Branch data 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) 1996-2026, PostgreSQL Global Development Group
14 : : * Portions Copyright (c) 1994, Regents of the University of California
15 : : *
16 : : *
17 : : * IDENTIFICATION
18 : : * src/backend/utils/adt/selfuncs.c
19 : : *
20 : : *-------------------------------------------------------------------------
21 : : */
22 : :
23 : : /*----------
24 : : * Operator selectivity estimation functions are called to estimate the
25 : : * selectivity of WHERE clauses whose top-level operator is their operator.
26 : : * We divide the problem into two cases:
27 : : * Restriction clause estimation: the clause involves vars of just
28 : : * one relation.
29 : : * Join clause estimation: the clause involves vars of multiple rels.
30 : : * Join selectivity estimation is far more difficult and usually less accurate
31 : : * than restriction estimation.
32 : : *
33 : : * When dealing with the inner scan of a nestloop join, we consider the
34 : : * join's joinclauses as restriction clauses for the inner relation, and
35 : : * treat vars of the outer relation as parameters (a/k/a constants of unknown
36 : : * values). So, restriction estimators need to be able to accept an argument
37 : : * telling which relation is to be treated as the variable.
38 : : *
39 : : * The call convention for a restriction estimator (oprrest function) is
40 : : *
41 : : * Selectivity oprrest (PlannerInfo *root,
42 : : * Oid operator,
43 : : * List *args,
44 : : * int varRelid);
45 : : *
46 : : * root: general information about the query (rtable and RelOptInfo lists
47 : : * are particularly important for the estimator).
48 : : * operator: OID of the specific operator in question.
49 : : * args: argument list from the operator clause.
50 : : * varRelid: if not zero, the relid (rtable index) of the relation to
51 : : * be treated as the variable relation. May be zero if the args list
52 : : * is known to contain vars of only one relation.
53 : : *
54 : : * This is represented at the SQL level (in pg_proc) as
55 : : *
56 : : * float8 oprrest (internal, oid, internal, int4);
57 : : *
58 : : * The result is a selectivity, that is, a fraction (0 to 1) of the rows
59 : : * of the relation that are expected to produce a TRUE result for the
60 : : * given operator.
61 : : *
62 : : * The call convention for a join estimator (oprjoin function) is similar
63 : : * except that varRelid is not needed, and instead join information is
64 : : * supplied:
65 : : *
66 : : * Selectivity oprjoin (PlannerInfo *root,
67 : : * Oid operator,
68 : : * List *args,
69 : : * JoinType jointype,
70 : : * SpecialJoinInfo *sjinfo);
71 : : *
72 : : * float8 oprjoin (internal, oid, internal, int2, internal);
73 : : *
74 : : * (Before Postgres 8.4, join estimators had only the first four of these
75 : : * parameters. That signature is still allowed, but deprecated.) The
76 : : * relationship between jointype and sjinfo is explained in the comments for
77 : : * clause_selectivity() --- the short version is that jointype is usually
78 : : * best ignored in favor of examining sjinfo.
79 : : *
80 : : * Join selectivity for regular inner and outer joins is defined as the
81 : : * fraction (0 to 1) of the cross product of the relations that is expected
82 : : * to produce a TRUE result for the given operator. For both semi and anti
83 : : * joins, however, the selectivity is defined as the fraction of the left-hand
84 : : * side relation's rows that are expected to have a match (ie, at least one
85 : : * row with a TRUE result) in the right-hand side.
86 : : *
87 : : * For both oprrest and oprjoin functions, the operator's input collation OID
88 : : * (if any) is passed using the standard fmgr mechanism, so that the estimator
89 : : * function can fetch it with PG_GET_COLLATION(). Note, however, that all
90 : : * statistics in pg_statistic are currently built using the relevant column's
91 : : * collation.
92 : : *----------
93 : : */
94 : :
95 : : #include "postgres.h"
96 : :
97 : : #include <ctype.h>
98 : : #include <math.h>
99 : :
100 : : #include "access/brin.h"
101 : : #include "access/brin_page.h"
102 : : #include "access/gin.h"
103 : : #include "access/table.h"
104 : : #include "access/tableam.h"
105 : : #include "access/visibilitymap.h"
106 : : #include "catalog/pg_collation.h"
107 : : #include "catalog/pg_operator.h"
108 : : #include "catalog/pg_statistic.h"
109 : : #include "catalog/pg_statistic_ext.h"
110 : : #include "executor/nodeAgg.h"
111 : : #include "miscadmin.h"
112 : : #include "nodes/makefuncs.h"
113 : : #include "nodes/nodeFuncs.h"
114 : : #include "optimizer/clauses.h"
115 : : #include "optimizer/cost.h"
116 : : #include "optimizer/optimizer.h"
117 : : #include "optimizer/pathnode.h"
118 : : #include "optimizer/paths.h"
119 : : #include "optimizer/plancat.h"
120 : : #include "parser/parse_clause.h"
121 : : #include "parser/parse_relation.h"
122 : : #include "parser/parsetree.h"
123 : : #include "rewrite/rewriteManip.h"
124 : : #include "statistics/statistics.h"
125 : : #include "storage/bufmgr.h"
126 : : #include "utils/acl.h"
127 : : #include "utils/array.h"
128 : : #include "utils/builtins.h"
129 : : #include "utils/date.h"
130 : : #include "utils/datum.h"
131 : : #include "utils/fmgroids.h"
132 : : #include "utils/index_selfuncs.h"
133 : : #include "utils/lsyscache.h"
134 : : #include "utils/memutils.h"
135 : : #include "utils/pg_locale.h"
136 : : #include "utils/rel.h"
137 : : #include "utils/selfuncs.h"
138 : : #include "utils/snapmgr.h"
139 : : #include "utils/spccache.h"
140 : : #include "utils/syscache.h"
141 : : #include "utils/timestamp.h"
142 : : #include "utils/typcache.h"
143 : :
144 : : #define DEFAULT_PAGE_CPU_MULTIPLIER 50.0
145 : :
146 : : /*
147 : : * In production builds, switch to hash-based MCV matching when the lists are
148 : : * large enough to amortize hash setup cost. (This threshold is compared to
149 : : * the sum of the lengths of the two MCV lists. This is simplistic but seems
150 : : * to work well enough.) In debug builds, we use a smaller threshold so that
151 : : * the regression tests cover both paths well.
152 : : */
153 : : #ifndef USE_ASSERT_CHECKING
154 : : #define EQJOINSEL_MCV_HASH_THRESHOLD 200
155 : : #else
156 : : #define EQJOINSEL_MCV_HASH_THRESHOLD 20
157 : : #endif
158 : :
159 : : /* Entries in the simplehash hash table used by eqjoinsel_find_matches */
160 : : typedef struct MCVHashEntry
161 : : {
162 : : Datum value; /* the value represented by this entry */
163 : : int index; /* its index in the relevant AttStatsSlot */
164 : : uint32 hash; /* hash code for the Datum */
165 : : char status; /* status code used by simplehash.h */
166 : : } MCVHashEntry;
167 : :
168 : : /* private_data for the simplehash hash table */
169 : : typedef struct MCVHashContext
170 : : {
171 : : FunctionCallInfo equal_fcinfo; /* the equality join operator */
172 : : FunctionCallInfo hash_fcinfo; /* the hash function to use */
173 : : bool op_is_reversed; /* equality compares hash type to probe type */
174 : : bool insert_mode; /* doing inserts or lookups? */
175 : : bool hash_typbyval; /* typbyval of hashed data type */
176 : : int16 hash_typlen; /* typlen of hashed data type */
177 : : } MCVHashContext;
178 : :
179 : : /* forward reference */
180 : : typedef struct MCVHashTable_hash MCVHashTable_hash;
181 : :
182 : : /* Hooks for plugins to get control when we ask for stats */
183 : : get_relation_stats_hook_type get_relation_stats_hook = NULL;
184 : : get_index_stats_hook_type get_index_stats_hook = NULL;
185 : :
186 : : static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
187 : : static double eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
188 : : Oid hashLeft, Oid hashRight,
189 : : VariableStatData *vardata1, VariableStatData *vardata2,
190 : : double nd1, double nd2,
191 : : bool isdefault1, bool isdefault2,
192 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
193 : : Form_pg_statistic stats1, Form_pg_statistic stats2,
194 : : bool have_mcvs1, bool have_mcvs2,
195 : : bool *hasmatch1, bool *hasmatch2,
196 : : int *p_nmatches);
197 : : static double eqjoinsel_semi(FmgrInfo *eqproc, Oid collation,
198 : : Oid hashLeft, Oid hashRight,
199 : : bool op_is_reversed,
200 : : VariableStatData *vardata1, VariableStatData *vardata2,
201 : : double nd1, double nd2,
202 : : bool isdefault1, bool isdefault2,
203 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
204 : : Form_pg_statistic stats1, Form_pg_statistic stats2,
205 : : bool have_mcvs1, bool have_mcvs2,
206 : : bool *hasmatch1, bool *hasmatch2,
207 : : int *p_nmatches,
208 : : RelOptInfo *inner_rel);
209 : : static void eqjoinsel_find_matches(FmgrInfo *eqproc, Oid collation,
210 : : Oid hashLeft, Oid hashRight,
211 : : bool op_is_reversed,
212 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
213 : : int nvalues1, int nvalues2,
214 : : bool *hasmatch1, bool *hasmatch2,
215 : : int *p_nmatches, double *p_matchprodfreq);
216 : : static uint32 hash_mcv(MCVHashTable_hash *tab, Datum key);
217 : : static bool mcvs_equal(MCVHashTable_hash *tab, Datum key0, Datum key1);
218 : : static bool estimate_multivariate_ndistinct(PlannerInfo *root,
219 : : RelOptInfo *rel, List **varinfos, double *ndistinct);
220 : : static bool convert_to_scalar(Datum value, Oid valuetypid, Oid collid,
221 : : double *scaledvalue,
222 : : Datum lobound, Datum hibound, Oid boundstypid,
223 : : double *scaledlobound, double *scaledhibound);
224 : : static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure);
225 : : static void convert_string_to_scalar(char *value,
226 : : double *scaledvalue,
227 : : char *lobound,
228 : : double *scaledlobound,
229 : : char *hibound,
230 : : double *scaledhibound);
231 : : static void convert_bytea_to_scalar(Datum value,
232 : : double *scaledvalue,
233 : : Datum lobound,
234 : : double *scaledlobound,
235 : : Datum hibound,
236 : : double *scaledhibound);
237 : : static double convert_one_string_to_scalar(char *value,
238 : : int rangelo, int rangehi);
239 : : static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
240 : : int rangelo, int rangehi);
241 : : static char *convert_string_datum(Datum value, Oid typid, Oid collid,
242 : : bool *failure);
243 : : static double convert_timevalue_to_scalar(Datum value, Oid typid,
244 : : bool *failure);
245 : : static Node *strip_all_phvs_deep(PlannerInfo *root, Node *node);
246 : : static bool contain_placeholder_walker(Node *node, void *context);
247 : : static Node *strip_all_phvs_mutator(Node *node, void *context);
248 : : static void examine_simple_variable(PlannerInfo *root, Var *var,
249 : : VariableStatData *vardata);
250 : : static void adjust_statstuple_for_grouping(PlannerInfo *subroot, Var *var,
251 : : VariableStatData *vardata);
252 : : static void examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
253 : : int indexcol, VariableStatData *vardata);
254 : : static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata,
255 : : Oid sortop, Oid collation,
256 : : Datum *min, Datum *max);
257 : : static void get_stats_slot_range(AttStatsSlot *sslot,
258 : : Oid opfuncoid, FmgrInfo *opproc,
259 : : Oid collation, int16 typLen, bool typByVal,
260 : : Datum *min, Datum *max, bool *p_have_data);
261 : : static bool get_actual_variable_range(PlannerInfo *root,
262 : : VariableStatData *vardata,
263 : : Oid sortop, Oid collation,
264 : : Datum *min, Datum *max);
265 : : static bool get_actual_variable_endpoint(Relation heapRel,
266 : : Relation indexRel,
267 : : ScanDirection indexscandir,
268 : : ScanKey scankeys,
269 : : int16 typLen,
270 : : bool typByVal,
271 : : TupleTableSlot *tableslot,
272 : : MemoryContext outercontext,
273 : : Datum *endpointDatum);
274 : : static RelOptInfo *find_join_input_rel(PlannerInfo *root, Relids relids);
275 : : static double btcost_correlation(IndexOptInfo *index,
276 : : VariableStatData *vardata);
277 : :
278 : : /* Define support routines for MCV hash tables */
279 : : #define SH_PREFIX MCVHashTable
280 : : #define SH_ELEMENT_TYPE MCVHashEntry
281 : : #define SH_KEY_TYPE Datum
282 : : #define SH_KEY value
283 : : #define SH_HASH_KEY(tab,key) hash_mcv(tab, key)
284 : : #define SH_EQUAL(tab,key0,key1) mcvs_equal(tab, key0, key1)
285 : : #define SH_SCOPE static inline
286 : : #define SH_STORE_HASH
287 : : #define SH_GET_HASH(tab,ent) (ent)->hash
288 : : #define SH_DEFINE
289 : : #define SH_DECLARE
290 : : #include "lib/simplehash.h"
291 : :
292 : :
293 : : /*
294 : : * eqsel - Selectivity of "=" for any data types.
295 : : *
296 : : * Note: this routine is also used to estimate selectivity for some
297 : : * operators that are not "=" but have comparable selectivity behavior,
298 : : * such as "~=" (geometric approximate-match). Even for "=", we must
299 : : * keep in mind that the left and right datatypes may differ.
300 : : */
301 : : Datum
302 : 541656 : eqsel(PG_FUNCTION_ARGS)
303 : : {
304 : 541656 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
305 : : }
306 : :
307 : : /*
308 : : * Common code for eqsel() and neqsel()
309 : : */
310 : : static double
311 : 571912 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
312 : : {
313 : 571912 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
314 : 571912 : Oid operator = PG_GETARG_OID(1);
315 : 571912 : List *args = (List *) PG_GETARG_POINTER(2);
316 : 571912 : int varRelid = PG_GETARG_INT32(3);
317 : 571912 : Oid collation = PG_GET_COLLATION();
318 : : VariableStatData vardata;
319 : : Node *other;
320 : : bool varonleft;
321 : : double selec;
322 : :
323 : : /*
324 : : * When asked about <>, we do the estimation using the corresponding =
325 : : * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
326 : : */
327 [ + + ]: 571912 : if (negate)
328 : : {
329 : 30256 : operator = get_negator(operator);
330 [ - + ]: 30256 : if (!OidIsValid(operator))
331 : : {
332 : : /* Use default selectivity (should we raise an error instead?) */
333 : 0 : return 1.0 - DEFAULT_EQ_SEL;
334 : : }
335 : : }
336 : :
337 : : /*
338 : : * If expression is not variable = something or something = variable, then
339 : : * punt and return a default estimate.
340 : : */
341 [ + + ]: 571912 : if (!get_restriction_variable(root, args, varRelid,
342 : : &vardata, &other, &varonleft))
343 [ + + ]: 3345 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
344 : :
345 : : /*
346 : : * We can do a lot better if the something is a constant. (Note: the
347 : : * Const might result from estimation rather than being a simple constant
348 : : * in the query.)
349 : : */
350 [ + + ]: 568563 : if (IsA(other, Const))
351 : 221451 : selec = var_eq_const(&vardata, operator, collation,
352 : 221451 : ((Const *) other)->constvalue,
353 : 221451 : ((Const *) other)->constisnull,
354 : : varonleft, negate);
355 : : else
356 : 347112 : selec = var_eq_non_const(&vardata, operator, collation, other,
357 : : varonleft, negate);
358 : :
359 [ + + ]: 568563 : ReleaseVariableStats(vardata);
360 : :
361 : 568563 : return selec;
362 : : }
363 : :
364 : : /*
365 : : * var_eq_const --- eqsel for var = const case
366 : : *
367 : : * This is exported so that some other estimation functions can use it.
368 : : */
369 : : double
370 : 250239 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
371 : : Datum constval, bool constisnull,
372 : : bool varonleft, bool negate)
373 : : {
374 : : double selec;
375 : 250239 : double nullfrac = 0.0;
376 : : bool isdefault;
377 : : Oid opfuncoid;
378 : :
379 : : /*
380 : : * If the constant is NULL, assume operator is strict and return zero, ie,
381 : : * operator will never return TRUE. (It's zero even for a negator op.)
382 : : */
383 [ + + ]: 250239 : if (constisnull)
384 : 271 : return 0.0;
385 : :
386 : : /*
387 : : * Grab the nullfrac for use below. Note we allow use of nullfrac
388 : : * regardless of security check.
389 : : */
390 [ + + ]: 249968 : if (HeapTupleIsValid(vardata->statsTuple))
391 : : {
392 : : Form_pg_statistic stats;
393 : :
394 : 176351 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
395 : 176351 : nullfrac = stats->stanullfrac;
396 : : }
397 : :
398 : : /*
399 : : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
400 : : * assume there is exactly one match regardless of anything else. (This
401 : : * is slightly bogus, since the index or clause's equality operator might
402 : : * be different from ours, but it's much more likely to be right than
403 : : * ignoring the information.)
404 : : */
405 [ + + + - : 249968 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
+ + ]
406 : : {
407 : 51370 : selec = 1.0 / vardata->rel->tuples;
408 : : }
409 [ + + + - ]: 333118 : else if (HeapTupleIsValid(vardata->statsTuple) &&
410 : 134520 : statistic_proc_security_check(vardata,
411 : 134520 : (opfuncoid = get_opcode(oproid))))
412 : 134520 : {
413 : : AttStatsSlot sslot;
414 : 134520 : bool match = false;
415 : : int i;
416 : :
417 : : /*
418 : : * Is the constant "=" to any of the column's most common values?
419 : : * (Although the given operator may not really be "=", we will assume
420 : : * that seeing whether it returns TRUE is an appropriate test. If you
421 : : * don't like this, maybe you shouldn't be using eqsel for your
422 : : * operator...)
423 : : */
424 [ + + ]: 134520 : if (get_attstatsslot(&sslot, vardata->statsTuple,
425 : : STATISTIC_KIND_MCV, InvalidOid,
426 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
427 : : {
428 : 117713 : LOCAL_FCINFO(fcinfo, 2);
429 : : FmgrInfo eqproc;
430 : :
431 : 117713 : fmgr_info(opfuncoid, &eqproc);
432 : :
433 : : /*
434 : : * Save a few cycles by setting up the fcinfo struct just once.
435 : : * Using FunctionCallInvoke directly also avoids failure if the
436 : : * eqproc returns NULL, though really equality functions should
437 : : * never do that.
438 : : */
439 : 117713 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
440 : : NULL, NULL);
441 : 117713 : fcinfo->args[0].isnull = false;
442 : 117713 : fcinfo->args[1].isnull = false;
443 : : /* be careful to apply operator right way 'round */
444 [ + + ]: 117713 : if (varonleft)
445 : 117687 : fcinfo->args[1].value = constval;
446 : : else
447 : 26 : fcinfo->args[0].value = constval;
448 : :
449 [ + + ]: 2050356 : for (i = 0; i < sslot.nvalues; i++)
450 : : {
451 : : Datum fresult;
452 : :
453 [ + + ]: 1991230 : if (varonleft)
454 : 1991184 : fcinfo->args[0].value = sslot.values[i];
455 : : else
456 : 46 : fcinfo->args[1].value = sslot.values[i];
457 : 1991230 : fcinfo->isnull = false;
458 : 1991230 : fresult = FunctionCallInvoke(fcinfo);
459 [ + - + + ]: 1991230 : if (!fcinfo->isnull && DatumGetBool(fresult))
460 : : {
461 : 58587 : match = true;
462 : 58587 : break;
463 : : }
464 : : }
465 : : }
466 : : else
467 : : {
468 : : /* no most-common-value info available */
469 : 16807 : i = 0; /* keep compiler quiet */
470 : : }
471 : :
472 [ + + ]: 134520 : if (match)
473 : : {
474 : : /*
475 : : * Constant is "=" to this common value. We know selectivity
476 : : * exactly (or as exactly as ANALYZE could calculate it, anyway).
477 : : */
478 : 58587 : selec = sslot.numbers[i];
479 : : }
480 : : else
481 : : {
482 : : /*
483 : : * Comparison is against a constant that is neither NULL nor any
484 : : * of the common values. Its selectivity cannot be more than
485 : : * this:
486 : : */
487 : 75933 : double sumcommon = 0.0;
488 : : double otherdistinct;
489 : :
490 [ + + ]: 1755084 : for (i = 0; i < sslot.nnumbers; i++)
491 : 1679151 : sumcommon += sslot.numbers[i];
492 : 75933 : selec = 1.0 - sumcommon - nullfrac;
493 [ + + - + ]: 75933 : CLAMP_PROBABILITY(selec);
494 : :
495 : : /*
496 : : * and in fact it's probably a good deal less. We approximate that
497 : : * all the not-common values share this remaining fraction
498 : : * equally, so we divide by the number of other distinct values.
499 : : */
500 : 75933 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
501 : 75933 : sslot.nnumbers;
502 [ + + ]: 75933 : if (otherdistinct > 1)
503 : 38364 : selec /= otherdistinct;
504 : :
505 : : /*
506 : : * Another cross-check: selectivity shouldn't be estimated as more
507 : : * than the least common "most common value".
508 : : */
509 [ + + - + ]: 75933 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
510 : 0 : selec = sslot.numbers[sslot.nnumbers - 1];
511 : : }
512 : :
513 : 134520 : free_attstatsslot(&sslot);
514 : : }
515 : : else
516 : : {
517 : : /*
518 : : * No ANALYZE stats available, so make a guess using estimated number
519 : : * of distinct values and assuming they are equally common. (The guess
520 : : * is unlikely to be very good, but we do know a few special cases.)
521 : : */
522 : 64078 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
523 : : }
524 : :
525 : : /* now adjust if we wanted <> rather than = */
526 [ + + ]: 249968 : if (negate)
527 : 24068 : selec = 1.0 - selec - nullfrac;
528 : :
529 : : /* result should be in range, but make sure... */
530 [ - + - + ]: 249968 : CLAMP_PROBABILITY(selec);
531 : :
532 : 249968 : return selec;
533 : : }
534 : :
535 : : /*
536 : : * var_eq_non_const --- eqsel for var = something-other-than-const case
537 : : *
538 : : * This is exported so that some other estimation functions can use it.
539 : : */
540 : : double
541 : 347112 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
542 : : Node *other,
543 : : bool varonleft, bool negate)
544 : : {
545 : : double selec;
546 : 347112 : double nullfrac = 0.0;
547 : : bool isdefault;
548 : :
549 : : /*
550 : : * Grab the nullfrac for use below.
551 : : */
552 [ + + ]: 347112 : if (HeapTupleIsValid(vardata->statsTuple))
553 : : {
554 : : Form_pg_statistic stats;
555 : :
556 : 203752 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
557 : 203752 : nullfrac = stats->stanullfrac;
558 : : }
559 : :
560 : : /*
561 : : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
562 : : * assume there is exactly one match regardless of anything else. (This
563 : : * is slightly bogus, since the index or clause's equality operator might
564 : : * be different from ours, but it's much more likely to be right than
565 : : * ignoring the information.)
566 : : */
567 [ + + + - : 347112 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
+ + ]
568 : : {
569 : 125083 : selec = 1.0 / vardata->rel->tuples;
570 : : }
571 [ + + ]: 222029 : else if (HeapTupleIsValid(vardata->statsTuple))
572 : : {
573 : : double ndistinct;
574 : : AttStatsSlot sslot;
575 : :
576 : : /*
577 : : * Search is for a value that we do not know a priori, but we will
578 : : * assume it is not NULL. Estimate the selectivity as non-null
579 : : * fraction divided by number of distinct values, so that we get a
580 : : * result averaged over all possible values whether common or
581 : : * uncommon. (Essentially, we are assuming that the not-yet-known
582 : : * comparison value is equally likely to be any of the possible
583 : : * values, regardless of their frequency in the table. Is that a good
584 : : * idea?)
585 : : */
586 : 99288 : selec = 1.0 - nullfrac;
587 : 99288 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
588 [ + + ]: 99288 : if (ndistinct > 1)
589 : 96833 : selec /= ndistinct;
590 : :
591 : : /*
592 : : * Cross-check: selectivity should never be estimated as more than the
593 : : * most common value's.
594 : : */
595 [ + + ]: 99288 : if (get_attstatsslot(&sslot, vardata->statsTuple,
596 : : STATISTIC_KIND_MCV, InvalidOid,
597 : : ATTSTATSSLOT_NUMBERS))
598 : : {
599 [ + - + + ]: 83654 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
600 : 483 : selec = sslot.numbers[0];
601 : 83654 : free_attstatsslot(&sslot);
602 : : }
603 : : }
604 : : else
605 : : {
606 : : /*
607 : : * No ANALYZE stats available, so make a guess using estimated number
608 : : * of distinct values and assuming they are equally common. (The guess
609 : : * is unlikely to be very good, but we do know a few special cases.)
610 : : */
611 : 122741 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
612 : : }
613 : :
614 : : /* now adjust if we wanted <> rather than = */
615 [ + + ]: 347112 : if (negate)
616 : 4740 : selec = 1.0 - selec - nullfrac;
617 : :
618 : : /* result should be in range, but make sure... */
619 [ - + - + ]: 347112 : CLAMP_PROBABILITY(selec);
620 : :
621 : 347112 : return selec;
622 : : }
623 : :
624 : : /*
625 : : * neqsel - Selectivity of "!=" for any data types.
626 : : *
627 : : * This routine is also used for some operators that are not "!="
628 : : * but have comparable selectivity behavior. See above comments
629 : : * for eqsel().
630 : : */
631 : : Datum
632 : 30256 : neqsel(PG_FUNCTION_ARGS)
633 : : {
634 : 30256 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
635 : : }
636 : :
637 : : /*
638 : : * scalarineqsel - Selectivity of "<", "<=", ">", ">=" for scalars.
639 : : *
640 : : * This is the guts of scalarltsel/scalarlesel/scalargtsel/scalargesel.
641 : : * The isgt and iseq flags distinguish which of the four cases apply.
642 : : *
643 : : * The caller has commuted the clause, if necessary, so that we can treat
644 : : * the variable as being on the left. The caller must also make sure that
645 : : * the other side of the clause is a non-null Const, and dissect that into
646 : : * a value and datatype. (This definition simplifies some callers that
647 : : * want to estimate against a computed value instead of a Const node.)
648 : : *
649 : : * This routine works for any datatype (or pair of datatypes) known to
650 : : * convert_to_scalar(). If it is applied to some other datatype,
651 : : * it will return an approximate estimate based on assuming that the constant
652 : : * value falls in the middle of the bin identified by binary search.
653 : : */
654 : : static double
655 : 238963 : scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
656 : : Oid collation,
657 : : VariableStatData *vardata, Datum constval, Oid consttype)
658 : : {
659 : : Form_pg_statistic stats;
660 : : FmgrInfo opproc;
661 : : double mcv_selec,
662 : : hist_selec,
663 : : sumcommon;
664 : : double selec;
665 : :
666 [ + + ]: 238963 : if (!HeapTupleIsValid(vardata->statsTuple))
667 : : {
668 : : /*
669 : : * No stats are available. Typically this means we have to fall back
670 : : * on the default estimate; but if the variable is CTID then we can
671 : : * make an estimate based on comparing the constant to the table size.
672 : : */
673 [ + - + + ]: 21366 : if (vardata->var && IsA(vardata->var, Var) &&
674 [ + + ]: 17338 : ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
675 : : {
676 : : ItemPointer itemptr;
677 : : double block;
678 : : double density;
679 : :
680 : : /*
681 : : * If the relation's empty, we're going to include all of it.
682 : : * (This is mostly to avoid divide-by-zero below.)
683 : : */
684 [ - + ]: 1679 : if (vardata->rel->pages == 0)
685 : 0 : return 1.0;
686 : :
687 : 1679 : itemptr = (ItemPointer) DatumGetPointer(constval);
688 : 1679 : block = ItemPointerGetBlockNumberNoCheck(itemptr);
689 : :
690 : : /*
691 : : * Determine the average number of tuples per page (density).
692 : : *
693 : : * Since the last page will, on average, be only half full, we can
694 : : * estimate it to have half as many tuples as earlier pages. So
695 : : * give it half the weight of a regular page.
696 : : */
697 : 1679 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
698 : :
699 : : /* If target is the last page, use half the density. */
700 [ + + ]: 1679 : if (block >= vardata->rel->pages - 1)
701 : 25 : density *= 0.5;
702 : :
703 : : /*
704 : : * Using the average tuples per page, calculate how far into the
705 : : * page the itemptr is likely to be and adjust block accordingly,
706 : : * by adding that fraction of a whole block (but never more than a
707 : : * whole block, no matter how high the itemptr's offset is). Here
708 : : * we are ignoring the possibility of dead-tuple line pointers,
709 : : * which is fairly bogus, but we lack the info to do better.
710 : : */
711 [ + - ]: 1679 : if (density > 0.0)
712 : : {
713 : 1679 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
714 : :
715 [ + + ]: 1679 : block += Min(offset / density, 1.0);
716 : : }
717 : :
718 : : /*
719 : : * Convert relative block number to selectivity. Again, the last
720 : : * page has only half weight.
721 : : */
722 : 1679 : selec = block / (vardata->rel->pages - 0.5);
723 : :
724 : : /*
725 : : * The calculation so far gave us a selectivity for the "<=" case.
726 : : * We'll have one fewer tuple for "<" and one additional tuple for
727 : : * ">=", the latter of which we'll reverse the selectivity for
728 : : * below, so we can simply subtract one tuple for both cases. The
729 : : * cases that need this adjustment can be identified by iseq being
730 : : * equal to isgt.
731 : : */
732 [ + + + - ]: 1679 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
733 : 1562 : selec -= (1.0 / vardata->rel->tuples);
734 : :
735 : : /* Finally, reverse the selectivity for the ">", ">=" cases. */
736 [ + + ]: 1679 : if (isgt)
737 : 1547 : selec = 1.0 - selec;
738 : :
739 [ + + - + ]: 1679 : CLAMP_PROBABILITY(selec);
740 : 1679 : return selec;
741 : : }
742 : :
743 : : /* no stats available, so default result */
744 : 19687 : return DEFAULT_INEQ_SEL;
745 : : }
746 : 217597 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
747 : :
748 : 217597 : fmgr_info(get_opcode(operator), &opproc);
749 : :
750 : : /*
751 : : * If we have most-common-values info, add up the fractions of the MCV
752 : : * entries that satisfy MCV OP CONST. These fractions contribute directly
753 : : * to the result selectivity. Also add up the total fraction represented
754 : : * by MCV entries.
755 : : */
756 : 217597 : mcv_selec = mcv_selectivity(vardata, &opproc, collation, constval, true,
757 : : &sumcommon);
758 : :
759 : : /*
760 : : * If there is a histogram, determine which bin the constant falls in, and
761 : : * compute the resulting contribution to selectivity.
762 : : */
763 : 217597 : hist_selec = ineq_histogram_selectivity(root, vardata,
764 : : operator, &opproc, isgt, iseq,
765 : : collation,
766 : : constval, consttype);
767 : :
768 : : /*
769 : : * Now merge the results from the MCV and histogram calculations,
770 : : * realizing that the histogram covers only the non-null values that are
771 : : * not listed in MCV.
772 : : */
773 : 217597 : selec = 1.0 - stats->stanullfrac - sumcommon;
774 : :
775 [ + + ]: 217597 : if (hist_selec >= 0.0)
776 : 130514 : selec *= hist_selec;
777 : : else
778 : : {
779 : : /*
780 : : * If no histogram but there are values not accounted for by MCV,
781 : : * arbitrarily assume half of them will match.
782 : : */
783 : 87083 : selec *= 0.5;
784 : : }
785 : :
786 : 217597 : selec += mcv_selec;
787 : :
788 : : /* result should be in range, but make sure... */
789 [ + + + + ]: 217597 : CLAMP_PROBABILITY(selec);
790 : :
791 : 217597 : return selec;
792 : : }
793 : :
794 : : /*
795 : : * mcv_selectivity - Examine the MCV list for selectivity estimates
796 : : *
797 : : * Determine the fraction of the variable's MCV population that satisfies
798 : : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft. Also
799 : : * compute the fraction of the total column population represented by the MCV
800 : : * list. This code will work for any boolean-returning predicate operator.
801 : : *
802 : : * The function result is the MCV selectivity, and the fraction of the
803 : : * total population is returned into *sumcommonp. Zeroes are returned
804 : : * if there is no MCV list.
805 : : */
806 : : double
807 : 221944 : mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation,
808 : : Datum constval, bool varonleft,
809 : : double *sumcommonp)
810 : : {
811 : : double mcv_selec,
812 : : sumcommon;
813 : : AttStatsSlot sslot;
814 : : int i;
815 : :
816 : 221944 : mcv_selec = 0.0;
817 : 221944 : sumcommon = 0.0;
818 : :
819 [ + + + + ]: 442085 : if (HeapTupleIsValid(vardata->statsTuple) &&
820 [ + + ]: 440007 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
821 : 219866 : get_attstatsslot(&sslot, vardata->statsTuple,
822 : : STATISTIC_KIND_MCV, InvalidOid,
823 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
824 : : {
825 : 123313 : LOCAL_FCINFO(fcinfo, 2);
826 : :
827 : : /*
828 : : * We invoke the opproc "by hand" so that we won't fail on NULL
829 : : * results. Such cases won't arise for normal comparison functions,
830 : : * but generic_restriction_selectivity could perhaps be used with
831 : : * operators that can return NULL. A small side benefit is to not
832 : : * need to re-initialize the fcinfo struct from scratch each time.
833 : : */
834 : 123313 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
835 : : NULL, NULL);
836 : 123313 : fcinfo->args[0].isnull = false;
837 : 123313 : fcinfo->args[1].isnull = false;
838 : : /* be careful to apply operator right way 'round */
839 [ + - ]: 123313 : if (varonleft)
840 : 123313 : fcinfo->args[1].value = constval;
841 : : else
842 : 0 : fcinfo->args[0].value = constval;
843 : :
844 [ + + ]: 2734115 : for (i = 0; i < sslot.nvalues; i++)
845 : : {
846 : : Datum fresult;
847 : :
848 [ + - ]: 2610802 : if (varonleft)
849 : 2610802 : fcinfo->args[0].value = sslot.values[i];
850 : : else
851 : 0 : fcinfo->args[1].value = sslot.values[i];
852 : 2610802 : fcinfo->isnull = false;
853 : 2610802 : fresult = FunctionCallInvoke(fcinfo);
854 [ + - + + ]: 2610802 : if (!fcinfo->isnull && DatumGetBool(fresult))
855 : 1056223 : mcv_selec += sslot.numbers[i];
856 : 2610802 : sumcommon += sslot.numbers[i];
857 : : }
858 : 123313 : free_attstatsslot(&sslot);
859 : : }
860 : :
861 : 221944 : *sumcommonp = sumcommon;
862 : 221944 : return mcv_selec;
863 : : }
864 : :
865 : : /*
866 : : * histogram_selectivity - Examine the histogram for selectivity estimates
867 : : *
868 : : * Determine the fraction of the variable's histogram entries that satisfy
869 : : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft.
870 : : *
871 : : * This code will work for any boolean-returning predicate operator, whether
872 : : * or not it has anything to do with the histogram sort operator. We are
873 : : * essentially using the histogram just as a representative sample. However,
874 : : * small histograms are unlikely to be all that representative, so the caller
875 : : * should be prepared to fall back on some other estimation approach when the
876 : : * histogram is missing or very small. It may also be prudent to combine this
877 : : * approach with another one when the histogram is small.
878 : : *
879 : : * If the actual histogram size is not at least min_hist_size, we won't bother
880 : : * to do the calculation at all. Also, if the n_skip parameter is > 0, we
881 : : * ignore the first and last n_skip histogram elements, on the grounds that
882 : : * they are outliers and hence not very representative. Typical values for
883 : : * these parameters are 10 and 1.
884 : : *
885 : : * The function result is the selectivity, or -1 if there is no histogram
886 : : * or it's smaller than min_hist_size.
887 : : *
888 : : * The output parameter *hist_size receives the actual histogram size,
889 : : * or zero if no histogram. Callers may use this number to decide how
890 : : * much faith to put in the function result.
891 : : *
892 : : * Note that the result disregards both the most-common-values (if any) and
893 : : * null entries. The caller is expected to combine this result with
894 : : * statistics for those portions of the column population. It may also be
895 : : * prudent to clamp the result range, ie, disbelieve exact 0 or 1 outputs.
896 : : */
897 : : double
898 : 4347 : histogram_selectivity(VariableStatData *vardata,
899 : : FmgrInfo *opproc, Oid collation,
900 : : Datum constval, bool varonleft,
901 : : int min_hist_size, int n_skip,
902 : : int *hist_size)
903 : : {
904 : : double result;
905 : : AttStatsSlot sslot;
906 : :
907 : : /* check sanity of parameters */
908 : : Assert(n_skip >= 0);
909 : : Assert(min_hist_size > 2 * n_skip);
910 : :
911 [ + + + + ]: 6891 : if (HeapTupleIsValid(vardata->statsTuple) &&
912 [ + + ]: 5083 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
913 : 2539 : get_attstatsslot(&sslot, vardata->statsTuple,
914 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
915 : : ATTSTATSSLOT_VALUES))
916 : : {
917 : 2462 : *hist_size = sslot.nvalues;
918 [ + + ]: 2462 : if (sslot.nvalues >= min_hist_size)
919 : : {
920 : 1336 : LOCAL_FCINFO(fcinfo, 2);
921 : 1336 : int nmatch = 0;
922 : : int i;
923 : :
924 : : /*
925 : : * We invoke the opproc "by hand" so that we won't fail on NULL
926 : : * results. Such cases won't arise for normal comparison
927 : : * functions, but generic_restriction_selectivity could perhaps be
928 : : * used with operators that can return NULL. A small side benefit
929 : : * is to not need to re-initialize the fcinfo struct from scratch
930 : : * each time.
931 : : */
932 : 1336 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
933 : : NULL, NULL);
934 : 1336 : fcinfo->args[0].isnull = false;
935 : 1336 : fcinfo->args[1].isnull = false;
936 : : /* be careful to apply operator right way 'round */
937 [ + - ]: 1336 : if (varonleft)
938 : 1336 : fcinfo->args[1].value = constval;
939 : : else
940 : 0 : fcinfo->args[0].value = constval;
941 : :
942 [ + + ]: 114971 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
943 : : {
944 : : Datum fresult;
945 : :
946 [ + - ]: 113635 : if (varonleft)
947 : 113635 : fcinfo->args[0].value = sslot.values[i];
948 : : else
949 : 0 : fcinfo->args[1].value = sslot.values[i];
950 : 113635 : fcinfo->isnull = false;
951 : 113635 : fresult = FunctionCallInvoke(fcinfo);
952 [ + - + + ]: 113635 : if (!fcinfo->isnull && DatumGetBool(fresult))
953 : 3780 : nmatch++;
954 : : }
955 : 1336 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
956 : : }
957 : : else
958 : 1126 : result = -1;
959 : 2462 : free_attstatsslot(&sslot);
960 : : }
961 : : else
962 : : {
963 : 1885 : *hist_size = 0;
964 : 1885 : result = -1;
965 : : }
966 : :
967 : 4347 : return result;
968 : : }
969 : :
970 : : /*
971 : : * generic_restriction_selectivity - Selectivity for almost anything
972 : : *
973 : : * This function estimates selectivity for operators that we don't have any
974 : : * special knowledge about, but are on data types that we collect standard
975 : : * MCV and/or histogram statistics for. (Additional assumptions are that
976 : : * the operator is strict and immutable, or at least stable.)
977 : : *
978 : : * If we have "VAR OP CONST" or "CONST OP VAR", selectivity is estimated by
979 : : * applying the operator to each element of the column's MCV and/or histogram
980 : : * stats, and merging the results using the assumption that the histogram is
981 : : * a reasonable random sample of the column's non-MCV population. Note that
982 : : * if the operator's semantics are related to the histogram ordering, this
983 : : * might not be such a great assumption; other functions such as
984 : : * scalarineqsel() are probably a better match in such cases.
985 : : *
986 : : * Otherwise, fall back to the default selectivity provided by the caller.
987 : : */
988 : : double
989 : 845 : generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation,
990 : : List *args, int varRelid,
991 : : double default_selectivity)
992 : : {
993 : : double selec;
994 : : VariableStatData vardata;
995 : : Node *other;
996 : : bool varonleft;
997 : :
998 : : /*
999 : : * If expression is not variable OP something or something OP variable,
1000 : : * then punt and return the default estimate.
1001 : : */
1002 [ - + ]: 845 : if (!get_restriction_variable(root, args, varRelid,
1003 : : &vardata, &other, &varonleft))
1004 : 0 : return default_selectivity;
1005 : :
1006 : : /*
1007 : : * If the something is a NULL constant, assume operator is strict and
1008 : : * return zero, ie, operator will never return TRUE.
1009 : : */
1010 [ + - ]: 845 : if (IsA(other, Const) &&
1011 [ - + ]: 845 : ((Const *) other)->constisnull)
1012 : : {
1013 [ # # ]: 0 : ReleaseVariableStats(vardata);
1014 : 0 : return 0.0;
1015 : : }
1016 : :
1017 [ + - ]: 845 : if (IsA(other, Const))
1018 : : {
1019 : : /* Variable is being compared to a known non-null constant */
1020 : 845 : Datum constval = ((Const *) other)->constvalue;
1021 : : FmgrInfo opproc;
1022 : : double mcvsum;
1023 : : double mcvsel;
1024 : : double nullfrac;
1025 : : int hist_size;
1026 : :
1027 : 845 : fmgr_info(get_opcode(oproid), &opproc);
1028 : :
1029 : : /*
1030 : : * Calculate the selectivity for the column's most common values.
1031 : : */
1032 : 845 : mcvsel = mcv_selectivity(&vardata, &opproc, collation,
1033 : : constval, varonleft,
1034 : : &mcvsum);
1035 : :
1036 : : /*
1037 : : * If the histogram is large enough, see what fraction of it matches
1038 : : * the query, and assume that's representative of the non-MCV
1039 : : * population. Otherwise use the default selectivity for the non-MCV
1040 : : * population.
1041 : : */
1042 : 845 : selec = histogram_selectivity(&vardata, &opproc, collation,
1043 : : constval, varonleft,
1044 : : 10, 1, &hist_size);
1045 [ + - ]: 845 : if (selec < 0)
1046 : : {
1047 : : /* Nope, fall back on default */
1048 : 845 : selec = default_selectivity;
1049 : : }
1050 [ # # ]: 0 : else if (hist_size < 100)
1051 : : {
1052 : : /*
1053 : : * For histogram sizes from 10 to 100, we combine the histogram
1054 : : * and default selectivities, putting increasingly more trust in
1055 : : * the histogram for larger sizes.
1056 : : */
1057 : 0 : double hist_weight = hist_size / 100.0;
1058 : :
1059 : 0 : selec = selec * hist_weight +
1060 : 0 : default_selectivity * (1.0 - hist_weight);
1061 : : }
1062 : :
1063 : : /* In any case, don't believe extremely small or large estimates. */
1064 [ - + ]: 845 : if (selec < 0.0001)
1065 : 0 : selec = 0.0001;
1066 [ - + ]: 845 : else if (selec > 0.9999)
1067 : 0 : selec = 0.9999;
1068 : :
1069 : : /* Don't forget to account for nulls. */
1070 [ + + ]: 845 : if (HeapTupleIsValid(vardata.statsTuple))
1071 : 70 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
1072 : : else
1073 : 775 : nullfrac = 0.0;
1074 : :
1075 : : /*
1076 : : * Now merge the results from the MCV and histogram calculations,
1077 : : * realizing that the histogram covers only the non-null values that
1078 : : * are not listed in MCV.
1079 : : */
1080 : 845 : selec *= 1.0 - nullfrac - mcvsum;
1081 : 845 : selec += mcvsel;
1082 : : }
1083 : : else
1084 : : {
1085 : : /* Comparison value is not constant, so we can't do anything */
1086 : 0 : selec = default_selectivity;
1087 : : }
1088 : :
1089 [ + + ]: 845 : ReleaseVariableStats(vardata);
1090 : :
1091 : : /* result should be in range, but make sure... */
1092 [ - + - + ]: 845 : CLAMP_PROBABILITY(selec);
1093 : :
1094 : 845 : return selec;
1095 : : }
1096 : :
1097 : : /*
1098 : : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
1099 : : *
1100 : : * Determine the fraction of the variable's histogram population that
1101 : : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
1102 : : * The isgt and iseq flags distinguish which of the four cases apply.
1103 : : *
1104 : : * While opproc could be looked up from the operator OID, common callers
1105 : : * also need to call it separately, so we make the caller pass both.
