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 690790 : eqsel(PG_FUNCTION_ARGS)
232 : {
233 690790 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
234 : }
235 :
236 : /*
237 : * Common code for eqsel() and neqsel()
238 : */
239 : static double
240 736880 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
241 : {
242 736880 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
243 736880 : Oid operator = PG_GETARG_OID(1);
244 736880 : List *args = (List *) PG_GETARG_POINTER(2);
245 736880 : int varRelid = PG_GETARG_INT32(3);
246 736880 : 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 736880 : if (negate)
257 : {
258 46090 : operator = get_negator(operator);
259 46090 : 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 736880 : if (!get_restriction_variable(root, args, varRelid,
271 : &vardata, &other, &varonleft))
272 5386 : 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 731488 : if (IsA(other, Const))
280 304096 : selec = var_eq_const(&vardata, operator, collation,
281 304096 : ((Const *) other)->constvalue,
282 304096 : ((Const *) other)->constisnull,
283 : varonleft, negate);
284 : else
285 427392 : selec = var_eq_non_const(&vardata, operator, collation, other,
286 : varonleft, negate);
287 :
288 731488 : ReleaseVariableStats(vardata);
289 :
290 731488 : 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 348496 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
300 : Datum constval, bool constisnull,
301 : bool varonleft, bool negate)
302 : {
303 : double selec;
304 348496 : 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 348496 : 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 348092 : if (HeapTupleIsValid(vardata->statsTuple))
320 : {
321 : Form_pg_statistic stats;
322 :
323 261640 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
324 261640 : 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 348092 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
335 : {
336 82738 : selec = 1.0 / vardata->rel->tuples;
337 : }
338 461254 : else if (HeapTupleIsValid(vardata->statsTuple) &&
339 195900 : statistic_proc_security_check(vardata,
340 195900 : (opfuncoid = get_opcode(oproid))))
341 195900 : {
342 : AttStatsSlot sslot;
343 195900 : 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 195900 : if (get_attstatsslot(&sslot, vardata->statsTuple,
354 : STATISTIC_KIND_MCV, InvalidOid,
355 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
356 : {
357 173172 : LOCAL_FCINFO(fcinfo, 2);
358 : FmgrInfo eqproc;
359 :
360 173172 : 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 173172 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
369 : NULL, NULL);
370 173172 : fcinfo->args[0].isnull = false;
371 173172 : fcinfo->args[1].isnull = false;
372 : /* be careful to apply operator right way 'round */
373 173172 : if (varonleft)
374 173140 : fcinfo->args[1].value = constval;
375 : else
376 32 : fcinfo->args[0].value = constval;
377 :
378 2893564 : for (i = 0; i < sslot.nvalues; i++)
379 : {
380 : Datum fresult;
381 :
382 2811238 : if (varonleft)
383 2811182 : fcinfo->args[0].value = sslot.values[i];
384 : else
385 56 : fcinfo->args[1].value = sslot.values[i];
386 2811238 : fcinfo->isnull = false;
387 2811238 : fresult = FunctionCallInvoke(fcinfo);
388 2811238 : if (!fcinfo->isnull && DatumGetBool(fresult))
389 : {
390 90846 : match = true;
391 90846 : break;
392 : }
393 : }
394 : }
395 : else
396 : {
397 : /* no most-common-value info available */
398 22728 : i = 0; /* keep compiler quiet */
399 : }
400 :
401 195900 : 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 90846 : 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 105054 : double sumcommon = 0.0;
417 : double otherdistinct;
418 :
419 2451572 : for (i = 0; i < sslot.nnumbers; i++)
420 2346518 : sumcommon += sslot.numbers[i];
421 105054 : selec = 1.0 - sumcommon - nullfrac;
422 105054 : 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 105054 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
430 105054 : sslot.nnumbers;
431 105054 : if (otherdistinct > 1)
432 52402 : 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 105054 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
439 0 : selec = sslot.numbers[sslot.nnumbers - 1];
440 : }
441 :
442 195900 : 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 69454 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
452 : }
453 :
454 : /* now adjust if we wanted <> rather than = */
455 348092 : if (negate)
456 37290 : selec = 1.0 - selec - nullfrac;
457 :
458 : /* result should be in range, but make sure... */
459 348092 : CLAMP_PROBABILITY(selec);
460 :
461 348092 : 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 427392 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
471 : Node *other,
472 : bool varonleft, bool negate)
473 : {
474 : double selec;
475 427392 : double nullfrac = 0.0;
476 : bool isdefault;
477 :
478 : /*
479 : * Grab the nullfrac for use below.
480 : */
481 427392 : if (HeapTupleIsValid(vardata->statsTuple))
482 : {
483 : Form_pg_statistic stats;
484 :
485 289648 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
486 289648 : 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 427392 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
497 : {
498 160846 : selec = 1.0 / vardata->rel->tuples;
499 : }
500 266546 : 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 144868 : selec = 1.0 - nullfrac;
516 144868 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
517 144868 : if (ndistinct > 1)
518 141144 : selec /= ndistinct;
519 :
520 : /*
521 : * Cross-check: selectivity should never be estimated as more than the
522 : * most common value's.
523 : */
524 144868 : if (get_attstatsslot(&sslot, vardata->statsTuple,
525 : STATISTIC_KIND_MCV, InvalidOid,
526 : ATTSTATSSLOT_NUMBERS))
527 : {
528 125780 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
529 558 : selec = sslot.numbers[0];
530 125780 : 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 121678 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
541 : }
542 :
543 : /* now adjust if we wanted <> rather than = */
544 427392 : if (negate)
545 6472 : selec = 1.0 - selec - nullfrac;
546 :
547 : /* result should be in range, but make sure... */
548 427392 : CLAMP_PROBABILITY(selec);
549 :
550 427392 : 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 46090 : neqsel(PG_FUNCTION_ARGS)
562 : {
563 46090 : 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 369246 : 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 369246 : 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 22656 : if (vardata->var && IsA(vardata->var, Var) &&
603 17702 : ((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 1960 : if (vardata->rel->pages == 0)
614 0 : return 1.0;
615 :
616 1960 : itemptr = (ItemPointer) DatumGetPointer(constval);
617 1960 : 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 1960 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
627 :
628 : /* If target is the last page, use half the density. */
629 1960 : 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 1960 : if (density > 0.0)
641 : {
642 1960 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
643 :
644 1960 : 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 1960 : 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 1960 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
662 1844 : selec -= (1.0 / vardata->rel->tuples);
663 :
664 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
665 1960 : if (isgt)
666 1838 : selec = 1.0 - selec;
667 :
668 1960 : CLAMP_PROBABILITY(selec);
669 1960 : return selec;
670 : }
671 :
672 : /* no stats available, so default result */
673 20696 : return DEFAULT_INEQ_SEL;
674 : }
675 346590 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
676 :
677 346590 : 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 346590 : 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 346590 : 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 346590 : selec = 1.0 - stats->stanullfrac - sumcommon;
703 :
704 346590 : if (hist_selec >= 0.0)
705 214194 : 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 132396 : selec *= 0.5;
713 : }
714 :
715 346590 : selec += mcv_selec;
716 :
717 : /* result should be in range, but make sure... */
718 346590 : CLAMP_PROBABILITY(selec);
719 :
720 346590 : 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 352764 : 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 352764 : mcv_selec = 0.0;
746 352764 : sumcommon = 0.0;
747 :
748 703072 : if (HeapTupleIsValid(vardata->statsTuple) &&
749 700286 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
750 349978 : get_attstatsslot(&sslot, vardata->statsTuple,
751 : STATISTIC_KIND_MCV, InvalidOid,
752 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
753 : {
754 193656 : 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 193656 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
764 : NULL, NULL);
765 193656 : fcinfo->args[0].isnull = false;
766 193656 : fcinfo->args[1].isnull = false;
767 : /* be careful to apply operator right way 'round */
768 193656 : if (varonleft)
769 193656 : fcinfo->args[1].value = constval;
770 : else
771 0 : fcinfo->args[0].value = constval;
772 :
773 4543348 : for (i = 0; i < sslot.nvalues; i++)
774 : {
775 : Datum fresult;
776 :
777 4349692 : if (varonleft)
778 4349692 : fcinfo->args[0].value = sslot.values[i];
779 : else
780 0 : fcinfo->args[1].value = sslot.values[i];
781 4349692 : fcinfo->isnull = false;
782 4349692 : fresult = FunctionCallInvoke(fcinfo);
783 4349692 : if (!fcinfo->isnull && DatumGetBool(fresult))
784 1654012 : mcv_selec += sslot.numbers[i];
785 4349692 : sumcommon += sslot.numbers[i];
786 : }
787 193656 : free_attstatsslot(&sslot);
788 : }
789 :
790 352764 : *sumcommonp = sumcommon;
791 352764 : 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 6174 : 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 9892 : 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 141818 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
872 : {
873 : Datum fresult;
874 :
875 140088 : if (varonleft)
876 140088 : fcinfo->args[0].value = sslot.values[i];
877 : else
878 0 : fcinfo->args[1].value = sslot.values[i];
879 140088 : fcinfo->isnull = false;
880 140088 : fresult = FunctionCallInvoke(fcinfo);
881 140088 : if (!fcinfo->isnull && DatumGetBool(fresult))
882 7386 : 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 2556 : *hist_size = 0;
893 2556 : result = -1;
894 : }
895 :
896 6174 : 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 1130 : 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 1130 : 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 1130 : if (IsA(other, Const) &&
940 1130 : ((Const *) other)->constisnull)
941 : {
942 0 : ReleaseVariableStats(vardata);
943 0 : return 0.0;
944 : }
945 :
946 1130 : if (IsA(other, Const))
947 : {
948 : /* Variable is being compared to a known non-null constant */
949 1130 : Datum constval = ((Const *) other)->constvalue;
950 : FmgrInfo opproc;
951 : double mcvsum;
952 : double mcvsel;
953 : double nullfrac;
954 : int hist_size;
955 :
956 1130 : fmgr_info(get_opcode(oproid), &opproc);
957 :
958 : /*
959 : * Calculate the selectivity for the column's most common values.
960 : */
961 1130 : 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 1130 : selec = histogram_selectivity(&vardata, &opproc, collation,
972 : constval, varonleft,
973 : 10, 1, &hist_size);
974 1130 : if (selec < 0)
975 : {
976 : /* Nope, fall back on default */
977 1130 : 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 1130 : if (selec < 0.0001)
994 0 : selec = 0.0001;
995 1130 : else if (selec > 0.9999)
996 0 : selec = 0.9999;
997 :
998 : /* Don't forget to account for nulls. */
999 1130 : if (HeapTupleIsValid(vardata.statsTuple))
1000 84 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
1001 : else
1002 1046 : 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 1130 : selec *= 1.0 - nullfrac - mcvsum;
1010 1130 : 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 1130 : ReleaseVariableStats(vardata);
1019 :
1020 : /* result should be in range, but make sure... */
1021 1130 : CLAMP_PROBABILITY(selec);
1022 :
1023 1130 : 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 351786 : 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 351786 : 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 702878 : if (HeapTupleIsValid(vardata->statsTuple) &&
1069 701860 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1070 350768 : get_attstatsslot(&sslot, vardata->statsTuple,
1071 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1072 : ATTSTATSSLOT_VALUES))
1073 : {
1074 218694 : if (sslot.nvalues > 1 &&
1075 437312 : sslot.stacoll == collation &&
1076 218618 : comparison_ops_are_compatible(sslot.staop, opoid))
1077 218510 : {
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 218510 : int lobound = 0; /* first possible slot to search */
1098 218510 : int hibound = sslot.nvalues; /* last+1 slot to search */
1099 218510 : 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 218510 : if (sslot.nvalues == 2)
1108 3790 : have_end = get_actual_variable_range(root,
1109 : vardata,
1110 : sslot.staop,
1111 : collation,
1112 : &sslot.values[0],
1113 3790 : &sslot.values[1]);
1114 :
1115 1441664 : while (lobound < hibound)
1116 : {
1117 1223154 : 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 1223154 : if (probe == 0 && sslot.nvalues > 2)
1126 108406 : have_end = get_actual_variable_range(root,
1127 : vardata,
1128 : sslot.staop,
1129 : collation,
1130 : &sslot.values[0],
1131 : NULL);
1132 1114748 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1133 73178 : have_end = get_actual_variable_range(root,
1134 : vardata,
1135 : sslot.staop,
1136 : collation,
1137 : NULL,
1138 73178 : &sslot.values[probe]);
1139 :
1140 1223154 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1141 : collation,
1142 1223154 : sslot.values[probe],
1143 : constval));
1144 1223154 : if (isgt)
1145 68240 : ltcmp = !ltcmp;
1146 1223154 : if (ltcmp)
1147 455688 : lobound = probe + 1;
1148 : else
1149 767466 : hibound = probe;
1150 : }
1151 :
1152 218510 : 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 94154 : histfrac = 0.0;
1163 : }
1164 124356 : else if (lobound >= sslot.nvalues)
1165 : {
1166 : /*
1167 : * Inverse case: constant is above upper histogram boundary.
1168 : */
1169 37286 : histfrac = 1.0;
1170 : }
1171 : else
1172 : {
1173 : /* We have values[i-1] <= constant <= values[i]. */
1174 87070 : int i = lobound;
1175 87070 : 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 87070 : if (i == 1 || isgt == iseq)
1198 : {
1199 : double otherdistinct;
1200 : bool isdefault;
1201 : AttStatsSlot mcvslot;
1202 :
1203 : /* Get estimated number of distinct values */
1204 37642 : otherdistinct = get_variable_numdistinct(vardata,
1205 : &isdefault);
1206 :
1207 : /* Subtract off the number of known MCVs */
1208 37642 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1209 : STATISTIC_KIND_MCV, InvalidOid,
1210 : ATTSTATSSLOT_NUMBERS))
1211 : {
1212 3764 : otherdistinct -= mcvslot.nnumbers;
1213 3764 : free_attstatsslot(&mcvslot);
1214 : }
1215 :
1216 : /* If result doesn't seem sane, leave eq_selec at 0 */
1217 37642 : if (otherdistinct > 1)
1218 37600 : 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 87070 : if (convert_to_scalar(constval, consttype, collation,
1227 : &val,
1228 87070 : sslot.values[i - 1], sslot.values[i],
1229 : vardata->vartype,
1230 : &low, &high))
1231 : {
1232 87070 : if (high <= low)
1233 : {
1234 : /* cope if bin boundaries appear identical */
1235 0 : binfrac = 0.5;
1236 : }
1237 87070 : else if (val <= low)
1238 20552 : binfrac = 0.0;
1239 66518 : else if (val >= high)
1240 2672 : binfrac = 1.0;
1241 : else
1242 : {
1243 63846 : 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 63846 : if (isnan(binfrac) ||
1252 63846 : 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 87070 : histfrac = (double) (i - 1) + binfrac;
1275 87070 : 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 87070 : if (i == 1)
1310 16582 : 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 87070 : if (isgt == iseq)
1318 28184 : 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 218510 : 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 218510 : if (have_end)
1337 123660 : CLAMP_PROBABILITY(hist_selec);
1338 : else
1339 : {
1340 94850 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1341 :
1342 94850 : if (hist_selec < cutoff)
1343 33906 : hist_selec = cutoff;
1344 60944 : else if (hist_selec > 1.0 - cutoff)
1345 22426 : 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 962484 : for (int i = 0; i < sslot.nvalues; i++)
1366 : {
1367 : Datum fresult;
1368 :
1369 962300 : fcinfo->args[0].value = sslot.values[i];
1370 962300 : fcinfo->isnull = false;
1371 962300 : fresult = FunctionCallInvoke(fcinfo);
1372 962300 : if (!fcinfo->isnull && DatumGetBool(fresult))
1373 2216 : 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 12 : hist_selec = cutoff;
1388 172 : else if (hist_selec > 1.0 - cutoff)
1389 12 : hist_selec = 1.0 - cutoff;
1390 : }
1391 : }
1392 :
1393 218694 : free_attstatsslot(&sslot);
1394 : }
1395 :
1396 351786 : return hist_selec;
1397 : }
1398 :
1399 : /*
1400 : * Common wrapper function for the selectivity estimators that simply
1401 : * invoke scalarineqsel().
1402 : */
1403 : static Datum
1404 45002 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1405 : {
1406 45002 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1407 45002 : Oid operator = PG_GETARG_OID(1);
1408 45002 : List *args = (List *) PG_GETARG_POINTER(2);
1409 45002 : int varRelid = PG_GETARG_INT32(3);
1410 45002 : 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 45002 : 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 44352 : if (!IsA(other, Const))
1430 : {
1431 2800 : ReleaseVariableStats(vardata);
1432 2800 : 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 41552 : if (((Const *) other)->constisnull)
1440 : {
1441 66 : ReleaseVariableStats(vardata);
1442 66 : PG_RETURN_FLOAT8(0.0);
1443 : }
1444 41486 : constval = ((Const *) other)->constvalue;
1445 41486 : consttype = ((Const *) other)->consttype;
1446 :
1447 : /*
1448 : * Force the var to be on the left to simplify logic in scalarineqsel.
1449 : */
1450 41486 : if (!varonleft)
1451 : {
1452 384 : operator = get_commutator(operator);
1453 384 : 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 384 : isgt = !isgt;
1460 : }
1461 :
1462 : /* The rest of the work is done by scalarineqsel(). */
1463 41486 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1464 : &vardata, constval, consttype);
1465 :
1466 41486 : ReleaseVariableStats(vardata);
1467 :
1468 41486 : PG_RETURN_FLOAT8((float8) selec);
1469 : }
1470 :
1471 : /*
1472 : * scalarltsel - Selectivity of "<" for scalars.
1473 : */
1474 : Datum
1475 15042 : scalarltsel(PG_FUNCTION_ARGS)
1476 : {
1477 15042 : return scalarineqsel_wrapper(fcinfo, false, false);
1478 : }
1479 :
1480 : /*
1481 : * scalarlesel - Selectivity of "<=" for scalars.
1482 : */
1483 : Datum
1484 4564 : scalarlesel(PG_FUNCTION_ARGS)
1485 : {
1486 4564 : return scalarineqsel_wrapper(fcinfo, false, true);
1487 : }
1488 :
1489 : /*
1490 : * scalargtsel - Selectivity of ">" for scalars.
1491 : */
1492 : Datum
1493 15402 : scalargtsel(PG_FUNCTION_ARGS)
1494 : {
1495 15402 : return scalarineqsel_wrapper(fcinfo, true, false);
1496 : }
1497 :
1498 : /*
1499 : * scalargesel - Selectivity of ">=" for scalars.
1500 : */
1501 : Datum
1502 9994 : scalargesel(PG_FUNCTION_ARGS)
1503 : {
1504 9994 : 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 56274 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1517 : {
1518 : VariableStatData vardata;
1519 : double selec;
1520 :
1521 56274 : examine_variable(root, arg, varRelid, &vardata);
1522 56274 : 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 35594 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1529 : BoolGetDatum(true), false, true, false);
1530 : }
1531 20680 : else if (is_funcclause(arg))
1532 : {
1533 : /*
1534 : * If we have no stats and it's a function call, estimate 0.3333333.
1535 : * This seems a pretty unprincipled choice, but Postgres has been
1536 : * using that estimate for function calls since 1992. The hoariness
1537 : * of this behavior suggests that we should not be in too much hurry
1538 : * to use another value.
1539 : */
1540 12182 : selec = 0.3333333;
1541 : }
1542 : else
1543 : {
1544 : /* Otherwise, the default estimate is 0.5 */
1545 8498 : selec = 0.5;
1546 : }
1547 56274 : ReleaseVariableStats(vardata);
1548 56274 : return selec;
1549 : }
1550 :
1551 : /*
1552 : * booltestsel - Selectivity of BooleanTest Node.
1553 : */
1554 : Selectivity
1555 886 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1556 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1557 : {
1558 : VariableStatData vardata;
1559 : double selec;
1560 :
1561 886 : examine_variable(root, arg, varRelid, &vardata);
1562 :
1563 886 : if (HeapTupleIsValid(vardata.statsTuple))
1564 : {
1565 : Form_pg_statistic stats;
1566 : double freq_null;
1567 : AttStatsSlot sslot;
1568 :
1569 0 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1570 0 : freq_null = stats->stanullfrac;
1571 :
1572 0 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1573 : STATISTIC_KIND_MCV, InvalidOid,
1574 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1575 0 : && sslot.nnumbers > 0)
1576 0 : {
1577 : double freq_true;
1578 : double freq_false;
1579 :
1580 : /*
1581 : * Get first MCV frequency and derive frequency for true.
1582 : */
1583 0 : if (DatumGetBool(sslot.values[0]))
1584 0 : freq_true = sslot.numbers[0];
1585 : else
1586 0 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1587 :
1588 : /*
1589 : * Next derive frequency for false. Then use these as appropriate
1590 : * to derive frequency for each case.
1591 : */
1592 0 : freq_false = 1.0 - freq_true - freq_null;
1593 :
1594 0 : switch (booltesttype)
1595 : {
1596 0 : case IS_UNKNOWN:
1597 : /* select only NULL values */
1598 0 : selec = freq_null;
1599 0 : break;
1600 0 : case IS_NOT_UNKNOWN:
1601 : /* select non-NULL values */
1602 0 : selec = 1.0 - freq_null;
1603 0 : break;
1604 0 : case IS_TRUE:
1605 : /* select only TRUE values */
1606 0 : selec = freq_true;
1607 0 : break;
1608 0 : case IS_NOT_TRUE:
1609 : /* select non-TRUE values */
1610 0 : selec = 1.0 - freq_true;
1611 0 : break;
1612 0 : case IS_FALSE:
1613 : /* select only FALSE values */
1614 0 : selec = freq_false;
1615 0 : break;
1616 0 : case IS_NOT_FALSE:
1617 : /* select non-FALSE values */
1618 0 : selec = 1.0 - freq_false;
1619 0 : break;
1620 0 : default:
1621 0 : elog(ERROR, "unrecognized booltesttype: %d",
1622 : (int) booltesttype);
1623 : selec = 0.0; /* Keep compiler quiet */
1624 : break;
1625 : }
1626 :
1627 0 : free_attstatsslot(&sslot);
1628 : }
1629 : else
1630 : {
1631 : /*
1632 : * No most-common-value info available. Still have null fraction
1633 : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1634 : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1635 : */
1636 0 : switch (booltesttype)
1637 : {
1638 0 : case IS_UNKNOWN:
1639 : /* select only NULL values */
1640 0 : selec = freq_null;
1641 0 : break;
1642 0 : case IS_NOT_UNKNOWN:
1643 : /* select non-NULL values */
1644 0 : selec = 1.0 - freq_null;
1645 0 : break;
1646 0 : case IS_TRUE:
1647 : case IS_FALSE:
1648 : /* Assume we select half of the non-NULL values */
1649 0 : selec = (1.0 - freq_null) / 2.0;
1650 0 : break;
1651 0 : case IS_NOT_TRUE:
1652 : case IS_NOT_FALSE:
1653 : /* Assume we select NULLs plus half of the non-NULLs */
1654 : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1655 0 : selec = (freq_null + 1.0) / 2.0;
1656 0 : break;
1657 0 : default:
1658 0 : elog(ERROR, "unrecognized booltesttype: %d",
1659 : (int) booltesttype);
1660 : selec = 0.0; /* Keep compiler quiet */
1661 : break;
1662 : }
1663 : }
1664 : }
1665 : else
1666 : {
1667 : /*
1668 : * If we can't get variable statistics for the argument, perhaps
1669 : * clause_selectivity can do something with it. We ignore the
1670 : * possibility of a NULL value when using clause_selectivity, and just
1671 : * assume the value is either TRUE or FALSE.
1672 : */
1673 886 : switch (booltesttype)
1674 : {
1675 48 : case IS_UNKNOWN:
1676 48 : selec = DEFAULT_UNK_SEL;
1677 48 : break;
1678 108 : case IS_NOT_UNKNOWN:
1679 108 : selec = DEFAULT_NOT_UNK_SEL;
1680 108 : break;
1681 252 : case IS_TRUE:
1682 : case IS_NOT_FALSE:
1683 252 : selec = (double) clause_selectivity(root, arg,
1684 : varRelid,
1685 : jointype, sjinfo);
1686 252 : break;
1687 478 : case IS_FALSE:
1688 : case IS_NOT_TRUE:
1689 478 : selec = 1.0 - (double) clause_selectivity(root, arg,
1690 : varRelid,
1691 : jointype, sjinfo);
1692 478 : break;
1693 0 : default:
1694 0 : elog(ERROR, "unrecognized booltesttype: %d",
1695 : (int) booltesttype);
1696 : selec = 0.0; /* Keep compiler quiet */
1697 : break;
1698 : }
1699 : }
1700 :
1701 886 : ReleaseVariableStats(vardata);
1702 :
1703 : /* result should be in range, but make sure... */
1704 886 : CLAMP_PROBABILITY(selec);
1705 :
1706 886 : return (Selectivity) selec;
1707 : }
1708 :
1709 : /*
1710 : * nulltestsel - Selectivity of NullTest Node.
1711 : */
1712 : Selectivity
1713 17442 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1714 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1715 : {
1716 : VariableStatData vardata;
1717 : double selec;
1718 :
1719 17442 : examine_variable(root, arg, varRelid, &vardata);
1720 :
1721 17442 : if (HeapTupleIsValid(vardata.statsTuple))
1722 : {
1723 : Form_pg_statistic stats;
1724 : double freq_null;
1725 :
1726 9820 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1727 9820 : freq_null = stats->stanullfrac;
1728 :
1729 9820 : switch (nulltesttype)
1730 : {
1731 7224 : case IS_NULL:
1732 :
1733 : /*
1734 : * Use freq_null directly.
1735 : */
1736 7224 : selec = freq_null;
1737 7224 : break;
1738 2596 : case IS_NOT_NULL:
1739 :
1740 : /*
1741 : * Select not unknown (not null) values. Calculate from
1742 : * freq_null.
1743 : */
1744 2596 : selec = 1.0 - freq_null;
1745 2596 : break;
1746 0 : default:
1747 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1748 : (int) nulltesttype);
1749 : return (Selectivity) 0; /* keep compiler quiet */
1750 : }
1751 : }
1752 7622 : else if (vardata.var && IsA(vardata.var, Var) &&
1753 6870 : ((Var *) vardata.var)->varattno < 0)
1754 : {
1755 : /*
1756 : * There are no stats for system columns, but we know they are never
1757 : * NULL.
1758 : */
1759 68 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1760 : }
1761 : else
1762 : {
1763 : /*
1764 : * No ANALYZE stats available, so make a guess
1765 : */
1766 7554 : switch (nulltesttype)
1767 : {
1768 2108 : case IS_NULL:
1769 2108 : selec = DEFAULT_UNK_SEL;
1770 2108 : break;
1771 5446 : case IS_NOT_NULL:
1772 5446 : selec = DEFAULT_NOT_UNK_SEL;
1773 5446 : break;
1774 0 : default:
1775 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1776 : (int) nulltesttype);
1777 : return (Selectivity) 0; /* keep compiler quiet */
1778 : }
1779 : }
1780 :
1781 17442 : ReleaseVariableStats(vardata);
1782 :
1783 : /* result should be in range, but make sure... */
1784 17442 : CLAMP_PROBABILITY(selec);
1785 :
1786 17442 : return (Selectivity) selec;
1787 : }
1788 :
1789 : /*
1790 : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1791 : *
1792 : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1793 : * but it seems possible that RelabelType might show up. Also, the planner
1794 : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1795 : * so we need to be ready to deal with more than one level.
1796 : */
1797 : static Node *
1798 126916 : strip_array_coercion(Node *node)
1799 : {
1800 : for (;;)
1801 : {
1802 126998 : if (node && IsA(node, ArrayCoerceExpr))
1803 82 : {
1804 2946 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1805 :
1806 : /*
1807 : * If the per-element expression is just a RelabelType on top of
1808 : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1809 : */
1810 2946 : if (IsA(acoerce->elemexpr, RelabelType) &&
1811 82 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1812 82 : node = (Node *) acoerce->arg;
1813 : else
1814 : break;
1815 : }
1816 124052 : else if (node && IsA(node, RelabelType))
1817 : {
1818 : /* We don't really expect this case, but may as well cope */
1819 0 : node = (Node *) ((RelabelType *) node)->arg;
1820 : }
1821 : else
1822 : break;
1823 : }
1824 126916 : return node;
1825 : }
1826 :
1827 : /*
1828 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1829 : */
1830 : Selectivity
1831 22168 : scalararraysel(PlannerInfo *root,
1832 : ScalarArrayOpExpr *clause,
1833 : bool is_join_clause,
1834 : int varRelid,
1835 : JoinType jointype,
1836 : SpecialJoinInfo *sjinfo)
1837 : {
1838 22168 : Oid operator = clause->opno;
1839 22168 : bool useOr = clause->useOr;
1840 22168 : bool isEquality = false;
1841 22168 : bool isInequality = false;
1842 : Node *leftop;
1843 : Node *rightop;
1844 : Oid nominal_element_type;
1845 : Oid nominal_element_collation;
1846 : TypeCacheEntry *typentry;
1847 : RegProcedure oprsel;
1848 : FmgrInfo oprselproc;
1849 : Selectivity s1;
1850 : Selectivity s1disjoint;
1851 :
1852 : /* First, deconstruct the expression */
1853 : Assert(list_length(clause->args) == 2);
1854 22168 : leftop = (Node *) linitial(clause->args);
1855 22168 : rightop = (Node *) lsecond(clause->args);
1856 :
1857 : /* aggressively reduce both sides to constants */
1858 22168 : leftop = estimate_expression_value(root, leftop);
1859 22168 : rightop = estimate_expression_value(root, rightop);
1860 :
1861 : /* get nominal (after relabeling) element type of rightop */
1862 22168 : nominal_element_type = get_base_element_type(exprType(rightop));
1863 22168 : if (!OidIsValid(nominal_element_type))
1864 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1865 : /* get nominal collation, too, for generating constants */
1866 22168 : nominal_element_collation = exprCollation(rightop);
1867 :
1868 : /* look through any binary-compatible relabeling of rightop */
1869 22168 : rightop = strip_array_coercion(rightop);
1870 :
1871 : /*
1872 : * Detect whether the operator is the default equality or inequality
1873 : * operator of the array element type.
1874 : */
1875 22168 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1876 22168 : if (OidIsValid(typentry->eq_opr))
1877 : {
1878 22164 : if (operator == typentry->eq_opr)
1879 18846 : isEquality = true;
1880 3318 : else if (get_negator(operator) == typentry->eq_opr)
1881 2752 : isInequality = true;
1882 : }
1883 :
1884 : /*
1885 : * If it is equality or inequality, we might be able to estimate this as a
1886 : * form of array containment; for instance "const = ANY(column)" can be
1887 : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1888 : * that, and returns the selectivity estimate if successful, or -1 if not.
1889 : */
1890 22168 : if ((isEquality || isInequality) && !is_join_clause)
1891 : {
1892 21598 : s1 = scalararraysel_containment(root, leftop, rightop,
1893 : nominal_element_type,
1894 : isEquality, useOr, varRelid);
1895 21598 : if (s1 >= 0.0)
1896 118 : return s1;
1897 : }
1898 :
1899 : /*
1900 : * Look up the underlying operator's selectivity estimator. Punt if it
1901 : * hasn't got one.
1902 : */
1903 22050 : if (is_join_clause)
1904 0 : oprsel = get_oprjoin(operator);
1905 : else
1906 22050 : oprsel = get_oprrest(operator);
1907 22050 : if (!oprsel)
1908 4 : return (Selectivity) 0.5;
1909 22046 : fmgr_info(oprsel, &oprselproc);
1910 :
1911 : /*
1912 : * In the array-containment check above, we must only believe that an
1913 : * operator is equality or inequality if it is the default btree equality
1914 : * operator (or its negator) for the element type, since those are the
1915 : * operators that array containment will use. But in what follows, we can
1916 : * be a little laxer, and also believe that any operators using eqsel() or
1917 : * neqsel() as selectivity estimator act like equality or inequality.
1918 : */
1919 22046 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1920 18924 : isEquality = true;
1921 3122 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1922 2642 : isInequality = true;
1923 :
1924 : /*
1925 : * We consider three cases:
1926 : *
1927 : * 1. rightop is an Array constant: deconstruct the array, apply the
1928 : * operator's selectivity function for each array element, and merge the
1929 : * results in the same way that clausesel.c does for AND/OR combinations.
1930 : *
1931 : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
1932 : * function for each element of the ARRAY[] construct, and merge.
1933 : *
1934 : * 3. otherwise, make a guess ...
1935 : */
1936 22046 : if (rightop && IsA(rightop, Const))
1937 17886 : {
1938 17916 : Datum arraydatum = ((Const *) rightop)->constvalue;
1939 17916 : bool arrayisnull = ((Const *) rightop)->constisnull;
1940 : ArrayType *arrayval;
1941 : int16 elmlen;
1942 : bool elmbyval;
1943 : char elmalign;
1944 : int num_elems;
1945 : Datum *elem_values;
1946 : bool *elem_nulls;
1947 : int i;
1948 :
1949 17916 : if (arrayisnull) /* qual can't succeed if null array */
1950 30 : return (Selectivity) 0.0;
1951 17886 : arrayval = DatumGetArrayTypeP(arraydatum);
1952 17886 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
1953 : &elmlen, &elmbyval, &elmalign);
1954 17886 : deconstruct_array(arrayval,
1955 : ARR_ELEMTYPE(arrayval),
1956 : elmlen, elmbyval, elmalign,
1957 : &elem_values, &elem_nulls, &num_elems);
1958 :
1959 : /*
1960 : * For generic operators, we assume the probability of success is
1961 : * independent for each array element. But for "= ANY" or "<> ALL",
1962 : * if the array elements are distinct (which'd typically be the case)
1963 : * then the probabilities are disjoint, and we should just sum them.
1964 : *
1965 : * If we were being really tense we would try to confirm that the
1966 : * elements are all distinct, but that would be expensive and it
1967 : * doesn't seem to be worth the cycles; it would amount to penalizing
1968 : * well-written queries in favor of poorly-written ones. However, we
1969 : * do protect ourselves a little bit by checking whether the
1970 : * disjointness assumption leads to an impossible (out of range)
1971 : * probability; if so, we fall back to the normal calculation.
1972 : */
1973 17886 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
1974 :
1975 77216 : for (i = 0; i < num_elems; i++)
1976 : {
1977 : List *args;
1978 : Selectivity s2;
1979 :
1980 59330 : args = list_make2(leftop,
1981 : makeConst(nominal_element_type,
1982 : -1,
1983 : nominal_element_collation,
1984 : elmlen,
1985 : elem_values[i],
1986 : elem_nulls[i],
1987 : elmbyval));
1988 59330 : if (is_join_clause)
1989 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1990 : clause->inputcollid,
1991 : PointerGetDatum(root),
1992 : ObjectIdGetDatum(operator),
1993 : PointerGetDatum(args),
1994 : Int16GetDatum(jointype),
1995 : PointerGetDatum(sjinfo)));
1996 : else
1997 59330 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1998 : clause->inputcollid,
1999 : PointerGetDatum(root),
2000 : ObjectIdGetDatum(operator),
2001 : PointerGetDatum(args),
2002 : Int32GetDatum(varRelid)));
2003 :
2004 59330 : if (useOr)
2005 : {
2006 50852 : s1 = s1 + s2 - s1 * s2;
2007 50852 : if (isEquality)
2008 49808 : s1disjoint += s2;
2009 : }
2010 : else
2011 : {
2012 8478 : s1 = s1 * s2;
2013 8478 : if (isInequality)
2014 8166 : s1disjoint += s2 - 1.0;
2015 : }
2016 : }
2017 :
2018 : /* accept disjoint-probability estimate if in range */
2019 17886 : if ((useOr ? isEquality : isInequality) &&
2020 17224 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2021 17194 : s1 = s1disjoint;
2022 : }
2023 4130 : else if (rightop && IsA(rightop, ArrayExpr) &&
2024 294 : !((ArrayExpr *) rightop)->multidims)
2025 294 : {
2026 294 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2027 : int16 elmlen;
2028 : bool elmbyval;
2029 : ListCell *l;
2030 :
2031 294 : get_typlenbyval(arrayexpr->element_typeid,
2032 : &elmlen, &elmbyval);
2033 :
2034 : /*
2035 : * We use the assumption of disjoint probabilities here too, although
2036 : * the odds of equal array elements are rather higher if the elements
2037 : * are not all constants (which they won't be, else constant folding
2038 : * would have reduced the ArrayExpr to a Const). In this path it's
2039 : * critical to have the sanity check on the s1disjoint estimate.
