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_am.h"
107 : #include "catalog/pg_collation.h"
108 : #include "catalog/pg_operator.h"
109 : #include "catalog/pg_statistic.h"
110 : #include "catalog/pg_statistic_ext.h"
111 : #include "executor/nodeAgg.h"
112 : #include "miscadmin.h"
113 : #include "nodes/makefuncs.h"
114 : #include "nodes/nodeFuncs.h"
115 : #include "optimizer/clauses.h"
116 : #include "optimizer/cost.h"
117 : #include "optimizer/optimizer.h"
118 : #include "optimizer/pathnode.h"
119 : #include "optimizer/paths.h"
120 : #include "optimizer/plancat.h"
121 : #include "parser/parse_clause.h"
122 : #include "parser/parse_relation.h"
123 : #include "parser/parsetree.h"
124 : #include "rewrite/rewriteManip.h"
125 : #include "statistics/statistics.h"
126 : #include "storage/bufmgr.h"
127 : #include "utils/acl.h"
128 : #include "utils/array.h"
129 : #include "utils/builtins.h"
130 : #include "utils/date.h"
131 : #include "utils/datum.h"
132 : #include "utils/fmgroids.h"
133 : #include "utils/index_selfuncs.h"
134 : #include "utils/lsyscache.h"
135 : #include "utils/memutils.h"
136 : #include "utils/pg_locale.h"
137 : #include "utils/rel.h"
138 : #include "utils/selfuncs.h"
139 : #include "utils/snapmgr.h"
140 : #include "utils/spccache.h"
141 : #include "utils/syscache.h"
142 : #include "utils/timestamp.h"
143 : #include "utils/typcache.h"
144 :
145 : #define DEFAULT_PAGE_CPU_MULTIPLIER 50.0
146 :
147 : /* Hooks for plugins to get control when we ask for stats */
148 : get_relation_stats_hook_type get_relation_stats_hook = NULL;
149 : get_index_stats_hook_type get_index_stats_hook = NULL;
150 :
151 : static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
152 : static double eqjoinsel_inner(Oid opfuncoid, Oid collation,
153 : VariableStatData *vardata1, VariableStatData *vardata2,
154 : double nd1, double nd2,
155 : bool isdefault1, bool isdefault2,
156 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
157 : Form_pg_statistic stats1, Form_pg_statistic stats2,
158 : bool have_mcvs1, bool have_mcvs2);
159 : static double eqjoinsel_semi(Oid opfuncoid, Oid collation,
160 : VariableStatData *vardata1, VariableStatData *vardata2,
161 : double nd1, double nd2,
162 : bool isdefault1, bool isdefault2,
163 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
164 : Form_pg_statistic stats1, Form_pg_statistic stats2,
165 : bool have_mcvs1, bool have_mcvs2,
166 : RelOptInfo *inner_rel);
167 : static bool estimate_multivariate_ndistinct(PlannerInfo *root,
168 : RelOptInfo *rel, List **varinfos, double *ndistinct);
169 : static bool convert_to_scalar(Datum value, Oid valuetypid, Oid collid,
170 : double *scaledvalue,
171 : Datum lobound, Datum hibound, Oid boundstypid,
172 : double *scaledlobound, double *scaledhibound);
173 : static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure);
174 : static void convert_string_to_scalar(char *value,
175 : double *scaledvalue,
176 : char *lobound,
177 : double *scaledlobound,
178 : char *hibound,
179 : double *scaledhibound);
180 : static void convert_bytea_to_scalar(Datum value,
181 : double *scaledvalue,
182 : Datum lobound,
183 : double *scaledlobound,
184 : Datum hibound,
185 : double *scaledhibound);
186 : static double convert_one_string_to_scalar(char *value,
187 : int rangelo, int rangehi);
188 : static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
189 : int rangelo, int rangehi);
190 : static char *convert_string_datum(Datum value, Oid typid, Oid collid,
191 : bool *failure);
192 : static double convert_timevalue_to_scalar(Datum value, Oid typid,
193 : bool *failure);
194 : static void examine_simple_variable(PlannerInfo *root, Var *var,
195 : VariableStatData *vardata);
196 : static void examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
197 : int indexcol, VariableStatData *vardata);
198 : static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata,
199 : Oid sortop, Oid collation,
200 : Datum *min, Datum *max);
201 : static void get_stats_slot_range(AttStatsSlot *sslot,
202 : Oid opfuncoid, FmgrInfo *opproc,
203 : Oid collation, int16 typLen, bool typByVal,
204 : Datum *min, Datum *max, bool *p_have_data);
205 : static bool get_actual_variable_range(PlannerInfo *root,
206 : VariableStatData *vardata,
207 : Oid sortop, Oid collation,
208 : Datum *min, Datum *max);
209 : static bool get_actual_variable_endpoint(Relation heapRel,
210 : Relation indexRel,
211 : ScanDirection indexscandir,
212 : ScanKey scankeys,
213 : int16 typLen,
214 : bool typByVal,
215 : TupleTableSlot *tableslot,
216 : MemoryContext outercontext,
217 : Datum *endpointDatum);
218 : static RelOptInfo *find_join_input_rel(PlannerInfo *root, Relids relids);
219 : static double btcost_correlation(IndexOptInfo *index,
220 : VariableStatData *vardata);
221 :
222 :
223 : /*
224 : * eqsel - Selectivity of "=" for any data types.
225 : *
226 : * Note: this routine is also used to estimate selectivity for some
227 : * operators that are not "=" but have comparable selectivity behavior,
228 : * such as "~=" (geometric approximate-match). Even for "=", we must
229 : * keep in mind that the left and right datatypes may differ.
230 : */
231 : Datum
232 660440 : eqsel(PG_FUNCTION_ARGS)
233 : {
234 660440 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
235 : }
236 :
237 : /*
238 : * Common code for eqsel() and neqsel()
239 : */
240 : static double
241 713406 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
242 : {
243 713406 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
244 713406 : Oid operator = PG_GETARG_OID(1);
245 713406 : List *args = (List *) PG_GETARG_POINTER(2);
246 713406 : int varRelid = PG_GETARG_INT32(3);
247 713406 : Oid collation = PG_GET_COLLATION();
248 : VariableStatData vardata;
249 : Node *other;
250 : bool varonleft;
251 : double selec;
252 :
253 : /*
254 : * When asked about <>, we do the estimation using the corresponding =
255 : * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
256 : */
257 713406 : if (negate)
258 : {
259 52966 : operator = get_negator(operator);
260 52966 : if (!OidIsValid(operator))
261 : {
262 : /* Use default selectivity (should we raise an error instead?) */
263 0 : return 1.0 - DEFAULT_EQ_SEL;
264 : }
265 : }
266 :
267 : /*
268 : * If expression is not variable = something or something = variable, then
269 : * punt and return a default estimate.
270 : */
271 713406 : if (!get_restriction_variable(root, args, varRelid,
272 : &vardata, &other, &varonleft))
273 5370 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
274 :
275 : /*
276 : * We can do a lot better if the something is a constant. (Note: the
277 : * Const might result from estimation rather than being a simple constant
278 : * in the query.)
279 : */
280 708030 : if (IsA(other, Const))
281 315888 : selec = var_eq_const(&vardata, operator, collation,
282 315888 : ((Const *) other)->constvalue,
283 315888 : ((Const *) other)->constisnull,
284 : varonleft, negate);
285 : else
286 392142 : selec = var_eq_non_const(&vardata, operator, collation, other,
287 : varonleft, negate);
288 :
289 708030 : ReleaseVariableStats(vardata);
290 :
291 708030 : return selec;
292 : }
293 :
294 : /*
295 : * var_eq_const --- eqsel for var = const case
296 : *
297 : * This is exported so that some other estimation functions can use it.
298 : */
299 : double
300 361010 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
301 : Datum constval, bool constisnull,
302 : bool varonleft, bool negate)
303 : {
304 : double selec;
305 361010 : double nullfrac = 0.0;
306 : bool isdefault;
307 : Oid opfuncoid;
308 :
309 : /*
310 : * If the constant is NULL, assume operator is strict and return zero, ie,
311 : * operator will never return TRUE. (It's zero even for a negator op.)
312 : */
313 361010 : if (constisnull)
314 404 : return 0.0;
315 :
316 : /*
317 : * Grab the nullfrac for use below. Note we allow use of nullfrac
318 : * regardless of security check.
319 : */
320 360606 : if (HeapTupleIsValid(vardata->statsTuple))
321 : {
322 : Form_pg_statistic stats;
323 :
324 271754 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
325 271754 : nullfrac = stats->stanullfrac;
326 : }
327 :
328 : /*
329 : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
330 : * assume there is exactly one match regardless of anything else. (This
331 : * is slightly bogus, since the index or clause's equality operator might
332 : * be different from ours, but it's much more likely to be right than
333 : * ignoring the information.)
334 : */
335 360606 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
336 : {
337 82692 : selec = 1.0 / vardata->rel->tuples;
338 : }
339 483888 : else if (HeapTupleIsValid(vardata->statsTuple) &&
340 205974 : statistic_proc_security_check(vardata,
341 205974 : (opfuncoid = get_opcode(oproid))))
342 205974 : {
343 : AttStatsSlot sslot;
344 205974 : bool match = false;
345 : int i;
346 :
347 : /*
348 : * Is the constant "=" to any of the column's most common values?
349 : * (Although the given operator may not really be "=", we will assume
350 : * that seeing whether it returns TRUE is an appropriate test. If you
351 : * don't like this, maybe you shouldn't be using eqsel for your
352 : * operator...)
353 : */
354 205974 : if (get_attstatsslot(&sslot, vardata->statsTuple,
355 : STATISTIC_KIND_MCV, InvalidOid,
356 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
357 : {
358 179124 : LOCAL_FCINFO(fcinfo, 2);
359 : FmgrInfo eqproc;
360 :
361 179124 : fmgr_info(opfuncoid, &eqproc);
362 :
363 : /*
364 : * Save a few cycles by setting up the fcinfo struct just once.
365 : * Using FunctionCallInvoke directly also avoids failure if the
366 : * eqproc returns NULL, though really equality functions should
367 : * never do that.
368 : */
369 179124 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
370 : NULL, NULL);
371 179124 : fcinfo->args[0].isnull = false;
372 179124 : fcinfo->args[1].isnull = false;
373 : /* be careful to apply operator right way 'round */
374 179124 : if (varonleft)
375 179092 : fcinfo->args[1].value = constval;
376 : else
377 32 : fcinfo->args[0].value = constval;
378 :
379 2857112 : for (i = 0; i < sslot.nvalues; i++)
380 : {
381 : Datum fresult;
382 :
383 2764008 : if (varonleft)
384 2763952 : fcinfo->args[0].value = sslot.values[i];
385 : else
386 56 : fcinfo->args[1].value = sslot.values[i];
387 2764008 : fcinfo->isnull = false;
388 2764008 : fresult = FunctionCallInvoke(fcinfo);
389 2764008 : if (!fcinfo->isnull && DatumGetBool(fresult))
390 : {
391 86020 : match = true;
392 86020 : break;
393 : }
394 : }
395 : }
396 : else
397 : {
398 : /* no most-common-value info available */
399 26850 : i = 0; /* keep compiler quiet */
400 : }
401 :
402 205974 : if (match)
403 : {
404 : /*
405 : * Constant is "=" to this common value. We know selectivity
406 : * exactly (or as exactly as ANALYZE could calculate it, anyway).
407 : */
408 86020 : selec = sslot.numbers[i];
409 : }
410 : else
411 : {
412 : /*
413 : * Comparison is against a constant that is neither NULL nor any
414 : * of the common values. Its selectivity cannot be more than
415 : * this:
416 : */
417 119954 : double sumcommon = 0.0;
418 : double otherdistinct;
419 :
420 2514738 : for (i = 0; i < sslot.nnumbers; i++)
421 2394784 : sumcommon += sslot.numbers[i];
422 119954 : selec = 1.0 - sumcommon - nullfrac;
423 119954 : CLAMP_PROBABILITY(selec);
424 :
425 : /*
426 : * and in fact it's probably a good deal less. We approximate that
427 : * all the not-common values share this remaining fraction
428 : * equally, so we divide by the number of other distinct values.
429 : */
430 119954 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
431 119954 : sslot.nnumbers;
432 119954 : if (otherdistinct > 1)
433 55694 : selec /= otherdistinct;
434 :
435 : /*
436 : * Another cross-check: selectivity shouldn't be estimated as more
437 : * than the least common "most common value".
438 : */
439 119954 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
440 0 : selec = sslot.numbers[sslot.nnumbers - 1];
441 : }
442 :
443 205974 : free_attstatsslot(&sslot);
444 : }
445 : else
446 : {
447 : /*
448 : * No ANALYZE stats available, so make a guess using estimated number
449 : * of distinct values and assuming they are equally common. (The guess
450 : * is unlikely to be very good, but we do know a few special cases.)
451 : */
452 71940 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
453 : }
454 :
455 : /* now adjust if we wanted <> rather than = */
456 360606 : if (negate)
457 43476 : selec = 1.0 - selec - nullfrac;
458 :
459 : /* result should be in range, but make sure... */
460 360606 : CLAMP_PROBABILITY(selec);
461 :
462 360606 : return selec;
463 : }
464 :
465 : /*
466 : * var_eq_non_const --- eqsel for var = something-other-than-const case
467 : *
468 : * This is exported so that some other estimation functions can use it.
469 : */
470 : double
471 392142 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
472 : Node *other,
473 : bool varonleft, bool negate)
474 : {
475 : double selec;
476 392142 : double nullfrac = 0.0;
477 : bool isdefault;
478 :
479 : /*
480 : * Grab the nullfrac for use below.
481 : */
482 392142 : if (HeapTupleIsValid(vardata->statsTuple))
483 : {
484 : Form_pg_statistic stats;
485 :
486 281934 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
487 281934 : nullfrac = stats->stanullfrac;
488 : }
489 :
490 : /*
491 : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
492 : * assume there is exactly one match regardless of anything else. (This
493 : * is slightly bogus, since the index or clause's equality operator might
494 : * be different from ours, but it's much more likely to be right than
495 : * ignoring the information.)
496 : */
497 392142 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
498 : {
499 155062 : selec = 1.0 / vardata->rel->tuples;
500 : }
501 237080 : else if (HeapTupleIsValid(vardata->statsTuple))
502 : {
503 : double ndistinct;
504 : AttStatsSlot sslot;
505 :
506 : /*
507 : * Search is for a value that we do not know a priori, but we will
508 : * assume it is not NULL. Estimate the selectivity as non-null
509 : * fraction divided by number of distinct values, so that we get a
510 : * result averaged over all possible values whether common or
511 : * uncommon. (Essentially, we are assuming that the not-yet-known
512 : * comparison value is equally likely to be any of the possible
513 : * values, regardless of their frequency in the table. Is that a good
514 : * idea?)
515 : */
516 141986 : selec = 1.0 - nullfrac;
517 141986 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
518 141986 : if (ndistinct > 1)
519 137922 : selec /= ndistinct;
520 :
521 : /*
522 : * Cross-check: selectivity should never be estimated as more than the
523 : * most common value's.
524 : */
525 141986 : if (get_attstatsslot(&sslot, vardata->statsTuple,
526 : STATISTIC_KIND_MCV, InvalidOid,
527 : ATTSTATSSLOT_NUMBERS))
528 : {
529 123766 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
530 558 : selec = sslot.numbers[0];
531 123766 : free_attstatsslot(&sslot);
532 : }
533 : }
534 : else
535 : {
536 : /*
537 : * No ANALYZE stats available, so make a guess using estimated number
538 : * of distinct values and assuming they are equally common. (The guess
539 : * is unlikely to be very good, but we do know a few special cases.)
540 : */
541 95094 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
542 : }
543 :
544 : /* now adjust if we wanted <> rather than = */
545 392142 : if (negate)
546 7070 : selec = 1.0 - selec - nullfrac;
547 :
548 : /* result should be in range, but make sure... */
549 392142 : CLAMP_PROBABILITY(selec);
550 :
551 392142 : return selec;
552 : }
553 :
554 : /*
555 : * neqsel - Selectivity of "!=" for any data types.
556 : *
557 : * This routine is also used for some operators that are not "!="
558 : * but have comparable selectivity behavior. See above comments
559 : * for eqsel().
560 : */
561 : Datum
562 52966 : neqsel(PG_FUNCTION_ARGS)
563 : {
564 52966 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
565 : }
566 :
567 : /*
568 : * scalarineqsel - Selectivity of "<", "<=", ">", ">=" for scalars.
569 : *
570 : * This is the guts of scalarltsel/scalarlesel/scalargtsel/scalargesel.
571 : * The isgt and iseq flags distinguish which of the four cases apply.
572 : *
573 : * The caller has commuted the clause, if necessary, so that we can treat
574 : * the variable as being on the left. The caller must also make sure that
575 : * the other side of the clause is a non-null Const, and dissect that into
576 : * a value and datatype. (This definition simplifies some callers that
577 : * want to estimate against a computed value instead of a Const node.)
578 : *
579 : * This routine works for any datatype (or pair of datatypes) known to
580 : * convert_to_scalar(). If it is applied to some other datatype,
581 : * it will return an approximate estimate based on assuming that the constant
582 : * value falls in the middle of the bin identified by binary search.
583 : */
584 : static double
585 312572 : scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
586 : Oid collation,
587 : VariableStatData *vardata, Datum constval, Oid consttype)
588 : {
589 : Form_pg_statistic stats;
590 : FmgrInfo opproc;
591 : double mcv_selec,
592 : hist_selec,
593 : sumcommon;
594 : double selec;
595 :
596 312572 : if (!HeapTupleIsValid(vardata->statsTuple))
597 : {
598 : /*
599 : * No stats are available. Typically this means we have to fall back
600 : * on the default estimate; but if the variable is CTID then we can
601 : * make an estimate based on comparing the constant to the table size.
602 : */
603 22066 : if (vardata->var && IsA(vardata->var, Var) &&
604 17314 : ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
605 : {
606 : ItemPointer itemptr;
607 : double block;
608 : double density;
609 :
610 : /*
611 : * If the relation's empty, we're going to include all of it.
612 : * (This is mostly to avoid divide-by-zero below.)
613 : */
614 1952 : if (vardata->rel->pages == 0)
615 0 : return 1.0;
616 :
617 1952 : itemptr = (ItemPointer) DatumGetPointer(constval);
618 1952 : block = ItemPointerGetBlockNumberNoCheck(itemptr);
619 :
620 : /*
621 : * Determine the average number of tuples per page (density).
622 : *
623 : * Since the last page will, on average, be only half full, we can
624 : * estimate it to have half as many tuples as earlier pages. So
625 : * give it half the weight of a regular page.
626 : */
627 1952 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
628 :
629 : /* If target is the last page, use half the density. */
630 1952 : if (block >= vardata->rel->pages - 1)
631 42 : density *= 0.5;
632 :
633 : /*
634 : * Using the average tuples per page, calculate how far into the
635 : * page the itemptr is likely to be and adjust block accordingly,
636 : * by adding that fraction of a whole block (but never more than a
637 : * whole block, no matter how high the itemptr's offset is). Here
638 : * we are ignoring the possibility of dead-tuple line pointers,
639 : * which is fairly bogus, but we lack the info to do better.
640 : */
641 1952 : if (density > 0.0)
642 : {
643 1952 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
644 :
645 1952 : block += Min(offset / density, 1.0);
646 : }
647 :
648 : /*
649 : * Convert relative block number to selectivity. Again, the last
650 : * page has only half weight.
651 : */
652 1952 : selec = block / (vardata->rel->pages - 0.5);
653 :
654 : /*
655 : * The calculation so far gave us a selectivity for the "<=" case.
656 : * We'll have one fewer tuple for "<" and one additional tuple for
657 : * ">=", the latter of which we'll reverse the selectivity for
658 : * below, so we can simply subtract one tuple for both cases. The
659 : * cases that need this adjustment can be identified by iseq being
660 : * equal to isgt.
661 : */
662 1952 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
663 1840 : selec -= (1.0 / vardata->rel->tuples);
664 :
665 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
666 1952 : if (isgt)
667 1834 : selec = 1.0 - selec;
668 :
669 1952 : CLAMP_PROBABILITY(selec);
670 1952 : return selec;
671 : }
672 :
673 : /* no stats available, so default result */
674 20114 : return DEFAULT_INEQ_SEL;
675 : }
676 290506 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
677 :
678 290506 : fmgr_info(get_opcode(operator), &opproc);
679 :
680 : /*
681 : * If we have most-common-values info, add up the fractions of the MCV
682 : * entries that satisfy MCV OP CONST. These fractions contribute directly
683 : * to the result selectivity. Also add up the total fraction represented
684 : * by MCV entries.
685 : */
686 290506 : mcv_selec = mcv_selectivity(vardata, &opproc, collation, constval, true,
687 : &sumcommon);
688 :
689 : /*
690 : * If there is a histogram, determine which bin the constant falls in, and
691 : * compute the resulting contribution to selectivity.
692 : */
693 290506 : hist_selec = ineq_histogram_selectivity(root, vardata,
694 : operator, &opproc, isgt, iseq,
695 : collation,
696 : constval, consttype);
697 :
698 : /*
699 : * Now merge the results from the MCV and histogram calculations,
700 : * realizing that the histogram covers only the non-null values that are
701 : * not listed in MCV.
702 : */
703 290506 : selec = 1.0 - stats->stanullfrac - sumcommon;
704 :
705 290506 : if (hist_selec >= 0.0)
706 215258 : selec *= hist_selec;
707 : else
708 : {
709 : /*
710 : * If no histogram but there are values not accounted for by MCV,
711 : * arbitrarily assume half of them will match.
712 : */
713 75248 : selec *= 0.5;
714 : }
715 :
716 290506 : selec += mcv_selec;
717 :
718 : /* result should be in range, but make sure... */
719 290506 : CLAMP_PROBABILITY(selec);
720 :
721 290506 : return selec;
722 : }
723 :
724 : /*
725 : * mcv_selectivity - Examine the MCV list for selectivity estimates
726 : *
727 : * Determine the fraction of the variable's MCV population that satisfies
728 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft. Also
729 : * compute the fraction of the total column population represented by the MCV
730 : * list. This code will work for any boolean-returning predicate operator.
731 : *
732 : * The function result is the MCV selectivity, and the fraction of the
733 : * total population is returned into *sumcommonp. Zeroes are returned
734 : * if there is no MCV list.
735 : */
736 : double
737 296440 : mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation,
738 : Datum constval, bool varonleft,
739 : double *sumcommonp)
740 : {
741 : double mcv_selec,
742 : sumcommon;
743 : AttStatsSlot sslot;
744 : int i;
745 :
746 296440 : mcv_selec = 0.0;
747 296440 : sumcommon = 0.0;
748 :
749 590506 : if (HeapTupleIsValid(vardata->statsTuple) &&
750 588078 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
751 294012 : get_attstatsslot(&sslot, vardata->statsTuple,
752 : STATISTIC_KIND_MCV, InvalidOid,
753 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
754 : {
755 139154 : LOCAL_FCINFO(fcinfo, 2);
756 :
757 : /*
758 : * We invoke the opproc "by hand" so that we won't fail on NULL
759 : * results. Such cases won't arise for normal comparison functions,
760 : * but generic_restriction_selectivity could perhaps be used with
761 : * operators that can return NULL. A small side benefit is to not
762 : * need to re-initialize the fcinfo struct from scratch each time.
763 : */
764 139154 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
765 : NULL, NULL);
766 139154 : fcinfo->args[0].isnull = false;
767 139154 : fcinfo->args[1].isnull = false;
768 : /* be careful to apply operator right way 'round */
769 139154 : if (varonleft)
770 139154 : fcinfo->args[1].value = constval;
771 : else
772 0 : fcinfo->args[0].value = constval;
773 :
774 4316824 : for (i = 0; i < sslot.nvalues; i++)
775 : {
776 : Datum fresult;
777 :
778 4177670 : if (varonleft)
779 4177670 : fcinfo->args[0].value = sslot.values[i];
780 : else
781 0 : fcinfo->args[1].value = sslot.values[i];
782 4177670 : fcinfo->isnull = false;
783 4177670 : fresult = FunctionCallInvoke(fcinfo);
784 4177670 : if (!fcinfo->isnull && DatumGetBool(fresult))
785 1527712 : mcv_selec += sslot.numbers[i];
786 4177670 : sumcommon += sslot.numbers[i];
787 : }
788 139154 : free_attstatsslot(&sslot);
789 : }
790 :
791 296440 : *sumcommonp = sumcommon;
792 296440 : return mcv_selec;
793 : }
794 :
795 : /*
796 : * histogram_selectivity - Examine the histogram for selectivity estimates
797 : *
798 : * Determine the fraction of the variable's histogram entries that satisfy
799 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft.
800 : *
801 : * This code will work for any boolean-returning predicate operator, whether
802 : * or not it has anything to do with the histogram sort operator. We are
803 : * essentially using the histogram just as a representative sample. However,
804 : * small histograms are unlikely to be all that representative, so the caller
805 : * should be prepared to fall back on some other estimation approach when the
806 : * histogram is missing or very small. It may also be prudent to combine this
807 : * approach with another one when the histogram is small.
808 : *
809 : * If the actual histogram size is not at least min_hist_size, we won't bother
810 : * to do the calculation at all. Also, if the n_skip parameter is > 0, we
811 : * ignore the first and last n_skip histogram elements, on the grounds that
812 : * they are outliers and hence not very representative. Typical values for
813 : * these parameters are 10 and 1.
814 : *
815 : * The function result is the selectivity, or -1 if there is no histogram
816 : * or it's smaller than min_hist_size.
817 : *
818 : * The output parameter *hist_size receives the actual histogram size,
819 : * or zero if no histogram. Callers may use this number to decide how
820 : * much faith to put in the function result.
821 : *
822 : * Note that the result disregards both the most-common-values (if any) and
823 : * null entries. The caller is expected to combine this result with
824 : * statistics for those portions of the column population. It may also be
825 : * prudent to clamp the result range, ie, disbelieve exact 0 or 1 outputs.
826 : */
827 : double
828 5934 : histogram_selectivity(VariableStatData *vardata,
829 : FmgrInfo *opproc, Oid collation,
830 : Datum constval, bool varonleft,
831 : int min_hist_size, int n_skip,
832 : int *hist_size)
833 : {
834 : double result;
835 : AttStatsSlot sslot;
836 :
837 : /* check sanity of parameters */
838 : Assert(n_skip >= 0);
839 : Assert(min_hist_size > 2 * n_skip);
840 :
841 9494 : if (HeapTupleIsValid(vardata->statsTuple) &&
842 7114 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
843 3554 : get_attstatsslot(&sslot, vardata->statsTuple,
844 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
845 : ATTSTATSSLOT_VALUES))
846 : {
847 3458 : *hist_size = sslot.nvalues;
848 3458 : if (sslot.nvalues >= min_hist_size)
849 : {
850 1736 : LOCAL_FCINFO(fcinfo, 2);
851 1736 : int nmatch = 0;
852 : int i;
853 :
854 : /*
855 : * We invoke the opproc "by hand" so that we won't fail on NULL
856 : * results. Such cases won't arise for normal comparison
857 : * functions, but generic_restriction_selectivity could perhaps be
858 : * used with operators that can return NULL. A small side benefit
859 : * is to not need to re-initialize the fcinfo struct from scratch
860 : * each time.
861 : */
862 1736 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
863 : NULL, NULL);
864 1736 : fcinfo->args[0].isnull = false;
865 1736 : fcinfo->args[1].isnull = false;
866 : /* be careful to apply operator right way 'round */
867 1736 : if (varonleft)
868 1736 : fcinfo->args[1].value = constval;
869 : else
870 0 : fcinfo->args[0].value = constval;
871 :
872 141544 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
873 : {
874 : Datum fresult;
875 :
876 139808 : if (varonleft)
877 139808 : fcinfo->args[0].value = sslot.values[i];
878 : else
879 0 : fcinfo->args[1].value = sslot.values[i];
880 139808 : fcinfo->isnull = false;
881 139808 : fresult = FunctionCallInvoke(fcinfo);
882 139808 : if (!fcinfo->isnull && DatumGetBool(fresult))
883 7422 : nmatch++;
884 : }
885 1736 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
886 : }
887 : else
888 1722 : result = -1;
889 3458 : free_attstatsslot(&sslot);
890 : }
891 : else
892 : {
893 2476 : *hist_size = 0;
894 2476 : result = -1;
895 : }
896 :
897 5934 : return result;
898 : }
899 :
900 : /*
901 : * generic_restriction_selectivity - Selectivity for almost anything
902 : *
903 : * This function estimates selectivity for operators that we don't have any
904 : * special knowledge about, but are on data types that we collect standard
905 : * MCV and/or histogram statistics for. (Additional assumptions are that
906 : * the operator is strict and immutable, or at least stable.)
907 : *
908 : * If we have "VAR OP CONST" or "CONST OP VAR", selectivity is estimated by
909 : * applying the operator to each element of the column's MCV and/or histogram
910 : * stats, and merging the results using the assumption that the histogram is
911 : * a reasonable random sample of the column's non-MCV population. Note that
912 : * if the operator's semantics are related to the histogram ordering, this
913 : * might not be such a great assumption; other functions such as
914 : * scalarineqsel() are probably a better match in such cases.
915 : *
916 : * Otherwise, fall back to the default selectivity provided by the caller.
917 : */
918 : double
919 1106 : generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation,
920 : List *args, int varRelid,
921 : double default_selectivity)
922 : {
923 : double selec;
924 : VariableStatData vardata;
925 : Node *other;
926 : bool varonleft;
927 :
928 : /*
929 : * If expression is not variable OP something or something OP variable,
930 : * then punt and return the default estimate.
931 : */
932 1106 : if (!get_restriction_variable(root, args, varRelid,
933 : &vardata, &other, &varonleft))
934 0 : return default_selectivity;
935 :
936 : /*
937 : * If the something is a NULL constant, assume operator is strict and
938 : * return zero, ie, operator will never return TRUE.
939 : */
940 1106 : if (IsA(other, Const) &&
941 1106 : ((Const *) other)->constisnull)
942 : {
943 0 : ReleaseVariableStats(vardata);
944 0 : return 0.0;
945 : }
946 :
947 1106 : if (IsA(other, Const))
948 : {
949 : /* Variable is being compared to a known non-null constant */
950 1106 : Datum constval = ((Const *) other)->constvalue;
951 : FmgrInfo opproc;
952 : double mcvsum;
953 : double mcvsel;
954 : double nullfrac;
955 : int hist_size;
956 :
957 1106 : fmgr_info(get_opcode(oproid), &opproc);
958 :
959 : /*
960 : * Calculate the selectivity for the column's most common values.
961 : */
962 1106 : mcvsel = mcv_selectivity(&vardata, &opproc, collation,
963 : constval, varonleft,
964 : &mcvsum);
965 :
966 : /*
967 : * If the histogram is large enough, see what fraction of it matches
968 : * the query, and assume that's representative of the non-MCV
969 : * population. Otherwise use the default selectivity for the non-MCV
970 : * population.
971 : */
972 1106 : selec = histogram_selectivity(&vardata, &opproc, collation,
973 : constval, varonleft,
974 : 10, 1, &hist_size);
975 1106 : if (selec < 0)
976 : {
977 : /* Nope, fall back on default */
978 1106 : selec = default_selectivity;
979 : }
980 0 : else if (hist_size < 100)
981 : {
982 : /*
983 : * For histogram sizes from 10 to 100, we combine the histogram
984 : * and default selectivities, putting increasingly more trust in
985 : * the histogram for larger sizes.
986 : */
987 0 : double hist_weight = hist_size / 100.0;
988 :
989 0 : selec = selec * hist_weight +
990 0 : default_selectivity * (1.0 - hist_weight);
991 : }
992 :
993 : /* In any case, don't believe extremely small or large estimates. */
994 1106 : if (selec < 0.0001)
995 0 : selec = 0.0001;
996 1106 : else if (selec > 0.9999)
997 0 : selec = 0.9999;
998 :
999 : /* Don't forget to account for nulls. */
1000 1106 : if (HeapTupleIsValid(vardata.statsTuple))
1001 84 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
1002 : else
1003 1022 : nullfrac = 0.0;
1004 :
1005 : /*
1006 : * Now merge the results from the MCV and histogram calculations,
1007 : * realizing that the histogram covers only the non-null values that
1008 : * are not listed in MCV.
1009 : */
1010 1106 : selec *= 1.0 - nullfrac - mcvsum;
1011 1106 : selec += mcvsel;
1012 : }
1013 : else
1014 : {
1015 : /* Comparison value is not constant, so we can't do anything */
1016 0 : selec = default_selectivity;
1017 : }
1018 :
1019 1106 : ReleaseVariableStats(vardata);
1020 :
1021 : /* result should be in range, but make sure... */
1022 1106 : CLAMP_PROBABILITY(selec);
1023 :
1024 1106 : return selec;
1025 : }
1026 :
1027 : /*
1028 : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
1029 : *
1030 : * Determine the fraction of the variable's histogram population that
1031 : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
1032 : * The isgt and iseq flags distinguish which of the four cases apply.
1033 : *
1034 : * While opproc could be looked up from the operator OID, common callers
1035 : * also need to call it separately, so we make the caller pass both.
1036 : *
1037 : * Returns -1 if there is no histogram (valid results will always be >= 0).
1038 : *
1039 : * Note that the result disregards both the most-common-values (if any) and
1040 : * null entries. The caller is expected to combine this result with
1041 : * statistics for those portions of the column population.
1042 : *
1043 : * This is exported so that some other estimation functions can use it.
1044 : */
1045 : double
1046 295328 : ineq_histogram_selectivity(PlannerInfo *root,
1047 : VariableStatData *vardata,
1048 : Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq,
1049 : Oid collation,
1050 : Datum constval, Oid consttype)
1051 : {
1052 : double hist_selec;
1053 : AttStatsSlot sslot;
1054 :
1055 295328 : hist_selec = -1.0;
1056 :
1057 : /*
1058 : * Someday, ANALYZE might store more than one histogram per rel/att,
1059 : * corresponding to more than one possible sort ordering defined for the
1060 : * column type. Right now, we know there is only one, so just grab it and
1061 : * see if it matches the query.
1062 : *
1063 : * Note that we can't use opoid as search argument; the staop appearing in
1064 : * pg_statistic will be for the relevant '<' operator, but what we have
1065 : * might be some other inequality operator such as '>='. (Even if opoid
1066 : * is a '<' operator, it could be cross-type.) Hence we must use
1067 : * comparison_ops_are_compatible() to see if the operators match.
1068 : */
1069 589990 : if (HeapTupleIsValid(vardata->statsTuple) &&
1070 589276 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1071 294614 : get_attstatsslot(&sslot, vardata->statsTuple,
1072 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1073 : ATTSTATSSLOT_VALUES))
1074 : {
1075 219410 : if (sslot.nvalues > 1 &&
1076 438744 : sslot.stacoll == collation &&
1077 219334 : comparison_ops_are_compatible(sslot.staop, opoid))
1078 219226 : {
1079 : /*
1080 : * Use binary search to find the desired location, namely the
1081 : * right end of the histogram bin containing the comparison value,
1082 : * which is the leftmost entry for which the comparison operator
1083 : * succeeds (if isgt) or fails (if !isgt).
1084 : *
1085 : * In this loop, we pay no attention to whether the operator iseq
1086 : * or not; that detail will be mopped up below. (We cannot tell,
1087 : * anyway, whether the operator thinks the values are equal.)
1088 : *
1089 : * If the binary search accesses the first or last histogram
1090 : * entry, we try to replace that endpoint with the true column min
1091 : * or max as found by get_actual_variable_range(). This
1092 : * ameliorates misestimates when the min or max is moving as a
1093 : * result of changes since the last ANALYZE. Note that this could
1094 : * result in effectively including MCVs into the histogram that
1095 : * weren't there before, but we don't try to correct for that.
1096 : */
1097 : double histfrac;
1098 219226 : int lobound = 0; /* first possible slot to search */
1099 219226 : int hibound = sslot.nvalues; /* last+1 slot to search */
1100 219226 : bool have_end = false;
1101 :
1102 : /*
1103 : * If there are only two histogram entries, we'll want up-to-date
1104 : * values for both. (If there are more than two, we need at most
1105 : * one of them to be updated, so we deal with that within the
1106 : * loop.)
1107 : */
1108 219226 : if (sslot.nvalues == 2)
1109 2270 : have_end = get_actual_variable_range(root,
1110 : vardata,
1111 : sslot.staop,
1112 : collation,
1113 : &sslot.values[0],
1114 2270 : &sslot.values[1]);
1115 :
1116 1419040 : while (lobound < hibound)
1117 : {
1118 1199814 : int probe = (lobound + hibound) / 2;
1119 : bool ltcmp;
1120 :
1121 : /*
1122 : * If we find ourselves about to compare to the first or last
1123 : * histogram entry, first try to replace it with the actual
1124 : * current min or max (unless we already did so above).
1125 : */
1126 1199814 : if (probe == 0 && sslot.nvalues > 2)
1127 110264 : have_end = get_actual_variable_range(root,
1128 : vardata,
1129 : sslot.staop,
1130 : collation,
1131 : &sslot.values[0],
1132 : NULL);
1133 1089550 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1134 73828 : have_end = get_actual_variable_range(root,
1135 : vardata,
1136 : sslot.staop,
1137 : collation,
1138 : NULL,
1139 73828 : &sslot.values[probe]);
1140 :
1141 1199814 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1142 : collation,
1143 1199814 : sslot.values[probe],
1144 : constval));
1145 1199814 : if (isgt)
1146 67800 : ltcmp = !ltcmp;
1147 1199814 : if (ltcmp)
1148 458128 : lobound = probe + 1;
1149 : else
1150 741686 : hibound = probe;
1151 : }
1152 :
1153 219226 : if (lobound <= 0)
1154 : {
1155 : /*
1156 : * Constant is below lower histogram boundary. More
1157 : * precisely, we have found that no entry in the histogram
1158 : * satisfies the inequality clause (if !isgt) or they all do
1159 : * (if isgt). We estimate that that's true of the entire
1160 : * table, so set histfrac to 0.0 (which we'll flip to 1.0
1161 : * below, if isgt).
1162 : */
1163 94992 : histfrac = 0.0;
1164 : }
1165 124234 : else if (lobound >= sslot.nvalues)
1166 : {
1167 : /*
1168 : * Inverse case: constant is above upper histogram boundary.
1169 : */
1170 32650 : histfrac = 1.0;
1171 : }
1172 : else
1173 : {
1174 : /* We have values[i-1] <= constant <= values[i]. */
1175 91584 : int i = lobound;
1176 91584 : double eq_selec = 0;
1177 : double val,
1178 : high,
1179 : low;
1180 : double binfrac;
1181 :
1182 : /*
1183 : * In the cases where we'll need it below, obtain an estimate
1184 : * of the selectivity of "x = constval". We use a calculation
1185 : * similar to what var_eq_const() does for a non-MCV constant,
1186 : * ie, estimate that all distinct non-MCV values occur equally
1187 : * often. But multiplication by "1.0 - sumcommon - nullfrac"
1188 : * will be done by our caller, so we shouldn't do that here.
1189 : * Therefore we can't try to clamp the estimate by reference
1190 : * to the least common MCV; the result would be too small.
1191 : *
1192 : * Note: since this is effectively assuming that constval
1193 : * isn't an MCV, it's logically dubious if constval in fact is
1194 : * one. But we have to apply *some* correction for equality,
1195 : * and anyway we cannot tell if constval is an MCV, since we
1196 : * don't have a suitable equality operator at hand.
