Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * selfuncs.c
4 : * Selectivity functions and index cost estimation functions for
5 : * standard operators and index access methods.
6 : *
7 : * Selectivity routines are registered in the pg_operator catalog
8 : * in the "oprrest" and "oprjoin" attributes.
9 : *
10 : * Index cost functions are located via the index AM's API struct,
11 : * which is obtained from the handler function registered in pg_am.
12 : *
13 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
14 : * Portions Copyright (c) 1994, Regents of the University of California
15 : *
16 : *
17 : * IDENTIFICATION
18 : * src/backend/utils/adt/selfuncs.c
19 : *
20 : *-------------------------------------------------------------------------
21 : */
22 :
23 : /*----------
24 : * Operator selectivity estimation functions are called to estimate the
25 : * selectivity of WHERE clauses whose top-level operator is their operator.
26 : * We divide the problem into two cases:
27 : * Restriction clause estimation: the clause involves vars of just
28 : * one relation.
29 : * Join clause estimation: the clause involves vars of multiple rels.
30 : * Join selectivity estimation is far more difficult and usually less accurate
31 : * than restriction estimation.
32 : *
33 : * When dealing with the inner scan of a nestloop join, we consider the
34 : * join's joinclauses as restriction clauses for the inner relation, and
35 : * treat vars of the outer relation as parameters (a/k/a constants of unknown
36 : * values). So, restriction estimators need to be able to accept an argument
37 : * telling which relation is to be treated as the variable.
38 : *
39 : * The call convention for a restriction estimator (oprrest function) is
40 : *
41 : * Selectivity oprrest (PlannerInfo *root,
42 : * Oid operator,
43 : * List *args,
44 : * int varRelid);
45 : *
46 : * root: general information about the query (rtable and RelOptInfo lists
47 : * are particularly important for the estimator).
48 : * operator: OID of the specific operator in question.
49 : * args: argument list from the operator clause.
50 : * varRelid: if not zero, the relid (rtable index) of the relation to
51 : * be treated as the variable relation. May be zero if the args list
52 : * is known to contain vars of only one relation.
53 : *
54 : * This is represented at the SQL level (in pg_proc) as
55 : *
56 : * float8 oprrest (internal, oid, internal, int4);
57 : *
58 : * The result is a selectivity, that is, a fraction (0 to 1) of the rows
59 : * of the relation that are expected to produce a TRUE result for the
60 : * given operator.
61 : *
62 : * The call convention for a join estimator (oprjoin function) is similar
63 : * except that varRelid is not needed, and instead join information is
64 : * supplied:
65 : *
66 : * Selectivity oprjoin (PlannerInfo *root,
67 : * Oid operator,
68 : * List *args,
69 : * JoinType jointype,
70 : * SpecialJoinInfo *sjinfo);
71 : *
72 : * float8 oprjoin (internal, oid, internal, int2, internal);
73 : *
74 : * (Before Postgres 8.4, join estimators had only the first four of these
75 : * parameters. That signature is still allowed, but deprecated.) The
76 : * relationship between jointype and sjinfo is explained in the comments for
77 : * clause_selectivity() --- the short version is that jointype is usually
78 : * best ignored in favor of examining sjinfo.
79 : *
80 : * Join selectivity for regular inner and outer joins is defined as the
81 : * fraction (0 to 1) of the cross product of the relations that is expected
82 : * to produce a TRUE result for the given operator. For both semi and anti
83 : * joins, however, the selectivity is defined as the fraction of the left-hand
84 : * side relation's rows that are expected to have a match (ie, at least one
85 : * row with a TRUE result) in the right-hand side.
86 : *
87 : * For both oprrest and oprjoin functions, the operator's input collation OID
88 : * (if any) is passed using the standard fmgr mechanism, so that the estimator
89 : * function can fetch it with PG_GET_COLLATION(). Note, however, that all
90 : * statistics in pg_statistic are currently built using the relevant column's
91 : * collation.
92 : *----------
93 : */
94 :
95 : #include "postgres.h"
96 :
97 : #include <ctype.h>
98 : #include <math.h>
99 :
100 : #include "access/brin.h"
101 : #include "access/brin_page.h"
102 : #include "access/gin.h"
103 : #include "access/table.h"
104 : #include "access/tableam.h"
105 : #include "access/visibilitymap.h"
106 : #include "catalog/pg_collation.h"
107 : #include "catalog/pg_operator.h"
108 : #include "catalog/pg_statistic.h"
109 : #include "catalog/pg_statistic_ext.h"
110 : #include "executor/nodeAgg.h"
111 : #include "miscadmin.h"
112 : #include "nodes/makefuncs.h"
113 : #include "nodes/nodeFuncs.h"
114 : #include "optimizer/clauses.h"
115 : #include "optimizer/cost.h"
116 : #include "optimizer/optimizer.h"
117 : #include "optimizer/pathnode.h"
118 : #include "optimizer/paths.h"
119 : #include "optimizer/plancat.h"
120 : #include "parser/parse_clause.h"
121 : #include "parser/parse_relation.h"
122 : #include "parser/parsetree.h"
123 : #include "rewrite/rewriteManip.h"
124 : #include "statistics/statistics.h"
125 : #include "storage/bufmgr.h"
126 : #include "utils/acl.h"
127 : #include "utils/array.h"
128 : #include "utils/builtins.h"
129 : #include "utils/date.h"
130 : #include "utils/datum.h"
131 : #include "utils/fmgroids.h"
132 : #include "utils/index_selfuncs.h"
133 : #include "utils/lsyscache.h"
134 : #include "utils/memutils.h"
135 : #include "utils/pg_locale.h"
136 : #include "utils/rel.h"
137 : #include "utils/selfuncs.h"
138 : #include "utils/snapmgr.h"
139 : #include "utils/spccache.h"
140 : #include "utils/syscache.h"
141 : #include "utils/timestamp.h"
142 : #include "utils/typcache.h"
143 :
144 : #define DEFAULT_PAGE_CPU_MULTIPLIER 50.0
145 :
146 : /* Hooks for plugins to get control when we ask for stats */
147 : get_relation_stats_hook_type get_relation_stats_hook = NULL;
148 : get_index_stats_hook_type get_index_stats_hook = NULL;
149 :
150 : static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
151 : static double eqjoinsel_inner(Oid opfuncoid, Oid collation,
152 : VariableStatData *vardata1, VariableStatData *vardata2,
153 : double nd1, double nd2,
154 : bool isdefault1, bool isdefault2,
155 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
156 : Form_pg_statistic stats1, Form_pg_statistic stats2,
157 : bool have_mcvs1, bool have_mcvs2);
158 : static double eqjoinsel_semi(Oid opfuncoid, Oid collation,
159 : VariableStatData *vardata1, VariableStatData *vardata2,
160 : double nd1, double nd2,
161 : bool isdefault1, bool isdefault2,
162 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
163 : Form_pg_statistic stats1, Form_pg_statistic stats2,
164 : bool have_mcvs1, bool have_mcvs2,
165 : RelOptInfo *inner_rel);
166 : static bool estimate_multivariate_ndistinct(PlannerInfo *root,
167 : RelOptInfo *rel, List **varinfos, double *ndistinct);
168 : static bool convert_to_scalar(Datum value, Oid valuetypid, Oid collid,
169 : double *scaledvalue,
170 : Datum lobound, Datum hibound, Oid boundstypid,
171 : double *scaledlobound, double *scaledhibound);
172 : static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure);
173 : static void convert_string_to_scalar(char *value,
174 : double *scaledvalue,
175 : char *lobound,
176 : double *scaledlobound,
177 : char *hibound,
178 : double *scaledhibound);
179 : static void convert_bytea_to_scalar(Datum value,
180 : double *scaledvalue,
181 : Datum lobound,
182 : double *scaledlobound,
183 : Datum hibound,
184 : double *scaledhibound);
185 : static double convert_one_string_to_scalar(char *value,
186 : int rangelo, int rangehi);
187 : static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
188 : int rangelo, int rangehi);
189 : static char *convert_string_datum(Datum value, Oid typid, Oid collid,
190 : bool *failure);
191 : static double convert_timevalue_to_scalar(Datum value, Oid typid,
192 : bool *failure);
193 : static void examine_simple_variable(PlannerInfo *root, Var *var,
194 : VariableStatData *vardata);
195 : static void examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
196 : int indexcol, VariableStatData *vardata);
197 : static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata,
198 : Oid sortop, Oid collation,
199 : Datum *min, Datum *max);
200 : static void get_stats_slot_range(AttStatsSlot *sslot,
201 : Oid opfuncoid, FmgrInfo *opproc,
202 : Oid collation, int16 typLen, bool typByVal,
203 : Datum *min, Datum *max, bool *p_have_data);
204 : static bool get_actual_variable_range(PlannerInfo *root,
205 : VariableStatData *vardata,
206 : Oid sortop, Oid collation,
207 : Datum *min, Datum *max);
208 : static bool get_actual_variable_endpoint(Relation heapRel,
209 : Relation indexRel,
210 : ScanDirection indexscandir,
211 : ScanKey scankeys,
212 : int16 typLen,
213 : bool typByVal,
214 : TupleTableSlot *tableslot,
215 : MemoryContext outercontext,
216 : Datum *endpointDatum);
217 : static RelOptInfo *find_join_input_rel(PlannerInfo *root, Relids relids);
218 : static double btcost_correlation(IndexOptInfo *index,
219 : VariableStatData *vardata);
220 :
221 :
222 : /*
223 : * eqsel - Selectivity of "=" for any data types.
224 : *
225 : * Note: this routine is also used to estimate selectivity for some
226 : * operators that are not "=" but have comparable selectivity behavior,
227 : * such as "~=" (geometric approximate-match). Even for "=", we must
228 : * keep in mind that the left and right datatypes may differ.
229 : */
230 : Datum
231 628006 : eqsel(PG_FUNCTION_ARGS)
232 : {
233 628006 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
234 : }
235 :
236 : /*
237 : * Common code for eqsel() and neqsel()
238 : */
239 : static double
240 673538 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
241 : {
242 673538 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
243 673538 : Oid operator = PG_GETARG_OID(1);
244 673538 : List *args = (List *) PG_GETARG_POINTER(2);
245 673538 : int varRelid = PG_GETARG_INT32(3);
246 673538 : Oid collation = PG_GET_COLLATION();
247 : VariableStatData vardata;
248 : Node *other;
249 : bool varonleft;
250 : double selec;
251 :
252 : /*
253 : * When asked about <>, we do the estimation using the corresponding =
254 : * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
255 : */
256 673538 : if (negate)
257 : {
258 45532 : operator = get_negator(operator);
259 45532 : if (!OidIsValid(operator))
260 : {
261 : /* Use default selectivity (should we raise an error instead?) */
262 0 : return 1.0 - DEFAULT_EQ_SEL;
263 : }
264 : }
265 :
266 : /*
267 : * If expression is not variable = something or something = variable, then
268 : * punt and return a default estimate.
269 : */
270 673538 : if (!get_restriction_variable(root, args, varRelid,
271 : &vardata, &other, &varonleft))
272 5162 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
273 :
274 : /*
275 : * We can do a lot better if the something is a constant. (Note: the
276 : * Const might result from estimation rather than being a simple constant
277 : * in the query.)
278 : */
279 668370 : if (IsA(other, Const))
280 297348 : selec = var_eq_const(&vardata, operator, collation,
281 297348 : ((Const *) other)->constvalue,
282 297348 : ((Const *) other)->constisnull,
283 : varonleft, negate);
284 : else
285 371022 : selec = var_eq_non_const(&vardata, operator, collation, other,
286 : varonleft, negate);
287 :
288 668370 : ReleaseVariableStats(vardata);
289 :
290 668370 : return selec;
291 : }
292 :
293 : /*
294 : * var_eq_const --- eqsel for var = const case
295 : *
296 : * This is exported so that some other estimation functions can use it.
297 : */
298 : double
299 341784 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
300 : Datum constval, bool constisnull,
301 : bool varonleft, bool negate)
302 : {
303 : double selec;
304 341784 : double nullfrac = 0.0;
305 : bool isdefault;
306 : Oid opfuncoid;
307 :
308 : /*
309 : * If the constant is NULL, assume operator is strict and return zero, ie,
310 : * operator will never return TRUE. (It's zero even for a negator op.)
311 : */
312 341784 : if (constisnull)
313 404 : return 0.0;
314 :
315 : /*
316 : * Grab the nullfrac for use below. Note we allow use of nullfrac
317 : * regardless of security check.
318 : */
319 341380 : if (HeapTupleIsValid(vardata->statsTuple))
320 : {
321 : Form_pg_statistic stats;
322 :
323 259386 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
324 259386 : nullfrac = stats->stanullfrac;
325 : }
326 :
327 : /*
328 : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
329 : * assume there is exactly one match regardless of anything else. (This
330 : * is slightly bogus, since the index or clause's equality operator might
331 : * be different from ours, but it's much more likely to be right than
332 : * ignoring the information.)
333 : */
334 341380 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
335 : {
336 82116 : selec = 1.0 / vardata->rel->tuples;
337 : }
338 453418 : else if (HeapTupleIsValid(vardata->statsTuple) &&
339 194154 : statistic_proc_security_check(vardata,
340 194154 : (opfuncoid = get_opcode(oproid))))
341 194154 : {
342 : AttStatsSlot sslot;
343 194154 : bool match = false;
344 : int i;
345 :
346 : /*
347 : * Is the constant "=" to any of the column's most common values?
348 : * (Although the given operator may not really be "=", we will assume
349 : * that seeing whether it returns TRUE is an appropriate test. If you
350 : * don't like this, maybe you shouldn't be using eqsel for your
351 : * operator...)
352 : */
353 194154 : if (get_attstatsslot(&sslot, vardata->statsTuple,
354 : STATISTIC_KIND_MCV, InvalidOid,
355 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
356 : {
357 172804 : LOCAL_FCINFO(fcinfo, 2);
358 : FmgrInfo eqproc;
359 :
360 172804 : fmgr_info(opfuncoid, &eqproc);
361 :
362 : /*
363 : * Save a few cycles by setting up the fcinfo struct just once.
364 : * Using FunctionCallInvoke directly also avoids failure if the
365 : * eqproc returns NULL, though really equality functions should
366 : * never do that.
367 : */
368 172804 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
369 : NULL, NULL);
370 172804 : fcinfo->args[0].isnull = false;
371 172804 : fcinfo->args[1].isnull = false;
372 : /* be careful to apply operator right way 'round */
373 172804 : if (varonleft)
374 172772 : fcinfo->args[1].value = constval;
375 : else
376 32 : fcinfo->args[0].value = constval;
377 :
378 2888490 : for (i = 0; i < sslot.nvalues; i++)
379 : {
380 : Datum fresult;
381 :
382 2804630 : if (varonleft)
383 2804574 : fcinfo->args[0].value = sslot.values[i];
384 : else
385 56 : fcinfo->args[1].value = sslot.values[i];
386 2804630 : fcinfo->isnull = false;
387 2804630 : fresult = FunctionCallInvoke(fcinfo);
388 2804630 : if (!fcinfo->isnull && DatumGetBool(fresult))
389 : {
390 88944 : match = true;
391 88944 : break;
392 : }
393 : }
394 : }
395 : else
396 : {
397 : /* no most-common-value info available */
398 21350 : i = 0; /* keep compiler quiet */
399 : }
400 :
401 194154 : if (match)
402 : {
403 : /*
404 : * Constant is "=" to this common value. We know selectivity
405 : * exactly (or as exactly as ANALYZE could calculate it, anyway).
406 : */
407 88944 : selec = sslot.numbers[i];
408 : }
409 : else
410 : {
411 : /*
412 : * Comparison is against a constant that is neither NULL nor any
413 : * of the common values. Its selectivity cannot be more than
414 : * this:
415 : */
416 105210 : double sumcommon = 0.0;
417 : double otherdistinct;
418 :
419 2463770 : for (i = 0; i < sslot.nnumbers; i++)
420 2358560 : sumcommon += sslot.numbers[i];
421 105210 : selec = 1.0 - sumcommon - nullfrac;
422 105210 : CLAMP_PROBABILITY(selec);
423 :
424 : /*
425 : * and in fact it's probably a good deal less. We approximate that
426 : * all the not-common values share this remaining fraction
427 : * equally, so we divide by the number of other distinct values.
428 : */
429 105210 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
430 105210 : sslot.nnumbers;
431 105210 : if (otherdistinct > 1)
432 51254 : selec /= otherdistinct;
433 :
434 : /*
435 : * Another cross-check: selectivity shouldn't be estimated as more
436 : * than the least common "most common value".
437 : */
438 105210 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
439 0 : selec = sslot.numbers[sslot.nnumbers - 1];
440 : }
441 :
442 194154 : free_attstatsslot(&sslot);
443 : }
444 : else
445 : {
446 : /*
447 : * No ANALYZE stats available, so make a guess using estimated number
448 : * of distinct values and assuming they are equally common. (The guess
449 : * is unlikely to be very good, but we do know a few special cases.)
450 : */
451 65110 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
452 : }
453 :
454 : /* now adjust if we wanted <> rather than = */
455 341380 : if (negate)
456 36758 : selec = 1.0 - selec - nullfrac;
457 :
458 : /* result should be in range, but make sure... */
459 341380 : CLAMP_PROBABILITY(selec);
460 :
461 341380 : return selec;
462 : }
463 :
464 : /*
465 : * var_eq_non_const --- eqsel for var = something-other-than-const case
466 : *
467 : * This is exported so that some other estimation functions can use it.
468 : */
469 : double
470 371022 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
471 : Node *other,
472 : bool varonleft, bool negate)
473 : {
474 : double selec;
475 371022 : double nullfrac = 0.0;
476 : bool isdefault;
477 :
478 : /*
479 : * Grab the nullfrac for use below.
480 : */
481 371022 : if (HeapTupleIsValid(vardata->statsTuple))
482 : {
483 : Form_pg_statistic stats;
484 :
485 273040 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
486 273040 : nullfrac = stats->stanullfrac;
487 : }
488 :
489 : /*
490 : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
491 : * assume there is exactly one match regardless of anything else. (This
492 : * is slightly bogus, since the index or clause's equality operator might
493 : * be different from ours, but it's much more likely to be right than
494 : * ignoring the information.)
495 : */
496 371022 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
497 : {
498 148382 : selec = 1.0 / vardata->rel->tuples;
499 : }
500 222640 : else if (HeapTupleIsValid(vardata->statsTuple))
501 : {
502 : double ndistinct;
503 : AttStatsSlot sslot;
504 :
505 : /*
506 : * Search is for a value that we do not know a priori, but we will
507 : * assume it is not NULL. Estimate the selectivity as non-null
508 : * fraction divided by number of distinct values, so that we get a
509 : * result averaged over all possible values whether common or
510 : * uncommon. (Essentially, we are assuming that the not-yet-known
511 : * comparison value is equally likely to be any of the possible
512 : * values, regardless of their frequency in the table. Is that a good
513 : * idea?)
514 : */
515 139168 : selec = 1.0 - nullfrac;
516 139168 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
517 139168 : if (ndistinct > 1)
518 135484 : selec /= ndistinct;
519 :
520 : /*
521 : * Cross-check: selectivity should never be estimated as more than the
522 : * most common value's.
523 : */
524 139168 : if (get_attstatsslot(&sslot, vardata->statsTuple,
525 : STATISTIC_KIND_MCV, InvalidOid,
526 : ATTSTATSSLOT_NUMBERS))
527 : {
528 120132 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
529 558 : selec = sslot.numbers[0];
530 120132 : free_attstatsslot(&sslot);
531 : }
532 : }
533 : else
534 : {
535 : /*
536 : * No ANALYZE stats available, so make a guess using estimated number
537 : * of distinct values and assuming they are equally common. (The guess
538 : * is unlikely to be very good, but we do know a few special cases.)
539 : */
540 83472 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
541 : }
542 :
543 : /* now adjust if we wanted <> rather than = */
544 371022 : if (negate)
545 6458 : selec = 1.0 - selec - nullfrac;
546 :
547 : /* result should be in range, but make sure... */
548 371022 : CLAMP_PROBABILITY(selec);
549 :
550 371022 : return selec;
551 : }
552 :
553 : /*
554 : * neqsel - Selectivity of "!=" for any data types.
555 : *
556 : * This routine is also used for some operators that are not "!="
557 : * but have comparable selectivity behavior. See above comments
558 : * for eqsel().
559 : */
560 : Datum
561 45532 : neqsel(PG_FUNCTION_ARGS)
562 : {
563 45532 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
564 : }
565 :
566 : /*
567 : * scalarineqsel - Selectivity of "<", "<=", ">", ">=" for scalars.
568 : *
569 : * This is the guts of scalarltsel/scalarlesel/scalargtsel/scalargesel.
570 : * The isgt and iseq flags distinguish which of the four cases apply.
571 : *
572 : * The caller has commuted the clause, if necessary, so that we can treat
573 : * the variable as being on the left. The caller must also make sure that
574 : * the other side of the clause is a non-null Const, and dissect that into
575 : * a value and datatype. (This definition simplifies some callers that
576 : * want to estimate against a computed value instead of a Const node.)
577 : *
578 : * This routine works for any datatype (or pair of datatypes) known to
579 : * convert_to_scalar(). If it is applied to some other datatype,
580 : * it will return an approximate estimate based on assuming that the constant
581 : * value falls in the middle of the bin identified by binary search.
582 : */
583 : static double
584 304288 : scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
585 : Oid collation,
586 : VariableStatData *vardata, Datum constval, Oid consttype)
587 : {
588 : Form_pg_statistic stats;
589 : FmgrInfo opproc;
590 : double mcv_selec,
591 : hist_selec,
592 : sumcommon;
593 : double selec;
594 :
595 304288 : if (!HeapTupleIsValid(vardata->statsTuple))
596 : {
597 : /*
598 : * No stats are available. Typically this means we have to fall back
599 : * on the default estimate; but if the variable is CTID then we can
600 : * make an estimate based on comparing the constant to the table size.
601 : */
602 22470 : if (vardata->var && IsA(vardata->var, Var) &&
603 17642 : ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
604 : {
605 : ItemPointer itemptr;
606 : double block;
607 : double density;
608 :
609 : /*
610 : * If the relation's empty, we're going to include all of it.
611 : * (This is mostly to avoid divide-by-zero below.)
612 : */
613 1952 : if (vardata->rel->pages == 0)
614 0 : return 1.0;
615 :
616 1952 : itemptr = (ItemPointer) DatumGetPointer(constval);
617 1952 : block = ItemPointerGetBlockNumberNoCheck(itemptr);
618 :
619 : /*
620 : * Determine the average number of tuples per page (density).
621 : *
622 : * Since the last page will, on average, be only half full, we can
623 : * estimate it to have half as many tuples as earlier pages. So
624 : * give it half the weight of a regular page.
625 : */
626 1952 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
627 :
628 : /* If target is the last page, use half the density. */
629 1952 : if (block >= vardata->rel->pages - 1)
630 30 : density *= 0.5;
631 :
632 : /*
633 : * Using the average tuples per page, calculate how far into the
634 : * page the itemptr is likely to be and adjust block accordingly,
635 : * by adding that fraction of a whole block (but never more than a
636 : * whole block, no matter how high the itemptr's offset is). Here
637 : * we are ignoring the possibility of dead-tuple line pointers,
638 : * which is fairly bogus, but we lack the info to do better.
639 : */
640 1952 : if (density > 0.0)
641 : {
642 1952 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
643 :
644 1952 : block += Min(offset / density, 1.0);
645 : }
646 :
647 : /*
648 : * Convert relative block number to selectivity. Again, the last
649 : * page has only half weight.
650 : */
651 1952 : selec = block / (vardata->rel->pages - 0.5);
652 :
653 : /*
654 : * The calculation so far gave us a selectivity for the "<=" case.
655 : * We'll have one fewer tuple for "<" and one additional tuple for
656 : * ">=", the latter of which we'll reverse the selectivity for
657 : * below, so we can simply subtract one tuple for both cases. The
658 : * cases that need this adjustment can be identified by iseq being
659 : * equal to isgt.
660 : */
661 1952 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
662 1840 : selec -= (1.0 / vardata->rel->tuples);
663 :
664 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
665 1952 : if (isgt)
666 1834 : selec = 1.0 - selec;
667 :
668 1952 : CLAMP_PROBABILITY(selec);
669 1952 : return selec;
670 : }
671 :
672 : /* no stats available, so default result */
673 20518 : return DEFAULT_INEQ_SEL;
674 : }
675 281818 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
676 :
677 281818 : fmgr_info(get_opcode(operator), &opproc);
678 :
679 : /*
680 : * If we have most-common-values info, add up the fractions of the MCV
681 : * entries that satisfy MCV OP CONST. These fractions contribute directly
682 : * to the result selectivity. Also add up the total fraction represented
683 : * by MCV entries.
684 : */
685 281818 : mcv_selec = mcv_selectivity(vardata, &opproc, collation, constval, true,
686 : &sumcommon);
687 :
688 : /*
689 : * If there is a histogram, determine which bin the constant falls in, and
690 : * compute the resulting contribution to selectivity.
691 : */
692 281818 : hist_selec = ineq_histogram_selectivity(root, vardata,
693 : operator, &opproc, isgt, iseq,
694 : collation,
695 : constval, consttype);
696 :
697 : /*
698 : * Now merge the results from the MCV and histogram calculations,
699 : * realizing that the histogram covers only the non-null values that are
700 : * not listed in MCV.
701 : */
702 281818 : selec = 1.0 - stats->stanullfrac - sumcommon;
703 :
704 281818 : if (hist_selec >= 0.0)
705 209352 : selec *= hist_selec;
706 : else
707 : {
708 : /*
709 : * If no histogram but there are values not accounted for by MCV,
710 : * arbitrarily assume half of them will match.
711 : */
712 72466 : selec *= 0.5;
713 : }
714 :
715 281818 : selec += mcv_selec;
716 :
717 : /* result should be in range, but make sure... */
718 281818 : CLAMP_PROBABILITY(selec);
719 :
720 281818 : return selec;
721 : }
722 :
723 : /*
724 : * mcv_selectivity - Examine the MCV list for selectivity estimates
725 : *
726 : * Determine the fraction of the variable's MCV population that satisfies
727 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft. Also
728 : * compute the fraction of the total column population represented by the MCV
729 : * list. This code will work for any boolean-returning predicate operator.
730 : *
731 : * The function result is the MCV selectivity, and the fraction of the
732 : * total population is returned into *sumcommonp. Zeroes are returned
733 : * if there is no MCV list.
734 : */
735 : double
736 287964 : mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation,
737 : Datum constval, bool varonleft,
738 : double *sumcommonp)
739 : {
740 : double mcv_selec,
741 : sumcommon;
742 : AttStatsSlot sslot;
743 : int i;
744 :
745 287964 : mcv_selec = 0.0;
746 287964 : sumcommon = 0.0;
747 :
748 573500 : if (HeapTupleIsValid(vardata->statsTuple) &&
749 570742 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
750 285206 : get_attstatsslot(&sslot, vardata->statsTuple,
751 : STATISTIC_KIND_MCV, InvalidOid,
752 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
753 : {
754 133352 : LOCAL_FCINFO(fcinfo, 2);
755 :
756 : /*
757 : * We invoke the opproc "by hand" so that we won't fail on NULL
758 : * results. Such cases won't arise for normal comparison functions,
759 : * but generic_restriction_selectivity could perhaps be used with
760 : * operators that can return NULL. A small side benefit is to not
761 : * need to re-initialize the fcinfo struct from scratch each time.
762 : */
763 133352 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
764 : NULL, NULL);
765 133352 : fcinfo->args[0].isnull = false;
766 133352 : fcinfo->args[1].isnull = false;
767 : /* be careful to apply operator right way 'round */
768 133352 : if (varonleft)
769 133352 : fcinfo->args[1].value = constval;
770 : else
771 0 : fcinfo->args[0].value = constval;
772 :
773 3972050 : for (i = 0; i < sslot.nvalues; i++)
774 : {
775 : Datum fresult;
776 :
777 3838698 : if (varonleft)
778 3838698 : fcinfo->args[0].value = sslot.values[i];
779 : else
780 0 : fcinfo->args[1].value = sslot.values[i];
781 3838698 : fcinfo->isnull = false;
782 3838698 : fresult = FunctionCallInvoke(fcinfo);
783 3838698 : if (!fcinfo->isnull && DatumGetBool(fresult))
784 1409198 : mcv_selec += sslot.numbers[i];
785 3838698 : sumcommon += sslot.numbers[i];
786 : }
787 133352 : free_attstatsslot(&sslot);
788 : }
789 :
790 287964 : *sumcommonp = sumcommon;
791 287964 : return mcv_selec;
792 : }
793 :
794 : /*
795 : * histogram_selectivity - Examine the histogram for selectivity estimates
796 : *
797 : * Determine the fraction of the variable's histogram entries that satisfy
798 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft.
799 : *
800 : * This code will work for any boolean-returning predicate operator, whether
801 : * or not it has anything to do with the histogram sort operator. We are
802 : * essentially using the histogram just as a representative sample. However,
803 : * small histograms are unlikely to be all that representative, so the caller
804 : * should be prepared to fall back on some other estimation approach when the
805 : * histogram is missing or very small. It may also be prudent to combine this
806 : * approach with another one when the histogram is small.
807 : *
808 : * If the actual histogram size is not at least min_hist_size, we won't bother
809 : * to do the calculation at all. Also, if the n_skip parameter is > 0, we
810 : * ignore the first and last n_skip histogram elements, on the grounds that
811 : * they are outliers and hence not very representative. Typical values for
812 : * these parameters are 10 and 1.
813 : *
814 : * The function result is the selectivity, or -1 if there is no histogram
815 : * or it's smaller than min_hist_size.
816 : *
817 : * The output parameter *hist_size receives the actual histogram size,
818 : * or zero if no histogram. Callers may use this number to decide how
819 : * much faith to put in the function result.
820 : *
821 : * Note that the result disregards both the most-common-values (if any) and
822 : * null entries. The caller is expected to combine this result with
823 : * statistics for those portions of the column population. It may also be
824 : * prudent to clamp the result range, ie, disbelieve exact 0 or 1 outputs.
825 : */
826 : double
827 6146 : histogram_selectivity(VariableStatData *vardata,
828 : FmgrInfo *opproc, Oid collation,
829 : Datum constval, bool varonleft,
830 : int min_hist_size, int n_skip,
831 : int *hist_size)
832 : {
833 : double result;
834 : AttStatsSlot sslot;
835 :
836 : /* check sanity of parameters */
837 : Assert(n_skip >= 0);
838 : Assert(min_hist_size > 2 * n_skip);
839 :
840 9864 : if (HeapTupleIsValid(vardata->statsTuple) &&
841 7430 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
842 3712 : get_attstatsslot(&sslot, vardata->statsTuple,
843 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
844 : ATTSTATSSLOT_VALUES))
845 : {
846 3618 : *hist_size = sslot.nvalues;
847 3618 : if (sslot.nvalues >= min_hist_size)
848 : {
849 1730 : LOCAL_FCINFO(fcinfo, 2);
850 1730 : int nmatch = 0;
851 : int i;
852 :
853 : /*
854 : * We invoke the opproc "by hand" so that we won't fail on NULL
855 : * results. Such cases won't arise for normal comparison
856 : * functions, but generic_restriction_selectivity could perhaps be
857 : * used with operators that can return NULL. A small side benefit
858 : * is to not need to re-initialize the fcinfo struct from scratch
859 : * each time.
860 : */
861 1730 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
862 : NULL, NULL);
863 1730 : fcinfo->args[0].isnull = false;
864 1730 : fcinfo->args[1].isnull = false;
865 : /* be careful to apply operator right way 'round */
866 1730 : if (varonleft)
867 1730 : fcinfo->args[1].value = constval;
868 : else
869 0 : fcinfo->args[0].value = constval;
870 :
871 141892 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
872 : {
873 : Datum fresult;
874 :
875 140162 : if (varonleft)
876 140162 : fcinfo->args[0].value = sslot.values[i];
877 : else
878 0 : fcinfo->args[1].value = sslot.values[i];
879 140162 : fcinfo->isnull = false;
880 140162 : fresult = FunctionCallInvoke(fcinfo);
881 140162 : if (!fcinfo->isnull && DatumGetBool(fresult))
882 7408 : nmatch++;
883 : }
884 1730 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
885 : }
886 : else
887 1888 : result = -1;
888 3618 : free_attstatsslot(&sslot);
889 : }
890 : else
891 : {
892 2528 : *hist_size = 0;
893 2528 : result = -1;
894 : }
895 :
896 6146 : return result;
897 : }
898 :
899 : /*
900 : * generic_restriction_selectivity - Selectivity for almost anything
901 : *
902 : * This function estimates selectivity for operators that we don't have any
903 : * special knowledge about, but are on data types that we collect standard
904 : * MCV and/or histogram statistics for. (Additional assumptions are that
905 : * the operator is strict and immutable, or at least stable.)
906 : *
907 : * If we have "VAR OP CONST" or "CONST OP VAR", selectivity is estimated by
908 : * applying the operator to each element of the column's MCV and/or histogram
909 : * stats, and merging the results using the assumption that the histogram is
910 : * a reasonable random sample of the column's non-MCV population. Note that
911 : * if the operator's semantics are related to the histogram ordering, this
912 : * might not be such a great assumption; other functions such as
913 : * scalarineqsel() are probably a better match in such cases.
914 : *
915 : * Otherwise, fall back to the default selectivity provided by the caller.
916 : */
917 : double
918 1106 : generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation,
919 : List *args, int varRelid,
920 : double default_selectivity)
921 : {
922 : double selec;
923 : VariableStatData vardata;
924 : Node *other;
925 : bool varonleft;
926 :
927 : /*
928 : * If expression is not variable OP something or something OP variable,
929 : * then punt and return the default estimate.
930 : */
931 1106 : if (!get_restriction_variable(root, args, varRelid,
932 : &vardata, &other, &varonleft))
933 0 : return default_selectivity;
934 :
935 : /*
936 : * If the something is a NULL constant, assume operator is strict and
937 : * return zero, ie, operator will never return TRUE.
938 : */
939 1106 : if (IsA(other, Const) &&
940 1106 : ((Const *) other)->constisnull)
941 : {
942 0 : ReleaseVariableStats(vardata);
943 0 : return 0.0;
944 : }
945 :
946 1106 : if (IsA(other, Const))
947 : {
948 : /* Variable is being compared to a known non-null constant */
949 1106 : Datum constval = ((Const *) other)->constvalue;
950 : FmgrInfo opproc;
951 : double mcvsum;
952 : double mcvsel;
953 : double nullfrac;
954 : int hist_size;
955 :
956 1106 : fmgr_info(get_opcode(oproid), &opproc);
957 :
958 : /*
959 : * Calculate the selectivity for the column's most common values.
960 : */
961 1106 : mcvsel = mcv_selectivity(&vardata, &opproc, collation,
962 : constval, varonleft,
963 : &mcvsum);
964 :
965 : /*
966 : * If the histogram is large enough, see what fraction of it matches
967 : * the query, and assume that's representative of the non-MCV
968 : * population. Otherwise use the default selectivity for the non-MCV
969 : * population.
970 : */
971 1106 : selec = histogram_selectivity(&vardata, &opproc, collation,
972 : constval, varonleft,
973 : 10, 1, &hist_size);
974 1106 : if (selec < 0)
975 : {
976 : /* Nope, fall back on default */
977 1106 : selec = default_selectivity;
978 : }
979 0 : else if (hist_size < 100)
980 : {
981 : /*
982 : * For histogram sizes from 10 to 100, we combine the histogram
983 : * and default selectivities, putting increasingly more trust in
984 : * the histogram for larger sizes.
985 : */
986 0 : double hist_weight = hist_size / 100.0;
987 :
988 0 : selec = selec * hist_weight +
989 0 : default_selectivity * (1.0 - hist_weight);
990 : }
991 :
992 : /* In any case, don't believe extremely small or large estimates. */
993 1106 : if (selec < 0.0001)
994 0 : selec = 0.0001;
995 1106 : else if (selec > 0.9999)
996 0 : selec = 0.9999;
997 :
998 : /* Don't forget to account for nulls. */
999 1106 : if (HeapTupleIsValid(vardata.statsTuple))
1000 84 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
1001 : else
1002 1022 : nullfrac = 0.0;
1003 :
1004 : /*
1005 : * Now merge the results from the MCV and histogram calculations,
1006 : * realizing that the histogram covers only the non-null values that
1007 : * are not listed in MCV.
1008 : */
1009 1106 : selec *= 1.0 - nullfrac - mcvsum;
1010 1106 : selec += mcvsel;
1011 : }
1012 : else
1013 : {
1014 : /* Comparison value is not constant, so we can't do anything */
1015 0 : selec = default_selectivity;
1016 : }
1017 :
1018 1106 : ReleaseVariableStats(vardata);
1019 :
1020 : /* result should be in range, but make sure... */
1021 1106 : CLAMP_PROBABILITY(selec);
1022 :
1023 1106 : return selec;
1024 : }
1025 :
1026 : /*
1027 : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
1028 : *
1029 : * Determine the fraction of the variable's histogram population that
1030 : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
1031 : * The isgt and iseq flags distinguish which of the four cases apply.
1032 : *
1033 : * While opproc could be looked up from the operator OID, common callers
1034 : * also need to call it separately, so we make the caller pass both.
1035 : *
1036 : * Returns -1 if there is no histogram (valid results will always be >= 0).
1037 : *
1038 : * Note that the result disregards both the most-common-values (if any) and
1039 : * null entries. The caller is expected to combine this result with
1040 : * statistics for those portions of the column population.
1041 : *
1042 : * This is exported so that some other estimation functions can use it.
1043 : */
1044 : double
1045 287008 : ineq_histogram_selectivity(PlannerInfo *root,
1046 : VariableStatData *vardata,
1047 : Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq,
1048 : Oid collation,
1049 : Datum constval, Oid consttype)
1050 : {
1051 : double hist_selec;
1052 : AttStatsSlot sslot;
1053 :
1054 287008 : hist_selec = -1.0;
1055 :
1056 : /*
1057 : * Someday, ANALYZE might store more than one histogram per rel/att,
1058 : * corresponding to more than one possible sort ordering defined for the
1059 : * column type. Right now, we know there is only one, so just grab it and
1060 : * see if it matches the query.
1061 : *
1062 : * Note that we can't use opoid as search argument; the staop appearing in
1063 : * pg_statistic will be for the relevant '<' operator, but what we have
1064 : * might be some other inequality operator such as '>='. (Even if opoid
1065 : * is a '<' operator, it could be cross-type.) Hence we must use
1066 : * comparison_ops_are_compatible() to see if the operators match.
1067 : */
1068 573328 : if (HeapTupleIsValid(vardata->statsTuple) &&
1069 572316 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1070 285996 : get_attstatsslot(&sslot, vardata->statsTuple,
1071 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1072 : ATTSTATSSLOT_VALUES))
1073 : {
1074 213852 : if (sslot.nvalues > 1 &&
1075 427628 : sslot.stacoll == collation &&
1076 213776 : comparison_ops_are_compatible(sslot.staop, opoid))
1077 213668 : {
1078 : /*
1079 : * Use binary search to find the desired location, namely the
1080 : * right end of the histogram bin containing the comparison value,
1081 : * which is the leftmost entry for which the comparison operator
1082 : * succeeds (if isgt) or fails (if !isgt).
1083 : *
1084 : * In this loop, we pay no attention to whether the operator iseq
1085 : * or not; that detail will be mopped up below. (We cannot tell,
1086 : * anyway, whether the operator thinks the values are equal.)
1087 : *
1088 : * If the binary search accesses the first or last histogram
1089 : * entry, we try to replace that endpoint with the true column min
1090 : * or max as found by get_actual_variable_range(). This
1091 : * ameliorates misestimates when the min or max is moving as a
1092 : * result of changes since the last ANALYZE. Note that this could
1093 : * result in effectively including MCVs into the histogram that
1094 : * weren't there before, but we don't try to correct for that.
1095 : */
1096 : double histfrac;
1097 213668 : int lobound = 0; /* first possible slot to search */
1098 213668 : int hibound = sslot.nvalues; /* last+1 slot to search */
1099 213668 : bool have_end = false;
1100 :
1101 : /*
1102 : * If there are only two histogram entries, we'll want up-to-date
1103 : * values for both. (If there are more than two, we need at most
1104 : * one of them to be updated, so we deal with that within the
1105 : * loop.)