2040 : */
2041 294 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2042 :
2043 1060 : foreach(l, arrayexpr->elements)
2044 : {
2045 766 : Node *elem = (Node *) lfirst(l);
2046 : List *args;
2047 : Selectivity s2;
2048 :
2049 : /*
2050 : * Theoretically, if elem isn't of nominal_element_type we should
2051 : * insert a RelabelType, but it seems unlikely that any operator
2052 : * estimation function would really care ...
2053 : */
2054 766 : args = list_make2(leftop, elem);
2055 766 : if (is_join_clause)
2056 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2057 : clause->inputcollid,
2058 : PointerGetDatum(root),
2059 : ObjectIdGetDatum(operator),
2060 : PointerGetDatum(args),
2061 : Int16GetDatum(jointype),
2062 : PointerGetDatum(sjinfo)));
2063 : else
2064 766 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2065 : clause->inputcollid,
2066 : PointerGetDatum(root),
2067 : ObjectIdGetDatum(operator),
2068 : PointerGetDatum(args),
2069 : Int32GetDatum(varRelid)));
2070 :
2071 766 : if (useOr)
2072 : {
2073 766 : s1 = s1 + s2 - s1 * s2;
2074 766 : if (isEquality)
2075 766 : s1disjoint += s2;
2076 : }
2077 : else
2078 : {
2079 0 : s1 = s1 * s2;
2080 0 : if (isInequality)
2081 0 : s1disjoint += s2 - 1.0;
2082 : }
2083 : }
2084 :
2085 : /* accept disjoint-probability estimate if in range */
2086 294 : if ((useOr ? isEquality : isInequality) &&
2087 294 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2088 294 : s1 = s1disjoint;
2089 : }
2090 : else
2091 : {
2092 : CaseTestExpr *dummyexpr;
2093 : List *args;
2094 : Selectivity s2;
2095 : int i;
2096 :
2097 : /*
2098 : * We need a dummy rightop to pass to the operator selectivity
2099 : * routine. It can be pretty much anything that doesn't look like a
2100 : * constant; CaseTestExpr is a convenient choice.
2101 : */
2102 3836 : dummyexpr = makeNode(CaseTestExpr);
2103 3836 : dummyexpr->typeId = nominal_element_type;
2104 3836 : dummyexpr->typeMod = -1;
2105 3836 : dummyexpr->collation = clause->inputcollid;
2106 3836 : args = list_make2(leftop, dummyexpr);
2107 3836 : if (is_join_clause)
2108 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2109 : clause->inputcollid,
2110 : PointerGetDatum(root),
2111 : ObjectIdGetDatum(operator),
2112 : PointerGetDatum(args),
2113 : Int16GetDatum(jointype),
2114 : PointerGetDatum(sjinfo)));
2115 : else
2116 3836 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2117 : clause->inputcollid,
2118 : PointerGetDatum(root),
2119 : ObjectIdGetDatum(operator),
2120 : PointerGetDatum(args),
2121 : Int32GetDatum(varRelid)));
2122 3836 : s1 = useOr ? 0.0 : 1.0;
2123 :
2124 : /*
2125 : * Arbitrarily assume 10 elements in the eventual array value (see
2126 : * also estimate_array_length). We don't risk an assumption of
2127 : * disjoint probabilities here.
2128 : */
2129 42196 : for (i = 0; i < 10; i++)
2130 : {
2131 38360 : if (useOr)
2132 38360 : s1 = s1 + s2 - s1 * s2;
2133 : else
2134 0 : s1 = s1 * s2;
2135 : }
2136 : }
2137 :
2138 : /* result should be in range, but make sure... */
2139 22016 : CLAMP_PROBABILITY(s1);
2140 :
2141 22016 : return s1;
2142 : }
2143 :
2144 : /*
2145 : * Estimate number of elements in the array yielded by an expression.
2146 : *
2147 : * Note: the result is integral, but we use "double" to avoid overflow
2148 : * concerns. Most callers will use it in double-type expressions anyway.
2149 : *
2150 : * Note: in some code paths root can be passed as NULL, resulting in
2151 : * slightly worse estimates.
2152 : */
2153 : double
2154 104748 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2155 : {
2156 : /* look through any binary-compatible relabeling of arrayexpr */
2157 104748 : arrayexpr = strip_array_coercion(arrayexpr);
2158 :
2159 104748 : if (arrayexpr && IsA(arrayexpr, Const))
2160 : {
2161 46958 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2162 46958 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2163 : ArrayType *arrayval;
2164 :
2165 46958 : if (arrayisnull)
2166 90 : return 0;
2167 46868 : arrayval = DatumGetArrayTypeP(arraydatum);
2168 46868 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2169 : }
2170 57790 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2171 498 : !((ArrayExpr *) arrayexpr)->multidims)
2172 : {
2173 498 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2174 : }
2175 57292 : else if (arrayexpr && root)
2176 : {
2177 : /* See if we can find any statistics about it */
2178 : VariableStatData vardata;
2179 : AttStatsSlot sslot;
2180 57268 : double nelem = 0;
2181 :
2182 57268 : examine_variable(root, arrayexpr, 0, &vardata);
2183 57268 : if (HeapTupleIsValid(vardata.statsTuple))
2184 : {
2185 : /*
2186 : * Found stats, so use the average element count, which is stored
2187 : * in the last stanumbers element of the DECHIST statistics.
2188 : * Actually that is the average count of *distinct* elements;
2189 : * perhaps we should scale it up somewhat?
2190 : */
2191 8336 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2192 : STATISTIC_KIND_DECHIST, InvalidOid,
2193 : ATTSTATSSLOT_NUMBERS))
2194 : {
2195 8222 : if (sslot.nnumbers > 0)
2196 8222 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2197 8222 : free_attstatsslot(&sslot);
2198 : }
2199 : }
2200 57268 : ReleaseVariableStats(vardata);
2201 :
2202 57268 : if (nelem > 0)
2203 8222 : return nelem;
2204 : }
2205 :
2206 : /* Else use a default guess --- this should match scalararraysel */
2207 49070 : return 10;
2208 : }
2209 :
2210 : /*
2211 : * rowcomparesel - Selectivity of RowCompareExpr Node.
2212 : *
2213 : * We estimate RowCompare selectivity by considering just the first (high
2214 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2215 : * this estimate could be refined by considering additional columns, it
2216 : * seems unlikely that we could do a lot better without multi-column
2217 : * statistics.
2218 : */
2219 : Selectivity
2220 252 : rowcomparesel(PlannerInfo *root,
2221 : RowCompareExpr *clause,
2222 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2223 : {
2224 : Selectivity s1;
2225 252 : Oid opno = linitial_oid(clause->opnos);
2226 252 : Oid inputcollid = linitial_oid(clause->inputcollids);
2227 : List *opargs;
2228 : bool is_join_clause;
2229 :
2230 : /* Build equivalent arg list for single operator */
2231 252 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2232 :
2233 : /*
2234 : * Decide if it's a join clause. This should match clausesel.c's
2235 : * treat_as_join_clause(), except that we intentionally consider only the
2236 : * leading columns and not the rest of the clause.
2237 : */
2238 252 : if (varRelid != 0)
2239 : {
2240 : /*
2241 : * Caller is forcing restriction mode (eg, because we are examining an
2242 : * inner indexscan qual).
2243 : */
2244 54 : is_join_clause = false;
2245 : }
2246 198 : else if (sjinfo == NULL)
2247 : {
2248 : /*
2249 : * It must be a restriction clause, since it's being evaluated at a
2250 : * scan node.
2251 : */
2252 186 : is_join_clause = false;
2253 : }
2254 : else
2255 : {
2256 : /*
2257 : * Otherwise, it's a join if there's more than one base relation used.
2258 : */
2259 12 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2260 : }
2261 :
2262 252 : if (is_join_clause)
2263 : {
2264 : /* Estimate selectivity for a join clause. */
2265 12 : s1 = join_selectivity(root, opno,
2266 : opargs,
2267 : inputcollid,
2268 : jointype,
2269 : sjinfo);
2270 : }
2271 : else
2272 : {
2273 : /* Estimate selectivity for a restriction clause. */
2274 240 : s1 = restriction_selectivity(root, opno,
2275 : opargs,
2276 : inputcollid,
2277 : varRelid);
2278 : }
2279 :
2280 252 : return s1;
2281 : }
2282 :
2283 : /*
2284 : * eqjoinsel - Join selectivity of "="
2285 : */
2286 : Datum
2287 263420 : eqjoinsel(PG_FUNCTION_ARGS)
2288 : {
2289 263420 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2290 263420 : Oid operator = PG_GETARG_OID(1);
2291 263420 : List *args = (List *) PG_GETARG_POINTER(2);
2292 :
2293 : #ifdef NOT_USED
2294 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2295 : #endif
2296 263420 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2297 263420 : Oid collation = PG_GET_COLLATION();
2298 : double selec;
2299 : double selec_inner;
2300 : VariableStatData vardata1;
2301 : VariableStatData vardata2;
2302 : double nd1;
2303 : double nd2;
2304 : bool isdefault1;
2305 : bool isdefault2;
2306 : Oid opfuncoid;
2307 : AttStatsSlot sslot1;
2308 : AttStatsSlot sslot2;
2309 263420 : Form_pg_statistic stats1 = NULL;
2310 263420 : Form_pg_statistic stats2 = NULL;
2311 263420 : bool have_mcvs1 = false;
2312 263420 : bool have_mcvs2 = false;
2313 : bool get_mcv_stats;
2314 : bool join_is_reversed;
2315 : RelOptInfo *inner_rel;
2316 :
2317 263420 : get_join_variables(root, args, sjinfo,
2318 : &vardata1, &vardata2, &join_is_reversed);
2319 :
2320 263420 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2321 263420 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2322 :
2323 263420 : opfuncoid = get_opcode(operator);
2324 :
2325 263420 : memset(&sslot1, 0, sizeof(sslot1));
2326 263420 : memset(&sslot2, 0, sizeof(sslot2));
2327 :
2328 : /*
2329 : * There is no use in fetching one side's MCVs if we lack MCVs for the
2330 : * other side, so do a quick check to verify that both stats exist.
2331 : */
2332 726234 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2333 354628 : HeapTupleIsValid(vardata2.statsTuple) &&
2334 155234 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2335 : STATISTIC_KIND_MCV, InvalidOid,
2336 462814 : 0) &&
2337 71198 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2338 : STATISTIC_KIND_MCV, InvalidOid,
2339 : 0));
2340 :
2341 263420 : if (HeapTupleIsValid(vardata1.statsTuple))
2342 : {
2343 : /* note we allow use of nullfrac regardless of security check */
2344 199394 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2345 227830 : if (get_mcv_stats &&
2346 28436 : statistic_proc_security_check(&vardata1, opfuncoid))
2347 28436 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2348 : STATISTIC_KIND_MCV, InvalidOid,
2349 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2350 : }
2351 :
2352 263420 : if (HeapTupleIsValid(vardata2.statsTuple))
2353 : {
2354 : /* note we allow use of nullfrac regardless of security check */
2355 175680 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2356 204116 : if (get_mcv_stats &&
2357 28436 : statistic_proc_security_check(&vardata2, opfuncoid))
2358 28436 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2359 : STATISTIC_KIND_MCV, InvalidOid,
2360 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2361 : }
2362 :
2363 : /* We need to compute the inner-join selectivity in all cases */
2364 263420 : selec_inner = eqjoinsel_inner(opfuncoid, collation,
2365 : &vardata1, &vardata2,
2366 : nd1, nd2,
2367 : isdefault1, isdefault2,
2368 : &sslot1, &sslot2,
2369 : stats1, stats2,
2370 : have_mcvs1, have_mcvs2);
2371 :
2372 263420 : switch (sjinfo->jointype)
2373 : {
2374 252718 : case JOIN_INNER:
2375 : case JOIN_LEFT:
2376 : case JOIN_FULL:
2377 252718 : selec = selec_inner;
2378 252718 : break;
2379 10702 : case JOIN_SEMI:
2380 : case JOIN_ANTI:
2381 :
2382 : /*
2383 : * Look up the join's inner relation. min_righthand is sufficient
2384 : * information because neither SEMI nor ANTI joins permit any
2385 : * reassociation into or out of their RHS, so the righthand will
2386 : * always be exactly that set of rels.
2387 : */
2388 10702 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2389 :
2390 10702 : if (!join_is_reversed)
2391 6588 : selec = eqjoinsel_semi(opfuncoid, collation,
2392 : &vardata1, &vardata2,
2393 : nd1, nd2,
2394 : isdefault1, isdefault2,
2395 : &sslot1, &sslot2,
2396 : stats1, stats2,
2397 : have_mcvs1, have_mcvs2,
2398 : inner_rel);
2399 : else
2400 : {
2401 4114 : Oid commop = get_commutator(operator);
2402 4114 : Oid commopfuncoid = OidIsValid(commop) ? get_opcode(commop) : InvalidOid;
2403 :
2404 4114 : selec = eqjoinsel_semi(commopfuncoid, collation,
2405 : &vardata2, &vardata1,
2406 : nd2, nd1,
2407 : isdefault2, isdefault1,
2408 : &sslot2, &sslot1,
2409 : stats2, stats1,
2410 : have_mcvs2, have_mcvs1,
2411 : inner_rel);
2412 : }
2413 :
2414 : /*
2415 : * We should never estimate the output of a semijoin to be more
2416 : * rows than we estimate for an inner join with the same input
2417 : * rels and join condition; it's obviously impossible for that to
2418 : * happen. The former estimate is N1 * Ssemi while the latter is
2419 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2420 : * this is worthwhile because of the shakier estimation rules we
2421 : * use in eqjoinsel_semi, particularly in cases where it has to
2422 : * punt entirely.
2423 : */
2424 10702 : selec = Min(selec, inner_rel->rows * selec_inner);
2425 10702 : break;
2426 0 : default:
2427 : /* other values not expected here */
2428 0 : elog(ERROR, "unrecognized join type: %d",
2429 : (int) sjinfo->jointype);
2430 : selec = 0; /* keep compiler quiet */
2431 : break;
2432 : }
2433 :
2434 263420 : free_attstatsslot(&sslot1);
2435 263420 : free_attstatsslot(&sslot2);
2436 :
2437 263420 : ReleaseVariableStats(vardata1);
2438 263420 : ReleaseVariableStats(vardata2);
2439 :
2440 263420 : CLAMP_PROBABILITY(selec);
2441 :
2442 263420 : PG_RETURN_FLOAT8((float8) selec);
2443 : }
2444 :
2445 : /*
2446 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2447 : *
2448 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2449 : * that it's worth trying to distinguish them here.
2450 : */
2451 : static double
2452 263420 : eqjoinsel_inner(Oid opfuncoid, Oid collation,
2453 : VariableStatData *vardata1, VariableStatData *vardata2,
2454 : double nd1, double nd2,
2455 : bool isdefault1, bool isdefault2,
2456 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2457 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2458 : bool have_mcvs1, bool have_mcvs2)
2459 : {
2460 : double selec;
2461 :
2462 263420 : if (have_mcvs1 && have_mcvs2)
2463 28436 : {
2464 : /*
2465 : * We have most-common-value lists for both relations. Run through
2466 : * the lists to see which MCVs actually join to each other with the
2467 : * given operator. This allows us to determine the exact join
2468 : * selectivity for the portion of the relations represented by the MCV
2469 : * lists. We still have to estimate for the remaining population, but
2470 : * in a skewed distribution this gives us a big leg up in accuracy.
2471 : * For motivation see the analysis in Y. Ioannidis and S.
2472 : * Christodoulakis, "On the propagation of errors in the size of join
2473 : * results", Technical Report 1018, Computer Science Dept., University
2474 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2475 : */
2476 28436 : LOCAL_FCINFO(fcinfo, 2);
2477 : FmgrInfo eqproc;
2478 : bool *hasmatch1;
2479 : bool *hasmatch2;
2480 28436 : double nullfrac1 = stats1->stanullfrac;
2481 28436 : double nullfrac2 = stats2->stanullfrac;
2482 : double matchprodfreq,
2483 : matchfreq1,
2484 : matchfreq2,
2485 : unmatchfreq1,
2486 : unmatchfreq2,
2487 : otherfreq1,
2488 : otherfreq2,
2489 : totalsel1,
2490 : totalsel2;
2491 : int i,
2492 : nmatches;
2493 :
2494 28436 : fmgr_info(opfuncoid, &eqproc);
2495 :
2496 : /*
2497 : * Save a few cycles by setting up the fcinfo struct just once. Using
2498 : * FunctionCallInvoke directly also avoids failure if the eqproc
2499 : * returns NULL, though really equality functions should never do
2500 : * that.
2501 : */
2502 28436 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2503 : NULL, NULL);
2504 28436 : fcinfo->args[0].isnull = false;
2505 28436 : fcinfo->args[1].isnull = false;
2506 :
2507 28436 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2508 28436 : hasmatch2 = (bool *) palloc0(sslot2->nvalues * sizeof(bool));
2509 :
2510 : /*
2511 : * Note we assume that each MCV will match at most one member of the
2512 : * other MCV list. If the operator isn't really equality, there could
2513 : * be multiple matches --- but we don't look for them, both for speed
2514 : * and because the math wouldn't add up...
2515 : */
2516 28436 : matchprodfreq = 0.0;
2517 28436 : nmatches = 0;
2518 705260 : for (i = 0; i < sslot1->nvalues; i++)
2519 : {
2520 : int j;
2521 :
2522 676824 : fcinfo->args[0].value = sslot1->values[i];
2523 :
2524 23499140 : for (j = 0; j < sslot2->nvalues; j++)
2525 : {
2526 : Datum fresult;
2527 :
2528 23114366 : if (hasmatch2[j])
2529 6233456 : continue;
2530 16880910 : fcinfo->args[1].value = sslot2->values[j];
2531 16880910 : fcinfo->isnull = false;
2532 16880910 : fresult = FunctionCallInvoke(fcinfo);
2533 16880910 : if (!fcinfo->isnull && DatumGetBool(fresult))
2534 : {
2535 292050 : hasmatch1[i] = hasmatch2[j] = true;
2536 292050 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
2537 292050 : nmatches++;
2538 292050 : break;
2539 : }
2540 : }
2541 : }
2542 28436 : CLAMP_PROBABILITY(matchprodfreq);
2543 : /* Sum up frequencies of matched and unmatched MCVs */
2544 28436 : matchfreq1 = unmatchfreq1 = 0.0;
2545 705260 : for (i = 0; i < sslot1->nvalues; i++)
2546 : {
2547 676824 : if (hasmatch1[i])
2548 292050 : matchfreq1 += sslot1->numbers[i];
2549 : else
2550 384774 : unmatchfreq1 += sslot1->numbers[i];
2551 : }
2552 28436 : CLAMP_PROBABILITY(matchfreq1);
2553 28436 : CLAMP_PROBABILITY(unmatchfreq1);
2554 28436 : matchfreq2 = unmatchfreq2 = 0.0;
2555 525950 : for (i = 0; i < sslot2->nvalues; i++)
2556 : {
2557 497514 : if (hasmatch2[i])
2558 292050 : matchfreq2 += sslot2->numbers[i];
2559 : else
2560 205464 : unmatchfreq2 += sslot2->numbers[i];
2561 : }
2562 28436 : CLAMP_PROBABILITY(matchfreq2);
2563 28436 : CLAMP_PROBABILITY(unmatchfreq2);
2564 28436 : pfree(hasmatch1);
2565 28436 : pfree(hasmatch2);
2566 :
2567 : /*
2568 : * Compute total frequency of non-null values that are not in the MCV
2569 : * lists.
2570 : */
2571 28436 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2572 28436 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2573 28436 : CLAMP_PROBABILITY(otherfreq1);
2574 28436 : CLAMP_PROBABILITY(otherfreq2);
2575 :
2576 : /*
2577 : * We can estimate the total selectivity from the point of view of
2578 : * relation 1 as: the known selectivity for matched MCVs, plus
2579 : * unmatched MCVs that are assumed to match against random members of
2580 : * relation 2's non-MCV population, plus non-MCV values that are
2581 : * assumed to match against random members of relation 2's unmatched
2582 : * MCVs plus non-MCV values.
2583 : */
2584 28436 : totalsel1 = matchprodfreq;
2585 28436 : if (nd2 > sslot2->nvalues)
2586 6216 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2587 28436 : if (nd2 > nmatches)
2588 11076 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2589 11076 : (nd2 - nmatches);
2590 : /* Same estimate from the point of view of relation 2. */
2591 28436 : totalsel2 = matchprodfreq;
2592 28436 : if (nd1 > sslot1->nvalues)
2593 7056 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2594 28436 : if (nd1 > nmatches)
2595 9804 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2596 9804 : (nd1 - nmatches);
2597 :
2598 : /*
2599 : * Use the smaller of the two estimates. This can be justified in
2600 : * essentially the same terms as given below for the no-stats case: to
2601 : * a first approximation, we are estimating from the point of view of
2602 : * the relation with smaller nd.
2603 : */
2604 28436 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2605 : }
2606 : else
2607 : {
2608 : /*
2609 : * We do not have MCV lists for both sides. Estimate the join
2610 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2611 : * is plausible if we assume that the join operator is strict and the
2612 : * non-null values are about equally distributed: a given non-null
2613 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2614 : * of rel2, so total join rows are at most
2615 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2616 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2617 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2618 : * with MIN() is an upper bound. Using the MIN() means we estimate
2619 : * from the point of view of the relation with smaller nd (since the
2620 : * larger nd is determining the MIN). It is reasonable to assume that
2621 : * most tuples in this rel will have join partners, so the bound is
2622 : * probably reasonably tight and should be taken as-is.
2623 : *
2624 : * XXX Can we be smarter if we have an MCV list for just one side? It
2625 : * seems that if we assume equal distribution for the other side, we
2626 : * end up with the same answer anyway.
2627 : */
2628 234984 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2629 234984 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2630 :
2631 234984 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2632 234984 : if (nd1 > nd2)
2633 122786 : selec /= nd1;
2634 : else
2635 112198 : selec /= nd2;
2636 : }
2637 :
2638 263420 : return selec;
2639 : }
2640 :
2641 : /*
2642 : * eqjoinsel_semi --- eqjoinsel for semi join
2643 : *
2644 : * (Also used for anti join, which we are supposed to estimate the same way.)
2645 : * Caller has ensured that vardata1 is the LHS variable.
2646 : * Unlike eqjoinsel_inner, we have to cope with opfuncoid being InvalidOid.
2647 : */
2648 : static double
2649 10702 : eqjoinsel_semi(Oid opfuncoid, Oid collation,
2650 : VariableStatData *vardata1, VariableStatData *vardata2,
2651 : double nd1, double nd2,
2652 : bool isdefault1, bool isdefault2,
2653 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2654 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2655 : bool have_mcvs1, bool have_mcvs2,
2656 : RelOptInfo *inner_rel)
2657 : {
2658 : double selec;
2659 :
2660 : /*
2661 : * We clamp nd2 to be not more than what we estimate the inner relation's
2662 : * size to be. This is intuitively somewhat reasonable since obviously
2663 : * there can't be more than that many distinct values coming from the
2664 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2665 : * likewise) is that this is the only pathway by which restriction clauses
2666 : * applied to the inner rel will affect the join result size estimate,
2667 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2668 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2669 : * the selectivity of outer-rel restrictions.
2670 : *
2671 : * We can apply this clamping both with respect to the base relation from
2672 : * which the join variable comes (if there is just one), and to the
2673 : * immediate inner input relation of the current join.
2674 : *
2675 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2676 : * great, maybe, but it didn't come out of nowhere either. This is most
2677 : * helpful when the inner relation is empty and consequently has no stats.
2678 : */
2679 10702 : if (vardata2->rel)
2680 : {
2681 10696 : if (nd2 >= vardata2->rel->rows)
2682 : {
2683 8574 : nd2 = vardata2->rel->rows;
2684 8574 : isdefault2 = false;
2685 : }
2686 : }
2687 10702 : if (nd2 >= inner_rel->rows)
2688 : {
2689 8516 : nd2 = inner_rel->rows;
2690 8516 : isdefault2 = false;
2691 : }
2692 :
2693 10702 : if (have_mcvs1 && have_mcvs2 && OidIsValid(opfuncoid))
2694 612 : {
2695 : /*
2696 : * We have most-common-value lists for both relations. Run through
2697 : * the lists to see which MCVs actually join to each other with the
2698 : * given operator. This allows us to determine the exact join
2699 : * selectivity for the portion of the relations represented by the MCV
2700 : * lists. We still have to estimate for the remaining population, but
2701 : * in a skewed distribution this gives us a big leg up in accuracy.
2702 : */
2703 612 : LOCAL_FCINFO(fcinfo, 2);
2704 : FmgrInfo eqproc;
2705 : bool *hasmatch1;
2706 : bool *hasmatch2;
2707 612 : double nullfrac1 = stats1->stanullfrac;
2708 : double matchfreq1,
2709 : uncertainfrac,
2710 : uncertain;
2711 : int i,
2712 : nmatches,
2713 : clamped_nvalues2;
2714 :
2715 : /*
2716 : * The clamping above could have resulted in nd2 being less than
2717 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2718 : * most common values in the relation will appear in the join input,
2719 : * and so compare to only the first nd2 members of the MCV list. Of
2720 : * course this is frequently wrong, but it's the best bet we can make.
2721 : */
2722 612 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2723 :
2724 612 : fmgr_info(opfuncoid, &eqproc);
2725 :
2726 : /*
2727 : * Save a few cycles by setting up the fcinfo struct just once. Using
2728 : * FunctionCallInvoke directly also avoids failure if the eqproc
2729 : * returns NULL, though really equality functions should never do
2730 : * that.
2731 : */
2732 612 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2733 : NULL, NULL);
2734 612 : fcinfo->args[0].isnull = false;
2735 612 : fcinfo->args[1].isnull = false;
2736 :
2737 612 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2738 612 : hasmatch2 = (bool *) palloc0(clamped_nvalues2 * sizeof(bool));
2739 :
2740 : /*
2741 : * Note we assume that each MCV will match at most one member of the
2742 : * other MCV list. If the operator isn't really equality, there could
2743 : * be multiple matches --- but we don't look for them, both for speed
2744 : * and because the math wouldn't add up...
2745 : */
2746 612 : nmatches = 0;
2747 12912 : for (i = 0; i < sslot1->nvalues; i++)
2748 : {
2749 : int j;
2750 :
2751 12300 : fcinfo->args[0].value = sslot1->values[i];
2752 :
2753 472754 : for (j = 0; j < clamped_nvalues2; j++)
2754 : {
2755 : Datum fresult;
2756 :
2757 471318 : if (hasmatch2[j])
2758 382292 : continue;
2759 89026 : fcinfo->args[1].value = sslot2->values[j];
2760 89026 : fcinfo->isnull = false;
2761 89026 : fresult = FunctionCallInvoke(fcinfo);
2762 89026 : if (!fcinfo->isnull && DatumGetBool(fresult))
2763 : {
2764 10864 : hasmatch1[i] = hasmatch2[j] = true;
2765 10864 : nmatches++;
2766 10864 : break;
2767 : }
2768 : }
2769 : }
2770 : /* Sum up frequencies of matched MCVs */
2771 612 : matchfreq1 = 0.0;
2772 12912 : for (i = 0; i < sslot1->nvalues; i++)
2773 : {
2774 12300 : if (hasmatch1[i])
2775 10864 : matchfreq1 += sslot1->numbers[i];
2776 : }
2777 612 : CLAMP_PROBABILITY(matchfreq1);
2778 612 : pfree(hasmatch1);
2779 612 : pfree(hasmatch2);
2780 :
2781 : /*
2782 : * Now we need to estimate the fraction of relation 1 that has at
2783 : * least one join partner. We know for certain that the matched MCVs
2784 : * do, so that gives us a lower bound, but we're really in the dark
2785 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2786 : * assume all non-null rel1 rows have join partners, else assume for
2787 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2788 : * can discount the known-matched MCVs from the distinct-values counts
2789 : * before doing the division.
2790 : *
2791 : * Crude as the above is, it's completely useless if we don't have
2792 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2793 : * nd2 is default, punt and assume half of the uncertain rows have
2794 : * join partners.
2795 : */
2796 612 : if (!isdefault1 && !isdefault2)
2797 : {
2798 612 : nd1 -= nmatches;
2799 612 : nd2 -= nmatches;
2800 612 : if (nd1 <= nd2 || nd2 < 0)
2801 582 : uncertainfrac = 1.0;
2802 : else
2803 30 : uncertainfrac = nd2 / nd1;
2804 : }
2805 : else
2806 0 : uncertainfrac = 0.5;
2807 612 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2808 612 : CLAMP_PROBABILITY(uncertain);
2809 612 : selec = matchfreq1 + uncertainfrac * uncertain;
2810 : }
2811 : else
2812 : {
2813 : /*
2814 : * Without MCV lists for both sides, we can only use the heuristic
2815 : * about nd1 vs nd2.
2816 : */
2817 10090 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2818 :
2819 10090 : if (!isdefault1 && !isdefault2)
2820 : {
2821 7734 : if (nd1 <= nd2 || nd2 < 0)
2822 4890 : selec = 1.0 - nullfrac1;
2823 : else
2824 2844 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2825 : }
2826 : else
2827 2356 : selec = 0.5 * (1.0 - nullfrac1);
2828 : }
2829 :
2830 10702 : return selec;
2831 : }
2832 :
2833 : /*
2834 : * neqjoinsel - Join selectivity of "!="
2835 : */
2836 : Datum
2837 3728 : neqjoinsel(PG_FUNCTION_ARGS)
2838 : {
2839 3728 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2840 3728 : Oid operator = PG_GETARG_OID(1);
2841 3728 : List *args = (List *) PG_GETARG_POINTER(2);
2842 3728 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2843 3728 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2844 3728 : Oid collation = PG_GET_COLLATION();
2845 : float8 result;
2846 :
2847 3728 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
2848 1262 : {
2849 : /*
2850 : * For semi-joins, if there is more than one distinct value in the RHS
2851 : * relation then every non-null LHS row must find a row to join since
2852 : * it can only be equal to one of them. We'll assume that there is
2853 : * always more than one distinct RHS value for the sake of stability,
2854 : * though in theory we could have special cases for empty RHS
2855 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
2856 : * fraction of LHS that has the same value as the single RHS value).
2857 : *
2858 : * For anti-joins, if we use the same assumption that there is more
2859 : * than one distinct key in the RHS relation, then every non-null LHS
2860 : * row must be suppressed by the anti-join.
2861 : *
2862 : * So either way, the selectivity estimate should be 1 - nullfrac.
2863 : */
2864 : VariableStatData leftvar;
2865 : VariableStatData rightvar;
2866 : bool reversed;
2867 : HeapTuple statsTuple;
2868 : double nullfrac;
2869 :
2870 1262 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
2871 1262 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
2872 1262 : if (HeapTupleIsValid(statsTuple))
2873 1030 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
2874 : else
2875 232 : nullfrac = 0.0;
2876 1262 : ReleaseVariableStats(leftvar);
2877 1262 : ReleaseVariableStats(rightvar);
2878 :
2879 1262 : result = 1.0 - nullfrac;
2880 : }
2881 : else
2882 : {
2883 : /*
2884 : * We want 1 - eqjoinsel() where the equality operator is the one
2885 : * associated with this != operator, that is, its negator.
2886 : */
2887 2466 : Oid eqop = get_negator(operator);
2888 :
2889 2466 : if (eqop)
2890 : {
2891 : result =
2892 2466 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
2893 : collation,
2894 : PointerGetDatum(root),
2895 : ObjectIdGetDatum(eqop),
2896 : PointerGetDatum(args),
2897 : Int16GetDatum(jointype),
2898 : PointerGetDatum(sjinfo)));
2899 : }
2900 : else
2901 : {
2902 : /* Use default selectivity (should we raise an error instead?) */
2903 0 : result = DEFAULT_EQ_SEL;
2904 : }
2905 2466 : result = 1.0 - result;
2906 : }
2907 :
2908 3728 : PG_RETURN_FLOAT8(result);
2909 : }
2910 :
2911 : /*
2912 : * scalarltjoinsel - Join selectivity of "<" for scalars
2913 : */
2914 : Datum
2915 324 : scalarltjoinsel(PG_FUNCTION_ARGS)
2916 : {
2917 324 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2918 : }
2919 :
2920 : /*
2921 : * scalarlejoinsel - Join selectivity of "<=" for scalars
2922 : */
2923 : Datum
2924 276 : scalarlejoinsel(PG_FUNCTION_ARGS)
2925 : {
2926 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2927 : }
2928 :
2929 : /*
2930 : * scalargtjoinsel - Join selectivity of ">" for scalars
2931 : */
2932 : Datum
2933 276 : scalargtjoinsel(PG_FUNCTION_ARGS)
2934 : {
2935 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2936 : }
2937 :
2938 : /*
2939 : * scalargejoinsel - Join selectivity of ">=" for scalars
2940 : */
2941 : Datum
2942 184 : scalargejoinsel(PG_FUNCTION_ARGS)
2943 : {
2944 184 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2945 : }
2946 :
2947 :
2948 : /*
2949 : * mergejoinscansel - Scan selectivity of merge join.
2950 : *
2951 : * A merge join will stop as soon as it exhausts either input stream.
2952 : * Therefore, if we can estimate the ranges of both input variables,
2953 : * we can estimate how much of the input will actually be read. This
2954 : * can have a considerable impact on the cost when using indexscans.
2955 : *
2956 : * Also, we can estimate how much of each input has to be read before the
2957 : * first join pair is found, which will affect the join's startup time.
2958 : *
2959 : * clause should be a clause already known to be mergejoinable. opfamily,
2960 : * cmptype, and nulls_first specify the sort ordering being used.
2961 : *
2962 : * The outputs are:
2963 : * *leftstart is set to the fraction of the left-hand variable expected
2964 : * to be scanned before the first join pair is found (0 to 1).
2965 : * *leftend is set to the fraction of the left-hand variable expected
2966 : * to be scanned before the join terminates (0 to 1).
2967 : * *rightstart, *rightend similarly for the right-hand variable.