1197 : */
1198 91584 : if (i == 1 || isgt == iseq)
1199 : {
1200 : double otherdistinct;
1201 : bool isdefault;
1202 : AttStatsSlot mcvslot;
1203 :
1204 : /* Get estimated number of distinct values */
1205 37186 : otherdistinct = get_variable_numdistinct(vardata,
1206 : &isdefault);
1207 :
1208 : /* Subtract off the number of known MCVs */
1209 37186 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1210 : STATISTIC_KIND_MCV, InvalidOid,
1211 : ATTSTATSSLOT_NUMBERS))
1212 : {
1213 3914 : otherdistinct -= mcvslot.nnumbers;
1214 3914 : free_attstatsslot(&mcvslot);
1215 : }
1216 :
1217 : /* If result doesn't seem sane, leave eq_selec at 0 */
1218 37186 : if (otherdistinct > 1)
1219 37144 : eq_selec = 1.0 / otherdistinct;
1220 : }
1221 :
1222 : /*
1223 : * Convert the constant and the two nearest bin boundary
1224 : * values to a uniform comparison scale, and do a linear
1225 : * interpolation within this bin.
1226 : */
1227 91584 : if (convert_to_scalar(constval, consttype, collation,
1228 : &val,
1229 91584 : sslot.values[i - 1], sslot.values[i],
1230 : vardata->vartype,
1231 : &low, &high))
1232 : {
1233 91584 : if (high <= low)
1234 : {
1235 : /* cope if bin boundaries appear identical */
1236 0 : binfrac = 0.5;
1237 : }
1238 91584 : else if (val <= low)
1239 22104 : binfrac = 0.0;
1240 69480 : else if (val >= high)
1241 3060 : binfrac = 1.0;
1242 : else
1243 : {
1244 66420 : binfrac = (val - low) / (high - low);
1245 :
1246 : /*
1247 : * Watch out for the possibility that we got a NaN or
1248 : * Infinity from the division. This can happen
1249 : * despite the previous checks, if for example "low"
1250 : * is -Infinity.
1251 : */
1252 66420 : if (isnan(binfrac) ||
1253 66420 : binfrac < 0.0 || binfrac > 1.0)
1254 0 : binfrac = 0.5;
1255 : }
1256 : }
1257 : else
1258 : {
1259 : /*
1260 : * Ideally we'd produce an error here, on the grounds that
1261 : * the given operator shouldn't have scalarXXsel
1262 : * registered as its selectivity func unless we can deal
1263 : * with its operand types. But currently, all manner of
1264 : * stuff is invoking scalarXXsel, so give a default
1265 : * estimate until that can be fixed.
1266 : */
1267 0 : binfrac = 0.5;
1268 : }
1269 :
1270 : /*
1271 : * Now, compute the overall selectivity across the values
1272 : * represented by the histogram. We have i-1 full bins and
1273 : * binfrac partial bin below the constant.
1274 : */
1275 91584 : histfrac = (double) (i - 1) + binfrac;
1276 91584 : histfrac /= (double) (sslot.nvalues - 1);
1277 :
1278 : /*
1279 : * At this point, histfrac is an estimate of the fraction of
1280 : * the population represented by the histogram that satisfies
1281 : * "x <= constval". Somewhat remarkably, this statement is
1282 : * true regardless of which operator we were doing the probes
1283 : * with, so long as convert_to_scalar() delivers reasonable
1284 : * results. If the probe constant is equal to some histogram
1285 : * entry, we would have considered the bin to the left of that
1286 : * entry if probing with "<" or ">=", or the bin to the right
1287 : * if probing with "<=" or ">"; but binfrac would have come
1288 : * out as 1.0 in the first case and 0.0 in the second, leading
1289 : * to the same histfrac in either case. For probe constants
1290 : * between histogram entries, we find the same bin and get the
1291 : * same estimate with any operator.
1292 : *
1293 : * The fact that the estimate corresponds to "x <= constval"
1294 : * and not "x < constval" is because of the way that ANALYZE
1295 : * constructs the histogram: each entry is, effectively, the
1296 : * rightmost value in its sample bucket. So selectivity
1297 : * values that are exact multiples of 1/(histogram_size-1)
1298 : * should be understood as estimates including a histogram
1299 : * entry plus everything to its left.
1300 : *
1301 : * However, that breaks down for the first histogram entry,
1302 : * which necessarily is the leftmost value in its sample
1303 : * bucket. That means the first histogram bin is slightly
1304 : * narrower than the rest, by an amount equal to eq_selec.
1305 : * Another way to say that is that we want "x <= leftmost" to
1306 : * be estimated as eq_selec not zero. So, if we're dealing
1307 : * with the first bin (i==1), rescale to make that true while
1308 : * adjusting the rest of that bin linearly.
1309 : */
1310 91584 : if (i == 1)
1311 16858 : histfrac += eq_selec * (1.0 - binfrac);
1312 :
1313 : /*
1314 : * "x <= constval" is good if we want an estimate for "<=" or
1315 : * ">", but if we are estimating for "<" or ">=", we now need
1316 : * to decrease the estimate by eq_selec.
1317 : */
1318 91584 : if (isgt == iseq)
1319 27742 : histfrac -= eq_selec;
1320 : }
1321 :
1322 : /*
1323 : * Now the estimate is finished for "<" and "<=" cases. If we are
1324 : * estimating for ">" or ">=", flip it.
1325 : */
1326 219226 : hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1327 :
1328 : /*
1329 : * The histogram boundaries are only approximate to begin with,
1330 : * and may well be out of date anyway. Therefore, don't believe
1331 : * extremely small or large selectivity estimates --- unless we
1332 : * got actual current endpoint values from the table, in which
1333 : * case just do the usual sanity clamp. Somewhat arbitrarily, we
1334 : * set the cutoff for other cases at a hundredth of the histogram
1335 : * resolution.
1336 : */
1337 219226 : if (have_end)
1338 122900 : CLAMP_PROBABILITY(hist_selec);
1339 : else
1340 : {
1341 96326 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1342 :
1343 96326 : if (hist_selec < cutoff)
1344 35008 : hist_selec = cutoff;
1345 61318 : else if (hist_selec > 1.0 - cutoff)
1346 21968 : hist_selec = 1.0 - cutoff;
1347 : }
1348 : }
1349 184 : else if (sslot.nvalues > 1)
1350 : {
1351 : /*
1352 : * If we get here, we have a histogram but it's not sorted the way
1353 : * we want. Do a brute-force search to see how many of the
1354 : * entries satisfy the comparison condition, and take that
1355 : * fraction as our estimate. (This is identical to the inner loop
1356 : * of histogram_selectivity; maybe share code?)
1357 : */
1358 184 : LOCAL_FCINFO(fcinfo, 2);
1359 184 : int nmatch = 0;
1360 :
1361 184 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1362 : NULL, NULL);
1363 184 : fcinfo->args[0].isnull = false;
1364 184 : fcinfo->args[1].isnull = false;
1365 184 : fcinfo->args[1].value = constval;
1366 962100 : for (int i = 0; i < sslot.nvalues; i++)
1367 : {
1368 : Datum fresult;
1369 :
1370 961916 : fcinfo->args[0].value = sslot.values[i];
1371 961916 : fcinfo->isnull = false;
1372 961916 : fresult = FunctionCallInvoke(fcinfo);
1373 961916 : if (!fcinfo->isnull && DatumGetBool(fresult))
1374 1932 : nmatch++;
1375 : }
1376 184 : hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1377 :
1378 : /*
1379 : * As above, clamp to a hundredth of the histogram resolution.
1380 : * This case is surely even less trustworthy than the normal one,
1381 : * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1382 : * clamp should be more restrictive in this case?)
1383 : */
1384 : {
1385 184 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1386 :
1387 184 : if (hist_selec < cutoff)
1388 12 : hist_selec = cutoff;
1389 172 : else if (hist_selec > 1.0 - cutoff)
1390 12 : hist_selec = 1.0 - cutoff;
1391 : }
1392 : }
1393 :
1394 219410 : free_attstatsslot(&sslot);
1395 : }
1396 :
1397 295328 : return hist_selec;
1398 : }
1399 :
1400 : /*
1401 : * Common wrapper function for the selectivity estimators that simply
1402 : * invoke scalarineqsel().
1403 : */
1404 : static Datum
1405 44204 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1406 : {
1407 44204 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1408 44204 : Oid operator = PG_GETARG_OID(1);
1409 44204 : List *args = (List *) PG_GETARG_POINTER(2);
1410 44204 : int varRelid = PG_GETARG_INT32(3);
1411 44204 : Oid collation = PG_GET_COLLATION();
1412 : VariableStatData vardata;
1413 : Node *other;
1414 : bool varonleft;
1415 : Datum constval;
1416 : Oid consttype;
1417 : double selec;
1418 :
1419 : /*
1420 : * If expression is not variable op something or something op variable,
1421 : * then punt and return a default estimate.
1422 : */
1423 44204 : if (!get_restriction_variable(root, args, varRelid,
1424 : &vardata, &other, &varonleft))
1425 624 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1426 :
1427 : /*
1428 : * Can't do anything useful if the something is not a constant, either.
1429 : */
1430 43580 : if (!IsA(other, Const))
1431 : {
1432 2686 : ReleaseVariableStats(vardata);
1433 2686 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1434 : }
1435 :
1436 : /*
1437 : * If the constant is NULL, assume operator is strict and return zero, ie,
1438 : * operator will never return TRUE.
1439 : */
1440 40894 : if (((Const *) other)->constisnull)
1441 : {
1442 66 : ReleaseVariableStats(vardata);
1443 66 : PG_RETURN_FLOAT8(0.0);
1444 : }
1445 40828 : constval = ((Const *) other)->constvalue;
1446 40828 : consttype = ((Const *) other)->consttype;
1447 :
1448 : /*
1449 : * Force the var to be on the left to simplify logic in scalarineqsel.
1450 : */
1451 40828 : if (!varonleft)
1452 : {
1453 366 : operator = get_commutator(operator);
1454 366 : if (!operator)
1455 : {
1456 : /* Use default selectivity (should we raise an error instead?) */
1457 0 : ReleaseVariableStats(vardata);
1458 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1459 : }
1460 366 : isgt = !isgt;
1461 : }
1462 :
1463 : /* The rest of the work is done by scalarineqsel(). */
1464 40828 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1465 : &vardata, constval, consttype);
1466 :
1467 40828 : ReleaseVariableStats(vardata);
1468 :
1469 40828 : PG_RETURN_FLOAT8((float8) selec);
1470 : }
1471 :
1472 : /*
1473 : * scalarltsel - Selectivity of "<" for scalars.
1474 : */
1475 : Datum
1476 14308 : scalarltsel(PG_FUNCTION_ARGS)
1477 : {
1478 14308 : return scalarineqsel_wrapper(fcinfo, false, false);
1479 : }
1480 :
1481 : /*
1482 : * scalarlesel - Selectivity of "<=" for scalars.
1483 : */
1484 : Datum
1485 4440 : scalarlesel(PG_FUNCTION_ARGS)
1486 : {
1487 4440 : return scalarineqsel_wrapper(fcinfo, false, true);
1488 : }
1489 :
1490 : /*
1491 : * scalargtsel - Selectivity of ">" for scalars.
1492 : */
1493 : Datum
1494 15520 : scalargtsel(PG_FUNCTION_ARGS)
1495 : {
1496 15520 : return scalarineqsel_wrapper(fcinfo, true, false);
1497 : }
1498 :
1499 : /*
1500 : * scalargesel - Selectivity of ">=" for scalars.
1501 : */
1502 : Datum
1503 9936 : scalargesel(PG_FUNCTION_ARGS)
1504 : {
1505 9936 : return scalarineqsel_wrapper(fcinfo, true, true);
1506 : }
1507 :
1508 : /*
1509 : * boolvarsel - Selectivity of Boolean variable.
1510 : *
1511 : * This can actually be called on any boolean-valued expression. If it
1512 : * involves only Vars of the specified relation, and if there are statistics
1513 : * about the Var or expression (the latter is possible if it's indexed) then
1514 : * we'll produce a real estimate; otherwise it's just a default.
1515 : */
1516 : Selectivity
1517 45210 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1518 : {
1519 : VariableStatData vardata;
1520 : double selec;
1521 :
1522 45210 : examine_variable(root, arg, varRelid, &vardata);
1523 45210 : if (HeapTupleIsValid(vardata.statsTuple))
1524 : {
1525 : /*
1526 : * A boolean variable V is equivalent to the clause V = 't', so we
1527 : * compute the selectivity as if that is what we have.
1528 : */
1529 36524 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1530 : BoolGetDatum(true), false, true, false);
1531 : }
1532 : else
1533 : {
1534 : /* Otherwise, the default estimate is 0.5 */
1535 8686 : selec = 0.5;
1536 : }
1537 45210 : ReleaseVariableStats(vardata);
1538 45210 : return selec;
1539 : }
1540 :
1541 : /*
1542 : * booltestsel - Selectivity of BooleanTest Node.
1543 : */
1544 : Selectivity
1545 872 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1546 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1547 : {
1548 : VariableStatData vardata;
1549 : double selec;
1550 :
1551 872 : examine_variable(root, arg, varRelid, &vardata);
1552 :
1553 872 : if (HeapTupleIsValid(vardata.statsTuple))
1554 : {
1555 : Form_pg_statistic stats;
1556 : double freq_null;
1557 : AttStatsSlot sslot;
1558 :
1559 0 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1560 0 : freq_null = stats->stanullfrac;
1561 :
1562 0 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1563 : STATISTIC_KIND_MCV, InvalidOid,
1564 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1565 0 : && sslot.nnumbers > 0)
1566 0 : {
1567 : double freq_true;
1568 : double freq_false;
1569 :
1570 : /*
1571 : * Get first MCV frequency and derive frequency for true.
1572 : */
1573 0 : if (DatumGetBool(sslot.values[0]))
1574 0 : freq_true = sslot.numbers[0];
1575 : else
1576 0 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1577 :
1578 : /*
1579 : * Next derive frequency for false. Then use these as appropriate
1580 : * to derive frequency for each case.
1581 : */
1582 0 : freq_false = 1.0 - freq_true - freq_null;
1583 :
1584 0 : switch (booltesttype)
1585 : {
1586 0 : case IS_UNKNOWN:
1587 : /* select only NULL values */
1588 0 : selec = freq_null;
1589 0 : break;
1590 0 : case IS_NOT_UNKNOWN:
1591 : /* select non-NULL values */
1592 0 : selec = 1.0 - freq_null;
1593 0 : break;
1594 0 : case IS_TRUE:
1595 : /* select only TRUE values */
1596 0 : selec = freq_true;
1597 0 : break;
1598 0 : case IS_NOT_TRUE:
1599 : /* select non-TRUE values */
1600 0 : selec = 1.0 - freq_true;
1601 0 : break;
1602 0 : case IS_FALSE:
1603 : /* select only FALSE values */
1604 0 : selec = freq_false;
1605 0 : break;
1606 0 : case IS_NOT_FALSE:
1607 : /* select non-FALSE values */
1608 0 : selec = 1.0 - freq_false;
1609 0 : break;
1610 0 : default:
1611 0 : elog(ERROR, "unrecognized booltesttype: %d",
1612 : (int) booltesttype);
1613 : selec = 0.0; /* Keep compiler quiet */
1614 : break;
1615 : }
1616 :
1617 0 : free_attstatsslot(&sslot);
1618 : }
1619 : else
1620 : {
1621 : /*
1622 : * No most-common-value info available. Still have null fraction
1623 : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1624 : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1625 : */
1626 0 : switch (booltesttype)
1627 : {
1628 0 : case IS_UNKNOWN:
1629 : /* select only NULL values */
1630 0 : selec = freq_null;
1631 0 : break;
1632 0 : case IS_NOT_UNKNOWN:
1633 : /* select non-NULL values */
1634 0 : selec = 1.0 - freq_null;
1635 0 : break;
1636 0 : case IS_TRUE:
1637 : case IS_FALSE:
1638 : /* Assume we select half of the non-NULL values */
1639 0 : selec = (1.0 - freq_null) / 2.0;
1640 0 : break;
1641 0 : case IS_NOT_TRUE:
1642 : case IS_NOT_FALSE:
1643 : /* Assume we select NULLs plus half of the non-NULLs */
1644 : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1645 0 : selec = (freq_null + 1.0) / 2.0;
1646 0 : break;
1647 0 : default:
1648 0 : elog(ERROR, "unrecognized booltesttype: %d",
1649 : (int) booltesttype);
1650 : selec = 0.0; /* Keep compiler quiet */
1651 : break;
1652 : }
1653 : }
1654 : }
1655 : else
1656 : {
1657 : /*
1658 : * If we can't get variable statistics for the argument, perhaps
1659 : * clause_selectivity can do something with it. We ignore the
1660 : * possibility of a NULL value when using clause_selectivity, and just
1661 : * assume the value is either TRUE or FALSE.
1662 : */
1663 872 : switch (booltesttype)
1664 : {
1665 48 : case IS_UNKNOWN:
1666 48 : selec = DEFAULT_UNK_SEL;
1667 48 : break;
1668 108 : case IS_NOT_UNKNOWN:
1669 108 : selec = DEFAULT_NOT_UNK_SEL;
1670 108 : break;
1671 252 : case IS_TRUE:
1672 : case IS_NOT_FALSE:
1673 252 : selec = (double) clause_selectivity(root, arg,
1674 : varRelid,
1675 : jointype, sjinfo);
1676 252 : break;
1677 464 : case IS_FALSE:
1678 : case IS_NOT_TRUE:
1679 464 : selec = 1.0 - (double) clause_selectivity(root, arg,
1680 : varRelid,
1681 : jointype, sjinfo);
1682 464 : break;
1683 0 : default:
1684 0 : elog(ERROR, "unrecognized booltesttype: %d",
1685 : (int) booltesttype);
1686 : selec = 0.0; /* Keep compiler quiet */
1687 : break;
1688 : }
1689 : }
1690 :
1691 872 : ReleaseVariableStats(vardata);
1692 :
1693 : /* result should be in range, but make sure... */
1694 872 : CLAMP_PROBABILITY(selec);
1695 :
1696 872 : return (Selectivity) selec;
1697 : }
1698 :
1699 : /*
1700 : * nulltestsel - Selectivity of NullTest Node.
1701 : */
1702 : Selectivity
1703 17746 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1704 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1705 : {
1706 : VariableStatData vardata;
1707 : double selec;
1708 :
1709 17746 : examine_variable(root, arg, varRelid, &vardata);
1710 :
1711 17746 : if (HeapTupleIsValid(vardata.statsTuple))
1712 : {
1713 : Form_pg_statistic stats;
1714 : double freq_null;
1715 :
1716 10136 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1717 10136 : freq_null = stats->stanullfrac;
1718 :
1719 10136 : switch (nulltesttype)
1720 : {
1721 7258 : case IS_NULL:
1722 :
1723 : /*
1724 : * Use freq_null directly.
1725 : */
1726 7258 : selec = freq_null;
1727 7258 : break;
1728 2878 : case IS_NOT_NULL:
1729 :
1730 : /*
1731 : * Select not unknown (not null) values. Calculate from
1732 : * freq_null.
1733 : */
1734 2878 : selec = 1.0 - freq_null;
1735 2878 : break;
1736 0 : default:
1737 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1738 : (int) nulltesttype);
1739 : return (Selectivity) 0; /* keep compiler quiet */
1740 : }
1741 : }
1742 7610 : else if (vardata.var && IsA(vardata.var, Var) &&
1743 6894 : ((Var *) vardata.var)->varattno < 0)
1744 : {
1745 : /*
1746 : * There are no stats for system columns, but we know they are never
1747 : * NULL.
1748 : */
1749 60 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1750 : }
1751 : else
1752 : {
1753 : /*
1754 : * No ANALYZE stats available, so make a guess
1755 : */
1756 7550 : switch (nulltesttype)
1757 : {
1758 2124 : case IS_NULL:
1759 2124 : selec = DEFAULT_UNK_SEL;
1760 2124 : break;
1761 5426 : case IS_NOT_NULL:
1762 5426 : selec = DEFAULT_NOT_UNK_SEL;
1763 5426 : break;
1764 0 : default:
1765 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1766 : (int) nulltesttype);
1767 : return (Selectivity) 0; /* keep compiler quiet */
1768 : }
1769 : }
1770 :
1771 17746 : ReleaseVariableStats(vardata);
1772 :
1773 : /* result should be in range, but make sure... */
1774 17746 : CLAMP_PROBABILITY(selec);
1775 :
1776 17746 : return (Selectivity) selec;
1777 : }
1778 :
1779 : /*
1780 : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1781 : *
1782 : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1783 : * but it seems possible that RelabelType might show up. Also, the planner
1784 : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1785 : * so we need to be ready to deal with more than one level.
1786 : */
1787 : static Node *
1788 136746 : strip_array_coercion(Node *node)
1789 : {
1790 : for (;;)
1791 : {
1792 136746 : if (node && IsA(node, ArrayCoerceExpr))
1793 66 : {
1794 3246 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1795 :
1796 : /*
1797 : * If the per-element expression is just a RelabelType on top of
1798 : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1799 : */
1800 3246 : if (IsA(acoerce->elemexpr, RelabelType) &&
1801 66 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1802 66 : node = (Node *) acoerce->arg;
1803 : else
1804 : break;
1805 : }
1806 133500 : else if (node && IsA(node, RelabelType))
1807 : {
1808 : /* We don't really expect this case, but may as well cope */
1809 0 : node = (Node *) ((RelabelType *) node)->arg;
1810 : }
1811 : else
1812 : break;
1813 : }
1814 136680 : return node;
1815 : }
1816 :
1817 : /*
1818 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1819 : */
1820 : Selectivity
1821 24232 : scalararraysel(PlannerInfo *root,
1822 : ScalarArrayOpExpr *clause,
1823 : bool is_join_clause,
1824 : int varRelid,
1825 : JoinType jointype,
1826 : SpecialJoinInfo *sjinfo)
1827 : {
1828 24232 : Oid operator = clause->opno;
1829 24232 : bool useOr = clause->useOr;
1830 24232 : bool isEquality = false;
1831 24232 : bool isInequality = false;
1832 : Node *leftop;
1833 : Node *rightop;
1834 : Oid nominal_element_type;
1835 : Oid nominal_element_collation;
1836 : TypeCacheEntry *typentry;
1837 : RegProcedure oprsel;
1838 : FmgrInfo oprselproc;
1839 : Selectivity s1;
1840 : Selectivity s1disjoint;
1841 :
1842 : /* First, deconstruct the expression */
1843 : Assert(list_length(clause->args) == 2);
1844 24232 : leftop = (Node *) linitial(clause->args);
1845 24232 : rightop = (Node *) lsecond(clause->args);
1846 :
1847 : /* aggressively reduce both sides to constants */
1848 24232 : leftop = estimate_expression_value(root, leftop);
1849 24232 : rightop = estimate_expression_value(root, rightop);
1850 :
1851 : /* get nominal (after relabeling) element type of rightop */
1852 24232 : nominal_element_type = get_base_element_type(exprType(rightop));
1853 24232 : if (!OidIsValid(nominal_element_type))
1854 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1855 : /* get nominal collation, too, for generating constants */
1856 24232 : nominal_element_collation = exprCollation(rightop);
1857 :
1858 : /* look through any binary-compatible relabeling of rightop */
1859 24232 : rightop = strip_array_coercion(rightop);
1860 :
1861 : /*
1862 : * Detect whether the operator is the default equality or inequality
1863 : * operator of the array element type.
1864 : */
1865 24232 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1866 24232 : if (OidIsValid(typentry->eq_opr))
1867 : {
1868 24230 : if (operator == typentry->eq_opr)
1869 18226 : isEquality = true;
1870 6004 : else if (get_negator(operator) == typentry->eq_opr)
1871 5440 : isInequality = true;
1872 : }
1873 :
1874 : /*
1875 : * If it is equality or inequality, we might be able to estimate this as a
1876 : * form of array containment; for instance "const = ANY(column)" can be
1877 : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1878 : * that, and returns the selectivity estimate if successful, or -1 if not.
1879 : */
1880 24232 : if ((isEquality || isInequality) && !is_join_clause)
1881 : {
1882 23666 : s1 = scalararraysel_containment(root, leftop, rightop,
1883 : nominal_element_type,
1884 : isEquality, useOr, varRelid);
1885 23666 : if (s1 >= 0.0)
1886 116 : return s1;
1887 : }
1888 :
1889 : /*
1890 : * Look up the underlying operator's selectivity estimator. Punt if it
1891 : * hasn't got one.
1892 : */
1893 24116 : if (is_join_clause)
1894 0 : oprsel = get_oprjoin(operator);
1895 : else
1896 24116 : oprsel = get_oprrest(operator);
1897 24116 : if (!oprsel)
1898 2 : return (Selectivity) 0.5;
1899 24114 : fmgr_info(oprsel, &oprselproc);
1900 :
1901 : /*
1902 : * In the array-containment check above, we must only believe that an
1903 : * operator is equality or inequality if it is the default btree equality
1904 : * operator (or its negator) for the element type, since those are the
1905 : * operators that array containment will use. But in what follows, we can
1906 : * be a little laxer, and also believe that any operators using eqsel() or
1907 : * neqsel() as selectivity estimator act like equality or inequality.
1908 : */
1909 24114 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1910 18304 : isEquality = true;
1911 5810 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1912 5330 : isInequality = true;
1913 :
1914 : /*
1915 : * We consider three cases:
1916 : *
1917 : * 1. rightop is an Array constant: deconstruct the array, apply the
1918 : * operator's selectivity function for each array element, and merge the
1919 : * results in the same way that clausesel.c does for AND/OR combinations.
1920 : *
1921 : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
1922 : * function for each element of the ARRAY[] construct, and merge.
1923 : *
1924 : * 3. otherwise, make a guess ...
1925 : */
1926 24114 : if (rightop && IsA(rightop, Const))
1927 19724 : {
1928 19754 : Datum arraydatum = ((Const *) rightop)->constvalue;
1929 19754 : bool arrayisnull = ((Const *) rightop)->constisnull;
1930 : ArrayType *arrayval;
1931 : int16 elmlen;
1932 : bool elmbyval;
1933 : char elmalign;
1934 : int num_elems;
1935 : Datum *elem_values;
1936 : bool *elem_nulls;
1937 : int i;
1938 :
1939 19754 : if (arrayisnull) /* qual can't succeed if null array */
1940 30 : return (Selectivity) 0.0;
1941 19724 : arrayval = DatumGetArrayTypeP(arraydatum);
1942 19724 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
1943 : &elmlen, &elmbyval, &elmalign);
1944 19724 : deconstruct_array(arrayval,
1945 : ARR_ELEMTYPE(arrayval),
1946 : elmlen, elmbyval, elmalign,
1947 : &elem_values, &elem_nulls, &num_elems);
1948 :
1949 : /*
1950 : * For generic operators, we assume the probability of success is
1951 : * independent for each array element. But for "= ANY" or "<> ALL",
1952 : * if the array elements are distinct (which'd typically be the case)
1953 : * then the probabilities are disjoint, and we should just sum them.
1954 : *
1955 : * If we were being really tense we would try to confirm that the
1956 : * elements are all distinct, but that would be expensive and it
1957 : * doesn't seem to be worth the cycles; it would amount to penalizing
1958 : * well-written queries in favor of poorly-written ones. However, we
1959 : * do protect ourselves a little bit by checking whether the
1960 : * disjointness assumption leads to an impossible (out of range)
1961 : * probability; if so, we fall back to the normal calculation.
1962 : */
1963 19724 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
1964 :
1965 86772 : for (i = 0; i < num_elems; i++)
1966 : {
1967 : List *args;
1968 : Selectivity s2;
1969 :
1970 67048 : args = list_make2(leftop,
1971 : makeConst(nominal_element_type,
1972 : -1,
1973 : nominal_element_collation,
1974 : elmlen,
1975 : elem_values[i],
1976 : elem_nulls[i],
1977 : elmbyval));
1978 67048 : if (is_join_clause)
1979 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1980 : clause->inputcollid,
1981 : PointerGetDatum(root),
1982 : ObjectIdGetDatum(operator),
1983 : PointerGetDatum(args),
1984 : Int16GetDatum(jointype),
1985 : PointerGetDatum(sjinfo)));
1986 : else
1987 67048 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1988 : clause->inputcollid,
1989 : PointerGetDatum(root),
1990 : ObjectIdGetDatum(operator),
1991 : PointerGetDatum(args),
1992 : Int32GetDatum(varRelid)));
1993 :
1994 67048 : if (useOr)
1995 : {
1996 53202 : s1 = s1 + s2 - s1 * s2;
1997 53202 : if (isEquality)
1998 52158 : s1disjoint += s2;
1999 : }
2000 : else
2001 : {
2002 13846 : s1 = s1 * s2;
2003 13846 : if (isInequality)
2004 13534 : s1disjoint += s2 - 1.0;
2005 : }
2006 : }
2007 :
2008 : /* accept disjoint-probability estimate if in range */
2009 19724 : if ((useOr ? isEquality : isInequality) &&
2010 19046 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2011 19016 : s1 = s1disjoint;
2012 : }
2013 4360 : else if (rightop && IsA(rightop, ArrayExpr) &&
2014 292 : !((ArrayExpr *) rightop)->multidims)
2015 292 : {
2016 292 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2017 : int16 elmlen;
2018 : bool elmbyval;
2019 : ListCell *l;
2020 :
2021 292 : get_typlenbyval(arrayexpr->element_typeid,
2022 : &elmlen, &elmbyval);
2023 :
2024 : /*
2025 : * We use the assumption of disjoint probabilities here too, although
2026 : * the odds of equal array elements are rather higher if the elements
2027 : * are not all constants (which they won't be, else constant folding
2028 : * would have reduced the ArrayExpr to a Const). In this path it's
2029 : * critical to have the sanity check on the s1disjoint estimate.
2030 : */
2031 292 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2032 :
2033 1038 : foreach(l, arrayexpr->elements)
2034 : {
2035 746 : Node *elem = (Node *) lfirst(l);
2036 : List *args;
2037 : Selectivity s2;
2038 :
2039 : /*
2040 : * Theoretically, if elem isn't of nominal_element_type we should
2041 : * insert a RelabelType, but it seems unlikely that any operator
2042 : * estimation function would really care ...
2043 : */
2044 746 : args = list_make2(leftop, elem);
2045 746 : if (is_join_clause)
2046 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2047 : clause->inputcollid,
2048 : PointerGetDatum(root),
2049 : ObjectIdGetDatum(operator),
2050 : PointerGetDatum(args),
2051 : Int16GetDatum(jointype),
2052 : PointerGetDatum(sjinfo)));
2053 : else
2054 746 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2055 : clause->inputcollid,
2056 : PointerGetDatum(root),
2057 : ObjectIdGetDatum(operator),
2058 : PointerGetDatum(args),
2059 : Int32GetDatum(varRelid)));
2060 :
2061 746 : if (useOr)
2062 : {
2063 746 : s1 = s1 + s2 - s1 * s2;
2064 746 : if (isEquality)
2065 746 : s1disjoint += s2;
2066 : }
2067 : else
2068 : {
2069 0 : s1 = s1 * s2;
2070 0 : if (isInequality)
2071 0 : s1disjoint += s2 - 1.0;
2072 : }
2073 : }
2074 :
2075 : /* accept disjoint-probability estimate if in range */
2076 292 : if ((useOr ? isEquality : isInequality) &&
2077 292 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2078 292 : s1 = s1disjoint;
2079 : }
2080 : else
2081 : {
2082 : CaseTestExpr *dummyexpr;
2083 : List *args;
2084 : Selectivity s2;
2085 : int i;
2086 :
2087 : /*
2088 : * We need a dummy rightop to pass to the operator selectivity
2089 : * routine. It can be pretty much anything that doesn't look like a
2090 : * constant; CaseTestExpr is a convenient choice.
2091 : */
2092 4068 : dummyexpr = makeNode(CaseTestExpr);
2093 4068 : dummyexpr->typeId = nominal_element_type;
2094 4068 : dummyexpr->typeMod = -1;
2095 4068 : dummyexpr->collation = clause->inputcollid;
2096 4068 : args = list_make2(leftop, dummyexpr);
2097 4068 : if (is_join_clause)
2098 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2099 : clause->inputcollid,
2100 : PointerGetDatum(root),
2101 : ObjectIdGetDatum(operator),
2102 : PointerGetDatum(args),
2103 : Int16GetDatum(jointype),
2104 : PointerGetDatum(sjinfo)));
2105 : else
2106 4068 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2107 : clause->inputcollid,
2108 : PointerGetDatum(root),
2109 : ObjectIdGetDatum(operator),
2110 : PointerGetDatum(args),
2111 : Int32GetDatum(varRelid)));
2112 4068 : s1 = useOr ? 0.0 : 1.0;
2113 :
2114 : /*
2115 : * Arbitrarily assume 10 elements in the eventual array value (see
2116 : * also estimate_array_length). We don't risk an assumption of
2117 : * disjoint probabilities here.
2118 : */
2119 44748 : for (i = 0; i < 10; i++)
2120 : {
2121 40680 : if (useOr)
2122 40680 : s1 = s1 + s2 - s1 * s2;
2123 : else
2124 0 : s1 = s1 * s2;
2125 : }
2126 : }
2127 :
2128 : /* result should be in range, but make sure... */
2129 24084 : CLAMP_PROBABILITY(s1);
2130 :
2131 24084 : return s1;
2132 : }
2133 :
2134 : /*
2135 : * Estimate number of elements in the array yielded by an expression.
2136 : *
2137 : * Note: the result is integral, but we use "double" to avoid overflow
2138 : * concerns. Most callers will use it in double-type expressions anyway.
2139 : *
2140 : * Note: in some code paths root can be passed as NULL, resulting in
2141 : * slightly worse estimates.
2142 : */
2143 : double
2144 112448 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2145 : {
2146 : /* look through any binary-compatible relabeling of arrayexpr */
2147 112448 : arrayexpr = strip_array_coercion(arrayexpr);
2148 :
2149 112448 : if (arrayexpr && IsA(arrayexpr, Const))
2150 : {
2151 52612 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2152 52612 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2153 : ArrayType *arrayval;
2154 :
2155 52612 : if (arrayisnull)
2156 90 : return 0;
2157 52522 : arrayval = DatumGetArrayTypeP(arraydatum);
2158 52522 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2159 : }
2160 59836 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2161 494 : !((ArrayExpr *) arrayexpr)->multidims)
2162 : {
2163 494 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2164 : }
2165 59342 : else if (arrayexpr && root)
2166 : {
2167 : /* See if we can find any statistics about it */
2168 : VariableStatData vardata;
2169 : AttStatsSlot sslot;
2170 59342 : double nelem = 0;
2171 :
2172 59342 : examine_variable(root, arrayexpr, 0, &vardata);
2173 59342 : if (HeapTupleIsValid(vardata.statsTuple))
2174 : {
2175 : /*
2176 : * Found stats, so use the average element count, which is stored
2177 : * in the last stanumbers element of the DECHIST statistics.
2178 : * Actually that is the average count of *distinct* elements;
2179 : * perhaps we should scale it up somewhat?
2180 : */
2181 8200 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2182 : STATISTIC_KIND_DECHIST, InvalidOid,
2183 : ATTSTATSSLOT_NUMBERS))
2184 : {
2185 8086 : if (sslot.nnumbers > 0)
2186 8086 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2187 8086 : free_attstatsslot(&sslot);
2188 : }
2189 : }
2190 59342 : ReleaseVariableStats(vardata);
2191 :
2192 59342 : if (nelem > 0)
2193 8086 : return nelem;
2194 : }
2195 :
2196 : /* Else use a default guess --- this should match scalararraysel */
2197 51256 : return 10;
2198 : }
2199 :
2200 : /*
2201 : * rowcomparesel - Selectivity of RowCompareExpr Node.
2202 : *
2203 : * We estimate RowCompare selectivity by considering just the first (high
2204 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2205 : * this estimate could be refined by considering additional columns, it
2206 : * seems unlikely that we could do a lot better without multi-column
2207 : * statistics.
2208 : */
2209 : Selectivity
2210 216 : rowcomparesel(PlannerInfo *root,
2211 : RowCompareExpr *clause,
2212 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2213 : {
2214 : Selectivity s1;
2215 216 : Oid opno = linitial_oid(clause->opnos);
2216 216 : Oid inputcollid = linitial_oid(clause->inputcollids);
2217 : List *opargs;
2218 : bool is_join_clause;
2219 :
2220 : /* Build equivalent arg list for single operator */
2221 216 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2222 :
2223 : /*
2224 : * Decide if it's a join clause. This should match clausesel.c's
2225 : * treat_as_join_clause(), except that we intentionally consider only the
2226 : * leading columns and not the rest of the clause.
2227 : */
2228 216 : if (varRelid != 0)
2229 : {
2230 : /*
2231 : * Caller is forcing restriction mode (eg, because we are examining an
2232 : * inner indexscan qual).
2233 : */
2234 54 : is_join_clause = false;
2235 : }
2236 162 : else if (sjinfo == NULL)
2237 : {
2238 : /*
2239 : * It must be a restriction clause, since it's being evaluated at a
2240 : * scan node.
2241 : */
2242 150 : is_join_clause = false;
2243 : }
2244 : else
2245 : {
2246 : /*
2247 : * Otherwise, it's a join if there's more than one base relation used.
2248 : */
2249 12 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2250 : }
2251 :
2252 216 : if (is_join_clause)
2253 : {
2254 : /* Estimate selectivity for a join clause. */
2255 12 : s1 = join_selectivity(root, opno,
2256 : opargs,
2257 : inputcollid,
2258 : jointype,
2259 : sjinfo);
2260 : }
2261 : else
2262 : {
2263 : /* Estimate selectivity for a restriction clause. */
2264 204 : s1 = restriction_selectivity(root, opno,
2265 : opargs,
2266 : inputcollid,
2267 : varRelid);
2268 : }
2269 :
2270 216 : return s1;
2271 : }
2272 :
2273 : /*
2274 : * eqjoinsel - Join selectivity of "="
2275 : */
2276 : Datum
2277 232882 : eqjoinsel(PG_FUNCTION_ARGS)
2278 : {
2279 232882 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2280 232882 : Oid operator = PG_GETARG_OID(1);
2281 232882 : List *args = (List *) PG_GETARG_POINTER(2);
2282 :
2283 : #ifdef NOT_USED
2284 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2285 : #endif
2286 232882 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2287 232882 : Oid collation = PG_GET_COLLATION();
2288 : double selec;
2289 : double selec_inner;
2290 : VariableStatData vardata1;
2291 : VariableStatData vardata2;
2292 : double nd1;
2293 : double nd2;
2294 : bool isdefault1;
2295 : bool isdefault2;
2296 : Oid opfuncoid;
2297 : AttStatsSlot sslot1;
2298 : AttStatsSlot sslot2;
2299 232882 : Form_pg_statistic stats1 = NULL;
2300 232882 : Form_pg_statistic stats2 = NULL;
2301 232882 : bool have_mcvs1 = false;
2302 232882 : bool have_mcvs2 = false;
2303 : bool get_mcv_stats;
2304 : bool join_is_reversed;
2305 : RelOptInfo *inner_rel;
2306 :
2307 232882 : get_join_variables(root, args, sjinfo,
2308 : &vardata1, &vardata2, &join_is_reversed);
2309 :
2310 232882 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2311 232882 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2312 :
2313 232882 : opfuncoid = get_opcode(operator);
2314 :
2315 232882 : memset(&sslot1, 0, sizeof(sslot1));
2316 232882 : memset(&sslot2, 0, sizeof(sslot2));
2317 :
2318 : /*
2319 : * There is no use in fetching one side's MCVs if we lack MCVs for the
2320 : * other side, so do a quick check to verify that both stats exist.
