->ndistinct;
3254 8534 : relvarcount++;
3255 : }
3256 :
3257 : /* we're done with this relation */
3258 6904 : relvarinfos = NIL;
3259 : }
3260 : }
3261 :
3262 : /*
3263 : * Sanity check --- don't divide by zero if empty relation.
3264 : */
3265 : Assert(IS_SIMPLE_REL(rel));
3266 6924 : if (rel->tuples > 0)
3267 : {
3268 : /*
3269 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
3270 : * fudge factor is because the Vars are probably correlated but we
3271 : * don't know by how much. We should never clamp to less than the
3272 : * largest ndistinct value for any of the Vars, though, since
3273 : * there will surely be at least that many groups.
3274 : */
3275 6882 : double clamp = rel->tuples;
3276 :
3277 6882 : if (relvarcount > 1)
3278 : {
3279 1188 : clamp *= 0.1;
3280 1188 : if (clamp < relmaxndistinct)
3281 : {
3282 626 : clamp = relmaxndistinct;
3283 : /* for sanity in case some ndistinct is too large: */
3284 626 : if (clamp > rel->tuples)
3285 8 : clamp = rel->tuples;
3286 : }
3287 : }
3288 6882 : if (reldistinct > clamp)
3289 800 : reldistinct = clamp;
3290 :
3291 : /*
3292 : * Update the estimate based on the restriction selectivity,
3293 : * guarding against division by zero when reldistinct is zero.
3294 : * Also skip this if we know that we are returning all rows.
3295 : */
3296 6882 : if (reldistinct > 0 && rel->rows < rel->tuples)
3297 : {
3298 : /*
3299 : * Given a table containing N rows with n distinct values in a
3300 : * uniform distribution, if we select p rows at random then
3301 : * the expected number of distinct values selected is
3302 : *
3303 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
3304 : *
3305 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
3306 : *
3307 : * See "Approximating block accesses in database
3308 : * organizations", S. B. Yao, Communications of the ACM,
3309 : * Volume 20 Issue 4, April 1977 Pages 260-261.
3310 : *
3311 : * Alternatively, re-arranging the terms from the factorials,
3312 : * this may be written as
3313 : *
3314 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
3315 : *
3316 : * This form of the formula is more efficient to compute in
3317 : * the common case where p is larger than N/n. Additionally,
3318 : * as pointed out by Dell'Era, if i << N for all terms in the
3319 : * product, it can be approximated by
3320 : *
3321 : * n * (1 - ((N-p)/N)^(N/n))
3322 : *
3323 : * See "Expected distinct values when selecting from a bag
3324 : * without replacement", Alberto Dell'Era,
3325 : * http://www.adellera.it/investigations/distinct_balls/.
3326 : *
3327 : * The condition i << N is equivalent to n >> 1, so this is a
3328 : * good approximation when the number of distinct values in
3329 : * the table is large. It turns out that this formula also
3330 : * works well even when n is small.
3331 : */
3332 1744 : reldistinct *=
3333 1744 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
3334 1744 : rel->tuples / reldistinct));
3335 : }
3336 6882 : reldistinct = clamp_row_est(reldistinct);
3337 :
3338 : /*
3339 : * Update estimate of total distinct groups.
3340 : */
3341 6882 : numdistinct *= reldistinct;
3342 : }
3343 :
3344 6924 : varinfos = newvarinfos;
3345 6924 : } while (varinfos != NIL);
3346 :
3347 : /* Now we can account for the effects of any SRFs */
3348 6704 : numdistinct *= srf_multiplier;
3349 :
3350 : /* Round off */
3351 6704 : numdistinct = ceil(numdistinct);
3352 :
3353 : /* Guard against out-of-range answers */
3354 6704 : if (numdistinct > input_rows)
3355 374 : numdistinct = input_rows;
3356 6704 : if (numdistinct < 1.0)
3357 0 : numdistinct = 1.0;
3358 :
3359 6704 : return numdistinct;
3360 : }
3361 :
3362 : /*
3363 : * Estimate hash bucket statistics when the specified expression is used
3364 : * as a hash key for the given number of buckets.
3365 : *
3366 : * This attempts to determine two values:
3367 : *
3368 : * 1. The frequency of the most common value of the expression (returns
3369 : * zero into *mcv_freq if we can't get that).
3370 : *
3371 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
3372 : * divided by total tuples in relation.
3373 : *
3374 : * XXX This is really pretty bogus since we're effectively assuming that the
3375 : * distribution of hash keys will be the same after applying restriction
3376 : * clauses as it was in the underlying relation. However, we are not nearly
3377 : * smart enough to figure out how the restrict clauses might change the
3378 : * distribution, so this will have to do for now.
3379 : *
3380 : * We are passed the number of buckets the executor will use for the given
3381 : * input relation. If the data were perfectly distributed, with the same
3382 : * number of tuples going into each available bucket, then the bucketsize
3383 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
3384 : * only if (a) there are at least nbuckets distinct data values, and (b)
3385 : * we have a not-too-skewed data distribution. Otherwise the buckets will
3386 : * be nonuniformly occupied. If the other relation in the join has a key
3387 : * distribution similar to this one's, then the most-loaded buckets are
3388 : * exactly those that will be probed most often. Therefore, the "average"
3389 : * bucket size for costing purposes should really be taken as something close
3390 : * to the "worst case" bucket size. We try to estimate this by adjusting the
3391 : * fraction if there are too few distinct data values, and then scaling up
3392 : * by the ratio of the most common value's frequency to the average frequency.
3393 : *
3394 : * If no statistics are available, use a default estimate of 0.1. This will
3395 : * discourage use of a hash rather strongly if the inner relation is large,
3396 : * which is what we want. We do not want to hash unless we know that the
3397 : * inner rel is well-dispersed (or the alternatives seem much worse).
3398 : *
3399 : * The caller should also check that the mcv_freq is not so large that the
3400 : * most common value would by itself require an impractically large bucket.
3401 : * In a hash join, the executor can split buckets if they get too big, but
3402 : * obviously that doesn't help for a bucket that contains many duplicates of
3403 : * the same value.
3404 : */
3405 : void
3406 76084 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
3407 : Selectivity *mcv_freq,
3408 : Selectivity *bucketsize_frac)
3409 : {
3410 : VariableStatData vardata;
3411 : double estfract,
3412 : ndistinct,
3413 : stanullfrac,
3414 : avgfreq;
3415 : bool isdefault;
3416 : AttStatsSlot sslot;
3417 :
3418 76084 : examine_variable(root, hashkey, 0, &vardata);
3419 :
3420 : /* Look up the frequency of the most common value, if available */
3421 76084 : *mcv_freq = 0.0;
3422 :
3423 76084 : if (HeapTupleIsValid(vardata.statsTuple))
3424 : {
3425 45676 : if (get_attstatsslot(&sslot, vardata.statsTuple,
3426 : STATISTIC_KIND_MCV, InvalidOid,
3427 : ATTSTATSSLOT_NUMBERS))
3428 : {
3429 : /*
3430 : * The first MCV stat is for the most common value.
3431 : */
3432 24692 : if (sslot.nnumbers > 0)
3433 24692 : *mcv_freq = sslot.numbers[0];
3434 24692 : free_attstatsslot(&sslot);
3435 : }
3436 : }
3437 :
3438 : /* Get number of distinct values */
3439 76084 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
3440 :
3441 : /*
3442 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
3443 : * mcv_freq is known to be even higher than that, use it instead.
3444 : */
3445 76084 : if (isdefault)
3446 : {
3447 12036 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
3448 12036 : ReleaseVariableStats(vardata);
3449 12036 : return;
3450 : }
3451 :
3452 : /* Get fraction that are null */
3453 64048 : if (HeapTupleIsValid(vardata.statsTuple))
3454 : {
3455 : Form_pg_statistic stats;
3456 :
3457 45664 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
3458 45664 : stanullfrac = stats->stanullfrac;
3459 : }
3460 : else
3461 18384 : stanullfrac = 0.0;
3462 :
3463 : /* Compute avg freq of all distinct data values in raw relation */
3464 64048 : avgfreq = (1.0 - stanullfrac) / ndistinct;
3465 :
3466 : /*
3467 : * Adjust ndistinct to account for restriction clauses. Observe we are
3468 : * assuming that the data distribution is affected uniformly by the
3469 : * restriction clauses!
3470 : *
3471 : * XXX Possibly better way, but much more expensive: multiply by
3472 : * selectivity of rel's restriction clauses that mention the target Var.
3473 : */
3474 64048 : if (vardata.rel && vardata.rel->tuples > 0)
3475 : {
3476 64040 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
3477 64040 : ndistinct = clamp_row_est(ndistinct);
3478 : }
3479 :
3480 : /*
3481 : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
3482 : * number of buckets is less than the expected number of distinct values;
3483 : * otherwise it is 1/ndistinct.
3484 : */
3485 64048 : if (ndistinct > nbuckets)
3486 40 : estfract = 1.0 / nbuckets;
3487 : else
3488 64008 : estfract = 1.0 / ndistinct;
3489 :
3490 : /*
3491 : * Adjust estimated bucketsize upward to account for skewed distribution.
3492 : */
3493 64048 : if (avgfreq > 0.0 && *mcv_freq > avgfreq)
3494 23824 : estfract *= *mcv_freq / avgfreq;
3495 :
3496 : /*
3497 : * Clamp bucketsize to sane range (the above adjustment could easily
3498 : * produce an out-of-range result). We set the lower bound a little above
3499 : * zero, since zero isn't a very sane result.
3500 : */
3501 64048 : if (estfract < 1.0e-6)
3502 0 : estfract = 1.0e-6;
3503 64048 : else if (estfract > 1.0)
3504 19162 : estfract = 1.0;
3505 :
3506 64048 : *bucketsize_frac = (Selectivity) estfract;
3507 :
3508 64048 : ReleaseVariableStats(vardata);
3509 : }
3510 :
3511 : /*
3512 : * estimate_hashagg_tablesize
3513 : * estimate the number of bytes that a hash aggregate hashtable will
3514 : * require based on the agg_costs, path width and number of groups.
3515 : *
3516 : * We return the result as "double" to forestall any possible overflow
3517 : * problem in the multiplication by dNumGroups.
3518 : *
3519 : * XXX this may be over-estimating the size now that hashagg knows to omit
3520 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
3521 : * grouping columns not in the hashed set are counted here even though hashagg
3522 : * won't store them. Is this a problem?
3523 : */
3524 : double
3525 4608 : estimate_hashagg_tablesize(Path *path, const AggClauseCosts *agg_costs,
3526 : double dNumGroups)
3527 : {
3528 : Size hashentrysize;
3529 :
3530 : /* Estimate per-hash-entry space at tuple width... */
3531 4608 : hashentrysize = MAXALIGN(path->pathtarget->width) +
3532 : MAXALIGN(SizeofMinimalTupleHeader);
3533 :
3534 : /* plus space for pass-by-ref transition values... */
3535 4608 : hashentrysize += agg_costs->transitionSpace;
3536 : /* plus the per-hash-entry overhead */
3537 4608 : hashentrysize += hash_agg_entry_size(agg_costs->numAggs);
3538 :
3539 : /*
3540 : * Note that this disregards the effect of fill-factor and growth policy
3541 : * of the hash table. That's probably ok, given that the default
3542 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
3543 : * "double-in-size" growth policies here.
3544 : */
3545 4608 : return hashentrysize * dNumGroups;
3546 : }
3547 :
3548 :
3549 : /*-------------------------------------------------------------------------
3550 : *
3551 : * Support routines
3552 : *
3553 : *-------------------------------------------------------------------------
3554 : */
3555 :
3556 : /*
3557 : * Find applicable ndistinct statistics for the given list of VarInfos (which
3558 : * must all belong to the given rel), and update *ndistinct to the estimate of
3559 : * the MVNDistinctItem that best matches. If a match it found, *varinfos is
3560 : * updated to remove the list of matched varinfos.
3561 : *
3562 : * Varinfos that aren't for simple Vars are ignored.
3563 : *
3564 : * Return true if we're able to find a match, false otherwise.
3565 : */
3566 : static bool
3567 6940 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
3568 : List **varinfos, double *ndistinct)
3569 : {
3570 : ListCell *lc;
3571 6940 : Bitmapset *attnums = NULL;
3572 : int nmatches;
3573 6940 : Oid statOid = InvalidOid;
3574 : MVNDistinct *stats;
3575 6940 : Bitmapset *matched = NULL;
3576 :
3577 : /* bail out immediately if the table has no extended statistics */
3578 6940 : if (!rel->statlist)
3579 6884 : return false;
3580 :
3581 : /* Determine the attnums we're looking for */
3582 184 : foreach(lc, *varinfos)
3583 : {
3584 128 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
3585 :
3586 : Assert(varinfo->rel == rel);
3587 :
3588 128 : if (IsA(varinfo->var, Var))
3589 : {
3590 128 : attnums = bms_add_member(attnums,
3591 128 : ((Var *) varinfo->var)->varattno);
3592 : }
3593 : }
3594 :
3595 : /* look for the ndistinct statistics matching the most vars */
3596 56 : nmatches = 1; /* we require at least two matches */
3597 224 : foreach(lc, rel->statlist)
3598 : {
3599 168 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
3600 : Bitmapset *shared;
3601 : int nshared;
3602 :
3603 : /* skip statistics of other kinds */
3604 168 : if (info->kind != STATS_EXT_NDISTINCT)
3605 112 : continue;
3606 :
3607 : /* compute attnums shared by the vars and the statistics object */
3608 56 : shared = bms_intersect(info->keys, attnums);
3609 56 : nshared = bms_num_members(shared);
3610 :
3611 : /*
3612 : * Does this statistics object match more columns than the currently
3613 : * best object? If so, use this one instead.
3614 : *
3615 : * XXX This should break ties using name of the object, or something
3616 : * like that, to make the outcome stable.
3617 : */
3618 56 : if (nshared > nmatches)
3619 : {
3620 36 : statOid = info->statOid;
3621 36 : nmatches = nshared;
3622 36 : matched = shared;
3623 : }
3624 : }
3625 :
3626 : /* No match? */
3627 56 : if (statOid == InvalidOid)
3628 20 : return false;
3629 : Assert(nmatches > 1 && matched != NULL);
3630 :
3631 36 : stats = statext_ndistinct_load(statOid);
3632 :
3633 : /*
3634 : * If we have a match, search it for the specific item that matches (there
3635 : * must be one), and construct the output values.
3636 : */
3637 36 : if (stats)
3638 : {
3639 : int i;
3640 36 : List *newlist = NIL;
3641 36 : MVNDistinctItem *item = NULL;
3642 :
3643 : /* Find the specific item that exactly matches the combination */
3644 108 : for (i = 0; i < stats->nitems; i++)
3645 : {
3646 108 : MVNDistinctItem *tmpitem = &stats->items[i];
3647 :
3648 108 : if (bms_subset_compare(tmpitem->attrs, matched) == BMS_EQUAL)
3649 : {
3650 36 : item = tmpitem;
3651 36 : break;
3652 : }
3653 : }
3654 :
3655 : /* make sure we found an item */
3656 36 : if (!item)
3657 0 : elog(ERROR, "corrupt MVNDistinct entry");
3658 :
3659 : /* Form the output varinfo list, keeping only unmatched ones */
3660 140 : foreach(lc, *varinfos)
3661 : {
3662 104 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
3663 : AttrNumber attnum;
3664 :
3665 104 : if (!IsA(varinfo->var, Var))
3666 : {
3667 0 : newlist = lappend(newlist, varinfo);
3668 0 : continue;
3669 : }
3670 :
3671 104 : attnum = ((Var *) varinfo->var)->varattno;
3672 104 : if (!bms_is_member(attnum, matched))
3673 16 : newlist = lappend(newlist, varinfo);
3674 : }
3675 :
3676 36 : *varinfos = newlist;
3677 36 : *ndistinct = item->ndistinct;
3678 36 : return true;
3679 : }
3680 :
3681 0 : return false;
3682 : }
3683 :
3684 : /*
3685 : * convert_to_scalar
3686 : * Convert non-NULL values of the indicated types to the comparison
3687 : * scale needed by scalarineqsel().
3688 : * Returns "true" if successful.
3689 : *
3690 : * XXX this routine is a hack: ideally we should look up the conversion
3691 : * subroutines in pg_type.
3692 : *
3693 : * All numeric datatypes are simply converted to their equivalent
3694 : * "double" values. (NUMERIC values that are outside the range of "double"
3695 : * are clamped to +/- HUGE_VAL.)
3696 : *
3697 : * String datatypes are converted by convert_string_to_scalar(),
3698 : * which is explained below. The reason why this routine deals with
3699 : * three values at a time, not just one, is that we need it for strings.
3700 : *
3701 : * The bytea datatype is just enough different from strings that it has
3702 : * to be treated separately.
3703 : *
3704 : * The several datatypes representing absolute times are all converted
3705 : * to Timestamp, which is actually a double, and then we just use that
3706 : * double value. Note this will give correct results even for the "special"
3707 : * values of Timestamp, since those are chosen to compare correctly;
3708 : * see timestamp_cmp.
3709 : *
3710 : * The several datatypes representing relative times (intervals) are all
3711 : * converted to measurements expressed in seconds.
3712 : */
3713 : static bool
3714 44936 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
3715 : Datum lobound, Datum hibound, Oid boundstypid,
3716 : double *scaledlobound, double *scaledhibound)
3717 : {
3718 44936 : bool failure = false;
3719 :
3720 : /*
3721 : * Both the valuetypid and the boundstypid should exactly match the
3722 : * declared input type(s) of the operator we are invoked for. However,
3723 : * extensions might try to use scalarineqsel as estimator for operators
3724 : * with input type(s) we don't handle here; in such cases, we want to
3725 : * return false, not fail. In any case, we mustn't assume that valuetypid
3726 : * and boundstypid are identical.