2968 : */
2969 : void
2970 140898 : mergejoinscansel(PlannerInfo *root, Node *clause,
2971 : Oid opfamily, CompareType cmptype, bool nulls_first,
2972 : Selectivity *leftstart, Selectivity *leftend,
2973 : Selectivity *rightstart, Selectivity *rightend)
2974 : {
2975 : Node *left,
2976 : *right;
2977 : VariableStatData leftvar,
2978 : rightvar;
2979 : Oid opmethod;
2980 : int op_strategy;
2981 : Oid op_lefttype;
2982 : Oid op_righttype;
2983 : Oid opno,
2984 : collation,
2985 : lsortop,
2986 : rsortop,
2987 : lstatop,
2988 : rstatop,
2989 : ltop,
2990 : leop,
2991 : revltop,
2992 : revleop;
2993 : StrategyNumber ltstrat,
2994 : lestrat,
2995 : gtstrat,
2996 : gestrat;
2997 : bool isgt;
2998 : Datum leftmin,
2999 : leftmax,
3000 : rightmin,
3001 : rightmax;
3002 : double selec;
3003 :
3004 : /* Set default results if we can't figure anything out. */
3005 : /* XXX should default "start" fraction be a bit more than 0? */
3006 140898 : *leftstart = *rightstart = 0.0;
3007 140898 : *leftend = *rightend = 1.0;
3008 :
3009 : /* Deconstruct the merge clause */
3010 140898 : if (!is_opclause(clause))
3011 0 : return; /* shouldn't happen */
3012 140898 : opno = ((OpExpr *) clause)->opno;
3013 140898 : collation = ((OpExpr *) clause)->inputcollid;
3014 140898 : left = get_leftop((Expr *) clause);
3015 140898 : right = get_rightop((Expr *) clause);
3016 140898 : if (!right)
3017 0 : return; /* shouldn't happen */
3018 :
3019 : /* Look for stats for the inputs */
3020 140898 : examine_variable(root, left, 0, &leftvar);
3021 140898 : examine_variable(root, right, 0, &rightvar);
3022 :
3023 140898 : opmethod = get_opfamily_method(opfamily);
3024 :
3025 : /* Extract the operator's declared left/right datatypes */
3026 140898 : get_op_opfamily_properties(opno, opfamily, false,
3027 : &op_strategy,
3028 : &op_lefttype,
3029 : &op_righttype);
3030 : Assert(IndexAmTranslateStrategy(op_strategy, opmethod, opfamily, true) == COMPARE_EQ);
3031 :
3032 : /*
3033 : * Look up the various operators we need. If we don't find them all, it
3034 : * probably means the opfamily is broken, but we just fail silently.
3035 : *
3036 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3037 : * operator, regardless of which sort direction we are considering.
3038 : */
3039 140898 : switch (cmptype)
3040 : {
3041 140826 : case COMPARE_LT:
3042 140826 : isgt = false;
3043 140826 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3044 140826 : lestrat = IndexAmTranslateCompareType(COMPARE_LE, opmethod, opfamily, true);
3045 140826 : if (op_lefttype == op_righttype)
3046 : {
3047 : /* easy case */
3048 139040 : ltop = get_opfamily_member(opfamily,
3049 : op_lefttype, op_righttype,
3050 : ltstrat);
3051 139040 : leop = get_opfamily_member(opfamily,
3052 : op_lefttype, op_righttype,
3053 : lestrat);
3054 139040 : lsortop = ltop;
3055 139040 : rsortop = ltop;
3056 139040 : lstatop = lsortop;
3057 139040 : rstatop = rsortop;
3058 139040 : revltop = ltop;
3059 139040 : revleop = leop;
3060 : }
3061 : else
3062 : {
3063 1786 : ltop = get_opfamily_member(opfamily,
3064 : op_lefttype, op_righttype,
3065 : ltstrat);
3066 1786 : leop = get_opfamily_member(opfamily,
3067 : op_lefttype, op_righttype,
3068 : lestrat);
3069 1786 : lsortop = get_opfamily_member(opfamily,
3070 : op_lefttype, op_lefttype,
3071 : ltstrat);
3072 1786 : rsortop = get_opfamily_member(opfamily,
3073 : op_righttype, op_righttype,
3074 : ltstrat);
3075 1786 : lstatop = lsortop;
3076 1786 : rstatop = rsortop;
3077 1786 : revltop = get_opfamily_member(opfamily,
3078 : op_righttype, op_lefttype,
3079 : ltstrat);
3080 1786 : revleop = get_opfamily_member(opfamily,
3081 : op_righttype, op_lefttype,
3082 : lestrat);
3083 : }
3084 140826 : break;
3085 72 : case COMPARE_GT:
3086 : /* descending-order case */
3087 72 : isgt = true;
3088 72 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3089 72 : gtstrat = IndexAmTranslateCompareType(COMPARE_GT, opmethod, opfamily, true);
3090 72 : gestrat = IndexAmTranslateCompareType(COMPARE_GE, opmethod, opfamily, true);
3091 72 : if (op_lefttype == op_righttype)
3092 : {
3093 : /* easy case */
3094 72 : ltop = get_opfamily_member(opfamily,
3095 : op_lefttype, op_righttype,
3096 : gtstrat);
3097 72 : leop = get_opfamily_member(opfamily,
3098 : op_lefttype, op_righttype,
3099 : gestrat);
3100 72 : lsortop = ltop;
3101 72 : rsortop = ltop;
3102 72 : lstatop = get_opfamily_member(opfamily,
3103 : op_lefttype, op_lefttype,
3104 : ltstrat);
3105 72 : rstatop = lstatop;
3106 72 : revltop = ltop;
3107 72 : revleop = leop;
3108 : }
3109 : else
3110 : {
3111 0 : ltop = get_opfamily_member(opfamily,
3112 : op_lefttype, op_righttype,
3113 : gtstrat);
3114 0 : leop = get_opfamily_member(opfamily,
3115 : op_lefttype, op_righttype,
3116 : gestrat);
3117 0 : lsortop = get_opfamily_member(opfamily,
3118 : op_lefttype, op_lefttype,
3119 : gtstrat);
3120 0 : rsortop = get_opfamily_member(opfamily,
3121 : op_righttype, op_righttype,
3122 : gtstrat);
3123 0 : lstatop = get_opfamily_member(opfamily,
3124 : op_lefttype, op_lefttype,
3125 : ltstrat);
3126 0 : rstatop = get_opfamily_member(opfamily,
3127 : op_righttype, op_righttype,
3128 : ltstrat);
3129 0 : revltop = get_opfamily_member(opfamily,
3130 : op_righttype, op_lefttype,
3131 : gtstrat);
3132 0 : revleop = get_opfamily_member(opfamily,
3133 : op_righttype, op_lefttype,
3134 : gestrat);
3135 : }
3136 72 : break;
3137 0 : default:
3138 0 : goto fail; /* shouldn't get here */
3139 : }
3140 :
3141 140898 : if (!OidIsValid(lsortop) ||
3142 140898 : !OidIsValid(rsortop) ||
3143 140898 : !OidIsValid(lstatop) ||
3144 140898 : !OidIsValid(rstatop) ||
3145 140886 : !OidIsValid(ltop) ||
3146 140886 : !OidIsValid(leop) ||
3147 140886 : !OidIsValid(revltop) ||
3148 : !OidIsValid(revleop))
3149 12 : goto fail; /* insufficient info in catalogs */
3150 :
3151 : /* Try to get ranges of both inputs */
3152 140886 : if (!isgt)
3153 : {
3154 140814 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3155 : &leftmin, &leftmax))
3156 34224 : goto fail; /* no range available from stats */
3157 106590 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3158 : &rightmin, &rightmax))
3159 24692 : goto fail; /* no range available from stats */
3160 : }
3161 : else
3162 : {
3163 : /* need to swap the max and min */
3164 72 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3165 : &leftmax, &leftmin))
3166 30 : goto fail; /* no range available from stats */
3167 42 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3168 : &rightmax, &rightmin))
3169 0 : goto fail; /* no range available from stats */
3170 : }
3171 :
3172 : /*
3173 : * Now, the fraction of the left variable that will be scanned is the
3174 : * fraction that's <= the right-side maximum value. But only believe
3175 : * non-default estimates, else stick with our 1.0.
3176 : */
3177 81940 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3178 : rightmax, op_righttype);
3179 81940 : if (selec != DEFAULT_INEQ_SEL)
3180 81934 : *leftend = selec;
3181 :
3182 : /* And similarly for the right variable. */
3183 81940 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3184 : leftmax, op_lefttype);
3185 81940 : if (selec != DEFAULT_INEQ_SEL)
3186 81940 : *rightend = selec;
3187 :
3188 : /*
3189 : * Only one of the two "end" fractions can really be less than 1.0;
3190 : * believe the smaller estimate and reset the other one to exactly 1.0. If
3191 : * we get exactly equal estimates (as can easily happen with self-joins),
3192 : * believe neither.
3193 : */
3194 81940 : if (*leftend > *rightend)
3195 24970 : *leftend = 1.0;
3196 56970 : else if (*leftend < *rightend)
3197 33080 : *rightend = 1.0;
3198 : else
3199 23890 : *leftend = *rightend = 1.0;
3200 :
3201 : /*
3202 : * Also, the fraction of the left variable that will be scanned before the
3203 : * first join pair is found is the fraction that's < the right-side
3204 : * minimum value. But only believe non-default estimates, else stick with
3205 : * our own default.
3206 : */
3207 81940 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3208 : rightmin, op_righttype);
3209 81940 : if (selec != DEFAULT_INEQ_SEL)
3210 81940 : *leftstart = selec;
3211 :
3212 : /* And similarly for the right variable. */
3213 81940 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3214 : leftmin, op_lefttype);
3215 81940 : if (selec != DEFAULT_INEQ_SEL)
3216 81940 : *rightstart = selec;
3217 :
3218 : /*
3219 : * Only one of the two "start" fractions can really be more than zero;
3220 : * believe the larger estimate and reset the other one to exactly 0.0. If
3221 : * we get exactly equal estimates (as can easily happen with self-joins),
3222 : * believe neither.
3223 : */
3224 81940 : if (*leftstart < *rightstart)
3225 16544 : *leftstart = 0.0;
3226 65396 : else if (*leftstart > *rightstart)
3227 22910 : *rightstart = 0.0;
3228 : else
3229 42486 : *leftstart = *rightstart = 0.0;
3230 :
3231 : /*
3232 : * If the sort order is nulls-first, we're going to have to skip over any
3233 : * nulls too. These would not have been counted by scalarineqsel, and we
3234 : * can safely add in this fraction regardless of whether we believe
3235 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3236 : */
3237 81940 : if (nulls_first)
3238 : {
3239 : Form_pg_statistic stats;
3240 :
3241 42 : if (HeapTupleIsValid(leftvar.statsTuple))
3242 : {
3243 42 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3244 42 : *leftstart += stats->stanullfrac;
3245 42 : CLAMP_PROBABILITY(*leftstart);
3246 42 : *leftend += stats->stanullfrac;
3247 42 : CLAMP_PROBABILITY(*leftend);
3248 : }
3249 42 : if (HeapTupleIsValid(rightvar.statsTuple))
3250 : {
3251 42 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3252 42 : *rightstart += stats->stanullfrac;
3253 42 : CLAMP_PROBABILITY(*rightstart);
3254 42 : *rightend += stats->stanullfrac;
3255 42 : CLAMP_PROBABILITY(*rightend);
3256 : }
3257 : }
3258 :
3259 : /* Disbelieve start >= end, just in case that can happen */
3260 81940 : if (*leftstart >= *leftend)
3261 : {
3262 204 : *leftstart = 0.0;
3263 204 : *leftend = 1.0;
3264 : }
3265 81940 : if (*rightstart >= *rightend)
3266 : {
3267 1044 : *rightstart = 0.0;
3268 1044 : *rightend = 1.0;
3269 : }
3270 :
3271 80896 : fail:
3272 140898 : ReleaseVariableStats(leftvar);
3273 140898 : ReleaseVariableStats(rightvar);
3274 : }
3275 :
3276 :
3277 : /*
3278 : * matchingsel -- generic matching-operator selectivity support
3279 : *
3280 : * Use these for any operators that (a) are on data types for which we collect
3281 : * standard statistics, and (b) have behavior for which the default estimate
3282 : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3283 : * operators.
3284 : */
3285 :
3286 : Datum
3287 1130 : matchingsel(PG_FUNCTION_ARGS)
3288 : {
3289 1130 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3290 1130 : Oid operator = PG_GETARG_OID(1);
3291 1130 : List *args = (List *) PG_GETARG_POINTER(2);
3292 1130 : int varRelid = PG_GETARG_INT32(3);
3293 1130 : Oid collation = PG_GET_COLLATION();
3294 : double selec;
3295 :
3296 : /* Use generic restriction selectivity logic. */
3297 1130 : selec = generic_restriction_selectivity(root, operator, collation,
3298 : args, varRelid,
3299 : DEFAULT_MATCHING_SEL);
3300 :
3301 1130 : PG_RETURN_FLOAT8((float8) selec);
3302 : }
3303 :
3304 : Datum
3305 6 : matchingjoinsel(PG_FUNCTION_ARGS)
3306 : {
3307 : /* Just punt, for the moment. */
3308 6 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3309 : }
3310 :
3311 :
3312 : /*
3313 : * Helper routine for estimate_num_groups: add an item to a list of
3314 : * GroupVarInfos, but only if it's not known equal to any of the existing
3315 : * entries.
3316 : */
3317 : typedef struct
3318 : {
3319 : Node *var; /* might be an expression, not just a Var */
3320 : RelOptInfo *rel; /* relation it belongs to */
3321 : double ndistinct; /* # distinct values */
3322 : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3323 : } GroupVarInfo;
3324 :
3325 : static List *
3326 384546 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3327 : Node *var, VariableStatData *vardata)
3328 : {
3329 : GroupVarInfo *varinfo;
3330 : double ndistinct;
3331 : bool isdefault;
3332 : ListCell *lc;
3333 :
3334 384546 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3335 :
3336 : /*
3337 : * The nullingrels bits within the var could cause the same var to be
3338 : * counted multiple times if it's marked with different nullingrels. They
3339 : * could also prevent us from matching the var to the expressions in
3340 : * extended statistics (see estimate_multivariate_ndistinct). So strip
3341 : * them out first.
3342 : */
3343 384546 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3344 :
3345 465700 : foreach(lc, varinfos)
3346 : {
3347 82264 : varinfo = (GroupVarInfo *) lfirst(lc);
3348 :
3349 : /* Drop exact duplicates */
3350 82264 : if (equal(var, varinfo->var))
3351 1110 : return varinfos;
3352 :
3353 : /*
3354 : * Drop known-equal vars, but only if they belong to different
3355 : * relations (see comments for estimate_num_groups). We aren't too
3356 : * fussy about the semantics of "equal" here.
3357 : */
3358 87198 : if (vardata->rel != varinfo->rel &&
3359 5768 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3360 : {
3361 300 : if (varinfo->ndistinct <= ndistinct)
3362 : {
3363 : /* Keep older item, forget new one */
3364 276 : return varinfos;
3365 : }
3366 : else
3367 : {
3368 : /* Delete the older item */
3369 24 : varinfos = foreach_delete_current(varinfos, lc);
3370 : }
3371 : }
3372 : }
3373 :
3374 383436 : varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
3375 :
3376 383436 : varinfo->var = var;
3377 383436 : varinfo->rel = vardata->rel;
3378 383436 : varinfo->ndistinct = ndistinct;
3379 383436 : varinfo->isdefault = isdefault;
3380 383436 : varinfos = lappend(varinfos, varinfo);
3381 383436 : return varinfos;
3382 : }
3383 :
3384 : /*
3385 : * estimate_num_groups - Estimate number of groups in a grouped query
3386 : *
3387 : * Given a query having a GROUP BY clause, estimate how many groups there
3388 : * will be --- ie, the number of distinct combinations of the GROUP BY
3389 : * expressions.
3390 : *
3391 : * This routine is also used to estimate the number of rows emitted by
3392 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3393 : * actually, we only use it for DISTINCT when there's no grouping or
3394 : * aggregation ahead of the DISTINCT.)
3395 : *
3396 : * Inputs:
3397 : * root - the query
3398 : * groupExprs - list of expressions being grouped by
3399 : * input_rows - number of rows estimated to arrive at the group/unique
3400 : * filter step
3401 : * pgset - NULL, or a List** pointing to a grouping set to filter the
3402 : * groupExprs against
3403 : *
3404 : * Outputs:
3405 : * estinfo - When passed as non-NULL, the function will set bits in the
3406 : * "flags" field in order to provide callers with additional information
3407 : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3408 : * bit if we used any default values in the estimation.
3409 : *
3410 : * Given the lack of any cross-correlation statistics in the system, it's
3411 : * impossible to do anything really trustworthy with GROUP BY conditions
3412 : * involving multiple Vars. We should however avoid assuming the worst
3413 : * case (all possible cross-product terms actually appear as groups) since
3414 : * very often the grouped-by Vars are highly correlated. Our current approach
3415 : * is as follows:
3416 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3417 : * independently of their content, and are ignored in the subsequent
3418 : * steps. This is mainly because tests like "col IS NULL" break the
3419 : * heuristic used in step 2 especially badly.
3420 : * 2. Reduce the given expressions to a list of unique Vars used. For
3421 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3422 : * It is clearly correct not to count the same Var more than once.
3423 : * It is also reasonable to treat f(x) the same as x: f() cannot
3424 : * increase the number of distinct values (unless it is volatile,
3425 : * which we consider unlikely for grouping), but it probably won't
3426 : * reduce the number of distinct values much either.
3427 : * As a special case, if a GROUP BY expression can be matched to an
3428 : * expressional index for which we have statistics, then we treat the
3429 : * whole expression as though it were just a Var.
3430 : * 3. If the list contains Vars of different relations that are known equal
3431 : * due to equivalence classes, then drop all but one of the Vars from each
3432 : * known-equal set, keeping the one with smallest estimated # of values
3433 : * (since the extra values of the others can't appear in joined rows).
3434 : * Note the reason we only consider Vars of different relations is that
3435 : * if we considered ones of the same rel, we'd be double-counting the
3436 : * restriction selectivity of the equality in the next step.
3437 : * 4. For Vars within a single source rel, we multiply together the numbers
3438 : * of values, clamp to the number of rows in the rel (divided by 10 if
3439 : * more than one Var), and then multiply by a factor based on the
3440 : * selectivity of the restriction clauses for that rel. When there's
3441 : * more than one Var, the initial product is probably too high (it's the
3442 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3443 : * helpful heuristic for not letting the estimate get out of hand. (The
3444 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3445 : * we multiply by to adjust for the restriction selectivity assumes that
3446 : * the restriction clauses are independent of the grouping, which may not
3447 : * be a valid assumption, but it's hard to do better.
3448 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3449 : * rel, and multiply the results together.
3450 : * Note that rels not containing grouped Vars are ignored completely, as are
3451 : * join clauses. Such rels cannot increase the number of groups, and we
3452 : * assume such clauses do not reduce the number either (somewhat bogus,
3453 : * but we don't have the info to do better).
3454 : */
3455 : double
3456 333956 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3457 : List **pgset, EstimationInfo *estinfo)
3458 : {
3459 333956 : List *varinfos = NIL;
3460 333956 : double srf_multiplier = 1.0;
3461 : double numdistinct;
3462 : ListCell *l;
3463 : int i;
3464 :
3465 : /* Zero the estinfo output parameter, if non-NULL */
3466 333956 : if (estinfo != NULL)
3467 286252 : memset(estinfo, 0, sizeof(EstimationInfo));
3468 :
3469 : /*
3470 : * We don't ever want to return an estimate of zero groups, as that tends
3471 : * to lead to division-by-zero and other unpleasantness. The input_rows
3472 : * estimate is usually already at least 1, but clamp it just in case it
3473 : * isn't.
3474 : */
3475 333956 : input_rows = clamp_row_est(input_rows);
3476 :
3477 : /*
3478 : * If no grouping columns, there's exactly one group. (This can't happen
3479 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3480 : * corner cases with set operations.)
3481 : */
3482 333956 : if (groupExprs == NIL || (pgset && *pgset == NIL))
3483 1074 : return 1.0;
3484 :
3485 : /*
3486 : * Count groups derived from boolean grouping expressions. For other
3487 : * expressions, find the unique Vars used, treating an expression as a Var
3488 : * if we can find stats for it. For each one, record the statistical
3489 : * estimate of number of distinct values (total in its table, without
3490 : * regard for filtering).
3491 : */
3492 332882 : numdistinct = 1.0;
3493 :
3494 332882 : i = 0;
3495 715678 : foreach(l, groupExprs)
3496 : {
3497 382844 : Node *groupexpr = (Node *) lfirst(l);
3498 : double this_srf_multiplier;
3499 : VariableStatData vardata;
3500 : List *varshere;
3501 : ListCell *l2;
3502 :
3503 : /* is expression in this grouping set? */
3504 382844 : if (pgset && !list_member_int(*pgset, i++))
3505 315930 : continue;
3506 :
3507 : /*
3508 : * Set-returning functions in grouping columns are a bit problematic.
3509 : * The code below will effectively ignore their SRF nature and come up
3510 : * with a numdistinct estimate as though they were scalar functions.
3511 : * We compensate by scaling up the end result by the largest SRF
3512 : * rowcount estimate. (This will be an overestimate if the SRF
3513 : * produces multiple copies of any output value, but it seems best to
3514 : * assume the SRF's outputs are distinct. In any case, it's probably
3515 : * pointless to worry too much about this without much better
3516 : * estimates for SRF output rowcounts than we have today.)
3517 : */
3518 382056 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3519 382056 : if (srf_multiplier < this_srf_multiplier)
3520 132 : srf_multiplier = this_srf_multiplier;
3521 :
3522 : /* Short-circuit for expressions returning boolean */
3523 382056 : if (exprType(groupexpr) == BOOLOID)
3524 : {
3525 208 : numdistinct *= 2.0;
3526 208 : continue;
3527 : }
3528 :
3529 : /*
3530 : * If examine_variable is able to deduce anything about the GROUP BY
3531 : * expression, treat it as a single variable even if it's really more
3532 : * complicated.
3533 : *
3534 : * XXX This has the consequence that if there's a statistics object on
3535 : * the expression, we don't split it into individual Vars. This
3536 : * affects our selection of statistics in
3537 : * estimate_multivariate_ndistinct, because it's probably better to
3538 : * use more accurate estimate for each expression and treat them as
3539 : * independent, than to combine estimates for the extracted variables
3540 : * when we don't know how that relates to the expressions.
3541 : */
3542 381848 : examine_variable(root, groupexpr, 0, &vardata);
3543 381848 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3544 : {
3545 314256 : varinfos = add_unique_group_var(root, varinfos,
3546 : groupexpr, &vardata);
3547 314256 : ReleaseVariableStats(vardata);
3548 314256 : continue;
3549 : }
3550 67592 : ReleaseVariableStats(vardata);
3551 :
3552 : /*
3553 : * Else pull out the component Vars. Handle PlaceHolderVars by
3554 : * recursing into their arguments (effectively assuming that the
3555 : * PlaceHolderVar doesn't change the number of groups, which boils
3556 : * down to ignoring the possible addition of nulls to the result set).
3557 : */
3558 67592 : varshere = pull_var_clause(groupexpr,
3559 : PVC_RECURSE_AGGREGATES |
3560 : PVC_RECURSE_WINDOWFUNCS |
3561 : PVC_RECURSE_PLACEHOLDERS);
3562 :
3563 : /*
3564 : * If we find any variable-free GROUP BY item, then either it is a
3565 : * constant (and we can ignore it) or it contains a volatile function;
3566 : * in the latter case we punt and assume that each input row will
3567 : * yield a distinct group.
3568 : */
3569 67592 : if (varshere == NIL)
3570 : {
3571 726 : if (contain_volatile_functions(groupexpr))
3572 48 : return input_rows;
3573 678 : continue;
3574 : }
3575 :
3576 : /*
3577 : * Else add variables to varinfos list
3578 : */
3579 137156 : foreach(l2, varshere)
3580 : {
3581 70290 : Node *var = (Node *) lfirst(l2);
3582 :
3583 70290 : examine_variable(root, var, 0, &vardata);
3584 70290 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3585 70290 : ReleaseVariableStats(vardata);
3586 : }
3587 : }
3588 :
3589 : /*
3590 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3591 : * list.
3592 : */
3593 332834 : if (varinfos == NIL)
3594 : {
3595 : /* Apply SRF multiplier as we would do in the long path */
3596 398 : numdistinct *= srf_multiplier;
3597 : /* Round off */
3598 398 : numdistinct = ceil(numdistinct);
3599 : /* Guard against out-of-range answers */
3600 398 : if (numdistinct > input_rows)
3601 44 : numdistinct = input_rows;
3602 398 : if (numdistinct < 1.0)
3603 0 : numdistinct = 1.0;
3604 398 : return numdistinct;
3605 : }
3606 :
3607 : /*
3608 : * Group Vars by relation and estimate total numdistinct.
3609 : *
3610 : * For each iteration of the outer loop, we process the frontmost Var in
3611 : * varinfos, plus all other Vars in the same relation. We remove these
3612 : * Vars from the newvarinfos list for the next iteration. This is the
3613 : * easiest way to group Vars of same rel together.
3614 : */
3615 : do
3616 : {
3617 334906 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3618 334906 : RelOptInfo *rel = varinfo1->rel;
3619 334906 : double reldistinct = 1;
3620 334906 : double relmaxndistinct = reldistinct;
3621 334906 : int relvarcount = 0;
3622 334906 : List *newvarinfos = NIL;
3623 334906 : List *relvarinfos = NIL;
3624 :
3625 : /*
3626 : * Split the list of varinfos in two - one for the current rel, one
3627 : * for remaining Vars on other rels.
3628 : */
3629 334906 : relvarinfos = lappend(relvarinfos, varinfo1);
3630 388190 : for_each_from(l, varinfos, 1)
3631 : {
3632 53284 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3633 :
3634 53284 : if (varinfo2->rel == varinfo1->rel)
3635 : {
3636 : /* varinfos on current rel */
3637 48506 : relvarinfos = lappend(relvarinfos, varinfo2);
3638 : }
3639 : else
3640 : {
3641 : /* not time to process varinfo2 yet */
3642 4778 : newvarinfos = lappend(newvarinfos, varinfo2);
3643 : }
3644 : }
3645 :
3646 : /*
3647 : * Get the numdistinct estimate for the Vars of this rel. We
3648 : * iteratively search for multivariate n-distinct with maximum number
3649 : * of vars; assuming that each var group is independent of the others,
3650 : * we multiply them together. Any remaining relvarinfos after no more
3651 : * multivariate matches are found are assumed independent too, so
3652 : * their individual ndistinct estimates are multiplied also.
3653 : *
3654 : * While iterating, count how many separate numdistinct values we
3655 : * apply. We apply a fudge factor below, but only if we multiplied
3656 : * more than one such values.
3657 : */
3658 669938 : while (relvarinfos)
3659 : {
3660 : double mvndistinct;
3661 :
3662 335032 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3663 : &mvndistinct))
3664 : {
3665 414 : reldistinct *= mvndistinct;
3666 414 : if (relmaxndistinct < mvndistinct)
3667 402 : relmaxndistinct = mvndistinct;
3668 414 : relvarcount++;
3669 : }
3670 : else
3671 : {
3672 717154 : foreach(l, relvarinfos)
3673 : {
3674 382536 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3675 :
3676 382536 : reldistinct *= varinfo2->ndistinct;
3677 382536 : if (relmaxndistinct < varinfo2->ndistinct)
3678 336300 : relmaxndistinct = varinfo2->ndistinct;
3679 382536 : relvarcount++;
3680 :
3681 : /*
3682 : * When varinfo2's isdefault is set then we'd better set
3683 : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3684 : */
3685 382536 : if (estinfo != NULL && varinfo2->isdefault)
3686 17944 : estinfo->flags |= SELFLAG_USED_DEFAULT;
3687 : }
3688 :
3689 : /* we're done with this relation */
3690 334618 : relvarinfos = NIL;
3691 : }
3692 : }
3693 :
3694 : /*
3695 : * Sanity check --- don't divide by zero if empty relation.
3696 : */
3697 : Assert(IS_SIMPLE_REL(rel));
3698 334906 : if (rel->tuples > 0)
3699 : {
3700 : /*
3701 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
3702 : * fudge factor is because the Vars are probably correlated but we
3703 : * don't know by how much. We should never clamp to less than the
3704 : * largest ndistinct value for any of the Vars, though, since
3705 : * there will surely be at least that many groups.
3706 : */
3707 333860 : double clamp = rel->tuples;
3708 :
3709 333860 : if (relvarcount > 1)
3710 : {
3711 43734 : clamp *= 0.1;
3712 43734 : if (clamp < relmaxndistinct)
3713 : {
3714 41074 : clamp = relmaxndistinct;
3715 : /* for sanity in case some ndistinct is too large: */
3716 41074 : if (clamp > rel->tuples)
3717 78 : clamp = rel->tuples;
3718 : }
3719 : }
3720 333860 : if (reldistinct > clamp)
3721 35694 : reldistinct = clamp;
3722 :
3723 : /*
3724 : * Update the estimate based on the restriction selectivity,
3725 : * guarding against division by zero when reldistinct is zero.
3726 : * Also skip this if we know that we are returning all rows.
3727 : */
3728 333860 : if (reldistinct > 0 && rel->rows < rel->tuples)
3729 : {
3730 : /*
3731 : * Given a table containing N rows with n distinct values in a
3732 : * uniform distribution, if we select p rows at random then
3733 : * the expected number of distinct values selected is
3734 : *
3735 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
3736 : *
3737 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
3738 : *
3739 : * See "Approximating block accesses in database
3740 : * organizations", S. B. Yao, Communications of the ACM,
3741 : * Volume 20 Issue 4, April 1977 Pages 260-261.
3742 : *
3743 : * Alternatively, re-arranging the terms from the factorials,
3744 : * this may be written as
3745 : *
3746 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
3747 : *
3748 : * This form of the formula is more efficient to compute in
3749 : * the common case where p is larger than N/n. Additionally,
3750 : * as pointed out by Dell'Era, if i << N for all terms in the
3751 : * product, it can be approximated by
3752 : *
3753 : * n * (1 - ((N-p)/N)^(N/n))
3754 : *
3755 : * See "Expected distinct values when selecting from a bag
3756 : * without replacement", Alberto Dell'Era,
3757 : * http://www.adellera.it/investigations/distinct_balls/.
3758 : *
3759 : * The condition i << N is equivalent to n >> 1, so this is a
3760 : * good approximation when the number of distinct values in
3761 : * the table is large. It turns out that this formula also
3762 : * works well even when n is small.
3763 : */
3764 104284 : reldistinct *=
3765 104284 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
3766 104284 : rel->tuples / reldistinct));
3767 : }
3768 333860 : reldistinct = clamp_row_est(reldistinct);
3769 :
3770 : /*
3771 : * Update estimate of total distinct groups.
3772 : */
3773 333860 : numdistinct *= reldistinct;
3774 : }
3775 :
3776 334906 : varinfos = newvarinfos;
3777 334906 : } while (varinfos != NIL);
3778 :
3779 : /* Now we can account for the effects of any SRFs */
3780 332436 : numdistinct *= srf_multiplier;
3781 :
3782 : /* Round off */
3783 332436 : numdistinct = ceil(numdistinct);
3784 :
3785 : /* Guard against out-of-range answers */
3786 332436 : if (numdistinct > input_rows)
3787 70780 : numdistinct = input_rows;
3788 332436 : if (numdistinct < 1.0)
3789 0 : numdistinct = 1.0;
3790 :
3791 332436 : return numdistinct;
3792 : }
3793 :
3794 : /*
3795 : * Try to estimate the bucket size of the hash join inner side when the join
3796 : * condition contains two or more clauses by employing extended statistics.
3797 : *
3798 : * The main idea of this approach is that the distinct value generated by
3799 : * multivariate estimation on two or more columns would provide less bucket size
3800 : * than estimation on one separate column.
3801 : *
3802 : * IMPORTANT: It is crucial to synchronize the approach of combining different
3803 : * estimations with the caller's method.
3804 : *
3805 : * Return a list of clauses that didn't fetch any extended statistics.
3806 : */
3807 : List *
3808 449228 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
3809 : List *hashclauses,
3810 : Selectivity *innerbucketsize)
3811 : {
3812 : List *clauses;
3813 : List *otherclauses;
3814 : double ndistinct;
3815 :
3816 449228 : if (list_length(hashclauses) <= 1)
3817 : {
3818 : /*
3819 : * Nothing to do for a single clause. Could we employ univariate
3820 : * extended stat here?
3821 : */
3822 412668 : return hashclauses;
3823 : }
3824 :
3825 : /* "clauses" is the list of hashclauses we've not dealt with yet */
3826 36560 : clauses = list_copy(hashclauses);
3827 : /* "otherclauses" holds clauses we are going to return to caller */
3828 36560 : otherclauses = NIL;
3829 : /* current estimate of ndistinct */
3830 36560 : ndistinct = 1.0;
3831 73132 : while (clauses != NIL)
3832 : {
3833 : ListCell *lc;
3834 36572 : int relid = -1;
3835 36572 : List *varinfos = NIL;
3836 36572 : List *origin_rinfos = NIL;
3837 : double mvndistinct;
3838 : List *origin_varinfos;
3839 36572 : int group_relid = -1;
3840 36572 : RelOptInfo *group_rel = NULL;
3841 : ListCell *lc1,
3842 : *lc2;
3843 :
3844 : /*
3845 : * Find clauses, referencing the same single base relation and try to
3846 : * estimate such a group with extended statistics. Create varinfo for
3847 : * an approved clause, push it to otherclauses, if it can't be
3848 : * estimated here or ignore to process at the next iteration.
3849 : */
3850 110328 : foreach(lc, clauses)
3851 : {
3852 73756 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
3853 : Node *expr;
3854 : Relids relids;
3855 : GroupVarInfo *varinfo;
3856 :
3857 : /*
3858 : * Find the inner side of the join, which we need to estimate the
3859 : * number of buckets. Use outer_is_left because the
3860 : * clause_sides_match_join routine has called on hash clauses.
3861 : */
3862 147512 : relids = rinfo->outer_is_left ?
3863 73756 : rinfo->right_relids : rinfo->left_relids;
3864 147512 : expr = rinfo->outer_is_left ?
3865 73756 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
3866 :
3867 73756 : if (bms_get_singleton_member(relids, &relid) &&
3868 71134 : root->simple_rel_array[relid]->statlist != NIL)
3869 48 : {
3870 60 : bool is_duplicate = false;
3871 :
3872 : /*
3873 : * This inner-side expression references only one relation.
3874 : * Extended statistics on this clause can exist.
3875 : */
3876 60 : if (group_relid < 0)
3877 : {
3878 30 : RangeTblEntry *rte = root->simple_rte_array[relid];
3879 :
3880 30 : if (!rte || (rte->relkind != RELKIND_RELATION &&
3881 0 : rte->relkind != RELKIND_MATVIEW &&
3882 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
3883 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
3884 : {
3885 : /* Extended statistics can't exist in principle */
3886 0 : otherclauses = lappend(otherclauses, rinfo);
3887 0 : clauses = foreach_delete_current(clauses, lc);
3888 0 : continue;
3889 : }
3890 :
3891 30 : group_relid = relid;
3892 30 : group_rel = root->simple_rel_array[relid];
3893 : }
3894 30 : else if (group_relid != relid)
3895 : {
3896 : /*
3897 : * Being in the group forming state we don't need other
3898 : * clauses.
3899 : */
3900 0 : continue;
3901 : }
3902 :
3903 : /*
3904 : * We're going to add the new clause to the varinfos list. We
3905 : * might re-use add_unique_group_var(), but we don't do so for
3906 : * two reasons.
3907 : *
3908 : * 1) We must keep the origin_rinfos list ordered exactly the
3909 : * same way as varinfos.
3910 : *
3911 : * 2) add_unique_group_var() is designed for
3912 : * estimate_num_groups(), where a larger number of groups is
3913 : * worse. While estimating the number of hash buckets, we
3914 : * have the opposite: a lesser number of groups is worse.
3915 : * Therefore, we don't have to remove "known equal" vars: the
3916 : * removed var may valuably contribute to the multivariate
3917 : * statistics to grow the number of groups.
3918 : */
3919 :
3920 : /*
3921 : * Clear nullingrels to correctly match hash keys. See
3922 : * add_unique_group_var()'s comment for details.
3923 : */
3924 60 : expr = remove_nulling_relids(expr, root->outer_join_rels, NULL);
3925 :
3926 : /*
3927 : * Detect and exclude exact duplicates from the list of hash
3928 : * keys (like add_unique_group_var does).
3929 : */
3930 84 : foreach(lc1, varinfos)
3931 : {
3932 36 : varinfo = (GroupVarInfo *) lfirst(lc1);
3933 :
3934 36 : if (!equal(expr, varinfo->var))
3935 24 : continue;
3936 :
3937 12 : is_duplicate = true;
3938 12 : break;
3939 : }
3940 :
3941 60 : if (is_duplicate)
3942 : {
3943 : /*
3944 : * Skip exact duplicates. Adding them to the otherclauses
3945 : * list also doesn't make sense.