1106 : */
1107 213668 : if (sslot.nvalues == 2)
1108 3722 : have_end = get_actual_variable_range(root,
1109 : vardata,
1110 : sslot.staop,
1111 : collation,
1112 : &sslot.values[0],
1113 3722 : &sslot.values[1]);
1114 :
1115 1413042 : while (lobound < hibound)
1116 : {
1117 1199374 : int probe = (lobound + hibound) / 2;
1118 : bool ltcmp;
1119 :
1120 : /*
1121 : * If we find ourselves about to compare to the first or last
1122 : * histogram entry, first try to replace it with the actual
1123 : * current min or max (unless we already did so above).
1124 : */
1125 1199374 : if (probe == 0 && sslot.nvalues > 2)
1126 105518 : have_end = get_actual_variable_range(root,
1127 : vardata,
1128 : sslot.staop,
1129 : collation,
1130 : &sslot.values[0],
1131 : NULL);
1132 1093856 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1133 71456 : have_end = get_actual_variable_range(root,
1134 : vardata,
1135 : sslot.staop,
1136 : collation,
1137 : NULL,
1138 71456 : &sslot.values[probe]);
1139 :
1140 1199374 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1141 : collation,
1142 1199374 : sslot.values[probe],
1143 : constval));
1144 1199374 : if (isgt)
1145 68162 : ltcmp = !ltcmp;
1146 1199374 : if (ltcmp)
1147 449094 : lobound = probe + 1;
1148 : else
1149 750280 : hibound = probe;
1150 : }
1151 :
1152 213668 : if (lobound <= 0)
1153 : {
1154 : /*
1155 : * Constant is below lower histogram boundary. More
1156 : * precisely, we have found that no entry in the histogram
1157 : * satisfies the inequality clause (if !isgt) or they all do
1158 : * (if isgt). We estimate that that's true of the entire
1159 : * table, so set histfrac to 0.0 (which we'll flip to 1.0
1160 : * below, if isgt).
1161 : */
1162 91448 : histfrac = 0.0;
1163 : }
1164 122220 : else if (lobound >= sslot.nvalues)
1165 : {
1166 : /*
1167 : * Inverse case: constant is above upper histogram boundary.
1168 : */
1169 32832 : histfrac = 1.0;
1170 : }
1171 : else
1172 : {
1173 : /* We have values[i-1] <= constant <= values[i]. */
1174 89388 : int i = lobound;
1175 89388 : double eq_selec = 0;
1176 : double val,
1177 : high,
1178 : low;
1179 : double binfrac;
1180 :
1181 : /*
1182 : * In the cases where we'll need it below, obtain an estimate
1183 : * of the selectivity of "x = constval". We use a calculation
1184 : * similar to what var_eq_const() does for a non-MCV constant,
1185 : * ie, estimate that all distinct non-MCV values occur equally
1186 : * often. But multiplication by "1.0 - sumcommon - nullfrac"
1187 : * will be done by our caller, so we shouldn't do that here.
1188 : * Therefore we can't try to clamp the estimate by reference
1189 : * to the least common MCV; the result would be too small.
1190 : *
1191 : * Note: since this is effectively assuming that constval
1192 : * isn't an MCV, it's logically dubious if constval in fact is
1193 : * one. But we have to apply *some* correction for equality,
1194 : * and anyway we cannot tell if constval is an MCV, since we
1195 : * don't have a suitable equality operator at hand.
1196 : */
1197 89388 : if (i == 1 || isgt == iseq)
1198 : {
1199 : double otherdistinct;
1200 : bool isdefault;
1201 : AttStatsSlot mcvslot;
1202 :
1203 : /* Get estimated number of distinct values */
1204 38030 : otherdistinct = get_variable_numdistinct(vardata,
1205 : &isdefault);
1206 :
1207 : /* Subtract off the number of known MCVs */
1208 38030 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1209 : STATISTIC_KIND_MCV, InvalidOid,
1210 : ATTSTATSSLOT_NUMBERS))
1211 : {
1212 4340 : otherdistinct -= mcvslot.nnumbers;
1213 4340 : free_attstatsslot(&mcvslot);
1214 : }
1215 :
1216 : /* If result doesn't seem sane, leave eq_selec at 0 */
1217 38030 : if (otherdistinct > 1)
1218 37988 : eq_selec = 1.0 / otherdistinct;
1219 : }
1220 :
1221 : /*
1222 : * Convert the constant and the two nearest bin boundary
1223 : * values to a uniform comparison scale, and do a linear
1224 : * interpolation within this bin.
1225 : */
1226 89388 : if (convert_to_scalar(constval, consttype, collation,
1227 : &val,
1228 89388 : sslot.values[i - 1], sslot.values[i],
1229 : vardata->vartype,
1230 : &low, &high))
1231 : {
1232 89388 : if (high <= low)
1233 : {
1234 : /* cope if bin boundaries appear identical */
1235 0 : binfrac = 0.5;
1236 : }
1237 89388 : else if (val <= low)
1238 20466 : binfrac = 0.0;
1239 68922 : else if (val >= high)
1240 3404 : binfrac = 1.0;
1241 : else
1242 : {
1243 65518 : binfrac = (val - low) / (high - low);
1244 :
1245 : /*
1246 : * Watch out for the possibility that we got a NaN or
1247 : * Infinity from the division. This can happen
1248 : * despite the previous checks, if for example "low"
1249 : * is -Infinity.
1250 : */
1251 65518 : if (isnan(binfrac) ||
1252 65518 : binfrac < 0.0 || binfrac > 1.0)
1253 0 : binfrac = 0.5;
1254 : }
1255 : }
1256 : else
1257 : {
1258 : /*
1259 : * Ideally we'd produce an error here, on the grounds that
1260 : * the given operator shouldn't have scalarXXsel
1261 : * registered as its selectivity func unless we can deal
1262 : * with its operand types. But currently, all manner of
1263 : * stuff is invoking scalarXXsel, so give a default
1264 : * estimate until that can be fixed.
1265 : */
1266 0 : binfrac = 0.5;
1267 : }
1268 :
1269 : /*
1270 : * Now, compute the overall selectivity across the values
1271 : * represented by the histogram. We have i-1 full bins and
1272 : * binfrac partial bin below the constant.
1273 : */
1274 89388 : histfrac = (double) (i - 1) + binfrac;
1275 89388 : histfrac /= (double) (sslot.nvalues - 1);
1276 :
1277 : /*
1278 : * At this point, histfrac is an estimate of the fraction of
1279 : * the population represented by the histogram that satisfies
1280 : * "x <= constval". Somewhat remarkably, this statement is
1281 : * true regardless of which operator we were doing the probes
1282 : * with, so long as convert_to_scalar() delivers reasonable
1283 : * results. If the probe constant is equal to some histogram
1284 : * entry, we would have considered the bin to the left of that
1285 : * entry if probing with "<" or ">=", or the bin to the right
1286 : * if probing with "<=" or ">"; but binfrac would have come
1287 : * out as 1.0 in the first case and 0.0 in the second, leading
1288 : * to the same histfrac in either case. For probe constants
1289 : * between histogram entries, we find the same bin and get the
1290 : * same estimate with any operator.
1291 : *
1292 : * The fact that the estimate corresponds to "x <= constval"
1293 : * and not "x < constval" is because of the way that ANALYZE
1294 : * constructs the histogram: each entry is, effectively, the
1295 : * rightmost value in its sample bucket. So selectivity
1296 : * values that are exact multiples of 1/(histogram_size-1)
1297 : * should be understood as estimates including a histogram
1298 : * entry plus everything to its left.
1299 : *
1300 : * However, that breaks down for the first histogram entry,
1301 : * which necessarily is the leftmost value in its sample
1302 : * bucket. That means the first histogram bin is slightly
1303 : * narrower than the rest, by an amount equal to eq_selec.
1304 : * Another way to say that is that we want "x <= leftmost" to
1305 : * be estimated as eq_selec not zero. So, if we're dealing
1306 : * with the first bin (i==1), rescale to make that true while
1307 : * adjusting the rest of that bin linearly.
1308 : */
1309 89388 : if (i == 1)
1310 16904 : histfrac += eq_selec * (1.0 - binfrac);
1311 :
1312 : /*
1313 : * "x <= constval" is good if we want an estimate for "<=" or
1314 : * ">", but if we are estimating for "<" or ">=", we now need
1315 : * to decrease the estimate by eq_selec.
1316 : */
1317 89388 : if (isgt == iseq)
1318 28252 : histfrac -= eq_selec;
1319 : }
1320 :
1321 : /*
1322 : * Now the estimate is finished for "<" and "<=" cases. If we are
1323 : * estimating for ">" or ">=", flip it.
1324 : */
1325 213668 : hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1326 :
1327 : /*
1328 : * The histogram boundaries are only approximate to begin with,
1329 : * and may well be out of date anyway. Therefore, don't believe
1330 : * extremely small or large selectivity estimates --- unless we
1331 : * got actual current endpoint values from the table, in which
1332 : * case just do the usual sanity clamp. Somewhat arbitrarily, we
1333 : * set the cutoff for other cases at a hundredth of the histogram
1334 : * resolution.
1335 : */
1336 213668 : if (have_end)
1337 119142 : CLAMP_PROBABILITY(hist_selec);
1338 : else
1339 : {
1340 94526 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1341 :
1342 94526 : if (hist_selec < cutoff)
1343 33552 : hist_selec = cutoff;
1344 60974 : else if (hist_selec > 1.0 - cutoff)
1345 21910 : hist_selec = 1.0 - cutoff;
1346 : }
1347 : }
1348 184 : else if (sslot.nvalues > 1)
1349 : {
1350 : /*
1351 : * If we get here, we have a histogram but it's not sorted the way
1352 : * we want. Do a brute-force search to see how many of the
1353 : * entries satisfy the comparison condition, and take that
1354 : * fraction as our estimate. (This is identical to the inner loop
1355 : * of histogram_selectivity; maybe share code?)
1356 : */
1357 184 : LOCAL_FCINFO(fcinfo, 2);
1358 184 : int nmatch = 0;
1359 :
1360 184 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1361 : NULL, NULL);
1362 184 : fcinfo->args[0].isnull = false;
1363 184 : fcinfo->args[1].isnull = false;
1364 184 : fcinfo->args[1].value = constval;
1365 962556 : for (int i = 0; i < sslot.nvalues; i++)
1366 : {
1367 : Datum fresult;
1368 :
1369 962372 : fcinfo->args[0].value = sslot.values[i];
1370 962372 : fcinfo->isnull = false;
1371 962372 : fresult = FunctionCallInvoke(fcinfo);
1372 962372 : if (!fcinfo->isnull && DatumGetBool(fresult))
1373 2260 : nmatch++;
1374 : }
1375 184 : hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1376 :
1377 : /*
1378 : * As above, clamp to a hundredth of the histogram resolution.
1379 : * This case is surely even less trustworthy than the normal one,
1380 : * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1381 : * clamp should be more restrictive in this case?)
1382 : */
1383 : {
1384 184 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1385 :
1386 184 : if (hist_selec < cutoff)
1387 8 : hist_selec = cutoff;
1388 176 : else if (hist_selec > 1.0 - cutoff)
1389 8 : hist_selec = 1.0 - cutoff;
1390 : }
1391 : }
1392 :
1393 213852 : free_attstatsslot(&sslot);
1394 : }
1395 :
1396 287008 : return hist_selec;
1397 : }
1398 :
1399 : /*
1400 : * Common wrapper function for the selectivity estimators that simply
1401 : * invoke scalarineqsel().
1402 : */
1403 : static Datum
1404 44788 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1405 : {
1406 44788 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1407 44788 : Oid operator = PG_GETARG_OID(1);
1408 44788 : List *args = (List *) PG_GETARG_POINTER(2);
1409 44788 : int varRelid = PG_GETARG_INT32(3);
1410 44788 : Oid collation = PG_GET_COLLATION();
1411 : VariableStatData vardata;
1412 : Node *other;
1413 : bool varonleft;
1414 : Datum constval;
1415 : Oid consttype;
1416 : double selec;
1417 :
1418 : /*
1419 : * If expression is not variable op something or something op variable,
1420 : * then punt and return a default estimate.
1421 : */
1422 44788 : if (!get_restriction_variable(root, args, varRelid,
1423 : &vardata, &other, &varonleft))
1424 650 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1425 :
1426 : /*
1427 : * Can't do anything useful if the something is not a constant, either.
1428 : */
1429 44138 : if (!IsA(other, Const))
1430 : {
1431 2776 : ReleaseVariableStats(vardata);
1432 2776 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1433 : }
1434 :
1435 : /*
1436 : * If the constant is NULL, assume operator is strict and return zero, ie,
1437 : * operator will never return TRUE.
1438 : */
1439 41362 : if (((Const *) other)->constisnull)
1440 : {
1441 66 : ReleaseVariableStats(vardata);
1442 66 : PG_RETURN_FLOAT8(0.0);
1443 : }
1444 41296 : constval = ((Const *) other)->constvalue;
1445 41296 : consttype = ((Const *) other)->consttype;
1446 :
1447 : /*
1448 : * Force the var to be on the left to simplify logic in scalarineqsel.
1449 : */
1450 41296 : if (!varonleft)
1451 : {
1452 366 : operator = get_commutator(operator);
1453 366 : if (!operator)
1454 : {
1455 : /* Use default selectivity (should we raise an error instead?) */
1456 0 : ReleaseVariableStats(vardata);
1457 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1458 : }
1459 366 : isgt = !isgt;
1460 : }
1461 :
1462 : /* The rest of the work is done by scalarineqsel(). */
1463 41296 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1464 : &vardata, constval, consttype);
1465 :
1466 41296 : ReleaseVariableStats(vardata);
1467 :
1468 41296 : PG_RETURN_FLOAT8((float8) selec);
1469 : }
1470 :
1471 : /*
1472 : * scalarltsel - Selectivity of "<" for scalars.
1473 : */
1474 : Datum
1475 15030 : scalarltsel(PG_FUNCTION_ARGS)
1476 : {
1477 15030 : return scalarineqsel_wrapper(fcinfo, false, false);
1478 : }
1479 :
1480 : /*
1481 : * scalarlesel - Selectivity of "<=" for scalars.
1482 : */
1483 : Datum
1484 4548 : scalarlesel(PG_FUNCTION_ARGS)
1485 : {
1486 4548 : return scalarineqsel_wrapper(fcinfo, false, true);
1487 : }
1488 :
1489 : /*
1490 : * scalargtsel - Selectivity of ">" for scalars.
1491 : */
1492 : Datum
1493 15264 : scalargtsel(PG_FUNCTION_ARGS)
1494 : {
1495 15264 : return scalarineqsel_wrapper(fcinfo, true, false);
1496 : }
1497 :
1498 : /*
1499 : * scalargesel - Selectivity of ">=" for scalars.
1500 : */
1501 : Datum
1502 9946 : scalargesel(PG_FUNCTION_ARGS)
1503 : {
1504 9946 : return scalarineqsel_wrapper(fcinfo, true, true);
1505 : }
1506 :
1507 : /*
1508 : * boolvarsel - Selectivity of Boolean variable.
1509 : *
1510 : * This can actually be called on any boolean-valued expression. If it
1511 : * involves only Vars of the specified relation, and if there are statistics
1512 : * about the Var or expression (the latter is possible if it's indexed) then
1513 : * we'll produce a real estimate; otherwise it's just a default.
1514 : */
1515 : Selectivity
1516 43798 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1517 : {
1518 : VariableStatData vardata;
1519 : double selec;
1520 :
1521 43798 : examine_variable(root, arg, varRelid, &vardata);
1522 43798 : if (HeapTupleIsValid(vardata.statsTuple))
1523 : {
1524 : /*
1525 : * A boolean variable V is equivalent to the clause V = 't', so we
1526 : * compute the selectivity as if that is what we have.
1527 : */
1528 35650 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1529 : BoolGetDatum(true), false, true, false);
1530 : }
1531 : else
1532 : {
1533 : /* Otherwise, the default estimate is 0.5 */
1534 8148 : selec = 0.5;
1535 : }
1536 43798 : ReleaseVariableStats(vardata);
1537 43798 : return selec;
1538 : }
1539 :
1540 : /*
1541 : * booltestsel - Selectivity of BooleanTest Node.
1542 : */
1543 : Selectivity
1544 886 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1545 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1546 : {
1547 : VariableStatData vardata;
1548 : double selec;
1549 :
1550 886 : examine_variable(root, arg, varRelid, &vardata);
1551 :
1552 886 : if (HeapTupleIsValid(vardata.statsTuple))
1553 : {
1554 : Form_pg_statistic stats;
1555 : double freq_null;
1556 : AttStatsSlot sslot;
1557 :
1558 0 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1559 0 : freq_null = stats->stanullfrac;
1560 :
1561 0 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1562 : STATISTIC_KIND_MCV, InvalidOid,
1563 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1564 0 : && sslot.nnumbers > 0)
1565 0 : {
1566 : double freq_true;
1567 : double freq_false;
1568 :
1569 : /*
1570 : * Get first MCV frequency and derive frequency for true.
1571 : */
1572 0 : if (DatumGetBool(sslot.values[0]))
1573 0 : freq_true = sslot.numbers[0];
1574 : else
1575 0 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1576 :
1577 : /*
1578 : * Next derive frequency for false. Then use these as appropriate
1579 : * to derive frequency for each case.
1580 : */
1581 0 : freq_false = 1.0 - freq_true - freq_null;
1582 :
1583 0 : switch (booltesttype)
1584 : {
1585 0 : case IS_UNKNOWN:
1586 : /* select only NULL values */
1587 0 : selec = freq_null;
1588 0 : break;
1589 0 : case IS_NOT_UNKNOWN:
1590 : /* select non-NULL values */
1591 0 : selec = 1.0 - freq_null;
1592 0 : break;
1593 0 : case IS_TRUE:
1594 : /* select only TRUE values */
1595 0 : selec = freq_true;
1596 0 : break;
1597 0 : case IS_NOT_TRUE:
1598 : /* select non-TRUE values */
1599 0 : selec = 1.0 - freq_true;
1600 0 : break;
1601 0 : case IS_FALSE:
1602 : /* select only FALSE values */
1603 0 : selec = freq_false;
1604 0 : break;
1605 0 : case IS_NOT_FALSE:
1606 : /* select non-FALSE values */
1607 0 : selec = 1.0 - freq_false;
1608 0 : break;
1609 0 : default:
1610 0 : elog(ERROR, "unrecognized booltesttype: %d",
1611 : (int) booltesttype);
1612 : selec = 0.0; /* Keep compiler quiet */
1613 : break;
1614 : }
1615 :
1616 0 : free_attstatsslot(&sslot);
1617 : }
1618 : else
1619 : {
1620 : /*
1621 : * No most-common-value info available. Still have null fraction
1622 : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1623 : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1624 : */
1625 0 : switch (booltesttype)
1626 : {
1627 0 : case IS_UNKNOWN:
1628 : /* select only NULL values */
1629 0 : selec = freq_null;
1630 0 : break;
1631 0 : case IS_NOT_UNKNOWN:
1632 : /* select non-NULL values */
1633 0 : selec = 1.0 - freq_null;
1634 0 : break;
1635 0 : case IS_TRUE:
1636 : case IS_FALSE:
1637 : /* Assume we select half of the non-NULL values */
1638 0 : selec = (1.0 - freq_null) / 2.0;
1639 0 : break;
1640 0 : case IS_NOT_TRUE:
1641 : case IS_NOT_FALSE:
1642 : /* Assume we select NULLs plus half of the non-NULLs */
1643 : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1644 0 : selec = (freq_null + 1.0) / 2.0;
1645 0 : break;
1646 0 : default:
1647 0 : elog(ERROR, "unrecognized booltesttype: %d",
1648 : (int) booltesttype);
1649 : selec = 0.0; /* Keep compiler quiet */
1650 : break;
1651 : }
1652 : }
1653 : }
1654 : else
1655 : {
1656 : /*
1657 : * If we can't get variable statistics for the argument, perhaps
1658 : * clause_selectivity can do something with it. We ignore the
1659 : * possibility of a NULL value when using clause_selectivity, and just
1660 : * assume the value is either TRUE or FALSE.
1661 : */
1662 886 : switch (booltesttype)
1663 : {
1664 48 : case IS_UNKNOWN:
1665 48 : selec = DEFAULT_UNK_SEL;
1666 48 : break;
1667 108 : case IS_NOT_UNKNOWN:
1668 108 : selec = DEFAULT_NOT_UNK_SEL;
1669 108 : break;
1670 252 : case IS_TRUE:
1671 : case IS_NOT_FALSE:
1672 252 : selec = (double) clause_selectivity(root, arg,
1673 : varRelid,
1674 : jointype, sjinfo);
1675 252 : break;
1676 478 : case IS_FALSE:
1677 : case IS_NOT_TRUE:
1678 478 : selec = 1.0 - (double) clause_selectivity(root, arg,
1679 : varRelid,
1680 : jointype, sjinfo);
1681 478 : break;
1682 0 : default:
1683 0 : elog(ERROR, "unrecognized booltesttype: %d",
1684 : (int) booltesttype);
1685 : selec = 0.0; /* Keep compiler quiet */
1686 : break;
1687 : }
1688 : }
1689 :
1690 886 : ReleaseVariableStats(vardata);
1691 :
1692 : /* result should be in range, but make sure... */
1693 886 : CLAMP_PROBABILITY(selec);
1694 :
1695 886 : return (Selectivity) selec;
1696 : }
1697 :
1698 : /*
1699 : * nulltestsel - Selectivity of NullTest Node.
1700 : */
1701 : Selectivity
1702 17378 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1703 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1704 : {
1705 : VariableStatData vardata;
1706 : double selec;
1707 :
1708 17378 : examine_variable(root, arg, varRelid, &vardata);
1709 :
1710 17378 : if (HeapTupleIsValid(vardata.statsTuple))
1711 : {
1712 : Form_pg_statistic stats;
1713 : double freq_null;
1714 :
1715 9768 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1716 9768 : freq_null = stats->stanullfrac;
1717 :
1718 9768 : switch (nulltesttype)
1719 : {
1720 7224 : case IS_NULL:
1721 :
1722 : /*
1723 : * Use freq_null directly.
1724 : */
1725 7224 : selec = freq_null;
1726 7224 : break;
1727 2544 : case IS_NOT_NULL:
1728 :
1729 : /*
1730 : * Select not unknown (not null) values. Calculate from
1731 : * freq_null.
1732 : */
1733 2544 : selec = 1.0 - freq_null;
1734 2544 : break;
1735 0 : default:
1736 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1737 : (int) nulltesttype);
1738 : return (Selectivity) 0; /* keep compiler quiet */
1739 : }
1740 : }
1741 7610 : else if (vardata.var && IsA(vardata.var, Var) &&
1742 6858 : ((Var *) vardata.var)->varattno < 0)
1743 : {
1744 : /*
1745 : * There are no stats for system columns, but we know they are never
1746 : * NULL.
1747 : */
1748 60 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1749 : }
1750 : else
1751 : {
1752 : /*
1753 : * No ANALYZE stats available, so make a guess
1754 : */
1755 7550 : switch (nulltesttype)
1756 : {
1757 2108 : case IS_NULL:
1758 2108 : selec = DEFAULT_UNK_SEL;
1759 2108 : break;
1760 5442 : case IS_NOT_NULL:
1761 5442 : selec = DEFAULT_NOT_UNK_SEL;
1762 5442 : break;
1763 0 : default:
1764 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1765 : (int) nulltesttype);
1766 : return (Selectivity) 0; /* keep compiler quiet */
1767 : }
1768 : }
1769 :
1770 17378 : ReleaseVariableStats(vardata);
1771 :
1772 : /* result should be in range, but make sure... */
1773 17378 : CLAMP_PROBABILITY(selec);
1774 :
1775 17378 : return (Selectivity) selec;
1776 : }
1777 :
1778 : /*
1779 : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1780 : *
1781 : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1782 : * but it seems possible that RelabelType might show up. Also, the planner
1783 : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1784 : * so we need to be ready to deal with more than one level.
1785 : */
1786 : static Node *
1787 124640 : strip_array_coercion(Node *node)
1788 : {
1789 : for (;;)
1790 : {
1791 124722 : if (node && IsA(node, ArrayCoerceExpr))
1792 82 : {
1793 2934 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1794 :
1795 : /*
1796 : * If the per-element expression is just a RelabelType on top of
1797 : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1798 : */
1799 2934 : if (IsA(acoerce->elemexpr, RelabelType) &&
1800 82 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1801 82 : node = (Node *) acoerce->arg;
1802 : else
1803 : break;
1804 : }
1805 121788 : else if (node && IsA(node, RelabelType))
1806 : {
1807 : /* We don't really expect this case, but may as well cope */
1808 0 : node = (Node *) ((RelabelType *) node)->arg;
1809 : }
1810 : else
1811 : break;
1812 : }
1813 124640 : return node;
1814 : }
1815 :
1816 : /*
1817 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1818 : */
1819 : Selectivity
1820 21260 : scalararraysel(PlannerInfo *root,
1821 : ScalarArrayOpExpr *clause,
1822 : bool is_join_clause,
1823 : int varRelid,
1824 : JoinType jointype,
1825 : SpecialJoinInfo *sjinfo)
1826 : {
1827 21260 : Oid operator = clause->opno;
1828 21260 : bool useOr = clause->useOr;
1829 21260 : bool isEquality = false;
1830 21260 : bool isInequality = false;
1831 : Node *leftop;
1832 : Node *rightop;
1833 : Oid nominal_element_type;
1834 : Oid nominal_element_collation;
1835 : TypeCacheEntry *typentry;
1836 : RegProcedure oprsel;
1837 : FmgrInfo oprselproc;
1838 : Selectivity s1;
1839 : Selectivity s1disjoint;
1840 :
1841 : /* First, deconstruct the expression */
1842 : Assert(list_length(clause->args) == 2);
1843 21260 : leftop = (Node *) linitial(clause->args);
1844 21260 : rightop = (Node *) lsecond(clause->args);
1845 :
1846 : /* aggressively reduce both sides to constants */
1847 21260 : leftop = estimate_expression_value(root, leftop);
1848 21260 : rightop = estimate_expression_value(root, rightop);
1849 :
1850 : /* get nominal (after relabeling) element type of rightop */
1851 21260 : nominal_element_type = get_base_element_type(exprType(rightop));
1852 21260 : if (!OidIsValid(nominal_element_type))
1853 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1854 : /* get nominal collation, too, for generating constants */
1855 21260 : nominal_element_collation = exprCollation(rightop);
1856 :
1857 : /* look through any binary-compatible relabeling of rightop */
1858 21260 : rightop = strip_array_coercion(rightop);
1859 :
1860 : /*
1861 : * Detect whether the operator is the default equality or inequality
1862 : * operator of the array element type.
1863 : */
1864 21260 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1865 21260 : if (OidIsValid(typentry->eq_opr))
1866 : {
1867 21256 : if (operator == typentry->eq_opr)
1868 17982 : isEquality = true;
1869 3274 : else if (get_negator(operator) == typentry->eq_opr)
1870 2708 : isInequality = true;
1871 : }
1872 :
1873 : /*
1874 : * If it is equality or inequality, we might be able to estimate this as a
1875 : * form of array containment; for instance "const = ANY(column)" can be
1876 : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1877 : * that, and returns the selectivity estimate if successful, or -1 if not.
1878 : */
1879 21260 : if ((isEquality || isInequality) && !is_join_clause)
1880 : {
1881 20690 : s1 = scalararraysel_containment(root, leftop, rightop,
1882 : nominal_element_type,
1883 : isEquality, useOr, varRelid);
1884 20690 : if (s1 >= 0.0)
1885 118 : return s1;
1886 : }
1887 :
1888 : /*
1889 : * Look up the underlying operator's selectivity estimator. Punt if it
1890 : * hasn't got one.
1891 : */
1892 21142 : if (is_join_clause)
1893 0 : oprsel = get_oprjoin(operator);
1894 : else
1895 21142 : oprsel = get_oprrest(operator);
1896 21142 : if (!oprsel)
1897 4 : return (Selectivity) 0.5;
1898 21138 : fmgr_info(oprsel, &oprselproc);
1899 :
1900 : /*
1901 : * In the array-containment check above, we must only believe that an
1902 : * operator is equality or inequality if it is the default btree equality
1903 : * operator (or its negator) for the element type, since those are the
1904 : * operators that array containment will use. But in what follows, we can
1905 : * be a little laxer, and also believe that any operators using eqsel() or
1906 : * neqsel() as selectivity estimator act like equality or inequality.
1907 : */
1908 21138 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1909 18060 : isEquality = true;
1910 3078 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1911 2598 : isInequality = true;
1912 :
1913 : /*
1914 : * We consider three cases:
1915 : *
1916 : * 1. rightop is an Array constant: deconstruct the array, apply the
1917 : * operator's selectivity function for each array element, and merge the
1918 : * results in the same way that clausesel.c does for AND/OR combinations.
1919 : *
1920 : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
1921 : * function for each element of the ARRAY[] construct, and merge.
1922 : *
1923 : * 3. otherwise, make a guess ...
1924 : */
1925 21138 : if (rightop && IsA(rightop, Const))
1926 16974 : {
1927 17004 : Datum arraydatum = ((Const *) rightop)->constvalue;
1928 17004 : bool arrayisnull = ((Const *) rightop)->constisnull;
1929 : ArrayType *arrayval;
1930 : int16 elmlen;
1931 : bool elmbyval;
1932 : char elmalign;
1933 : int num_elems;
1934 : Datum *elem_values;
1935 : bool *elem_nulls;
1936 : int i;
1937 :
1938 17004 : if (arrayisnull) /* qual can't succeed if null array */
1939 30 : return (Selectivity) 0.0;
1940 16974 : arrayval = DatumGetArrayTypeP(arraydatum);
1941 16974 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
1942 : &elmlen, &elmbyval, &elmalign);
1943 16974 : deconstruct_array(arrayval,
1944 : ARR_ELEMTYPE(arrayval),
1945 : elmlen, elmbyval, elmalign,
1946 : &elem_values, &elem_nulls, &num_elems);
1947 :
1948 : /*
1949 : * For generic operators, we assume the probability of success is
1950 : * independent for each array element. But for "= ANY" or "<> ALL",
1951 : * if the array elements are distinct (which'd typically be the case)
1952 : * then the probabilities are disjoint, and we should just sum them.
1953 : *
1954 : * If we were being really tense we would try to confirm that the
1955 : * elements are all distinct, but that would be expensive and it
1956 : * doesn't seem to be worth the cycles; it would amount to penalizing
1957 : * well-written queries in favor of poorly-written ones. However, we
1958 : * do protect ourselves a little bit by checking whether the
1959 : * disjointness assumption leads to an impossible (out of range)
1960 : * probability; if so, we fall back to the normal calculation.
1961 : */
1962 16974 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
1963 :
1964 73822 : for (i = 0; i < num_elems; i++)
1965 : {
1966 : List *args;
1967 : Selectivity s2;
1968 :
1969 56848 : args = list_make2(leftop,
1970 : makeConst(nominal_element_type,
1971 : -1,
1972 : nominal_element_collation,
1973 : elmlen,
1974 : elem_values[i],
1975 : elem_nulls[i],
1976 : elmbyval));
1977 56848 : if (is_join_clause)
1978 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1979 : clause->inputcollid,
1980 : PointerGetDatum(root),
1981 : ObjectIdGetDatum(operator),
1982 : PointerGetDatum(args),
1983 : Int16GetDatum(jointype),
1984 : PointerGetDatum(sjinfo)));
1985 : else
1986 56848 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1987 : clause->inputcollid,
1988 : PointerGetDatum(root),
1989 : ObjectIdGetDatum(operator),
1990 : PointerGetDatum(args),
1991 : Int32GetDatum(varRelid)));
1992 :
1993 56848 : if (useOr)
1994 : {
1995 48466 : s1 = s1 + s2 - s1 * s2;
1996 48466 : if (isEquality)
1997 47422 : s1disjoint += s2;
1998 : }
1999 : else
2000 : {
2001 8382 : s1 = s1 * s2;
2002 8382 : if (isInequality)
2003 8070 : s1disjoint += s2 - 1.0;
2004 : }
2005 : }
2006 :
2007 : /* accept disjoint-probability estimate if in range */
2008 16974 : if ((useOr ? isEquality : isInequality) &&
2009 16334 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2010 16304 : s1 = s1disjoint;
2011 : }
2012 4134 : else if (rightop && IsA(rightop, ArrayExpr) &&
2013 294 : !((ArrayExpr *) rightop)->multidims)
2014 294 : {
2015 294 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2016 : int16 elmlen;
2017 : bool elmbyval;
2018 : ListCell *l;
2019 :
2020 294 : get_typlenbyval(arrayexpr->element_typeid,
2021 : &elmlen, &elmbyval);
2022 :
2023 : /*
2024 : * We use the assumption of disjoint probabilities here too, although
2025 : * the odds of equal array elements are rather higher if the elements
2026 : * are not all constants (which they won't be, else constant folding
2027 : * would have reduced the ArrayExpr to a Const). In this path it's
2028 : * critical to have the sanity check on the s1disjoint estimate.
2029 : */
2030 294 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2031 :
2032 1060 : foreach(l, arrayexpr->elements)
2033 : {
2034 766 : Node *elem = (Node *) lfirst(l);
2035 : List *args;
2036 : Selectivity s2;
2037 :
2038 : /*
2039 : * Theoretically, if elem isn't of nominal_element_type we should
2040 : * insert a RelabelType, but it seems unlikely that any operator
2041 : * estimation function would really care ...
2042 : */
2043 766 : args = list_make2(leftop, elem);
2044 766 : if (is_join_clause)
2045 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2046 : clause->inputcollid,
2047 : PointerGetDatum(root),
2048 : ObjectIdGetDatum(operator),
2049 : PointerGetDatum(args),
2050 : Int16GetDatum(jointype),
2051 : PointerGetDatum(sjinfo)));
2052 : else
2053 766 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2054 : clause->inputcollid,
2055 : PointerGetDatum(root),
2056 : ObjectIdGetDatum(operator),
2057 : PointerGetDatum(args),
2058 : Int32GetDatum(varRelid)));
2059 :
2060 766 : if (useOr)
2061 : {
2062 766 : s1 = s1 + s2 - s1 * s2;
2063 766 : if (isEquality)
2064 766 : s1disjoint += s2;
2065 : }
2066 : else
2067 : {
2068 0 : s1 = s1 * s2;
2069 0 : if (isInequality)
2070 0 : s1disjoint += s2 - 1.0;
2071 : }
2072 : }
2073 :
2074 : /* accept disjoint-probability estimate if in range */
2075 294 : if ((useOr ? isEquality : isInequality) &&
2076 294 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2077 294 : s1 = s1disjoint;
2078 : }
2079 : else
2080 : {
2081 : CaseTestExpr *dummyexpr;
2082 : List *args;
2083 : Selectivity s2;
2084 : int i;
2085 :
2086 : /*
2087 : * We need a dummy rightop to pass to the operator selectivity
2088 : * routine. It can be pretty much anything that doesn't look like a
2089 : * constant; CaseTestExpr is a convenient choice.
2090 : */
2091 3840 : dummyexpr = makeNode(CaseTestExpr);
2092 3840 : dummyexpr->typeId = nominal_element_type;
2093 3840 : dummyexpr->typeMod = -1;
2094 3840 : dummyexpr->collation = clause->inputcollid;
2095 3840 : args = list_make2(leftop, dummyexpr);
2096 3840 : if (is_join_clause)
2097 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2098 : clause->inputcollid,
2099 : PointerGetDatum(root),
2100 : ObjectIdGetDatum(operator),
2101 : PointerGetDatum(args),
2102 : Int16GetDatum(jointype),
2103 : PointerGetDatum(sjinfo)));
2104 : else
2105 3840 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2106 : clause->inputcollid,
2107 : PointerGetDatum(root),
2108 : ObjectIdGetDatum(operator),
2109 : PointerGetDatum(args),
2110 : Int32GetDatum(varRelid)));
2111 3840 : s1 = useOr ? 0.0 : 1.0;
2112 :
2113 : /*
2114 : * Arbitrarily assume 10 elements in the eventual array value (see
2115 : * also estimate_array_length). We don't risk an assumption of
2116 : * disjoint probabilities here.
2117 : */
2118 42240 : for (i = 0; i < 10; i++)
2119 : {
2120 38400 : if (useOr)
2121 38400 : s1 = s1 + s2 - s1 * s2;
2122 : else
2123 0 : s1 = s1 * s2;
2124 : }
2125 : }
2126 :
2127 : /* result should be in range, but make sure... */
2128 21108 : CLAMP_PROBABILITY(s1);
2129 :
2130 21108 : return s1;
2131 : }
2132 :
2133 : /*
2134 : * Estimate number of elements in the array yielded by an expression.
2135 : *
2136 : * Note: the result is integral, but we use "double" to avoid overflow
2137 : * concerns. Most callers will use it in double-type expressions anyway.
2138 : *
2139 : * Note: in some code paths root can be passed as NULL, resulting in
2140 : * slightly worse estimates.
2141 : */
2142 : double
2143 103380 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2144 : {
2145 : /* look through any binary-compatible relabeling of arrayexpr */
2146 103380 : arrayexpr = strip_array_coercion(arrayexpr);
2147 :
2148 103380 : if (arrayexpr && IsA(arrayexpr, Const))
2149 : {
2150 45532 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2151 45532 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2152 : ArrayType *arrayval;
2153 :
2154 45532 : if (arrayisnull)
2155 90 : return 0;
2156 45442 : arrayval = DatumGetArrayTypeP(arraydatum);
2157 45442 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2158 : }
2159 57848 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2160 498 : !((ArrayExpr *) arrayexpr)->multidims)
2161 : {
2162 498 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2163 : }
2164 57350 : else if (arrayexpr && root)
2165 : {
2166 : /* See if we can find any statistics about it */
2167 : VariableStatData vardata;
2168 : AttStatsSlot sslot;
2169 57350 : double nelem = 0;
2170 :
2171 57350 : examine_variable(root, arrayexpr, 0, &vardata);
2172 57350 : if (HeapTupleIsValid(vardata.statsTuple))
2173 : {
2174 : /*
2175 : * Found stats, so use the average element count, which is stored
2176 : * in the last stanumbers element of the DECHIST statistics.
2177 : * Actually that is the average count of *distinct* elements;
2178 : * perhaps we should scale it up somewhat?
2179 : */
2180 6960 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2181 : STATISTIC_KIND_DECHIST, InvalidOid,
2182 : ATTSTATSSLOT_NUMBERS))
2183 : {
2184 6846 : if (sslot.nnumbers > 0)
2185 6846 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2186 6846 : free_attstatsslot(&sslot);
2187 : }
2188 : }
2189 57350 : ReleaseVariableStats(vardata);
2190 :
2191 57350 : if (nelem > 0)
2192 6846 : return nelem;
2193 : }
2194 :
2195 : /* Else use a default guess --- this should match scalararraysel */
2196 50504 : return 10;
2197 : }
2198 :
2199 : /*
2200 : * rowcomparesel - Selectivity of RowCompareExpr Node.
2201 : *
2202 : * We estimate RowCompare selectivity by considering just the first (high
2203 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2204 : * this estimate could be refined by considering additional columns, it
2205 : * seems unlikely that we could do a lot better without multi-column
2206 : * statistics.
2207 : */
2208 : Selectivity
2209 252 : rowcomparesel(PlannerInfo *root,
2210 : RowCompareExpr *clause,
2211 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2212 : {
2213 : Selectivity s1;
2214 252 : Oid opno = linitial_oid(clause->opnos);
2215 252 : Oid inputcollid = linitial_oid(clause->inputcollids);
2216 : List *opargs;
2217 : bool is_join_clause;
2218 :
2219 : /* Build equivalent arg list for single operator */
2220 252 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2221 :
2222 : /*
2223 : * Decide if it's a join clause. This should match clausesel.c's
2224 : * treat_as_join_clause(), except that we intentionally consider only the
2225 : * leading columns and not the rest of the clause.
2226 : */
2227 252 : if (varRelid != 0)
2228 : {
2229 : /*
2230 : * Caller is forcing restriction mode (eg, because we are examining an
2231 : * inner indexscan qual).
2232 : */
2233 54 : is_join_clause = false;
2234 : }
2235 198 : else if (sjinfo == NULL)
2236 : {
2237 : /*
2238 : * It must be a restriction clause, since it's being evaluated at a
2239 : * scan node.
2240 : */
2241 186 : is_join_clause = false;
2242 : }
2243 : else
2244 : {
2245 : /*
2246 : * Otherwise, it's a join if there's more than one base relation used.