3727 : *
3728 : * XXX The histogram we are interpolating between points of could belong
3729 : * to a column that's only binary-compatible with the declared type. In
3730 : * essence we are assuming that the semantics of binary-compatible types
3731 : * are enough alike that we can use a histogram generated with one type's
3732 : * operators to estimate selectivity for the other's. This is outright
3733 : * wrong in some cases --- in particular signed versus unsigned
3734 : * interpretation could trip us up. But it's useful enough in the
3735 : * majority of cases that we do it anyway. Should think about more
3736 : * rigorous ways to do it.
3737 : */
3738 44936 : switch (valuetypid)
3739 : {
3740 : /*
3741 : * Built-in numeric types
3742 : */
3743 : case BOOLOID:
3744 : case INT2OID:
3745 : case INT4OID:
3746 : case INT8OID:
3747 : case FLOAT4OID:
3748 : case FLOAT8OID:
3749 : case NUMERICOID:
3750 : case OIDOID:
3751 : case REGPROCOID:
3752 : case REGPROCEDUREOID:
3753 : case REGOPEROID:
3754 : case REGOPERATOROID:
3755 : case REGCLASSOID:
3756 : case REGTYPEOID:
3757 : case REGCONFIGOID:
3758 : case REGDICTIONARYOID:
3759 : case REGROLEOID:
3760 : case REGNAMESPACEOID:
3761 42744 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
3762 : &failure);
3763 42744 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
3764 : &failure);
3765 42744 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
3766 : &failure);
3767 42744 : return !failure;
3768 :
3769 : /*
3770 : * Built-in string types
3771 : */
3772 : case CHAROID:
3773 : case BPCHAROID:
3774 : case VARCHAROID:
3775 : case TEXTOID:
3776 : case NAMEOID:
3777 : {
3778 2188 : char *valstr = convert_string_datum(value, valuetypid,
3779 : collid, &failure);
3780 2188 : char *lostr = convert_string_datum(lobound, boundstypid,
3781 : collid, &failure);
3782 2188 : char *histr = convert_string_datum(hibound, boundstypid,
3783 : collid, &failure);
3784 :
3785 : /*
3786 : * Bail out if any of the values is not of string type. We
3787 : * might leak converted strings for the other value(s), but
3788 : * that's not worth troubling over.
3789 : */
3790 2188 : if (failure)
3791 0 : return false;
3792 :
3793 2188 : convert_string_to_scalar(valstr, scaledvalue,
3794 : lostr, scaledlobound,
3795 : histr, scaledhibound);
3796 2188 : pfree(valstr);
3797 2188 : pfree(lostr);
3798 2188 : pfree(histr);
3799 2188 : return true;
3800 : }
3801 :
3802 : /*
3803 : * Built-in bytea type
3804 : */
3805 : case BYTEAOID:
3806 : {
3807 : /* We only support bytea vs bytea comparison */
3808 0 : if (boundstypid != BYTEAOID)
3809 0 : return false;
3810 0 : convert_bytea_to_scalar(value, scaledvalue,
3811 : lobound, scaledlobound,
3812 : hibound, scaledhibound);
3813 0 : return true;
3814 : }
3815 :
3816 : /*
3817 : * Built-in time types
3818 : */
3819 : case TIMESTAMPOID:
3820 : case TIMESTAMPTZOID:
3821 : case DATEOID:
3822 : case INTERVALOID:
3823 : case TIMEOID:
3824 : case TIMETZOID:
3825 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
3826 : &failure);
3827 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
3828 : &failure);
3829 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
3830 : &failure);
3831 0 : return !failure;
3832 :
3833 : /*
3834 : * Built-in network types
3835 : */
3836 : case INETOID:
3837 : case CIDROID:
3838 : case MACADDROID:
3839 : case MACADDR8OID:
3840 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
3841 : &failure);
3842 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
3843 : &failure);
3844 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
3845 : &failure);
3846 0 : return !failure;
3847 : }
3848 : /* Don't know how to convert */
3849 4 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
3850 4 : return false;
3851 : }
3852 :
3853 : /*
3854 : * Do convert_to_scalar()'s work for any numeric data type.
3855 : *
3856 : * On failure (e.g., unsupported typid), set *failure to true;
3857 : * otherwise, that variable is not changed.
3858 : */
3859 : static double
3860 128232 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
3861 : {
3862 128232 : switch (typid)
3863 : {
3864 : case BOOLOID:
3865 0 : return (double) DatumGetBool(value);
3866 : case INT2OID:
3867 8 : return (double) DatumGetInt16(value);
3868 : case INT4OID:
3869 12154 : return (double) DatumGetInt32(value);
3870 : case INT8OID:
3871 0 : return (double) DatumGetInt64(value);
3872 : case FLOAT4OID:
3873 0 : return (double) DatumGetFloat4(value);
3874 : case FLOAT8OID:
3875 24 : return (double) DatumGetFloat8(value);
3876 : case NUMERICOID:
3877 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
3878 0 : return (double)
3879 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
3880 : value));
3881 : case OIDOID:
3882 : case REGPROCOID:
3883 : case REGPROCEDUREOID:
3884 : case REGOPEROID:
3885 : case REGOPERATOROID:
3886 : case REGCLASSOID:
3887 : case REGTYPEOID:
3888 : case REGCONFIGOID:
3889 : case REGDICTIONARYOID:
3890 : case REGROLEOID:
3891 : case REGNAMESPACEOID:
3892 : /* we can treat OIDs as integers... */
3893 116046 : return (double) DatumGetObjectId(value);
3894 : }
3895 :
3896 0 : *failure = true;
3897 0 : return 0;
3898 : }
3899 :
3900 : /*
3901 : * Do convert_to_scalar()'s work for any character-string data type.
3902 : *
3903 : * String datatypes are converted to a scale that ranges from 0 to 1,
3904 : * where we visualize the bytes of the string as fractional digits.
3905 : *
3906 : * We do not want the base to be 256, however, since that tends to
3907 : * generate inflated selectivity estimates; few databases will have
3908 : * occurrences of all 256 possible byte values at each position.
3909 : * Instead, use the smallest and largest byte values seen in the bounds
3910 : * as the estimated range for each byte, after some fudging to deal with
3911 : * the fact that we probably aren't going to see the full range that way.
3912 : *
3913 : * An additional refinement is that we discard any common prefix of the
3914 : * three strings before computing the scaled values. This allows us to
3915 : * "zoom in" when we encounter a narrow data range. An example is a phone
3916 : * number database where all the values begin with the same area code.
3917 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
3918 : * so this is more likely to happen than you might think.)
3919 : */
3920 : static void
3921 2188 : convert_string_to_scalar(char *value,
3922 : double *scaledvalue,
3923 : char *lobound,
3924 : double *scaledlobound,
3925 : char *hibound,
3926 : double *scaledhibound)
3927 : {
3928 : int rangelo,
3929 : rangehi;
3930 : char *sptr;
3931 :
3932 2188 : rangelo = rangehi = (unsigned char) hibound[0];
3933 30466 : for (sptr = lobound; *sptr; sptr++)
3934 : {
3935 28278 : if (rangelo > (unsigned char) *sptr)
3936 5752 : rangelo = (unsigned char) *sptr;
3937 28278 : if (rangehi < (unsigned char) *sptr)
3938 2522 : rangehi = (unsigned char) *sptr;
3939 : }
3940 34306 : for (sptr = hibound; *sptr; sptr++)
3941 : {
3942 32118 : if (rangelo > (unsigned char) *sptr)
3943 654 : rangelo = (unsigned char) *sptr;
3944 32118 : if (rangehi < (unsigned char) *sptr)
3945 2926 : rangehi = (unsigned char) *sptr;
3946 : }
3947 : /* If range includes any upper-case ASCII chars, make it include all */
3948 2188 : if (rangelo <= 'Z' && rangehi >= 'A')
3949 : {
3950 1172 : if (rangelo > 'A')
3951 56 : rangelo = 'A';
3952 1172 : if (rangehi < 'Z')
3953 328 : rangehi = 'Z';
3954 : }
3955 : /* Ditto lower-case */
3956 2188 : if (rangelo <= 'z' && rangehi >= 'a')
3957 : {
3958 1828 : if (rangelo > 'a')
3959 0 : rangelo = 'a';
3960 1828 : if (rangehi < 'z')
3961 1650 : rangehi = 'z';
3962 : }
3963 : /* Ditto digits */
3964 2188 : if (rangelo <= '9' && rangehi >= '0')
3965 : {
3966 796 : if (rangelo > '0')
3967 788 : rangelo = '0';
3968 796 : if (rangehi < '9')
3969 4 : rangehi = '9';
3970 : }
3971 :
3972 : /*
3973 : * If range includes less than 10 chars, assume we have not got enough
3974 : * data, and make it include regular ASCII set.
3975 : */
3976 2188 : if (rangehi - rangelo < 9)
3977 : {
3978 0 : rangelo = ' ';
3979 0 : rangehi = 127;
3980 : }
3981 :
3982 : /*
3983 : * Now strip any common prefix of the three strings.
3984 : */
3985 6426 : while (*lobound)
3986 : {
3987 4154 : if (*lobound != *hibound || *lobound != *value)
3988 : break;
3989 2050 : lobound++, hibound++, value++;
3990 : }
3991 :
3992 : /*
3993 : * Now we can do the conversions.
3994 : */
3995 2188 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
3996 2188 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
3997 2188 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
3998 2188 : }
3999 :
4000 : static double
4001 6564 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
4002 : {
4003 6564 : int slen = strlen(value);
4004 : double num,
4005 : denom,
4006 : base;
4007 :
4008 6564 : if (slen <= 0)
4009 84 : return 0.0; /* empty string has scalar value 0 */
4010 :
4011 : /*
4012 : * There seems little point in considering more than a dozen bytes from
4013 : * the string. Since base is at least 10, that will give us nominal
4014 : * resolution of at least 12 decimal digits, which is surely far more
4015 : * precision than this estimation technique has got anyway (especially in
4016 : * non-C locales). Also, even with the maximum possible base of 256, this
4017 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
4018 : * overflow on any known machine.
4019 : */
4020 6480 : if (slen > 12)
4021 1954 : slen = 12;
4022 :
4023 : /* Convert initial characters to fraction */
4024 6480 : base = rangehi - rangelo + 1;
4025 6480 : num = 0.0;
4026 6480 : denom = base;
4027 62146 : while (slen-- > 0)
4028 : {
4029 49186 : int ch = (unsigned char) *value++;
4030 :
4031 49186 : if (ch < rangelo)
4032 80 : ch = rangelo - 1;
4033 49106 : else if (ch > rangehi)
4034 0 : ch = rangehi + 1;
4035 49186 : num += ((double) (ch - rangelo)) / denom;
4036 49186 : denom *= base;
4037 : }
4038 :
4039 6480 : return num;
4040 : }
4041 :
4042 : /*
4043 : * Convert a string-type Datum into a palloc'd, null-terminated string.
4044 : *
4045 : * On failure (e.g., unsupported typid), set *failure to true;
4046 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
4047 : *
4048 : * When using a non-C locale, we must pass the string through strxfrm()
4049 : * before continuing, so as to generate correct locale-specific results.
4050 : */
4051 : static char *
4052 6564 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
4053 : {
4054 : char *val;
4055 :
4056 6564 : switch (typid)
4057 : {
4058 : case CHAROID:
4059 0 : val = (char *) palloc(2);
4060 0 : val[0] = DatumGetChar(value);
4061 0 : val[1] = '\0';
4062 0 : break;
4063 : case BPCHAROID:
4064 : case VARCHAROID:
4065 : case TEXTOID:
4066 1140 : val = TextDatumGetCString(value);
4067 1140 : break;
4068 : case NAMEOID:
4069 : {
4070 5424 : NameData *nm = (NameData *) DatumGetPointer(value);
4071 :
4072 5424 : val = pstrdup(NameStr(*nm));
4073 5424 : break;
4074 : }
4075 : default:
4076 0 : *failure = true;
4077 0 : return NULL;
4078 : }
4079 :
4080 6564 : if (!lc_collate_is_c(collid))
4081 : {
4082 : char *xfrmstr;
4083 : size_t xfrmlen;
4084 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
4085 :
4086 : /*
4087 : * XXX: We could guess at a suitable output buffer size and only call
4088 : * strxfrm twice if our guess is too small.
4089 : *
4090 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
4091 : * bogus data or set an error. This is not really a problem unless it
4092 : * crashes since it will only give an estimation error and nothing
4093 : * fatal.
4094 : */
4095 : #if _MSC_VER == 1400 /* VS.Net 2005 */
4096 :
4097 : /*
4098 : *
4099 : * http://connect.microsoft.com/VisualStudio/feedback/ViewFeedback.aspx?FeedbackID=99694
4100 : */
4101 : {
4102 : char x[1];
4103 :
4104 : xfrmlen = strxfrm(x, val, 0);
4105 : }
4106 : #else
4107 72 : xfrmlen = strxfrm(NULL, val, 0);
4108 : #endif
4109 : #ifdef WIN32
4110 :
4111 : /*
4112 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
4113 : * of trying to allocate this much memory (and fail), just return the
4114 : * original string unmodified as if we were in the C locale.
4115 : */
4116 : if (xfrmlen == INT_MAX)
4117 : return val;
4118 : #endif
4119 72 : xfrmstr = (char *) palloc(xfrmlen + 1);
4120 72 : xfrmlen2 = strxfrm(xfrmstr, val, xfrmlen + 1);
4121 :
4122 : /*
4123 : * Some systems (e.g., glibc) can return a smaller value from the
4124 : * second call than the first; thus the Assert must be <= not ==.
4125 : */
4126 : Assert(xfrmlen2 <= xfrmlen);
4127 72 : pfree(val);
4128 72 : val = xfrmstr;
4129 : }
4130 :
4131 6564 : return val;
4132 : }
4133 :
4134 : /*
4135 : * Do convert_to_scalar()'s work for any bytea data type.
4136 : *
4137 : * Very similar to convert_string_to_scalar except we can't assume
4138 : * null-termination and therefore pass explicit lengths around.
4139 : *
4140 : * Also, assumptions about likely "normal" ranges of characters have been
4141 : * removed - a data range of 0..255 is always used, for now. (Perhaps
4142 : * someday we will add information about actual byte data range to
4143 : * pg_statistic.)
4144 : */
4145 : static void
4146 0 : convert_bytea_to_scalar(Datum value,
4147 : double *scaledvalue,
4148 : Datum lobound,
4149 : double *scaledlobound,
4150 : Datum hibound,
4151 : double *scaledhibound)
4152 : {
4153 0 : bytea *valuep = DatumGetByteaPP(value);
4154 0 : bytea *loboundp = DatumGetByteaPP(lobound);
4155 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
4156 : int rangelo,
4157 : rangehi,
4158 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
4159 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
4160 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
4161 : i,
4162 : minlen;
4163 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
4164 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
4165 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
4166 :
4167 : /*
4168 : * Assume bytea data is uniformly distributed across all byte values.
4169 : */
4170 0 : rangelo = 0;
4171 0 : rangehi = 255;
4172 :
4173 : /*
4174 : * Now strip any common prefix of the three strings.
4175 : */
4176 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
4177 0 : for (i = 0; i < minlen; i++)
4178 : {
4179 0 : if (*lostr != *histr || *lostr != *valstr)
4180 : break;
4181 0 : lostr++, histr++, valstr++;
4182 0 : loboundlen--, hiboundlen--, valuelen--;
4183 : }
4184 :
4185 : /*
4186 : * Now we can do the conversions.
4187 : */
4188 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
4189 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
4190 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
4191 0 : }
4192 :
4193 : static double
4194 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
4195 : int rangelo, int rangehi)
4196 : {
4197 : double num,
4198 : denom,
4199 : base;
4200 :
4201 0 : if (valuelen <= 0)
4202 0 : return 0.0; /* empty string has scalar value 0 */
4203 :
4204 : /*
4205 : * Since base is 256, need not consider more than about 10 chars (even
4206 : * this many seems like overkill)
4207 : */
4208 0 : if (valuelen > 10)
4209 0 : valuelen = 10;
4210 :
4211 : /* Convert initial characters to fraction */
4212 0 : base = rangehi - rangelo + 1;
4213 0 : num = 0.0;
4214 0 : denom = base;
4215 0 : while (valuelen-- > 0)
4216 : {
4217 0 : int ch = *value++;
4218 :
4219 0 : if (ch < rangelo)
4220 0 : ch = rangelo - 1;
4221 0 : else if (ch > rangehi)
4222 0 : ch = rangehi + 1;
4223 0 : num += ((double) (ch - rangelo)) / denom;
4224 0 : denom *= base;
4225 : }
4226 :
4227 0 : return num;
4228 : }
4229 :
4230 : /*
4231 : * Do convert_to_scalar()'s work for any timevalue data type.
4232 : *
4233 : * On failure (e.g., unsupported typid), set *failure to true;
4234 : * otherwise, that variable is not changed.
4235 : */
4236 : static double
4237 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
4238 : {
4239 0 : switch (typid)
4240 : {
4241 : case TIMESTAMPOID:
4242 0 : return DatumGetTimestamp(value);
4243 : case TIMESTAMPTZOID:
4244 0 : return DatumGetTimestampTz(value);
4245 : case DATEOID:
4246 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
4247 : case INTERVALOID:
4248 : {
4249 0 : Interval *interval = DatumGetIntervalP(value);
4250 :
4251 : /*
4252 : * Convert the month part of Interval to days using assumed
4253 : * average month length of 365.25/12.0 days. Not too
4254 : * accurate, but plenty good enough for our purposes.