3946 : */
3947 12 : continue;
3948 : }
3949 :
3950 : /*
3951 : * Initialize GroupVarInfo. We only use it to call
3952 : * estimate_multivariate_ndistinct(), which doesn't care about
3953 : * ndistinct and isdefault fields. Thus, skip these fields.
3954 : */
3955 48 : varinfo = (GroupVarInfo *) palloc0(sizeof(GroupVarInfo));
3956 48 : varinfo->var = expr;
3957 48 : varinfo->rel = root->simple_rel_array[relid];
3958 48 : varinfos = lappend(varinfos, varinfo);
3959 :
3960 : /*
3961 : * Remember the link to RestrictInfo for the case the clause
3962 : * is failed to be estimated.
3963 : */
3964 48 : origin_rinfos = lappend(origin_rinfos, rinfo);
3965 : }
3966 : else
3967 : {
3968 : /* This clause can't be estimated with extended statistics */
3969 73696 : otherclauses = lappend(otherclauses, rinfo);
3970 : }
3971 :
3972 73744 : clauses = foreach_delete_current(clauses, lc);
3973 : }
3974 :
3975 36572 : if (list_length(varinfos) < 2)
3976 : {
3977 : /*
3978 : * Multivariate statistics doesn't apply to single columns except
3979 : * for expressions, but it has not been implemented yet.
3980 : */
3981 36560 : otherclauses = list_concat(otherclauses, origin_rinfos);
3982 36560 : list_free_deep(varinfos);
3983 36560 : list_free(origin_rinfos);
3984 36560 : continue;
3985 : }
3986 :
3987 : Assert(group_rel != NULL);
3988 :
3989 : /* Employ the extended statistics. */
3990 12 : origin_varinfos = varinfos;
3991 : for (;;)
3992 12 : {
3993 24 : bool estimated = estimate_multivariate_ndistinct(root,
3994 : group_rel,
3995 : &varinfos,
3996 : &mvndistinct);
3997 :
3998 24 : if (!estimated)
3999 12 : break;
4000 :
4001 : /*
4002 : * We've got an estimation. Use ndistinct value in a consistent
4003 : * way - according to the caller's logic (see
4004 : * final_cost_hashjoin).
4005 : */
4006 12 : if (ndistinct < mvndistinct)
4007 12 : ndistinct = mvndistinct;
4008 : Assert(ndistinct >= 1.0);
4009 : }
4010 :
4011 : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
4012 :
4013 : /* Collect unmatched clauses as otherclauses. */
4014 42 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
4015 : {
4016 30 : GroupVarInfo *vinfo = lfirst(lc1);
4017 :
4018 30 : if (!list_member_ptr(varinfos, vinfo))
4019 : /* Already estimated */
4020 30 : continue;
4021 :
4022 : /* Can't be estimated here - push to the returning list */
4023 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
4024 : }
4025 : }
4026 :
4027 36560 : *innerbucketsize = 1.0 / ndistinct;
4028 36560 : return otherclauses;
4029 : }
4030 :
4031 : /*
4032 : * Estimate hash bucket statistics when the specified expression is used
4033 : * as a hash key for the given number of buckets.
4034 : *
4035 : * This attempts to determine two values:
4036 : *
4037 : * 1. The frequency of the most common value of the expression (returns
4038 : * zero into *mcv_freq if we can't get that).
4039 : *
4040 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
4041 : * divided by total tuples in relation.
4042 : *
4043 : * XXX This is really pretty bogus since we're effectively assuming that the
4044 : * distribution of hash keys will be the same after applying restriction
4045 : * clauses as it was in the underlying relation. However, we are not nearly
4046 : * smart enough to figure out how the restrict clauses might change the
4047 : * distribution, so this will have to do for now.
4048 : *
4049 : * We are passed the number of buckets the executor will use for the given
4050 : * input relation. If the data were perfectly distributed, with the same
4051 : * number of tuples going into each available bucket, then the bucketsize
4052 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
4053 : * only if (a) there are at least nbuckets distinct data values, and (b)
4054 : * we have a not-too-skewed data distribution. Otherwise the buckets will
4055 : * be nonuniformly occupied. If the other relation in the join has a key
4056 : * distribution similar to this one's, then the most-loaded buckets are
4057 : * exactly those that will be probed most often. Therefore, the "average"
4058 : * bucket size for costing purposes should really be taken as something close
4059 : * to the "worst case" bucket size. We try to estimate this by adjusting the
4060 : * fraction if there are too few distinct data values, and then scaling up
4061 : * by the ratio of the most common value's frequency to the average frequency.
4062 : *
4063 : * If no statistics are available, use a default estimate of 0.1. This will
4064 : * discourage use of a hash rather strongly if the inner relation is large,
4065 : * which is what we want. We do not want to hash unless we know that the
4066 : * inner rel is well-dispersed (or the alternatives seem much worse).
4067 : *
4068 : * The caller should also check that the mcv_freq is not so large that the
4069 : * most common value would by itself require an impractically large bucket.
4070 : * In a hash join, the executor can split buckets if they get too big, but
4071 : * obviously that doesn't help for a bucket that contains many duplicates of
4072 : * the same value.
4073 : */
4074 : void
4075 204728 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
4076 : Selectivity *mcv_freq,
4077 : Selectivity *bucketsize_frac)
4078 : {
4079 : VariableStatData vardata;
4080 : double estfract,
4081 : ndistinct,
4082 : stanullfrac,
4083 : avgfreq;
4084 : bool isdefault;
4085 : AttStatsSlot sslot;
4086 :
4087 204728 : examine_variable(root, hashkey, 0, &vardata);
4088 :
4089 : /* Look up the frequency of the most common value, if available */
4090 204728 : *mcv_freq = 0.0;
4091 :
4092 204728 : if (HeapTupleIsValid(vardata.statsTuple))
4093 : {
4094 146846 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4095 : STATISTIC_KIND_MCV, InvalidOid,
4096 : ATTSTATSSLOT_NUMBERS))
4097 : {
4098 : /*
4099 : * The first MCV stat is for the most common value.
4100 : */
4101 84902 : if (sslot.nnumbers > 0)
4102 84902 : *mcv_freq = sslot.numbers[0];
4103 84902 : free_attstatsslot(&sslot);
4104 : }
4105 : }
4106 :
4107 : /* Get number of distinct values */
4108 204728 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4109 :
4110 : /*
4111 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4112 : * mcv_freq is known to be even higher than that, use it instead.
4113 : */
4114 204728 : if (isdefault)
4115 : {
4116 25566 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4117 25566 : ReleaseVariableStats(vardata);
4118 25566 : return;
4119 : }
4120 :
4121 : /* Get fraction that are null */
4122 179162 : if (HeapTupleIsValid(vardata.statsTuple))
4123 : {
4124 : Form_pg_statistic stats;
4125 :
4126 146828 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
4127 146828 : stanullfrac = stats->stanullfrac;
4128 : }
4129 : else
4130 32334 : stanullfrac = 0.0;
4131 :
4132 : /* Compute avg freq of all distinct data values in raw relation */
4133 179162 : avgfreq = (1.0 - stanullfrac) / ndistinct;
4134 :
4135 : /*
4136 : * Adjust ndistinct to account for restriction clauses. Observe we are
4137 : * assuming that the data distribution is affected uniformly by the
4138 : * restriction clauses!
4139 : *
4140 : * XXX Possibly better way, but much more expensive: multiply by
4141 : * selectivity of rel's restriction clauses that mention the target Var.
4142 : */
4143 179162 : if (vardata.rel && vardata.rel->tuples > 0)
4144 : {
4145 179116 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
4146 179116 : ndistinct = clamp_row_est(ndistinct);
4147 : }
4148 :
4149 : /*
4150 : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
4151 : * number of buckets is less than the expected number of distinct values;
4152 : * otherwise it is 1/ndistinct.
4153 : */
4154 179162 : if (ndistinct > nbuckets)
4155 84 : estfract = 1.0 / nbuckets;
4156 : else
4157 179078 : estfract = 1.0 / ndistinct;
4158 :
4159 : /*
4160 : * Adjust estimated bucketsize upward to account for skewed distribution.
4161 : */
4162 179162 : if (avgfreq > 0.0 && *mcv_freq > avgfreq)
4163 77706 : estfract *= *mcv_freq / avgfreq;
4164 :
4165 : /*
4166 : * Clamp bucketsize to sane range (the above adjustment could easily
4167 : * produce an out-of-range result). We set the lower bound a little above
4168 : * zero, since zero isn't a very sane result.
4169 : */
4170 179162 : if (estfract < 1.0e-6)
4171 0 : estfract = 1.0e-6;
4172 179162 : else if (estfract > 1.0)
4173 34502 : estfract = 1.0;
4174 :
4175 179162 : *bucketsize_frac = (Selectivity) estfract;
4176 :
4177 179162 : ReleaseVariableStats(vardata);
4178 : }
4179 :
4180 : /*
4181 : * estimate_hashagg_tablesize
4182 : * estimate the number of bytes that a hash aggregate hashtable will
4183 : * require based on the agg_costs, path width and number of groups.
4184 : *
4185 : * We return the result as "double" to forestall any possible overflow
4186 : * problem in the multiplication by dNumGroups.
4187 : *
4188 : * XXX this may be over-estimating the size now that hashagg knows to omit
4189 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4190 : * grouping columns not in the hashed set are counted here even though hashagg
4191 : * won't store them. Is this a problem?
4192 : */
4193 : double
4194 2350 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4195 : const AggClauseCosts *agg_costs, double dNumGroups)
4196 : {
4197 : Size hashentrysize;
4198 :
4199 2350 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4200 2350 : path->pathtarget->width,
4201 2350 : agg_costs->transitionSpace);
4202 :
4203 : /*
4204 : * Note that this disregards the effect of fill-factor and growth policy
4205 : * of the hash table. That's probably ok, given that the default
4206 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4207 : * "double-in-size" growth policies here.
4208 : */
4209 2350 : return hashentrysize * dNumGroups;
4210 : }
4211 :
4212 :
4213 : /*-------------------------------------------------------------------------
4214 : *
4215 : * Support routines
4216 : *
4217 : *-------------------------------------------------------------------------
4218 : */
4219 :
4220 : /*
4221 : * Find the best matching ndistinct extended statistics for the given list of
4222 : * GroupVarInfos.
4223 : *
4224 : * Callers must ensure that the given GroupVarInfos all belong to 'rel' and
4225 : * the GroupVarInfos list does not contain any duplicate Vars or expressions.
4226 : *
4227 : * When statistics are found that match > 1 of the given GroupVarInfo, the
4228 : * *ndistinct parameter is set according to the ndistinct estimate and a new
4229 : * list is built with the matching GroupVarInfos removed, which is output via
4230 : * the *varinfos parameter before returning true. When no matching stats are
4231 : * found, false is returned and the *varinfos and *ndistinct parameters are
4232 : * left untouched.
4233 : */
4234 : static bool
4235 335056 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4236 : List **varinfos, double *ndistinct)
4237 : {
4238 : ListCell *lc;
4239 : int nmatches_vars;
4240 : int nmatches_exprs;
4241 335056 : Oid statOid = InvalidOid;
4242 : MVNDistinct *stats;
4243 335056 : StatisticExtInfo *matched_info = NULL;
4244 335056 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4245 :
4246 : /* bail out immediately if the table has no extended statistics */
4247 335056 : if (!rel->statlist)
4248 334492 : return false;
4249 :
4250 : /* look for the ndistinct statistics object matching the most vars */
4251 564 : nmatches_vars = 0; /* we require at least two matches */
4252 564 : nmatches_exprs = 0;
4253 2244 : foreach(lc, rel->statlist)
4254 : {
4255 : ListCell *lc2;
4256 1680 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
4257 1680 : int nshared_vars = 0;
4258 1680 : int nshared_exprs = 0;
4259 :
4260 : /* skip statistics of other kinds */
4261 1680 : if (info->kind != STATS_EXT_NDISTINCT)
4262 792 : continue;
4263 :
4264 : /* skip statistics with mismatching stxdinherit value */
4265 888 : if (info->inherit != rte->inh)
4266 30 : continue;
4267 :
4268 : /*
4269 : * Determine how many expressions (and variables in non-matched
4270 : * expressions) match. We'll then use these numbers to pick the
4271 : * statistics object that best matches the clauses.
4272 : */
4273 2718 : foreach(lc2, *varinfos)
4274 : {
4275 : ListCell *lc3;
4276 1860 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4277 : AttrNumber attnum;
4278 :
4279 : Assert(varinfo->rel == rel);
4280 :
4281 : /* simple Var, search in statistics keys directly */
4282 1860 : if (IsA(varinfo->var, Var))
4283 : {
4284 1494 : attnum = ((Var *) varinfo->var)->varattno;
4285 :
4286 : /*
4287 : * Ignore system attributes - we don't support statistics on
4288 : * them, so can't match them (and it'd fail as the values are
4289 : * negative).
4290 : */
4291 1494 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4292 12 : continue;
4293 :
4294 1482 : if (bms_is_member(attnum, info->keys))
4295 876 : nshared_vars++;
4296 :
4297 1482 : continue;
4298 : }
4299 :
4300 : /* expression - see if it's in the statistics object */
4301 660 : foreach(lc3, info->exprs)
4302 : {
4303 528 : Node *expr = (Node *) lfirst(lc3);
4304 :
4305 528 : if (equal(varinfo->var, expr))
4306 : {
4307 234 : nshared_exprs++;
4308 234 : break;
4309 : }
4310 : }
4311 : }
4312 :
4313 : /*
4314 : * The ndistinct extended statistics contain estimates for a minimum
4315 : * of pairs of columns which the statistics are defined on and
4316 : * certainly not single columns. Here we skip unless we managed to
4317 : * match to at least two columns.
4318 : */
4319 858 : if (nshared_vars + nshared_exprs < 2)
4320 396 : continue;
4321 :
4322 : /*
4323 : * Check if these statistics are a better match than the previous best
4324 : * match and if so, take note of the StatisticExtInfo.
4325 : *
4326 : * The statslist is sorted by statOid, so the StatisticExtInfo we
4327 : * select as the best match is deterministic even when multiple sets
4328 : * of statistics match equally as well.
4329 : */
4330 462 : if ((nshared_exprs > nmatches_exprs) ||
4331 354 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4332 : {
4333 438 : statOid = info->statOid;
4334 438 : nmatches_vars = nshared_vars;
4335 438 : nmatches_exprs = nshared_exprs;
4336 438 : matched_info = info;
4337 : }
4338 : }
4339 :
4340 : /* No match? */
4341 564 : if (statOid == InvalidOid)
4342 138 : return false;
4343 :
4344 : Assert(nmatches_vars + nmatches_exprs > 1);
4345 :
4346 426 : stats = statext_ndistinct_load(statOid, rte->inh);
4347 :
4348 : /*
4349 : * If we have a match, search it for the specific item that matches (there
4350 : * must be one), and construct the output values.
4351 : */
4352 426 : if (stats)
4353 : {
4354 : int i;
4355 426 : List *newlist = NIL;
4356 426 : MVNDistinctItem *item = NULL;
4357 : ListCell *lc2;
4358 426 : Bitmapset *matched = NULL;
4359 : AttrNumber attnum_offset;
4360 :
4361 : /*
4362 : * How much we need to offset the attnums? If there are no
4363 : * expressions, no offset is needed. Otherwise offset enough to move
4364 : * the lowest one (which is equal to number of expressions) to 1.
4365 : */
4366 426 : if (matched_info->exprs)
4367 150 : attnum_offset = (list_length(matched_info->exprs) + 1);
4368 : else
4369 276 : attnum_offset = 0;
4370 :
4371 : /* see what actually matched */
4372 1488 : foreach(lc2, *varinfos)
4373 : {
4374 : ListCell *lc3;
4375 : int idx;
4376 1062 : bool found = false;
4377 :
4378 1062 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4379 :
4380 : /*
4381 : * Process a simple Var expression, by matching it to keys
4382 : * directly. If there's a matching expression, we'll try matching
4383 : * it later.
4384 : */
4385 1062 : if (IsA(varinfo->var, Var))
4386 : {
4387 876 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4388 :
4389 : /*
4390 : * Ignore expressions on system attributes. Can't rely on the
4391 : * bms check for negative values.
4392 : */
4393 876 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4394 6 : continue;
4395 :
4396 : /* Is the variable covered by the statistics object? */
4397 870 : if (!bms_is_member(attnum, matched_info->keys))
4398 120 : continue;
4399 :
4400 750 : attnum = attnum + attnum_offset;
4401 :
4402 : /* ensure sufficient offset */
4403 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4404 :
4405 750 : matched = bms_add_member(matched, attnum);
4406 :
4407 750 : found = true;
4408 : }
4409 :
4410 : /*
4411 : * XXX Maybe we should allow searching the expressions even if we
4412 : * found an attribute matching the expression? That would handle
4413 : * trivial expressions like "(a)" but it seems fairly useless.
4414 : */
4415 936 : if (found)
4416 750 : continue;
4417 :
4418 : /* expression - see if it's in the statistics object */
4419 186 : idx = 0;
4420 306 : foreach(lc3, matched_info->exprs)
4421 : {
4422 276 : Node *expr = (Node *) lfirst(lc3);
4423 :
4424 276 : if (equal(varinfo->var, expr))
4425 : {
4426 156 : AttrNumber attnum = -(idx + 1);
4427 :
4428 156 : attnum = attnum + attnum_offset;
4429 :
4430 : /* ensure sufficient offset */
4431 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4432 :
4433 156 : matched = bms_add_member(matched, attnum);
4434 :
4435 : /* there should be just one matching expression */
4436 156 : break;
4437 : }
4438 :
4439 120 : idx++;
4440 : }
4441 : }
4442 :
4443 : /* Find the specific item that exactly matches the combination */
4444 864 : for (i = 0; i < stats->nitems; i++)
4445 : {
4446 : int j;
4447 864 : MVNDistinctItem *tmpitem = &stats->items[i];
4448 :
4449 864 : if (tmpitem->nattributes != bms_num_members(matched))
4450 162 : continue;
4451 :
4452 : /* assume it's the right item */
4453 702 : item = tmpitem;
4454 :
4455 : /* check that all item attributes/expressions fit the match */
4456 1692 : for (j = 0; j < tmpitem->nattributes; j++)
4457 : {
4458 1266 : AttrNumber attnum = tmpitem->attributes[j];
4459 :
4460 : /*
4461 : * Thanks to how we constructed the matched bitmap above, we
4462 : * can just offset all attnums the same way.
4463 : */
4464 1266 : attnum = attnum + attnum_offset;
4465 :
4466 1266 : if (!bms_is_member(attnum, matched))
4467 : {
4468 : /* nah, it's not this item */
4469 276 : item = NULL;
4470 276 : break;
4471 : }
4472 : }
4473 :
4474 : /*
4475 : * If the item has all the matched attributes, we know it's the
4476 : * right one - there can't be a better one. matching more.
4477 : */
4478 702 : if (item)
4479 426 : break;
4480 : }
4481 :
4482 : /*
4483 : * Make sure we found an item. There has to be one, because ndistinct
4484 : * statistics includes all combinations of attributes.
4485 : */
4486 426 : if (!item)
4487 0 : elog(ERROR, "corrupt MVNDistinct entry");
4488 :
4489 : /* Form the output varinfo list, keeping only unmatched ones */
4490 1488 : foreach(lc, *varinfos)
4491 : {
4492 1062 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4493 : ListCell *lc3;
4494 1062 : bool found = false;
4495 :
4496 : /*
4497 : * Let's look at plain variables first, because it's the most
4498 : * common case and the check is quite cheap. We can simply get the
4499 : * attnum and check (with an offset) matched bitmap.
4500 : */
4501 1062 : if (IsA(varinfo->var, Var))
4502 870 : {
4503 876 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4504 :
4505 : /*
4506 : * If it's a system attribute, we're done. We don't support
4507 : * extended statistics on system attributes, so it's clearly
4508 : * not matched. Just keep the expression and continue.
4509 : */
4510 876 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4511 : {
4512 6 : newlist = lappend(newlist, varinfo);
4513 6 : continue;
4514 : }
4515 :
4516 : /* apply the same offset as above */
4517 870 : attnum += attnum_offset;
4518 :
4519 : /* if it's not matched, keep the varinfo */
4520 870 : if (!bms_is_member(attnum, matched))
4521 120 : newlist = lappend(newlist, varinfo);
4522 :
4523 : /* The rest of the loop deals with complex expressions. */
4524 870 : continue;
4525 : }
4526 :
4527 : /*
4528 : * Process complex expressions, not just simple Vars.
4529 : *
4530 : * First, we search for an exact match of an expression. If we
4531 : * find one, we can just discard the whole GroupVarInfo, with all
4532 : * the variables we extracted from it.
4533 : *
4534 : * Otherwise we inspect the individual vars, and try matching it
4535 : * to variables in the item.
4536 : */
4537 306 : foreach(lc3, matched_info->exprs)
4538 : {
4539 276 : Node *expr = (Node *) lfirst(lc3);
4540 :
4541 276 : if (equal(varinfo->var, expr))
4542 : {
4543 156 : found = true;
4544 156 : break;
4545 : }
4546 : }
4547 :
4548 : /* found exact match, skip */
4549 186 : if (found)
4550 156 : continue;
4551 :
4552 30 : newlist = lappend(newlist, varinfo);
4553 : }
4554 :
4555 426 : *varinfos = newlist;
4556 426 : *ndistinct = item->ndistinct;
4557 426 : return true;
4558 : }
4559 :
4560 0 : return false;
4561 : }
4562 :
4563 : /*
4564 : * convert_to_scalar
4565 : * Convert non-NULL values of the indicated types to the comparison
4566 : * scale needed by scalarineqsel().
4567 : * Returns "true" if successful.
4568 : *
4569 : * XXX this routine is a hack: ideally we should look up the conversion
4570 : * subroutines in pg_type.
4571 : *
4572 : * All numeric datatypes are simply converted to their equivalent
4573 : * "double" values. (NUMERIC values that are outside the range of "double"
4574 : * are clamped to +/- HUGE_VAL.)
4575 : *
4576 : * String datatypes are converted by convert_string_to_scalar(),
4577 : * which is explained below. The reason why this routine deals with
4578 : * three values at a time, not just one, is that we need it for strings.
4579 : *
4580 : * The bytea datatype is just enough different from strings that it has
4581 : * to be treated separately.
4582 : *
4583 : * The several datatypes representing absolute times are all converted
4584 : * to Timestamp, which is actually an int64, and then we promote that to
4585 : * a double. Note this will give correct results even for the "special"
4586 : * values of Timestamp, since those are chosen to compare correctly;
4587 : * see timestamp_cmp.
4588 : *
4589 : * The several datatypes representing relative times (intervals) are all
4590 : * converted to measurements expressed in seconds.
4591 : */
4592 : static bool
4593 87070 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4594 : Datum lobound, Datum hibound, Oid boundstypid,
4595 : double *scaledlobound, double *scaledhibound)
4596 : {
4597 87070 : bool failure = false;
4598 :
4599 : /*
4600 : * Both the valuetypid and the boundstypid should exactly match the
4601 : * declared input type(s) of the operator we are invoked for. However,
4602 : * extensions might try to use scalarineqsel as estimator for operators
4603 : * with input type(s) we don't handle here; in such cases, we want to
4604 : * return false, not fail. In any case, we mustn't assume that valuetypid
4605 : * and boundstypid are identical.
4606 : *
4607 : * XXX The histogram we are interpolating between points of could belong
4608 : * to a column that's only binary-compatible with the declared type. In
4609 : * essence we are assuming that the semantics of binary-compatible types
4610 : * are enough alike that we can use a histogram generated with one type's
4611 : * operators to estimate selectivity for the other's. This is outright
4612 : * wrong in some cases --- in particular signed versus unsigned
4613 : * interpretation could trip us up. But it's useful enough in the
4614 : * majority of cases that we do it anyway. Should think about more
4615 : * rigorous ways to do it.
4616 : */
4617 87070 : switch (valuetypid)
4618 : {
4619 : /*
4620 : * Built-in numeric types
4621 : */
4622 80118 : case BOOLOID:
4623 : case INT2OID:
4624 : case INT4OID:
4625 : case INT8OID:
4626 : case FLOAT4OID:
4627 : case FLOAT8OID:
4628 : case NUMERICOID:
4629 : case OIDOID:
4630 : case REGPROCOID:
4631 : case REGPROCEDUREOID:
4632 : case REGOPEROID:
4633 : case REGOPERATOROID:
4634 : case REGCLASSOID:
4635 : case REGTYPEOID:
4636 : case REGCOLLATIONOID:
4637 : case REGCONFIGOID:
4638 : case REGDICTIONARYOID:
4639 : case REGROLEOID:
4640 : case REGNAMESPACEOID:
4641 : case REGDATABASEOID:
4642 80118 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4643 : &failure);
4644 80118 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4645 : &failure);
4646 80118 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4647 : &failure);
4648 80118 : return !failure;
4649 :
4650 : /*
4651 : * Built-in string types
4652 : */
4653 6952 : case CHAROID:
4654 : case BPCHAROID:
4655 : case VARCHAROID:
4656 : case TEXTOID:
4657 : case NAMEOID:
4658 : {
4659 6952 : char *valstr = convert_string_datum(value, valuetypid,
4660 : collid, &failure);
4661 6952 : char *lostr = convert_string_datum(lobound, boundstypid,
4662 : collid, &failure);
4663 6952 : char *histr = convert_string_datum(hibound, boundstypid,
4664 : collid, &failure);
4665 :
4666 : /*
4667 : * Bail out if any of the values is not of string type. We
4668 : * might leak converted strings for the other value(s), but
4669 : * that's not worth troubling over.
4670 : */
4671 6952 : if (failure)
4672 0 : return false;
4673 :
4674 6952 : convert_string_to_scalar(valstr, scaledvalue,
4675 : lostr, scaledlobound,
4676 : histr, scaledhibound);
4677 6952 : pfree(valstr);
4678 6952 : pfree(lostr);
4679 6952 : pfree(histr);
4680 6952 : return true;
4681 : }
4682 :
4683 : /*
4684 : * Built-in bytea type
4685 : */
4686 0 : case BYTEAOID:
4687 : {
4688 : /* We only support bytea vs bytea comparison */
4689 0 : if (boundstypid != BYTEAOID)
4690 0 : return false;
4691 0 : convert_bytea_to_scalar(value, scaledvalue,
4692 : lobound, scaledlobound,
4693 : hibound, scaledhibound);
4694 0 : return true;
4695 : }
4696 :
4697 : /*
4698 : * Built-in time types
4699 : */
4700 0 : case TIMESTAMPOID:
4701 : case TIMESTAMPTZOID:
4702 : case DATEOID:
4703 : case INTERVALOID:
4704 : case TIMEOID:
4705 : case TIMETZOID:
4706 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
4707 : &failure);
4708 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
4709 : &failure);
4710 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
4711 : &failure);
4712 0 : return !failure;
4713 :
4714 : /*
4715 : * Built-in network types
4716 : */
4717 0 : case INETOID:
4718 : case CIDROID:
4719 : case MACADDROID:
4720 : case MACADDR8OID:
4721 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
4722 : &failure);
4723 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
4724 : &failure);
4725 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
4726 : &failure);
4727 0 : return !failure;
4728 : }
4729 : /* Don't know how to convert */
4730 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
4731 0 : return false;
4732 : }
4733 :
4734 : /*
4735 : * Do convert_to_scalar()'s work for any numeric data type.
4736 : *
4737 : * On failure (e.g., unsupported typid), set *failure to true;
4738 : * otherwise, that variable is not changed.
4739 : */
4740 : static double
4741 240354 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
4742 : {
4743 240354 : switch (typid)
4744 : {
4745 0 : case BOOLOID:
4746 0 : return (double) DatumGetBool(value);
4747 12 : case INT2OID:
4748 12 : return (double) DatumGetInt16(value);
4749 31770 : case INT4OID:
4750 31770 : return (double) DatumGetInt32(value);
4751 0 : case INT8OID:
4752 0 : return (double) DatumGetInt64(value);
4753 0 : case FLOAT4OID:
4754 0 : return (double) DatumGetFloat4(value);
4755 54 : case FLOAT8OID:
4756 54 : return (double) DatumGetFloat8(value);
4757 0 : case NUMERICOID:
4758 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
4759 0 : return (double)
4760 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
4761 : value));
4762 208518 : case OIDOID:
4763 : case REGPROCOID:
4764 : case REGPROCEDUREOID:
4765 : case REGOPEROID:
4766 : case REGOPERATOROID:
4767 : case REGCLASSOID:
4768 : case REGTYPEOID:
4769 : case REGCOLLATIONOID:
4770 : case REGCONFIGOID:
4771 : case REGDICTIONARYOID:
4772 : case REGROLEOID:
4773 : case REGNAMESPACEOID:
4774 : case REGDATABASEOID:
4775 : /* we can treat OIDs as integers... */
4776 208518 : return (double) DatumGetObjectId(value);
4777 : }
4778 :
4779 0 : *failure = true;
4780 0 : return 0;
4781 : }
4782 :
4783 : /*
4784 : * Do convert_to_scalar()'s work for any character-string data type.
4785 : *
4786 : * String datatypes are converted to a scale that ranges from 0 to 1,
4787 : * where we visualize the bytes of the string as fractional digits.
4788 : *
4789 : * We do not want the base to be 256, however, since that tends to
4790 : * generate inflated selectivity estimates; few databases will have
4791 : * occurrences of all 256 possible byte values at each position.
4792 : * Instead, use the smallest and largest byte values seen in the bounds
4793 : * as the estimated range for each byte, after some fudging to deal with
4794 : * the fact that we probably aren't going to see the full range that way.
4795 : *
4796 : * An additional refinement is that we discard any common prefix of the
4797 : * three strings before computing the scaled values. This allows us to
4798 : * "zoom in" when we encounter a narrow data range. An example is a phone
4799 : * number database where all the values begin with the same area code.
4800 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
4801 : * so this is more likely to happen than you might think.)
4802 : */
4803 : static void
4804 6952 : convert_string_to_scalar(char *value,
4805 : double *scaledvalue,
4806 : char *lobound,
4807 : double *scaledlobound,
4808 : char *hibound,
4809 : double *scaledhibound)
4810 : {
4811 : int rangelo,
4812 : rangehi;
4813 : char *sptr;
4814 :
4815 6952 : rangelo = rangehi = (unsigned char) hibound[0];
4816 86568 : for (sptr = lobound; *sptr; sptr++)
4817 : {
4818 79616 : if (rangelo > (unsigned char) *sptr)
4819 16218 : rangelo = (unsigned char) *sptr;
4820 79616 : if (rangehi < (unsigned char) *sptr)
4821 8740 : rangehi = (unsigned char) *sptr;
4822 : }
4823 85430 : for (sptr = hibound; *sptr; sptr++)
4824 : {
4825 78478 : if (rangelo > (unsigned char) *sptr)
4826 1280 : rangelo = (unsigned char) *sptr;
4827 78478 : if (rangehi < (unsigned char) *sptr)
4828 3338 : rangehi = (unsigned char) *sptr;
4829 : }
4830 : /* If range includes any upper-case ASCII chars, make it include all */
4831 6952 : if (rangelo <= 'Z' && rangehi >= 'A')
4832 : {
4833 1202 : if (rangelo > 'A')
4834 222 : rangelo = 'A';
4835 1202 : if (rangehi < 'Z')
4836 480 : rangehi = 'Z';
4837 : }
4838 : /* Ditto lower-case */
4839 6952 : if (rangelo <= 'z' && rangehi >= 'a')
4840 : {
4841 6450 : if (rangelo > 'a')
4842 30 : rangelo = 'a';
4843 6450 : if (rangehi < 'z')
4844 6368 : rangehi = 'z';
4845 : }
4846 : /* Ditto digits */
4847 6952 : if (rangelo <= '9' && rangehi >= '0')
4848 : {
4849 522 : if (rangelo > '0')
4850 412 : rangelo = '0';
4851 522 : if (rangehi < '9')
4852 14 : rangehi = '9';
4853 : }
4854 :
4855 : /*
4856 : * If range includes less than 10 chars, assume we have not got enough
4857 : * data, and make it include regular ASCII set.
4858 : */
4859 6952 : if (rangehi - rangelo < 9)
4860 : {
4861 0 : rangelo = ' ';
4862 0 : rangehi = 127;
4863 : }
4864 :
4865 : /*
4866 : * Now strip any common prefix of the three strings.
4867 : */
4868 14538 : while (*lobound)
4869 : {
4870 14538 : if (*lobound != *hibound || *lobound != *value)
4871 : break;
4872 7586 : lobound++, hibound++, value++;
4873 : }
4874 :
4875 : /*
4876 : * Now we can do the conversions.
4877 : */
4878 6952 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
4879 6952 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
4880 6952 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
4881 6952 : }
4882 :
4883 : static double
4884 20856 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
4885 : {
4886 20856 : int slen = strlen(value);
4887 : double num,
4888 : denom,
4889 : base;
4890 :
4891 20856 : if (slen <= 0)
4892 0 : return 0.0; /* empty string has scalar value 0 */
4893 :
4894 : /*
4895 : * There seems little point in considering more than a dozen bytes from
4896 : * the string. Since base is at least 10, that will give us nominal
4897 : * resolution of at least 12 decimal digits, which is surely far more
4898 : * precision than this estimation technique has got anyway (especially in
4899 : * non-C locales). Also, even with the maximum possible base of 256, this
4900 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
4901 : * overflow on any known machine.
4902 : */
4903 20856 : if (slen > 12)
4904 5482 : slen = 12;
4905 :
4906 : /* Convert initial characters to fraction */
4907 20856 : base = rangehi - rangelo + 1;
4908 20856 : num = 0.0;
4909 20856 : denom = base;
4910 173768 : while (slen-- > 0)
4911 : {
4912 152912 : int ch = (unsigned char) *value++;
4913 :
4914 152912 : if (ch < rangelo)
4915 212 : ch = rangelo - 1;
4916 152700 : else if (ch > rangehi)
4917 0 : ch = rangehi + 1;
4918 152912 : num += ((double) (ch - rangelo)) / denom;
4919 152912 : denom *= base;
4920 : }
4921 :
4922 20856 : return num;
4923 : }
4924 :
4925 : /*
4926 : * Convert a string-type Datum into a palloc'd, null-terminated string.
4927 : *
4928 : * On failure (e.g., unsupported typid), set *failure to true;
4929 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
4930 : *
4931 : * When using a non-C locale, we must pass the string through pg_strxfrm()
4932 : * before continuing, so as to generate correct locale-specific results.
4933 : */
4934 : static char *
4935 20856 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
4936 : {
4937 : char *val;
4938 : pg_locale_t mylocale;
4939 :
4940 20856 : switch (typid)
4941 : {
4942 0 : case CHAROID:
4943 0 : val = (char *) palloc(2);
4944 0 : val[0] = DatumGetChar(value);
4945 0 : val[1] = '\0';
4946 0 : break;
4947 6416 : case BPCHAROID:
4948 : case VARCHAROID:
4949 : case TEXTOID:
4950 6416 : val = TextDatumGetCString(value);
4951 6416 : break;
4952 14440 : case NAMEOID:
4953 : {
4954 14440 : NameData *nm = (NameData *) DatumGetPointer(value);
4955 :
4956 14440 : val = pstrdup(NameStr(*nm));
4957 14440 : break;
4958 : }
4959 0 : default:
4960 0 : *failure = true;
4961 0 : return NULL;
4962 : }
4963 :
4964 20856 : mylocale = pg_newlocale_from_collation(collid);
4965 :
4966 20856 : if (!mylocale->collate_is_c)
4967 : {
4968 : char *xfrmstr;
4969 : size_t xfrmlen;
4970 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
4971 :
4972 : /*
4973 : * XXX: We could guess at a suitable output buffer size and only call
4974 : * pg_strxfrm() twice if our guess is too small.