2247 : */
2248 12 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2249 : }
2250 :
2251 252 : if (is_join_clause)
2252 : {
2253 : /* Estimate selectivity for a join clause. */
2254 12 : s1 = join_selectivity(root, opno,
2255 : opargs,
2256 : inputcollid,
2257 : jointype,
2258 : sjinfo);
2259 : }
2260 : else
2261 : {
2262 : /* Estimate selectivity for a restriction clause. */
2263 240 : s1 = restriction_selectivity(root, opno,
2264 : opargs,
2265 : inputcollid,
2266 : varRelid);
2267 : }
2268 :
2269 252 : return s1;
2270 : }
2271 :
2272 : /*
2273 : * eqjoinsel - Join selectivity of "="
2274 : */
2275 : Datum
2276 224080 : eqjoinsel(PG_FUNCTION_ARGS)
2277 : {
2278 224080 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2279 224080 : Oid operator = PG_GETARG_OID(1);
2280 224080 : List *args = (List *) PG_GETARG_POINTER(2);
2281 :
2282 : #ifdef NOT_USED
2283 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2284 : #endif
2285 224080 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2286 224080 : Oid collation = PG_GET_COLLATION();
2287 : double selec;
2288 : double selec_inner;
2289 : VariableStatData vardata1;
2290 : VariableStatData vardata2;
2291 : double nd1;
2292 : double nd2;
2293 : bool isdefault1;
2294 : bool isdefault2;
2295 : Oid opfuncoid;
2296 : AttStatsSlot sslot1;
2297 : AttStatsSlot sslot2;
2298 224080 : Form_pg_statistic stats1 = NULL;
2299 224080 : Form_pg_statistic stats2 = NULL;
2300 224080 : bool have_mcvs1 = false;
2301 224080 : bool have_mcvs2 = false;
2302 : bool get_mcv_stats;
2303 : bool join_is_reversed;
2304 : RelOptInfo *inner_rel;
2305 :
2306 224080 : get_join_variables(root, args, sjinfo,
2307 : &vardata1, &vardata2, &join_is_reversed);
2308 :
2309 224080 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2310 224080 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2311 :
2312 224080 : opfuncoid = get_opcode(operator);
2313 :
2314 224080 : memset(&sslot1, 0, sizeof(sslot1));
2315 224080 : memset(&sslot2, 0, sizeof(sslot2));
2316 :
2317 : /*
2318 : * There is no use in fetching one side's MCVs if we lack MCVs for the
2319 : * other side, so do a quick check to verify that both stats exist.
2320 : */
2321 629044 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2322 318198 : HeapTupleIsValid(vardata2.statsTuple) &&
2323 137314 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2324 : STATISTIC_KIND_MCV, InvalidOid,
2325 404964 : 0) &&
2326 54030 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2327 : STATISTIC_KIND_MCV, InvalidOid,
2328 : 0));
2329 :
2330 224080 : if (HeapTupleIsValid(vardata1.statsTuple))
2331 : {
2332 : /* note we allow use of nullfrac regardless of security check */
2333 180884 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2334 195328 : if (get_mcv_stats &&
2335 14444 : statistic_proc_security_check(&vardata1, opfuncoid))
2336 14444 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2337 : STATISTIC_KIND_MCV, InvalidOid,
2338 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2339 : }
2340 :
2341 224080 : if (HeapTupleIsValid(vardata2.statsTuple))
2342 : {
2343 : /* note we allow use of nullfrac regardless of security check */
2344 149360 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2345 163804 : if (get_mcv_stats &&
2346 14444 : statistic_proc_security_check(&vardata2, opfuncoid))
2347 14444 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2348 : STATISTIC_KIND_MCV, InvalidOid,
2349 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2350 : }
2351 :
2352 : /* We need to compute the inner-join selectivity in all cases */
2353 224080 : selec_inner = eqjoinsel_inner(opfuncoid, collation,
2354 : &vardata1, &vardata2,
2355 : nd1, nd2,
2356 : isdefault1, isdefault2,
2357 : &sslot1, &sslot2,
2358 : stats1, stats2,
2359 : have_mcvs1, have_mcvs2);
2360 :
2361 224080 : switch (sjinfo->jointype)
2362 : {
2363 213390 : case JOIN_INNER:
2364 : case JOIN_LEFT:
2365 : case JOIN_FULL:
2366 213390 : selec = selec_inner;
2367 213390 : break;
2368 10690 : case JOIN_SEMI:
2369 : case JOIN_ANTI:
2370 :
2371 : /*
2372 : * Look up the join's inner relation. min_righthand is sufficient
2373 : * information because neither SEMI nor ANTI joins permit any
2374 : * reassociation into or out of their RHS, so the righthand will
2375 : * always be exactly that set of rels.
2376 : */
2377 10690 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2378 :
2379 10690 : if (!join_is_reversed)
2380 6580 : selec = eqjoinsel_semi(opfuncoid, collation,
2381 : &vardata1, &vardata2,
2382 : nd1, nd2,
2383 : isdefault1, isdefault2,
2384 : &sslot1, &sslot2,
2385 : stats1, stats2,
2386 : have_mcvs1, have_mcvs2,
2387 : inner_rel);
2388 : else
2389 : {
2390 4110 : Oid commop = get_commutator(operator);
2391 4110 : Oid commopfuncoid = OidIsValid(commop) ? get_opcode(commop) : InvalidOid;
2392 :
2393 4110 : selec = eqjoinsel_semi(commopfuncoid, collation,
2394 : &vardata2, &vardata1,
2395 : nd2, nd1,
2396 : isdefault2, isdefault1,
2397 : &sslot2, &sslot1,
2398 : stats2, stats1,
2399 : have_mcvs2, have_mcvs1,
2400 : inner_rel);
2401 : }
2402 :
2403 : /*
2404 : * We should never estimate the output of a semijoin to be more
2405 : * rows than we estimate for an inner join with the same input
2406 : * rels and join condition; it's obviously impossible for that to
2407 : * happen. The former estimate is N1 * Ssemi while the latter is
2408 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2409 : * this is worthwhile because of the shakier estimation rules we
2410 : * use in eqjoinsel_semi, particularly in cases where it has to
2411 : * punt entirely.
2412 : */
2413 10690 : selec = Min(selec, inner_rel->rows * selec_inner);
2414 10690 : break;
2415 0 : default:
2416 : /* other values not expected here */
2417 0 : elog(ERROR, "unrecognized join type: %d",
2418 : (int) sjinfo->jointype);
2419 : selec = 0; /* keep compiler quiet */
2420 : break;
2421 : }
2422 :
2423 224080 : free_attstatsslot(&sslot1);
2424 224080 : free_attstatsslot(&sslot2);
2425 :
2426 224080 : ReleaseVariableStats(vardata1);
2427 224080 : ReleaseVariableStats(vardata2);
2428 :
2429 224080 : CLAMP_PROBABILITY(selec);
2430 :
2431 224080 : PG_RETURN_FLOAT8((float8) selec);
2432 : }
2433 :
2434 : /*
2435 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2436 : *
2437 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2438 : * that it's worth trying to distinguish them here.
2439 : */
2440 : static double
2441 224080 : eqjoinsel_inner(Oid opfuncoid, Oid collation,
2442 : VariableStatData *vardata1, VariableStatData *vardata2,
2443 : double nd1, double nd2,
2444 : bool isdefault1, bool isdefault2,
2445 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2446 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2447 : bool have_mcvs1, bool have_mcvs2)
2448 : {
2449 : double selec;
2450 :
2451 224080 : if (have_mcvs1 && have_mcvs2)
2452 14444 : {
2453 : /*
2454 : * We have most-common-value lists for both relations. Run through
2455 : * the lists to see which MCVs actually join to each other with the
2456 : * given operator. This allows us to determine the exact join
2457 : * selectivity for the portion of the relations represented by the MCV
2458 : * lists. We still have to estimate for the remaining population, but
2459 : * in a skewed distribution this gives us a big leg up in accuracy.
2460 : * For motivation see the analysis in Y. Ioannidis and S.
2461 : * Christodoulakis, "On the propagation of errors in the size of join
2462 : * results", Technical Report 1018, Computer Science Dept., University
2463 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2464 : */
2465 14444 : LOCAL_FCINFO(fcinfo, 2);
2466 : FmgrInfo eqproc;
2467 : bool *hasmatch1;
2468 : bool *hasmatch2;
2469 14444 : double nullfrac1 = stats1->stanullfrac;
2470 14444 : double nullfrac2 = stats2->stanullfrac;
2471 : double matchprodfreq,
2472 : matchfreq1,
2473 : matchfreq2,
2474 : unmatchfreq1,
2475 : unmatchfreq2,
2476 : otherfreq1,
2477 : otherfreq2,
2478 : totalsel1,
2479 : totalsel2;
2480 : int i,
2481 : nmatches;
2482 :
2483 14444 : fmgr_info(opfuncoid, &eqproc);
2484 :
2485 : /*
2486 : * Save a few cycles by setting up the fcinfo struct just once. Using
2487 : * FunctionCallInvoke directly also avoids failure if the eqproc
2488 : * returns NULL, though really equality functions should never do
2489 : * that.
2490 : */
2491 14444 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2492 : NULL, NULL);
2493 14444 : fcinfo->args[0].isnull = false;
2494 14444 : fcinfo->args[1].isnull = false;
2495 :
2496 14444 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2497 14444 : hasmatch2 = (bool *) palloc0(sslot2->nvalues * sizeof(bool));
2498 :
2499 : /*
2500 : * Note we assume that each MCV will match at most one member of the
2501 : * other MCV list. If the operator isn't really equality, there could
2502 : * be multiple matches --- but we don't look for them, both for speed
2503 : * and because the math wouldn't add up...
2504 : */
2505 14444 : matchprodfreq = 0.0;
2506 14444 : nmatches = 0;
2507 588716 : for (i = 0; i < sslot1->nvalues; i++)
2508 : {
2509 : int j;
2510 :
2511 574272 : fcinfo->args[0].value = sslot1->values[i];
2512 :
2513 22590482 : for (j = 0; j < sslot2->nvalues; j++)
2514 : {
2515 : Datum fresult;
2516 :
2517 22210998 : if (hasmatch2[j])
2518 5799468 : continue;
2519 16411530 : fcinfo->args[1].value = sslot2->values[j];
2520 16411530 : fcinfo->isnull = false;
2521 16411530 : fresult = FunctionCallInvoke(fcinfo);
2522 16411530 : if (!fcinfo->isnull && DatumGetBool(fresult))
2523 : {
2524 194788 : hasmatch1[i] = hasmatch2[j] = true;
2525 194788 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
2526 194788 : nmatches++;
2527 194788 : break;
2528 : }
2529 : }
2530 : }
2531 14444 : CLAMP_PROBABILITY(matchprodfreq);
2532 : /* Sum up frequencies of matched and unmatched MCVs */
2533 14444 : matchfreq1 = unmatchfreq1 = 0.0;
2534 588716 : for (i = 0; i < sslot1->nvalues; i++)
2535 : {
2536 574272 : if (hasmatch1[i])
2537 194788 : matchfreq1 += sslot1->numbers[i];
2538 : else
2539 379484 : unmatchfreq1 += sslot1->numbers[i];
2540 : }
2541 14444 : CLAMP_PROBABILITY(matchfreq1);
2542 14444 : CLAMP_PROBABILITY(unmatchfreq1);
2543 14444 : matchfreq2 = unmatchfreq2 = 0.0;
2544 413826 : for (i = 0; i < sslot2->nvalues; i++)
2545 : {
2546 399382 : if (hasmatch2[i])
2547 194788 : matchfreq2 += sslot2->numbers[i];
2548 : else
2549 204594 : unmatchfreq2 += sslot2->numbers[i];
2550 : }
2551 14444 : CLAMP_PROBABILITY(matchfreq2);
2552 14444 : CLAMP_PROBABILITY(unmatchfreq2);
2553 14444 : pfree(hasmatch1);
2554 14444 : pfree(hasmatch2);
2555 :
2556 : /*
2557 : * Compute total frequency of non-null values that are not in the MCV
2558 : * lists.
2559 : */
2560 14444 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2561 14444 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2562 14444 : CLAMP_PROBABILITY(otherfreq1);
2563 14444 : CLAMP_PROBABILITY(otherfreq2);
2564 :
2565 : /*
2566 : * We can estimate the total selectivity from the point of view of
2567 : * relation 1 as: the known selectivity for matched MCVs, plus
2568 : * unmatched MCVs that are assumed to match against random members of
2569 : * relation 2's non-MCV population, plus non-MCV values that are
2570 : * assumed to match against random members of relation 2's unmatched
2571 : * MCVs plus non-MCV values.
2572 : */
2573 14444 : totalsel1 = matchprodfreq;
2574 14444 : if (nd2 > sslot2->nvalues)
2575 6292 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2576 14444 : if (nd2 > nmatches)
2577 11148 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2578 11148 : (nd2 - nmatches);
2579 : /* Same estimate from the point of view of relation 2. */
2580 14444 : totalsel2 = matchprodfreq;
2581 14444 : if (nd1 > sslot1->nvalues)
2582 6940 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2583 14444 : if (nd1 > nmatches)
2584 9884 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2585 9884 : (nd1 - nmatches);
2586 :
2587 : /*
2588 : * Use the smaller of the two estimates. This can be justified in
2589 : * essentially the same terms as given below for the no-stats case: to
2590 : * a first approximation, we are estimating from the point of view of
2591 : * the relation with smaller nd.
2592 : */
2593 14444 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2594 : }
2595 : else
2596 : {
2597 : /*
2598 : * We do not have MCV lists for both sides. Estimate the join
2599 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2600 : * is plausible if we assume that the join operator is strict and the
2601 : * non-null values are about equally distributed: a given non-null
2602 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2603 : * of rel2, so total join rows are at most
2604 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2605 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2606 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2607 : * with MIN() is an upper bound. Using the MIN() means we estimate
2608 : * from the point of view of the relation with smaller nd (since the
2609 : * larger nd is determining the MIN). It is reasonable to assume that
2610 : * most tuples in this rel will have join partners, so the bound is
2611 : * probably reasonably tight and should be taken as-is.
2612 : *
2613 : * XXX Can we be smarter if we have an MCV list for just one side? It
2614 : * seems that if we assume equal distribution for the other side, we
2615 : * end up with the same answer anyway.
2616 : */
2617 209636 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2618 209636 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2619 :
2620 209636 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2621 209636 : if (nd1 > nd2)
2622 107742 : selec /= nd1;
2623 : else
2624 101894 : selec /= nd2;
2625 : }
2626 :
2627 224080 : return selec;
2628 : }
2629 :
2630 : /*
2631 : * eqjoinsel_semi --- eqjoinsel for semi join
2632 : *
2633 : * (Also used for anti join, which we are supposed to estimate the same way.)
2634 : * Caller has ensured that vardata1 is the LHS variable.
2635 : * Unlike eqjoinsel_inner, we have to cope with opfuncoid being InvalidOid.
2636 : */
2637 : static double
2638 10690 : eqjoinsel_semi(Oid opfuncoid, Oid collation,
2639 : VariableStatData *vardata1, VariableStatData *vardata2,
2640 : double nd1, double nd2,
2641 : bool isdefault1, bool isdefault2,
2642 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2643 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2644 : bool have_mcvs1, bool have_mcvs2,
2645 : RelOptInfo *inner_rel)
2646 : {
2647 : double selec;
2648 :
2649 : /*
2650 : * We clamp nd2 to be not more than what we estimate the inner relation's
2651 : * size to be. This is intuitively somewhat reasonable since obviously
2652 : * there can't be more than that many distinct values coming from the
2653 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2654 : * likewise) is that this is the only pathway by which restriction clauses
2655 : * applied to the inner rel will affect the join result size estimate,
2656 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2657 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2658 : * the selectivity of outer-rel restrictions.
2659 : *
2660 : * We can apply this clamping both with respect to the base relation from
2661 : * which the join variable comes (if there is just one), and to the
2662 : * immediate inner input relation of the current join.
2663 : *
2664 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2665 : * great, maybe, but it didn't come out of nowhere either. This is most
2666 : * helpful when the inner relation is empty and consequently has no stats.
2667 : */
2668 10690 : if (vardata2->rel)
2669 : {
2670 10684 : if (nd2 >= vardata2->rel->rows)
2671 : {
2672 8562 : nd2 = vardata2->rel->rows;
2673 8562 : isdefault2 = false;
2674 : }
2675 : }
2676 10690 : if (nd2 >= inner_rel->rows)
2677 : {
2678 8504 : nd2 = inner_rel->rows;
2679 8504 : isdefault2 = false;
2680 : }
2681 :
2682 10690 : if (have_mcvs1 && have_mcvs2 && OidIsValid(opfuncoid))
2683 612 : {
2684 : /*
2685 : * We have most-common-value lists for both relations. Run through
2686 : * the lists to see which MCVs actually join to each other with the
2687 : * given operator. This allows us to determine the exact join
2688 : * selectivity for the portion of the relations represented by the MCV
2689 : * lists. We still have to estimate for the remaining population, but
2690 : * in a skewed distribution this gives us a big leg up in accuracy.
2691 : */
2692 612 : LOCAL_FCINFO(fcinfo, 2);
2693 : FmgrInfo eqproc;
2694 : bool *hasmatch1;
2695 : bool *hasmatch2;
2696 612 : double nullfrac1 = stats1->stanullfrac;
2697 : double matchfreq1,
2698 : uncertainfrac,
2699 : uncertain;
2700 : int i,
2701 : nmatches,
2702 : clamped_nvalues2;
2703 :
2704 : /*
2705 : * The clamping above could have resulted in nd2 being less than
2706 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2707 : * most common values in the relation will appear in the join input,
2708 : * and so compare to only the first nd2 members of the MCV list. Of
2709 : * course this is frequently wrong, but it's the best bet we can make.
2710 : */
2711 612 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2712 :
2713 612 : fmgr_info(opfuncoid, &eqproc);
2714 :
2715 : /*
2716 : * Save a few cycles by setting up the fcinfo struct just once. Using
2717 : * FunctionCallInvoke directly also avoids failure if the eqproc
2718 : * returns NULL, though really equality functions should never do
2719 : * that.
2720 : */
2721 612 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2722 : NULL, NULL);
2723 612 : fcinfo->args[0].isnull = false;
2724 612 : fcinfo->args[1].isnull = false;
2725 :
2726 612 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2727 612 : hasmatch2 = (bool *) palloc0(clamped_nvalues2 * sizeof(bool));
2728 :
2729 : /*
2730 : * Note we assume that each MCV will match at most one member of the
2731 : * other MCV list. If the operator isn't really equality, there could
2732 : * be multiple matches --- but we don't look for them, both for speed
2733 : * and because the math wouldn't add up...
2734 : */
2735 612 : nmatches = 0;
2736 12772 : for (i = 0; i < sslot1->nvalues; i++)
2737 : {
2738 : int j;
2739 :
2740 12160 : fcinfo->args[0].value = sslot1->values[i];
2741 :
2742 457932 : for (j = 0; j < clamped_nvalues2; j++)
2743 : {
2744 : Datum fresult;
2745 :
2746 456650 : if (hasmatch2[j])
2747 382244 : continue;
2748 74406 : fcinfo->args[1].value = sslot2->values[j];
2749 74406 : fcinfo->isnull = false;
2750 74406 : fresult = FunctionCallInvoke(fcinfo);
2751 74406 : if (!fcinfo->isnull && DatumGetBool(fresult))
2752 : {
2753 10878 : hasmatch1[i] = hasmatch2[j] = true;
2754 10878 : nmatches++;
2755 10878 : break;
2756 : }
2757 : }
2758 : }
2759 : /* Sum up frequencies of matched MCVs */
2760 612 : matchfreq1 = 0.0;
2761 12772 : for (i = 0; i < sslot1->nvalues; i++)
2762 : {
2763 12160 : if (hasmatch1[i])
2764 10878 : matchfreq1 += sslot1->numbers[i];
2765 : }
2766 612 : CLAMP_PROBABILITY(matchfreq1);
2767 612 : pfree(hasmatch1);
2768 612 : pfree(hasmatch2);
2769 :
2770 : /*
2771 : * Now we need to estimate the fraction of relation 1 that has at
2772 : * least one join partner. We know for certain that the matched MCVs
2773 : * do, so that gives us a lower bound, but we're really in the dark
2774 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2775 : * assume all non-null rel1 rows have join partners, else assume for
2776 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2777 : * can discount the known-matched MCVs from the distinct-values counts
2778 : * before doing the division.
2779 : *
2780 : * Crude as the above is, it's completely useless if we don't have
2781 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2782 : * nd2 is default, punt and assume half of the uncertain rows have
2783 : * join partners.
2784 : */
2785 612 : if (!isdefault1 && !isdefault2)
2786 : {
2787 612 : nd1 -= nmatches;
2788 612 : nd2 -= nmatches;
2789 612 : if (nd1 <= nd2 || nd2 < 0)
2790 582 : uncertainfrac = 1.0;
2791 : else
2792 30 : uncertainfrac = nd2 / nd1;
2793 : }
2794 : else
2795 0 : uncertainfrac = 0.5;
2796 612 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2797 612 : CLAMP_PROBABILITY(uncertain);
2798 612 : selec = matchfreq1 + uncertainfrac * uncertain;
2799 : }
2800 : else
2801 : {
2802 : /*
2803 : * Without MCV lists for both sides, we can only use the heuristic
2804 : * about nd1 vs nd2.
2805 : */
2806 10078 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2807 :
2808 10078 : if (!isdefault1 && !isdefault2)
2809 : {
2810 7728 : if (nd1 <= nd2 || nd2 < 0)
2811 5084 : selec = 1.0 - nullfrac1;
2812 : else
2813 2644 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2814 : }
2815 : else
2816 2350 : selec = 0.5 * (1.0 - nullfrac1);
2817 : }
2818 :
2819 10690 : return selec;
2820 : }
2821 :
2822 : /*
2823 : * neqjoinsel - Join selectivity of "!="
2824 : */
2825 : Datum
2826 3700 : neqjoinsel(PG_FUNCTION_ARGS)
2827 : {
2828 3700 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2829 3700 : Oid operator = PG_GETARG_OID(1);
2830 3700 : List *args = (List *) PG_GETARG_POINTER(2);
2831 3700 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2832 3700 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2833 3700 : Oid collation = PG_GET_COLLATION();
2834 : float8 result;
2835 :
2836 3700 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
2837 1254 : {
2838 : /*
2839 : * For semi-joins, if there is more than one distinct value in the RHS
2840 : * relation then every non-null LHS row must find a row to join since
2841 : * it can only be equal to one of them. We'll assume that there is
2842 : * always more than one distinct RHS value for the sake of stability,
2843 : * though in theory we could have special cases for empty RHS
2844 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
2845 : * fraction of LHS that has the same value as the single RHS value).
2846 : *
2847 : * For anti-joins, if we use the same assumption that there is more
2848 : * than one distinct key in the RHS relation, then every non-null LHS
2849 : * row must be suppressed by the anti-join.
2850 : *
2851 : * So either way, the selectivity estimate should be 1 - nullfrac.
2852 : */
2853 : VariableStatData leftvar;
2854 : VariableStatData rightvar;
2855 : bool reversed;
2856 : HeapTuple statsTuple;
2857 : double nullfrac;
2858 :
2859 1254 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
2860 1254 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
2861 1254 : if (HeapTupleIsValid(statsTuple))
2862 1022 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
2863 : else
2864 232 : nullfrac = 0.0;
2865 1254 : ReleaseVariableStats(leftvar);
2866 1254 : ReleaseVariableStats(rightvar);
2867 :
2868 1254 : result = 1.0 - nullfrac;
2869 : }
2870 : else
2871 : {
2872 : /*
2873 : * We want 1 - eqjoinsel() where the equality operator is the one
2874 : * associated with this != operator, that is, its negator.
2875 : */
2876 2446 : Oid eqop = get_negator(operator);
2877 :
2878 2446 : if (eqop)
2879 : {
2880 : result =
2881 2446 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
2882 : collation,
2883 : PointerGetDatum(root),
2884 : ObjectIdGetDatum(eqop),
2885 : PointerGetDatum(args),
2886 : Int16GetDatum(jointype),
2887 : PointerGetDatum(sjinfo)));
2888 : }
2889 : else
2890 : {
2891 : /* Use default selectivity (should we raise an error instead?) */
2892 0 : result = DEFAULT_EQ_SEL;
2893 : }
2894 2446 : result = 1.0 - result;
2895 : }
2896 :
2897 3700 : PG_RETURN_FLOAT8(result);
2898 : }
2899 :
2900 : /*
2901 : * scalarltjoinsel - Join selectivity of "<" for scalars
2902 : */
2903 : Datum
2904 324 : scalarltjoinsel(PG_FUNCTION_ARGS)
2905 : {
2906 324 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2907 : }
2908 :
2909 : /*
2910 : * scalarlejoinsel - Join selectivity of "<=" for scalars
2911 : */
2912 : Datum
2913 276 : scalarlejoinsel(PG_FUNCTION_ARGS)
2914 : {
2915 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2916 : }
2917 :
2918 : /*
2919 : * scalargtjoinsel - Join selectivity of ">" for scalars
2920 : */
2921 : Datum
2922 276 : scalargtjoinsel(PG_FUNCTION_ARGS)
2923 : {
2924 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2925 : }
2926 :
2927 : /*
2928 : * scalargejoinsel - Join selectivity of ">=" for scalars
2929 : */
2930 : Datum
2931 184 : scalargejoinsel(PG_FUNCTION_ARGS)
2932 : {
2933 184 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2934 : }
2935 :
2936 :
2937 : /*
2938 : * mergejoinscansel - Scan selectivity of merge join.
2939 : *
2940 : * A merge join will stop as soon as it exhausts either input stream.
2941 : * Therefore, if we can estimate the ranges of both input variables,
2942 : * we can estimate how much of the input will actually be read. This
2943 : * can have a considerable impact on the cost when using indexscans.
2944 : *
2945 : * Also, we can estimate how much of each input has to be read before the
2946 : * first join pair is found, which will affect the join's startup time.
2947 : *
2948 : * clause should be a clause already known to be mergejoinable. opfamily,
2949 : * cmptype, and nulls_first specify the sort ordering being used.
2950 : *
2951 : * The outputs are:
2952 : * *leftstart is set to the fraction of the left-hand variable expected
2953 : * to be scanned before the first join pair is found (0 to 1).
2954 : * *leftend is set to the fraction of the left-hand variable expected
2955 : * to be scanned before the join terminates (0 to 1).
2956 : * *rightstart, *rightend similarly for the right-hand variable.
2957 : */
2958 : void
2959 114118 : mergejoinscansel(PlannerInfo *root, Node *clause,
2960 : Oid opfamily, CompareType cmptype, bool nulls_first,
2961 : Selectivity *leftstart, Selectivity *leftend,
2962 : Selectivity *rightstart, Selectivity *rightend)
2963 : {
2964 : Node *left,
2965 : *right;
2966 : VariableStatData leftvar,
2967 : rightvar;
2968 : Oid opmethod;
2969 : int op_strategy;
2970 : Oid op_lefttype;
2971 : Oid op_righttype;
2972 : Oid opno,
2973 : collation,
2974 : lsortop,
2975 : rsortop,
2976 : lstatop,
2977 : rstatop,
2978 : ltop,
2979 : leop,
2980 : revltop,
2981 : revleop;
2982 : StrategyNumber ltstrat,
2983 : lestrat,
2984 : gtstrat,
2985 : gestrat;
2986 : bool isgt;
2987 : Datum leftmin,
2988 : leftmax,
2989 : rightmin,
2990 : rightmax;
2991 : double selec;
2992 :
2993 : /* Set default results if we can't figure anything out. */
2994 : /* XXX should default "start" fraction be a bit more than 0? */
2995 114118 : *leftstart = *rightstart = 0.0;
2996 114118 : *leftend = *rightend = 1.0;
2997 :
2998 : /* Deconstruct the merge clause */
2999 114118 : if (!is_opclause(clause))
3000 0 : return; /* shouldn't happen */
3001 114118 : opno = ((OpExpr *) clause)->opno;
3002 114118 : collation = ((OpExpr *) clause)->inputcollid;
3003 114118 : left = get_leftop((Expr *) clause);
3004 114118 : right = get_rightop((Expr *) clause);
3005 114118 : if (!right)
3006 0 : return; /* shouldn't happen */
3007 :
3008 : /* Look for stats for the inputs */
3009 114118 : examine_variable(root, left, 0, &leftvar);
3010 114118 : examine_variable(root, right, 0, &rightvar);
3011 :
3012 114118 : opmethod = get_opfamily_method(opfamily);
3013 :
3014 : /* Extract the operator's declared left/right datatypes */
3015 114118 : get_op_opfamily_properties(opno, opfamily, false,
3016 : &op_strategy,
3017 : &op_lefttype,
3018 : &op_righttype);
3019 : Assert(IndexAmTranslateStrategy(op_strategy, opmethod, opfamily, true) == COMPARE_EQ);
3020 :
3021 : /*
3022 : * Look up the various operators we need. If we don't find them all, it
3023 : * probably means the opfamily is broken, but we just fail silently.
3024 : *
3025 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3026 : * operator, regardless of which sort direction we are considering.
3027 : */
3028 114118 : switch (cmptype)
3029 : {
3030 114046 : case COMPARE_LT:
3031 114046 : isgt = false;
3032 114046 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3033 114046 : lestrat = IndexAmTranslateCompareType(COMPARE_LE, opmethod, opfamily, true);
3034 114046 : if (op_lefttype == op_righttype)
3035 : {
3036 : /* easy case */
3037 112664 : ltop = get_opfamily_member(opfamily,
3038 : op_lefttype, op_righttype,
3039 : ltstrat);
3040 112664 : leop = get_opfamily_member(opfamily,
3041 : op_lefttype, op_righttype,
3042 : lestrat);
3043 112664 : lsortop = ltop;
3044 112664 : rsortop = ltop;
3045 112664 : lstatop = lsortop;
3046 112664 : rstatop = rsortop;
3047 112664 : revltop = ltop;
3048 112664 : revleop = leop;
3049 : }
3050 : else
3051 : {
3052 1382 : ltop = get_opfamily_member(opfamily,
3053 : op_lefttype, op_righttype,
3054 : ltstrat);
3055 1382 : leop = get_opfamily_member(opfamily,
3056 : op_lefttype, op_righttype,
3057 : lestrat);
3058 1382 : lsortop = get_opfamily_member(opfamily,
3059 : op_lefttype, op_lefttype,
3060 : ltstrat);
3061 1382 : rsortop = get_opfamily_member(opfamily,
3062 : op_righttype, op_righttype,
3063 : ltstrat);
3064 1382 : lstatop = lsortop;
3065 1382 : rstatop = rsortop;
3066 1382 : revltop = get_opfamily_member(opfamily,
3067 : op_righttype, op_lefttype,
3068 : ltstrat);
3069 1382 : revleop = get_opfamily_member(opfamily,
3070 : op_righttype, op_lefttype,
3071 : lestrat);
3072 : }
3073 114046 : break;
3074 72 : case COMPARE_GT:
3075 : /* descending-order case */
3076 72 : isgt = true;
3077 72 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3078 72 : gtstrat = IndexAmTranslateCompareType(COMPARE_GT, opmethod, opfamily, true);
3079 72 : gestrat = IndexAmTranslateCompareType(COMPARE_GE, opmethod, opfamily, true);
3080 72 : if (op_lefttype == op_righttype)
3081 : {
3082 : /* easy case */
3083 72 : ltop = get_opfamily_member(opfamily,
3084 : op_lefttype, op_righttype,
3085 : gtstrat);
3086 72 : leop = get_opfamily_member(opfamily,
3087 : op_lefttype, op_righttype,
3088 : gestrat);
3089 72 : lsortop = ltop;
3090 72 : rsortop = ltop;
3091 72 : lstatop = get_opfamily_member(opfamily,
3092 : op_lefttype, op_lefttype,
3093 : ltstrat);
3094 72 : rstatop = lstatop;
3095 72 : revltop = ltop;
3096 72 : revleop = leop;
3097 : }
3098 : else
3099 : {
3100 0 : ltop = get_opfamily_member(opfamily,
3101 : op_lefttype, op_righttype,
3102 : gtstrat);
3103 0 : leop = get_opfamily_member(opfamily,
3104 : op_lefttype, op_righttype,
3105 : gestrat);
3106 0 : lsortop = get_opfamily_member(opfamily,
3107 : op_lefttype, op_lefttype,
3108 : gtstrat);
3109 0 : rsortop = get_opfamily_member(opfamily,
3110 : op_righttype, op_righttype,
3111 : gtstrat);
3112 0 : lstatop = get_opfamily_member(opfamily,
3113 : op_lefttype, op_lefttype,
3114 : ltstrat);
3115 0 : rstatop = get_opfamily_member(opfamily,
3116 : op_righttype, op_righttype,
3117 : ltstrat);
3118 0 : revltop = get_opfamily_member(opfamily,
3119 : op_righttype, op_lefttype,
3120 : gtstrat);
3121 0 : revleop = get_opfamily_member(opfamily,
3122 : op_righttype, op_lefttype,
3123 : gestrat);
3124 : }
3125 72 : break;
3126 0 : default:
3127 0 : goto fail; /* shouldn't get here */
3128 : }
3129 :
3130 114118 : if (!OidIsValid(lsortop) ||
3131 114118 : !OidIsValid(rsortop) ||
3132 114118 : !OidIsValid(lstatop) ||
3133 114118 : !OidIsValid(rstatop) ||
3134 114106 : !OidIsValid(ltop) ||
3135 114106 : !OidIsValid(leop) ||
3136 114106 : !OidIsValid(revltop) ||
3137 : !OidIsValid(revleop))
3138 12 : goto fail; /* insufficient info in catalogs */
3139 :
3140 : /* Try to get ranges of both inputs */
3141 114106 : if (!isgt)
3142 : {
3143 114034 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3144 : &leftmin, &leftmax))
3145 24074 : goto fail; /* no range available from stats */
3146 89960 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3147 : &rightmin, &rightmax))
3148 24254 : goto fail; /* no range available from stats */
3149 : }
3150 : else
3151 : {
3152 : /* need to swap the max and min */
3153 72 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3154 : &leftmax, &leftmin))
3155 30 : goto fail; /* no range available from stats */
3156 42 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3157 : &rightmax, &rightmin))
3158 0 : goto fail; /* no range available from stats */
3159 : }
3160 :
3161 : /*
3162 : * Now, the fraction of the left variable that will be scanned is the
3163 : * fraction that's <= the right-side maximum value. But only believe
3164 : * non-default estimates, else stick with our 1.0.
3165 : */
3166 65748 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3167 : rightmax, op_righttype);
3168 65748 : if (selec != DEFAULT_INEQ_SEL)
3169 65742 : *leftend = selec;
3170 :
3171 : /* And similarly for the right variable. */
3172 65748 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3173 : leftmax, op_lefttype);
3174 65748 : if (selec != DEFAULT_INEQ_SEL)
3175 65748 : *rightend = selec;
3176 :
3177 : /*
3178 : * Only one of the two "end" fractions can really be less than 1.0;
3179 : * believe the smaller estimate and reset the other one to exactly 1.0. If
3180 : * we get exactly equal estimates (as can easily happen with self-joins),
3181 : * believe neither.
3182 : */
3183 65748 : if (*leftend > *rightend)
3184 24784 : *leftend = 1.0;
3185 40964 : else if (*leftend < *rightend)
3186 32728 : *rightend = 1.0;
3187 : else
3188 8236 : *leftend = *rightend = 1.0;
3189 :
3190 : /*
3191 : * Also, the fraction of the left variable that will be scanned before the
3192 : * first join pair is found is the fraction that's < the right-side
3193 : * minimum value. But only believe non-default estimates, else stick with
3194 : * our own default.
3195 : */
3196 65748 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3197 : rightmin, op_righttype);
3198 65748 : if (selec != DEFAULT_INEQ_SEL)
3199 65748 : *leftstart = selec;
3200 :
3201 : /* And similarly for the right variable. */
3202 65748 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3203 : leftmin, op_lefttype);
3204 65748 : if (selec != DEFAULT_INEQ_SEL)
3205 65748 : *rightstart = selec;
3206 :
3207 : /*
3208 : * Only one of the two "start" fractions can really be more than zero;
3209 : * believe the larger estimate and reset the other one to exactly 0.0. If
3210 : * we get exactly equal estimates (as can easily happen with self-joins),
3211 : * believe neither.
3212 : */
3213 65748 : if (*leftstart < *rightstart)
3214 17388 : *leftstart = 0.0;
3215 48360 : else if (*leftstart > *rightstart)
3216 23378 : *rightstart = 0.0;
3217 : else
3218 24982 : *leftstart = *rightstart = 0.0;
3219 :
3220 : /*
3221 : * If the sort order is nulls-first, we're going to have to skip over any
3222 : * nulls too. These would not have been counted by scalarineqsel, and we
3223 : * can safely add in this fraction regardless of whether we believe
3224 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3225 : */
3226 65748 : if (nulls_first)
3227 : {
3228 : Form_pg_statistic stats;
3229 :
3230 42 : if (HeapTupleIsValid(leftvar.statsTuple))
3231 : {
3232 42 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3233 42 : *leftstart += stats->stanullfrac;
3234 42 : CLAMP_PROBABILITY(*leftstart);
3235 42 : *leftend += stats->stanullfrac;
3236 42 : CLAMP_PROBABILITY(*leftend);
3237 : }
3238 42 : if (HeapTupleIsValid(rightvar.statsTuple))
3239 : {
3240 42 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3241 42 : *rightstart += stats->stanullfrac;
3242 42 : CLAMP_PROBABILITY(*rightstart);
3243 42 : *rightend += stats->stanullfrac;
3244 42 : CLAMP_PROBABILITY(*rightend);
3245 : }
3246 : }
3247 :
3248 : /* Disbelieve start >= end, just in case that can happen */
3249 65748 : if (*leftstart >= *leftend)
3250 : {
3251 202 : *leftstart = 0.0;
3252 202 : *leftend = 1.0;
3253 : }
3254 65748 : if (*rightstart >= *rightend)
3255 : {
3256 1044 : *rightstart = 0.0;
3257 1044 : *rightend = 1.0;
3258 : }
3259 :
3260 64704 : fail:
3261 114118 : ReleaseVariableStats(leftvar);
3262 114118 : ReleaseVariableStats(rightvar);
3263 : }
3264 :
3265 :
3266 : /*
3267 : * matchingsel -- generic matching-operator selectivity support
3268 : *
3269 : * Use these for any operators that (a) are on data types for which we collect
3270 : * standard statistics, and (b) have behavior for which the default estimate
3271 : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3272 : * operators.
3273 : */
3274 :
3275 : Datum
3276 1106 : matchingsel(PG_FUNCTION_ARGS)
3277 : {
3278 1106 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3279 1106 : Oid operator = PG_GETARG_OID(1);
3280 1106 : List *args = (List *) PG_GETARG_POINTER(2);
3281 1106 : int varRelid = PG_GETARG_INT32(3);
3282 1106 : Oid collation = PG_GET_COLLATION();
3283 : double selec;
3284 :
3285 : /* Use generic restriction selectivity logic. */
3286 1106 : selec = generic_restriction_selectivity(root, operator, collation,
3287 : args, varRelid,
3288 : DEFAULT_MATCHING_SEL);
3289 :
3290 1106 : PG_RETURN_FLOAT8((float8) selec);
3291 : }
3292 :
3293 : Datum
3294 6 : matchingjoinsel(PG_FUNCTION_ARGS)
3295 : {
3296 : /* Just punt, for the moment. */
3297 6 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3298 : }
3299 :
3300 :
3301 : /*
3302 : * Helper routine for estimate_num_groups: add an item to a list of
3303 : * GroupVarInfos, but only if it's not known equal to any of the existing
3304 : * entries.
3305 : */
3306 : typedef struct
3307 : {
3308 : Node *var; /* might be an expression, not just a Var */
3309 : RelOptInfo *rel; /* relation it belongs to */
3310 : double ndistinct; /* # distinct values */
3311 : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3312 : } GroupVarInfo;
3313 :
3314 : static List *
3315 363656 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3316 : Node *var, VariableStatData *vardata)
3317 : {
3318 : GroupVarInfo *varinfo;
3319 : double ndistinct;
3320 : bool isdefault;
3321 : ListCell *lc;
3322 :
3323 363656 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3324 :
3325 : /*
3326 : * The nullingrels bits within the var could cause the same var to be
3327 : * counted multiple times if it's marked with different nullingrels. They
3328 : * could also prevent us from matching the var to the expressions in
3329 : * extended statistics (see estimate_multivariate_ndistinct). So strip
3330 : * them out first.