2321 : */
2322 650996 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2323 325832 : HeapTupleIsValid(vardata2.statsTuple) &&
2324 140600 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2325 : STATISTIC_KIND_MCV, InvalidOid,
2326 418114 : 0) &&
2327 57514 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2328 : STATISTIC_KIND_MCV, InvalidOid,
2329 : 0));
2330 :
2331 232882 : if (HeapTupleIsValid(vardata1.statsTuple))
2332 : {
2333 : /* note we allow use of nullfrac regardless of security check */
2334 185232 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2335 200400 : if (get_mcv_stats &&
2336 15168 : statistic_proc_security_check(&vardata1, opfuncoid))
2337 15168 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2338 : STATISTIC_KIND_MCV, InvalidOid,
2339 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2340 : }
2341 :
2342 232882 : if (HeapTupleIsValid(vardata2.statsTuple))
2343 : {
2344 : /* note we allow use of nullfrac regardless of security check */
2345 155098 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2346 170266 : if (get_mcv_stats &&
2347 15168 : statistic_proc_security_check(&vardata2, opfuncoid))
2348 15168 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2349 : STATISTIC_KIND_MCV, InvalidOid,
2350 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2351 : }
2352 :
2353 : /* We need to compute the inner-join selectivity in all cases */
2354 232882 : selec_inner = eqjoinsel_inner(opfuncoid, collation,
2355 : &vardata1, &vardata2,
2356 : nd1, nd2,
2357 : isdefault1, isdefault2,
2358 : &sslot1, &sslot2,
2359 : stats1, stats2,
2360 : have_mcvs1, have_mcvs2);
2361 :
2362 232882 : switch (sjinfo->jointype)
2363 : {
2364 222012 : case JOIN_INNER:
2365 : case JOIN_LEFT:
2366 : case JOIN_FULL:
2367 222012 : selec = selec_inner;
2368 222012 : break;
2369 10870 : case JOIN_SEMI:
2370 : case JOIN_ANTI:
2371 :
2372 : /*
2373 : * Look up the join's inner relation. min_righthand is sufficient
2374 : * information because neither SEMI nor ANTI joins permit any
2375 : * reassociation into or out of their RHS, so the righthand will
2376 : * always be exactly that set of rels.
2377 : */
2378 10870 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2379 :
2380 10870 : if (!join_is_reversed)
2381 6620 : selec = eqjoinsel_semi(opfuncoid, collation,
2382 : &vardata1, &vardata2,
2383 : nd1, nd2,
2384 : isdefault1, isdefault2,
2385 : &sslot1, &sslot2,
2386 : stats1, stats2,
2387 : have_mcvs1, have_mcvs2,
2388 : inner_rel);
2389 : else
2390 : {
2391 4250 : Oid commop = get_commutator(operator);
2392 4250 : Oid commopfuncoid = OidIsValid(commop) ? get_opcode(commop) : InvalidOid;
2393 :
2394 4250 : selec = eqjoinsel_semi(commopfuncoid, collation,
2395 : &vardata2, &vardata1,
2396 : nd2, nd1,
2397 : isdefault2, isdefault1,
2398 : &sslot2, &sslot1,
2399 : stats2, stats1,
2400 : have_mcvs2, have_mcvs1,
2401 : inner_rel);
2402 : }
2403 :
2404 : /*
2405 : * We should never estimate the output of a semijoin to be more
2406 : * rows than we estimate for an inner join with the same input
2407 : * rels and join condition; it's obviously impossible for that to
2408 : * happen. The former estimate is N1 * Ssemi while the latter is
2409 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2410 : * this is worthwhile because of the shakier estimation rules we
2411 : * use in eqjoinsel_semi, particularly in cases where it has to
2412 : * punt entirely.
2413 : */
2414 10870 : selec = Min(selec, inner_rel->rows * selec_inner);
2415 10870 : break;
2416 0 : default:
2417 : /* other values not expected here */
2418 0 : elog(ERROR, "unrecognized join type: %d",
2419 : (int) sjinfo->jointype);
2420 : selec = 0; /* keep compiler quiet */
2421 : break;
2422 : }
2423 :
2424 232882 : free_attstatsslot(&sslot1);
2425 232882 : free_attstatsslot(&sslot2);
2426 :
2427 232882 : ReleaseVariableStats(vardata1);
2428 232882 : ReleaseVariableStats(vardata2);
2429 :
2430 232882 : CLAMP_PROBABILITY(selec);
2431 :
2432 232882 : PG_RETURN_FLOAT8((float8) selec);
2433 : }
2434 :
2435 : /*
2436 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2437 : *
2438 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2439 : * that it's worth trying to distinguish them here.
2440 : */
2441 : static double
2442 232882 : eqjoinsel_inner(Oid opfuncoid, Oid collation,
2443 : VariableStatData *vardata1, VariableStatData *vardata2,
2444 : double nd1, double nd2,
2445 : bool isdefault1, bool isdefault2,
2446 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2447 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2448 : bool have_mcvs1, bool have_mcvs2)
2449 : {
2450 : double selec;
2451 :
2452 232882 : if (have_mcvs1 && have_mcvs2)
2453 15168 : {
2454 : /*
2455 : * We have most-common-value lists for both relations. Run through
2456 : * the lists to see which MCVs actually join to each other with the
2457 : * given operator. This allows us to determine the exact join
2458 : * selectivity for the portion of the relations represented by the MCV
2459 : * lists. We still have to estimate for the remaining population, but
2460 : * in a skewed distribution this gives us a big leg up in accuracy.
2461 : * For motivation see the analysis in Y. Ioannidis and S.
2462 : * Christodoulakis, "On the propagation of errors in the size of join
2463 : * results", Technical Report 1018, Computer Science Dept., University
2464 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2465 : */
2466 15168 : LOCAL_FCINFO(fcinfo, 2);
2467 : FmgrInfo eqproc;
2468 : bool *hasmatch1;
2469 : bool *hasmatch2;
2470 15168 : double nullfrac1 = stats1->stanullfrac;
2471 15168 : double nullfrac2 = stats2->stanullfrac;
2472 : double matchprodfreq,
2473 : matchfreq1,
2474 : matchfreq2,
2475 : unmatchfreq1,
2476 : unmatchfreq2,
2477 : otherfreq1,
2478 : otherfreq2,
2479 : totalsel1,
2480 : totalsel2;
2481 : int i,
2482 : nmatches;
2483 :
2484 15168 : fmgr_info(opfuncoid, &eqproc);
2485 :
2486 : /*
2487 : * Save a few cycles by setting up the fcinfo struct just once. Using
2488 : * FunctionCallInvoke directly also avoids failure if the eqproc
2489 : * returns NULL, though really equality functions should never do
2490 : * that.
2491 : */
2492 15168 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2493 : NULL, NULL);
2494 15168 : fcinfo->args[0].isnull = false;
2495 15168 : fcinfo->args[1].isnull = false;
2496 :
2497 15168 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2498 15168 : hasmatch2 = (bool *) palloc0(sslot2->nvalues * sizeof(bool));
2499 :
2500 : /*
2501 : * Note we assume that each MCV will match at most one member of the
2502 : * other MCV list. If the operator isn't really equality, there could
2503 : * be multiple matches --- but we don't look for them, both for speed
2504 : * and because the math wouldn't add up...
2505 : */
2506 15168 : matchprodfreq = 0.0;
2507 15168 : nmatches = 0;
2508 694824 : for (i = 0; i < sslot1->nvalues; i++)
2509 : {
2510 : int j;
2511 :
2512 679656 : fcinfo->args[0].value = sslot1->values[i];
2513 :
2514 26790386 : for (j = 0; j < sslot2->nvalues; j++)
2515 : {
2516 : Datum fresult;
2517 :
2518 26327818 : if (hasmatch2[j])
2519 6528872 : continue;
2520 19798946 : fcinfo->args[1].value = sslot2->values[j];
2521 19798946 : fcinfo->isnull = false;
2522 19798946 : fresult = FunctionCallInvoke(fcinfo);
2523 19798946 : if (!fcinfo->isnull && DatumGetBool(fresult))
2524 : {
2525 217088 : hasmatch1[i] = hasmatch2[j] = true;
2526 217088 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
2527 217088 : nmatches++;
2528 217088 : break;
2529 : }
2530 : }
2531 : }
2532 15168 : CLAMP_PROBABILITY(matchprodfreq);
2533 : /* Sum up frequencies of matched and unmatched MCVs */
2534 15168 : matchfreq1 = unmatchfreq1 = 0.0;
2535 694824 : for (i = 0; i < sslot1->nvalues; i++)
2536 : {
2537 679656 : if (hasmatch1[i])
2538 217088 : matchfreq1 += sslot1->numbers[i];
2539 : else
2540 462568 : unmatchfreq1 += sslot1->numbers[i];
2541 : }
2542 15168 : CLAMP_PROBABILITY(matchfreq1);
2543 15168 : CLAMP_PROBABILITY(unmatchfreq1);
2544 15168 : matchfreq2 = unmatchfreq2 = 0.0;
2545 460286 : for (i = 0; i < sslot2->nvalues; i++)
2546 : {
2547 445118 : if (hasmatch2[i])
2548 217088 : matchfreq2 += sslot2->numbers[i];
2549 : else
2550 228030 : unmatchfreq2 += sslot2->numbers[i];
2551 : }
2552 15168 : CLAMP_PROBABILITY(matchfreq2);
2553 15168 : CLAMP_PROBABILITY(unmatchfreq2);
2554 15168 : pfree(hasmatch1);
2555 15168 : pfree(hasmatch2);
2556 :
2557 : /*
2558 : * Compute total frequency of non-null values that are not in the MCV
2559 : * lists.
2560 : */
2561 15168 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2562 15168 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2563 15168 : CLAMP_PROBABILITY(otherfreq1);
2564 15168 : CLAMP_PROBABILITY(otherfreq2);
2565 :
2566 : /*
2567 : * We can estimate the total selectivity from the point of view of
2568 : * relation 1 as: the known selectivity for matched MCVs, plus
2569 : * unmatched MCVs that are assumed to match against random members of
2570 : * relation 2's non-MCV population, plus non-MCV values that are
2571 : * assumed to match against random members of relation 2's unmatched
2572 : * MCVs plus non-MCV values.
2573 : */
2574 15168 : totalsel1 = matchprodfreq;
2575 15168 : if (nd2 > sslot2->nvalues)
2576 6568 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2577 15168 : if (nd2 > nmatches)
2578 12128 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2579 12128 : (nd2 - nmatches);
2580 : /* Same estimate from the point of view of relation 2. */
2581 15168 : totalsel2 = matchprodfreq;
2582 15168 : if (nd1 > sslot1->nvalues)
2583 8240 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2584 15168 : if (nd1 > nmatches)
2585 10652 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2586 10652 : (nd1 - nmatches);
2587 :
2588 : /*
2589 : * Use the smaller of the two estimates. This can be justified in
2590 : * essentially the same terms as given below for the no-stats case: to
2591 : * a first approximation, we are estimating from the point of view of
2592 : * the relation with smaller nd.
2593 : */
2594 15168 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2595 : }
2596 : else
2597 : {
2598 : /*
2599 : * We do not have MCV lists for both sides. Estimate the join
2600 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2601 : * is plausible if we assume that the join operator is strict and the
2602 : * non-null values are about equally distributed: a given non-null
2603 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2604 : * of rel2, so total join rows are at most
2605 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2606 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2607 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2608 : * with MIN() is an upper bound. Using the MIN() means we estimate
2609 : * from the point of view of the relation with smaller nd (since the
2610 : * larger nd is determining the MIN). It is reasonable to assume that
2611 : * most tuples in this rel will have join partners, so the bound is
2612 : * probably reasonably tight and should be taken as-is.
2613 : *
2614 : * XXX Can we be smarter if we have an MCV list for just one side? It
2615 : * seems that if we assume equal distribution for the other side, we
2616 : * end up with the same answer anyway.
2617 : */
2618 217714 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2619 217714 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2620 :
2621 217714 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2622 217714 : if (nd1 > nd2)
2623 107530 : selec /= nd1;
2624 : else
2625 110184 : selec /= nd2;
2626 : }
2627 :
2628 232882 : return selec;
2629 : }
2630 :
2631 : /*
2632 : * eqjoinsel_semi --- eqjoinsel for semi join
2633 : *
2634 : * (Also used for anti join, which we are supposed to estimate the same way.)
2635 : * Caller has ensured that vardata1 is the LHS variable.
2636 : * Unlike eqjoinsel_inner, we have to cope with opfuncoid being InvalidOid.
2637 : */
2638 : static double
2639 10870 : eqjoinsel_semi(Oid opfuncoid, Oid collation,
2640 : VariableStatData *vardata1, VariableStatData *vardata2,
2641 : double nd1, double nd2,
2642 : bool isdefault1, bool isdefault2,
2643 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2644 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2645 : bool have_mcvs1, bool have_mcvs2,
2646 : RelOptInfo *inner_rel)
2647 : {
2648 : double selec;
2649 :
2650 : /*
2651 : * We clamp nd2 to be not more than what we estimate the inner relation's
2652 : * size to be. This is intuitively somewhat reasonable since obviously
2653 : * there can't be more than that many distinct values coming from the
2654 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2655 : * likewise) is that this is the only pathway by which restriction clauses
2656 : * applied to the inner rel will affect the join result size estimate,
2657 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2658 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2659 : * the selectivity of outer-rel restrictions.
2660 : *
2661 : * We can apply this clamping both with respect to the base relation from
2662 : * which the join variable comes (if there is just one), and to the
2663 : * immediate inner input relation of the current join.
2664 : *
2665 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2666 : * great, maybe, but it didn't come out of nowhere either. This is most
2667 : * helpful when the inner relation is empty and consequently has no stats.
2668 : */
2669 10870 : if (vardata2->rel)
2670 : {
2671 10864 : if (nd2 >= vardata2->rel->rows)
2672 : {
2673 8868 : nd2 = vardata2->rel->rows;
2674 8868 : isdefault2 = false;
2675 : }
2676 : }
2677 10870 : if (nd2 >= inner_rel->rows)
2678 : {
2679 8810 : nd2 = inner_rel->rows;
2680 8810 : isdefault2 = false;
2681 : }
2682 :
2683 10870 : if (have_mcvs1 && have_mcvs2 && OidIsValid(opfuncoid))
2684 516 : {
2685 : /*
2686 : * We have most-common-value lists for both relations. Run through
2687 : * the lists to see which MCVs actually join to each other with the
2688 : * given operator. This allows us to determine the exact join
2689 : * selectivity for the portion of the relations represented by the MCV
2690 : * lists. We still have to estimate for the remaining population, but
2691 : * in a skewed distribution this gives us a big leg up in accuracy.
2692 : */
2693 516 : LOCAL_FCINFO(fcinfo, 2);
2694 : FmgrInfo eqproc;
2695 : bool *hasmatch1;
2696 : bool *hasmatch2;
2697 516 : double nullfrac1 = stats1->stanullfrac;
2698 : double matchfreq1,
2699 : uncertainfrac,
2700 : uncertain;
2701 : int i,
2702 : nmatches,
2703 : clamped_nvalues2;
2704 :
2705 : /*
2706 : * The clamping above could have resulted in nd2 being less than
2707 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2708 : * most common values in the relation will appear in the join input,
2709 : * and so compare to only the first nd2 members of the MCV list. Of
2710 : * course this is frequently wrong, but it's the best bet we can make.
2711 : */
2712 516 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2713 :
2714 516 : fmgr_info(opfuncoid, &eqproc);
2715 :
2716 : /*
2717 : * Save a few cycles by setting up the fcinfo struct just once. Using
2718 : * FunctionCallInvoke directly also avoids failure if the eqproc
2719 : * returns NULL, though really equality functions should never do
2720 : * that.
2721 : */
2722 516 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2723 : NULL, NULL);
2724 516 : fcinfo->args[0].isnull = false;
2725 516 : fcinfo->args[1].isnull = false;
2726 :
2727 516 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2728 516 : hasmatch2 = (bool *) palloc0(clamped_nvalues2 * sizeof(bool));
2729 :
2730 : /*
2731 : * Note we assume that each MCV will match at most one member of the
2732 : * other MCV list. If the operator isn't really equality, there could
2733 : * be multiple matches --- but we don't look for them, both for speed
2734 : * and because the math wouldn't add up...
2735 : */
2736 516 : nmatches = 0;
2737 12076 : for (i = 0; i < sslot1->nvalues; i++)
2738 : {
2739 : int j;
2740 :
2741 11560 : fcinfo->args[0].value = sslot1->values[i];
2742 :
2743 457936 : for (j = 0; j < clamped_nvalues2; j++)
2744 : {
2745 : Datum fresult;
2746 :
2747 456616 : if (hasmatch2[j])
2748 379692 : continue;
2749 76924 : fcinfo->args[1].value = sslot2->values[j];
2750 76924 : fcinfo->isnull = false;
2751 76924 : fresult = FunctionCallInvoke(fcinfo);
2752 76924 : if (!fcinfo->isnull && DatumGetBool(fresult))
2753 : {
2754 10240 : hasmatch1[i] = hasmatch2[j] = true;
2755 10240 : nmatches++;
2756 10240 : break;
2757 : }
2758 : }
2759 : }
2760 : /* Sum up frequencies of matched MCVs */
2761 516 : matchfreq1 = 0.0;
2762 12076 : for (i = 0; i < sslot1->nvalues; i++)
2763 : {
2764 11560 : if (hasmatch1[i])
2765 10240 : matchfreq1 += sslot1->numbers[i];
2766 : }
2767 516 : CLAMP_PROBABILITY(matchfreq1);
2768 516 : pfree(hasmatch1);
2769 516 : pfree(hasmatch2);
2770 :
2771 : /*
2772 : * Now we need to estimate the fraction of relation 1 that has at
2773 : * least one join partner. We know for certain that the matched MCVs
2774 : * do, so that gives us a lower bound, but we're really in the dark
2775 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2776 : * assume all non-null rel1 rows have join partners, else assume for
2777 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2778 : * can discount the known-matched MCVs from the distinct-values counts
2779 : * before doing the division.
2780 : *
2781 : * Crude as the above is, it's completely useless if we don't have
2782 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2783 : * nd2 is default, punt and assume half of the uncertain rows have
2784 : * join partners.
2785 : */
2786 516 : if (!isdefault1 && !isdefault2)
2787 : {
2788 516 : nd1 -= nmatches;
2789 516 : nd2 -= nmatches;
2790 516 : if (nd1 <= nd2 || nd2 < 0)
2791 486 : uncertainfrac = 1.0;
2792 : else
2793 30 : uncertainfrac = nd2 / nd1;
2794 : }
2795 : else
2796 0 : uncertainfrac = 0.5;
2797 516 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2798 516 : CLAMP_PROBABILITY(uncertain);
2799 516 : selec = matchfreq1 + uncertainfrac * uncertain;
2800 : }
2801 : else
2802 : {
2803 : /*
2804 : * Without MCV lists for both sides, we can only use the heuristic
2805 : * about nd1 vs nd2.
2806 : */
2807 10354 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2808 :
2809 10354 : if (!isdefault1 && !isdefault2)
2810 : {
2811 7998 : if (nd1 <= nd2 || nd2 < 0)
2812 4766 : selec = 1.0 - nullfrac1;
2813 : else
2814 3232 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2815 : }
2816 : else
2817 2356 : selec = 0.5 * (1.0 - nullfrac1);
2818 : }
2819 :
2820 10870 : return selec;
2821 : }
2822 :
2823 : /*
2824 : * neqjoinsel - Join selectivity of "!="
2825 : */
2826 : Datum
2827 4424 : neqjoinsel(PG_FUNCTION_ARGS)
2828 : {
2829 4424 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2830 4424 : Oid operator = PG_GETARG_OID(1);
2831 4424 : List *args = (List *) PG_GETARG_POINTER(2);
2832 4424 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2833 4424 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2834 4424 : Oid collation = PG_GET_COLLATION();
2835 : float8 result;
2836 :
2837 4424 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
2838 1452 : {
2839 : /*
2840 : * For semi-joins, if there is more than one distinct value in the RHS
2841 : * relation then every non-null LHS row must find a row to join since
2842 : * it can only be equal to one of them. We'll assume that there is
2843 : * always more than one distinct RHS value for the sake of stability,
2844 : * though in theory we could have special cases for empty RHS
2845 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
2846 : * fraction of LHS that has the same value as the single RHS value).
2847 : *
2848 : * For anti-joins, if we use the same assumption that there is more
2849 : * than one distinct key in the RHS relation, then every non-null LHS
2850 : * row must be suppressed by the anti-join.
2851 : *
2852 : * So either way, the selectivity estimate should be 1 - nullfrac.
2853 : */
2854 : VariableStatData leftvar;
2855 : VariableStatData rightvar;
2856 : bool reversed;
2857 : HeapTuple statsTuple;
2858 : double nullfrac;
2859 :
2860 1452 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
2861 1452 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
2862 1452 : if (HeapTupleIsValid(statsTuple))
2863 1230 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
2864 : else
2865 222 : nullfrac = 0.0;
2866 1452 : ReleaseVariableStats(leftvar);
2867 1452 : ReleaseVariableStats(rightvar);
2868 :
2869 1452 : result = 1.0 - nullfrac;
2870 : }
2871 : else
2872 : {
2873 : /*
2874 : * We want 1 - eqjoinsel() where the equality operator is the one
2875 : * associated with this != operator, that is, its negator.
2876 : */
2877 2972 : Oid eqop = get_negator(operator);
2878 :
2879 2972 : if (eqop)
2880 : {
2881 : result =
2882 2972 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
2883 : collation,
2884 : PointerGetDatum(root),
2885 : ObjectIdGetDatum(eqop),
2886 : PointerGetDatum(args),
2887 : Int16GetDatum(jointype),
2888 : PointerGetDatum(sjinfo)));
2889 : }
2890 : else
2891 : {
2892 : /* Use default selectivity (should we raise an error instead?) */
2893 0 : result = DEFAULT_EQ_SEL;
2894 : }
2895 2972 : result = 1.0 - result;
2896 : }
2897 :
2898 4424 : PG_RETURN_FLOAT8(result);
2899 : }
2900 :
2901 : /*
2902 : * scalarltjoinsel - Join selectivity of "<" for scalars
2903 : */
2904 : Datum
2905 318 : scalarltjoinsel(PG_FUNCTION_ARGS)
2906 : {
2907 318 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2908 : }
2909 :
2910 : /*
2911 : * scalarlejoinsel - Join selectivity of "<=" for scalars
2912 : */
2913 : Datum
2914 276 : scalarlejoinsel(PG_FUNCTION_ARGS)
2915 : {
2916 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2917 : }
2918 :
2919 : /*
2920 : * scalargtjoinsel - Join selectivity of ">" for scalars
2921 : */
2922 : Datum
2923 240 : scalargtjoinsel(PG_FUNCTION_ARGS)
2924 : {
2925 240 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2926 : }
2927 :
2928 : /*
2929 : * scalargejoinsel - Join selectivity of ">=" for scalars
2930 : */
2931 : Datum
2932 184 : scalargejoinsel(PG_FUNCTION_ARGS)
2933 : {
2934 184 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2935 : }
2936 :
2937 :
2938 : /*
2939 : * mergejoinscansel - Scan selectivity of merge join.
2940 : *
2941 : * A merge join will stop as soon as it exhausts either input stream.
2942 : * Therefore, if we can estimate the ranges of both input variables,
2943 : * we can estimate how much of the input will actually be read. This
2944 : * can have a considerable impact on the cost when using indexscans.
2945 : *
2946 : * Also, we can estimate how much of each input has to be read before the
2947 : * first join pair is found, which will affect the join's startup time.
2948 : *
2949 : * clause should be a clause already known to be mergejoinable. opfamily,
2950 : * cmptype, and nulls_first specify the sort ordering being used.
2951 : *
2952 : * The outputs are:
2953 : * *leftstart is set to the fraction of the left-hand variable expected
2954 : * to be scanned before the first join pair is found (0 to 1).
2955 : * *leftend is set to the fraction of the left-hand variable expected
2956 : * to be scanned before the join terminates (0 to 1).
2957 : * *rightstart, *rightend similarly for the right-hand variable.
2958 : */
2959 : void
2960 118526 : mergejoinscansel(PlannerInfo *root, Node *clause,
2961 : Oid opfamily, CompareType cmptype, bool nulls_first,
2962 : Selectivity *leftstart, Selectivity *leftend,
2963 : Selectivity *rightstart, Selectivity *rightend)
2964 : {
2965 : Node *left,
2966 : *right;
2967 : VariableStatData leftvar,
2968 : rightvar;
2969 : Oid opmethod;
2970 : int op_strategy;
2971 : Oid op_lefttype;
2972 : Oid op_righttype;
2973 : Oid opno,
2974 : collation,
2975 : lsortop,
2976 : rsortop,
2977 : lstatop,
2978 : rstatop,
2979 : ltop,
2980 : leop,
2981 : revltop,
2982 : revleop;
2983 : StrategyNumber ltstrat,
2984 : lestrat,
2985 : gtstrat,
2986 : gestrat;
2987 : bool isgt;
2988 : Datum leftmin,
2989 : leftmax,
2990 : rightmin,
2991 : rightmax;
2992 : double selec;
2993 :
2994 : /* Set default results if we can't figure anything out. */
2995 : /* XXX should default "start" fraction be a bit more than 0? */
2996 118526 : *leftstart = *rightstart = 0.0;
2997 118526 : *leftend = *rightend = 1.0;
2998 :
2999 : /* Deconstruct the merge clause */
3000 118526 : if (!is_opclause(clause))
3001 0 : return; /* shouldn't happen */
3002 118526 : opno = ((OpExpr *) clause)->opno;
3003 118526 : collation = ((OpExpr *) clause)->inputcollid;
3004 118526 : left = get_leftop((Expr *) clause);
3005 118526 : right = get_rightop((Expr *) clause);
3006 118526 : if (!right)
3007 0 : return; /* shouldn't happen */
3008 :
3009 : /* Look for stats for the inputs */
3010 118526 : examine_variable(root, left, 0, &leftvar);
3011 118526 : examine_variable(root, right, 0, &rightvar);
3012 :
3013 118526 : opmethod = get_opfamily_method(opfamily);
3014 :
3015 : /* Extract the operator's declared left/right datatypes */
3016 118526 : get_op_opfamily_properties(opno, opfamily, false,
3017 : &op_strategy,
3018 : &op_lefttype,
3019 : &op_righttype);
3020 : Assert(IndexAmTranslateStrategy(op_strategy, opmethod, opfamily, true) == COMPARE_EQ);
3021 :
3022 : /*
3023 : * Look up the various operators we need. If we don't find them all, it
3024 : * probably means the opfamily is broken, but we just fail silently.
3025 : *
3026 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3027 : * operator, regardless of which sort direction we are considering.
3028 : */
3029 118526 : switch (cmptype)
3030 : {
3031 118472 : case COMPARE_LT:
3032 118472 : isgt = false;
3033 118472 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3034 118472 : lestrat = IndexAmTranslateCompareType(COMPARE_LE, opmethod, opfamily, true);
3035 118472 : if (op_lefttype == op_righttype)
3036 : {
3037 : /* easy case */
3038 116968 : ltop = get_opfamily_member(opfamily,
3039 : op_lefttype, op_righttype,
3040 : ltstrat);
3041 116968 : leop = get_opfamily_member(opfamily,
3042 : op_lefttype, op_righttype,
3043 : lestrat);
3044 116968 : lsortop = ltop;
3045 116968 : rsortop = ltop;
3046 116968 : lstatop = lsortop;
3047 116968 : rstatop = rsortop;
3048 116968 : revltop = ltop;
3049 116968 : revleop = leop;
3050 : }
3051 : else
3052 : {
3053 1504 : ltop = get_opfamily_member(opfamily,
3054 : op_lefttype, op_righttype,
3055 : ltstrat);
3056 1504 : leop = get_opfamily_member(opfamily,
3057 : op_lefttype, op_righttype,
3058 : lestrat);
3059 1504 : lsortop = get_opfamily_member(opfamily,
3060 : op_lefttype, op_lefttype,
3061 : ltstrat);
3062 1504 : rsortop = get_opfamily_member(opfamily,
3063 : op_righttype, op_righttype,
3064 : ltstrat);
3065 1504 : lstatop = lsortop;
3066 1504 : rstatop = rsortop;
3067 1504 : revltop = get_opfamily_member(opfamily,
3068 : op_righttype, op_lefttype,
3069 : ltstrat);
3070 1504 : revleop = get_opfamily_member(opfamily,
3071 : op_righttype, op_lefttype,
3072 : lestrat);
3073 : }
3074 118472 : break;
3075 54 : case COMPARE_GT:
3076 : /* descending-order case */
3077 54 : isgt = true;
3078 54 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3079 54 : gtstrat = IndexAmTranslateCompareType(COMPARE_GT, opmethod, opfamily, true);
3080 54 : gestrat = IndexAmTranslateCompareType(COMPARE_GE, opmethod, opfamily, true);
3081 54 : if (op_lefttype == op_righttype)
3082 : {
3083 : /* easy case */
3084 54 : ltop = get_opfamily_member(opfamily,
3085 : op_lefttype, op_righttype,
3086 : gtstrat);
3087 54 : leop = get_opfamily_member(opfamily,
3088 : op_lefttype, op_righttype,
3089 : gestrat);
3090 54 : lsortop = ltop;
3091 54 : rsortop = ltop;
3092 54 : lstatop = get_opfamily_member(opfamily,
3093 : op_lefttype, op_lefttype,
3094 : ltstrat);
3095 54 : rstatop = lstatop;
3096 54 : revltop = ltop;
3097 54 : revleop = leop;
3098 : }
3099 : else
3100 : {
3101 0 : ltop = get_opfamily_member(opfamily,
3102 : op_lefttype, op_righttype,
3103 : gtstrat);
3104 0 : leop = get_opfamily_member(opfamily,
3105 : op_lefttype, op_righttype,
3106 : gestrat);
3107 0 : lsortop = get_opfamily_member(opfamily,
3108 : op_lefttype, op_lefttype,
3109 : gtstrat);
3110 0 : rsortop = get_opfamily_member(opfamily,
3111 : op_righttype, op_righttype,
3112 : gtstrat);
3113 0 : lstatop = get_opfamily_member(opfamily,
3114 : op_lefttype, op_lefttype,
3115 : ltstrat);
3116 0 : rstatop = get_opfamily_member(opfamily,
3117 : op_righttype, op_righttype,
3118 : ltstrat);
3119 0 : revltop = get_opfamily_member(opfamily,
3120 : op_righttype, op_lefttype,
3121 : gtstrat);
3122 0 : revleop = get_opfamily_member(opfamily,
3123 : op_righttype, op_lefttype,
3124 : gestrat);
3125 : }
3126 54 : break;
3127 0 : default:
3128 0 : goto fail; /* shouldn't get here */
3129 : }
3130 :
3131 118526 : if (!OidIsValid(lsortop) ||
3132 118526 : !OidIsValid(rsortop) ||
3133 118526 : !OidIsValid(lstatop) ||
3134 118526 : !OidIsValid(rstatop) ||
3135 118514 : !OidIsValid(ltop) ||
3136 118514 : !OidIsValid(leop) ||
3137 118514 : !OidIsValid(revltop) ||
3138 : !OidIsValid(revleop))
3139 12 : goto fail; /* insufficient info in catalogs */
3140 :
3141 : /* Try to get ranges of both inputs */
3142 118514 : if (!isgt)
3143 : {
3144 118460 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3145 : &leftmin, &leftmax))
3146 25920 : goto fail; /* no range available from stats */
3147 92540 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3148 : &rightmin, &rightmax))
3149 24628 : goto fail; /* no range available from stats */
3150 : }
3151 : else
3152 : {
3153 : /* need to swap the max and min */
3154 54 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3155 : &leftmax, &leftmin))
3156 30 : goto fail; /* no range available from stats */
3157 24 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3158 : &rightmax, &rightmin))
3159 0 : goto fail; /* no range available from stats */
3160 : }
3161 :
3162 : /*
3163 : * Now, the fraction of the left variable that will be scanned is the
3164 : * fraction that's <= the right-side maximum value. But only believe
3165 : * non-default estimates, else stick with our 1.0.
3166 : */
3167 67936 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3168 : rightmax, op_righttype);
3169 67936 : if (selec != DEFAULT_INEQ_SEL)
3170 67930 : *leftend = selec;
3171 :
3172 : /* And similarly for the right variable. */
3173 67936 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3174 : leftmax, op_lefttype);
3175 67936 : if (selec != DEFAULT_INEQ_SEL)
3176 67936 : *rightend = selec;
3177 :
3178 : /*
3179 : * Only one of the two "end" fractions can really be less than 1.0;
3180 : * believe the smaller estimate and reset the other one to exactly 1.0. If
3181 : * we get exactly equal estimates (as can easily happen with self-joins),
3182 : * believe neither.
3183 : */
3184 67936 : if (*leftend > *rightend)
3185 26562 : *leftend = 1.0;
3186 41374 : else if (*leftend < *rightend)
3187 34892 : *rightend = 1.0;
3188 : else
3189 6482 : *leftend = *rightend = 1.0;
3190 :
3191 : /*
3192 : * Also, the fraction of the left variable that will be scanned before the
3193 : * first join pair is found is the fraction that's < the right-side
3194 : * minimum value. But only believe non-default estimates, else stick with
3195 : * our own default.
3196 : */
3197 67936 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3198 : rightmin, op_righttype);
3199 67936 : if (selec != DEFAULT_INEQ_SEL)
3200 67936 : *leftstart = selec;
3201 :
3202 : /* And similarly for the right variable. */
3203 67936 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3204 : leftmin, op_lefttype);
3205 67936 : if (selec != DEFAULT_INEQ_SEL)
3206 67936 : *rightstart = selec;
3207 :
3208 : /*
3209 : * Only one of the two "start" fractions can really be more than zero;
3210 : * believe the larger estimate and reset the other one to exactly 0.0. If
3211 : * we get exactly equal estimates (as can easily happen with self-joins),
3212 : * believe neither.
3213 : */
3214 67936 : if (*leftstart < *rightstart)
3215 22348 : *leftstart = 0.0;
3216 45588 : else if (*leftstart > *rightstart)
3217 26354 : *rightstart = 0.0;
3218 : else
3219 19234 : *leftstart = *rightstart = 0.0;
3220 :
3221 : /*
3222 : * If the sort order is nulls-first, we're going to have to skip over any
3223 : * nulls too. These would not have been counted by scalarineqsel, and we
3224 : * can safely add in this fraction regardless of whether we believe
3225 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3226 : */
3227 67936 : if (nulls_first)
3228 : {
3229 : Form_pg_statistic stats;
3230 :
3231 24 : if (HeapTupleIsValid(leftvar.statsTuple))
3232 : {
3233 24 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3234 24 : *leftstart += stats->stanullfrac;
3235 24 : CLAMP_PROBABILITY(*leftstart);
3236 24 : *leftend += stats->stanullfrac;
3237 24 : CLAMP_PROBABILITY(*leftend);
3238 : }
3239 24 : if (HeapTupleIsValid(rightvar.statsTuple))
3240 : {
3241 24 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3242 24 : *rightstart += stats->stanullfrac;
3243 24 : CLAMP_PROBABILITY(*rightstart);
3244 24 : *rightend += stats->stanullfrac;
3245 24 : CLAMP_PROBABILITY(*rightend);
3246 : }
3247 : }
3248 :
3249 : /* Disbelieve start >= end, just in case that can happen */
3250 67936 : if (*leftstart >= *leftend)
3251 : {
3252 202 : *leftstart = 0.0;
3253 202 : *leftend = 1.0;
3254 : }
3255 67936 : if (*rightstart >= *rightend)
3256 : {
3257 1430 : *rightstart = 0.0;
3258 1430 : *rightend = 1.0;
3259 : }
3260 :
3261 66506 : fail:
3262 118526 : ReleaseVariableStats(leftvar);
3263 118526 : ReleaseVariableStats(rightvar);
3264 : }
3265 :
3266 :
3267 : /*
3268 : * matchingsel -- generic matching-operator selectivity support
3269 : *
3270 : * Use these for any operators that (a) are on data types for which we collect
3271 : * standard statistics, and (b) have behavior for which the default estimate
3272 : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3273 : * operators.
3274 : */
3275 :
3276 : Datum
3277 1106 : matchingsel(PG_FUNCTION_ARGS)
3278 : {
3279 1106 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3280 1106 : Oid operator = PG_GETARG_OID(1);
3281 1106 : List *args = (List *) PG_GETARG_POINTER(2);
3282 1106 : int varRelid = PG_GETARG_INT32(3);
3283 1106 : Oid collation = PG_GET_COLLATION();
3284 : double selec;
3285 :
3286 : /* Use generic restriction selectivity logic. */
3287 1106 : selec = generic_restriction_selectivity(root, operator, collation,
3288 : args, varRelid,
3289 : DEFAULT_MATCHING_SEL);
3290 :
3291 1106 : PG_RETURN_FLOAT8((float8) selec);
3292 : }
3293 :
3294 : Datum
3295 6 : matchingjoinsel(PG_FUNCTION_ARGS)
3296 : {
3297 : /* Just punt, for the moment. */
3298 6 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3299 : }
3300 :
3301 :
3302 : /*
3303 : * Helper routine for estimate_num_groups: add an item to a list of
3304 : * GroupVarInfos, but only if it's not known equal to any of the existing
3305 : * entries.
3306 : */
3307 : typedef struct
3308 : {
3309 : Node *var; /* might be an expression, not just a Var */
3310 : RelOptInfo *rel; /* relation it belongs to */
3311 : double ndistinct; /* # distinct values */
3312 : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3313 : } GroupVarInfo;
3314 :
3315 : static List *
3316 376390 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3317 : Node *var, VariableStatData *vardata)
3318 : {
3319 : GroupVarInfo *varinfo;
3320 : double ndistinct;
3321 : bool isdefault;
3322 : ListCell *lc;
3323 :
3324 376390 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3325 :
3326 : /*
3327 : * The nullingrels bits within the var could cause the same var to be
3328 : * counted multiple times if it's marked with different nullingrels. They
3329 : * could also prevent us from matching the var to the expressions in
3330 : * extended statistics (see estimate_multivariate_ndistinct). So strip
3331 : * them out first.
3332 : */
3333 376390 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3334 :
3335 455140 : foreach(lc, varinfos)
3336 : {
3337 79588 : varinfo = (GroupVarInfo *) lfirst(lc);
3338 :
3339 : /* Drop exact duplicates */
3340 79588 : if (equal(var, varinfo->var))
3341 838 : return varinfos;
3342 :
3343 : /*
3344 : * Drop known-equal vars, but only if they belong to different
3345 : * relations (see comments for estimate_num_groups). We aren't too
3346 : * fussy about the semantics of "equal" here.
3347 : */
3348 83432 : if (vardata->rel != varinfo->rel &&
3349 4562 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3350 : {
3351 120 : if (varinfo->ndistinct <= ndistinct)
3352 : {
3353 : /* Keep older item, forget new one */
3354 120 : return varinfos;
3355 : }
3356 : else
3357 : {
3358 : /* Delete the older item */
3359 0 : varinfos = foreach_delete_current(varinfos, lc);
3360 : }
3361 : }
3362 : }
3363 :
3364 375552 : varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
3365 :
3366 375552 : varinfo->var = var;
3367 375552 : varinfo->rel = vardata->rel;
3368 375552 : varinfo->ndistinct = ndistinct;
3369 375552 : varinfo->isdefault = isdefault;
3370 375552 : varinfos = lappend(varinfos, varinfo);
3371 375552 : return varinfos;
3372 : }
3373 :
3374 : /*
3375 : * estimate_num_groups - Estimate number of groups in a grouped query
3376 : *
3377 : * Given a query having a GROUP BY clause, estimate how many groups there
3378 : * will be --- ie, the number of distinct combinations of the GROUP BY
3379 : * expressions.