4975 : *
4976 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
4977 : * bogus data or set an error. This is not really a problem unless it
4978 : * crashes since it will only give an estimation error and nothing
4979 : * fatal.
4980 : *
4981 : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
4982 : * some cases, libc strxfrm() may return the wrong results, but that
4983 : * will only lead to an estimation error.
4984 : */
4985 72 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
4986 : #ifdef WIN32
4987 :
4988 : /*
4989 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
4990 : * of trying to allocate this much memory (and fail), just return the
4991 : * original string unmodified as if we were in the C locale.
4992 : */
4993 : if (xfrmlen == INT_MAX)
4994 : return val;
4995 : #endif
4996 72 : xfrmstr = (char *) palloc(xfrmlen + 1);
4997 72 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
4998 :
4999 : /*
5000 : * Some systems (e.g., glibc) can return a smaller value from the
5001 : * second call than the first; thus the Assert must be <= not ==.
5002 : */
5003 : Assert(xfrmlen2 <= xfrmlen);
5004 72 : pfree(val);
5005 72 : val = xfrmstr;
5006 : }
5007 :
5008 20856 : return val;
5009 : }
5010 :
5011 : /*
5012 : * Do convert_to_scalar()'s work for any bytea data type.
5013 : *
5014 : * Very similar to convert_string_to_scalar except we can't assume
5015 : * null-termination and therefore pass explicit lengths around.
5016 : *
5017 : * Also, assumptions about likely "normal" ranges of characters have been
5018 : * removed - a data range of 0..255 is always used, for now. (Perhaps
5019 : * someday we will add information about actual byte data range to
5020 : * pg_statistic.)
5021 : */
5022 : static void
5023 0 : convert_bytea_to_scalar(Datum value,
5024 : double *scaledvalue,
5025 : Datum lobound,
5026 : double *scaledlobound,
5027 : Datum hibound,
5028 : double *scaledhibound)
5029 : {
5030 0 : bytea *valuep = DatumGetByteaPP(value);
5031 0 : bytea *loboundp = DatumGetByteaPP(lobound);
5032 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
5033 : int rangelo,
5034 : rangehi,
5035 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
5036 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
5037 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
5038 : i,
5039 : minlen;
5040 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
5041 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
5042 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
5043 :
5044 : /*
5045 : * Assume bytea data is uniformly distributed across all byte values.
5046 : */
5047 0 : rangelo = 0;
5048 0 : rangehi = 255;
5049 :
5050 : /*
5051 : * Now strip any common prefix of the three strings.
5052 : */
5053 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
5054 0 : for (i = 0; i < minlen; i++)
5055 : {
5056 0 : if (*lostr != *histr || *lostr != *valstr)
5057 : break;
5058 0 : lostr++, histr++, valstr++;
5059 0 : loboundlen--, hiboundlen--, valuelen--;
5060 : }
5061 :
5062 : /*
5063 : * Now we can do the conversions.
5064 : */
5065 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
5066 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
5067 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
5068 0 : }
5069 :
5070 : static double
5071 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
5072 : int rangelo, int rangehi)
5073 : {
5074 : double num,
5075 : denom,
5076 : base;
5077 :
5078 0 : if (valuelen <= 0)
5079 0 : return 0.0; /* empty string has scalar value 0 */
5080 :
5081 : /*
5082 : * Since base is 256, need not consider more than about 10 chars (even
5083 : * this many seems like overkill)
5084 : */
5085 0 : if (valuelen > 10)
5086 0 : valuelen = 10;
5087 :
5088 : /* Convert initial characters to fraction */
5089 0 : base = rangehi - rangelo + 1;
5090 0 : num = 0.0;
5091 0 : denom = base;
5092 0 : while (valuelen-- > 0)
5093 : {
5094 0 : int ch = *value++;
5095 :
5096 0 : if (ch < rangelo)
5097 0 : ch = rangelo - 1;
5098 0 : else if (ch > rangehi)
5099 0 : ch = rangehi + 1;
5100 0 : num += ((double) (ch - rangelo)) / denom;
5101 0 : denom *= base;
5102 : }
5103 :
5104 0 : return num;
5105 : }
5106 :
5107 : /*
5108 : * Do convert_to_scalar()'s work for any timevalue data type.
5109 : *
5110 : * On failure (e.g., unsupported typid), set *failure to true;
5111 : * otherwise, that variable is not changed.
5112 : */
5113 : static double
5114 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5115 : {
5116 0 : switch (typid)
5117 : {
5118 0 : case TIMESTAMPOID:
5119 0 : return DatumGetTimestamp(value);
5120 0 : case TIMESTAMPTZOID:
5121 0 : return DatumGetTimestampTz(value);
5122 0 : case DATEOID:
5123 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
5124 0 : case INTERVALOID:
5125 : {
5126 0 : Interval *interval = DatumGetIntervalP(value);
5127 :
5128 : /*
5129 : * Convert the month part of Interval to days using assumed
5130 : * average month length of 365.25/12.0 days. Not too
5131 : * accurate, but plenty good enough for our purposes.
5132 : *
5133 : * This also works for infinite intervals, which just have all
5134 : * fields set to INT_MIN/INT_MAX, and so will produce a result
5135 : * smaller/larger than any finite interval.
5136 : */
5137 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5138 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5139 : }
5140 0 : case TIMEOID:
5141 0 : return DatumGetTimeADT(value);
5142 0 : case TIMETZOID:
5143 : {
5144 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5145 :
5146 : /* use GMT-equivalent time */
5147 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5148 : }
5149 : }
5150 :
5151 0 : *failure = true;
5152 0 : return 0;
5153 : }
5154 :
5155 :
5156 : /*
5157 : * get_restriction_variable
5158 : * Examine the args of a restriction clause to see if it's of the
5159 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5160 : * where "variable" could be either a Var or an expression in vars of a
5161 : * single relation. If so, extract information about the variable,
5162 : * and also indicate which side it was on and the other argument.
5163 : *
5164 : * Inputs:
5165 : * root: the planner info
5166 : * args: clause argument list
5167 : * varRelid: see specs for restriction selectivity functions
5168 : *
5169 : * Outputs: (these are valid only if true is returned)
5170 : * *vardata: gets information about variable (see examine_variable)
5171 : * *other: gets other clause argument, aggressively reduced to a constant
5172 : * *varonleft: set true if variable is on the left, false if on the right
5173 : *
5174 : * Returns true if a variable is identified, otherwise false.
5175 : *
5176 : * Note: if there are Vars on both sides of the clause, we must fail, because
5177 : * callers are expecting that the other side will act like a pseudoconstant.
5178 : */
5179 : bool
5180 801052 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5181 : VariableStatData *vardata, Node **other,
5182 : bool *varonleft)
5183 : {
5184 : Node *left,
5185 : *right;
5186 : VariableStatData rdata;
5187 :
5188 : /* Fail if not a binary opclause (probably shouldn't happen) */
5189 801052 : if (list_length(args) != 2)
5190 0 : return false;
5191 :
5192 801052 : left = (Node *) linitial(args);
5193 801052 : right = (Node *) lsecond(args);
5194 :
5195 : /*
5196 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5197 : * relations will be treated as pseudoconstants.
5198 : */
5199 801052 : examine_variable(root, left, varRelid, vardata);
5200 801052 : examine_variable(root, right, varRelid, &rdata);
5201 :
5202 : /*
5203 : * If one side is a variable and the other not, we win.
5204 : */
5205 801052 : if (vardata->rel && rdata.rel == NULL)
5206 : {
5207 714308 : *varonleft = true;
5208 714308 : *other = estimate_expression_value(root, rdata.var);
5209 : /* Assume we need no ReleaseVariableStats(rdata) here */
5210 714302 : return true;
5211 : }
5212 :
5213 86744 : if (vardata->rel == NULL && rdata.rel)
5214 : {
5215 80288 : *varonleft = false;
5216 80288 : *other = estimate_expression_value(root, vardata->var);
5217 : /* Assume we need no ReleaseVariableStats(*vardata) here */
5218 80288 : *vardata = rdata;
5219 80288 : return true;
5220 : }
5221 :
5222 : /* Oops, clause has wrong structure (probably var op var) */
5223 6456 : ReleaseVariableStats(*vardata);
5224 6456 : ReleaseVariableStats(rdata);
5225 :
5226 6456 : return false;
5227 : }
5228 :
5229 : /*
5230 : * get_join_variables
5231 : * Apply examine_variable() to each side of a join clause.
5232 : * Also, attempt to identify whether the join clause has the same
5233 : * or reversed sense compared to the SpecialJoinInfo.
5234 : *
5235 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5236 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5237 : * where we can't tell for sure, we default to assuming it's normal.
5238 : */
5239 : void
5240 264682 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5241 : VariableStatData *vardata1, VariableStatData *vardata2,
5242 : bool *join_is_reversed)
5243 : {
5244 : Node *left,
5245 : *right;
5246 :
5247 264682 : if (list_length(args) != 2)
5248 0 : elog(ERROR, "join operator should take two arguments");
5249 :
5250 264682 : left = (Node *) linitial(args);
5251 264682 : right = (Node *) lsecond(args);
5252 :
5253 264682 : examine_variable(root, left, 0, vardata1);
5254 264682 : examine_variable(root, right, 0, vardata2);
5255 :
5256 529184 : if (vardata1->rel &&
5257 264502 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
5258 89904 : *join_is_reversed = true; /* var1 is on RHS */
5259 349410 : else if (vardata2->rel &&
5260 174632 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
5261 132 : *join_is_reversed = true; /* var2 is on LHS */
5262 : else
5263 174646 : *join_is_reversed = false;
5264 264682 : }
5265 :
5266 : /* statext_expressions_load copies the tuple, so just pfree it. */
5267 : static void
5268 1650 : ReleaseDummy(HeapTuple tuple)
5269 : {
5270 1650 : pfree(tuple);
5271 1650 : }
5272 :
5273 : /*
5274 : * examine_variable
5275 : * Try to look up statistical data about an expression.
5276 : * Fill in a VariableStatData struct to describe the expression.
5277 : *
5278 : * Inputs:
5279 : * root: the planner info
5280 : * node: the expression tree to examine
5281 : * varRelid: see specs for restriction selectivity functions
5282 : *
5283 : * Outputs: *vardata is filled as follows:
5284 : * var: the input expression (with any binary relabeling stripped, if
5285 : * it is or contains a variable; but otherwise the type is preserved)
5286 : * rel: RelOptInfo for relation containing variable; NULL if expression
5287 : * contains no Vars (NOTE this could point to a RelOptInfo of a
5288 : * subquery, not one in the current query).
5289 : * statsTuple: the pg_statistic entry for the variable, if one exists;
5290 : * otherwise NULL.
5291 : * freefunc: pointer to a function to release statsTuple with.
5292 : * vartype: exposed type of the expression; this should always match
5293 : * the declared input type of the operator we are estimating for.
5294 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5295 : * commonly the same as the exposed type of the variable argument,
5296 : * but can be different in binary-compatible-type cases.
5297 : * isunique: true if we were able to match the var to a unique index, a
5298 : * single-column DISTINCT or GROUP-BY clause, implying its values are
5299 : * unique for this query. (Caution: this should be trusted for
5300 : * statistical purposes only, since we do not check indimmediate nor
5301 : * verify that the exact same definition of equality applies.)
5302 : * acl_ok: true if current user has permission to read all table rows from
5303 : * the column(s) underlying the pg_statistic entry. This is consulted by
5304 : * statistic_proc_security_check().
5305 : *
5306 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5307 : */
5308 : void
5309 3223598 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5310 : VariableStatData *vardata)
5311 : {
5312 : Node *basenode;
5313 : Relids varnos;
5314 : Relids basevarnos;
5315 : RelOptInfo *onerel;
5316 :
5317 : /* Make sure we don't return dangling pointers in vardata */
5318 22565186 : MemSet(vardata, 0, sizeof(VariableStatData));
5319 :
5320 : /* Save the exposed type of the expression */
5321 3223598 : vardata->vartype = exprType(node);
5322 :
5323 : /* Look inside any binary-compatible relabeling */
5324 :
5325 3223598 : if (IsA(node, RelabelType))
5326 46926 : basenode = (Node *) ((RelabelType *) node)->arg;
5327 : else
5328 3176672 : basenode = node;
5329 :
5330 : /* Fast path for a simple Var */
5331 :
5332 3223598 : if (IsA(basenode, Var) &&
5333 778710 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5334 : {
5335 2309834 : Var *var = (Var *) basenode;
5336 :
5337 : /* Set up result fields other than the stats tuple */
5338 2309834 : vardata->var = basenode; /* return Var without relabeling */
5339 2309834 : vardata->rel = find_base_rel(root, var->varno);
5340 2309834 : vardata->atttype = var->vartype;
5341 2309834 : vardata->atttypmod = var->vartypmod;
5342 2309834 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5343 :
5344 : /* Try to locate some stats */
5345 2309834 : examine_simple_variable(root, var, vardata);
5346 :
5347 2309834 : return;
5348 : }
5349 :
5350 : /*
5351 : * Okay, it's a more complicated expression. Determine variable
5352 : * membership. Note that when varRelid isn't zero, only vars of that
5353 : * relation are considered "real" vars.
5354 : */
5355 913764 : varnos = pull_varnos(root, basenode);
5356 913764 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5357 :
5358 913764 : onerel = NULL;
5359 :
5360 913764 : if (bms_is_empty(basevarnos))
5361 : {
5362 : /* No Vars at all ... must be pseudo-constant clause */
5363 : }
5364 : else
5365 : {
5366 : int relid;
5367 :
5368 : /* Check if the expression is in vars of a single base relation */
5369 466856 : if (bms_get_singleton_member(basevarnos, &relid))
5370 : {
5371 459954 : if (varRelid == 0 || varRelid == relid)
5372 : {
5373 69092 : onerel = find_base_rel(root, relid);
5374 69092 : vardata->rel = onerel;
5375 69092 : node = basenode; /* strip any relabeling */
5376 : }
5377 : /* else treat it as a constant */
5378 : }
5379 : else
5380 : {
5381 : /* varnos has multiple relids */
5382 6902 : if (varRelid == 0)
5383 : {
5384 : /* treat it as a variable of a join relation */
5385 5430 : vardata->rel = find_join_rel(root, varnos);
5386 5430 : node = basenode; /* strip any relabeling */
5387 : }
5388 1472 : else if (bms_is_member(varRelid, varnos))
5389 : {
5390 : /* ignore the vars belonging to other relations */
5391 1298 : vardata->rel = find_base_rel(root, varRelid);
5392 1298 : node = basenode; /* strip any relabeling */
5393 : /* note: no point in expressional-index search here */
5394 : }
5395 : /* else treat it as a constant */
5396 : }
5397 : }
5398 :
5399 913764 : bms_free(basevarnos);
5400 :
5401 913764 : vardata->var = node;
5402 913764 : vardata->atttype = exprType(node);
5403 913764 : vardata->atttypmod = exprTypmod(node);
5404 :
5405 913764 : if (onerel)
5406 : {
5407 : /*
5408 : * We have an expression in vars of a single relation. Try to match
5409 : * it to expressional index columns, in hopes of finding some
5410 : * statistics.
5411 : *
5412 : * Note that we consider all index columns including INCLUDE columns,
5413 : * since there could be stats for such columns. But the test for
5414 : * uniqueness needs to be warier.
5415 : *
5416 : * XXX it's conceivable that there are multiple matches with different
5417 : * index opfamilies; if so, we need to pick one that matches the
5418 : * operator we are estimating for. FIXME later.
5419 : */
5420 : ListCell *ilist;
5421 : ListCell *slist;
5422 :
5423 : /*
5424 : * The nullingrels bits within the expression could prevent us from
5425 : * matching it to expressional index columns or to the expressions in
5426 : * extended statistics. So strip them out first.
5427 : */
5428 69092 : if (bms_overlap(varnos, root->outer_join_rels))
5429 3064 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5430 :
5431 148842 : foreach(ilist, onerel->indexlist)
5432 : {
5433 82708 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5434 : ListCell *indexpr_item;
5435 : int pos;
5436 :
5437 82708 : indexpr_item = list_head(index->indexprs);
5438 82708 : if (indexpr_item == NULL)
5439 77818 : continue; /* no expressions here... */
5440 :
5441 6894 : for (pos = 0; pos < index->ncolumns; pos++)
5442 : {
5443 4962 : if (index->indexkeys[pos] == 0)
5444 : {
5445 : Node *indexkey;
5446 :
5447 4890 : if (indexpr_item == NULL)
5448 0 : elog(ERROR, "too few entries in indexprs list");
5449 4890 : indexkey = (Node *) lfirst(indexpr_item);
5450 4890 : if (indexkey && IsA(indexkey, RelabelType))
5451 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5452 4890 : if (equal(node, indexkey))
5453 : {
5454 : /*
5455 : * Found a match ... is it a unique index? Tests here
5456 : * should match has_unique_index().
5457 : */
5458 3594 : if (index->unique &&
5459 438 : index->nkeycolumns == 1 &&
5460 438 : pos == 0 &&
5461 438 : (index->indpred == NIL || index->predOK))
5462 438 : vardata->isunique = true;
5463 :
5464 : /*
5465 : * Has it got stats? We only consider stats for
5466 : * non-partial indexes, since partial indexes probably
5467 : * don't reflect whole-relation statistics; the above
5468 : * check for uniqueness is the only info we take from
5469 : * a partial index.
5470 : *
5471 : * An index stats hook, however, must make its own
5472 : * decisions about what to do with partial indexes.
5473 : */
5474 3594 : if (get_index_stats_hook &&
5475 0 : (*get_index_stats_hook) (root, index->indexoid,
5476 0 : pos + 1, vardata))
5477 : {
5478 : /*
5479 : * The hook took control of acquiring a stats
5480 : * tuple. If it did supply a tuple, it'd better
5481 : * have supplied a freefunc.
5482 : */
5483 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5484 0 : !vardata->freefunc)
5485 0 : elog(ERROR, "no function provided to release variable stats with");
5486 : }
5487 3594 : else if (index->indpred == NIL)
5488 : {
5489 3594 : vardata->statsTuple =
5490 7188 : SearchSysCache3(STATRELATTINH,
5491 : ObjectIdGetDatum(index->indexoid),
5492 3594 : Int16GetDatum(pos + 1),
5493 : BoolGetDatum(false));
5494 3594 : vardata->freefunc = ReleaseSysCache;
5495 :
5496 3594 : if (HeapTupleIsValid(vardata->statsTuple))
5497 : {
5498 : /*
5499 : * Test if user has permission to access all
5500 : * rows from the index's table.
5501 : *
5502 : * For simplicity, we insist on the whole
5503 : * table being selectable, rather than trying
5504 : * to identify which column(s) the index
5505 : * depends on.
5506 : *
5507 : * Note that for an inheritance child,
5508 : * permissions are checked on the inheritance
5509 : * root parent, and whole-table select
5510 : * privilege on the parent doesn't quite
5511 : * guarantee that the user could read all
5512 : * columns of the child. But in practice it's
5513 : * unlikely that any interesting security
5514 : * violation could result from allowing access
5515 : * to the expression index's stats, so we
5516 : * allow it anyway. See similar code in
5517 : * examine_simple_variable() for additional
5518 : * comments.
5519 : */
5520 2958 : vardata->acl_ok =
5521 2958 : all_rows_selectable(root,
5522 2958 : index->rel->relid,
5523 : NULL);
5524 : }
5525 : else
5526 : {
5527 : /* suppress leakproofness checks later */
5528 636 : vardata->acl_ok = true;
5529 : }
5530 : }
5531 3594 : if (vardata->statsTuple)
5532 2958 : break;
5533 : }
5534 1932 : indexpr_item = lnext(index->indexprs, indexpr_item);
5535 : }
5536 : }
5537 4890 : if (vardata->statsTuple)
5538 2958 : break;
5539 : }
5540 :
5541 : /*
5542 : * Search extended statistics for one with a matching expression.
5543 : * There might be multiple ones, so just grab the first one. In the
5544 : * future, we might consider the statistics target (and pick the most
5545 : * accurate statistics) and maybe some other parameters.
5546 : */
5547 73208 : foreach(slist, onerel->statlist)
5548 : {
5549 4404 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5550 4404 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5551 : ListCell *expr_item;
5552 : int pos;
5553 :
5554 : /*
5555 : * Stop once we've found statistics for the expression (either
5556 : * from extended stats, or for an index in the preceding loop).
5557 : */
5558 4404 : if (vardata->statsTuple)
5559 288 : break;
5560 :
5561 : /* skip stats without per-expression stats */
5562 4116 : if (info->kind != STATS_EXT_EXPRESSIONS)
5563 2106 : continue;
5564 :
5565 : /* skip stats with mismatching stxdinherit value */
5566 2010 : if (info->inherit != rte->inh)
5567 6 : continue;
5568 :
5569 2004 : pos = 0;
5570 3306 : foreach(expr_item, info->exprs)
5571 : {
5572 2952 : Node *expr = (Node *) lfirst(expr_item);
5573 :
5574 : Assert(expr);
5575 :
5576 : /* strip RelabelType before comparing it */
5577 2952 : if (expr && IsA(expr, RelabelType))
5578 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5579 :
5580 : /* found a match, see if we can extract pg_statistic row */
5581 2952 : if (equal(node, expr))
5582 : {
5583 : /*
5584 : * XXX Not sure if we should cache the tuple somewhere.
5585 : * Now we just create a new copy every time.
5586 : */
5587 1650 : vardata->statsTuple =
5588 1650 : statext_expressions_load(info->statOid, rte->inh, pos);
5589 :
5590 1650 : vardata->freefunc = ReleaseDummy;
5591 :
5592 : /*
5593 : * Test if user has permission to access all rows from the
5594 : * table.
5595 : *
5596 : * For simplicity, we insist on the whole table being
5597 : * selectable, rather than trying to identify which
5598 : * column(s) the statistics object depends on.
5599 : *
5600 : * Note that for an inheritance child, permissions are
5601 : * checked on the inheritance root parent, and whole-table
5602 : * select privilege on the parent doesn't quite guarantee
5603 : * that the user could read all columns of the child. But
5604 : * in practice it's unlikely that any interesting security
5605 : * violation could result from allowing access to the
5606 : * expression stats, so we allow it anyway. See similar
5607 : * code in examine_simple_variable() for additional
5608 : * comments.
5609 : */
5610 1650 : vardata->acl_ok = all_rows_selectable(root,
5611 : onerel->relid,
5612 : NULL);
5613 :
5614 1650 : break;
5615 : }
5616 :
5617 1302 : pos++;
5618 : }
5619 : }
5620 : }
5621 :
5622 913764 : bms_free(varnos);
5623 : }
5624 :
5625 : /*
5626 : * examine_simple_variable
5627 : * Handle a simple Var for examine_variable
5628 : *
5629 : * This is split out as a subroutine so that we can recurse to deal with
5630 : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
5631 : *
5632 : * We already filled in all the fields of *vardata except for the stats tuple.
5633 : */
5634 : static void
5635 2316078 : examine_simple_variable(PlannerInfo *root, Var *var,
5636 : VariableStatData *vardata)
5637 : {
5638 2316078 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
5639 :
5640 : Assert(IsA(rte, RangeTblEntry));
5641 :
5642 2316078 : if (get_relation_stats_hook &&
5643 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
5644 : {
5645 : /*
5646 : * The hook took control of acquiring a stats tuple. If it did supply
5647 : * a tuple, it'd better have supplied a freefunc.
5648 : */
5649 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5650 0 : !vardata->freefunc)
5651 0 : elog(ERROR, "no function provided to release variable stats with");
5652 : }
5653 2316078 : else if (rte->rtekind == RTE_RELATION)
5654 : {
5655 : /*
5656 : * Plain table or parent of an inheritance appendrel, so look up the
5657 : * column in pg_statistic
5658 : */
5659 2199454 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
5660 : ObjectIdGetDatum(rte->relid),
5661 2199454 : Int16GetDatum(var->varattno),
5662 2199454 : BoolGetDatum(rte->inh));
5663 2199454 : vardata->freefunc = ReleaseSysCache;
5664 :
5665 2199454 : if (HeapTupleIsValid(vardata->statsTuple))
5666 : {
5667 : /*
5668 : * Test if user has permission to read all rows from this column.
5669 : *
5670 : * This requires that the user has the appropriate SELECT
5671 : * privileges and that there are no securityQuals from security
5672 : * barrier views or RLS policies. If that's not the case, then we
5673 : * only permit leakproof functions to be passed pg_statistic data
5674 : * in vardata, otherwise the functions might reveal data that the
5675 : * user doesn't have permission to see --- see
5676 : * statistic_proc_security_check().
5677 : */
5678 1621958 : vardata->acl_ok =
5679 1621958 : all_rows_selectable(root, var->varno,
5680 1621958 : bms_make_singleton(var->varattno - FirstLowInvalidHeapAttributeNumber));
5681 : }
5682 : else
5683 : {
5684 : /* suppress any possible leakproofness checks later */
5685 577496 : vardata->acl_ok = true;
5686 : }
5687 : }
5688 116624 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
5689 106388 : (rte->rtekind == RTE_CTE && !rte->self_reference))
5690 : {
5691 : /*
5692 : * Plain subquery (not one that was converted to an appendrel) or
5693 : * non-recursive CTE. In either case, we can try to find out what the
5694 : * Var refers to within the subquery. We skip this for appendrel and
5695 : * recursive-CTE cases because any column stats we did find would
5696 : * likely not be very relevant.
5697 : */
5698 : PlannerInfo *subroot;
5699 : Query *subquery;
5700 : List *subtlist;
5701 : TargetEntry *ste;
5702 :
5703 : /*
5704 : * Punt if it's a whole-row var rather than a plain column reference.
5705 : */
5706 17298 : if (var->varattno == InvalidAttrNumber)
5707 0 : return;
5708 :
5709 : /*
5710 : * Otherwise, find the subquery's planner subroot.
5711 : */
5712 17298 : if (rte->rtekind == RTE_SUBQUERY)
5713 : {
5714 : RelOptInfo *rel;
5715 :
5716 : /*
5717 : * Fetch RelOptInfo for subquery. Note that we don't change the
5718 : * rel returned in vardata, since caller expects it to be a rel of
5719 : * the caller's query level. Because we might already be
5720 : * recursing, we can't use that rel pointer either, but have to
5721 : * look up the Var's rel afresh.
5722 : */
5723 10236 : rel = find_base_rel(root, var->varno);
5724 :
5725 10236 : subroot = rel->subroot;
5726 : }
5727 : else
5728 : {
5729 : /* CTE case is more difficult */
5730 : PlannerInfo *cteroot;
5731 : Index levelsup;
5732 : int ndx;
5733 : int plan_id;
5734 : ListCell *lc;
5735 :
5736 : /*
5737 : * Find the referenced CTE, and locate the subroot previously made
5738 : * for it.
5739 : */
5740 7062 : levelsup = rte->ctelevelsup;
5741 7062 : cteroot = root;
5742 13312 : while (levelsup-- > 0)
5743 : {
5744 6250 : cteroot = cteroot->parent_root;
5745 6250 : if (!cteroot) /* shouldn't happen */
5746 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
5747 : }
5748 :
5749 : /*
5750 : * Note: cte_plan_ids can be shorter than cteList, if we are still
5751 : * working on planning the CTEs (ie, this is a side-reference from
5752 : * another CTE). So we mustn't use forboth here.
5753 : */
5754 7062 : ndx = 0;
5755 9284 : foreach(lc, cteroot->parse->cteList)
5756 : {
5757 9284 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
5758 :
5759 9284 : if (strcmp(cte->ctename, rte->ctename) == 0)
5760 7062 : break;
5761 2222 : ndx++;
5762 : }
5763 7062 : if (lc == NULL) /* shouldn't happen */
5764 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
5765 7062 : if (ndx >= list_length(cteroot->cte_plan_ids))
5766 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
5767 7062 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
5768 7062 : if (plan_id <= 0)
5769 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
5770 7062 : subroot = list_nth(root->glob->subroots, plan_id - 1);
5771 : }
5772 :
5773 : /* If the subquery hasn't been planned yet, we have to punt */
5774 17298 : if (subroot == NULL)
5775 0 : return;
5776 : Assert(IsA(subroot, PlannerInfo));
5777 :
5778 : /*
5779 : * We must use the subquery parsetree as mangled by the planner, not
5780 : * the raw version from the RTE, because we need a Var that will refer
5781 : * to the subroot's live RelOptInfos. For instance, if any subquery
5782 : * pullup happened during planning, Vars in the targetlist might have
5783 : * gotten replaced, and we need to see the replacement expressions.
5784 : */
5785 17298 : subquery = subroot->parse;
5786 : Assert(IsA(subquery, Query));
5787 :
5788 : /*
5789 : * Punt if subquery uses set operations or grouping sets, as these
5790 : * will mash underlying columns' stats beyond recognition. (Set ops
5791 : * are particularly nasty; if we forged ahead, we would return stats
5792 : * relevant to only the leftmost subselect...) DISTINCT is also
5793 : * problematic, but we check that later because there is a possibility
5794 : * of learning something even with it.
5795 : */
5796 17298 : if (subquery->setOperations ||
5797 15078 : subquery->groupingSets)
5798 2244 : return;
5799 :
5800 : /* Get the subquery output expression referenced by the upper Var */
5801 15054 : if (subquery->returningList)
5802 206 : subtlist = subquery->returningList;
5803 : else
5804 14848 : subtlist = subquery->targetList;
5805 15054 : ste = get_tle_by_resno(subtlist, var->varattno);
5806 15054 : if (ste == NULL || ste->resjunk)
5807 0 : elog(ERROR, "subquery %s does not have attribute %d",
5808 : rte->eref->aliasname, var->varattno);
5809 15054 : var = (Var *) ste->expr;
5810 :
5811 : /*
5812 : * If subquery uses DISTINCT, we can't make use of any stats for the
5813 : * variable ... but, if it's the only DISTINCT column, we are entitled
5814 : * to consider it unique. We do the test this way so that it works
5815 : * for cases involving DISTINCT ON.
5816 : */
5817 15054 : if (subquery->distinctClause)
5818 : {
5819 1838 : if (list_length(subquery->distinctClause) == 1 &&
5820 616 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
5821 308 : vardata->isunique = true;
5822 : /* cannot go further */
5823 1222 : return;
5824 : }
5825 :
5826 : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
5827 13832 : if (subquery->groupClause)
5828 : {
5829 1080 : if (list_length(subquery->groupClause) == 1 &&
5830 450 : targetIsInSortList(ste, InvalidOid, subquery->groupClause))
5831 338 : vardata->isunique = true;
5832 : /* cannot go further */
5833 630 : return;
5834 : }
5835 :
5836 : /*
5837 : * If the sub-query originated from a view with the security_barrier
5838 : * attribute, we must not look at the variable's statistics, though it
5839 : * seems all right to notice the existence of a DISTINCT clause. So
5840 : * stop here.
5841 : *
5842 : * This is probably a harsher restriction than necessary; it's
5843 : * certainly OK for the selectivity estimator (which is a C function,
5844 : * and therefore omnipotent anyway) to look at the statistics. But
5845 : * many selectivity estimators will happily *invoke the operator
5846 : * function* to try to work out a good estimate - and that's not OK.
5847 : * So for now, don't dig down for stats.
5848 : */
5849 13202 : if (rte->security_barrier)
5850 1422 : return;
5851 :
5852 : /* Can only handle a simple Var of subquery's query level */
5853 11780 : if (var && IsA(var, Var) &&
5854 6244 : var->varlevelsup == 0)
5855 : {
5856 : /*
5857 : * OK, recurse into the subquery. Note that the original setting
5858 : * of vardata->isunique (which will surely be false) is left
5859 : * unchanged in this situation. That's what we want, since even
5860 : * if the underlying column is unique, the subquery may have
5861 : * joined to other tables in a way that creates duplicates.
5862 : */
5863 6244 : examine_simple_variable(subroot, var, vardata);
5864 : }
5865 : }
5866 : else
5867 : {
5868 : /*
5869 : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
5870 : * see RTE_JOIN here because join alias Vars have already been
5871 : * flattened.) There's not much we can do with function outputs, but
5872 : * maybe someday try to be smarter about VALUES.
5873 : */
5874 : }
5875 : }
5876 :
5877 : /*
5878 : * all_rows_selectable
5879 : * Test whether the user has permission to select all rows from a given
5880 : * relation.
5881 : *
5882 : * Inputs:
5883 : * root: the planner info
5884 : * varno: the index of the relation (assumed to be an RTE_RELATION)
5885 : * varattnos: the attributes for which permission is required, or NULL if
5886 : * whole-table access is required
5887 : *
5888 : * Returns true if the user has the required select permissions, and there are
5889 : * no securityQuals from security barrier views or RLS policies.
5890 : *
5891 : * Note that if the relation is an inheritance child relation, securityQuals
5892 : * and access permissions are checked against the inheritance root parent (the
5893 : * relation actually mentioned in the query) --- see the comments in
5894 : * expand_single_inheritance_child() for an explanation of why it has to be
5895 : * done this way.
5896 : *
5897 : * If varattnos is non-NULL, its attribute numbers should be offset by
5898 : * FirstLowInvalidHeapAttributeNumber so that system attributes can be
5899 : * checked. If varattnos is NULL, only table-level SELECT privileges are
5900 : * checked, not any column-level privileges.
5901 : *
5902 : * Note: if the relation is accessed via a view, this function actually tests
5903 : * whether the view owner has permission to select from the relation. To
5904 : * ensure that the current user has permission, it is also necessary to check
5905 : * that the current user has permission to select from the view, which we do
5906 : * at planner-startup --- see subquery_planner().
5907 : *
5908 : * This is exported so that other estimation functions can use it.
5909 : */
5910 : bool
5911 1626818 : all_rows_selectable(PlannerInfo *root, Index varno, Bitmapset *varattnos)
5912 : {
5913 1626818 : RelOptInfo *rel = find_base_rel_noerr(root, varno);
5914 1626818 : RangeTblEntry *rte = planner_rt_fetch(varno, root);
5915 : Oid userid;
5916 : int varattno;
5917 :
5918 : Assert(rte->rtekind == RTE_RELATION);
5919 :
5920 : /*
5921 : * Determine the user ID to use for privilege checks (either the current
5922 : * user or the view owner, if we're accessing the table via a view).
5923 : *
5924 : * Normally the relation will have an associated RelOptInfo from which we
5925 : * can find the userid, but it might not if it's a RETURNING Var for an
5926 : * INSERT target relation. In that case use the RTEPermissionInfo
5927 : * associated with the RTE.
5928 : *
5929 : * If we navigate up to a parent relation, we keep using the same userid,
5930 : * since it's the same in all relations of a given inheritance tree.
5931 : */
5932 1626818 : if (rel)
5933 1626776 : userid = rel->userid;
5934 : else
5935 : {
5936 : RTEPermissionInfo *perminfo;
5937 :
5938 42 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
5939 42 : userid = perminfo->checkAsUser;
5940 : }
5941 1626818 : if (!OidIsValid(userid))
5942 1463132 : userid = GetUserId();
5943 :
5944 : /*
5945 : * Permissions and securityQuals must be checked on the table actually
5946 : * mentioned in the query, so if this is an inheritance child, navigate up
5947 : * to the inheritance root parent. If the user can read the whole table
5948 : * or the required columns there, then they can read from the child table
5949 : * too. For per-column checks, we must find out which of the root
5950 : * parent's attributes the child relation's attributes correspond to.
5951 : */
5952 1626818 : if (root->append_rel_array != NULL)
5953 : {
5954 : AppendRelInfo *appinfo;
5955 :
5956 232594 : appinfo = root->append_rel_array[varno];
5957 :
5958 : /*
5959 : * Partitions are mapped to their immediate parent, not the root
5960 : * parent, so must be ready to walk up multiple AppendRelInfos. But
5961 : * stop if we hit a parent that is not RTE_RELATION --- that's a
5962 : * flattened UNION ALL subquery, not an inheritance parent.