4255 : */
4256 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
4257 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
4258 : }
4259 : case TIMEOID:
4260 0 : return DatumGetTimeADT(value);
4261 : case TIMETZOID:
4262 : {
4263 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
4264 :
4265 : /* use GMT-equivalent time */
4266 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
4267 : }
4268 : }
4269 :
4270 0 : *failure = true;
4271 0 : return 0;
4272 : }
4273 :
4274 :
4275 : /*
4276 : * get_restriction_variable
4277 : * Examine the args of a restriction clause to see if it's of the
4278 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
4279 : * where "variable" could be either a Var or an expression in vars of a
4280 : * single relation. If so, extract information about the variable,
4281 : * and also indicate which side it was on and the other argument.
4282 : *
4283 : * Inputs:
4284 : * root: the planner info
4285 : * args: clause argument list
4286 : * varRelid: see specs for restriction selectivity functions
4287 : *
4288 : * Outputs: (these are valid only if true is returned)
4289 : * *vardata: gets information about variable (see examine_variable)
4290 : * *other: gets other clause argument, aggressively reduced to a constant
4291 : * *varonleft: set true if variable is on the left, false if on the right
4292 : *
4293 : * Returns true if a variable is identified, otherwise false.
4294 : *
4295 : * Note: if there are Vars on both sides of the clause, we must fail, because
4296 : * callers are expecting that the other side will act like a pseudoconstant.
4297 : */
4298 : bool
4299 346586 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
4300 : VariableStatData *vardata, Node **other,
4301 : bool *varonleft)
4302 : {
4303 : Node *left,
4304 : *right;
4305 : VariableStatData rdata;
4306 :
4307 : /* Fail if not a binary opclause (probably shouldn't happen) */
4308 346586 : if (list_length(args) != 2)
4309 0 : return false;
4310 :
4311 346586 : left = (Node *) linitial(args);
4312 346586 : right = (Node *) lsecond(args);
4313 :
4314 : /*
4315 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
4316 : * relations will be treated as pseudoconstants.
4317 : */
4318 346586 : examine_variable(root, left, varRelid, vardata);
4319 346586 : examine_variable(root, right, varRelid, &rdata);
4320 :
4321 : /*
4322 : * If one side is a variable and the other not, we win.
4323 : */
4324 346586 : if (vardata->rel && rdata.rel == NULL)
4325 : {
4326 286006 : *varonleft = true;
4327 286006 : *other = estimate_expression_value(root, rdata.var);
4328 : /* Assume we need no ReleaseVariableStats(rdata) here */
4329 286006 : return true;
4330 : }
4331 :
4332 60580 : if (vardata->rel == NULL && rdata.rel)
4333 : {
4334 58808 : *varonleft = false;
4335 58808 : *other = estimate_expression_value(root, vardata->var);
4336 : /* Assume we need no ReleaseVariableStats(*vardata) here */
4337 58808 : *vardata = rdata;
4338 58808 : return true;
4339 : }
4340 :
4341 : /* Oops, clause has wrong structure (probably var op var) */
4342 1772 : ReleaseVariableStats(*vardata);
4343 1772 : ReleaseVariableStats(rdata);
4344 :
4345 1772 : return false;
4346 : }
4347 :
4348 : /*
4349 : * get_join_variables
4350 : * Apply examine_variable() to each side of a join clause.
4351 : * Also, attempt to identify whether the join clause has the same
4352 : * or reversed sense compared to the SpecialJoinInfo.
4353 : *
4354 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
4355 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
4356 : * where we can't tell for sure, we default to assuming it's normal.
4357 : */
4358 : void
4359 113546 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
4360 : VariableStatData *vardata1, VariableStatData *vardata2,
4361 : bool *join_is_reversed)
4362 : {
4363 : Node *left,
4364 : *right;
4365 :
4366 113546 : if (list_length(args) != 2)
4367 0 : elog(ERROR, "join operator should take two arguments");
4368 :
4369 113546 : left = (Node *) linitial(args);
4370 113546 : right = (Node *) lsecond(args);
4371 :
4372 113546 : examine_variable(root, left, 0, vardata1);
4373 113546 : examine_variable(root, right, 0, vardata2);
4374 :
4375 227020 : if (vardata1->rel &&
4376 113474 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
4377 18696 : *join_is_reversed = true; /* var1 is on RHS */
4378 189648 : else if (vardata2->rel &&
4379 94798 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
4380 80 : *join_is_reversed = true; /* var2 is on LHS */
4381 : else
4382 94770 : *join_is_reversed = false;
4383 113546 : }
4384 :
4385 : /*
4386 : * examine_variable
4387 : * Try to look up statistical data about an expression.
4388 : * Fill in a VariableStatData struct to describe the expression.
4389 : *
4390 : * Inputs:
4391 : * root: the planner info
4392 : * node: the expression tree to examine
4393 : * varRelid: see specs for restriction selectivity functions
4394 : *
4395 : * Outputs: *vardata is filled as follows:
4396 : * var: the input expression (with any binary relabeling stripped, if
4397 : * it is or contains a variable; but otherwise the type is preserved)
4398 : * rel: RelOptInfo for relation containing variable; NULL if expression
4399 : * contains no Vars (NOTE this could point to a RelOptInfo of a
4400 : * subquery, not one in the current query).
4401 : * statsTuple: the pg_statistic entry for the variable, if one exists;
4402 : * otherwise NULL.
4403 : * freefunc: pointer to a function to release statsTuple with.
4404 : * vartype: exposed type of the expression; this should always match
4405 : * the declared input type of the operator we are estimating for.
4406 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
4407 : * commonly the same as the exposed type of the variable argument,
4408 : * but can be different in binary-compatible-type cases.
4409 : * isunique: true if we were able to match the var to a unique index or a
4410 : * single-column DISTINCT clause, implying its values are unique for
4411 : * this query. (Caution: this should be trusted for statistical
4412 : * purposes only, since we do not check indimmediate nor verify that
4413 : * the exact same definition of equality applies.)
4414 : * acl_ok: true if current user has permission to read the column(s)
4415 : * underlying the pg_statistic entry. This is consulted by
4416 : * statistic_proc_security_check().
4417 : *
4418 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
4419 : */
4420 : void
4421 1153566 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
4422 : VariableStatData *vardata)
4423 : {
4424 : Node *basenode;
4425 : Relids varnos;
4426 : RelOptInfo *onerel;
4427 :
4428 : /* Make sure we don't return dangling pointers in vardata */
4429 1153566 : MemSet(vardata, 0, sizeof(VariableStatData));
4430 :
4431 : /* Save the exposed type of the expression */
4432 1153566 : vardata->vartype = exprType(node);
4433 :
4434 : /* Look inside any binary-compatible relabeling */
4435 :
4436 1153566 : if (IsA(node, RelabelType))
4437 15272 : basenode = (Node *) ((RelabelType *) node)->arg;
4438 : else
4439 1138294 : basenode = node;
4440 :
4441 : /* Fast path for a simple Var */
4442 :
4443 1153566 : if (IsA(basenode, Var) &&
4444 284490 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
4445 : {
4446 777346 : Var *var = (Var *) basenode;
4447 :
4448 : /* Set up result fields other than the stats tuple */
4449 777346 : vardata->var = basenode; /* return Var without relabeling */
4450 777346 : vardata->rel = find_base_rel(root, var->varno);
4451 777346 : vardata->atttype = var->vartype;
4452 777346 : vardata->atttypmod = var->vartypmod;
4453 777346 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
4454 :
4455 : /* Try to locate some stats */
4456 777346 : examine_simple_variable(root, var, vardata);
4457 :
4458 777346 : return;
4459 : }
4460 :
4461 : /*
4462 : * Okay, it's a more complicated expression. Determine variable
4463 : * membership. Note that when varRelid isn't zero, only vars of that
4464 : * relation are considered "real" vars.
4465 : */
4466 376220 : varnos = pull_varnos(basenode);
4467 :
4468 376220 : onerel = NULL;
4469 :
4470 376220 : switch (bms_membership(varnos))
4471 : {
4472 : case BMS_EMPTY_SET:
4473 : /* No Vars at all ... must be pseudo-constant clause */
4474 220362 : break;
4475 : case BMS_SINGLETON:
4476 153054 : if (varRelid == 0 || bms_is_member(varRelid, varnos))
4477 : {
4478 13574 : onerel = find_base_rel(root,
4479 : (varRelid ? varRelid : bms_singleton_member(varnos)));
4480 13574 : vardata->rel = onerel;
4481 13574 : node = basenode; /* strip any relabeling */
4482 : }
4483 : /* else treat it as a constant */
4484 153054 : break;
4485 : case BMS_MULTIPLE:
4486 2804 : if (varRelid == 0)
4487 : {
4488 : /* treat it as a variable of a join relation */
4489 2692 : vardata->rel = find_join_rel(root, varnos);
4490 2692 : node = basenode; /* strip any relabeling */
4491 : }
4492 112 : else if (bms_is_member(varRelid, varnos))
4493 : {
4494 : /* ignore the vars belonging to other relations */
4495 20 : vardata->rel = find_base_rel(root, varRelid);
4496 20 : node = basenode; /* strip any relabeling */
4497 : /* note: no point in expressional-index search here */
4498 : }
4499 : /* else treat it as a constant */
4500 2804 : break;
4501 : }
4502 :
4503 376220 : bms_free(varnos);
4504 :
4505 376220 : vardata->var = node;
4506 376220 : vardata->atttype = exprType(node);
4507 376220 : vardata->atttypmod = exprTypmod(node);
4508 :
4509 376220 : if (onerel)
4510 : {
4511 : /*
4512 : * We have an expression in vars of a single relation. Try to match
4513 : * it to expressional index columns, in hopes of finding some
4514 : * statistics.
4515 : *
4516 : * Note that we consider all index columns including INCLUDE columns,
4517 : * since there could be stats for such columns. But the test for
4518 : * uniqueness needs to be warier.
4519 : *
4520 : * XXX it's conceivable that there are multiple matches with different
4521 : * index opfamilies; if so, we need to pick one that matches the
4522 : * operator we are estimating for. FIXME later.
4523 : */
4524 : ListCell *ilist;
4525 :
4526 21970 : foreach(ilist, onerel->indexlist)
4527 : {
4528 10256 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
4529 : ListCell *indexpr_item;
4530 : int pos;
4531 :
4532 10256 : indexpr_item = list_head(index->indexprs);
4533 10256 : if (indexpr_item == NULL)
4534 7276 : continue; /* no expressions here... */
4535 :
4536 4100 : for (pos = 0; pos < index->ncolumns; pos++)
4537 : {
4538 2980 : if (index->indexkeys[pos] == 0)
4539 : {
4540 : Node *indexkey;
4541 :
4542 2980 : if (indexpr_item == NULL)
4543 0 : elog(ERROR, "too few entries in indexprs list");
4544 2980 : indexkey = (Node *) lfirst(indexpr_item);
4545 2980 : if (indexkey && IsA(indexkey, RelabelType))
4546 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
4547 2980 : if (equal(node, indexkey))
4548 : {
4549 : /*
4550 : * Found a match ... is it a unique index? Tests here
4551 : * should match has_unique_index().
4552 : */
4553 2444 : if (index->unique &&
4554 520 : index->nkeycolumns == 1 &&
4555 260 : pos == 0 &&
4556 260 : (index->indpred == NIL || index->predOK))
4557 260 : vardata->isunique = true;
4558 :
4559 : /*
4560 : * Has it got stats? We only consider stats for
4561 : * non-partial indexes, since partial indexes probably
4562 : * don't reflect whole-relation statistics; the above
4563 : * check for uniqueness is the only info we take from
4564 : * a partial index.
4565 : *
4566 : * An index stats hook, however, must make its own
4567 : * decisions about what to do with partial indexes.
4568 : */
4569 2184 : if (get_index_stats_hook &&
4570 0 : (*get_index_stats_hook) (root, index->indexoid,
4571 0 : pos + 1, vardata))
4572 : {
4573 : /*
4574 : * The hook took control of acquiring a stats
4575 : * tuple. If it did supply a tuple, it'd better
4576 : * have supplied a freefunc.
4577 : */
4578 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
4579 0 : !vardata->freefunc)
4580 0 : elog(ERROR, "no function provided to release variable stats with");
4581 : }
4582 2184 : else if (index->indpred == NIL)
4583 : {
4584 2184 : vardata->statsTuple =
4585 4368 : SearchSysCache3(STATRELATTINH,
4586 2184 : ObjectIdGetDatum(index->indexoid),
4587 2184 : Int16GetDatum(pos + 1),
4588 : BoolGetDatum(false));
4589 2184 : vardata->freefunc = ReleaseSysCache;
4590 :
4591 2184 : if (HeapTupleIsValid(vardata->statsTuple))
4592 : {
4593 : /* Get index's table for permission check */
4594 : RangeTblEntry *rte;
4595 : Oid userid;
4596 :
4597 1860 : rte = planner_rt_fetch(index->rel->relid, root);
4598 : Assert(rte->rtekind == RTE_RELATION);
4599 :
4600 : /*
4601 : * Use checkAsUser if it's set, in case we're
4602 : * accessing the table via a view.
4603 : */
4604 1860 : userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
4605 :
4606 : /*
4607 : * For simplicity, we insist on the whole
4608 : * table being selectable, rather than trying
4609 : * to identify which column(s) the index
4610 : * depends on. Also require all rows to be
4611 : * selectable --- there must be no
4612 : * securityQuals from security barrier views
4613 : * or RLS policies.
4614 : */
4615 1860 : vardata->acl_ok =
4616 3720 : rte->securityQuals == NIL &&
4617 1860 : (pg_class_aclcheck(rte->relid, userid,
4618 : ACL_SELECT) == ACLCHECK_OK);
4619 : }
4620 : else
4621 : {
4622 : /* suppress leakproofness checks later */
4623 324 : vardata->acl_ok = true;
4624 : }
4625 : }
4626 2184 : if (vardata->statsTuple)
4627 1860 : break;
4628 : }
4629 1120 : indexpr_item = lnext(index->indexprs, indexpr_item);
4630 : }
4631 : }
4632 2980 : if (vardata->statsTuple)
4633 1860 : break;
4634 : }
4635 : }
4636 : }
4637 :
4638 : /*
4639 : * examine_simple_variable
4640 : * Handle a simple Var for examine_variable
4641 : *
4642 : * This is split out as a subroutine so that we can recurse to deal with
4643 : * Vars referencing subqueries.
4644 : *
4645 : * We already filled in all the fields of *vardata except for the stats tuple.
4646 : */
4647 : static void
4648 777894 : examine_simple_variable(PlannerInfo *root, Var *var,
4649 : VariableStatData *vardata)
4650 : {
4651 777894 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
4652 :
4653 : Assert(IsA(rte, RangeTblEntry));
4654 :
4655 777894 : if (get_relation_stats_hook &&
4656 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
4657 : {
4658 : /*
4659 : * The hook took control of acquiring a stats tuple. If it did supply
4660 : * a tuple, it'd better have supplied a freefunc.
4661 : */
4662 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
4663 0 : !vardata->freefunc)
4664 0 : elog(ERROR, "no function provided to release variable stats with");
4665 : }
4666 777894 : else if (rte->rtekind == RTE_RELATION)
4667 : {
4668 : /*
4669 : * Plain table or parent of an inheritance appendrel, so look up the
4670 : * column in pg_statistic
4671 : */
4672 2196474 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
4673 732158 : ObjectIdGetDatum(rte->relid),
4674 732158 : Int16GetDatum(var->varattno),
4675 732158 : BoolGetDatum(rte->inh));
4676 732158 : vardata->freefunc = ReleaseSysCache;
4677 :
4678 732158 : if (HeapTupleIsValid(vardata->statsTuple))
4679 : {
4680 : Oid userid;
4681 :
4682 : /*
4683 : * Check if user has permission to read this column. We require
4684 : * all rows to be accessible, so there must be no securityQuals
4685 : * from security barrier views or RLS policies. Use checkAsUser
4686 : * if it's set, in case we're accessing the table via a view.
4687 : */
4688 521942 : userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
4689 :
4690 521942 : vardata->acl_ok =
4691 1565206 : rte->securityQuals == NIL &&
4692 521642 : ((pg_class_aclcheck(rte->relid, userid,
4693 120 : ACL_SELECT) == ACLCHECK_OK) ||
4694 120 : (pg_attribute_aclcheck(rte->relid, var->varattno, userid,
4695 : ACL_SELECT) == ACLCHECK_OK));
4696 : }
4697 : else
4698 : {
4699 : /* suppress any possible leakproofness checks later */
4700 210216 : vardata->acl_ok = true;
4701 : }
4702 : }
4703 45736 : else if (rte->rtekind == RTE_SUBQUERY && !rte->inh)
4704 : {
4705 : /*
4706 : * Plain subquery (not one that was converted to an appendrel).
4707 : */
4708 2216 : Query *subquery = rte->subquery;
4709 : RelOptInfo *rel;
4710 : TargetEntry *ste;
4711 :
4712 : /*
4713 : * Punt if it's a whole-row var rather than a plain column reference.
4714 : */
4715 2216 : if (var->varattno == InvalidAttrNumber)
4716 0 : return;
4717 :
4718 : /*
4719 : * Punt if subquery uses set operations or GROUP BY, as these will
4720 : * mash underlying columns' stats beyond recognition. (Set ops are
4721 : * particularly nasty; if we forged ahead, we would return stats
4722 : * relevant to only the leftmost subselect...) DISTINCT is also
4723 : * problematic, but we check that later because there is a possibility
4724 : * of learning something even with it.
4725 : */
4726 4316 : if (subquery->setOperations ||
4727 2100 : subquery->groupClause)
4728 364 : return;
4729 :
4730 : /*
4731 : * OK, fetch RelOptInfo for subquery. Note that we don't change the
4732 : * rel returned in vardata, since caller expects it to be a rel of the
4733 : * caller's query level. Because we might already be recursing, we
4734 : * can't use that rel pointer either, but have to look up the Var's
4735 : * rel afresh.