1106 : : *
1107 : : * Returns -1 if there is no histogram (valid results will always be >= 0).
1108 : : *
1109 : : * Note that the result disregards both the most-common-values (if any) and
1110 : : * null entries. The caller is expected to combine this result with
1111 : : * statistics for those portions of the column population.
1112 : : *
1113 : : * This is exported so that some other estimation functions can use it.
1114 : : */
1115 : : double
1116 : 220763 : ineq_histogram_selectivity(PlannerInfo *root,
1117 : : VariableStatData *vardata,
1118 : : Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq,
1119 : : Oid collation,
1120 : : Datum constval, Oid consttype)
1121 : : {
1122 : : double hist_selec;
1123 : : AttStatsSlot sslot;
1124 : :
1125 : 220763 : hist_selec = -1.0;
1126 : :
1127 : : /*
1128 : : * Someday, ANALYZE might store more than one histogram per rel/att,
1129 : : * corresponding to more than one possible sort ordering defined for the
1130 : : * column type. Right now, we know there is only one, so just grab it and
1131 : : * see if it matches the query.
1132 : : *
1133 : : * Note that we can't use opoid as search argument; the staop appearing in
1134 : : * pg_statistic will be for the relevant '<' operator, but what we have
1135 : : * might be some other inequality operator such as '>='. (Even if opoid
1136 : : * is a '<' operator, it could be cross-type.) Hence we must use
1137 : : * comparison_ops_are_compatible() to see if the operators match.
1138 : : */
1139 [ + + + + ]: 441015 : if (HeapTupleIsValid(vardata->statsTuple) &&
1140 [ + + ]: 440234 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1141 : 219982 : get_attstatsslot(&sslot, vardata->statsTuple,
1142 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1143 : : ATTSTATSSLOT_VALUES))
1144 : : {
1145 [ + - ]: 133168 : if (sslot.nvalues > 1 &&
1146 [ + + + + ]: 266274 : sslot.stacoll == collation &&
1147 : 133106 : comparison_ops_are_compatible(sslot.staop, opoid))
1148 : 133016 : {
1149 : : /*
1150 : : * Use binary search to find the desired location, namely the
1151 : : * right end of the histogram bin containing the comparison value,
1152 : : * which is the leftmost entry for which the comparison operator
1153 : : * succeeds (if isgt) or fails (if !isgt).
1154 : : *
1155 : : * In this loop, we pay no attention to whether the operator iseq
1156 : : * or not; that detail will be mopped up below. (We cannot tell,
1157 : : * anyway, whether the operator thinks the values are equal.)
1158 : : *
1159 : : * If the binary search accesses the first or last histogram
1160 : : * entry, we try to replace that endpoint with the true column min
1161 : : * or max as found by get_actual_variable_range(). This
1162 : : * ameliorates misestimates when the min or max is moving as a
1163 : : * result of changes since the last ANALYZE. Note that this could
1164 : : * result in effectively including MCVs into the histogram that
1165 : : * weren't there before, but we don't try to correct for that.
1166 : : */
1167 : : double histfrac;
1168 : 133016 : int lobound = 0; /* first possible slot to search */
1169 : 133016 : int hibound = sslot.nvalues; /* last+1 slot to search */
1170 : 133016 : bool have_end = false;
1171 : :
1172 : : /*
1173 : : * If there are only two histogram entries, we'll want up-to-date
1174 : : * values for both. (If there are more than two, we need at most
1175 : : * one of them to be updated, so we deal with that within the
1176 : : * loop.)
1177 : : */
1178 [ + + ]: 133016 : if (sslot.nvalues == 2)
1179 : 2574 : have_end = get_actual_variable_range(root,
1180 : : vardata,
1181 : : sslot.staop,
1182 : : collation,
1183 : : &sslot.values[0],
1184 : 2574 : &sslot.values[1]);
1185 : :
1186 [ + + ]: 871288 : while (lobound < hibound)
1187 : : {
1188 : 738272 : int probe = (lobound + hibound) / 2;
1189 : : bool ltcmp;
1190 : :
1191 : : /*
1192 : : * If we find ourselves about to compare to the first or last
1193 : : * histogram entry, first try to replace it with the actual
1194 : : * current min or max (unless we already did so above).
1195 : : */
1196 [ + + + + ]: 738272 : if (probe == 0 && sslot.nvalues > 2)
1197 : 65653 : have_end = get_actual_variable_range(root,
1198 : : vardata,
1199 : : sslot.staop,
1200 : : collation,
1201 : : &sslot.values[0],
1202 : : NULL);
1203 [ + + + + ]: 672619 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1204 : 45213 : have_end = get_actual_variable_range(root,
1205 : : vardata,
1206 : : sslot.staop,
1207 : : collation,
1208 : : NULL,
1209 : 45213 : &sslot.values[probe]);
1210 : :
1211 : 738272 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1212 : : collation,
1213 : 738272 : sslot.values[probe],
1214 : : constval));
1215 [ + + ]: 738272 : if (isgt)
1216 : 49867 : ltcmp = !ltcmp;
1217 [ + + ]: 738272 : if (ltcmp)
1218 : 274074 : lobound = probe + 1;
1219 : : else
1220 : 464198 : hibound = probe;
1221 : : }
1222 : :
1223 [ + + ]: 133016 : if (lobound <= 0)
1224 : : {
1225 : : /*
1226 : : * Constant is below lower histogram boundary. More
1227 : : * precisely, we have found that no entry in the histogram
1228 : : * satisfies the inequality clause (if !isgt) or they all do
1229 : : * (if isgt). We estimate that that's true of the entire
1230 : : * table, so set histfrac to 0.0 (which we'll flip to 1.0
1231 : : * below, if isgt).
1232 : : */
1233 : 57860 : histfrac = 0.0;
1234 : : }
1235 [ + + ]: 75156 : else if (lobound >= sslot.nvalues)
1236 : : {
1237 : : /*
1238 : : * Inverse case: constant is above upper histogram boundary.
1239 : : */
1240 : 22048 : histfrac = 1.0;
1241 : : }
1242 : : else
1243 : : {
1244 : : /* We have values[i-1] <= constant <= values[i]. */
1245 : 53108 : int i = lobound;
1246 : 53108 : double eq_selec = 0;
1247 : : double val,
1248 : : high,
1249 : : low;
1250 : : double binfrac;
1251 : :
1252 : : /*
1253 : : * In the cases where we'll need it below, obtain an estimate
1254 : : * of the selectivity of "x = constval". We use a calculation
1255 : : * similar to what var_eq_const() does for a non-MCV constant,
1256 : : * ie, estimate that all distinct non-MCV values occur equally
1257 : : * often. But multiplication by "1.0 - sumcommon - nullfrac"
1258 : : * will be done by our caller, so we shouldn't do that here.
1259 : : * Therefore we can't try to clamp the estimate by reference
1260 : : * to the least common MCV; the result would be too small.
1261 : : *
1262 : : * Note: since this is effectively assuming that constval
1263 : : * isn't an MCV, it's logically dubious if constval in fact is
1264 : : * one. But we have to apply *some* correction for equality,
1265 : : * and anyway we cannot tell if constval is an MCV, since we
1266 : : * don't have a suitable equality operator at hand.
1267 : : */
1268 [ + + + + ]: 53108 : if (i == 1 || isgt == iseq)
1269 : : {
1270 : : double otherdistinct;
1271 : : bool isdefault;
1272 : : AttStatsSlot mcvslot;
1273 : :
1274 : : /* Get estimated number of distinct values */
1275 : 22236 : otherdistinct = get_variable_numdistinct(vardata,
1276 : : &isdefault);
1277 : :
1278 : : /* Subtract off the number of known MCVs */
1279 [ + + ]: 22236 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1280 : : STATISTIC_KIND_MCV, InvalidOid,
1281 : : ATTSTATSSLOT_NUMBERS))
1282 : : {
1283 : 2310 : otherdistinct -= mcvslot.nnumbers;
1284 : 2310 : free_attstatsslot(&mcvslot);
1285 : : }
1286 : :
1287 : : /* If result doesn't seem sane, leave eq_selec at 0 */
1288 [ + + ]: 22236 : if (otherdistinct > 1)
1289 : 22215 : eq_selec = 1.0 / otherdistinct;
1290 : : }
1291 : :
1292 : : /*
1293 : : * Convert the constant and the two nearest bin boundary
1294 : : * values to a uniform comparison scale, and do a linear
1295 : : * interpolation within this bin.
1296 : : */
1297 [ + - ]: 53108 : if (convert_to_scalar(constval, consttype, collation,
1298 : : &val,
1299 : 53108 : sslot.values[i - 1], sslot.values[i],
1300 : : vardata->vartype,
1301 : : &low, &high))
1302 : : {
1303 [ - + ]: 53108 : if (high <= low)
1304 : : {
1305 : : /* cope if bin boundaries appear identical */
1306 : 0 : binfrac = 0.5;
1307 : : }
1308 [ + + ]: 53108 : else if (val <= low)
1309 : 11300 : binfrac = 0.0;
1310 [ + + ]: 41808 : else if (val >= high)
1311 : 1742 : binfrac = 1.0;
1312 : : else
1313 : : {
1314 : 40066 : binfrac = (val - low) / (high - low);
1315 : :
1316 : : /*
1317 : : * Watch out for the possibility that we got a NaN or
1318 : : * Infinity from the division. This can happen
1319 : : * despite the previous checks, if for example "low"
1320 : : * is -Infinity.
1321 : : */
1322 [ + - + - ]: 40066 : if (isnan(binfrac) ||
1323 [ - + ]: 40066 : binfrac < 0.0 || binfrac > 1.0)
1324 : 0 : binfrac = 0.5;
1325 : : }
1326 : : }
1327 : : else
1328 : : {
1329 : : /*
1330 : : * Ideally we'd produce an error here, on the grounds that
1331 : : * the given operator shouldn't have scalarXXsel
1332 : : * registered as its selectivity func unless we can deal
1333 : : * with its operand types. But currently, all manner of
1334 : : * stuff is invoking scalarXXsel, so give a default
1335 : : * estimate until that can be fixed.
1336 : : */
1337 : 0 : binfrac = 0.5;
1338 : : }
1339 : :
1340 : : /*
1341 : : * Now, compute the overall selectivity across the values
1342 : : * represented by the histogram. We have i-1 full bins and
1343 : : * binfrac partial bin below the constant.
1344 : : */
1345 : 53108 : histfrac = (double) (i - 1) + binfrac;
1346 : 53108 : histfrac /= (double) (sslot.nvalues - 1);
1347 : :
1348 : : /*
1349 : : * At this point, histfrac is an estimate of the fraction of
1350 : : * the population represented by the histogram that satisfies
1351 : : * "x <= constval". Somewhat remarkably, this statement is
1352 : : * true regardless of which operator we were doing the probes
1353 : : * with, so long as convert_to_scalar() delivers reasonable
1354 : : * results. If the probe constant is equal to some histogram
1355 : : * entry, we would have considered the bin to the left of that
1356 : : * entry if probing with "<" or ">=", or the bin to the right
1357 : : * if probing with "<=" or ">"; but binfrac would have come
1358 : : * out as 1.0 in the first case and 0.0 in the second, leading
1359 : : * to the same histfrac in either case. For probe constants
1360 : : * between histogram entries, we find the same bin and get the
1361 : : * same estimate with any operator.
1362 : : *
1363 : : * The fact that the estimate corresponds to "x <= constval"
1364 : : * and not "x < constval" is because of the way that ANALYZE
1365 : : * constructs the histogram: each entry is, effectively, the
1366 : : * rightmost value in its sample bucket. So selectivity
1367 : : * values that are exact multiples of 1/(histogram_size-1)
1368 : : * should be understood as estimates including a histogram
1369 : : * entry plus everything to its left.
1370 : : *
1371 : : * However, that breaks down for the first histogram entry,
1372 : : * which necessarily is the leftmost value in its sample
1373 : : * bucket. That means the first histogram bin is slightly
1374 : : * narrower than the rest, by an amount equal to eq_selec.
1375 : : * Another way to say that is that we want "x <= leftmost" to
1376 : : * be estimated as eq_selec not zero. So, if we're dealing
1377 : : * with the first bin (i==1), rescale to make that true while
1378 : : * adjusting the rest of that bin linearly.
1379 : : */
1380 [ + + ]: 53108 : if (i == 1)
1381 : 9452 : histfrac += eq_selec * (1.0 - binfrac);
1382 : :
1383 : : /*
1384 : : * "x <= constval" is good if we want an estimate for "<=" or
1385 : : * ">", but if we are estimating for "<" or ">=", we now need
1386 : : * to decrease the estimate by eq_selec.
1387 : : */
1388 [ + + ]: 53108 : if (isgt == iseq)
1389 : 17166 : histfrac -= eq_selec;
1390 : : }
1391 : :
1392 : : /*
1393 : : * Now the estimate is finished for "<" and "<=" cases. If we are
1394 : : * estimating for ">" or ">=", flip it.
1395 : : */
1396 [ + + ]: 133016 : hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1397 : :
1398 : : /*
1399 : : * The histogram boundaries are only approximate to begin with,
1400 : : * and may well be out of date anyway. Therefore, don't believe
1401 : : * extremely small or large selectivity estimates --- unless we
1402 : : * got actual current endpoint values from the table, in which
1403 : : * case just do the usual sanity clamp. Somewhat arbitrarily, we
1404 : : * set the cutoff for other cases at a hundredth of the histogram
1405 : : * resolution.
1406 : : */
1407 [ + + ]: 133016 : if (have_end)
1408 [ - + - + ]: 75870 : CLAMP_PROBABILITY(hist_selec);
1409 : : else
1410 : : {
1411 : 57146 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1412 : :
1413 [ + + ]: 57146 : if (hist_selec < cutoff)
1414 : 19636 : hist_selec = cutoff;
1415 [ + + ]: 37510 : else if (hist_selec > 1.0 - cutoff)
1416 : 14918 : hist_selec = 1.0 - cutoff;
1417 : : }
1418 : : }
1419 [ + - ]: 152 : else if (sslot.nvalues > 1)
1420 : : {
1421 : : /*
1422 : : * If we get here, we have a histogram but it's not sorted the way
1423 : : * we want. Do a brute-force search to see how many of the
1424 : : * entries satisfy the comparison condition, and take that
1425 : : * fraction as our estimate. (This is identical to the inner loop
1426 : : * of histogram_selectivity; maybe share code?)
1427 : : */
1428 : 152 : LOCAL_FCINFO(fcinfo, 2);
1429 : 152 : int nmatch = 0;
1430 : :
1431 : 152 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1432 : : NULL, NULL);
1433 : 152 : fcinfo->args[0].isnull = false;
1434 : 152 : fcinfo->args[1].isnull = false;
1435 : 152 : fcinfo->args[1].value = constval;
1436 [ + + ]: 801772 : for (int i = 0; i < sslot.nvalues; i++)
1437 : : {
1438 : : Datum fresult;
1439 : :
1440 : 801620 : fcinfo->args[0].value = sslot.values[i];
1441 : 801620 : fcinfo->isnull = false;
1442 : 801620 : fresult = FunctionCallInvoke(fcinfo);
1443 [ + - + + ]: 801620 : if (!fcinfo->isnull && DatumGetBool(fresult))
1444 : 1626 : nmatch++;
1445 : : }
1446 : 152 : hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1447 : :
1448 : : /*
1449 : : * As above, clamp to a hundredth of the histogram resolution.
1450 : : * This case is surely even less trustworthy than the normal one,
1451 : : * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1452 : : * clamp should be more restrictive in this case?)
1453 : : */
1454 : : {
1455 : 152 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1456 : :
1457 [ + + ]: 152 : if (hist_selec < cutoff)
1458 : 10 : hist_selec = cutoff;
1459 [ + + ]: 142 : else if (hist_selec > 1.0 - cutoff)
1460 : 10 : hist_selec = 1.0 - cutoff;
1461 : : }
1462 : : }
1463 : :
1464 : 133168 : free_attstatsslot(&sslot);
1465 : : }
1466 : :
1467 : 220763 : return hist_selec;
1468 : : }
1469 : :
1470 : : /*
1471 : : * Common wrapper function for the selectivity estimators that simply
1472 : : * invoke scalarineqsel().
1473 : : */
1474 : : static Datum
1475 : 38641 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1476 : : {
1477 : 38641 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1478 : 38641 : Oid operator = PG_GETARG_OID(1);
1479 : 38641 : List *args = (List *) PG_GETARG_POINTER(2);
1480 : 38641 : int varRelid = PG_GETARG_INT32(3);
1481 : 38641 : Oid collation = PG_GET_COLLATION();
1482 : : VariableStatData vardata;
1483 : : Node *other;
1484 : : bool varonleft;
1485 : : Datum constval;
1486 : : Oid consttype;
1487 : : double selec;
1488 : :
1489 : : /*
1490 : : * If expression is not variable op something or something op variable,
1491 : : * then punt and return a default estimate.
1492 : : */
1493 [ + + ]: 38641 : if (!get_restriction_variable(root, args, varRelid,
1494 : : &vardata, &other, &varonleft))
1495 : 462 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1496 : :
1497 : : /*
1498 : : * Can't do anything useful if the something is not a constant, either.
1499 : : */
1500 [ + + ]: 38179 : if (!IsA(other, Const))
1501 : : {
1502 [ + + ]: 2365 : ReleaseVariableStats(vardata);
1503 : 2365 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1504 : : }
1505 : :
1506 : : /*
1507 : : * If the constant is NULL, assume operator is strict and return zero, ie,
1508 : : * operator will never return TRUE.
1509 : : */
1510 [ + + ]: 35814 : if (((Const *) other)->constisnull)
1511 : : {
1512 [ + + ]: 55 : ReleaseVariableStats(vardata);
1513 : 55 : PG_RETURN_FLOAT8(0.0);
1514 : : }
1515 : 35759 : constval = ((Const *) other)->constvalue;
1516 : 35759 : consttype = ((Const *) other)->consttype;
1517 : :
1518 : : /*
1519 : : * Force the var to be on the left to simplify logic in scalarineqsel.
1520 : : */
1521 [ + + ]: 35759 : if (!varonleft)
1522 : : {
1523 : 318 : operator = get_commutator(operator);
1524 [ - + ]: 318 : if (!operator)
1525 : : {
1526 : : /* Use default selectivity (should we raise an error instead?) */
1527 [ # # ]: 0 : ReleaseVariableStats(vardata);
1528 : 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1529 : : }
1530 : 318 : isgt = !isgt;
1531 : : }
1532 : :
1533 : : /* The rest of the work is done by scalarineqsel(). */
1534 : 35759 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1535 : : &vardata, constval, consttype);
1536 : :
1537 [ + + ]: 35759 : ReleaseVariableStats(vardata);
1538 : :
1539 : 35759 : PG_RETURN_FLOAT8((float8) selec);
1540 : : }
1541 : :
1542 : : /*
1543 : : * scalarltsel - Selectivity of "<" for scalars.
1544 : : */
1545 : : Datum
1546 : 12199 : scalarltsel(PG_FUNCTION_ARGS)
1547 : : {
1548 : 12199 : return scalarineqsel_wrapper(fcinfo, false, false);
1549 : : }
1550 : :
1551 : : /*
1552 : : * scalarlesel - Selectivity of "<=" for scalars.
1553 : : */
1554 : : Datum
1555 : 3842 : scalarlesel(PG_FUNCTION_ARGS)
1556 : : {
1557 : 3842 : return scalarineqsel_wrapper(fcinfo, false, true);
1558 : : }
1559 : :
1560 : : /*
1561 : : * scalargtsel - Selectivity of ">" for scalars.
1562 : : */
1563 : : Datum
1564 : 12137 : scalargtsel(PG_FUNCTION_ARGS)
1565 : : {
1566 : 12137 : return scalarineqsel_wrapper(fcinfo, true, false);
1567 : : }
1568 : :
1569 : : /*
1570 : : * scalargesel - Selectivity of ">=" for scalars.
1571 : : */
1572 : : Datum
1573 : 10463 : scalargesel(PG_FUNCTION_ARGS)
1574 : : {
1575 : 10463 : return scalarineqsel_wrapper(fcinfo, true, true);
1576 : : }
1577 : :
1578 : : /*
1579 : : * boolvarsel - Selectivity of Boolean variable.
1580 : : *
1581 : : * This can actually be called on any boolean-valued expression. If it
1582 : : * involves only Vars of the specified relation, and if there are statistics
1583 : : * about the Var or expression (the latter is possible if it's indexed) then
1584 : : * we'll produce a real estimate; otherwise it's just a default.
1585 : : */
1586 : : Selectivity
1587 : 46553 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1588 : : {
1589 : : VariableStatData vardata;
1590 : : double selec;
1591 : :
1592 : 46553 : examine_variable(root, arg, varRelid, &vardata);
1593 [ + + ]: 46553 : if (HeapTupleIsValid(vardata.statsTuple))
1594 : : {
1595 : : /*
1596 : : * A boolean variable V is equivalent to the clause V = 't', so we
1597 : : * compute the selectivity as if that is what we have.
1598 : : */
1599 : 21750 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1600 : : BoolGetDatum(true), false, true, false);
1601 : : }
1602 [ + + ]: 24803 : else if (is_funcclause(arg))
1603 : : {
1604 : : /*
1605 : : * If we have no stats and it's a function call, estimate 0.3333333.
1606 : : * This seems a pretty unprincipled choice, but Postgres has been
1607 : : * using that estimate for function calls since 1992. The hoariness
1608 : : * of this behavior suggests that we should not be in too much hurry
1609 : : * to use another value.
1610 : : */
1611 : 11525 : selec = 0.3333333;
1612 : : }
1613 : : else
1614 : : {
1615 : : /* Otherwise, the default estimate is 0.5 */
1616 : 13278 : selec = 0.5;
1617 : : }
1618 [ + + ]: 46553 : ReleaseVariableStats(vardata);
1619 : 46553 : return selec;
1620 : : }
1621 : :
1622 : : /*
1623 : : * booltestsel - Selectivity of BooleanTest Node.
1624 : : */
1625 : : Selectivity
1626 : 791 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1627 : : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1628 : : {
1629 : : VariableStatData vardata;
1630 : : double selec;
1631 : :
1632 : 791 : examine_variable(root, arg, varRelid, &vardata);
1633 : :
1634 [ + + ]: 791 : if (HeapTupleIsValid(vardata.statsTuple))
1635 : : {
1636 : : Form_pg_statistic stats;
1637 : : double freq_null;
1638 : : AttStatsSlot sslot;
1639 : :
1640 : 20 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1641 : 20 : freq_null = stats->stanullfrac;
1642 : :
1643 [ + + ]: 20 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1644 : : STATISTIC_KIND_MCV, InvalidOid,
1645 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1646 [ + - ]: 10 : && sslot.nnumbers > 0)
1647 : 10 : {
1648 : : double freq_true;
1649 : : double freq_false;
1650 : :
1651 : : /*
1652 : : * Get first MCV frequency and derive frequency for true.
1653 : : */
1654 [ - + ]: 10 : if (DatumGetBool(sslot.values[0]))
1655 : 0 : freq_true = sslot.numbers[0];
1656 : : else
1657 : 10 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1658 : :
1659 : : /*
1660 : : * Next derive frequency for false. Then use these as appropriate
1661 : : * to derive frequency for each case.
1662 : : */
1663 : 10 : freq_false = 1.0 - freq_true - freq_null;
1664 : :
1665 [ - - + - : 10 : switch (booltesttype)
- - - ]
1666 : : {
1667 : 0 : case IS_UNKNOWN:
1668 : : /* select only NULL values */
1669 : 0 : selec = freq_null;
1670 : 0 : break;
1671 : 0 : case IS_NOT_UNKNOWN:
1672 : : /* select non-NULL values */
1673 : 0 : selec = 1.0 - freq_null;
1674 : 0 : break;
1675 : 10 : case IS_TRUE:
1676 : : /* select only TRUE values */
1677 : 10 : selec = freq_true;
1678 : 10 : break;
1679 : 0 : case IS_NOT_TRUE:
1680 : : /* select non-TRUE values */
1681 : 0 : selec = 1.0 - freq_true;
1682 : 0 : break;
1683 : 0 : case IS_FALSE:
1684 : : /* select only FALSE values */
1685 : 0 : selec = freq_false;
1686 : 0 : break;
1687 : 0 : case IS_NOT_FALSE:
1688 : : /* select non-FALSE values */
1689 : 0 : selec = 1.0 - freq_false;
1690 : 0 : break;
1691 : 0 : default:
1692 [ # # ]: 0 : elog(ERROR, "unrecognized booltesttype: %d",
1693 : : (int) booltesttype);
1694 : : selec = 0.0; /* Keep compiler quiet */
1695 : : break;
1696 : : }
1697 : :
1698 : 10 : free_attstatsslot(&sslot);
1699 : : }
1700 : : else
1701 : : {
1702 : : /*
1703 : : * No most-common-value info available. Still have null fraction
1704 : : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1705 : : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1706 : : */
1707 [ + - - - : 10 : switch (booltesttype)
- ]
1708 : : {
1709 : 10 : case IS_UNKNOWN:
1710 : : /* select only NULL values */
1711 : 10 : selec = freq_null;
1712 : 10 : break;
1713 : 0 : case IS_NOT_UNKNOWN:
1714 : : /* select non-NULL values */
1715 : 0 : selec = 1.0 - freq_null;
1716 : 0 : break;
1717 : 0 : case IS_TRUE:
1718 : : case IS_FALSE:
1719 : : /* Assume we select half of the non-NULL values */
1720 : 0 : selec = (1.0 - freq_null) / 2.0;
1721 : 0 : break;
1722 : 0 : case IS_NOT_TRUE:
1723 : : case IS_NOT_FALSE:
1724 : : /* Assume we select NULLs plus half of the non-NULLs */
1725 : : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1726 : 0 : selec = (freq_null + 1.0) / 2.0;
1727 : 0 : break;
1728 : 0 : default:
1729 [ # # ]: 0 : elog(ERROR, "unrecognized booltesttype: %d",
1730 : : (int) booltesttype);
1731 : : selec = 0.0; /* Keep compiler quiet */
1732 : : break;
1733 : : }
1734 : : }
1735 : : }
1736 : : else
1737 : : {
1738 : : /*
1739 : : * If we can't get variable statistics for the argument, perhaps
1740 : : * clause_selectivity can do something with it. We ignore the
1741 : : * possibility of a NULL value when using clause_selectivity, and just
1742 : : * assume the value is either TRUE or FALSE.
1743 : : */
1744 [ + + + + : 771 : switch (booltesttype)
- ]
1745 : : {
1746 : 40 : case IS_UNKNOWN:
1747 : 40 : selec = DEFAULT_UNK_SEL;
1748 : 40 : break;
1749 : 90 : case IS_NOT_UNKNOWN:
1750 : 90 : selec = DEFAULT_NOT_UNK_SEL;
1751 : 90 : break;
1752 : 220 : case IS_TRUE:
1753 : : case IS_NOT_FALSE:
1754 : 220 : selec = (double) clause_selectivity(root, arg,
1755 : : varRelid,
1756 : : jointype, sjinfo);
1757 : 220 : break;
1758 : 421 : case IS_FALSE:
1759 : : case IS_NOT_TRUE:
1760 : 421 : selec = 1.0 - (double) clause_selectivity(root, arg,
1761 : : varRelid,
1762 : : jointype, sjinfo);
1763 : 421 : break;
1764 : 0 : default:
1765 [ # # ]: 0 : elog(ERROR, "unrecognized booltesttype: %d",
1766 : : (int) booltesttype);
1767 : : selec = 0.0; /* Keep compiler quiet */
1768 : : break;
1769 : : }
1770 : : }
1771 : :
1772 [ + + ]: 791 : ReleaseVariableStats(vardata);
1773 : :
1774 : : /* result should be in range, but make sure... */
1775 [ - + - + ]: 791 : CLAMP_PROBABILITY(selec);
1776 : :
1777 : 791 : return (Selectivity) selec;
1778 : : }
1779 : :
1780 : : /*
1781 : : * nulltestsel - Selectivity of NullTest Node.
1782 : : */
1783 : : Selectivity
1784 : 14113 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1785 : : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1786 : : {
1787 : : VariableStatData vardata;
1788 : : double selec;
1789 : :
1790 : 14113 : examine_variable(root, arg, varRelid, &vardata);
1791 : :
1792 [ + + ]: 14113 : if (HeapTupleIsValid(vardata.statsTuple))
1793 : : {
1794 : : Form_pg_statistic stats;
1795 : : double freq_null;
1796 : :
1797 : 7566 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1798 : 7566 : freq_null = stats->stanullfrac;
1799 : :
1800 [ + + - ]: 7566 : switch (nulltesttype)
1801 : : {
1802 : 5972 : case IS_NULL:
1803 : :
1804 : : /*
1805 : : * Use freq_null directly.
1806 : : */
1807 : 5972 : selec = freq_null;
1808 : 5972 : break;
1809 : 1594 : case IS_NOT_NULL:
1810 : :
1811 : : /*
1812 : : * Select not unknown (not null) values. Calculate from
1813 : : * freq_null.
1814 : : */
1815 : 1594 : selec = 1.0 - freq_null;
1816 : 1594 : break;
1817 : 0 : default:
1818 [ # # ]: 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1819 : : (int) nulltesttype);
1820 : : return (Selectivity) 0; /* keep compiler quiet */
1821 : : }
1822 : : }
1823 [ + - + + ]: 6547 : else if (vardata.var && IsA(vardata.var, Var) &&
1824 [ + + ]: 6013 : ((Var *) vardata.var)->varattno < 0)
1825 : : {
1826 : : /*
1827 : : * There are no stats for system columns, but we know they are never
1828 : : * NULL.
1829 : : */
1830 [ + - ]: 89 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1831 : : }
1832 : : else
1833 : : {
1834 : : /*
1835 : : * No ANALYZE stats available, so make a guess
1836 : : */
1837 [ + + - ]: 6458 : switch (nulltesttype)
1838 : : {
1839 : 1648 : case IS_NULL:
1840 : 1648 : selec = DEFAULT_UNK_SEL;
1841 : 1648 : break;
1842 : 4810 : case IS_NOT_NULL:
1843 : 4810 : selec = DEFAULT_NOT_UNK_SEL;
1844 : 4810 : break;
1845 : 0 : default:
1846 [ # # ]: 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1847 : : (int) nulltesttype);
1848 : : return (Selectivity) 0; /* keep compiler quiet */
1849 : : }
1850 : : }
1851 : :
1852 [ + + ]: 14113 : ReleaseVariableStats(vardata);
1853 : :
1854 : : /* result should be in range, but make sure... */
1855 [ - + - + ]: 14113 : CLAMP_PROBABILITY(selec);
1856 : :
1857 : 14113 : return (Selectivity) selec;
1858 : : }
1859 : :
1860 : : /*
1861 : : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1862 : : *
1863 : : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1864 : : * but it seems possible that RelabelType might show up. Also, the planner
1865 : : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1866 : : * so we need to be ready to deal with more than one level.
1867 : : */
1868 : : static Node *
1869 : 102421 : strip_array_coercion(Node *node)
1870 : : {
1871 : : for (;;)
1872 : : {
1873 [ + - + + ]: 102509 : if (node && IsA(node, ArrayCoerceExpr))
1874 : 88 : {
1875 : 1763 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1876 : :
1877 : : /*
1878 : : * If the per-element expression is just a RelabelType on top of
1879 : : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1880 : : */
1881 [ + + ]: 1763 : if (IsA(acoerce->elemexpr, RelabelType) &&
1882 [ + - ]: 88 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1883 : 88 : node = (Node *) acoerce->arg;
1884 : : else
1885 : : break;
1886 : : }
1887 [ + - - + ]: 100746 : else if (node && IsA(node, RelabelType))
1888 : : {
1889 : : /* We don't really expect this case, but may as well cope */
1890 : 0 : node = (Node *) ((RelabelType *) node)->arg;
1891 : : }
1892 : : else
1893 : : break;
1894 : : }
1895 : 102421 : return node;
1896 : : }
1897 : :
1898 : : /*
1899 : : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1900 : : */
1901 : : Selectivity
1902 : 17593 : scalararraysel(PlannerInfo *root,
1903 : : ScalarArrayOpExpr *clause,
1904 : : bool is_join_clause,
1905 : : int varRelid,
1906 : : JoinType jointype,
1907 : : SpecialJoinInfo *sjinfo)
1908 : : {
1909 : 17593 : Oid operator = clause->opno;
1910 : 17593 : bool useOr = clause->useOr;
1911 : 17593 : bool isEquality = false;
1912 : 17593 : bool isInequality = false;
1913 : : Node *leftop;
1914 : : Node *rightop;
1915 : : Oid nominal_element_type;
1916 : : Oid nominal_element_collation;
1917 : : TypeCacheEntry *typentry;
1918 : : RegProcedure oprsel;
1919 : : FmgrInfo oprselproc;
1920 : : Selectivity s1;
1921 : : Selectivity s1disjoint;
1922 : :
1923 : : /* First, deconstruct the expression */
1924 : : Assert(list_length(clause->args) == 2);
1925 : 17593 : leftop = (Node *) linitial(clause->args);
1926 : 17593 : rightop = (Node *) lsecond(clause->args);
1927 : :
1928 : : /* aggressively reduce both sides to constants */
1929 : 17593 : leftop = estimate_expression_value(root, leftop);
1930 : 17593 : rightop = estimate_expression_value(root, rightop);
1931 : :
1932 : : /* get nominal (after relabeling) element type of rightop */
1933 : 17593 : nominal_element_type = get_base_element_type(exprType(rightop));
1934 [ - + ]: 17593 : if (!OidIsValid(nominal_element_type))
1935 : 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1936 : : /* get nominal collation, too, for generating constants */
1937 : 17593 : nominal_element_collation = exprCollation(rightop);
1938 : :
1939 : : /* look through any binary-compatible relabeling of rightop */
1940 : 17593 : rightop = strip_array_coercion(rightop);
1941 : :
1942 : : /*
1943 : : * Detect whether the operator is the default equality or inequality
1944 : : * operator of the array element type.
1945 : : */
1946 : 17593 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1947 [ + + ]: 17593 : if (OidIsValid(typentry->eq_opr))
1948 : : {
1949 [ + + ]: 17591 : if (operator == typentry->eq_opr)
1950 : 15412 : isEquality = true;
1951 [ + + ]: 2179 : else if (get_negator(operator) == typentry->eq_opr)
1952 : 1680 : isInequality = true;
1953 : : }
1954 : :
1955 : : /*
1956 : : * If it is equality or inequality, we might be able to estimate this as a
1957 : : * form of array containment; for instance "const = ANY(column)" can be
1958 : : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1959 : : * that, and returns the selectivity estimate if successful, or -1 if not.
1960 : : */
1961 [ + + + + : 17593 : if ((isEquality || isInequality) && !is_join_clause)
+ + ]
1962 : : {
1963 : 17091 : s1 = scalararraysel_containment(root, leftop, rightop,
1964 : : nominal_element_type,
1965 : : isEquality, useOr, varRelid);
1966 [ + + ]: 17091 : if (s1 >= 0.0)
1967 : 95 : return s1;
1968 : : }
1969 : :
1970 : : /*
1971 : : * Look up the underlying operator's selectivity estimator. Punt if it
1972 : : * hasn't got one.
1973 : : */
1974 [ + + ]: 17498 : if (is_join_clause)
1975 : 1 : oprsel = get_oprjoin(operator);
1976 : : else
1977 : 17497 : oprsel = get_oprrest(operator);
1978 [ + + ]: 17498 : if (!oprsel)
1979 : 2 : return (Selectivity) 0.5;
1980 : 17496 : fmgr_info(oprsel, &oprselproc);
1981 : :
1982 : : /*
1983 : : * In the array-containment check above, we must only believe that an
1984 : : * operator is equality or inequality if it is the default btree equality
1985 : : * operator (or its negator) for the element type, since those are the
1986 : : * operators that array containment will use. But in what follows, we can
1987 : : * be a little laxer, and also believe that any operators using eqsel() or
1988 : : * neqsel() as selectivity estimator act like equality or inequality.
1989 : : */
1990 [ + + + + ]: 17496 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1991 : 15487 : isEquality = true;
1992 [ + + - + ]: 2009 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1993 : 1609 : isInequality = true;
1994 : :
1995 : : /*
1996 : : * We consider three cases:
1997 : : *
1998 : : * 1. rightop is an Array constant: deconstruct the array, apply the
1999 : : * operator's selectivity function for each array element, and merge the
2000 : : * results in the same way that clausesel.c does for AND/OR combinations.
2001 : : *
2002 : : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
2003 : : * function for each element of the ARRAY[] construct, and merge.
2004 : : *
2005 : : * 3. otherwise, make a guess ...
2006 : : */
2007 [ + - + + ]: 17496 : if (rightop && IsA(rightop, Const))
2008 : 14135 : {
2009 : 14165 : Datum arraydatum = ((Const *) rightop)->constvalue;
2010 : 14165 : bool arrayisnull = ((Const *) rightop)->constisnull;
2011 : : ArrayType *arrayval;
2012 : : int16 elmlen;
2013 : : bool elmbyval;
2014 : : char elmalign;
2015 : : int num_elems;
2016 : : Datum *elem_values;
2017 : : bool *elem_nulls;
2018 : : int i;
2019 : :
2020 [ + + ]: 14165 : if (arrayisnull) /* qual can't succeed if null array */
2021 : 30 : return (Selectivity) 0.0;
2022 : 14140 : arrayval = DatumGetArrayTypeP(arraydatum);
2023 : :
2024 : : /*
2025 : : * When the array contains a NULL constant, same as var_eq_const, we
2026 : : * assume the operator is strict and nothing will match, thus return
2027 : : * 0.0.
2028 : : */
2029 [ + + + + ]: 14140 : if (!useOr && array_contains_nulls(arrayval))
2030 : 5 : return (Selectivity) 0.0;
2031 : :
2032 : 14135 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
2033 : : &elmlen, &elmbyval, &elmalign);
2034 : 14135 : deconstruct_array(arrayval,
2035 : : ARR_ELEMTYPE(arrayval),
2036 : : elmlen, elmbyval, elmalign,
2037 : : &elem_values, &elem_nulls, &num_elems);
2038 : :
2039 : : /*
2040 : : * For generic operators, we assume the probability of success is
2041 : : * independent for each array element. But for "= ANY" or "<> ALL",
2042 : : * if the array elements are distinct (which'd typically be the case)
2043 : : * then the probabilities are disjoint, and we should just sum them.
2044 : : *
2045 : : * If we were being really tense we would try to confirm that the
2046 : : * elements are all distinct, but that would be expensive and it
2047 : : * doesn't seem to be worth the cycles; it would amount to penalizing
2048 : : * well-written queries in favor of poorly-written ones. However, we
2049 : : * do protect ourselves a little bit by checking whether the
2050 : : * disjointness assumption leads to an impossible (out of range)
2051 : : * probability; if so, we fall back to the normal calculation.
2052 : : */
2053 [ + + ]: 14135 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2054 : :
2055 [ + + ]: 54820 : for (i = 0; i < num_elems; i++)
2056 : : {
2057 : : List *args;
2058 : : Selectivity s2;
2059 : :
2060 : 40685 : args = list_make2(leftop,
2061 : : makeConst(nominal_element_type,
2062 : : -1,
2063 : : nominal_element_collation,
2064 : : elmlen,
2065 : : elem_values[i],
2066 : : elem_nulls[i],
2067 : : elmbyval));
2068 [ - + ]: 40685 : if (is_join_clause)
2069 : 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2070 : : clause->inputcollid,
2071 : : PointerGetDatum(root),
2072 : : ObjectIdGetDatum(operator),
2073 : : PointerGetDatum(args),
2074 : : Int16GetDatum(jointype),
2075 : : PointerGetDatum(sjinfo)));
2076 : : else
2077 : 40685 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2078 : : clause->inputcollid,
2079 : : PointerGetDatum(root),
2080 : : ObjectIdGetDatum(operator),
2081 : : PointerGetDatum(args),
2082 : : Int32GetDatum(varRelid)));
2083 : :
2084 [ + + ]: 40685 : if (useOr)
2085 : : {
2086 : 35475 : s1 = s1 + s2 - s1 * s2;
2087 [ + + ]: 35475 : if (isEquality)
2088 : 34605 : s1disjoint += s2;
2089 : : }
2090 : : else
2091 : : {
2092 : 5210 : s1 = s1 * s2;
2093 [ + + ]: 5210 : if (isInequality)
2094 : 4950 : s1disjoint += s2 - 1.0;
2095 : : }
2096 : : }
2097 : :
2098 : : /* accept disjoint-probability estimate if in range */
2099 [ + + + + : 14135 : if ((useOr ? isEquality : isInequality) &&
+ + ]
2100 [ + + ]: 13610 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2101 : 13563 : s1 = s1disjoint;
2102 : : }
2103 [ + - + + ]: 3331 : else if (rightop && IsA(rightop, ArrayExpr) &&
2104 [ + - ]: 269 : !((ArrayExpr *) rightop)->multidims)
2105 : 264 : {
2106 : 269 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2107 : : int16 elmlen;
2108 : : bool elmbyval;
2109 : : ListCell *l;
2110 : :
2111 : 269 : get_typlenbyval(arrayexpr->element_typeid,
2112 : : &elmlen, &elmbyval);
2113 : :
2114 : : /*
2115 : : * We use the assumption of disjoint probabilities here too, although
2116 : : * the odds of equal array elements are rather higher if the elements
2117 : : * are not all constants (which they won't be, else constant folding
2118 : : * would have reduced the ArrayExpr to a Const). In this path it's
2119 : : * critical to have the sanity check on the s1disjoint estimate.
2120 : : */
2121 [ + + ]: 269 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2122 : :
2123 [ + - + + : 967 : foreach(l, arrayexpr->elements)
+ + ]
2124 : : {
2125 : 703 : Node *elem = (Node *) lfirst(l);
2126 : : List *args;
2127 : : Selectivity s2;
2128 : :
2129 : : /*
2130 : : * When the array contains a NULL constant, same as var_eq_const,
2131 : : * we assume the operator is strict and nothing will match, thus
2132 : : * return 0.0.
2133 : : */
2134 [ + + + + : 703 : if (!useOr && IsA(elem, Const) && ((Const *) elem)->constisnull)
+ + ]
2135 : 5 : return (Selectivity) 0.0;
2136 : :
2137 : : /*
2138 : : * Theoretically, if elem isn't of nominal_element_type we should
2139 : : * insert a RelabelType, but it seems unlikely that any operator
2140 : : * estimation function would really care ...
2141 : : */
2142 : 698 : args = list_make2(leftop, elem);
2143 [ + + ]: 698 : if (is_join_clause)
2144 : 3 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2145 : : clause->inputcollid,
2146 : : PointerGetDatum(root),
2147 : : ObjectIdGetDatum(operator),
2148 : : PointerGetDatum(args),
2149 : : Int16GetDatum(jointype),
2150 : : PointerGetDatum(sjinfo)));
2151 : : else
2152 : 695 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2153 : : clause->inputcollid,
2154 : : PointerGetDatum(root),
2155 : : ObjectIdGetDatum(operator),
2156 : : PointerGetDatum(args),
2157 : : Int32GetDatum(varRelid)));
2158 : :
2159 [ + + ]: 698 : if (useOr)
2160 : : {
2161 : 678 : s1 = s1 + s2 - s1 * s2;
2162 [ + - ]: 678 : if (isEquality)
2163 : 678 : s1disjoint += s2;
2164 : : }
2165 : : else
2166 : : {
2167 : 20 : s1 = s1 * s2;
2168 [ + - ]: 20 : if (isInequality)
2169 : 20 : s1disjoint += s2 - 1.0;
2170 : : }
2171 : : }
2172 : :
2173 : : /* accept disjoint-probability estimate if in range */
2174 [ + - + - : 264 : if ((useOr ? isEquality : isInequality) &&
+ - ]
2175 [ + - ]: 264 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2176 : 264 : s1 = s1disjoint;
2177 : : }
2178 : : else
2179 : : {
2180 : : CaseTestExpr *dummyexpr;
2181 : : List *args;
2182 : : Selectivity s2;
2183 : : int i;
2184 : :
2185 : : /*
2186 : : * We need a dummy rightop to pass to the operator selectivity
2187 : : * routine. It can be pretty much anything that doesn't look like a
2188 : : * constant; CaseTestExpr is a convenient choice.