3380 : *
3381 : * This routine is also used to estimate the number of rows emitted by
3382 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3383 : * actually, we only use it for DISTINCT when there's no grouping or
3384 : * aggregation ahead of the DISTINCT.)
3385 : *
3386 : * Inputs:
3387 : * root - the query
3388 : * groupExprs - list of expressions being grouped by
3389 : * input_rows - number of rows estimated to arrive at the group/unique
3390 : * filter step
3391 : * pgset - NULL, or a List** pointing to a grouping set to filter the
3392 : * groupExprs against
3393 : *
3394 : * Outputs:
3395 : * estinfo - When passed as non-NULL, the function will set bits in the
3396 : * "flags" field in order to provide callers with additional information
3397 : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3398 : * bit if we used any default values in the estimation.
3399 : *
3400 : * Given the lack of any cross-correlation statistics in the system, it's
3401 : * impossible to do anything really trustworthy with GROUP BY conditions
3402 : * involving multiple Vars. We should however avoid assuming the worst
3403 : * case (all possible cross-product terms actually appear as groups) since
3404 : * very often the grouped-by Vars are highly correlated. Our current approach
3405 : * is as follows:
3406 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3407 : * independently of their content, and are ignored in the subsequent
3408 : * steps. This is mainly because tests like "col IS NULL" break the
3409 : * heuristic used in step 2 especially badly.
3410 : * 2. Reduce the given expressions to a list of unique Vars used. For
3411 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3412 : * It is clearly correct not to count the same Var more than once.
3413 : * It is also reasonable to treat f(x) the same as x: f() cannot
3414 : * increase the number of distinct values (unless it is volatile,
3415 : * which we consider unlikely for grouping), but it probably won't
3416 : * reduce the number of distinct values much either.
3417 : * As a special case, if a GROUP BY expression can be matched to an
3418 : * expressional index for which we have statistics, then we treat the
3419 : * whole expression as though it were just a Var.
3420 : * 3. If the list contains Vars of different relations that are known equal
3421 : * due to equivalence classes, then drop all but one of the Vars from each
3422 : * known-equal set, keeping the one with smallest estimated # of values
3423 : * (since the extra values of the others can't appear in joined rows).
3424 : * Note the reason we only consider Vars of different relations is that
3425 : * if we considered ones of the same rel, we'd be double-counting the
3426 : * restriction selectivity of the equality in the next step.
3427 : * 4. For Vars within a single source rel, we multiply together the numbers
3428 : * of values, clamp to the number of rows in the rel (divided by 10 if
3429 : * more than one Var), and then multiply by a factor based on the
3430 : * selectivity of the restriction clauses for that rel. When there's
3431 : * more than one Var, the initial product is probably too high (it's the
3432 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3433 : * helpful heuristic for not letting the estimate get out of hand. (The
3434 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3435 : * we multiply by to adjust for the restriction selectivity assumes that
3436 : * the restriction clauses are independent of the grouping, which may not
3437 : * be a valid assumption, but it's hard to do better.
3438 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3439 : * rel, and multiply the results together.
3440 : * Note that rels not containing grouped Vars are ignored completely, as are
3441 : * join clauses. Such rels cannot increase the number of groups, and we
3442 : * assume such clauses do not reduce the number either (somewhat bogus,
3443 : * but we don't have the info to do better).
3444 : */
3445 : double
3446 327626 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3447 : List **pgset, EstimationInfo *estinfo)
3448 : {
3449 327626 : List *varinfos = NIL;
3450 327626 : double srf_multiplier = 1.0;
3451 : double numdistinct;
3452 : ListCell *l;
3453 : int i;
3454 :
3455 : /* Zero the estinfo output parameter, if non-NULL */
3456 327626 : if (estinfo != NULL)
3457 292028 : memset(estinfo, 0, sizeof(EstimationInfo));
3458 :
3459 : /*
3460 : * We don't ever want to return an estimate of zero groups, as that tends
3461 : * to lead to division-by-zero and other unpleasantness. The input_rows
3462 : * estimate is usually already at least 1, but clamp it just in case it
3463 : * isn't.
3464 : */
3465 327626 : input_rows = clamp_row_est(input_rows);
3466 :
3467 : /*
3468 : * If no grouping columns, there's exactly one group. (This can't happen
3469 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3470 : * corner cases with set operations.)
3471 : */
3472 327626 : if (groupExprs == NIL || (pgset && *pgset == NIL))
3473 1012 : return 1.0;
3474 :
3475 : /*
3476 : * Count groups derived from boolean grouping expressions. For other
3477 : * expressions, find the unique Vars used, treating an expression as a Var
3478 : * if we can find stats for it. For each one, record the statistical
3479 : * estimate of number of distinct values (total in its table, without
3480 : * regard for filtering).
3481 : */
3482 326614 : numdistinct = 1.0;
3483 :
3484 326614 : i = 0;
3485 701406 : foreach(l, groupExprs)
3486 : {
3487 374840 : Node *groupexpr = (Node *) lfirst(l);
3488 : double this_srf_multiplier;
3489 : VariableStatData vardata;
3490 : List *varshere;
3491 : ListCell *l2;
3492 :
3493 : /* is expression in this grouping set? */
3494 374840 : if (pgset && !list_member_int(*pgset, i++))
3495 319188 : continue;
3496 :
3497 : /*
3498 : * Set-returning functions in grouping columns are a bit problematic.
3499 : * The code below will effectively ignore their SRF nature and come up
3500 : * with a numdistinct estimate as though they were scalar functions.
3501 : * We compensate by scaling up the end result by the largest SRF
3502 : * rowcount estimate. (This will be an overestimate if the SRF
3503 : * produces multiple copies of any output value, but it seems best to
3504 : * assume the SRF's outputs are distinct. In any case, it's probably
3505 : * pointless to worry too much about this without much better
3506 : * estimates for SRF output rowcounts than we have today.)
3507 : */
3508 374052 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3509 374052 : if (srf_multiplier < this_srf_multiplier)
3510 132 : srf_multiplier = this_srf_multiplier;
3511 :
3512 : /* Short-circuit for expressions returning boolean */
3513 374052 : if (exprType(groupexpr) == BOOLOID)
3514 : {
3515 192 : numdistinct *= 2.0;
3516 192 : continue;
3517 : }
3518 :
3519 : /*
3520 : * If examine_variable is able to deduce anything about the GROUP BY
3521 : * expression, treat it as a single variable even if it's really more
3522 : * complicated.
3523 : *
3524 : * XXX This has the consequence that if there's a statistics object on
3525 : * the expression, we don't split it into individual Vars. This
3526 : * affects our selection of statistics in
3527 : * estimate_multivariate_ndistinct, because it's probably better to
3528 : * use more accurate estimate for each expression and treat them as
3529 : * independent, than to combine estimates for the extracted variables
3530 : * when we don't know how that relates to the expressions.
3531 : */
3532 373860 : examine_variable(root, groupexpr, 0, &vardata);
3533 373860 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3534 : {
3535 317620 : varinfos = add_unique_group_var(root, varinfos,
3536 : groupexpr, &vardata);
3537 317620 : ReleaseVariableStats(vardata);
3538 317620 : continue;
3539 : }
3540 56240 : ReleaseVariableStats(vardata);
3541 :
3542 : /*
3543 : * Else pull out the component Vars. Handle PlaceHolderVars by
3544 : * recursing into their arguments (effectively assuming that the
3545 : * PlaceHolderVar doesn't change the number of groups, which boils
3546 : * down to ignoring the possible addition of nulls to the result set).
3547 : */
3548 56240 : varshere = pull_var_clause(groupexpr,
3549 : PVC_RECURSE_AGGREGATES |
3550 : PVC_RECURSE_WINDOWFUNCS |
3551 : PVC_RECURSE_PLACEHOLDERS);
3552 :
3553 : /*
3554 : * If we find any variable-free GROUP BY item, then either it is a
3555 : * constant (and we can ignore it) or it contains a volatile function;
3556 : * in the latter case we punt and assume that each input row will
3557 : * yield a distinct group.
3558 : */
3559 56240 : if (varshere == NIL)
3560 : {
3561 636 : if (contain_volatile_functions(groupexpr))
3562 48 : return input_rows;
3563 588 : continue;
3564 : }
3565 :
3566 : /*
3567 : * Else add variables to varinfos list
3568 : */
3569 114374 : foreach(l2, varshere)
3570 : {
3571 58770 : Node *var = (Node *) lfirst(l2);
3572 :
3573 58770 : examine_variable(root, var, 0, &vardata);
3574 58770 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3575 58770 : ReleaseVariableStats(vardata);
3576 : }
3577 : }
3578 :
3579 : /*
3580 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3581 : * list.
3582 : */
3583 326566 : if (varinfos == NIL)
3584 : {
3585 : /* Apply SRF multiplier as we would do in the long path */
3586 346 : numdistinct *= srf_multiplier;
3587 : /* Round off */
3588 346 : numdistinct = ceil(numdistinct);
3589 : /* Guard against out-of-range answers */
3590 346 : if (numdistinct > input_rows)
3591 38 : numdistinct = input_rows;
3592 346 : if (numdistinct < 1.0)
3593 0 : numdistinct = 1.0;
3594 346 : return numdistinct;
3595 : }
3596 :
3597 : /*
3598 : * Group Vars by relation and estimate total numdistinct.
3599 : *
3600 : * For each iteration of the outer loop, we process the frontmost Var in
3601 : * varinfos, plus all other Vars in the same relation. We remove these
3602 : * Vars from the newvarinfos list for the next iteration. This is the
3603 : * easiest way to group Vars of same rel together.
3604 : */
3605 : do
3606 : {
3607 328130 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3608 328130 : RelOptInfo *rel = varinfo1->rel;
3609 328130 : double reldistinct = 1;
3610 328130 : double relmaxndistinct = reldistinct;
3611 328130 : int relvarcount = 0;
3612 328130 : List *newvarinfos = NIL;
3613 328130 : List *relvarinfos = NIL;
3614 :
3615 : /*
3616 : * Split the list of varinfos in two - one for the current rel, one
3617 : * for remaining Vars on other rels.
3618 : */
3619 328130 : relvarinfos = lappend(relvarinfos, varinfo1);
3620 379376 : for_each_from(l, varinfos, 1)
3621 : {
3622 51246 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3623 :
3624 51246 : if (varinfo2->rel == varinfo1->rel)
3625 : {
3626 : /* varinfos on current rel */
3627 47422 : relvarinfos = lappend(relvarinfos, varinfo2);
3628 : }
3629 : else
3630 : {
3631 : /* not time to process varinfo2 yet */
3632 3824 : newvarinfos = lappend(newvarinfos, varinfo2);
3633 : }
3634 : }
3635 :
3636 : /*
3637 : * Get the numdistinct estimate for the Vars of this rel. We
3638 : * iteratively search for multivariate n-distinct with maximum number
3639 : * of vars; assuming that each var group is independent of the others,
3640 : * we multiply them together. Any remaining relvarinfos after no more
3641 : * multivariate matches are found are assumed independent too, so
3642 : * their individual ndistinct estimates are multiplied also.
3643 : *
3644 : * While iterating, count how many separate numdistinct values we
3645 : * apply. We apply a fudge factor below, but only if we multiplied
3646 : * more than one such values.
3647 : */
3648 656386 : while (relvarinfos)
3649 : {
3650 : double mvndistinct;
3651 :
3652 328256 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3653 : &mvndistinct))
3654 : {
3655 402 : reldistinct *= mvndistinct;
3656 402 : if (relmaxndistinct < mvndistinct)
3657 390 : relmaxndistinct = mvndistinct;
3658 402 : relvarcount++;
3659 : }
3660 : else
3661 : {
3662 702554 : foreach(l, relvarinfos)
3663 : {
3664 374700 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3665 :
3666 374700 : reldistinct *= varinfo2->ndistinct;
3667 374700 : if (relmaxndistinct < varinfo2->ndistinct)
3668 333098 : relmaxndistinct = varinfo2->ndistinct;
3669 374700 : relvarcount++;
3670 :
3671 : /*
3672 : * When varinfo2's isdefault is set then we'd better set
3673 : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3674 : */
3675 374700 : if (estinfo != NULL && varinfo2->isdefault)
3676 17326 : estinfo->flags |= SELFLAG_USED_DEFAULT;
3677 : }
3678 :
3679 : /* we're done with this relation */
3680 327854 : relvarinfos = NIL;
3681 : }
3682 : }
3683 :
3684 : /*
3685 : * Sanity check --- don't divide by zero if empty relation.
3686 : */
3687 : Assert(IS_SIMPLE_REL(rel));
3688 328130 : if (rel->tuples > 0)
3689 : {
3690 : /*
3691 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
3692 : * fudge factor is because the Vars are probably correlated but we
3693 : * don't know by how much. We should never clamp to less than the
3694 : * largest ndistinct value for any of the Vars, though, since
3695 : * there will surely be at least that many groups.
3696 : */
3697 326892 : double clamp = rel->tuples;
3698 :
3699 326892 : if (relvarcount > 1)
3700 : {
3701 42920 : clamp *= 0.1;
3702 42920 : if (clamp < relmaxndistinct)
3703 : {
3704 41308 : clamp = relmaxndistinct;
3705 : /* for sanity in case some ndistinct is too large: */
3706 41308 : if (clamp > rel->tuples)
3707 78 : clamp = rel->tuples;
3708 : }
3709 : }
3710 326892 : if (reldistinct > clamp)
3711 41004 : reldistinct = clamp;
3712 :
3713 : /*
3714 : * Update the estimate based on the restriction selectivity,
3715 : * guarding against division by zero when reldistinct is zero.
3716 : * Also skip this if we know that we are returning all rows.
3717 : */
3718 326892 : if (reldistinct > 0 && rel->rows < rel->tuples)
3719 : {
3720 : /*
3721 : * Given a table containing N rows with n distinct values in a
3722 : * uniform distribution, if we select p rows at random then
3723 : * the expected number of distinct values selected is
3724 : *
3725 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
3726 : *
3727 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
3728 : *
3729 : * See "Approximating block accesses in database
3730 : * organizations", S. B. Yao, Communications of the ACM,
3731 : * Volume 20 Issue 4, April 1977 Pages 260-261.
3732 : *
3733 : * Alternatively, re-arranging the terms from the factorials,
3734 : * this may be written as
3735 : *
3736 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
3737 : *
3738 : * This form of the formula is more efficient to compute in
3739 : * the common case where p is larger than N/n. Additionally,
3740 : * as pointed out by Dell'Era, if i << N for all terms in the
3741 : * product, it can be approximated by
3742 : *
3743 : * n * (1 - ((N-p)/N)^(N/n))
3744 : *
3745 : * See "Expected distinct values when selecting from a bag
3746 : * without replacement", Alberto Dell'Era,
3747 : * http://www.adellera.it/investigations/distinct_balls/.
3748 : *
3749 : * The condition i << N is equivalent to n >> 1, so this is a
3750 : * good approximation when the number of distinct values in
3751 : * the table is large. It turns out that this formula also
3752 : * works well even when n is small.
3753 : */
3754 109732 : reldistinct *=
3755 109732 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
3756 109732 : rel->tuples / reldistinct));
3757 : }
3758 326892 : reldistinct = clamp_row_est(reldistinct);
3759 :
3760 : /*
3761 : * Update estimate of total distinct groups.
3762 : */
3763 326892 : numdistinct *= reldistinct;
3764 : }
3765 :
3766 328130 : varinfos = newvarinfos;
3767 328130 : } while (varinfos != NIL);
3768 :
3769 : /* Now we can account for the effects of any SRFs */
3770 326220 : numdistinct *= srf_multiplier;
3771 :
3772 : /* Round off */
3773 326220 : numdistinct = ceil(numdistinct);
3774 :
3775 : /* Guard against out-of-range answers */
3776 326220 : if (numdistinct > input_rows)
3777 69226 : numdistinct = input_rows;
3778 326220 : if (numdistinct < 1.0)
3779 0 : numdistinct = 1.0;
3780 :
3781 326220 : return numdistinct;
3782 : }
3783 :
3784 : /*
3785 : * Try to estimate the bucket size of the hash join inner side when the join
3786 : * condition contains two or more clauses by employing extended statistics.
3787 : *
3788 : * The main idea of this approach is that the distinct value generated by
3789 : * multivariate estimation on two or more columns would provide less bucket size
3790 : * than estimation on one separate column.
3791 : *
3792 : * IMPORTANT: It is crucial to synchronize the approach of combining different
3793 : * estimations with the caller's method.
3794 : *
3795 : * Return a list of clauses that didn't fetch any extended statistics.
3796 : */
3797 : List *
3798 291050 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
3799 : List *hashclauses,
3800 : Selectivity *innerbucketsize)
3801 : {
3802 291050 : List *clauses = list_copy(hashclauses);
3803 291050 : List *otherclauses = NIL;
3804 291050 : double ndistinct = 1.0;
3805 :
3806 291050 : if (list_length(hashclauses) <= 1)
3807 :
3808 : /*
3809 : * Nothing to do for a single clause. Could we employ univariate
3810 : * extended stat here?
3811 : */
3812 260746 : return hashclauses;
3813 :
3814 60620 : while (clauses != NIL)
3815 : {
3816 : ListCell *lc;
3817 30316 : int relid = -1;
3818 30316 : List *varinfos = NIL;
3819 30316 : List *origin_rinfos = NIL;
3820 : double mvndistinct;
3821 : List *origin_varinfos;
3822 30316 : int group_relid = -1;
3823 30316 : RelOptInfo *group_rel = NULL;
3824 : ListCell *lc1,
3825 : *lc2;
3826 :
3827 : /*
3828 : * Find clauses, referencing the same single base relation and try to
3829 : * estimate such a group with extended statistics. Create varinfo for
3830 : * an approved clause, push it to otherclauses, if it can't be
3831 : * estimated here or ignore to process at the next iteration.
3832 : */
3833 91560 : foreach(lc, clauses)
3834 : {
3835 61244 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
3836 : Node *expr;
3837 : Relids relids;
3838 : GroupVarInfo *varinfo;
3839 :
3840 : /*
3841 : * Find the inner side of the join, which we need to estimate the
3842 : * number of buckets. Use outer_is_left because the
3843 : * clause_sides_match_join routine has called on hash clauses.
3844 : */
3845 122488 : relids = rinfo->outer_is_left ?
3846 61244 : rinfo->right_relids : rinfo->left_relids;
3847 122488 : expr = rinfo->outer_is_left ?
3848 61244 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
3849 :
3850 61244 : if (bms_get_singleton_member(relids, &relid) &&
3851 60590 : root->simple_rel_array[relid]->statlist != NIL)
3852 48 : {
3853 60 : bool is_duplicate = false;
3854 :
3855 : /*
3856 : * This inner-side expression references only one relation.
3857 : * Extended statistics on this clause can exist.
3858 : */
3859 60 : if (group_relid < 0)
3860 : {
3861 30 : RangeTblEntry *rte = root->simple_rte_array[relid];
3862 :
3863 30 : if (!rte || (rte->relkind != RELKIND_RELATION &&
3864 0 : rte->relkind != RELKIND_MATVIEW &&
3865 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
3866 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
3867 : {
3868 : /* Extended statistics can't exist in principle */
3869 0 : otherclauses = lappend(otherclauses, rinfo);
3870 0 : clauses = foreach_delete_current(clauses, lc);
3871 0 : continue;
3872 : }
3873 :
3874 30 : group_relid = relid;
3875 30 : group_rel = root->simple_rel_array[relid];
3876 : }
3877 30 : else if (group_relid != relid)
3878 :
3879 : /*
3880 : * Being in the group forming state we don't need other
3881 : * clauses.
3882 : */
3883 0 : continue;
3884 :
3885 : /*
3886 : * We're going to add the new clause to the varinfos list. We
3887 : * might re-use add_unique_group_var(), but we don't do so for
3888 : * two reasons.
3889 : *
3890 : * 1) We must keep the origin_rinfos list ordered exactly the
3891 : * same way as varinfos.
3892 : *
3893 : * 2) add_unique_group_var() is designed for
3894 : * estimate_num_groups(), where a larger number of groups is
3895 : * worse. While estimating the number of hash buckets, we
3896 : * have the opposite: a lesser number of groups is worse.
3897 : * Therefore, we don't have to remove "known equal" vars: the
3898 : * removed var may valuably contribute to the multivariate
3899 : * statistics to grow the number of groups.
3900 : */
3901 :
3902 : /*
3903 : * Clear nullingrels to correctly match hash keys. See
3904 : * add_unique_group_var()'s comment for details.
3905 : */
3906 60 : expr = remove_nulling_relids(expr, root->outer_join_rels, NULL);
3907 :
3908 : /*
3909 : * Detect and exclude exact duplicates from the list of hash
3910 : * keys (like add_unique_group_var does).
3911 : */
3912 84 : foreach(lc1, varinfos)
3913 : {
3914 36 : varinfo = (GroupVarInfo *) lfirst(lc1);
3915 :
3916 36 : if (!equal(expr, varinfo->var))
3917 24 : continue;
3918 :
3919 12 : is_duplicate = true;
3920 12 : break;
3921 : }
3922 :
3923 60 : if (is_duplicate)
3924 : {
3925 : /*
3926 : * Skip exact duplicates. Adding them to the otherclauses
3927 : * list also doesn't make sense.
3928 : */
3929 12 : continue;
3930 : }
3931 :
3932 : /*
3933 : * Initialize GroupVarInfo. We only use it to call
3934 : * estimate_multivariate_ndistinct(), which doesn't care about
3935 : * ndistinct and isdefault fields. Thus, skip these fields.
3936 : */
3937 48 : varinfo = (GroupVarInfo *) palloc0(sizeof(GroupVarInfo));
3938 48 : varinfo->var = expr;
3939 48 : varinfo->rel = root->simple_rel_array[relid];
3940 48 : varinfos = lappend(varinfos, varinfo);
3941 :
3942 : /*
3943 : * Remember the link to RestrictInfo for the case the clause
3944 : * is failed to be estimated.
3945 : */
3946 48 : origin_rinfos = lappend(origin_rinfos, rinfo);
3947 : }
3948 : else
3949 : {
3950 : /* This clause can't be estimated with extended statistics */
3951 61184 : otherclauses = lappend(otherclauses, rinfo);
3952 : }
3953 :
3954 61232 : clauses = foreach_delete_current(clauses, lc);
3955 : }
3956 :
3957 30316 : if (list_length(varinfos) < 2)
3958 : {
3959 : /*
3960 : * Multivariate statistics doesn't apply to single columns except
3961 : * for expressions, but it has not been implemented yet.
3962 : */
3963 30304 : otherclauses = list_concat(otherclauses, origin_rinfos);
3964 30304 : list_free_deep(varinfos);
3965 30304 : list_free(origin_rinfos);
3966 30304 : continue;
3967 : }
3968 :
3969 : Assert(group_rel != NULL);
3970 :
3971 : /* Employ the extended statistics. */
3972 12 : origin_varinfos = varinfos;
3973 : for (;;)
3974 12 : {
3975 24 : bool estimated = estimate_multivariate_ndistinct(root,
3976 : group_rel,
3977 : &varinfos,
3978 : &mvndistinct);
3979 :
3980 24 : if (!estimated)
3981 12 : break;
3982 :
3983 : /*
3984 : * We've got an estimation. Use ndistinct value in a consistent
3985 : * way - according to the caller's logic (see
3986 : * final_cost_hashjoin).
3987 : */
3988 12 : if (ndistinct < mvndistinct)
3989 12 : ndistinct = mvndistinct;
3990 : Assert(ndistinct >= 1.0);
3991 : }
3992 :
3993 : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
3994 :
3995 : /* Collect unmatched clauses as otherclauses. */
3996 42 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
3997 : {
3998 30 : GroupVarInfo *vinfo = lfirst(lc1);
3999 :
4000 30 : if (!list_member_ptr(varinfos, vinfo))
4001 : /* Already estimated */
4002 30 : continue;
4003 :
4004 : /* Can't be estimated here - push to the returning list */
4005 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
4006 : }
4007 : }
4008 :
4009 30304 : *innerbucketsize = 1.0 / ndistinct;
4010 30304 : return otherclauses;
4011 : }
4012 :
4013 : /*
4014 : * Estimate hash bucket statistics when the specified expression is used
4015 : * as a hash key for the given number of buckets.
4016 : *
4017 : * This attempts to determine two values:
4018 : *
4019 : * 1. The frequency of the most common value of the expression (returns
4020 : * zero into *mcv_freq if we can't get that).
4021 : *
4022 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
4023 : * divided by total tuples in relation.
4024 : *
4025 : * XXX This is really pretty bogus since we're effectively assuming that the
4026 : * distribution of hash keys will be the same after applying restriction
4027 : * clauses as it was in the underlying relation. However, we are not nearly
4028 : * smart enough to figure out how the restrict clauses might change the
4029 : * distribution, so this will have to do for now.
4030 : *
4031 : * We are passed the number of buckets the executor will use for the given
4032 : * input relation. If the data were perfectly distributed, with the same
4033 : * number of tuples going into each available bucket, then the bucketsize
4034 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
4035 : * only if (a) there are at least nbuckets distinct data values, and (b)
4036 : * we have a not-too-skewed data distribution. Otherwise the buckets will
4037 : * be nonuniformly occupied. If the other relation in the join has a key
4038 : * distribution similar to this one's, then the most-loaded buckets are
4039 : * exactly those that will be probed most often. Therefore, the "average"
4040 : * bucket size for costing purposes should really be taken as something close
4041 : * to the "worst case" bucket size. We try to estimate this by adjusting the
4042 : * fraction if there are too few distinct data values, and then scaling up
4043 : * by the ratio of the most common value's frequency to the average frequency.
4044 : *
4045 : * If no statistics are available, use a default estimate of 0.1. This will
4046 : * discourage use of a hash rather strongly if the inner relation is large,
4047 : * which is what we want. We do not want to hash unless we know that the
4048 : * inner rel is well-dispersed (or the alternatives seem much worse).
4049 : *
4050 : * The caller should also check that the mcv_freq is not so large that the
4051 : * most common value would by itself require an impractically large bucket.
4052 : * In a hash join, the executor can split buckets if they get too big, but
4053 : * obviously that doesn't help for a bucket that contains many duplicates of
4054 : * the same value.
4055 : */
4056 : void
4057 169330 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
4058 : Selectivity *mcv_freq,
4059 : Selectivity *bucketsize_frac)
4060 : {
4061 : VariableStatData vardata;
4062 : double estfract,
4063 : ndistinct,
4064 : stanullfrac,
4065 : avgfreq;
4066 : bool isdefault;
4067 : AttStatsSlot sslot;
4068 :
4069 169330 : examine_variable(root, hashkey, 0, &vardata);
4070 :
4071 : /* Look up the frequency of the most common value, if available */
4072 169330 : *mcv_freq = 0.0;
4073 :
4074 169330 : if (HeapTupleIsValid(vardata.statsTuple))
4075 : {
4076 116662 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4077 : STATISTIC_KIND_MCV, InvalidOid,
4078 : ATTSTATSSLOT_NUMBERS))
4079 : {
4080 : /*
4081 : * The first MCV stat is for the most common value.
4082 : */
4083 57400 : if (sslot.nnumbers > 0)
4084 57400 : *mcv_freq = sslot.numbers[0];
4085 57400 : free_attstatsslot(&sslot);
4086 : }
4087 : }
4088 :
4089 : /* Get number of distinct values */
4090 169330 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4091 :
4092 : /*
4093 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4094 : * mcv_freq is known to be even higher than that, use it instead.
4095 : */
4096 169330 : if (isdefault)
4097 : {
4098 22998 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4099 22998 : ReleaseVariableStats(vardata);
4100 22998 : return;
4101 : }
4102 :
4103 : /* Get fraction that are null */
4104 146332 : if (HeapTupleIsValid(vardata.statsTuple))
4105 : {
4106 : Form_pg_statistic stats;
4107 :
4108 116644 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
4109 116644 : stanullfrac = stats->stanullfrac;
4110 : }
4111 : else
4112 29688 : stanullfrac = 0.0;
4113 :
4114 : /* Compute avg freq of all distinct data values in raw relation */
4115 146332 : avgfreq = (1.0 - stanullfrac) / ndistinct;
4116 :
4117 : /*
4118 : * Adjust ndistinct to account for restriction clauses. Observe we are
4119 : * assuming that the data distribution is affected uniformly by the
4120 : * restriction clauses!
4121 : *
4122 : * XXX Possibly better way, but much more expensive: multiply by
4123 : * selectivity of rel's restriction clauses that mention the target Var.
4124 : */
4125 146332 : if (vardata.rel && vardata.rel->tuples > 0)
4126 : {
4127 146318 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
4128 146318 : ndistinct = clamp_row_est(ndistinct);
4129 : }
4130 :
4131 : /*
4132 : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
4133 : * number of buckets is less than the expected number of distinct values;
4134 : * otherwise it is 1/ndistinct.
4135 : */
4136 146332 : if (ndistinct > nbuckets)
4137 88 : estfract = 1.0 / nbuckets;
4138 : else
4139 146244 : estfract = 1.0 / ndistinct;
4140 :
4141 : /*
4142 : * Adjust estimated bucketsize upward to account for skewed distribution.
4143 : */
4144 146332 : if (avgfreq > 0.0 && *mcv_freq > avgfreq)
4145 49998 : estfract *= *mcv_freq / avgfreq;
4146 :
4147 : /*
4148 : * Clamp bucketsize to sane range (the above adjustment could easily
4149 : * produce an out-of-range result). We set the lower bound a little above
4150 : * zero, since zero isn't a very sane result.
4151 : */
4152 146332 : if (estfract < 1.0e-6)
4153 0 : estfract = 1.0e-6;
4154 146332 : else if (estfract > 1.0)
4155 34628 : estfract = 1.0;
4156 :
4157 146332 : *bucketsize_frac = (Selectivity) estfract;
4158 :
4159 146332 : ReleaseVariableStats(vardata);
4160 : }
4161 :
4162 : /*
4163 : * estimate_hashagg_tablesize
4164 : * estimate the number of bytes that a hash aggregate hashtable will
4165 : * require based on the agg_costs, path width and number of groups.
4166 : *
4167 : * We return the result as "double" to forestall any possible overflow
4168 : * problem in the multiplication by dNumGroups.
4169 : *
4170 : * XXX this may be over-estimating the size now that hashagg knows to omit
4171 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4172 : * grouping columns not in the hashed set are counted here even though hashagg
4173 : * won't store them. Is this a problem?
4174 : */
4175 : double
4176 2338 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4177 : const AggClauseCosts *agg_costs, double dNumGroups)
4178 : {
4179 : Size hashentrysize;
4180 :
4181 2338 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4182 2338 : path->pathtarget->width,
4183 : agg_costs->transitionSpace);
4184 :
4185 : /*
4186 : * Note that this disregards the effect of fill-factor and growth policy
4187 : * of the hash table. That's probably ok, given that the default
4188 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4189 : * "double-in-size" growth policies here.
4190 : */
4191 2338 : return hashentrysize * dNumGroups;
4192 : }
4193 :
4194 :
4195 : /*-------------------------------------------------------------------------
4196 : *
4197 : * Support routines
4198 : *
4199 : *-------------------------------------------------------------------------
4200 : */
4201 :
4202 : /*
4203 : * Find the best matching ndistinct extended statistics for the given list of
4204 : * GroupVarInfos.
4205 : *
4206 : * Callers must ensure that the given GroupVarInfos all belong to 'rel' and
4207 : * the GroupVarInfos list does not contain any duplicate Vars or expressions.
4208 : *
4209 : * When statistics are found that match > 1 of the given GroupVarInfo, the
4210 : * *ndistinct parameter is set according to the ndistinct estimate and a new
4211 : * list is built with the matching GroupVarInfos removed, which is output via
4212 : * the *varinfos parameter before returning true. When no matching stats are
4213 : * found, false is returned and the *varinfos and *ndistinct parameters are
4214 : * left untouched.
4215 : */
4216 : static bool
4217 328280 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4218 : List **varinfos, double *ndistinct)
4219 : {
4220 : ListCell *lc;
4221 : int nmatches_vars;
4222 : int nmatches_exprs;
4223 328280 : Oid statOid = InvalidOid;
4224 : MVNDistinct *stats;
4225 328280 : StatisticExtInfo *matched_info = NULL;
4226 328280 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4227 :
4228 : /* bail out immediately if the table has no extended statistics */
4229 328280 : if (!rel->statlist)
4230 327728 : return false;
4231 :
4232 : /* look for the ndistinct statistics object matching the most vars */
4233 552 : nmatches_vars = 0; /* we require at least two matches */
4234 552 : nmatches_exprs = 0;
4235 2172 : foreach(lc, rel->statlist)
4236 : {
4237 : ListCell *lc2;
4238 1620 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
4239 1620 : int nshared_vars = 0;
4240 1620 : int nshared_exprs = 0;
4241 :
4242 : /* skip statistics of other kinds */
4243 1620 : if (info->kind != STATS_EXT_NDISTINCT)
4244 750 : continue;
4245 :
4246 : /* skip statistics with mismatching stxdinherit value */
4247 870 : if (info->inherit != rte->inh)
4248 24 : continue;
4249 :
4250 : /*
4251 : * Determine how many expressions (and variables in non-matched
4252 : * expressions) match. We'll then use these numbers to pick the
4253 : * statistics object that best matches the clauses.
4254 : */
4255 2682 : foreach(lc2, *varinfos)
4256 : {
4257 : ListCell *lc3;
4258 1836 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4259 : AttrNumber attnum;
4260 :
4261 : Assert(varinfo->rel == rel);
4262 :
4263 : /* simple Var, search in statistics keys directly */
4264 1836 : if (IsA(varinfo->var, Var))
4265 : {
4266 1470 : attnum = ((Var *) varinfo->var)->varattno;
4267 :
4268 : /*
4269 : * Ignore system attributes - we don't support statistics on
4270 : * them, so can't match them (and it'd fail as the values are
4271 : * negative).
4272 : */
4273 1470 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4274 12 : continue;
4275 :
4276 1458 : if (bms_is_member(attnum, info->keys))
4277 852 : nshared_vars++;
4278 :
4279 1458 : continue;
4280 : }
4281 :
4282 : /* expression - see if it's in the statistics object */
4283 660 : foreach(lc3, info->exprs)
4284 : {
4285 528 : Node *expr = (Node *) lfirst(lc3);
4286 :
4287 528 : if (equal(varinfo->var, expr))
4288 : {
4289 234 : nshared_exprs++;
4290 234 : break;
4291 : }
4292 : }
4293 : }
4294 :
4295 : /*
4296 : * The ndistinct extended statistics contain estimates for a minimum
4297 : * of pairs of columns which the statistics are defined on and
4298 : * certainly not single columns. Here we skip unless we managed to
4299 : * match to at least two columns.
4300 : */
4301 846 : if (nshared_vars + nshared_exprs < 2)
4302 396 : continue;
4303 :
4304 : /*
4305 : * Check if these statistics are a better match than the previous best
4306 : * match and if so, take note of the StatisticExtInfo.
4307 : *
4308 : * The statslist is sorted by statOid, so the StatisticExtInfo we
4309 : * select as the best match is deterministic even when multiple sets
4310 : * of statistics match equally as well.
4311 : */
4312 450 : if ((nshared_exprs > nmatches_exprs) ||
4313 342 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4314 : {
4315 426 : statOid = info->statOid;
4316 426 : nmatches_vars = nshared_vars;
4317 426 : nmatches_exprs = nshared_exprs;
4318 426 : matched_info = info;
4319 : }
4320 : }
4321 :
4322 : /* No match? */
4323 552 : if (statOid == InvalidOid)
4324 138 : return false;
4325 :
4326 : Assert(nmatches_vars + nmatches_exprs > 1);
4327 :
4328 414 : stats = statext_ndistinct_load(statOid, rte->inh);
4329 :
4330 : /*
4331 : * If we have a match, search it for the specific item that matches (there
4332 : * must be one), and construct the output values.
4333 : */
4334 414 : if (stats)
4335 : {
4336 : int i;
4337 414 : List *newlist = NIL;
4338 414 : MVNDistinctItem *item = NULL;
4339 : ListCell *lc2;
4340 414 : Bitmapset *matched = NULL;
4341 : AttrNumber attnum_offset;
4342 :
4343 : /*
4344 : * How much we need to offset the attnums? If there are no
4345 : * expressions, no offset is needed. Otherwise offset enough to move
4346 : * the lowest one (which is equal to number of expressions) to 1.
4347 : */
4348 414 : if (matched_info->exprs)
4349 144 : attnum_offset = (list_length(matched_info->exprs) + 1);
4350 : else
4351 270 : attnum_offset = 0;
4352 :
4353 : /* see what actually matched */
4354 1452 : foreach(lc2, *varinfos)
4355 : {
4356 : ListCell *lc3;
4357 : int idx;
4358 1038 : bool found = false;
4359 :
4360 1038 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4361 :
4362 : /*
4363 : * Process a simple Var expression, by matching it to keys
4364 : * directly. If there's a matching expression, we'll try matching
4365 : * it later.
4366 : */
4367 1038 : if (IsA(varinfo->var, Var))
4368 : {
4369 852 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4370 :
4371 : /*
4372 : * Ignore expressions on system attributes. Can't rely on the
4373 : * bms check for negative values.
4374 : */
4375 852 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4376 6 : continue;
4377 :
4378 : /* Is the variable covered by the statistics object? */
4379 846 : if (!bms_is_member(attnum, matched_info->keys))
4380 120 : continue;
4381 :
4382 726 : attnum = attnum + attnum_offset;
4383 :
4384 : /* ensure sufficient offset */
4385 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4386 :
4387 726 : matched = bms_add_member(matched, attnum);
4388 :
4389 726 : found = true;
4390 : }
4391 :
4392 : /*
4393 : * XXX Maybe we should allow searching the expressions even if we
4394 : * found an attribute matching the expression? That would handle
4395 : * trivial expressions like "(a)" but it seems fairly useless.
4396 : */
4397 912 : if (found)
4398 726 : continue;
4399 :
4400 : /* expression - see if it's in the statistics object */
4401 186 : idx = 0;
4402 306 : foreach(lc3, matched_info->exprs)
4403 : {
4404 276 : Node *expr = (Node *) lfirst(lc3);
4405 :
4406 276 : if (equal(varinfo->var, expr))
4407 : {
4408 156 : AttrNumber attnum = -(idx + 1);
4409 :
4410 156 : attnum = attnum + attnum_offset;
4411 :
4412 : /* ensure sufficient offset */
4413 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4414 :
4415 156 : matched = bms_add_member(matched, attnum);
4416 :
4417 : /* there should be just one matching expression */
4418 156 : break;
4419 : }
4420 :
4421 120 : idx++;
4422 : }
4423 : }
4424 :
4425 : /* Find the specific item that exactly matches the combination */
4426 852 : for (i = 0; i < stats->nitems; i++)
4427 : {
4428 : int j;
4429 852 : MVNDistinctItem *tmpitem = &stats->items[i];
4430 :
4431 852 : if (tmpitem->nattributes != bms_num_members(matched))
4432 162 : continue;
4433 :
4434 : /* assume it's the right item */
4435 690 : item = tmpitem;
4436 :
4437 : /* check that all item attributes/expressions fit the match */
4438 1656 : for (j = 0; j < tmpitem->nattributes; j++)
4439 : {
4440 1242 : AttrNumber attnum = tmpitem->attributes[j];
4441 :
4442 : /*
4443 : * Thanks to how we constructed the matched bitmap above, we
4444 : * can just offset all attnums the same way.
4445 : */
4446 1242 : attnum = attnum + attnum_offset;
4447 :
4448 1242 : if (!bms_is_member(attnum, matched))
4449 : {
4450 : /* nah, it's not this item */
4451 276 : item = NULL;
4452 276 : break;
4453 : }
4454 : }
4455 :
4456 : /*
4457 : * If the item has all the matched attributes, we know it's the
4458 : * right one - there can't be a better one. matching more.