4736 : */
4737 1852 : rel = find_base_rel(root, var->varno);
4738 :
4739 : /* If the subquery hasn't been planned yet, we have to punt */
4740 1852 : if (rel->subroot == NULL)
4741 0 : return;
4742 : Assert(IsA(rel->subroot, PlannerInfo));
4743 :
4744 : /*
4745 : * Switch our attention to the subquery as mangled by the planner. It
4746 : * was okay to look at the pre-planning version for the tests above,
4747 : * but now we need a Var that will refer to the subroot's live
4748 : * RelOptInfos. For instance, if any subquery pullup happened during
4749 : * planning, Vars in the targetlist might have gotten replaced, and we
4750 : * need to see the replacement expressions.
4751 : */
4752 1852 : subquery = rel->subroot->parse;
4753 : Assert(IsA(subquery, Query));
4754 :
4755 : /* Get the subquery output expression referenced by the upper Var */
4756 1852 : ste = get_tle_by_resno(subquery->targetList, var->varattno);
4757 1852 : if (ste == NULL || ste->resjunk)
4758 0 : elog(ERROR, "subquery %s does not have attribute %d",
4759 : rte->eref->aliasname, var->varattno);
4760 1852 : var = (Var *) ste->expr;
4761 :
4762 : /*
4763 : * If subquery uses DISTINCT, we can't make use of any stats for the
4764 : * variable ... but, if it's the only DISTINCT column, we are entitled
4765 : * to consider it unique. We do the test this way so that it works
4766 : * for cases involving DISTINCT ON.
4767 : */
4768 1852 : if (subquery->distinctClause)
4769 : {
4770 200 : if (list_length(subquery->distinctClause) == 1 &&
4771 68 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
4772 68 : vardata->isunique = true;
4773 : /* cannot go further */
4774 132 : return;
4775 : }
4776 :
4777 : /*
4778 : * If the sub-query originated from a view with the security_barrier
4779 : * attribute, we must not look at the variable's statistics, though it
4780 : * seems all right to notice the existence of a DISTINCT clause. So
4781 : * stop here.
4782 : *
4783 : * This is probably a harsher restriction than necessary; it's
4784 : * certainly OK for the selectivity estimator (which is a C function,
4785 : * and therefore omnipotent anyway) to look at the statistics. But
4786 : * many selectivity estimators will happily *invoke the operator
4787 : * function* to try to work out a good estimate - and that's not OK.
4788 : * So for now, don't dig down for stats.
4789 : */
4790 1720 : if (rte->security_barrier)
4791 140 : return;
4792 :
4793 : /* Can only handle a simple Var of subquery's query level */
4794 2128 : if (var && IsA(var, Var) &&
4795 548 : var->varlevelsup == 0)
4796 : {
4797 : /*
4798 : * OK, recurse into the subquery. Note that the original setting
4799 : * of vardata->isunique (which will surely be false) is left
4800 : * unchanged in this situation. That's what we want, since even
4801 : * if the underlying column is unique, the subquery may have
4802 : * joined to other tables in a way that creates duplicates.
4803 : */
4804 548 : examine_simple_variable(rel->subroot, var, vardata);
4805 : }
4806 : }
4807 : else
4808 : {
4809 : /*
4810 : * Otherwise, the Var comes from a FUNCTION, VALUES, or CTE RTE. (We
4811 : * won't see RTE_JOIN here because join alias Vars have already been
4812 : * flattened.) There's not much we can do with function outputs, but
4813 : * maybe someday try to be smarter about VALUES and/or CTEs.
4814 : */
4815 : }
4816 : }
4817 :
4818 : /*
4819 : * Check whether it is permitted to call func_oid passing some of the
4820 : * pg_statistic data in vardata. We allow this either if the user has SELECT
4821 : * privileges on the table or column underlying the pg_statistic data or if
4822 : * the function is marked leak-proof.
4823 : */
4824 : bool
4825 601232 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
4826 : {
4827 601232 : if (vardata->acl_ok)
4828 600840 : return true;
4829 :
4830 392 : if (!OidIsValid(func_oid))
4831 0 : return false;
4832 :
4833 392 : if (get_func_leakproof(func_oid))
4834 288 : return true;
4835 :
4836 104 : ereport(DEBUG2,
4837 : (errmsg_internal("not using statistics because function \"%s\" is not leak-proof",
4838 : get_func_name(func_oid))));
4839 104 : return false;
4840 : }
4841 :
4842 : /*
4843 : * get_variable_numdistinct
4844 : * Estimate the number of distinct values of a variable.
4845 : *
4846 : * vardata: results of examine_variable
4847 : * *isdefault: set to true if the result is a default rather than based on
4848 : * anything meaningful.
4849 : *
4850 : * NB: be careful to produce a positive integral result, since callers may
4851 : * compare the result to exact integer counts, or might divide by it.
4852 : */
4853 : double
4854 495168 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
4855 : {
4856 : double stadistinct;
4857 495168 : double stanullfrac = 0.0;
4858 : double ntuples;
4859 :
4860 495168 : *isdefault = false;
4861 :
4862 : /*
4863 : * Determine the stadistinct value to use. There are cases where we can
4864 : * get an estimate even without a pg_statistic entry, or can get a better
4865 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
4866 : * (otherwise, assume no nulls, for lack of any better idea).
4867 : */
4868 495168 : if (HeapTupleIsValid(vardata->statsTuple))
4869 : {
4870 : /* Use the pg_statistic entry */
4871 : Form_pg_statistic stats;
4872 :
4873 315260 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
4874 315260 : stadistinct = stats->stadistinct;
4875 315260 : stanullfrac = stats->stanullfrac;
4876 : }
4877 179908 : else if (vardata->vartype == BOOLOID)
4878 : {
4879 : /*
4880 : * Special-case boolean columns: presumably, two distinct values.
4881 : *
4882 : * Are there any other datatypes we should wire in special estimates
4883 : * for?
4884 : */
4885 140 : stadistinct = 2.0;
4886 : }
4887 179768 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
4888 : {
4889 : /*
4890 : * If the Var represents a column of a VALUES RTE, assume it's unique.
4891 : * This could of course be very wrong, but it should tend to be true
4892 : * in well-written queries. We could consider examining the VALUES'
4893 : * contents to get some real statistics; but that only works if the
4894 : * entries are all constants, and it would be pretty expensive anyway.
4895 : */
4896 702 : stadistinct = -1.0; /* unique (and all non null) */
4897 : }
4898 : else
4899 : {
4900 : /*
4901 : * We don't keep statistics for system columns, but in some cases we
4902 : * can infer distinctness anyway.
4903 : */
4904 179066 : if (vardata->var && IsA(vardata->var, Var))
4905 : {
4906 173748 : switch (((Var *) vardata->var)->varattno)
4907 : {
4908 : case SelfItemPointerAttributeNumber:
4909 474 : stadistinct = -1.0; /* unique (and all non null) */
4910 474 : break;
4911 : case TableOidAttributeNumber:
4912 1942 : stadistinct = 1.0; /* only 1 value */
4913 1942 : break;
4914 : default:
4915 171332 : stadistinct = 0.0; /* means "unknown" */
4916 171332 : break;
4917 : }
4918 173748 : }
4919 : else
4920 5318 : stadistinct = 0.0; /* means "unknown" */
4921 :
4922 : /*
4923 : * XXX consider using estimate_num_groups on expressions?
4924 : */
4925 : }
4926 :
4927 : /*
4928 : * If there is a unique index or DISTINCT clause for the variable, assume
4929 : * it is unique no matter what pg_statistic says; the statistics could be
4930 : * out of date, or we might have found a partial unique index that proves
4931 : * the var is unique for this query. However, we'd better still believe
4932 : * the null-fraction statistic.
4933 : */
4934 495168 : if (vardata->isunique)
4935 112868 : stadistinct = -1.0 * (1.0 - stanullfrac);
4936 :
4937 : /*
4938 : * If we had an absolute estimate, use that.
4939 : */
4940 495168 : if (stadistinct > 0.0)
4941 81706 : return clamp_row_est(stadistinct);
4942 :
4943 : /*
4944 : * Otherwise we need to get the relation size; punt if not available.
4945 : */
4946 413462 : if (vardata->rel == NULL)
4947 : {
4948 124 : *isdefault = true;
4949 124 : return DEFAULT_NUM_DISTINCT;
4950 : }
4951 413338 : ntuples = vardata->rel->tuples;
4952 413338 : if (ntuples <= 0.0)
4953 : {
4954 1288 : *isdefault = true;
4955 1288 : return DEFAULT_NUM_DISTINCT;
4956 : }
4957 :
4958 : /*
4959 : * If we had a relative estimate, use that.
4960 : */
4961 412050 : if (stadistinct < 0.0)
4962 260390 : return clamp_row_est(-stadistinct * ntuples);
4963 :
4964 : /*
4965 : * With no data, estimate ndistinct = ntuples if the table is small, else
4966 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
4967 : * that the behavior isn't discontinuous.
4968 : */
4969 151660 : if (ntuples < DEFAULT_NUM_DISTINCT)
4970 52910 : return clamp_row_est(ntuples);
4971 :
4972 98750 : *isdefault = true;
4973 98750 : return DEFAULT_NUM_DISTINCT;
4974 : }
4975 :
4976 : /*
4977 : * get_variable_range
4978 : * Estimate the minimum and maximum value of the specified variable.
4979 : * If successful, store values in *min and *max, and return true.
4980 : * If no data available, return false.
4981 : *
4982 : * sortop is the "<" comparison operator to use. This should generally
4983 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
4984 : */
4985 : static bool
4986 88978 : get_variable_range(PlannerInfo *root, VariableStatData *vardata, Oid sortop,
4987 : Datum *min, Datum *max)
4988 : {
4989 88978 : Datum tmin = 0;
4990 88978 : Datum tmax = 0;
4991 88978 : bool have_data = false;
4992 : int16 typLen;
4993 : bool typByVal;
4994 : Oid opfuncoid;
4995 : AttStatsSlot sslot;
4996 : int i;
4997 :
4998 : /*
4999 : * XXX It's very tempting to try to use the actual column min and max, if
5000 : * we can get them relatively-cheaply with an index probe. However, since
5001 : * this function is called many times during join planning, that could
5002 : * have unpleasant effects on planning speed. Need more investigation
5003 : * before enabling this.
5004 : */
5005 : #ifdef NOT_USED
5006 : if (get_actual_variable_range(root, vardata, sortop, min, max))
5007 : return true;
5008 : #endif
5009 :
5010 88978 : if (!HeapTupleIsValid(vardata->statsTuple))
5011 : {
5012 : /* no stats available, so default result */
5013 18372 : return false;
5014 : }
5015 :
5016 : /*
5017 : * If we can't apply the sortop to the stats data, just fail. In
5018 : * principle, if there's a histogram and no MCVs, we could return the
5019 : * histogram endpoints without ever applying the sortop ... but it's
5020 : * probably not worth trying, because whatever the caller wants to do with
5021 : * the endpoints would likely fail the security check too.
5022 : */
5023 70606 : if (!statistic_proc_security_check(vardata,
5024 : (opfuncoid = get_opcode(sortop))))
5025 0 : return false;
5026 :
5027 70606 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
5028 :
5029 : /*
5030 : * If there is a histogram, grab the first and last values.
5031 : *
5032 : * If there is a histogram that is sorted with some other operator than
5033 : * the one we want, fail --- this suggests that there is data we can't
5034 : * use.
5035 : */
5036 70606 : if (get_attstatsslot(&sslot, vardata->statsTuple,
5037 : STATISTIC_KIND_HISTOGRAM, sortop,
5038 : ATTSTATSSLOT_VALUES))
5039 : {
5040 60946 : if (sslot.nvalues > 0)
5041 : {
5042 60946 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
5043 60946 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
5044 60946 : have_data = true;
5045 : }
5046 60946 : free_attstatsslot(&sslot);
5047 : }
5048 9660 : else if (get_attstatsslot(&sslot, vardata->statsTuple,
5049 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
5050 : 0))
5051 : {
5052 0 : free_attstatsslot(&sslot);
5053 0 : return false;
5054 : }
5055 :
5056 : /*
5057 : * If we have most-common-values info, look for extreme MCVs. This is
5058 : * needed even if we also have a histogram, since the histogram excludes
5059 : * the MCVs. However, usually the MCVs will not be the extreme values, so
5060 : * avoid unnecessary data copying.
5061 : */
5062 70606 : if (get_attstatsslot(&sslot, vardata->statsTuple,
5063 : STATISTIC_KIND_MCV, InvalidOid,
5064 : ATTSTATSSLOT_VALUES))
5065 : {
5066 33774 : bool tmin_is_mcv = false;
5067 33774 : bool tmax_is_mcv = false;
5068 : FmgrInfo opproc;
5069 :
5070 33774 : fmgr_info(opfuncoid, &opproc);
5071 :
5072 1032698 : for (i = 0; i < sslot.nvalues; i++)
5073 : {
5074 998924 : if (!have_data)
5075 : {
5076 9322 : tmin = tmax = sslot.values[i];
5077 9322 : tmin_is_mcv = tmax_is_mcv = have_data = true;
5078 9322 : continue;
5079 : }
5080 989602 : if (DatumGetBool(FunctionCall2Coll(&opproc,
5081 : sslot.stacoll,
5082 : sslot.values[i], tmin)))
5083 : {
5084 35636 : tmin = sslot.values[i];
5085 35636 : tmin_is_mcv = true;
5086 : }
5087 989602 : if (DatumGetBool(FunctionCall2Coll(&opproc,
5088 : sslot.stacoll,
5089 : tmax, sslot.values[i])))
5090 : {
5091 32004 : tmax = sslot.values[i];
5092 32004 : tmax_is_mcv = true;
5093 : }
5094 : }
5095 33774 : if (tmin_is_mcv)
5096 28316 : tmin = datumCopy(tmin, typByVal, typLen);
5097 33774 : if (tmax_is_mcv)
5098 13146 : tmax = datumCopy(tmax, typByVal, typLen);
5099 33774 : free_attstatsslot(&sslot);
5100 : }
5101 :
5102 70606 : *min = tmin;
5103 70606 : *max = tmax;
5104 70606 : return have_data;
5105 : }
5106 :
5107 :
5108 : /*
5109 : * get_actual_variable_range
5110 : * Attempt to identify the current *actual* minimum and/or maximum
5111 : * of the specified variable, by looking for a suitable btree index
5112 : * and fetching its low and/or high values.
5113 : * If successful, store values in *min and *max, and return true.
5114 : * (Either pointer can be NULL if that endpoint isn't needed.)
5115 : * If no data available, return false.
5116 : *
5117 : * sortop is the "<" comparison operator to use.
5118 : */
5119 : static bool
5120 109510 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
5121 : Oid sortop,
5122 : Datum *min, Datum *max)
5123 : {
5124 109510 : bool have_data = false;
5125 109510 : RelOptInfo *rel = vardata->rel;
5126 : RangeTblEntry *rte;
5127 : ListCell *lc;
5128 :
5129 : /* No hope if no relation or it doesn't have indexes */
5130 109510 : if (rel == NULL || rel->indexlist == NIL)
5131 4518 : return false;
5132 : /* If it has indexes it must be a plain relation */
5133 104992 : rte = root->simple_rte_array[rel->relid];
5134 : Assert(rte->rtekind == RTE_RELATION);
5135 :
5136 : /* Search through the indexes to see if any match our problem */
5137 239470 : foreach(lc, rel->indexlist)
5138 : {
5139 200010 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
5140 : ScanDirection indexscandir;
5141 :
5142 : /* Ignore non-btree indexes */
5143 200010 : if (index->relam != BTREE_AM_OID)
5144 0 : continue;
5145 :
5146 : /*
5147 : * Ignore partial indexes --- we only want stats that cover the entire
5148 : * relation.
5149 : */
5150 200010 : if (index->indpred != NIL)
5151 68 : continue;
5152 :
5153 : /*
5154 : * The index list might include hypothetical indexes inserted by a
5155 : * get_relation_info hook --- don't try to access them.
5156 : */
5157 199942 : if (index->hypothetical)
5158 0 : continue;
5159 :
5160 : /*
5161 : * The first index column must match the desired variable and sort
5162 : * operator --- but we can use a descending-order index.
5163 : */
5164 199942 : if (!match_index_to_operand(vardata->var, 0, index))
5165 134410 : continue;
5166 65532 : switch (get_op_opfamily_strategy(sortop, index->sortopfamily[0]))
5167 : {
5168 : case BTLessStrategyNumber:
5169 65532 : if (index->reverse_sort[0])
5170 0 : indexscandir = BackwardScanDirection;
5171 : else
5172 65532 : indexscandir = ForwardScanDirection;
5173 65532 : break;
5174 : case BTGreaterStrategyNumber:
5175 0 : if (index->reverse_sort[0])
5176 0 : indexscandir = ForwardScanDirection;
5177 : else
5178 0 : indexscandir = BackwardScanDirection;
5179 0 : break;
5180 : default:
5181 : /* index doesn't match the sortop */
5182 0 : continue;
5183 : }
5184 :
5185 : /*
5186 : * Found a suitable index to extract data from. Set up some data that
5187 : * can be used by both invocations of get_actual_variable_endpoint.
5188 : */
5189 : {
5190 : MemoryContext tmpcontext;
5191 : MemoryContext oldcontext;
5192 : Relation heapRel;
5193 : Relation indexRel;
5194 : TupleTableSlot *slot;
5195 : int16 typLen;
5196 : bool typByVal;
5197 : ScanKeyData scankeys[1];
5198 :
5199 : /* Make sure any cruft gets recycled when we're done */
5200 65532 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
5201 : "get_actual_variable_range workspace",
5202 : ALLOCSET_DEFAULT_SIZES);
5203 65532 : oldcontext = MemoryContextSwitchTo(tmpcontext);
5204 :
5205 : /*
5206 : * Open the table and index so we can read from them. We should
5207 : * already have some type of lock on each.