3331 : */
3332 363656 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3333 :
3334 442044 : foreach(lc, varinfos)
3335 : {
3336 79440 : varinfo = (GroupVarInfo *) lfirst(lc);
3337 :
3338 : /* Drop exact duplicates */
3339 79440 : if (equal(var, varinfo->var))
3340 1052 : return varinfos;
3341 :
3342 : /*
3343 : * Drop known-equal vars, but only if they belong to different
3344 : * relations (see comments for estimate_num_groups). We aren't too
3345 : * fussy about the semantics of "equal" here.
3346 : */
3347 83460 : if (vardata->rel != varinfo->rel &&
3348 4856 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3349 : {
3350 240 : if (varinfo->ndistinct <= ndistinct)
3351 : {
3352 : /* Keep older item, forget new one */
3353 216 : return varinfos;
3354 : }
3355 : else
3356 : {
3357 : /* Delete the older item */
3358 24 : varinfos = foreach_delete_current(varinfos, lc);
3359 : }
3360 : }
3361 : }
3362 :
3363 362604 : varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
3364 :
3365 362604 : varinfo->var = var;
3366 362604 : varinfo->rel = vardata->rel;
3367 362604 : varinfo->ndistinct = ndistinct;
3368 362604 : varinfo->isdefault = isdefault;
3369 362604 : varinfos = lappend(varinfos, varinfo);
3370 362604 : return varinfos;
3371 : }
3372 :
3373 : /*
3374 : * estimate_num_groups - Estimate number of groups in a grouped query
3375 : *
3376 : * Given a query having a GROUP BY clause, estimate how many groups there
3377 : * will be --- ie, the number of distinct combinations of the GROUP BY
3378 : * expressions.
3379 : *
3380 : * This routine is also used to estimate the number of rows emitted by
3381 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3382 : * actually, we only use it for DISTINCT when there's no grouping or
3383 : * aggregation ahead of the DISTINCT.)
3384 : *
3385 : * Inputs:
3386 : * root - the query
3387 : * groupExprs - list of expressions being grouped by
3388 : * input_rows - number of rows estimated to arrive at the group/unique
3389 : * filter step
3390 : * pgset - NULL, or a List** pointing to a grouping set to filter the
3391 : * groupExprs against
3392 : *
3393 : * Outputs:
3394 : * estinfo - When passed as non-NULL, the function will set bits in the
3395 : * "flags" field in order to provide callers with additional information
3396 : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3397 : * bit if we used any default values in the estimation.
3398 : *
3399 : * Given the lack of any cross-correlation statistics in the system, it's
3400 : * impossible to do anything really trustworthy with GROUP BY conditions
3401 : * involving multiple Vars. We should however avoid assuming the worst
3402 : * case (all possible cross-product terms actually appear as groups) since
3403 : * very often the grouped-by Vars are highly correlated. Our current approach
3404 : * is as follows:
3405 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3406 : * independently of their content, and are ignored in the subsequent
3407 : * steps. This is mainly because tests like "col IS NULL" break the
3408 : * heuristic used in step 2 especially badly.
3409 : * 2. Reduce the given expressions to a list of unique Vars used. For
3410 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3411 : * It is clearly correct not to count the same Var more than once.
3412 : * It is also reasonable to treat f(x) the same as x: f() cannot
3413 : * increase the number of distinct values (unless it is volatile,
3414 : * which we consider unlikely for grouping), but it probably won't
3415 : * reduce the number of distinct values much either.
3416 : * As a special case, if a GROUP BY expression can be matched to an
3417 : * expressional index for which we have statistics, then we treat the
3418 : * whole expression as though it were just a Var.
3419 : * 3. If the list contains Vars of different relations that are known equal
3420 : * due to equivalence classes, then drop all but one of the Vars from each
3421 : * known-equal set, keeping the one with smallest estimated # of values
3422 : * (since the extra values of the others can't appear in joined rows).
3423 : * Note the reason we only consider Vars of different relations is that
3424 : * if we considered ones of the same rel, we'd be double-counting the
3425 : * restriction selectivity of the equality in the next step.
3426 : * 4. For Vars within a single source rel, we multiply together the numbers
3427 : * of values, clamp to the number of rows in the rel (divided by 10 if
3428 : * more than one Var), and then multiply by a factor based on the
3429 : * selectivity of the restriction clauses for that rel. When there's
3430 : * more than one Var, the initial product is probably too high (it's the
3431 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3432 : * helpful heuristic for not letting the estimate get out of hand. (The
3433 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3434 : * we multiply by to adjust for the restriction selectivity assumes that
3435 : * the restriction clauses are independent of the grouping, which may not
3436 : * be a valid assumption, but it's hard to do better.
3437 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3438 : * rel, and multiply the results together.
3439 : * Note that rels not containing grouped Vars are ignored completely, as are
3440 : * join clauses. Such rels cannot increase the number of groups, and we
3441 : * assume such clauses do not reduce the number either (somewhat bogus,
3442 : * but we don't have the info to do better).
3443 : */
3444 : double
3445 315302 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3446 : List **pgset, EstimationInfo *estinfo)
3447 : {
3448 315302 : List *varinfos = NIL;
3449 315302 : double srf_multiplier = 1.0;
3450 : double numdistinct;
3451 : ListCell *l;
3452 : int i;
3453 :
3454 : /* Zero the estinfo output parameter, if non-NULL */
3455 315302 : if (estinfo != NULL)
3456 273314 : memset(estinfo, 0, sizeof(EstimationInfo));
3457 :
3458 : /*
3459 : * We don't ever want to return an estimate of zero groups, as that tends
3460 : * to lead to division-by-zero and other unpleasantness. The input_rows
3461 : * estimate is usually already at least 1, but clamp it just in case it
3462 : * isn't.
3463 : */
3464 315302 : input_rows = clamp_row_est(input_rows);
3465 :
3466 : /*
3467 : * If no grouping columns, there's exactly one group. (This can't happen
3468 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3469 : * corner cases with set operations.)
3470 : */
3471 315302 : if (groupExprs == NIL || (pgset && *pgset == NIL))
3472 1070 : return 1.0;
3473 :
3474 : /*
3475 : * Count groups derived from boolean grouping expressions. For other
3476 : * expressions, find the unique Vars used, treating an expression as a Var
3477 : * if we can find stats for it. For each one, record the statistical
3478 : * estimate of number of distinct values (total in its table, without
3479 : * regard for filtering).
3480 : */
3481 314232 : numdistinct = 1.0;
3482 :
3483 314232 : i = 0;
3484 676154 : foreach(l, groupExprs)
3485 : {
3486 361970 : Node *groupexpr = (Node *) lfirst(l);
3487 : double this_srf_multiplier;
3488 : VariableStatData vardata;
3489 : List *varshere;
3490 : ListCell *l2;
3491 :
3492 : /* is expression in this grouping set? */
3493 361970 : if (pgset && !list_member_int(*pgset, i++))
3494 295006 : continue;
3495 :
3496 : /*
3497 : * Set-returning functions in grouping columns are a bit problematic.
3498 : * The code below will effectively ignore their SRF nature and come up
3499 : * with a numdistinct estimate as though they were scalar functions.
3500 : * We compensate by scaling up the end result by the largest SRF
3501 : * rowcount estimate. (This will be an overestimate if the SRF
3502 : * produces multiple copies of any output value, but it seems best to
3503 : * assume the SRF's outputs are distinct. In any case, it's probably
3504 : * pointless to worry too much about this without much better
3505 : * estimates for SRF output rowcounts than we have today.)
3506 : */
3507 361182 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3508 361182 : if (srf_multiplier < this_srf_multiplier)
3509 132 : srf_multiplier = this_srf_multiplier;
3510 :
3511 : /* Short-circuit for expressions returning boolean */
3512 361182 : if (exprType(groupexpr) == BOOLOID)
3513 : {
3514 198 : numdistinct *= 2.0;
3515 198 : continue;
3516 : }
3517 :
3518 : /*
3519 : * If examine_variable is able to deduce anything about the GROUP BY
3520 : * expression, treat it as a single variable even if it's really more
3521 : * complicated.
3522 : *
3523 : * XXX This has the consequence that if there's a statistics object on
3524 : * the expression, we don't split it into individual Vars. This
3525 : * affects our selection of statistics in
3526 : * estimate_multivariate_ndistinct, because it's probably better to
3527 : * use more accurate estimate for each expression and treat them as
3528 : * independent, than to combine estimates for the extracted variables
3529 : * when we don't know how that relates to the expressions.
3530 : */
3531 360984 : examine_variable(root, groupexpr, 0, &vardata);
3532 360984 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3533 : {
3534 293342 : varinfos = add_unique_group_var(root, varinfos,
3535 : groupexpr, &vardata);
3536 293342 : ReleaseVariableStats(vardata);
3537 293342 : continue;
3538 : }
3539 67642 : ReleaseVariableStats(vardata);
3540 :
3541 : /*
3542 : * Else pull out the component Vars. Handle PlaceHolderVars by
3543 : * recursing into their arguments (effectively assuming that the
3544 : * PlaceHolderVar doesn't change the number of groups, which boils
3545 : * down to ignoring the possible addition of nulls to the result set).
3546 : */
3547 67642 : varshere = pull_var_clause(groupexpr,
3548 : PVC_RECURSE_AGGREGATES |
3549 : PVC_RECURSE_WINDOWFUNCS |
3550 : PVC_RECURSE_PLACEHOLDERS);
3551 :
3552 : /*
3553 : * If we find any variable-free GROUP BY item, then either it is a
3554 : * constant (and we can ignore it) or it contains a volatile function;
3555 : * in the latter case we punt and assume that each input row will
3556 : * yield a distinct group.
3557 : */
3558 67642 : if (varshere == NIL)
3559 : {
3560 726 : if (contain_volatile_functions(groupexpr))
3561 48 : return input_rows;
3562 678 : continue;
3563 : }
3564 :
3565 : /*
3566 : * Else add variables to varinfos list
3567 : */
3568 137230 : foreach(l2, varshere)
3569 : {
3570 70314 : Node *var = (Node *) lfirst(l2);
3571 :
3572 70314 : examine_variable(root, var, 0, &vardata);
3573 70314 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3574 70314 : ReleaseVariableStats(vardata);
3575 : }
3576 : }
3577 :
3578 : /*
3579 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3580 : * list.
3581 : */
3582 314184 : if (varinfos == NIL)
3583 : {
3584 : /* Apply SRF multiplier as we would do in the long path */
3585 394 : numdistinct *= srf_multiplier;
3586 : /* Round off */
3587 394 : numdistinct = ceil(numdistinct);
3588 : /* Guard against out-of-range answers */
3589 394 : if (numdistinct > input_rows)
3590 44 : numdistinct = input_rows;
3591 394 : if (numdistinct < 1.0)
3592 0 : numdistinct = 1.0;
3593 394 : return numdistinct;
3594 : }
3595 :
3596 : /*
3597 : * Group Vars by relation and estimate total numdistinct.
3598 : *
3599 : * For each iteration of the outer loop, we process the frontmost Var in
3600 : * varinfos, plus all other Vars in the same relation. We remove these
3601 : * Vars from the newvarinfos list for the next iteration. This is the
3602 : * easiest way to group Vars of same rel together.
3603 : */
3604 : do
3605 : {
3606 315772 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3607 315772 : RelOptInfo *rel = varinfo1->rel;
3608 315772 : double reldistinct = 1;
3609 315772 : double relmaxndistinct = reldistinct;
3610 315772 : int relvarcount = 0;
3611 315772 : List *newvarinfos = NIL;
3612 315772 : List *relvarinfos = NIL;
3613 :
3614 : /*
3615 : * Split the list of varinfos in two - one for the current rel, one
3616 : * for remaining Vars on other rels.
3617 : */
3618 315772 : relvarinfos = lappend(relvarinfos, varinfo1);
3619 366530 : for_each_from(l, varinfos, 1)
3620 : {
3621 50758 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3622 :
3623 50758 : if (varinfo2->rel == varinfo1->rel)
3624 : {
3625 : /* varinfos on current rel */
3626 46808 : relvarinfos = lappend(relvarinfos, varinfo2);
3627 : }
3628 : else
3629 : {
3630 : /* not time to process varinfo2 yet */
3631 3950 : newvarinfos = lappend(newvarinfos, varinfo2);
3632 : }
3633 : }
3634 :
3635 : /*
3636 : * Get the numdistinct estimate for the Vars of this rel. We
3637 : * iteratively search for multivariate n-distinct with maximum number
3638 : * of vars; assuming that each var group is independent of the others,
3639 : * we multiply them together. Any remaining relvarinfos after no more
3640 : * multivariate matches are found are assumed independent too, so
3641 : * their individual ndistinct estimates are multiplied also.
3642 : *
3643 : * While iterating, count how many separate numdistinct values we
3644 : * apply. We apply a fudge factor below, but only if we multiplied
3645 : * more than one such values.
3646 : */
3647 631670 : while (relvarinfos)
3648 : {
3649 : double mvndistinct;
3650 :
3651 315898 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3652 : &mvndistinct))
3653 : {
3654 402 : reldistinct *= mvndistinct;
3655 402 : if (relmaxndistinct < mvndistinct)
3656 390 : relmaxndistinct = mvndistinct;
3657 402 : relvarcount++;
3658 : }
3659 : else
3660 : {
3661 677224 : foreach(l, relvarinfos)
3662 : {
3663 361728 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3664 :
3665 361728 : reldistinct *= varinfo2->ndistinct;
3666 361728 : if (relmaxndistinct < varinfo2->ndistinct)
3667 317208 : relmaxndistinct = varinfo2->ndistinct;
3668 361728 : relvarcount++;
3669 :
3670 : /*
3671 : * When varinfo2's isdefault is set then we'd better set
3672 : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3673 : */
3674 361728 : if (estinfo != NULL && varinfo2->isdefault)
3675 17706 : estinfo->flags |= SELFLAG_USED_DEFAULT;
3676 : }
3677 :
3678 : /* we're done with this relation */
3679 315496 : relvarinfos = NIL;
3680 : }
3681 : }
3682 :
3683 : /*
3684 : * Sanity check --- don't divide by zero if empty relation.
3685 : */
3686 : Assert(IS_SIMPLE_REL(rel));
3687 315772 : if (rel->tuples > 0)
3688 : {
3689 : /*
3690 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
3691 : * fudge factor is because the Vars are probably correlated but we
3692 : * don't know by how much. We should never clamp to less than the
3693 : * largest ndistinct value for any of the Vars, though, since
3694 : * there will surely be at least that many groups.
3695 : */
3696 314744 : double clamp = rel->tuples;
3697 :
3698 314744 : if (relvarcount > 1)
3699 : {
3700 42264 : clamp *= 0.1;
3701 42264 : if (clamp < relmaxndistinct)
3702 : {
3703 40168 : clamp = relmaxndistinct;
3704 : /* for sanity in case some ndistinct is too large: */
3705 40168 : if (clamp > rel->tuples)
3706 78 : clamp = rel->tuples;
3707 : }
3708 : }
3709 314744 : if (reldistinct > clamp)
3710 34754 : reldistinct = clamp;
3711 :
3712 : /*
3713 : * Update the estimate based on the restriction selectivity,
3714 : * guarding against division by zero when reldistinct is zero.
3715 : * Also skip this if we know that we are returning all rows.
3716 : */
3717 314744 : if (reldistinct > 0 && rel->rows < rel->tuples)
3718 : {
3719 : /*
3720 : * Given a table containing N rows with n distinct values in a
3721 : * uniform distribution, if we select p rows at random then
3722 : * the expected number of distinct values selected is
3723 : *
3724 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
3725 : *
3726 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
3727 : *
3728 : * See "Approximating block accesses in database
3729 : * organizations", S. B. Yao, Communications of the ACM,
3730 : * Volume 20 Issue 4, April 1977 Pages 260-261.
3731 : *
3732 : * Alternatively, re-arranging the terms from the factorials,
3733 : * this may be written as
3734 : *
3735 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
3736 : *
3737 : * This form of the formula is more efficient to compute in
3738 : * the common case where p is larger than N/n. Additionally,
3739 : * as pointed out by Dell'Era, if i << N for all terms in the
3740 : * product, it can be approximated by
3741 : *
3742 : * n * (1 - ((N-p)/N)^(N/n))
3743 : *
3744 : * See "Expected distinct values when selecting from a bag
3745 : * without replacement", Alberto Dell'Era,
3746 : * http://www.adellera.it/investigations/distinct_balls/.
3747 : *
3748 : * The condition i << N is equivalent to n >> 1, so this is a
3749 : * good approximation when the number of distinct values in
3750 : * the table is large. It turns out that this formula also
3751 : * works well even when n is small.
3752 : */
3753 103430 : reldistinct *=
3754 103430 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
3755 103430 : rel->tuples / reldistinct));
3756 : }
3757 314744 : reldistinct = clamp_row_est(reldistinct);
3758 :
3759 : /*
3760 : * Update estimate of total distinct groups.
3761 : */
3762 314744 : numdistinct *= reldistinct;
3763 : }
3764 :
3765 315772 : varinfos = newvarinfos;
3766 315772 : } while (varinfos != NIL);
3767 :
3768 : /* Now we can account for the effects of any SRFs */
3769 313790 : numdistinct *= srf_multiplier;
3770 :
3771 : /* Round off */
3772 313790 : numdistinct = ceil(numdistinct);
3773 :
3774 : /* Guard against out-of-range answers */
3775 313790 : if (numdistinct > input_rows)
3776 68730 : numdistinct = input_rows;
3777 313790 : if (numdistinct < 1.0)
3778 0 : numdistinct = 1.0;
3779 :
3780 313790 : return numdistinct;
3781 : }
3782 :
3783 : /*
3784 : * Try to estimate the bucket size of the hash join inner side when the join
3785 : * condition contains two or more clauses by employing extended statistics.
3786 : *
3787 : * The main idea of this approach is that the distinct value generated by
3788 : * multivariate estimation on two or more columns would provide less bucket size
3789 : * than estimation on one separate column.
3790 : *
3791 : * IMPORTANT: It is crucial to synchronize the approach of combining different
3792 : * estimations with the caller's method.
3793 : *
3794 : * Return a list of clauses that didn't fetch any extended statistics.
3795 : */
3796 : List *
3797 274872 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
3798 : List *hashclauses,
3799 : Selectivity *innerbucketsize)
3800 : {
3801 : List *clauses;
3802 : List *otherclauses;
3803 : double ndistinct;
3804 :
3805 274872 : if (list_length(hashclauses) <= 1)
3806 : {
3807 : /*
3808 : * Nothing to do for a single clause. Could we employ univariate
3809 : * extended stat here?
3810 : */
3811 242988 : return hashclauses;
3812 : }
3813 :
3814 : /* "clauses" is the list of hashclauses we've not dealt with yet */
3815 31884 : clauses = list_copy(hashclauses);
3816 : /* "otherclauses" holds clauses we are going to return to caller */
3817 31884 : otherclauses = NIL;
3818 : /* current estimate of ndistinct */
3819 31884 : ndistinct = 1.0;
3820 63780 : while (clauses != NIL)
3821 : {
3822 : ListCell *lc;
3823 31896 : int relid = -1;
3824 31896 : List *varinfos = NIL;
3825 31896 : List *origin_rinfos = NIL;
3826 : double mvndistinct;
3827 : List *origin_varinfos;
3828 31896 : int group_relid = -1;
3829 31896 : RelOptInfo *group_rel = NULL;
3830 : ListCell *lc1,
3831 : *lc2;
3832 :
3833 : /*
3834 : * Find clauses, referencing the same single base relation and try to
3835 : * estimate such a group with extended statistics. Create varinfo for
3836 : * an approved clause, push it to otherclauses, if it can't be
3837 : * estimated here or ignore to process at the next iteration.
3838 : */
3839 96300 : foreach(lc, clauses)
3840 : {
3841 64404 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
3842 : Node *expr;
3843 : Relids relids;
3844 : GroupVarInfo *varinfo;
3845 :
3846 : /*
3847 : * Find the inner side of the join, which we need to estimate the
3848 : * number of buckets. Use outer_is_left because the
3849 : * clause_sides_match_join routine has called on hash clauses.
3850 : */
3851 128808 : relids = rinfo->outer_is_left ?
3852 64404 : rinfo->right_relids : rinfo->left_relids;
3853 128808 : expr = rinfo->outer_is_left ?
3854 64404 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
3855 :
3856 64404 : if (bms_get_singleton_member(relids, &relid) &&
3857 61838 : root->simple_rel_array[relid]->statlist != NIL)
3858 48 : {
3859 60 : bool is_duplicate = false;
3860 :
3861 : /*
3862 : * This inner-side expression references only one relation.
3863 : * Extended statistics on this clause can exist.
3864 : */
3865 60 : if (group_relid < 0)
3866 : {
3867 30 : RangeTblEntry *rte = root->simple_rte_array[relid];
3868 :
3869 30 : if (!rte || (rte->relkind != RELKIND_RELATION &&
3870 0 : rte->relkind != RELKIND_MATVIEW &&
3871 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
3872 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
3873 : {
3874 : /* Extended statistics can't exist in principle */
3875 0 : otherclauses = lappend(otherclauses, rinfo);
3876 0 : clauses = foreach_delete_current(clauses, lc);
3877 0 : continue;
3878 : }
3879 :
3880 30 : group_relid = relid;
3881 30 : group_rel = root->simple_rel_array[relid];
3882 : }
3883 30 : else if (group_relid != relid)
3884 : {
3885 : /*
3886 : * Being in the group forming state we don't need other
3887 : * clauses.
3888 : */
3889 0 : continue;
3890 : }
3891 :
3892 : /*
3893 : * We're going to add the new clause to the varinfos list. We
3894 : * might re-use add_unique_group_var(), but we don't do so for
3895 : * two reasons.
3896 : *
3897 : * 1) We must keep the origin_rinfos list ordered exactly the
3898 : * same way as varinfos.
3899 : *
3900 : * 2) add_unique_group_var() is designed for
3901 : * estimate_num_groups(), where a larger number of groups is
3902 : * worse. While estimating the number of hash buckets, we
3903 : * have the opposite: a lesser number of groups is worse.
3904 : * Therefore, we don't have to remove "known equal" vars: the
3905 : * removed var may valuably contribute to the multivariate
3906 : * statistics to grow the number of groups.
3907 : */
3908 :
3909 : /*
3910 : * Clear nullingrels to correctly match hash keys. See
3911 : * add_unique_group_var()'s comment for details.
3912 : */
3913 60 : expr = remove_nulling_relids(expr, root->outer_join_rels, NULL);
3914 :
3915 : /*
3916 : * Detect and exclude exact duplicates from the list of hash
3917 : * keys (like add_unique_group_var does).
3918 : */
3919 84 : foreach(lc1, varinfos)
3920 : {
3921 36 : varinfo = (GroupVarInfo *) lfirst(lc1);
3922 :
3923 36 : if (!equal(expr, varinfo->var))
3924 24 : continue;
3925 :
3926 12 : is_duplicate = true;
3927 12 : break;
3928 : }
3929 :
3930 60 : if (is_duplicate)
3931 : {
3932 : /*
3933 : * Skip exact duplicates. Adding them to the otherclauses
3934 : * list also doesn't make sense.
3935 : */
3936 12 : continue;
3937 : }
3938 :
3939 : /*
3940 : * Initialize GroupVarInfo. We only use it to call
3941 : * estimate_multivariate_ndistinct(), which doesn't care about
3942 : * ndistinct and isdefault fields. Thus, skip these fields.
3943 : */
3944 48 : varinfo = (GroupVarInfo *) palloc0(sizeof(GroupVarInfo));
3945 48 : varinfo->var = expr;
3946 48 : varinfo->rel = root->simple_rel_array[relid];
3947 48 : varinfos = lappend(varinfos, varinfo);
3948 :
3949 : /*
3950 : * Remember the link to RestrictInfo for the case the clause
3951 : * is failed to be estimated.
3952 : */
3953 48 : origin_rinfos = lappend(origin_rinfos, rinfo);
3954 : }
3955 : else
3956 : {
3957 : /* This clause can't be estimated with extended statistics */
3958 64344 : otherclauses = lappend(otherclauses, rinfo);
3959 : }
3960 :
3961 64392 : clauses = foreach_delete_current(clauses, lc);
3962 : }
3963 :
3964 31896 : if (list_length(varinfos) < 2)
3965 : {
3966 : /*
3967 : * Multivariate statistics doesn't apply to single columns except
3968 : * for expressions, but it has not been implemented yet.
3969 : */
3970 31884 : otherclauses = list_concat(otherclauses, origin_rinfos);
3971 31884 : list_free_deep(varinfos);
3972 31884 : list_free(origin_rinfos);
3973 31884 : continue;
3974 : }
3975 :
3976 : Assert(group_rel != NULL);
3977 :
3978 : /* Employ the extended statistics. */
3979 12 : origin_varinfos = varinfos;
3980 : for (;;)
3981 12 : {
3982 24 : bool estimated = estimate_multivariate_ndistinct(root,
3983 : group_rel,
3984 : &varinfos,
3985 : &mvndistinct);
3986 :
3987 24 : if (!estimated)
3988 12 : break;
3989 :
3990 : /*
3991 : * We've got an estimation. Use ndistinct value in a consistent
3992 : * way - according to the caller's logic (see
3993 : * final_cost_hashjoin).
3994 : */
3995 12 : if (ndistinct < mvndistinct)
3996 12 : ndistinct = mvndistinct;
3997 : Assert(ndistinct >= 1.0);
3998 : }
3999 :
4000 : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
4001 :
4002 : /* Collect unmatched clauses as otherclauses. */
4003 42 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
4004 : {
4005 30 : GroupVarInfo *vinfo = lfirst(lc1);
4006 :
4007 30 : if (!list_member_ptr(varinfos, vinfo))
4008 : /* Already estimated */
4009 30 : continue;
4010 :
4011 : /* Can't be estimated here - push to the returning list */
4012 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
4013 : }
4014 : }
4015 :
4016 31884 : *innerbucketsize = 1.0 / ndistinct;
4017 31884 : return otherclauses;
4018 : }
4019 :
4020 : /*
4021 : * Estimate hash bucket statistics when the specified expression is used
4022 : * as a hash key for the given number of buckets.
4023 : *
4024 : * This attempts to determine two values:
4025 : *
4026 : * 1. The frequency of the most common value of the expression (returns
4027 : * zero into *mcv_freq if we can't get that).
4028 : *
4029 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
4030 : * divided by total tuples in relation.
4031 : *
4032 : * XXX This is really pretty bogus since we're effectively assuming that the
4033 : * distribution of hash keys will be the same after applying restriction
4034 : * clauses as it was in the underlying relation. However, we are not nearly
4035 : * smart enough to figure out how the restrict clauses might change the
4036 : * distribution, so this will have to do for now.
4037 : *
4038 : * We are passed the number of buckets the executor will use for the given
4039 : * input relation. If the data were perfectly distributed, with the same
4040 : * number of tuples going into each available bucket, then the bucketsize
4041 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
4042 : * only if (a) there are at least nbuckets distinct data values, and (b)
4043 : * we have a not-too-skewed data distribution. Otherwise the buckets will
4044 : * be nonuniformly occupied. If the other relation in the join has a key
4045 : * distribution similar to this one's, then the most-loaded buckets are
4046 : * exactly those that will be probed most often. Therefore, the "average"
4047 : * bucket size for costing purposes should really be taken as something close
4048 : * to the "worst case" bucket size. We try to estimate this by adjusting the
4049 : * fraction if there are too few distinct data values, and then scaling up
4050 : * by the ratio of the most common value's frequency to the average frequency.
4051 : *
4052 : * If no statistics are available, use a default estimate of 0.1. This will
4053 : * discourage use of a hash rather strongly if the inner relation is large,
4054 : * which is what we want. We do not want to hash unless we know that the
4055 : * inner rel is well-dispersed (or the alternatives seem much worse).
4056 : *
4057 : * The caller should also check that the mcv_freq is not so large that the
4058 : * most common value would by itself require an impractically large bucket.
4059 : * In a hash join, the executor can split buckets if they get too big, but
4060 : * obviously that doesn't help for a bucket that contains many duplicates of
4061 : * the same value.
4062 : */
4063 : void
4064 163000 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
4065 : Selectivity *mcv_freq,
4066 : Selectivity *bucketsize_frac)
4067 : {
4068 : VariableStatData vardata;
4069 : double estfract,
4070 : ndistinct,
4071 : stanullfrac,
4072 : avgfreq;
4073 : bool isdefault;
4074 : AttStatsSlot sslot;
4075 :
4076 163000 : examine_variable(root, hashkey, 0, &vardata);
4077 :
4078 : /* Look up the frequency of the most common value, if available */
4079 163000 : *mcv_freq = 0.0;
4080 :
4081 163000 : if (HeapTupleIsValid(vardata.statsTuple))
4082 : {
4083 113542 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4084 : STATISTIC_KIND_MCV, InvalidOid,
4085 : ATTSTATSSLOT_NUMBERS))
4086 : {
4087 : /*
4088 : * The first MCV stat is for the most common value.
4089 : */
4090 56436 : if (sslot.nnumbers > 0)
4091 56436 : *mcv_freq = sslot.numbers[0];
4092 56436 : free_attstatsslot(&sslot);
4093 : }
4094 : }
4095 :
4096 : /* Get number of distinct values */
4097 163000 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4098 :
4099 : /*
4100 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4101 : * mcv_freq is known to be even higher than that, use it instead.
4102 : */
4103 163000 : if (isdefault)
4104 : {
4105 20824 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4106 20824 : ReleaseVariableStats(vardata);
4107 20824 : return;
4108 : }
4109 :
4110 : /* Get fraction that are null */
4111 142176 : if (HeapTupleIsValid(vardata.statsTuple))
4112 : {
4113 : Form_pg_statistic stats;
4114 :
4115 113524 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
4116 113524 : stanullfrac = stats->stanullfrac;
4117 : }
4118 : else
4119 28652 : stanullfrac = 0.0;
4120 :
4121 : /* Compute avg freq of all distinct data values in raw relation */
4122 142176 : avgfreq = (1.0 - stanullfrac) / ndistinct;
4123 :
4124 : /*
4125 : * Adjust ndistinct to account for restriction clauses. Observe we are
4126 : * assuming that the data distribution is affected uniformly by the
4127 : * restriction clauses!
4128 : *
4129 : * XXX Possibly better way, but much more expensive: multiply by
4130 : * selectivity of rel's restriction clauses that mention the target Var.
4131 : */
4132 142176 : if (vardata.rel && vardata.rel->tuples > 0)
4133 : {
4134 142160 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
4135 142160 : ndistinct = clamp_row_est(ndistinct);
4136 : }
4137 :
4138 : /*
4139 : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
4140 : * number of buckets is less than the expected number of distinct values;
4141 : * otherwise it is 1/ndistinct.
4142 : */
4143 142176 : if (ndistinct > nbuckets)
4144 88 : estfract = 1.0 / nbuckets;
4145 : else
4146 142088 : estfract = 1.0 / ndistinct;
4147 :
4148 : /*
4149 : * Adjust estimated bucketsize upward to account for skewed distribution.
4150 : */
4151 142176 : if (avgfreq > 0.0 && *mcv_freq > avgfreq)
4152 49296 : estfract *= *mcv_freq / avgfreq;
4153 :
4154 : /*
4155 : * Clamp bucketsize to sane range (the above adjustment could easily
4156 : * produce an out-of-range result). We set the lower bound a little above
4157 : * zero, since zero isn't a very sane result.
4158 : */
4159 142176 : if (estfract < 1.0e-6)
4160 0 : estfract = 1.0e-6;
4161 142176 : else if (estfract > 1.0)
4162 34672 : estfract = 1.0;
4163 :
4164 142176 : *bucketsize_frac = (Selectivity) estfract;
4165 :
4166 142176 : ReleaseVariableStats(vardata);
4167 : }
4168 :
4169 : /*
4170 : * estimate_hashagg_tablesize
4171 : * estimate the number of bytes that a hash aggregate hashtable will
4172 : * require based on the agg_costs, path width and number of groups.
4173 : *
4174 : * We return the result as "double" to forestall any possible overflow
4175 : * problem in the multiplication by dNumGroups.
4176 : *
4177 : * XXX this may be over-estimating the size now that hashagg knows to omit
4178 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4179 : * grouping columns not in the hashed set are counted here even though hashagg
4180 : * won't store them. Is this a problem?
4181 : */
4182 : double
4183 2350 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4184 : const AggClauseCosts *agg_costs, double dNumGroups)
4185 : {
4186 : Size hashentrysize;
4187 :
4188 2350 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4189 2350 : path->pathtarget->width,
4190 2350 : agg_costs->transitionSpace);
4191 :
4192 : /*
4193 : * Note that this disregards the effect of fill-factor and growth policy
4194 : * of the hash table. That's probably ok, given that the default
4195 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4196 : * "double-in-size" growth policies here.
4197 : */
4198 2350 : return hashentrysize * dNumGroups;
4199 : }
4200 :
4201 :
4202 : /*-------------------------------------------------------------------------
4203 : *
4204 : * Support routines
4205 : *
4206 : *-------------------------------------------------------------------------
4207 : */
4208 :
4209 : /*
4210 : * Find the best matching ndistinct extended statistics for the given list of
4211 : * GroupVarInfos.
4212 : *
4213 : * Callers must ensure that the given GroupVarInfos all belong to 'rel' and
4214 : * the GroupVarInfos list does not contain any duplicate Vars or expressions.
4215 : *
4216 : * When statistics are found that match > 1 of the given GroupVarInfo, the
4217 : * *ndistinct parameter is set according to the ndistinct estimate and a new
4218 : * list is built with the matching GroupVarInfos removed, which is output via
4219 : * the *varinfos parameter before returning true. When no matching stats are
4220 : * found, false is returned and the *varinfos and *ndistinct parameters are
4221 : * left untouched.
4222 : */
4223 : static bool
4224 315922 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4225 : List **varinfos, double *ndistinct)
4226 : {
4227 : ListCell *lc;
4228 : int nmatches_vars;
4229 : int nmatches_exprs;
4230 315922 : Oid statOid = InvalidOid;
4231 : MVNDistinct *stats;
4232 315922 : StatisticExtInfo *matched_info = NULL;
4233 315922 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4234 :
4235 : /* bail out immediately if the table has no extended statistics */
4236 315922 : if (!rel->statlist)
4237 315370 : return false;
4238 :
4239 : /* look for the ndistinct statistics object matching the most vars */
4240 552 : nmatches_vars = 0; /* we require at least two matches */
4241 552 : nmatches_exprs = 0;
4242 2172 : foreach(lc, rel->statlist)
4243 : {
4244 : ListCell *lc2;
4245 1620 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
4246 1620 : int nshared_vars = 0;
4247 1620 : int nshared_exprs = 0;
4248 :
4249 : /* skip statistics of other kinds */
4250 1620 : if (info->kind != STATS_EXT_NDISTINCT)
4251 750 : continue;
4252 :
4253 : /* skip statistics with mismatching stxdinherit value */
4254 870 : if (info->inherit != rte->inh)
4255 24 : continue;
4256 :
4257 : /*
4258 : * Determine how many expressions (and variables in non-matched
4259 : * expressions) match. We'll then use these numbers to pick the
4260 : * statistics object that best matches the clauses.
4261 : */
4262 2682 : foreach(lc2, *varinfos)
4263 : {
4264 : ListCell *lc3;
4265 1836 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4266 : AttrNumber attnum;
4267 :
4268 : Assert(varinfo->rel == rel);
4269 :
4270 : /* simple Var, search in statistics keys directly */
4271 1836 : if (IsA(varinfo->var, Var))
4272 : {
4273 1470 : attnum = ((Var *) varinfo->var)->varattno;
4274 :
4275 : /*
4276 : * Ignore system attributes - we don't support statistics on
4277 : * them, so can't match them (and it'd fail as the values are
4278 : * negative).
4279 : */
4280 1470 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4281 12 : continue;
4282 :
4283 1458 : if (bms_is_member(attnum, info->keys))
4284 852 : nshared_vars++;
4285 :
4286 1458 : continue;
4287 : }
4288 :
4289 : /* expression - see if it's in the statistics object */
4290 660 : foreach(lc3, info->exprs)
4291 : {
4292 528 : Node *expr = (Node *) lfirst(lc3);
4293 :
4294 528 : if (equal(varinfo->var, expr))
4295 : {
4296 234 : nshared_exprs++;
4297 234 : break;
4298 : }
4299 : }
4300 : }
4301 :
4302 : /*
4303 : * The ndistinct extended statistics contain estimates for a minimum
4304 : * of pairs of columns which the statistics are defined on and
4305 : * certainly not single columns. Here we skip unless we managed to
4306 : * match to at least two columns.
4307 : */
4308 846 : if (nshared_vars + nshared_exprs < 2)
4309 396 : continue;
4310 :
4311 : /*
4312 : * Check if these statistics are a better match than the previous best
4313 : * match and if so, take note of the StatisticExtInfo.
4314 : *
4315 : * The statslist is sorted by statOid, so the StatisticExtInfo we
4316 : * select as the best match is deterministic even when multiple sets
4317 : * of statistics match equally as well.
4318 : */
4319 450 : if ((nshared_exprs > nmatches_exprs) ||
4320 342 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4321 : {
4322 426 : statOid = info->statOid;
4323 426 : nmatches_vars = nshared_vars;
4324 426 : nmatches_exprs = nshared_exprs;
4325 426 : matched_info = info;
4326 : }
4327 : }
4328 :
4329 : /* No match? */
4330 552 : if (statOid == InvalidOid)
4331 138 : return false;
4332 :
4333 : Assert(nmatches_vars + nmatches_exprs > 1);
4334 :
4335 414 : stats = statext_ndistinct_load(statOid, rte->inh);
4336 :
4337 : /*
4338 : * If we have a match, search it for the specific item that matches (there
4339 : * must be one), and construct the output values.
4340 : */
4341 414 : if (stats)
4342 : {
4343 : int i;
4344 414 : List *newlist = NIL;
4345 414 : MVNDistinctItem *item = NULL;
4346 : ListCell *lc2;
4347 414 : Bitmapset *matched = NULL;
4348 : AttrNumber attnum_offset;
4349 :
4350 : /*
4351 : * How much we need to offset the attnums? If there are no
4352 : * expressions, no offset is needed. Otherwise offset enough to move
4353 : * the lowest one (which is equal to number of expressions) to 1.
4354 : */
4355 414 : if (matched_info->exprs)
4356 144 : attnum_offset = (list_length(matched_info->exprs) + 1);
4357 : else
4358 270 : attnum_offset = 0;
4359 :
4360 : /* see what actually matched */
4361 1452 : foreach(lc2, *varinfos)
4362 : {
4363 : ListCell *lc3;
4364 : int idx;
4365 1038 : bool found = false;
4366 :
4367 1038 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4368 :
4369 : /*
4370 : * Process a simple Var expression, by matching it to keys
4371 : * directly. If there's a matching expression, we'll try matching
4372 : * it later.
4373 : */
4374 1038 : if (IsA(varinfo->var, Var))
4375 : {
4376 852 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4377 :
4378 : /*
4379 : * Ignore expressions on system attributes. Can't rely on the
4380 : * bms check for negative values.
4381 : */
4382 852 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4383 6 : continue;
4384 :
4385 : /* Is the variable covered by the statistics object? */
4386 846 : if (!bms_is_member(attnum, matched_info->keys))
4387 120 : continue;
4388 :
4389 726 : attnum = attnum + attnum_offset;
4390 :
4391 : /* ensure sufficient offset */
4392 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4393 :
4394 726 : matched = bms_add_member(matched, attnum);
4395 :
4396 726 : found = true;
4397 : }
4398 :
4399 : /*
4400 : * XXX Maybe we should allow searching the expressions even if we
4401 : * found an attribute matching the expression? That would handle
4402 : * trivial expressions like "(a)" but it seems fairly useless.