4459 : */
4460 690 : if (item)
4461 414 : break;
4462 : }
4463 :
4464 : /*
4465 : * Make sure we found an item. There has to be one, because ndistinct
4466 : * statistics includes all combinations of attributes.
4467 : */
4468 414 : if (!item)
4469 0 : elog(ERROR, "corrupt MVNDistinct entry");
4470 :
4471 : /* Form the output varinfo list, keeping only unmatched ones */
4472 1452 : foreach(lc, *varinfos)
4473 : {
4474 1038 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4475 : ListCell *lc3;
4476 1038 : bool found = false;
4477 :
4478 : /*
4479 : * Let's look at plain variables first, because it's the most
4480 : * common case and the check is quite cheap. We can simply get the
4481 : * attnum and check (with an offset) matched bitmap.
4482 : */
4483 1038 : if (IsA(varinfo->var, Var))
4484 : {
4485 852 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4486 :
4487 : /*
4488 : * If it's a system attribute, we're done. We don't support
4489 : * extended statistics on system attributes, so it's clearly
4490 : * not matched. Just keep the expression and continue.
4491 : */
4492 852 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4493 : {
4494 6 : newlist = lappend(newlist, varinfo);
4495 6 : continue;
4496 : }
4497 :
4498 : /* apply the same offset as above */
4499 846 : attnum += attnum_offset;
4500 :
4501 : /* if it's not matched, keep the varinfo */
4502 846 : if (!bms_is_member(attnum, matched))
4503 120 : newlist = lappend(newlist, varinfo);
4504 :
4505 : /* The rest of the loop deals with complex expressions. */
4506 846 : continue;
4507 : }
4508 :
4509 : /*
4510 : * Process complex expressions, not just simple Vars.
4511 : *
4512 : * First, we search for an exact match of an expression. If we
4513 : * find one, we can just discard the whole GroupVarInfo, with all
4514 : * the variables we extracted from it.
4515 : *
4516 : * Otherwise we inspect the individual vars, and try matching it
4517 : * to variables in the item.
4518 : */
4519 306 : foreach(lc3, matched_info->exprs)
4520 : {
4521 276 : Node *expr = (Node *) lfirst(lc3);
4522 :
4523 276 : if (equal(varinfo->var, expr))
4524 : {
4525 156 : found = true;
4526 156 : break;
4527 : }
4528 : }
4529 :
4530 : /* found exact match, skip */
4531 186 : if (found)
4532 156 : continue;
4533 :
4534 30 : newlist = lappend(newlist, varinfo);
4535 : }
4536 :
4537 414 : *varinfos = newlist;
4538 414 : *ndistinct = item->ndistinct;
4539 414 : return true;
4540 : }
4541 :
4542 0 : return false;
4543 : }
4544 :
4545 : /*
4546 : * convert_to_scalar
4547 : * Convert non-NULL values of the indicated types to the comparison
4548 : * scale needed by scalarineqsel().
4549 : * Returns "true" if successful.
4550 : *
4551 : * XXX this routine is a hack: ideally we should look up the conversion
4552 : * subroutines in pg_type.
4553 : *
4554 : * All numeric datatypes are simply converted to their equivalent
4555 : * "double" values. (NUMERIC values that are outside the range of "double"
4556 : * are clamped to +/- HUGE_VAL.)
4557 : *
4558 : * String datatypes are converted by convert_string_to_scalar(),
4559 : * which is explained below. The reason why this routine deals with
4560 : * three values at a time, not just one, is that we need it for strings.
4561 : *
4562 : * The bytea datatype is just enough different from strings that it has
4563 : * to be treated separately.
4564 : *
4565 : * The several datatypes representing absolute times are all converted
4566 : * to Timestamp, which is actually an int64, and then we promote that to
4567 : * a double. Note this will give correct results even for the "special"
4568 : * values of Timestamp, since those are chosen to compare correctly;
4569 : * see timestamp_cmp.
4570 : *
4571 : * The several datatypes representing relative times (intervals) are all
4572 : * converted to measurements expressed in seconds.
4573 : */
4574 : static bool
4575 91584 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4576 : Datum lobound, Datum hibound, Oid boundstypid,
4577 : double *scaledlobound, double *scaledhibound)
4578 : {
4579 91584 : bool failure = false;
4580 :
4581 : /*
4582 : * Both the valuetypid and the boundstypid should exactly match the
4583 : * declared input type(s) of the operator we are invoked for. However,
4584 : * extensions might try to use scalarineqsel as estimator for operators
4585 : * with input type(s) we don't handle here; in such cases, we want to
4586 : * return false, not fail. In any case, we mustn't assume that valuetypid
4587 : * and boundstypid are identical.
4588 : *
4589 : * XXX The histogram we are interpolating between points of could belong
4590 : * to a column that's only binary-compatible with the declared type. In
4591 : * essence we are assuming that the semantics of binary-compatible types
4592 : * are enough alike that we can use a histogram generated with one type's
4593 : * operators to estimate selectivity for the other's. This is outright
4594 : * wrong in some cases --- in particular signed versus unsigned
4595 : * interpretation could trip us up. But it's useful enough in the
4596 : * majority of cases that we do it anyway. Should think about more
4597 : * rigorous ways to do it.
4598 : */
4599 91584 : switch (valuetypid)
4600 : {
4601 : /*
4602 : * Built-in numeric types
4603 : */
4604 85126 : case BOOLOID:
4605 : case INT2OID:
4606 : case INT4OID:
4607 : case INT8OID:
4608 : case FLOAT4OID:
4609 : case FLOAT8OID:
4610 : case NUMERICOID:
4611 : case OIDOID:
4612 : case REGPROCOID:
4613 : case REGPROCEDUREOID:
4614 : case REGOPEROID:
4615 : case REGOPERATOROID:
4616 : case REGCLASSOID:
4617 : case REGTYPEOID:
4618 : case REGCOLLATIONOID:
4619 : case REGCONFIGOID:
4620 : case REGDICTIONARYOID:
4621 : case REGROLEOID:
4622 : case REGNAMESPACEOID:
4623 85126 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4624 : &failure);
4625 85126 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4626 : &failure);
4627 85126 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4628 : &failure);
4629 85126 : return !failure;
4630 :
4631 : /*
4632 : * Built-in string types
4633 : */
4634 6458 : case CHAROID:
4635 : case BPCHAROID:
4636 : case VARCHAROID:
4637 : case TEXTOID:
4638 : case NAMEOID:
4639 : {
4640 6458 : char *valstr = convert_string_datum(value, valuetypid,
4641 : collid, &failure);
4642 6458 : char *lostr = convert_string_datum(lobound, boundstypid,
4643 : collid, &failure);
4644 6458 : char *histr = convert_string_datum(hibound, boundstypid,
4645 : collid, &failure);
4646 :
4647 : /*
4648 : * Bail out if any of the values is not of string type. We
4649 : * might leak converted strings for the other value(s), but
4650 : * that's not worth troubling over.
4651 : */
4652 6458 : if (failure)
4653 0 : return false;
4654 :
4655 6458 : convert_string_to_scalar(valstr, scaledvalue,
4656 : lostr, scaledlobound,
4657 : histr, scaledhibound);
4658 6458 : pfree(valstr);
4659 6458 : pfree(lostr);
4660 6458 : pfree(histr);
4661 6458 : return true;
4662 : }
4663 :
4664 : /*
4665 : * Built-in bytea type
4666 : */
4667 0 : case BYTEAOID:
4668 : {
4669 : /* We only support bytea vs bytea comparison */
4670 0 : if (boundstypid != BYTEAOID)
4671 0 : return false;
4672 0 : convert_bytea_to_scalar(value, scaledvalue,
4673 : lobound, scaledlobound,
4674 : hibound, scaledhibound);
4675 0 : return true;
4676 : }
4677 :
4678 : /*
4679 : * Built-in time types
4680 : */
4681 0 : case TIMESTAMPOID:
4682 : case TIMESTAMPTZOID:
4683 : case DATEOID:
4684 : case INTERVALOID:
4685 : case TIMEOID:
4686 : case TIMETZOID:
4687 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
4688 : &failure);
4689 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
4690 : &failure);
4691 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
4692 : &failure);
4693 0 : return !failure;
4694 :
4695 : /*
4696 : * Built-in network types
4697 : */
4698 0 : case INETOID:
4699 : case CIDROID:
4700 : case MACADDROID:
4701 : case MACADDR8OID:
4702 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
4703 : &failure);
4704 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
4705 : &failure);
4706 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
4707 : &failure);
4708 0 : return !failure;
4709 : }
4710 : /* Don't know how to convert */
4711 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
4712 0 : return false;
4713 : }
4714 :
4715 : /*
4716 : * Do convert_to_scalar()'s work for any numeric data type.
4717 : *
4718 : * On failure (e.g., unsupported typid), set *failure to true;
4719 : * otherwise, that variable is not changed.
4720 : */
4721 : static double
4722 255378 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
4723 : {
4724 255378 : switch (typid)
4725 : {
4726 0 : case BOOLOID:
4727 0 : return (double) DatumGetBool(value);
4728 12 : case INT2OID:
4729 12 : return (double) DatumGetInt16(value);
4730 31392 : case INT4OID:
4731 31392 : return (double) DatumGetInt32(value);
4732 0 : case INT8OID:
4733 0 : return (double) DatumGetInt64(value);
4734 0 : case FLOAT4OID:
4735 0 : return (double) DatumGetFloat4(value);
4736 54 : case FLOAT8OID:
4737 54 : return (double) DatumGetFloat8(value);
4738 0 : case NUMERICOID:
4739 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
4740 0 : return (double)
4741 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
4742 : value));
4743 223920 : case OIDOID:
4744 : case REGPROCOID:
4745 : case REGPROCEDUREOID:
4746 : case REGOPEROID:
4747 : case REGOPERATOROID:
4748 : case REGCLASSOID:
4749 : case REGTYPEOID:
4750 : case REGCOLLATIONOID:
4751 : case REGCONFIGOID:
4752 : case REGDICTIONARYOID:
4753 : case REGROLEOID:
4754 : case REGNAMESPACEOID:
4755 : /* we can treat OIDs as integers... */
4756 223920 : return (double) DatumGetObjectId(value);
4757 : }
4758 :
4759 0 : *failure = true;
4760 0 : return 0;
4761 : }
4762 :
4763 : /*
4764 : * Do convert_to_scalar()'s work for any character-string data type.
4765 : *
4766 : * String datatypes are converted to a scale that ranges from 0 to 1,
4767 : * where we visualize the bytes of the string as fractional digits.
4768 : *
4769 : * We do not want the base to be 256, however, since that tends to
4770 : * generate inflated selectivity estimates; few databases will have
4771 : * occurrences of all 256 possible byte values at each position.
4772 : * Instead, use the smallest and largest byte values seen in the bounds
4773 : * as the estimated range for each byte, after some fudging to deal with
4774 : * the fact that we probably aren't going to see the full range that way.
4775 : *
4776 : * An additional refinement is that we discard any common prefix of the
4777 : * three strings before computing the scaled values. This allows us to
4778 : * "zoom in" when we encounter a narrow data range. An example is a phone
4779 : * number database where all the values begin with the same area code.
4780 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
4781 : * so this is more likely to happen than you might think.)
4782 : */
4783 : static void
4784 6458 : convert_string_to_scalar(char *value,
4785 : double *scaledvalue,
4786 : char *lobound,
4787 : double *scaledlobound,
4788 : char *hibound,
4789 : double *scaledhibound)
4790 : {
4791 : int rangelo,
4792 : rangehi;
4793 : char *sptr;
4794 :
4795 6458 : rangelo = rangehi = (unsigned char) hibound[0];
4796 81556 : for (sptr = lobound; *sptr; sptr++)
4797 : {
4798 75098 : if (rangelo > (unsigned char) *sptr)
4799 15324 : rangelo = (unsigned char) *sptr;
4800 75098 : if (rangehi < (unsigned char) *sptr)
4801 8022 : rangehi = (unsigned char) *sptr;
4802 : }
4803 80576 : for (sptr = hibound; *sptr; sptr++)
4804 : {
4805 74118 : if (rangelo > (unsigned char) *sptr)
4806 1008 : rangelo = (unsigned char) *sptr;
4807 74118 : if (rangehi < (unsigned char) *sptr)
4808 3172 : rangehi = (unsigned char) *sptr;
4809 : }
4810 : /* If range includes any upper-case ASCII chars, make it include all */
4811 6458 : if (rangelo <= 'Z' && rangehi >= 'A')
4812 : {
4813 1038 : if (rangelo > 'A')
4814 138 : rangelo = 'A';
4815 1038 : if (rangehi < 'Z')
4816 480 : rangehi = 'Z';
4817 : }
4818 : /* Ditto lower-case */
4819 6458 : if (rangelo <= 'z' && rangehi >= 'a')
4820 : {
4821 5956 : if (rangelo > 'a')
4822 54 : rangelo = 'a';
4823 5956 : if (rangehi < 'z')
4824 5886 : rangehi = 'z';
4825 : }
4826 : /* Ditto digits */
4827 6458 : if (rangelo <= '9' && rangehi >= '0')
4828 : {
4829 442 : if (rangelo > '0')
4830 348 : rangelo = '0';
4831 442 : if (rangehi < '9')
4832 14 : rangehi = '9';
4833 : }
4834 :
4835 : /*
4836 : * If range includes less than 10 chars, assume we have not got enough
4837 : * data, and make it include regular ASCII set.
4838 : */
4839 6458 : if (rangehi - rangelo < 9)
4840 : {
4841 0 : rangelo = ' ';
4842 0 : rangehi = 127;
4843 : }
4844 :
4845 : /*
4846 : * Now strip any common prefix of the three strings.
4847 : */
4848 13092 : while (*lobound)
4849 : {
4850 13092 : if (*lobound != *hibound || *lobound != *value)
4851 : break;
4852 6634 : lobound++, hibound++, value++;
4853 : }
4854 :
4855 : /*
4856 : * Now we can do the conversions.
4857 : */
4858 6458 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
4859 6458 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
4860 6458 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
4861 6458 : }
4862 :
4863 : static double
4864 19374 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
4865 : {
4866 19374 : int slen = strlen(value);
4867 : double num,
4868 : denom,
4869 : base;
4870 :
4871 19374 : if (slen <= 0)
4872 0 : return 0.0; /* empty string has scalar value 0 */
4873 :
4874 : /*
4875 : * There seems little point in considering more than a dozen bytes from
4876 : * the string. Since base is at least 10, that will give us nominal
4877 : * resolution of at least 12 decimal digits, which is surely far more
4878 : * precision than this estimation technique has got anyway (especially in
4879 : * non-C locales). Also, even with the maximum possible base of 256, this
4880 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
4881 : * overflow on any known machine.
4882 : */
4883 19374 : if (slen > 12)
4884 5114 : slen = 12;
4885 :
4886 : /* Convert initial characters to fraction */
4887 19374 : base = rangehi - rangelo + 1;
4888 19374 : num = 0.0;
4889 19374 : denom = base;
4890 162036 : while (slen-- > 0)
4891 : {
4892 142662 : int ch = (unsigned char) *value++;
4893 :
4894 142662 : if (ch < rangelo)
4895 208 : ch = rangelo - 1;
4896 142454 : else if (ch > rangehi)
4897 0 : ch = rangehi + 1;
4898 142662 : num += ((double) (ch - rangelo)) / denom;
4899 142662 : denom *= base;
4900 : }
4901 :
4902 19374 : return num;
4903 : }
4904 :
4905 : /*
4906 : * Convert a string-type Datum into a palloc'd, null-terminated string.
4907 : *
4908 : * On failure (e.g., unsupported typid), set *failure to true;
4909 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
4910 : *
4911 : * When using a non-C locale, we must pass the string through pg_strxfrm()
4912 : * before continuing, so as to generate correct locale-specific results.
4913 : */
4914 : static char *
4915 19374 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
4916 : {
4917 : char *val;
4918 : pg_locale_t mylocale;
4919 :
4920 19374 : switch (typid)
4921 : {
4922 0 : case CHAROID:
4923 0 : val = (char *) palloc(2);
4924 0 : val[0] = DatumGetChar(value);
4925 0 : val[1] = '\0';
4926 0 : break;
4927 6006 : case BPCHAROID:
4928 : case VARCHAROID:
4929 : case TEXTOID:
4930 6006 : val = TextDatumGetCString(value);
4931 6006 : break;
4932 13368 : case NAMEOID:
4933 : {
4934 13368 : NameData *nm = (NameData *) DatumGetPointer(value);
4935 :
4936 13368 : val = pstrdup(NameStr(*nm));
4937 13368 : break;
4938 : }
4939 0 : default:
4940 0 : *failure = true;
4941 0 : return NULL;
4942 : }
4943 :
4944 19374 : mylocale = pg_newlocale_from_collation(collid);
4945 :
4946 19374 : if (!mylocale->collate_is_c)
4947 : {
4948 : char *xfrmstr;
4949 : size_t xfrmlen;
4950 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
4951 :
4952 : /*
4953 : * XXX: We could guess at a suitable output buffer size and only call
4954 : * pg_strxfrm() twice if our guess is too small.
4955 : *
4956 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
4957 : * bogus data or set an error. This is not really a problem unless it
4958 : * crashes since it will only give an estimation error and nothing
4959 : * fatal.
4960 : *
4961 : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
4962 : * some cases, libc strxfrm() may return the wrong results, but that
4963 : * will only lead to an estimation error.
4964 : */
4965 72 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
4966 : #ifdef WIN32
4967 :
4968 : /*
4969 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
4970 : * of trying to allocate this much memory (and fail), just return the
4971 : * original string unmodified as if we were in the C locale.
4972 : */
4973 : if (xfrmlen == INT_MAX)
4974 : return val;
4975 : #endif
4976 72 : xfrmstr = (char *) palloc(xfrmlen + 1);
4977 72 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
4978 :
4979 : /*
4980 : * Some systems (e.g., glibc) can return a smaller value from the
4981 : * second call than the first; thus the Assert must be <= not ==.
4982 : */
4983 : Assert(xfrmlen2 <= xfrmlen);
4984 72 : pfree(val);
4985 72 : val = xfrmstr;
4986 : }
4987 :
4988 19374 : return val;
4989 : }
4990 :
4991 : /*
4992 : * Do convert_to_scalar()'s work for any bytea data type.
4993 : *
4994 : * Very similar to convert_string_to_scalar except we can't assume
4995 : * null-termination and therefore pass explicit lengths around.
4996 : *
4997 : * Also, assumptions about likely "normal" ranges of characters have been
4998 : * removed - a data range of 0..255 is always used, for now. (Perhaps
4999 : * someday we will add information about actual byte data range to
5000 : * pg_statistic.)
5001 : */
5002 : static void
5003 0 : convert_bytea_to_scalar(Datum value,
5004 : double *scaledvalue,
5005 : Datum lobound,
5006 : double *scaledlobound,
5007 : Datum hibound,
5008 : double *scaledhibound)
5009 : {
5010 0 : bytea *valuep = DatumGetByteaPP(value);
5011 0 : bytea *loboundp = DatumGetByteaPP(lobound);
5012 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
5013 : int rangelo,
5014 : rangehi,
5015 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
5016 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
5017 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
5018 : i,
5019 : minlen;
5020 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
5021 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
5022 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
5023 :
5024 : /*
5025 : * Assume bytea data is uniformly distributed across all byte values.
5026 : */
5027 0 : rangelo = 0;
5028 0 : rangehi = 255;
5029 :
5030 : /*
5031 : * Now strip any common prefix of the three strings.
5032 : */
5033 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
5034 0 : for (i = 0; i < minlen; i++)
5035 : {
5036 0 : if (*lostr != *histr || *lostr != *valstr)
5037 : break;
5038 0 : lostr++, histr++, valstr++;
5039 0 : loboundlen--, hiboundlen--, valuelen--;
5040 : }
5041 :
5042 : /*
5043 : * Now we can do the conversions.
5044 : */
5045 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
5046 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
5047 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
5048 0 : }
5049 :
5050 : static double
5051 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
5052 : int rangelo, int rangehi)
5053 : {
5054 : double num,
5055 : denom,
5056 : base;
5057 :
5058 0 : if (valuelen <= 0)
5059 0 : return 0.0; /* empty string has scalar value 0 */
5060 :
5061 : /*
5062 : * Since base is 256, need not consider more than about 10 chars (even
5063 : * this many seems like overkill)
5064 : */
5065 0 : if (valuelen > 10)
5066 0 : valuelen = 10;
5067 :
5068 : /* Convert initial characters to fraction */
5069 0 : base = rangehi - rangelo + 1;
5070 0 : num = 0.0;
5071 0 : denom = base;
5072 0 : while (valuelen-- > 0)
5073 : {
5074 0 : int ch = *value++;
5075 :
5076 0 : if (ch < rangelo)
5077 0 : ch = rangelo - 1;
5078 0 : else if (ch > rangehi)
5079 0 : ch = rangehi + 1;
5080 0 : num += ((double) (ch - rangelo)) / denom;
5081 0 : denom *= base;
5082 : }
5083 :
5084 0 : return num;
5085 : }
5086 :
5087 : /*
5088 : * Do convert_to_scalar()'s work for any timevalue data type.
5089 : *
5090 : * On failure (e.g., unsupported typid), set *failure to true;
5091 : * otherwise, that variable is not changed.
5092 : */
5093 : static double
5094 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5095 : {
5096 0 : switch (typid)
5097 : {
5098 0 : case TIMESTAMPOID:
5099 0 : return DatumGetTimestamp(value);
5100 0 : case TIMESTAMPTZOID:
5101 0 : return DatumGetTimestampTz(value);
5102 0 : case DATEOID:
5103 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
5104 0 : case INTERVALOID:
5105 : {
5106 0 : Interval *interval = DatumGetIntervalP(value);
5107 :
5108 : /*
5109 : * Convert the month part of Interval to days using assumed
5110 : * average month length of 365.25/12.0 days. Not too
5111 : * accurate, but plenty good enough for our purposes.
5112 : *
5113 : * This also works for infinite intervals, which just have all
5114 : * fields set to INT_MIN/INT_MAX, and so will produce a result
5115 : * smaller/larger than any finite interval.
5116 : */
5117 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5118 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5119 : }
5120 0 : case TIMEOID:
5121 0 : return DatumGetTimeADT(value);
5122 0 : case TIMETZOID:
5123 : {
5124 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5125 :
5126 : /* use GMT-equivalent time */
5127 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5128 : }
5129 : }
5130 :
5131 0 : *failure = true;
5132 0 : return 0;
5133 : }
5134 :
5135 :
5136 : /*
5137 : * get_restriction_variable
5138 : * Examine the args of a restriction clause to see if it's of the
5139 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5140 : * where "variable" could be either a Var or an expression in vars of a
5141 : * single relation. If so, extract information about the variable,
5142 : * and also indicate which side it was on and the other argument.
5143 : *
5144 : * Inputs:
5145 : * root: the planner info
5146 : * args: clause argument list
5147 : * varRelid: see specs for restriction selectivity functions
5148 : *
5149 : * Outputs: (these are valid only if true is returned)
5150 : * *vardata: gets information about variable (see examine_variable)
5151 : * *other: gets other clause argument, aggressively reduced to a constant
5152 : * *varonleft: set true if variable is on the left, false if on the right
5153 : *
5154 : * Returns true if a variable is identified, otherwise false.
5155 : *
5156 : * Note: if there are Vars on both sides of the clause, we must fail, because
5157 : * callers are expecting that the other side will act like a pseudoconstant.
5158 : */
5159 : bool
5160 776314 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5161 : VariableStatData *vardata, Node **other,
5162 : bool *varonleft)
5163 : {
5164 : Node *left,
5165 : *right;
5166 : VariableStatData rdata;
5167 :
5168 : /* Fail if not a binary opclause (probably shouldn't happen) */
5169 776314 : if (list_length(args) != 2)
5170 0 : return false;
5171 :
5172 776314 : left = (Node *) linitial(args);
5173 776314 : right = (Node *) lsecond(args);
5174 :
5175 : /*
5176 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5177 : * relations will be treated as pseudoconstants.
5178 : */
5179 776314 : examine_variable(root, left, varRelid, vardata);
5180 776314 : examine_variable(root, right, varRelid, &rdata);
5181 :
5182 : /*
5183 : * If one side is a variable and the other not, we win.
5184 : */
5185 776314 : if (vardata->rel && rdata.rel == NULL)
5186 : {
5187 698042 : *varonleft = true;
5188 698042 : *other = estimate_expression_value(root, rdata.var);
5189 : /* Assume we need no ReleaseVariableStats(rdata) here */
5190 698036 : return true;
5191 : }
5192 :
5193 78272 : if (vardata->rel == NULL && rdata.rel)
5194 : {
5195 71910 : *varonleft = false;
5196 71910 : *other = estimate_expression_value(root, vardata->var);
5197 : /* Assume we need no ReleaseVariableStats(*vardata) here */
5198 71910 : *vardata = rdata;
5199 71910 : return true;
5200 : }
5201 :
5202 : /* Oops, clause has wrong structure (probably var op var) */
5203 6362 : ReleaseVariableStats(*vardata);
5204 6362 : ReleaseVariableStats(rdata);
5205 :
5206 6362 : return false;
5207 : }
5208 :
5209 : /*
5210 : * get_join_variables
5211 : * Apply examine_variable() to each side of a join clause.
5212 : * Also, attempt to identify whether the join clause has the same
5213 : * or reversed sense compared to the SpecialJoinInfo.
5214 : *
5215 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5216 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5217 : * where we can't tell for sure, we default to assuming it's normal.
5218 : */
5219 : void
5220 234334 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5221 : VariableStatData *vardata1, VariableStatData *vardata2,
5222 : bool *join_is_reversed)
5223 : {
5224 : Node *left,
5225 : *right;
5226 :
5227 234334 : if (list_length(args) != 2)
5228 0 : elog(ERROR, "join operator should take two arguments");
5229 :
5230 234334 : left = (Node *) linitial(args);
5231 234334 : right = (Node *) lsecond(args);
5232 :
5233 234334 : examine_variable(root, left, 0, vardata1);
5234 234334 : examine_variable(root, right, 0, vardata2);
5235 :
5236 468494 : if (vardata1->rel &&
5237 234160 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
5238 84912 : *join_is_reversed = true; /* var1 is on RHS */
5239 298704 : else if (vardata2->rel &&
5240 149282 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
5241 126 : *join_is_reversed = true; /* var2 is on LHS */
5242 : else
5243 149296 : *join_is_reversed = false;
5244 234334 : }
5245 :
5246 : /* statext_expressions_load copies the tuple, so just pfree it. */
5247 : static void
5248 1644 : ReleaseDummy(HeapTuple tuple)
5249 : {
5250 1644 : pfree(tuple);
5251 1644 : }
5252 :
5253 : /*
5254 : * examine_variable
5255 : * Try to look up statistical data about an expression.
5256 : * Fill in a VariableStatData struct to describe the expression.
5257 : *
5258 : * Inputs:
5259 : * root: the planner info
5260 : * node: the expression tree to examine
5261 : * varRelid: see specs for restriction selectivity functions
5262 : *
5263 : * Outputs: *vardata is filled as follows:
5264 : * var: the input expression (with any binary relabeling stripped, if
5265 : * it is or contains a variable; but otherwise the type is preserved)
5266 : * rel: RelOptInfo for relation containing variable; NULL if expression
5267 : * contains no Vars (NOTE this could point to a RelOptInfo of a
5268 : * subquery, not one in the current query).
5269 : * statsTuple: the pg_statistic entry for the variable, if one exists;
5270 : * otherwise NULL.
5271 : * freefunc: pointer to a function to release statsTuple with.
5272 : * vartype: exposed type of the expression; this should always match
5273 : * the declared input type of the operator we are estimating for.
5274 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5275 : * commonly the same as the exposed type of the variable argument,
5276 : * but can be different in binary-compatible-type cases.
5277 : * isunique: true if we were able to match the var to a unique index, a
5278 : * single-column DISTINCT or GROUP-BY clause, implying its values are
5279 : * unique for this query. (Caution: this should be trusted for
5280 : * statistical purposes only, since we do not check indimmediate nor
5281 : * verify that the exact same definition of equality applies.)
5282 : * acl_ok: true if current user has permission to read the column(s)
5283 : * underlying the pg_statistic entry. This is consulted by
5284 : * statistic_proc_security_check().
5285 : *
5286 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5287 : */
5288 : void
5289 3007144 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5290 : VariableStatData *vardata)
5291 : {
5292 : Node *basenode;
5293 : Relids varnos;
5294 : Relids basevarnos;
5295 : RelOptInfo *onerel;
5296 :
5297 : /* Make sure we don't return dangling pointers in vardata */
5298 21050008 : MemSet(vardata, 0, sizeof(VariableStatData));
5299 :
5300 : /* Save the exposed type of the expression */
5301 3007144 : vardata->vartype = exprType(node);
5302 :
5303 : /* Look inside any binary-compatible relabeling */
5304 :
5305 3007144 : if (IsA(node, RelabelType))
5306 40814 : basenode = (Node *) ((RelabelType *) node)->arg;
5307 : else
5308 2966330 : basenode = node;
5309 :
5310 : /* Fast path for a simple Var */
5311 :
5312 3007144 : if (IsA(basenode, Var) &&
5313 703706 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5314 : {
5315 2127012 : Var *var = (Var *) basenode;
5316 :
5317 : /* Set up result fields other than the stats tuple */
5318 2127012 : vardata->var = basenode; /* return Var without relabeling */
5319 2127012 : vardata->rel = find_base_rel(root, var->varno);
5320 2127012 : vardata->atttype = var->vartype;
5321 2127012 : vardata->atttypmod = var->vartypmod;
5322 2127012 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5323 :
5324 : /* Try to locate some stats */
5325 2127012 : examine_simple_variable(root, var, vardata);
5326 :
5327 2127012 : return;
5328 : }
5329 :
5330 : /*
5331 : * Okay, it's a more complicated expression. Determine variable
5332 : * membership. Note that when varRelid isn't zero, only vars of that
5333 : * relation are considered "real" vars.
5334 : */
5335 880132 : varnos = pull_varnos(root, basenode);
5336 880132 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5337 :
5338 880132 : onerel = NULL;
5339 :
5340 880132 : if (bms_is_empty(basevarnos))
5341 : {
5342 : /* No Vars at all ... must be pseudo-constant clause */
5343 : }
5344 : else
5345 : {
5346 : int relid;
5347 :
5348 : /* Check if the expression is in vars of a single base relation */
5349 417924 : if (bms_get_singleton_member(basevarnos, &relid))
5350 : {
5351 413922 : if (varRelid == 0 || varRelid == relid)
5352 : {
5353 60816 : onerel = find_base_rel(root, relid);
5354 60816 : vardata->rel = onerel;
5355 60816 : node = basenode; /* strip any relabeling */
5356 : }
5357 : /* else treat it as a constant */
5358 : }
5359 : else
5360 : {
5361 : /* varnos has multiple relids */
5362 4002 : if (varRelid == 0)
5363 : {
5364 : /* treat it as a variable of a join relation */
5365 3676 : vardata->rel = find_join_rel(root, varnos);
5366 3676 : node = basenode; /* strip any relabeling */
5367 : }
5368 326 : else if (bms_is_member(varRelid, varnos))
5369 : {
5370 : /* ignore the vars belonging to other relations */
5371 164 : vardata->rel = find_base_rel(root, varRelid);
5372 164 : node = basenode; /* strip any relabeling */
5373 : /* note: no point in expressional-index search here */
5374 : }
5375 : /* else treat it as a constant */
5376 : }
5377 : }
5378 :
5379 880132 : bms_free(basevarnos);
5380 :
5381 880132 : vardata->var = node;
5382 880132 : vardata->atttype = exprType(node);
5383 880132 : vardata->atttypmod = exprTypmod(node);
5384 :
5385 880132 : if (onerel)
5386 : {
5387 : /*
5388 : * We have an expression in vars of a single relation. Try to match
5389 : * it to expressional index columns, in hopes of finding some
5390 : * statistics.
5391 : *
5392 : * Note that we consider all index columns including INCLUDE columns,
5393 : * since there could be stats for such columns. But the test for
5394 : * uniqueness needs to be warier.
5395 : *
5396 : * XXX it's conceivable that there are multiple matches with different
5397 : * index opfamilies; if so, we need to pick one that matches the
5398 : * operator we are estimating for. FIXME later.
5399 : */
5400 : ListCell *ilist;
5401 : ListCell *slist;
5402 : Oid userid;
5403 :
5404 : /*
5405 : * The nullingrels bits within the expression could prevent us from
5406 : * matching it to expressional index columns or to the expressions in
5407 : * extended statistics. So strip them out first.
5408 : */
5409 60816 : if (bms_overlap(varnos, root->outer_join_rels))
5410 3044 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5411 :
5412 : /*
5413 : * Determine the user ID to use for privilege checks: either
5414 : * onerel->userid if it's set (e.g., in case we're accessing the table
5415 : * via a view), or the current user otherwise.
5416 : *
5417 : * If we drill down to child relations, we keep using the same userid:
5418 : * it's going to be the same anyway, due to how we set up the relation
5419 : * tree (q.v. build_simple_rel).
5420 : */
5421 60816 : userid = OidIsValid(onerel->userid) ? onerel->userid : GetUserId();
5422 :
5423 122912 : foreach(ilist, onerel->indexlist)
5424 : {
5425 64910 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5426 : ListCell *indexpr_item;
5427 : int pos;
5428 :
5429 64910 : indexpr_item = list_head(index->indexprs);
5430 64910 : if (indexpr_item == NULL)
5431 60164 : continue; /* no expressions here... */
5432 :
5433 6750 : for (pos = 0; pos < index->ncolumns; pos++)
5434 : {
5435 4818 : if (index->indexkeys[pos] == 0)
5436 : {
5437 : Node *indexkey;
5438 :
5439 4746 : if (indexpr_item == NULL)
5440 0 : elog(ERROR, "too few entries in indexprs list");
5441 4746 : indexkey = (Node *) lfirst(indexpr_item);
5442 4746 : if (indexkey && IsA(indexkey, RelabelType))
5443 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5444 4746 : if (equal(node, indexkey))
5445 : {
5446 : /*
5447 : * Found a match ... is it a unique index? Tests here
5448 : * should match has_unique_index().
5449 : */
5450 3450 : if (index->unique &&
5451 438 : index->nkeycolumns == 1 &&
5452 438 : pos == 0 &&
5453 438 : (index->indpred == NIL || index->predOK))
5454 438 : vardata->isunique = true;
5455 :
5456 : /*
5457 : * Has it got stats? We only consider stats for
5458 : * non-partial indexes, since partial indexes probably
5459 : * don't reflect whole-relation statistics; the above
5460 : * check for uniqueness is the only info we take from
5461 : * a partial index.
5462 : *
5463 : * An index stats hook, however, must make its own
5464 : * decisions about what to do with partial indexes.
5465 : */
5466 3450 : if (get_index_stats_hook &&
5467 0 : (*get_index_stats_hook) (root, index->indexoid,
5468 0 : pos + 1, vardata))
5469 : {
5470 : /*
5471 : * The hook took control of acquiring a stats
5472 : * tuple. If it did supply a tuple, it'd better
5473 : * have supplied a freefunc.
5474 : */
5475 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5476 0 : !vardata->freefunc)
5477 0 : elog(ERROR, "no function provided to release variable stats with");
5478 : }
5479 3450 : else if (index->indpred == NIL)
5480 : {
5481 3450 : vardata->statsTuple =
5482 6900 : SearchSysCache3(STATRELATTINH,
5483 : ObjectIdGetDatum(index->indexoid),
5484 3450 : Int16GetDatum(pos + 1),
5485 : BoolGetDatum(false));
5486 3450 : vardata->freefunc = ReleaseSysCache;
5487 :
5488 3450 : if (HeapTupleIsValid(vardata->statsTuple))
5489 : {
5490 : /* Get index's table for permission check */
5491 : RangeTblEntry *rte;
5492 :
5493 2814 : rte = planner_rt_fetch(index->rel->relid, root);
5494 : Assert(rte->rtekind == RTE_RELATION);
5495 :
5496 : /*
5497 : * For simplicity, we insist on the whole
5498 : * table being selectable, rather than trying
5499 : * to identify which column(s) the index
5500 : * depends on. Also require all rows to be
5501 : * selectable --- there must be no
5502 : * securityQuals from security barrier views
5503 : * or RLS policies.
5504 : */
5505 2814 : vardata->acl_ok =
5506 5628 : rte->securityQuals == NIL &&
5507 2814 : (pg_class_aclcheck(rte->relid, userid,
5508 : ACL_SELECT) == ACLCHECK_OK);
5509 :
5510 : /*
5511 : * If the user doesn't have permissions to
5512 : * access an inheritance child relation, check
5513 : * the permissions of the table actually
5514 : * mentioned in the query, since most likely
5515 : * the user does have that permission. Note
5516 : * that whole-table select privilege on the
5517 : * parent doesn't quite guarantee that the
5518 : * user could read all columns of the child.
5519 : * But in practice it's unlikely that any
5520 : * interesting security violation could result
5521 : * from allowing access to the expression
5522 : * index's stats, so we allow it anyway. See
5523 : * similar code in examine_simple_variable()
5524 : * for additional comments.
5525 : */
5526 2814 : if (!vardata->acl_ok &&
5527 18 : root->append_rel_array != NULL)
5528 : {
5529 : AppendRelInfo *appinfo;
5530 12 : Index varno = index->rel->relid;
5531 :
5532 12 : appinfo = root->append_rel_array[varno];
5533 36 : while (appinfo &&
5534 24 : planner_rt_fetch(appinfo->parent_relid,
5535 24 : root)->rtekind == RTE_RELATION)
5536 : {
5537 24 : varno = appinfo->parent_relid;
5538 24 : appinfo = root->append_rel_array[varno];
5539 : }
5540 12 : if (varno != index->rel->relid)
5541 : {
5542 : /* Repeat access check on this rel */
5543 12 : rte = planner_rt_fetch(varno, root);
5544 : Assert(rte->rtekind == RTE_RELATION);
5545 :
5546 12 : vardata->acl_ok =
5547 24 : rte->securityQuals == NIL &&
5548 12 : (pg_class_aclcheck(rte->relid,
5549 : userid,
5550 : ACL_SELECT) == ACLCHECK_OK);
5551 : }
5552 : }
5553 : }
5554 : else
5555 : {
5556 : /* suppress leakproofness checks later */
5557 636 : vardata->acl_ok = true;
5558 : }
5559 : }
5560 3450 : if (vardata->statsTuple)
5561 2814 : break;
5562 : }
5563 1932 : indexpr_item = lnext(index->indexprs, indexpr_item);
5564 : }
5565 : }
5566 4746 : if (vardata->statsTuple)
5567 2814 : break;
5568 : }
5569 :
5570 : /*
5571 : * Search extended statistics for one with a matching expression.
5572 : * There might be multiple ones, so just grab the first one. In the
5573 : * future, we might consider the statistics target (and pick the most
5574 : * accurate statistics) and maybe some other parameters.
5575 : */
5576 64842 : foreach(slist, onerel->statlist)
5577 : {
5578 4314 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5579 4314 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5580 : ListCell *expr_item;
5581 : int pos;
5582 :
5583 : /*
5584 : * Stop once we've found statistics for the expression (either
5585 : * from extended stats, or for an index in the preceding loop).