5963 : */
5964 434338 : while (appinfo &&
5965 202116 : planner_rt_fetch(appinfo->parent_relid,
5966 202116 : root)->rtekind == RTE_RELATION)
5967 : {
5968 201744 : Bitmapset *parent_varattnos = NULL;
5969 :
5970 : /*
5971 : * For each child attribute, find the corresponding parent
5972 : * attribute. In rare cases, the attribute may be local to the
5973 : * child table, in which case, we've got to live with having no
5974 : * access to this column.
5975 : */
5976 201744 : varattno = -1;
5977 400638 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
5978 : {
5979 : AttrNumber attno;
5980 : AttrNumber parent_attno;
5981 :
5982 198894 : attno = varattno + FirstLowInvalidHeapAttributeNumber;
5983 :
5984 198894 : if (attno == InvalidAttrNumber)
5985 : {
5986 : /*
5987 : * Whole-row reference, so must map each column of the
5988 : * child to the parent table.
5989 : */
5990 36 : for (attno = 1; attno <= appinfo->num_child_cols; attno++)
5991 : {
5992 24 : parent_attno = appinfo->parent_colnos[attno - 1];
5993 24 : if (parent_attno == 0)
5994 0 : return false; /* attr is local to child */
5995 : parent_varattnos =
5996 24 : bms_add_member(parent_varattnos,
5997 : parent_attno - FirstLowInvalidHeapAttributeNumber);
5998 : }
5999 : }
6000 : else
6001 : {
6002 198882 : if (attno < 0)
6003 : {
6004 : /* System attnos are the same in all tables */
6005 0 : parent_attno = attno;
6006 : }
6007 : else
6008 : {
6009 198882 : if (attno > appinfo->num_child_cols)
6010 0 : return false; /* safety check */
6011 198882 : parent_attno = appinfo->parent_colnos[attno - 1];
6012 198882 : if (parent_attno == 0)
6013 0 : return false; /* attr is local to child */
6014 : }
6015 : parent_varattnos =
6016 198882 : bms_add_member(parent_varattnos,
6017 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6018 : }
6019 : }
6020 :
6021 : /* If the parent is itself a child, continue up */
6022 201744 : varno = appinfo->parent_relid;
6023 201744 : varattnos = parent_varattnos;
6024 201744 : appinfo = root->append_rel_array[varno];
6025 : }
6026 :
6027 : /* Perform the access check on this parent rel */
6028 232594 : rte = planner_rt_fetch(varno, root);
6029 : Assert(rte->rtekind == RTE_RELATION);
6030 : }
6031 :
6032 : /*
6033 : * For all rows to be accessible, there must be no securityQuals from
6034 : * security barrier views or RLS policies.
6035 : */
6036 1626818 : if (rte->securityQuals != NIL)
6037 828 : return false;
6038 :
6039 : /*
6040 : * Test for table-level SELECT privilege.
6041 : *
6042 : * If varattnos is non-NULL, this is sufficient to give access to all
6043 : * requested attributes, even for a child table, since we have verified
6044 : * that all required child columns have matching parent columns.
6045 : *
6046 : * If varattnos is NULL (whole-table access requested), this doesn't
6047 : * necessarily guarantee that the user can read all columns of a child
6048 : * table, but we allow it anyway (see comments in examine_variable()) and
6049 : * don't bother checking any column privileges.
6050 : */
6051 1625990 : if (pg_class_aclcheck(rte->relid, userid, ACL_SELECT) == ACLCHECK_OK)
6052 1625538 : return true;
6053 :
6054 452 : if (varattnos == NULL)
6055 12 : return false; /* whole-table access requested */
6056 :
6057 : /*
6058 : * Don't have table-level SELECT privilege, so check per-column
6059 : * privileges.
6060 : */
6061 440 : varattno = -1;
6062 646 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6063 : {
6064 440 : AttrNumber attno = varattno + FirstLowInvalidHeapAttributeNumber;
6065 :
6066 440 : if (attno == InvalidAttrNumber)
6067 : {
6068 : /* Whole-row reference, so must have access to all columns */
6069 6 : if (pg_attribute_aclcheck_all(rte->relid, userid, ACL_SELECT,
6070 : ACLMASK_ALL) != ACLCHECK_OK)
6071 6 : return false;
6072 : }
6073 : else
6074 : {
6075 434 : if (pg_attribute_aclcheck(rte->relid, attno, userid,
6076 : ACL_SELECT) != ACLCHECK_OK)
6077 228 : return false;
6078 : }
6079 : }
6080 :
6081 : /* If we reach here, have all required column privileges */
6082 206 : return true;
6083 : }
6084 :
6085 : /*
6086 : * examine_indexcol_variable
6087 : * Try to look up statistical data about an index column/expression.
6088 : * Fill in a VariableStatData struct to describe the column.
6089 : *
6090 : * Inputs:
6091 : * root: the planner info
6092 : * index: the index whose column we're interested in
6093 : * indexcol: 0-based index column number (subscripts index->indexkeys[])
6094 : *
6095 : * Outputs: *vardata is filled as follows:
6096 : * var: the input expression (with any binary relabeling stripped, if
6097 : * it is or contains a variable; but otherwise the type is preserved)
6098 : * rel: RelOptInfo for table relation containing variable.
6099 : * statsTuple: the pg_statistic entry for the variable, if one exists;
6100 : * otherwise NULL.
6101 : * freefunc: pointer to a function to release statsTuple with.
6102 : *
6103 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
6104 : */
6105 : static void
6106 784014 : examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
6107 : int indexcol, VariableStatData *vardata)
6108 : {
6109 : AttrNumber colnum;
6110 : Oid relid;
6111 :
6112 784014 : if (index->indexkeys[indexcol] != 0)
6113 : {
6114 : /* Simple variable --- look to stats for the underlying table */
6115 781824 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6116 :
6117 : Assert(rte->rtekind == RTE_RELATION);
6118 781824 : relid = rte->relid;
6119 : Assert(relid != InvalidOid);
6120 781824 : colnum = index->indexkeys[indexcol];
6121 781824 : vardata->rel = index->rel;
6122 :
6123 781824 : if (get_relation_stats_hook &&
6124 0 : (*get_relation_stats_hook) (root, rte, colnum, vardata))
6125 : {
6126 : /*
6127 : * The hook took control of acquiring a stats tuple. If it did
6128 : * supply a tuple, it'd better have supplied a freefunc.
6129 : */
6130 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6131 0 : !vardata->freefunc)
6132 0 : elog(ERROR, "no function provided to release variable stats with");
6133 : }
6134 : else
6135 : {
6136 781824 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6137 : ObjectIdGetDatum(relid),
6138 : Int16GetDatum(colnum),
6139 781824 : BoolGetDatum(rte->inh));
6140 781824 : vardata->freefunc = ReleaseSysCache;
6141 : }
6142 : }
6143 : else
6144 : {
6145 : /* Expression --- maybe there are stats for the index itself */
6146 2190 : relid = index->indexoid;
6147 2190 : colnum = indexcol + 1;
6148 :
6149 2190 : if (get_index_stats_hook &&
6150 0 : (*get_index_stats_hook) (root, relid, colnum, vardata))
6151 : {
6152 : /*
6153 : * The hook took control of acquiring a stats tuple. If it did
6154 : * supply a tuple, it'd better have supplied a freefunc.
6155 : */
6156 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6157 0 : !vardata->freefunc)
6158 0 : elog(ERROR, "no function provided to release variable stats with");
6159 : }
6160 : else
6161 : {
6162 2190 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6163 : ObjectIdGetDatum(relid),
6164 : Int16GetDatum(colnum),
6165 : BoolGetDatum(false));
6166 2190 : vardata->freefunc = ReleaseSysCache;
6167 : }
6168 : }
6169 784014 : }
6170 :
6171 : /*
6172 : * Check whether it is permitted to call func_oid passing some of the
6173 : * pg_statistic data in vardata. We allow this if either of the following
6174 : * conditions is met: (1) the user has SELECT privileges on the table or
6175 : * column underlying the pg_statistic data and there are no securityQuals from
6176 : * security barrier views or RLS policies, or (2) the function is marked
6177 : * leakproof.
6178 : */
6179 : bool
6180 1154398 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
6181 : {
6182 1154398 : if (vardata->acl_ok)
6183 1152548 : return true; /* have SELECT privs and no securityQuals */
6184 :
6185 1850 : if (!OidIsValid(func_oid))
6186 0 : return false;
6187 :
6188 1850 : if (get_func_leakproof(func_oid))
6189 916 : return true;
6190 :
6191 934 : ereport(DEBUG2,
6192 : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
6193 : get_func_name(func_oid))));
6194 934 : return false;
6195 : }
6196 :
6197 : /*
6198 : * get_variable_numdistinct
6199 : * Estimate the number of distinct values of a variable.
6200 : *
6201 : * vardata: results of examine_variable
6202 : * *isdefault: set to true if the result is a default rather than based on
6203 : * anything meaningful.
6204 : *
6205 : * NB: be careful to produce a positive integral result, since callers may
6206 : * compare the result to exact integer counts, or might divide by it.
6207 : */
6208 : double
6209 1634506 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
6210 : {
6211 : double stadistinct;
6212 1634506 : double stanullfrac = 0.0;
6213 : double ntuples;
6214 :
6215 1634506 : *isdefault = false;
6216 :
6217 : /*
6218 : * Determine the stadistinct value to use. There are cases where we can
6219 : * get an estimate even without a pg_statistic entry, or can get a better
6220 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
6221 : * (otherwise, assume no nulls, for lack of any better idea).
6222 : */
6223 1634506 : if (HeapTupleIsValid(vardata->statsTuple))
6224 : {
6225 : /* Use the pg_statistic entry */
6226 : Form_pg_statistic stats;
6227 :
6228 1141224 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6229 1141224 : stadistinct = stats->stadistinct;
6230 1141224 : stanullfrac = stats->stanullfrac;
6231 : }
6232 493282 : else if (vardata->vartype == BOOLOID)
6233 : {
6234 : /*
6235 : * Special-case boolean columns: presumably, two distinct values.
6236 : *
6237 : * Are there any other datatypes we should wire in special estimates
6238 : * for?
6239 : */
6240 592 : stadistinct = 2.0;
6241 : }
6242 492690 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6243 : {
6244 : /*
6245 : * If the Var represents a column of a VALUES RTE, assume it's unique.
6246 : * This could of course be very wrong, but it should tend to be true
6247 : * in well-written queries. We could consider examining the VALUES'
6248 : * contents to get some real statistics; but that only works if the
6249 : * entries are all constants, and it would be pretty expensive anyway.
6250 : */
6251 3484 : stadistinct = -1.0; /* unique (and all non null) */
6252 : }
6253 : else
6254 : {
6255 : /*
6256 : * We don't keep statistics for system columns, but in some cases we
6257 : * can infer distinctness anyway.
6258 : */
6259 489206 : if (vardata->var && IsA(vardata->var, Var))
6260 : {
6261 451962 : switch (((Var *) vardata->var)->varattno)
6262 : {
6263 1200 : case SelfItemPointerAttributeNumber:
6264 1200 : stadistinct = -1.0; /* unique (and all non null) */
6265 1200 : break;
6266 26076 : case TableOidAttributeNumber:
6267 26076 : stadistinct = 1.0; /* only 1 value */
6268 26076 : break;
6269 424686 : default:
6270 424686 : stadistinct = 0.0; /* means "unknown" */
6271 424686 : break;
6272 : }
6273 : }
6274 : else
6275 37244 : stadistinct = 0.0; /* means "unknown" */
6276 :
6277 : /*
6278 : * XXX consider using estimate_num_groups on expressions?
6279 : */
6280 : }
6281 :
6282 : /*
6283 : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6284 : * variable, assume it is unique no matter what pg_statistic says; the
6285 : * statistics could be out of date, or we might have found a partial
6286 : * unique index that proves the var is unique for this query. However,
6287 : * we'd better still believe the null-fraction statistic.
6288 : */
6289 1634506 : if (vardata->isunique)
6290 402374 : stadistinct = -1.0 * (1.0 - stanullfrac);
6291 :
6292 : /*
6293 : * If we had an absolute estimate, use that.
6294 : */
6295 1634506 : if (stadistinct > 0.0)
6296 428066 : return clamp_row_est(stadistinct);
6297 :
6298 : /*
6299 : * Otherwise we need to get the relation size; punt if not available.
6300 : */
6301 1206440 : if (vardata->rel == NULL)
6302 : {
6303 416 : *isdefault = true;
6304 416 : return DEFAULT_NUM_DISTINCT;
6305 : }
6306 1206024 : ntuples = vardata->rel->tuples;
6307 1206024 : if (ntuples <= 0.0)
6308 : {
6309 116538 : *isdefault = true;
6310 116538 : return DEFAULT_NUM_DISTINCT;
6311 : }
6312 :
6313 : /*
6314 : * If we had a relative estimate, use that.
6315 : */
6316 1089486 : if (stadistinct < 0.0)
6317 782268 : return clamp_row_est(-stadistinct * ntuples);
6318 :
6319 : /*
6320 : * With no data, estimate ndistinct = ntuples if the table is small, else
6321 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6322 : * that the behavior isn't discontinuous.
6323 : */
6324 307218 : if (ntuples < DEFAULT_NUM_DISTINCT)
6325 148194 : return clamp_row_est(ntuples);
6326 :
6327 159024 : *isdefault = true;
6328 159024 : return DEFAULT_NUM_DISTINCT;
6329 : }
6330 :
6331 : /*
6332 : * get_variable_range
6333 : * Estimate the minimum and maximum value of the specified variable.
6334 : * If successful, store values in *min and *max, and return true.
6335 : * If no data available, return false.
6336 : *
6337 : * sortop is the "<" comparison operator to use. This should generally
6338 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6339 : * The collation must be specified too.
6340 : */
6341 : static bool
6342 247518 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6343 : Oid sortop, Oid collation,
6344 : Datum *min, Datum *max)
6345 : {
6346 247518 : Datum tmin = 0;
6347 247518 : Datum tmax = 0;
6348 247518 : bool have_data = false;
6349 : int16 typLen;
6350 : bool typByVal;
6351 : Oid opfuncoid;
6352 : FmgrInfo opproc;
6353 : AttStatsSlot sslot;
6354 :
6355 : /*
6356 : * XXX It's very tempting to try to use the actual column min and max, if
6357 : * we can get them relatively-cheaply with an index probe. However, since
6358 : * this function is called many times during join planning, that could
6359 : * have unpleasant effects on planning speed. Need more investigation
6360 : * before enabling this.
6361 : */
6362 : #ifdef NOT_USED
6363 : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6364 : return true;
6365 : #endif
6366 :
6367 247518 : if (!HeapTupleIsValid(vardata->statsTuple))
6368 : {
6369 : /* no stats available, so default result */
6370 55882 : return false;
6371 : }
6372 :
6373 : /*
6374 : * If we can't apply the sortop to the stats data, just fail. In
6375 : * principle, if there's a histogram and no MCVs, we could return the
6376 : * histogram endpoints without ever applying the sortop ... but it's
6377 : * probably not worth trying, because whatever the caller wants to do with
6378 : * the endpoints would likely fail the security check too.
6379 : */
6380 191636 : if (!statistic_proc_security_check(vardata,
6381 191636 : (opfuncoid = get_opcode(sortop))))
6382 0 : return false;
6383 :
6384 191636 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6385 :
6386 191636 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6387 :
6388 : /*
6389 : * If there is a histogram with the ordering we want, grab the first and
6390 : * last values.
6391 : */
6392 191636 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6393 : STATISTIC_KIND_HISTOGRAM, sortop,
6394 : ATTSTATSSLOT_VALUES))
6395 : {
6396 119896 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6397 : {
6398 119896 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6399 119896 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6400 119896 : have_data = true;
6401 : }
6402 119896 : free_attstatsslot(&sslot);
6403 : }
6404 :
6405 : /*
6406 : * Otherwise, if there is a histogram with some other ordering, scan it
6407 : * and get the min and max values according to the ordering we want. This
6408 : * of course may not find values that are really extremal according to our
6409 : * ordering, but it beats ignoring available data.
6410 : */
6411 263376 : if (!have_data &&
6412 71740 : get_attstatsslot(&sslot, vardata->statsTuple,
6413 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6414 : ATTSTATSSLOT_VALUES))
6415 : {
6416 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6417 : collation, typLen, typByVal,
6418 : &tmin, &tmax, &have_data);
6419 0 : free_attstatsslot(&sslot);
6420 : }
6421 :
6422 : /*
6423 : * If we have most-common-values info, look for extreme MCVs. This is
6424 : * needed even if we also have a histogram, since the histogram excludes
6425 : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6426 : * be pretty wary of deciding that that is a full representation of the
6427 : * data. Proceed only if the MCVs represent the whole table (to within
6428 : * roundoff error).
6429 : */
6430 191636 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6431 : STATISTIC_KIND_MCV, InvalidOid,
6432 191636 : have_data ? ATTSTATSSLOT_VALUES :
6433 : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6434 : {
6435 107394 : bool use_mcvs = have_data;
6436 :
6437 107394 : if (!have_data)
6438 : {
6439 70336 : double sumcommon = 0.0;
6440 : double nullfrac;
6441 : int i;
6442 :
6443 530108 : for (i = 0; i < sslot.nnumbers; i++)
6444 459772 : sumcommon += sslot.numbers[i];
6445 70336 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6446 70336 : if (sumcommon + nullfrac > 0.99999)
6447 68676 : use_mcvs = true;
6448 : }
6449 :
6450 107394 : if (use_mcvs)
6451 105734 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6452 : collation, typLen, typByVal,
6453 : &tmin, &tmax, &have_data);
6454 107394 : free_attstatsslot(&sslot);
6455 : }
6456 :
6457 191636 : *min = tmin;
6458 191636 : *max = tmax;
6459 191636 : return have_data;
6460 : }
6461 :
6462 : /*
6463 : * get_stats_slot_range: scan sslot for min/max values
6464 : *
6465 : * Subroutine for get_variable_range: update min/max/have_data according
6466 : * to what we find in the statistics array.
6467 : */
6468 : static void
6469 105734 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6470 : Oid collation, int16 typLen, bool typByVal,
6471 : Datum *min, Datum *max, bool *p_have_data)
6472 : {
6473 105734 : Datum tmin = *min;
6474 105734 : Datum tmax = *max;
6475 105734 : bool have_data = *p_have_data;
6476 105734 : bool found_tmin = false;
6477 105734 : bool found_tmax = false;
6478 :
6479 : /* Look up the comparison function, if we didn't already do so */
6480 105734 : if (opproc->fn_oid != opfuncoid)
6481 105734 : fmgr_info(opfuncoid, opproc);
6482 :
6483 : /* Scan all the slot's values */
6484 2534720 : for (int i = 0; i < sslot->nvalues; i++)
6485 : {
6486 2428986 : if (!have_data)
6487 : {
6488 68676 : tmin = tmax = sslot->values[i];
6489 68676 : found_tmin = found_tmax = true;
6490 68676 : *p_have_data = have_data = true;
6491 68676 : continue;
6492 : }
6493 2360310 : if (DatumGetBool(FunctionCall2Coll(opproc,
6494 : collation,
6495 2360310 : sslot->values[i], tmin)))
6496 : {
6497 58244 : tmin = sslot->values[i];
6498 58244 : found_tmin = true;
6499 : }
6500 2360310 : if (DatumGetBool(FunctionCall2Coll(opproc,
6501 : collation,
6502 2360310 : tmax, sslot->values[i])))
6503 : {
6504 259962 : tmax = sslot->values[i];
6505 259962 : found_tmax = true;
6506 : }
6507 : }
6508 :
6509 : /*
6510 : * Copy the slot's values, if we found new extreme values.
6511 : */
6512 105734 : if (found_tmin)
6513 90638 : *min = datumCopy(tmin, typByVal, typLen);
6514 105734 : if (found_tmax)
6515 72532 : *max = datumCopy(tmax, typByVal, typLen);
6516 105734 : }
6517 :
6518 :
6519 : /*
6520 : * get_actual_variable_range
6521 : * Attempt to identify the current *actual* minimum and/or maximum
6522 : * of the specified variable, by looking for a suitable btree index
6523 : * and fetching its low and/or high values.
6524 : * If successful, store values in *min and *max, and return true.
6525 : * (Either pointer can be NULL if that endpoint isn't needed.)
6526 : * If unsuccessful, return false.
6527 : *
6528 : * sortop is the "<" comparison operator to use.
6529 : * collation is the required collation.
6530 : */
6531 : static bool
6532 185374 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6533 : Oid sortop, Oid collation,
6534 : Datum *min, Datum *max)
6535 : {
6536 185374 : bool have_data = false;
6537 185374 : RelOptInfo *rel = vardata->rel;
6538 : RangeTblEntry *rte;
6539 : ListCell *lc;
6540 :
6541 : /* No hope if no relation or it doesn't have indexes */
6542 185374 : if (rel == NULL || rel->indexlist == NIL)
6543 13692 : return false;
6544 : /* If it has indexes it must be a plain relation */
6545 171682 : rte = root->simple_rte_array[rel->relid];
6546 : Assert(rte->rtekind == RTE_RELATION);
6547 :
6548 : /* ignore partitioned tables. Any indexes here are not real indexes */
6549 171682 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6550 876 : return false;
6551 :
6552 : /* Search through the indexes to see if any match our problem */
6553 329746 : foreach(lc, rel->indexlist)
6554 : {
6555 282600 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6556 : ScanDirection indexscandir;
6557 : StrategyNumber strategy;
6558 :
6559 : /* Ignore non-ordering indexes */
6560 282600 : if (index->sortopfamily == NULL)
6561 0 : continue;
6562 :
6563 : /*
6564 : * Ignore partial indexes --- we only want stats that cover the entire
6565 : * relation.
6566 : */
6567 282600 : if (index->indpred != NIL)
6568 288 : continue;
6569 :
6570 : /*
6571 : * The index list might include hypothetical indexes inserted by a
6572 : * get_relation_info hook --- don't try to access them.
6573 : */
6574 282312 : if (index->hypothetical)
6575 0 : continue;
6576 :
6577 : /*
6578 : * The first index column must match the desired variable, sortop, and
6579 : * collation --- but we can use a descending-order index.
6580 : */
6581 282312 : if (collation != index->indexcollations[0])
6582 36280 : continue; /* test first 'cause it's cheapest */
6583 246032 : if (!match_index_to_operand(vardata->var, 0, index))
6584 122372 : continue;
6585 123660 : strategy = get_op_opfamily_strategy(sortop, index->sortopfamily[0]);
6586 123660 : switch (IndexAmTranslateStrategy(strategy, index->relam, index->sortopfamily[0], true))
6587 : {
6588 123660 : case COMPARE_LT:
6589 123660 : if (index->reverse_sort[0])
6590 0 : indexscandir = BackwardScanDirection;
6591 : else
6592 123660 : indexscandir = ForwardScanDirection;
6593 123660 : break;
6594 0 : case COMPARE_GT:
6595 0 : if (index->reverse_sort[0])
6596 0 : indexscandir = ForwardScanDirection;
6597 : else
6598 0 : indexscandir = BackwardScanDirection;
6599 0 : break;
6600 0 : default:
6601 : /* index doesn't match the sortop */
6602 0 : continue;
6603 : }
6604 :
6605 : /*
6606 : * Found a suitable index to extract data from. Set up some data that
6607 : * can be used by both invocations of get_actual_variable_endpoint.
6608 : */
6609 : {
6610 : MemoryContext tmpcontext;
6611 : MemoryContext oldcontext;
6612 : Relation heapRel;
6613 : Relation indexRel;
6614 : TupleTableSlot *slot;
6615 : int16 typLen;
6616 : bool typByVal;
6617 : ScanKeyData scankeys[1];
6618 :
6619 : /* Make sure any cruft gets recycled when we're done */
6620 123660 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
6621 : "get_actual_variable_range workspace",
6622 : ALLOCSET_DEFAULT_SIZES);
6623 123660 : oldcontext = MemoryContextSwitchTo(tmpcontext);
6624 :
6625 : /*
6626 : * Open the table and index so we can read from them. We should
6627 : * already have some type of lock on each.
6628 : */
6629 123660 : heapRel = table_open(rte->relid, NoLock);
6630 123660 : indexRel = index_open(index->indexoid, NoLock);
6631 :
6632 : /* build some stuff needed for indexscan execution */
6633 123660 : slot = table_slot_create(heapRel, NULL);
6634 123660 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6635 :
6636 : /* set up an IS NOT NULL scan key so that we ignore nulls */
6637 123660 : ScanKeyEntryInitialize(&scankeys[0],
6638 : SK_ISNULL | SK_SEARCHNOTNULL,
6639 : 1, /* index col to scan */
6640 : InvalidStrategy, /* no strategy */
6641 : InvalidOid, /* no strategy subtype */
6642 : InvalidOid, /* no collation */
6643 : InvalidOid, /* no reg proc for this */
6644 : (Datum) 0); /* constant */
6645 :
6646 : /* If min is requested ... */
6647 123660 : if (min)
6648 : {
6649 70340 : have_data = get_actual_variable_endpoint(heapRel,
6650 : indexRel,
6651 : indexscandir,
6652 : scankeys,
6653 : typLen,
6654 : typByVal,
6655 : slot,
6656 : oldcontext,
6657 : min);
6658 : }
6659 : else
6660 : {
6661 : /* If min not requested, still want to fetch max */
6662 53320 : have_data = true;
6663 : }
6664 :
6665 : /* If max is requested, and we didn't already fail ... */
6666 123660 : if (max && have_data)
6667 : {
6668 : /* scan in the opposite direction; all else is the same */
6669 54964 : have_data = get_actual_variable_endpoint(heapRel,
6670 : indexRel,
6671 54964 : -indexscandir,
6672 : scankeys,
6673 : typLen,
6674 : typByVal,
6675 : slot,
6676 : oldcontext,
6677 : max);
6678 : }
6679 :
6680 : /* Clean everything up */
6681 123660 : ExecDropSingleTupleTableSlot(slot);
6682 :
6683 123660 : index_close(indexRel, NoLock);
6684 123660 : table_close(heapRel, NoLock);
6685 :
6686 123660 : MemoryContextSwitchTo(oldcontext);
6687 123660 : MemoryContextDelete(tmpcontext);
6688 :
6689 : /* And we're done */
6690 123660 : break;
6691 : }
6692 : }
6693 :
6694 170806 : return have_data;
6695 : }
6696 :
6697 : /*
6698 : * Get one endpoint datum (min or max depending on indexscandir) from the
6699 : * specified index. Return true if successful, false if not.
6700 : * On success, endpoint value is stored to *endpointDatum (and copied into
6701 : * outercontext).
6702 : *
6703 : * scankeys is a 1-element scankey array set up to reject nulls.
6704 : * typLen/typByVal describe the datatype of the index's first column.
6705 : * tableslot is a slot suitable to hold table tuples, in case we need
6706 : * to probe the heap.
6707 : * (We could compute these values locally, but that would mean computing them
6708 : * twice when get_actual_variable_range needs both the min and the max.)
6709 : *
6710 : * Failure occurs either when the index is empty, or we decide that it's
6711 : * taking too long to find a suitable tuple.
6712 : */
6713 : static bool
6714 125304 : get_actual_variable_endpoint(Relation heapRel,
6715 : Relation indexRel,
6716 : ScanDirection indexscandir,
6717 : ScanKey scankeys,
6718 : int16 typLen,
6719 : bool typByVal,
6720 : TupleTableSlot *tableslot,
6721 : MemoryContext outercontext,
6722 : Datum *endpointDatum)
6723 : {
6724 125304 : bool have_data = false;
6725 : SnapshotData SnapshotNonVacuumable;
6726 : IndexScanDesc index_scan;
6727 125304 : Buffer vmbuffer = InvalidBuffer;
6728 125304 : BlockNumber last_heap_block = InvalidBlockNumber;
6729 125304 : int n_visited_heap_pages = 0;
6730 : ItemPointer tid;
6731 : Datum values[INDEX_MAX_KEYS];
6732 : bool isnull[INDEX_MAX_KEYS];
6733 : MemoryContext oldcontext;
6734 :
6735 : /*
6736 : * We use the index-only-scan machinery for this. With mostly-static
6737 : * tables that's a win because it avoids a heap visit. It's also a win
6738 : * for dynamic data, but the reason is less obvious; read on for details.
6739 : *
6740 : * In principle, we should scan the index with our current active
6741 : * snapshot, which is the best approximation we've got to what the query
6742 : * will see when executed. But that won't be exact if a new snap is taken
6743 : * before running the query, and it can be very expensive if a lot of
6744 : * recently-dead or uncommitted rows exist at the beginning or end of the
6745 : * index (because we'll laboriously fetch each one and reject it).
6746 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
6747 : * and uncommitted rows as well as normal visible rows. On the other
6748 : * hand, it will reject known-dead rows, and thus not give a bogus answer
6749 : * when the extreme value has been deleted (unless the deletion was quite
6750 : * recent); that case motivates not using SnapshotAny here.
6751 : *
6752 : * A crucial point here is that SnapshotNonVacuumable, with
6753 : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
6754 : * condition that the indexscan will use to decide that index entries are
6755 : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
6756 : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
6757 : * have to continue scanning past it, we know that the indexscan will mark
6758 : * that index entry killed. That means that the next
6759 : * get_actual_variable_endpoint() call will not have to re-consider that
6760 : * index entry. In this way we avoid repetitive work when this function
6761 : * is used a lot during planning.
6762 : *
6763 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
6764 : * recently-created index, some index entries may point at "broken" HOT
6765 : * chains in which not all the tuple versions contain data matching the
6766 : * index entry. The live tuple version(s) certainly do match the index,
6767 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
6768 : * don't match. Hence, if we took data from the selected heap tuple, we
6769 : * might get a bogus answer that's not close to the index extremal value,
6770 : * or could even be NULL. We avoid this hazard because we take the data
6771 : * from the index entry not the heap.
6772 : *
6773 : * Despite all this care, there are situations where we might find many
6774 : * non-visible tuples near the end of the index. We don't want to expend
6775 : * a huge amount of time here, so we give up once we've read too many heap
6776 : * pages. When we fail for that reason, the caller will end up using
6777 : * whatever extremal value is recorded in pg_statistic.
6778 : */
6779 125304 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
6780 : GlobalVisTestFor(heapRel));
6781 :
6782 125304 : index_scan = index_beginscan(heapRel, indexRel,
6783 : &SnapshotNonVacuumable, NULL,
6784 : 1, 0);
6785 : /* Set it up for index-only scan */
6786 125304 : index_scan->xs_want_itup = true;
6787 125304 : index_rescan(index_scan, scankeys, 1, NULL, 0);
6788 :
6789 : /* Fetch first/next tuple in specified direction */
6790 145196 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
6791 : {
6792 145196 : BlockNumber block = ItemPointerGetBlockNumber(tid);
6793 :
6794 145196 : if (!VM_ALL_VISIBLE(heapRel,
6795 : block,
6796 : &vmbuffer))
6797 : {
6798 : /* Rats, we have to visit the heap to check visibility */
6799 100964 : if (!index_fetch_heap(index_scan, tableslot))
6800 : {
6801 : /*
6802 : * No visible tuple for this index entry, so we need to
6803 : * advance to the next entry. Before doing so, count heap
6804 : * page fetches and give up if we've done too many.
6805 : *
6806 : * We don't charge a page fetch if this is the same heap page
6807 : * as the previous tuple. This is on the conservative side,
6808 : * since other recently-accessed pages are probably still in
6809 : * buffers too; but it's good enough for this heuristic.
6810 : */
6811 : #define VISITED_PAGES_LIMIT 100
6812 :
6813 19892 : if (block != last_heap_block)
6814 : {
6815 2566 : last_heap_block = block;
6816 2566 : n_visited_heap_pages++;
6817 2566 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
6818 0 : break;
6819 : }
6820 :
6821 19892 : continue; /* no visible tuple, try next index entry */
6822 : }
6823 :
6824 : /* We don't actually need the heap tuple for anything */
6825 81072 : ExecClearTuple(tableslot);
6826 :
6827 : /*
6828 : * We don't care whether there's more than one visible tuple in
6829 : * the HOT chain; if any are visible, that's good enough.
6830 : */
6831 : }
6832 :
6833 : /*
6834 : * We expect that the index will return data in IndexTuple not
6835 : * HeapTuple format.
6836 : */
6837 125304 : if (!index_scan->xs_itup)
6838 0 : elog(ERROR, "no data returned for index-only scan");
6839 :
6840 : /*
6841 : * We do not yet support recheck here.
6842 : */
6843 125304 : if (index_scan->xs_recheck)
6844 0 : break;
6845 :
6846 : /* OK to deconstruct the index tuple */
6847 125304 : index_deform_tuple(index_scan->xs_itup,
6848 : index_scan->xs_itupdesc,
6849 : values, isnull);
6850 :
6851 : /* Shouldn't have got a null, but be careful */
6852 125304 : if (isnull[0])
6853 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
6854 : RelationGetRelationName(indexRel));
6855 :
6856 : /* Copy the index column value out to caller's context */
6857 125304 : oldcontext = MemoryContextSwitchTo(outercontext);
6858 125304 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
6859 125304 : MemoryContextSwitchTo(oldcontext);
6860 125304 : have_data = true;
6861 125304 : break;
6862 : }
6863 :
6864 125304 : if (vmbuffer != InvalidBuffer)
6865 112286 : ReleaseBuffer(vmbuffer);
6866 125304 : index_endscan(index_scan);
6867 :
6868 125304 : return have_data;
6869 : }
6870 :
6871 : /*
6872 : * find_join_input_rel
6873 : * Look up the input relation for a join.
6874 : *
6875 : * We assume that the input relation's RelOptInfo must have been constructed
6876 : * already.
6877 : */
6878 : static RelOptInfo *
6879 10702 : find_join_input_rel(PlannerInfo *root, Relids relids)
6880 : {
6881 10702 : RelOptInfo *rel = NULL;
6882 :
6883 10702 : if (!bms_is_empty(relids))
6884 : {
6885 : int relid;
6886 :
6887 10702 : if (bms_get_singleton_member(relids, &relid))
6888 10386 : rel = find_base_rel(root, relid);
6889 : else
6890 316 : rel = find_join_rel(root, relids);
6891 : }
6892 :
6893 10702 : if (rel == NULL)
6894 0 : elog(ERROR, "could not find RelOptInfo for given relids");
6895 :
6896 10702 : return rel;
6897 : }
6898 :
6899 :
6900 : /*-------------------------------------------------------------------------
6901 : *
6902 : * Index cost estimation functions
6903 : *
6904 : *-------------------------------------------------------------------------
6905 : */
6906 :
6907 : /*
6908 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
6909 : */
6910 : List *
6911 801248 : get_quals_from_indexclauses(List *indexclauses)
6912 : {
6913 801248 : List *result = NIL;
6914 : ListCell *lc;
6915 :
6916 1411956 : foreach(lc, indexclauses)
6917 : {
6918 610708 : IndexClause *iclause = lfirst_node(IndexClause, lc);
6919 : ListCell *lc2;
6920 :
6921 1224328 : foreach(lc2, iclause->indexquals)
6922 : {
6923 613620 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6924 :
6925 613620 : result = lappend(result, rinfo);
6926 : }
6927 : }
6928 801248 : return result;
6929 : }
6930 :
6931 : /*
6932 : * Compute the total evaluation cost of the comparison operands in a list
6933 : * of index qual expressions. Since we know these will be evaluated just
6934 : * once per scan, there's no need to distinguish startup from per-row cost.
6935 : *
6936 : * This can be used either on the result of get_quals_from_indexclauses(),
6937 : * or directly on an indexorderbys list. In both cases, we expect that the
6938 : * index key expression is on the left side of binary clauses.