2189 : : */
2190 : 3062 : dummyexpr = makeNode(CaseTestExpr);
2191 : 3062 : dummyexpr->typeId = nominal_element_type;
2192 : 3062 : dummyexpr->typeMod = -1;
2193 : 3062 : dummyexpr->collation = clause->inputcollid;
2194 : 3062 : args = list_make2(leftop, dummyexpr);
2195 [ - + ]: 3062 : if (is_join_clause)
2196 : 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2197 : : clause->inputcollid,
2198 : : PointerGetDatum(root),
2199 : : ObjectIdGetDatum(operator),
2200 : : PointerGetDatum(args),
2201 : : Int16GetDatum(jointype),
2202 : : PointerGetDatum(sjinfo)));
2203 : : else
2204 : 3062 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2205 : : clause->inputcollid,
2206 : : PointerGetDatum(root),
2207 : : ObjectIdGetDatum(operator),
2208 : : PointerGetDatum(args),
2209 : : Int32GetDatum(varRelid)));
2210 [ + - ]: 3062 : s1 = useOr ? 0.0 : 1.0;
2211 : :
2212 : : /*
2213 : : * Arbitrarily assume 10 elements in the eventual array value (see
2214 : : * also estimate_array_length). We don't risk an assumption of
2215 : : * disjoint probabilities here.
2216 : : */
2217 [ + + ]: 33682 : for (i = 0; i < 10; i++)
2218 : : {
2219 [ + - ]: 30620 : if (useOr)
2220 : 30620 : s1 = s1 + s2 - s1 * s2;
2221 : : else
2222 : 0 : s1 = s1 * s2;
2223 : : }
2224 : : }
2225 : :
2226 : : /* result should be in range, but make sure... */
2227 [ - + - + ]: 17461 : CLAMP_PROBABILITY(s1);
2228 : :
2229 : 17461 : return s1;
2230 : : }
2231 : :
2232 : : /*
2233 : : * Estimate number of elements in the array yielded by an expression.
2234 : : *
2235 : : * Note: the result is integral, but we use "double" to avoid overflow
2236 : : * concerns. Most callers will use it in double-type expressions anyway.
2237 : : *
2238 : : * Note: in some code paths root can be passed as NULL, resulting in
2239 : : * slightly worse estimates.
2240 : : */
2241 : : double
2242 : 84828 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2243 : : {
2244 : : /* look through any binary-compatible relabeling of arrayexpr */
2245 : 84828 : arrayexpr = strip_array_coercion(arrayexpr);
2246 : :
2247 [ + - + + ]: 84828 : if (arrayexpr && IsA(arrayexpr, Const))
2248 : : {
2249 : 37013 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2250 : 37013 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2251 : : ArrayType *arrayval;
2252 : :
2253 [ + + ]: 37013 : if (arrayisnull)
2254 : 75 : return 0;
2255 : 36938 : arrayval = DatumGetArrayTypeP(arraydatum);
2256 : 36938 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2257 : : }
2258 [ + - + + ]: 47815 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2259 [ + - ]: 503 : !((ArrayExpr *) arrayexpr)->multidims)
2260 : : {
2261 : 503 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2262 : : }
2263 [ + - + + ]: 47312 : else if (arrayexpr && root)
2264 : : {
2265 : : /* See if we can find any statistics about it */
2266 : : VariableStatData vardata;
2267 : : AttStatsSlot sslot;
2268 : 47292 : double nelem = 0;
2269 : :
2270 : : /*
2271 : : * Skip calling examine_variable for Var with varno 0, which has no
2272 : : * valid relation entry and would error in find_base_rel. Such a Var
2273 : : * can appear when a nested set operation's output type doesn't match
2274 : : * the parent's expected type, because recurse_set_operations builds a
2275 : : * projection target list using generate_setop_tlist with varno 0, and
2276 : : * if the required type coercion involves an ArrayCoerceExpr, we can
2277 : : * be called on that Var.
2278 : : */
2279 [ + + + + ]: 47292 : if (IsA(arrayexpr, Var) && ((Var *) arrayexpr)->varno == 0)
2280 : 4485 : return 10; /* default guess, should match scalararraysel */
2281 : :
2282 : 47287 : examine_variable(root, arrayexpr, 0, &vardata);
2283 [ + + ]: 47287 : if (HeapTupleIsValid(vardata.statsTuple))
2284 : : {
2285 : : /*
2286 : : * Found stats, so use the average element count, which is stored
2287 : : * in the last stanumbers element of the DECHIST statistics.
2288 : : * Actually that is the average count of *distinct* elements;
2289 : : * perhaps we should scale it up somewhat?
2290 : : */
2291 [ + + ]: 4575 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2292 : : STATISTIC_KIND_DECHIST, InvalidOid,
2293 : : ATTSTATSSLOT_NUMBERS))
2294 : : {
2295 [ + - ]: 4480 : if (sslot.nnumbers > 0)
2296 : 4480 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2297 : 4480 : free_attstatsslot(&sslot);
2298 : : }
2299 : : }
2300 [ + + ]: 47287 : ReleaseVariableStats(vardata);
2301 : :
2302 [ + + ]: 47287 : if (nelem > 0)
2303 : 4480 : return nelem;
2304 : : }
2305 : :
2306 : : /* Else use a default guess --- this should match scalararraysel */
2307 : 42827 : return 10;
2308 : : }
2309 : :
2310 : : /*
2311 : : * rowcomparesel - Selectivity of RowCompareExpr Node.
2312 : : *
2313 : : * We estimate RowCompare selectivity by considering just the first (high
2314 : : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2315 : : * this estimate could be refined by considering additional columns, it
2316 : : * seems unlikely that we could do a lot better without multi-column
2317 : : * statistics.
2318 : : */
2319 : : Selectivity
2320 : 240 : rowcomparesel(PlannerInfo *root,
2321 : : RowCompareExpr *clause,
2322 : : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2323 : : {
2324 : : Selectivity s1;
2325 : 240 : Oid opno = linitial_oid(clause->opnos);
2326 : 240 : Oid inputcollid = linitial_oid(clause->inputcollids);
2327 : : List *opargs;
2328 : : bool is_join_clause;
2329 : :
2330 : : /* Build equivalent arg list for single operator */
2331 : 240 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2332 : :
2333 : : /*
2334 : : * Decide if it's a join clause. This should match clausesel.c's
2335 : : * treat_as_join_clause(), except that we intentionally consider only the
2336 : : * leading columns and not the rest of the clause.
2337 : : */
2338 [ + + ]: 240 : if (varRelid != 0)
2339 : : {
2340 : : /*
2341 : : * Caller is forcing restriction mode (eg, because we are examining an
2342 : : * inner indexscan qual).
2343 : : */
2344 : 45 : is_join_clause = false;
2345 : : }
2346 [ + + ]: 195 : else if (sjinfo == NULL)
2347 : : {
2348 : : /*
2349 : : * It must be a restriction clause, since it's being evaluated at a
2350 : : * scan node.
2351 : : */
2352 : 175 : is_join_clause = false;
2353 : : }
2354 : : else
2355 : : {
2356 : : /*
2357 : : * Otherwise, it's a join if there's more than one base relation used.
2358 : : */
2359 : 20 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2360 : : }
2361 : :
2362 [ + + ]: 240 : if (is_join_clause)
2363 : : {
2364 : : /* Estimate selectivity for a join clause. */
2365 : 20 : s1 = join_selectivity(root, opno,
2366 : : opargs,
2367 : : inputcollid,
2368 : : jointype,
2369 : : sjinfo);
2370 : : }
2371 : : else
2372 : : {
2373 : : /* Estimate selectivity for a restriction clause. */
2374 : 220 : s1 = restriction_selectivity(root, opno,
2375 : : opargs,
2376 : : inputcollid,
2377 : : varRelid);
2378 : : }
2379 : :
2380 : 240 : return s1;
2381 : : }
2382 : :
2383 : : /*
2384 : : * eqjoinsel - Join selectivity of "="
2385 : : */
2386 : : Datum
2387 : 208629 : eqjoinsel(PG_FUNCTION_ARGS)
2388 : : {
2389 : 208629 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2390 : 208629 : Oid operator = PG_GETARG_OID(1);
2391 : 208629 : List *args = (List *) PG_GETARG_POINTER(2);
2392 : :
2393 : : #ifdef NOT_USED
2394 : : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2395 : : #endif
2396 : 208629 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2397 : 208629 : Oid collation = PG_GET_COLLATION();
2398 : : double selec;
2399 : : double selec_inner;
2400 : : VariableStatData vardata1;
2401 : : VariableStatData vardata2;
2402 : : double nd1;
2403 : : double nd2;
2404 : : bool isdefault1;
2405 : : bool isdefault2;
2406 : : Oid opfuncoid;
2407 : : FmgrInfo eqproc;
2408 : 208629 : Oid hashLeft = InvalidOid;
2409 : 208629 : Oid hashRight = InvalidOid;
2410 : : AttStatsSlot sslot1;
2411 : : AttStatsSlot sslot2;
2412 : 208629 : Form_pg_statistic stats1 = NULL;
2413 : 208629 : Form_pg_statistic stats2 = NULL;
2414 : 208629 : bool have_mcvs1 = false;
2415 : 208629 : bool have_mcvs2 = false;
2416 : 208629 : bool *hasmatch1 = NULL;
2417 : 208629 : bool *hasmatch2 = NULL;
2418 : 208629 : int nmatches = 0;
2419 : : bool get_mcv_stats;
2420 : : bool join_is_reversed;
2421 : : RelOptInfo *inner_rel;
2422 : :
2423 : 208629 : get_join_variables(root, args, sjinfo,
2424 : : &vardata1, &vardata2, &join_is_reversed);
2425 : :
2426 : 208629 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2427 : 208629 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2428 : :
2429 : 208629 : opfuncoid = get_opcode(operator);
2430 : :
2431 : 208629 : memset(&sslot1, 0, sizeof(sslot1));
2432 : 208629 : memset(&sslot2, 0, sizeof(sslot2));
2433 : :
2434 : : /*
2435 : : * There is no use in fetching one side's MCVs if we lack MCVs for the
2436 : : * other side, so do a quick check to verify that both stats exist.
2437 : : */
2438 : 561559 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2439 [ + + + + ]: 249718 : HeapTupleIsValid(vardata2.statsTuple) &&
2440 : 105417 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2441 : : STATISTIC_KIND_MCV, InvalidOid,
2442 [ + + + + ]: 352930 : 0) &&
2443 : 52526 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2444 : : STATISTIC_KIND_MCV, InvalidOid,
2445 : : 0));
2446 : :
2447 [ + + ]: 208629 : if (HeapTupleIsValid(vardata1.statsTuple))
2448 : : {
2449 : : /* note we allow use of nullfrac regardless of security check */
2450 : 144301 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2451 [ + + + - ]: 165684 : if (get_mcv_stats &&
2452 : 21383 : statistic_proc_security_check(&vardata1, opfuncoid))
2453 : 21383 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2454 : : STATISTIC_KIND_MCV, InvalidOid,
2455 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2456 : : }
2457 : :
2458 [ + + ]: 208629 : if (HeapTupleIsValid(vardata2.statsTuple))
2459 : : {
2460 : : /* note we allow use of nullfrac regardless of security check */
2461 : 121765 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2462 [ + + + - ]: 143148 : if (get_mcv_stats &&
2463 : 21383 : statistic_proc_security_check(&vardata2, opfuncoid))
2464 : 21383 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2465 : : STATISTIC_KIND_MCV, InvalidOid,
2466 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2467 : : }
2468 : :
2469 : : /* Prepare info usable by both eqjoinsel_inner and eqjoinsel_semi */
2470 [ + + + - ]: 208629 : if (have_mcvs1 && have_mcvs2)
2471 : : {
2472 : 21383 : fmgr_info(opfuncoid, &eqproc);
2473 : 21383 : hasmatch1 = (bool *) palloc0(sslot1.nvalues * sizeof(bool));
2474 : 21383 : hasmatch2 = (bool *) palloc0(sslot2.nvalues * sizeof(bool));
2475 : :
2476 : : /*
2477 : : * If the MCV lists are long enough to justify hashing, try to look up
2478 : : * hash functions for the join operator.
2479 : : */
2480 [ + + ]: 21383 : if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
2481 : 1143 : (void) get_op_hash_functions_ext(operator,
2482 : 1143 : exprType((Node *) linitial(args)),
2483 : : &hashLeft, &hashRight);
2484 : : }
2485 : : else
2486 : 187246 : memset(&eqproc, 0, sizeof(eqproc)); /* silence uninit-var warnings */
2487 : :
2488 : : /* We need to compute the inner-join selectivity in all cases */
2489 : 208629 : selec_inner = eqjoinsel_inner(&eqproc, collation,
2490 : : hashLeft, hashRight,
2491 : : &vardata1, &vardata2,
2492 : : nd1, nd2,
2493 : : isdefault1, isdefault2,
2494 : : &sslot1, &sslot2,
2495 : : stats1, stats2,
2496 : : have_mcvs1, have_mcvs2,
2497 : : hasmatch1, hasmatch2,
2498 : : &nmatches);
2499 : :
2500 [ + + - ]: 208629 : switch (sjinfo->jointype)
2501 : : {
2502 : 197420 : case JOIN_INNER:
2503 : : case JOIN_LEFT:
2504 : : case JOIN_FULL:
2505 : 197420 : selec = selec_inner;
2506 : 197420 : break;
2507 : 11209 : case JOIN_SEMI:
2508 : : case JOIN_ANTI:
2509 : :
2510 : : /*
2511 : : * Look up the join's inner relation. min_righthand is sufficient
2512 : : * information because neither SEMI nor ANTI joins permit any
2513 : : * reassociation into or out of their RHS, so the righthand will
2514 : : * always be exactly that set of rels.
2515 : : */
2516 : 11209 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2517 : :
2518 [ + + ]: 11209 : if (!join_is_reversed)
2519 : 5283 : selec = eqjoinsel_semi(&eqproc, collation,
2520 : : hashLeft, hashRight,
2521 : : false,
2522 : : &vardata1, &vardata2,
2523 : : nd1, nd2,
2524 : : isdefault1, isdefault2,
2525 : : &sslot1, &sslot2,
2526 : : stats1, stats2,
2527 : : have_mcvs1, have_mcvs2,
2528 : : hasmatch1, hasmatch2,
2529 : : &nmatches,
2530 : : inner_rel);
2531 : : else
2532 : 5926 : selec = eqjoinsel_semi(&eqproc, collation,
2533 : : hashLeft, hashRight,
2534 : : true,
2535 : : &vardata2, &vardata1,
2536 : : nd2, nd1,
2537 : : isdefault2, isdefault1,
2538 : : &sslot2, &sslot1,
2539 : : stats2, stats1,
2540 : : have_mcvs2, have_mcvs1,
2541 : : hasmatch2, hasmatch1,
2542 : : &nmatches,
2543 : : inner_rel);
2544 : :
2545 : : /*
2546 : : * We should never estimate the output of a semijoin to be more
2547 : : * rows than we estimate for an inner join with the same input
2548 : : * rels and join condition; it's obviously impossible for that to
2549 : : * happen. The former estimate is N1 * Ssemi while the latter is
2550 : : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2551 : : * this is worthwhile because of the shakier estimation rules we
2552 : : * use in eqjoinsel_semi, particularly in cases where it has to
2553 : : * punt entirely.
2554 : : */
2555 [ + + ]: 11209 : selec = Min(selec, inner_rel->rows * selec_inner);
2556 : 11209 : break;
2557 : 0 : default:
2558 : : /* other values not expected here */
2559 [ # # ]: 0 : elog(ERROR, "unrecognized join type: %d",
2560 : : (int) sjinfo->jointype);
2561 : : selec = 0; /* keep compiler quiet */
2562 : : break;
2563 : : }
2564 : :
2565 : 208629 : free_attstatsslot(&sslot1);
2566 : 208629 : free_attstatsslot(&sslot2);
2567 : :
2568 [ + + ]: 208629 : ReleaseVariableStats(vardata1);
2569 [ + + ]: 208629 : ReleaseVariableStats(vardata2);
2570 : :
2571 [ + + ]: 208629 : if (hasmatch1)
2572 : 21383 : pfree(hasmatch1);
2573 [ + + ]: 208629 : if (hasmatch2)
2574 : 21383 : pfree(hasmatch2);
2575 : :
2576 [ - + - + ]: 208629 : CLAMP_PROBABILITY(selec);
2577 : :
2578 : 208629 : PG_RETURN_FLOAT8((float8) selec);
2579 : : }
2580 : :
2581 : : /*
2582 : : * eqjoinsel_inner --- eqjoinsel for normal inner join
2583 : : *
2584 : : * In addition to computing the selectivity estimate, this will fill
2585 : : * hasmatch1[], hasmatch2[], and *p_nmatches (if have_mcvs1 && have_mcvs2).
2586 : : * We may be able to re-use that data in eqjoinsel_semi.
2587 : : *
2588 : : * We also use this for LEFT/FULL outer joins; it's not presently clear
2589 : : * that it's worth trying to distinguish them here.
2590 : : */
2591 : : static double
2592 : 208629 : eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
2593 : : Oid hashLeft, Oid hashRight,
2594 : : VariableStatData *vardata1, VariableStatData *vardata2,
2595 : : double nd1, double nd2,
2596 : : bool isdefault1, bool isdefault2,
2597 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2598 : : Form_pg_statistic stats1, Form_pg_statistic stats2,
2599 : : bool have_mcvs1, bool have_mcvs2,
2600 : : bool *hasmatch1, bool *hasmatch2,
2601 : : int *p_nmatches)
2602 : : {
2603 : : double selec;
2604 : :
2605 [ + + + - ]: 208629 : if (have_mcvs1 && have_mcvs2)
2606 : 21383 : {
2607 : : /*
2608 : : * We have most-common-value lists for both relations. Run through
2609 : : * the lists to see which MCVs actually join to each other with the
2610 : : * given operator. This allows us to determine the exact join
2611 : : * selectivity for the portion of the relations represented by the MCV
2612 : : * lists. We still have to estimate for the remaining population, but
2613 : : * in a skewed distribution this gives us a big leg up in accuracy.
2614 : : * For motivation see the analysis in Y. Ioannidis and S.
2615 : : * Christodoulakis, "On the propagation of errors in the size of join
2616 : : * results", Technical Report 1018, Computer Science Dept., University
2617 : : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2618 : : */
2619 : 21383 : double nullfrac1 = stats1->stanullfrac;
2620 : 21383 : double nullfrac2 = stats2->stanullfrac;
2621 : : double matchprodfreq,
2622 : : matchfreq1,
2623 : : matchfreq2,
2624 : : unmatchfreq1,
2625 : : unmatchfreq2,
2626 : : otherfreq1,
2627 : : otherfreq2,
2628 : : totalsel1,
2629 : : totalsel2;
2630 : : int i,
2631 : : nmatches;
2632 : :
2633 : : /* Fill the match arrays */
2634 : 21383 : eqjoinsel_find_matches(eqproc, collation,
2635 : : hashLeft, hashRight,
2636 : : false,
2637 : : sslot1, sslot2,
2638 : : sslot1->nvalues, sslot2->nvalues,
2639 : : hasmatch1, hasmatch2,
2640 : : p_nmatches, &matchprodfreq);
2641 : 21383 : nmatches = *p_nmatches;
2642 [ - + - + ]: 21383 : CLAMP_PROBABILITY(matchprodfreq);
2643 : :
2644 : : /* Sum up frequencies of matched and unmatched MCVs */
2645 : 21383 : matchfreq1 = unmatchfreq1 = 0.0;
2646 [ + + ]: 455004 : for (i = 0; i < sslot1->nvalues; i++)
2647 : : {
2648 [ + + ]: 433621 : if (hasmatch1[i])
2649 : 220107 : matchfreq1 += sslot1->numbers[i];
2650 : : else
2651 : 213514 : unmatchfreq1 += sslot1->numbers[i];
2652 : : }
2653 [ - + + + ]: 21383 : CLAMP_PROBABILITY(matchfreq1);
2654 [ - + - + ]: 21383 : CLAMP_PROBABILITY(unmatchfreq1);
2655 : 21383 : matchfreq2 = unmatchfreq2 = 0.0;
2656 [ + + ]: 364586 : for (i = 0; i < sslot2->nvalues; i++)
2657 : : {
2658 [ + + ]: 343203 : if (hasmatch2[i])
2659 : 220107 : matchfreq2 += sslot2->numbers[i];
2660 : : else
2661 : 123096 : unmatchfreq2 += sslot2->numbers[i];
2662 : : }
2663 [ - + + + ]: 21383 : CLAMP_PROBABILITY(matchfreq2);
2664 [ - + - + ]: 21383 : CLAMP_PROBABILITY(unmatchfreq2);
2665 : :
2666 : : /*
2667 : : * Compute total frequency of non-null values that are not in the MCV
2668 : : * lists.
2669 : : */
2670 : 21383 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2671 : 21383 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2672 [ + + - + ]: 21383 : CLAMP_PROBABILITY(otherfreq1);
2673 [ + + - + ]: 21383 : CLAMP_PROBABILITY(otherfreq2);
2674 : :
2675 : : /*
2676 : : * We can estimate the total selectivity from the point of view of
2677 : : * relation 1 as: the known selectivity for matched MCVs, plus
2678 : : * unmatched MCVs that are assumed to match against random members of
2679 : : * relation 2's non-MCV population, plus non-MCV values that are
2680 : : * assumed to match against random members of relation 2's unmatched
2681 : : * MCVs plus non-MCV values.
2682 : : */
2683 : 21383 : totalsel1 = matchprodfreq;
2684 [ + + ]: 21383 : if (nd2 > sslot2->nvalues)
2685 : 3625 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2686 [ + + ]: 21383 : if (nd2 > nmatches)
2687 : 7068 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2688 : 7068 : (nd2 - nmatches);
2689 : : /* Same estimate from the point of view of relation 2. */
2690 : 21383 : totalsel2 = matchprodfreq;
2691 [ + + ]: 21383 : if (nd1 > sslot1->nvalues)
2692 : 4199 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2693 [ + + ]: 21383 : if (nd1 > nmatches)
2694 : 6184 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2695 : 6184 : (nd1 - nmatches);
2696 : :
2697 : : /*
2698 : : * Use the smaller of the two estimates. This can be justified in
2699 : : * essentially the same terms as given below for the no-stats case: to
2700 : : * a first approximation, we are estimating from the point of view of
2701 : : * the relation with smaller nd.
2702 : : */
2703 [ + + ]: 21383 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2704 : : }
2705 : : else
2706 : : {
2707 : : /*
2708 : : * We do not have MCV lists for both sides. Estimate the join
2709 : : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2710 : : * is plausible if we assume that the join operator is strict and the
2711 : : * non-null values are about equally distributed: a given non-null
2712 : : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2713 : : * of rel2, so total join rows are at most
2714 : : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2715 : : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2716 : : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2717 : : * with MIN() is an upper bound. Using the MIN() means we estimate
2718 : : * from the point of view of the relation with smaller nd (since the
2719 : : * larger nd is determining the MIN). It is reasonable to assume that
2720 : : * most tuples in this rel will have join partners, so the bound is
2721 : : * probably reasonably tight and should be taken as-is.
2722 : : *
2723 : : * XXX Can we be smarter if we have an MCV list for just one side? It
2724 : : * seems that if we assume equal distribution for the other side, we
2725 : : * end up with the same answer anyway.
2726 : : */
2727 [ + + ]: 187246 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2728 [ + + ]: 187246 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2729 : :
2730 : 187246 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2731 [ + + ]: 187246 : if (nd1 > nd2)
2732 : 89896 : selec /= nd1;
2733 : : else
2734 : 97350 : selec /= nd2;
2735 : : }
2736 : :
2737 : 208629 : return selec;
2738 : : }
2739 : :
2740 : : /*
2741 : : * eqjoinsel_semi --- eqjoinsel for semi join
2742 : : *
2743 : : * (Also used for anti join, which we are supposed to estimate the same way.)
2744 : : * Caller has ensured that vardata1 is the LHS variable; however, eqproc
2745 : : * is for the original join operator, which might now need to have the inputs
2746 : : * swapped in order to apply correctly. Also, if have_mcvs1 && have_mcvs2
2747 : : * then hasmatch1[], hasmatch2[], and *p_nmatches were filled by
2748 : : * eqjoinsel_inner.
2749 : : */
2750 : : static double
2751 : 11209 : eqjoinsel_semi(FmgrInfo *eqproc, Oid collation,
2752 : : Oid hashLeft, Oid hashRight,
2753 : : bool op_is_reversed,
2754 : : VariableStatData *vardata1, VariableStatData *vardata2,
2755 : : double nd1, double nd2,
2756 : : bool isdefault1, bool isdefault2,
2757 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2758 : : Form_pg_statistic stats1, Form_pg_statistic stats2,
2759 : : bool have_mcvs1, bool have_mcvs2,
2760 : : bool *hasmatch1, bool *hasmatch2,
2761 : : int *p_nmatches,
2762 : : RelOptInfo *inner_rel)
2763 : : {
2764 : : double selec;
2765 : :
2766 : : /*
2767 : : * We clamp nd2 to be not more than what we estimate the inner relation's
2768 : : * size to be. This is intuitively somewhat reasonable since obviously
2769 : : * there can't be more than that many distinct values coming from the
2770 : : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2771 : : * likewise) is that this is the only pathway by which restriction clauses
2772 : : * applied to the inner rel will affect the join result size estimate,
2773 : : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2774 : : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2775 : : * the selectivity of outer-rel restrictions.
2776 : : *
2777 : : * We can apply this clamping both with respect to the base relation from
2778 : : * which the join variable comes (if there is just one), and to the
2779 : : * immediate inner input relation of the current join.
2780 : : *
2781 : : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2782 : : * great, maybe, but it didn't come out of nowhere either. This is most
2783 : : * helpful when the inner relation is empty and consequently has no stats.
2784 : : */
2785 [ + + ]: 11209 : if (vardata2->rel)
2786 : : {
2787 [ + + ]: 11204 : if (nd2 >= vardata2->rel->rows)
2788 : : {
2789 : 9444 : nd2 = vardata2->rel->rows;
2790 : 9444 : isdefault2 = false;
2791 : : }
2792 : : }
2793 [ + + ]: 11209 : if (nd2 >= inner_rel->rows)
2794 : : {
2795 : 9394 : nd2 = inner_rel->rows;
2796 : 9394 : isdefault2 = false;
2797 : : }
2798 : :
2799 [ + + + - ]: 11209 : if (have_mcvs1 && have_mcvs2)
2800 : 513 : {
2801 : : /*
2802 : : * We have most-common-value lists for both relations. Run through
2803 : : * the lists to see which MCVs actually join to each other with the
2804 : : * given operator. This allows us to determine the exact join
2805 : : * selectivity for the portion of the relations represented by the MCV
2806 : : * lists. We still have to estimate for the remaining population, but
2807 : : * in a skewed distribution this gives us a big leg up in accuracy.
2808 : : */
2809 : 513 : double nullfrac1 = stats1->stanullfrac;
2810 : : double matchprodfreq,
2811 : : matchfreq1,
2812 : : uncertainfrac,
2813 : : uncertain;
2814 : : int i,
2815 : : nmatches,
2816 : : clamped_nvalues2;
2817 : :
2818 : : /*
2819 : : * The clamping above could have resulted in nd2 being less than
2820 : : * sslot2->nvalues; in which case, we assume that precisely the nd2
2821 : : * most common values in the relation will appear in the join input,
2822 : : * and so compare to only the first nd2 members of the MCV list. Of
2823 : : * course this is frequently wrong, but it's the best bet we can make.
2824 : : */
2825 [ + + ]: 513 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2826 : :
2827 : : /*
2828 : : * If we did not set clamped_nvalues2 to less than sslot2->nvalues,
2829 : : * then the hasmatch1[] and hasmatch2[] match flags computed by
2830 : : * eqjoinsel_inner are still perfectly applicable, so we need not
2831 : : * re-do the matching work. Note that it does not matter if
2832 : : * op_is_reversed: we'd get the same answers.
2833 : : *
2834 : : * If we did clamp, then a different set of sslot2 values is to be
2835 : : * compared, so we have to re-do the matching.
2836 : : */
2837 [ - + ]: 513 : if (clamped_nvalues2 != sslot2->nvalues)
2838 : : {
2839 : : /* Must re-zero the arrays */
2840 : 0 : memset(hasmatch1, 0, sslot1->nvalues * sizeof(bool));
2841 : 0 : memset(hasmatch2, 0, clamped_nvalues2 * sizeof(bool));
2842 : : /* Re-fill the match arrays */
2843 : 0 : eqjoinsel_find_matches(eqproc, collation,
2844 : : hashLeft, hashRight,
2845 : : op_is_reversed,
2846 : : sslot1, sslot2,
2847 : : sslot1->nvalues, clamped_nvalues2,
2848 : : hasmatch1, hasmatch2,
2849 : : p_nmatches, &matchprodfreq);
2850 : : }
2851 : 513 : nmatches = *p_nmatches;
2852 : :
2853 : : /* Sum up frequencies of matched MCVs */
2854 : 513 : matchfreq1 = 0.0;
2855 [ + + ]: 11143 : for (i = 0; i < sslot1->nvalues; i++)
2856 : : {
2857 [ + + ]: 10630 : if (hasmatch1[i])
2858 : 9563 : matchfreq1 += sslot1->numbers[i];
2859 : : }
2860 [ - + + + ]: 513 : CLAMP_PROBABILITY(matchfreq1);
2861 : :
2862 : : /*
2863 : : * Now we need to estimate the fraction of relation 1 that has at
2864 : : * least one join partner. We know for certain that the matched MCVs
2865 : : * do, so that gives us a lower bound, but we're really in the dark
2866 : : * about everything else. Our crude approach is: if nd1 <= nd2 then
2867 : : * assume all non-null rel1 rows have join partners, else assume for
2868 : : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2869 : : * can discount the known-matched MCVs from the distinct-values counts
2870 : : * before doing the division.
2871 : : *
2872 : : * Crude as the above is, it's completely useless if we don't have
2873 : : * reliable ndistinct values for both sides. Hence, if either nd1 or
2874 : : * nd2 is default, punt and assume half of the uncertain rows have
2875 : : * join partners.
2876 : : */
2877 [ + - + - ]: 513 : if (!isdefault1 && !isdefault2)
2878 : : {
2879 : 513 : nd1 -= nmatches;
2880 : 513 : nd2 -= nmatches;
2881 [ + + - + ]: 513 : if (nd1 <= nd2 || nd2 < 0)
2882 : 483 : uncertainfrac = 1.0;
2883 : : else
2884 : 30 : uncertainfrac = nd2 / nd1;
2885 : : }
2886 : : else
2887 : 0 : uncertainfrac = 0.5;
2888 : 513 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2889 [ - + - + ]: 513 : CLAMP_PROBABILITY(uncertain);
2890 : 513 : selec = matchfreq1 + uncertainfrac * uncertain;
2891 : : }
2892 : : else
2893 : : {
2894 : : /*
2895 : : * Without MCV lists for both sides, we can only use the heuristic
2896 : : * about nd1 vs nd2.
2897 : : */
2898 [ + + ]: 10696 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2899 : :
2900 [ + + + + ]: 10696 : if (!isdefault1 && !isdefault2)
2901 : : {
2902 [ + + - + ]: 7880 : if (nd1 <= nd2 || nd2 < 0)
2903 : 4335 : selec = 1.0 - nullfrac1;
2904 : : else
2905 : 3545 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2906 : : }
2907 : : else
2908 : 2816 : selec = 0.5 * (1.0 - nullfrac1);
2909 : : }
2910 : :
2911 : 11209 : return selec;
2912 : : }
2913 : :
2914 : : /*
2915 : : * Identify matching MCVs for eqjoinsel_inner or eqjoinsel_semi.
2916 : : *
2917 : : * Inputs:
2918 : : * eqproc: FmgrInfo for equality function to use (might be reversed)
2919 : : * collation: OID of collation to use
2920 : : * hashLeft, hashRight: OIDs of hash functions associated with equality op,
2921 : : * or InvalidOid if we're not to use hashing
2922 : : * op_is_reversed: indicates that eqproc compares right type to left type
2923 : : * sslot1, sslot2: MCV values for the lefthand and righthand inputs
2924 : : * nvalues1, nvalues2: number of values to be considered (can be less than
2925 : : * sslotN->nvalues, but not more)
2926 : : * Outputs:
2927 : : * hasmatch1[], hasmatch2[]: pre-zeroed arrays of lengths nvalues1, nvalues2;
2928 : : * entries are set to true if that MCV has a match on the other side
2929 : : * *p_nmatches: receives number of MCV pairs that match
2930 : : * *p_matchprodfreq: receives sum(sslot1->numbers[i] * sslot2->numbers[j])
2931 : : * for matching MCVs
2932 : : *
2933 : : * Note that hashLeft is for the eqproc's left-hand input type, hashRight
2934 : : * for its right, regardless of op_is_reversed.
2935 : : *
2936 : : * Note we assume that each MCV will match at most one member of the other
2937 : : * MCV list. If the operator isn't really equality, there could be multiple
2938 : : * matches --- but we don't look for them, both for speed and because the
2939 : : * math wouldn't add up...
2940 : : */
2941 : : static void
2942 : 21383 : eqjoinsel_find_matches(FmgrInfo *eqproc, Oid collation,
2943 : : Oid hashLeft, Oid hashRight,
2944 : : bool op_is_reversed,
2945 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2946 : : int nvalues1, int nvalues2,
2947 : : bool *hasmatch1, bool *hasmatch2,
2948 : : int *p_nmatches, double *p_matchprodfreq)
2949 : : {
2950 : 21383 : LOCAL_FCINFO(fcinfo, 2);
2951 : 21383 : double matchprodfreq = 0.0;
2952 : 21383 : int nmatches = 0;
2953 : :
2954 : : /*
2955 : : * Save a few cycles by setting up the fcinfo struct just once. Using
2956 : : * FunctionCallInvoke directly also avoids failure if the eqproc returns
2957 : : * NULL, though really equality functions should never do that.
2958 : : */
2959 : 21383 : InitFunctionCallInfoData(*fcinfo, eqproc, 2, collation,
2960 : : NULL, NULL);
2961 : 21383 : fcinfo->args[0].isnull = false;
2962 : 21383 : fcinfo->args[1].isnull = false;
2963 : :
2964 [ + + + - ]: 21383 : if (OidIsValid(hashLeft) && OidIsValid(hashRight))
2965 : 1143 : {
2966 : : /* Use a hash table to speed up the matching */
2967 : 1143 : LOCAL_FCINFO(hash_fcinfo, 1);
2968 : : FmgrInfo hash_proc;
2969 : : MCVHashContext hashContext;
2970 : : MCVHashTable_hash *hashTable;
2971 : : AttStatsSlot *statsProbe;
2972 : : AttStatsSlot *statsHash;
2973 : : bool *hasMatchProbe;
2974 : : bool *hasMatchHash;
2975 : : int nvaluesProbe;
2976 : : int nvaluesHash;
2977 : :
2978 : : /* Make sure we build the hash table on the smaller array. */
2979 [ + - ]: 1143 : if (sslot1->nvalues >= sslot2->nvalues)
2980 : : {
2981 : 1143 : statsProbe = sslot1;
2982 : 1143 : statsHash = sslot2;
2983 : 1143 : hasMatchProbe = hasmatch1;
2984 : 1143 : hasMatchHash = hasmatch2;
2985 : 1143 : nvaluesProbe = nvalues1;
2986 : 1143 : nvaluesHash = nvalues2;
2987 : : }
2988 : : else
2989 : : {
2990 : : /* We'll have to reverse the direction of use of the operator. */
2991 : 0 : op_is_reversed = !op_is_reversed;
2992 : 0 : statsProbe = sslot2;
2993 : 0 : statsHash = sslot1;
2994 : 0 : hasMatchProbe = hasmatch2;
2995 : 0 : hasMatchHash = hasmatch1;
2996 : 0 : nvaluesProbe = nvalues2;
2997 : 0 : nvaluesHash = nvalues1;
2998 : : }
2999 : :
3000 : : /*
3001 : : * Build the hash table on the smaller array, using the appropriate
3002 : : * hash function for its data type.
3003 : : */
3004 [ - + ]: 1143 : fmgr_info(op_is_reversed ? hashLeft : hashRight, &hash_proc);
3005 : 1143 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
3006 : : NULL, NULL);
3007 : 1143 : hash_fcinfo->args[0].isnull = false;
3008 : :
3009 : 1143 : hashContext.equal_fcinfo = fcinfo;
3010 : 1143 : hashContext.hash_fcinfo = hash_fcinfo;
3011 : 1143 : hashContext.op_is_reversed = op_is_reversed;
3012 : 1143 : hashContext.insert_mode = true;
3013 : 1143 : get_typlenbyval(statsHash->valuetype,
3014 : : &hashContext.hash_typlen,
3015 : : &hashContext.hash_typbyval);
3016 : :
3017 : 1143 : hashTable = MCVHashTable_create(CurrentMemoryContext,
3018 : : nvaluesHash,
3019 : : &hashContext);
3020 : :
3021 [ + + ]: 115443 : for (int i = 0; i < nvaluesHash; i++)
3022 : : {
3023 : 114300 : bool found = false;
3024 : 114300 : MCVHashEntry *entry = MCVHashTable_insert(hashTable,
3025 : 114300 : statsHash->values[i],
3026 : : &found);
3027 : :
3028 : : /*
3029 : : * MCVHashTable_insert will only report "found" if the new value
3030 : : * is equal to some previous one per datum_image_eq(). That
3031 : : * probably shouldn't happen, since we're not expecting duplicates
3032 : : * in the MCV list. If we do find a dup, just ignore it, leaving
3033 : : * the hash entry's index pointing at the first occurrence. That
3034 : : * matches the behavior that the non-hashed code path would have.
3035 : : */
3036 [ + - ]: 114300 : if (likely(!found))
3037 : 114300 : entry->index = i;
3038 : : }
3039 : :
3040 : : /*
3041 : : * Prepare to probe the hash table. If the probe values are of a
3042 : : * different data type, then we need to change hash functions. (This
3043 : : * code relies on the assumption that since we defined SH_STORE_HASH,
3044 : : * simplehash.h will never need to compute hash values for existing
3045 : : * hash table entries.)
3046 : : */
3047 : 1143 : hashContext.insert_mode = false;
3048 [ - + ]: 1143 : if (hashLeft != hashRight)
3049 : : {
3050 [ # # ]: 0 : fmgr_info(op_is_reversed ? hashRight : hashLeft, &hash_proc);
3051 : : /* Resetting hash_fcinfo is probably unnecessary, but be safe */
3052 : 0 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
3053 : : NULL, NULL);
3054 : 0 : hash_fcinfo->args[0].isnull = false;
3055 : : }
3056 : :
3057 : : /* Look up each probe value in turn. */
3058 [ + + ]: 115443 : for (int i = 0; i < nvaluesProbe; i++)
3059 : : {
3060 : 114300 : MCVHashEntry *entry = MCVHashTable_lookup(hashTable,
3061 : 114300 : statsProbe->values[i]);
3062 : :
3063 : : /* As in the other code path, skip already-matched hash entries */
3064 [ + + + - ]: 114300 : if (entry != NULL && !hasMatchHash[entry->index])
3065 : : {
3066 : 53265 : hasMatchHash[entry->index] = hasMatchProbe[i] = true;
3067 : 53265 : nmatches++;
3068 : 53265 : matchprodfreq += statsHash->numbers[entry->index] * statsProbe->numbers[i];
3069 : : }
3070 : : }
3071 : :
3072 : 1143 : MCVHashTable_destroy(hashTable);
3073 : : }
3074 : : else
3075 : : {
3076 : : /* We're not to use hashing, so do it the O(N^2) way */
3077 : : int index1,
3078 : : index2;
3079 : :
3080 : : /* Set up to supply the values in the order the operator expects */
3081 [ - + ]: 20240 : if (op_is_reversed)
3082 : : {
3083 : 0 : index1 = 1;
3084 : 0 : index2 = 0;
3085 : : }
3086 : : else
3087 : : {
3088 : 20240 : index1 = 0;
3089 : 20240 : index2 = 1;
3090 : : }
3091 : :
3092 [ + + ]: 339561 : for (int i = 0; i < nvalues1; i++)
3093 : : {
3094 : 319321 : fcinfo->args[index1].value = sslot1->values[i];
3095 : :
3096 [ + + ]: 6257729 : for (int j = 0; j < nvalues2; j++)
3097 : : {
3098 : : Datum fresult;
3099 : :
3100 [ + + ]: 6105250 : if (hasmatch2[j])
3101 : 2072731 : continue;
3102 : 4032519 : fcinfo->args[index2].value = sslot2->values[j];
3103 : 4032519 : fcinfo->isnull = false;
3104 : 4032519 : fresult = FunctionCallInvoke(fcinfo);
3105 [ + - + + ]: 4032519 : if (!fcinfo->isnull && DatumGetBool(fresult))
3106 : : {
3107 : 166842 : hasmatch1[i] = hasmatch2[j] = true;
3108 : 166842 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
3109 : 166842 : nmatches++;
3110 : 166842 : break;
3111 : : }
3112 : : }
3113 : : }
3114 : : }
3115 : :
3116 : 21383 : *p_nmatches = nmatches;
3117 : 21383 : *p_matchprodfreq = matchprodfreq;
3118 : 21383 : }
3119 : :
3120 : : /*
3121 : : * Support functions for the hash tables used by eqjoinsel_find_matches
3122 : : */
3123 : : static uint32
3124 : 228600 : hash_mcv(MCVHashTable_hash *tab, Datum key)
3125 : : {
3126 : 228600 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3127 : 228600 : FunctionCallInfo fcinfo = context->hash_fcinfo;
3128 : : Datum fresult;
3129 : :
3130 : 228600 : fcinfo->args[0].value = key;
3131 : 228600 : fcinfo->isnull = false;
3132 : 228600 : fresult = FunctionCallInvoke(fcinfo);
3133 : : Assert(!fcinfo->isnull);
3134 : 228600 : return DatumGetUInt32(fresult);
3135 : : }
3136 : :
3137 : : static bool
3138 : 53265 : mcvs_equal(MCVHashTable_hash *tab, Datum key0, Datum key1)
3139 : : {
3140 : 53265 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3141 : :
3142 [ - + ]: 53265 : if (context->insert_mode)
3143 : : {
3144 : : /*
3145 : : * During the insertion step, any comparisons will be between two
3146 : : * Datums of the hash table's data type, so if the given operator is
3147 : : * cross-type it will be the wrong thing to use. Fortunately, we can
3148 : : * use datum_image_eq instead. The MCV values should all be distinct
3149 : : * anyway, so it's mostly pro-forma to compare them at all.
3150 : : */
3151 : 0 : return datum_image_eq(key0, key1,
3152 : 0 : context->hash_typbyval, context->hash_typlen);
3153 : : }
3154 : : else
3155 : : {
3156 : 53265 : FunctionCallInfo fcinfo = context->equal_fcinfo;
3157 : : Datum fresult;
3158 : :
3159 : : /*
3160 : : * Apply the operator the correct way around. Although simplehash.h
3161 : : * doesn't document this explicitly, during lookups key0 is from the
3162 : : * hash table while key1 is the probe value, so we should compare them
3163 : : * in that order only if op_is_reversed.
3164 : : */
3165 [ - + ]: 53265 : if (context->op_is_reversed)
3166 : : {
3167 : 0 : fcinfo->args[0].value = key0;
3168 : 0 : fcinfo->args[1].value = key1;
3169 : : }
3170 : : else
3171 : : {
3172 : 53265 : fcinfo->args[0].value = key1;
3173 : 53265 : fcinfo->args[1].value = key0;
3174 : : }
3175 : 53265 : fcinfo->isnull = false;
3176 : 53265 : fresult = FunctionCallInvoke(fcinfo);
3177 [ + - + - ]: 53265 : return (!fcinfo->isnull && DatumGetBool(fresult));
3178 : : }
3179 : : }
3180 : :
3181 : : /*
3182 : : * neqjoinsel - Join selectivity of "!="
3183 : : */
3184 : : Datum
3185 : 2459 : neqjoinsel(PG_FUNCTION_ARGS)
3186 : : {
3187 : 2459 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3188 : 2459 : Oid operator = PG_GETARG_OID(1);
3189 : 2459 : List *args = (List *) PG_GETARG_POINTER(2);
3190 : 2459 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
3191 : 2459 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
3192 : 2459 : Oid collation = PG_GET_COLLATION();
3193 : : float8 result;
3194 : :
3195 [ + + - + ]: 2459 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
3196 : 879 : {
3197 : : /*
3198 : : * For semi-joins, if there is more than one distinct value in the RHS
3199 : : * relation then every non-null LHS row must find a row to join since
3200 : : * it can only be equal to one of them. We'll assume that there is
3201 : : * always more than one distinct RHS value for the sake of stability,
3202 : : * though in theory we could have special cases for empty RHS
3203 : : * (selectivity = 0) and single-distinct-value RHS (selectivity =
3204 : : * fraction of LHS that has the same value as the single RHS value).
3205 : : *
3206 : : * For anti-joins, if we use the same assumption that there is more
3207 : : * than one distinct key in the RHS relation, then every non-null LHS
3208 : : * row must be suppressed by the anti-join.
3209 : : *
3210 : : * So either way, the selectivity estimate should be 1 - nullfrac.