5586 : */
5587 4314 : if (vardata->statsTuple)
5588 288 : break;
5589 :
5590 : /* skip stats without per-expression stats */
5591 4026 : if (info->kind != STATS_EXT_EXPRESSIONS)
5592 2022 : continue;
5593 :
5594 : /* skip stats with mismatching stxdinherit value */
5595 2004 : if (info->inherit != rte->inh)
5596 6 : continue;
5597 :
5598 1998 : pos = 0;
5599 3300 : foreach(expr_item, info->exprs)
5600 : {
5601 2946 : Node *expr = (Node *) lfirst(expr_item);
5602 :
5603 : Assert(expr);
5604 :
5605 : /* strip RelabelType before comparing it */
5606 2946 : if (expr && IsA(expr, RelabelType))
5607 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5608 :
5609 : /* found a match, see if we can extract pg_statistic row */
5610 2946 : if (equal(node, expr))
5611 : {
5612 : /*
5613 : * XXX Not sure if we should cache the tuple somewhere.
5614 : * Now we just create a new copy every time.
5615 : */
5616 1644 : vardata->statsTuple =
5617 1644 : statext_expressions_load(info->statOid, rte->inh, pos);
5618 :
5619 1644 : vardata->freefunc = ReleaseDummy;
5620 :
5621 : /*
5622 : * For simplicity, we insist on the whole table being
5623 : * selectable, rather than trying to identify which
5624 : * column(s) the statistics object depends on. Also
5625 : * require all rows to be selectable --- there must be no
5626 : * securityQuals from security barrier views or RLS
5627 : * policies.
5628 : */
5629 1644 : vardata->acl_ok =
5630 3288 : rte->securityQuals == NIL &&
5631 1644 : (pg_class_aclcheck(rte->relid, userid,
5632 : ACL_SELECT) == ACLCHECK_OK);
5633 :
5634 : /*
5635 : * If the user doesn't have permissions to access an
5636 : * inheritance child relation, check the permissions of
5637 : * the table actually mentioned in the query, since most
5638 : * likely the user does have that permission. Note that
5639 : * whole-table select privilege on the parent doesn't
5640 : * quite guarantee that the user could read all columns of
5641 : * the child. But in practice it's unlikely that any
5642 : * interesting security violation could result from
5643 : * allowing access to the expression stats, so we allow it
5644 : * anyway. See similar code in examine_simple_variable()
5645 : * for additional comments.
5646 : */
5647 1644 : if (!vardata->acl_ok &&
5648 0 : root->append_rel_array != NULL)
5649 : {
5650 : AppendRelInfo *appinfo;
5651 0 : Index varno = onerel->relid;
5652 :
5653 0 : appinfo = root->append_rel_array[varno];
5654 0 : while (appinfo &&
5655 0 : planner_rt_fetch(appinfo->parent_relid,
5656 0 : root)->rtekind == RTE_RELATION)
5657 : {
5658 0 : varno = appinfo->parent_relid;
5659 0 : appinfo = root->append_rel_array[varno];
5660 : }
5661 0 : if (varno != onerel->relid)
5662 : {
5663 : /* Repeat access check on this rel */
5664 0 : rte = planner_rt_fetch(varno, root);
5665 : Assert(rte->rtekind == RTE_RELATION);
5666 :
5667 0 : vardata->acl_ok =
5668 0 : rte->securityQuals == NIL &&
5669 0 : (pg_class_aclcheck(rte->relid,
5670 : userid,
5671 : ACL_SELECT) == ACLCHECK_OK);
5672 : }
5673 : }
5674 :
5675 1644 : break;
5676 : }
5677 :
5678 1302 : pos++;
5679 : }
5680 : }
5681 : }
5682 :
5683 880132 : bms_free(varnos);
5684 : }
5685 :
5686 : /*
5687 : * examine_simple_variable
5688 : * Handle a simple Var for examine_variable
5689 : *
5690 : * This is split out as a subroutine so that we can recurse to deal with
5691 : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
5692 : *
5693 : * We already filled in all the fields of *vardata except for the stats tuple.
5694 : */
5695 : static void
5696 2132716 : examine_simple_variable(PlannerInfo *root, Var *var,
5697 : VariableStatData *vardata)
5698 : {
5699 2132716 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
5700 :
5701 : Assert(IsA(rte, RangeTblEntry));
5702 :
5703 2132716 : if (get_relation_stats_hook &&
5704 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
5705 : {
5706 : /*
5707 : * The hook took control of acquiring a stats tuple. If it did supply
5708 : * a tuple, it'd better have supplied a freefunc.
5709 : */
5710 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5711 0 : !vardata->freefunc)
5712 0 : elog(ERROR, "no function provided to release variable stats with");
5713 : }
5714 2132716 : else if (rte->rtekind == RTE_RELATION)
5715 : {
5716 : /*
5717 : * Plain table or parent of an inheritance appendrel, so look up the
5718 : * column in pg_statistic
5719 : */
5720 2026838 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
5721 : ObjectIdGetDatum(rte->relid),
5722 2026838 : Int16GetDatum(var->varattno),
5723 2026838 : BoolGetDatum(rte->inh));
5724 2026838 : vardata->freefunc = ReleaseSysCache;
5725 :
5726 2026838 : if (HeapTupleIsValid(vardata->statsTuple))
5727 : {
5728 1531734 : RelOptInfo *onerel = find_base_rel_noerr(root, var->varno);
5729 : Oid userid;
5730 :
5731 : /*
5732 : * Check if user has permission to read this column. We require
5733 : * all rows to be accessible, so there must be no securityQuals
5734 : * from security barrier views or RLS policies.
5735 : *
5736 : * Normally the Var will have an associated RelOptInfo from which
5737 : * we can find out which userid to do the check as; but it might
5738 : * not if it's a RETURNING Var for an INSERT target relation. In
5739 : * that case use the RTEPermissionInfo associated with the RTE.
5740 : */
5741 1531734 : if (onerel)
5742 1531692 : userid = onerel->userid;
5743 : else
5744 : {
5745 : RTEPermissionInfo *perminfo;
5746 :
5747 42 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
5748 42 : userid = perminfo->checkAsUser;
5749 : }
5750 1531734 : if (!OidIsValid(userid))
5751 1399388 : userid = GetUserId();
5752 :
5753 1531734 : vardata->acl_ok =
5754 3063728 : rte->securityQuals == NIL &&
5755 1531734 : ((pg_class_aclcheck(rte->relid, userid,
5756 260 : ACL_SELECT) == ACLCHECK_OK) ||
5757 260 : (pg_attribute_aclcheck(rte->relid, var->varattno, userid,
5758 : ACL_SELECT) == ACLCHECK_OK));
5759 :
5760 : /*
5761 : * If the user doesn't have permissions to access an inheritance
5762 : * child relation or specifically this attribute, check the
5763 : * permissions of the table/column actually mentioned in the
5764 : * query, since most likely the user does have that permission
5765 : * (else the query will fail at runtime), and if the user can read
5766 : * the column there then he can get the values of the child table
5767 : * too. To do that, we must find out which of the root parent's
5768 : * attributes the child relation's attribute corresponds to.
5769 : */
5770 1531734 : if (!vardata->acl_ok && var->varattno > 0 &&
5771 60 : root->append_rel_array != NULL)
5772 : {
5773 : AppendRelInfo *appinfo;
5774 12 : Index varno = var->varno;
5775 12 : int varattno = var->varattno;
5776 12 : bool found = false;
5777 :
5778 12 : appinfo = root->append_rel_array[varno];
5779 :
5780 : /*
5781 : * Partitions are mapped to their immediate parent, not the
5782 : * root parent, so must be ready to walk up multiple
5783 : * AppendRelInfos. But stop if we hit a parent that is not
5784 : * RTE_RELATION --- that's a flattened UNION ALL subquery, not
5785 : * an inheritance parent.
5786 : */
5787 36 : while (appinfo &&
5788 24 : planner_rt_fetch(appinfo->parent_relid,
5789 24 : root)->rtekind == RTE_RELATION)
5790 : {
5791 : int parent_varattno;
5792 :
5793 24 : found = false;
5794 24 : if (varattno <= 0 || varattno > appinfo->num_child_cols)
5795 : break; /* safety check */
5796 24 : parent_varattno = appinfo->parent_colnos[varattno - 1];
5797 24 : if (parent_varattno == 0)
5798 0 : break; /* Var is local to child */
5799 :
5800 24 : varno = appinfo->parent_relid;
5801 24 : varattno = parent_varattno;
5802 24 : found = true;
5803 :
5804 : /* If the parent is itself a child, continue up. */
5805 24 : appinfo = root->append_rel_array[varno];
5806 : }
5807 :
5808 : /*
5809 : * In rare cases, the Var may be local to the child table, in
5810 : * which case, we've got to live with having no access to this
5811 : * column's stats.
5812 : */
5813 12 : if (!found)
5814 0 : return;
5815 :
5816 : /* Repeat the access check on this parent rel & column */
5817 12 : rte = planner_rt_fetch(varno, root);
5818 : Assert(rte->rtekind == RTE_RELATION);
5819 :
5820 : /*
5821 : * Fine to use the same userid as it's the same in all
5822 : * relations of a given inheritance tree.
5823 : */
5824 12 : vardata->acl_ok =
5825 30 : rte->securityQuals == NIL &&
5826 12 : ((pg_class_aclcheck(rte->relid, userid,
5827 6 : ACL_SELECT) == ACLCHECK_OK) ||
5828 6 : (pg_attribute_aclcheck(rte->relid, varattno, userid,
5829 : ACL_SELECT) == ACLCHECK_OK));
5830 : }
5831 : }
5832 : else
5833 : {
5834 : /* suppress any possible leakproofness checks later */
5835 495104 : vardata->acl_ok = true;
5836 : }
5837 : }
5838 105878 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
5839 94998 : (rte->rtekind == RTE_CTE && !rte->self_reference))
5840 : {
5841 : /*
5842 : * Plain subquery (not one that was converted to an appendrel) or
5843 : * non-recursive CTE. In either case, we can try to find out what the
5844 : * Var refers to within the subquery. We skip this for appendrel and
5845 : * recursive-CTE cases because any column stats we did find would
5846 : * likely not be very relevant.
5847 : */
5848 : PlannerInfo *subroot;
5849 : Query *subquery;
5850 : List *subtlist;
5851 : TargetEntry *ste;
5852 :
5853 : /*
5854 : * Punt if it's a whole-row var rather than a plain column reference.
5855 : */
5856 17358 : if (var->varattno == InvalidAttrNumber)
5857 0 : return;
5858 :
5859 : /*
5860 : * Otherwise, find the subquery's planner subroot.
5861 : */
5862 17358 : if (rte->rtekind == RTE_SUBQUERY)
5863 : {
5864 : RelOptInfo *rel;
5865 :
5866 : /*
5867 : * Fetch RelOptInfo for subquery. Note that we don't change the
5868 : * rel returned in vardata, since caller expects it to be a rel of
5869 : * the caller's query level. Because we might already be
5870 : * recursing, we can't use that rel pointer either, but have to
5871 : * look up the Var's rel afresh.
5872 : */
5873 10880 : rel = find_base_rel(root, var->varno);
5874 :
5875 10880 : subroot = rel->subroot;
5876 : }
5877 : else
5878 : {
5879 : /* CTE case is more difficult */
5880 : PlannerInfo *cteroot;
5881 : Index levelsup;
5882 : int ndx;
5883 : int plan_id;
5884 : ListCell *lc;
5885 :
5886 : /*
5887 : * Find the referenced CTE, and locate the subroot previously made
5888 : * for it.
5889 : */
5890 6478 : levelsup = rte->ctelevelsup;
5891 6478 : cteroot = root;
5892 12086 : while (levelsup-- > 0)
5893 : {
5894 5608 : cteroot = cteroot->parent_root;
5895 5608 : if (!cteroot) /* shouldn't happen */
5896 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
5897 : }
5898 :
5899 : /*
5900 : * Note: cte_plan_ids can be shorter than cteList, if we are still
5901 : * working on planning the CTEs (ie, this is a side-reference from
5902 : * another CTE). So we mustn't use forboth here.
5903 : */
5904 6478 : ndx = 0;
5905 8700 : foreach(lc, cteroot->parse->cteList)
5906 : {
5907 8700 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
5908 :
5909 8700 : if (strcmp(cte->ctename, rte->ctename) == 0)
5910 6478 : break;
5911 2222 : ndx++;
5912 : }
5913 6478 : if (lc == NULL) /* shouldn't happen */
5914 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
5915 6478 : if (ndx >= list_length(cteroot->cte_plan_ids))
5916 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
5917 6478 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
5918 6478 : if (plan_id <= 0)
5919 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
5920 6478 : subroot = list_nth(root->glob->subroots, plan_id - 1);
5921 : }
5922 :
5923 : /* If the subquery hasn't been planned yet, we have to punt */
5924 17358 : if (subroot == NULL)
5925 0 : return;
5926 : Assert(IsA(subroot, PlannerInfo));
5927 :
5928 : /*
5929 : * We must use the subquery parsetree as mangled by the planner, not
5930 : * the raw version from the RTE, because we need a Var that will refer
5931 : * to the subroot's live RelOptInfos. For instance, if any subquery
5932 : * pullup happened during planning, Vars in the targetlist might have
5933 : * gotten replaced, and we need to see the replacement expressions.
5934 : */
5935 17358 : subquery = subroot->parse;
5936 : Assert(IsA(subquery, Query));
5937 :
5938 : /*
5939 : * Punt if subquery uses set operations or grouping sets, as these
5940 : * will mash underlying columns' stats beyond recognition. (Set ops
5941 : * are particularly nasty; if we forged ahead, we would return stats
5942 : * relevant to only the leftmost subselect...) DISTINCT is also
5943 : * problematic, but we check that later because there is a possibility
5944 : * of learning something even with it.
5945 : */
5946 17358 : if (subquery->setOperations ||
5947 15550 : subquery->groupingSets)
5948 1832 : return;
5949 :
5950 : /* Get the subquery output expression referenced by the upper Var */
5951 15526 : if (subquery->returningList)
5952 206 : subtlist = subquery->returningList;
5953 : else
5954 15320 : subtlist = subquery->targetList;
5955 15526 : ste = get_tle_by_resno(subtlist, var->varattno);
5956 15526 : if (ste == NULL || ste->resjunk)
5957 0 : elog(ERROR, "subquery %s does not have attribute %d",
5958 : rte->eref->aliasname, var->varattno);
5959 15526 : var = (Var *) ste->expr;
5960 :
5961 : /*
5962 : * If subquery uses DISTINCT, we can't make use of any stats for the
5963 : * variable ... but, if it's the only DISTINCT column, we are entitled
5964 : * to consider it unique. We do the test this way so that it works
5965 : * for cases involving DISTINCT ON.
5966 : */
5967 15526 : if (subquery->distinctClause)
5968 : {
5969 1766 : if (list_length(subquery->distinctClause) == 1 &&
5970 580 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
5971 272 : vardata->isunique = true;
5972 : /* cannot go further */
5973 1186 : return;
5974 : }
5975 :
5976 : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
5977 14340 : if (subquery->groupClause)
5978 : {
5979 1040 : if (list_length(subquery->groupClause) == 1 &&
5980 430 : targetIsInSortList(ste, InvalidOid, subquery->groupClause))
5981 334 : vardata->isunique = true;
5982 : /* cannot go further */
5983 610 : return;
5984 : }
5985 :
5986 : /*
5987 : * If the sub-query originated from a view with the security_barrier
5988 : * attribute, we must not look at the variable's statistics, though it
5989 : * seems all right to notice the existence of a DISTINCT clause. So
5990 : * stop here.
5991 : *
5992 : * This is probably a harsher restriction than necessary; it's
5993 : * certainly OK for the selectivity estimator (which is a C function,
5994 : * and therefore omnipotent anyway) to look at the statistics. But
5995 : * many selectivity estimators will happily *invoke the operator
5996 : * function* to try to work out a good estimate - and that's not OK.
5997 : * So for now, don't dig down for stats.
5998 : */
5999 13730 : if (rte->security_barrier)
6000 2874 : return;
6001 :
6002 : /* Can only handle a simple Var of subquery's query level */
6003 10856 : if (var && IsA(var, Var) &&
6004 5704 : var->varlevelsup == 0)
6005 : {
6006 : /*
6007 : * OK, recurse into the subquery. Note that the original setting
6008 : * of vardata->isunique (which will surely be false) is left
6009 : * unchanged in this situation. That's what we want, since even
6010 : * if the underlying column is unique, the subquery may have
6011 : * joined to other tables in a way that creates duplicates.
6012 : */
6013 5704 : examine_simple_variable(subroot, var, vardata);
6014 : }
6015 : }
6016 : else
6017 : {
6018 : /*
6019 : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
6020 : * see RTE_JOIN here because join alias Vars have already been
6021 : * flattened.) There's not much we can do with function outputs, but
6022 : * maybe someday try to be smarter about VALUES.
6023 : */
6024 : }
6025 : }
6026 :
6027 : /*
6028 : * examine_indexcol_variable
6029 : * Try to look up statistical data about an index column/expression.
6030 : * Fill in a VariableStatData struct to describe the column.
6031 : *
6032 : * Inputs:
6033 : * root: the planner info
6034 : * index: the index whose column we're interested in
6035 : * indexcol: 0-based index column number (subscripts index->indexkeys[])
6036 : *
6037 : * Outputs: *vardata is filled as follows:
6038 : * var: the input expression (with any binary relabeling stripped, if
6039 : * it is or contains a variable; but otherwise the type is preserved)
6040 : * rel: RelOptInfo for table relation containing variable.
6041 : * statsTuple: the pg_statistic entry for the variable, if one exists;
6042 : * otherwise NULL.
6043 : * freefunc: pointer to a function to release statsTuple with.
6044 : *
6045 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
6046 : */
6047 : static void
6048 771592 : examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
6049 : int indexcol, VariableStatData *vardata)
6050 : {
6051 : AttrNumber colnum;
6052 : Oid relid;
6053 :
6054 771592 : if (index->indexkeys[indexcol] != 0)
6055 : {
6056 : /* Simple variable --- look to stats for the underlying table */
6057 769402 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6058 :
6059 : Assert(rte->rtekind == RTE_RELATION);
6060 769402 : relid = rte->relid;
6061 : Assert(relid != InvalidOid);
6062 769402 : colnum = index->indexkeys[indexcol];
6063 769402 : vardata->rel = index->rel;
6064 :
6065 769402 : if (get_relation_stats_hook &&
6066 0 : (*get_relation_stats_hook) (root, rte, colnum, vardata))
6067 : {
6068 : /*
6069 : * The hook took control of acquiring a stats tuple. If it did
6070 : * supply a tuple, it'd better have supplied a freefunc.
6071 : */
6072 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6073 0 : !vardata->freefunc)
6074 0 : elog(ERROR, "no function provided to release variable stats with");
6075 : }
6076 : else
6077 : {
6078 769402 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6079 : ObjectIdGetDatum(relid),
6080 : Int16GetDatum(colnum),
6081 769402 : BoolGetDatum(rte->inh));
6082 769402 : vardata->freefunc = ReleaseSysCache;
6083 : }
6084 : }
6085 : else
6086 : {
6087 : /* Expression --- maybe there are stats for the index itself */
6088 2190 : relid = index->indexoid;
6089 2190 : colnum = indexcol + 1;
6090 :
6091 2190 : if (get_index_stats_hook &&
6092 0 : (*get_index_stats_hook) (root, relid, colnum, vardata))
6093 : {
6094 : /*
6095 : * The hook took control of acquiring a stats tuple. If it did
6096 : * supply a tuple, it'd better have supplied a freefunc.
6097 : */
6098 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6099 0 : !vardata->freefunc)
6100 0 : elog(ERROR, "no function provided to release variable stats with");
6101 : }
6102 : else
6103 : {
6104 2190 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6105 : ObjectIdGetDatum(relid),
6106 : Int16GetDatum(colnum),
6107 : BoolGetDatum(false));
6108 2190 : vardata->freefunc = ReleaseSysCache;
6109 : }
6110 : }
6111 771592 : }
6112 :
6113 : /*
6114 : * Check whether it is permitted to call func_oid passing some of the
6115 : * pg_statistic data in vardata. We allow this either if the user has SELECT
6116 : * privileges on the table or column underlying the pg_statistic data or if
6117 : * the function is marked leakproof.
6118 : */
6119 : bool
6120 995920 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
6121 : {
6122 995920 : if (vardata->acl_ok)
6123 995326 : return true;
6124 :
6125 594 : if (!OidIsValid(func_oid))
6126 0 : return false;
6127 :
6128 594 : if (get_func_leakproof(func_oid))
6129 246 : return true;
6130 :
6131 348 : ereport(DEBUG2,
6132 : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
6133 : get_func_name(func_oid))));
6134 348 : return false;
6135 : }
6136 :
6137 : /*
6138 : * get_variable_numdistinct
6139 : * Estimate the number of distinct values of a variable.
6140 : *
6141 : * vardata: results of examine_variable
6142 : * *isdefault: set to true if the result is a default rather than based on
6143 : * anything meaningful.
6144 : *
6145 : * NB: be careful to produce a positive integral result, since callers may
6146 : * compare the result to exact integer counts, or might divide by it.
6147 : */
6148 : double
6149 1510806 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
6150 : {
6151 : double stadistinct;
6152 1510806 : double stanullfrac = 0.0;
6153 : double ntuples;
6154 :
6155 1510806 : *isdefault = false;
6156 :
6157 : /*
6158 : * Determine the stadistinct value to use. There are cases where we can
6159 : * get an estimate even without a pg_statistic entry, or can get a better
6160 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
6161 : * (otherwise, assume no nulls, for lack of any better idea).
6162 : */
6163 1510806 : if (HeapTupleIsValid(vardata->statsTuple))
6164 : {
6165 : /* Use the pg_statistic entry */
6166 : Form_pg_statistic stats;
6167 :
6168 1091272 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6169 1091272 : stadistinct = stats->stadistinct;
6170 1091272 : stanullfrac = stats->stanullfrac;
6171 : }
6172 419534 : else if (vardata->vartype == BOOLOID)
6173 : {
6174 : /*
6175 : * Special-case boolean columns: presumably, two distinct values.
6176 : *
6177 : * Are there any other datatypes we should wire in special estimates
6178 : * for?
6179 : */
6180 580 : stadistinct = 2.0;
6181 : }
6182 418954 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6183 : {
6184 : /*
6185 : * If the Var represents a column of a VALUES RTE, assume it's unique.
6186 : * This could of course be very wrong, but it should tend to be true
6187 : * in well-written queries. We could consider examining the VALUES'
6188 : * contents to get some real statistics; but that only works if the
6189 : * entries are all constants, and it would be pretty expensive anyway.
6190 : */
6191 3442 : stadistinct = -1.0; /* unique (and all non null) */
6192 : }
6193 : else
6194 : {
6195 : /*
6196 : * We don't keep statistics for system columns, but in some cases we
6197 : * can infer distinctness anyway.
6198 : */
6199 415512 : if (vardata->var && IsA(vardata->var, Var))
6200 : {
6201 384094 : switch (((Var *) vardata->var)->varattno)
6202 : {
6203 1178 : case SelfItemPointerAttributeNumber:
6204 1178 : stadistinct = -1.0; /* unique (and all non null) */
6205 1178 : break;
6206 2518 : case TableOidAttributeNumber:
6207 2518 : stadistinct = 1.0; /* only 1 value */
6208 2518 : break;
6209 380398 : default:
6210 380398 : stadistinct = 0.0; /* means "unknown" */
6211 380398 : break;
6212 : }
6213 : }
6214 : else
6215 31418 : stadistinct = 0.0; /* means "unknown" */
6216 :
6217 : /*
6218 : * XXX consider using estimate_num_groups on expressions?
6219 : */
6220 : }
6221 :
6222 : /*
6223 : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6224 : * variable, assume it is unique no matter what pg_statistic says; the
6225 : * statistics could be out of date, or we might have found a partial
6226 : * unique index that proves the var is unique for this query. However,
6227 : * we'd better still believe the null-fraction statistic.
6228 : */
6229 1510806 : if (vardata->isunique)
6230 394878 : stadistinct = -1.0 * (1.0 - stanullfrac);
6231 :
6232 : /*
6233 : * If we had an absolute estimate, use that.
6234 : */
6235 1510806 : if (stadistinct > 0.0)
6236 348234 : return clamp_row_est(stadistinct);
6237 :
6238 : /*
6239 : * Otherwise we need to get the relation size; punt if not available.
6240 : */
6241 1162572 : if (vardata->rel == NULL)
6242 : {
6243 404 : *isdefault = true;
6244 404 : return DEFAULT_NUM_DISTINCT;
6245 : }
6246 1162168 : ntuples = vardata->rel->tuples;
6247 1162168 : if (ntuples <= 0.0)
6248 : {
6249 63402 : *isdefault = true;
6250 63402 : return DEFAULT_NUM_DISTINCT;
6251 : }
6252 :
6253 : /*
6254 : * If we had a relative estimate, use that.
6255 : */
6256 1098766 : if (stadistinct < 0.0)
6257 786262 : return clamp_row_est(-stadistinct * ntuples);
6258 :
6259 : /*
6260 : * With no data, estimate ndistinct = ntuples if the table is small, else
6261 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6262 : * that the behavior isn't discontinuous.
6263 : */
6264 312504 : if (ntuples < DEFAULT_NUM_DISTINCT)
6265 152646 : return clamp_row_est(ntuples);
6266 :
6267 159858 : *isdefault = true;
6268 159858 : return DEFAULT_NUM_DISTINCT;
6269 : }
6270 :
6271 : /*
6272 : * get_variable_range
6273 : * Estimate the minimum and maximum value of the specified variable.
6274 : * If successful, store values in *min and *max, and return true.
6275 : * If no data available, return false.
6276 : *
6277 : * sortop is the "<" comparison operator to use. This should generally
6278 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6279 : * The collation must be specified too.
6280 : */
6281 : static bool
6282 211078 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6283 : Oid sortop, Oid collation,
6284 : Datum *min, Datum *max)
6285 : {
6286 211078 : Datum tmin = 0;
6287 211078 : Datum tmax = 0;
6288 211078 : bool have_data = false;
6289 : int16 typLen;
6290 : bool typByVal;
6291 : Oid opfuncoid;
6292 : FmgrInfo opproc;
6293 : AttStatsSlot sslot;
6294 :
6295 : /*
6296 : * XXX It's very tempting to try to use the actual column min and max, if
6297 : * we can get them relatively-cheaply with an index probe. However, since
6298 : * this function is called many times during join planning, that could
6299 : * have unpleasant effects on planning speed. Need more investigation
6300 : * before enabling this.
6301 : */
6302 : #ifdef NOT_USED
6303 : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6304 : return true;
6305 : #endif
6306 :
6307 211078 : if (!HeapTupleIsValid(vardata->statsTuple))
6308 : {
6309 : /* no stats available, so default result */
6310 48348 : return false;
6311 : }
6312 :
6313 : /*
6314 : * If we can't apply the sortop to the stats data, just fail. In
6315 : * principle, if there's a histogram and no MCVs, we could return the
6316 : * histogram endpoints without ever applying the sortop ... but it's
6317 : * probably not worth trying, because whatever the caller wants to do with
6318 : * the endpoints would likely fail the security check too.
6319 : */
6320 162730 : if (!statistic_proc_security_check(vardata,
6321 162730 : (opfuncoid = get_opcode(sortop))))
6322 0 : return false;
6323 :
6324 162730 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6325 :
6326 162730 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6327 :
6328 : /*
6329 : * If there is a histogram with the ordering we want, grab the first and
6330 : * last values.
6331 : */
6332 162730 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6333 : STATISTIC_KIND_HISTOGRAM, sortop,
6334 : ATTSTATSSLOT_VALUES))
6335 : {
6336 120006 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6337 : {
6338 120006 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6339 120006 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6340 120006 : have_data = true;
6341 : }
6342 120006 : free_attstatsslot(&sslot);
6343 : }
6344 :
6345 : /*
6346 : * Otherwise, if there is a histogram with some other ordering, scan it
6347 : * and get the min and max values according to the ordering we want. This
6348 : * of course may not find values that are really extremal according to our
6349 : * ordering, but it beats ignoring available data.
6350 : */
6351 205454 : if (!have_data &&
6352 42724 : get_attstatsslot(&sslot, vardata->statsTuple,
6353 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6354 : ATTSTATSSLOT_VALUES))
6355 : {
6356 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6357 : collation, typLen, typByVal,
6358 : &tmin, &tmax, &have_data);
6359 0 : free_attstatsslot(&sslot);
6360 : }
6361 :
6362 : /*
6363 : * If we have most-common-values info, look for extreme MCVs. This is
6364 : * needed even if we also have a histogram, since the histogram excludes
6365 : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6366 : * be pretty wary of deciding that that is a full representation of the
6367 : * data. Proceed only if the MCVs represent the whole table (to within
6368 : * roundoff error).
6369 : */
6370 162730 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6371 : STATISTIC_KIND_MCV, InvalidOid,
6372 162730 : have_data ? ATTSTATSSLOT_VALUES :
6373 : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6374 : {
6375 79146 : bool use_mcvs = have_data;
6376 :
6377 79146 : if (!have_data)
6378 : {
6379 41226 : double sumcommon = 0.0;
6380 : double nullfrac;
6381 : int i;
6382 :
6383 326008 : for (i = 0; i < sslot.nnumbers; i++)
6384 284782 : sumcommon += sslot.numbers[i];
6385 41226 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6386 41226 : if (sumcommon + nullfrac > 0.99999)
6387 40494 : use_mcvs = true;
6388 : }
6389 :
6390 79146 : if (use_mcvs)
6391 78414 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6392 : collation, typLen, typByVal,
6393 : &tmin, &tmax, &have_data);
6394 79146 : free_attstatsslot(&sslot);
6395 : }
6396 :
6397 162730 : *min = tmin;
6398 162730 : *max = tmax;
6399 162730 : return have_data;
6400 : }
6401 :
6402 : /*
6403 : * get_stats_slot_range: scan sslot for min/max values
6404 : *
6405 : * Subroutine for get_variable_range: update min/max/have_data according
6406 : * to what we find in the statistics array.
6407 : */
6408 : static void
6409 78414 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6410 : Oid collation, int16 typLen, bool typByVal,
6411 : Datum *min, Datum *max, bool *p_have_data)
6412 : {
6413 78414 : Datum tmin = *min;
6414 78414 : Datum tmax = *max;
6415 78414 : bool have_data = *p_have_data;
6416 78414 : bool found_tmin = false;
6417 78414 : bool found_tmax = false;
6418 :
6419 : /* Look up the comparison function, if we didn't already do so */
6420 78414 : if (opproc->fn_oid != opfuncoid)
6421 78414 : fmgr_info(opfuncoid, opproc);
6422 :
6423 : /* Scan all the slot's values */
6424 2428276 : for (int i = 0; i < sslot->nvalues; i++)
6425 : {
6426 2349862 : if (!have_data)
6427 : {
6428 40494 : tmin = tmax = sslot->values[i];
6429 40494 : found_tmin = found_tmax = true;
6430 40494 : *p_have_data = have_data = true;
6431 40494 : continue;
6432 : }
6433 2309368 : if (DatumGetBool(FunctionCall2Coll(opproc,
6434 : collation,
6435 2309368 : sslot->values[i], tmin)))
6436 : {
6437 52962 : tmin = sslot->values[i];
6438 52962 : found_tmin = true;
6439 : }
6440 2309368 : if (DatumGetBool(FunctionCall2Coll(opproc,
6441 : collation,
6442 2309368 : tmax, sslot->values[i])))
6443 : {
6444 99882 : tmax = sslot->values[i];
6445 99882 : found_tmax = true;
6446 : }
6447 : }
6448 :
6449 : /*
6450 : * Copy the slot's values, if we found new extreme values.
6451 : */
6452 78414 : if (found_tmin)
6453 62114 : *min = datumCopy(tmin, typByVal, typLen);
6454 78414 : if (found_tmax)
6455 43146 : *max = datumCopy(tmax, typByVal, typLen);
6456 78414 : }
6457 :
6458 :
6459 : /*
6460 : * get_actual_variable_range
6461 : * Attempt to identify the current *actual* minimum and/or maximum
6462 : * of the specified variable, by looking for a suitable btree index
6463 : * and fetching its low and/or high values.
6464 : * If successful, store values in *min and *max, and return true.
6465 : * (Either pointer can be NULL if that endpoint isn't needed.)
6466 : * If unsuccessful, return false.
6467 : *
6468 : * sortop is the "<" comparison operator to use.
6469 : * collation is the required collation.
6470 : */
6471 : static bool
6472 186362 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6473 : Oid sortop, Oid collation,
6474 : Datum *min, Datum *max)
6475 : {
6476 186362 : bool have_data = false;
6477 186362 : RelOptInfo *rel = vardata->rel;
6478 : RangeTblEntry *rte;
6479 : ListCell *lc;
6480 :
6481 : /* No hope if no relation or it doesn't have indexes */
6482 186362 : if (rel == NULL || rel->indexlist == NIL)
6483 13400 : return false;
6484 : /* If it has indexes it must be a plain relation */
6485 172962 : rte = root->simple_rte_array[rel->relid];
6486 : Assert(rte->rtekind == RTE_RELATION);
6487 :
6488 : /* ignore partitioned tables. Any indexes here are not real indexes */
6489 172962 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6490 828 : return false;
6491 :
6492 : /* Search through the indexes to see if any match our problem */
6493 333840 : foreach(lc, rel->indexlist)
6494 : {
6495 284606 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6496 : ScanDirection indexscandir;
6497 : StrategyNumber strategy;
6498 :
6499 : /* Ignore non-ordering indexes */
6500 284606 : if (index->sortopfamily == NULL)
6501 0 : continue;
6502 :
6503 : /*
6504 : * Ignore partial indexes --- we only want stats that cover the entire
6505 : * relation.
6506 : */
6507 284606 : if (index->indpred != NIL)
6508 288 : continue;
6509 :
6510 : /*
6511 : * The index list might include hypothetical indexes inserted by a
6512 : * get_relation_info hook --- don't try to access them.
6513 : */
6514 284318 : if (index->hypothetical)
6515 0 : continue;
6516 :
6517 : /*
6518 : * The first index column must match the desired variable, sortop, and
6519 : * collation --- but we can use a descending-order index.
6520 : */
6521 284318 : if (collation != index->indexcollations[0])
6522 35646 : continue; /* test first 'cause it's cheapest */
6523 248672 : if (!match_index_to_operand(vardata->var, 0, index))
6524 125772 : continue;
6525 122900 : strategy = get_op_opfamily_strategy(sortop, index->sortopfamily[0]);
6526 122900 : switch (IndexAmTranslateStrategy(strategy, index->relam, index->sortopfamily[0], true))
6527 : {
6528 122900 : case COMPARE_LT:
6529 122900 : if (index->reverse_sort[0])
6530 0 : indexscandir = BackwardScanDirection;
6531 : else
6532 122900 : indexscandir = ForwardScanDirection;
6533 122900 : break;
6534 0 : case COMPARE_GT:
6535 0 : if (index->reverse_sort[0])
6536 0 : indexscandir = ForwardScanDirection;
6537 : else
6538 0 : indexscandir = BackwardScanDirection;
6539 0 : break;
6540 0 : default:
6541 : /* index doesn't match the sortop */
6542 0 : continue;
6543 : }
6544 :
6545 : /*
6546 : * Found a suitable index to extract data from. Set up some data that
6547 : * can be used by both invocations of get_actual_variable_endpoint.
6548 : */
6549 : {
6550 : MemoryContext tmpcontext;
6551 : MemoryContext oldcontext;
6552 : Relation heapRel;
6553 : Relation indexRel;
6554 : TupleTableSlot *slot;
6555 : int16 typLen;
6556 : bool typByVal;
6557 : ScanKeyData scankeys[1];
6558 :
6559 : /* Make sure any cruft gets recycled when we're done */
6560 122900 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
6561 : "get_actual_variable_range workspace",
6562 : ALLOCSET_DEFAULT_SIZES);
6563 122900 : oldcontext = MemoryContextSwitchTo(tmpcontext);
6564 :
6565 : /*
6566 : * Open the table and index so we can read from them. We should
6567 : * already have some type of lock on each.
6568 : */
6569 122900 : heapRel = table_open(rte->relid, NoLock);
6570 122900 : indexRel = index_open(index->indexoid, NoLock);
6571 :
6572 : /* build some stuff needed for indexscan execution */
6573 122900 : slot = table_slot_create(heapRel, NULL);
6574 122900 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6575 :
6576 : /* set up an IS NOT NULL scan key so that we ignore nulls */
6577 122900 : ScanKeyEntryInitialize(&scankeys[0],
6578 : SK_ISNULL | SK_SEARCHNOTNULL,
6579 : 1, /* index col to scan */
6580 : InvalidStrategy, /* no strategy */
6581 : InvalidOid, /* no strategy subtype */
6582 : InvalidOid, /* no collation */
6583 : InvalidOid, /* no reg proc for this */
6584 : (Datum) 0); /* constant */
6585 :
6586 : /* If min is requested ... */
6587 122900 : if (min)
6588 : {
6589 70464 : have_data = get_actual_variable_endpoint(heapRel,
6590 : indexRel,
6591 : indexscandir,
6592 : scankeys,
6593 : typLen,
6594 : typByVal,
6595 : slot,
6596 : oldcontext,
6597 : min);
6598 : }
6599 : else
6600 : {
6601 : /* If min not requested, still want to fetch max */
6602 52436 : have_data = true;
6603 : }
6604 :
6605 : /* If max is requested, and we didn't already fail ... */
6606 122900 : if (max && have_data)
6607 : {
6608 : /* scan in the opposite direction; all else is the same */
6609 54272 : have_data = get_actual_variable_endpoint(heapRel,
6610 : indexRel,
6611 54272 : -indexscandir,
6612 : scankeys,
6613 : typLen,
6614 : typByVal,
6615 : slot,
6616 : oldcontext,
6617 : max);
6618 : }
6619 :
6620 : /* Clean everything up */
6621 122900 : ExecDropSingleTupleTableSlot(slot);
6622 :
6623 122900 : index_close(indexRel, NoLock);
6624 122900 : table_close(heapRel, NoLock);
6625 :
6626 122900 : MemoryContextSwitchTo(oldcontext);
6627 122900 : MemoryContextDelete(tmpcontext);
6628 :
6629 : /* And we're done */
6630 122900 : break;
6631 : }
6632 : }
6633 :
6634 172134 : return have_data;
6635 : }
6636 :
6637 : /*
6638 : * Get one endpoint datum (min or max depending on indexscandir) from the
6639 : * specified index. Return true if successful, false if not.
6640 : * On success, endpoint value is stored to *endpointDatum (and copied into
6641 : * outercontext).
6642 : *
6643 : * scankeys is a 1-element scankey array set up to reject nulls.
6644 : * typLen/typByVal describe the datatype of the index's first column.
6645 : * tableslot is a slot suitable to hold table tuples, in case we need
6646 : * to probe the heap.
6647 : * (We could compute these values locally, but that would mean computing them
6648 : * twice when get_actual_variable_range needs both the min and the max.)
6649 : *
6650 : * Failure occurs either when the index is empty, or we decide that it's
6651 : * taking too long to find a suitable tuple.
6652 : */
6653 : static bool
6654 124736 : get_actual_variable_endpoint(Relation heapRel,
6655 : Relation indexRel,
6656 : ScanDirection indexscandir,
6657 : ScanKey scankeys,
6658 : int16 typLen,
6659 : bool typByVal,
6660 : TupleTableSlot *tableslot,
6661 : MemoryContext outercontext,
6662 : Datum *endpointDatum)
6663 : {
6664 124736 : bool have_data = false;
6665 : SnapshotData SnapshotNonVacuumable;
6666 : IndexScanDesc index_scan;
6667 124736 : Buffer vmbuffer = InvalidBuffer;
6668 124736 : BlockNumber last_heap_block = InvalidBlockNumber;
6669 124736 : int n_visited_heap_pages = 0;
6670 : ItemPointer tid;
6671 : Datum values[INDEX_MAX_KEYS];
6672 : bool isnull[INDEX_MAX_KEYS];
6673 : MemoryContext oldcontext;
6674 :
6675 : /*
6676 : * We use the index-only-scan machinery for this. With mostly-static
6677 : * tables that's a win because it avoids a heap visit. It's also a win
6678 : * for dynamic data, but the reason is less obvious; read on for details.