5208 : */
5209 65532 : heapRel = table_open(rte->relid, NoLock);
5210 65532 : indexRel = index_open(index->indexoid, NoLock);
5211 :
5212 : /* build some stuff needed for indexscan execution */
5213 65532 : slot = table_slot_create(heapRel, NULL);
5214 65532 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
5215 :
5216 : /* set up an IS NOT NULL scan key so that we ignore nulls */
5217 65532 : ScanKeyEntryInitialize(&scankeys[0],
5218 : SK_ISNULL | SK_SEARCHNOTNULL,
5219 : 1, /* index col to scan */
5220 : InvalidStrategy, /* no strategy */
5221 : InvalidOid, /* no strategy subtype */
5222 : InvalidOid, /* no collation */
5223 : InvalidOid, /* no reg proc for this */
5224 : (Datum) 0); /* constant */
5225 :
5226 : /* If min is requested ... */
5227 65532 : if (min)
5228 : {
5229 38292 : have_data = get_actual_variable_endpoint(heapRel,
5230 : indexRel,
5231 : indexscandir,
5232 : scankeys,
5233 : typLen,
5234 : typByVal,
5235 : slot,
5236 : oldcontext,
5237 : min);
5238 : }
5239 : else
5240 : {
5241 : /* If min not requested, assume index is nonempty */
5242 27240 : have_data = true;
5243 : }
5244 :
5245 : /* If max is requested, and we didn't find the index is empty */
5246 65532 : if (max && have_data)
5247 : {
5248 : /* scan in the opposite direction; all else is the same */
5249 55056 : have_data = get_actual_variable_endpoint(heapRel,
5250 : indexRel,
5251 27528 : -indexscandir,
5252 : scankeys,
5253 : typLen,
5254 : typByVal,
5255 : slot,
5256 : oldcontext,
5257 : max);
5258 : }
5259 :
5260 : /* Clean everything up */
5261 65532 : ExecDropSingleTupleTableSlot(slot);
5262 :
5263 65532 : index_close(indexRel, NoLock);
5264 65532 : table_close(heapRel, NoLock);
5265 :
5266 65532 : MemoryContextSwitchTo(oldcontext);
5267 65532 : MemoryContextDelete(tmpcontext);
5268 :
5269 : /* And we're done */
5270 65532 : break;
5271 : }
5272 : }
5273 :
5274 104992 : return have_data;
5275 : }
5276 :
5277 : /*
5278 : * Get one endpoint datum (min or max depending on indexscandir) from the
5279 : * specified index. Return true if successful, false if index is empty.
5280 : * On success, endpoint value is stored to *endpointDatum (and copied into
5281 : * outercontext).
5282 : *
5283 : * scankeys is a 1-element scankey array set up to reject nulls.
5284 : * typLen/typByVal describe the datatype of the index's first column.
5285 : * tableslot is a slot suitable to hold table tuples, in case we need
5286 : * to probe the heap.
5287 : * (We could compute these values locally, but that would mean computing them
5288 : * twice when get_actual_variable_range needs both the min and the max.)
5289 : */
5290 : static bool
5291 65820 : get_actual_variable_endpoint(Relation heapRel,
5292 : Relation indexRel,
5293 : ScanDirection indexscandir,
5294 : ScanKey scankeys,
5295 : int16 typLen,
5296 : bool typByVal,
5297 : TupleTableSlot *tableslot,
5298 : MemoryContext outercontext,
5299 : Datum *endpointDatum)
5300 : {
5301 65820 : bool have_data = false;
5302 : SnapshotData SnapshotNonVacuumable;
5303 : IndexScanDesc index_scan;
5304 65820 : Buffer vmbuffer = InvalidBuffer;
5305 : ItemPointer tid;
5306 : Datum values[INDEX_MAX_KEYS];
5307 : bool isnull[INDEX_MAX_KEYS];
5308 : MemoryContext oldcontext;
5309 :
5310 : /*
5311 : * We use the index-only-scan machinery for this. With mostly-static
5312 : * tables that's a win because it avoids a heap visit. It's also a win
5313 : * for dynamic data, but the reason is less obvious; read on for details.
5314 : *
5315 : * In principle, we should scan the index with our current active
5316 : * snapshot, which is the best approximation we've got to what the query
5317 : * will see when executed. But that won't be exact if a new snap is taken
5318 : * before running the query, and it can be very expensive if a lot of
5319 : * recently-dead or uncommitted rows exist at the beginning or end of the
5320 : * index (because we'll laboriously fetch each one and reject it).
5321 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
5322 : * and uncommitted rows as well as normal visible rows. On the other
5323 : * hand, it will reject known-dead rows, and thus not give a bogus answer
5324 : * when the extreme value has been deleted (unless the deletion was quite
5325 : * recent); that case motivates not using SnapshotAny here.
5326 : *
5327 : * A crucial point here is that SnapshotNonVacuumable, with
5328 : * RecentGlobalXmin as horizon, yields the inverse of the condition that
5329 : * the indexscan will use to decide that index entries are killable (see
5330 : * heap_hot_search_buffer()). Therefore, if the snapshot rejects a tuple
5331 : * (or more precisely, all tuples of a HOT chain) and we have to continue
5332 : * scanning past it, we know that the indexscan will mark that index entry
5333 : * killed. That means that the next get_actual_variable_endpoint() call
5334 : * will not have to re-consider that index entry. In this way we avoid
5335 : * repetitive work when this function is used a lot during planning.
5336 : *
5337 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
5338 : * recently-created index, some index entries may point at "broken" HOT
5339 : * chains in which not all the tuple versions contain data matching the
5340 : * index entry. The live tuple version(s) certainly do match the index,
5341 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
5342 : * don't match. Hence, if we took data from the selected heap tuple, we
5343 : * might get a bogus answer that's not close to the index extremal value,
5344 : * or could even be NULL. We avoid this hazard because we take the data
5345 : * from the index entry not the heap.
5346 : */
5347 65820 : InitNonVacuumableSnapshot(SnapshotNonVacuumable, RecentGlobalXmin);
5348 :
5349 65820 : index_scan = index_beginscan(heapRel, indexRel,
5350 : &SnapshotNonVacuumable,
5351 : 1, 0);
5352 : /* Set it up for index-only scan */
5353 65820 : index_scan->xs_want_itup = true;
5354 65820 : index_rescan(index_scan, scankeys, 1, NULL, 0);
5355 :
5356 : /* Fetch first/next tuple in specified direction */
5357 65820 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
5358 : {
5359 70012 : if (!VM_ALL_VISIBLE(heapRel,
5360 : ItemPointerGetBlockNumber(tid),
5361 : &vmbuffer))
5362 : {
5363 : /* Rats, we have to visit the heap to check visibility */
5364 39366 : if (!index_fetch_heap(index_scan, tableslot))
5365 4192 : continue; /* no visible tuple, try next index entry */
5366 :
5367 : /* We don't actually need the heap tuple for anything */
5368 35174 : ExecClearTuple(tableslot);
5369 :
5370 : /*
5371 : * We don't care whether there's more than one visible tuple in
5372 : * the HOT chain; if any are visible, that's good enough.
5373 : */
5374 : }
5375 :
5376 : /*
5377 : * We expect that btree will return data in IndexTuple not HeapTuple
5378 : * format. It's not lossy either.
5379 : */
5380 65820 : if (!index_scan->xs_itup)
5381 0 : elog(ERROR, "no data returned for index-only scan");
5382 65820 : if (index_scan->xs_recheck)
5383 0 : elog(ERROR, "unexpected recheck indication from btree");
5384 :
5385 : /* OK to deconstruct the index tuple */
5386 65820 : index_deform_tuple(index_scan->xs_itup,
5387 : index_scan->xs_itupdesc,
5388 : values, isnull);
5389 :
5390 : /* Shouldn't have got a null, but be careful */
5391 65820 : if (isnull[0])
5392 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
5393 : RelationGetRelationName(indexRel));
5394 :
5395 : /* Copy the index column value out to caller's context */
5396 65820 : oldcontext = MemoryContextSwitchTo(outercontext);
5397 65820 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
5398 65820 : MemoryContextSwitchTo(oldcontext);
5399 65820 : have_data = true;
5400 65820 : break;
5401 : }
5402 :
5403 65820 : if (vmbuffer != InvalidBuffer)
5404 59064 : ReleaseBuffer(vmbuffer);
5405 65820 : index_endscan(index_scan);
5406 :
5407 65820 : return have_data;
5408 : }
5409 :
5410 : /*
5411 : * find_join_input_rel
5412 : * Look up the input relation for a join.
5413 : *
5414 : * We assume that the input relation's RelOptInfo must have been constructed
5415 : * already.
5416 : */
5417 : static RelOptInfo *
5418 9128 : find_join_input_rel(PlannerInfo *root, Relids relids)
5419 : {
5420 9128 : RelOptInfo *rel = NULL;
5421 :
5422 9128 : switch (bms_membership(relids))
5423 : {
5424 : case BMS_EMPTY_SET:
5425 : /* should not happen */
5426 0 : break;
5427 : case BMS_SINGLETON:
5428 8976 : rel = find_base_rel(root, bms_singleton_member(relids));
5429 8976 : break;
5430 : case BMS_MULTIPLE:
5431 152 : rel = find_join_rel(root, relids);
5432 152 : break;
5433 : }
5434 :
5435 9128 : if (rel == NULL)
5436 0 : elog(ERROR, "could not find RelOptInfo for given relids");
5437 :
5438 9128 : return rel;
5439 : }
5440 :
5441 :
5442 : /*-------------------------------------------------------------------------
5443 : *
5444 : * Index cost estimation functions
5445 : *
5446 : *-------------------------------------------------------------------------
5447 : */
5448 :
5449 : /*
5450 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
5451 : */
5452 : List *
5453 368942 : get_quals_from_indexclauses(List *indexclauses)
5454 : {
5455 368942 : List *result = NIL;
5456 : ListCell *lc;
5457 :
5458 678050 : foreach(lc, indexclauses)
5459 : {
5460 309108 : IndexClause *iclause = lfirst_node(IndexClause, lc);
5461 : ListCell *lc2;
5462 :
5463 619926 : foreach(lc2, iclause->indexquals)
5464 : {
5465 310818 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
5466 :
5467 310818 : result = lappend(result, rinfo);
5468 : }
5469 : }
5470 368942 : return result;
5471 : }
5472 :
5473 : /*
5474 : * Compute the total evaluation cost of the comparison operands in a list
5475 : * of index qual expressions. Since we know these will be evaluated just
5476 : * once per scan, there's no need to distinguish startup from per-row cost.
5477 : *
5478 : * This can be used either on the result of get_quals_from_indexclauses(),
5479 : * or directly on an indexorderbys list. In both cases, we expect that the
5480 : * index key expression is on the left side of binary clauses.
5481 : */
5482 : Cost
5483 732786 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
5484 : {
5485 732786 : Cost qual_arg_cost = 0;
5486 : ListCell *lc;
5487 :
5488 1043904 : foreach(lc, indexquals)
5489 : {
5490 311118 : Expr *clause = (Expr *) lfirst(lc);
5491 : Node *other_operand;
5492 : QualCost index_qual_cost;
5493 :
5494 : /*
5495 : * Index quals will have RestrictInfos, indexorderbys won't. Look
5496 : * through RestrictInfo if present.
5497 : */
5498 311118 : if (IsA(clause, RestrictInfo))
5499 310810 : clause = ((RestrictInfo *) clause)->clause;
5500 :
5501 311118 : if (IsA(clause, OpExpr))
5502 : {
5503 302056 : OpExpr *op = (OpExpr *) clause;
5504 :
5505 302056 : other_operand = (Node *) lsecond(op->args);
5506 : }
5507 9062 : else if (IsA(clause, RowCompareExpr))
5508 : {
5509 88 : RowCompareExpr *rc = (RowCompareExpr *) clause;
5510 :
5511 88 : other_operand = (Node *) rc->rargs;
5512 : }
5513 8974 : else if (IsA(clause, ScalarArrayOpExpr))
5514 : {
5515 7076 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
5516 :
5517 7076 : other_operand = (Node *) lsecond(saop->args);
5518 : }
5519 1898 : else if (IsA(clause, NullTest))
5520 : {
5521 1898 : other_operand = NULL;
5522 : }
5523 : else
5524 : {
5525 0 : elog(ERROR, "unsupported indexqual type: %d",
5526 : (int) nodeTag(clause));
5527 : other_operand = NULL; /* keep compiler quiet */
5528 : }
5529 :
5530 311118 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
5531 311118 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
5532 : }
5533 732786 : return qual_arg_cost;
5534 : }
5535 :
5536 : void
5537 363852 : genericcostestimate(PlannerInfo *root,
5538 : IndexPath *path,
5539 : double loop_count,
5540 : GenericCosts *costs)
5541 : {
5542 363852 : IndexOptInfo *index = path->indexinfo;
5543 363852 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
5544 363852 : List *indexOrderBys = path->indexorderbys;
5545 : Cost indexStartupCost;
5546 : Cost indexTotalCost;
5547 : Selectivity indexSelectivity;
5548 : double indexCorrelation;
5549 : double numIndexPages;
5550 : double numIndexTuples;
5551 : double spc_random_page_cost;
5552 : double num_sa_scans;
5553 : double num_outer_scans;
5554 : double num_scans;
5555 : double qual_op_cost;
5556 : double qual_arg_cost;
5557 : List *selectivityQuals;
5558 : ListCell *l;
5559 :
5560 : /*
5561 : * If the index is partial, AND the index predicate with the explicitly
5562 : * given indexquals to produce a more accurate idea of the index
5563 : * selectivity.
5564 : */
5565 363852 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
5566 :
5567 : /*
5568 : * Check for ScalarArrayOpExpr index quals, and estimate the number of
5569 : * index scans that will be performed.
5570 : */
5571 363852 : num_sa_scans = 1;
5572 669572 : foreach(l, indexQuals)
5573 : {
5574 305720 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
5575 :
5576 305720 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
5577 : {
5578 7072 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
5579 7072 : int alength = estimate_array_length(lsecond(saop->args));
5580 :
5581 7072 : if (alength > 1)
5582 7072 : num_sa_scans *= alength;
5583 : }
5584 : }
5585 :
5586 : /* Estimate the fraction of main-table tuples that will be visited */
5587 363852 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
5588 363852 : index->rel->relid,
5589 : JOIN_INNER,
5590 : NULL);
5591 :
5592 : /*
5593 : * If caller didn't give us an estimate, estimate the number of index
5594 : * tuples that will be visited. We do it in this rather peculiar-looking
5595 : * way in order to get the right answer for partial indexes.
5596 : */
5597 363852 : numIndexTuples = costs->numIndexTuples;
5598 363852 : if (numIndexTuples <= 0.0)
5599 : {
5600 20466 : numIndexTuples = indexSelectivity * index->rel->tuples;
5601 :
5602 : /*
5603 : * The above calculation counts all the tuples visited across all
5604 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
5605 : * average per-indexscan number, so adjust. This is a handy place to
5606 : * round to integer, too. (If caller supplied tuple estimate, it's
5607 : * responsible for handling these considerations.)
5608 : */
5609 20466 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
5610 : }
5611 :
5612 : /*
5613 : * We can bound the number of tuples by the index size in any case. Also,
5614 : * always estimate at least one tuple is touched, even when
5615 : * indexSelectivity estimate is tiny.
5616 : */
5617 363852 : if (numIndexTuples > index->tuples)
5618 202 : numIndexTuples = index->tuples;
5619 363852 : if (numIndexTuples < 1.0)
5620 17836 : numIndexTuples = 1.0;
5621 :
5622 : /*
5623 : * Estimate the number of index pages that will be retrieved.
5624 : *
5625 : * We use the simplistic method of taking a pro-rata fraction of the total
5626 : * number of index pages. In effect, this counts only leaf pages and not
5627 : * any overhead such as index metapage or upper tree levels.
5628 : *
5629 : * In practice access to upper index levels is often nearly free because
5630 : * those tend to stay in cache under load; moreover, the cost involved is
5631 : * highly dependent on index type. We therefore ignore such costs here
5632 : * and leave it to the caller to add a suitable charge if needed.
5633 : */
5634 363852 : if (index->pages > 1 && index->tuples > 1)
5635 345344 : numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
5636 : else
5637 18508 : numIndexPages = 1.0;
5638 :
5639 : /* fetch estimated page cost for tablespace containing index */
5640 363852 : get_tablespace_page_costs(index->reltablespace,
5641 : &spc_random_page_cost,
5642 : NULL);
5643 :
5644 : /*
5645 : * Now compute the disk access costs.
5646 : *
5647 : * The above calculations are all per-index-scan. However, if we are in a
5648 : * nestloop inner scan, we can expect the scan to be repeated (with
5649 : * different search keys) for each row of the outer relation. Likewise,
5650 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
5651 : * the potential for cache effects to reduce the number of disk page
5652 : * fetches needed. We want to estimate the average per-scan I/O cost in
5653 : * the presence of caching.
5654 : *
5655 : * We use the Mackert-Lohman formula (see costsize.c for details) to
5656 : * estimate the total number of page fetches that occur. While this
5657 : * wasn't what it was designed for, it seems a reasonable model anyway.
5658 : * Note that we are counting pages not tuples anymore, so we take N = T =
5659 : * index size, as if there were one "tuple" per page.