3211 : : */
3212 : : VariableStatData leftvar;
3213 : : VariableStatData rightvar;
3214 : : bool reversed;
3215 : : HeapTuple statsTuple;
3216 : : double nullfrac;
3217 : :
3218 : 879 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
3219 [ + + ]: 879 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
3220 [ + + ]: 879 : if (HeapTupleIsValid(statsTuple))
3221 : 702 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
3222 : : else
3223 : 177 : nullfrac = 0.0;
3224 [ + + ]: 879 : ReleaseVariableStats(leftvar);
3225 [ + + ]: 879 : ReleaseVariableStats(rightvar);
3226 : :
3227 : 879 : result = 1.0 - nullfrac;
3228 : : }
3229 : : else
3230 : : {
3231 : : /*
3232 : : * We want 1 - eqjoinsel() where the equality operator is the one
3233 : : * associated with this != operator, that is, its negator.
3234 : : */
3235 : 1580 : Oid eqop = get_negator(operator);
3236 : :
3237 [ + - ]: 1580 : if (eqop)
3238 : : {
3239 : : result =
3240 : 1580 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
3241 : : collation,
3242 : : PointerGetDatum(root),
3243 : : ObjectIdGetDatum(eqop),
3244 : : PointerGetDatum(args),
3245 : : Int16GetDatum(jointype),
3246 : : PointerGetDatum(sjinfo)));
3247 : : }
3248 : : else
3249 : : {
3250 : : /* Use default selectivity (should we raise an error instead?) */
3251 : 0 : result = DEFAULT_EQ_SEL;
3252 : : }
3253 : 1580 : result = 1.0 - result;
3254 : : }
3255 : :
3256 : 2459 : PG_RETURN_FLOAT8(result);
3257 : : }
3258 : :
3259 : : /*
3260 : : * scalarltjoinsel - Join selectivity of "<" for scalars
3261 : : */
3262 : : Datum
3263 : 270 : scalarltjoinsel(PG_FUNCTION_ARGS)
3264 : : {
3265 : 270 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3266 : : }
3267 : :
3268 : : /*
3269 : : * scalarlejoinsel - Join selectivity of "<=" for scalars
3270 : : */
3271 : : Datum
3272 : 198 : scalarlejoinsel(PG_FUNCTION_ARGS)
3273 : : {
3274 : 198 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3275 : : }
3276 : :
3277 : : /*
3278 : : * scalargtjoinsel - Join selectivity of ">" for scalars
3279 : : */
3280 : : Datum
3281 : 240 : scalargtjoinsel(PG_FUNCTION_ARGS)
3282 : : {
3283 : 240 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3284 : : }
3285 : :
3286 : : /*
3287 : : * scalargejoinsel - Join selectivity of ">=" for scalars
3288 : : */
3289 : : Datum
3290 : 152 : scalargejoinsel(PG_FUNCTION_ARGS)
3291 : : {
3292 : 152 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3293 : : }
3294 : :
3295 : :
3296 : : /*
3297 : : * mergejoinscansel - Scan selectivity of merge join.
3298 : : *
3299 : : * A merge join will stop as soon as it exhausts either input stream.
3300 : : * Therefore, if we can estimate the ranges of both input variables,
3301 : : * we can estimate how much of the input will actually be read. This
3302 : : * can have a considerable impact on the cost when using indexscans.
3303 : : *
3304 : : * Also, we can estimate how much of each input has to be read before the
3305 : : * first join pair is found, which will affect the join's startup time.
3306 : : *
3307 : : * clause should be a clause already known to be mergejoinable. opfamily,
3308 : : * cmptype, and nulls_first specify the sort ordering being used.
3309 : : *
3310 : : * The outputs are:
3311 : : * *leftstart is set to the fraction of the left-hand variable expected
3312 : : * to be scanned before the first join pair is found (0 to 1).
3313 : : * *leftend is set to the fraction of the left-hand variable expected
3314 : : * to be scanned before the join terminates (0 to 1).
3315 : : * *rightstart, *rightend similarly for the right-hand variable.
3316 : : */
3317 : : void
3318 : 97814 : mergejoinscansel(PlannerInfo *root, Node *clause,
3319 : : Oid opfamily, CompareType cmptype, bool nulls_first,
3320 : : Selectivity *leftstart, Selectivity *leftend,
3321 : : Selectivity *rightstart, Selectivity *rightend)
3322 : : {
3323 : : Node *left,
3324 : : *right;
3325 : : VariableStatData leftvar,
3326 : : rightvar;
3327 : : Oid opmethod;
3328 : : int op_strategy;
3329 : : Oid op_lefttype;
3330 : : Oid op_righttype;
3331 : : Oid opno,
3332 : : collation,
3333 : : lsortop,
3334 : : rsortop,
3335 : : lstatop,
3336 : : rstatop,
3337 : : ltop,
3338 : : leop,
3339 : : revltop,
3340 : : revleop;
3341 : : StrategyNumber ltstrat,
3342 : : lestrat,
3343 : : gtstrat,
3344 : : gestrat;
3345 : : bool isgt;
3346 : : Datum leftmin,
3347 : : leftmax,
3348 : : rightmin,
3349 : : rightmax;
3350 : : double selec;
3351 : :
3352 : : /* Set default results if we can't figure anything out. */
3353 : : /* XXX should default "start" fraction be a bit more than 0? */
3354 : 97814 : *leftstart = *rightstart = 0.0;
3355 : 97814 : *leftend = *rightend = 1.0;
3356 : :
3357 : : /* Deconstruct the merge clause */
3358 [ - + ]: 97814 : if (!is_opclause(clause))
3359 : 0 : return; /* shouldn't happen */
3360 : 97814 : opno = ((OpExpr *) clause)->opno;
3361 : 97814 : collation = ((OpExpr *) clause)->inputcollid;
3362 : 97814 : left = get_leftop((Expr *) clause);
3363 : 97814 : right = get_rightop((Expr *) clause);
3364 [ - + ]: 97814 : if (!right)
3365 : 0 : return; /* shouldn't happen */
3366 : :
3367 : : /* Look for stats for the inputs */
3368 : 97814 : examine_variable(root, left, 0, &leftvar);
3369 : 97814 : examine_variable(root, right, 0, &rightvar);
3370 : :
3371 : 97814 : opmethod = get_opfamily_method(opfamily);
3372 : :
3373 : : /* Extract the operator's declared left/right datatypes */
3374 : 97814 : get_op_opfamily_properties(opno, opfamily, false,
3375 : : &op_strategy,
3376 : : &op_lefttype,
3377 : : &op_righttype);
3378 : : Assert(IndexAmTranslateStrategy(op_strategy, opmethod, opfamily, true) == COMPARE_EQ);
3379 : :
3380 : : /*
3381 : : * Look up the various operators we need. If we don't find them all, it
3382 : : * probably means the opfamily is broken, but we just fail silently.
3383 : : *
3384 : : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3385 : : * operator, regardless of which sort direction we are considering.
3386 : : */
3387 [ + + - ]: 97814 : switch (cmptype)
3388 : : {
3389 : 97785 : case COMPARE_LT:
3390 : 97785 : isgt = false;
3391 : 97785 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3392 : 97785 : lestrat = IndexAmTranslateCompareType(COMPARE_LE, opmethod, opfamily, true);
3393 [ + + ]: 97785 : if (op_lefttype == op_righttype)
3394 : : {
3395 : : /* easy case */
3396 : 96719 : ltop = get_opfamily_member(opfamily,
3397 : : op_lefttype, op_righttype,
3398 : : ltstrat);
3399 : 96719 : leop = get_opfamily_member(opfamily,
3400 : : op_lefttype, op_righttype,
3401 : : lestrat);
3402 : 96719 : lsortop = ltop;
3403 : 96719 : rsortop = ltop;
3404 : 96719 : lstatop = lsortop;
3405 : 96719 : rstatop = rsortop;
3406 : 96719 : revltop = ltop;
3407 : 96719 : revleop = leop;
3408 : : }
3409 : : else
3410 : : {
3411 : 1066 : ltop = get_opfamily_member(opfamily,
3412 : : op_lefttype, op_righttype,
3413 : : ltstrat);
3414 : 1066 : leop = get_opfamily_member(opfamily,
3415 : : op_lefttype, op_righttype,
3416 : : lestrat);
3417 : 1066 : lsortop = get_opfamily_member(opfamily,
3418 : : op_lefttype, op_lefttype,
3419 : : ltstrat);
3420 : 1066 : rsortop = get_opfamily_member(opfamily,
3421 : : op_righttype, op_righttype,
3422 : : ltstrat);
3423 : 1066 : lstatop = lsortop;
3424 : 1066 : rstatop = rsortop;
3425 : 1066 : revltop = get_opfamily_member(opfamily,
3426 : : op_righttype, op_lefttype,
3427 : : ltstrat);
3428 : 1066 : revleop = get_opfamily_member(opfamily,
3429 : : op_righttype, op_lefttype,
3430 : : lestrat);
3431 : : }
3432 : 97785 : break;
3433 : 29 : case COMPARE_GT:
3434 : : /* descending-order case */
3435 : 29 : isgt = true;
3436 : 29 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3437 : 29 : gtstrat = IndexAmTranslateCompareType(COMPARE_GT, opmethod, opfamily, true);
3438 : 29 : gestrat = IndexAmTranslateCompareType(COMPARE_GE, opmethod, opfamily, true);
3439 [ + - ]: 29 : if (op_lefttype == op_righttype)
3440 : : {
3441 : : /* easy case */
3442 : 29 : ltop = get_opfamily_member(opfamily,
3443 : : op_lefttype, op_righttype,
3444 : : gtstrat);
3445 : 29 : leop = get_opfamily_member(opfamily,
3446 : : op_lefttype, op_righttype,
3447 : : gestrat);
3448 : 29 : lsortop = ltop;
3449 : 29 : rsortop = ltop;
3450 : 29 : lstatop = get_opfamily_member(opfamily,
3451 : : op_lefttype, op_lefttype,
3452 : : ltstrat);
3453 : 29 : rstatop = lstatop;
3454 : 29 : revltop = ltop;
3455 : 29 : revleop = leop;
3456 : : }
3457 : : else
3458 : : {
3459 : 0 : ltop = get_opfamily_member(opfamily,
3460 : : op_lefttype, op_righttype,
3461 : : gtstrat);
3462 : 0 : leop = get_opfamily_member(opfamily,
3463 : : op_lefttype, op_righttype,
3464 : : gestrat);
3465 : 0 : lsortop = get_opfamily_member(opfamily,
3466 : : op_lefttype, op_lefttype,
3467 : : gtstrat);
3468 : 0 : rsortop = get_opfamily_member(opfamily,
3469 : : op_righttype, op_righttype,
3470 : : gtstrat);
3471 : 0 : lstatop = get_opfamily_member(opfamily,
3472 : : op_lefttype, op_lefttype,
3473 : : ltstrat);
3474 : 0 : rstatop = get_opfamily_member(opfamily,
3475 : : op_righttype, op_righttype,
3476 : : ltstrat);
3477 : 0 : revltop = get_opfamily_member(opfamily,
3478 : : op_righttype, op_lefttype,
3479 : : gtstrat);
3480 : 0 : revleop = get_opfamily_member(opfamily,
3481 : : op_righttype, op_lefttype,
3482 : : gestrat);
3483 : : }
3484 : 29 : break;
3485 : 0 : default:
3486 : 0 : goto fail; /* shouldn't get here */
3487 : : }
3488 : :
3489 [ + - + - ]: 97814 : if (!OidIsValid(lsortop) ||
3490 [ + - ]: 97814 : !OidIsValid(rsortop) ||
3491 [ + - ]: 97814 : !OidIsValid(lstatop) ||
3492 [ + + ]: 97814 : !OidIsValid(rstatop) ||
3493 [ + - ]: 97804 : !OidIsValid(ltop) ||
3494 [ + - ]: 97804 : !OidIsValid(leop) ||
3495 [ - + ]: 97804 : !OidIsValid(revltop) ||
3496 : : !OidIsValid(revleop))
3497 : 10 : goto fail; /* insufficient info in catalogs */
3498 : :
3499 : : /* Try to get ranges of both inputs */
3500 [ + + ]: 97804 : if (!isgt)
3501 : : {
3502 [ + + ]: 97775 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3503 : : &leftmin, &leftmax))
3504 : 29622 : goto fail; /* no range available from stats */
3505 [ + + ]: 68153 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3506 : : &rightmin, &rightmax))
3507 : 17357 : goto fail; /* no range available from stats */
3508 : : }
3509 : : else
3510 : : {
3511 : : /* need to swap the max and min */
3512 [ + + ]: 29 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3513 : : &leftmax, &leftmin))
3514 : 24 : goto fail; /* no range available from stats */
3515 [ - + ]: 5 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3516 : : &rightmax, &rightmin))
3517 : 0 : goto fail; /* no range available from stats */
3518 : : }
3519 : :
3520 : : /*
3521 : : * Now, the fraction of the left variable that will be scanned is the
3522 : : * fraction that's <= the right-side maximum value. But only believe
3523 : : * non-default estimates, else stick with our 1.0.
3524 : : */
3525 : 50801 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3526 : : rightmax, op_righttype);
3527 [ + + ]: 50801 : if (selec != DEFAULT_INEQ_SEL)
3528 : 50797 : *leftend = selec;
3529 : :
3530 : : /* And similarly for the right variable. */
3531 : 50801 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3532 : : leftmax, op_lefttype);
3533 [ + - ]: 50801 : if (selec != DEFAULT_INEQ_SEL)
3534 : 50801 : *rightend = selec;
3535 : :
3536 : : /*
3537 : : * Only one of the two "end" fractions can really be less than 1.0;
3538 : : * believe the smaller estimate and reset the other one to exactly 1.0. If
3539 : : * we get exactly equal estimates (as can easily happen with self-joins),
3540 : : * believe neither.
3541 : : */
3542 [ + + ]: 50801 : if (*leftend > *rightend)
3543 : 16342 : *leftend = 1.0;
3544 [ + + ]: 34459 : else if (*leftend < *rightend)
3545 : 18444 : *rightend = 1.0;
3546 : : else
3547 : 16015 : *leftend = *rightend = 1.0;
3548 : :
3549 : : /*
3550 : : * Also, the fraction of the left variable that will be scanned before the
3551 : : * first join pair is found is the fraction that's < the right-side
3552 : : * minimum value. But only believe non-default estimates, else stick with
3553 : : * our own default.
3554 : : */
3555 : 50801 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3556 : : rightmin, op_righttype);
3557 [ + - ]: 50801 : if (selec != DEFAULT_INEQ_SEL)
3558 : 50801 : *leftstart = selec;
3559 : :
3560 : : /* And similarly for the right variable. */
3561 : 50801 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3562 : : leftmin, op_lefttype);
3563 [ + - ]: 50801 : if (selec != DEFAULT_INEQ_SEL)
3564 : 50801 : *rightstart = selec;
3565 : :
3566 : : /*
3567 : : * Only one of the two "start" fractions can really be more than zero;
3568 : : * believe the larger estimate and reset the other one to exactly 0.0. If
3569 : : * we get exactly equal estimates (as can easily happen with self-joins),
3570 : : * believe neither.
3571 : : */
3572 [ + + ]: 50801 : if (*leftstart < *rightstart)
3573 : 9604 : *leftstart = 0.0;
3574 [ + + ]: 41197 : else if (*leftstart > *rightstart)
3575 : 13554 : *rightstart = 0.0;
3576 : : else
3577 : 27643 : *leftstart = *rightstart = 0.0;
3578 : :
3579 : : /*
3580 : : * If the sort order is nulls-first, we're going to have to skip over any
3581 : : * nulls too. These would not have been counted by scalarineqsel, and we
3582 : : * can safely add in this fraction regardless of whether we believe
3583 : : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3584 : : */
3585 [ + + ]: 50801 : if (nulls_first)
3586 : : {
3587 : : Form_pg_statistic stats;
3588 : :
3589 [ + - ]: 5 : if (HeapTupleIsValid(leftvar.statsTuple))
3590 : : {
3591 : 5 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3592 : 5 : *leftstart += stats->stanullfrac;
3593 [ - + - + ]: 5 : CLAMP_PROBABILITY(*leftstart);
3594 : 5 : *leftend += stats->stanullfrac;
3595 [ - + - + ]: 5 : CLAMP_PROBABILITY(*leftend);
3596 : : }
3597 [ + - ]: 5 : if (HeapTupleIsValid(rightvar.statsTuple))
3598 : : {
3599 : 5 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3600 : 5 : *rightstart += stats->stanullfrac;
3601 [ - + - + ]: 5 : CLAMP_PROBABILITY(*rightstart);
3602 : 5 : *rightend += stats->stanullfrac;
3603 [ - + - + ]: 5 : CLAMP_PROBABILITY(*rightend);
3604 : : }
3605 : : }
3606 : :
3607 : : /* Disbelieve start >= end, just in case that can happen */
3608 [ + + ]: 50801 : if (*leftstart >= *leftend)
3609 : : {
3610 : 124 : *leftstart = 0.0;
3611 : 124 : *leftend = 1.0;
3612 : : }
3613 [ + + ]: 50801 : if (*rightstart >= *rightend)
3614 : : {
3615 : 581 : *rightstart = 0.0;
3616 : 581 : *rightend = 1.0;
3617 : : }
3618 : :
3619 : 50220 : fail:
3620 [ + + ]: 97814 : ReleaseVariableStats(leftvar);
3621 [ + + ]: 97814 : ReleaseVariableStats(rightvar);
3622 : : }
3623 : :
3624 : :
3625 : : /*
3626 : : * matchingsel -- generic matching-operator selectivity support
3627 : : *
3628 : : * Use these for any operators that (a) are on data types for which we collect
3629 : : * standard statistics, and (b) have behavior for which the default estimate
3630 : : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3631 : : * operators.
3632 : : */
3633 : :
3634 : : Datum
3635 : 845 : matchingsel(PG_FUNCTION_ARGS)
3636 : : {
3637 : 845 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3638 : 845 : Oid operator = PG_GETARG_OID(1);
3639 : 845 : List *args = (List *) PG_GETARG_POINTER(2);
3640 : 845 : int varRelid = PG_GETARG_INT32(3);
3641 : 845 : Oid collation = PG_GET_COLLATION();
3642 : : double selec;
3643 : :
3644 : : /* Use generic restriction selectivity logic. */
3645 : 845 : selec = generic_restriction_selectivity(root, operator, collation,
3646 : : args, varRelid,
3647 : : DEFAULT_MATCHING_SEL);
3648 : :
3649 : 845 : PG_RETURN_FLOAT8((float8) selec);
3650 : : }
3651 : :
3652 : : Datum
3653 : 5 : matchingjoinsel(PG_FUNCTION_ARGS)
3654 : : {
3655 : : /* Just punt, for the moment. */
3656 : 5 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3657 : : }
3658 : :
3659 : :
3660 : : /*
3661 : : * Helper routine for estimate_num_groups: add an item to a list of
3662 : : * GroupVarInfos, but only if it's not known equal to any of the existing
3663 : : * entries.
3664 : : */
3665 : : typedef struct
3666 : : {
3667 : : Node *var; /* might be an expression, not just a Var */
3668 : : RelOptInfo *rel; /* relation it belongs to */
3669 : : double ndistinct; /* # distinct values */
3670 : : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3671 : : } GroupVarInfo;
3672 : :
3673 : : static List *
3674 : 279682 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3675 : : Node *var, VariableStatData *vardata)
3676 : : {
3677 : : GroupVarInfo *varinfo;
3678 : : double ndistinct;
3679 : : bool isdefault;
3680 : : ListCell *lc;
3681 : :
3682 : 279682 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3683 : :
3684 : : /*
3685 : : * The nullingrels bits within the var could cause the same var to be
3686 : : * counted multiple times if it's marked with different nullingrels. They
3687 : : * could also prevent us from matching the var to the expressions in
3688 : : * extended statistics (see estimate_multivariate_ndistinct). So strip
3689 : : * them out first.
3690 : : */
3691 : 279682 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3692 : :
3693 [ + + + + : 372900 : foreach(lc, varinfos)
+ + ]
3694 : : {
3695 : 96582 : varinfo = (GroupVarInfo *) lfirst(lc);
3696 : :
3697 : : /* Drop exact duplicates */
3698 [ + + ]: 96582 : if (equal(var, varinfo->var))
3699 : 3364 : return varinfos;
3700 : :
3701 : : /*
3702 : : * Drop known-equal vars, but only if they belong to different
3703 : : * relations (see comments for estimate_num_groups). We aren't too
3704 : : * fussy about the semantics of "equal" here.
3705 : : */
3706 [ + + + + ]: 118466 : if (vardata->rel != varinfo->rel &&
3707 : 25054 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3708 : : {
3709 [ + + ]: 2801 : if (varinfo->ndistinct <= ndistinct)
3710 : : {
3711 : : /* Keep older item, forget new one */
3712 : 194 : return varinfos;
3713 : : }
3714 : : else
3715 : : {
3716 : : /* Delete the older item */
3717 : 2607 : varinfos = foreach_delete_current(varinfos, lc);
3718 : : }
3719 : : }
3720 : : }
3721 : :
3722 : 276318 : varinfo = palloc_object(GroupVarInfo);
3723 : :
3724 : 276318 : varinfo->var = var;
3725 : 276318 : varinfo->rel = vardata->rel;
3726 : 276318 : varinfo->ndistinct = ndistinct;
3727 : 276318 : varinfo->isdefault = isdefault;
3728 : 276318 : varinfos = lappend(varinfos, varinfo);
3729 : 276318 : return varinfos;
3730 : : }
3731 : :
3732 : : /*
3733 : : * estimate_num_groups - Estimate number of groups in a grouped query
3734 : : *
3735 : : * Given a query having a GROUP BY clause, estimate how many groups there
3736 : : * will be --- ie, the number of distinct combinations of the GROUP BY
3737 : : * expressions.
3738 : : *
3739 : : * This routine is also used to estimate the number of rows emitted by
3740 : : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3741 : : * actually, we only use it for DISTINCT when there's no grouping or
3742 : : * aggregation ahead of the DISTINCT.)
3743 : : *
3744 : : * Inputs:
3745 : : * root - the query
3746 : : * groupExprs - list of expressions being grouped by
3747 : : * input_rows - number of rows estimated to arrive at the group/unique
3748 : : * filter step
3749 : : * pgset - NULL, or a List** pointing to a grouping set to filter the
3750 : : * groupExprs against
3751 : : *
3752 : : * Outputs:
3753 : : * estinfo - When passed as non-NULL, the function will set bits in the
3754 : : * "flags" field in order to provide callers with additional information
3755 : : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3756 : : * bit if we used any default values in the estimation.
3757 : : *
3758 : : * Given the lack of any cross-correlation statistics in the system, it's
3759 : : * impossible to do anything really trustworthy with GROUP BY conditions
3760 : : * involving multiple Vars. We should however avoid assuming the worst
3761 : : * case (all possible cross-product terms actually appear as groups) since
3762 : : * very often the grouped-by Vars are highly correlated. Our current approach
3763 : : * is as follows:
3764 : : * 1. Expressions yielding boolean are assumed to contribute two groups,
3765 : : * independently of their content, and are ignored in the subsequent
3766 : : * steps. This is mainly because tests like "col IS NULL" break the
3767 : : * heuristic used in step 2 especially badly.
3768 : : * 2. Reduce the given expressions to a list of unique Vars used. For
3769 : : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3770 : : * It is clearly correct not to count the same Var more than once.
3771 : : * It is also reasonable to treat f(x) the same as x: f() cannot
3772 : : * increase the number of distinct values (unless it is volatile,
3773 : : * which we consider unlikely for grouping), but it probably won't
3774 : : * reduce the number of distinct values much either.
3775 : : * As a special case, if a GROUP BY expression can be matched to an
3776 : : * expressional index for which we have statistics, then we treat the
3777 : : * whole expression as though it were just a Var.
3778 : : * 3. If the list contains Vars of different relations that are known equal
3779 : : * due to equivalence classes, then drop all but one of the Vars from each
3780 : : * known-equal set, keeping the one with smallest estimated # of values
3781 : : * (since the extra values of the others can't appear in joined rows).
3782 : : * Note the reason we only consider Vars of different relations is that
3783 : : * if we considered ones of the same rel, we'd be double-counting the
3784 : : * restriction selectivity of the equality in the next step.
3785 : : * 4. For Vars within a single source rel, we multiply together the numbers
3786 : : * of values, clamp to the number of rows in the rel (divided by 10 if
3787 : : * more than one Var), and then multiply by a factor based on the
3788 : : * selectivity of the restriction clauses for that rel. When there's
3789 : : * more than one Var, the initial product is probably too high (it's the
3790 : : * worst case) but clamping to a fraction of the rel's rows seems to be a
3791 : : * helpful heuristic for not letting the estimate get out of hand. (The
3792 : : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3793 : : * we multiply by to adjust for the restriction selectivity assumes that
3794 : : * the restriction clauses are independent of the grouping, which may not
3795 : : * be a valid assumption, but it's hard to do better.
3796 : : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3797 : : * rel, and multiply the results together.
3798 : : * Note that rels not containing grouped Vars are ignored completely, as are
3799 : : * join clauses. Such rels cannot increase the number of groups, and we
3800 : : * assume such clauses do not reduce the number either (somewhat bogus,
3801 : : * but we don't have the info to do better).
3802 : : */
3803 : : double
3804 : 234700 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3805 : : List **pgset, EstimationInfo *estinfo)
3806 : : {
3807 : 234700 : List *varinfos = NIL;
3808 : 234700 : double srf_multiplier = 1.0;
3809 : : double numdistinct;
3810 : : ListCell *l;
3811 : : int i;
3812 : :
3813 : : /* Zero the estinfo output parameter, if non-NULL */
3814 [ + + ]: 234700 : if (estinfo != NULL)
3815 : 184577 : memset(estinfo, 0, sizeof(EstimationInfo));
3816 : :
3817 : : /*
3818 : : * We don't ever want to return an estimate of zero groups, as that tends
3819 : : * to lead to division-by-zero and other unpleasantness. The input_rows
3820 : : * estimate is usually already at least 1, but clamp it just in case it
3821 : : * isn't.
3822 : : */
3823 : 234700 : input_rows = clamp_row_est(input_rows);
3824 : :
3825 : : /*
3826 : : * If no grouping columns, there's exactly one group. (This can't happen
3827 : : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3828 : : * corner cases with set operations.)
3829 : : */
3830 [ + + + + : 234700 : if (groupExprs == NIL || (pgset && *pgset == NIL))
+ + ]
3831 : 1025 : return 1.0;
3832 : :
3833 : : /*
3834 : : * Count groups derived from boolean grouping expressions. For other
3835 : : * expressions, find the unique Vars used, treating an expression as a Var
3836 : : * if we can find stats for it. For each one, record the statistical
3837 : : * estimate of number of distinct values (total in its table, without
3838 : : * regard for filtering).
3839 : : */
3840 : 233675 : numdistinct = 1.0;
3841 : :
3842 : 233675 : i = 0;
3843 [ + - + + : 516866 : foreach(l, groupExprs)
+ + ]
3844 : : {
3845 : 283241 : Node *groupexpr = (Node *) lfirst(l);
3846 : : double this_srf_multiplier;
3847 : : VariableStatData vardata;
3848 : : List *varshere;
3849 : : ListCell *l2;
3850 : :
3851 : : /* is expression in this grouping set? */
3852 [ + + + + ]: 283241 : if (pgset && !list_member_int(*pgset, i++))
3853 : 219039 : continue;
3854 : :
3855 : : /*
3856 : : * Set-returning functions in grouping columns are a bit problematic.
3857 : : * The code below will effectively ignore their SRF nature and come up
3858 : : * with a numdistinct estimate as though they were scalar functions.
3859 : : * We compensate by scaling up the end result by the largest SRF
3860 : : * rowcount estimate. (This will be an overestimate if the SRF
3861 : : * produces multiple copies of any output value, but it seems best to
3862 : : * assume the SRF's outputs are distinct. In any case, it's probably
3863 : : * pointless to worry too much about this without much better
3864 : : * estimates for SRF output rowcounts than we have today.)
3865 : : */
3866 : 282573 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3867 [ + + ]: 282573 : if (srf_multiplier < this_srf_multiplier)
3868 : 170 : srf_multiplier = this_srf_multiplier;
3869 : :
3870 : : /* Short-circuit for expressions returning boolean */
3871 [ + + ]: 282573 : if (exprType(groupexpr) == BOOLOID)
3872 : : {
3873 : 732 : numdistinct *= 2.0;
3874 : 732 : continue;
3875 : : }
3876 : :
3877 : : /*
3878 : : * If examine_variable is able to deduce anything about the GROUP BY
3879 : : * expression, treat it as a single variable even if it's really more
3880 : : * complicated.
3881 : : *
3882 : : * XXX This has the consequence that if there's a statistics object on
3883 : : * the expression, we don't split it into individual Vars. This
3884 : : * affects our selection of statistics in
3885 : : * estimate_multivariate_ndistinct, because it's probably better to
3886 : : * use more accurate estimate for each expression and treat them as
3887 : : * independent, than to combine estimates for the extracted variables
3888 : : * when we don't know how that relates to the expressions.
3889 : : */
3890 : 281841 : examine_variable(root, groupexpr, 0, &vardata);
3891 [ + + + + ]: 281841 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3892 : : {
3893 : 205677 : varinfos = add_unique_group_var(root, varinfos,
3894 : : groupexpr, &vardata);
3895 [ + + ]: 205677 : ReleaseVariableStats(vardata);
3896 : 205677 : continue;
3897 : : }
3898 [ - + ]: 76164 : ReleaseVariableStats(vardata);
3899 : :
3900 : : /*
3901 : : * Else pull out the component Vars. Handle PlaceHolderVars by
3902 : : * recursing into their arguments (effectively assuming that the
3903 : : * PlaceHolderVar doesn't change the number of groups, which boils
3904 : : * down to ignoring the possible addition of nulls to the result set).
3905 : : */
3906 : 76164 : varshere = pull_var_clause(groupexpr,
3907 : : PVC_RECURSE_AGGREGATES |
3908 : : PVC_RECURSE_WINDOWFUNCS |
3909 : : PVC_RECURSE_PLACEHOLDERS);
3910 : :
3911 : : /*
3912 : : * If we find any variable-free GROUP BY item, then either it is a
3913 : : * constant (and we can ignore it) or it contains a volatile function;
3914 : : * in the latter case we punt and assume that each input row will
3915 : : * yield a distinct group.
3916 : : */
3917 [ + + ]: 76164 : if (varshere == NIL)
3918 : : {
3919 [ + + ]: 12012 : if (contain_volatile_functions(groupexpr))
3920 : 50 : return input_rows;
3921 : 11962 : continue;
3922 : : }
3923 : :
3924 : : /*
3925 : : * Else add variables to varinfos list
3926 : : */
3927 [ + - + + : 138157 : foreach(l2, varshere)
+ + ]
3928 : : {
3929 : 74005 : Node *var = (Node *) lfirst(l2);
3930 : :
3931 : 74005 : examine_variable(root, var, 0, &vardata);
3932 : 74005 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3933 [ + + ]: 74005 : ReleaseVariableStats(vardata);
3934 : : }
3935 : : }
3936 : :
3937 : : /*
3938 : : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3939 : : * list.
3940 : : */
3941 [ + + ]: 233625 : if (varinfos == NIL)
3942 : : {
3943 : : /* Apply SRF multiplier as we would do in the long path */
3944 : 911 : numdistinct *= srf_multiplier;
3945 : : /* Round off */
3946 : 911 : numdistinct = ceil(numdistinct);
3947 : : /* Guard against out-of-range answers */
3948 [ + + ]: 911 : if (numdistinct > input_rows)
3949 : 61 : numdistinct = input_rows;
3950 [ - + ]: 911 : if (numdistinct < 1.0)
3951 : 0 : numdistinct = 1.0;
3952 : 911 : return numdistinct;
3953 : : }
3954 : :
3955 : : /*
3956 : : * Group Vars by relation and estimate total numdistinct.
3957 : : *
3958 : : * For each iteration of the outer loop, we process the frontmost Var in
3959 : : * varinfos, plus all other Vars in the same relation. We remove these
3960 : : * Vars from the newvarinfos list for the next iteration. This is the
3961 : : * easiest way to group Vars of same rel together.
3962 : : */
3963 : : do
3964 : : {
3965 : 237523 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3966 : 237523 : RelOptInfo *rel = varinfo1->rel;
3967 : 237523 : double reldistinct = 1;
3968 : 237523 : double relmaxndistinct = reldistinct;
3969 : 237523 : int relvarcount = 0;
3970 : 237523 : List *newvarinfos = NIL;
3971 : 237523 : List *relvarinfos = NIL;
3972 : :
3973 : : /*
3974 : : * Split the list of varinfos in two - one for the current rel, one
3975 : : * for remaining Vars on other rels.
3976 : : */
3977 : 237523 : relvarinfos = lappend(relvarinfos, varinfo1);
3978 [ + - + + : 284711 : for_each_from(l, varinfos, 1)
+ + ]
3979 : : {
3980 : 47188 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3981 : :
3982 [ + + ]: 47188 : if (varinfo2->rel == varinfo1->rel)
3983 : : {
3984 : : /* varinfos on current rel */
3985 : 36188 : relvarinfos = lappend(relvarinfos, varinfo2);
3986 : : }
3987 : : else
3988 : : {
3989 : : /* not time to process varinfo2 yet */
3990 : 11000 : newvarinfos = lappend(newvarinfos, varinfo2);
3991 : : }
3992 : : }
3993 : :
3994 : : /*
3995 : : * Get the numdistinct estimate for the Vars of this rel. We
3996 : : * iteratively search for multivariate n-distinct with maximum number
3997 : : * of vars; assuming that each var group is independent of the others,
3998 : : * we multiply them together. Any remaining relvarinfos after no more
3999 : : * multivariate matches are found are assumed independent too, so
4000 : : * their individual ndistinct estimates are multiplied also.
4001 : : *
4002 : : * While iterating, count how many separate numdistinct values we
4003 : : * apply. We apply a fudge factor below, but only if we multiplied
4004 : : * more than one such values.
4005 : : */
4006 [ + + ]: 475151 : while (relvarinfos)
4007 : : {
4008 : : double mvndistinct;
4009 : :
4010 [ + + ]: 237628 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
4011 : : &mvndistinct))
4012 : : {
4013 : 345 : reldistinct *= mvndistinct;
4014 [ + + ]: 345 : if (relmaxndistinct < mvndistinct)
4015 : 335 : relmaxndistinct = mvndistinct;
4016 : 345 : relvarcount++;
4017 : : }
4018 : : else
4019 : : {
4020 [ + - + + : 510264 : foreach(l, relvarinfos)
+ + ]
4021 : : {
4022 : 272981 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
4023 : :
4024 : 272981 : reldistinct *= varinfo2->ndistinct;
4025 [ + + ]: 272981 : if (relmaxndistinct < varinfo2->ndistinct)
4026 : 238029 : relmaxndistinct = varinfo2->ndistinct;
4027 : 272981 : relvarcount++;
4028 : :
4029 : : /*
4030 : : * When varinfo2's isdefault is set then we'd better set
4031 : : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
4032 : : */
4033 [ + + + + ]: 272981 : if (estinfo != NULL && varinfo2->isdefault)
4034 : 18236 : estinfo->flags |= SELFLAG_USED_DEFAULT;
4035 : : }
4036 : :
4037 : : /* we're done with this relation */
4038 : 237283 : relvarinfos = NIL;
4039 : : }
4040 : : }
4041 : :
4042 : : /*
4043 : : * Sanity check --- don't divide by zero if empty relation.
4044 : : */
4045 : : Assert(IS_SIMPLE_REL(rel));
4046 [ + + ]: 237523 : if (rel->tuples > 0)
4047 : : {
4048 : : /*
4049 : : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
4050 : : * fudge factor is because the Vars are probably correlated but we
4051 : : * don't know by how much. We should never clamp to less than the
4052 : : * largest ndistinct value for any of the Vars, though, since
4053 : : * there will surely be at least that many groups.
4054 : : */
4055 : 233313 : double clamp = rel->tuples;
4056 : :
4057 [ + + ]: 233313 : if (relvarcount > 1)
4058 : : {
4059 : 27936 : clamp *= 0.1;
4060 [ + + ]: 27936 : if (clamp < relmaxndistinct)
4061 : : {
4062 : 25750 : clamp = relmaxndistinct;
4063 : : /* for sanity in case some ndistinct is too large: */
4064 [ + + ]: 25750 : if (clamp > rel->tuples)
4065 : 110 : clamp = rel->tuples;
4066 : : }
4067 : : }
4068 [ + + ]: 233313 : if (reldistinct > clamp)
4069 : 23258 : reldistinct = clamp;
4070 : :
4071 : : /*
4072 : : * Update the estimate based on the restriction selectivity,
4073 : : * guarding against division by zero when reldistinct is zero.
4074 : : * Also skip this if we know that we are returning all rows.
4075 : : */
4076 [ + - + + ]: 233313 : if (reldistinct > 0 && rel->rows < rel->tuples)
4077 : : {
4078 : : /*
4079 : : * Given a table containing N rows with n distinct values in a
4080 : : * uniform distribution, if we select p rows at random then
4081 : : * the expected number of distinct values selected is
4082 : : *
4083 : : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
4084 : : *
4085 : : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
4086 : : *
4087 : : * See "Approximating block accesses in database
4088 : : * organizations", S. B. Yao, Communications of the ACM,
4089 : : * Volume 20 Issue 4, April 1977 Pages 260-261.
4090 : : *
4091 : : * Alternatively, re-arranging the terms from the factorials,
4092 : : * this may be written as
4093 : : *
4094 : : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
4095 : : *
4096 : : * This form of the formula is more efficient to compute in
4097 : : * the common case where p is larger than N/n. Additionally,
4098 : : * as pointed out by Dell'Era, if i << N for all terms in the
4099 : : * product, it can be approximated by
4100 : : *
4101 : : * n * (1 - ((N-p)/N)^(N/n))
4102 : : *
4103 : : * See "Expected distinct values when selecting from a bag
4104 : : * without replacement", Alberto Dell'Era,
4105 : : * http://www.adellera.it/investigations/distinct_balls/.
4106 : : *
4107 : : * The condition i << N is equivalent to n >> 1, so this is a
4108 : : * good approximation when the number of distinct values in
4109 : : * the table is large. It turns out that this formula also
4110 : : * works well even when n is small.
4111 : : */
4112 : 66393 : reldistinct *=
4113 : 66393 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
4114 : 66393 : rel->tuples / reldistinct));
4115 : : }
4116 : 233313 : reldistinct = clamp_row_est(reldistinct);
4117 : :
4118 : : /*
4119 : : * Update estimate of total distinct groups.
4120 : : */
4121 : 233313 : numdistinct *= reldistinct;
4122 : : }
4123 : :
4124 : 237523 : varinfos = newvarinfos;
4125 [ + + ]: 237523 : } while (varinfos != NIL);
4126 : :
4127 : : /* Now we can account for the effects of any SRFs */
4128 : 232714 : numdistinct *= srf_multiplier;
4129 : :
4130 : : /* Round off */
4131 : 232714 : numdistinct = ceil(numdistinct);
4132 : :
4133 : : /* Guard against out-of-range answers */
4134 [ + + ]: 232714 : if (numdistinct > input_rows)
4135 : 49941 : numdistinct = input_rows;
4136 [ - + ]: 232714 : if (numdistinct < 1.0)
4137 : 0 : numdistinct = 1.0;
4138 : :
4139 : 232714 : return numdistinct;
4140 : : }
4141 : :
4142 : : /*
4143 : : * Try to estimate the bucket size of the hash join inner side when the join
4144 : : * condition contains two or more clauses by employing extended statistics.
4145 : : *
4146 : : * The main idea of this approach is that the distinct value generated by
4147 : : * multivariate estimation on two or more columns would provide less bucket size
4148 : : * than estimation on one separate column.
4149 : : *
4150 : : * IMPORTANT: It is crucial to synchronize the approach of combining different
4151 : : * estimations with the caller's method.
4152 : : *
4153 : : * Return a list of clauses that didn't fetch any extended statistics.
4154 : : */
4155 : : List *
4156 : 335102 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
4157 : : List *hashclauses,
4158 : : Selectivity *innerbucketsize)
4159 : : {
4160 : : List *clauses;
4161 : : List *otherclauses;
4162 : : double ndistinct;
4163 : :
4164 [ + + ]: 335102 : if (list_length(hashclauses) <= 1)
4165 : : {
4166 : : /*
4167 : : * Nothing to do for a single clause. Could we employ univariate
4168 : : * extended stat here?
4169 : : */
4170 : 308321 : return hashclauses;
4171 : : }
4172 : :
4173 : : /* "clauses" is the list of hashclauses we've not dealt with yet */
4174 : 26781 : clauses = list_copy(hashclauses);
4175 : : /* "otherclauses" holds clauses we are going to return to caller */
4176 : 26781 : otherclauses = NIL;
4177 : : /* current estimate of ndistinct */
4178 : 26781 : ndistinct = 1.0;
4179 [ + + ]: 53572 : while (clauses != NIL)
4180 : : {
4181 : : ListCell *lc;
4182 : 26791 : int relid = -1;
4183 : 26791 : List *varinfos = NIL;
4184 : 26791 : List *origin_rinfos = NIL;
4185 : : double mvndistinct;
4186 : : List *origin_varinfos;
4187 : 26791 : int group_relid = -1;
4188 : 26791 : RelOptInfo *group_rel = NULL;
4189 : : ListCell *lc1,
4190 : : *lc2;
4191 : :
4192 : : /*
4193 : : * Find clauses, referencing the same single base relation and try to
4194 : : * estimate such a group with extended statistics. Create varinfo for
4195 : : * an approved clause, push it to otherclauses, if it can't be
4196 : : * estimated here or ignore to process at the next iteration.
4197 : : */
4198 [ + + + + : 82412 : foreach(lc, clauses)
+ + ]
4199 : : {
4200 : 55621 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
4201 : : Node *expr;
4202 : : Relids relids;
4203 : : GroupVarInfo *varinfo;
4204 : :
4205 : : /*
4206 : : * Find the inner side of the join, which we need to estimate the
4207 : : * number of buckets. Use outer_is_left because the
4208 : : * clause_sides_match_join routine has called on hash clauses.
4209 : : */
4210 : 111242 : relids = rinfo->outer_is_left ?
4211 [ + + ]: 55621 : rinfo->right_relids : rinfo->left_relids;
4212 : 111242 : expr = rinfo->outer_is_left ?
4213 [ + + ]: 55621 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
4214 : :
4215 [ + + ]: 55621 : if (bms_get_singleton_member(relids, &relid) &&
4216 [ + + ]: 54305 : root->simple_rel_array[relid]->statlist != NIL)
4217 : 40 : {
4218 : 50 : bool is_duplicate = false;
4219 : :
4220 : : /*
4221 : : * This inner-side expression references only one relation.
4222 : : * Extended statistics on this clause can exist.
4223 : : */
4224 [ + + ]: 50 : if (group_relid < 0)
4225 : : {
4226 : 25 : RangeTblEntry *rte = root->simple_rte_array[relid];
4227 : :
4228 [ + - - + ]: 25 : if (!rte || (rte->relkind != RELKIND_RELATION &&
4229 [ # # ]: 0 : rte->relkind != RELKIND_MATVIEW &&
4230 [ # # ]: 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
4231 [ # # ]: 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
4232 : : {
4233 : : /* Extended statistics can't exist in principle */
4234 : 0 : otherclauses = lappend(otherclauses, rinfo);
4235 : 0 : clauses = foreach_delete_current(clauses, lc);
4236 : 0 : continue;
4237 : : }
4238 : :
4239 : 25 : group_relid = relid;
4240 : 25 : group_rel = root->simple_rel_array[relid];
4241 : : }
4242 [ - + ]: 25 : else if (group_relid != relid)
4243 : : {
4244 : : /*
4245 : : * Being in the group forming state we don't need other
4246 : : * clauses.
4247 : : */
4248 : 0 : continue;
4249 : : }
4250 : :
4251 : : /*
4252 : : * We're going to add the new clause to the varinfos list. We
4253 : : * might re-use add_unique_group_var(), but we don't do so for
4254 : : * two reasons.
4255 : : *
4256 : : * 1) We must keep the origin_rinfos list ordered exactly the
4257 : : * same way as varinfos.
4258 : : *
4259 : : * 2) add_unique_group_var() is designed for
4260 : : * estimate_num_groups(), where a larger number of groups is
4261 : : * worse. While estimating the number of hash buckets, we
4262 : : * have the opposite: a lesser number of groups is worse.
4263 : : * Therefore, we don't have to remove "known equal" vars: the
4264 : : * removed var may valuably contribute to the multivariate
4265 : : * statistics to grow the number of groups.
4266 : : */
4267 : :
4268 : : /*
4269 : : * Clear nullingrels to correctly match hash keys. See
4270 : : * add_unique_group_var()'s comment for details.