6679 : *
6680 : * In principle, we should scan the index with our current active
6681 : * snapshot, which is the best approximation we've got to what the query
6682 : * will see when executed. But that won't be exact if a new snap is taken
6683 : * before running the query, and it can be very expensive if a lot of
6684 : * recently-dead or uncommitted rows exist at the beginning or end of the
6685 : * index (because we'll laboriously fetch each one and reject it).
6686 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
6687 : * and uncommitted rows as well as normal visible rows. On the other
6688 : * hand, it will reject known-dead rows, and thus not give a bogus answer
6689 : * when the extreme value has been deleted (unless the deletion was quite
6690 : * recent); that case motivates not using SnapshotAny here.
6691 : *
6692 : * A crucial point here is that SnapshotNonVacuumable, with
6693 : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
6694 : * condition that the indexscan will use to decide that index entries are
6695 : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
6696 : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
6697 : * have to continue scanning past it, we know that the indexscan will mark
6698 : * that index entry killed. That means that the next
6699 : * get_actual_variable_endpoint() call will not have to re-consider that
6700 : * index entry. In this way we avoid repetitive work when this function
6701 : * is used a lot during planning.
6702 : *
6703 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
6704 : * recently-created index, some index entries may point at "broken" HOT
6705 : * chains in which not all the tuple versions contain data matching the
6706 : * index entry. The live tuple version(s) certainly do match the index,
6707 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
6708 : * don't match. Hence, if we took data from the selected heap tuple, we
6709 : * might get a bogus answer that's not close to the index extremal value,
6710 : * or could even be NULL. We avoid this hazard because we take the data
6711 : * from the index entry not the heap.
6712 : *
6713 : * Despite all this care, there are situations where we might find many
6714 : * non-visible tuples near the end of the index. We don't want to expend
6715 : * a huge amount of time here, so we give up once we've read too many heap
6716 : * pages. When we fail for that reason, the caller will end up using
6717 : * whatever extremal value is recorded in pg_statistic.
6718 : */
6719 124736 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
6720 : GlobalVisTestFor(heapRel));
6721 :
6722 124736 : index_scan = index_beginscan(heapRel, indexRel,
6723 : &SnapshotNonVacuumable, NULL,
6724 : 1, 0);
6725 : /* Set it up for index-only scan */
6726 124736 : index_scan->xs_want_itup = true;
6727 124736 : index_rescan(index_scan, scankeys, 1, NULL, 0);
6728 :
6729 : /* Fetch first/next tuple in specified direction */
6730 155356 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
6731 : {
6732 155356 : BlockNumber block = ItemPointerGetBlockNumber(tid);
6733 :
6734 155356 : if (!VM_ALL_VISIBLE(heapRel,
6735 : block,
6736 : &vmbuffer))
6737 : {
6738 : /* Rats, we have to visit the heap to check visibility */
6739 110386 : if (!index_fetch_heap(index_scan, tableslot))
6740 : {
6741 : /*
6742 : * No visible tuple for this index entry, so we need to
6743 : * advance to the next entry. Before doing so, count heap
6744 : * page fetches and give up if we've done too many.
6745 : *
6746 : * We don't charge a page fetch if this is the same heap page
6747 : * as the previous tuple. This is on the conservative side,
6748 : * since other recently-accessed pages are probably still in
6749 : * buffers too; but it's good enough for this heuristic.
6750 : */
6751 : #define VISITED_PAGES_LIMIT 100
6752 :
6753 30620 : if (block != last_heap_block)
6754 : {
6755 2962 : last_heap_block = block;
6756 2962 : n_visited_heap_pages++;
6757 2962 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
6758 0 : break;
6759 : }
6760 :
6761 30620 : continue; /* no visible tuple, try next index entry */
6762 : }
6763 :
6764 : /* We don't actually need the heap tuple for anything */
6765 79766 : ExecClearTuple(tableslot);
6766 :
6767 : /*
6768 : * We don't care whether there's more than one visible tuple in
6769 : * the HOT chain; if any are visible, that's good enough.
6770 : */
6771 : }
6772 :
6773 : /*
6774 : * We expect that the index will return data in IndexTuple not
6775 : * HeapTuple format.
6776 : */
6777 124736 : if (!index_scan->xs_itup)
6778 0 : elog(ERROR, "no data returned for index-only scan");
6779 :
6780 : /*
6781 : * We do not yet support recheck here.
6782 : */
6783 124736 : if (index_scan->xs_recheck)
6784 0 : break;
6785 :
6786 : /* OK to deconstruct the index tuple */
6787 124736 : index_deform_tuple(index_scan->xs_itup,
6788 : index_scan->xs_itupdesc,
6789 : values, isnull);
6790 :
6791 : /* Shouldn't have got a null, but be careful */
6792 124736 : if (isnull[0])
6793 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
6794 : RelationGetRelationName(indexRel));
6795 :
6796 : /* Copy the index column value out to caller's context */
6797 124736 : oldcontext = MemoryContextSwitchTo(outercontext);
6798 124736 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
6799 124736 : MemoryContextSwitchTo(oldcontext);
6800 124736 : have_data = true;
6801 124736 : break;
6802 : }
6803 :
6804 124736 : if (vmbuffer != InvalidBuffer)
6805 111834 : ReleaseBuffer(vmbuffer);
6806 124736 : index_endscan(index_scan);
6807 :
6808 124736 : return have_data;
6809 : }
6810 :
6811 : /*
6812 : * find_join_input_rel
6813 : * Look up the input relation for a join.
6814 : *
6815 : * We assume that the input relation's RelOptInfo must have been constructed
6816 : * already.
6817 : */
6818 : static RelOptInfo *
6819 10870 : find_join_input_rel(PlannerInfo *root, Relids relids)
6820 : {
6821 10870 : RelOptInfo *rel = NULL;
6822 :
6823 10870 : if (!bms_is_empty(relids))
6824 : {
6825 : int relid;
6826 :
6827 10870 : if (bms_get_singleton_member(relids, &relid))
6828 10566 : rel = find_base_rel(root, relid);
6829 : else
6830 304 : rel = find_join_rel(root, relids);
6831 : }
6832 :
6833 10870 : if (rel == NULL)
6834 0 : elog(ERROR, "could not find RelOptInfo for given relids");
6835 :
6836 10870 : return rel;
6837 : }
6838 :
6839 :
6840 : /*-------------------------------------------------------------------------
6841 : *
6842 : * Index cost estimation functions
6843 : *
6844 : *-------------------------------------------------------------------------
6845 : */
6846 :
6847 : /*
6848 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
6849 : */
6850 : List *
6851 788474 : get_quals_from_indexclauses(List *indexclauses)
6852 : {
6853 788474 : List *result = NIL;
6854 : ListCell *lc;
6855 :
6856 1390558 : foreach(lc, indexclauses)
6857 : {
6858 602084 : IndexClause *iclause = lfirst_node(IndexClause, lc);
6859 : ListCell *lc2;
6860 :
6861 1207088 : foreach(lc2, iclause->indexquals)
6862 : {
6863 605004 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6864 :
6865 605004 : result = lappend(result, rinfo);
6866 : }
6867 : }
6868 788474 : return result;
6869 : }
6870 :
6871 : /*
6872 : * Compute the total evaluation cost of the comparison operands in a list
6873 : * of index qual expressions. Since we know these will be evaluated just
6874 : * once per scan, there's no need to distinguish startup from per-row cost.
6875 : *
6876 : * This can be used either on the result of get_quals_from_indexclauses(),
6877 : * or directly on an indexorderbys list. In both cases, we expect that the
6878 : * index key expression is on the left side of binary clauses.
6879 : */
6880 : Cost
6881 1564336 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
6882 : {
6883 1564336 : Cost qual_arg_cost = 0;
6884 : ListCell *lc;
6885 :
6886 2169802 : foreach(lc, indexquals)
6887 : {
6888 605466 : Expr *clause = (Expr *) lfirst(lc);
6889 : Node *other_operand;
6890 : QualCost index_qual_cost;
6891 :
6892 : /*
6893 : * Index quals will have RestrictInfos, indexorderbys won't. Look
6894 : * through RestrictInfo if present.
6895 : */
6896 605466 : if (IsA(clause, RestrictInfo))
6897 604992 : clause = ((RestrictInfo *) clause)->clause;
6898 :
6899 605466 : if (IsA(clause, OpExpr))
6900 : {
6901 590836 : OpExpr *op = (OpExpr *) clause;
6902 :
6903 590836 : other_operand = (Node *) lsecond(op->args);
6904 : }
6905 14630 : else if (IsA(clause, RowCompareExpr))
6906 : {
6907 300 : RowCompareExpr *rc = (RowCompareExpr *) clause;
6908 :
6909 300 : other_operand = (Node *) rc->rargs;
6910 : }
6911 14330 : else if (IsA(clause, ScalarArrayOpExpr))
6912 : {
6913 11378 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6914 :
6915 11378 : other_operand = (Node *) lsecond(saop->args);
6916 : }
6917 2952 : else if (IsA(clause, NullTest))
6918 : {
6919 2952 : other_operand = NULL;
6920 : }
6921 : else
6922 : {
6923 0 : elog(ERROR, "unsupported indexqual type: %d",
6924 : (int) nodeTag(clause));
6925 : other_operand = NULL; /* keep compiler quiet */
6926 : }
6927 :
6928 605466 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
6929 605466 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
6930 : }
6931 1564336 : return qual_arg_cost;
6932 : }
6933 :
6934 : void
6935 775874 : genericcostestimate(PlannerInfo *root,
6936 : IndexPath *path,
6937 : double loop_count,
6938 : GenericCosts *costs)
6939 : {
6940 775874 : IndexOptInfo *index = path->indexinfo;
6941 775874 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
6942 775874 : List *indexOrderBys = path->indexorderbys;
6943 : Cost indexStartupCost;
6944 : Cost indexTotalCost;
6945 : Selectivity indexSelectivity;
6946 : double indexCorrelation;
6947 : double numIndexPages;
6948 : double numIndexTuples;
6949 : double spc_random_page_cost;
6950 : double num_sa_scans;
6951 : double num_outer_scans;
6952 : double num_scans;
6953 : double qual_op_cost;
6954 : double qual_arg_cost;
6955 : List *selectivityQuals;
6956 : ListCell *l;
6957 :
6958 : /*
6959 : * If the index is partial, AND the index predicate with the explicitly
6960 : * given indexquals to produce a more accurate idea of the index
6961 : * selectivity.
6962 : */
6963 775874 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
6964 :
6965 : /*
6966 : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
6967 : * just assume that the number of index descents is the number of distinct
6968 : * combinations of array elements from all of the scan's SAOP clauses.
6969 : */
6970 775874 : num_sa_scans = costs->num_sa_scans;
6971 775874 : if (num_sa_scans < 1)
6972 : {
6973 7762 : num_sa_scans = 1;
6974 16238 : foreach(l, indexQuals)
6975 : {
6976 8476 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
6977 :
6978 8476 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
6979 : {
6980 26 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
6981 26 : double alength = estimate_array_length(root, lsecond(saop->args));
6982 :
6983 26 : if (alength > 1)
6984 26 : num_sa_scans *= alength;
6985 : }
6986 : }
6987 : }
6988 :
6989 : /* Estimate the fraction of main-table tuples that will be visited */
6990 775874 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
6991 775874 : index->rel->relid,
6992 : JOIN_INNER,
6993 : NULL);
6994 :
6995 : /*
6996 : * If caller didn't give us an estimate, estimate the number of index
6997 : * tuples that will be visited. We do it in this rather peculiar-looking
6998 : * way in order to get the right answer for partial indexes.
6999 : */
7000 775874 : numIndexTuples = costs->numIndexTuples;
7001 775874 : if (numIndexTuples <= 0.0)
7002 : {
7003 89224 : numIndexTuples = indexSelectivity * index->rel->tuples;
7004 :
7005 : /*
7006 : * The above calculation counts all the tuples visited across all
7007 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
7008 : * average per-indexscan number, so adjust. This is a handy place to
7009 : * round to integer, too. (If caller supplied tuple estimate, it's
7010 : * responsible for handling these considerations.)
7011 : */
7012 89224 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7013 : }
7014 :
7015 : /*
7016 : * We can bound the number of tuples by the index size in any case. Also,
7017 : * always estimate at least one tuple is touched, even when
7018 : * indexSelectivity estimate is tiny.
7019 : */
7020 775874 : if (numIndexTuples > index->tuples)
7021 7412 : numIndexTuples = index->tuples;
7022 775874 : if (numIndexTuples < 1.0)
7023 90854 : numIndexTuples = 1.0;
7024 :
7025 : /*
7026 : * Estimate the number of index pages that will be retrieved.
7027 : *
7028 : * We use the simplistic method of taking a pro-rata fraction of the total
7029 : * number of index pages. In effect, this counts only leaf pages and not
7030 : * any overhead such as index metapage or upper tree levels.
7031 : *
7032 : * In practice access to upper index levels is often nearly free because
7033 : * those tend to stay in cache under load; moreover, the cost involved is
7034 : * highly dependent on index type. We therefore ignore such costs here
7035 : * and leave it to the caller to add a suitable charge if needed.
7036 : */
7037 775874 : if (index->pages > 1 && index->tuples > 1)
7038 720476 : numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
7039 : else
7040 55398 : numIndexPages = 1.0;
7041 :
7042 : /* fetch estimated page cost for tablespace containing index */
7043 775874 : get_tablespace_page_costs(index->reltablespace,
7044 : &spc_random_page_cost,
7045 : NULL);
7046 :
7047 : /*
7048 : * Now compute the disk access costs.
7049 : *
7050 : * The above calculations are all per-index-scan. However, if we are in a
7051 : * nestloop inner scan, we can expect the scan to be repeated (with
7052 : * different search keys) for each row of the outer relation. Likewise,
7053 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
7054 : * the potential for cache effects to reduce the number of disk page
7055 : * fetches needed. We want to estimate the average per-scan I/O cost in
7056 : * the presence of caching.
7057 : *
7058 : * We use the Mackert-Lohman formula (see costsize.c for details) to
7059 : * estimate the total number of page fetches that occur. While this
7060 : * wasn't what it was designed for, it seems a reasonable model anyway.
7061 : * Note that we are counting pages not tuples anymore, so we take N = T =
7062 : * index size, as if there were one "tuple" per page.
7063 : */
7064 775874 : num_outer_scans = loop_count;
7065 775874 : num_scans = num_sa_scans * num_outer_scans;
7066 :
7067 775874 : if (num_scans > 1)
7068 : {
7069 : double pages_fetched;
7070 :
7071 : /* total page fetches ignoring cache effects */
7072 88716 : pages_fetched = numIndexPages * num_scans;
7073 :
7074 : /* use Mackert and Lohman formula to adjust for cache effects */
7075 88716 : pages_fetched = index_pages_fetched(pages_fetched,
7076 : index->pages,
7077 88716 : (double) index->pages,
7078 : root);
7079 :
7080 : /*
7081 : * Now compute the total disk access cost, and then report a pro-rated
7082 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
7083 : * since that's internal to the indexscan.)
7084 : */
7085 88716 : indexTotalCost = (pages_fetched * spc_random_page_cost)
7086 : / num_outer_scans;
7087 : }
7088 : else
7089 : {
7090 : /*
7091 : * For a single index scan, we just charge spc_random_page_cost per
7092 : * page touched.
7093 : */
7094 687158 : indexTotalCost = numIndexPages * spc_random_page_cost;
7095 : }
7096 :
7097 : /*
7098 : * CPU cost: any complex expressions in the indexquals will need to be
7099 : * evaluated once at the start of the scan to reduce them to runtime keys
7100 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
7101 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
7102 : * indexqual operator. Because we have numIndexTuples as a per-scan
7103 : * number, we have to multiply by num_sa_scans to get the correct result
7104 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
7105 : * ORDER BY expressions.
7106 : *
7107 : * Note: this neglects the possible costs of rechecking lossy operators.
7108 : * Detecting that that might be needed seems more expensive than it's
7109 : * worth, though, considering all the other inaccuracies here ...
7110 : */
7111 775874 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
7112 775874 : index_other_operands_eval_cost(root, indexOrderBys);
7113 775874 : qual_op_cost = cpu_operator_cost *
7114 775874 : (list_length(indexQuals) + list_length(indexOrderBys));
7115 :
7116 775874 : indexStartupCost = qual_arg_cost;
7117 775874 : indexTotalCost += qual_arg_cost;
7118 775874 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
7119 :
7120 : /*
7121 : * Generic assumption about index correlation: there isn't any.
7122 : */
7123 775874 : indexCorrelation = 0.0;
7124 :
7125 : /*
7126 : * Return everything to caller.
7127 : */
7128 775874 : costs->indexStartupCost = indexStartupCost;
7129 775874 : costs->indexTotalCost = indexTotalCost;
7130 775874 : costs->indexSelectivity = indexSelectivity;
7131 775874 : costs->indexCorrelation = indexCorrelation;
7132 775874 : costs->numIndexPages = numIndexPages;
7133 775874 : costs->numIndexTuples = numIndexTuples;
7134 775874 : costs->spc_random_page_cost = spc_random_page_cost;
7135 775874 : costs->num_sa_scans = num_sa_scans;
7136 775874 : }
7137 :
7138 : /*
7139 : * If the index is partial, add its predicate to the given qual list.
7140 : *
7141 : * ANDing the index predicate with the explicitly given indexquals produces
7142 : * a more accurate idea of the index's selectivity. However, we need to be
7143 : * careful not to insert redundant clauses, because clauselist_selectivity()
7144 : * is easily fooled into computing a too-low selectivity estimate. Our
7145 : * approach is to add only the predicate clause(s) that cannot be proven to
7146 : * be implied by the given indexquals. This successfully handles cases such
7147 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
7148 : * There are many other cases where we won't detect redundancy, leading to a
7149 : * too-low selectivity estimate, which will bias the system in favor of using
7150 : * partial indexes where possible. That is not necessarily bad though.
7151 : *
7152 : * Note that indexQuals contains RestrictInfo nodes while the indpred
7153 : * does not, so the output list will be mixed. This is OK for both
7154 : * predicate_implied_by() and clauselist_selectivity(), but might be
7155 : * problematic if the result were passed to other things.
7156 : */
7157 : List *
7158 1303436 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
7159 : {
7160 1303436 : List *predExtraQuals = NIL;
7161 : ListCell *lc;
7162 :
7163 1303436 : if (index->indpred == NIL)
7164 1301410 : return indexQuals;
7165 :
7166 4064 : foreach(lc, index->indpred)
7167 : {
7168 2038 : Node *predQual = (Node *) lfirst(lc);
7169 2038 : List *oneQual = list_make1(predQual);
7170 :
7171 2038 : if (!predicate_implied_by(oneQual, indexQuals, false))
7172 1836 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7173 : }
7174 2026 : return list_concat(predExtraQuals, indexQuals);
7175 : }
7176 :
7177 : /*
7178 : * Estimate correlation of btree index's first column.
7179 : *
7180 : * If we can get an estimate of the first column's ordering correlation C
7181 : * from pg_statistic, estimate the index correlation as C for a single-column
7182 : * index, or C * 0.75 for multiple columns. The idea here is that multiple
7183 : * columns dilute the importance of the first column's ordering, but don't
7184 : * negate it entirely.
7185 : *
7186 : * We already filled in the stats tuple for *vardata when called.
7187 : */
7188 : static double
7189 577238 : btcost_correlation(IndexOptInfo *index, VariableStatData *vardata)
7190 : {
7191 : Oid sortop;
7192 : AttStatsSlot sslot;
7193 577238 : double indexCorrelation = 0;
7194 :
7195 : Assert(HeapTupleIsValid(vardata->statsTuple));
7196 :
7197 577238 : sortop = get_opfamily_member(index->opfamily[0],
7198 577238 : index->opcintype[0],
7199 577238 : index->opcintype[0],
7200 : BTLessStrategyNumber);
7201 1154476 : if (OidIsValid(sortop) &&
7202 577238 : get_attstatsslot(&sslot, vardata->statsTuple,
7203 : STATISTIC_KIND_CORRELATION, sortop,
7204 : ATTSTATSSLOT_NUMBERS))
7205 : {
7206 : double varCorrelation;
7207 :
7208 : Assert(sslot.nnumbers == 1);
7209 569906 : varCorrelation = sslot.numbers[0];
7210 :
7211 569906 : if (index->reverse_sort[0])
7212 0 : varCorrelation = -varCorrelation;
7213 :
7214 569906 : if (index->nkeycolumns > 1)
7215 202510 : indexCorrelation = varCorrelation * 0.75;
7216 : else
7217 367396 : indexCorrelation = varCorrelation;
7218 :
7219 569906 : free_attstatsslot(&sslot);
7220 : }
7221 :
7222 577238 : return indexCorrelation;
7223 : }
7224 :
7225 : void
7226 768112 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7227 : Cost *indexStartupCost, Cost *indexTotalCost,
7228 : Selectivity *indexSelectivity, double *indexCorrelation,
7229 : double *indexPages)
7230 : {
7231 768112 : IndexOptInfo *index = path->indexinfo;
7232 768112 : GenericCosts costs = {0};
7233 768112 : VariableStatData vardata = {0};
7234 : double numIndexTuples;
7235 : Cost descentCost;
7236 : List *indexBoundQuals;
7237 : List *indexSkipQuals;
7238 : int indexcol;
7239 : bool eqQualHere;
7240 : bool found_row_compare;
7241 : bool found_array;
7242 : bool found_is_null_op;
7243 768112 : bool have_correlation = false;
7244 : double num_sa_scans;
7245 768112 : double correlation = 0.0;
7246 : ListCell *lc;
7247 :
7248 : /*
7249 : * For a btree scan, only leading '=' quals plus inequality quals for the
7250 : * immediately next attribute contribute to index selectivity (these are
7251 : * the "boundary quals" that determine the starting and stopping points of
7252 : * the index scan). Additional quals can suppress visits to the heap, so
7253 : * it's OK to count them in indexSelectivity, but they should not count
7254 : * for estimating numIndexTuples. So we must examine the given indexquals
7255 : * to find out which ones count as boundary quals. We rely on the
7256 : * knowledge that they are given in index column order. Note that nbtree
7257 : * preprocessing can add skip arrays that act as leading '=' quals in the
7258 : * absence of ordinary input '=' quals, so in practice _most_ input quals
7259 : * are able to act as index bound quals (which we take into account here).
7260 : *
7261 : * For a RowCompareExpr, we consider only the first column, just as
7262 : * rowcomparesel() does.
7263 : *
7264 : * If there's a SAOP or skip array in the quals, we'll actually perform up
7265 : * to N index descents (not just one), but the underlying array key's
7266 : * operator can be considered to act the same as it normally does.
7267 : */
7268 768112 : indexBoundQuals = NIL;
7269 768112 : indexSkipQuals = NIL;
7270 768112 : indexcol = 0;
7271 768112 : eqQualHere = false;
7272 768112 : found_row_compare = false;
7273 768112 : found_array = false;
7274 768112 : found_is_null_op = false;
7275 768112 : num_sa_scans = 1;
7276 1316184 : foreach(lc, path->indexclauses)
7277 : {
7278 577912 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7279 : ListCell *lc2;
7280 :
7281 577912 : if (indexcol < iclause->indexcol)
7282 : {
7283 104716 : double num_sa_scans_prev_cols = num_sa_scans;
7284 :
7285 : /*
7286 : * Beginning of a new column's quals.
7287 : *
7288 : * Skip scans use skip arrays, which are ScalarArrayOp style
7289 : * arrays that generate their elements procedurally and on demand.
7290 : * Given a multi-column index on "(a, b)", and an SQL WHERE clause
7291 : * "WHERE b = 42", a skip scan will effectively use an indexqual
7292 : * "WHERE a = ANY('{every col a value}') AND b = 42". (Obviously,
7293 : * the array on "a" must also return "IS NULL" matches, since our
7294 : * WHERE clause used no strict operator on "a").
7295 : *
7296 : * Here we consider how nbtree will backfill skip arrays for any
7297 : * index columns that lacked an '=' qual. This maintains our
7298 : * num_sa_scans estimate, and determines if this new column (the
7299 : * "iclause->indexcol" column, not the prior "indexcol" column)
7300 : * can have its RestrictInfos/quals added to indexBoundQuals.
7301 : *
7302 : * We'll need to handle columns that have inequality quals, where
7303 : * the skip array generates values from a range constrained by the
7304 : * quals (not every possible value). We've been maintaining
7305 : * indexSkipQuals to help with this; it will now contain all of
7306 : * the prior column's quals (that is, indexcol's quals) when they
7307 : * might be used for this.
7308 : */
7309 104716 : if (found_row_compare)
7310 : {
7311 : /*
7312 : * Skip arrays can't be added after a RowCompare input qual
7313 : * due to limitations in nbtree
7314 : */
7315 0 : break;
7316 : }
7317 104716 : if (eqQualHere)
7318 : {
7319 : /*
7320 : * Don't need to add a skip array for an indexcol that already
7321 : * has an '=' qual/equality constraint
7322 : */
7323 75022 : indexcol++;
7324 75022 : indexSkipQuals = NIL;
7325 : }
7326 104716 : eqQualHere = false;
7327 :
7328 108038 : while (indexcol < iclause->indexcol)
7329 : {
7330 : double ndistinct;
7331 33162 : bool isdefault = true;
7332 :
7333 33162 : found_array = true;
7334 :
7335 : /*
7336 : * A skipped attribute's ndistinct forms the basis of our
7337 : * estimate of the total number of "array elements" used by
7338 : * its skip array at runtime. Look that up first.
7339 : */
7340 33162 : examine_indexcol_variable(root, index, indexcol, &vardata);
7341 33162 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
7342 :
7343 33162 : if (indexcol == 0)
7344 : {
7345 : /*
7346 : * Get an estimate of the leading column's correlation in
7347 : * passing (avoids rereading variable stats below)
7348 : */
7349 29682 : if (HeapTupleIsValid(vardata.statsTuple))
7350 22790 : correlation = btcost_correlation(index, &vardata);
7351 29682 : have_correlation = true;
7352 : }
7353 :
7354 33162 : ReleaseVariableStats(vardata);
7355 :
7356 : /*
7357 : * If ndistinct is a default estimate, conservatively assume
7358 : * that no skipping will happen at runtime
7359 : */
7360 33162 : if (isdefault)
7361 : {
7362 5350 : num_sa_scans = num_sa_scans_prev_cols;
7363 29840 : break; /* done building indexBoundQuals */
7364 : }
7365 :
7366 : /*
7367 : * Apply indexcol's indexSkipQuals selectivity to ndistinct
7368 : */
7369 27812 : if (indexSkipQuals != NIL)
7370 : {
7371 : List *partialSkipQuals;
7372 : Selectivity ndistinctfrac;
7373 :
7374 : /*
7375 : * If the index is partial, AND the index predicate with
7376 : * the index-bound quals to produce a more accurate idea
7377 : * of the number of distinct values for prior indexcol
7378 : */
7379 664 : partialSkipQuals = add_predicate_to_index_quals(index,
7380 : indexSkipQuals);
7381 :
7382 664 : ndistinctfrac = clauselist_selectivity(root, partialSkipQuals,
7383 664 : index->rel->relid,
7384 : JOIN_INNER,
7385 : NULL);
7386 :
7387 : /*
7388 : * If ndistinctfrac is selective (on its own), the scan is
7389 : * unlikely to benefit from repositioning itself using
7390 : * later quals. Do not allow iclause->indexcol's quals to
7391 : * be added to indexBoundQuals (it would increase descent
7392 : * costs, without lowering numIndexTuples costs by much).
7393 : */
7394 664 : if (ndistinctfrac < DEFAULT_RANGE_INEQ_SEL)
7395 : {
7396 374 : num_sa_scans = num_sa_scans_prev_cols;
7397 374 : break; /* done building indexBoundQuals */
7398 : }
7399 :
7400 : /* Adjust ndistinct downward */
7401 290 : ndistinct = rint(ndistinct * ndistinctfrac);
7402 290 : ndistinct = Max(ndistinct, 1);
7403 : }
7404 :
7405 : /*
7406 : * When there's no inequality quals, account for the need to
7407 : * find an initial value by counting -inf/+inf as a value.
7408 : *
7409 : * We don't charge anything extra for possible next/prior key
7410 : * index probes, which are sometimes used to find the next
7411 : * valid skip array element (ahead of using the located
7412 : * element value to relocate the scan to the next position
7413 : * that might contain matching tuples). It seems hard to do
7414 : * better here. Use of the skip support infrastructure often
7415 : * avoids most next/prior key probes. But even when it can't,
7416 : * there's a decent chance that most individual next/prior key
7417 : * probes will locate a leaf page whose key space overlaps all
7418 : * of the scan's keys (even the lower-order keys) -- which
7419 : * also avoids the need for a separate, extra index descent.
7420 : * Note also that these probes are much cheaper than non-probe
7421 : * primitive index scans: they're reliably very selective.
7422 : */
7423 27438 : if (indexSkipQuals == NIL)
7424 27148 : ndistinct += 1;
7425 :
7426 : /*
7427 : * Update num_sa_scans estimate by multiplying by ndistinct.
7428 : *
7429 : * We make the pessimistic assumption that there is no
7430 : * naturally occurring cross-column correlation. This is
7431 : * often wrong, but it seems best to err on the side of not
7432 : * expecting skipping to be helpful...
7433 : */
7434 27438 : num_sa_scans *= ndistinct;
7435 :
7436 : /*
7437 : * ...but back out of adding this latest group of 1 or more
7438 : * skip arrays when num_sa_scans exceeds the total number of
7439 : * index pages (revert to num_sa_scans from before indexcol).
7440 : * This causes a sharp discontinuity in cost (as a function of
7441 : * the indexcol's ndistinct), but that is representative of
7442 : * actual runtime costs.
7443 : *
7444 : * Note that skipping is helpful when each primitive index
7445 : * scan only manages to skip over 1 or 2 irrelevant leaf pages
7446 : * on average. Skip arrays bring savings in CPU costs due to
7447 : * the scan not needing to evaluate indexquals against every
7448 : * tuple, which can greatly exceed any savings in I/O costs.
7449 : * This test is a test of whether num_sa_scans implies that
7450 : * we're past the point where the ability to skip ceases to
7451 : * lower the scan's costs (even qual evaluation CPU costs).
7452 : */
7453 27438 : if (index->pages < num_sa_scans)
7454 : {
7455 24116 : num_sa_scans = num_sa_scans_prev_cols;
7456 24116 : break; /* done building indexBoundQuals */
7457 : }
7458 :
7459 3322 : indexcol++;
7460 3322 : indexSkipQuals = NIL;
7461 : }
7462 :
7463 : /*
7464 : * Finished considering the need to add skip arrays to bridge an
7465 : * initial eqQualHere gap between the old and new index columns
7466 : * (or there was no initial eqQualHere gap in the first place).
7467 : *
7468 : * If an initial gap could not be bridged, then new column's quals
7469 : * (i.e. iclause->indexcol's quals) won't go into indexBoundQuals,
7470 : * and so won't affect our final numIndexTuples estimate.
7471 : */
7472 104716 : if (indexcol != iclause->indexcol)
7473 29840 : break; /* done building indexBoundQuals */
7474 : }
7475 :
7476 : Assert(indexcol == iclause->indexcol);
7477 :
7478 : /* Examine each indexqual associated with this index clause */
7479 1098892 : foreach(lc2, iclause->indexquals)
7480 : {
7481 550820 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7482 550820 : Expr *clause = rinfo->clause;
7483 550820 : Oid clause_op = InvalidOid;
7484 : int op_strategy;
7485 :
7486 550820 : if (IsA(clause, OpExpr))
7487 : {
7488 537214 : OpExpr *op = (OpExpr *) clause;
7489 :
7490 537214 : clause_op = op->opno;
7491 : }
7492 13606 : else if (IsA(clause, RowCompareExpr))
7493 : {
7494 300 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7495 :
7496 300 : clause_op = linitial_oid(rc->opnos);
7497 300 : found_row_compare = true;
7498 : }
7499 13306 : else if (IsA(clause, ScalarArrayOpExpr))
7500 : {
7501 10998 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7502 10998 : Node *other_operand = (Node *) lsecond(saop->args);
7503 10998 : double alength = estimate_array_length(root, other_operand);
7504 :
7505 10998 : clause_op = saop->opno;
7506 10998 : found_array = true;
7507 : /* estimate SA descents by indexBoundQuals only */
7508 10998 : if (alength > 1)
7509 10770 : num_sa_scans *= alength;
7510 : }
7511 2308 : else if (IsA(clause, NullTest))
7512 : {
7513 2308 : NullTest *nt = (NullTest *) clause;
7514 :
7515 2308 : if (nt->nulltesttype == IS_NULL)
7516 : {
7517 258 : found_is_null_op = true;
7518 : /* IS NULL is like = for selectivity/skip scan purposes */
7519 258 : eqQualHere = true;
7520 : }
7521 : }
7522 : else
7523 0 : elog(ERROR, "unsupported indexqual type: %d",
7524 : (int) nodeTag(clause));
7525 :
7526 : /* check for equality operator */
7527 550820 : if (OidIsValid(clause_op))
7528 : {
7529 548512 : op_strategy = get_op_opfamily_strategy(clause_op,
7530 548512 : index->opfamily[indexcol]);
7531 : Assert(op_strategy != 0); /* not a member of opfamily?? */
7532 548512 : if (op_strategy == BTEqualStrategyNumber)
7533 522050 : eqQualHere = true;
7534 : }
7535 :
7536 550820 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
7537 :
7538 : /*
7539 : * We apply inequality selectivities to estimate index descent
7540 : * costs with scans that use skip arrays. Save this indexcol's
7541 : * RestrictInfos if it looks like they'll be needed for that.
7542 : */
7543 550820 : if (!eqQualHere && !found_row_compare &&
7544 27570 : indexcol < index->nkeycolumns - 1)
7545 5642 : indexSkipQuals = lappend(indexSkipQuals, rinfo);
7546 : }
7547 : }
7548 :
7549 : /*
7550 : * If index is unique and we found an '=' clause for each column, we can
7551 : * just assume numIndexTuples = 1 and skip the expensive
7552 : * clauselist_selectivity calculations. However, an array or NullTest
7553 : * always invalidates that theory (even when eqQualHere has been set).
7554 : */
7555 768112 : if (index->unique &&
7556 618854 : indexcol == index->nkeycolumns - 1 &&
7557 249204 : eqQualHere &&
7558 249204 : !found_array &&
7559 243150 : !found_is_null_op)
7560 243084 : numIndexTuples = 1.0;
7561 : else
7562 : {
7563 : List *selectivityQuals;
7564 : Selectivity btreeSelectivity;
7565 :
7566 : /*
7567 : * If the index is partial, AND the index predicate with the
7568 : * index-bound quals to produce a more accurate idea of the number of
7569 : * rows covered by the bound conditions.
7570 : */
7571 525028 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
7572 :
7573 525028 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
7574 525028 : index->rel->relid,
7575 : JOIN_INNER,
7576 : NULL);
7577 525028 : numIndexTuples = btreeSelectivity * index->rel->tuples;
7578 :
7579 : /*
7580 : * btree automatically combines individual array element primitive
7581 : * index scans whenever the tuples covered by the next set of array
7582 : * keys are close to tuples covered by the current set. That puts a
7583 : * natural ceiling on the worst case number of descents -- there
7584 : * cannot possibly be more than one descent per leaf page scanned.
7585 : *
7586 : * Clamp the number of descents to at most 1/3 the number of index
7587 : * pages. This avoids implausibly high estimates with low selectivity
7588 : * paths, where scans usually require only one or two descents. This
7589 : * is most likely to help when there are several SAOP clauses, where
7590 : * naively accepting the total number of distinct combinations of
7591 : * array elements as the number of descents would frequently lead to
7592 : * wild overestimates.
7593 : *
7594 : * We somewhat arbitrarily don't just make the cutoff the total number
7595 : * of leaf pages (we make it 1/3 the total number of pages instead) to
7596 : * give the btree code credit for its ability to continue on the leaf
7597 : * level with low selectivity scans.
7598 : *
7599 : * Note: num_sa_scans includes both ScalarArrayOp array elements and
7600 : * skip array elements whose qual affects our numIndexTuples estimate.
7601 : */
7602 525028 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
7603 525028 : num_sa_scans = Max(num_sa_scans, 1);
7604 :
7605 : /*
7606 : * As in genericcostestimate(), we have to adjust for any array quals
7607 : * included in indexBoundQuals, and then round to integer.
7608 : *
7609 : * It is tempting to make genericcostestimate behave as if array
7610 : * clauses work in almost the same way as scalar operators during
7611 : * btree scans, making the top-level scan look like a continuous scan
7612 : * (as opposed to num_sa_scans-many primitive index scans). After
7613 : * all, btree scans mostly work like that at runtime. However, such a
7614 : * scheme would badly bias genericcostestimate's simplistic approach
7615 : * to calculating numIndexPages through prorating.
7616 : *
7617 : * Stick with the approach taken by non-native SAOP scans for now.
7618 : * genericcostestimate will use the Mackert-Lohman formula to
7619 : * compensate for repeat page fetches, even though that definitely
7620 : * won't happen during btree scans (not for leaf pages, at least).
7621 : * We're usually very pessimistic about the number of primitive index
7622 : * scans that will be required, but it's not clear how to do better.
7623 : */
7624 525028 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7625 : }
7626 :
7627 : /*
7628 : * Now do generic index cost estimation.
7629 : */
7630 768112 : costs.numIndexTuples = numIndexTuples;
7631 768112 : costs.num_sa_scans = num_sa_scans;
7632 :
7633 768112 : genericcostestimate(root, path, loop_count, &costs);
7634 :
7635 : /*
7636 : * Add a CPU-cost component to represent the costs of initial btree
7637 : * descent. We don't charge any I/O cost for touching upper btree levels,
7638 : * since they tend to stay in cache, but we still have to do about log2(N)
7639 : * comparisons to descend a btree of N leaf tuples. We charge one
7640 : * cpu_operator_cost per comparison.
7641 : *
7642 : * If there are SAOP or skip array keys, charge this once per estimated
7643 : * index descent. The ones after the first one are not startup cost so
7644 : * far as the overall plan goes, so just add them to "total" cost.
7645 : */
7646 768112 : if (index->tuples > 1) /* avoid computing log(0) */
7647 : {
7648 721428 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
7649 721428 : costs.indexStartupCost += descentCost;
7650 721428 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7651 : }
7652 :
7653 : /*
7654 : * Even though we're not charging I/O cost for touching upper btree pages,
7655 : * it's still reasonable to charge some CPU cost per page descended
7656 : * through. Moreover, if we had no such charge at all, bloated indexes
7657 : * would appear to have the same search cost as unbloated ones, at least
7658 : * in cases where only a single leaf page is expected to be visited. This
7659 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
7660 : * touched. The number of such pages is btree tree height plus one (ie,
7661 : * we charge for the leaf page too). As above, charge once per estimated
7662 : * SAOP/skip array descent.