5660 : */
5661 363852 : num_outer_scans = loop_count;
5662 363852 : num_scans = num_sa_scans * num_outer_scans;
5663 :
5664 363852 : if (num_scans > 1)
5665 : {
5666 : double pages_fetched;
5667 :
5668 : /* total page fetches ignoring cache effects */
5669 48330 : pages_fetched = numIndexPages * num_scans;
5670 :
5671 : /* use Mackert and Lohman formula to adjust for cache effects */
5672 48330 : pages_fetched = index_pages_fetched(pages_fetched,
5673 : index->pages,
5674 48330 : (double) index->pages,
5675 : root);
5676 :
5677 : /*
5678 : * Now compute the total disk access cost, and then report a pro-rated
5679 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
5680 : * since that's internal to the indexscan.)
5681 : */
5682 48330 : indexTotalCost = (pages_fetched * spc_random_page_cost)
5683 : / num_outer_scans;
5684 : }
5685 : else
5686 : {
5687 : /*
5688 : * For a single index scan, we just charge spc_random_page_cost per
5689 : * page touched.
5690 : */
5691 315522 : indexTotalCost = numIndexPages * spc_random_page_cost;
5692 : }
5693 :
5694 : /*
5695 : * CPU cost: any complex expressions in the indexquals will need to be
5696 : * evaluated once at the start of the scan to reduce them to runtime keys
5697 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
5698 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
5699 : * indexqual operator. Because we have numIndexTuples as a per-scan
5700 : * number, we have to multiply by num_sa_scans to get the correct result
5701 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
5702 : * ORDER BY expressions.
5703 : *
5704 : * Note: this neglects the possible costs of rechecking lossy operators.
5705 : * Detecting that that might be needed seems more expensive than it's
5706 : * worth, though, considering all the other inaccuracies here ...
5707 : */
5708 727704 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
5709 363852 : index_other_operands_eval_cost(root, indexOrderBys);
5710 363852 : qual_op_cost = cpu_operator_cost *
5711 363852 : (list_length(indexQuals) + list_length(indexOrderBys));
5712 :
5713 363852 : indexStartupCost = qual_arg_cost;
5714 363852 : indexTotalCost += qual_arg_cost;
5715 363852 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
5716 :
5717 : /*
5718 : * Generic assumption about index correlation: there isn't any.
5719 : */
5720 363852 : indexCorrelation = 0.0;
5721 :
5722 : /*
5723 : * Return everything to caller.
5724 : */
5725 363852 : costs->indexStartupCost = indexStartupCost;
5726 363852 : costs->indexTotalCost = indexTotalCost;
5727 363852 : costs->indexSelectivity = indexSelectivity;
5728 363852 : costs->indexCorrelation = indexCorrelation;
5729 363852 : costs->numIndexPages = numIndexPages;
5730 363852 : costs->numIndexTuples = numIndexTuples;
5731 363852 : costs->spc_random_page_cost = spc_random_page_cost;
5732 363852 : costs->num_sa_scans = num_sa_scans;
5733 363852 : }
5734 :
5735 : /*
5736 : * If the index is partial, add its predicate to the given qual list.
5737 : *
5738 : * ANDing the index predicate with the explicitly given indexquals produces
5739 : * a more accurate idea of the index's selectivity. However, we need to be
5740 : * careful not to insert redundant clauses, because clauselist_selectivity()
5741 : * is easily fooled into computing a too-low selectivity estimate. Our
5742 : * approach is to add only the predicate clause(s) that cannot be proven to
5743 : * be implied by the given indexquals. This successfully handles cases such
5744 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
5745 : * There are many other cases where we won't detect redundancy, leading to a
5746 : * too-low selectivity estimate, which will bias the system in favor of using
5747 : * partial indexes where possible. That is not necessarily bad though.
5748 : *
5749 : * Note that indexQuals contains RestrictInfo nodes while the indpred
5750 : * does not, so the output list will be mixed. This is OK for both
5751 : * predicate_implied_by() and clauselist_selectivity(), but might be
5752 : * problematic if the result were passed to other things.
5753 : */
5754 : List *
5755 620402 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
5756 : {
5757 620402 : List *predExtraQuals = NIL;
5758 : ListCell *lc;
5759 :
5760 620402 : if (index->indpred == NIL)
5761 619164 : return indexQuals;
5762 :
5763 2484 : foreach(lc, index->indpred)
5764 : {
5765 1246 : Node *predQual = (Node *) lfirst(lc);
5766 1246 : List *oneQual = list_make1(predQual);
5767 :
5768 1246 : if (!predicate_implied_by(oneQual, indexQuals, false))
5769 1096 : predExtraQuals = list_concat(predExtraQuals, oneQual);
5770 : }
5771 1238 : return list_concat(predExtraQuals, indexQuals);
5772 : }
5773 :
5774 :
5775 : void
5776 360842 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
5777 : Cost *indexStartupCost, Cost *indexTotalCost,
5778 : Selectivity *indexSelectivity, double *indexCorrelation,
5779 : double *indexPages)
5780 : {
5781 360842 : IndexOptInfo *index = path->indexinfo;
5782 : GenericCosts costs;
5783 : Oid relid;
5784 : AttrNumber colnum;
5785 : VariableStatData vardata;
5786 : double numIndexTuples;
5787 : Cost descentCost;
5788 : List *indexBoundQuals;
5789 : int indexcol;
5790 : bool eqQualHere;
5791 : bool found_saop;
5792 : bool found_is_null_op;
5793 : double num_sa_scans;
5794 : ListCell *lc;
5795 :
5796 : /*
5797 : * For a btree scan, only leading '=' quals plus inequality quals for the
5798 : * immediately next attribute contribute to index selectivity (these are
5799 : * the "boundary quals" that determine the starting and stopping points of
5800 : * the index scan). Additional quals can suppress visits to the heap, so
5801 : * it's OK to count them in indexSelectivity, but they should not count
5802 : * for estimating numIndexTuples. So we must examine the given indexquals
5803 : * to find out which ones count as boundary quals. We rely on the
5804 : * knowledge that they are given in index column order.
5805 : *
5806 : * For a RowCompareExpr, we consider only the first column, just as
5807 : * rowcomparesel() does.
5808 : *
5809 : * If there's a ScalarArrayOpExpr in the quals, we'll actually perform N
5810 : * index scans not one, but the ScalarArrayOpExpr's operator can be
5811 : * considered to act the same as it normally does.
5812 : */
5813 360842 : indexBoundQuals = NIL;
5814 360842 : indexcol = 0;
5815 360842 : eqQualHere = false;
5816 360842 : found_saop = false;
5817 360842 : found_is_null_op = false;
5818 360842 : num_sa_scans = 1;
5819 649736 : foreach(lc, path->indexclauses)
5820 : {
5821 297994 : IndexClause *iclause = lfirst_node(IndexClause, lc);
5822 : ListCell *lc2;
5823 :
5824 297994 : if (indexcol != iclause->indexcol)
5825 : {
5826 : /* Beginning of a new column's quals */
5827 48472 : if (!eqQualHere)
5828 8560 : break; /* done if no '=' qual for indexcol */
5829 39912 : eqQualHere = false;
5830 39912 : indexcol++;
5831 39912 : if (indexcol != iclause->indexcol)
5832 540 : break; /* no quals at all for indexcol */
5833 : }
5834 :
5835 : /* Examine each indexqual associated with this index clause */
5836 579458 : foreach(lc2, iclause->indexquals)
5837 : {
5838 290564 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
5839 290564 : Expr *clause = rinfo->clause;
5840 290564 : Oid clause_op = InvalidOid;
5841 : int op_strategy;
5842 :
5843 290564 : if (IsA(clause, OpExpr))
5844 : {
5845 281890 : OpExpr *op = (OpExpr *) clause;
5846 :
5847 281890 : clause_op = op->opno;
5848 : }
5849 8674 : else if (IsA(clause, RowCompareExpr))
5850 : {
5851 88 : RowCompareExpr *rc = (RowCompareExpr *) clause;
5852 :
5853 88 : clause_op = linitial_oid(rc->opnos);
5854 : }
5855 8586 : else if (IsA(clause, ScalarArrayOpExpr))
5856 : {
5857 7012 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
5858 7012 : Node *other_operand = (Node *) lsecond(saop->args);
5859 7012 : int alength = estimate_array_length(other_operand);
5860 :
5861 7012 : clause_op = saop->opno;
5862 7012 : found_saop = true;
5863 : /* count number of SA scans induced by indexBoundQuals only */
5864 7012 : if (alength > 1)
5865 7012 : num_sa_scans *= alength;
5866 : }
5867 1574 : else if (IsA(clause, NullTest))
5868 : {
5869 1574 : NullTest *nt = (NullTest *) clause;
5870 :
5871 1574 : if (nt->nulltesttype == IS_NULL)
5872 : {
5873 88 : found_is_null_op = true;
5874 : /* IS NULL is like = for selectivity purposes */
5875 88 : eqQualHere = true;
5876 : }
5877 : }
5878 : else
5879 0 : elog(ERROR, "unsupported indexqual type: %d",
5880 : (int) nodeTag(clause));
5881 :
5882 : /* check for equality operator */
5883 290564 : if (OidIsValid(clause_op))
5884 : {
5885 288990 : op_strategy = get_op_opfamily_strategy(clause_op,
5886 288990 : index->opfamily[indexcol]);
5887 : Assert(op_strategy != 0); /* not a member of opfamily?? */
5888 288990 : if (op_strategy == BTEqualStrategyNumber)
5889 267626 : eqQualHere = true;
5890 : }
5891 :
5892 290564 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
5893 : }
5894 : }
5895 :
5896 : /*
5897 : * If index is unique and we found an '=' clause for each column, we can
5898 : * just assume numIndexTuples = 1 and skip the expensive
5899 : * clauselist_selectivity calculations. However, a ScalarArrayOp or
5900 : * NullTest invalidates that theory, even though it sets eqQualHere.
5901 : */
5902 673082 : if (index->unique &&
5903 515848 : indexcol == index->nkeycolumns - 1 &&
5904 112198 : eqQualHere &&
5905 217640 : !found_saop &&
5906 105442 : !found_is_null_op)
5907 105402 : numIndexTuples = 1.0;
5908 : else
5909 : {
5910 : List *selectivityQuals;
5911 : Selectivity btreeSelectivity;
5912 :
5913 : /*
5914 : * If the index is partial, AND the index predicate with the
5915 : * index-bound quals to produce a more accurate idea of the number of
5916 : * rows covered by the bound conditions.
5917 : */
5918 255440 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
5919 :
5920 255440 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
5921 255440 : index->rel->relid,
5922 : JOIN_INNER,
5923 : NULL);
5924 255440 : numIndexTuples = btreeSelectivity * index->rel->tuples;
5925 :
5926 : /*
5927 : * As in genericcostestimate(), we have to adjust for any
5928 : * ScalarArrayOpExpr quals included in indexBoundQuals, and then round
5929 : * to integer.
5930 : */
5931 255440 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
5932 : }
5933 :
5934 : /*
5935 : * Now do generic index cost estimation.
5936 : */
5937 360842 : MemSet(&costs, 0, sizeof(costs));
5938 360842 : costs.numIndexTuples = numIndexTuples;
5939 :
5940 360842 : genericcostestimate(root, path, loop_count, &costs);
5941 :
5942 : /*
5943 : * Add a CPU-cost component to represent the costs of initial btree
5944 : * descent. We don't charge any I/O cost for touching upper btree levels,
5945 : * since they tend to stay in cache, but we still have to do about log2(N)
5946 : * comparisons to descend a btree of N leaf tuples. We charge one
5947 : * cpu_operator_cost per comparison.
5948 : *
5949 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
5950 : * ones after the first one are not startup cost so far as the overall
5951 : * plan is concerned, so add them only to "total" cost.
5952 : */
5953 360842 : if (index->tuples > 1) /* avoid computing log(0) */
5954 : {
5955 358976 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
5956 358976 : costs.indexStartupCost += descentCost;
5957 358976 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
5958 : }
5959 :
5960 : /*
5961 : * Even though we're not charging I/O cost for touching upper btree pages,
5962 : * it's still reasonable to charge some CPU cost per page descended
5963 : * through. Moreover, if we had no such charge at all, bloated indexes
5964 : * would appear to have the same search cost as unbloated ones, at least
5965 : * in cases where only a single leaf page is expected to be visited. This
5966 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
5967 : * touched. The number of such pages is btree tree height plus one (ie,
5968 : * we charge for the leaf page too). As above, charge once per SA scan.
5969 : */
5970 360842 : descentCost = (index->tree_height + 1) * 50.0 * cpu_operator_cost;
5971 360842 : costs.indexStartupCost += descentCost;
5972 360842 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
5973 :
5974 : /*
5975 : * If we can get an estimate of the first column's ordering correlation C
5976 : * from pg_statistic, estimate the index correlation as C for a
5977 : * single-column index, or C * 0.75 for multiple columns. (The idea here
5978 : * is that multiple columns dilute the importance of the first column's
5979 : * ordering, but don't negate it entirely. Before 8.0 we divided the
5980 : * correlation by the number of columns, but that seems too strong.)
5981 : */
5982 360842 : MemSet(&vardata, 0, sizeof(vardata));
5983 :
5984 360842 : if (index->indexkeys[0] != 0)
5985 : {
5986 : /* Simple variable --- look to stats for the underlying table */
5987 359646 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
5988 :
5989 : Assert(rte->rtekind == RTE_RELATION);
5990 359646 : relid = rte->relid;
5991 : Assert(relid != InvalidOid);
5992 359646 : colnum = index->indexkeys[0];
5993 :
5994 359646 : if (get_relation_stats_hook &&
5995 0 : (*get_relation_stats_hook) (root, rte, colnum, &vardata))
5996 : {
5997 : /*
5998 : * The hook took control of acquiring a stats tuple. If it did
5999 : * supply a tuple, it'd better have supplied a freefunc.
6000 : */
6001 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
6002 0 : !vardata.freefunc)
6003 0 : elog(ERROR, "no function provided to release variable stats with");
6004 : }
6005 : else
6006 : {
6007 359646 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
6008 : ObjectIdGetDatum(relid),
6009 : Int16GetDatum(colnum),
6010 359646 : BoolGetDatum(rte->inh));
6011 359646 : vardata.freefunc = ReleaseSysCache;
6012 : }
6013 : }
6014 : else
6015 : {
6016 : /* Expression --- maybe there are stats for the index itself */
6017 1196 : relid = index->indexoid;
6018 1196 : colnum = 1;
6019 :
6020 1196 : if (get_index_stats_hook &&
6021 0 : (*get_index_stats_hook) (root, relid, colnum, &vardata))
6022 : {
6023 : /*
6024 : * The hook took control of acquiring a stats tuple. If it did
6025 : * supply a tuple, it'd better have supplied a freefunc.
6026 : */
6027 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
6028 0 : !vardata.freefunc)
6029 0 : elog(ERROR, "no function provided to release variable stats with");
6030 : }
6031 : else
6032 : {
6033 1196 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
6034 : ObjectIdGetDatum(relid),
6035 : Int16GetDatum(colnum),
6036 : BoolGetDatum(false));
6037 1196 : vardata.freefunc = ReleaseSysCache;
6038 : }
6039 : }
6040 :
6041 360842 : if (HeapTupleIsValid(vardata.statsTuple))
6042 : {
6043 : Oid sortop;
6044 : AttStatsSlot sslot;
6045 :
6046 548912 : sortop = get_opfamily_member(index->opfamily[0],
6047 274456 : index->opcintype[0],
6048 274456 : index->opcintype[0],
6049 : BTLessStrategyNumber);
6050 548912 : if (OidIsValid(sortop) &&
6051 274456 : get_attstatsslot(&sslot, vardata.statsTuple,
6052 : STATISTIC_KIND_CORRELATION, sortop,
6053 : ATTSTATSSLOT_NUMBERS))
6054 : {
6055 : double varCorrelation;
6056 :
6057 : Assert(sslot.nnumbers == 1);
6058 272792 : varCorrelation = sslot.numbers[0];
6059 :
6060 272792 : if (index->reverse_sort[0])
6061 0 : varCorrelation = -varCorrelation;
6062 :
6063 272792 : if (index->nkeycolumns > 1)
6064 114372 : costs.indexCorrelation = varCorrelation * 0.75;
6065 : else
6066 158420 : costs.indexCorrelation = varCorrelation;
6067 :
6068 272792 : free_attstatsslot(&sslot);
6069 : }
6070 : }
6071 :
6072 360842 : ReleaseVariableStats(vardata);
6073 :
6074 360842 : *indexStartupCost = costs.indexStartupCost;
6075 360842 : *indexTotalCost = costs.indexTotalCost;
6076 360842 : *indexSelectivity = costs.indexSelectivity;
6077 360842 : *indexCorrelation = costs.indexCorrelation;
6078 360842 : *indexPages = costs.numIndexPages;
6079 360842 : }
6080 :
6081 : void
6082 270 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
6083 : Cost *indexStartupCost, Cost *indexTotalCost,
6084 : Selectivity *indexSelectivity, double *indexCorrelation,
6085 : double *indexPages)
6086 : {
6087 : GenericCosts costs;
6088 :
6089 270 : MemSet(&costs, 0, sizeof(costs));
6090 :
6091 270 : genericcostestimate(root, path, loop_count, &costs);
6092 :
6093 : /*
6094 : * A hash index has no descent costs as such, since the index AM can go
6095 : * directly to the target bucket after computing the hash value. There
6096 : * are a couple of other hash-specific costs that we could conceivably add
6097 : * here, though:
6098 : *
6099 : * Ideally we'd charge spc_random_page_cost for each page in the target
6100 : * bucket, not just the numIndexPages pages that genericcostestimate
6101 : * thought we'd visit. However in most cases we don't know which bucket
6102 : * that will be. There's no point in considering the average bucket size
6103 : * because the hash AM makes sure that's always one page.
6104 : *
6105 : * Likewise, we could consider charging some CPU for each index tuple in
6106 : * the bucket, if we knew how many there were. But the per-tuple cost is
6107 : * just a hash value comparison, not a general datatype-dependent
6108 : * comparison, so any such charge ought to be quite a bit less than
6109 : * cpu_operator_cost; which makes it probably not worth worrying about.