4271 : : */
4272 : 50 : expr = remove_nulling_relids(expr, root->outer_join_rels, NULL);
4273 : :
4274 : : /*
4275 : : * Detect and exclude exact duplicates from the list of hash
4276 : : * keys (like add_unique_group_var does).
4277 : : */
4278 [ + + + + : 70 : foreach(lc1, varinfos)
+ + ]
4279 : : {
4280 : 30 : varinfo = (GroupVarInfo *) lfirst(lc1);
4281 : :
4282 [ + + ]: 30 : if (!equal(expr, varinfo->var))
4283 : 20 : continue;
4284 : :
4285 : 10 : is_duplicate = true;
4286 : 10 : break;
4287 : : }
4288 : :
4289 [ + + ]: 50 : if (is_duplicate)
4290 : : {
4291 : : /*
4292 : : * Skip exact duplicates. Adding them to the otherclauses
4293 : : * list also doesn't make sense.
4294 : : */
4295 : 10 : continue;
4296 : : }
4297 : :
4298 : : /*
4299 : : * Initialize GroupVarInfo. We only use it to call
4300 : : * estimate_multivariate_ndistinct(), which doesn't care about
4301 : : * ndistinct and isdefault fields. Thus, skip these fields.
4302 : : */
4303 : 40 : varinfo = palloc0_object(GroupVarInfo);
4304 : 40 : varinfo->var = expr;
4305 : 40 : varinfo->rel = root->simple_rel_array[relid];
4306 : 40 : varinfos = lappend(varinfos, varinfo);
4307 : :
4308 : : /*
4309 : : * Remember the link to RestrictInfo for the case the clause
4310 : : * is failed to be estimated.
4311 : : */
4312 : 40 : origin_rinfos = lappend(origin_rinfos, rinfo);
4313 : : }
4314 : : else
4315 : : {
4316 : : /* This clause can't be estimated with extended statistics */
4317 : 55571 : otherclauses = lappend(otherclauses, rinfo);
4318 : : }
4319 : :
4320 : 55611 : clauses = foreach_delete_current(clauses, lc);
4321 : : }
4322 : :
4323 [ + + ]: 26791 : if (list_length(varinfos) < 2)
4324 : : {
4325 : : /*
4326 : : * Multivariate statistics doesn't apply to single columns except
4327 : : * for expressions, but it has not been implemented yet.
4328 : : */
4329 : 26781 : otherclauses = list_concat(otherclauses, origin_rinfos);
4330 : 26781 : list_free_deep(varinfos);
4331 : 26781 : list_free(origin_rinfos);
4332 : 26781 : continue;
4333 : : }
4334 : :
4335 : : Assert(group_rel != NULL);
4336 : :
4337 : : /* Employ the extended statistics. */
4338 : 10 : origin_varinfos = varinfos;
4339 : : for (;;)
4340 : 10 : {
4341 : 20 : bool estimated = estimate_multivariate_ndistinct(root,
4342 : : group_rel,
4343 : : &varinfos,
4344 : : &mvndistinct);
4345 : :
4346 [ + + ]: 20 : if (!estimated)
4347 : 10 : break;
4348 : :
4349 : : /*
4350 : : * We've got an estimation. Use ndistinct value in a consistent
4351 : : * way - according to the caller's logic (see
4352 : : * final_cost_hashjoin).
4353 : : */
4354 [ + - ]: 10 : if (ndistinct < mvndistinct)
4355 : 10 : ndistinct = mvndistinct;
4356 : : Assert(ndistinct >= 1.0);
4357 : : }
4358 : :
4359 : : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
4360 : :
4361 : : /* Collect unmatched clauses as otherclauses. */
4362 [ + - + + : 35 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
+ - + + +
+ + - +
+ ]
4363 : : {
4364 : 25 : GroupVarInfo *vinfo = lfirst(lc1);
4365 : :
4366 [ + - ]: 25 : if (!list_member_ptr(varinfos, vinfo))
4367 : : /* Already estimated */
4368 : 25 : continue;
4369 : :
4370 : : /* Can't be estimated here - push to the returning list */
4371 : 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
4372 : : }
4373 : : }
4374 : :
4375 : 26781 : *innerbucketsize = 1.0 / ndistinct;
4376 : 26781 : return otherclauses;
4377 : : }
4378 : :
4379 : : /*
4380 : : * Estimate hash bucket statistics when the specified expression is used
4381 : : * as a hash key for the given number of buckets.
4382 : : *
4383 : : * This attempts to determine two values:
4384 : : *
4385 : : * 1. The frequency of the most common value of the expression (returns
4386 : : * zero into *mcv_freq if we can't get that). This will be frequency
4387 : : * relative to the entire underlying table.
4388 : : *
4389 : : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
4390 : : * divided by total number of tuples to be hashed.
4391 : : *
4392 : : * XXX This is really pretty bogus since we're effectively assuming that the
4393 : : * distribution of hash keys will be the same after applying restriction
4394 : : * clauses as it was in the underlying relation. However, we are not nearly
4395 : : * smart enough to figure out how the restrict clauses might change the
4396 : : * distribution, so this will have to do for now.
4397 : : *
4398 : : * We are passed the number of buckets the executor will use for the given
4399 : : * input relation. If the data were perfectly distributed, with the same
4400 : : * number of tuples going into each available bucket, then the bucketsize
4401 : : * fraction would be 1/nbuckets. But this happy state of affairs will occur
4402 : : * only if (a) there are at least nbuckets distinct data values, and (b)
4403 : : * we have a not-too-skewed data distribution. Otherwise the buckets will
4404 : : * be nonuniformly occupied. If the other relation in the join has a key
4405 : : * distribution similar to this one's, then the most-loaded buckets are
4406 : : * exactly those that will be probed most often. Therefore, the "average"
4407 : : * bucket size for costing purposes should really be taken as something close
4408 : : * to the "worst case" bucket size. We try to estimate this by adjusting the
4409 : : * fraction if there are too few distinct data values, and then clamping to
4410 : : * at least the bucket size implied by the most common value's frequency.
4411 : : *
4412 : : * If no statistics are available, use a default estimate of 0.1. This will
4413 : : * discourage use of a hash rather strongly if the inner relation is large,
4414 : : * which is what we want. We do not want to hash unless we know that the
4415 : : * inner rel is well-dispersed (or the alternatives seem much worse).
4416 : : *
4417 : : * The caller should also check that the mcv_freq is not so large that the
4418 : : * most common value would by itself require an impractically large bucket.
4419 : : * In a hash join, the executor can split buckets if they get too big, but
4420 : : * obviously that doesn't help for a bucket that contains many duplicates of
4421 : : * the same value.
4422 : : */
4423 : : void
4424 : 149999 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
4425 : : Selectivity *mcv_freq,
4426 : : Selectivity *bucketsize_frac)
4427 : : {
4428 : : VariableStatData vardata;
4429 : : double estfract,
4430 : : ndistinct;
4431 : : bool isdefault;
4432 : : AttStatsSlot sslot;
4433 : :
4434 : 149999 : examine_variable(root, hashkey, 0, &vardata);
4435 : :
4436 : : /* Initialize *mcv_freq to "unknown" */
4437 : 149999 : *mcv_freq = 0.0;
4438 : :
4439 : : /* Look up the frequency of the most common value, if available */
4440 [ + + ]: 149999 : if (HeapTupleIsValid(vardata.statsTuple))
4441 : : {
4442 [ + + ]: 97068 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4443 : : STATISTIC_KIND_MCV, InvalidOid,
4444 : : ATTSTATSSLOT_NUMBERS))
4445 : : {
4446 : : /*
4447 : : * The first MCV stat is for the most common value.
4448 : : */
4449 [ + - ]: 57944 : if (sslot.nnumbers > 0)
4450 : 57944 : *mcv_freq = sslot.numbers[0];
4451 : 57944 : free_attstatsslot(&sslot);
4452 : : }
4453 [ + + ]: 39124 : else if (get_attstatsslot(&sslot, vardata.statsTuple,
4454 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
4455 : : 0))
4456 : : {
4457 : : /*
4458 : : * If there are no recorded MCVs, but we do have a histogram, then
4459 : : * assume that ANALYZE determined that the column is unique.
4460 : : */
4461 [ + - + + ]: 37692 : if (vardata.rel && vardata.rel->tuples > 0)
4462 : 37677 : *mcv_freq = 1.0 / vardata.rel->tuples;
4463 : : }
4464 : : }
4465 : :
4466 : : /* Get number of distinct values */
4467 : 149999 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4468 : :
4469 : : /*
4470 : : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4471 : : * mcv_freq is known to be even higher than that, use it instead.
4472 : : */
4473 [ + + ]: 149999 : if (isdefault)
4474 : : {
4475 [ + - ]: 24787 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4476 [ + + ]: 24787 : ReleaseVariableStats(vardata);
4477 : 24787 : return;
4478 : : }
4479 : :
4480 : : /*
4481 : : * Adjust ndistinct to account for restriction clauses. Observe we are
4482 : : * assuming that the data distribution is affected uniformly by the
4483 : : * restriction clauses!
4484 : : *
4485 : : * XXX Possibly better way, but much more expensive: multiply by
4486 : : * selectivity of rel's restriction clauses that mention the target Var.
4487 : : */
4488 [ + - + + ]: 125212 : if (vardata.rel && vardata.rel->tuples > 0)
4489 : : {
4490 : 125181 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
4491 : 125181 : ndistinct = clamp_row_est(ndistinct);
4492 : : }
4493 : :
4494 : : /*
4495 : : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
4496 : : * number of buckets is less than the expected number of distinct values;
4497 : : * otherwise it is 1/ndistinct.
4498 : : */
4499 [ + + ]: 125212 : if (ndistinct > nbuckets)
4500 : 94 : estfract = 1.0 / nbuckets;
4501 : : else
4502 : 125118 : estfract = 1.0 / ndistinct;
4503 : :
4504 : : /*
4505 : : * Clamp the bucketsize fraction to be not less than the MCV frequency,
4506 : : * since whichever bucket the MCV values end up in will have at least that
4507 : : * size. This has no effect if *mcv_freq is still zero.
4508 : : */
4509 [ + + ]: 125212 : estfract = Max(estfract, *mcv_freq);
4510 : :
4511 : 125212 : *bucketsize_frac = (Selectivity) estfract;
4512 : :
4513 [ + + ]: 125212 : ReleaseVariableStats(vardata);
4514 : : }
4515 : :
4516 : : /*
4517 : : * estimate_hashagg_tablesize
4518 : : * estimate the number of bytes that a hash aggregate hashtable will
4519 : : * require based on the agg_costs, path width and number of groups.
4520 : : *
4521 : : * We return the result as "double" to forestall any possible overflow
4522 : : * problem in the multiplication by dNumGroups.
4523 : : *
4524 : : * XXX this may be over-estimating the size now that hashagg knows to omit
4525 : : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4526 : : * grouping columns not in the hashed set are counted here even though hashagg
4527 : : * won't store them. Is this a problem?
4528 : : */
4529 : : double
4530 : 2424 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4531 : : const AggClauseCosts *agg_costs, double dNumGroups)
4532 : : {
4533 : : Size hashentrysize;
4534 : :
4535 : 2424 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4536 : 2424 : path->pathtarget->width,
4537 : 2424 : agg_costs->transitionSpace);
4538 : :
4539 : : /*
4540 : : * Note that this disregards the effect of fill-factor and growth policy
4541 : : * of the hash table. That's probably ok, given that the default
4542 : : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4543 : : * "double-in-size" growth policies here.
4544 : : */
4545 : 2424 : return hashentrysize * dNumGroups;
4546 : : }
4547 : :
4548 : :
4549 : : /*-------------------------------------------------------------------------
4550 : : *
4551 : : * Support routines
4552 : : *
4553 : : *-------------------------------------------------------------------------
4554 : : */
4555 : :
4556 : : /*
4557 : : * Find the best matching ndistinct extended statistics for the given list of
4558 : : * GroupVarInfos.
4559 : : *
4560 : : * Callers must ensure that the given GroupVarInfos all belong to 'rel' and
4561 : : * the GroupVarInfos list does not contain any duplicate Vars or expressions.
4562 : : *
4563 : : * When statistics are found that match > 1 of the given GroupVarInfo, the
4564 : : * *ndistinct parameter is set according to the ndistinct estimate and a new
4565 : : * list is built with the matching GroupVarInfos removed, which is output via
4566 : : * the *varinfos parameter before returning true. When no matching stats are
4567 : : * found, false is returned and the *varinfos and *ndistinct parameters are
4568 : : * left untouched.
4569 : : */
4570 : : static bool
4571 : 237648 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4572 : : List **varinfos, double *ndistinct)
4573 : : {
4574 : : ListCell *lc;
4575 : : int nmatches_vars;
4576 : : int nmatches_exprs;
4577 : 237648 : Oid statOid = InvalidOid;
4578 : : MVNDistinct *stats;
4579 : 237648 : StatisticExtInfo *matched_info = NULL;
4580 [ + - ]: 237648 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4581 : :
4582 : : /* bail out immediately if the table has no extended statistics */
4583 [ + + ]: 237648 : if (!rel->statlist)
4584 : 237177 : return false;
4585 : :
4586 : : /* look for the ndistinct statistics object matching the most vars */
4587 : 471 : nmatches_vars = 0; /* we require at least two matches */
4588 : 471 : nmatches_exprs = 0;
4589 [ + - + + : 1875 : foreach(lc, rel->statlist)
+ + ]
4590 : : {
4591 : : ListCell *lc2;
4592 : 1404 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
4593 : 1404 : int nshared_vars = 0;
4594 : 1404 : int nshared_exprs = 0;
4595 : :
4596 : : /* skip statistics of other kinds */
4597 [ + + ]: 1404 : if (info->kind != STATS_EXT_NDISTINCT)
4598 : 663 : continue;
4599 : :
4600 : : /* skip statistics with mismatching stxdinherit value */
4601 [ + + ]: 741 : if (info->inherit != rte->inh)
4602 : 25 : continue;
4603 : :
4604 : : /*
4605 : : * Determine how many expressions (and variables in non-matched
4606 : : * expressions) match. We'll then use these numbers to pick the
4607 : : * statistics object that best matches the clauses.
4608 : : */
4609 [ + + + + : 2267 : foreach(lc2, *varinfos)
+ + ]
4610 : : {
4611 : : ListCell *lc3;
4612 : 1551 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4613 : : AttrNumber attnum;
4614 : :
4615 : : Assert(varinfo->rel == rel);
4616 : :
4617 : : /* simple Var, search in statistics keys directly */
4618 [ + + ]: 1551 : if (IsA(varinfo->var, Var))
4619 : : {
4620 : 1246 : attnum = ((Var *) varinfo->var)->varattno;
4621 : :
4622 : : /*
4623 : : * Ignore system attributes - we don't support statistics on
4624 : : * them, so can't match them (and it'd fail as the values are
4625 : : * negative).
4626 : : */
4627 [ + + ]: 1246 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4628 : 10 : continue;
4629 : :
4630 [ + + ]: 1236 : if (bms_is_member(attnum, info->keys))
4631 : 730 : nshared_vars++;
4632 : :
4633 : 1236 : continue;
4634 : : }
4635 : :
4636 : : /* expression - see if it's in the statistics object */
4637 [ + + + + : 550 : foreach(lc3, info->exprs)
+ + ]
4638 : : {
4639 : 440 : Node *expr = (Node *) lfirst(lc3);
4640 : :
4641 [ + + ]: 440 : if (equal(varinfo->var, expr))
4642 : : {
4643 : 195 : nshared_exprs++;
4644 : 195 : break;
4645 : : }
4646 : : }
4647 : : }
4648 : :
4649 : : /*
4650 : : * The ndistinct extended statistics contain estimates for a minimum
4651 : : * of pairs of columns which the statistics are defined on and
4652 : : * certainly not single columns. Here we skip unless we managed to
4653 : : * match to at least two columns.
4654 : : */
4655 [ + + ]: 716 : if (nshared_vars + nshared_exprs < 2)
4656 : 331 : continue;
4657 : :
4658 : : /*
4659 : : * Check if these statistics are a better match than the previous best
4660 : : * match and if so, take note of the StatisticExtInfo.
4661 : : *
4662 : : * The statslist is sorted by statOid, so the StatisticExtInfo we
4663 : : * select as the best match is deterministic even when multiple sets
4664 : : * of statistics match equally as well.
4665 : : */
4666 [ + + + - ]: 385 : if ((nshared_exprs > nmatches_exprs) ||
4667 [ + + ]: 295 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4668 : : {
4669 : 365 : statOid = info->statOid;
4670 : 365 : nmatches_vars = nshared_vars;
4671 : 365 : nmatches_exprs = nshared_exprs;
4672 : 365 : matched_info = info;
4673 : : }
4674 : : }
4675 : :
4676 : : /* No match? */
4677 [ + + ]: 471 : if (statOid == InvalidOid)
4678 : 116 : return false;
4679 : :
4680 : : Assert(nmatches_vars + nmatches_exprs > 1);
4681 : :
4682 : 355 : stats = statext_ndistinct_load(statOid, rte->inh);
4683 : :
4684 : : /*
4685 : : * If we have a match, search it for the specific item that matches (there
4686 : : * must be one), and construct the output values.
4687 : : */
4688 [ + - ]: 355 : if (stats)
4689 : : {
4690 : : int i;
4691 : 355 : List *newlist = NIL;
4692 : 355 : MVNDistinctItem *item = NULL;
4693 : : ListCell *lc2;
4694 : 355 : Bitmapset *matched = NULL;
4695 : : AttrNumber attnum_offset;
4696 : :
4697 : : /*
4698 : : * How much we need to offset the attnums? If there are no
4699 : : * expressions, no offset is needed. Otherwise offset enough to move
4700 : : * the lowest one (which is equal to number of expressions) to 1.
4701 : : */
4702 [ + + ]: 355 : if (matched_info->exprs)
4703 : 125 : attnum_offset = (list_length(matched_info->exprs) + 1);
4704 : : else
4705 : 230 : attnum_offset = 0;
4706 : :
4707 : : /* see what actually matched */
4708 [ + - + + : 1240 : foreach(lc2, *varinfos)
+ + ]
4709 : : {
4710 : : ListCell *lc3;
4711 : : int idx;
4712 : 885 : bool found = false;
4713 : :
4714 : 885 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4715 : :
4716 : : /*
4717 : : * Process a simple Var expression, by matching it to keys
4718 : : * directly. If there's a matching expression, we'll try matching
4719 : : * it later.
4720 : : */
4721 [ + + ]: 885 : if (IsA(varinfo->var, Var))
4722 : : {
4723 : 730 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4724 : :
4725 : : /*
4726 : : * Ignore expressions on system attributes. Can't rely on the
4727 : : * bms check for negative values.
4728 : : */
4729 [ + + ]: 730 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4730 : 5 : continue;
4731 : :
4732 : : /* Is the variable covered by the statistics object? */
4733 [ + + ]: 725 : if (!bms_is_member(attnum, matched_info->keys))
4734 : 100 : continue;
4735 : :
4736 : 625 : attnum = attnum + attnum_offset;
4737 : :
4738 : : /* ensure sufficient offset */
4739 : : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4740 : :
4741 : 625 : matched = bms_add_member(matched, attnum);
4742 : :
4743 : 625 : found = true;
4744 : : }
4745 : :
4746 : : /*
4747 : : * XXX Maybe we should allow searching the expressions even if we
4748 : : * found an attribute matching the expression? That would handle
4749 : : * trivial expressions like "(a)" but it seems fairly useless.
4750 : : */
4751 [ + + ]: 780 : if (found)
4752 : 625 : continue;
4753 : :
4754 : : /* expression - see if it's in the statistics object */
4755 : 155 : idx = 0;
4756 [ + + + + : 255 : foreach(lc3, matched_info->exprs)
+ + ]
4757 : : {
4758 : 230 : Node *expr = (Node *) lfirst(lc3);
4759 : :
4760 [ + + ]: 230 : if (equal(varinfo->var, expr))
4761 : : {
4762 : 130 : AttrNumber attnum = -(idx + 1);
4763 : :
4764 : 130 : attnum = attnum + attnum_offset;
4765 : :
4766 : : /* ensure sufficient offset */
4767 : : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4768 : :
4769 : 130 : matched = bms_add_member(matched, attnum);
4770 : :
4771 : : /* there should be just one matching expression */
4772 : 130 : break;
4773 : : }
4774 : :
4775 : 100 : idx++;
4776 : : }
4777 : : }
4778 : :
4779 : : /* Find the specific item that exactly matches the combination */
4780 [ + - ]: 720 : for (i = 0; i < stats->nitems; i++)
4781 : : {
4782 : : int j;
4783 : 720 : MVNDistinctItem *tmpitem = &stats->items[i];
4784 : :
4785 [ + + ]: 720 : if (tmpitem->nattributes != bms_num_members(matched))
4786 : 135 : continue;
4787 : :
4788 : : /* assume it's the right item */
4789 : 585 : item = tmpitem;
4790 : :
4791 : : /* check that all item attributes/expressions fit the match */
4792 [ + + ]: 1410 : for (j = 0; j < tmpitem->nattributes; j++)
4793 : : {
4794 : 1055 : AttrNumber attnum = tmpitem->attributes[j];
4795 : :
4796 : : /*
4797 : : * Thanks to how we constructed the matched bitmap above, we
4798 : : * can just offset all attnums the same way.
4799 : : */
4800 : 1055 : attnum = attnum + attnum_offset;
4801 : :
4802 [ + + ]: 1055 : if (!bms_is_member(attnum, matched))
4803 : : {
4804 : : /* nah, it's not this item */
4805 : 230 : item = NULL;
4806 : 230 : break;
4807 : : }
4808 : : }
4809 : :
4810 : : /*
4811 : : * If the item has all the matched attributes, we know it's the
4812 : : * right one - there can't be a better one. matching more.
4813 : : */
4814 [ + + ]: 585 : if (item)
4815 : 355 : break;
4816 : : }
4817 : :
4818 : : /*
4819 : : * Make sure we found an item. There has to be one, because ndistinct
4820 : : * statistics includes all combinations of attributes.
4821 : : */
4822 [ - + ]: 355 : if (!item)
4823 [ # # ]: 0 : elog(ERROR, "corrupt MVNDistinct entry");
4824 : :
4825 : : /* Form the output varinfo list, keeping only unmatched ones */
4826 [ + - + + : 1240 : foreach(lc, *varinfos)
+ + ]
4827 : : {
4828 : 885 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4829 : : ListCell *lc3;
4830 : 885 : bool found = false;
4831 : :
4832 : : /*
4833 : : * Let's look at plain variables first, because it's the most
4834 : : * common case and the check is quite cheap. We can simply get the
4835 : : * attnum and check (with an offset) matched bitmap.
4836 : : */
4837 [ + + ]: 885 : if (IsA(varinfo->var, Var))
4838 : 725 : {
4839 : 730 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4840 : :
4841 : : /*
4842 : : * If it's a system attribute, we're done. We don't support
4843 : : * extended statistics on system attributes, so it's clearly
4844 : : * not matched. Just keep the expression and continue.
4845 : : */
4846 [ + + ]: 730 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4847 : : {
4848 : 5 : newlist = lappend(newlist, varinfo);
4849 : 5 : continue;
4850 : : }
4851 : :
4852 : : /* apply the same offset as above */
4853 : 725 : attnum += attnum_offset;
4854 : :
4855 : : /* if it's not matched, keep the varinfo */
4856 [ + + ]: 725 : if (!bms_is_member(attnum, matched))
4857 : 100 : newlist = lappend(newlist, varinfo);
4858 : :
4859 : : /* The rest of the loop deals with complex expressions. */
4860 : 725 : continue;
4861 : : }
4862 : :
4863 : : /*
4864 : : * Process complex expressions, not just simple Vars.
4865 : : *
4866 : : * First, we search for an exact match of an expression. If we
4867 : : * find one, we can just discard the whole GroupVarInfo, with all
4868 : : * the variables we extracted from it.
4869 : : *
4870 : : * Otherwise we inspect the individual vars, and try matching it
4871 : : * to variables in the item.
4872 : : */
4873 [ + + + + : 255 : foreach(lc3, matched_info->exprs)
+ + ]
4874 : : {
4875 : 230 : Node *expr = (Node *) lfirst(lc3);
4876 : :
4877 [ + + ]: 230 : if (equal(varinfo->var, expr))
4878 : : {
4879 : 130 : found = true;
4880 : 130 : break;
4881 : : }
4882 : : }
4883 : :
4884 : : /* found exact match, skip */
4885 [ + + ]: 155 : if (found)
4886 : 130 : continue;
4887 : :
4888 : 25 : newlist = lappend(newlist, varinfo);
4889 : : }
4890 : :
4891 : 355 : *varinfos = newlist;
4892 : 355 : *ndistinct = item->ndistinct;
4893 : 355 : return true;
4894 : : }
4895 : :
4896 : 0 : return false;
4897 : : }
4898 : :
4899 : : /*
4900 : : * convert_to_scalar
4901 : : * Convert non-NULL values of the indicated types to the comparison
4902 : : * scale needed by scalarineqsel().
4903 : : * Returns "true" if successful.
4904 : : *
4905 : : * XXX this routine is a hack: ideally we should look up the conversion
4906 : : * subroutines in pg_type.
4907 : : *
4908 : : * All numeric datatypes are simply converted to their equivalent
4909 : : * "double" values. (NUMERIC values that are outside the range of "double"
4910 : : * are clamped to +/- HUGE_VAL.)
4911 : : *
4912 : : * String datatypes are converted by convert_string_to_scalar(),
4913 : : * which is explained below. The reason why this routine deals with
4914 : : * three values at a time, not just one, is that we need it for strings.
4915 : : *
4916 : : * The bytea datatype is just enough different from strings that it has
4917 : : * to be treated separately.
4918 : : *
4919 : : * The several datatypes representing absolute times are all converted
4920 : : * to Timestamp, which is actually an int64, and then we promote that to
4921 : : * a double. Note this will give correct results even for the "special"
4922 : : * values of Timestamp, since those are chosen to compare correctly;
4923 : : * see timestamp_cmp.
4924 : : *
4925 : : * The several datatypes representing relative times (intervals) are all
4926 : : * converted to measurements expressed in seconds.
4927 : : */
4928 : : static bool
4929 : 53108 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4930 : : Datum lobound, Datum hibound, Oid boundstypid,
4931 : : double *scaledlobound, double *scaledhibound)
4932 : : {
4933 : 53108 : bool failure = false;
4934 : :
4935 : : /*
4936 : : * Both the valuetypid and the boundstypid should exactly match the
4937 : : * declared input type(s) of the operator we are invoked for. However,
4938 : : * extensions might try to use scalarineqsel as estimator for operators
4939 : : * with input type(s) we don't handle here; in such cases, we want to
4940 : : * return false, not fail. In any case, we mustn't assume that valuetypid
4941 : : * and boundstypid are identical.
4942 : : *
4943 : : * XXX The histogram we are interpolating between points of could belong
4944 : : * to a column that's only binary-compatible with the declared type. In
4945 : : * essence we are assuming that the semantics of binary-compatible types
4946 : : * are enough alike that we can use a histogram generated with one type's
4947 : : * operators to estimate selectivity for the other's. This is outright
4948 : : * wrong in some cases --- in particular signed versus unsigned
4949 : : * interpretation could trip us up. But it's useful enough in the
4950 : : * majority of cases that we do it anyway. Should think about more
4951 : : * rigorous ways to do it.
4952 : : */
4953 [ + + - - : 53108 : switch (valuetypid)
- - ]
4954 : : {
4955 : : /*
4956 : : * Built-in numeric types
4957 : : */
4958 : 48306 : case BOOLOID:
4959 : : case INT2OID:
4960 : : case INT4OID:
4961 : : case INT8OID:
4962 : : case FLOAT4OID:
4963 : : case FLOAT8OID:
4964 : : case NUMERICOID:
4965 : : case OIDOID:
4966 : : case REGPROCOID:
4967 : : case REGPROCEDUREOID:
4968 : : case REGOPEROID:
4969 : : case REGOPERATOROID:
4970 : : case REGCLASSOID:
4971 : : case REGTYPEOID:
4972 : : case REGCOLLATIONOID:
4973 : : case REGCONFIGOID:
4974 : : case REGDICTIONARYOID:
4975 : : case REGROLEOID:
4976 : : case REGNAMESPACEOID:
4977 : : case REGDATABASEOID:
4978 : 48306 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4979 : : &failure);
4980 : 48306 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4981 : : &failure);
4982 : 48306 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4983 : : &failure);
4984 : 48306 : return !failure;
4985 : :
4986 : : /*
4987 : : * Built-in string types
4988 : : */
4989 : 4802 : case CHAROID:
4990 : : case BPCHAROID:
4991 : : case VARCHAROID:
4992 : : case TEXTOID:
4993 : : case NAMEOID:
4994 : : {
4995 : 4802 : char *valstr = convert_string_datum(value, valuetypid,
4996 : : collid, &failure);
4997 : 4802 : char *lostr = convert_string_datum(lobound, boundstypid,
4998 : : collid, &failure);
4999 : 4802 : char *histr = convert_string_datum(hibound, boundstypid,
5000 : : collid, &failure);
5001 : :
5002 : : /*
5003 : : * Bail out if any of the values is not of string type. We
5004 : : * might leak converted strings for the other value(s), but
5005 : : * that's not worth troubling over.
5006 : : */
5007 [ - + ]: 4802 : if (failure)
5008 : 0 : return false;
5009 : :
5010 : 4802 : convert_string_to_scalar(valstr, scaledvalue,
5011 : : lostr, scaledlobound,
5012 : : histr, scaledhibound);
5013 : 4802 : pfree(valstr);
5014 : 4802 : pfree(lostr);
5015 : 4802 : pfree(histr);
5016 : 4802 : return true;
5017 : : }
5018 : :
5019 : : /*
5020 : : * Built-in bytea type
5021 : : */
5022 : 0 : case BYTEAOID:
5023 : : {
5024 : : /* We only support bytea vs bytea comparison */
5025 [ # # ]: 0 : if (boundstypid != BYTEAOID)
5026 : 0 : return false;
5027 : 0 : convert_bytea_to_scalar(value, scaledvalue,
5028 : : lobound, scaledlobound,
5029 : : hibound, scaledhibound);
5030 : 0 : return true;
5031 : : }
5032 : :
5033 : : /*
5034 : : * Built-in time types
5035 : : */
5036 : 0 : case TIMESTAMPOID:
5037 : : case TIMESTAMPTZOID:
5038 : : case DATEOID:
5039 : : case INTERVALOID:
5040 : : case TIMEOID:
5041 : : case TIMETZOID:
5042 : 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
5043 : : &failure);
5044 : 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
5045 : : &failure);
5046 : 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
5047 : : &failure);
5048 : 0 : return !failure;
5049 : :
5050 : : /*
5051 : : * Built-in network types
5052 : : */
5053 : 0 : case INETOID:
5054 : : case CIDROID:
5055 : : case MACADDROID:
5056 : : case MACADDR8OID:
5057 : 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
5058 : : &failure);
5059 : 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
5060 : : &failure);
5061 : 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
5062 : : &failure);
5063 : 0 : return !failure;
5064 : : }
5065 : : /* Don't know how to convert */
5066 : 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
5067 : 0 : return false;
5068 : : }
5069 : :
5070 : : /*
5071 : : * Do convert_to_scalar()'s work for any numeric data type.
5072 : : *
5073 : : * On failure (e.g., unsupported typid), set *failure to true;
5074 : : * otherwise, that variable is not changed.
5075 : : */
5076 : : static double
5077 : 144918 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
5078 : : {
5079 [ - + + - : 144918 : switch (typid)
- + - +
- ]
5080 : : {
5081 : 0 : case BOOLOID:
5082 : 0 : return (double) DatumGetBool(value);
5083 : 10 : case INT2OID:
5084 : 10 : return (double) DatumGetInt16(value);
5085 : 22169 : case INT4OID:
5086 : 22169 : return (double) DatumGetInt32(value);
5087 : 0 : case INT8OID:
5088 : 0 : return (double) DatumGetInt64(value);
5089 : 0 : case FLOAT4OID:
5090 : 0 : return (double) DatumGetFloat4(value);
5091 : 45 : case FLOAT8OID:
5092 : 45 : return (double) DatumGetFloat8(value);
5093 : 0 : case NUMERICOID:
5094 : : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
5095 : 0 : return (double)
5096 : 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
5097 : : value));
5098 : 122694 : case OIDOID:
5099 : : case REGPROCOID:
5100 : : case REGPROCEDUREOID:
5101 : : case REGOPEROID:
5102 : : case REGOPERATOROID:
5103 : : case REGCLASSOID:
5104 : : case REGTYPEOID:
5105 : : case REGCOLLATIONOID:
5106 : : case REGCONFIGOID:
5107 : : case REGDICTIONARYOID:
5108 : : case REGROLEOID:
5109 : : case REGNAMESPACEOID:
5110 : : case REGDATABASEOID:
5111 : : /* we can treat OIDs as integers... */
5112 : 122694 : return (double) DatumGetObjectId(value);
5113 : : }
5114 : :
5115 : 0 : *failure = true;
5116 : 0 : return 0;
5117 : : }
5118 : :
5119 : : /*
5120 : : * Do convert_to_scalar()'s work for any character-string data type.
5121 : : *
5122 : : * String datatypes are converted to a scale that ranges from 0 to 1,
5123 : : * where we visualize the bytes of the string as fractional digits.
5124 : : *
5125 : : * We do not want the base to be 256, however, since that tends to
5126 : : * generate inflated selectivity estimates; few databases will have
5127 : : * occurrences of all 256 possible byte values at each position.
5128 : : * Instead, use the smallest and largest byte values seen in the bounds
5129 : : * as the estimated range for each byte, after some fudging to deal with
5130 : : * the fact that we probably aren't going to see the full range that way.
5131 : : *
5132 : : * An additional refinement is that we discard any common prefix of the
5133 : : * three strings before computing the scaled values. This allows us to
5134 : : * "zoom in" when we encounter a narrow data range. An example is a phone
5135 : : * number database where all the values begin with the same area code.
5136 : : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
5137 : : * so this is more likely to happen than you might think.)
5138 : : */
5139 : : static void
5140 : 4802 : convert_string_to_scalar(char *value,
5141 : : double *scaledvalue,
5142 : : char *lobound,
5143 : : double *scaledlobound,
5144 : : char *hibound,
5145 : : double *scaledhibound)
5146 : : {
5147 : : int rangelo,
5148 : : rangehi;
5149 : : char *sptr;
5150 : :
5151 : 4802 : rangelo = rangehi = (unsigned char) hibound[0];
5152 [ + + ]: 69452 : for (sptr = lobound; *sptr; sptr++)
5153 : : {
5154 [ + + ]: 64650 : if (rangelo > (unsigned char) *sptr)
5155 : 11850 : rangelo = (unsigned char) *sptr;
5156 [ + + ]: 64650 : if (rangehi < (unsigned char) *sptr)
5157 : 6064 : rangehi = (unsigned char) *sptr;
5158 : : }
5159 [ + + ]: 55856 : for (sptr = hibound; *sptr; sptr++)
5160 : : {
5161 [ + + ]: 51054 : if (rangelo > (unsigned char) *sptr)
5162 : 499 : rangelo = (unsigned char) *sptr;
5163 [ + + ]: 51054 : if (rangehi < (unsigned char) *sptr)
5164 : 1851 : rangehi = (unsigned char) *sptr;
5165 : : }
5166 : : /* If range includes any upper-case ASCII chars, make it include all */
5167 [ + + + + ]: 4802 : if (rangelo <= 'Z' && rangehi >= 'A')
5168 : : {
5169 [ + + ]: 976 : if (rangelo > 'A')
5170 : 190 : rangelo = 'A';
5171 [ + + ]: 976 : if (rangehi < 'Z')
5172 : 430 : rangehi = 'Z';
5173 : : }
5174 : : /* Ditto lower-case */
5175 [ + - + + ]: 4802 : if (rangelo <= 'z' && rangehi >= 'a')
5176 : : {
5177 [ + + ]: 4357 : if (rangelo > 'a')
5178 : 20 : rangelo = 'a';
5179 [ + + ]: 4357 : if (rangehi < 'z')
5180 : 4299 : rangehi = 'z';
5181 : : }
5182 : : /* Ditto digits */
5183 [ + + + - ]: 4802 : if (rangelo <= '9' && rangehi >= '0')
5184 : : {
5185 [ + + ]: 376 : if (rangelo > '0')
5186 : 303 : rangelo = '0';
5187 [ + + ]: 376 : if (rangehi < '9')
5188 : 9 : rangehi = '9';
5189 : : }
5190 : :
5191 : : /*
5192 : : * If range includes less than 10 chars, assume we have not got enough
5193 : : * data, and make it include regular ASCII set.
5194 : : */
5195 [ - + ]: 4802 : if (rangehi - rangelo < 9)
5196 : : {
5197 : 0 : rangelo = ' ';
5198 : 0 : rangehi = 127;
5199 : : }
5200 : :
5201 : : /*
5202 : : * Now strip any common prefix of the three strings.
5203 : : */
5204 [ + - ]: 9239 : while (*lobound)
5205 : : {
5206 [ + + + - ]: 9239 : if (*lobound != *hibound || *lobound != *value)
5207 : : break;
5208 : 4437 : lobound++, hibound++, value++;
5209 : : }
5210 : :
5211 : : /*
5212 : : * Now we can do the conversions.
5213 : : */
5214 : 4802 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
5215 : 4802 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
5216 : 4802 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
5217 : 4802 : }
5218 : :
5219 : : static double
5220 : 14406 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
5221 : : {
5222 : 14406 : int slen = strlen(value);
5223 : : double num,
5224 : : denom,
5225 : : base;
5226 : :
5227 [ - + ]: 14406 : if (slen <= 0)
5228 : 0 : return 0.0; /* empty string has scalar value 0 */
5229 : :
5230 : : /*
5231 : : * There seems little point in considering more than a dozen bytes from
5232 : : * the string. Since base is at least 10, that will give us nominal
5233 : : * resolution of at least 12 decimal digits, which is surely far more
5234 : : * precision than this estimation technique has got anyway (especially in
5235 : : * non-C locales). Also, even with the maximum possible base of 256, this
5236 : : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
5237 : : * overflow on any known machine.
5238 : : */
5239 [ + + ]: 14406 : if (slen > 12)
5240 : 3969 : slen = 12;
5241 : :
5242 : : /* Convert initial characters to fraction */
5243 : 14406 : base = rangehi - rangelo + 1;
5244 : 14406 : num = 0.0;
5245 : 14406 : denom = base;
5246 [ + + ]: 124657 : while (slen-- > 0)
5247 : : {
5248 : 110251 : int ch = (unsigned char) *value++;
5249 : :
5250 [ + + ]: 110251 : if (ch < rangelo)
5251 : 180 : ch = rangelo - 1;
5252 [ - + ]: 110071 : else if (ch > rangehi)
5253 : 0 : ch = rangehi + 1;
5254 : 110251 : num += ((double) (ch - rangelo)) / denom;
5255 : 110251 : denom *= base;
5256 : : }
5257 : :
5258 : 14406 : return num;
5259 : : }
5260 : :
5261 : : /*
5262 : : * Convert a string-type Datum into a palloc'd, null-terminated string.
5263 : : *
5264 : : * On failure (e.g., unsupported typid), set *failure to true;
5265 : : * otherwise, that variable is not changed. (We'll return NULL on failure.)
5266 : : *
5267 : : * When using a non-C locale, we must pass the string through pg_strxfrm()
5268 : : * before continuing, so as to generate correct locale-specific results.
5269 : : */
5270 : : static char *
5271 : 14406 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
5272 : : {
5273 : : char *val;
5274 : : pg_locale_t mylocale;
5275 : :
5276 [ + + + - ]: 14406 : switch (typid)
5277 : : {
5278 : 15 : case CHAROID:
5279 : 15 : val = (char *) palloc(2);
5280 : 15 : val[0] = DatumGetChar(value);
5281 : 15 : val[1] = '\0';
5282 : 15 : break;
5283 : 4328 : case BPCHAROID:
5284 : : case VARCHAROID:
5285 : : case TEXTOID:
5286 : 4328 : val = TextDatumGetCString(value);
5287 : 4328 : break;
5288 : 10063 : case NAMEOID:
5289 : : {
5290 : 10063 : NameData *nm = (NameData *) DatumGetPointer(value);
5291 : :
5292 : 10063 : val = pstrdup(NameStr(*nm));
5293 : 10063 : break;
5294 : : }
5295 : 0 : default:
5296 : 0 : *failure = true;
5297 : 0 : return NULL;
5298 : : }
5299 : :
5300 : : /*
5301 : : * If we don't have a collation, act as though it's "C". This would
5302 : : * normally happen only for the "char" type, but perhaps there are other
5303 : : * cases.
5304 : : */
5305 [ + + ]: 14406 : if (!OidIsValid(collid))
5306 : 15 : return val;
5307 : :
5308 : 14391 : mylocale = pg_newlocale_from_collation(collid);
5309 : :
5310 [ + + ]: 14391 : if (!mylocale->collate_is_c)
5311 : : {
5312 : : char *xfrmstr;
5313 : : size_t xfrmlen;
5314 : : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
5315 : :
5316 : : /*
5317 : : * XXX: We could guess at a suitable output buffer size and only call
5318 : : * pg_strxfrm() twice if our guess is too small.
5319 : : *
5320 : : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
5321 : : * bogus data or set an error. This is not really a problem unless it
5322 : : * crashes since it will only give an estimation error and nothing
5323 : : * fatal.
5324 : : *
5325 : : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
5326 : : * some cases, libc strxfrm() may return the wrong results, but that
5327 : : * will only lead to an estimation error.
5328 : : */
5329 : 66 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
5330 : : #ifdef WIN32
5331 : :
5332 : : /*
5333 : : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
5334 : : * of trying to allocate this much memory (and fail), just return the
5335 : : * original string unmodified as if we were in the C locale.
5336 : : */
5337 : : if (xfrmlen == INT_MAX)
5338 : : return val;
5339 : : #endif
5340 : 66 : xfrmstr = (char *) palloc(xfrmlen + 1);
5341 : 66 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
5342 : :
5343 : : /*
5344 : : * Some systems (e.g., glibc) can return a smaller value from the
5345 : : * second call than the first; thus the Assert must be <= not ==.
5346 : : */
5347 : : Assert(xfrmlen2 <= xfrmlen);
5348 : 66 : pfree(val);
5349 : 66 : val = xfrmstr;
5350 : : }
5351 : :
5352 : 14391 : return val;
5353 : : }
5354 : :
5355 : : /*
5356 : : * Do convert_to_scalar()'s work for any bytea data type.
5357 : : *
5358 : : * Very similar to convert_string_to_scalar except we can't assume
5359 : : * null-termination and therefore pass explicit lengths around.
5360 : : *
5361 : : * Also, assumptions about likely "normal" ranges of characters have been
5362 : : * removed - a data range of 0..255 is always used, for now. (Perhaps
5363 : : * someday we will add information about actual byte data range to
5364 : : * pg_statistic.)
5365 : : */
5366 : : static void
5367 : 0 : convert_bytea_to_scalar(Datum value,
5368 : : double *scaledvalue,
5369 : : Datum lobound,
5370 : : double *scaledlobound,
5371 : : Datum hibound,
5372 : : double *scaledhibound)
5373 : : {
5374 : 0 : bytea *valuep = DatumGetByteaPP(value);
5375 : 0 : bytea *loboundp = DatumGetByteaPP(lobound);
5376 : 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
5377 : : int rangelo,
5378 : : rangehi,
5379 : 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
5380 : 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
5381 : 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
5382 : : i,
5383 : : minlen;
5384 : 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
5385 : 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
5386 : 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
5387 : :
5388 : : /*
5389 : : * Assume bytea data is uniformly distributed across all byte values.
5390 : : */
5391 : 0 : rangelo = 0;
5392 : 0 : rangehi = 255;
5393 : :
5394 : : /*
5395 : : * Now strip any common prefix of the three strings.
5396 : : */
5397 : 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
5398 [ # # ]: 0 : for (i = 0; i < minlen; i++)
5399 : : {
5400 [ # # # # ]: 0 : if (*lostr != *histr || *lostr != *valstr)
5401 : : break;
5402 : 0 : lostr++, histr++, valstr++;
5403 : 0 : loboundlen--, hiboundlen--, valuelen--;
5404 : : }
5405 : :
5406 : : /*
5407 : : * Now we can do the conversions.