7663 : */
7664 768112 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7665 768112 : costs.indexStartupCost += descentCost;
7666 768112 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7667 :
7668 768112 : if (!have_correlation)
7669 : {
7670 738430 : examine_indexcol_variable(root, index, 0, &vardata);
7671 738430 : if (HeapTupleIsValid(vardata.statsTuple))
7672 554448 : costs.indexCorrelation = btcost_correlation(index, &vardata);
7673 738430 : ReleaseVariableStats(vardata);
7674 : }
7675 : else
7676 : {
7677 : /* btcost_correlation already called earlier on */
7678 29682 : costs.indexCorrelation = correlation;
7679 : }
7680 :
7681 768112 : *indexStartupCost = costs.indexStartupCost;
7682 768112 : *indexTotalCost = costs.indexTotalCost;
7683 768112 : *indexSelectivity = costs.indexSelectivity;
7684 768112 : *indexCorrelation = costs.indexCorrelation;
7685 768112 : *indexPages = costs.numIndexPages;
7686 768112 : }
7687 :
7688 : void
7689 410 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7690 : Cost *indexStartupCost, Cost *indexTotalCost,
7691 : Selectivity *indexSelectivity, double *indexCorrelation,
7692 : double *indexPages)
7693 : {
7694 410 : GenericCosts costs = {0};
7695 :
7696 410 : genericcostestimate(root, path, loop_count, &costs);
7697 :
7698 : /*
7699 : * A hash index has no descent costs as such, since the index AM can go
7700 : * directly to the target bucket after computing the hash value. There
7701 : * are a couple of other hash-specific costs that we could conceivably add
7702 : * here, though:
7703 : *
7704 : * Ideally we'd charge spc_random_page_cost for each page in the target
7705 : * bucket, not just the numIndexPages pages that genericcostestimate
7706 : * thought we'd visit. However in most cases we don't know which bucket
7707 : * that will be. There's no point in considering the average bucket size
7708 : * because the hash AM makes sure that's always one page.
7709 : *
7710 : * Likewise, we could consider charging some CPU for each index tuple in
7711 : * the bucket, if we knew how many there were. But the per-tuple cost is
7712 : * just a hash value comparison, not a general datatype-dependent
7713 : * comparison, so any such charge ought to be quite a bit less than
7714 : * cpu_operator_cost; which makes it probably not worth worrying about.
7715 : *
7716 : * A bigger issue is that chance hash-value collisions will result in
7717 : * wasted probes into the heap. We don't currently attempt to model this
7718 : * cost on the grounds that it's rare, but maybe it's not rare enough.
7719 : * (Any fix for this ought to consider the generic lossy-operator problem,
7720 : * though; it's not entirely hash-specific.)
7721 : */
7722 :
7723 410 : *indexStartupCost = costs.indexStartupCost;
7724 410 : *indexTotalCost = costs.indexTotalCost;
7725 410 : *indexSelectivity = costs.indexSelectivity;
7726 410 : *indexCorrelation = costs.indexCorrelation;
7727 410 : *indexPages = costs.numIndexPages;
7728 410 : }
7729 :
7730 : void
7731 4756 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7732 : Cost *indexStartupCost, Cost *indexTotalCost,
7733 : Selectivity *indexSelectivity, double *indexCorrelation,
7734 : double *indexPages)
7735 : {
7736 4756 : IndexOptInfo *index = path->indexinfo;
7737 4756 : GenericCosts costs = {0};
7738 : Cost descentCost;
7739 :
7740 4756 : genericcostestimate(root, path, loop_count, &costs);
7741 :
7742 : /*
7743 : * We model index descent costs similarly to those for btree, but to do
7744 : * that we first need an idea of the tree height. We somewhat arbitrarily
7745 : * assume that the fanout is 100, meaning the tree height is at most
7746 : * log100(index->pages).
7747 : *
7748 : * Although this computation isn't really expensive enough to require
7749 : * caching, we might as well use index->tree_height to cache it.
7750 : */
7751 4756 : if (index->tree_height < 0) /* unknown? */
7752 : {
7753 4742 : if (index->pages > 1) /* avoid computing log(0) */
7754 2712 : index->tree_height = (int) (log(index->pages) / log(100.0));
7755 : else
7756 2030 : index->tree_height = 0;
7757 : }
7758 :
7759 : /*
7760 : * Add a CPU-cost component to represent the costs of initial descent. We
7761 : * just use log(N) here not log2(N) since the branching factor isn't
7762 : * necessarily two anyway. As for btree, charge once per SA scan.
7763 : */
7764 4756 : if (index->tuples > 1) /* avoid computing log(0) */
7765 : {
7766 4756 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7767 4756 : costs.indexStartupCost += descentCost;
7768 4756 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7769 : }
7770 :
7771 : /*
7772 : * Likewise add a per-page charge, calculated the same as for btrees.
7773 : */
7774 4756 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7775 4756 : costs.indexStartupCost += descentCost;
7776 4756 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7777 :
7778 4756 : *indexStartupCost = costs.indexStartupCost;
7779 4756 : *indexTotalCost = costs.indexTotalCost;
7780 4756 : *indexSelectivity = costs.indexSelectivity;
7781 4756 : *indexCorrelation = costs.indexCorrelation;
7782 4756 : *indexPages = costs.numIndexPages;
7783 4756 : }
7784 :
7785 : void
7786 1784 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7787 : Cost *indexStartupCost, Cost *indexTotalCost,
7788 : Selectivity *indexSelectivity, double *indexCorrelation,
7789 : double *indexPages)
7790 : {
7791 1784 : IndexOptInfo *index = path->indexinfo;
7792 1784 : GenericCosts costs = {0};
7793 : Cost descentCost;
7794 :
7795 1784 : genericcostestimate(root, path, loop_count, &costs);
7796 :
7797 : /*
7798 : * We model index descent costs similarly to those for btree, but to do
7799 : * that we first need an idea of the tree height. We somewhat arbitrarily
7800 : * assume that the fanout is 100, meaning the tree height is at most
7801 : * log100(index->pages).
7802 : *
7803 : * Although this computation isn't really expensive enough to require
7804 : * caching, we might as well use index->tree_height to cache it.
7805 : */
7806 1784 : if (index->tree_height < 0) /* unknown? */
7807 : {
7808 1778 : if (index->pages > 1) /* avoid computing log(0) */
7809 1778 : index->tree_height = (int) (log(index->pages) / log(100.0));
7810 : else
7811 0 : index->tree_height = 0;
7812 : }
7813 :
7814 : /*
7815 : * Add a CPU-cost component to represent the costs of initial descent. We
7816 : * just use log(N) here not log2(N) since the branching factor isn't
7817 : * necessarily two anyway. As for btree, charge once per SA scan.
7818 : */
7819 1784 : if (index->tuples > 1) /* avoid computing log(0) */
7820 : {
7821 1784 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7822 1784 : costs.indexStartupCost += descentCost;
7823 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7824 : }
7825 :
7826 : /*
7827 : * Likewise add a per-page charge, calculated the same as for btrees.
7828 : */
7829 1784 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7830 1784 : costs.indexStartupCost += descentCost;
7831 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7832 :
7833 1784 : *indexStartupCost = costs.indexStartupCost;
7834 1784 : *indexTotalCost = costs.indexTotalCost;
7835 1784 : *indexSelectivity = costs.indexSelectivity;
7836 1784 : *indexCorrelation = costs.indexCorrelation;
7837 1784 : *indexPages = costs.numIndexPages;
7838 1784 : }
7839 :
7840 :
7841 : /*
7842 : * Support routines for gincostestimate
7843 : */
7844 :
7845 : typedef struct
7846 : {
7847 : bool attHasFullScan[INDEX_MAX_KEYS];
7848 : bool attHasNormalScan[INDEX_MAX_KEYS];
7849 : double partialEntries;
7850 : double exactEntries;
7851 : double searchEntries;
7852 : double arrayScans;
7853 : } GinQualCounts;
7854 :
7855 : /*
7856 : * Estimate the number of index terms that need to be searched for while
7857 : * testing the given GIN query, and increment the counts in *counts
7858 : * appropriately. If the query is unsatisfiable, return false.
7859 : */
7860 : static bool
7861 2074 : gincost_pattern(IndexOptInfo *index, int indexcol,
7862 : Oid clause_op, Datum query,
7863 : GinQualCounts *counts)
7864 : {
7865 : FmgrInfo flinfo;
7866 : Oid extractProcOid;
7867 : Oid collation;
7868 : int strategy_op;
7869 : Oid lefttype,
7870 : righttype;
7871 2074 : int32 nentries = 0;
7872 2074 : bool *partial_matches = NULL;
7873 2074 : Pointer *extra_data = NULL;
7874 2074 : bool *nullFlags = NULL;
7875 2074 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
7876 : int32 i;
7877 :
7878 : Assert(indexcol < index->nkeycolumns);
7879 :
7880 : /*
7881 : * Get the operator's strategy number and declared input data types within
7882 : * the index opfamily. (We don't need the latter, but we use
7883 : * get_op_opfamily_properties because it will throw error if it fails to
7884 : * find a matching pg_amop entry.)
7885 : */
7886 2074 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
7887 : &strategy_op, &lefttype, &righttype);
7888 :
7889 : /*
7890 : * GIN always uses the "default" support functions, which are those with
7891 : * lefttype == righttype == the opclass' opcintype (see
7892 : * IndexSupportInitialize in relcache.c).
7893 : */
7894 2074 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
7895 2074 : index->opcintype[indexcol],
7896 2074 : index->opcintype[indexcol],
7897 : GIN_EXTRACTQUERY_PROC);
7898 :
7899 2074 : if (!OidIsValid(extractProcOid))
7900 : {
7901 : /* should not happen; throw same error as index_getprocinfo */
7902 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
7903 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
7904 : get_rel_name(index->indexoid));
7905 : }
7906 :
7907 : /*
7908 : * Choose collation to pass to extractProc (should match initGinState).
7909 : */
7910 2074 : if (OidIsValid(index->indexcollations[indexcol]))
7911 364 : collation = index->indexcollations[indexcol];
7912 : else
7913 1710 : collation = DEFAULT_COLLATION_OID;
7914 :
7915 2074 : fmgr_info(extractProcOid, &flinfo);
7916 :
7917 2074 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
7918 :
7919 2074 : FunctionCall7Coll(&flinfo,
7920 : collation,
7921 : query,
7922 : PointerGetDatum(&nentries),
7923 : UInt16GetDatum(strategy_op),
7924 : PointerGetDatum(&partial_matches),
7925 : PointerGetDatum(&extra_data),
7926 : PointerGetDatum(&nullFlags),
7927 : PointerGetDatum(&searchMode));
7928 :
7929 2074 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
7930 : {
7931 : /* No match is possible */
7932 12 : return false;
7933 : }
7934 :
7935 8804 : for (i = 0; i < nentries; i++)
7936 : {
7937 : /*
7938 : * For partial match we haven't any information to estimate number of
7939 : * matched entries in index, so, we just estimate it as 100
7940 : */
7941 6742 : if (partial_matches && partial_matches[i])
7942 316 : counts->partialEntries += 100;
7943 : else
7944 6426 : counts->exactEntries++;
7945 :
7946 6742 : counts->searchEntries++;
7947 : }
7948 :
7949 2062 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
7950 : {
7951 1590 : counts->attHasNormalScan[indexcol] = true;
7952 : }
7953 472 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
7954 : {
7955 : /* Treat "include empty" like an exact-match item */
7956 44 : counts->attHasNormalScan[indexcol] = true;
7957 44 : counts->exactEntries++;
7958 44 : counts->searchEntries++;
7959 : }
7960 : else
7961 : {
7962 : /* It's GIN_SEARCH_MODE_ALL */
7963 428 : counts->attHasFullScan[indexcol] = true;
7964 : }
7965 :
7966 2062 : return true;
7967 : }
7968 :
7969 : /*
7970 : * Estimate the number of index terms that need to be searched for while
7971 : * testing the given GIN index clause, and increment the counts in *counts
7972 : * appropriately. If the query is unsatisfiable, return false.
7973 : */
7974 : static bool
7975 2062 : gincost_opexpr(PlannerInfo *root,
7976 : IndexOptInfo *index,
7977 : int indexcol,
7978 : OpExpr *clause,
7979 : GinQualCounts *counts)
7980 : {
7981 2062 : Oid clause_op = clause->opno;
7982 2062 : Node *operand = (Node *) lsecond(clause->args);
7983 :
7984 : /* aggressively reduce to a constant, and look through relabeling */
7985 2062 : operand = estimate_expression_value(root, operand);
7986 :
7987 2062 : if (IsA(operand, RelabelType))
7988 0 : operand = (Node *) ((RelabelType *) operand)->arg;
7989 :
7990 : /*
7991 : * It's impossible to call extractQuery method for unknown operand. So
7992 : * unless operand is a Const we can't do much; just assume there will be
7993 : * one ordinary search entry from the operand at runtime.
7994 : */
7995 2062 : if (!IsA(operand, Const))
7996 : {
7997 0 : counts->exactEntries++;
7998 0 : counts->searchEntries++;
7999 0 : return true;
8000 : }
8001 :
8002 : /* If Const is null, there can be no matches */
8003 2062 : if (((Const *) operand)->constisnull)
8004 0 : return false;
8005 :
8006 : /* Otherwise, apply extractQuery and get the actual term counts */
8007 2062 : return gincost_pattern(index, indexcol, clause_op,
8008 : ((Const *) operand)->constvalue,
8009 : counts);
8010 : }
8011 :
8012 : /*
8013 : * Estimate the number of index terms that need to be searched for while
8014 : * testing the given GIN index clause, and increment the counts in *counts
8015 : * appropriately. If the query is unsatisfiable, return false.
8016 : *
8017 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
8018 : * each of which involves one value from the RHS array, plus all the
8019 : * non-array quals (if any). To model this, we average the counts across
8020 : * the RHS elements, and add the averages to the counts in *counts (which
8021 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
8022 : * by N, causing gincostestimate to scale up its estimates accordingly.
8023 : */
8024 : static bool
8025 6 : gincost_scalararrayopexpr(PlannerInfo *root,
8026 : IndexOptInfo *index,
8027 : int indexcol,
8028 : ScalarArrayOpExpr *clause,
8029 : double numIndexEntries,
8030 : GinQualCounts *counts)
8031 : {
8032 6 : Oid clause_op = clause->opno;
8033 6 : Node *rightop = (Node *) lsecond(clause->args);
8034 : ArrayType *arrayval;
8035 : int16 elmlen;
8036 : bool elmbyval;
8037 : char elmalign;
8038 : int numElems;
8039 : Datum *elemValues;
8040 : bool *elemNulls;
8041 : GinQualCounts arraycounts;
8042 6 : int numPossible = 0;
8043 : int i;
8044 :
8045 : Assert(clause->useOr);
8046 :
8047 : /* aggressively reduce to a constant, and look through relabeling */
8048 6 : rightop = estimate_expression_value(root, rightop);
8049 :
8050 6 : if (IsA(rightop, RelabelType))
8051 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
8052 :
8053 : /*
8054 : * It's impossible to call extractQuery method for unknown operand. So
8055 : * unless operand is a Const we can't do much; just assume there will be
8056 : * one ordinary search entry from each array entry at runtime, and fall
8057 : * back on a probably-bad estimate of the number of array entries.
8058 : */
8059 6 : if (!IsA(rightop, Const))
8060 : {
8061 0 : counts->exactEntries++;
8062 0 : counts->searchEntries++;
8063 0 : counts->arrayScans *= estimate_array_length(root, rightop);
8064 0 : return true;
8065 : }
8066 :
8067 : /* If Const is null, there can be no matches */
8068 6 : if (((Const *) rightop)->constisnull)
8069 0 : return false;
8070 :
8071 : /* Otherwise, extract the array elements and iterate over them */
8072 6 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
8073 6 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
8074 : &elmlen, &elmbyval, &elmalign);
8075 6 : deconstruct_array(arrayval,
8076 : ARR_ELEMTYPE(arrayval),
8077 : elmlen, elmbyval, elmalign,
8078 : &elemValues, &elemNulls, &numElems);
8079 :
8080 6 : memset(&arraycounts, 0, sizeof(arraycounts));
8081 :
8082 18 : for (i = 0; i < numElems; i++)
8083 : {
8084 : GinQualCounts elemcounts;
8085 :
8086 : /* NULL can't match anything, so ignore, as the executor will */
8087 12 : if (elemNulls[i])
8088 0 : continue;
8089 :
8090 : /* Otherwise, apply extractQuery and get the actual term counts */
8091 12 : memset(&elemcounts, 0, sizeof(elemcounts));
8092 :
8093 12 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
8094 : &elemcounts))
8095 : {
8096 : /* We ignore array elements that are unsatisfiable patterns */
8097 12 : numPossible++;
8098 :
8099 12 : if (elemcounts.attHasFullScan[indexcol] &&
8100 0 : !elemcounts.attHasNormalScan[indexcol])
8101 : {
8102 : /*
8103 : * Full index scan will be required. We treat this as if
8104 : * every key in the index had been listed in the query; is
8105 : * that reasonable?
8106 : */
8107 0 : elemcounts.partialEntries = 0;
8108 0 : elemcounts.exactEntries = numIndexEntries;
8109 0 : elemcounts.searchEntries = numIndexEntries;
8110 : }
8111 12 : arraycounts.partialEntries += elemcounts.partialEntries;
8112 12 : arraycounts.exactEntries += elemcounts.exactEntries;
8113 12 : arraycounts.searchEntries += elemcounts.searchEntries;
8114 : }
8115 : }
8116 :
8117 6 : if (numPossible == 0)
8118 : {
8119 : /* No satisfiable patterns in the array */
8120 0 : return false;
8121 : }
8122 :
8123 : /*
8124 : * Now add the averages to the global counts. This will give us an
8125 : * estimate of the average number of terms searched for in each indexscan,
8126 : * including contributions from both array and non-array quals.
8127 : */
8128 6 : counts->partialEntries += arraycounts.partialEntries / numPossible;
8129 6 : counts->exactEntries += arraycounts.exactEntries / numPossible;
8130 6 : counts->searchEntries += arraycounts.searchEntries / numPossible;
8131 :
8132 6 : counts->arrayScans *= numPossible;
8133 :
8134 6 : return true;
8135 : }
8136 :
8137 : /*
8138 : * GIN has search behavior completely different from other index types
8139 : */
8140 : void
8141 1870 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8142 : Cost *indexStartupCost, Cost *indexTotalCost,
8143 : Selectivity *indexSelectivity, double *indexCorrelation,
8144 : double *indexPages)
8145 : {
8146 1870 : IndexOptInfo *index = path->indexinfo;
8147 1870 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8148 : List *selectivityQuals;
8149 1870 : double numPages = index->pages,
8150 1870 : numTuples = index->tuples;
8151 : double numEntryPages,
8152 : numDataPages,
8153 : numPendingPages,
8154 : numEntries;
8155 : GinQualCounts counts;
8156 : bool matchPossible;
8157 : bool fullIndexScan;
8158 : double partialScale;
8159 : double entryPagesFetched,
8160 : dataPagesFetched,
8161 : dataPagesFetchedBySel;
8162 : double qual_op_cost,
8163 : qual_arg_cost,
8164 : spc_random_page_cost,
8165 : outer_scans;
8166 : Cost descentCost;
8167 : Relation indexRel;
8168 : GinStatsData ginStats;
8169 : ListCell *lc;
8170 : int i;
8171 :
8172 : /*
8173 : * Obtain statistical information from the meta page, if possible. Else
8174 : * set ginStats to zeroes, and we'll cope below.
8175 : */
8176 1870 : if (!index->hypothetical)
8177 : {
8178 : /* Lock should have already been obtained in plancat.c */
8179 1870 : indexRel = index_open(index->indexoid, NoLock);
8180 1870 : ginGetStats(indexRel, &ginStats);
8181 1870 : index_close(indexRel, NoLock);
8182 : }
8183 : else
8184 : {
8185 0 : memset(&ginStats, 0, sizeof(ginStats));
8186 : }
8187 :
8188 : /*
8189 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
8190 : * trusted, but the other fields are data as of the last VACUUM. We can
8191 : * scale them up to account for growth since then, but that method only
8192 : * goes so far; in the worst case, the stats might be for a completely
8193 : * empty index, and scaling them will produce pretty bogus numbers.
8194 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
8195 : * it's grown more than that, fall back to estimating things only from the
8196 : * assumed-accurate index size. But we'll trust nPendingPages in any case
8197 : * so long as it's not clearly insane, ie, more than the index size.
8198 : */
8199 1870 : if (ginStats.nPendingPages < numPages)
8200 1870 : numPendingPages = ginStats.nPendingPages;
8201 : else
8202 0 : numPendingPages = 0;
8203 :
8204 1870 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
8205 1870 : ginStats.nTotalPages > numPages / 4 &&
8206 1818 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
8207 1560 : {
8208 : /*
8209 : * OK, the stats seem close enough to sane to be trusted. But we
8210 : * still need to scale them by the ratio numPages / nTotalPages to
8211 : * account for growth since the last VACUUM.
8212 : */
8213 1560 : double scale = numPages / ginStats.nTotalPages;
8214 :
8215 1560 : numEntryPages = ceil(ginStats.nEntryPages * scale);
8216 1560 : numDataPages = ceil(ginStats.nDataPages * scale);
8217 1560 : numEntries = ceil(ginStats.nEntries * scale);
8218 : /* ensure we didn't round up too much */
8219 1560 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
8220 1560 : numDataPages = Min(numDataPages,
8221 : numPages - numPendingPages - numEntryPages);
8222 : }
8223 : else
8224 : {
8225 : /*
8226 : * We might get here because it's a hypothetical index, or an index
8227 : * created pre-9.1 and never vacuumed since upgrading (in which case
8228 : * its stats would read as zeroes), or just because it's grown too
8229 : * much since the last VACUUM for us to put our faith in scaling.
8230 : *
8231 : * Invent some plausible internal statistics based on the index page
8232 : * count (and clamp that to at least 10 pages, just in case). We
8233 : * estimate that 90% of the index is entry pages, and the rest is data
8234 : * pages. Estimate 100 entries per entry page; this is rather bogus
8235 : * since it'll depend on the size of the keys, but it's more robust
8236 : * than trying to predict the number of entries per heap tuple.
8237 : */
8238 310 : numPages = Max(numPages, 10);
8239 310 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
8240 310 : numDataPages = numPages - numPendingPages - numEntryPages;
8241 310 : numEntries = floor(numEntryPages * 100);
8242 : }
8243 :
8244 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
8245 1870 : if (numEntries < 1)
8246 0 : numEntries = 1;
8247 :
8248 : /*
8249 : * If the index is partial, AND the index predicate with the index-bound
8250 : * quals to produce a more accurate idea of the number of rows covered by
8251 : * the bound conditions.
8252 : */
8253 1870 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
8254 :
8255 : /* Estimate the fraction of main-table tuples that will be visited */
8256 3740 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
8257 1870 : index->rel->relid,
8258 : JOIN_INNER,
8259 : NULL);
8260 :
8261 : /* fetch estimated page cost for tablespace containing index */
8262 1870 : get_tablespace_page_costs(index->reltablespace,
8263 : &spc_random_page_cost,
8264 : NULL);
8265 :
8266 : /*
8267 : * Generic assumption about index correlation: there isn't any.
8268 : */
8269 1870 : *indexCorrelation = 0.0;
8270 :
8271 : /*
8272 : * Examine quals to estimate number of search entries & partial matches
8273 : */
8274 1870 : memset(&counts, 0, sizeof(counts));
8275 1870 : counts.arrayScans = 1;
8276 1870 : matchPossible = true;
8277 :
8278 3938 : foreach(lc, path->indexclauses)
8279 : {
8280 2068 : IndexClause *iclause = lfirst_node(IndexClause, lc);
8281 : ListCell *lc2;
8282 :
8283 4124 : foreach(lc2, iclause->indexquals)
8284 : {
8285 2068 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8286 2068 : Expr *clause = rinfo->clause;
8287 :
8288 2068 : if (IsA(clause, OpExpr))
8289 : {
8290 2062 : matchPossible = gincost_opexpr(root,
8291 : index,
8292 2062 : iclause->indexcol,
8293 : (OpExpr *) clause,
8294 : &counts);
8295 2062 : if (!matchPossible)
8296 12 : break;
8297 : }
8298 6 : else if (IsA(clause, ScalarArrayOpExpr))
8299 : {
8300 6 : matchPossible = gincost_scalararrayopexpr(root,
8301 : index,
8302 6 : iclause->indexcol,
8303 : (ScalarArrayOpExpr *) clause,
8304 : numEntries,
8305 : &counts);
8306 6 : if (!matchPossible)
8307 0 : break;
8308 : }
8309 : else
8310 : {
8311 : /* shouldn't be anything else for a GIN index */
8312 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
8313 : (int) nodeTag(clause));
8314 : }
8315 : }
8316 : }
8317 :
8318 : /* Fall out if there were any provably-unsatisfiable quals */
8319 1870 : if (!matchPossible)
8320 : {
8321 12 : *indexStartupCost = 0;
8322 12 : *indexTotalCost = 0;
8323 12 : *indexSelectivity = 0;
8324 12 : return;
8325 : }
8326 :
8327 : /*
8328 : * If attribute has a full scan and at the same time doesn't have normal
8329 : * scan, then we'll have to scan all non-null entries of that attribute.
8330 : * Currently, we don't have per-attribute statistics for GIN. Thus, we
8331 : * must assume the whole GIN index has to be scanned in this case.
8332 : */
8333 1858 : fullIndexScan = false;
8334 3606 : for (i = 0; i < index->nkeycolumns; i++)
8335 : {
8336 2086 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
8337 : {
8338 338 : fullIndexScan = true;
8339 338 : break;
8340 : }
8341 : }
8342 :
8343 1858 : if (fullIndexScan || indexQuals == NIL)
8344 : {
8345 : /*
8346 : * Full index scan will be required. We treat this as if every key in
8347 : * the index had been listed in the query; is that reasonable?
8348 : */
8349 338 : counts.partialEntries = 0;
8350 338 : counts.exactEntries = numEntries;
8351 338 : counts.searchEntries = numEntries;
8352 : }
8353 :
8354 : /* Will we have more than one iteration of a nestloop scan? */
8355 1858 : outer_scans = loop_count;
8356 :
8357 : /*
8358 : * Compute cost to begin scan, first of all, pay attention to pending
8359 : * list.
8360 : */
8361 1858 : entryPagesFetched = numPendingPages;
8362 :
8363 : /*
8364 : * Estimate number of entry pages read. We need to do
8365 : * counts.searchEntries searches. Use a power function as it should be,
8366 : * but tuples on leaf pages usually is much greater. Here we include all
8367 : * searches in entry tree, including search of first entry in partial
8368 : * match algorithm
8369 : */
8370 1858 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8371 :
8372 : /*
8373 : * Add an estimate of entry pages read by partial match algorithm. It's a
8374 : * scan over leaf pages in entry tree. We haven't any useful stats here,
8375 : * so estimate it as proportion. Because counts.partialEntries is really
8376 : * pretty bogus (see code above), it's possible that it is more than
8377 : * numEntries; clamp the proportion to ensure sanity.
8378 : */
8379 1858 : partialScale = counts.partialEntries / numEntries;
8380 1858 : partialScale = Min(partialScale, 1.0);
8381 :
8382 1858 : entryPagesFetched += ceil(numEntryPages * partialScale);
8383 :
8384 : /*
8385 : * Partial match algorithm reads all data pages before doing actual scan,
8386 : * so it's a startup cost. Again, we haven't any useful stats here, so
8387 : * estimate it as proportion.
8388 : */
8389 1858 : dataPagesFetched = ceil(numDataPages * partialScale);
8390 :
8391 1858 : *indexStartupCost = 0;
8392 1858 : *indexTotalCost = 0;
8393 :
8394 : /*
8395 : * Add a CPU-cost component to represent the costs of initial entry btree
8396 : * descent. We don't charge any I/O cost for touching upper btree levels,
8397 : * since they tend to stay in cache, but we still have to do about log2(N)
8398 : * comparisons to descend a btree of N leaf tuples. We charge one
8399 : * cpu_operator_cost per comparison.
8400 : *
8401 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
8402 : * ones after the first one are not startup cost so far as the overall
8403 : * plan is concerned, so add them only to "total" cost.
8404 : */
8405 1858 : if (numEntries > 1) /* avoid computing log(0) */
8406 : {
8407 1858 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
8408 1858 : *indexStartupCost += descentCost * counts.searchEntries;
8409 1858 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
8410 : }
8411 :
8412 : /*
8413 : * Add a cpu cost per entry-page fetched. This is not amortized over a
8414 : * loop.
8415 : */
8416 1858 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8417 1858 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8418 :
8419 : /*
8420 : * Add a cpu cost per data-page fetched. This is also not amortized over a
8421 : * loop. Since those are the data pages from the partial match algorithm,
8422 : * charge them as startup cost.
8423 : */
8424 1858 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
8425 :
8426 : /*
8427 : * Since we add the startup cost to the total cost later on, remove the
8428 : * initial arrayscan from the total.
8429 : */
8430 1858 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8431 :
8432 : /*
8433 : * Calculate cache effects if more than one scan due to nestloops or array
8434 : * quals. The result is pro-rated per nestloop scan, but the array qual
8435 : * factor shouldn't be pro-rated (compare genericcostestimate).
8436 : */
8437 1858 : if (outer_scans > 1 || counts.arrayScans > 1)
8438 : {
8439 6 : entryPagesFetched *= outer_scans * counts.arrayScans;
8440 6 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
8441 : (BlockNumber) numEntryPages,
8442 : numEntryPages, root);
8443 6 : entryPagesFetched /= outer_scans;
8444 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8445 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8446 : (BlockNumber) numDataPages,
8447 : numDataPages, root);
8448 6 : dataPagesFetched /= outer_scans;
8449 : }
8450 :
8451 : /*
8452 : * Here we use random page cost because logically-close pages could be far
8453 : * apart on disk.
8454 : */
8455 1858 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
8456 :
8457 : /*
8458 : * Now compute the number of data pages fetched during the scan.
8459 : *
8460 : * We assume every entry to have the same number of items, and that there
8461 : * is no overlap between them. (XXX: tsvector and array opclasses collect
8462 : * statistics on the frequency of individual keys; it would be nice to use
8463 : * those here.)
8464 : */
8465 1858 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
8466 :
8467 : /*
8468 : * If there is a lot of overlap among the entries, in particular if one of
8469 : * the entries is very frequent, the above calculation can grossly
8470 : * under-estimate. As a simple cross-check, calculate a lower bound based
8471 : * on the overall selectivity of the quals. At a minimum, we must read
8472 : * one item pointer for each matching entry.
8473 : *
8474 : * The width of each item pointer varies, based on the level of
8475 : * compression. We don't have statistics on that, but an average of
8476 : * around 3 bytes per item is fairly typical.
8477 : */
8478 1858 : dataPagesFetchedBySel = ceil(*indexSelectivity *
8479 1858 : (numTuples / (BLCKSZ / 3)));
8480 1858 : if (dataPagesFetchedBySel > dataPagesFetched)
8481 1478 : dataPagesFetched = dataPagesFetchedBySel;
8482 :
8483 : /* Add one page cpu-cost to the startup cost */
8484 1858 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
8485 :
8486 : /*
8487 : * Add once again a CPU-cost for those data pages, before amortizing for
8488 : * cache.
8489 : */
8490 1858 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8491 :
8492 : /* Account for cache effects, the same as above */
8493 1858 : if (outer_scans > 1 || counts.arrayScans > 1)
8494 : {
8495 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8496 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8497 : (BlockNumber) numDataPages,
8498 : numDataPages, root);
8499 6 : dataPagesFetched /= outer_scans;
8500 : }
8501 :
8502 : /* And apply random_page_cost as the cost per page */
8503 1858 : *indexTotalCost += *indexStartupCost +
8504 1858 : dataPagesFetched * spc_random_page_cost;
8505 :
8506 : /*
8507 : * Add on index qual eval costs, much as in genericcostestimate. We charge
8508 : * cpu but we can disregard indexorderbys, since GIN doesn't support
8509 : * those.
8510 : */
8511 1858 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8512 1858 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
8513 :
8514 1858 : *indexStartupCost += qual_arg_cost;
8515 1858 : *indexTotalCost += qual_arg_cost;
8516 :
8517 : /*
8518 : * Add a cpu cost per search entry, corresponding to the actual visited
8519 : * entries.
8520 : */
8521 1858 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
8522 : /* Now add a cpu cost per tuple in the posting lists / trees */
8523 1858 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
8524 1858 : *indexPages = dataPagesFetched;
8525 : }
8526 :
8527 : /*
8528 : * BRIN has search behavior completely different from other index types
8529 : */
8530 : void
8531 10730 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8532 : Cost *indexStartupCost, Cost *indexTotalCost,
8533 : Selectivity *indexSelectivity, double *indexCorrelation,
8534 : double *indexPages)
8535 : {
8536 10730 : IndexOptInfo *index = path->indexinfo;
8537 10730 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8538 10730 : double numPages = index->pages;
8539 10730 : RelOptInfo *baserel = index->rel;
8540 10730 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
8541 : Cost spc_seq_page_cost;
8542 : Cost spc_random_page_cost;
8543 : double qual_arg_cost;
8544 : double qualSelectivity;
8545 : BrinStatsData statsData;
8546 : double indexRanges;
8547 : double minimalRanges;
8548 : double estimatedRanges;
8549 : double selec;
8550 : Relation indexRel;
8551 : ListCell *l;
8552 : VariableStatData vardata;
8553 :
8554 : Assert(rte->rtekind == RTE_RELATION);
8555 :
8556 : /* fetch estimated page cost for the tablespace containing the index */
8557 10730 : get_tablespace_page_costs(index->reltablespace,
8558 : &spc_random_page_cost,
8559 : &spc_seq_page_cost);
8560 :
8561 : /*
8562 : * Obtain some data from the index itself, if possible. Otherwise invent
8563 : * some plausible internal statistics based on the relation page count.
8564 : */
8565 10730 : if (!index->hypothetical)
8566 : {
8567 : /*
8568 : * A lock should have already been obtained on the index in plancat.c.
8569 : */
8570 10730 : indexRel = index_open(index->indexoid, NoLock);
8571 10730 : brinGetStats(indexRel, &statsData);
8572 10730 : index_close(indexRel, NoLock);
8573 :
8574 : /* work out the actual number of ranges in the index */
8575 10730 : indexRanges = Max(ceil((double) baserel->pages /
8576 : statsData.pagesPerRange), 1.0);
8577 : }
8578 : else
8579 : {
8580 : /*
8581 : * Assume default number of pages per range, and estimate the number
8582 : * of ranges based on that.
8583 : */
8584 0 : indexRanges = Max(ceil((double) baserel->pages /
8585 : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
8586 :
8587 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
8588 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
8589 : }
8590 :
8591 : /*
8592 : * Compute index correlation
8593 : *
8594 : * Because we can use all index quals equally when scanning, we can use
8595 : * the largest correlation (in absolute value) among columns used by the
8596 : * query. Start at zero, the worst possible case. If we cannot find any
8597 : * correlation statistics, we will keep it as 0.
8598 : */
8599 10730 : *indexCorrelation = 0;
8600 :
8601 21462 : foreach(l, path->indexclauses)
8602 : {
8603 10732 : IndexClause *iclause = lfirst_node(IndexClause, l);
8604 10732 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
8605 :
8606 : /* attempt to lookup stats in relation for this index column */
8607 10732 : if (attnum != 0)
8608 : {
8609 : /* Simple variable -- look to stats for the underlying table */
8610 10732 : if (get_relation_stats_hook &&
8611 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
8612 : {
8613 : /*
8614 : * The hook took control of acquiring a stats tuple. If it
8615 : * did supply a tuple, it'd better have supplied a freefunc.
8616 : */
8617 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
8618 0 : elog(ERROR,
8619 : "no function provided to release variable stats with");
8620 : }
8621 : else
8622 : {
8623 10732 : vardata.statsTuple =
8624 10732 : SearchSysCache3(STATRELATTINH,
8625 : ObjectIdGetDatum(rte->relid),
8626 : Int16GetDatum(attnum),
8627 : BoolGetDatum(false));
8628 10732 : vardata.freefunc = ReleaseSysCache;
8629 : }
8630 : }
8631 : else
8632 : {
8633 : /*
8634 : * Looks like we've found an expression column in the index. Let's
8635 : * see if there's any stats for it.
8636 : */
8637 :
8638 : /* get the attnum from the 0-based index. */
8639 0 : attnum = iclause->indexcol + 1;
8640 :
8641 0 : if (get_index_stats_hook &&
8642 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
8643 : {
8644 : /*
8645 : * The hook took control of acquiring a stats tuple. If it
8646 : * did supply a tuple, it'd better have supplied a freefunc.
8647 : */
8648 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
8649 0 : !vardata.freefunc)
8650 0 : elog(ERROR, "no function provided to release variable stats with");
8651 : }
8652 : else
8653 : {
8654 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
8655 : ObjectIdGetDatum(index->indexoid),
8656 : Int16GetDatum(attnum),
8657 : BoolGetDatum(false));
8658 0 : vardata.freefunc = ReleaseSysCache;
8659 : }
8660 : }
8661 :
8662 10732 : if (HeapTupleIsValid(vardata.statsTuple))
8663 : {
8664 : AttStatsSlot sslot;
8665 :
8666 36 : if (get_attstatsslot(&sslot, vardata.statsTuple,
8667 : STATISTIC_KIND_CORRELATION, InvalidOid,
8668 : ATTSTATSSLOT_NUMBERS))
8669 : {
8670 36 : double varCorrelation = 0.0;
8671 :
8672 36 : if (sslot.nnumbers > 0)
8673 36 : varCorrelation = fabs(sslot.numbers[0]);
8674 :
8675 36 : if (varCorrelation > *indexCorrelation)
8676 36 : *indexCorrelation = varCorrelation;
8677 :
8678 36 : free_attstatsslot(&sslot);
8679 : }
8680 : }
8681 :
8682 10732 : ReleaseVariableStats(vardata);
8683 : }
8684 :
8685 10730 : qualSelectivity = clauselist_selectivity(root, indexQuals,
8686 10730 : baserel->relid,
8687 : JOIN_INNER, NULL);
8688 :
8689 : /*
8690 : * Now calculate the minimum possible ranges we could match with if all of
8691 : * the rows were in the perfect order in the table's heap.
8692 : */
8693 10730 : minimalRanges = ceil(indexRanges * qualSelectivity);
8694 :
8695 : /*
8696 : * Now estimate the number of ranges that we'll touch by using the
8697 : * indexCorrelation from the stats. Careful not to divide by zero (note
8698 : * we're using the absolute value of the correlation).
8699 : */
8700 10730 : if (*indexCorrelation < 1.0e-10)
8701 10694 : estimatedRanges = indexRanges;
8702 : else
8703 36 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
8704 :
8705 : /* we expect to visit this portion of the table */
8706 10730 : selec = estimatedRanges / indexRanges;
8707 :
8708 10730 : CLAMP_PROBABILITY(selec);
8709 :
8710 10730 : *indexSelectivity = selec;
8711 :
8712 : /*
8713 : * Compute the index qual costs, much as in genericcostestimate, to add to
8714 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
8715 : * support those.
8716 : */
8717 10730 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8718 :
8719 : /*
8720 : * Compute the startup cost as the cost to read the whole revmap
8721 : * sequentially, including the cost to execute the index quals.
8722 : */
8723 10730 : *indexStartupCost =
8724 10730 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
8725 10730 : *indexStartupCost += qual_arg_cost;
8726 :
8727 : /*
8728 : * To read a BRIN index there might be a bit of back and forth over
8729 : * regular pages, as revmap might point to them out of sequential order;
8730 : * calculate the total cost as reading the whole index in random order.
8731 : */
8732 10730 : *indexTotalCost = *indexStartupCost +
8733 10730 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
8734 :
8735 : /*
8736 : * Charge a small amount per range tuple which we expect to match to. This
8737 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
8738 : * will set a bit for each page in the range when we find a matching
8739 : * range, so we must multiply the charge by the number of pages in the
8740 : * range.
8741 : */
8742 10730 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
8743 10730 : statsData.pagesPerRange;
8744 :
8745 10730 : *indexPages = index->pages;
8746 10730 : }
|