4403 : */
4404 912 : if (found)
4405 726 : continue;
4406 :
4407 : /* expression - see if it's in the statistics object */
4408 186 : idx = 0;
4409 306 : foreach(lc3, matched_info->exprs)
4410 : {
4411 276 : Node *expr = (Node *) lfirst(lc3);
4412 :
4413 276 : if (equal(varinfo->var, expr))
4414 : {
4415 156 : AttrNumber attnum = -(idx + 1);
4416 :
4417 156 : attnum = attnum + attnum_offset;
4418 :
4419 : /* ensure sufficient offset */
4420 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4421 :
4422 156 : matched = bms_add_member(matched, attnum);
4423 :
4424 : /* there should be just one matching expression */
4425 156 : break;
4426 : }
4427 :
4428 120 : idx++;
4429 : }
4430 : }
4431 :
4432 : /* Find the specific item that exactly matches the combination */
4433 852 : for (i = 0; i < stats->nitems; i++)
4434 : {
4435 : int j;
4436 852 : MVNDistinctItem *tmpitem = &stats->items[i];
4437 :
4438 852 : if (tmpitem->nattributes != bms_num_members(matched))
4439 162 : continue;
4440 :
4441 : /* assume it's the right item */
4442 690 : item = tmpitem;
4443 :
4444 : /* check that all item attributes/expressions fit the match */
4445 1656 : for (j = 0; j < tmpitem->nattributes; j++)
4446 : {
4447 1242 : AttrNumber attnum = tmpitem->attributes[j];
4448 :
4449 : /*
4450 : * Thanks to how we constructed the matched bitmap above, we
4451 : * can just offset all attnums the same way.
4452 : */
4453 1242 : attnum = attnum + attnum_offset;
4454 :
4455 1242 : if (!bms_is_member(attnum, matched))
4456 : {
4457 : /* nah, it's not this item */
4458 276 : item = NULL;
4459 276 : break;
4460 : }
4461 : }
4462 :
4463 : /*
4464 : * If the item has all the matched attributes, we know it's the
4465 : * right one - there can't be a better one. matching more.
4466 : */
4467 690 : if (item)
4468 414 : break;
4469 : }
4470 :
4471 : /*
4472 : * Make sure we found an item. There has to be one, because ndistinct
4473 : * statistics includes all combinations of attributes.
4474 : */
4475 414 : if (!item)
4476 0 : elog(ERROR, "corrupt MVNDistinct entry");
4477 :
4478 : /* Form the output varinfo list, keeping only unmatched ones */
4479 1452 : foreach(lc, *varinfos)
4480 : {
4481 1038 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4482 : ListCell *lc3;
4483 1038 : bool found = false;
4484 :
4485 : /*
4486 : * Let's look at plain variables first, because it's the most
4487 : * common case and the check is quite cheap. We can simply get the
4488 : * attnum and check (with an offset) matched bitmap.
4489 : */
4490 1038 : if (IsA(varinfo->var, Var))
4491 846 : {
4492 852 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4493 :
4494 : /*
4495 : * If it's a system attribute, we're done. We don't support
4496 : * extended statistics on system attributes, so it's clearly
4497 : * not matched. Just keep the expression and continue.
4498 : */
4499 852 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4500 : {
4501 6 : newlist = lappend(newlist, varinfo);
4502 6 : continue;
4503 : }
4504 :
4505 : /* apply the same offset as above */
4506 846 : attnum += attnum_offset;
4507 :
4508 : /* if it's not matched, keep the varinfo */
4509 846 : if (!bms_is_member(attnum, matched))
4510 120 : newlist = lappend(newlist, varinfo);
4511 :
4512 : /* The rest of the loop deals with complex expressions. */
4513 846 : continue;
4514 : }
4515 :
4516 : /*
4517 : * Process complex expressions, not just simple Vars.
4518 : *
4519 : * First, we search for an exact match of an expression. If we
4520 : * find one, we can just discard the whole GroupVarInfo, with all
4521 : * the variables we extracted from it.
4522 : *
4523 : * Otherwise we inspect the individual vars, and try matching it
4524 : * to variables in the item.
4525 : */
4526 306 : foreach(lc3, matched_info->exprs)
4527 : {
4528 276 : Node *expr = (Node *) lfirst(lc3);
4529 :
4530 276 : if (equal(varinfo->var, expr))
4531 : {
4532 156 : found = true;
4533 156 : break;
4534 : }
4535 : }
4536 :
4537 : /* found exact match, skip */
4538 186 : if (found)
4539 156 : continue;
4540 :
4541 30 : newlist = lappend(newlist, varinfo);
4542 : }
4543 :
4544 414 : *varinfos = newlist;
4545 414 : *ndistinct = item->ndistinct;
4546 414 : return true;
4547 : }
4548 :
4549 0 : return false;
4550 : }
4551 :
4552 : /*
4553 : * convert_to_scalar
4554 : * Convert non-NULL values of the indicated types to the comparison
4555 : * scale needed by scalarineqsel().
4556 : * Returns "true" if successful.
4557 : *
4558 : * XXX this routine is a hack: ideally we should look up the conversion
4559 : * subroutines in pg_type.
4560 : *
4561 : * All numeric datatypes are simply converted to their equivalent
4562 : * "double" values. (NUMERIC values that are outside the range of "double"
4563 : * are clamped to +/- HUGE_VAL.)
4564 : *
4565 : * String datatypes are converted by convert_string_to_scalar(),
4566 : * which is explained below. The reason why this routine deals with
4567 : * three values at a time, not just one, is that we need it for strings.
4568 : *
4569 : * The bytea datatype is just enough different from strings that it has
4570 : * to be treated separately.
4571 : *
4572 : * The several datatypes representing absolute times are all converted
4573 : * to Timestamp, which is actually an int64, and then we promote that to
4574 : * a double. Note this will give correct results even for the "special"
4575 : * values of Timestamp, since those are chosen to compare correctly;
4576 : * see timestamp_cmp.
4577 : *
4578 : * The several datatypes representing relative times (intervals) are all
4579 : * converted to measurements expressed in seconds.
4580 : */
4581 : static bool
4582 89388 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4583 : Datum lobound, Datum hibound, Oid boundstypid,
4584 : double *scaledlobound, double *scaledhibound)
4585 : {
4586 89388 : bool failure = false;
4587 :
4588 : /*
4589 : * Both the valuetypid and the boundstypid should exactly match the
4590 : * declared input type(s) of the operator we are invoked for. However,
4591 : * extensions might try to use scalarineqsel as estimator for operators
4592 : * with input type(s) we don't handle here; in such cases, we want to
4593 : * return false, not fail. In any case, we mustn't assume that valuetypid
4594 : * and boundstypid are identical.
4595 : *
4596 : * XXX The histogram we are interpolating between points of could belong
4597 : * to a column that's only binary-compatible with the declared type. In
4598 : * essence we are assuming that the semantics of binary-compatible types
4599 : * are enough alike that we can use a histogram generated with one type's
4600 : * operators to estimate selectivity for the other's. This is outright
4601 : * wrong in some cases --- in particular signed versus unsigned
4602 : * interpretation could trip us up. But it's useful enough in the
4603 : * majority of cases that we do it anyway. Should think about more
4604 : * rigorous ways to do it.
4605 : */
4606 89388 : switch (valuetypid)
4607 : {
4608 : /*
4609 : * Built-in numeric types
4610 : */
4611 82470 : case BOOLOID:
4612 : case INT2OID:
4613 : case INT4OID:
4614 : case INT8OID:
4615 : case FLOAT4OID:
4616 : case FLOAT8OID:
4617 : case NUMERICOID:
4618 : case OIDOID:
4619 : case REGPROCOID:
4620 : case REGPROCEDUREOID:
4621 : case REGOPEROID:
4622 : case REGOPERATOROID:
4623 : case REGCLASSOID:
4624 : case REGTYPEOID:
4625 : case REGCOLLATIONOID:
4626 : case REGCONFIGOID:
4627 : case REGDICTIONARYOID:
4628 : case REGROLEOID:
4629 : case REGNAMESPACEOID:
4630 : case REGDATABASEOID:
4631 82470 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4632 : &failure);
4633 82470 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4634 : &failure);
4635 82470 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4636 : &failure);
4637 82470 : return !failure;
4638 :
4639 : /*
4640 : * Built-in string types
4641 : */
4642 6918 : case CHAROID:
4643 : case BPCHAROID:
4644 : case VARCHAROID:
4645 : case TEXTOID:
4646 : case NAMEOID:
4647 : {
4648 6918 : char *valstr = convert_string_datum(value, valuetypid,
4649 : collid, &failure);
4650 6918 : char *lostr = convert_string_datum(lobound, boundstypid,
4651 : collid, &failure);
4652 6918 : char *histr = convert_string_datum(hibound, boundstypid,
4653 : collid, &failure);
4654 :
4655 : /*
4656 : * Bail out if any of the values is not of string type. We
4657 : * might leak converted strings for the other value(s), but
4658 : * that's not worth troubling over.
4659 : */
4660 6918 : if (failure)
4661 0 : return false;
4662 :
4663 6918 : convert_string_to_scalar(valstr, scaledvalue,
4664 : lostr, scaledlobound,
4665 : histr, scaledhibound);
4666 6918 : pfree(valstr);
4667 6918 : pfree(lostr);
4668 6918 : pfree(histr);
4669 6918 : return true;
4670 : }
4671 :
4672 : /*
4673 : * Built-in bytea type
4674 : */
4675 0 : case BYTEAOID:
4676 : {
4677 : /* We only support bytea vs bytea comparison */
4678 0 : if (boundstypid != BYTEAOID)
4679 0 : return false;
4680 0 : convert_bytea_to_scalar(value, scaledvalue,
4681 : lobound, scaledlobound,
4682 : hibound, scaledhibound);
4683 0 : return true;
4684 : }
4685 :
4686 : /*
4687 : * Built-in time types
4688 : */
4689 0 : case TIMESTAMPOID:
4690 : case TIMESTAMPTZOID:
4691 : case DATEOID:
4692 : case INTERVALOID:
4693 : case TIMEOID:
4694 : case TIMETZOID:
4695 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
4696 : &failure);
4697 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
4698 : &failure);
4699 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
4700 : &failure);
4701 0 : return !failure;
4702 :
4703 : /*
4704 : * Built-in network types
4705 : */
4706 0 : case INETOID:
4707 : case CIDROID:
4708 : case MACADDROID:
4709 : case MACADDR8OID:
4710 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
4711 : &failure);
4712 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
4713 : &failure);
4714 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
4715 : &failure);
4716 0 : return !failure;
4717 : }
4718 : /* Don't know how to convert */
4719 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
4720 0 : return false;
4721 : }
4722 :
4723 : /*
4724 : * Do convert_to_scalar()'s work for any numeric data type.
4725 : *
4726 : * On failure (e.g., unsupported typid), set *failure to true;
4727 : * otherwise, that variable is not changed.
4728 : */
4729 : static double
4730 247410 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
4731 : {
4732 247410 : switch (typid)
4733 : {
4734 0 : case BOOLOID:
4735 0 : return (double) DatumGetBool(value);
4736 12 : case INT2OID:
4737 12 : return (double) DatumGetInt16(value);
4738 31410 : case INT4OID:
4739 31410 : return (double) DatumGetInt32(value);
4740 0 : case INT8OID:
4741 0 : return (double) DatumGetInt64(value);
4742 0 : case FLOAT4OID:
4743 0 : return (double) DatumGetFloat4(value);
4744 54 : case FLOAT8OID:
4745 54 : return (double) DatumGetFloat8(value);
4746 0 : case NUMERICOID:
4747 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
4748 0 : return (double)
4749 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
4750 : value));
4751 215934 : case OIDOID:
4752 : case REGPROCOID:
4753 : case REGPROCEDUREOID:
4754 : case REGOPEROID:
4755 : case REGOPERATOROID:
4756 : case REGCLASSOID:
4757 : case REGTYPEOID:
4758 : case REGCOLLATIONOID:
4759 : case REGCONFIGOID:
4760 : case REGDICTIONARYOID:
4761 : case REGROLEOID:
4762 : case REGNAMESPACEOID:
4763 : case REGDATABASEOID:
4764 : /* we can treat OIDs as integers... */
4765 215934 : return (double) DatumGetObjectId(value);
4766 : }
4767 :
4768 0 : *failure = true;
4769 0 : return 0;
4770 : }
4771 :
4772 : /*
4773 : * Do convert_to_scalar()'s work for any character-string data type.
4774 : *
4775 : * String datatypes are converted to a scale that ranges from 0 to 1,
4776 : * where we visualize the bytes of the string as fractional digits.
4777 : *
4778 : * We do not want the base to be 256, however, since that tends to
4779 : * generate inflated selectivity estimates; few databases will have
4780 : * occurrences of all 256 possible byte values at each position.
4781 : * Instead, use the smallest and largest byte values seen in the bounds
4782 : * as the estimated range for each byte, after some fudging to deal with
4783 : * the fact that we probably aren't going to see the full range that way.
4784 : *
4785 : * An additional refinement is that we discard any common prefix of the
4786 : * three strings before computing the scaled values. This allows us to
4787 : * "zoom in" when we encounter a narrow data range. An example is a phone
4788 : * number database where all the values begin with the same area code.
4789 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
4790 : * so this is more likely to happen than you might think.)
4791 : */
4792 : static void
4793 6918 : convert_string_to_scalar(char *value,
4794 : double *scaledvalue,
4795 : char *lobound,
4796 : double *scaledlobound,
4797 : char *hibound,
4798 : double *scaledhibound)
4799 : {
4800 : int rangelo,
4801 : rangehi;
4802 : char *sptr;
4803 :
4804 6918 : rangelo = rangehi = (unsigned char) hibound[0];
4805 85996 : for (sptr = lobound; *sptr; sptr++)
4806 : {
4807 79078 : if (rangelo > (unsigned char) *sptr)
4808 16456 : rangelo = (unsigned char) *sptr;
4809 79078 : if (rangehi < (unsigned char) *sptr)
4810 8676 : rangehi = (unsigned char) *sptr;
4811 : }
4812 83440 : for (sptr = hibound; *sptr; sptr++)
4813 : {
4814 76522 : if (rangelo > (unsigned char) *sptr)
4815 1030 : rangelo = (unsigned char) *sptr;
4816 76522 : if (rangehi < (unsigned char) *sptr)
4817 3318 : rangehi = (unsigned char) *sptr;
4818 : }
4819 : /* If range includes any upper-case ASCII chars, make it include all */
4820 6918 : if (rangelo <= 'Z' && rangehi >= 'A')
4821 : {
4822 1210 : if (rangelo > 'A')
4823 222 : rangelo = 'A';
4824 1210 : if (rangehi < 'Z')
4825 480 : rangehi = 'Z';
4826 : }
4827 : /* Ditto lower-case */
4828 6918 : if (rangelo <= 'z' && rangehi >= 'a')
4829 : {
4830 6416 : if (rangelo > 'a')
4831 30 : rangelo = 'a';
4832 6416 : if (rangehi < 'z')
4833 6346 : rangehi = 'z';
4834 : }
4835 : /* Ditto digits */
4836 6918 : if (rangelo <= '9' && rangehi >= '0')
4837 : {
4838 530 : if (rangelo > '0')
4839 426 : rangelo = '0';
4840 530 : if (rangehi < '9')
4841 14 : rangehi = '9';
4842 : }
4843 :
4844 : /*
4845 : * If range includes less than 10 chars, assume we have not got enough
4846 : * data, and make it include regular ASCII set.
4847 : */
4848 6918 : if (rangehi - rangelo < 9)
4849 : {
4850 0 : rangelo = ' ';
4851 0 : rangehi = 127;
4852 : }
4853 :
4854 : /*
4855 : * Now strip any common prefix of the three strings.
4856 : */
4857 14450 : while (*lobound)
4858 : {
4859 14450 : if (*lobound != *hibound || *lobound != *value)
4860 : break;
4861 7532 : lobound++, hibound++, value++;
4862 : }
4863 :
4864 : /*
4865 : * Now we can do the conversions.
4866 : */
4867 6918 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
4868 6918 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
4869 6918 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
4870 6918 : }
4871 :
4872 : static double
4873 20754 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
4874 : {
4875 20754 : int slen = strlen(value);
4876 : double num,
4877 : denom,
4878 : base;
4879 :
4880 20754 : if (slen <= 0)
4881 0 : return 0.0; /* empty string has scalar value 0 */
4882 :
4883 : /*
4884 : * There seems little point in considering more than a dozen bytes from
4885 : * the string. Since base is at least 10, that will give us nominal
4886 : * resolution of at least 12 decimal digits, which is surely far more
4887 : * precision than this estimation technique has got anyway (especially in
4888 : * non-C locales). Also, even with the maximum possible base of 256, this
4889 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
4890 : * overflow on any known machine.
4891 : */
4892 20754 : if (slen > 12)
4893 5326 : slen = 12;
4894 :
4895 : /* Convert initial characters to fraction */
4896 20754 : base = rangehi - rangelo + 1;
4897 20754 : num = 0.0;
4898 20754 : denom = base;
4899 172586 : while (slen-- > 0)
4900 : {
4901 151832 : int ch = (unsigned char) *value++;
4902 :
4903 151832 : if (ch < rangelo)
4904 188 : ch = rangelo - 1;
4905 151644 : else if (ch > rangehi)
4906 0 : ch = rangehi + 1;
4907 151832 : num += ((double) (ch - rangelo)) / denom;
4908 151832 : denom *= base;
4909 : }
4910 :
4911 20754 : return num;
4912 : }
4913 :
4914 : /*
4915 : * Convert a string-type Datum into a palloc'd, null-terminated string.
4916 : *
4917 : * On failure (e.g., unsupported typid), set *failure to true;
4918 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
4919 : *
4920 : * When using a non-C locale, we must pass the string through pg_strxfrm()
4921 : * before continuing, so as to generate correct locale-specific results.
4922 : */
4923 : static char *
4924 20754 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
4925 : {
4926 : char *val;
4927 : pg_locale_t mylocale;
4928 :
4929 20754 : switch (typid)
4930 : {
4931 0 : case CHAROID:
4932 0 : val = (char *) palloc(2);
4933 0 : val[0] = DatumGetChar(value);
4934 0 : val[1] = '\0';
4935 0 : break;
4936 6382 : case BPCHAROID:
4937 : case VARCHAROID:
4938 : case TEXTOID:
4939 6382 : val = TextDatumGetCString(value);
4940 6382 : break;
4941 14372 : case NAMEOID:
4942 : {
4943 14372 : NameData *nm = (NameData *) DatumGetPointer(value);
4944 :
4945 14372 : val = pstrdup(NameStr(*nm));
4946 14372 : break;
4947 : }
4948 0 : default:
4949 0 : *failure = true;
4950 0 : return NULL;
4951 : }
4952 :
4953 20754 : mylocale = pg_newlocale_from_collation(collid);
4954 :
4955 20754 : if (!mylocale->collate_is_c)
4956 : {
4957 : char *xfrmstr;
4958 : size_t xfrmlen;
4959 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
4960 :
4961 : /*
4962 : * XXX: We could guess at a suitable output buffer size and only call
4963 : * pg_strxfrm() twice if our guess is too small.
4964 : *
4965 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
4966 : * bogus data or set an error. This is not really a problem unless it
4967 : * crashes since it will only give an estimation error and nothing
4968 : * fatal.
4969 : *
4970 : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
4971 : * some cases, libc strxfrm() may return the wrong results, but that
4972 : * will only lead to an estimation error.
4973 : */
4974 72 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
4975 : #ifdef WIN32
4976 :
4977 : /*
4978 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
4979 : * of trying to allocate this much memory (and fail), just return the
4980 : * original string unmodified as if we were in the C locale.
4981 : */
4982 : if (xfrmlen == INT_MAX)
4983 : return val;
4984 : #endif
4985 72 : xfrmstr = (char *) palloc(xfrmlen + 1);
4986 72 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
4987 :
4988 : /*
4989 : * Some systems (e.g., glibc) can return a smaller value from the
4990 : * second call than the first; thus the Assert must be <= not ==.
4991 : */
4992 : Assert(xfrmlen2 <= xfrmlen);
4993 72 : pfree(val);
4994 72 : val = xfrmstr;
4995 : }
4996 :
4997 20754 : return val;
4998 : }
4999 :
5000 : /*
5001 : * Do convert_to_scalar()'s work for any bytea data type.
5002 : *
5003 : * Very similar to convert_string_to_scalar except we can't assume
5004 : * null-termination and therefore pass explicit lengths around.
5005 : *
5006 : * Also, assumptions about likely "normal" ranges of characters have been
5007 : * removed - a data range of 0..255 is always used, for now. (Perhaps
5008 : * someday we will add information about actual byte data range to
5009 : * pg_statistic.)
5010 : */
5011 : static void
5012 0 : convert_bytea_to_scalar(Datum value,
5013 : double *scaledvalue,
5014 : Datum lobound,
5015 : double *scaledlobound,
5016 : Datum hibound,
5017 : double *scaledhibound)
5018 : {
5019 0 : bytea *valuep = DatumGetByteaPP(value);
5020 0 : bytea *loboundp = DatumGetByteaPP(lobound);
5021 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
5022 : int rangelo,
5023 : rangehi,
5024 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
5025 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
5026 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
5027 : i,
5028 : minlen;
5029 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
5030 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
5031 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
5032 :
5033 : /*
5034 : * Assume bytea data is uniformly distributed across all byte values.
5035 : */
5036 0 : rangelo = 0;
5037 0 : rangehi = 255;
5038 :
5039 : /*
5040 : * Now strip any common prefix of the three strings.
5041 : */
5042 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
5043 0 : for (i = 0; i < minlen; i++)
5044 : {
5045 0 : if (*lostr != *histr || *lostr != *valstr)
5046 : break;
5047 0 : lostr++, histr++, valstr++;
5048 0 : loboundlen--, hiboundlen--, valuelen--;
5049 : }
5050 :
5051 : /*
5052 : * Now we can do the conversions.
5053 : */
5054 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
5055 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
5056 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
5057 0 : }
5058 :
5059 : static double
5060 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
5061 : int rangelo, int rangehi)
5062 : {
5063 : double num,
5064 : denom,
5065 : base;
5066 :
5067 0 : if (valuelen <= 0)
5068 0 : return 0.0; /* empty string has scalar value 0 */
5069 :
5070 : /*
5071 : * Since base is 256, need not consider more than about 10 chars (even
5072 : * this many seems like overkill)
5073 : */
5074 0 : if (valuelen > 10)
5075 0 : valuelen = 10;
5076 :
5077 : /* Convert initial characters to fraction */
5078 0 : base = rangehi - rangelo + 1;
5079 0 : num = 0.0;
5080 0 : denom = base;
5081 0 : while (valuelen-- > 0)
5082 : {
5083 0 : int ch = *value++;
5084 :
5085 0 : if (ch < rangelo)
5086 0 : ch = rangelo - 1;
5087 0 : else if (ch > rangehi)
5088 0 : ch = rangehi + 1;
5089 0 : num += ((double) (ch - rangelo)) / denom;
5090 0 : denom *= base;
5091 : }
5092 :
5093 0 : return num;
5094 : }
5095 :
5096 : /*
5097 : * Do convert_to_scalar()'s work for any timevalue data type.
5098 : *
5099 : * On failure (e.g., unsupported typid), set *failure to true;
5100 : * otherwise, that variable is not changed.
5101 : */
5102 : static double
5103 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5104 : {
5105 0 : switch (typid)
5106 : {
5107 0 : case TIMESTAMPOID:
5108 0 : return DatumGetTimestamp(value);
5109 0 : case TIMESTAMPTZOID:
5110 0 : return DatumGetTimestampTz(value);
5111 0 : case DATEOID:
5112 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
5113 0 : case INTERVALOID:
5114 : {
5115 0 : Interval *interval = DatumGetIntervalP(value);
5116 :
5117 : /*
5118 : * Convert the month part of Interval to days using assumed
5119 : * average month length of 365.25/12.0 days. Not too
5120 : * accurate, but plenty good enough for our purposes.
5121 : *
5122 : * This also works for infinite intervals, which just have all
5123 : * fields set to INT_MIN/INT_MAX, and so will produce a result
5124 : * smaller/larger than any finite interval.
5125 : */
5126 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5127 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5128 : }
5129 0 : case TIMEOID:
5130 0 : return DatumGetTimeADT(value);
5131 0 : case TIMETZOID:
5132 : {
5133 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5134 :
5135 : /* use GMT-equivalent time */
5136 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5137 : }
5138 : }
5139 :
5140 0 : *failure = true;
5141 0 : return 0;
5142 : }
5143 :
5144 :
5145 : /*
5146 : * get_restriction_variable
5147 : * Examine the args of a restriction clause to see if it's of the
5148 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5149 : * where "variable" could be either a Var or an expression in vars of a
5150 : * single relation. If so, extract information about the variable,
5151 : * and also indicate which side it was on and the other argument.
5152 : *
5153 : * Inputs:
5154 : * root: the planner info
5155 : * args: clause argument list
5156 : * varRelid: see specs for restriction selectivity functions
5157 : *
5158 : * Outputs: (these are valid only if true is returned)
5159 : * *vardata: gets information about variable (see examine_variable)
5160 : * *other: gets other clause argument, aggressively reduced to a constant
5161 : * *varonleft: set true if variable is on the left, false if on the right
5162 : *
5163 : * Returns true if a variable is identified, otherwise false.
5164 : *
5165 : * Note: if there are Vars on both sides of the clause, we must fail, because
5166 : * callers are expecting that the other side will act like a pseudoconstant.
5167 : */
5168 : bool
5169 737324 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5170 : VariableStatData *vardata, Node **other,
5171 : bool *varonleft)
5172 : {
5173 : Node *left,
5174 : *right;
5175 : VariableStatData rdata;
5176 :
5177 : /* Fail if not a binary opclause (probably shouldn't happen) */
5178 737324 : if (list_length(args) != 2)
5179 0 : return false;
5180 :
5181 737324 : left = (Node *) linitial(args);
5182 737324 : right = (Node *) lsecond(args);
5183 :
5184 : /*
5185 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5186 : * relations will be treated as pseudoconstants.
5187 : */
5188 737324 : examine_variable(root, left, varRelid, vardata);
5189 737324 : examine_variable(root, right, varRelid, &rdata);
5190 :
5191 : /*
5192 : * If one side is a variable and the other not, we win.
5193 : */
5194 737324 : if (vardata->rel && rdata.rel == NULL)
5195 : {
5196 663098 : *varonleft = true;
5197 663098 : *other = estimate_expression_value(root, rdata.var);
5198 : /* Assume we need no ReleaseVariableStats(rdata) here */
5199 663092 : return true;
5200 : }
5201 :
5202 74226 : if (vardata->rel == NULL && rdata.rel)
5203 : {
5204 67994 : *varonleft = false;
5205 67994 : *other = estimate_expression_value(root, vardata->var);
5206 : /* Assume we need no ReleaseVariableStats(*vardata) here */
5207 67994 : *vardata = rdata;
5208 67994 : return true;
5209 : }
5210 :
5211 : /* Oops, clause has wrong structure (probably var op var) */
5212 6232 : ReleaseVariableStats(*vardata);
5213 6232 : ReleaseVariableStats(rdata);
5214 :
5215 6232 : return false;
5216 : }
5217 :
5218 : /*
5219 : * get_join_variables
5220 : * Apply examine_variable() to each side of a join clause.
5221 : * Also, attempt to identify whether the join clause has the same
5222 : * or reversed sense compared to the SpecialJoinInfo.
5223 : *
5224 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5225 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5226 : * where we can't tell for sure, we default to assuming it's normal.
5227 : */
5228 : void
5229 225334 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5230 : VariableStatData *vardata1, VariableStatData *vardata2,
5231 : bool *join_is_reversed)
5232 : {
5233 : Node *left,
5234 : *right;
5235 :
5236 225334 : if (list_length(args) != 2)
5237 0 : elog(ERROR, "join operator should take two arguments");
5238 :
5239 225334 : left = (Node *) linitial(args);
5240 225334 : right = (Node *) lsecond(args);
5241 :
5242 225334 : examine_variable(root, left, 0, vardata1);
5243 225334 : examine_variable(root, right, 0, vardata2);
5244 :
5245 450488 : if (vardata1->rel &&
5246 225154 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
5247 82964 : *join_is_reversed = true; /* var1 is on RHS */
5248 284594 : else if (vardata2->rel &&
5249 142224 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
5250 132 : *join_is_reversed = true; /* var2 is on LHS */
5251 : else
5252 142238 : *join_is_reversed = false;
5253 225334 : }
5254 :
5255 : /* statext_expressions_load copies the tuple, so just pfree it. */
5256 : static void
5257 1644 : ReleaseDummy(HeapTuple tuple)
5258 : {
5259 1644 : pfree(tuple);
5260 1644 : }
5261 :
5262 : /*
5263 : * examine_variable
5264 : * Try to look up statistical data about an expression.
5265 : * Fill in a VariableStatData struct to describe the expression.
5266 : *
5267 : * Inputs:
5268 : * root: the planner info
5269 : * node: the expression tree to examine
5270 : * varRelid: see specs for restriction selectivity functions
5271 : *
5272 : * Outputs: *vardata is filled as follows:
5273 : * var: the input expression (with any binary relabeling stripped, if
5274 : * it is or contains a variable; but otherwise the type is preserved)
5275 : * rel: RelOptInfo for relation containing variable; NULL if expression
5276 : * contains no Vars (NOTE this could point to a RelOptInfo of a
5277 : * subquery, not one in the current query).
5278 : * statsTuple: the pg_statistic entry for the variable, if one exists;
5279 : * otherwise NULL.
5280 : * freefunc: pointer to a function to release statsTuple with.
5281 : * vartype: exposed type of the expression; this should always match
5282 : * the declared input type of the operator we are estimating for.
5283 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5284 : * commonly the same as the exposed type of the variable argument,
5285 : * but can be different in binary-compatible-type cases.
5286 : * isunique: true if we were able to match the var to a unique index, a
5287 : * single-column DISTINCT or GROUP-BY clause, implying its values are
5288 : * unique for this query. (Caution: this should be trusted for
5289 : * statistical purposes only, since we do not check indimmediate nor
5290 : * verify that the exact same definition of equality applies.)
5291 : * acl_ok: true if current user has permission to read all table rows from
5292 : * the column(s) underlying the pg_statistic entry. This is consulted by
5293 : * statistic_proc_security_check().
5294 : *
5295 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5296 : */
5297 : void
5298 2887952 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5299 : VariableStatData *vardata)
5300 : {
5301 : Node *basenode;
5302 : Relids varnos;
5303 : Relids basevarnos;
5304 : RelOptInfo *onerel;
5305 :
5306 : /* Make sure we don't return dangling pointers in vardata */
5307 20215664 : MemSet(vardata, 0, sizeof(VariableStatData));
5308 :
5309 : /* Save the exposed type of the expression */
5310 2887952 : vardata->vartype = exprType(node);
5311 :
5312 : /* Look inside any binary-compatible relabeling */
5313 :
5314 2887952 : if (IsA(node, RelabelType))
5315 46996 : basenode = (Node *) ((RelabelType *) node)->arg;
5316 : else
5317 2840956 : basenode = node;
5318 :
5319 : /* Fast path for a simple Var */
5320 :
5321 2887952 : if (IsA(basenode, Var) &&
5322 666102 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5323 : {
5324 2051508 : Var *var = (Var *) basenode;
5325 :
5326 : /* Set up result fields other than the stats tuple */
5327 2051508 : vardata->var = basenode; /* return Var without relabeling */
5328 2051508 : vardata->rel = find_base_rel(root, var->varno);
5329 2051508 : vardata->atttype = var->vartype;
5330 2051508 : vardata->atttypmod = var->vartypmod;
5331 2051508 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5332 :
5333 : /* Try to locate some stats */
5334 2051508 : examine_simple_variable(root, var, vardata);
5335 :
5336 2051508 : return;
5337 : }
5338 :
5339 : /*
5340 : * Okay, it's a more complicated expression. Determine variable
5341 : * membership. Note that when varRelid isn't zero, only vars of that
5342 : * relation are considered "real" vars.
5343 : */
5344 836444 : varnos = pull_varnos(root, basenode);
5345 836444 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5346 :
5347 836444 : onerel = NULL;
5348 :
5349 836444 : if (bms_is_empty(basevarnos))
5350 : {
5351 : /* No Vars at all ... must be pseudo-constant clause */
5352 : }
5353 : else
5354 : {
5355 : int relid;
5356 :
5357 : /* Check if the expression is in vars of a single base relation */
5358 398020 : if (bms_get_singleton_member(basevarnos, &relid))
5359 : {
5360 393820 : if (varRelid == 0 || varRelid == relid)
5361 : {
5362 59480 : onerel = find_base_rel(root, relid);
5363 59480 : vardata->rel = onerel;
5364 59480 : node = basenode; /* strip any relabeling */
5365 : }
5366 : /* else treat it as a constant */
5367 : }
5368 : else
5369 : {
5370 : /* varnos has multiple relids */
5371 4200 : if (varRelid == 0)
5372 : {
5373 : /* treat it as a variable of a join relation */
5374 3852 : vardata->rel = find_join_rel(root, varnos);
5375 3852 : node = basenode; /* strip any relabeling */
5376 : }
5377 348 : else if (bms_is_member(varRelid, varnos))
5378 : {
5379 : /* ignore the vars belonging to other relations */
5380 174 : vardata->rel = find_base_rel(root, varRelid);
5381 174 : node = basenode; /* strip any relabeling */
5382 : /* note: no point in expressional-index search here */
5383 : }
5384 : /* else treat it as a constant */
5385 : }
5386 : }
5387 :
5388 836444 : bms_free(basevarnos);
5389 :
5390 836444 : vardata->var = node;
5391 836444 : vardata->atttype = exprType(node);
5392 836444 : vardata->atttypmod = exprTypmod(node);
5393 :
5394 836444 : if (onerel)
5395 : {
5396 : /*
5397 : * We have an expression in vars of a single relation. Try to match
5398 : * it to expressional index columns, in hopes of finding some
5399 : * statistics.
5400 : *
5401 : * Note that we consider all index columns including INCLUDE columns,
5402 : * since there could be stats for such columns. But the test for
5403 : * uniqueness needs to be warier.
5404 : *
5405 : * XXX it's conceivable that there are multiple matches with different
5406 : * index opfamilies; if so, we need to pick one that matches the
5407 : * operator we are estimating for. FIXME later.
5408 : */
5409 : ListCell *ilist;
5410 : ListCell *slist;
5411 :
5412 : /*
5413 : * The nullingrels bits within the expression could prevent us from
5414 : * matching it to expressional index columns or to the expressions in
5415 : * extended statistics. So strip them out first.
5416 : */
5417 59480 : if (bms_overlap(varnos, root->outer_join_rels))
5418 3052 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5419 :
5420 119044 : foreach(ilist, onerel->indexlist)
5421 : {
5422 62522 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5423 : ListCell *indexpr_item;
5424 : int pos;
5425 :
5426 62522 : indexpr_item = list_head(index->indexprs);
5427 62522 : if (indexpr_item == NULL)
5428 57632 : continue; /* no expressions here... */
5429 :
5430 6894 : for (pos = 0; pos < index->ncolumns; pos++)
5431 : {
5432 4962 : if (index->indexkeys[pos] == 0)
5433 : {
5434 : Node *indexkey;
5435 :
5436 4890 : if (indexpr_item == NULL)
5437 0 : elog(ERROR, "too few entries in indexprs list");
5438 4890 : indexkey = (Node *) lfirst(indexpr_item);
5439 4890 : if (indexkey && IsA(indexkey, RelabelType))
5440 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5441 4890 : if (equal(node, indexkey))
5442 : {
5443 : /*
5444 : * Found a match ... is it a unique index? Tests here
5445 : * should match has_unique_index().
5446 : */
5447 3594 : if (index->unique &&
5448 438 : index->nkeycolumns == 1 &&
5449 438 : pos == 0 &&
5450 438 : (index->indpred == NIL || index->predOK))
5451 438 : vardata->isunique = true;
5452 :
5453 : /*
5454 : * Has it got stats? We only consider stats for
5455 : * non-partial indexes, since partial indexes probably
5456 : * don't reflect whole-relation statistics; the above
5457 : * check for uniqueness is the only info we take from
5458 : * a partial index.
5459 : *
5460 : * An index stats hook, however, must make its own
5461 : * decisions about what to do with partial indexes.
5462 : */
5463 3594 : if (get_index_stats_hook &&
5464 0 : (*get_index_stats_hook) (root, index->indexoid,
5465 0 : pos + 1, vardata))
5466 : {
5467 : /*
5468 : * The hook took control of acquiring a stats
5469 : * tuple. If it did supply a tuple, it'd better
5470 : * have supplied a freefunc.
5471 : */
5472 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5473 0 : !vardata->freefunc)
5474 0 : elog(ERROR, "no function provided to release variable stats with");
5475 : }
5476 3594 : else if (index->indpred == NIL)
5477 : {
5478 3594 : vardata->statsTuple =
5479 7188 : SearchSysCache3(STATRELATTINH,
5480 : ObjectIdGetDatum(index->indexoid),
5481 3594 : Int16GetDatum(pos + 1),
5482 : BoolGetDatum(false));
5483 3594 : vardata->freefunc = ReleaseSysCache;
5484 :
5485 3594 : if (HeapTupleIsValid(vardata->statsTuple))
5486 : {
5487 : /*
5488 : * Test if user has permission to access all
5489 : * rows from the index's table.
5490 : *
5491 : * For simplicity, we insist on the whole
5492 : * table being selectable, rather than trying
5493 : * to identify which column(s) the index
5494 : * depends on.
5495 : *
5496 : * Note that for an inheritance child,
5497 : * permissions are checked on the inheritance
5498 : * root parent, and whole-table select
5499 : * privilege on the parent doesn't quite
5500 : * guarantee that the user could read all
5501 : * columns of the child. But in practice it's
5502 : * unlikely that any interesting security
5503 : * violation could result from allowing access
5504 : * to the expression index's stats, so we
5505 : * allow it anyway. See similar code in
5506 : * examine_simple_variable() for additional
5507 : * comments.
5508 : */
5509 2958 : vardata->acl_ok =
5510 2958 : all_rows_selectable(root,
5511 2958 : index->rel->relid,
5512 : NULL);
5513 : }
5514 : else
5515 : {
5516 : /* suppress leakproofness checks later */
5517 636 : vardata->acl_ok = true;
5518 : }
5519 : }
5520 3594 : if (vardata->statsTuple)
5521 2958 : break;
5522 : }
5523 1932 : indexpr_item = lnext(index->indexprs, indexpr_item);
5524 : }
5525 : }
5526 4890 : if (vardata->statsTuple)
5527 2958 : break;
5528 : }
5529 :
5530 : /*
5531 : * Search extended statistics for one with a matching expression.
5532 : * There might be multiple ones, so just grab the first one. In the
5533 : * future, we might consider the statistics target (and pick the most
5534 : * accurate statistics) and maybe some other parameters.
5535 : */
5536 63578 : foreach(slist, onerel->statlist)
5537 : {
5538 4386 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5539 4386 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5540 : ListCell *expr_item;
5541 : int pos;
5542 :
5543 : /*
5544 : * Stop once we've found statistics for the expression (either
5545 : * from extended stats, or for an index in the preceding loop).
5546 : */
5547 4386 : if (vardata->statsTuple)
5548 288 : break;
5549 :
5550 : /* skip stats without per-expression stats */
5551 4098 : if (info->kind != STATS_EXT_EXPRESSIONS)
5552 2094 : continue;
5553 :
5554 : /* skip stats with mismatching stxdinherit value */
5555 2004 : if (info->inherit != rte->inh)
5556 6 : continue;
5557 :
5558 1998 : pos = 0;
5559 3300 : foreach(expr_item, info->exprs)
5560 : {
5561 2946 : Node *expr = (Node *) lfirst(expr_item);
5562 :
5563 : Assert(expr);
5564 :
5565 : /* strip RelabelType before comparing it */
5566 2946 : if (expr && IsA(expr, RelabelType))
5567 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5568 :
5569 : /* found a match, see if we can extract pg_statistic row */
5570 2946 : if (equal(node, expr))
5571 : {
5572 : /*
5573 : * XXX Not sure if we should cache the tuple somewhere.
5574 : * Now we just create a new copy every time.
5575 : */
5576 1644 : vardata->statsTuple =
5577 1644 : statext_expressions_load(info->statOid, rte->inh, pos);
5578 :
5579 1644 : vardata->freefunc = ReleaseDummy;
5580 :
5581 : /*
5582 : * Test if user has permission to access all rows from the
5583 : * table.
5584 : *
5585 : * For simplicity, we insist on the whole table being
5586 : * selectable, rather than trying to identify which
5587 : * column(s) the statistics object depends on.