5408 : : */
5409 : 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
5410 : 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
5411 : 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
5412 : 0 : }
5413 : :
5414 : : static double
5415 : 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
5416 : : int rangelo, int rangehi)
5417 : : {
5418 : : double num,
5419 : : denom,
5420 : : base;
5421 : :
5422 [ # # ]: 0 : if (valuelen <= 0)
5423 : 0 : return 0.0; /* empty string has scalar value 0 */
5424 : :
5425 : : /*
5426 : : * Since base is 256, need not consider more than about 10 chars (even
5427 : : * this many seems like overkill)
5428 : : */
5429 [ # # ]: 0 : if (valuelen > 10)
5430 : 0 : valuelen = 10;
5431 : :
5432 : : /* Convert initial characters to fraction */
5433 : 0 : base = rangehi - rangelo + 1;
5434 : 0 : num = 0.0;
5435 : 0 : denom = base;
5436 [ # # ]: 0 : while (valuelen-- > 0)
5437 : : {
5438 : 0 : int ch = *value++;
5439 : :
5440 [ # # ]: 0 : if (ch < rangelo)
5441 : 0 : ch = rangelo - 1;
5442 [ # # ]: 0 : else if (ch > rangehi)
5443 : 0 : ch = rangehi + 1;
5444 : 0 : num += ((double) (ch - rangelo)) / denom;
5445 : 0 : denom *= base;
5446 : : }
5447 : :
5448 : 0 : return num;
5449 : : }
5450 : :
5451 : : /*
5452 : : * Do convert_to_scalar()'s work for any timevalue data type.
5453 : : *
5454 : : * On failure (e.g., unsupported typid), set *failure to true;
5455 : : * otherwise, that variable is not changed.
5456 : : */
5457 : : static double
5458 : 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5459 : : {
5460 [ # # # # : 0 : switch (typid)
# # # ]
5461 : : {
5462 : 0 : case TIMESTAMPOID:
5463 : 0 : return DatumGetTimestamp(value);
5464 : 0 : case TIMESTAMPTZOID:
5465 : 0 : return DatumGetTimestampTz(value);
5466 : 0 : case DATEOID:
5467 : 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
5468 : 0 : case INTERVALOID:
5469 : : {
5470 : 0 : Interval *interval = DatumGetIntervalP(value);
5471 : :
5472 : : /*
5473 : : * Convert the month part of Interval to days using assumed
5474 : : * average month length of 365.25/12.0 days. Not too
5475 : : * accurate, but plenty good enough for our purposes.
5476 : : *
5477 : : * This also works for infinite intervals, which just have all
5478 : : * fields set to INT_MIN/INT_MAX, and so will produce a result
5479 : : * smaller/larger than any finite interval.
5480 : : */
5481 : 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5482 : 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5483 : : }
5484 : 0 : case TIMEOID:
5485 : 0 : return DatumGetTimeADT(value);
5486 : 0 : case TIMETZOID:
5487 : : {
5488 : 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5489 : :
5490 : : /* use GMT-equivalent time */
5491 : 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5492 : : }
5493 : : }
5494 : :
5495 : 0 : *failure = true;
5496 : 0 : return 0;
5497 : : }
5498 : :
5499 : :
5500 : : /*
5501 : : * get_restriction_variable
5502 : : * Examine the args of a restriction clause to see if it's of the
5503 : : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5504 : : * where "variable" could be either a Var or an expression in vars of a
5505 : : * single relation. If so, extract information about the variable,
5506 : : * and also indicate which side it was on and the other argument.
5507 : : *
5508 : : * Inputs:
5509 : : * root: the planner info
5510 : : * args: clause argument list
5511 : : * varRelid: see specs for restriction selectivity functions
5512 : : *
5513 : : * Outputs: (these are valid only if true is returned)
5514 : : * *vardata: gets information about variable (see examine_variable)
5515 : : * *other: gets other clause argument, aggressively reduced to a constant
5516 : : * *varonleft: set true if variable is on the left, false if on the right
5517 : : *
5518 : : * Returns true if a variable is identified, otherwise false.
5519 : : *
5520 : : * Note: if there are Vars on both sides of the clause, we must fail, because
5521 : : * callers are expecting that the other side will act like a pseudoconstant.
5522 : : */
5523 : : bool
5524 : 626923 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5525 : : VariableStatData *vardata, Node **other,
5526 : : bool *varonleft)
5527 : : {
5528 : : Node *left,
5529 : : *right;
5530 : : VariableStatData rdata;
5531 : :
5532 : : /* Fail if not a binary opclause (probably shouldn't happen) */
5533 [ - + ]: 626923 : if (list_length(args) != 2)
5534 : 0 : return false;
5535 : :
5536 : 626923 : left = (Node *) linitial(args);
5537 : 626923 : right = (Node *) lsecond(args);
5538 : :
5539 : : /*
5540 : : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5541 : : * relations will be treated as pseudoconstants.
5542 : : */
5543 : 626923 : examine_variable(root, left, varRelid, vardata);
5544 : 626923 : examine_variable(root, right, varRelid, &rdata);
5545 : :
5546 : : /*
5547 : : * If one side is a variable and the other not, we win.
5548 : : */
5549 [ + + + + ]: 626923 : if (vardata->rel && rdata.rel == NULL)
5550 : : {
5551 : 559828 : *varonleft = true;
5552 : 559828 : *other = estimate_expression_value(root, rdata.var);
5553 : : /* Assume we need no ReleaseVariableStats(rdata) here */
5554 : 559824 : return true;
5555 : : }
5556 : :
5557 [ + + + + ]: 67095 : if (vardata->rel == NULL && rdata.rel)
5558 : : {
5559 : 63050 : *varonleft = false;
5560 : 63050 : *other = estimate_expression_value(root, vardata->var);
5561 : : /* Assume we need no ReleaseVariableStats(*vardata) here */
5562 : 63050 : *vardata = rdata;
5563 : 63050 : return true;
5564 : : }
5565 : :
5566 : : /* Oops, clause has wrong structure (probably var op var) */
5567 [ + + ]: 4045 : ReleaseVariableStats(*vardata);
5568 [ + + ]: 4045 : ReleaseVariableStats(rdata);
5569 : :
5570 : 4045 : return false;
5571 : : }
5572 : :
5573 : : /*
5574 : : * get_join_variables
5575 : : * Apply examine_variable() to each side of a join clause.
5576 : : * Also, attempt to identify whether the join clause has the same
5577 : : * or reversed sense compared to the SpecialJoinInfo.
5578 : : *
5579 : : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5580 : : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5581 : : * where we can't tell for sure, we default to assuming it's normal.
5582 : : */
5583 : : void
5584 : 209508 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5585 : : VariableStatData *vardata1, VariableStatData *vardata2,
5586 : : bool *join_is_reversed)
5587 : : {
5588 : : Node *left,
5589 : : *right;
5590 : :
5591 [ - + ]: 209508 : if (list_length(args) != 2)
5592 [ # # ]: 0 : elog(ERROR, "join operator should take two arguments");
5593 : :
5594 : 209508 : left = (Node *) linitial(args);
5595 : 209508 : right = (Node *) lsecond(args);
5596 : :
5597 : 209508 : examine_variable(root, left, 0, vardata1);
5598 : 209508 : examine_variable(root, right, 0, vardata2);
5599 : :
5600 [ + + + + ]: 418806 : if (vardata1->rel &&
5601 : 209298 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
5602 : 69071 : *join_is_reversed = true; /* var1 is on RHS */
5603 [ + + + + ]: 280664 : else if (vardata2->rel &&
5604 : 140227 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
5605 : 304 : *join_is_reversed = true; /* var2 is on LHS */
5606 : : else
5607 : 140133 : *join_is_reversed = false;
5608 : 209508 : }
5609 : :
5610 : : /* statext_expressions_load copies the tuple, so just pfree it. */
5611 : : static void
5612 : 1415 : ReleaseDummy(HeapTuple tuple)
5613 : : {
5614 : 1415 : pfree(tuple);
5615 : 1415 : }
5616 : :
5617 : : /*
5618 : : * examine_variable
5619 : : * Try to look up statistical data about an expression.
5620 : : * Fill in a VariableStatData struct to describe the expression.
5621 : : *
5622 : : * Inputs:
5623 : : * root: the planner info
5624 : : * node: the expression tree to examine
5625 : : * varRelid: see specs for restriction selectivity functions
5626 : : *
5627 : : * Outputs: *vardata is filled as follows:
5628 : : * var: the input expression (with any phvs or binary relabeling stripped,
5629 : : * if it is or contains a variable; but otherwise unchanged)
5630 : : * rel: RelOptInfo for relation containing variable; NULL if expression
5631 : : * contains no Vars (NOTE this could point to a RelOptInfo of a
5632 : : * subquery, not one in the current query).
5633 : : * statsTuple: the pg_statistic entry for the variable, if one exists;
5634 : : * otherwise NULL.
5635 : : * freefunc: pointer to a function to release statsTuple with.
5636 : : * vartype: exposed type of the expression; this should always match
5637 : : * the declared input type of the operator we are estimating for.
5638 : : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5639 : : * commonly the same as the exposed type of the variable argument,
5640 : : * but can be different in binary-compatible-type cases.
5641 : : * isunique: true if we were able to match the var to a unique index, a
5642 : : * single-column DISTINCT or GROUP-BY clause, implying its values are
5643 : : * unique for this query. (Caution: this should be trusted for
5644 : : * statistical purposes only, since we do not check indimmediate nor
5645 : : * verify that the exact same definition of equality applies.)
5646 : : * acl_ok: true if current user has permission to read all table rows from
5647 : : * the column(s) underlying the pg_statistic entry. This is consulted by
5648 : : * statistic_proc_security_check().
5649 : : *
5650 : : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5651 : : */
5652 : : void
5653 : 2500170 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5654 : : VariableStatData *vardata)
5655 : : {
5656 : : Node *basenode;
5657 : : Relids varnos;
5658 : : Relids basevarnos;
5659 : : RelOptInfo *onerel;
5660 : :
5661 : : /* Make sure we don't return dangling pointers in vardata */
5662 [ + - + - : 17501190 : MemSet(vardata, 0, sizeof(VariableStatData));
+ - + - +
+ ]
5663 : :
5664 : : /* Save the exposed type of the expression */
5665 : 2500170 : vardata->vartype = exprType(node);
5666 : :
5667 : : /*
5668 : : * PlaceHolderVars are transparent for the purpose of statistics lookup;
5669 : : * they do not alter the value distribution of the underlying expression.
5670 : : * However, they can obscure the structure, preventing us from recognizing
5671 : : * matches to base columns, index expressions, or extended statistics. So
5672 : : * strip them out first.
5673 : : */
5674 : 2500170 : basenode = strip_all_phvs_deep(root, node);
5675 : :
5676 : : /*
5677 : : * Look inside any binary-compatible relabeling. We need to handle nested
5678 : : * RelabelType nodes here, because the prior stripping of PlaceHolderVars
5679 : : * may have brought separate RelabelTypes into adjacency.
5680 : : */
5681 [ + + ]: 2540625 : while (IsA(basenode, RelabelType))
5682 : 40455 : basenode = (Node *) ((RelabelType *) basenode)->arg;
5683 : :
5684 : : /* Fast path for a simple Var */
5685 [ + + + + ]: 2500170 : if (IsA(basenode, Var) &&
5686 [ + + ]: 642916 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5687 : : {
5688 : 1769110 : Var *var = (Var *) basenode;
5689 : :
5690 : : /* Set up result fields other than the stats tuple */
5691 : 1769110 : vardata->var = basenode; /* return Var without phvs or relabeling */
5692 : 1769110 : vardata->rel = find_base_rel(root, var->varno);
5693 : 1769110 : vardata->atttype = var->vartype;
5694 : 1769110 : vardata->atttypmod = var->vartypmod;
5695 : 1769110 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5696 : :
5697 : : /* Try to locate some stats */
5698 : 1769110 : examine_simple_variable(root, var, vardata);
5699 : :
5700 : 1769110 : return;
5701 : : }
5702 : :
5703 : : /*
5704 : : * Okay, it's a more complicated expression. Determine variable
5705 : : * membership. Note that when varRelid isn't zero, only vars of that
5706 : : * relation are considered "real" vars.
5707 : : */
5708 : 731060 : varnos = pull_varnos(root, basenode);
5709 : 731060 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5710 : :
5711 : 731060 : onerel = NULL;
5712 : :
5713 [ + + ]: 731060 : if (bms_is_empty(basevarnos))
5714 : : {
5715 : : /* No Vars at all ... must be pseudo-constant clause */
5716 : : }
5717 : : else
5718 : : {
5719 : : int relid;
5720 : :
5721 : : /* Check if the expression is in vars of a single base relation */
5722 [ + + ]: 379678 : if (bms_get_singleton_member(basevarnos, &relid))
5723 : : {
5724 [ + + + + ]: 373398 : if (varRelid == 0 || varRelid == relid)
5725 : : {
5726 : 55679 : onerel = find_base_rel(root, relid);
5727 : 55679 : vardata->rel = onerel;
5728 : 55679 : node = basenode; /* strip any phvs or relabeling */
5729 : : }
5730 : : /* else treat it as a constant */
5731 : : }
5732 : : else
5733 : : {
5734 : : /* varnos has multiple relids */
5735 [ + + ]: 6280 : if (varRelid == 0)
5736 : : {
5737 : : /* treat it as a variable of a join relation */
5738 : 5284 : vardata->rel = find_join_rel(root, varnos);
5739 : 5284 : node = basenode; /* strip any phvs or relabeling */
5740 : : }
5741 [ + + ]: 996 : else if (bms_is_member(varRelid, varnos))
5742 : : {
5743 : : /* ignore the vars belonging to other relations */
5744 : 901 : vardata->rel = find_base_rel(root, varRelid);
5745 : 901 : node = basenode; /* strip any phvs or relabeling */
5746 : : /* note: no point in expressional-index search here */
5747 : : }
5748 : : /* else treat it as a constant */
5749 : : }
5750 : : }
5751 : :
5752 : 731060 : bms_free(basevarnos);
5753 : :
5754 : 731060 : vardata->var = node;
5755 : 731060 : vardata->atttype = exprType(node);
5756 : 731060 : vardata->atttypmod = exprTypmod(node);
5757 : :
5758 [ + + ]: 731060 : if (onerel)
5759 : : {
5760 : : /*
5761 : : * We have an expression in vars of a single relation. Try to match
5762 : : * it to expressional index columns, in hopes of finding some
5763 : : * statistics.
5764 : : *
5765 : : * Note that we consider all index columns including INCLUDE columns,
5766 : : * since there could be stats for such columns. But the test for
5767 : : * uniqueness needs to be warier.
5768 : : *
5769 : : * XXX it's conceivable that there are multiple matches with different
5770 : : * index opfamilies; if so, we need to pick one that matches the
5771 : : * operator we are estimating for. FIXME later.
5772 : : */
5773 : : ListCell *ilist;
5774 : : ListCell *slist;
5775 : :
5776 : : /*
5777 : : * The nullingrels bits within the expression could prevent us from
5778 : : * matching it to expressional index columns or to the expressions in
5779 : : * extended statistics. So strip them out first.
5780 : : */
5781 [ + + ]: 55679 : if (bms_overlap(varnos, root->outer_join_rels))
5782 : 1573 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5783 : :
5784 [ + + + + : 127009 : foreach(ilist, onerel->indexlist)
+ + ]
5785 : : {
5786 : 73707 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5787 : : ListCell *indexpr_item;
5788 : : int pos;
5789 : :
5790 : 73707 : indexpr_item = list_head(index->indexprs);
5791 [ + + ]: 73707 : if (indexpr_item == NULL)
5792 : 69746 : continue; /* no expressions here... */
5793 : :
5794 [ + + ]: 5609 : for (pos = 0; pos < index->ncolumns; pos++)
5795 : : {
5796 [ + + ]: 4025 : if (index->indexkeys[pos] == 0)
5797 : : {
5798 : : Node *indexkey;
5799 : :
5800 [ - + ]: 3961 : if (indexpr_item == NULL)
5801 [ # # ]: 0 : elog(ERROR, "too few entries in indexprs list");
5802 : 3961 : indexkey = (Node *) lfirst(indexpr_item);
5803 [ + - - + ]: 3961 : if (indexkey && IsA(indexkey, RelabelType))
5804 : 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5805 [ + + ]: 3961 : if (equal(node, indexkey))
5806 : : {
5807 : : /*
5808 : : * Found a match ... is it a unique index? Tests here
5809 : : * should match has_unique_index().
5810 : : */
5811 [ + + ]: 2916 : if (index->unique &&
5812 [ + - + - ]: 365 : index->nkeycolumns == 1 &&
5813 : 365 : pos == 0 &&
5814 [ - + - - ]: 365 : (index->indpred == NIL || index->predOK))
5815 : 365 : vardata->isunique = true;
5816 : :
5817 : : /*
5818 : : * Has it got stats? We only consider stats for
5819 : : * non-partial indexes, since partial indexes probably
5820 : : * don't reflect whole-relation statistics; the above
5821 : : * check for uniqueness is the only info we take from
5822 : : * a partial index.
5823 : : *
5824 : : * An index stats hook, however, must make its own
5825 : : * decisions about what to do with partial indexes.
5826 : : */
5827 [ - + - - ]: 2916 : if (get_index_stats_hook &&
5828 : 0 : (*get_index_stats_hook) (root, index->indexoid,
5829 : 0 : pos + 1, vardata))
5830 : : {
5831 : : /*
5832 : : * The hook took control of acquiring a stats
5833 : : * tuple. If it did supply a tuple, it'd better
5834 : : * have supplied a freefunc.
5835 : : */
5836 [ # # ]: 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5837 [ # # ]: 0 : !vardata->freefunc)
5838 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
5839 : : }
5840 [ + - ]: 2916 : else if (index->indpred == NIL)
5841 : : {
5842 : 2916 : vardata->statsTuple =
5843 : 5832 : SearchSysCache3(STATRELATTINH,
5844 : : ObjectIdGetDatum(index->indexoid),
5845 : 2916 : Int16GetDatum(pos + 1),
5846 : : BoolGetDatum(false));
5847 : 2916 : vardata->freefunc = ReleaseSysCache;
5848 : :
5849 [ + + ]: 2916 : if (HeapTupleIsValid(vardata->statsTuple))
5850 : : {
5851 : : /*
5852 : : * Test if user has permission to access all
5853 : : * rows from the index's table.
5854 : : *
5855 : : * For simplicity, we insist on the whole
5856 : : * table being selectable, rather than trying
5857 : : * to identify which column(s) the index
5858 : : * depends on.
5859 : : *
5860 : : * Note that for an inheritance child,
5861 : : * permissions are checked on the inheritance
5862 : : * root parent, and whole-table select
5863 : : * privilege on the parent doesn't quite
5864 : : * guarantee that the user could read all
5865 : : * columns of the child. But in practice it's
5866 : : * unlikely that any interesting security
5867 : : * violation could result from allowing access
5868 : : * to the expression index's stats, so we
5869 : : * allow it anyway. See similar code in
5870 : : * examine_simple_variable() for additional
5871 : : * comments.
5872 : : */
5873 : 2377 : vardata->acl_ok =
5874 : 2377 : all_rows_selectable(root,
5875 : 2377 : index->rel->relid,
5876 : : NULL);
5877 : : }
5878 : : else
5879 : : {
5880 : : /* suppress leakproofness checks later */
5881 : 539 : vardata->acl_ok = true;
5882 : : }
5883 : : }
5884 [ + + ]: 2916 : if (vardata->statsTuple)
5885 : 2377 : break;
5886 : : }
5887 : 1584 : indexpr_item = lnext(index->indexprs, indexpr_item);
5888 : : }
5889 : : }
5890 [ + + ]: 3961 : if (vardata->statsTuple)
5891 : 2377 : break;
5892 : : }
5893 : :
5894 : : /*
5895 : : * Search extended statistics for one with a matching expression.
5896 : : * There might be multiple ones, so just grab the first one. In the
5897 : : * future, we might consider the statistics target (and pick the most
5898 : : * accurate statistics) and maybe some other parameters.
5899 : : */
5900 [ + + + + : 59208 : foreach(slist, onerel->statlist)
+ + ]
5901 : : {
5902 : 3774 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5903 [ + - ]: 3774 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5904 : : ListCell *expr_item;
5905 : : int pos;
5906 : :
5907 : : /*
5908 : : * Stop once we've found statistics for the expression (either
5909 : : * from extended stats, or for an index in the preceding loop).
5910 : : */
5911 [ + + ]: 3774 : if (vardata->statsTuple)
5912 : 245 : break;
5913 : :
5914 : : /* skip stats without per-expression stats */
5915 [ + + ]: 3529 : if (info->kind != STATS_EXT_EXPRESSIONS)
5916 : 1808 : continue;
5917 : :
5918 : : /* skip stats with mismatching stxdinherit value */
5919 [ + + ]: 1721 : if (info->inherit != rte->inh)
5920 : 5 : continue;
5921 : :
5922 : 1716 : pos = 0;
5923 [ + - + + : 2836 : foreach(expr_item, info->exprs)
+ + ]
5924 : : {
5925 : 2536 : Node *expr = (Node *) lfirst(expr_item);
5926 : :
5927 : : Assert(expr);
5928 : :
5929 : : /* strip RelabelType before comparing it */
5930 [ + - - + ]: 2536 : if (expr && IsA(expr, RelabelType))
5931 : 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5932 : :
5933 : : /* found a match, see if we can extract pg_statistic row */
5934 [ + + ]: 2536 : if (equal(node, expr))
5935 : : {
5936 : : /*
5937 : : * XXX Not sure if we should cache the tuple somewhere.
5938 : : * Now we just create a new copy every time.
5939 : : */
5940 : 1416 : vardata->statsTuple =
5941 : 1416 : statext_expressions_load(info->statOid, rte->inh, pos);
5942 : :
5943 : : /* Nothing to release if no data found */
5944 [ + + ]: 1416 : if (vardata->statsTuple != NULL)
5945 : : {
5946 : 1415 : vardata->freefunc = ReleaseDummy;
5947 : : }
5948 : :
5949 : : /*
5950 : : * Test if user has permission to access all rows from the
5951 : : * table.
5952 : : *
5953 : : * For simplicity, we insist on the whole table being
5954 : : * selectable, rather than trying to identify which
5955 : : * column(s) the statistics object depends on.
5956 : : *
5957 : : * Note that for an inheritance child, permissions are
5958 : : * checked on the inheritance root parent, and whole-table
5959 : : * select privilege on the parent doesn't quite guarantee
5960 : : * that the user could read all columns of the child. But
5961 : : * in practice it's unlikely that any interesting security
5962 : : * violation could result from allowing access to the
5963 : : * expression stats, so we allow it anyway. See similar
5964 : : * code in examine_simple_variable() for additional
5965 : : * comments.
5966 : : */
5967 : 1416 : vardata->acl_ok = all_rows_selectable(root,
5968 : : onerel->relid,
5969 : : NULL);
5970 : :
5971 : 1416 : break;
5972 : : }
5973 : :
5974 : 1120 : pos++;
5975 : : }
5976 : : }
5977 : : }
5978 : :
5979 : 731060 : bms_free(varnos);
5980 : : }
5981 : :
5982 : : /*
5983 : : * strip_all_phvs_deep
5984 : : * Deeply strip all PlaceHolderVars in an expression.
5985 : : *
5986 : : * As a performance optimization, we first use a lightweight walker to check
5987 : : * for the presence of any PlaceHolderVars. The expensive mutator is invoked
5988 : : * only if a PlaceHolderVar is found, avoiding unnecessary memory allocation
5989 : : * and tree copying in the common case where no PlaceHolderVars are present.
5990 : : */
5991 : : static Node *
5992 : 2500170 : strip_all_phvs_deep(PlannerInfo *root, Node *node)
5993 : : {
5994 : : /* If there are no PHVs anywhere, we needn't work hard */
5995 [ + + ]: 2500170 : if (root->glob->lastPHId == 0)
5996 : 2474034 : return node;
5997 : :
5998 [ + + ]: 26136 : if (!contain_placeholder_walker(node, NULL))
5999 : 22398 : return node;
6000 : 3738 : return strip_all_phvs_mutator(node, NULL);
6001 : : }
6002 : :
6003 : : /*
6004 : : * contain_placeholder_walker
6005 : : * Lightweight walker to check if an expression contains any
6006 : : * PlaceHolderVars
6007 : : */
6008 : : static bool
6009 : 29440 : contain_placeholder_walker(Node *node, void *context)
6010 : : {
6011 [ + + ]: 29440 : if (node == NULL)
6012 : 109 : return false;
6013 [ + + ]: 29331 : if (IsA(node, PlaceHolderVar))
6014 : 3738 : return true;
6015 : :
6016 : 25593 : return expression_tree_walker(node, contain_placeholder_walker, context);
6017 : : }
6018 : :
6019 : : /*
6020 : : * strip_all_phvs_mutator
6021 : : * Mutator to deeply strip all PlaceHolderVars
6022 : : */
6023 : : static Node *
6024 : 9929 : strip_all_phvs_mutator(Node *node, void *context)
6025 : : {
6026 [ + + ]: 9929 : if (node == NULL)
6027 : 34 : return NULL;
6028 [ + + ]: 9895 : if (IsA(node, PlaceHolderVar))
6029 : : {
6030 : : /* Strip it and recurse into its contained expression */
6031 : 3858 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6032 : :
6033 : 3858 : return strip_all_phvs_mutator((Node *) phv->phexpr, context);
6034 : : }
6035 : :
6036 : 6037 : return expression_tree_mutator(node, strip_all_phvs_mutator, context);
6037 : : }
6038 : :
6039 : : /*
6040 : : * examine_simple_variable
6041 : : * Handle a simple Var for examine_variable
6042 : : *
6043 : : * This is split out as a subroutine so that we can recurse to deal with
6044 : : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
6045 : : *
6046 : : * We already filled in all the fields of *vardata except for the stats tuple.
6047 : : */
6048 : : static void
6049 : 1778559 : examine_simple_variable(PlannerInfo *root, Var *var,
6050 : : VariableStatData *vardata)
6051 : : {
6052 : 1778559 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
6053 : :
6054 : : Assert(IsA(rte, RangeTblEntry));
6055 : :
6056 [ - + - - ]: 1778559 : if (get_relation_stats_hook &&
6057 : 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
6058 : : {
6059 : : /*
6060 : : * The hook took control of acquiring a stats tuple. If it did supply
6061 : : * a tuple, it'd better have supplied a freefunc.
6062 : : */
6063 [ # # ]: 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6064 [ # # ]: 0 : !vardata->freefunc)
6065 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
6066 : : }
6067 [ + + ]: 1778559 : else if (rte->rtekind == RTE_RELATION)
6068 : : {
6069 : : /*
6070 : : * Plain table or parent of an inheritance appendrel, so look up the
6071 : : * column in pg_statistic
6072 : : */
6073 : 1687224 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6074 : : ObjectIdGetDatum(rte->relid),
6075 : 1687224 : Int16GetDatum(var->varattno),
6076 : 1687224 : BoolGetDatum(rte->inh));
6077 : 1687224 : vardata->freefunc = ReleaseSysCache;
6078 : :
6079 [ + + ]: 1687224 : if (HeapTupleIsValid(vardata->statsTuple))
6080 : : {
6081 : : /*
6082 : : * Test if user has permission to read all rows from this column.
6083 : : *
6084 : : * This requires that the user has the appropriate SELECT
6085 : : * privileges and that there are no securityQuals from security
6086 : : * barrier views or RLS policies. If that's not the case, then we
6087 : : * only permit leakproof functions to be passed pg_statistic data
6088 : : * in vardata, otherwise the functions might reveal data that the
6089 : : * user doesn't have permission to see --- see
6090 : : * statistic_proc_security_check().
6091 : : */
6092 : 1087622 : vardata->acl_ok =
6093 : 1087622 : all_rows_selectable(root, var->varno,
6094 : 1087622 : bms_make_singleton(var->varattno - FirstLowInvalidHeapAttributeNumber));
6095 : : }
6096 : : else
6097 : : {
6098 : : /* suppress any possible leakproofness checks later */
6099 : 599602 : vardata->acl_ok = true;
6100 : : }
6101 : : }
6102 [ + + + + ]: 91335 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
6103 [ + + + + ]: 83210 : (rte->rtekind == RTE_CTE && !rte->self_reference))
6104 : : {
6105 : : /*
6106 : : * Plain subquery (not one that was converted to an appendrel) or
6107 : : * non-recursive CTE. In either case, we can try to find out what the
6108 : : * Var refers to within the subquery. We skip this for appendrel and
6109 : : * recursive-CTE cases because any column stats we did find would
6110 : : * likely not be very relevant.
6111 : : */
6112 : : PlannerInfo *subroot;
6113 : : Query *subquery;
6114 : : List *subtlist;
6115 : : TargetEntry *ste;
6116 : 17052 : bool have_grouping = false;
6117 : :
6118 : : /*
6119 : : * Punt if it's a whole-row var rather than a plain column reference.
6120 : : */
6121 [ - + ]: 17052 : if (var->varattno == InvalidAttrNumber)
6122 : 0 : return;
6123 : :
6124 : : /*
6125 : : * Otherwise, find the subquery's planner subroot.
6126 : : */
6127 [ + + ]: 17052 : if (rte->rtekind == RTE_SUBQUERY)
6128 : : {
6129 : : RelOptInfo *rel;
6130 : :
6131 : : /*
6132 : : * Fetch RelOptInfo for subquery. Note that we don't change the
6133 : : * rel returned in vardata, since caller expects it to be a rel of
6134 : : * the caller's query level. Because we might already be
6135 : : * recursing, we can't use that rel pointer either, but have to
6136 : : * look up the Var's rel afresh.
6137 : : */
6138 : 8125 : rel = find_base_rel(root, var->varno);
6139 : :
6140 : 8125 : subroot = rel->subroot;
6141 : : }
6142 : : else
6143 : : {
6144 : : /* CTE case is more difficult */
6145 : : PlannerInfo *cteroot;
6146 : : Index levelsup;
6147 : : int ndx;
6148 : : int plan_id;
6149 : : ListCell *lc;
6150 : :
6151 : : /*
6152 : : * Find the referenced CTE, and locate the subroot previously made
6153 : : * for it.
6154 : : */
6155 : 8927 : levelsup = rte->ctelevelsup;
6156 : 8927 : cteroot = root;
6157 [ + + ]: 21363 : while (levelsup-- > 0)
6158 : : {
6159 : 12436 : cteroot = cteroot->parent_root;
6160 [ - + ]: 12436 : if (!cteroot) /* shouldn't happen */
6161 [ # # ]: 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
6162 : : }
6163 : :
6164 : : /*
6165 : : * Note: cte_plan_ids can be shorter than cteList, if we are still
6166 : : * working on planning the CTEs (ie, this is a side-reference from
6167 : : * another CTE). So we mustn't use forboth here.
6168 : : */
6169 : 8927 : ndx = 0;
6170 [ + - + - : 13105 : foreach(lc, cteroot->parse->cteList)
+ - ]
6171 : : {
6172 : 13105 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
6173 : :
6174 [ + + ]: 13105 : if (strcmp(cte->ctename, rte->ctename) == 0)
6175 : 8927 : break;
6176 : 4178 : ndx++;
6177 : : }
6178 [ - + ]: 8927 : if (lc == NULL) /* shouldn't happen */
6179 [ # # ]: 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
6180 [ - + ]: 8927 : if (ndx >= list_length(cteroot->cte_plan_ids))
6181 [ # # ]: 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
6182 : 8927 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
6183 [ - + ]: 8927 : if (plan_id <= 0)
6184 [ # # ]: 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
6185 : 8927 : subroot = list_nth(root->glob->subroots, plan_id - 1);
6186 : : }
6187 : :
6188 : : /* If the subquery hasn't been planned yet, we have to punt */
6189 [ - + ]: 17052 : if (subroot == NULL)
6190 : 0 : return;
6191 : : Assert(IsA(subroot, PlannerInfo));
6192 : :
6193 : : /*
6194 : : * We must use the subquery parsetree as mangled by the planner, not
6195 : : * the raw version from the RTE, because we need a Var that will refer
6196 : : * to the subroot's live RelOptInfos. For instance, if any subquery
6197 : : * pullup happened during planning, Vars in the targetlist might have
6198 : : * gotten replaced, and we need to see the replacement expressions.
6199 : : */
6200 : 17052 : subquery = subroot->parse;
6201 : : Assert(IsA(subquery, Query));
6202 : :
6203 : : /*
6204 : : * Punt if subquery uses set operations or grouping sets, as these
6205 : : * will mash underlying columns' stats beyond recognition. (Set ops
6206 : : * are particularly nasty; if we forged ahead, we would return stats
6207 : : * relevant to only the leftmost subselect...) DISTINCT and GROUP BY
6208 : : * are also problematic, but we check those later because there is a
6209 : : * possibility of learning something even with them: we can detect
6210 : : * uniqueness for single-column cases, and for key columns that are
6211 : : * simple Vars, we can obtain a useful stadistinct from the underlying
6212 : : * base table.
6213 : : */
6214 [ + + ]: 17052 : if (subquery->setOperations ||
6215 [ + + ]: 15329 : subquery->groupingSets)
6216 : 1800 : return;
6217 : :
6218 : : /* Get the subquery output expression referenced by the upper Var */
6219 [ + + ]: 15252 : if (subquery->returningList)
6220 : 179 : subtlist = subquery->returningList;
6221 : : else
6222 : 15073 : subtlist = subquery->targetList;
6223 : 15252 : ste = get_tle_by_resno(subtlist, var->varattno);
6224 [ + - - + ]: 15252 : if (ste == NULL || ste->resjunk)
6225 [ # # ]: 0 : elog(ERROR, "subquery %s does not have attribute %d",
6226 : : rte->eref->aliasname, var->varattno);
6227 : 15252 : var = (Var *) ste->expr;
6228 : :
6229 : : /*
6230 : : * If subquery uses DISTINCT, we can't make full use of stats for the
6231 : : * variable ... but, if it's the only DISTINCT column, we are entitled
6232 : : * to consider it unique. We do the test this way so that it works
6233 : : * for cases involving DISTINCT ON.
6234 : : *
6235 : : * If the target is a DISTINCT key that is a simple Var, we can still
6236 : : * obtain a useful stadistinct from the base table, though the
6237 : : * frequency-dependent stats must be adjusted since DISTINCT changes
6238 : : * the frequency distribution. We set have_grouping and fall through
6239 : : * to the simple-Var recursion below. Non-key columns cannot go
6240 : : * further.
6241 : : */
6242 [ + + ]: 15252 : if (subquery->distinctClause)
6243 : : {
6244 [ + + ]: 1491 : if (targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
6245 : : {
6246 : 845 : have_grouping = true;
6247 : :
6248 [ + + ]: 845 : if (list_length(subquery->distinctClause) == 1)
6249 : 374 : vardata->isunique = true;
6250 : : }
6251 : : else
6252 : 646 : return;
6253 : : }
6254 : :
6255 : : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
6256 [ + + ]: 14606 : if (subquery->groupClause)
6257 : : {
6258 [ + + ]: 619 : if (targetIsInSortList(ste, InvalidOid, subquery->groupClause))
6259 : : {
6260 : 485 : have_grouping = true;
6261 : :
6262 [ + + ]: 485 : if (list_length(subquery->groupClause) == 1)
6263 : 322 : vardata->isunique = true;
6264 : : }
6265 [ + - ]: 134 : else if (!have_grouping)
6266 : 134 : return;
6267 : : }
6268 : :
6269 : : /*
6270 : : * If the sub-query originated from a view with the security_barrier
6271 : : * attribute, we must not look at the variable's statistics, though it
6272 : : * seems all right to notice the existence of a DISTINCT clause. So
6273 : : * stop here.
6274 : : *
6275 : : * This is probably a harsher restriction than necessary; it's
6276 : : * certainly OK for the selectivity estimator (which is a C function,
6277 : : * and therefore omnipotent anyway) to look at the statistics. But
6278 : : * many selectivity estimators will happily *invoke the operator
6279 : : * function* to try to work out a good estimate - and that's not OK.
6280 : : * So for now, don't dig down for stats.
6281 : : */
6282 [ + + ]: 14472 : if (rte->security_barrier)
6283 : 488 : return;
6284 : :
6285 : : /* Can only handle a simple Var of subquery's query level */
6286 [ + - + + ]: 13984 : if (var && IsA(var, Var) &&
6287 [ + - ]: 9449 : var->varlevelsup == 0)
6288 : : {
6289 : : /*
6290 : : * OK, recurse into the subquery. Note that the original setting
6291 : : * of vardata->isunique (which will surely be false) is left
6292 : : * unchanged in this situation. That's what we want, since even
6293 : : * if the underlying column is unique, the subquery may have
6294 : : * joined to other tables in a way that creates duplicates.
6295 : : */
6296 : 9449 : examine_simple_variable(subroot, var, vardata);
6297 : :
6298 : : /*
6299 : : * If the subquery uses DISTINCT or GROUP BY and we got here
6300 : : * because the target is a key column, adjust the recursively
6301 : : * obtained stats tuple for the grouped context.
6302 : : */
6303 [ + + ]: 9449 : if (have_grouping)
6304 : 1195 : adjust_statstuple_for_grouping(subroot, var, vardata);
6305 : : }
6306 : : }
6307 : : else
6308 : : {
6309 : : /*
6310 : : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
6311 : : * see RTE_JOIN here because join alias Vars have already been
6312 : : * flattened.) There's not much we can do with function outputs, but
6313 : : * maybe someday try to be smarter about VALUES.
6314 : : */
6315 : : }
6316 : : }
6317 : :
6318 : : /*
6319 : : * adjust_statstuple_for_grouping
6320 : : * Adjust a stats tuple for use in a grouped or distinct context.
6321 : : *
6322 : : * This is used when the stats tuple was obtained by recursing into a subquery,
6323 : : * but the subquery's output invalidates frequency-related statistics (e.g. due
6324 : : * to GROUP BY or DISTINCT). The set of distinct values is preserved by such
6325 : : * operations, so stadistinct remains valid, but MCV frequencies, histograms,
6326 : : * and correlation data are not. Zeroing all stats slots causes callers (e.g.
6327 : : * var_eq_const) to fall through to the 1/ndistinct estimate instead.
6328 : : *
6329 : : * stanullfrac must also be adjusted. When this column is the only GROUP BY or
6330 : : * DISTINCT column, its NULLs are collapsed into one group, so the null
6331 : : * fraction is 1/(ndistinct+1) if the base column had NULLs. With multiple
6332 : : * grouping columns a NULL can pair with many combinations of the other keys,
6333 : : * so the null fraction depends on their joint distribution, which we don't
6334 : : * have. We approximate it as zero: NULLs collapse far more aggressively than
6335 : : * non-NULLs, so the output fraction is well below the base table's, and erring
6336 : : * low keeps estimates on the hash-join-favoring side.
6337 : : *
6338 : : * If stadistinct is negative (a fraction of the base table's row count), we
6339 : : * convert it to an absolute count, since it would otherwise be misinterpreted
6340 : : * relative to the subquery output's row count.
6341 : : */
6342 : : static void
6343 : 1195 : adjust_statstuple_for_grouping(PlannerInfo *subroot, Var *var,
6344 : : VariableStatData *vardata)
6345 : : {
6346 : : HeapTuple copy;
6347 : : Form_pg_statistic stats;
6348 : :
6349 [ + + ]: 1195 : if (!HeapTupleIsValid(vardata->statsTuple))
6350 : 621 : return;
6351 : :
6352 : 574 : copy = heap_copytuple(vardata->statsTuple);
6353 : 574 : stats = (Form_pg_statistic) GETSTRUCT(copy);
6354 : :
6355 : : /* Convert negative stadistinct to absolute count */
6356 [ + + ]: 574 : if (stats->stadistinct < 0)
6357 : : {
6358 : 340 : RelOptInfo *baserel = find_base_rel(subroot, var->varno);
6359 : :
6360 [ + - ]: 340 : if (baserel->tuples > 0)
6361 : : {
6362 : 340 : stats->stadistinct = (float4)
6363 : 340 : clamp_row_est(-stats->stadistinct * baserel->tuples);
6364 : : }
6365 : : }
6366 : :
6367 : : /* Zero out all stats slots */
6368 [ + + ]: 3444 : for (int k = 0; k < STATISTIC_NUM_SLOTS; k++)
6369 : 2870 : (&stats->stakind1)[k] = 0;
6370 : :
6371 : : /* Adjust the null fraction (see comment above). */
6372 [ + + - + : 574 : if (vardata->isunique && stats->stanullfrac > 0.0 && stats->stadistinct > 0)
- - ]
6373 : 0 : stats->stanullfrac = 1.0 / (stats->stadistinct + 1.0);
6374 : : else
6375 : 574 : stats->stanullfrac = 0.0;
6376 : :
6377 : : /* Replace original with our modified copy */
6378 : 574 : vardata->freefunc(vardata->statsTuple);
6379 : 574 : vardata->statsTuple = copy;
6380 : 574 : vardata->freefunc = heap_freetuple;
6381 : : }
6382 : :
6383 : : /*
6384 : : * all_rows_selectable
6385 : : * Test whether the user has permission to select all rows from a given
6386 : : * relation.
6387 : : *
6388 : : * Inputs:
6389 : : * root: the planner info
6390 : : * varno: the index of the relation (assumed to be an RTE_RELATION)
6391 : : * varattnos: the attributes for which permission is required, or NULL if
6392 : : * whole-table access is required
6393 : : *
6394 : : * Returns true if the user has the required select permissions, and there are
6395 : : * no securityQuals from security barrier views or RLS policies.
6396 : : *
6397 : : * Note that if the relation is an inheritance child relation, securityQuals
6398 : : * and access permissions are checked against the inheritance root parent (the
6399 : : * relation actually mentioned in the query) --- see the comments in
6400 : : * expand_single_inheritance_child() for an explanation of why it has to be
6401 : : * done this way.
6402 : : *
6403 : : * If varattnos is non-NULL, its attribute numbers should be offset by
6404 : : * FirstLowInvalidHeapAttributeNumber so that system attributes can be
6405 : : * checked. If varattnos is NULL, only table-level SELECT privileges are
6406 : : * checked, not any column-level privileges.
6407 : : *
6408 : : * Note: if the relation is accessed via a view, this function actually tests
6409 : : * whether the view owner has permission to select from the relation. To
6410 : : * ensure that the current user has permission, it is also necessary to check
6411 : : * that the current user has permission to select from the view, which we do
6412 : : * at planner-startup --- see subquery_planner().
6413 : : *
6414 : : * This is exported so that other estimation functions can use it.
6415 : : */
6416 : : bool
6417 : 1091625 : all_rows_selectable(PlannerInfo *root, Index varno, Bitmapset *varattnos)
6418 : : {
6419 : 1091625 : RelOptInfo *rel = find_base_rel_noerr(root, varno);
6420 [ + - ]: 1091625 : RangeTblEntry *rte = planner_rt_fetch(varno, root);
6421 : : Oid userid;
6422 : : int varattno;
6423 : :
6424 : : Assert(rte->rtekind == RTE_RELATION);
6425 : :
6426 : : /*
6427 : : * Determine the user ID to use for privilege checks (either the current
6428 : : * user or the view owner, if we're accessing the table via a view).
6429 : : *
6430 : : * Normally the relation will have an associated RelOptInfo from which we
6431 : : * can find the userid, but it might not if it's a RETURNING Var for an
6432 : : * INSERT target relation. In that case use the RTEPermissionInfo
6433 : : * associated with the RTE.
6434 : : *
6435 : : * If we navigate up to a parent relation, we keep using the same userid,
6436 : : * since it's the same in all relations of a given inheritance tree.
6437 : : */
6438 [ + + ]: 1091625 : if (rel)
6439 : 1091592 : userid = rel->userid;
6440 : : else
6441 : : {
6442 : : RTEPermissionInfo *perminfo;
6443 : :
6444 : 33 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
6445 : 33 : userid = perminfo->checkAsUser;
6446 : : }
6447 [ + + ]: 1091625 : if (!OidIsValid(userid))
6448 : 970757 : userid = GetUserId();
6449 : :
6450 : : /*
6451 : : * Permissions and securityQuals must be checked on the table actually
6452 : : * mentioned in the query, so if this is an inheritance child, navigate up
6453 : : * to the inheritance root parent. If the user can read the whole table
6454 : : * or the required columns there, then they can read from the child table
6455 : : * too. For per-column checks, we must find out which of the root
6456 : : * parent's attributes the child relation's attributes correspond to.
6457 : : */
6458 [ + + ]: 1091625 : if (root->append_rel_array != NULL)
6459 : : {
6460 : : AppendRelInfo *appinfo;
6461 : :
6462 : 177748 : appinfo = root->append_rel_array[varno];
6463 : :
6464 : : /*
6465 : : * Partitions are mapped to their immediate parent, not the root
6466 : : * parent, so must be ready to walk up multiple AppendRelInfos. But
6467 : : * stop if we hit a parent that is not RTE_RELATION --- that's a
6468 : : * flattened UNION ALL subquery, not an inheritance parent.