6939 : */
6940 : Cost
6941 1589490 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
6942 : {
6943 1589490 : Cost qual_arg_cost = 0;
6944 : ListCell *lc;
6945 :
6946 2203572 : foreach(lc, indexquals)
6947 : {
6948 614082 : Expr *clause = (Expr *) lfirst(lc);
6949 : Node *other_operand;
6950 : QualCost index_qual_cost;
6951 :
6952 : /*
6953 : * Index quals will have RestrictInfos, indexorderbys won't. Look
6954 : * through RestrictInfo if present.
6955 : */
6956 614082 : if (IsA(clause, RestrictInfo))
6957 613608 : clause = ((RestrictInfo *) clause)->clause;
6958 :
6959 614082 : if (IsA(clause, OpExpr))
6960 : {
6961 599556 : OpExpr *op = (OpExpr *) clause;
6962 :
6963 599556 : other_operand = (Node *) lsecond(op->args);
6964 : }
6965 14526 : else if (IsA(clause, RowCompareExpr))
6966 : {
6967 396 : RowCompareExpr *rc = (RowCompareExpr *) clause;
6968 :
6969 396 : other_operand = (Node *) rc->rargs;
6970 : }
6971 14130 : else if (IsA(clause, ScalarArrayOpExpr))
6972 : {
6973 11204 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6974 :
6975 11204 : other_operand = (Node *) lsecond(saop->args);
6976 : }
6977 2926 : else if (IsA(clause, NullTest))
6978 : {
6979 2926 : other_operand = NULL;
6980 : }
6981 : else
6982 : {
6983 0 : elog(ERROR, "unsupported indexqual type: %d",
6984 : (int) nodeTag(clause));
6985 : other_operand = NULL; /* keep compiler quiet */
6986 : }
6987 :
6988 614082 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
6989 614082 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
6990 : }
6991 1589490 : return qual_arg_cost;
6992 : }
6993 :
6994 : void
6995 788254 : genericcostestimate(PlannerInfo *root,
6996 : IndexPath *path,
6997 : double loop_count,
6998 : GenericCosts *costs)
6999 : {
7000 788254 : IndexOptInfo *index = path->indexinfo;
7001 788254 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7002 788254 : List *indexOrderBys = path->indexorderbys;
7003 : Cost indexStartupCost;
7004 : Cost indexTotalCost;
7005 : Selectivity indexSelectivity;
7006 : double indexCorrelation;
7007 : double numIndexPages;
7008 : double numIndexTuples;
7009 : double spc_random_page_cost;
7010 : double num_sa_scans;
7011 : double num_outer_scans;
7012 : double num_scans;
7013 : double qual_op_cost;
7014 : double qual_arg_cost;
7015 : List *selectivityQuals;
7016 : ListCell *l;
7017 :
7018 : /*
7019 : * If the index is partial, AND the index predicate with the explicitly
7020 : * given indexquals to produce a more accurate idea of the index
7021 : * selectivity.
7022 : */
7023 788254 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7024 :
7025 : /*
7026 : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
7027 : * just assume that the number of index descents is the number of distinct
7028 : * combinations of array elements from all of the scan's SAOP clauses.
7029 : */
7030 788254 : num_sa_scans = costs->num_sa_scans;
7031 788254 : if (num_sa_scans < 1)
7032 : {
7033 7892 : num_sa_scans = 1;
7034 16568 : foreach(l, indexQuals)
7035 : {
7036 8676 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
7037 :
7038 8676 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
7039 : {
7040 26 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
7041 26 : double alength = estimate_array_length(root, lsecond(saop->args));
7042 :
7043 26 : if (alength > 1)
7044 26 : num_sa_scans *= alength;
7045 : }
7046 : }
7047 : }
7048 :
7049 : /* Estimate the fraction of main-table tuples that will be visited */
7050 788254 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7051 788254 : index->rel->relid,
7052 : JOIN_INNER,
7053 : NULL);
7054 :
7055 : /*
7056 : * If caller didn't give us an estimate, estimate the number of index
7057 : * tuples that will be visited. We do it in this rather peculiar-looking
7058 : * way in order to get the right answer for partial indexes.
7059 : */
7060 788254 : numIndexTuples = costs->numIndexTuples;
7061 788254 : if (numIndexTuples <= 0.0)
7062 : {
7063 91852 : numIndexTuples = indexSelectivity * index->rel->tuples;
7064 :
7065 : /*
7066 : * The above calculation counts all the tuples visited across all
7067 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
7068 : * average per-indexscan number, so adjust. This is a handy place to
7069 : * round to integer, too. (If caller supplied tuple estimate, it's
7070 : * responsible for handling these considerations.)
7071 : */
7072 91852 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7073 : }
7074 :
7075 : /*
7076 : * We can bound the number of tuples by the index size in any case. Also,
7077 : * always estimate at least one tuple is touched, even when
7078 : * indexSelectivity estimate is tiny.
7079 : */
7080 788254 : if (numIndexTuples > index->tuples)
7081 6446 : numIndexTuples = index->tuples;
7082 788254 : if (numIndexTuples < 1.0)
7083 92456 : numIndexTuples = 1.0;
7084 :
7085 : /*
7086 : * Estimate the number of index pages that will be retrieved.
7087 : *
7088 : * We use the simplistic method of taking a pro-rata fraction of the total
7089 : * number of index pages. In effect, this counts only leaf pages and not
7090 : * any overhead such as index metapage or upper tree levels.
7091 : *
7092 : * In practice access to upper index levels is often nearly free because
7093 : * those tend to stay in cache under load; moreover, the cost involved is
7094 : * highly dependent on index type. We therefore ignore such costs here
7095 : * and leave it to the caller to add a suitable charge if needed.
7096 : */
7097 788254 : if (index->pages > 1 && index->tuples > 1)
7098 724646 : numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
7099 : else
7100 63608 : numIndexPages = 1.0;
7101 :
7102 : /* fetch estimated page cost for tablespace containing index */
7103 788254 : get_tablespace_page_costs(index->reltablespace,
7104 : &spc_random_page_cost,
7105 : NULL);
7106 :
7107 : /*
7108 : * Now compute the disk access costs.
7109 : *
7110 : * The above calculations are all per-index-scan. However, if we are in a
7111 : * nestloop inner scan, we can expect the scan to be repeated (with
7112 : * different search keys) for each row of the outer relation. Likewise,
7113 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
7114 : * the potential for cache effects to reduce the number of disk page
7115 : * fetches needed. We want to estimate the average per-scan I/O cost in
7116 : * the presence of caching.
7117 : *
7118 : * We use the Mackert-Lohman formula (see costsize.c for details) to
7119 : * estimate the total number of page fetches that occur. While this
7120 : * wasn't what it was designed for, it seems a reasonable model anyway.
7121 : * Note that we are counting pages not tuples anymore, so we take N = T =
7122 : * index size, as if there were one "tuple" per page.
7123 : */
7124 788254 : num_outer_scans = loop_count;
7125 788254 : num_scans = num_sa_scans * num_outer_scans;
7126 :
7127 788254 : if (num_scans > 1)
7128 : {
7129 : double pages_fetched;
7130 :
7131 : /* total page fetches ignoring cache effects */
7132 93492 : pages_fetched = numIndexPages * num_scans;
7133 :
7134 : /* use Mackert and Lohman formula to adjust for cache effects */
7135 93492 : pages_fetched = index_pages_fetched(pages_fetched,
7136 : index->pages,
7137 93492 : (double) index->pages,
7138 : root);
7139 :
7140 : /*
7141 : * Now compute the total disk access cost, and then report a pro-rated
7142 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
7143 : * since that's internal to the indexscan.)
7144 : */
7145 93492 : indexTotalCost = (pages_fetched * spc_random_page_cost)
7146 : / num_outer_scans;
7147 : }
7148 : else
7149 : {
7150 : /*
7151 : * For a single index scan, we just charge spc_random_page_cost per
7152 : * page touched.
7153 : */
7154 694762 : indexTotalCost = numIndexPages * spc_random_page_cost;
7155 : }
7156 :
7157 : /*
7158 : * CPU cost: any complex expressions in the indexquals will need to be
7159 : * evaluated once at the start of the scan to reduce them to runtime keys
7160 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
7161 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
7162 : * indexqual operator. Because we have numIndexTuples as a per-scan
7163 : * number, we have to multiply by num_sa_scans to get the correct result
7164 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
7165 : * ORDER BY expressions.
7166 : *
7167 : * Note: this neglects the possible costs of rechecking lossy operators.
7168 : * Detecting that that might be needed seems more expensive than it's
7169 : * worth, though, considering all the other inaccuracies here ...
7170 : */
7171 788254 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
7172 788254 : index_other_operands_eval_cost(root, indexOrderBys);
7173 788254 : qual_op_cost = cpu_operator_cost *
7174 788254 : (list_length(indexQuals) + list_length(indexOrderBys));
7175 :
7176 788254 : indexStartupCost = qual_arg_cost;
7177 788254 : indexTotalCost += qual_arg_cost;
7178 788254 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
7179 :
7180 : /*
7181 : * Generic assumption about index correlation: there isn't any.
7182 : */
7183 788254 : indexCorrelation = 0.0;
7184 :
7185 : /*
7186 : * Return everything to caller.
7187 : */
7188 788254 : costs->indexStartupCost = indexStartupCost;
7189 788254 : costs->indexTotalCost = indexTotalCost;
7190 788254 : costs->indexSelectivity = indexSelectivity;
7191 788254 : costs->indexCorrelation = indexCorrelation;
7192 788254 : costs->numIndexPages = numIndexPages;
7193 788254 : costs->numIndexTuples = numIndexTuples;
7194 788254 : costs->spc_random_page_cost = spc_random_page_cost;
7195 788254 : costs->num_sa_scans = num_sa_scans;
7196 788254 : }
7197 :
7198 : /*
7199 : * If the index is partial, add its predicate to the given qual list.
7200 : *
7201 : * ANDing the index predicate with the explicitly given indexquals produces
7202 : * a more accurate idea of the index's selectivity. However, we need to be
7203 : * careful not to insert redundant clauses, because clauselist_selectivity()
7204 : * is easily fooled into computing a too-low selectivity estimate. Our
7205 : * approach is to add only the predicate clause(s) that cannot be proven to
7206 : * be implied by the given indexquals. This successfully handles cases such
7207 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
7208 : * There are many other cases where we won't detect redundancy, leading to a
7209 : * too-low selectivity estimate, which will bias the system in favor of using
7210 : * partial indexes where possible. That is not necessarily bad though.
7211 : *
7212 : * Note that indexQuals contains RestrictInfo nodes while the indpred
7213 : * does not, so the output list will be mixed. This is OK for both
7214 : * predicate_implied_by() and clauselist_selectivity(), but might be
7215 : * problematic if the result were passed to other things.
7216 : */
7217 : List *
7218 1328744 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
7219 : {
7220 1328744 : List *predExtraQuals = NIL;
7221 : ListCell *lc;
7222 :
7223 1328744 : if (index->indpred == NIL)
7224 1326718 : return indexQuals;
7225 :
7226 4064 : foreach(lc, index->indpred)
7227 : {
7228 2038 : Node *predQual = (Node *) lfirst(lc);
7229 2038 : List *oneQual = list_make1(predQual);
7230 :
7231 2038 : if (!predicate_implied_by(oneQual, indexQuals, false))
7232 1836 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7233 : }
7234 2026 : return list_concat(predExtraQuals, indexQuals);
7235 : }
7236 :
7237 : /*
7238 : * Estimate correlation of btree index's first column.
7239 : *
7240 : * If we can get an estimate of the first column's ordering correlation C
7241 : * from pg_statistic, estimate the index correlation as C for a single-column
7242 : * index, or C * 0.75 for multiple columns. The idea here is that multiple
7243 : * columns dilute the importance of the first column's ordering, but don't
7244 : * negate it entirely.
7245 : *
7246 : * We already filled in the stats tuple for *vardata when called.
7247 : */
7248 : static double
7249 582006 : btcost_correlation(IndexOptInfo *index, VariableStatData *vardata)
7250 : {
7251 : Oid sortop;
7252 : AttStatsSlot sslot;
7253 582006 : double indexCorrelation = 0;
7254 :
7255 : Assert(HeapTupleIsValid(vardata->statsTuple));
7256 :
7257 582006 : sortop = get_opfamily_member(index->opfamily[0],
7258 582006 : index->opcintype[0],
7259 582006 : index->opcintype[0],
7260 : BTLessStrategyNumber);
7261 1164012 : if (OidIsValid(sortop) &&
7262 582006 : get_attstatsslot(&sslot, vardata->statsTuple,
7263 : STATISTIC_KIND_CORRELATION, sortop,
7264 : ATTSTATSSLOT_NUMBERS))
7265 : {
7266 : double varCorrelation;
7267 :
7268 : Assert(sslot.nnumbers == 1);
7269 573980 : varCorrelation = sslot.numbers[0];
7270 :
7271 573980 : if (index->reverse_sort[0])
7272 0 : varCorrelation = -varCorrelation;
7273 :
7274 573980 : if (index->nkeycolumns > 1)
7275 199926 : indexCorrelation = varCorrelation * 0.75;
7276 : else
7277 374054 : indexCorrelation = varCorrelation;
7278 :
7279 573980 : free_attstatsslot(&sslot);
7280 : }
7281 :
7282 582006 : return indexCorrelation;
7283 : }
7284 :
7285 : void
7286 780362 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7287 : Cost *indexStartupCost, Cost *indexTotalCost,
7288 : Selectivity *indexSelectivity, double *indexCorrelation,
7289 : double *indexPages)
7290 : {
7291 780362 : IndexOptInfo *index = path->indexinfo;
7292 780362 : GenericCosts costs = {0};
7293 780362 : VariableStatData vardata = {0};
7294 : double numIndexTuples;
7295 : Cost descentCost;
7296 : List *indexBoundQuals;
7297 : List *indexSkipQuals;
7298 : int indexcol;
7299 : bool eqQualHere;
7300 : bool found_row_compare;
7301 : bool found_array;
7302 : bool found_is_null_op;
7303 780362 : bool have_correlation = false;
7304 : double num_sa_scans;
7305 780362 : double correlation = 0.0;
7306 : ListCell *lc;
7307 :
7308 : /*
7309 : * For a btree scan, only leading '=' quals plus inequality quals for the
7310 : * immediately next attribute contribute to index selectivity (these are
7311 : * the "boundary quals" that determine the starting and stopping points of
7312 : * the index scan). Additional quals can suppress visits to the heap, so
7313 : * it's OK to count them in indexSelectivity, but they should not count
7314 : * for estimating numIndexTuples. So we must examine the given indexquals
7315 : * to find out which ones count as boundary quals. We rely on the
7316 : * knowledge that they are given in index column order. Note that nbtree
7317 : * preprocessing can add skip arrays that act as leading '=' quals in the
7318 : * absence of ordinary input '=' quals, so in practice _most_ input quals
7319 : * are able to act as index bound quals (which we take into account here).
7320 : *
7321 : * For a RowCompareExpr, we consider only the first column, just as
7322 : * rowcomparesel() does.
7323 : *
7324 : * If there's a SAOP or skip array in the quals, we'll actually perform up
7325 : * to N index descents (not just one), but the underlying array key's
7326 : * operator can be considered to act the same as it normally does.
7327 : */
7328 780362 : indexBoundQuals = NIL;
7329 780362 : indexSkipQuals = NIL;
7330 780362 : indexcol = 0;
7331 780362 : eqQualHere = false;
7332 780362 : found_row_compare = false;
7333 780362 : found_array = false;
7334 780362 : found_is_null_op = false;
7335 780362 : num_sa_scans = 1;
7336 1329382 : foreach(lc, path->indexclauses)
7337 : {
7338 585786 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7339 : ListCell *lc2;
7340 :
7341 585786 : if (indexcol < iclause->indexcol)
7342 : {
7343 115298 : double num_sa_scans_prev_cols = num_sa_scans;
7344 :
7345 : /*
7346 : * Beginning of a new column's quals.
7347 : *
7348 : * Skip scans use skip arrays, which are ScalarArrayOp style
7349 : * arrays that generate their elements procedurally and on demand.
7350 : * Given a multi-column index on "(a, b)", and an SQL WHERE clause
7351 : * "WHERE b = 42", a skip scan will effectively use an indexqual
7352 : * "WHERE a = ANY('{every col a value}') AND b = 42". (Obviously,
7353 : * the array on "a" must also return "IS NULL" matches, since our
7354 : * WHERE clause used no strict operator on "a").
7355 : *
7356 : * Here we consider how nbtree will backfill skip arrays for any
7357 : * index columns that lacked an '=' qual. This maintains our
7358 : * num_sa_scans estimate, and determines if this new column (the
7359 : * "iclause->indexcol" column, not the prior "indexcol" column)
7360 : * can have its RestrictInfos/quals added to indexBoundQuals.
7361 : *
7362 : * We'll need to handle columns that have inequality quals, where
7363 : * the skip array generates values from a range constrained by the
7364 : * quals (not every possible value). We've been maintaining
7365 : * indexSkipQuals to help with this; it will now contain all of
7366 : * the prior column's quals (that is, indexcol's quals) when they
7367 : * might be used for this.
7368 : */
7369 115298 : if (found_row_compare)
7370 : {
7371 : /*
7372 : * Skip arrays can't be added after a RowCompare input qual
7373 : * due to limitations in nbtree
7374 : */
7375 24 : break;
7376 : }
7377 115274 : if (eqQualHere)
7378 : {
7379 : /*
7380 : * Don't need to add a skip array for an indexcol that already
7381 : * has an '=' qual/equality constraint
7382 : */
7383 79218 : indexcol++;
7384 79218 : indexSkipQuals = NIL;
7385 : }
7386 115274 : eqQualHere = false;
7387 :
7388 118228 : while (indexcol < iclause->indexcol)
7389 : {
7390 : double ndistinct;
7391 39696 : bool isdefault = true;
7392 :
7393 39696 : found_array = true;
7394 :
7395 : /*
7396 : * A skipped attribute's ndistinct forms the basis of our
7397 : * estimate of the total number of "array elements" used by
7398 : * its skip array at runtime. Look that up first.
7399 : */
7400 39696 : examine_indexcol_variable(root, index, indexcol, &vardata);
7401 39696 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
7402 :
7403 39696 : if (indexcol == 0)
7404 : {
7405 : /*
7406 : * Get an estimate of the leading column's correlation in
7407 : * passing (avoids rereading variable stats below)
7408 : */
7409 36044 : if (HeapTupleIsValid(vardata.statsTuple))
7410 23292 : correlation = btcost_correlation(index, &vardata);
7411 36044 : have_correlation = true;
7412 : }
7413 :
7414 39696 : ReleaseVariableStats(vardata);
7415 :
7416 : /*
7417 : * If ndistinct is a default estimate, conservatively assume
7418 : * that no skipping will happen at runtime
7419 : */
7420 39696 : if (isdefault)
7421 : {
7422 11382 : num_sa_scans = num_sa_scans_prev_cols;
7423 36742 : break; /* done building indexBoundQuals */
7424 : }
7425 :
7426 : /*
7427 : * Apply indexcol's indexSkipQuals selectivity to ndistinct
7428 : */
7429 28314 : if (indexSkipQuals != NIL)
7430 : {
7431 : List *partialSkipQuals;
7432 : Selectivity ndistinctfrac;
7433 :
7434 : /*
7435 : * If the index is partial, AND the index predicate with
7436 : * the index-bound quals to produce a more accurate idea
7437 : * of the number of distinct values for prior indexcol
7438 : */
7439 664 : partialSkipQuals = add_predicate_to_index_quals(index,
7440 : indexSkipQuals);
7441 :
7442 664 : ndistinctfrac = clauselist_selectivity(root, partialSkipQuals,
7443 664 : index->rel->relid,
7444 : JOIN_INNER,
7445 : NULL);
7446 :
7447 : /*
7448 : * If ndistinctfrac is selective (on its own), the scan is
7449 : * unlikely to benefit from repositioning itself using
7450 : * later quals. Do not allow iclause->indexcol's quals to
7451 : * be added to indexBoundQuals (it would increase descent
7452 : * costs, without lowering numIndexTuples costs by much).
7453 : */
7454 664 : if (ndistinctfrac < DEFAULT_RANGE_INEQ_SEL)
7455 : {
7456 374 : num_sa_scans = num_sa_scans_prev_cols;
7457 374 : break; /* done building indexBoundQuals */
7458 : }
7459 :
7460 : /* Adjust ndistinct downward */
7461 290 : ndistinct = rint(ndistinct * ndistinctfrac);
7462 290 : ndistinct = Max(ndistinct, 1);
7463 : }
7464 :
7465 : /*
7466 : * When there's no inequality quals, account for the need to
7467 : * find an initial value by counting -inf/+inf as a value.
7468 : *
7469 : * We don't charge anything extra for possible next/prior key
7470 : * index probes, which are sometimes used to find the next
7471 : * valid skip array element (ahead of using the located
7472 : * element value to relocate the scan to the next position
7473 : * that might contain matching tuples). It seems hard to do
7474 : * better here. Use of the skip support infrastructure often
7475 : * avoids most next/prior key probes. But even when it can't,
7476 : * there's a decent chance that most individual next/prior key
7477 : * probes will locate a leaf page whose key space overlaps all
7478 : * of the scan's keys (even the lower-order keys) -- which
7479 : * also avoids the need for a separate, extra index descent.
7480 : * Note also that these probes are much cheaper than non-probe
7481 : * primitive index scans: they're reliably very selective.
7482 : */
7483 27940 : if (indexSkipQuals == NIL)
7484 27650 : ndistinct += 1;
7485 :
7486 : /*
7487 : * Update num_sa_scans estimate by multiplying by ndistinct.
7488 : *
7489 : * We make the pessimistic assumption that there is no
7490 : * naturally occurring cross-column correlation. This is
7491 : * often wrong, but it seems best to err on the side of not
7492 : * expecting skipping to be helpful...
7493 : */
7494 27940 : num_sa_scans *= ndistinct;
7495 :
7496 : /*
7497 : * ...but back out of adding this latest group of 1 or more
7498 : * skip arrays when num_sa_scans exceeds the total number of
7499 : * index pages (revert to num_sa_scans from before indexcol).
7500 : * This causes a sharp discontinuity in cost (as a function of
7501 : * the indexcol's ndistinct), but that is representative of
7502 : * actual runtime costs.
7503 : *
7504 : * Note that skipping is helpful when each primitive index
7505 : * scan only manages to skip over 1 or 2 irrelevant leaf pages
7506 : * on average. Skip arrays bring savings in CPU costs due to
7507 : * the scan not needing to evaluate indexquals against every
7508 : * tuple, which can greatly exceed any savings in I/O costs.
7509 : * This test is a test of whether num_sa_scans implies that
7510 : * we're past the point where the ability to skip ceases to
7511 : * lower the scan's costs (even qual evaluation CPU costs).
7512 : */
7513 27940 : if (index->pages < num_sa_scans)
7514 : {
7515 24986 : num_sa_scans = num_sa_scans_prev_cols;
7516 24986 : break; /* done building indexBoundQuals */
7517 : }
7518 :
7519 2954 : indexcol++;
7520 2954 : indexSkipQuals = NIL;
7521 : }
7522 :
7523 : /*
7524 : * Finished considering the need to add skip arrays to bridge an
7525 : * initial eqQualHere gap between the old and new index columns
7526 : * (or there was no initial eqQualHere gap in the first place).
7527 : *
7528 : * If an initial gap could not be bridged, then new column's quals
7529 : * (i.e. iclause->indexcol's quals) won't go into indexBoundQuals,
7530 : * and so won't affect our final numIndexTuples estimate.
7531 : */
7532 115274 : if (indexcol != iclause->indexcol)
7533 36742 : break; /* done building indexBoundQuals */
7534 : }
7535 :
7536 : Assert(indexcol == iclause->indexcol);
7537 :
7538 : /* Examine each indexqual associated with this index clause */
7539 1100780 : foreach(lc2, iclause->indexquals)
7540 : {
7541 551760 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7542 551760 : Expr *clause = rinfo->clause;
7543 551760 : Oid clause_op = InvalidOid;
7544 : int op_strategy;
7545 :
7546 551760 : if (IsA(clause, OpExpr))
7547 : {
7548 538304 : OpExpr *op = (OpExpr *) clause;
7549 :
7550 538304 : clause_op = op->opno;
7551 : }
7552 13456 : else if (IsA(clause, RowCompareExpr))
7553 : {
7554 396 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7555 :
7556 396 : clause_op = linitial_oid(rc->opnos);
7557 396 : found_row_compare = true;
7558 : }
7559 13060 : else if (IsA(clause, ScalarArrayOpExpr))
7560 : {
7561 10776 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7562 10776 : Node *other_operand = (Node *) lsecond(saop->args);
7563 10776 : double alength = estimate_array_length(root, other_operand);
7564 :
7565 10776 : clause_op = saop->opno;
7566 10776 : found_array = true;
7567 : /* estimate SA descents by indexBoundQuals only */
7568 10776 : if (alength > 1)
7569 10576 : num_sa_scans *= alength;
7570 : }
7571 2284 : else if (IsA(clause, NullTest))
7572 : {
7573 2284 : NullTest *nt = (NullTest *) clause;
7574 :
7575 2284 : if (nt->nulltesttype == IS_NULL)
7576 : {
7577 240 : found_is_null_op = true;
7578 : /* IS NULL is like = for selectivity/skip scan purposes */
7579 240 : eqQualHere = true;
7580 : }
7581 : }
7582 : else
7583 0 : elog(ERROR, "unsupported indexqual type: %d",
7584 : (int) nodeTag(clause));
7585 :
7586 : /* check for equality operator */
7587 551760 : if (OidIsValid(clause_op))
7588 : {
7589 549476 : op_strategy = get_op_opfamily_strategy(clause_op,
7590 549476 : index->opfamily[indexcol]);
7591 : Assert(op_strategy != 0); /* not a member of opfamily?? */
7592 549476 : if (op_strategy == BTEqualStrategyNumber)
7593 523132 : eqQualHere = true;
7594 : }
7595 :
7596 551760 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
7597 :
7598 : /*
7599 : * We apply inequality selectivities to estimate index descent
7600 : * costs with scans that use skip arrays. Save this indexcol's
7601 : * RestrictInfos if it looks like they'll be needed for that.
7602 : */
7603 551760 : if (!eqQualHere && !found_row_compare &&
7604 27290 : indexcol < index->nkeycolumns - 1)
7605 5704 : indexSkipQuals = lappend(indexSkipQuals, rinfo);
7606 : }
7607 : }
7608 :
7609 : /*
7610 : * If index is unique and we found an '=' clause for each column, we can
7611 : * just assume numIndexTuples = 1 and skip the expensive
7612 : * clauselist_selectivity calculations. However, an array or NullTest
7613 : * always invalidates that theory (even when eqQualHere has been set).
7614 : */
7615 780362 : if (index->unique &&
7616 635546 : indexcol == index->nkeycolumns - 1 &&
7617 248804 : eqQualHere &&
7618 248804 : !found_array &&
7619 242848 : !found_is_null_op)
7620 242800 : numIndexTuples = 1.0;
7621 : else
7622 : {
7623 : List *selectivityQuals;
7624 : Selectivity btreeSelectivity;
7625 :
7626 : /*
7627 : * If the index is partial, AND the index predicate with the
7628 : * index-bound quals to produce a more accurate idea of the number of
7629 : * rows covered by the bound conditions.
7630 : */
7631 537562 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
7632 :
7633 537562 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
7634 537562 : index->rel->relid,
7635 : JOIN_INNER,
7636 : NULL);
7637 537562 : numIndexTuples = btreeSelectivity * index->rel->tuples;
7638 :
7639 : /*
7640 : * btree automatically combines individual array element primitive
7641 : * index scans whenever the tuples covered by the next set of array
7642 : * keys are close to tuples covered by the current set. That puts a
7643 : * natural ceiling on the worst case number of descents -- there
7644 : * cannot possibly be more than one descent per leaf page scanned.
7645 : *
7646 : * Clamp the number of descents to at most 1/3 the number of index
7647 : * pages. This avoids implausibly high estimates with low selectivity
7648 : * paths, where scans usually require only one or two descents. This
7649 : * is most likely to help when there are several SAOP clauses, where
7650 : * naively accepting the total number of distinct combinations of
7651 : * array elements as the number of descents would frequently lead to
7652 : * wild overestimates.
7653 : *
7654 : * We somewhat arbitrarily don't just make the cutoff the total number
7655 : * of leaf pages (we make it 1/3 the total number of pages instead) to
7656 : * give the btree code credit for its ability to continue on the leaf
7657 : * level with low selectivity scans.
7658 : *
7659 : * Note: num_sa_scans includes both ScalarArrayOp array elements and
7660 : * skip array elements whose qual affects our numIndexTuples estimate.
7661 : */
7662 537562 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
7663 537562 : num_sa_scans = Max(num_sa_scans, 1);
7664 :
7665 : /*
7666 : * As in genericcostestimate(), we have to adjust for any array quals
7667 : * included in indexBoundQuals, and then round to integer.
7668 : *
7669 : * It is tempting to make genericcostestimate behave as if array
7670 : * clauses work in almost the same way as scalar operators during
7671 : * btree scans, making the top-level scan look like a continuous scan
7672 : * (as opposed to num_sa_scans-many primitive index scans). After
7673 : * all, btree scans mostly work like that at runtime. However, such a
7674 : * scheme would badly bias genericcostestimate's simplistic approach
7675 : * to calculating numIndexPages through prorating.
7676 : *
7677 : * Stick with the approach taken by non-native SAOP scans for now.
7678 : * genericcostestimate will use the Mackert-Lohman formula to
7679 : * compensate for repeat page fetches, even though that definitely
7680 : * won't happen during btree scans (not for leaf pages, at least).
7681 : * We're usually very pessimistic about the number of primitive index
7682 : * scans that will be required, but it's not clear how to do better.
7683 : */
7684 537562 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7685 : }
7686 :
7687 : /*
7688 : * Now do generic index cost estimation.
7689 : */
7690 780362 : costs.numIndexTuples = numIndexTuples;
7691 780362 : costs.num_sa_scans = num_sa_scans;
7692 :
7693 780362 : genericcostestimate(root, path, loop_count, &costs);
7694 :
7695 : /*
7696 : * Add a CPU-cost component to represent the costs of initial btree
7697 : * descent. We don't charge any I/O cost for touching upper btree levels,
7698 : * since they tend to stay in cache, but we still have to do about log2(N)
7699 : * comparisons to descend a btree of N leaf tuples. We charge one
7700 : * cpu_operator_cost per comparison.
7701 : *
7702 : * If there are SAOP or skip array keys, charge this once per estimated
7703 : * index descent. The ones after the first one are not startup cost so
7704 : * far as the overall plan goes, so just add them to "total" cost.
7705 : */
7706 780362 : if (index->tuples > 1) /* avoid computing log(0) */
7707 : {
7708 725534 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
7709 725534 : costs.indexStartupCost += descentCost;
7710 725534 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7711 : }
7712 :
7713 : /*
7714 : * Even though we're not charging I/O cost for touching upper btree pages,
7715 : * it's still reasonable to charge some CPU cost per page descended
7716 : * through. Moreover, if we had no such charge at all, bloated indexes
7717 : * would appear to have the same search cost as unbloated ones, at least
7718 : * in cases where only a single leaf page is expected to be visited. This
7719 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
7720 : * touched. The number of such pages is btree tree height plus one (ie,
7721 : * we charge for the leaf page too). As above, charge once per estimated
7722 : * SAOP/skip array descent.
7723 : */
7724 780362 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7725 780362 : costs.indexStartupCost += descentCost;
7726 780362 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7727 :
7728 780362 : if (!have_correlation)
7729 : {
7730 744318 : examine_indexcol_variable(root, index, 0, &vardata);
7731 744318 : if (HeapTupleIsValid(vardata.statsTuple))
7732 558714 : costs.indexCorrelation = btcost_correlation(index, &vardata);
7733 744318 : ReleaseVariableStats(vardata);
7734 : }
7735 : else
7736 : {
7737 : /* btcost_correlation already called earlier on */
7738 36044 : costs.indexCorrelation = correlation;
7739 : }
7740 :
7741 780362 : *indexStartupCost = costs.indexStartupCost;
7742 780362 : *indexTotalCost = costs.indexTotalCost;
7743 780362 : *indexSelectivity = costs.indexSelectivity;
7744 780362 : *indexCorrelation = costs.indexCorrelation;
7745 780362 : *indexPages = costs.numIndexPages;
7746 780362 : }
7747 :
7748 : void
7749 418 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7750 : Cost *indexStartupCost, Cost *indexTotalCost,
7751 : Selectivity *indexSelectivity, double *indexCorrelation,
7752 : double *indexPages)
7753 : {
7754 418 : GenericCosts costs = {0};
7755 :
7756 418 : genericcostestimate(root, path, loop_count, &costs);
7757 :
7758 : /*
7759 : * A hash index has no descent costs as such, since the index AM can go
7760 : * directly to the target bucket after computing the hash value. There
7761 : * are a couple of other hash-specific costs that we could conceivably add
7762 : * here, though:
7763 : *
7764 : * Ideally we'd charge spc_random_page_cost for each page in the target
7765 : * bucket, not just the numIndexPages pages that genericcostestimate
7766 : * thought we'd visit. However in most cases we don't know which bucket
7767 : * that will be. There's no point in considering the average bucket size
7768 : * because the hash AM makes sure that's always one page.
7769 : *
7770 : * Likewise, we could consider charging some CPU for each index tuple in
7771 : * the bucket, if we knew how many there were. But the per-tuple cost is
7772 : * just a hash value comparison, not a general datatype-dependent
7773 : * comparison, so any such charge ought to be quite a bit less than
7774 : * cpu_operator_cost; which makes it probably not worth worrying about.
7775 : *
7776 : * A bigger issue is that chance hash-value collisions will result in
7777 : * wasted probes into the heap. We don't currently attempt to model this
7778 : * cost on the grounds that it's rare, but maybe it's not rare enough.
7779 : * (Any fix for this ought to consider the generic lossy-operator problem,
7780 : * though; it's not entirely hash-specific.)
7781 : */
7782 :
7783 418 : *indexStartupCost = costs.indexStartupCost;
7784 418 : *indexTotalCost = costs.indexTotalCost;
7785 418 : *indexSelectivity = costs.indexSelectivity;
7786 418 : *indexCorrelation = costs.indexCorrelation;
7787 418 : *indexPages = costs.numIndexPages;
7788 418 : }
7789 :
7790 : void
7791 4878 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7792 : Cost *indexStartupCost, Cost *indexTotalCost,
7793 : Selectivity *indexSelectivity, double *indexCorrelation,
7794 : double *indexPages)
7795 : {
7796 4878 : IndexOptInfo *index = path->indexinfo;
7797 4878 : GenericCosts costs = {0};
7798 : Cost descentCost;
7799 :
7800 4878 : genericcostestimate(root, path, loop_count, &costs);
7801 :
7802 : /*
7803 : * We model index descent costs similarly to those for btree, but to do
7804 : * that we first need an idea of the tree height. We somewhat arbitrarily
7805 : * assume that the fanout is 100, meaning the tree height is at most
7806 : * log100(index->pages).
7807 : *
7808 : * Although this computation isn't really expensive enough to require
7809 : * caching, we might as well use index->tree_height to cache it.
7810 : */
7811 4878 : if (index->tree_height < 0) /* unknown? */
7812 : {
7813 4864 : if (index->pages > 1) /* avoid computing log(0) */
7814 2720 : index->tree_height = (int) (log(index->pages) / log(100.0));
7815 : else
7816 2144 : index->tree_height = 0;
7817 : }
7818 :
7819 : /*
7820 : * Add a CPU-cost component to represent the costs of initial descent. We
7821 : * just use log(N) here not log2(N) since the branching factor isn't
7822 : * necessarily two anyway. As for btree, charge once per SA scan.