5588 : *
5589 : * Note that for an inheritance child, permissions are
5590 : * checked on the inheritance root parent, and whole-table
5591 : * select privilege on the parent doesn't quite guarantee
5592 : * that the user could read all columns of the child. But
5593 : * in practice it's unlikely that any interesting security
5594 : * violation could result from allowing access to the
5595 : * expression stats, so we allow it anyway. See similar
5596 : * code in examine_simple_variable() for additional
5597 : * comments.
5598 : */
5599 1644 : vardata->acl_ok = all_rows_selectable(root,
5600 : onerel->relid,
5601 : NULL);
5602 :
5603 1644 : break;
5604 : }
5605 :
5606 1302 : pos++;
5607 : }
5608 : }
5609 : }
5610 :
5611 836444 : bms_free(varnos);
5612 : }
5613 :
5614 : /*
5615 : * examine_simple_variable
5616 : * Handle a simple Var for examine_variable
5617 : *
5618 : * This is split out as a subroutine so that we can recurse to deal with
5619 : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
5620 : *
5621 : * We already filled in all the fields of *vardata except for the stats tuple.
5622 : */
5623 : static void
5624 2057684 : examine_simple_variable(PlannerInfo *root, Var *var,
5625 : VariableStatData *vardata)
5626 : {
5627 2057684 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
5628 :
5629 : Assert(IsA(rte, RangeTblEntry));
5630 :
5631 2057684 : if (get_relation_stats_hook &&
5632 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
5633 : {
5634 : /*
5635 : * The hook took control of acquiring a stats tuple. If it did supply
5636 : * a tuple, it'd better have supplied a freefunc.
5637 : */
5638 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5639 0 : !vardata->freefunc)
5640 0 : elog(ERROR, "no function provided to release variable stats with");
5641 : }
5642 2057684 : else if (rte->rtekind == RTE_RELATION)
5643 : {
5644 : /*
5645 : * Plain table or parent of an inheritance appendrel, so look up the
5646 : * column in pg_statistic
5647 : */
5648 1942518 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
5649 : ObjectIdGetDatum(rte->relid),
5650 1942518 : Int16GetDatum(var->varattno),
5651 1942518 : BoolGetDatum(rte->inh));
5652 1942518 : vardata->freefunc = ReleaseSysCache;
5653 :
5654 1942518 : if (HeapTupleIsValid(vardata->statsTuple))
5655 : {
5656 : /*
5657 : * Test if user has permission to read all rows from this column.
5658 : *
5659 : * This requires that the user has the appropriate SELECT
5660 : * privileges and that there are no securityQuals from security
5661 : * barrier views or RLS policies. If that's not the case, then we
5662 : * only permit leakproof functions to be passed pg_statistic data
5663 : * in vardata, otherwise the functions might reveal data that the
5664 : * user doesn't have permission to see --- see
5665 : * statistic_proc_security_check().
5666 : */
5667 1466694 : vardata->acl_ok =
5668 1466694 : all_rows_selectable(root, var->varno,
5669 1466694 : bms_make_singleton(var->varattno - FirstLowInvalidHeapAttributeNumber));
5670 : }
5671 : else
5672 : {
5673 : /* suppress any possible leakproofness checks later */
5674 475824 : vardata->acl_ok = true;
5675 : }
5676 : }
5677 115166 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
5678 105370 : (rte->rtekind == RTE_CTE && !rte->self_reference))
5679 : {
5680 : /*
5681 : * Plain subquery (not one that was converted to an appendrel) or
5682 : * non-recursive CTE. In either case, we can try to find out what the
5683 : * Var refers to within the subquery. We skip this for appendrel and
5684 : * recursive-CTE cases because any column stats we did find would
5685 : * likely not be very relevant.
5686 : */
5687 : PlannerInfo *subroot;
5688 : Query *subquery;
5689 : List *subtlist;
5690 : TargetEntry *ste;
5691 :
5692 : /*
5693 : * Punt if it's a whole-row var rather than a plain column reference.
5694 : */
5695 16762 : if (var->varattno == InvalidAttrNumber)
5696 0 : return;
5697 :
5698 : /*
5699 : * Otherwise, find the subquery's planner subroot.
5700 : */
5701 16762 : if (rte->rtekind == RTE_SUBQUERY)
5702 : {
5703 : RelOptInfo *rel;
5704 :
5705 : /*
5706 : * Fetch RelOptInfo for subquery. Note that we don't change the
5707 : * rel returned in vardata, since caller expects it to be a rel of
5708 : * the caller's query level. Because we might already be
5709 : * recursing, we can't use that rel pointer either, but have to
5710 : * look up the Var's rel afresh.
5711 : */
5712 9796 : rel = find_base_rel(root, var->varno);
5713 :
5714 9796 : subroot = rel->subroot;
5715 : }
5716 : else
5717 : {
5718 : /* CTE case is more difficult */
5719 : PlannerInfo *cteroot;
5720 : Index levelsup;
5721 : int ndx;
5722 : int plan_id;
5723 : ListCell *lc;
5724 :
5725 : /*
5726 : * Find the referenced CTE, and locate the subroot previously made
5727 : * for it.
5728 : */
5729 6966 : levelsup = rte->ctelevelsup;
5730 6966 : cteroot = root;
5731 13000 : while (levelsup-- > 0)
5732 : {
5733 6034 : cteroot = cteroot->parent_root;
5734 6034 : if (!cteroot) /* shouldn't happen */
5735 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
5736 : }
5737 :
5738 : /*
5739 : * Note: cte_plan_ids can be shorter than cteList, if we are still
5740 : * working on planning the CTEs (ie, this is a side-reference from
5741 : * another CTE). So we mustn't use forboth here.
5742 : */
5743 6966 : ndx = 0;
5744 9188 : foreach(lc, cteroot->parse->cteList)
5745 : {
5746 9188 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
5747 :
5748 9188 : if (strcmp(cte->ctename, rte->ctename) == 0)
5749 6966 : break;
5750 2222 : ndx++;
5751 : }
5752 6966 : if (lc == NULL) /* shouldn't happen */
5753 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
5754 6966 : if (ndx >= list_length(cteroot->cte_plan_ids))
5755 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
5756 6966 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
5757 6966 : if (plan_id <= 0)
5758 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
5759 6966 : subroot = list_nth(root->glob->subroots, plan_id - 1);
5760 : }
5761 :
5762 : /* If the subquery hasn't been planned yet, we have to punt */
5763 16762 : if (subroot == NULL)
5764 0 : return;
5765 : Assert(IsA(subroot, PlannerInfo));
5766 :
5767 : /*
5768 : * We must use the subquery parsetree as mangled by the planner, not
5769 : * the raw version from the RTE, because we need a Var that will refer
5770 : * to the subroot's live RelOptInfos. For instance, if any subquery
5771 : * pullup happened during planning, Vars in the targetlist might have
5772 : * gotten replaced, and we need to see the replacement expressions.
5773 : */
5774 16762 : subquery = subroot->parse;
5775 : Assert(IsA(subquery, Query));
5776 :
5777 : /*
5778 : * Punt if subquery uses set operations or grouping sets, as these
5779 : * will mash underlying columns' stats beyond recognition. (Set ops
5780 : * are particularly nasty; if we forged ahead, we would return stats
5781 : * relevant to only the leftmost subselect...) DISTINCT is also
5782 : * problematic, but we check that later because there is a possibility
5783 : * of learning something even with it.
5784 : */
5785 16762 : if (subquery->setOperations ||
5786 14874 : subquery->groupingSets)
5787 1912 : return;
5788 :
5789 : /* Get the subquery output expression referenced by the upper Var */
5790 14850 : if (subquery->returningList)
5791 206 : subtlist = subquery->returningList;
5792 : else
5793 14644 : subtlist = subquery->targetList;
5794 14850 : ste = get_tle_by_resno(subtlist, var->varattno);
5795 14850 : if (ste == NULL || ste->resjunk)
5796 0 : elog(ERROR, "subquery %s does not have attribute %d",
5797 : rte->eref->aliasname, var->varattno);
5798 14850 : var = (Var *) ste->expr;
5799 :
5800 : /*
5801 : * If subquery uses DISTINCT, we can't make use of any stats for the
5802 : * variable ... but, if it's the only DISTINCT column, we are entitled
5803 : * to consider it unique. We do the test this way so that it works
5804 : * for cases involving DISTINCT ON.
5805 : */
5806 14850 : if (subquery->distinctClause)
5807 : {
5808 1838 : if (list_length(subquery->distinctClause) == 1 &&
5809 616 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
5810 308 : vardata->isunique = true;
5811 : /* cannot go further */
5812 1222 : return;
5813 : }
5814 :
5815 : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
5816 13628 : if (subquery->groupClause)
5817 : {
5818 1040 : if (list_length(subquery->groupClause) == 1 &&
5819 430 : targetIsInSortList(ste, InvalidOid, subquery->groupClause))
5820 334 : vardata->isunique = true;
5821 : /* cannot go further */
5822 610 : return;
5823 : }
5824 :
5825 : /*
5826 : * If the sub-query originated from a view with the security_barrier
5827 : * attribute, we must not look at the variable's statistics, though it
5828 : * seems all right to notice the existence of a DISTINCT clause. So
5829 : * stop here.
5830 : *
5831 : * This is probably a harsher restriction than necessary; it's
5832 : * certainly OK for the selectivity estimator (which is a C function,
5833 : * and therefore omnipotent anyway) to look at the statistics. But
5834 : * many selectivity estimators will happily *invoke the operator
5835 : * function* to try to work out a good estimate - and that's not OK.
5836 : * So for now, don't dig down for stats.
5837 : */
5838 13018 : if (rte->security_barrier)
5839 1350 : return;
5840 :
5841 : /* Can only handle a simple Var of subquery's query level */
5842 11668 : if (var && IsA(var, Var) &&
5843 6176 : var->varlevelsup == 0)
5844 : {
5845 : /*
5846 : * OK, recurse into the subquery. Note that the original setting
5847 : * of vardata->isunique (which will surely be false) is left
5848 : * unchanged in this situation. That's what we want, since even
5849 : * if the underlying column is unique, the subquery may have
5850 : * joined to other tables in a way that creates duplicates.
5851 : */
5852 6176 : examine_simple_variable(subroot, var, vardata);
5853 : }
5854 : }
5855 : else
5856 : {
5857 : /*
5858 : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
5859 : * see RTE_JOIN here because join alias Vars have already been
5860 : * flattened.) There's not much we can do with function outputs, but
5861 : * maybe someday try to be smarter about VALUES.
5862 : */
5863 : }
5864 : }
5865 :
5866 : /*
5867 : * all_rows_selectable
5868 : * Test whether the user has permission to select all rows from a given
5869 : * relation.
5870 : *
5871 : * Inputs:
5872 : * root: the planner info
5873 : * varno: the index of the relation (assumed to be an RTE_RELATION)
5874 : * varattnos: the attributes for which permission is required, or NULL if
5875 : * whole-table access is required
5876 : *
5877 : * Returns true if the user has the required select permissions, and there are
5878 : * no securityQuals from security barrier views or RLS policies.
5879 : *
5880 : * Note that if the relation is an inheritance child relation, securityQuals
5881 : * and access permissions are checked against the inheritance root parent (the
5882 : * relation actually mentioned in the query) --- see the comments in
5883 : * expand_single_inheritance_child() for an explanation of why it has to be
5884 : * done this way.
5885 : *
5886 : * If varattnos is non-NULL, its attribute numbers should be offset by
5887 : * FirstLowInvalidHeapAttributeNumber so that system attributes can be
5888 : * checked. If varattnos is NULL, only table-level SELECT privileges are
5889 : * checked, not any column-level privileges.
5890 : *
5891 : * Note: if the relation is accessed via a view, this function actually tests
5892 : * whether the view owner has permission to select from the relation. To
5893 : * ensure that the current user has permission, it is also necessary to check
5894 : * that the current user has permission to select from the view, which we do
5895 : * at planner-startup --- see subquery_planner().
5896 : *
5897 : * This is exported so that other estimation functions can use it.
5898 : */
5899 : bool
5900 1471548 : all_rows_selectable(PlannerInfo *root, Index varno, Bitmapset *varattnos)
5901 : {
5902 1471548 : RelOptInfo *rel = find_base_rel_noerr(root, varno);
5903 1471548 : RangeTblEntry *rte = planner_rt_fetch(varno, root);
5904 : Oid userid;
5905 : int varattno;
5906 :
5907 : Assert(rte->rtekind == RTE_RELATION);
5908 :
5909 : /*
5910 : * Determine the user ID to use for privilege checks (either the current
5911 : * user or the view owner, if we're accessing the table via a view).
5912 : *
5913 : * Normally the relation will have an associated RelOptInfo from which we
5914 : * can find the userid, but it might not if it's a RETURNING Var for an
5915 : * INSERT target relation. In that case use the RTEPermissionInfo
5916 : * associated with the RTE.
5917 : *
5918 : * If we navigate up to a parent relation, we keep using the same userid,
5919 : * since it's the same in all relations of a given inheritance tree.
5920 : */
5921 1471548 : if (rel)
5922 1471506 : userid = rel->userid;
5923 : else
5924 : {
5925 : RTEPermissionInfo *perminfo;
5926 :
5927 42 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
5928 42 : userid = perminfo->checkAsUser;
5929 : }
5930 1471548 : if (!OidIsValid(userid))
5931 1363224 : userid = GetUserId();
5932 :
5933 : /*
5934 : * Permissions and securityQuals must be checked on the table actually
5935 : * mentioned in the query, so if this is an inheritance child, navigate up
5936 : * to the inheritance root parent. If the user can read the whole table
5937 : * or the required columns there, then they can read from the child table
5938 : * too. For per-column checks, we must find out which of the root
5939 : * parent's attributes the child relation's attributes correspond to.
5940 : */
5941 1471548 : if (root->append_rel_array != NULL)
5942 : {
5943 : AppendRelInfo *appinfo;
5944 :
5945 141468 : appinfo = root->append_rel_array[varno];
5946 :
5947 : /*
5948 : * Partitions are mapped to their immediate parent, not the root
5949 : * parent, so must be ready to walk up multiple AppendRelInfos. But
5950 : * stop if we hit a parent that is not RTE_RELATION --- that's a
5951 : * flattened UNION ALL subquery, not an inheritance parent.
5952 : */
5953 219996 : while (appinfo &&
5954 78900 : planner_rt_fetch(appinfo->parent_relid,
5955 78900 : root)->rtekind == RTE_RELATION)
5956 : {
5957 78528 : Bitmapset *parent_varattnos = NULL;
5958 :
5959 : /*
5960 : * For each child attribute, find the corresponding parent
5961 : * attribute. In rare cases, the attribute may be local to the
5962 : * child table, in which case, we've got to live with having no
5963 : * access to this column.
5964 : */
5965 78528 : varattno = -1;
5966 154206 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
5967 : {
5968 : AttrNumber attno;
5969 : AttrNumber parent_attno;
5970 :
5971 75678 : attno = varattno + FirstLowInvalidHeapAttributeNumber;
5972 :
5973 75678 : if (attno == InvalidAttrNumber)
5974 : {
5975 : /*
5976 : * Whole-row reference, so must map each column of the
5977 : * child to the parent table.
5978 : */
5979 36 : for (attno = 1; attno <= appinfo->num_child_cols; attno++)
5980 : {
5981 24 : parent_attno = appinfo->parent_colnos[attno - 1];
5982 24 : if (parent_attno == 0)
5983 0 : return false; /* attr is local to child */
5984 : parent_varattnos =
5985 24 : bms_add_member(parent_varattnos,
5986 : parent_attno - FirstLowInvalidHeapAttributeNumber);
5987 : }
5988 : }
5989 : else
5990 : {
5991 75666 : if (attno < 0)
5992 : {
5993 : /* System attnos are the same in all tables */
5994 0 : parent_attno = attno;
5995 : }
5996 : else
5997 : {
5998 75666 : if (attno > appinfo->num_child_cols)
5999 0 : return false; /* safety check */
6000 75666 : parent_attno = appinfo->parent_colnos[attno - 1];
6001 75666 : if (parent_attno == 0)
6002 0 : return false; /* attr is local to child */
6003 : }
6004 : parent_varattnos =
6005 75666 : bms_add_member(parent_varattnos,
6006 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6007 : }
6008 : }
6009 :
6010 : /* If the parent is itself a child, continue up */
6011 78528 : varno = appinfo->parent_relid;
6012 78528 : varattnos = parent_varattnos;
6013 78528 : appinfo = root->append_rel_array[varno];
6014 : }
6015 :
6016 : /* Perform the access check on this parent rel */
6017 141468 : rte = planner_rt_fetch(varno, root);
6018 : Assert(rte->rtekind == RTE_RELATION);
6019 : }
6020 :
6021 : /*
6022 : * For all rows to be accessible, there must be no securityQuals from
6023 : * security barrier views or RLS policies.
6024 : */
6025 1471548 : if (rte->securityQuals != NIL)
6026 828 : return false;
6027 :
6028 : /*
6029 : * Test for table-level SELECT privilege.
6030 : *
6031 : * If varattnos is non-NULL, this is sufficient to give access to all
6032 : * requested attributes, even for a child table, since we have verified
6033 : * that all required child columns have matching parent columns.
6034 : *
6035 : * If varattnos is NULL (whole-table access requested), this doesn't
6036 : * necessarily guarantee that the user can read all columns of a child
6037 : * table, but we allow it anyway (see comments in examine_variable()) and
6038 : * don't bother checking any column privileges.
6039 : */
6040 1470720 : if (pg_class_aclcheck(rte->relid, userid, ACL_SELECT) == ACLCHECK_OK)
6041 1470268 : return true;
6042 :
6043 452 : if (varattnos == NULL)
6044 12 : return false; /* whole-table access requested */
6045 :
6046 : /*
6047 : * Don't have table-level SELECT privilege, so check per-column
6048 : * privileges.
6049 : */
6050 440 : varattno = -1;
6051 646 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6052 : {
6053 440 : AttrNumber attno = varattno + FirstLowInvalidHeapAttributeNumber;
6054 :
6055 440 : if (attno == InvalidAttrNumber)
6056 : {
6057 : /* Whole-row reference, so must have access to all columns */
6058 6 : if (pg_attribute_aclcheck_all(rte->relid, userid, ACL_SELECT,
6059 : ACLMASK_ALL) != ACLCHECK_OK)
6060 6 : return false;
6061 : }
6062 : else
6063 : {
6064 434 : if (pg_attribute_aclcheck(rte->relid, attno, userid,
6065 : ACL_SELECT) != ACLCHECK_OK)
6066 228 : return false;
6067 : }
6068 : }
6069 :
6070 : /* If we reach here, have all required column privileges */
6071 206 : return true;
6072 : }
6073 :
6074 : /*
6075 : * examine_indexcol_variable
6076 : * Try to look up statistical data about an index column/expression.
6077 : * Fill in a VariableStatData struct to describe the column.
6078 : *
6079 : * Inputs:
6080 : * root: the planner info
6081 : * index: the index whose column we're interested in
6082 : * indexcol: 0-based index column number (subscripts index->indexkeys[])
6083 : *
6084 : * Outputs: *vardata is filled as follows:
6085 : * var: the input expression (with any binary relabeling stripped, if
6086 : * it is or contains a variable; but otherwise the type is preserved)
6087 : * rel: RelOptInfo for table relation containing variable.
6088 : * statsTuple: the pg_statistic entry for the variable, if one exists;
6089 : * otherwise NULL.
6090 : * freefunc: pointer to a function to release statsTuple with.
6091 : *
6092 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
6093 : */
6094 : static void
6095 740648 : examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
6096 : int indexcol, VariableStatData *vardata)
6097 : {
6098 : AttrNumber colnum;
6099 : Oid relid;
6100 :
6101 740648 : if (index->indexkeys[indexcol] != 0)
6102 : {
6103 : /* Simple variable --- look to stats for the underlying table */
6104 738458 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6105 :
6106 : Assert(rte->rtekind == RTE_RELATION);
6107 738458 : relid = rte->relid;
6108 : Assert(relid != InvalidOid);
6109 738458 : colnum = index->indexkeys[indexcol];
6110 738458 : vardata->rel = index->rel;
6111 :
6112 738458 : if (get_relation_stats_hook &&
6113 0 : (*get_relation_stats_hook) (root, rte, colnum, vardata))
6114 : {
6115 : /*
6116 : * The hook took control of acquiring a stats tuple. If it did
6117 : * supply a tuple, it'd better have supplied a freefunc.
6118 : */
6119 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6120 0 : !vardata->freefunc)
6121 0 : elog(ERROR, "no function provided to release variable stats with");
6122 : }
6123 : else
6124 : {
6125 738458 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6126 : ObjectIdGetDatum(relid),
6127 : Int16GetDatum(colnum),
6128 738458 : BoolGetDatum(rte->inh));
6129 738458 : vardata->freefunc = ReleaseSysCache;
6130 : }
6131 : }
6132 : else
6133 : {
6134 : /* Expression --- maybe there are stats for the index itself */
6135 2190 : relid = index->indexoid;
6136 2190 : colnum = indexcol + 1;
6137 :
6138 2190 : if (get_index_stats_hook &&
6139 0 : (*get_index_stats_hook) (root, relid, colnum, vardata))
6140 : {
6141 : /*
6142 : * The hook took control of acquiring a stats tuple. If it did
6143 : * supply a tuple, it'd better have supplied a freefunc.
6144 : */
6145 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6146 0 : !vardata->freefunc)
6147 0 : elog(ERROR, "no function provided to release variable stats with");
6148 : }
6149 : else
6150 : {
6151 2190 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6152 : ObjectIdGetDatum(relid),
6153 : Int16GetDatum(colnum),
6154 : BoolGetDatum(false));
6155 2190 : vardata->freefunc = ReleaseSysCache;
6156 : }
6157 : }
6158 740648 : }
6159 :
6160 : /*
6161 : * Check whether it is permitted to call func_oid passing some of the
6162 : * pg_statistic data in vardata. We allow this if either of the following
6163 : * conditions is met: (1) the user has SELECT privileges on the table or
6164 : * column underlying the pg_statistic data and there are no securityQuals from
6165 : * security barrier views or RLS policies, or (2) the function is marked
6166 : * leakproof.
6167 : */
6168 : bool
6169 962394 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
6170 : {
6171 962394 : if (vardata->acl_ok)
6172 960616 : return true; /* have SELECT privs and no securityQuals */
6173 :
6174 1778 : if (!OidIsValid(func_oid))
6175 0 : return false;
6176 :
6177 1778 : if (get_func_leakproof(func_oid))
6178 880 : return true;
6179 :
6180 898 : ereport(DEBUG2,
6181 : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
6182 : get_func_name(func_oid))));
6183 898 : return false;
6184 : }
6185 :
6186 : /*
6187 : * get_variable_numdistinct
6188 : * Estimate the number of distinct values of a variable.
6189 : *
6190 : * vardata: results of examine_variable
6191 : * *isdefault: set to true if the result is a default rather than based on
6192 : * anything meaningful.
6193 : *
6194 : * NB: be careful to produce a positive integral result, since callers may
6195 : * compare the result to exact integer counts, or might divide by it.
6196 : */
6197 : double
6198 1437300 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
6199 : {
6200 : double stadistinct;
6201 1437300 : double stanullfrac = 0.0;
6202 : double ntuples;
6203 :
6204 1437300 : *isdefault = false;
6205 :
6206 : /*
6207 : * Determine the stadistinct value to use. There are cases where we can
6208 : * get an estimate even without a pg_statistic entry, or can get a better
6209 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
6210 : * (otherwise, assume no nulls, for lack of any better idea).
6211 : */
6212 1437300 : if (HeapTupleIsValid(vardata->statsTuple))
6213 : {
6214 : /* Use the pg_statistic entry */
6215 : Form_pg_statistic stats;
6216 :
6217 1036048 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6218 1036048 : stadistinct = stats->stadistinct;
6219 1036048 : stanullfrac = stats->stanullfrac;
6220 : }
6221 401252 : else if (vardata->vartype == BOOLOID)
6222 : {
6223 : /*
6224 : * Special-case boolean columns: presumably, two distinct values.
6225 : *
6226 : * Are there any other datatypes we should wire in special estimates
6227 : * for?
6228 : */
6229 592 : stadistinct = 2.0;
6230 : }
6231 400660 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6232 : {
6233 : /*
6234 : * If the Var represents a column of a VALUES RTE, assume it's unique.
6235 : * This could of course be very wrong, but it should tend to be true
6236 : * in well-written queries. We could consider examining the VALUES'
6237 : * contents to get some real statistics; but that only works if the
6238 : * entries are all constants, and it would be pretty expensive anyway.
6239 : */
6240 3466 : stadistinct = -1.0; /* unique (and all non null) */
6241 : }
6242 : else
6243 : {
6244 : /*
6245 : * We don't keep statistics for system columns, but in some cases we
6246 : * can infer distinctness anyway.
6247 : */
6248 397194 : if (vardata->var && IsA(vardata->var, Var))
6249 : {
6250 366436 : switch (((Var *) vardata->var)->varattno)
6251 : {
6252 1186 : case SelfItemPointerAttributeNumber:
6253 1186 : stadistinct = -1.0; /* unique (and all non null) */
6254 1186 : break;
6255 10192 : case TableOidAttributeNumber:
6256 10192 : stadistinct = 1.0; /* only 1 value */
6257 10192 : break;
6258 355058 : default:
6259 355058 : stadistinct = 0.0; /* means "unknown" */
6260 355058 : break;
6261 : }
6262 : }
6263 : else
6264 30758 : stadistinct = 0.0; /* means "unknown" */
6265 :
6266 : /*
6267 : * XXX consider using estimate_num_groups on expressions?
6268 : */
6269 : }
6270 :
6271 : /*
6272 : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6273 : * variable, assume it is unique no matter what pg_statistic says; the
6274 : * statistics could be out of date, or we might have found a partial
6275 : * unique index that proves the var is unique for this query. However,
6276 : * we'd better still believe the null-fraction statistic.
6277 : */
6278 1437300 : if (vardata->isunique)
6279 373056 : stadistinct = -1.0 * (1.0 - stanullfrac);
6280 :
6281 : /*
6282 : * If we had an absolute estimate, use that.
6283 : */
6284 1437300 : if (stadistinct > 0.0)
6285 334770 : return clamp_row_est(stadistinct);
6286 :
6287 : /*
6288 : * Otherwise we need to get the relation size; punt if not available.
6289 : */
6290 1102530 : if (vardata->rel == NULL)
6291 : {
6292 416 : *isdefault = true;
6293 416 : return DEFAULT_NUM_DISTINCT;
6294 : }
6295 1102114 : ntuples = vardata->rel->tuples;
6296 1102114 : if (ntuples <= 0.0)
6297 : {
6298 49864 : *isdefault = true;
6299 49864 : return DEFAULT_NUM_DISTINCT;
6300 : }
6301 :
6302 : /*
6303 : * If we had a relative estimate, use that.
6304 : */
6305 1052250 : if (stadistinct < 0.0)
6306 751458 : return clamp_row_est(-stadistinct * ntuples);
6307 :
6308 : /*
6309 : * With no data, estimate ndistinct = ntuples if the table is small, else
6310 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6311 : * that the behavior isn't discontinuous.
6312 : */
6313 300792 : if (ntuples < DEFAULT_NUM_DISTINCT)
6314 144308 : return clamp_row_est(ntuples);
6315 :
6316 156484 : *isdefault = true;
6317 156484 : return DEFAULT_NUM_DISTINCT;
6318 : }
6319 :
6320 : /*
6321 : * get_variable_range
6322 : * Estimate the minimum and maximum value of the specified variable.
6323 : * If successful, store values in *min and *max, and return true.
6324 : * If no data available, return false.
6325 : *
6326 : * sortop is the "<" comparison operator to use. This should generally
6327 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6328 : * The collation must be specified too.
6329 : */
6330 : static bool
6331 204108 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6332 : Oid sortop, Oid collation,
6333 : Datum *min, Datum *max)
6334 : {
6335 204108 : Datum tmin = 0;
6336 204108 : Datum tmax = 0;
6337 204108 : bool have_data = false;
6338 : int16 typLen;
6339 : bool typByVal;
6340 : Oid opfuncoid;
6341 : FmgrInfo opproc;
6342 : AttStatsSlot sslot;
6343 :
6344 : /*
6345 : * XXX It's very tempting to try to use the actual column min and max, if
6346 : * we can get them relatively-cheaply with an index probe. However, since
6347 : * this function is called many times during join planning, that could
6348 : * have unpleasant effects on planning speed. Need more investigation
6349 : * before enabling this.
6350 : */
6351 : #ifdef NOT_USED
6352 : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6353 : return true;
6354 : #endif
6355 :
6356 204108 : if (!HeapTupleIsValid(vardata->statsTuple))
6357 : {
6358 : /* no stats available, so default result */
6359 44954 : return false;
6360 : }
6361 :
6362 : /*
6363 : * If we can't apply the sortop to the stats data, just fail. In
6364 : * principle, if there's a histogram and no MCVs, we could return the
6365 : * histogram endpoints without ever applying the sortop ... but it's
6366 : * probably not worth trying, because whatever the caller wants to do with
6367 : * the endpoints would likely fail the security check too.
6368 : */
6369 159154 : if (!statistic_proc_security_check(vardata,
6370 159154 : (opfuncoid = get_opcode(sortop))))
6371 0 : return false;
6372 :
6373 159154 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6374 :
6375 159154 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6376 :
6377 : /*
6378 : * If there is a histogram with the ordering we want, grab the first and
6379 : * last values.
6380 : */
6381 159154 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6382 : STATISTIC_KIND_HISTOGRAM, sortop,
6383 : ATTSTATSSLOT_VALUES))
6384 : {
6385 117210 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6386 : {
6387 117210 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6388 117210 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6389 117210 : have_data = true;
6390 : }
6391 117210 : free_attstatsslot(&sslot);
6392 : }
6393 :
6394 : /*
6395 : * Otherwise, if there is a histogram with some other ordering, scan it
6396 : * and get the min and max values according to the ordering we want. This
6397 : * of course may not find values that are really extremal according to our
6398 : * ordering, but it beats ignoring available data.
6399 : */
6400 201098 : if (!have_data &&
6401 41944 : get_attstatsslot(&sslot, vardata->statsTuple,
6402 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6403 : ATTSTATSSLOT_VALUES))
6404 : {
6405 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6406 : collation, typLen, typByVal,
6407 : &tmin, &tmax, &have_data);
6408 0 : free_attstatsslot(&sslot);
6409 : }
6410 :
6411 : /*
6412 : * If we have most-common-values info, look for extreme MCVs. This is
6413 : * needed even if we also have a histogram, since the histogram excludes
6414 : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6415 : * be pretty wary of deciding that that is a full representation of the
6416 : * data. Proceed only if the MCVs represent the whole table (to within
6417 : * roundoff error).
6418 : */
6419 159154 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6420 : STATISTIC_KIND_MCV, InvalidOid,
6421 159154 : have_data ? ATTSTATSSLOT_VALUES :
6422 : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6423 : {
6424 77350 : bool use_mcvs = have_data;
6425 :
6426 77350 : if (!have_data)
6427 : {
6428 40550 : double sumcommon = 0.0;
6429 : double nullfrac;
6430 : int i;
6431 :
6432 302236 : for (i = 0; i < sslot.nnumbers; i++)
6433 261686 : sumcommon += sslot.numbers[i];
6434 40550 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6435 40550 : if (sumcommon + nullfrac > 0.99999)
6436 38540 : use_mcvs = true;
6437 : }
6438 :
6439 77350 : if (use_mcvs)
6440 75340 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6441 : collation, typLen, typByVal,
6442 : &tmin, &tmax, &have_data);
6443 77350 : free_attstatsslot(&sslot);
6444 : }
6445 :
6446 159154 : *min = tmin;
6447 159154 : *max = tmax;
6448 159154 : return have_data;
6449 : }
6450 :
6451 : /*
6452 : * get_stats_slot_range: scan sslot for min/max values
6453 : *
6454 : * Subroutine for get_variable_range: update min/max/have_data according
6455 : * to what we find in the statistics array.
6456 : */
6457 : static void
6458 75340 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6459 : Oid collation, int16 typLen, bool typByVal,
6460 : Datum *min, Datum *max, bool *p_have_data)
6461 : {
6462 75340 : Datum tmin = *min;
6463 75340 : Datum tmax = *max;
6464 75340 : bool have_data = *p_have_data;
6465 75340 : bool found_tmin = false;
6466 75340 : bool found_tmax = false;
6467 :
6468 : /* Look up the comparison function, if we didn't already do so */
6469 75340 : if (opproc->fn_oid != opfuncoid)
6470 75340 : fmgr_info(opfuncoid, opproc);
6471 :
6472 : /* Scan all the slot's values */
6473 2249052 : for (int i = 0; i < sslot->nvalues; i++)
6474 : {
6475 2173712 : if (!have_data)
6476 : {
6477 38540 : tmin = tmax = sslot->values[i];
6478 38540 : found_tmin = found_tmax = true;
6479 38540 : *p_have_data = have_data = true;
6480 38540 : continue;
6481 : }
6482 2135172 : if (DatumGetBool(FunctionCall2Coll(opproc,
6483 : collation,
6484 2135172 : sslot->values[i], tmin)))
6485 : {
6486 52928 : tmin = sslot->values[i];
6487 52928 : found_tmin = true;
6488 : }
6489 2135172 : if (DatumGetBool(FunctionCall2Coll(opproc,
6490 : collation,
6491 2135172 : tmax, sslot->values[i])))
6492 : {
6493 99662 : tmax = sslot->values[i];
6494 99662 : found_tmax = true;
6495 : }
6496 : }
6497 :
6498 : /*
6499 : * Copy the slot's values, if we found new extreme values.
6500 : */
6501 75340 : if (found_tmin)
6502 60758 : *min = datumCopy(tmin, typByVal, typLen);
6503 75340 : if (found_tmax)
6504 42982 : *max = datumCopy(tmax, typByVal, typLen);
6505 75340 : }
6506 :
6507 :
6508 : /*
6509 : * get_actual_variable_range
6510 : * Attempt to identify the current *actual* minimum and/or maximum
6511 : * of the specified variable, by looking for a suitable btree index
6512 : * and fetching its low and/or high values.
6513 : * If successful, store values in *min and *max, and return true.
6514 : * (Either pointer can be NULL if that endpoint isn't needed.)
6515 : * If unsuccessful, return false.
6516 : *
6517 : * sortop is the "<" comparison operator to use.
6518 : * collation is the required collation.
6519 : */
6520 : static bool
6521 180696 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6522 : Oid sortop, Oid collation,
6523 : Datum *min, Datum *max)
6524 : {
6525 180696 : bool have_data = false;
6526 180696 : RelOptInfo *rel = vardata->rel;
6527 : RangeTblEntry *rte;
6528 : ListCell *lc;
6529 :
6530 : /* No hope if no relation or it doesn't have indexes */
6531 180696 : if (rel == NULL || rel->indexlist == NIL)
6532 13452 : return false;
6533 : /* If it has indexes it must be a plain relation */
6534 167244 : rte = root->simple_rte_array[rel->relid];
6535 : Assert(rte->rtekind == RTE_RELATION);
6536 :
6537 : /* ignore partitioned tables. Any indexes here are not real indexes */
6538 167244 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6539 876 : return false;
6540 :
6541 : /* Search through the indexes to see if any match our problem */
6542 324428 : foreach(lc, rel->indexlist)
6543 : {
6544 277202 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6545 : ScanDirection indexscandir;
6546 : StrategyNumber strategy;
6547 :
6548 : /* Ignore non-ordering indexes */
6549 277202 : if (index->sortopfamily == NULL)
6550 0 : continue;
6551 :
6552 : /*
6553 : * Ignore partial indexes --- we only want stats that cover the entire
6554 : * relation.
6555 : */
6556 277202 : if (index->indpred != NIL)
6557 288 : continue;
6558 :
6559 : /*
6560 : * The index list might include hypothetical indexes inserted by a
6561 : * get_relation_info hook --- don't try to access them.
6562 : */
6563 276914 : if (index->hypothetical)
6564 0 : continue;
6565 :
6566 : /*
6567 : * The first index column must match the desired variable, sortop, and
6568 : * collation --- but we can use a descending-order index.
6569 : */
6570 276914 : if (collation != index->indexcollations[0])
6571 35866 : continue; /* test first 'cause it's cheapest */
6572 241048 : if (!match_index_to_operand(vardata->var, 0, index))
6573 121906 : continue;
6574 119142 : strategy = get_op_opfamily_strategy(sortop, index->sortopfamily[0]);
6575 119142 : switch (IndexAmTranslateStrategy(strategy, index->relam, index->sortopfamily[0], true))
6576 : {
6577 119142 : case COMPARE_LT:
6578 119142 : if (index->reverse_sort[0])
6579 0 : indexscandir = BackwardScanDirection;
6580 : else
6581 119142 : indexscandir = ForwardScanDirection;
6582 119142 : break;
6583 0 : case COMPARE_GT:
6584 0 : if (index->reverse_sort[0])
6585 0 : indexscandir = ForwardScanDirection;
6586 : else
6587 0 : indexscandir = BackwardScanDirection;
6588 0 : break;
6589 0 : default:
6590 : /* index doesn't match the sortop */
6591 0 : continue;
6592 : }
6593 :
6594 : /*
6595 : * Found a suitable index to extract data from. Set up some data that
6596 : * can be used by both invocations of get_actual_variable_endpoint.
6597 : */
6598 : {
6599 : MemoryContext tmpcontext;
6600 : MemoryContext oldcontext;
6601 : Relation heapRel;
6602 : Relation indexRel;
6603 : TupleTableSlot *slot;
6604 : int16 typLen;
6605 : bool typByVal;
6606 : ScanKeyData scankeys[1];
6607 :
6608 : /* Make sure any cruft gets recycled when we're done */
6609 119142 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
6610 : "get_actual_variable_range workspace",
6611 : ALLOCSET_DEFAULT_SIZES);
6612 119142 : oldcontext = MemoryContextSwitchTo(tmpcontext);
6613 :
6614 : /*
6615 : * Open the table and index so we can read from them. We should
6616 : * already have some type of lock on each.
6617 : */
6618 119142 : heapRel = table_open(rte->relid, NoLock);
6619 119142 : indexRel = index_open(index->indexoid, NoLock);
6620 :
6621 : /* build some stuff needed for indexscan execution */
6622 119142 : slot = table_slot_create(heapRel, NULL);
6623 119142 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6624 :
6625 : /* set up an IS NOT NULL scan key so that we ignore nulls */
6626 119142 : ScanKeyEntryInitialize(&scankeys[0],
6627 : SK_ISNULL | SK_SEARCHNOTNULL,
6628 : 1, /* index col to scan */
6629 : InvalidStrategy, /* no strategy */
6630 : InvalidOid, /* no strategy subtype */
6631 : InvalidOid, /* no collation */
6632 : InvalidOid, /* no reg proc for this */
6633 : (Datum) 0); /* constant */
6634 :
6635 : /* If min is requested ... */
6636 119142 : if (min)
6637 : {
6638 67878 : have_data = get_actual_variable_endpoint(heapRel,
6639 : indexRel,
6640 : indexscandir,
6641 : scankeys,
6642 : typLen,
6643 : typByVal,
6644 : slot,
6645 : oldcontext,
6646 : min);
6647 : }
6648 : else
6649 : {
6650 : /* If min not requested, still want to fetch max */
6651 51264 : have_data = true;
6652 : }
6653 :
6654 : /* If max is requested, and we didn't already fail ... */
6655 119142 : if (max && have_data)
6656 : {
6657 : /* scan in the opposite direction; all else is the same */
6658 52904 : have_data = get_actual_variable_endpoint(heapRel,
6659 : indexRel,
6660 52904 : -indexscandir,
6661 : scankeys,
6662 : typLen,
6663 : typByVal,
6664 : slot,
6665 : oldcontext,
6666 : max);
6667 : }
6668 :
6669 : /* Clean everything up */
6670 119142 : ExecDropSingleTupleTableSlot(slot);
6671 :
6672 119142 : index_close(indexRel, NoLock);
6673 119142 : table_close(heapRel, NoLock);
6674 :
6675 119142 : MemoryContextSwitchTo(oldcontext);
6676 119142 : MemoryContextDelete(tmpcontext);
6677 :
6678 : /* And we're done */
6679 119142 : break;
6680 : }
6681 : }
6682 :
6683 166368 : return have_data;
6684 : }
6685 :
6686 : /*
6687 : * Get one endpoint datum (min or max depending on indexscandir) from the
6688 : * specified index. Return true if successful, false if not.