6469 : : */
6470 [ + + ]: 334090 : while (appinfo &&
6471 [ + - ]: 156727 : planner_rt_fetch(appinfo->parent_relid,
6472 [ + + ]: 156727 : root)->rtekind == RTE_RELATION)
6473 : : {
6474 : 156342 : Bitmapset *parent_varattnos = NULL;
6475 : :
6476 : : /*
6477 : : * For each child attribute, find the corresponding parent
6478 : : * attribute. In rare cases, the attribute may be local to the
6479 : : * child table, in which case, we've got to live with having no
6480 : : * access to this column.
6481 : : */
6482 : 156342 : varattno = -1;
6483 [ + + ]: 310417 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6484 : : {
6485 : : AttrNumber attno;
6486 : : AttrNumber parent_attno;
6487 : :
6488 : 154075 : attno = varattno + FirstLowInvalidHeapAttributeNumber;
6489 : :
6490 [ + + ]: 154075 : if (attno == InvalidAttrNumber)
6491 : : {
6492 : : /*
6493 : : * Whole-row reference, so must map each column of the
6494 : : * child to the parent table.
6495 : : */
6496 [ + + ]: 30 : for (attno = 1; attno <= appinfo->num_child_cols; attno++)
6497 : : {
6498 : 20 : parent_attno = appinfo->parent_colnos[attno - 1];
6499 [ - + ]: 20 : if (parent_attno == 0)
6500 : 0 : return false; /* attr is local to child */
6501 : : parent_varattnos =
6502 : 20 : bms_add_member(parent_varattnos,
6503 : : parent_attno - FirstLowInvalidHeapAttributeNumber);
6504 : : }
6505 : : }
6506 : : else
6507 : : {
6508 [ - + ]: 154065 : if (attno < 0)
6509 : : {
6510 : : /* System attnos are the same in all tables */
6511 : 0 : parent_attno = attno;
6512 : : }
6513 : : else
6514 : : {
6515 [ - + ]: 154065 : if (attno > appinfo->num_child_cols)
6516 : 0 : return false; /* safety check */
6517 : 154065 : parent_attno = appinfo->parent_colnos[attno - 1];
6518 [ - + ]: 154065 : if (parent_attno == 0)
6519 : 0 : return false; /* attr is local to child */
6520 : : }
6521 : : parent_varattnos =
6522 : 154065 : bms_add_member(parent_varattnos,
6523 : : parent_attno - FirstLowInvalidHeapAttributeNumber);
6524 : : }
6525 : : }
6526 : :
6527 : : /* If the parent is itself a child, continue up */
6528 : 156342 : varno = appinfo->parent_relid;
6529 : 156342 : varattnos = parent_varattnos;
6530 : 156342 : appinfo = root->append_rel_array[varno];
6531 : : }
6532 : :
6533 : : /* Perform the access check on this parent rel */
6534 [ + - ]: 177748 : rte = planner_rt_fetch(varno, root);
6535 : : Assert(rte->rtekind == RTE_RELATION);
6536 : : }
6537 : :
6538 : : /*
6539 : : * For all rows to be accessible, there must be no securityQuals from
6540 : : * security barrier views or RLS policies.
6541 : : */
6542 [ + + ]: 1091625 : if (rte->securityQuals != NIL)
6543 : 690 : return false;
6544 : :
6545 : : /*
6546 : : * Test for table-level SELECT privilege.
6547 : : *
6548 : : * If varattnos is non-NULL, this is sufficient to give access to all
6549 : : * requested attributes, even for a child table, since we have verified
6550 : : * that all required child columns have matching parent columns.
6551 : : *
6552 : : * If varattnos is NULL (whole-table access requested), this doesn't
6553 : : * necessarily guarantee that the user can read all columns of a child
6554 : : * table, but we allow it anyway (see comments in examine_variable()) and
6555 : : * don't bother checking any column privileges.
6556 : : */
6557 [ + + ]: 1090935 : if (pg_class_aclcheck(rte->relid, userid, ACL_SELECT) == ACLCHECK_OK)
6558 : 1090592 : return true;
6559 : :
6560 [ + + ]: 343 : if (varattnos == NULL)
6561 : 10 : return false; /* whole-table access requested */
6562 : :
6563 : : /*
6564 : : * Don't have table-level SELECT privilege, so check per-column
6565 : : * privileges.
6566 : : */
6567 : 333 : varattno = -1;
6568 [ + + ]: 471 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6569 : : {
6570 : 333 : AttrNumber attno = varattno + FirstLowInvalidHeapAttributeNumber;
6571 : :
6572 [ + + ]: 333 : if (attno == InvalidAttrNumber)
6573 : : {
6574 : : /* Whole-row reference, so must have access to all columns */
6575 [ + - ]: 5 : if (pg_attribute_aclcheck_all(rte->relid, userid, ACL_SELECT,
6576 : : ACLMASK_ALL) != ACLCHECK_OK)
6577 : 5 : return false;
6578 : : }
6579 : : else
6580 : : {
6581 [ + + ]: 328 : if (pg_attribute_aclcheck(rte->relid, attno, userid,
6582 : : ACL_SELECT) != ACLCHECK_OK)
6583 : 190 : return false;
6584 : : }
6585 : : }
6586 : :
6587 : : /* If we reach here, have all required column privileges */
6588 : 138 : return true;
6589 : : }
6590 : :
6591 : : /*
6592 : : * examine_indexcol_variable
6593 : : * Try to look up statistical data about an index column/expression.
6594 : : * Fill in a VariableStatData struct to describe the column.
6595 : : *
6596 : : * Inputs:
6597 : : * root: the planner info
6598 : : * index: the index whose column we're interested in
6599 : : * indexcol: 0-based index column number (subscripts index->indexkeys[])
6600 : : *
6601 : : * Outputs: *vardata is filled as follows:
6602 : : * var: the input expression (with any binary relabeling stripped, if
6603 : : * it is or contains a variable; but otherwise the type is preserved)
6604 : : * rel: RelOptInfo for table relation containing variable.
6605 : : * statsTuple: the pg_statistic entry for the variable, if one exists;
6606 : : * otherwise NULL.
6607 : : * freefunc: pointer to a function to release statsTuple with.
6608 : : *
6609 : : * Caller is responsible for doing ReleaseVariableStats() before exiting.
6610 : : */
6611 : : static void
6612 : 638341 : examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
6613 : : int indexcol, VariableStatData *vardata)
6614 : : {
6615 : : AttrNumber colnum;
6616 : : Oid relid;
6617 : :
6618 [ + + ]: 638341 : if (index->indexkeys[indexcol] != 0)
6619 : : {
6620 : : /* Simple variable --- look to stats for the underlying table */
6621 [ + - ]: 636508 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6622 : :
6623 : : Assert(rte->rtekind == RTE_RELATION);
6624 : 636508 : relid = rte->relid;
6625 : : Assert(relid != InvalidOid);
6626 : 636508 : colnum = index->indexkeys[indexcol];
6627 : 636508 : vardata->rel = index->rel;
6628 : :
6629 [ - + - - ]: 636508 : if (get_relation_stats_hook &&
6630 : 0 : (*get_relation_stats_hook) (root, rte, colnum, vardata))
6631 : : {
6632 : : /*
6633 : : * The hook took control of acquiring a stats tuple. If it did
6634 : : * supply a tuple, it'd better have supplied a freefunc.
6635 : : */
6636 [ # # ]: 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6637 [ # # ]: 0 : !vardata->freefunc)
6638 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
6639 : : }
6640 : : else
6641 : : {
6642 : 636508 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6643 : : ObjectIdGetDatum(relid),
6644 : : Int16GetDatum(colnum),
6645 : 636508 : BoolGetDatum(rte->inh));
6646 : 636508 : vardata->freefunc = ReleaseSysCache;
6647 : : }
6648 : : }
6649 : : else
6650 : : {
6651 : : /* Expression --- maybe there are stats for the index itself */
6652 : 1833 : relid = index->indexoid;
6653 : 1833 : colnum = indexcol + 1;
6654 : :
6655 [ - + - - ]: 1833 : if (get_index_stats_hook &&
6656 : 0 : (*get_index_stats_hook) (root, relid, colnum, vardata))
6657 : : {
6658 : : /*
6659 : : * The hook took control of acquiring a stats tuple. If it did
6660 : : * supply a tuple, it'd better have supplied a freefunc.
6661 : : */
6662 [ # # ]: 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6663 [ # # ]: 0 : !vardata->freefunc)
6664 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
6665 : : }
6666 : : else
6667 : : {
6668 : 1833 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6669 : : ObjectIdGetDatum(relid),
6670 : : Int16GetDatum(colnum),
6671 : : BoolGetDatum(false));
6672 : 1833 : vardata->freefunc = ReleaseSysCache;
6673 : : }
6674 : : }
6675 : 638341 : }
6676 : :
6677 : : /*
6678 : : * Check whether it is permitted to call func_oid passing some of the
6679 : : * pg_statistic data in vardata. We allow this if either of the following
6680 : : * conditions is met: (1) the user has SELECT privileges on the table or
6681 : : * column underlying the pg_statistic data and there are no securityQuals from
6682 : : * security barrier views or RLS policies, or (2) the function is marked
6683 : : * leakproof.
6684 : : */
6685 : : bool
6686 : 746356 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
6687 : : {
6688 [ + + ]: 746356 : if (vardata->acl_ok)
6689 : 744845 : return true; /* have SELECT privs and no securityQuals */
6690 : :
6691 [ - + ]: 1511 : if (!OidIsValid(func_oid))
6692 : 0 : return false;
6693 : :
6694 [ + + ]: 1511 : if (get_func_leakproof(func_oid))
6695 : 748 : return true;
6696 : :
6697 [ - + ]: 763 : ereport(DEBUG2,
6698 : : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
6699 : : get_func_name(func_oid))));
6700 : 763 : return false;
6701 : : }
6702 : :
6703 : : /*
6704 : : * get_variable_numdistinct
6705 : : * Estimate the number of distinct values of a variable.
6706 : : *
6707 : : * vardata: results of examine_variable
6708 : : * *isdefault: set to true if the result is a default rather than based on
6709 : : * anything meaningful.
6710 : : *
6711 : : * NB: be careful to produce a positive integral result, since callers may
6712 : : * compare the result to exact integer counts, or might divide by it.
6713 : : */
6714 : : double
6715 : 1258298 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
6716 : : {
6717 : : double stadistinct;
6718 : 1258298 : double stanullfrac = 0.0;
6719 : : double ntuples;
6720 : :
6721 : 1258298 : *isdefault = false;
6722 : :
6723 : : /*
6724 : : * Determine the stadistinct value to use. There are cases where we can
6725 : : * get an estimate even without a pg_statistic entry, or can get a better
6726 : : * value than is in pg_statistic. Grab stanullfrac too if we can find it
6727 : : * (otherwise, assume no nulls, for lack of any better idea).
6728 : : */
6729 [ + + ]: 1258298 : if (HeapTupleIsValid(vardata->statsTuple))
6730 : : {
6731 : : /* Use the pg_statistic entry */
6732 : : Form_pg_statistic stats;
6733 : :
6734 : 764483 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6735 : 764483 : stadistinct = stats->stadistinct;
6736 : 764483 : stanullfrac = stats->stanullfrac;
6737 : : }
6738 [ + + ]: 493815 : else if (vardata->vartype == BOOLOID)
6739 : : {
6740 : : /*
6741 : : * Special-case boolean columns: presumably, two distinct values.
6742 : : *
6743 : : * Are there any other datatypes we should wire in special estimates
6744 : : * for?
6745 : : */
6746 : 685 : stadistinct = 2.0;
6747 : : }
6748 [ + + + + ]: 493130 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6749 : : {
6750 : : /*
6751 : : * If the Var represents a column of a VALUES RTE, assume it's unique.
6752 : : * This could of course be very wrong, but it should tend to be true
6753 : : * in well-written queries. We could consider examining the VALUES'
6754 : : * contents to get some real statistics; but that only works if the
6755 : : * entries are all constants, and it would be pretty expensive anyway.
6756 : : */
6757 : 3163 : stadistinct = -1.0; /* unique (and all non null) */
6758 : : }
6759 : : else
6760 : : {
6761 : : /*
6762 : : * We don't keep statistics for system columns, but in some cases we
6763 : : * can infer distinctness anyway.
6764 : : */
6765 [ + + + + ]: 489967 : if (vardata->var && IsA(vardata->var, Var))
6766 : : {
6767 [ + + + ]: 459827 : switch (((Var *) vardata->var)->varattno)
6768 : : {
6769 : 858 : case SelfItemPointerAttributeNumber:
6770 : 858 : stadistinct = -1.0; /* unique (and all non null) */
6771 : 858 : break;
6772 : 14053 : case TableOidAttributeNumber:
6773 : 14053 : stadistinct = 1.0; /* only 1 value */
6774 : 14053 : break;
6775 : 444916 : default:
6776 : 444916 : stadistinct = 0.0; /* means "unknown" */
6777 : 444916 : break;
6778 : : }
6779 : : }
6780 : : else
6781 : 30140 : stadistinct = 0.0; /* means "unknown" */
6782 : :
6783 : : /*
6784 : : * XXX consider using estimate_num_groups on expressions?
6785 : : */
6786 : : }
6787 : :
6788 : : /*
6789 : : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6790 : : * variable, assume it is unique no matter what pg_statistic says; the
6791 : : * statistics could be out of date, or we might have found a partial
6792 : : * unique index that proves the var is unique for this query. However,
6793 : : * we'd better still believe the null-fraction statistic.
6794 : : */
6795 [ + + ]: 1258298 : if (vardata->isunique)
6796 : 301000 : stadistinct = -1.0 * (1.0 - stanullfrac);
6797 : :
6798 : : /*
6799 : : * If we had an absolute estimate, use that.
6800 : : */
6801 [ + + ]: 1258298 : if (stadistinct > 0.0)
6802 : 303970 : return clamp_row_est(stadistinct);
6803 : :
6804 : : /*
6805 : : * Otherwise we need to get the relation size; punt if not available.
6806 : : */
6807 [ + + ]: 954328 : if (vardata->rel == NULL)
6808 : : {
6809 : 589 : *isdefault = true;
6810 : 589 : return DEFAULT_NUM_DISTINCT;
6811 : : }
6812 : 953739 : ntuples = vardata->rel->tuples;
6813 [ + + ]: 953739 : if (ntuples <= 0.0)
6814 : : {
6815 : 99090 : *isdefault = true;
6816 : 99090 : return DEFAULT_NUM_DISTINCT;
6817 : : }
6818 : :
6819 : : /*
6820 : : * If we had a relative estimate, use that.
6821 : : */
6822 [ + + ]: 854649 : if (stadistinct < 0.0)
6823 : 532228 : return clamp_row_est(-stadistinct * ntuples);
6824 : :
6825 : : /*
6826 : : * With no data, estimate ndistinct = ntuples if the table is small, else
6827 : : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6828 : : * that the behavior isn't discontinuous.
6829 : : */
6830 [ + + ]: 322421 : if (ntuples < DEFAULT_NUM_DISTINCT)
6831 : 162752 : return clamp_row_est(ntuples);
6832 : :
6833 : 159669 : *isdefault = true;
6834 : 159669 : return DEFAULT_NUM_DISTINCT;
6835 : : }
6836 : :
6837 : : /*
6838 : : * get_variable_range
6839 : : * Estimate the minimum and maximum value of the specified variable.
6840 : : * If successful, store values in *min and *max, and return true.
6841 : : * If no data available, return false.
6842 : : *
6843 : : * sortop is the "<" comparison operator to use. This should generally
6844 : : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6845 : : * The collation must be specified too.
6846 : : */
6847 : : static bool
6848 : 165962 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6849 : : Oid sortop, Oid collation,
6850 : : Datum *min, Datum *max)
6851 : : {
6852 : 165962 : Datum tmin = 0;
6853 : 165962 : Datum tmax = 0;
6854 : 165962 : bool have_data = false;
6855 : : int16 typLen;
6856 : : bool typByVal;
6857 : : Oid opfuncoid;
6858 : : FmgrInfo opproc;
6859 : : AttStatsSlot sslot;
6860 : :
6861 : : /*
6862 : : * XXX It's very tempting to try to use the actual column min and max, if
6863 : : * we can get them relatively-cheaply with an index probe. However, since
6864 : : * this function is called many times during join planning, that could
6865 : : * have unpleasant effects on planning speed. Need more investigation
6866 : : * before enabling this.
6867 : : */
6868 : : #ifdef NOT_USED
6869 : : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6870 : : return true;
6871 : : #endif
6872 : :
6873 [ + + ]: 165962 : if (!HeapTupleIsValid(vardata->statsTuple))
6874 : : {
6875 : : /* no stats available, so default result */
6876 : 45915 : return false;
6877 : : }
6878 : :
6879 : : /*
6880 : : * If we can't apply the sortop to the stats data, just fail. In
6881 : : * principle, if there's a histogram and no MCVs, we could return the
6882 : : * histogram endpoints without ever applying the sortop ... but it's
6883 : : * probably not worth trying, because whatever the caller wants to do with
6884 : : * the endpoints would likely fail the security check too.
6885 : : */
6886 [ - + ]: 120047 : if (!statistic_proc_security_check(vardata,
6887 : 120047 : (opfuncoid = get_opcode(sortop))))
6888 : 0 : return false;
6889 : :
6890 : 120047 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6891 : :
6892 : 120047 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6893 : :
6894 : : /*
6895 : : * If there is a histogram with the ordering we want, grab the first and
6896 : : * last values.
6897 : : */
6898 [ + + ]: 120047 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6899 : : STATISTIC_KIND_HISTOGRAM, sortop,
6900 : : ATTSTATSSLOT_VALUES))
6901 : : {
6902 [ + - + - ]: 73620 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6903 : : {
6904 : 73620 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6905 : 73620 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6906 : 73620 : have_data = true;
6907 : : }
6908 : 73620 : free_attstatsslot(&sslot);
6909 : : }
6910 : :
6911 : : /*
6912 : : * Otherwise, if there is a histogram with some other ordering, scan it
6913 : : * and get the min and max values according to the ordering we want. This
6914 : : * of course may not find values that are really extremal according to our
6915 : : * ordering, but it beats ignoring available data.
6916 : : */
6917 [ + + - + ]: 166474 : if (!have_data &&
6918 : 46427 : get_attstatsslot(&sslot, vardata->statsTuple,
6919 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6920 : : ATTSTATSSLOT_VALUES))
6921 : : {
6922 : 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6923 : : collation, typLen, typByVal,
6924 : : &tmin, &tmax, &have_data);
6925 : 0 : free_attstatsslot(&sslot);
6926 : : }
6927 : :
6928 : : /*
6929 : : * If we have most-common-values info, look for extreme MCVs. This is
6930 : : * needed even if we also have a histogram, since the histogram excludes
6931 : : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6932 : : * be pretty wary of deciding that that is a full representation of the
6933 : : * data. Proceed only if the MCVs represent the whole table (to within
6934 : : * roundoff error).
6935 : : */
6936 [ + + ]: 120047 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6937 : : STATISTIC_KIND_MCV, InvalidOid,
6938 [ + + ]: 120047 : have_data ? ATTSTATSSLOT_VALUES :
6939 : : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6940 : : {
6941 : 67164 : bool use_mcvs = have_data;
6942 : :
6943 [ + + ]: 67164 : if (!have_data)
6944 : : {
6945 : 45483 : double sumcommon = 0.0;
6946 : : double nullfrac;
6947 : : int i;
6948 : :
6949 [ + + ]: 358379 : for (i = 0; i < sslot.nnumbers; i++)
6950 : 312896 : sumcommon += sslot.numbers[i];
6951 : 45483 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6952 [ + + ]: 45483 : if (sumcommon + nullfrac > 0.99999)
6953 : 45339 : use_mcvs = true;
6954 : : }
6955 : :
6956 [ + + ]: 67164 : if (use_mcvs)
6957 : 67020 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6958 : : collation, typLen, typByVal,
6959 : : &tmin, &tmax, &have_data);
6960 : 67164 : free_attstatsslot(&sslot);
6961 : : }
6962 : :
6963 : 120047 : *min = tmin;
6964 : 120047 : *max = tmax;
6965 : 120047 : return have_data;
6966 : : }
6967 : :
6968 : : /*
6969 : : * get_stats_slot_range: scan sslot for min/max values
6970 : : *
6971 : : * Subroutine for get_variable_range: update min/max/have_data according
6972 : : * to what we find in the statistics array.
6973 : : */
6974 : : static void
6975 : 67020 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6976 : : Oid collation, int16 typLen, bool typByVal,
6977 : : Datum *min, Datum *max, bool *p_have_data)
6978 : : {
6979 : 67020 : Datum tmin = *min;
6980 : 67020 : Datum tmax = *max;
6981 : 67020 : bool have_data = *p_have_data;
6982 : 67020 : bool found_tmin = false;
6983 : 67020 : bool found_tmax = false;
6984 : :
6985 : : /* Look up the comparison function, if we didn't already do so */
6986 [ + - ]: 67020 : if (opproc->fn_oid != opfuncoid)
6987 : 67020 : fmgr_info(opfuncoid, opproc);
6988 : :
6989 : : /* Scan all the slot's values */
6990 [ + + ]: 1512749 : for (int i = 0; i < sslot->nvalues; i++)
6991 : : {
6992 [ + + ]: 1445729 : if (!have_data)
6993 : : {
6994 : 45339 : tmin = tmax = sslot->values[i];
6995 : 45339 : found_tmin = found_tmax = true;
6996 : 45339 : *p_have_data = have_data = true;
6997 : 45339 : continue;
6998 : : }
6999 [ + + ]: 1400390 : if (DatumGetBool(FunctionCall2Coll(opproc,
7000 : : collation,
7001 : 1400390 : sslot->values[i], tmin)))
7002 : : {
7003 : 36811 : tmin = sslot->values[i];
7004 : 36811 : found_tmin = true;
7005 : : }
7006 [ + + ]: 1400390 : if (DatumGetBool(FunctionCall2Coll(opproc,
7007 : : collation,
7008 : 1400390 : tmax, sslot->values[i])))
7009 : : {
7010 : 185733 : tmax = sslot->values[i];
7011 : 185733 : found_tmax = true;
7012 : : }
7013 : : }
7014 : :
7015 : : /*
7016 : : * Copy the slot's values, if we found new extreme values.
7017 : : */
7018 [ + + ]: 67020 : if (found_tmin)
7019 : 58956 : *min = datumCopy(tmin, typByVal, typLen);
7020 [ + + ]: 67020 : if (found_tmax)
7021 : 47232 : *max = datumCopy(tmax, typByVal, typLen);
7022 : 67020 : }
7023 : :
7024 : :
7025 : : /*
7026 : : * get_actual_variable_range
7027 : : * Attempt to identify the current *actual* minimum and/or maximum
7028 : : * of the specified variable, by looking for a suitable btree index
7029 : : * and fetching its low and/or high values.
7030 : : * If successful, store values in *min and *max, and return true.
7031 : : * (Either pointer can be NULL if that endpoint isn't needed.)
7032 : : * If unsuccessful, return false.
7033 : : *
7034 : : * sortop is the "<" comparison operator to use.
7035 : : * collation is the required collation.
7036 : : */
7037 : : static bool
7038 : 113440 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
7039 : : Oid sortop, Oid collation,
7040 : : Datum *min, Datum *max)
7041 : : {
7042 : 113440 : bool have_data = false;
7043 : 113440 : RelOptInfo *rel = vardata->rel;
7044 : : RangeTblEntry *rte;
7045 : : ListCell *lc;
7046 : :
7047 : : /* No hope if no relation or it doesn't have indexes */
7048 [ + - + + ]: 113440 : if (rel == NULL || rel->indexlist == NIL)
7049 : 8979 : return false;
7050 : : /* If it has indexes it must be a plain relation */
7051 : 104461 : rte = root->simple_rte_array[rel->relid];
7052 : : Assert(rte->rtekind == RTE_RELATION);
7053 : :
7054 : : /* ignore partitioned tables. Any indexes here are not real indexes */
7055 [ + + ]: 104461 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
7056 : 560 : return false;
7057 : :
7058 : : /* Search through the indexes to see if any match our problem */
7059 [ + - + + : 201957 : foreach(lc, rel->indexlist)
+ + ]
7060 : : {
7061 : 173926 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
7062 : : ScanDirection indexscandir;
7063 : : StrategyNumber strategy;
7064 : :
7065 : : /* Ignore non-ordering indexes */
7066 [ + + ]: 173926 : if (index->sortopfamily == NULL)
7067 : 3 : continue;
7068 : :
7069 : : /*
7070 : : * Ignore partial indexes --- we only want stats that cover the entire
7071 : : * relation.
7072 : : */
7073 [ + + ]: 173923 : if (index->indpred != NIL)
7074 : 240 : continue;
7075 : :
7076 : : /*
7077 : : * The index list might include hypothetical indexes inserted by a
7078 : : * get_relation_info hook --- don't try to access them.
7079 : : */
7080 [ - + ]: 173683 : if (index->hypothetical)
7081 : 0 : continue;
7082 : :
7083 : : /*
7084 : : * get_actual_variable_endpoint uses the index-only-scan machinery, so
7085 : : * ignore indexes that can't use it on their first column.
7086 : : */
7087 [ - + ]: 173683 : if (!index->canreturn[0])
7088 : 0 : continue;
7089 : :
7090 : : /*
7091 : : * The first index column must match the desired variable, sortop, and
7092 : : * collation --- but we can use a descending-order index.
7093 : : */
7094 [ + + ]: 173683 : if (collation != index->indexcollations[0])
7095 : 23826 : continue; /* test first 'cause it's cheapest */
7096 [ + + ]: 149857 : if (!match_index_to_operand(vardata->var, 0, index))
7097 : 73987 : continue;
7098 : 75870 : strategy = get_op_opfamily_strategy(sortop, index->sortopfamily[0]);
7099 [ + - - ]: 75870 : switch (IndexAmTranslateStrategy(strategy, index->relam, index->sortopfamily[0], true))
7100 : : {
7101 : 75870 : case COMPARE_LT:
7102 [ - + ]: 75870 : if (index->reverse_sort[0])
7103 : 0 : indexscandir = BackwardScanDirection;
7104 : : else
7105 : 75870 : indexscandir = ForwardScanDirection;
7106 : 75870 : break;
7107 : 0 : case COMPARE_GT:
7108 [ # # ]: 0 : if (index->reverse_sort[0])
7109 : 0 : indexscandir = ForwardScanDirection;
7110 : : else
7111 : 0 : indexscandir = BackwardScanDirection;
7112 : 0 : break;
7113 : 0 : default:
7114 : : /* index doesn't match the sortop */
7115 : 0 : continue;
7116 : : }
7117 : :
7118 : : /*
7119 : : * Found a suitable index to extract data from. Set up some data that
7120 : : * can be used by both invocations of get_actual_variable_endpoint.
7121 : : */
7122 : : {
7123 : : MemoryContext tmpcontext;
7124 : : MemoryContext oldcontext;
7125 : : Relation heapRel;
7126 : : Relation indexRel;
7127 : : TupleTableSlot *slot;
7128 : : int16 typLen;
7129 : : bool typByVal;
7130 : : ScanKeyData scankeys[1];
7131 : :
7132 : : /* Make sure any cruft gets recycled when we're done */
7133 : 75870 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
7134 : : "get_actual_variable_range workspace",
7135 : : ALLOCSET_DEFAULT_SIZES);
7136 : 75870 : oldcontext = MemoryContextSwitchTo(tmpcontext);
7137 : :
7138 : : /*
7139 : : * Open the table and index so we can read from them. We should
7140 : : * already have some type of lock on each.
7141 : : */
7142 : 75870 : heapRel = table_open(rte->relid, NoLock);
7143 : 75870 : indexRel = index_open(index->indexoid, NoLock);
7144 : :
7145 : : /* build some stuff needed for indexscan execution */
7146 : 75870 : slot = table_slot_create(heapRel, NULL);
7147 : 75870 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
7148 : :
7149 : : /* set up an IS NOT NULL scan key so that we ignore nulls */
7150 : 75870 : ScanKeyEntryInitialize(&scankeys[0],
7151 : : SK_ISNULL | SK_SEARCHNOTNULL,
7152 : : 1, /* index col to scan */
7153 : : InvalidStrategy, /* no strategy */
7154 : : InvalidOid, /* no strategy subtype */
7155 : : InvalidOid, /* no collation */
7156 : : InvalidOid, /* no reg proc for this */
7157 : : (Datum) 0); /* constant */
7158 : :
7159 : : /* If min is requested ... */
7160 [ + + ]: 75870 : if (min)
7161 : : {
7162 : 42780 : have_data = get_actual_variable_endpoint(heapRel,
7163 : : indexRel,
7164 : : indexscandir,
7165 : : scankeys,
7166 : : typLen,
7167 : : typByVal,
7168 : : slot,
7169 : : oldcontext,
7170 : : min);
7171 : : }
7172 : : else
7173 : : {
7174 : : /* If min not requested, still want to fetch max */
7175 : 33090 : have_data = true;
7176 : : }
7177 : :
7178 : : /* If max is requested, and we didn't already fail ... */
7179 [ + + + - ]: 75870 : if (max && have_data)
7180 : : {
7181 : : /* scan in the opposite direction; all else is the same */
7182 : 34120 : have_data = get_actual_variable_endpoint(heapRel,
7183 : : indexRel,
7184 : 34120 : -indexscandir,
7185 : : scankeys,
7186 : : typLen,
7187 : : typByVal,
7188 : : slot,
7189 : : oldcontext,
7190 : : max);
7191 : : }
7192 : :
7193 : : /* Clean everything up */
7194 : 75870 : ExecDropSingleTupleTableSlot(slot);
7195 : :
7196 : 75870 : index_close(indexRel, NoLock);
7197 : 75870 : table_close(heapRel, NoLock);
7198 : :
7199 : 75870 : MemoryContextSwitchTo(oldcontext);
7200 : 75870 : MemoryContextDelete(tmpcontext);
7201 : :
7202 : : /* And we're done */
7203 : 75870 : break;
7204 : : }
7205 : : }
7206 : :
7207 : 103901 : return have_data;
7208 : : }
7209 : :
7210 : : /*
7211 : : * Get one endpoint datum (min or max depending on indexscandir) from the
7212 : : * specified index. Return true if successful, false if not.
7213 : : * On success, endpoint value is stored to *endpointDatum (and copied into
7214 : : * outercontext).
7215 : : *
7216 : : * scankeys is a 1-element scankey array set up to reject nulls.
7217 : : * typLen/typByVal describe the datatype of the index's first column.
7218 : : * tableslot is a slot suitable to hold table tuples, in case we need
7219 : : * to probe the heap.
7220 : : * (We could compute these values locally, but that would mean computing them
7221 : : * twice when get_actual_variable_range needs both the min and the max.)
7222 : : *
7223 : : * Failure occurs either when the index is empty, or we decide that it's
7224 : : * taking too long to find a suitable tuple.
7225 : : */
7226 : : static bool
7227 : 76900 : get_actual_variable_endpoint(Relation heapRel,
7228 : : Relation indexRel,
7229 : : ScanDirection indexscandir,
7230 : : ScanKey scankeys,
7231 : : int16 typLen,
7232 : : bool typByVal,
7233 : : TupleTableSlot *tableslot,
7234 : : MemoryContext outercontext,
7235 : : Datum *endpointDatum)
7236 : : {
7237 : 76900 : bool have_data = false;
7238 : : SnapshotData SnapshotNonVacuumable;
7239 : : IndexScanDesc index_scan;
7240 : 76900 : Buffer vmbuffer = InvalidBuffer;
7241 : 76900 : BlockNumber last_heap_block = InvalidBlockNumber;
7242 : 76900 : int n_visited_heap_pages = 0;
7243 : : ItemPointer tid;
7244 : : Datum values[INDEX_MAX_KEYS];
7245 : : bool isnull[INDEX_MAX_KEYS];
7246 : : MemoryContext oldcontext;
7247 : :
7248 : : /*
7249 : : * We use the index-only-scan machinery for this. With mostly-static
7250 : : * tables that's a win because it avoids a heap visit. It's also a win
7251 : : * for dynamic data, but the reason is less obvious; read on for details.
7252 : : *
7253 : : * In principle, we should scan the index with our current active
7254 : : * snapshot, which is the best approximation we've got to what the query
7255 : : * will see when executed. But that won't be exact if a new snap is taken
7256 : : * before running the query, and it can be very expensive if a lot of
7257 : : * recently-dead or uncommitted rows exist at the beginning or end of the
7258 : : * index (because we'll laboriously fetch each one and reject it).
7259 : : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
7260 : : * and uncommitted rows as well as normal visible rows. On the other
7261 : : * hand, it will reject known-dead rows, and thus not give a bogus answer
7262 : : * when the extreme value has been deleted (unless the deletion was quite
7263 : : * recent); that case motivates not using SnapshotAny here.
7264 : : *
7265 : : * A crucial point here is that SnapshotNonVacuumable, with
7266 : : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
7267 : : * condition that the indexscan will use to decide that index entries are
7268 : : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
7269 : : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
7270 : : * have to continue scanning past it, we know that the indexscan will mark
7271 : : * that index entry killed. That means that the next
7272 : : * get_actual_variable_endpoint() call will not have to re-consider that
7273 : : * index entry. In this way we avoid repetitive work when this function
7274 : : * is used a lot during planning.
7275 : : *
7276 : : * But using SnapshotNonVacuumable creates a hazard of its own. In a
7277 : : * recently-created index, some index entries may point at "broken" HOT
7278 : : * chains in which not all the tuple versions contain data matching the
7279 : : * index entry. The live tuple version(s) certainly do match the index,
7280 : : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
7281 : : * don't match. Hence, if we took data from the selected heap tuple, we
7282 : : * might get a bogus answer that's not close to the index extremal value,
7283 : : * or could even be NULL. We avoid this hazard because we take the data
7284 : : * from the index entry not the heap.
7285 : : *
7286 : : * Despite all this care, there are situations where we might find many
7287 : : * non-visible tuples near the end of the index. We don't want to expend
7288 : : * a huge amount of time here, so we give up once we've read too many heap
7289 : : * pages. When we fail for that reason, the caller will end up using
7290 : : * whatever extremal value is recorded in pg_statistic.
7291 : : */
7292 : 76900 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
7293 : : GlobalVisTestFor(heapRel));
7294 : :
7295 : 76900 : index_scan = index_beginscan(heapRel, indexRel,
7296 : : &SnapshotNonVacuumable, NULL,
7297 : : 1, 0,
7298 : : SO_NONE);
7299 : : /* Set it up for index-only scan */
7300 : 76900 : index_scan->xs_want_itup = true;
7301 : 76900 : index_rescan(index_scan, scankeys, 1, NULL, 0);
7302 : :
7303 : : /* Fetch first/next tuple in specified direction */
7304 [ + - ]: 91499 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
7305 : : {
7306 : 91499 : BlockNumber block = ItemPointerGetBlockNumber(tid);
7307 : :
7308 [ + + ]: 91499 : if (!VM_ALL_VISIBLE(heapRel,
7309 : : block,
7310 : : &vmbuffer))
7311 : : {
7312 : : /* Rats, we have to visit the heap to check visibility */
7313 [ + + ]: 63417 : if (!index_fetch_heap(index_scan, tableslot))
7314 : : {
7315 : : /*
7316 : : * No visible tuple for this index entry, so we need to
7317 : : * advance to the next entry. Before doing so, count heap
7318 : : * page fetches and give up if we've done too many.
7319 : : *
7320 : : * We don't charge a page fetch if this is the same heap page
7321 : : * as the previous tuple. This is on the conservative side,
7322 : : * since other recently-accessed pages are probably still in
7323 : : * buffers too; but it's good enough for this heuristic.
7324 : : */
7325 : : #define VISITED_PAGES_LIMIT 100
7326 : :
7327 [ + + ]: 14599 : if (block != last_heap_block)
7328 : : {
7329 : 1564 : last_heap_block = block;
7330 : 1564 : n_visited_heap_pages++;
7331 [ - + ]: 1564 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
7332 : 0 : break;
7333 : : }
7334 : :
7335 : 14599 : continue; /* no visible tuple, try next index entry */
7336 : : }
7337 : :
7338 : : /* We don't actually need the heap tuple for anything */
7339 : 48818 : ExecClearTuple(tableslot);
7340 : :
7341 : : /*
7342 : : * We don't care whether there's more than one visible tuple in
7343 : : * the HOT chain; if any are visible, that's good enough.
7344 : : */
7345 : : }
7346 : :
7347 : : /*
7348 : : * We expect that the index will return data in IndexTuple not
7349 : : * HeapTuple format.
7350 : : */
7351 [ - + ]: 76900 : if (!index_scan->xs_itup)
7352 [ # # ]: 0 : elog(ERROR, "no data returned for index-only scan");
7353 : :
7354 : : /*
7355 : : * We do not yet support recheck here.
7356 : : */
7357 [ - + ]: 76900 : if (index_scan->xs_recheck)
7358 : 0 : break;
7359 : :
7360 : : /* OK to deconstruct the index tuple */
7361 : 76900 : index_deform_tuple(index_scan->xs_itup,
7362 : : index_scan->xs_itupdesc,
7363 : : values, isnull);
7364 : :
7365 : : /* Shouldn't have got a null, but be careful */
7366 [ - + ]: 76900 : if (isnull[0])
7367 [ # # ]: 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
7368 : : RelationGetRelationName(indexRel));
7369 : :
7370 : : /* Copy the index column value out to caller's context */
7371 : 76900 : oldcontext = MemoryContextSwitchTo(outercontext);
7372 : 76900 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
7373 : 76900 : MemoryContextSwitchTo(oldcontext);
7374 : 76900 : have_data = true;
7375 : 76900 : break;
7376 : : }
7377 : :
7378 [ + + ]: 76900 : if (vmbuffer != InvalidBuffer)
7379 : 69971 : ReleaseBuffer(vmbuffer);
7380 : 76900 : index_endscan(index_scan);
7381 : :
7382 : 76900 : return have_data;
7383 : : }
7384 : :
7385 : : /*
7386 : : * find_join_input_rel
7387 : : * Look up the input relation for a join.
7388 : : *
7389 : : * We assume that the input relation's RelOptInfo must have been constructed
7390 : : * already.
7391 : : */
7392 : : static RelOptInfo *
7393 : 11209 : find_join_input_rel(PlannerInfo *root, Relids relids)
7394 : : {
7395 : 11209 : RelOptInfo *rel = NULL;
7396 : :
7397 [ + - ]: 11209 : if (!bms_is_empty(relids))
7398 : : {
7399 : : int relid;
7400 : :
7401 [ + + ]: 11209 : if (bms_get_singleton_member(relids, &relid))
7402 : 10933 : rel = find_base_rel(root, relid);
7403 : : else
7404 : 276 : rel = find_join_rel(root, relids);
7405 : : }
7406 : :
7407 [ - + ]: 11209 : if (rel == NULL)
7408 [ # # ]: 0 : elog(ERROR, "could not find RelOptInfo for given relids");
7409 : :
7410 : 11209 : return rel;
7411 : : }
7412 : :
7413 : :
7414 : : /*-------------------------------------------------------------------------
7415 : : *
7416 : : * Index cost estimation functions
7417 : : *
7418 : : *-------------------------------------------------------------------------
7419 : : */
7420 : :
7421 : : /*
7422 : : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
7423 : : */
7424 : : List *
7425 : 653877 : get_quals_from_indexclauses(List *indexclauses)
7426 : : {
7427 : 653877 : List *result = NIL;
7428 : : ListCell *lc;
7429 : :
7430 [ + + + + : 1138772 : foreach(lc, indexclauses)
+ + ]
7431 : : {
7432 : 484895 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7433 : : ListCell *lc2;
7434 : :
7435 [ + - + + : 972219 : foreach(lc2, iclause->indexquals)
+ + ]
7436 : : {
7437 : 487324 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7438 : :
7439 : 487324 : result = lappend(result, rinfo);
7440 : : }
7441 : : }
7442 : 653877 : return result;
7443 : : }
7444 : :
7445 : : /*
7446 : : * Compute the total evaluation cost of the comparison operands in a list
7447 : : * of index qual expressions. Since we know these will be evaluated just
7448 : : * once per scan, there's no need to distinguish startup from per-row cost.
7449 : : *
7450 : : * This can be used either on the result of get_quals_from_indexclauses(),
7451 : : * or directly on an indexorderbys list. In both cases, we expect that the
7452 : : * index key expression is on the left side of binary clauses.
7453 : : */
7454 : : Cost
7455 : 1297254 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
7456 : : {
7457 : 1297254 : Cost qual_arg_cost = 0;
7458 : : ListCell *lc;
7459 : :
7460 [ + + + + : 1784927 : foreach(lc, indexquals)
+ + ]
7461 : : {
7462 : 487673 : Expr *clause = (Expr *) lfirst(lc);
7463 : : Node *other_operand;
7464 : : QualCost index_qual_cost;
7465 : :
7466 : : /*
7467 : : * Index quals will have RestrictInfos, indexorderbys won't. Look
7468 : : * through RestrictInfo if present.
7469 : : */
7470 [ + + ]: 487673 : if (IsA(clause, RestrictInfo))
7471 : 487314 : clause = ((RestrictInfo *) clause)->clause;
7472 : :
7473 [ + + ]: 487673 : if (IsA(clause, OpExpr))
7474 : : {
7475 : 472291 : OpExpr *op = (OpExpr *) clause;
7476 : :
7477 : 472291 : other_operand = (Node *) lsecond(op->args);
7478 : : }
7479 [ + + ]: 15382 : else if (IsA(clause, RowCompareExpr))
7480 : : {
7481 : 330 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7482 : :
7483 : 330 : other_operand = (Node *) rc->rargs;
7484 : : }
7485 [ + + ]: 15052 : else if (IsA(clause, ScalarArrayOpExpr))
7486 : : {
7487 : 12613 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7488 : :
7489 : 12613 : other_operand = (Node *) lsecond(saop->args);
7490 : : }
7491 [ + - ]: 2439 : else if (IsA(clause, NullTest))
7492 : : {
7493 : 2439 : other_operand = NULL;
7494 : : }
7495 : : else
7496 : : {
7497 [ # # ]: 0 : elog(ERROR, "unsupported indexqual type: %d",
7498 : : (int) nodeTag(clause));
7499 : : other_operand = NULL; /* keep compiler quiet */
7500 : : }
7501 : :
7502 : 487673 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
7503 : 487673 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
7504 : : }
7505 : 1297254 : return qual_arg_cost;
7506 : : }
7507 : :
7508 : : /*
7509 : : * Compute generic index access cost estimates.
7510 : : *
7511 : : * See struct GenericCosts in selfuncs.h for more info.
7512 : : */
7513 : : void
7514 : 643387 : genericcostestimate(PlannerInfo *root,
7515 : : IndexPath *path,
7516 : : double loop_count,
7517 : : GenericCosts *costs)
7518 : : {
7519 : 643387 : IndexOptInfo *index = path->indexinfo;
7520 : 643387 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7521 : 643387 : List *indexOrderBys = path->indexorderbys;
7522 : : Cost indexStartupCost;
7523 : : Cost indexTotalCost;
7524 : : Selectivity indexSelectivity;
7525 : : double indexCorrelation;
7526 : : double numIndexPages;
7527 : : double numIndexTuples;
7528 : : double spc_random_page_cost;
7529 : : double num_sa_scans;
7530 : : double num_outer_scans;
7531 : : double num_scans;
7532 : : double qual_op_cost;
7533 : : double qual_arg_cost;
7534 : : List *selectivityQuals;
7535 : : ListCell *l;
7536 : :
7537 : : /*
7538 : : * If the index is partial, AND the index predicate with the explicitly
7539 : : * given indexquals to produce a more accurate idea of the index
7540 : : * selectivity.
7541 : : */
7542 : 643387 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7543 : :
7544 : : /*
7545 : : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
7546 : : * just assume that the number of index descents is the number of distinct
7547 : : * combinations of array elements from all of the scan's SAOP clauses.
7548 : : */
7549 : 643387 : num_sa_scans = costs->num_sa_scans;
7550 [ + + ]: 643387 : if (num_sa_scans < 1)
7551 : : {
7552 : 6974 : num_sa_scans = 1;
7553 [ + + + + : 15252 : foreach(l, indexQuals)
+ + ]
7554 : : {
7555 : 8278 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
7556 : :
7557 [ + + ]: 8278 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
7558 : : {
7559 : 46 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
7560 : 46 : double alength = estimate_array_length(root, lsecond(saop->args));
7561 : :
7562 [ + - ]: 46 : if (alength > 1)
7563 : 46 : num_sa_scans *= alength;
7564 : : }
7565 : : }
7566 : : }
7567 : :
7568 : : /* Estimate the fraction of main-table tuples that will be visited */
7569 : 643387 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7570 : 643387 : index->rel->relid,
7571 : : JOIN_INNER,
7572 : : NULL);
7573 : :
7574 : : /*
7575 : : * If caller didn't give us an estimate, estimate the number of index
7576 : : * tuples that will be visited. We do it in this rather peculiar-looking
7577 : : * way in order to get the right answer for partial indexes.