7823 : */
7824 4878 : if (index->tuples > 1) /* avoid computing log(0) */
7825 : {
7826 4878 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7827 4878 : costs.indexStartupCost += descentCost;
7828 4878 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7829 : }
7830 :
7831 : /*
7832 : * Likewise add a per-page charge, calculated the same as for btrees.
7833 : */
7834 4878 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7835 4878 : costs.indexStartupCost += descentCost;
7836 4878 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7837 :
7838 4878 : *indexStartupCost = costs.indexStartupCost;
7839 4878 : *indexTotalCost = costs.indexTotalCost;
7840 4878 : *indexSelectivity = costs.indexSelectivity;
7841 4878 : *indexCorrelation = costs.indexCorrelation;
7842 4878 : *indexPages = costs.numIndexPages;
7843 4878 : }
7844 :
7845 : void
7846 1784 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7847 : Cost *indexStartupCost, Cost *indexTotalCost,
7848 : Selectivity *indexSelectivity, double *indexCorrelation,
7849 : double *indexPages)
7850 : {
7851 1784 : IndexOptInfo *index = path->indexinfo;
7852 1784 : GenericCosts costs = {0};
7853 : Cost descentCost;
7854 :
7855 1784 : genericcostestimate(root, path, loop_count, &costs);
7856 :
7857 : /*
7858 : * We model index descent costs similarly to those for btree, but to do
7859 : * that we first need an idea of the tree height. We somewhat arbitrarily
7860 : * assume that the fanout is 100, meaning the tree height is at most
7861 : * log100(index->pages).
7862 : *
7863 : * Although this computation isn't really expensive enough to require
7864 : * caching, we might as well use index->tree_height to cache it.
7865 : */
7866 1784 : if (index->tree_height < 0) /* unknown? */
7867 : {
7868 1778 : if (index->pages > 1) /* avoid computing log(0) */
7869 1778 : index->tree_height = (int) (log(index->pages) / log(100.0));
7870 : else
7871 0 : index->tree_height = 0;
7872 : }
7873 :
7874 : /*
7875 : * Add a CPU-cost component to represent the costs of initial descent. We
7876 : * just use log(N) here not log2(N) since the branching factor isn't
7877 : * necessarily two anyway. As for btree, charge once per SA scan.
7878 : */
7879 1784 : if (index->tuples > 1) /* avoid computing log(0) */
7880 : {
7881 1784 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7882 1784 : costs.indexStartupCost += descentCost;
7883 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7884 : }
7885 :
7886 : /*
7887 : * Likewise add a per-page charge, calculated the same as for btrees.
7888 : */
7889 1784 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7890 1784 : costs.indexStartupCost += descentCost;
7891 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7892 :
7893 1784 : *indexStartupCost = costs.indexStartupCost;
7894 1784 : *indexTotalCost = costs.indexTotalCost;
7895 1784 : *indexSelectivity = costs.indexSelectivity;
7896 1784 : *indexCorrelation = costs.indexCorrelation;
7897 1784 : *indexPages = costs.numIndexPages;
7898 1784 : }
7899 :
7900 :
7901 : /*
7902 : * Support routines for gincostestimate
7903 : */
7904 :
7905 : typedef struct
7906 : {
7907 : bool attHasFullScan[INDEX_MAX_KEYS];
7908 : bool attHasNormalScan[INDEX_MAX_KEYS];
7909 : double partialEntries;
7910 : double exactEntries;
7911 : double searchEntries;
7912 : double arrayScans;
7913 : } GinQualCounts;
7914 :
7915 : /*
7916 : * Estimate the number of index terms that need to be searched for while
7917 : * testing the given GIN query, and increment the counts in *counts
7918 : * appropriately. If the query is unsatisfiable, return false.
7919 : */
7920 : static bool
7921 2480 : gincost_pattern(IndexOptInfo *index, int indexcol,
7922 : Oid clause_op, Datum query,
7923 : GinQualCounts *counts)
7924 : {
7925 : FmgrInfo flinfo;
7926 : Oid extractProcOid;
7927 : Oid collation;
7928 : int strategy_op;
7929 : Oid lefttype,
7930 : righttype;
7931 2480 : int32 nentries = 0;
7932 2480 : bool *partial_matches = NULL;
7933 2480 : Pointer *extra_data = NULL;
7934 2480 : bool *nullFlags = NULL;
7935 2480 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
7936 : int32 i;
7937 :
7938 : Assert(indexcol < index->nkeycolumns);
7939 :
7940 : /*
7941 : * Get the operator's strategy number and declared input data types within
7942 : * the index opfamily. (We don't need the latter, but we use
7943 : * get_op_opfamily_properties because it will throw error if it fails to
7944 : * find a matching pg_amop entry.)
7945 : */
7946 2480 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
7947 : &strategy_op, &lefttype, &righttype);
7948 :
7949 : /*
7950 : * GIN always uses the "default" support functions, which are those with
7951 : * lefttype == righttype == the opclass' opcintype (see
7952 : * IndexSupportInitialize in relcache.c).
7953 : */
7954 2480 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
7955 2480 : index->opcintype[indexcol],
7956 2480 : index->opcintype[indexcol],
7957 : GIN_EXTRACTQUERY_PROC);
7958 :
7959 2480 : if (!OidIsValid(extractProcOid))
7960 : {
7961 : /* should not happen; throw same error as index_getprocinfo */
7962 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
7963 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
7964 : get_rel_name(index->indexoid));
7965 : }
7966 :
7967 : /*
7968 : * Choose collation to pass to extractProc (should match initGinState).
7969 : */
7970 2480 : if (OidIsValid(index->indexcollations[indexcol]))
7971 414 : collation = index->indexcollations[indexcol];
7972 : else
7973 2066 : collation = DEFAULT_COLLATION_OID;
7974 :
7975 2480 : fmgr_info(extractProcOid, &flinfo);
7976 :
7977 2480 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
7978 :
7979 2480 : FunctionCall7Coll(&flinfo,
7980 : collation,
7981 : query,
7982 : PointerGetDatum(&nentries),
7983 : UInt16GetDatum(strategy_op),
7984 : PointerGetDatum(&partial_matches),
7985 : PointerGetDatum(&extra_data),
7986 : PointerGetDatum(&nullFlags),
7987 : PointerGetDatum(&searchMode));
7988 :
7989 2480 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
7990 : {
7991 : /* No match is possible */
7992 12 : return false;
7993 : }
7994 :
7995 9676 : for (i = 0; i < nentries; i++)
7996 : {
7997 : /*
7998 : * For partial match we haven't any information to estimate number of
7999 : * matched entries in index, so, we just estimate it as 100
8000 : */
8001 7208 : if (partial_matches && partial_matches[i])
8002 694 : counts->partialEntries += 100;
8003 : else
8004 6514 : counts->exactEntries++;
8005 :
8006 7208 : counts->searchEntries++;
8007 : }
8008 :
8009 2468 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
8010 : {
8011 1984 : counts->attHasNormalScan[indexcol] = true;
8012 : }
8013 484 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
8014 : {
8015 : /* Treat "include empty" like an exact-match item */
8016 44 : counts->attHasNormalScan[indexcol] = true;
8017 44 : counts->exactEntries++;
8018 44 : counts->searchEntries++;
8019 : }
8020 : else
8021 : {
8022 : /* It's GIN_SEARCH_MODE_ALL */
8023 440 : counts->attHasFullScan[indexcol] = true;
8024 : }
8025 :
8026 2468 : return true;
8027 : }
8028 :
8029 : /*
8030 : * Estimate the number of index terms that need to be searched for while
8031 : * testing the given GIN index clause, and increment the counts in *counts
8032 : * appropriately. If the query is unsatisfiable, return false.
8033 : */
8034 : static bool
8035 2468 : gincost_opexpr(PlannerInfo *root,
8036 : IndexOptInfo *index,
8037 : int indexcol,
8038 : OpExpr *clause,
8039 : GinQualCounts *counts)
8040 : {
8041 2468 : Oid clause_op = clause->opno;
8042 2468 : Node *operand = (Node *) lsecond(clause->args);
8043 :
8044 : /* aggressively reduce to a constant, and look through relabeling */
8045 2468 : operand = estimate_expression_value(root, operand);
8046 :
8047 2468 : if (IsA(operand, RelabelType))
8048 0 : operand = (Node *) ((RelabelType *) operand)->arg;
8049 :
8050 : /*
8051 : * It's impossible to call extractQuery method for unknown operand. So
8052 : * unless operand is a Const we can't do much; just assume there will be
8053 : * one ordinary search entry from the operand at runtime.
8054 : */
8055 2468 : if (!IsA(operand, Const))
8056 : {
8057 0 : counts->exactEntries++;
8058 0 : counts->searchEntries++;
8059 0 : return true;
8060 : }
8061 :
8062 : /* If Const is null, there can be no matches */
8063 2468 : if (((Const *) operand)->constisnull)
8064 0 : return false;
8065 :
8066 : /* Otherwise, apply extractQuery and get the actual term counts */
8067 2468 : return gincost_pattern(index, indexcol, clause_op,
8068 : ((Const *) operand)->constvalue,
8069 : counts);
8070 : }
8071 :
8072 : /*
8073 : * Estimate the number of index terms that need to be searched for while
8074 : * testing the given GIN index clause, and increment the counts in *counts
8075 : * appropriately. If the query is unsatisfiable, return false.
8076 : *
8077 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
8078 : * each of which involves one value from the RHS array, plus all the
8079 : * non-array quals (if any). To model this, we average the counts across
8080 : * the RHS elements, and add the averages to the counts in *counts (which
8081 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
8082 : * by N, causing gincostestimate to scale up its estimates accordingly.
8083 : */
8084 : static bool
8085 6 : gincost_scalararrayopexpr(PlannerInfo *root,
8086 : IndexOptInfo *index,
8087 : int indexcol,
8088 : ScalarArrayOpExpr *clause,
8089 : double numIndexEntries,
8090 : GinQualCounts *counts)
8091 : {
8092 6 : Oid clause_op = clause->opno;
8093 6 : Node *rightop = (Node *) lsecond(clause->args);
8094 : ArrayType *arrayval;
8095 : int16 elmlen;
8096 : bool elmbyval;
8097 : char elmalign;
8098 : int numElems;
8099 : Datum *elemValues;
8100 : bool *elemNulls;
8101 : GinQualCounts arraycounts;
8102 6 : int numPossible = 0;
8103 : int i;
8104 :
8105 : Assert(clause->useOr);
8106 :
8107 : /* aggressively reduce to a constant, and look through relabeling */
8108 6 : rightop = estimate_expression_value(root, rightop);
8109 :
8110 6 : if (IsA(rightop, RelabelType))
8111 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
8112 :
8113 : /*
8114 : * It's impossible to call extractQuery method for unknown operand. So
8115 : * unless operand is a Const we can't do much; just assume there will be
8116 : * one ordinary search entry from each array entry at runtime, and fall
8117 : * back on a probably-bad estimate of the number of array entries.
8118 : */
8119 6 : if (!IsA(rightop, Const))
8120 : {
8121 0 : counts->exactEntries++;
8122 0 : counts->searchEntries++;
8123 0 : counts->arrayScans *= estimate_array_length(root, rightop);
8124 0 : return true;
8125 : }
8126 :
8127 : /* If Const is null, there can be no matches */
8128 6 : if (((Const *) rightop)->constisnull)
8129 0 : return false;
8130 :
8131 : /* Otherwise, extract the array elements and iterate over them */
8132 6 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
8133 6 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
8134 : &elmlen, &elmbyval, &elmalign);
8135 6 : deconstruct_array(arrayval,
8136 : ARR_ELEMTYPE(arrayval),
8137 : elmlen, elmbyval, elmalign,
8138 : &elemValues, &elemNulls, &numElems);
8139 :
8140 6 : memset(&arraycounts, 0, sizeof(arraycounts));
8141 :
8142 18 : for (i = 0; i < numElems; i++)
8143 : {
8144 : GinQualCounts elemcounts;
8145 :
8146 : /* NULL can't match anything, so ignore, as the executor will */
8147 12 : if (elemNulls[i])
8148 0 : continue;
8149 :
8150 : /* Otherwise, apply extractQuery and get the actual term counts */
8151 12 : memset(&elemcounts, 0, sizeof(elemcounts));
8152 :
8153 12 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
8154 : &elemcounts))
8155 : {
8156 : /* We ignore array elements that are unsatisfiable patterns */
8157 12 : numPossible++;
8158 :
8159 12 : if (elemcounts.attHasFullScan[indexcol] &&
8160 0 : !elemcounts.attHasNormalScan[indexcol])
8161 : {
8162 : /*
8163 : * Full index scan will be required. We treat this as if
8164 : * every key in the index had been listed in the query; is
8165 : * that reasonable?
8166 : */
8167 0 : elemcounts.partialEntries = 0;
8168 0 : elemcounts.exactEntries = numIndexEntries;
8169 0 : elemcounts.searchEntries = numIndexEntries;
8170 : }
8171 12 : arraycounts.partialEntries += elemcounts.partialEntries;
8172 12 : arraycounts.exactEntries += elemcounts.exactEntries;
8173 12 : arraycounts.searchEntries += elemcounts.searchEntries;
8174 : }
8175 : }
8176 :
8177 6 : if (numPossible == 0)
8178 : {
8179 : /* No satisfiable patterns in the array */
8180 0 : return false;
8181 : }
8182 :
8183 : /*
8184 : * Now add the averages to the global counts. This will give us an
8185 : * estimate of the average number of terms searched for in each indexscan,
8186 : * including contributions from both array and non-array quals.
8187 : */
8188 6 : counts->partialEntries += arraycounts.partialEntries / numPossible;
8189 6 : counts->exactEntries += arraycounts.exactEntries / numPossible;
8190 6 : counts->searchEntries += arraycounts.searchEntries / numPossible;
8191 :
8192 6 : counts->arrayScans *= numPossible;
8193 :
8194 6 : return true;
8195 : }
8196 :
8197 : /*
8198 : * GIN has search behavior completely different from other index types
8199 : */
8200 : void
8201 2264 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8202 : Cost *indexStartupCost, Cost *indexTotalCost,
8203 : Selectivity *indexSelectivity, double *indexCorrelation,
8204 : double *indexPages)
8205 : {
8206 2264 : IndexOptInfo *index = path->indexinfo;
8207 2264 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8208 : List *selectivityQuals;
8209 2264 : double numPages = index->pages,
8210 2264 : numTuples = index->tuples;
8211 : double numEntryPages,
8212 : numDataPages,
8213 : numPendingPages,
8214 : numEntries;
8215 : GinQualCounts counts;
8216 : bool matchPossible;
8217 : bool fullIndexScan;
8218 : double partialScale;
8219 : double entryPagesFetched,
8220 : dataPagesFetched,
8221 : dataPagesFetchedBySel;
8222 : double qual_op_cost,
8223 : qual_arg_cost,
8224 : spc_random_page_cost,
8225 : outer_scans;
8226 : Cost descentCost;
8227 : Relation indexRel;
8228 : GinStatsData ginStats;
8229 : ListCell *lc;
8230 : int i;
8231 :
8232 : /*
8233 : * Obtain statistical information from the meta page, if possible. Else
8234 : * set ginStats to zeroes, and we'll cope below.
8235 : */
8236 2264 : if (!index->hypothetical)
8237 : {
8238 : /* Lock should have already been obtained in plancat.c */
8239 2264 : indexRel = index_open(index->indexoid, NoLock);
8240 2264 : ginGetStats(indexRel, &ginStats);
8241 2264 : index_close(indexRel, NoLock);
8242 : }
8243 : else
8244 : {
8245 0 : memset(&ginStats, 0, sizeof(ginStats));
8246 : }
8247 :
8248 : /*
8249 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
8250 : * trusted, but the other fields are data as of the last VACUUM. We can
8251 : * scale them up to account for growth since then, but that method only
8252 : * goes so far; in the worst case, the stats might be for a completely
8253 : * empty index, and scaling them will produce pretty bogus numbers.
8254 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
8255 : * it's grown more than that, fall back to estimating things only from the
8256 : * assumed-accurate index size. But we'll trust nPendingPages in any case
8257 : * so long as it's not clearly insane, ie, more than the index size.
8258 : */
8259 2264 : if (ginStats.nPendingPages < numPages)
8260 2264 : numPendingPages = ginStats.nPendingPages;
8261 : else
8262 0 : numPendingPages = 0;
8263 :
8264 2264 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
8265 2264 : ginStats.nTotalPages > numPages / 4 &&
8266 2212 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
8267 1948 : {
8268 : /*
8269 : * OK, the stats seem close enough to sane to be trusted. But we
8270 : * still need to scale them by the ratio numPages / nTotalPages to
8271 : * account for growth since the last VACUUM.
8272 : */
8273 1948 : double scale = numPages / ginStats.nTotalPages;
8274 :
8275 1948 : numEntryPages = ceil(ginStats.nEntryPages * scale);
8276 1948 : numDataPages = ceil(ginStats.nDataPages * scale);
8277 1948 : numEntries = ceil(ginStats.nEntries * scale);
8278 : /* ensure we didn't round up too much */
8279 1948 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
8280 1948 : numDataPages = Min(numDataPages,
8281 : numPages - numPendingPages - numEntryPages);
8282 : }
8283 : else
8284 : {
8285 : /*
8286 : * We might get here because it's a hypothetical index, or an index
8287 : * created pre-9.1 and never vacuumed since upgrading (in which case
8288 : * its stats would read as zeroes), or just because it's grown too
8289 : * much since the last VACUUM for us to put our faith in scaling.
8290 : *
8291 : * Invent some plausible internal statistics based on the index page
8292 : * count (and clamp that to at least 10 pages, just in case). We
8293 : * estimate that 90% of the index is entry pages, and the rest is data
8294 : * pages. Estimate 100 entries per entry page; this is rather bogus
8295 : * since it'll depend on the size of the keys, but it's more robust
8296 : * than trying to predict the number of entries per heap tuple.
8297 : */
8298 316 : numPages = Max(numPages, 10);
8299 316 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
8300 316 : numDataPages = numPages - numPendingPages - numEntryPages;
8301 316 : numEntries = floor(numEntryPages * 100);
8302 : }
8303 :
8304 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
8305 2264 : if (numEntries < 1)
8306 0 : numEntries = 1;
8307 :
8308 : /*
8309 : * If the index is partial, AND the index predicate with the index-bound
8310 : * quals to produce a more accurate idea of the number of rows covered by
8311 : * the bound conditions.
8312 : */
8313 2264 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
8314 :
8315 : /* Estimate the fraction of main-table tuples that will be visited */
8316 4528 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
8317 2264 : index->rel->relid,
8318 : JOIN_INNER,
8319 : NULL);
8320 :
8321 : /* fetch estimated page cost for tablespace containing index */
8322 2264 : get_tablespace_page_costs(index->reltablespace,
8323 : &spc_random_page_cost,
8324 : NULL);
8325 :
8326 : /*
8327 : * Generic assumption about index correlation: there isn't any.
8328 : */
8329 2264 : *indexCorrelation = 0.0;
8330 :
8331 : /*
8332 : * Examine quals to estimate number of search entries & partial matches
8333 : */
8334 2264 : memset(&counts, 0, sizeof(counts));
8335 2264 : counts.arrayScans = 1;
8336 2264 : matchPossible = true;
8337 :
8338 4738 : foreach(lc, path->indexclauses)
8339 : {
8340 2474 : IndexClause *iclause = lfirst_node(IndexClause, lc);
8341 : ListCell *lc2;
8342 :
8343 4936 : foreach(lc2, iclause->indexquals)
8344 : {
8345 2474 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8346 2474 : Expr *clause = rinfo->clause;
8347 :
8348 2474 : if (IsA(clause, OpExpr))
8349 : {
8350 2468 : matchPossible = gincost_opexpr(root,
8351 : index,
8352 2468 : iclause->indexcol,
8353 : (OpExpr *) clause,
8354 : &counts);
8355 2468 : if (!matchPossible)
8356 12 : break;
8357 : }
8358 6 : else if (IsA(clause, ScalarArrayOpExpr))
8359 : {
8360 6 : matchPossible = gincost_scalararrayopexpr(root,
8361 : index,
8362 6 : iclause->indexcol,
8363 : (ScalarArrayOpExpr *) clause,
8364 : numEntries,
8365 : &counts);
8366 6 : if (!matchPossible)
8367 0 : break;
8368 : }
8369 : else
8370 : {
8371 : /* shouldn't be anything else for a GIN index */
8372 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
8373 : (int) nodeTag(clause));
8374 : }
8375 : }
8376 : }
8377 :
8378 : /* Fall out if there were any provably-unsatisfiable quals */
8379 2264 : if (!matchPossible)
8380 : {
8381 12 : *indexStartupCost = 0;
8382 12 : *indexTotalCost = 0;
8383 12 : *indexSelectivity = 0;
8384 12 : return;
8385 : }
8386 :
8387 : /*
8388 : * If attribute has a full scan and at the same time doesn't have normal
8389 : * scan, then we'll have to scan all non-null entries of that attribute.
8390 : * Currently, we don't have per-attribute statistics for GIN. Thus, we
8391 : * must assume the whole GIN index has to be scanned in this case.
8392 : */
8393 2252 : fullIndexScan = false;
8394 4394 : for (i = 0; i < index->nkeycolumns; i++)
8395 : {
8396 2480 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
8397 : {
8398 338 : fullIndexScan = true;
8399 338 : break;
8400 : }
8401 : }
8402 :
8403 2252 : if (fullIndexScan || indexQuals == NIL)
8404 : {
8405 : /*
8406 : * Full index scan will be required. We treat this as if every key in
8407 : * the index had been listed in the query; is that reasonable?
8408 : */
8409 338 : counts.partialEntries = 0;
8410 338 : counts.exactEntries = numEntries;
8411 338 : counts.searchEntries = numEntries;
8412 : }
8413 :
8414 : /* Will we have more than one iteration of a nestloop scan? */
8415 2252 : outer_scans = loop_count;
8416 :
8417 : /*
8418 : * Compute cost to begin scan, first of all, pay attention to pending
8419 : * list.
8420 : */
8421 2252 : entryPagesFetched = numPendingPages;
8422 :
8423 : /*
8424 : * Estimate number of entry pages read. We need to do
8425 : * counts.searchEntries searches. Use a power function as it should be,
8426 : * but tuples on leaf pages usually is much greater. Here we include all
8427 : * searches in entry tree, including search of first entry in partial
8428 : * match algorithm
8429 : */
8430 2252 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8431 :
8432 : /*
8433 : * Add an estimate of entry pages read by partial match algorithm. It's a
8434 : * scan over leaf pages in entry tree. We haven't any useful stats here,
8435 : * so estimate it as proportion. Because counts.partialEntries is really
8436 : * pretty bogus (see code above), it's possible that it is more than
8437 : * numEntries; clamp the proportion to ensure sanity.
8438 : */
8439 2252 : partialScale = counts.partialEntries / numEntries;
8440 2252 : partialScale = Min(partialScale, 1.0);
8441 :
8442 2252 : entryPagesFetched += ceil(numEntryPages * partialScale);
8443 :
8444 : /*
8445 : * Partial match algorithm reads all data pages before doing actual scan,
8446 : * so it's a startup cost. Again, we haven't any useful stats here, so
8447 : * estimate it as proportion.
8448 : */
8449 2252 : dataPagesFetched = ceil(numDataPages * partialScale);
8450 :
8451 2252 : *indexStartupCost = 0;
8452 2252 : *indexTotalCost = 0;
8453 :
8454 : /*
8455 : * Add a CPU-cost component to represent the costs of initial entry btree
8456 : * descent. We don't charge any I/O cost for touching upper btree levels,
8457 : * since they tend to stay in cache, but we still have to do about log2(N)
8458 : * comparisons to descend a btree of N leaf tuples. We charge one
8459 : * cpu_operator_cost per comparison.
8460 : *
8461 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
8462 : * ones after the first one are not startup cost so far as the overall
8463 : * plan is concerned, so add them only to "total" cost.
8464 : */
8465 2252 : if (numEntries > 1) /* avoid computing log(0) */
8466 : {
8467 2252 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
8468 2252 : *indexStartupCost += descentCost * counts.searchEntries;
8469 2252 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
8470 : }
8471 :
8472 : /*
8473 : * Add a cpu cost per entry-page fetched. This is not amortized over a
8474 : * loop.
8475 : */
8476 2252 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8477 2252 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8478 :
8479 : /*
8480 : * Add a cpu cost per data-page fetched. This is also not amortized over a
8481 : * loop. Since those are the data pages from the partial match algorithm,
8482 : * charge them as startup cost.
8483 : */
8484 2252 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
8485 :
8486 : /*
8487 : * Since we add the startup cost to the total cost later on, remove the
8488 : * initial arrayscan from the total.
8489 : */
8490 2252 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8491 :
8492 : /*
8493 : * Calculate cache effects if more than one scan due to nestloops or array
8494 : * quals. The result is pro-rated per nestloop scan, but the array qual
8495 : * factor shouldn't be pro-rated (compare genericcostestimate).
8496 : */
8497 2252 : if (outer_scans > 1 || counts.arrayScans > 1)
8498 : {
8499 6 : entryPagesFetched *= outer_scans * counts.arrayScans;
8500 6 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
8501 : (BlockNumber) numEntryPages,
8502 : numEntryPages, root);
8503 6 : entryPagesFetched /= outer_scans;
8504 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8505 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8506 : (BlockNumber) numDataPages,
8507 : numDataPages, root);
8508 6 : dataPagesFetched /= outer_scans;
8509 : }
8510 :
8511 : /*
8512 : * Here we use random page cost because logically-close pages could be far
8513 : * apart on disk.
8514 : */
8515 2252 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
8516 :
8517 : /*
8518 : * Now compute the number of data pages fetched during the scan.
8519 : *
8520 : * We assume every entry to have the same number of items, and that there
8521 : * is no overlap between them. (XXX: tsvector and array opclasses collect
8522 : * statistics on the frequency of individual keys; it would be nice to use
8523 : * those here.)
8524 : */
8525 2252 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
8526 :
8527 : /*
8528 : * If there is a lot of overlap among the entries, in particular if one of
8529 : * the entries is very frequent, the above calculation can grossly
8530 : * under-estimate. As a simple cross-check, calculate a lower bound based
8531 : * on the overall selectivity of the quals. At a minimum, we must read
8532 : * one item pointer for each matching entry.
8533 : *
8534 : * The width of each item pointer varies, based on the level of
8535 : * compression. We don't have statistics on that, but an average of
8536 : * around 3 bytes per item is fairly typical.
8537 : */
8538 2252 : dataPagesFetchedBySel = ceil(*indexSelectivity *
8539 2252 : (numTuples / (BLCKSZ / 3)));
8540 2252 : if (dataPagesFetchedBySel > dataPagesFetched)
8541 1866 : dataPagesFetched = dataPagesFetchedBySel;
8542 :
8543 : /* Add one page cpu-cost to the startup cost */
8544 2252 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
8545 :
8546 : /*
8547 : * Add once again a CPU-cost for those data pages, before amortizing for
8548 : * cache.
8549 : */
8550 2252 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8551 :
8552 : /* Account for cache effects, the same as above */
8553 2252 : if (outer_scans > 1 || counts.arrayScans > 1)
8554 : {
8555 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8556 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8557 : (BlockNumber) numDataPages,
8558 : numDataPages, root);
8559 6 : dataPagesFetched /= outer_scans;
8560 : }
8561 :
8562 : /* And apply random_page_cost as the cost per page */
8563 2252 : *indexTotalCost += *indexStartupCost +
8564 2252 : dataPagesFetched * spc_random_page_cost;
8565 :
8566 : /*
8567 : * Add on index qual eval costs, much as in genericcostestimate. We charge
8568 : * cpu but we can disregard indexorderbys, since GIN doesn't support
8569 : * those.
8570 : */
8571 2252 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8572 2252 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
8573 :
8574 2252 : *indexStartupCost += qual_arg_cost;
8575 2252 : *indexTotalCost += qual_arg_cost;
8576 :
8577 : /*
8578 : * Add a cpu cost per search entry, corresponding to the actual visited
8579 : * entries.
8580 : */
8581 2252 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
8582 : /* Now add a cpu cost per tuple in the posting lists / trees */
8583 2252 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
8584 2252 : *indexPages = dataPagesFetched;
8585 : }
8586 :
8587 : /*
8588 : * BRIN has search behavior completely different from other index types
8589 : */
8590 : void
8591 10730 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8592 : Cost *indexStartupCost, Cost *indexTotalCost,
8593 : Selectivity *indexSelectivity, double *indexCorrelation,
8594 : double *indexPages)
8595 : {
8596 10730 : IndexOptInfo *index = path->indexinfo;
8597 10730 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8598 10730 : double numPages = index->pages;
8599 10730 : RelOptInfo *baserel = index->rel;
8600 10730 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
8601 : Cost spc_seq_page_cost;
8602 : Cost spc_random_page_cost;
8603 : double qual_arg_cost;
8604 : double qualSelectivity;
8605 : BrinStatsData statsData;
8606 : double indexRanges;
8607 : double minimalRanges;
8608 : double estimatedRanges;
8609 : double selec;
8610 : Relation indexRel;
8611 : ListCell *l;
8612 : VariableStatData vardata;
8613 :
8614 : Assert(rte->rtekind == RTE_RELATION);
8615 :
8616 : /* fetch estimated page cost for the tablespace containing the index */
8617 10730 : get_tablespace_page_costs(index->reltablespace,
8618 : &spc_random_page_cost,
8619 : &spc_seq_page_cost);
8620 :
8621 : /*
8622 : * Obtain some data from the index itself, if possible. Otherwise invent
8623 : * some plausible internal statistics based on the relation page count.
8624 : */
8625 10730 : if (!index->hypothetical)
8626 : {
8627 : /*
8628 : * A lock should have already been obtained on the index in plancat.c.
8629 : */
8630 10730 : indexRel = index_open(index->indexoid, NoLock);
8631 10730 : brinGetStats(indexRel, &statsData);
8632 10730 : index_close(indexRel, NoLock);
8633 :
8634 : /* work out the actual number of ranges in the index */
8635 10730 : indexRanges = Max(ceil((double) baserel->pages /
8636 : statsData.pagesPerRange), 1.0);
8637 : }
8638 : else
8639 : {
8640 : /*
8641 : * Assume default number of pages per range, and estimate the number
8642 : * of ranges based on that.
8643 : */
8644 0 : indexRanges = Max(ceil((double) baserel->pages /
8645 : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
8646 :
8647 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
8648 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
8649 : }
8650 :
8651 : /*
8652 : * Compute index correlation
8653 : *
8654 : * Because we can use all index quals equally when scanning, we can use
8655 : * the largest correlation (in absolute value) among columns used by the
8656 : * query. Start at zero, the worst possible case. If we cannot find any
8657 : * correlation statistics, we will keep it as 0.
8658 : */
8659 10730 : *indexCorrelation = 0;
8660 :
8661 21462 : foreach(l, path->indexclauses)
8662 : {
8663 10732 : IndexClause *iclause = lfirst_node(IndexClause, l);
8664 10732 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
8665 :
8666 : /* attempt to lookup stats in relation for this index column */
8667 10732 : if (attnum != 0)
8668 : {
8669 : /* Simple variable -- look to stats for the underlying table */
8670 10732 : if (get_relation_stats_hook &&
8671 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
8672 : {
8673 : /*
8674 : * The hook took control of acquiring a stats tuple. If it
8675 : * did supply a tuple, it'd better have supplied a freefunc.
8676 : */
8677 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
8678 0 : elog(ERROR,
8679 : "no function provided to release variable stats with");
8680 : }
8681 : else
8682 : {
8683 10732 : vardata.statsTuple =
8684 10732 : SearchSysCache3(STATRELATTINH,
8685 : ObjectIdGetDatum(rte->relid),
8686 : Int16GetDatum(attnum),
8687 : BoolGetDatum(false));
8688 10732 : vardata.freefunc = ReleaseSysCache;
8689 : }
8690 : }
8691 : else
8692 : {
8693 : /*
8694 : * Looks like we've found an expression column in the index. Let's
8695 : * see if there's any stats for it.
8696 : */
8697 :
8698 : /* get the attnum from the 0-based index. */
8699 0 : attnum = iclause->indexcol + 1;
8700 :
8701 0 : if (get_index_stats_hook &&
8702 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
8703 : {
8704 : /*
8705 : * The hook took control of acquiring a stats tuple. If it
8706 : * did supply a tuple, it'd better have supplied a freefunc.
8707 : */
8708 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
8709 0 : !vardata.freefunc)
8710 0 : elog(ERROR, "no function provided to release variable stats with");
8711 : }
8712 : else
8713 : {
8714 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
8715 : ObjectIdGetDatum(index->indexoid),
8716 : Int16GetDatum(attnum),
8717 : BoolGetDatum(false));
8718 0 : vardata.freefunc = ReleaseSysCache;
8719 : }
8720 : }
8721 :
8722 10732 : if (HeapTupleIsValid(vardata.statsTuple))
8723 : {
8724 : AttStatsSlot sslot;
8725 :
8726 36 : if (get_attstatsslot(&sslot, vardata.statsTuple,
8727 : STATISTIC_KIND_CORRELATION, InvalidOid,
8728 : ATTSTATSSLOT_NUMBERS))
8729 : {
8730 36 : double varCorrelation = 0.0;
8731 :
8732 36 : if (sslot.nnumbers > 0)
8733 36 : varCorrelation = fabs(sslot.numbers[0]);
8734 :
8735 36 : if (varCorrelation > *indexCorrelation)
8736 36 : *indexCorrelation = varCorrelation;
8737 :
8738 36 : free_attstatsslot(&sslot);
8739 : }
8740 : }
8741 :
8742 10732 : ReleaseVariableStats(vardata);
8743 : }
8744 :
8745 10730 : qualSelectivity = clauselist_selectivity(root, indexQuals,
8746 10730 : baserel->relid,
8747 : JOIN_INNER, NULL);
8748 :
8749 : /*
8750 : * Now calculate the minimum possible ranges we could match with if all of
8751 : * the rows were in the perfect order in the table's heap.
8752 : */
8753 10730 : minimalRanges = ceil(indexRanges * qualSelectivity);
8754 :
8755 : /*
8756 : * Now estimate the number of ranges that we'll touch by using the
8757 : * indexCorrelation from the stats. Careful not to divide by zero (note
8758 : * we're using the absolute value of the correlation).
8759 : */
8760 10730 : if (*indexCorrelation < 1.0e-10)
8761 10694 : estimatedRanges = indexRanges;
8762 : else
8763 36 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
8764 :
8765 : /* we expect to visit this portion of the table */
8766 10730 : selec = estimatedRanges / indexRanges;
8767 :
8768 10730 : CLAMP_PROBABILITY(selec);
8769 :
8770 10730 : *indexSelectivity = selec;
8771 :
8772 : /*
8773 : * Compute the index qual costs, much as in genericcostestimate, to add to
8774 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
8775 : * support those.
8776 : */
8777 10730 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8778 :
8779 : /*
8780 : * Compute the startup cost as the cost to read the whole revmap
8781 : * sequentially, including the cost to execute the index quals.
8782 : */
8783 10730 : *indexStartupCost =
8784 10730 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
8785 10730 : *indexStartupCost += qual_arg_cost;
8786 :
8787 : /*
8788 : * To read a BRIN index there might be a bit of back and forth over
8789 : * regular pages, as revmap might point to them out of sequential order;
8790 : * calculate the total cost as reading the whole index in random order.
8791 : */
8792 10730 : *indexTotalCost = *indexStartupCost +
8793 10730 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
8794 :
8795 : /*
8796 : * Charge a small amount per range tuple which we expect to match to. This
8797 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
8798 : * will set a bit for each page in the range when we find a matching
8799 : * range, so we must multiply the charge by the number of pages in the
8800 : * range.
8801 : */
8802 10730 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
8803 10730 : statsData.pagesPerRange;
8804 :
8805 10730 : *indexPages = index->pages;
8806 10730 : }
|