6689 : * On success, endpoint value is stored to *endpointDatum (and copied into
6690 : * outercontext).
6691 : *
6692 : * scankeys is a 1-element scankey array set up to reject nulls.
6693 : * typLen/typByVal describe the datatype of the index's first column.
6694 : * tableslot is a slot suitable to hold table tuples, in case we need
6695 : * to probe the heap.
6696 : * (We could compute these values locally, but that would mean computing them
6697 : * twice when get_actual_variable_range needs both the min and the max.)
6698 : *
6699 : * Failure occurs either when the index is empty, or we decide that it's
6700 : * taking too long to find a suitable tuple.
6701 : */
6702 : static bool
6703 120782 : get_actual_variable_endpoint(Relation heapRel,
6704 : Relation indexRel,
6705 : ScanDirection indexscandir,
6706 : ScanKey scankeys,
6707 : int16 typLen,
6708 : bool typByVal,
6709 : TupleTableSlot *tableslot,
6710 : MemoryContext outercontext,
6711 : Datum *endpointDatum)
6712 : {
6713 120782 : bool have_data = false;
6714 : SnapshotData SnapshotNonVacuumable;
6715 : IndexScanDesc index_scan;
6716 120782 : Buffer vmbuffer = InvalidBuffer;
6717 120782 : BlockNumber last_heap_block = InvalidBlockNumber;
6718 120782 : int n_visited_heap_pages = 0;
6719 : ItemPointer tid;
6720 : Datum values[INDEX_MAX_KEYS];
6721 : bool isnull[INDEX_MAX_KEYS];
6722 : MemoryContext oldcontext;
6723 :
6724 : /*
6725 : * We use the index-only-scan machinery for this. With mostly-static
6726 : * tables that's a win because it avoids a heap visit. It's also a win
6727 : * for dynamic data, but the reason is less obvious; read on for details.
6728 : *
6729 : * In principle, we should scan the index with our current active
6730 : * snapshot, which is the best approximation we've got to what the query
6731 : * will see when executed. But that won't be exact if a new snap is taken
6732 : * before running the query, and it can be very expensive if a lot of
6733 : * recently-dead or uncommitted rows exist at the beginning or end of the
6734 : * index (because we'll laboriously fetch each one and reject it).
6735 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
6736 : * and uncommitted rows as well as normal visible rows. On the other
6737 : * hand, it will reject known-dead rows, and thus not give a bogus answer
6738 : * when the extreme value has been deleted (unless the deletion was quite
6739 : * recent); that case motivates not using SnapshotAny here.
6740 : *
6741 : * A crucial point here is that SnapshotNonVacuumable, with
6742 : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
6743 : * condition that the indexscan will use to decide that index entries are
6744 : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
6745 : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
6746 : * have to continue scanning past it, we know that the indexscan will mark
6747 : * that index entry killed. That means that the next
6748 : * get_actual_variable_endpoint() call will not have to re-consider that
6749 : * index entry. In this way we avoid repetitive work when this function
6750 : * is used a lot during planning.
6751 : *
6752 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
6753 : * recently-created index, some index entries may point at "broken" HOT
6754 : * chains in which not all the tuple versions contain data matching the
6755 : * index entry. The live tuple version(s) certainly do match the index,
6756 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
6757 : * don't match. Hence, if we took data from the selected heap tuple, we
6758 : * might get a bogus answer that's not close to the index extremal value,
6759 : * or could even be NULL. We avoid this hazard because we take the data
6760 : * from the index entry not the heap.
6761 : *
6762 : * Despite all this care, there are situations where we might find many
6763 : * non-visible tuples near the end of the index. We don't want to expend
6764 : * a huge amount of time here, so we give up once we've read too many heap
6765 : * pages. When we fail for that reason, the caller will end up using
6766 : * whatever extremal value is recorded in pg_statistic.
6767 : */
6768 120782 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
6769 : GlobalVisTestFor(heapRel));
6770 :
6771 120782 : index_scan = index_beginscan(heapRel, indexRel,
6772 : &SnapshotNonVacuumable, NULL,
6773 : 1, 0);
6774 : /* Set it up for index-only scan */
6775 120782 : index_scan->xs_want_itup = true;
6776 120782 : index_rescan(index_scan, scankeys, 1, NULL, 0);
6777 :
6778 : /* Fetch first/next tuple in specified direction */
6779 154164 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
6780 : {
6781 154164 : BlockNumber block = ItemPointerGetBlockNumber(tid);
6782 :
6783 154164 : if (!VM_ALL_VISIBLE(heapRel,
6784 : block,
6785 : &vmbuffer))
6786 : {
6787 : /* Rats, we have to visit the heap to check visibility */
6788 112882 : if (!index_fetch_heap(index_scan, tableslot))
6789 : {
6790 : /*
6791 : * No visible tuple for this index entry, so we need to
6792 : * advance to the next entry. Before doing so, count heap
6793 : * page fetches and give up if we've done too many.
6794 : *
6795 : * We don't charge a page fetch if this is the same heap page
6796 : * as the previous tuple. This is on the conservative side,
6797 : * since other recently-accessed pages are probably still in
6798 : * buffers too; but it's good enough for this heuristic.
6799 : */
6800 : #define VISITED_PAGES_LIMIT 100
6801 :
6802 33382 : if (block != last_heap_block)
6803 : {
6804 3472 : last_heap_block = block;
6805 3472 : n_visited_heap_pages++;
6806 3472 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
6807 0 : break;
6808 : }
6809 :
6810 33382 : continue; /* no visible tuple, try next index entry */
6811 : }
6812 :
6813 : /* We don't actually need the heap tuple for anything */
6814 79500 : ExecClearTuple(tableslot);
6815 :
6816 : /*
6817 : * We don't care whether there's more than one visible tuple in
6818 : * the HOT chain; if any are visible, that's good enough.
6819 : */
6820 : }
6821 :
6822 : /*
6823 : * We expect that the index will return data in IndexTuple not
6824 : * HeapTuple format.
6825 : */
6826 120782 : if (!index_scan->xs_itup)
6827 0 : elog(ERROR, "no data returned for index-only scan");
6828 :
6829 : /*
6830 : * We do not yet support recheck here.
6831 : */
6832 120782 : if (index_scan->xs_recheck)
6833 0 : break;
6834 :
6835 : /* OK to deconstruct the index tuple */
6836 120782 : index_deform_tuple(index_scan->xs_itup,
6837 : index_scan->xs_itupdesc,
6838 : values, isnull);
6839 :
6840 : /* Shouldn't have got a null, but be careful */
6841 120782 : if (isnull[0])
6842 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
6843 : RelationGetRelationName(indexRel));
6844 :
6845 : /* Copy the index column value out to caller's context */
6846 120782 : oldcontext = MemoryContextSwitchTo(outercontext);
6847 120782 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
6848 120782 : MemoryContextSwitchTo(oldcontext);
6849 120782 : have_data = true;
6850 120782 : break;
6851 : }
6852 :
6853 120782 : if (vmbuffer != InvalidBuffer)
6854 107734 : ReleaseBuffer(vmbuffer);
6855 120782 : index_endscan(index_scan);
6856 :
6857 120782 : return have_data;
6858 : }
6859 :
6860 : /*
6861 : * find_join_input_rel
6862 : * Look up the input relation for a join.
6863 : *
6864 : * We assume that the input relation's RelOptInfo must have been constructed
6865 : * already.
6866 : */
6867 : static RelOptInfo *
6868 10690 : find_join_input_rel(PlannerInfo *root, Relids relids)
6869 : {
6870 10690 : RelOptInfo *rel = NULL;
6871 :
6872 10690 : if (!bms_is_empty(relids))
6873 : {
6874 : int relid;
6875 :
6876 10690 : if (bms_get_singleton_member(relids, &relid))
6877 10386 : rel = find_base_rel(root, relid);
6878 : else
6879 304 : rel = find_join_rel(root, relids);
6880 : }
6881 :
6882 10690 : if (rel == NULL)
6883 0 : elog(ERROR, "could not find RelOptInfo for given relids");
6884 :
6885 10690 : return rel;
6886 : }
6887 :
6888 :
6889 : /*-------------------------------------------------------------------------
6890 : *
6891 : * Index cost estimation functions
6892 : *
6893 : *-------------------------------------------------------------------------
6894 : */
6895 :
6896 : /*
6897 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
6898 : */
6899 : List *
6900 758182 : get_quals_from_indexclauses(List *indexclauses)
6901 : {
6902 758182 : List *result = NIL;
6903 : ListCell *lc;
6904 :
6905 1335150 : foreach(lc, indexclauses)
6906 : {
6907 576968 : IndexClause *iclause = lfirst_node(IndexClause, lc);
6908 : ListCell *lc2;
6909 :
6910 1156848 : foreach(lc2, iclause->indexquals)
6911 : {
6912 579880 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6913 :
6914 579880 : result = lappend(result, rinfo);
6915 : }
6916 : }
6917 758182 : return result;
6918 : }
6919 :
6920 : /*
6921 : * Compute the total evaluation cost of the comparison operands in a list
6922 : * of index qual expressions. Since we know these will be evaluated just
6923 : * once per scan, there's no need to distinguish startup from per-row cost.
6924 : *
6925 : * This can be used either on the result of get_quals_from_indexclauses(),
6926 : * or directly on an indexorderbys list. In both cases, we expect that the
6927 : * index key expression is on the left side of binary clauses.
6928 : */
6929 : Cost
6930 1503370 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
6931 : {
6932 1503370 : Cost qual_arg_cost = 0;
6933 : ListCell *lc;
6934 :
6935 2083712 : foreach(lc, indexquals)
6936 : {
6937 580342 : Expr *clause = (Expr *) lfirst(lc);
6938 : Node *other_operand;
6939 : QualCost index_qual_cost;
6940 :
6941 : /*
6942 : * Index quals will have RestrictInfos, indexorderbys won't. Look
6943 : * through RestrictInfo if present.
6944 : */
6945 580342 : if (IsA(clause, RestrictInfo))
6946 579868 : clause = ((RestrictInfo *) clause)->clause;
6947 :
6948 580342 : if (IsA(clause, OpExpr))
6949 : {
6950 566180 : OpExpr *op = (OpExpr *) clause;
6951 :
6952 566180 : other_operand = (Node *) lsecond(op->args);
6953 : }
6954 14162 : else if (IsA(clause, RowCompareExpr))
6955 : {
6956 396 : RowCompareExpr *rc = (RowCompareExpr *) clause;
6957 :
6958 396 : other_operand = (Node *) rc->rargs;
6959 : }
6960 13766 : else if (IsA(clause, ScalarArrayOpExpr))
6961 : {
6962 10840 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6963 :
6964 10840 : other_operand = (Node *) lsecond(saop->args);
6965 : }
6966 2926 : else if (IsA(clause, NullTest))
6967 : {
6968 2926 : other_operand = NULL;
6969 : }
6970 : else
6971 : {
6972 0 : elog(ERROR, "unsupported indexqual type: %d",
6973 : (int) nodeTag(clause));
6974 : other_operand = NULL; /* keep compiler quiet */
6975 : }
6976 :
6977 580342 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
6978 580342 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
6979 : }
6980 1503370 : return qual_arg_cost;
6981 : }
6982 :
6983 : void
6984 745200 : genericcostestimate(PlannerInfo *root,
6985 : IndexPath *path,
6986 : double loop_count,
6987 : GenericCosts *costs)
6988 : {
6989 745200 : IndexOptInfo *index = path->indexinfo;
6990 745200 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
6991 745200 : List *indexOrderBys = path->indexorderbys;
6992 : Cost indexStartupCost;
6993 : Cost indexTotalCost;
6994 : Selectivity indexSelectivity;
6995 : double indexCorrelation;
6996 : double numIndexPages;
6997 : double numIndexTuples;
6998 : double spc_random_page_cost;
6999 : double num_sa_scans;
7000 : double num_outer_scans;
7001 : double num_scans;
7002 : double qual_op_cost;
7003 : double qual_arg_cost;
7004 : List *selectivityQuals;
7005 : ListCell *l;
7006 :
7007 : /*
7008 : * If the index is partial, AND the index predicate with the explicitly
7009 : * given indexquals to produce a more accurate idea of the index
7010 : * selectivity.
7011 : */
7012 745200 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7013 :
7014 : /*
7015 : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
7016 : * just assume that the number of index descents is the number of distinct
7017 : * combinations of array elements from all of the scan's SAOP clauses.
7018 : */
7019 745200 : num_sa_scans = costs->num_sa_scans;
7020 745200 : if (num_sa_scans < 1)
7021 : {
7022 7804 : num_sa_scans = 1;
7023 16330 : foreach(l, indexQuals)
7024 : {
7025 8526 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
7026 :
7027 8526 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
7028 : {
7029 26 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
7030 26 : double alength = estimate_array_length(root, lsecond(saop->args));
7031 :
7032 26 : if (alength > 1)
7033 26 : num_sa_scans *= alength;
7034 : }
7035 : }
7036 : }
7037 :
7038 : /* Estimate the fraction of main-table tuples that will be visited */
7039 745200 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7040 745200 : index->rel->relid,
7041 : JOIN_INNER,
7042 : NULL);
7043 :
7044 : /*
7045 : * If caller didn't give us an estimate, estimate the number of index
7046 : * tuples that will be visited. We do it in this rather peculiar-looking
7047 : * way in order to get the right answer for partial indexes.
7048 : */
7049 745200 : numIndexTuples = costs->numIndexTuples;
7050 745200 : if (numIndexTuples <= 0.0)
7051 : {
7052 75480 : numIndexTuples = indexSelectivity * index->rel->tuples;
7053 :
7054 : /*
7055 : * The above calculation counts all the tuples visited across all
7056 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
7057 : * average per-indexscan number, so adjust. This is a handy place to
7058 : * round to integer, too. (If caller supplied tuple estimate, it's
7059 : * responsible for handling these considerations.)
7060 : */
7061 75480 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7062 : }
7063 :
7064 : /*
7065 : * We can bound the number of tuples by the index size in any case. Also,
7066 : * always estimate at least one tuple is touched, even when
7067 : * indexSelectivity estimate is tiny.
7068 : */
7069 745200 : if (numIndexTuples > index->tuples)
7070 5424 : numIndexTuples = index->tuples;
7071 745200 : if (numIndexTuples < 1.0)
7072 75088 : numIndexTuples = 1.0;
7073 :
7074 : /*
7075 : * Estimate the number of index pages that will be retrieved.
7076 : *
7077 : * We use the simplistic method of taking a pro-rata fraction of the total
7078 : * number of index pages. In effect, this counts only leaf pages and not
7079 : * any overhead such as index metapage or upper tree levels.
7080 : *
7081 : * In practice access to upper index levels is often nearly free because
7082 : * those tend to stay in cache under load; moreover, the cost involved is
7083 : * highly dependent on index type. We therefore ignore such costs here
7084 : * and leave it to the caller to add a suitable charge if needed.
7085 : */
7086 745200 : if (index->pages > 1 && index->tuples > 1)
7087 699214 : numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
7088 : else
7089 45986 : numIndexPages = 1.0;
7090 :
7091 : /* fetch estimated page cost for tablespace containing index */
7092 745200 : get_tablespace_page_costs(index->reltablespace,
7093 : &spc_random_page_cost,
7094 : NULL);
7095 :
7096 : /*
7097 : * Now compute the disk access costs.
7098 : *
7099 : * The above calculations are all per-index-scan. However, if we are in a
7100 : * nestloop inner scan, we can expect the scan to be repeated (with
7101 : * different search keys) for each row of the outer relation. Likewise,
7102 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
7103 : * the potential for cache effects to reduce the number of disk page
7104 : * fetches needed. We want to estimate the average per-scan I/O cost in
7105 : * the presence of caching.
7106 : *
7107 : * We use the Mackert-Lohman formula (see costsize.c for details) to
7108 : * estimate the total number of page fetches that occur. While this
7109 : * wasn't what it was designed for, it seems a reasonable model anyway.
7110 : * Note that we are counting pages not tuples anymore, so we take N = T =
7111 : * index size, as if there were one "tuple" per page.
7112 : */
7113 745200 : num_outer_scans = loop_count;
7114 745200 : num_scans = num_sa_scans * num_outer_scans;
7115 :
7116 745200 : if (num_scans > 1)
7117 : {
7118 : double pages_fetched;
7119 :
7120 : /* total page fetches ignoring cache effects */
7121 83972 : pages_fetched = numIndexPages * num_scans;
7122 :
7123 : /* use Mackert and Lohman formula to adjust for cache effects */
7124 83972 : pages_fetched = index_pages_fetched(pages_fetched,
7125 : index->pages,
7126 83972 : (double) index->pages,
7127 : root);
7128 :
7129 : /*
7130 : * Now compute the total disk access cost, and then report a pro-rated
7131 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
7132 : * since that's internal to the indexscan.)
7133 : */
7134 83972 : indexTotalCost = (pages_fetched * spc_random_page_cost)
7135 : / num_outer_scans;
7136 : }
7137 : else
7138 : {
7139 : /*
7140 : * For a single index scan, we just charge spc_random_page_cost per
7141 : * page touched.
7142 : */
7143 661228 : indexTotalCost = numIndexPages * spc_random_page_cost;
7144 : }
7145 :
7146 : /*
7147 : * CPU cost: any complex expressions in the indexquals will need to be
7148 : * evaluated once at the start of the scan to reduce them to runtime keys
7149 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
7150 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
7151 : * indexqual operator. Because we have numIndexTuples as a per-scan
7152 : * number, we have to multiply by num_sa_scans to get the correct result
7153 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
7154 : * ORDER BY expressions.
7155 : *
7156 : * Note: this neglects the possible costs of rechecking lossy operators.
7157 : * Detecting that that might be needed seems more expensive than it's
7158 : * worth, though, considering all the other inaccuracies here ...
7159 : */
7160 745200 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
7161 745200 : index_other_operands_eval_cost(root, indexOrderBys);
7162 745200 : qual_op_cost = cpu_operator_cost *
7163 745200 : (list_length(indexQuals) + list_length(indexOrderBys));
7164 :
7165 745200 : indexStartupCost = qual_arg_cost;
7166 745200 : indexTotalCost += qual_arg_cost;
7167 745200 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
7168 :
7169 : /*
7170 : * Generic assumption about index correlation: there isn't any.
7171 : */
7172 745200 : indexCorrelation = 0.0;
7173 :
7174 : /*
7175 : * Return everything to caller.
7176 : */
7177 745200 : costs->indexStartupCost = indexStartupCost;
7178 745200 : costs->indexTotalCost = indexTotalCost;
7179 745200 : costs->indexSelectivity = indexSelectivity;
7180 745200 : costs->indexCorrelation = indexCorrelation;
7181 745200 : costs->numIndexPages = numIndexPages;
7182 745200 : costs->numIndexTuples = numIndexTuples;
7183 745200 : costs->spc_random_page_cost = spc_random_page_cost;
7184 745200 : costs->num_sa_scans = num_sa_scans;
7185 745200 : }
7186 :
7187 : /*
7188 : * If the index is partial, add its predicate to the given qual list.
7189 : *
7190 : * ANDing the index predicate with the explicitly given indexquals produces
7191 : * a more accurate idea of the index's selectivity. However, we need to be
7192 : * careful not to insert redundant clauses, because clauselist_selectivity()
7193 : * is easily fooled into computing a too-low selectivity estimate. Our
7194 : * approach is to add only the predicate clause(s) that cannot be proven to
7195 : * be implied by the given indexquals. This successfully handles cases such
7196 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
7197 : * There are many other cases where we won't detect redundancy, leading to a
7198 : * too-low selectivity estimate, which will bias the system in favor of using
7199 : * partial indexes where possible. That is not necessarily bad though.
7200 : *
7201 : * Note that indexQuals contains RestrictInfo nodes while the indpred
7202 : * does not, so the output list will be mixed. This is OK for both
7203 : * predicate_implied_by() and clauselist_selectivity(), but might be
7204 : * problematic if the result were passed to other things.
7205 : */
7206 : List *
7207 1251412 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
7208 : {
7209 1251412 : List *predExtraQuals = NIL;
7210 : ListCell *lc;
7211 :
7212 1251412 : if (index->indpred == NIL)
7213 1249386 : return indexQuals;
7214 :
7215 4064 : foreach(lc, index->indpred)
7216 : {
7217 2038 : Node *predQual = (Node *) lfirst(lc);
7218 2038 : List *oneQual = list_make1(predQual);
7219 :
7220 2038 : if (!predicate_implied_by(oneQual, indexQuals, false))
7221 1836 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7222 : }
7223 2026 : return list_concat(predExtraQuals, indexQuals);
7224 : }
7225 :
7226 : /*
7227 : * Estimate correlation of btree index's first column.
7228 : *
7229 : * If we can get an estimate of the first column's ordering correlation C
7230 : * from pg_statistic, estimate the index correlation as C for a single-column
7231 : * index, or C * 0.75 for multiple columns. The idea here is that multiple
7232 : * columns dilute the importance of the first column's ordering, but don't
7233 : * negate it entirely.
7234 : *
7235 : * We already filled in the stats tuple for *vardata when called.
7236 : */
7237 : static double
7238 562522 : btcost_correlation(IndexOptInfo *index, VariableStatData *vardata)
7239 : {
7240 : Oid sortop;
7241 : AttStatsSlot sslot;
7242 562522 : double indexCorrelation = 0;
7243 :
7244 : Assert(HeapTupleIsValid(vardata->statsTuple));
7245 :
7246 562522 : sortop = get_opfamily_member(index->opfamily[0],
7247 562522 : index->opcintype[0],
7248 562522 : index->opcintype[0],
7249 : BTLessStrategyNumber);
7250 1125044 : if (OidIsValid(sortop) &&
7251 562522 : get_attstatsslot(&sslot, vardata->statsTuple,
7252 : STATISTIC_KIND_CORRELATION, sortop,
7253 : ATTSTATSSLOT_NUMBERS))
7254 : {
7255 : double varCorrelation;
7256 :
7257 : Assert(sslot.nnumbers == 1);
7258 555726 : varCorrelation = sslot.numbers[0];
7259 :
7260 555726 : if (index->reverse_sort[0])
7261 0 : varCorrelation = -varCorrelation;
7262 :
7263 555726 : if (index->nkeycolumns > 1)
7264 195524 : indexCorrelation = varCorrelation * 0.75;
7265 : else
7266 360202 : indexCorrelation = varCorrelation;
7267 :
7268 555726 : free_attstatsslot(&sslot);
7269 : }
7270 :
7271 562522 : return indexCorrelation;
7272 : }
7273 :
7274 : void
7275 737396 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7276 : Cost *indexStartupCost, Cost *indexTotalCost,
7277 : Selectivity *indexSelectivity, double *indexCorrelation,
7278 : double *indexPages)
7279 : {
7280 737396 : IndexOptInfo *index = path->indexinfo;
7281 737396 : GenericCosts costs = {0};
7282 737396 : VariableStatData vardata = {0};
7283 : double numIndexTuples;
7284 : Cost descentCost;
7285 : List *indexBoundQuals;
7286 : List *indexSkipQuals;
7287 : int indexcol;
7288 : bool eqQualHere;
7289 : bool found_row_compare;
7290 : bool found_array;
7291 : bool found_is_null_op;
7292 737396 : bool have_correlation = false;
7293 : double num_sa_scans;
7294 737396 : double correlation = 0.0;
7295 : ListCell *lc;
7296 :
7297 : /*
7298 : * For a btree scan, only leading '=' quals plus inequality quals for the
7299 : * immediately next attribute contribute to index selectivity (these are
7300 : * the "boundary quals" that determine the starting and stopping points of
7301 : * the index scan). Additional quals can suppress visits to the heap, so
7302 : * it's OK to count them in indexSelectivity, but they should not count
7303 : * for estimating numIndexTuples. So we must examine the given indexquals
7304 : * to find out which ones count as boundary quals. We rely on the
7305 : * knowledge that they are given in index column order. Note that nbtree
7306 : * preprocessing can add skip arrays that act as leading '=' quals in the
7307 : * absence of ordinary input '=' quals, so in practice _most_ input quals
7308 : * are able to act as index bound quals (which we take into account here).
7309 : *
7310 : * For a RowCompareExpr, we consider only the first column, just as
7311 : * rowcomparesel() does.
7312 : *
7313 : * If there's a SAOP or skip array in the quals, we'll actually perform up
7314 : * to N index descents (not just one), but the underlying array key's
7315 : * operator can be considered to act the same as it normally does.
7316 : */
7317 737396 : indexBoundQuals = NIL;
7318 737396 : indexSkipQuals = NIL;
7319 737396 : indexcol = 0;
7320 737396 : eqQualHere = false;
7321 737396 : found_row_compare = false;
7322 737396 : found_array = false;
7323 737396 : found_is_null_op = false;
7324 737396 : num_sa_scans = 1;
7325 1261946 : foreach(lc, path->indexclauses)
7326 : {
7327 552980 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7328 : ListCell *lc2;
7329 :
7330 552980 : if (indexcol < iclause->indexcol)
7331 : {
7332 98284 : double num_sa_scans_prev_cols = num_sa_scans;
7333 :
7334 : /*
7335 : * Beginning of a new column's quals.
7336 : *
7337 : * Skip scans use skip arrays, which are ScalarArrayOp style
7338 : * arrays that generate their elements procedurally and on demand.
7339 : * Given a multi-column index on "(a, b)", and an SQL WHERE clause
7340 : * "WHERE b = 42", a skip scan will effectively use an indexqual
7341 : * "WHERE a = ANY('{every col a value}') AND b = 42". (Obviously,
7342 : * the array on "a" must also return "IS NULL" matches, since our
7343 : * WHERE clause used no strict operator on "a").
7344 : *
7345 : * Here we consider how nbtree will backfill skip arrays for any
7346 : * index columns that lacked an '=' qual. This maintains our
7347 : * num_sa_scans estimate, and determines if this new column (the
7348 : * "iclause->indexcol" column, not the prior "indexcol" column)
7349 : * can have its RestrictInfos/quals added to indexBoundQuals.
7350 : *
7351 : * We'll need to handle columns that have inequality quals, where
7352 : * the skip array generates values from a range constrained by the
7353 : * quals (not every possible value). We've been maintaining
7354 : * indexSkipQuals to help with this; it will now contain all of
7355 : * the prior column's quals (that is, indexcol's quals) when they
7356 : * might be used for this.
7357 : */
7358 98284 : if (found_row_compare)
7359 : {
7360 : /*
7361 : * Skip arrays can't be added after a RowCompare input qual
7362 : * due to limitations in nbtree
7363 : */
7364 24 : break;
7365 : }
7366 98260 : if (eqQualHere)
7367 : {
7368 : /*
7369 : * Don't need to add a skip array for an indexcol that already
7370 : * has an '=' qual/equality constraint
7371 : */
7372 70006 : indexcol++;
7373 70006 : indexSkipQuals = NIL;
7374 : }
7375 98260 : eqQualHere = false;
7376 :
7377 101348 : while (indexcol < iclause->indexcol)
7378 : {
7379 : double ndistinct;
7380 31494 : bool isdefault = true;
7381 :
7382 31494 : found_array = true;
7383 :
7384 : /*
7385 : * A skipped attribute's ndistinct forms the basis of our
7386 : * estimate of the total number of "array elements" used by
7387 : * its skip array at runtime. Look that up first.
7388 : */
7389 31494 : examine_indexcol_variable(root, index, indexcol, &vardata);
7390 31494 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
7391 :
7392 31494 : if (indexcol == 0)
7393 : {
7394 : /*
7395 : * Get an estimate of the leading column's correlation in
7396 : * passing (avoids rereading variable stats below)
7397 : */
7398 28242 : if (HeapTupleIsValid(vardata.statsTuple))
7399 21698 : correlation = btcost_correlation(index, &vardata);
7400 28242 : have_correlation = true;
7401 : }
7402 :
7403 31494 : ReleaseVariableStats(vardata);
7404 :
7405 : /*
7406 : * If ndistinct is a default estimate, conservatively assume
7407 : * that no skipping will happen at runtime
7408 : */
7409 31494 : if (isdefault)
7410 : {
7411 5050 : num_sa_scans = num_sa_scans_prev_cols;
7412 28406 : break; /* done building indexBoundQuals */
7413 : }
7414 :
7415 : /*
7416 : * Apply indexcol's indexSkipQuals selectivity to ndistinct
7417 : */
7418 26444 : if (indexSkipQuals != NIL)
7419 : {
7420 : List *partialSkipQuals;
7421 : Selectivity ndistinctfrac;
7422 :
7423 : /*
7424 : * If the index is partial, AND the index predicate with
7425 : * the index-bound quals to produce a more accurate idea
7426 : * of the number of distinct values for prior indexcol
7427 : */
7428 664 : partialSkipQuals = add_predicate_to_index_quals(index,
7429 : indexSkipQuals);
7430 :
7431 664 : ndistinctfrac = clauselist_selectivity(root, partialSkipQuals,
7432 664 : index->rel->relid,
7433 : JOIN_INNER,
7434 : NULL);
7435 :
7436 : /*
7437 : * If ndistinctfrac is selective (on its own), the scan is
7438 : * unlikely to benefit from repositioning itself using
7439 : * later quals. Do not allow iclause->indexcol's quals to
7440 : * be added to indexBoundQuals (it would increase descent
7441 : * costs, without lowering numIndexTuples costs by much).
7442 : */
7443 664 : if (ndistinctfrac < DEFAULT_RANGE_INEQ_SEL)
7444 : {
7445 374 : num_sa_scans = num_sa_scans_prev_cols;
7446 374 : break; /* done building indexBoundQuals */
7447 : }
7448 :
7449 : /* Adjust ndistinct downward */
7450 290 : ndistinct = rint(ndistinct * ndistinctfrac);
7451 290 : ndistinct = Max(ndistinct, 1);
7452 : }
7453 :
7454 : /*
7455 : * When there's no inequality quals, account for the need to
7456 : * find an initial value by counting -inf/+inf as a value.
7457 : *
7458 : * We don't charge anything extra for possible next/prior key
7459 : * index probes, which are sometimes used to find the next
7460 : * valid skip array element (ahead of using the located
7461 : * element value to relocate the scan to the next position
7462 : * that might contain matching tuples). It seems hard to do
7463 : * better here. Use of the skip support infrastructure often
7464 : * avoids most next/prior key probes. But even when it can't,
7465 : * there's a decent chance that most individual next/prior key
7466 : * probes will locate a leaf page whose key space overlaps all
7467 : * of the scan's keys (even the lower-order keys) -- which
7468 : * also avoids the need for a separate, extra index descent.
7469 : * Note also that these probes are much cheaper than non-probe
7470 : * primitive index scans: they're reliably very selective.
7471 : */
7472 26070 : if (indexSkipQuals == NIL)
7473 25780 : ndistinct += 1;
7474 :
7475 : /*
7476 : * Update num_sa_scans estimate by multiplying by ndistinct.
7477 : *
7478 : * We make the pessimistic assumption that there is no
7479 : * naturally occurring cross-column correlation. This is
7480 : * often wrong, but it seems best to err on the side of not
7481 : * expecting skipping to be helpful...
7482 : */
7483 26070 : num_sa_scans *= ndistinct;
7484 :
7485 : /*
7486 : * ...but back out of adding this latest group of 1 or more
7487 : * skip arrays when num_sa_scans exceeds the total number of
7488 : * index pages (revert to num_sa_scans from before indexcol).
7489 : * This causes a sharp discontinuity in cost (as a function of
7490 : * the indexcol's ndistinct), but that is representative of
7491 : * actual runtime costs.
7492 : *
7493 : * Note that skipping is helpful when each primitive index
7494 : * scan only manages to skip over 1 or 2 irrelevant leaf pages
7495 : * on average. Skip arrays bring savings in CPU costs due to
7496 : * the scan not needing to evaluate indexquals against every
7497 : * tuple, which can greatly exceed any savings in I/O costs.
7498 : * This test is a test of whether num_sa_scans implies that
7499 : * we're past the point where the ability to skip ceases to
7500 : * lower the scan's costs (even qual evaluation CPU costs).
7501 : */
7502 26070 : if (index->pages < num_sa_scans)
7503 : {
7504 22982 : num_sa_scans = num_sa_scans_prev_cols;
7505 22982 : break; /* done building indexBoundQuals */
7506 : }
7507 :
7508 3088 : indexcol++;
7509 3088 : indexSkipQuals = NIL;
7510 : }
7511 :
7512 : /*
7513 : * Finished considering the need to add skip arrays to bridge an
7514 : * initial eqQualHere gap between the old and new index columns
7515 : * (or there was no initial eqQualHere gap in the first place).
7516 : *
7517 : * If an initial gap could not be bridged, then new column's quals
7518 : * (i.e. iclause->indexcol's quals) won't go into indexBoundQuals,
7519 : * and so won't affect our final numIndexTuples estimate.
7520 : */
7521 98260 : if (indexcol != iclause->indexcol)
7522 28406 : break; /* done building indexBoundQuals */
7523 : }
7524 :
7525 : Assert(indexcol == iclause->indexcol);
7526 :
7527 : /* Examine each indexqual associated with this index clause */
7528 1051840 : foreach(lc2, iclause->indexquals)
7529 : {
7530 527290 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7531 527290 : Expr *clause = rinfo->clause;
7532 527290 : Oid clause_op = InvalidOid;
7533 : int op_strategy;
7534 :
7535 527290 : if (IsA(clause, OpExpr))
7536 : {
7537 514194 : OpExpr *op = (OpExpr *) clause;
7538 :
7539 514194 : clause_op = op->opno;
7540 : }
7541 13096 : else if (IsA(clause, RowCompareExpr))
7542 : {
7543 396 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7544 :
7545 396 : clause_op = linitial_oid(rc->opnos);
7546 396 : found_row_compare = true;
7547 : }
7548 12700 : else if (IsA(clause, ScalarArrayOpExpr))
7549 : {
7550 10416 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7551 10416 : Node *other_operand = (Node *) lsecond(saop->args);
7552 10416 : double alength = estimate_array_length(root, other_operand);
7553 :
7554 10416 : clause_op = saop->opno;
7555 10416 : found_array = true;
7556 : /* estimate SA descents by indexBoundQuals only */
7557 10416 : if (alength > 1)
7558 10216 : num_sa_scans *= alength;
7559 : }
7560 2284 : else if (IsA(clause, NullTest))
7561 : {
7562 2284 : NullTest *nt = (NullTest *) clause;
7563 :
7564 2284 : if (nt->nulltesttype == IS_NULL)
7565 : {
7566 240 : found_is_null_op = true;
7567 : /* IS NULL is like = for selectivity/skip scan purposes */
7568 240 : eqQualHere = true;
7569 : }
7570 : }
7571 : else
7572 0 : elog(ERROR, "unsupported indexqual type: %d",
7573 : (int) nodeTag(clause));
7574 :
7575 : /* check for equality operator */
7576 527290 : if (OidIsValid(clause_op))
7577 : {
7578 525006 : op_strategy = get_op_opfamily_strategy(clause_op,
7579 525006 : index->opfamily[indexcol]);
7580 : Assert(op_strategy != 0); /* not a member of opfamily?? */
7581 525006 : if (op_strategy == BTEqualStrategyNumber)
7582 498698 : eqQualHere = true;
7583 : }
7584 :
7585 527290 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
7586 :
7587 : /*
7588 : * We apply inequality selectivities to estimate index descent
7589 : * costs with scans that use skip arrays. Save this indexcol's
7590 : * RestrictInfos if it looks like they'll be needed for that.
7591 : */
7592 527290 : if (!eqQualHere && !found_row_compare &&
7593 27254 : indexcol < index->nkeycolumns - 1)
7594 5704 : indexSkipQuals = lappend(indexSkipQuals, rinfo);
7595 : }
7596 : }
7597 :
7598 : /*
7599 : * If index is unique and we found an '=' clause for each column, we can
7600 : * just assume numIndexTuples = 1 and skip the expensive
7601 : * clauselist_selectivity calculations. However, an array or NullTest
7602 : * always invalidates that theory (even when eqQualHere has been set).
7603 : */
7604 737396 : if (index->unique &&
7605 592988 : indexcol == index->nkeycolumns - 1 &&
7606 240114 : eqQualHere &&
7607 240114 : !found_array &&
7608 234148 : !found_is_null_op)
7609 234100 : numIndexTuples = 1.0;
7610 : else
7611 : {
7612 : List *selectivityQuals;
7613 : Selectivity btreeSelectivity;
7614 :
7615 : /*
7616 : * If the index is partial, AND the index predicate with the
7617 : * index-bound quals to produce a more accurate idea of the number of
7618 : * rows covered by the bound conditions.
7619 : */
7620 503296 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
7621 :
7622 503296 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
7623 503296 : index->rel->relid,
7624 : JOIN_INNER,
7625 : NULL);
7626 503296 : numIndexTuples = btreeSelectivity * index->rel->tuples;
7627 :
7628 : /*
7629 : * btree automatically combines individual array element primitive
7630 : * index scans whenever the tuples covered by the next set of array
7631 : * keys are close to tuples covered by the current set. That puts a
7632 : * natural ceiling on the worst case number of descents -- there
7633 : * cannot possibly be more than one descent per leaf page scanned.
7634 : *
7635 : * Clamp the number of descents to at most 1/3 the number of index
7636 : * pages. This avoids implausibly high estimates with low selectivity
7637 : * paths, where scans usually require only one or two descents. This
7638 : * is most likely to help when there are several SAOP clauses, where
7639 : * naively accepting the total number of distinct combinations of
7640 : * array elements as the number of descents would frequently lead to
7641 : * wild overestimates.
7642 : *
7643 : * We somewhat arbitrarily don't just make the cutoff the total number
7644 : * of leaf pages (we make it 1/3 the total number of pages instead) to
7645 : * give the btree code credit for its ability to continue on the leaf
7646 : * level with low selectivity scans.
7647 : *
7648 : * Note: num_sa_scans includes both ScalarArrayOp array elements and
7649 : * skip array elements whose qual affects our numIndexTuples estimate.
7650 : */
7651 503296 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
7652 503296 : num_sa_scans = Max(num_sa_scans, 1);
7653 :
7654 : /*
7655 : * As in genericcostestimate(), we have to adjust for any array quals
7656 : * included in indexBoundQuals, and then round to integer.
7657 : *
7658 : * It is tempting to make genericcostestimate behave as if array
7659 : * clauses work in almost the same way as scalar operators during
7660 : * btree scans, making the top-level scan look like a continuous scan
7661 : * (as opposed to num_sa_scans-many primitive index scans). After
7662 : * all, btree scans mostly work like that at runtime. However, such a
7663 : * scheme would badly bias genericcostestimate's simplistic approach
7664 : * to calculating numIndexPages through prorating.
7665 : *
7666 : * Stick with the approach taken by non-native SAOP scans for now.
7667 : * genericcostestimate will use the Mackert-Lohman formula to
7668 : * compensate for repeat page fetches, even though that definitely
7669 : * won't happen during btree scans (not for leaf pages, at least).
7670 : * We're usually very pessimistic about the number of primitive index
7671 : * scans that will be required, but it's not clear how to do better.
7672 : */
7673 503296 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7674 : }
7675 :
7676 : /*
7677 : * Now do generic index cost estimation.
7678 : */
7679 737396 : costs.numIndexTuples = numIndexTuples;
7680 737396 : costs.num_sa_scans = num_sa_scans;
7681 :
7682 737396 : genericcostestimate(root, path, loop_count, &costs);
7683 :
7684 : /*
7685 : * Add a CPU-cost component to represent the costs of initial btree
7686 : * descent. We don't charge any I/O cost for touching upper btree levels,
7687 : * since they tend to stay in cache, but we still have to do about log2(N)
7688 : * comparisons to descend a btree of N leaf tuples. We charge one
7689 : * cpu_operator_cost per comparison.
7690 : *
7691 : * If there are SAOP or skip array keys, charge this once per estimated
7692 : * index descent. The ones after the first one are not startup cost so
7693 : * far as the overall plan goes, so just add them to "total" cost.
7694 : */
7695 737396 : if (index->tuples > 1) /* avoid computing log(0) */
7696 : {
7697 700060 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
7698 700060 : costs.indexStartupCost += descentCost;
7699 700060 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7700 : }
7701 :
7702 : /*
7703 : * Even though we're not charging I/O cost for touching upper btree pages,
7704 : * it's still reasonable to charge some CPU cost per page descended
7705 : * through. Moreover, if we had no such charge at all, bloated indexes
7706 : * would appear to have the same search cost as unbloated ones, at least
7707 : * in cases where only a single leaf page is expected to be visited. This
7708 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
7709 : * touched. The number of such pages is btree tree height plus one (ie,
7710 : * we charge for the leaf page too). As above, charge once per estimated
7711 : * SAOP/skip array descent.