6110 : *
6111 : * A bigger issue is that chance hash-value collisions will result in
6112 : * wasted probes into the heap. We don't currently attempt to model this
6113 : * cost on the grounds that it's rare, but maybe it's not rare enough.
6114 : * (Any fix for this ought to consider the generic lossy-operator problem,
6115 : * though; it's not entirely hash-specific.)
6116 : */
6117 :
6118 270 : *indexStartupCost = costs.indexStartupCost;
6119 270 : *indexTotalCost = costs.indexTotalCost;
6120 270 : *indexSelectivity = costs.indexSelectivity;
6121 270 : *indexCorrelation = costs.indexCorrelation;
6122 270 : *indexPages = costs.numIndexPages;
6123 270 : }
6124 :
6125 : void
6126 1556 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
6127 : Cost *indexStartupCost, Cost *indexTotalCost,
6128 : Selectivity *indexSelectivity, double *indexCorrelation,
6129 : double *indexPages)
6130 : {
6131 1556 : IndexOptInfo *index = path->indexinfo;
6132 : GenericCosts costs;
6133 : Cost descentCost;
6134 :
6135 1556 : MemSet(&costs, 0, sizeof(costs));
6136 :
6137 1556 : genericcostestimate(root, path, loop_count, &costs);
6138 :
6139 : /*
6140 : * We model index descent costs similarly to those for btree, but to do
6141 : * that we first need an idea of the tree height. We somewhat arbitrarily
6142 : * assume that the fanout is 100, meaning the tree height is at most
6143 : * log100(index->pages).
6144 : *
6145 : * Although this computation isn't really expensive enough to require
6146 : * caching, we might as well use index->tree_height to cache it.
6147 : */
6148 1556 : if (index->tree_height < 0) /* unknown? */
6149 : {
6150 1548 : if (index->pages > 1) /* avoid computing log(0) */
6151 1216 : index->tree_height = (int) (log(index->pages) / log(100.0));
6152 : else
6153 332 : index->tree_height = 0;
6154 : }
6155 :
6156 : /*
6157 : * Add a CPU-cost component to represent the costs of initial descent. We
6158 : * just use log(N) here not log2(N) since the branching factor isn't
6159 : * necessarily two anyway. As for btree, charge once per SA scan.
6160 : */
6161 1556 : if (index->tuples > 1) /* avoid computing log(0) */
6162 : {
6163 1556 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
6164 1556 : costs.indexStartupCost += descentCost;
6165 1556 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
6166 : }
6167 :
6168 : /*
6169 : * Likewise add a per-page charge, calculated the same as for btrees.
6170 : */
6171 1556 : descentCost = (index->tree_height + 1) * 50.0 * cpu_operator_cost;
6172 1556 : costs.indexStartupCost += descentCost;
6173 1556 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
6174 :
6175 1556 : *indexStartupCost = costs.indexStartupCost;
6176 1556 : *indexTotalCost = costs.indexTotalCost;
6177 1556 : *indexSelectivity = costs.indexSelectivity;
6178 1556 : *indexCorrelation = costs.indexCorrelation;
6179 1556 : *indexPages = costs.numIndexPages;
6180 1556 : }
6181 :
6182 : void
6183 1136 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
6184 : Cost *indexStartupCost, Cost *indexTotalCost,
6185 : Selectivity *indexSelectivity, double *indexCorrelation,
6186 : double *indexPages)
6187 : {
6188 1136 : IndexOptInfo *index = path->indexinfo;
6189 : GenericCosts costs;
6190 : Cost descentCost;
6191 :
6192 1136 : MemSet(&costs, 0, sizeof(costs));
6193 :
6194 1136 : genericcostestimate(root, path, loop_count, &costs);
6195 :
6196 : /*
6197 : * We model index descent costs similarly to those for btree, but to do
6198 : * that we first need an idea of the tree height. We somewhat arbitrarily
6199 : * assume that the fanout is 100, meaning the tree height is at most
6200 : * log100(index->pages).
6201 : *
6202 : * Although this computation isn't really expensive enough to require
6203 : * caching, we might as well use index->tree_height to cache it.
6204 : */
6205 1136 : if (index->tree_height < 0) /* unknown? */
6206 : {
6207 1132 : if (index->pages > 1) /* avoid computing log(0) */
6208 1132 : index->tree_height = (int) (log(index->pages) / log(100.0));
6209 : else
6210 0 : index->tree_height = 0;
6211 : }
6212 :
6213 : /*
6214 : * Add a CPU-cost component to represent the costs of initial descent. We
6215 : * just use log(N) here not log2(N) since the branching factor isn't
6216 : * necessarily two anyway. As for btree, charge once per SA scan.
6217 : */
6218 1136 : if (index->tuples > 1) /* avoid computing log(0) */
6219 : {
6220 1136 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
6221 1136 : costs.indexStartupCost += descentCost;
6222 1136 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
6223 : }
6224 :
6225 : /*
6226 : * Likewise add a per-page charge, calculated the same as for btrees.
6227 : */
6228 1136 : descentCost = (index->tree_height + 1) * 50.0 * cpu_operator_cost;
6229 1136 : costs.indexStartupCost += descentCost;
6230 1136 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
6231 :
6232 1136 : *indexStartupCost = costs.indexStartupCost;
6233 1136 : *indexTotalCost = costs.indexTotalCost;
6234 1136 : *indexSelectivity = costs.indexSelectivity;
6235 1136 : *indexCorrelation = costs.indexCorrelation;
6236 1136 : *indexPages = costs.numIndexPages;
6237 1136 : }
6238 :
6239 :
6240 : /*
6241 : * Support routines for gincostestimate
6242 : */
6243 :
6244 : typedef struct
6245 : {
6246 : bool haveFullScan;
6247 : double partialEntries;
6248 : double exactEntries;
6249 : double searchEntries;
6250 : double arrayScans;
6251 : } GinQualCounts;
6252 :
6253 : /*
6254 : * Estimate the number of index terms that need to be searched for while
6255 : * testing the given GIN query, and increment the counts in *counts
6256 : * appropriately. If the query is unsatisfiable, return false.
6257 : */
6258 : static bool
6259 1122 : gincost_pattern(IndexOptInfo *index, int indexcol,
6260 : Oid clause_op, Datum query,
6261 : GinQualCounts *counts)
6262 : {
6263 : Oid extractProcOid;
6264 : Oid collation;
6265 : int strategy_op;
6266 : Oid lefttype,
6267 : righttype;
6268 1122 : int32 nentries = 0;
6269 1122 : bool *partial_matches = NULL;
6270 1122 : Pointer *extra_data = NULL;
6271 1122 : bool *nullFlags = NULL;
6272 1122 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
6273 : int32 i;
6274 :
6275 : Assert(indexcol < index->nkeycolumns);
6276 :
6277 : /*
6278 : * Get the operator's strategy number and declared input data types within
6279 : * the index opfamily. (We don't need the latter, but we use
6280 : * get_op_opfamily_properties because it will throw error if it fails to
6281 : * find a matching pg_amop entry.)
6282 : */
6283 1122 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
6284 : &strategy_op, &lefttype, &righttype);
6285 :
6286 : /*
6287 : * GIN always uses the "default" support functions, which are those with
6288 : * lefttype == righttype == the opclass' opcintype (see
6289 : * IndexSupportInitialize in relcache.c).
6290 : */
6291 2244 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
6292 1122 : index->opcintype[indexcol],
6293 1122 : index->opcintype[indexcol],
6294 : GIN_EXTRACTQUERY_PROC);
6295 :
6296 1122 : if (!OidIsValid(extractProcOid))
6297 : {
6298 : /* should not happen; throw same error as index_getprocinfo */
6299 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
6300 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
6301 : get_rel_name(index->indexoid));
6302 : }
6303 :
6304 : /*
6305 : * Choose collation to pass to extractProc (should match initGinState).
6306 : */
6307 1122 : if (OidIsValid(index->indexcollations[indexcol]))
6308 250 : collation = index->indexcollations[indexcol];
6309 : else
6310 872 : collation = DEFAULT_COLLATION_OID;
6311 :
6312 1122 : OidFunctionCall7Coll(extractProcOid,
6313 : collation,
6314 : query,
6315 : PointerGetDatum(&nentries),
6316 : UInt16GetDatum(strategy_op),
6317 : PointerGetDatum(&partial_matches),
6318 : PointerGetDatum(&extra_data),
6319 : PointerGetDatum(&nullFlags),
6320 : PointerGetDatum(&searchMode));
6321 :
6322 1122 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
6323 : {
6324 : /* No match is possible */
6325 8 : return false;
6326 : }
6327 :
6328 3772 : for (i = 0; i < nentries; i++)
6329 : {
6330 : /*
6331 : * For partial match we haven't any information to estimate number of
6332 : * matched entries in index, so, we just estimate it as 100
6333 : */
6334 2658 : if (partial_matches && partial_matches[i])
6335 302 : counts->partialEntries += 100;
6336 : else
6337 2356 : counts->exactEntries++;
6338 :
6339 2658 : counts->searchEntries++;
6340 : }
6341 :
6342 1114 : if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
6343 : {
6344 : /* Treat "include empty" like an exact-match item */
6345 30 : counts->exactEntries++;
6346 30 : counts->searchEntries++;
6347 : }
6348 1084 : else if (searchMode != GIN_SEARCH_MODE_DEFAULT)
6349 : {
6350 : /* It's GIN_SEARCH_MODE_ALL */
6351 80 : counts->haveFullScan = true;
6352 : }
6353 :
6354 1114 : return true;
6355 : }
6356 :
6357 : /*
6358 : * Estimate the number of index terms that need to be searched for while
6359 : * testing the given GIN index clause, and increment the counts in *counts
6360 : * appropriately. If the query is unsatisfiable, return false.
6361 : */
6362 : static bool
6363 1114 : gincost_opexpr(PlannerInfo *root,
6364 : IndexOptInfo *index,
6365 : int indexcol,
6366 : OpExpr *clause,
6367 : GinQualCounts *counts)
6368 : {
6369 1114 : Oid clause_op = clause->opno;
6370 1114 : Node *operand = (Node *) lsecond(clause->args);
6371 :
6372 : /* aggressively reduce to a constant, and look through relabeling */
6373 1114 : operand = estimate_expression_value(root, operand);
6374 :
6375 1114 : if (IsA(operand, RelabelType))
6376 0 : operand = (Node *) ((RelabelType *) operand)->arg;
6377 :
6378 : /*
6379 : * It's impossible to call extractQuery method for unknown operand. So
6380 : * unless operand is a Const we can't do much; just assume there will be
6381 : * one ordinary search entry from the operand at runtime.
6382 : */
6383 1114 : if (!IsA(operand, Const))
6384 : {
6385 0 : counts->exactEntries++;
6386 0 : counts->searchEntries++;
6387 0 : return true;
6388 : }
6389 :
6390 : /* If Const is null, there can be no matches */
6391 1114 : if (((Const *) operand)->constisnull)
6392 0 : return false;
6393 :
6394 : /* Otherwise, apply extractQuery and get the actual term counts */
6395 1114 : return gincost_pattern(index, indexcol, clause_op,
6396 : ((Const *) operand)->constvalue,
6397 : counts);
6398 : }
6399 :
6400 : /*
6401 : * Estimate the number of index terms that need to be searched for while
6402 : * testing the given GIN index clause, and increment the counts in *counts
6403 : * appropriately. If the query is unsatisfiable, return false.
6404 : *
6405 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
6406 : * each of which involves one value from the RHS array, plus all the
6407 : * non-array quals (if any). To model this, we average the counts across
6408 : * the RHS elements, and add the averages to the counts in *counts (which
6409 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
6410 : * by N, causing gincostestimate to scale up its estimates accordingly.
6411 : */
6412 : static bool
6413 4 : gincost_scalararrayopexpr(PlannerInfo *root,
6414 : IndexOptInfo *index,
6415 : int indexcol,
6416 : ScalarArrayOpExpr *clause,
6417 : double numIndexEntries,
6418 : GinQualCounts *counts)
6419 : {
6420 4 : Oid clause_op = clause->opno;
6421 4 : Node *rightop = (Node *) lsecond(clause->args);
6422 : ArrayType *arrayval;
6423 : int16 elmlen;
6424 : bool elmbyval;
6425 : char elmalign;
6426 : int numElems;
6427 : Datum *elemValues;
6428 : bool *elemNulls;
6429 : GinQualCounts arraycounts;
6430 4 : int numPossible = 0;
6431 : int i;
6432 :
6433 : Assert(clause->useOr);
6434 :
6435 : /* aggressively reduce to a constant, and look through relabeling */
6436 4 : rightop = estimate_expression_value(root, rightop);
6437 :
6438 4 : if (IsA(rightop, RelabelType))
6439 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
6440 :
6441 : /*
6442 : * It's impossible to call extractQuery method for unknown operand. So
6443 : * unless operand is a Const we can't do much; just assume there will be
6444 : * one ordinary search entry from each array entry at runtime, and fall
6445 : * back on a probably-bad estimate of the number of array entries.
6446 : */
6447 4 : if (!IsA(rightop, Const))
6448 : {
6449 0 : counts->exactEntries++;
6450 0 : counts->searchEntries++;
6451 0 : counts->arrayScans *= estimate_array_length(rightop);
6452 0 : return true;
6453 : }
6454 :
6455 : /* If Const is null, there can be no matches */
6456 4 : if (((Const *) rightop)->constisnull)
6457 0 : return false;
6458 :
6459 : /* Otherwise, extract the array elements and iterate over them */
6460 4 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
6461 4 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
6462 : &elmlen, &elmbyval, &elmalign);
6463 4 : deconstruct_array(arrayval,
6464 : ARR_ELEMTYPE(arrayval),
6465 : elmlen, elmbyval, elmalign,
6466 : &elemValues, &elemNulls, &numElems);
6467 :
6468 4 : memset(&arraycounts, 0, sizeof(arraycounts));
6469 :
6470 12 : for (i = 0; i < numElems; i++)
6471 : {
6472 : GinQualCounts elemcounts;
6473 :
6474 : /* NULL can't match anything, so ignore, as the executor will */
6475 8 : if (elemNulls[i])
6476 0 : continue;
6477 :
6478 : /* Otherwise, apply extractQuery and get the actual term counts */
6479 8 : memset(&elemcounts, 0, sizeof(elemcounts));
6480 :
6481 8 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
6482 : &elemcounts))
6483 : {
6484 : /* We ignore array elements that are unsatisfiable patterns */
6485 8 : numPossible++;
6486 :
6487 8 : if (elemcounts.haveFullScan)
6488 : {
6489 : /*
6490 : * Full index scan will be required. We treat this as if
6491 : * every key in the index had been listed in the query; is
6492 : * that reasonable?
6493 : */
6494 0 : elemcounts.partialEntries = 0;
6495 0 : elemcounts.exactEntries = numIndexEntries;
6496 0 : elemcounts.searchEntries = numIndexEntries;
6497 : }
6498 8 : arraycounts.partialEntries += elemcounts.partialEntries;
6499 8 : arraycounts.exactEntries += elemcounts.exactEntries;
6500 8 : arraycounts.searchEntries += elemcounts.searchEntries;
6501 : }
6502 : }
6503 :
6504 4 : if (numPossible == 0)
6505 : {
6506 : /* No satisfiable patterns in the array */
6507 0 : return false;
6508 : }
6509 :
6510 : /*
6511 : * Now add the averages to the global counts. This will give us an
6512 : * estimate of the average number of terms searched for in each indexscan,
6513 : * including contributions from both array and non-array quals.
6514 : */
6515 4 : counts->partialEntries += arraycounts.partialEntries / numPossible;
6516 4 : counts->exactEntries += arraycounts.exactEntries / numPossible;
6517 4 : counts->searchEntries += arraycounts.searchEntries / numPossible;
6518 :
6519 4 : counts->arrayScans *= numPossible;
6520 :
6521 4 : return true;
6522 : }
6523 :
6524 : /*
6525 : * GIN has search behavior completely different from other index types
6526 : */
6527 : void
6528 1110 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
6529 : Cost *indexStartupCost, Cost *indexTotalCost,
6530 : Selectivity *indexSelectivity, double *indexCorrelation,
6531 : double *indexPages)
6532 : {
6533 1110 : IndexOptInfo *index = path->indexinfo;
6534 1110 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
6535 : List *selectivityQuals;
6536 1110 : double numPages = index->pages,
6537 1110 : numTuples = index->tuples;
6538 : double numEntryPages,
6539 : numDataPages,
6540 : numPendingPages,
6541 : numEntries;
6542 : GinQualCounts counts;
6543 : bool matchPossible;
6544 : double partialScale;
6545 : double entryPagesFetched,
6546 : dataPagesFetched,
6547 : dataPagesFetchedBySel;
6548 : double qual_op_cost,
6549 : qual_arg_cost,
6550 : spc_random_page_cost,
6551 : outer_scans;
6552 : Relation indexRel;
6553 : GinStatsData ginStats;
6554 : ListCell *lc;
6555 :
6556 : /*
6557 : * Obtain statistical information from the meta page, if possible. Else
6558 : * set ginStats to zeroes, and we'll cope below.
6559 : */
6560 1110 : if (!index->hypothetical)
6561 : {
6562 : /* Lock should have already been obtained in plancat.c */
6563 1110 : indexRel = index_open(index->indexoid, NoLock);
6564 1110 : ginGetStats(indexRel, &ginStats);
6565 1110 : index_close(indexRel, NoLock);
6566 : }
6567 : else
6568 : {
6569 0 : memset(&ginStats, 0, sizeof(ginStats));
6570 : }
6571 :
6572 : /*
6573 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
6574 : * trusted, but the other fields are data as of the last VACUUM. We can
6575 : * scale them up to account for growth since then, but that method only
6576 : * goes so far; in the worst case, the stats might be for a completely
6577 : * empty index, and scaling them will produce pretty bogus numbers.
6578 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
6579 : * it's grown more than that, fall back to estimating things only from the
6580 : * assumed-accurate index size. But we'll trust nPendingPages in any case
6581 : * so long as it's not clearly insane, ie, more than the index size.
6582 : */
6583 1110 : if (ginStats.nPendingPages < numPages)
6584 1110 : numPendingPages = ginStats.nPendingPages;
6585 : else
6586 0 : numPendingPages = 0;
6587 :
6588 2220 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
6589 2220 : ginStats.nTotalPages > numPages / 4 &&
6590 2220 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
6591 1090 : {
6592 : /*
6593 : * OK, the stats seem close enough to sane to be trusted. But we
6594 : * still need to scale them by the ratio numPages / nTotalPages to
6595 : * account for growth since the last VACUUM.
6596 : */
6597 1090 : double scale = numPages / ginStats.nTotalPages;
6598 :
6599 1090 : numEntryPages = ceil(ginStats.nEntryPages * scale);
6600 1090 : numDataPages = ceil(ginStats.nDataPages * scale);
6601 1090 : numEntries = ceil(ginStats.nEntries * scale);
6602 : /* ensure we didn't round up too much */
6603 1090 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
6604 1090 : numDataPages = Min(numDataPages,
6605 : numPages - numPendingPages - numEntryPages);
6606 : }
6607 : else
6608 : {
6609 : /*
6610 : * We might get here because it's a hypothetical index, or an index
6611 : * created pre-9.1 and never vacuumed since upgrading (in which case
6612 : * its stats would read as zeroes), or just because it's grown too
6613 : * much since the last VACUUM for us to put our faith in scaling.
6614 : *
6615 : * Invent some plausible internal statistics based on the index page
6616 : * count (and clamp that to at least 10 pages, just in case). We
6617 : * estimate that 90% of the index is entry pages, and the rest is data
6618 : * pages. Estimate 100 entries per entry page; this is rather bogus
6619 : * since it'll depend on the size of the keys, but it's more robust
6620 : * than trying to predict the number of entries per heap tuple.
6621 : */
6622 20 : numPages = Max(numPages, 10);
6623 20 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
6624 20 : numDataPages = numPages - numPendingPages - numEntryPages;
6625 20 : numEntries = floor(numEntryPages * 100);
6626 : }
6627 :
6628 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
6629 1110 : if (numEntries < 1)
6630 0 : numEntries = 1;
6631 :
6632 : /*
6633 : * If the index is partial, AND the index predicate with the index-bound
6634 : * quals to produce a more accurate idea of the number of rows covered by
6635 : * the bound conditions.
6636 : */
6637 1110 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
6638 :
6639 : /* Estimate the fraction of main-table tuples that will be visited */
6640 1110 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
6641 1110 : index->rel->relid,
6642 : JOIN_INNER,
6643 : NULL);
6644 :
6645 : /* fetch estimated page cost for tablespace containing index */
6646 1110 : get_tablespace_page_costs(index->reltablespace,
6647 : &spc_random_page_cost,
6648 : NULL);
6649 :
6650 : /*
6651 : * Generic assumption about index correlation: there isn't any.
6652 : */
6653 1110 : *indexCorrelation = 0.0;
6654 :
6655 : /*
6656 : * Examine quals to estimate number of search entries & partial matches
6657 : */
6658 1110 : memset(&counts, 0, sizeof(counts));
6659 1110 : counts.arrayScans = 1;
6660 1110 : matchPossible = true;
6661 :
6662 2228 : foreach(lc, path->indexclauses)
6663 : {
6664 1118 : IndexClause *iclause = lfirst_node(IndexClause, lc);
6665 : ListCell *lc2;
6666 :
6667 2228 : foreach(lc2, iclause->indexquals)
6668 : {
6669 1118 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6670 1118 : Expr *clause = rinfo->clause;
6671 :
6672 1118 : if (IsA(clause, OpExpr))
6673 : {
6674 1114 : matchPossible = gincost_opexpr(root,
6675 : index,
6676 1114 : iclause->indexcol,
6677 : (OpExpr *) clause,
6678 : &counts);
6679 1114 : if (!matchPossible)
6680 8 : break;
6681 : }
6682 4 : else if (IsA(clause, ScalarArrayOpExpr))
6683 : {
6684 4 : matchPossible = gincost_scalararrayopexpr(root,
6685 : index,
6686 4 : iclause->indexcol,
6687 : (ScalarArrayOpExpr *) clause,
6688 : numEntries,
6689 : &counts);
6690 4 : if (!matchPossible)
6691 0 : break;
6692 : }
6693 : else
6694 : {
6695 : /* shouldn't be anything else for a GIN index */
6696 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
6697 : (int) nodeTag(clause));
6698 : }
6699 : }
6700 : }
6701 :
6702 : /* Fall out if there were any provably-unsatisfiable quals */
6703 1110 : if (!matchPossible)
6704 : {
6705 8 : *indexStartupCost = 0;
6706 8 : *indexTotalCost = 0;
6707 8 : *indexSelectivity = 0;
6708 8 : return;
6709 : }
6710 :
6711 1102 : if (counts.haveFullScan || indexQuals == NIL)
6712 : {
6713 : /*
6714 : * Full index scan will be required. We treat this as if every key in
6715 : * the index had been listed in the query; is that reasonable?
6716 : */
6717 80 : counts.partialEntries = 0;
6718 80 : counts.exactEntries = numEntries;
6719 80 : counts.searchEntries = numEntries;
6720 : }
6721 :
6722 : /* Will we have more than one iteration of a nestloop scan? */
6723 1102 : outer_scans = loop_count;
6724 :
6725 : /*
6726 : * Compute cost to begin scan, first of all, pay attention to pending
6727 : * list.
6728 : */
6729 1102 : entryPagesFetched = numPendingPages;
6730 :
6731 : /*
6732 : * Estimate number of entry pages read. We need to do
6733 : * counts.searchEntries searches. Use a power function as it should be,
6734 : * but tuples on leaf pages usually is much greater. Here we include all
6735 : * searches in entry tree, including search of first entry in partial
6736 : * match algorithm
6737 : */
6738 1102 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
6739 :
6740 : /*
6741 : * Add an estimate of entry pages read by partial match algorithm. It's a
6742 : * scan over leaf pages in entry tree. We haven't any useful stats here,
6743 : * so estimate it as proportion. Because counts.partialEntries is really
6744 : * pretty bogus (see code above), it's possible that it is more than
6745 : * numEntries; clamp the proportion to ensure sanity.
6746 : */
6747 1102 : partialScale = counts.partialEntries / numEntries;
6748 1102 : partialScale = Min(partialScale, 1.0);
6749 :
6750 1102 : entryPagesFetched += ceil(numEntryPages * partialScale);
6751 :
6752 : /*
6753 : * Partial match algorithm reads all data pages before doing actual scan,
6754 : * so it's a startup cost. Again, we haven't any useful stats here, so
6755 : * estimate it as proportion.
6756 : */
6757 1102 : dataPagesFetched = ceil(numDataPages * partialScale);
6758 :
6759 : /*
6760 : * Calculate cache effects if more than one scan due to nestloops or array
6761 : * quals. The result is pro-rated per nestloop scan, but the array qual
6762 : * factor shouldn't be pro-rated (compare genericcostestimate).
6763 : */
6764 1102 : if (outer_scans > 1 || counts.arrayScans > 1)
6765 : {
6766 4 : entryPagesFetched *= outer_scans * counts.arrayScans;
6767 4 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
6768 : (BlockNumber) numEntryPages,
6769 : numEntryPages, root);
6770 4 : entryPagesFetched /= outer_scans;
6771 4 : dataPagesFetched *= outer_scans * counts.arrayScans;
6772 4 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
6773 : (BlockNumber) numDataPages,
6774 : numDataPages, root);
6775 4 : dataPagesFetched /= outer_scans;
6776 : }
6777 :
6778 : /*
6779 : * Here we use random page cost because logically-close pages could be far
6780 : * apart on disk.
6781 : */
6782 1102 : *indexStartupCost = (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
6783 :
6784 : /*
6785 : * Now compute the number of data pages fetched during the scan.
6786 : *
6787 : * We assume every entry to have the same number of items, and that there
6788 : * is no overlap between them. (XXX: tsvector and array opclasses collect
6789 : * statistics on the frequency of individual keys; it would be nice to use
6790 : * those here.)
6791 : */
6792 1102 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
6793 :
6794 : /*
6795 : * If there is a lot of overlap among the entries, in particular if one of
6796 : * the entries is very frequent, the above calculation can grossly
6797 : * under-estimate. As a simple cross-check, calculate a lower bound based
6798 : * on the overall selectivity of the quals. At a minimum, we must read
6799 : * one item pointer for each matching entry.
6800 : *
6801 : * The width of each item pointer varies, based on the level of
6802 : * compression. We don't have statistics on that, but an average of
6803 : * around 3 bytes per item is fairly typical.
6804 : */
6805 2204 : dataPagesFetchedBySel = ceil(*indexSelectivity *
6806 1102 : (numTuples / (BLCKSZ / 3)));
6807 1102 : if (dataPagesFetchedBySel > dataPagesFetched)
6808 1028 : dataPagesFetched = dataPagesFetchedBySel;
6809 :
6810 : /* Account for cache effects, the same as above */
6811 1102 : if (outer_scans > 1 || counts.arrayScans > 1)
6812 : {
6813 4 : dataPagesFetched *= outer_scans * counts.arrayScans;
6814 4 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
6815 : (BlockNumber) numDataPages,
6816 : numDataPages, root);
6817 4 : dataPagesFetched /= outer_scans;
6818 : }
6819 :
6820 : /* And apply random_page_cost as the cost per page */
6821 2204 : *indexTotalCost = *indexStartupCost +
6822 1102 : dataPagesFetched * spc_random_page_cost;
6823 :
6824 : /*
6825 : * Add on index qual eval costs, much as in genericcostestimate. But we
6826 : * can disregard indexorderbys, since GIN doesn't support those.
6827 : */
6828 1102 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
6829 1102 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
6830 :
6831 1102 : *indexStartupCost += qual_arg_cost;
6832 1102 : *indexTotalCost += qual_arg_cost;
6833 1102 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost + qual_op_cost);
6834 1102 : *indexPages = dataPagesFetched;
6835 : }
6836 :
6837 : /*
6838 : * BRIN has search behavior completely different from other index types
6839 : */
6840 : void
6841 3980 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
6842 : Cost *indexStartupCost, Cost *indexTotalCost,
6843 : Selectivity *indexSelectivity, double *indexCorrelation,
6844 : double *indexPages)
6845 : {
6846 3980 : IndexOptInfo *index = path->indexinfo;
6847 3980 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
6848 3980 : double numPages = index->pages;
6849 3980 : RelOptInfo *baserel = index->rel;
6850 3980 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
6851 : Cost spc_seq_page_cost;
6852 : Cost spc_random_page_cost;
6853 : double qual_arg_cost;
6854 : double qualSelectivity;
6855 : BrinStatsData statsData;
6856 : double indexRanges;
6857 : double minimalRanges;
6858 : double estimatedRanges;
6859 : double selec;
6860 : Relation indexRel;
6861 : ListCell *l;
6862 : VariableStatData vardata;
6863 :
6864 : Assert(rte->rtekind == RTE_RELATION);
6865 :
6866 : /* fetch estimated page cost for the tablespace containing the index */
6867 3980 : get_tablespace_page_costs(index->reltablespace,
6868 : &spc_random_page_cost,
6869 : &spc_seq_page_cost);
6870 :
6871 : /*
6872 : * Obtain some data from the index itself. A lock should have already
6873 : * been obtained on the index in plancat.c.
6874 : */
6875 3980 : indexRel = index_open(index->indexoid, NoLock);
6876 3980 : brinGetStats(indexRel, &statsData);
6877 3980 : index_close(indexRel, NoLock);
6878 :
6879 : /*
6880 : * Compute index correlation
6881 : *
6882 : * Because we can use all index quals equally when scanning, we can use
6883 : * the largest correlation (in absolute value) among columns used by the
6884 : * query. Start at zero, the worst possible case. If we cannot find any
6885 : * correlation statistics, we will keep it as 0.
6886 : */
6887 3980 : *indexCorrelation = 0;
6888 :
6889 7960 : foreach(l, path->indexclauses)
6890 : {
6891 3980 : IndexClause *iclause = lfirst_node(IndexClause, l);
6892 3980 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
6893 :
6894 : /* attempt to lookup stats in relation for this index column */
6895 3980 : if (attnum != 0)
6896 : {
6897 : /* Simple variable -- look to stats for the underlying table */
6898 3980 : if (get_relation_stats_hook &&
6899 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
6900 : {
6901 : /*
6902 : * The hook took control of acquiring a stats tuple. If it
6903 : * did supply a tuple, it'd better have supplied a freefunc.
6904 : */
6905 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
6906 0 : elog(ERROR,
6907 : "no function provided to release variable stats with");
6908 : }
6909 : else
6910 : {
6911 3980 : vardata.statsTuple =
6912 7960 : SearchSysCache3(STATRELATTINH,
6913 3980 : ObjectIdGetDatum(rte->relid),
6914 : Int16GetDatum(attnum),
6915 : BoolGetDatum(false));
6916 3980 : vardata.freefunc = ReleaseSysCache;
6917 : }
6918 : }
6919 : else
6920 : {
6921 : /*
6922 : * Looks like we've found an expression column in the index. Let's
6923 : * see if there's any stats for it.
6924 : */
6925 :
6926 : /* get the attnum from the 0-based index. */
6927 0 : attnum = iclause->indexcol + 1;
6928 :
6929 0 : if (get_index_stats_hook &&
6930 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
6931 : {
6932 : /*
6933 : * The hook took control of acquiring a stats tuple. If it
6934 : * did supply a tuple, it'd better have supplied a freefunc.
6935 : */
6936 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
6937 0 : !vardata.freefunc)
6938 0 : elog(ERROR, "no function provided to release variable stats with");
6939 : }
6940 : else
6941 : {
6942 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
6943 0 : ObjectIdGetDatum(index->indexoid),
6944 : Int16GetDatum(attnum),
6945 : BoolGetDatum(false));
6946 0 : vardata.freefunc = ReleaseSysCache;
6947 : }
6948 : }
6949 :
6950 3980 : if (HeapTupleIsValid(vardata.statsTuple))
6951 : {
6952 : AttStatsSlot sslot;
6953 :
6954 8 : if (get_attstatsslot(&sslot, vardata.statsTuple,
6955 : STATISTIC_KIND_CORRELATION, InvalidOid,
6956 : ATTSTATSSLOT_NUMBERS))
6957 : {
6958 8 : double varCorrelation = 0.0;
6959 :
6960 8 : if (sslot.nnumbers > 0)
6961 8 : varCorrelation = Abs(sslot.numbers[0]);
6962 :
6963 8 : if (varCorrelation > *indexCorrelation)
6964 8 : *indexCorrelation = varCorrelation;
6965 :
6966 8 : free_attstatsslot(&sslot);
6967 : }
6968 : }
6969 :
6970 3980 : ReleaseVariableStats(vardata);
6971 : }
6972 :
6973 3980 : qualSelectivity = clauselist_selectivity(root, indexQuals,
6974 3980 : baserel->relid,
6975 : JOIN_INNER, NULL);
6976 :
6977 : /* work out the actual number of ranges in the index */
6978 3980 : indexRanges = Max(ceil((double) baserel->pages / statsData.pagesPerRange),
6979 : 1.0);
6980 :
6981 : /*
6982 : * Now calculate the minimum possible ranges we could match with if all of
6983 : * the rows were in the perfect order in the table's heap.
6984 : */
6985 3980 : minimalRanges = ceil(indexRanges * qualSelectivity);
6986 :
6987 : /*
6988 : * Now estimate the number of ranges that we'll touch by using the
6989 : * indexCorrelation from the stats. Careful not to divide by zero (note
6990 : * we're using the absolute value of the correlation).
6991 : */
6992 3980 : if (*indexCorrelation < 1.0e-10)
6993 3972 : estimatedRanges = indexRanges;
6994 : else
6995 8 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
6996 :
6997 : /* we expect to visit this portion of the table */
6998 3980 : selec = estimatedRanges / indexRanges;
6999 :
7000 3980 : CLAMP_PROBABILITY(selec);
7001 :
7002 3980 : *indexSelectivity = selec;
7003 :
7004 : /*
7005 : * Compute the index qual costs, much as in genericcostestimate, to add to
7006 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
7007 : * support those.
7008 : */
7009 3980 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
7010 :
7011 : /*
7012 : * Compute the startup cost as the cost to read the whole revmap
7013 : * sequentially, including the cost to execute the index quals.
7014 : */
7015 3980 : *indexStartupCost =
7016 3980 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
7017 3980 : *indexStartupCost += qual_arg_cost;
7018 :
7019 : /*
7020 : * To read a BRIN index there might be a bit of back and forth over
7021 : * regular pages, as revmap might point to them out of sequential order;
7022 : * calculate the total cost as reading the whole index in random order.
7023 : */
7024 7960 : *indexTotalCost = *indexStartupCost +
7025 3980 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
7026 :
7027 : /*
7028 : * Charge a small amount per range tuple which we expect to match to. This
7029 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
7030 : * will set a bit for each page in the range when we find a matching
7031 : * range, so we must multiply the charge by the number of pages in the
7032 : * range.
7033 : */
7034 7960 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
7035 3980 : statsData.pagesPerRange;
7036 :
7037 3980 : *indexPages = index->pages;
7038 3980 : }
|