7712 : */
7713 737396 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7714 737396 : costs.indexStartupCost += descentCost;
7715 737396 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7716 :
7717 737396 : if (!have_correlation)
7718 : {
7719 709154 : examine_indexcol_variable(root, index, 0, &vardata);
7720 709154 : if (HeapTupleIsValid(vardata.statsTuple))
7721 540824 : costs.indexCorrelation = btcost_correlation(index, &vardata);
7722 709154 : ReleaseVariableStats(vardata);
7723 : }
7724 : else
7725 : {
7726 : /* btcost_correlation already called earlier on */
7727 28242 : costs.indexCorrelation = correlation;
7728 : }
7729 :
7730 737396 : *indexStartupCost = costs.indexStartupCost;
7731 737396 : *indexTotalCost = costs.indexTotalCost;
7732 737396 : *indexSelectivity = costs.indexSelectivity;
7733 737396 : *indexCorrelation = costs.indexCorrelation;
7734 737396 : *indexPages = costs.numIndexPages;
7735 737396 : }
7736 :
7737 : void
7738 418 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7739 : Cost *indexStartupCost, Cost *indexTotalCost,
7740 : Selectivity *indexSelectivity, double *indexCorrelation,
7741 : double *indexPages)
7742 : {
7743 418 : GenericCosts costs = {0};
7744 :
7745 418 : genericcostestimate(root, path, loop_count, &costs);
7746 :
7747 : /*
7748 : * A hash index has no descent costs as such, since the index AM can go
7749 : * directly to the target bucket after computing the hash value. There
7750 : * are a couple of other hash-specific costs that we could conceivably add
7751 : * here, though:
7752 : *
7753 : * Ideally we'd charge spc_random_page_cost for each page in the target
7754 : * bucket, not just the numIndexPages pages that genericcostestimate
7755 : * thought we'd visit. However in most cases we don't know which bucket
7756 : * that will be. There's no point in considering the average bucket size
7757 : * because the hash AM makes sure that's always one page.
7758 : *
7759 : * Likewise, we could consider charging some CPU for each index tuple in
7760 : * the bucket, if we knew how many there were. But the per-tuple cost is
7761 : * just a hash value comparison, not a general datatype-dependent
7762 : * comparison, so any such charge ought to be quite a bit less than
7763 : * cpu_operator_cost; which makes it probably not worth worrying about.
7764 : *
7765 : * A bigger issue is that chance hash-value collisions will result in
7766 : * wasted probes into the heap. We don't currently attempt to model this
7767 : * cost on the grounds that it's rare, but maybe it's not rare enough.
7768 : * (Any fix for this ought to consider the generic lossy-operator problem,
7769 : * though; it's not entirely hash-specific.)
7770 : */
7771 :
7772 418 : *indexStartupCost = costs.indexStartupCost;
7773 418 : *indexTotalCost = costs.indexTotalCost;
7774 418 : *indexSelectivity = costs.indexSelectivity;
7775 418 : *indexCorrelation = costs.indexCorrelation;
7776 418 : *indexPages = costs.numIndexPages;
7777 418 : }
7778 :
7779 : void
7780 4790 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7781 : Cost *indexStartupCost, Cost *indexTotalCost,
7782 : Selectivity *indexSelectivity, double *indexCorrelation,
7783 : double *indexPages)
7784 : {
7785 4790 : IndexOptInfo *index = path->indexinfo;
7786 4790 : GenericCosts costs = {0};
7787 : Cost descentCost;
7788 :
7789 4790 : genericcostestimate(root, path, loop_count, &costs);
7790 :
7791 : /*
7792 : * We model index descent costs similarly to those for btree, but to do
7793 : * that we first need an idea of the tree height. We somewhat arbitrarily
7794 : * assume that the fanout is 100, meaning the tree height is at most
7795 : * log100(index->pages).
7796 : *
7797 : * Although this computation isn't really expensive enough to require
7798 : * caching, we might as well use index->tree_height to cache it.
7799 : */
7800 4790 : if (index->tree_height < 0) /* unknown? */
7801 : {
7802 4776 : if (index->pages > 1) /* avoid computing log(0) */
7803 2720 : index->tree_height = (int) (log(index->pages) / log(100.0));
7804 : else
7805 2056 : index->tree_height = 0;
7806 : }
7807 :
7808 : /*
7809 : * Add a CPU-cost component to represent the costs of initial descent. We
7810 : * just use log(N) here not log2(N) since the branching factor isn't
7811 : * necessarily two anyway. As for btree, charge once per SA scan.
7812 : */
7813 4790 : if (index->tuples > 1) /* avoid computing log(0) */
7814 : {
7815 4790 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7816 4790 : costs.indexStartupCost += descentCost;
7817 4790 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7818 : }
7819 :
7820 : /*
7821 : * Likewise add a per-page charge, calculated the same as for btrees.
7822 : */
7823 4790 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7824 4790 : costs.indexStartupCost += descentCost;
7825 4790 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7826 :
7827 4790 : *indexStartupCost = costs.indexStartupCost;
7828 4790 : *indexTotalCost = costs.indexTotalCost;
7829 4790 : *indexSelectivity = costs.indexSelectivity;
7830 4790 : *indexCorrelation = costs.indexCorrelation;
7831 4790 : *indexPages = costs.numIndexPages;
7832 4790 : }
7833 :
7834 : void
7835 1784 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7836 : Cost *indexStartupCost, Cost *indexTotalCost,
7837 : Selectivity *indexSelectivity, double *indexCorrelation,
7838 : double *indexPages)
7839 : {
7840 1784 : IndexOptInfo *index = path->indexinfo;
7841 1784 : GenericCosts costs = {0};
7842 : Cost descentCost;
7843 :
7844 1784 : genericcostestimate(root, path, loop_count, &costs);
7845 :
7846 : /*
7847 : * We model index descent costs similarly to those for btree, but to do
7848 : * that we first need an idea of the tree height. We somewhat arbitrarily
7849 : * assume that the fanout is 100, meaning the tree height is at most
7850 : * log100(index->pages).
7851 : *
7852 : * Although this computation isn't really expensive enough to require
7853 : * caching, we might as well use index->tree_height to cache it.
7854 : */
7855 1784 : if (index->tree_height < 0) /* unknown? */
7856 : {
7857 1778 : if (index->pages > 1) /* avoid computing log(0) */
7858 1778 : index->tree_height = (int) (log(index->pages) / log(100.0));
7859 : else
7860 0 : index->tree_height = 0;
7861 : }
7862 :
7863 : /*
7864 : * Add a CPU-cost component to represent the costs of initial descent. We
7865 : * just use log(N) here not log2(N) since the branching factor isn't
7866 : * necessarily two anyway. As for btree, charge once per SA scan.
7867 : */
7868 1784 : if (index->tuples > 1) /* avoid computing log(0) */
7869 : {
7870 1784 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7871 1784 : costs.indexStartupCost += descentCost;
7872 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7873 : }
7874 :
7875 : /*
7876 : * Likewise add a per-page charge, calculated the same as for btrees.
7877 : */
7878 1784 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7879 1784 : costs.indexStartupCost += descentCost;
7880 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7881 :
7882 1784 : *indexStartupCost = costs.indexStartupCost;
7883 1784 : *indexTotalCost = costs.indexTotalCost;
7884 1784 : *indexSelectivity = costs.indexSelectivity;
7885 1784 : *indexCorrelation = costs.indexCorrelation;
7886 1784 : *indexPages = costs.numIndexPages;
7887 1784 : }
7888 :
7889 :
7890 : /*
7891 : * Support routines for gincostestimate
7892 : */
7893 :
7894 : typedef struct
7895 : {
7896 : bool attHasFullScan[INDEX_MAX_KEYS];
7897 : bool attHasNormalScan[INDEX_MAX_KEYS];
7898 : double partialEntries;
7899 : double exactEntries;
7900 : double searchEntries;
7901 : double arrayScans;
7902 : } GinQualCounts;
7903 :
7904 : /*
7905 : * Estimate the number of index terms that need to be searched for while
7906 : * testing the given GIN query, and increment the counts in *counts
7907 : * appropriately. If the query is unsatisfiable, return false.
7908 : */
7909 : static bool
7910 2456 : gincost_pattern(IndexOptInfo *index, int indexcol,
7911 : Oid clause_op, Datum query,
7912 : GinQualCounts *counts)
7913 : {
7914 : FmgrInfo flinfo;
7915 : Oid extractProcOid;
7916 : Oid collation;
7917 : int strategy_op;
7918 : Oid lefttype,
7919 : righttype;
7920 2456 : int32 nentries = 0;
7921 2456 : bool *partial_matches = NULL;
7922 2456 : Pointer *extra_data = NULL;
7923 2456 : bool *nullFlags = NULL;
7924 2456 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
7925 : int32 i;
7926 :
7927 : Assert(indexcol < index->nkeycolumns);
7928 :
7929 : /*
7930 : * Get the operator's strategy number and declared input data types within
7931 : * the index opfamily. (We don't need the latter, but we use
7932 : * get_op_opfamily_properties because it will throw error if it fails to
7933 : * find a matching pg_amop entry.)
7934 : */
7935 2456 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
7936 : &strategy_op, &lefttype, &righttype);
7937 :
7938 : /*
7939 : * GIN always uses the "default" support functions, which are those with
7940 : * lefttype == righttype == the opclass' opcintype (see
7941 : * IndexSupportInitialize in relcache.c).
7942 : */
7943 2456 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
7944 2456 : index->opcintype[indexcol],
7945 2456 : index->opcintype[indexcol],
7946 : GIN_EXTRACTQUERY_PROC);
7947 :
7948 2456 : if (!OidIsValid(extractProcOid))
7949 : {
7950 : /* should not happen; throw same error as index_getprocinfo */
7951 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
7952 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
7953 : get_rel_name(index->indexoid));
7954 : }
7955 :
7956 : /*
7957 : * Choose collation to pass to extractProc (should match initGinState).
7958 : */
7959 2456 : if (OidIsValid(index->indexcollations[indexcol]))
7960 390 : collation = index->indexcollations[indexcol];
7961 : else
7962 2066 : collation = DEFAULT_COLLATION_OID;
7963 :
7964 2456 : fmgr_info(extractProcOid, &flinfo);
7965 :
7966 2456 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
7967 :
7968 2456 : FunctionCall7Coll(&flinfo,
7969 : collation,
7970 : query,
7971 : PointerGetDatum(&nentries),
7972 : UInt16GetDatum(strategy_op),
7973 : PointerGetDatum(&partial_matches),
7974 : PointerGetDatum(&extra_data),
7975 : PointerGetDatum(&nullFlags),
7976 : PointerGetDatum(&searchMode));
7977 :
7978 2456 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
7979 : {
7980 : /* No match is possible */
7981 12 : return false;
7982 : }
7983 :
7984 9568 : for (i = 0; i < nentries; i++)
7985 : {
7986 : /*
7987 : * For partial match we haven't any information to estimate number of
7988 : * matched entries in index, so, we just estimate it as 100
7989 : */
7990 7124 : if (partial_matches && partial_matches[i])
7991 694 : counts->partialEntries += 100;
7992 : else
7993 6430 : counts->exactEntries++;
7994 :
7995 7124 : counts->searchEntries++;
7996 : }
7997 :
7998 2444 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
7999 : {
8000 1972 : counts->attHasNormalScan[indexcol] = true;
8001 : }
8002 472 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
8003 : {
8004 : /* Treat "include empty" like an exact-match item */
8005 44 : counts->attHasNormalScan[indexcol] = true;
8006 44 : counts->exactEntries++;
8007 44 : counts->searchEntries++;
8008 : }
8009 : else
8010 : {
8011 : /* It's GIN_SEARCH_MODE_ALL */
8012 428 : counts->attHasFullScan[indexcol] = true;
8013 : }
8014 :
8015 2444 : return true;
8016 : }
8017 :
8018 : /*
8019 : * Estimate the number of index terms that need to be searched for while
8020 : * testing the given GIN index clause, and increment the counts in *counts
8021 : * appropriately. If the query is unsatisfiable, return false.
8022 : */
8023 : static bool
8024 2444 : gincost_opexpr(PlannerInfo *root,
8025 : IndexOptInfo *index,
8026 : int indexcol,
8027 : OpExpr *clause,
8028 : GinQualCounts *counts)
8029 : {
8030 2444 : Oid clause_op = clause->opno;
8031 2444 : Node *operand = (Node *) lsecond(clause->args);
8032 :
8033 : /* aggressively reduce to a constant, and look through relabeling */
8034 2444 : operand = estimate_expression_value(root, operand);
8035 :
8036 2444 : if (IsA(operand, RelabelType))
8037 0 : operand = (Node *) ((RelabelType *) operand)->arg;
8038 :
8039 : /*
8040 : * It's impossible to call extractQuery method for unknown operand. So
8041 : * unless operand is a Const we can't do much; just assume there will be
8042 : * one ordinary search entry from the operand at runtime.
8043 : */
8044 2444 : if (!IsA(operand, Const))
8045 : {
8046 0 : counts->exactEntries++;
8047 0 : counts->searchEntries++;
8048 0 : return true;
8049 : }
8050 :
8051 : /* If Const is null, there can be no matches */
8052 2444 : if (((Const *) operand)->constisnull)
8053 0 : return false;
8054 :
8055 : /* Otherwise, apply extractQuery and get the actual term counts */
8056 2444 : return gincost_pattern(index, indexcol, clause_op,
8057 : ((Const *) operand)->constvalue,
8058 : counts);
8059 : }
8060 :
8061 : /*
8062 : * Estimate the number of index terms that need to be searched for while
8063 : * testing the given GIN index clause, and increment the counts in *counts
8064 : * appropriately. If the query is unsatisfiable, return false.
8065 : *
8066 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
8067 : * each of which involves one value from the RHS array, plus all the
8068 : * non-array quals (if any). To model this, we average the counts across
8069 : * the RHS elements, and add the averages to the counts in *counts (which
8070 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
8071 : * by N, causing gincostestimate to scale up its estimates accordingly.
8072 : */
8073 : static bool
8074 6 : gincost_scalararrayopexpr(PlannerInfo *root,
8075 : IndexOptInfo *index,
8076 : int indexcol,
8077 : ScalarArrayOpExpr *clause,
8078 : double numIndexEntries,
8079 : GinQualCounts *counts)
8080 : {
8081 6 : Oid clause_op = clause->opno;
8082 6 : Node *rightop = (Node *) lsecond(clause->args);
8083 : ArrayType *arrayval;
8084 : int16 elmlen;
8085 : bool elmbyval;
8086 : char elmalign;
8087 : int numElems;
8088 : Datum *elemValues;
8089 : bool *elemNulls;
8090 : GinQualCounts arraycounts;
8091 6 : int numPossible = 0;
8092 : int i;
8093 :
8094 : Assert(clause->useOr);
8095 :
8096 : /* aggressively reduce to a constant, and look through relabeling */
8097 6 : rightop = estimate_expression_value(root, rightop);
8098 :
8099 6 : if (IsA(rightop, RelabelType))
8100 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
8101 :
8102 : /*
8103 : * It's impossible to call extractQuery method for unknown operand. So
8104 : * unless operand is a Const we can't do much; just assume there will be
8105 : * one ordinary search entry from each array entry at runtime, and fall
8106 : * back on a probably-bad estimate of the number of array entries.
8107 : */
8108 6 : if (!IsA(rightop, Const))
8109 : {
8110 0 : counts->exactEntries++;
8111 0 : counts->searchEntries++;
8112 0 : counts->arrayScans *= estimate_array_length(root, rightop);
8113 0 : return true;
8114 : }
8115 :
8116 : /* If Const is null, there can be no matches */
8117 6 : if (((Const *) rightop)->constisnull)
8118 0 : return false;
8119 :
8120 : /* Otherwise, extract the array elements and iterate over them */
8121 6 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
8122 6 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
8123 : &elmlen, &elmbyval, &elmalign);
8124 6 : deconstruct_array(arrayval,
8125 : ARR_ELEMTYPE(arrayval),
8126 : elmlen, elmbyval, elmalign,
8127 : &elemValues, &elemNulls, &numElems);
8128 :
8129 6 : memset(&arraycounts, 0, sizeof(arraycounts));
8130 :
8131 18 : for (i = 0; i < numElems; i++)
8132 : {
8133 : GinQualCounts elemcounts;
8134 :
8135 : /* NULL can't match anything, so ignore, as the executor will */
8136 12 : if (elemNulls[i])
8137 0 : continue;
8138 :
8139 : /* Otherwise, apply extractQuery and get the actual term counts */
8140 12 : memset(&elemcounts, 0, sizeof(elemcounts));
8141 :
8142 12 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
8143 : &elemcounts))
8144 : {
8145 : /* We ignore array elements that are unsatisfiable patterns */
8146 12 : numPossible++;
8147 :
8148 12 : if (elemcounts.attHasFullScan[indexcol] &&
8149 0 : !elemcounts.attHasNormalScan[indexcol])
8150 : {
8151 : /*
8152 : * Full index scan will be required. We treat this as if
8153 : * every key in the index had been listed in the query; is
8154 : * that reasonable?
8155 : */
8156 0 : elemcounts.partialEntries = 0;
8157 0 : elemcounts.exactEntries = numIndexEntries;
8158 0 : elemcounts.searchEntries = numIndexEntries;
8159 : }
8160 12 : arraycounts.partialEntries += elemcounts.partialEntries;
8161 12 : arraycounts.exactEntries += elemcounts.exactEntries;
8162 12 : arraycounts.searchEntries += elemcounts.searchEntries;
8163 : }
8164 : }
8165 :
8166 6 : if (numPossible == 0)
8167 : {
8168 : /* No satisfiable patterns in the array */
8169 0 : return false;
8170 : }
8171 :
8172 : /*
8173 : * Now add the averages to the global counts. This will give us an
8174 : * estimate of the average number of terms searched for in each indexscan,
8175 : * including contributions from both array and non-array quals.
8176 : */
8177 6 : counts->partialEntries += arraycounts.partialEntries / numPossible;
8178 6 : counts->exactEntries += arraycounts.exactEntries / numPossible;
8179 6 : counts->searchEntries += arraycounts.searchEntries / numPossible;
8180 :
8181 6 : counts->arrayScans *= numPossible;
8182 :
8183 6 : return true;
8184 : }
8185 :
8186 : /*
8187 : * GIN has search behavior completely different from other index types
8188 : */
8189 : void
8190 2252 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8191 : Cost *indexStartupCost, Cost *indexTotalCost,
8192 : Selectivity *indexSelectivity, double *indexCorrelation,
8193 : double *indexPages)
8194 : {
8195 2252 : IndexOptInfo *index = path->indexinfo;
8196 2252 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8197 : List *selectivityQuals;
8198 2252 : double numPages = index->pages,
8199 2252 : numTuples = index->tuples;
8200 : double numEntryPages,
8201 : numDataPages,
8202 : numPendingPages,
8203 : numEntries;
8204 : GinQualCounts counts;
8205 : bool matchPossible;
8206 : bool fullIndexScan;
8207 : double partialScale;
8208 : double entryPagesFetched,
8209 : dataPagesFetched,
8210 : dataPagesFetchedBySel;
8211 : double qual_op_cost,
8212 : qual_arg_cost,
8213 : spc_random_page_cost,
8214 : outer_scans;
8215 : Cost descentCost;
8216 : Relation indexRel;
8217 : GinStatsData ginStats;
8218 : ListCell *lc;
8219 : int i;
8220 :
8221 : /*
8222 : * Obtain statistical information from the meta page, if possible. Else
8223 : * set ginStats to zeroes, and we'll cope below.
8224 : */
8225 2252 : if (!index->hypothetical)
8226 : {
8227 : /* Lock should have already been obtained in plancat.c */
8228 2252 : indexRel = index_open(index->indexoid, NoLock);
8229 2252 : ginGetStats(indexRel, &ginStats);
8230 2252 : index_close(indexRel, NoLock);
8231 : }
8232 : else
8233 : {
8234 0 : memset(&ginStats, 0, sizeof(ginStats));
8235 : }
8236 :
8237 : /*
8238 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
8239 : * trusted, but the other fields are data as of the last VACUUM. We can
8240 : * scale them up to account for growth since then, but that method only
8241 : * goes so far; in the worst case, the stats might be for a completely
8242 : * empty index, and scaling them will produce pretty bogus numbers.
8243 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
8244 : * it's grown more than that, fall back to estimating things only from the
8245 : * assumed-accurate index size. But we'll trust nPendingPages in any case
8246 : * so long as it's not clearly insane, ie, more than the index size.
8247 : */
8248 2252 : if (ginStats.nPendingPages < numPages)
8249 2252 : numPendingPages = ginStats.nPendingPages;
8250 : else
8251 0 : numPendingPages = 0;
8252 :
8253 2252 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
8254 2252 : ginStats.nTotalPages > numPages / 4 &&
8255 2200 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
8256 1942 : {
8257 : /*
8258 : * OK, the stats seem close enough to sane to be trusted. But we
8259 : * still need to scale them by the ratio numPages / nTotalPages to
8260 : * account for growth since the last VACUUM.
8261 : */
8262 1942 : double scale = numPages / ginStats.nTotalPages;
8263 :
8264 1942 : numEntryPages = ceil(ginStats.nEntryPages * scale);
8265 1942 : numDataPages = ceil(ginStats.nDataPages * scale);
8266 1942 : numEntries = ceil(ginStats.nEntries * scale);
8267 : /* ensure we didn't round up too much */
8268 1942 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
8269 1942 : numDataPages = Min(numDataPages,
8270 : numPages - numPendingPages - numEntryPages);
8271 : }
8272 : else
8273 : {
8274 : /*
8275 : * We might get here because it's a hypothetical index, or an index
8276 : * created pre-9.1 and never vacuumed since upgrading (in which case
8277 : * its stats would read as zeroes), or just because it's grown too
8278 : * much since the last VACUUM for us to put our faith in scaling.
8279 : *
8280 : * Invent some plausible internal statistics based on the index page
8281 : * count (and clamp that to at least 10 pages, just in case). We
8282 : * estimate that 90% of the index is entry pages, and the rest is data
8283 : * pages. Estimate 100 entries per entry page; this is rather bogus
8284 : * since it'll depend on the size of the keys, but it's more robust
8285 : * than trying to predict the number of entries per heap tuple.
8286 : */
8287 310 : numPages = Max(numPages, 10);
8288 310 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
8289 310 : numDataPages = numPages - numPendingPages - numEntryPages;
8290 310 : numEntries = floor(numEntryPages * 100);
8291 : }
8292 :
8293 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
8294 2252 : if (numEntries < 1)
8295 0 : numEntries = 1;
8296 :
8297 : /*
8298 : * If the index is partial, AND the index predicate with the index-bound
8299 : * quals to produce a more accurate idea of the number of rows covered by
8300 : * the bound conditions.
8301 : */
8302 2252 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
8303 :
8304 : /* Estimate the fraction of main-table tuples that will be visited */
8305 4504 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
8306 2252 : index->rel->relid,
8307 : JOIN_INNER,
8308 : NULL);
8309 :
8310 : /* fetch estimated page cost for tablespace containing index */
8311 2252 : get_tablespace_page_costs(index->reltablespace,
8312 : &spc_random_page_cost,
8313 : NULL);
8314 :
8315 : /*
8316 : * Generic assumption about index correlation: there isn't any.
8317 : */
8318 2252 : *indexCorrelation = 0.0;
8319 :
8320 : /*
8321 : * Examine quals to estimate number of search entries & partial matches
8322 : */
8323 2252 : memset(&counts, 0, sizeof(counts));
8324 2252 : counts.arrayScans = 1;
8325 2252 : matchPossible = true;
8326 :
8327 4702 : foreach(lc, path->indexclauses)
8328 : {
8329 2450 : IndexClause *iclause = lfirst_node(IndexClause, lc);
8330 : ListCell *lc2;
8331 :
8332 4888 : foreach(lc2, iclause->indexquals)
8333 : {
8334 2450 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8335 2450 : Expr *clause = rinfo->clause;
8336 :
8337 2450 : if (IsA(clause, OpExpr))
8338 : {
8339 2444 : matchPossible = gincost_opexpr(root,
8340 : index,
8341 2444 : iclause->indexcol,
8342 : (OpExpr *) clause,
8343 : &counts);
8344 2444 : if (!matchPossible)
8345 12 : break;
8346 : }
8347 6 : else if (IsA(clause, ScalarArrayOpExpr))
8348 : {
8349 6 : matchPossible = gincost_scalararrayopexpr(root,
8350 : index,
8351 6 : iclause->indexcol,
8352 : (ScalarArrayOpExpr *) clause,
8353 : numEntries,
8354 : &counts);
8355 6 : if (!matchPossible)
8356 0 : break;
8357 : }
8358 : else
8359 : {
8360 : /* shouldn't be anything else for a GIN index */
8361 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
8362 : (int) nodeTag(clause));
8363 : }
8364 : }
8365 : }
8366 :
8367 : /* Fall out if there were any provably-unsatisfiable quals */
8368 2252 : if (!matchPossible)
8369 : {
8370 12 : *indexStartupCost = 0;
8371 12 : *indexTotalCost = 0;
8372 12 : *indexSelectivity = 0;
8373 12 : return;
8374 : }
8375 :
8376 : /*
8377 : * If attribute has a full scan and at the same time doesn't have normal
8378 : * scan, then we'll have to scan all non-null entries of that attribute.
8379 : * Currently, we don't have per-attribute statistics for GIN. Thus, we
8380 : * must assume the whole GIN index has to be scanned in this case.
8381 : */
8382 2240 : fullIndexScan = false;
8383 4370 : for (i = 0; i < index->nkeycolumns; i++)
8384 : {
8385 2468 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
8386 : {
8387 338 : fullIndexScan = true;
8388 338 : break;
8389 : }
8390 : }
8391 :
8392 2240 : if (fullIndexScan || indexQuals == NIL)
8393 : {
8394 : /*
8395 : * Full index scan will be required. We treat this as if every key in
8396 : * the index had been listed in the query; is that reasonable?
8397 : */
8398 338 : counts.partialEntries = 0;
8399 338 : counts.exactEntries = numEntries;
8400 338 : counts.searchEntries = numEntries;
8401 : }
8402 :
8403 : /* Will we have more than one iteration of a nestloop scan? */
8404 2240 : outer_scans = loop_count;
8405 :
8406 : /*
8407 : * Compute cost to begin scan, first of all, pay attention to pending
8408 : * list.
8409 : */
8410 2240 : entryPagesFetched = numPendingPages;
8411 :
8412 : /*
8413 : * Estimate number of entry pages read. We need to do
8414 : * counts.searchEntries searches. Use a power function as it should be,
8415 : * but tuples on leaf pages usually is much greater. Here we include all
8416 : * searches in entry tree, including search of first entry in partial
8417 : * match algorithm
8418 : */
8419 2240 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8420 :
8421 : /*
8422 : * Add an estimate of entry pages read by partial match algorithm. It's a
8423 : * scan over leaf pages in entry tree. We haven't any useful stats here,
8424 : * so estimate it as proportion. Because counts.partialEntries is really
8425 : * pretty bogus (see code above), it's possible that it is more than
8426 : * numEntries; clamp the proportion to ensure sanity.
8427 : */
8428 2240 : partialScale = counts.partialEntries / numEntries;
8429 2240 : partialScale = Min(partialScale, 1.0);
8430 :
8431 2240 : entryPagesFetched += ceil(numEntryPages * partialScale);
8432 :
8433 : /*
8434 : * Partial match algorithm reads all data pages before doing actual scan,
8435 : * so it's a startup cost. Again, we haven't any useful stats here, so
8436 : * estimate it as proportion.
8437 : */
8438 2240 : dataPagesFetched = ceil(numDataPages * partialScale);
8439 :
8440 2240 : *indexStartupCost = 0;
8441 2240 : *indexTotalCost = 0;
8442 :
8443 : /*
8444 : * Add a CPU-cost component to represent the costs of initial entry btree
8445 : * descent. We don't charge any I/O cost for touching upper btree levels,
8446 : * since they tend to stay in cache, but we still have to do about log2(N)
8447 : * comparisons to descend a btree of N leaf tuples. We charge one
8448 : * cpu_operator_cost per comparison.
8449 : *
8450 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
8451 : * ones after the first one are not startup cost so far as the overall
8452 : * plan is concerned, so add them only to "total" cost.
8453 : */
8454 2240 : if (numEntries > 1) /* avoid computing log(0) */
8455 : {
8456 2240 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
8457 2240 : *indexStartupCost += descentCost * counts.searchEntries;
8458 2240 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
8459 : }
8460 :
8461 : /*
8462 : * Add a cpu cost per entry-page fetched. This is not amortized over a
8463 : * loop.
8464 : */
8465 2240 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8466 2240 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8467 :
8468 : /*
8469 : * Add a cpu cost per data-page fetched. This is also not amortized over a
8470 : * loop. Since those are the data pages from the partial match algorithm,
8471 : * charge them as startup cost.
8472 : */
8473 2240 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
8474 :
8475 : /*
8476 : * Since we add the startup cost to the total cost later on, remove the
8477 : * initial arrayscan from the total.
8478 : */
8479 2240 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8480 :
8481 : /*
8482 : * Calculate cache effects if more than one scan due to nestloops or array
8483 : * quals. The result is pro-rated per nestloop scan, but the array qual
8484 : * factor shouldn't be pro-rated (compare genericcostestimate).
8485 : */
8486 2240 : if (outer_scans > 1 || counts.arrayScans > 1)
8487 : {
8488 6 : entryPagesFetched *= outer_scans * counts.arrayScans;
8489 6 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
8490 : (BlockNumber) numEntryPages,
8491 : numEntryPages, root);
8492 6 : entryPagesFetched /= outer_scans;
8493 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8494 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8495 : (BlockNumber) numDataPages,
8496 : numDataPages, root);
8497 6 : dataPagesFetched /= outer_scans;
8498 : }
8499 :
8500 : /*
8501 : * Here we use random page cost because logically-close pages could be far
8502 : * apart on disk.
8503 : */
8504 2240 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
8505 :
8506 : /*
8507 : * Now compute the number of data pages fetched during the scan.
8508 : *
8509 : * We assume every entry to have the same number of items, and that there
8510 : * is no overlap between them. (XXX: tsvector and array opclasses collect
8511 : * statistics on the frequency of individual keys; it would be nice to use
8512 : * those here.)
8513 : */
8514 2240 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
8515 :
8516 : /*
8517 : * If there is a lot of overlap among the entries, in particular if one of
8518 : * the entries is very frequent, the above calculation can grossly
8519 : * under-estimate. As a simple cross-check, calculate a lower bound based
8520 : * on the overall selectivity of the quals. At a minimum, we must read
8521 : * one item pointer for each matching entry.
8522 : *
8523 : * The width of each item pointer varies, based on the level of
8524 : * compression. We don't have statistics on that, but an average of
8525 : * around 3 bytes per item is fairly typical.
8526 : */
8527 2240 : dataPagesFetchedBySel = ceil(*indexSelectivity *
8528 2240 : (numTuples / (BLCKSZ / 3)));
8529 2240 : if (dataPagesFetchedBySel > dataPagesFetched)
8530 1860 : dataPagesFetched = dataPagesFetchedBySel;
8531 :
8532 : /* Add one page cpu-cost to the startup cost */
8533 2240 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
8534 :
8535 : /*
8536 : * Add once again a CPU-cost for those data pages, before amortizing for
8537 : * cache.
8538 : */
8539 2240 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8540 :
8541 : /* Account for cache effects, the same as above */
8542 2240 : if (outer_scans > 1 || counts.arrayScans > 1)
8543 : {
8544 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8545 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8546 : (BlockNumber) numDataPages,
8547 : numDataPages, root);
8548 6 : dataPagesFetched /= outer_scans;
8549 : }
8550 :
8551 : /* And apply random_page_cost as the cost per page */
8552 2240 : *indexTotalCost += *indexStartupCost +
8553 2240 : dataPagesFetched * spc_random_page_cost;
8554 :
8555 : /*
8556 : * Add on index qual eval costs, much as in genericcostestimate. We charge
8557 : * cpu but we can disregard indexorderbys, since GIN doesn't support
8558 : * those.
8559 : */
8560 2240 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8561 2240 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
8562 :
8563 2240 : *indexStartupCost += qual_arg_cost;
8564 2240 : *indexTotalCost += qual_arg_cost;
8565 :
8566 : /*
8567 : * Add a cpu cost per search entry, corresponding to the actual visited
8568 : * entries.
8569 : */
8570 2240 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
8571 : /* Now add a cpu cost per tuple in the posting lists / trees */
8572 2240 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
8573 2240 : *indexPages = dataPagesFetched;
8574 : }
8575 :
8576 : /*
8577 : * BRIN has search behavior completely different from other index types
8578 : */
8579 : void
8580 10730 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8581 : Cost *indexStartupCost, Cost *indexTotalCost,
8582 : Selectivity *indexSelectivity, double *indexCorrelation,
8583 : double *indexPages)
8584 : {
8585 10730 : IndexOptInfo *index = path->indexinfo;
8586 10730 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8587 10730 : double numPages = index->pages;
8588 10730 : RelOptInfo *baserel = index->rel;
8589 10730 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
8590 : Cost spc_seq_page_cost;
8591 : Cost spc_random_page_cost;
8592 : double qual_arg_cost;
8593 : double qualSelectivity;
8594 : BrinStatsData statsData;
8595 : double indexRanges;
8596 : double minimalRanges;
8597 : double estimatedRanges;
8598 : double selec;
8599 : Relation indexRel;
8600 : ListCell *l;
8601 : VariableStatData vardata;
8602 :
8603 : Assert(rte->rtekind == RTE_RELATION);
8604 :
8605 : /* fetch estimated page cost for the tablespace containing the index */
8606 10730 : get_tablespace_page_costs(index->reltablespace,
8607 : &spc_random_page_cost,
8608 : &spc_seq_page_cost);
8609 :
8610 : /*
8611 : * Obtain some data from the index itself, if possible. Otherwise invent
8612 : * some plausible internal statistics based on the relation page count.
8613 : */
8614 10730 : if (!index->hypothetical)
8615 : {
8616 : /*
8617 : * A lock should have already been obtained on the index in plancat.c.
8618 : */
8619 10730 : indexRel = index_open(index->indexoid, NoLock);
8620 10730 : brinGetStats(indexRel, &statsData);
8621 10730 : index_close(indexRel, NoLock);
8622 :
8623 : /* work out the actual number of ranges in the index */
8624 10730 : indexRanges = Max(ceil((double) baserel->pages /
8625 : statsData.pagesPerRange), 1.0);
8626 : }
8627 : else
8628 : {
8629 : /*
8630 : * Assume default number of pages per range, and estimate the number
8631 : * of ranges based on that.
8632 : */
8633 0 : indexRanges = Max(ceil((double) baserel->pages /
8634 : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
8635 :
8636 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
8637 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
8638 : }
8639 :
8640 : /*
8641 : * Compute index correlation
8642 : *
8643 : * Because we can use all index quals equally when scanning, we can use
8644 : * the largest correlation (in absolute value) among columns used by the
8645 : * query. Start at zero, the worst possible case. If we cannot find any
8646 : * correlation statistics, we will keep it as 0.
8647 : */
8648 10730 : *indexCorrelation = 0;
8649 :
8650 21462 : foreach(l, path->indexclauses)
8651 : {
8652 10732 : IndexClause *iclause = lfirst_node(IndexClause, l);
8653 10732 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
8654 :
8655 : /* attempt to lookup stats in relation for this index column */
8656 10732 : if (attnum != 0)
8657 : {
8658 : /* Simple variable -- look to stats for the underlying table */
8659 10732 : if (get_relation_stats_hook &&
8660 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
8661 : {
8662 : /*
8663 : * The hook took control of acquiring a stats tuple. If it
8664 : * did supply a tuple, it'd better have supplied a freefunc.
8665 : */
8666 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
8667 0 : elog(ERROR,
8668 : "no function provided to release variable stats with");
8669 : }
8670 : else
8671 : {
8672 10732 : vardata.statsTuple =
8673 10732 : SearchSysCache3(STATRELATTINH,
8674 : ObjectIdGetDatum(rte->relid),
8675 : Int16GetDatum(attnum),
8676 : BoolGetDatum(false));
8677 10732 : vardata.freefunc = ReleaseSysCache;
8678 : }
8679 : }
8680 : else
8681 : {
8682 : /*
8683 : * Looks like we've found an expression column in the index. Let's
8684 : * see if there's any stats for it.
8685 : */
8686 :
8687 : /* get the attnum from the 0-based index. */
8688 0 : attnum = iclause->indexcol + 1;
8689 :
8690 0 : if (get_index_stats_hook &&
8691 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
8692 : {
8693 : /*
8694 : * The hook took control of acquiring a stats tuple. If it
8695 : * did supply a tuple, it'd better have supplied a freefunc.
8696 : */
8697 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
8698 0 : !vardata.freefunc)
8699 0 : elog(ERROR, "no function provided to release variable stats with");
8700 : }
8701 : else
8702 : {
8703 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
8704 : ObjectIdGetDatum(index->indexoid),
8705 : Int16GetDatum(attnum),
8706 : BoolGetDatum(false));
8707 0 : vardata.freefunc = ReleaseSysCache;
8708 : }
8709 : }
8710 :
8711 10732 : if (HeapTupleIsValid(vardata.statsTuple))
8712 : {
8713 : AttStatsSlot sslot;
8714 :
8715 36 : if (get_attstatsslot(&sslot, vardata.statsTuple,
8716 : STATISTIC_KIND_CORRELATION, InvalidOid,
8717 : ATTSTATSSLOT_NUMBERS))
8718 : {
8719 36 : double varCorrelation = 0.0;
8720 :
8721 36 : if (sslot.nnumbers > 0)
8722 36 : varCorrelation = fabs(sslot.numbers[0]);
8723 :
8724 36 : if (varCorrelation > *indexCorrelation)
8725 36 : *indexCorrelation = varCorrelation;
8726 :
8727 36 : free_attstatsslot(&sslot);
8728 : }
8729 : }
8730 :
8731 10732 : ReleaseVariableStats(vardata);
8732 : }
8733 :
8734 10730 : qualSelectivity = clauselist_selectivity(root, indexQuals,
8735 10730 : baserel->relid,
8736 : JOIN_INNER, NULL);
8737 :
8738 : /*
8739 : * Now calculate the minimum possible ranges we could match with if all of
8740 : * the rows were in the perfect order in the table's heap.
8741 : */
8742 10730 : minimalRanges = ceil(indexRanges * qualSelectivity);
8743 :
8744 : /*
8745 : * Now estimate the number of ranges that we'll touch by using the
8746 : * indexCorrelation from the stats. Careful not to divide by zero (note
8747 : * we're using the absolute value of the correlation).
8748 : */
8749 10730 : if (*indexCorrelation < 1.0e-10)
8750 10694 : estimatedRanges = indexRanges;
8751 : else
8752 36 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
8753 :
8754 : /* we expect to visit this portion of the table */
8755 10730 : selec = estimatedRanges / indexRanges;
8756 :
8757 10730 : CLAMP_PROBABILITY(selec);
8758 :
8759 10730 : *indexSelectivity = selec;
8760 :
8761 : /*
8762 : * Compute the index qual costs, much as in genericcostestimate, to add to
8763 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
8764 : * support those.
8765 : */
8766 10730 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8767 :
8768 : /*
8769 : * Compute the startup cost as the cost to read the whole revmap
8770 : * sequentially, including the cost to execute the index quals.
8771 : */
8772 10730 : *indexStartupCost =
8773 10730 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
8774 10730 : *indexStartupCost += qual_arg_cost;
8775 :
8776 : /*
8777 : * To read a BRIN index there might be a bit of back and forth over
8778 : * regular pages, as revmap might point to them out of sequential order;
8779 : * calculate the total cost as reading the whole index in random order.
8780 : */
8781 10730 : *indexTotalCost = *indexStartupCost +
8782 10730 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
8783 :
8784 : /*
8785 : * Charge a small amount per range tuple which we expect to match to. This
8786 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
8787 : * will set a bit for each page in the range when we find a matching
8788 : * range, so we must multiply the charge by the number of pages in the
8789 : * range.
8790 : */
8791 10730 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
8792 10730 : statsData.pagesPerRange;
8793 :
8794 10730 : *indexPages = index->pages;
8795 10730 : }
|