7578 : : */
7579 : 643387 : numIndexTuples = costs->numIndexTuples;
7580 [ + + ]: 643387 : if (numIndexTuples <= 0.0)
7581 : : {
7582 : 77193 : numIndexTuples = indexSelectivity * index->rel->tuples;
7583 : :
7584 : : /*
7585 : : * The above calculation counts all the tuples visited across all
7586 : : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
7587 : : * average per-indexscan number, so adjust. This is a handy place to
7588 : : * round to integer, too. (If caller supplied tuple estimate, it's
7589 : : * responsible for handling these considerations.)
7590 : : */
7591 : 77193 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7592 : : }
7593 : :
7594 : : /*
7595 : : * We can bound the number of tuples by the index size in any case. Also,
7596 : : * always estimate at least one tuple is touched, even when
7597 : : * indexSelectivity estimate is tiny.
7598 : : */
7599 [ + + ]: 643387 : if (numIndexTuples > index->tuples)
7600 : 7382 : numIndexTuples = index->tuples;
7601 [ + + ]: 643387 : if (numIndexTuples < 1.0)
7602 : 79739 : numIndexTuples = 1.0;
7603 : :
7604 : : /*
7605 : : * Estimate the number of index pages that will be retrieved.
7606 : : *
7607 : : * We use the simplistic method of taking a pro-rata fraction of the total
7608 : : * number of index leaf pages. We disregard any overhead such as index
7609 : : * metapages or upper tree levels.
7610 : : *
7611 : : * In practice access to upper index levels is often nearly free because
7612 : : * those tend to stay in cache under load; moreover, the cost involved is
7613 : : * highly dependent on index type. We therefore ignore such costs here
7614 : : * and leave it to the caller to add a suitable charge if needed.
7615 : : */
7616 [ + + + + ]: 643387 : if (index->pages > costs->numNonLeafPages && index->tuples > 1)
7617 : 583740 : numIndexPages =
7618 : 583740 : ceil(numIndexTuples * (index->pages - costs->numNonLeafPages)
7619 : 583740 : / index->tuples);
7620 : : else
7621 : 59647 : numIndexPages = 1.0;
7622 : :
7623 : : /* fetch estimated page cost for tablespace containing index */
7624 : 643387 : get_tablespace_page_costs(index->reltablespace,
7625 : : &spc_random_page_cost,
7626 : : NULL);
7627 : :
7628 : : /*
7629 : : * Now compute the disk access costs.
7630 : : *
7631 : : * The above calculations are all per-index-scan. However, if we are in a
7632 : : * nestloop inner scan, we can expect the scan to be repeated (with
7633 : : * different search keys) for each row of the outer relation. Likewise,
7634 : : * ScalarArrayOpExpr quals result in multiple index scans. This creates
7635 : : * the potential for cache effects to reduce the number of disk page
7636 : : * fetches needed. We want to estimate the average per-scan I/O cost in
7637 : : * the presence of caching.
7638 : : *
7639 : : * We use the Mackert-Lohman formula (see costsize.c for details) to
7640 : : * estimate the total number of page fetches that occur. While this
7641 : : * wasn't what it was designed for, it seems a reasonable model anyway.
7642 : : * Note that we are counting pages not tuples anymore, so we take N = T =
7643 : : * index size, as if there were one "tuple" per page.
7644 : : */
7645 : 643387 : num_outer_scans = loop_count;
7646 : 643387 : num_scans = num_sa_scans * num_outer_scans;
7647 : :
7648 [ + + ]: 643387 : if (num_scans > 1)
7649 : : {
7650 : : double pages_fetched;
7651 : :
7652 : : /* total page fetches ignoring cache effects */
7653 : 77015 : pages_fetched = numIndexPages * num_scans;
7654 : :
7655 : : /* use Mackert and Lohman formula to adjust for cache effects */
7656 : 77015 : pages_fetched = index_pages_fetched(pages_fetched,
7657 : : index->pages,
7658 : 77015 : (double) index->pages,
7659 : : root);
7660 : :
7661 : : /*
7662 : : * Now compute the total disk access cost, and then report a pro-rated
7663 : : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
7664 : : * since that's internal to the indexscan.)
7665 : : */
7666 : 77015 : indexTotalCost = (pages_fetched * spc_random_page_cost)
7667 : : / num_outer_scans;
7668 : : }
7669 : : else
7670 : : {
7671 : : /*
7672 : : * For a single index scan, we just charge spc_random_page_cost per
7673 : : * page touched.
7674 : : */
7675 : 566372 : indexTotalCost = numIndexPages * spc_random_page_cost;
7676 : : }
7677 : :
7678 : : /*
7679 : : * CPU cost: any complex expressions in the indexquals will need to be
7680 : : * evaluated once at the start of the scan to reduce them to runtime keys
7681 : : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
7682 : : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
7683 : : * indexqual operator. Because we have numIndexTuples as a per-scan
7684 : : * number, we have to multiply by num_sa_scans to get the correct result
7685 : : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
7686 : : * ORDER BY expressions.
7687 : : *
7688 : : * Note: this neglects the possible costs of rechecking lossy operators.
7689 : : * Detecting that that might be needed seems more expensive than it's
7690 : : * worth, though, considering all the other inaccuracies here ...
7691 : : */
7692 : 643387 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
7693 : 643387 : index_other_operands_eval_cost(root, indexOrderBys);
7694 : 643387 : qual_op_cost = cpu_operator_cost *
7695 : 643387 : (list_length(indexQuals) + list_length(indexOrderBys));
7696 : :
7697 : 643387 : indexStartupCost = qual_arg_cost;
7698 : 643387 : indexTotalCost += qual_arg_cost;
7699 : 643387 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
7700 : :
7701 : : /*
7702 : : * Generic assumption about index correlation: there isn't any.
7703 : : */
7704 : 643387 : indexCorrelation = 0.0;
7705 : :
7706 : : /*
7707 : : * Return everything to caller.
7708 : : */
7709 : 643387 : costs->indexStartupCost = indexStartupCost;
7710 : 643387 : costs->indexTotalCost = indexTotalCost;
7711 : 643387 : costs->indexSelectivity = indexSelectivity;
7712 : 643387 : costs->indexCorrelation = indexCorrelation;
7713 : 643387 : costs->numIndexPages = numIndexPages;
7714 : 643387 : costs->numIndexTuples = numIndexTuples;
7715 : 643387 : costs->spc_random_page_cost = spc_random_page_cost;
7716 : 643387 : costs->num_sa_scans = num_sa_scans;
7717 : 643387 : }
7718 : :
7719 : : /*
7720 : : * If the index is partial, add its predicate to the given qual list.
7721 : : *
7722 : : * ANDing the index predicate with the explicitly given indexquals produces
7723 : : * a more accurate idea of the index's selectivity. However, we need to be
7724 : : * careful not to insert redundant clauses, because clauselist_selectivity()
7725 : : * is easily fooled into computing a too-low selectivity estimate. Our
7726 : : * approach is to add only the predicate clause(s) that cannot be proven to
7727 : : * be implied by the given indexquals. This successfully handles cases such
7728 : : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
7729 : : * There are many other cases where we won't detect redundancy, leading to a
7730 : : * too-low selectivity estimate, which will bias the system in favor of using
7731 : : * partial indexes where possible. That is not necessarily bad though.
7732 : : *
7733 : : * Note that indexQuals contains RestrictInfo nodes while the indpred
7734 : : * does not, so the output list will be mixed. This is OK for both
7735 : : * predicate_implied_by() and clauselist_selectivity(), but might be
7736 : : * problematic if the result were passed to other things.
7737 : : */
7738 : : List *
7739 : 1104207 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
7740 : : {
7741 : 1104207 : List *predExtraQuals = NIL;
7742 : : ListCell *lc;
7743 : :
7744 [ + + ]: 1104207 : if (index->indpred == NIL)
7745 : 1102671 : return indexQuals;
7746 : :
7747 [ + - + + : 3082 : foreach(lc, index->indpred)
+ + ]
7748 : : {
7749 : 1546 : Node *predQual = (Node *) lfirst(lc);
7750 : 1546 : List *oneQual = list_make1(predQual);
7751 : :
7752 [ + + ]: 1546 : if (!predicate_implied_by(oneQual, indexQuals, false))
7753 : 1370 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7754 : : }
7755 : 1536 : return list_concat(predExtraQuals, indexQuals);
7756 : : }
7757 : :
7758 : : /*
7759 : : * Estimate correlation of btree index's first column.
7760 : : *
7761 : : * If we can get an estimate of the first column's ordering correlation C
7762 : : * from pg_statistic, estimate the index correlation as C for a single-column
7763 : : * index, or C * 0.75 for multiple columns. The idea here is that multiple
7764 : : * columns dilute the importance of the first column's ordering, but don't
7765 : : * negate it entirely.
7766 : : *
7767 : : * We already filled in the stats tuple for *vardata when called.
7768 : : */
7769 : : static double
7770 : 413277 : btcost_correlation(IndexOptInfo *index, VariableStatData *vardata)
7771 : : {
7772 : : Oid sortop;
7773 : : AttStatsSlot sslot;
7774 : 413277 : double indexCorrelation = 0;
7775 : :
7776 : : Assert(HeapTupleIsValid(vardata->statsTuple));
7777 : :
7778 : 413277 : sortop = get_opfamily_member(index->opfamily[0],
7779 : 413277 : index->opcintype[0],
7780 : 413277 : index->opcintype[0],
7781 : : BTLessStrategyNumber);
7782 [ + - + + ]: 826554 : if (OidIsValid(sortop) &&
7783 : 413277 : get_attstatsslot(&sslot, vardata->statsTuple,
7784 : : STATISTIC_KIND_CORRELATION, sortop,
7785 : : ATTSTATSSLOT_NUMBERS))
7786 : : {
7787 : : double varCorrelation;
7788 : :
7789 : : Assert(sslot.nnumbers == 1);
7790 : 408448 : varCorrelation = sslot.numbers[0];
7791 : :
7792 [ - + ]: 408448 : if (index->reverse_sort[0])
7793 : 0 : varCorrelation = -varCorrelation;
7794 : :
7795 [ + + ]: 408448 : if (index->nkeycolumns > 1)
7796 : 144117 : indexCorrelation = varCorrelation * 0.75;
7797 : : else
7798 : 264331 : indexCorrelation = varCorrelation;
7799 : :
7800 : 408448 : free_attstatsslot(&sslot);
7801 : : }
7802 : :
7803 : 413277 : return indexCorrelation;
7804 : : }
7805 : :
7806 : : void
7807 : 636413 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7808 : : Cost *indexStartupCost, Cost *indexTotalCost,
7809 : : Selectivity *indexSelectivity, double *indexCorrelation,
7810 : : double *indexPages)
7811 : : {
7812 : 636413 : IndexOptInfo *index = path->indexinfo;
7813 : 636413 : GenericCosts costs = {0};
7814 : 636413 : VariableStatData vardata = {0};
7815 : : double numIndexTuples;
7816 : : Cost descentCost;
7817 : : List *indexBoundQuals;
7818 : : List *indexSkipQuals;
7819 : : int indexcol;
7820 : : bool eqQualHere;
7821 : : bool found_row_compare;
7822 : : bool found_array;
7823 : : bool found_is_null_op;
7824 : 636413 : bool have_correlation = false;
7825 : : double num_sa_scans;
7826 : 636413 : double correlation = 0.0;
7827 : : ListCell *lc;
7828 : :
7829 : : /*
7830 : : * For a btree scan, only leading '=' quals plus inequality quals for the
7831 : : * immediately next attribute contribute to index selectivity (these are
7832 : : * the "boundary quals" that determine the starting and stopping points of
7833 : : * the index scan). Additional quals can suppress visits to the heap, so
7834 : : * it's OK to count them in indexSelectivity, but they should not count
7835 : : * for estimating numIndexTuples. So we must examine the given indexquals
7836 : : * to find out which ones count as boundary quals. We rely on the
7837 : : * knowledge that they are given in index column order. Note that nbtree
7838 : : * preprocessing can add skip arrays that act as leading '=' quals in the
7839 : : * absence of ordinary input '=' quals, so in practice _most_ input quals
7840 : : * are able to act as index bound quals (which we take into account here).
7841 : : *
7842 : : * For a RowCompareExpr, we consider only the first column, just as
7843 : : * rowcomparesel() does.
7844 : : *
7845 : : * If there's a SAOP or skip array in the quals, we'll actually perform up
7846 : : * to N index descents (not just one), but the underlying array key's
7847 : : * operator can be considered to act the same as it normally does.
7848 : : */
7849 : 636413 : indexBoundQuals = NIL;
7850 : 636413 : indexSkipQuals = NIL;
7851 : 636413 : indexcol = 0;
7852 : 636413 : eqQualHere = false;
7853 : 636413 : found_row_compare = false;
7854 : 636413 : found_array = false;
7855 : 636413 : found_is_null_op = false;
7856 : 636413 : num_sa_scans = 1;
7857 [ + + + + : 1075064 : foreach(lc, path->indexclauses)
+ + ]
7858 : : {
7859 : 464278 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7860 : : ListCell *lc2;
7861 : :
7862 [ + + ]: 464278 : if (indexcol < iclause->indexcol)
7863 : : {
7864 : 87452 : double num_sa_scans_prev_cols = num_sa_scans;
7865 : :
7866 : : /*
7867 : : * Beginning of a new column's quals.
7868 : : *
7869 : : * Skip scans use skip arrays, which are ScalarArrayOp style
7870 : : * arrays that generate their elements procedurally and on demand.
7871 : : * Given a multi-column index on "(a, b)", and an SQL WHERE clause
7872 : : * "WHERE b = 42", a skip scan will effectively use an indexqual
7873 : : * "WHERE a = ANY('{every col a value}') AND b = 42". (Obviously,
7874 : : * the array on "a" must also return "IS NULL" matches, since our
7875 : : * WHERE clause used no strict operator on "a").
7876 : : *
7877 : : * Here we consider how nbtree will backfill skip arrays for any
7878 : : * index columns that lacked an '=' qual. This maintains our
7879 : : * num_sa_scans estimate, and determines if this new column (the
7880 : : * "iclause->indexcol" column, not the prior "indexcol" column)
7881 : : * can have its RestrictInfos/quals added to indexBoundQuals.
7882 : : *
7883 : : * We'll need to handle columns that have inequality quals, where
7884 : : * the skip array generates values from a range constrained by the
7885 : : * quals (not every possible value). We've been maintaining
7886 : : * indexSkipQuals to help with this; it will now contain all of
7887 : : * the prior column's quals (that is, indexcol's quals) when they
7888 : : * might be used for this.
7889 : : */
7890 [ + + ]: 87452 : if (found_row_compare)
7891 : : {
7892 : : /*
7893 : : * Skip arrays can't be added after a RowCompare input qual
7894 : : * due to limitations in nbtree
7895 : : */
7896 : 20 : break;
7897 : : }
7898 [ + + ]: 87432 : if (eqQualHere)
7899 : : {
7900 : : /*
7901 : : * Don't need to add a skip array for an indexcol that already
7902 : : * has an '=' qual/equality constraint
7903 : : */
7904 : 62267 : indexcol++;
7905 : 62267 : indexSkipQuals = NIL;
7906 : : }
7907 : 87432 : eqQualHere = false;
7908 : :
7909 [ + + ]: 88908 : while (indexcol < iclause->indexcol)
7910 : : {
7911 : : double ndistinct;
7912 : 27083 : bool isdefault = true;
7913 : :
7914 : 27083 : found_array = true;
7915 : :
7916 : : /*
7917 : : * A skipped attribute's ndistinct forms the basis of our
7918 : : * estimate of the total number of "array elements" used by
7919 : : * its skip array at runtime. Look that up first.
7920 : : */
7921 : 27083 : examine_indexcol_variable(root, index, indexcol, &vardata);
7922 : 27083 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
7923 : :
7924 [ + + ]: 27083 : if (indexcol == 0)
7925 : : {
7926 : : /*
7927 : : * Get an estimate of the leading column's correlation in
7928 : : * passing (avoids rereading variable stats below)
7929 : : */
7930 [ + + ]: 25155 : if (HeapTupleIsValid(vardata.statsTuple))
7931 : 14569 : correlation = btcost_correlation(index, &vardata);
7932 : 25155 : have_correlation = true;
7933 : : }
7934 : :
7935 [ + + ]: 27083 : ReleaseVariableStats(vardata);
7936 : :
7937 : : /*
7938 : : * If ndistinct is a default estimate, conservatively assume
7939 : : * that no skipping will happen at runtime
7940 : : */
7941 [ + + ]: 27083 : if (isdefault)
7942 : : {
7943 : 8026 : num_sa_scans = num_sa_scans_prev_cols;
7944 : 25607 : break; /* done building indexBoundQuals */
7945 : : }
7946 : :
7947 : : /*
7948 : : * Apply indexcol's indexSkipQuals selectivity to ndistinct
7949 : : */
7950 [ + + ]: 19057 : if (indexSkipQuals != NIL)
7951 : : {
7952 : : List *partialSkipQuals;
7953 : : Selectivity ndistinctfrac;
7954 : :
7955 : : /*
7956 : : * If the index is partial, AND the index predicate with
7957 : : * the index-bound quals to produce a more accurate idea
7958 : : * of the number of distinct values for prior indexcol
7959 : : */
7960 : 552 : partialSkipQuals = add_predicate_to_index_quals(index,
7961 : : indexSkipQuals);
7962 : :
7963 : 552 : ndistinctfrac = clauselist_selectivity(root, partialSkipQuals,
7964 : 552 : index->rel->relid,
7965 : : JOIN_INNER,
7966 : : NULL);
7967 : :
7968 : : /*
7969 : : * If ndistinctfrac is selective (on its own), the scan is
7970 : : * unlikely to benefit from repositioning itself using
7971 : : * later quals. Do not allow iclause->indexcol's quals to
7972 : : * be added to indexBoundQuals (it would increase descent
7973 : : * costs, without lowering numIndexTuples costs by much).
7974 : : */
7975 [ + + ]: 552 : if (ndistinctfrac < DEFAULT_RANGE_INEQ_SEL)
7976 : : {
7977 : 311 : num_sa_scans = num_sa_scans_prev_cols;
7978 : 311 : break; /* done building indexBoundQuals */
7979 : : }
7980 : :
7981 : : /* Adjust ndistinct downward */
7982 : 241 : ndistinct = rint(ndistinct * ndistinctfrac);
7983 [ + - ]: 241 : ndistinct = Max(ndistinct, 1);
7984 : : }
7985 : :
7986 : : /*
7987 : : * When there's no inequality quals, account for the need to
7988 : : * find an initial value by counting -inf/+inf as a value.
7989 : : *
7990 : : * We don't charge anything extra for possible next/prior key
7991 : : * index probes, which are sometimes used to find the next
7992 : : * valid skip array element (ahead of using the located
7993 : : * element value to relocate the scan to the next position
7994 : : * that might contain matching tuples). It seems hard to do
7995 : : * better here. Use of the skip support infrastructure often
7996 : : * avoids most next/prior key probes. But even when it can't,
7997 : : * there's a decent chance that most individual next/prior key
7998 : : * probes will locate a leaf page whose key space overlaps all
7999 : : * of the scan's keys (even the lower-order keys) -- which
8000 : : * also avoids the need for a separate, extra index descent.
8001 : : * Note also that these probes are much cheaper than non-probe
8002 : : * primitive index scans: they're reliably very selective.
8003 : : */
8004 [ + + ]: 18746 : if (indexSkipQuals == NIL)
8005 : 18505 : ndistinct += 1;
8006 : :
8007 : : /*
8008 : : * Update num_sa_scans estimate by multiplying by ndistinct.
8009 : : *
8010 : : * We make the pessimistic assumption that there is no
8011 : : * naturally occurring cross-column correlation. This is
8012 : : * often wrong, but it seems best to err on the side of not
8013 : : * expecting skipping to be helpful...
8014 : : */
8015 : 18746 : num_sa_scans *= ndistinct;
8016 : :
8017 : : /*
8018 : : * ...but back out of adding this latest group of 1 or more
8019 : : * skip arrays when num_sa_scans exceeds the total number of
8020 : : * index pages (revert to num_sa_scans from before indexcol).
8021 : : * This causes a sharp discontinuity in cost (as a function of
8022 : : * the indexcol's ndistinct), but that is representative of
8023 : : * actual runtime costs.
8024 : : *
8025 : : * Note that skipping is helpful when each primitive index
8026 : : * scan only manages to skip over 1 or 2 irrelevant leaf pages
8027 : : * on average. Skip arrays bring savings in CPU costs due to
8028 : : * the scan not needing to evaluate indexquals against every
8029 : : * tuple, which can greatly exceed any savings in I/O costs.
8030 : : * This test is a test of whether num_sa_scans implies that
8031 : : * we're past the point where the ability to skip ceases to
8032 : : * lower the scan's costs (even qual evaluation CPU costs).
8033 : : */
8034 [ + + ]: 18746 : if (index->pages < num_sa_scans)
8035 : : {
8036 : 17270 : num_sa_scans = num_sa_scans_prev_cols;
8037 : 17270 : break; /* done building indexBoundQuals */
8038 : : }
8039 : :
8040 : 1476 : indexcol++;
8041 : 1476 : indexSkipQuals = NIL;
8042 : : }
8043 : :
8044 : : /*
8045 : : * Finished considering the need to add skip arrays to bridge an
8046 : : * initial eqQualHere gap between the old and new index columns
8047 : : * (or there was no initial eqQualHere gap in the first place).
8048 : : *
8049 : : * If an initial gap could not be bridged, then new column's quals
8050 : : * (i.e. iclause->indexcol's quals) won't go into indexBoundQuals,
8051 : : * and so won't affect our final numIndexTuples estimate.
8052 : : */
8053 [ + + ]: 87432 : if (indexcol != iclause->indexcol)
8054 : 25607 : break; /* done building indexBoundQuals */
8055 : : }
8056 : :
8057 : : Assert(indexcol == iclause->indexcol);
8058 : :
8059 : : /* Examine each indexqual associated with this index clause */
8060 [ + - + + : 879605 : foreach(lc2, iclause->indexquals)
+ + ]
8061 : : {
8062 : 440954 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8063 : 440954 : Expr *clause = rinfo->clause;
8064 : 440954 : Oid clause_op = InvalidOid;
8065 : : int op_strategy;
8066 : :
8067 [ + + ]: 440954 : if (IsA(clause, OpExpr))
8068 : : {
8069 : 426480 : OpExpr *op = (OpExpr *) clause;
8070 : :
8071 : 426480 : clause_op = op->opno;
8072 : : }
8073 [ + + ]: 14474 : else if (IsA(clause, RowCompareExpr))
8074 : : {
8075 : 330 : RowCompareExpr *rc = (RowCompareExpr *) clause;
8076 : :
8077 : 330 : clause_op = linitial_oid(rc->opnos);
8078 : 330 : found_row_compare = true;
8079 : : }
8080 [ + + ]: 14144 : else if (IsA(clause, ScalarArrayOpExpr))
8081 : : {
8082 : 12246 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
8083 : 12246 : Node *other_operand = (Node *) lsecond(saop->args);
8084 : 12246 : double alength = estimate_array_length(root, other_operand);
8085 : :
8086 : 12246 : clause_op = saop->opno;
8087 : 12246 : found_array = true;
8088 : : /* estimate SA descents by indexBoundQuals only */
8089 [ + + ]: 12246 : if (alength > 1)
8090 : 12022 : num_sa_scans *= alength;
8091 : : }
8092 [ + - ]: 1898 : else if (IsA(clause, NullTest))
8093 : : {
8094 : 1898 : NullTest *nt = (NullTest *) clause;
8095 : :
8096 [ + + ]: 1898 : if (nt->nulltesttype == IS_NULL)
8097 : : {
8098 : 200 : found_is_null_op = true;
8099 : : /* IS NULL is like = for selectivity/skip scan purposes */
8100 : 200 : eqQualHere = true;
8101 : : }
8102 : : }
8103 : : else
8104 [ # # ]: 0 : elog(ERROR, "unsupported indexqual type: %d",
8105 : : (int) nodeTag(clause));
8106 : :
8107 : : /* check for equality operator */
8108 [ + + ]: 440954 : if (OidIsValid(clause_op))
8109 : : {
8110 : 439056 : op_strategy = get_op_opfamily_strategy(clause_op,
8111 : 439056 : index->opfamily[indexcol]);
8112 : : Assert(op_strategy != 0); /* not a member of opfamily?? */
8113 [ + + ]: 439056 : if (op_strategy == BTEqualStrategyNumber)
8114 : 415547 : eqQualHere = true;
8115 : : }
8116 : :
8117 : 440954 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
8118 : :
8119 : : /*
8120 : : * We apply inequality selectivities to estimate index descent
8121 : : * costs with scans that use skip arrays. Save this indexcol's
8122 : : * RestrictInfos if it looks like they'll be needed for that.
8123 : : */
8124 [ + + + + ]: 440954 : if (!eqQualHere && !found_row_compare &&
8125 [ + + ]: 24310 : indexcol < index->nkeycolumns - 1)
8126 : 4772 : indexSkipQuals = lappend(indexSkipQuals, rinfo);
8127 : : }
8128 : : }
8129 : :
8130 : : /*
8131 : : * If index is unique and we found an '=' clause for each column, we can
8132 : : * just assume numIndexTuples = 1 and skip the expensive
8133 : : * clauselist_selectivity calculations. However, an array or NullTest
8134 : : * always invalidates that theory (even when eqQualHere has been set).
8135 : : */
8136 [ + + ]: 636413 : if (index->unique &&
8137 [ + + + + ]: 508393 : indexcol == index->nkeycolumns - 1 &&
8138 : 183137 : eqQualHere &&
8139 [ + + ]: 183137 : !found_array &&
8140 [ + + ]: 177731 : !found_is_null_op)
8141 : 177691 : numIndexTuples = 1.0;
8142 : : else
8143 : : {
8144 : : List *selectivityQuals;
8145 : : Selectivity btreeSelectivity;
8146 : :
8147 : : /*
8148 : : * If the index is partial, AND the index predicate with the
8149 : : * index-bound quals to produce a more accurate idea of the number of
8150 : : * rows covered by the bound conditions.
8151 : : */
8152 : 458722 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
8153 : :
8154 : 458722 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
8155 : 458722 : index->rel->relid,
8156 : : JOIN_INNER,
8157 : : NULL);
8158 : 458722 : numIndexTuples = btreeSelectivity * index->rel->tuples;
8159 : :
8160 : : /*
8161 : : * btree automatically combines individual array element primitive
8162 : : * index scans whenever the tuples covered by the next set of array
8163 : : * keys are close to tuples covered by the current set. That puts a
8164 : : * natural ceiling on the worst case number of descents -- there
8165 : : * cannot possibly be more than one descent per leaf page scanned.
8166 : : *
8167 : : * Clamp the number of descents to at most 1/3 the number of index
8168 : : * pages. This avoids implausibly high estimates with low selectivity
8169 : : * paths, where scans usually require only one or two descents. This
8170 : : * is most likely to help when there are several SAOP clauses, where
8171 : : * naively accepting the total number of distinct combinations of
8172 : : * array elements as the number of descents would frequently lead to
8173 : : * wild overestimates.
8174 : : *
8175 : : * We somewhat arbitrarily don't just make the cutoff the total number
8176 : : * of leaf pages (we make it 1/3 the total number of pages instead) to
8177 : : * give the btree code credit for its ability to continue on the leaf
8178 : : * level with low selectivity scans.
8179 : : *
8180 : : * Note: num_sa_scans includes both ScalarArrayOp array elements and
8181 : : * skip array elements whose qual affects our numIndexTuples estimate.
8182 : : */
8183 [ + + ]: 458722 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
8184 [ + + ]: 458722 : num_sa_scans = Max(num_sa_scans, 1);
8185 : :
8186 : : /*
8187 : : * As in genericcostestimate(), we have to adjust for any array quals
8188 : : * included in indexBoundQuals, and then round to integer.
8189 : : *
8190 : : * It is tempting to make genericcostestimate behave as if array
8191 : : * clauses work in almost the same way as scalar operators during
8192 : : * btree scans, making the top-level scan look like a continuous scan
8193 : : * (as opposed to num_sa_scans-many primitive index scans). After
8194 : : * all, btree scans mostly work like that at runtime. However, such a
8195 : : * scheme would badly bias genericcostestimate's simplistic approach
8196 : : * to calculating numIndexPages through prorating.
8197 : : *
8198 : : * Stick with the approach taken by non-native SAOP scans for now.
8199 : : * genericcostestimate will use the Mackert-Lohman formula to
8200 : : * compensate for repeat page fetches, even though that definitely
8201 : : * won't happen during btree scans (not for leaf pages, at least).
8202 : : * We're usually very pessimistic about the number of primitive index
8203 : : * scans that will be required, but it's not clear how to do better.
8204 : : */
8205 : 458722 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
8206 : : }
8207 : :
8208 : : /*
8209 : : * Now do generic index cost estimation.
8210 : : *
8211 : : * While we expended effort to make realistic estimates of numIndexTuples
8212 : : * and num_sa_scans, we are content to count only the btree metapage as
8213 : : * non-leaf. btree fanout is typically high enough that upper pages are
8214 : : * few relative to leaf pages, so accounting for them would move the
8215 : : * estimates at most a percent or two. Given the uncertainty in just how
8216 : : * many upper pages exist in a particular index, we'll skip trying to
8217 : : * handle that.
8218 : : */
8219 : 636413 : costs.numIndexTuples = numIndexTuples;
8220 : 636413 : costs.num_sa_scans = num_sa_scans;
8221 : 636413 : costs.numNonLeafPages = 1;
8222 : :
8223 : 636413 : genericcostestimate(root, path, loop_count, &costs);
8224 : :
8225 : : /*
8226 : : * Add a CPU-cost component to represent the costs of initial btree
8227 : : * descent. We don't charge any I/O cost for touching upper btree levels,
8228 : : * since they tend to stay in cache, but we still have to do about log2(N)
8229 : : * comparisons to descend a btree of N leaf tuples. We charge one
8230 : : * cpu_operator_cost per comparison.
8231 : : *
8232 : : * If there are SAOP or skip array keys, charge this once per estimated
8233 : : * index descent. The ones after the first one are not startup cost so
8234 : : * far as the overall plan goes, so just add them to "total" cost.
8235 : : */
8236 [ + + ]: 636413 : if (index->tuples > 1) /* avoid computing log(0) */
8237 : : {
8238 : 582406 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
8239 : 582406 : costs.indexStartupCost += descentCost;
8240 : 582406 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8241 : : }
8242 : :
8243 : : /*
8244 : : * Even though we're not charging I/O cost for touching upper btree pages,
8245 : : * it's still reasonable to charge some CPU cost per page descended
8246 : : * through. Moreover, if we had no such charge at all, bloated indexes
8247 : : * would appear to have the same search cost as unbloated ones, at least
8248 : : * in cases where only a single leaf page is expected to be visited. This
8249 : : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
8250 : : * touched. The number of such pages is btree tree height plus one (ie,
8251 : : * we charge for the leaf page too). As above, charge once per estimated
8252 : : * SAOP/skip array descent.
8253 : : */
8254 : 636413 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8255 : 636413 : costs.indexStartupCost += descentCost;
8256 : 636413 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8257 : :
8258 [ + + ]: 636413 : if (!have_correlation)
8259 : : {
8260 : 611258 : examine_indexcol_variable(root, index, 0, &vardata);
8261 [ + + ]: 611258 : if (HeapTupleIsValid(vardata.statsTuple))
8262 : 398708 : costs.indexCorrelation = btcost_correlation(index, &vardata);
8263 [ + + ]: 611258 : ReleaseVariableStats(vardata);
8264 : : }
8265 : : else
8266 : : {
8267 : : /* btcost_correlation already called earlier on */
8268 : 25155 : costs.indexCorrelation = correlation;
8269 : : }
8270 : :
8271 : 636413 : *indexStartupCost = costs.indexStartupCost;
8272 : 636413 : *indexTotalCost = costs.indexTotalCost;
8273 : 636413 : *indexSelectivity = costs.indexSelectivity;
8274 : 636413 : *indexCorrelation = costs.indexCorrelation;
8275 : 636413 : *indexPages = costs.numIndexPages;
8276 : 636413 : }
8277 : :
8278 : : void
8279 : 308 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8280 : : Cost *indexStartupCost, Cost *indexTotalCost,
8281 : : Selectivity *indexSelectivity, double *indexCorrelation,
8282 : : double *indexPages)
8283 : : {
8284 : 308 : GenericCosts costs = {0};
8285 : :
8286 : : /* As in btcostestimate, count only the metapage as non-leaf */
8287 : 308 : costs.numNonLeafPages = 1;
8288 : :
8289 : 308 : genericcostestimate(root, path, loop_count, &costs);
8290 : :
8291 : : /*
8292 : : * A hash index has no descent costs as such, since the index AM can go
8293 : : * directly to the target bucket after computing the hash value. There
8294 : : * are a couple of other hash-specific costs that we could conceivably add
8295 : : * here, though:
8296 : : *
8297 : : * Ideally we'd charge spc_random_page_cost for each page in the target
8298 : : * bucket, not just the numIndexPages pages that genericcostestimate
8299 : : * thought we'd visit. However in most cases we don't know which bucket
8300 : : * that will be. There's no point in considering the average bucket size
8301 : : * because the hash AM makes sure that's always one page.
8302 : : *
8303 : : * Likewise, we could consider charging some CPU for each index tuple in
8304 : : * the bucket, if we knew how many there were. But the per-tuple cost is
8305 : : * just a hash value comparison, not a general datatype-dependent
8306 : : * comparison, so any such charge ought to be quite a bit less than
8307 : : * cpu_operator_cost; which makes it probably not worth worrying about.
8308 : : *
8309 : : * A bigger issue is that chance hash-value collisions will result in
8310 : : * wasted probes into the heap. We don't currently attempt to model this
8311 : : * cost on the grounds that it's rare, but maybe it's not rare enough.
8312 : : * (Any fix for this ought to consider the generic lossy-operator problem,
8313 : : * though; it's not entirely hash-specific.)
8314 : : */
8315 : :
8316 : 308 : *indexStartupCost = costs.indexStartupCost;
8317 : 308 : *indexTotalCost = costs.indexTotalCost;
8318 : 308 : *indexSelectivity = costs.indexSelectivity;
8319 : 308 : *indexCorrelation = costs.indexCorrelation;
8320 : 308 : *indexPages = costs.numIndexPages;
8321 : 308 : }
8322 : :
8323 : : void
8324 : 4778 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8325 : : Cost *indexStartupCost, Cost *indexTotalCost,
8326 : : Selectivity *indexSelectivity, double *indexCorrelation,
8327 : : double *indexPages)
8328 : : {
8329 : 4778 : IndexOptInfo *index = path->indexinfo;
8330 : 4778 : GenericCosts costs = {0};
8331 : : Cost descentCost;
8332 : :
8333 : : /* GiST has no metapage, so we treat all pages as leaf pages */
8334 : :
8335 : 4778 : genericcostestimate(root, path, loop_count, &costs);
8336 : :
8337 : : /*
8338 : : * We model index descent costs similarly to those for btree, but to do
8339 : : * that we first need an idea of the tree height. We somewhat arbitrarily
8340 : : * assume that the fanout is 100, meaning the tree height is at most
8341 : : * log100(index->pages).
8342 : : *
8343 : : * Although this computation isn't really expensive enough to require
8344 : : * caching, we might as well use index->tree_height to cache it.
8345 : : */
8346 [ + + ]: 4778 : if (index->tree_height < 0) /* unknown? */
8347 : : {
8348 [ + + ]: 4751 : if (index->pages > 1) /* avoid computing log(0) */
8349 : 1964 : index->tree_height = (int) (log(index->pages) / log(100.0));
8350 : : else
8351 : 2787 : index->tree_height = 0;
8352 : : }
8353 : :
8354 : : /*
8355 : : * Add a CPU-cost component to represent the costs of initial descent. We
8356 : : * just use log(N) here not log2(N) since the branching factor isn't
8357 : : * necessarily two anyway. As for btree, charge once per SA scan.
8358 : : */
8359 [ + - ]: 4778 : if (index->tuples > 1) /* avoid computing log(0) */
8360 : : {
8361 : 4778 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8362 : 4778 : costs.indexStartupCost += descentCost;
8363 : 4778 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8364 : : }
8365 : :
8366 : : /*
8367 : : * Likewise add a per-page charge, calculated the same as for btrees.
8368 : : */
8369 : 4778 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8370 : 4778 : costs.indexStartupCost += descentCost;
8371 : 4778 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8372 : :
8373 : 4778 : *indexStartupCost = costs.indexStartupCost;
8374 : 4778 : *indexTotalCost = costs.indexTotalCost;
8375 : 4778 : *indexSelectivity = costs.indexSelectivity;
8376 : 4778 : *indexCorrelation = costs.indexCorrelation;
8377 : 4778 : *indexPages = costs.numIndexPages;
8378 : 4778 : }
8379 : :
8380 : : void
8381 : 1482 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8382 : : Cost *indexStartupCost, Cost *indexTotalCost,
8383 : : Selectivity *indexSelectivity, double *indexCorrelation,
8384 : : double *indexPages)
8385 : : {
8386 : 1482 : IndexOptInfo *index = path->indexinfo;
8387 : 1482 : GenericCosts costs = {0};
8388 : : Cost descentCost;
8389 : :
8390 : : /* As in btcostestimate, count only the metapage as non-leaf */
8391 : 1482 : costs.numNonLeafPages = 1;
8392 : :
8393 : 1482 : genericcostestimate(root, path, loop_count, &costs);
8394 : :
8395 : : /*
8396 : : * We model index descent costs similarly to those for btree, but to do
8397 : : * that we first need an idea of the tree height. We somewhat arbitrarily
8398 : : * assume that the fanout is 100, meaning the tree height is at most
8399 : : * log100(index->pages).
8400 : : *
8401 : : * Although this computation isn't really expensive enough to require
8402 : : * caching, we might as well use index->tree_height to cache it.
8403 : : */
8404 [ + + ]: 1482 : if (index->tree_height < 0) /* unknown? */
8405 : : {
8406 [ + - ]: 1477 : if (index->pages > 1) /* avoid computing log(0) */
8407 : 1477 : index->tree_height = (int) (log(index->pages) / log(100.0));
8408 : : else
8409 : 0 : index->tree_height = 0;
8410 : : }
8411 : :
8412 : : /*
8413 : : * Add a CPU-cost component to represent the costs of initial descent. We
8414 : : * just use log(N) here not log2(N) since the branching factor isn't
8415 : : * necessarily two anyway. As for btree, charge once per SA scan.
8416 : : */
8417 [ + - ]: 1482 : if (index->tuples > 1) /* avoid computing log(0) */
8418 : : {
8419 : 1482 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8420 : 1482 : costs.indexStartupCost += descentCost;
8421 : 1482 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8422 : : }
8423 : :
8424 : : /*
8425 : : * Likewise add a per-page charge, calculated the same as for btrees.
8426 : : */
8427 : 1482 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8428 : 1482 : costs.indexStartupCost += descentCost;
8429 : 1482 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8430 : :
8431 : 1482 : *indexStartupCost = costs.indexStartupCost;
8432 : 1482 : *indexTotalCost = costs.indexTotalCost;
8433 : 1482 : *indexSelectivity = costs.indexSelectivity;
8434 : 1482 : *indexCorrelation = costs.indexCorrelation;
8435 : 1482 : *indexPages = costs.numIndexPages;
8436 : 1482 : }
8437 : :
8438 : :
8439 : : /*
8440 : : * Support routines for gincostestimate
8441 : : */
8442 : :
8443 : : typedef struct
8444 : : {
8445 : : bool attHasFullScan[INDEX_MAX_KEYS];
8446 : : bool attHasNormalScan[INDEX_MAX_KEYS];
8447 : : double partialEntries;
8448 : : double exactEntries;
8449 : : double searchEntries;
8450 : : double arrayScans;
8451 : : } GinQualCounts;
8452 : :
8453 : : /*
8454 : : * Estimate the number of index terms that need to be searched for while
8455 : : * testing the given GIN query, and increment the counts in *counts
8456 : : * appropriately. If the query is unsatisfiable, return false.
8457 : : */
8458 : : static bool
8459 : 1714 : gincost_pattern(IndexOptInfo *index, int indexcol,
8460 : : Oid clause_op, Datum query,
8461 : : GinQualCounts *counts)
8462 : : {
8463 : : FmgrInfo flinfo;
8464 : : Oid extractProcOid;
8465 : : Oid collation;
8466 : : int strategy_op;
8467 : : Oid lefttype,
8468 : : righttype;
8469 : 1714 : int32 nentries = 0;
8470 : 1714 : bool *partial_matches = NULL;
8471 : 1714 : Pointer *extra_data = NULL;
8472 : 1714 : bool *nullFlags = NULL;
8473 : 1714 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
8474 : : int32 i;
8475 : :
8476 : : Assert(indexcol < index->nkeycolumns);
8477 : :
8478 : : /*
8479 : : * Get the operator's strategy number and declared input data types within
8480 : : * the index opfamily. (We don't need the latter, but we use
8481 : : * get_op_opfamily_properties because it will throw error if it fails to
8482 : : * find a matching pg_amop entry.)
8483 : : */
8484 : 1714 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
8485 : : &strategy_op, &lefttype, &righttype);
8486 : :
8487 : : /*
8488 : : * GIN always uses the "default" support functions, which are those with
8489 : : * lefttype == righttype == the opclass' opcintype (see
8490 : : * IndexSupportInitialize in relcache.c).
8491 : : */
8492 : 1714 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
8493 : 1714 : index->opcintype[indexcol],
8494 : 1714 : index->opcintype[indexcol],
8495 : : GIN_EXTRACTQUERY_PROC);
8496 : :
8497 [ - + ]: 1714 : if (!OidIsValid(extractProcOid))
8498 : : {
8499 : : /* should not happen; throw same error as index_getprocinfo */
8500 [ # # ]: 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
8501 : : GIN_EXTRACTQUERY_PROC, indexcol + 1,
8502 : : get_rel_name(index->indexoid));
8503 : : }
8504 : :
8505 : : /*
8506 : : * Choose collation to pass to extractProc (should match initGinState).
8507 : : */
8508 [ + + ]: 1714 : if (OidIsValid(index->indexcollations[indexcol]))
8509 : 243 : collation = index->indexcollations[indexcol];
8510 : : else
8511 : 1471 : collation = DEFAULT_COLLATION_OID;
8512 : :
8513 : 1714 : fmgr_info(extractProcOid, &flinfo);
8514 : :
8515 : 1714 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
8516 : :
8517 : 1714 : FunctionCall7Coll(&flinfo,
8518 : : collation,
8519 : : query,
8520 : : PointerGetDatum(&nentries),
8521 : : UInt16GetDatum(strategy_op),
8522 : : PointerGetDatum(&partial_matches),
8523 : : PointerGetDatum(&extra_data),
8524 : : PointerGetDatum(&nullFlags),
8525 : : PointerGetDatum(&searchMode));
8526 : :
8527 [ + + + + ]: 1714 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
8528 : : {
8529 : : /* No match is possible */
8530 : 10 : return false;
8531 : : }
8532 : :
8533 [ + + ]: 5854 : for (i = 0; i < nentries; i++)
8534 : : {
8535 : : /*
8536 : : * For partial match we haven't any information to estimate number of
8537 : : * matched entries in index, so, we just estimate it as 100
8538 : : */
8539 [ + + + + ]: 4150 : if (partial_matches && partial_matches[i])
8540 : 361 : counts->partialEntries += 100;
8541 : : else
8542 : 3789 : counts->exactEntries++;
8543 : :
8544 : 4150 : counts->searchEntries++;
8545 : : }
8546 : :
8547 [ + + ]: 1704 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
8548 : : {
8549 : 1322 : counts->attHasNormalScan[indexcol] = true;
8550 : : }
8551 [ + + ]: 382 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
8552 : : {
8553 : : /* Treat "include empty" like an exact-match item */
8554 : 36 : counts->attHasNormalScan[indexcol] = true;
8555 : 36 : counts->exactEntries++;
8556 : 36 : counts->searchEntries++;
8557 : : }
8558 : : else
8559 : : {
8560 : : /* It's GIN_SEARCH_MODE_ALL */
8561 : 346 : counts->attHasFullScan[indexcol] = true;
8562 : : }
8563 : :
8564 : 1704 : return true;
8565 : : }
8566 : :
8567 : : /*
8568 : : * Estimate the number of index terms that need to be searched for while
8569 : : * testing the given GIN index clause, and increment the counts in *counts
8570 : : * appropriately. If the query is unsatisfiable, return false.
8571 : : */
8572 : : static bool
8573 : 1704 : gincost_opexpr(PlannerInfo *root,
8574 : : IndexOptInfo *index,
8575 : : int indexcol,
8576 : : OpExpr *clause,
8577 : : GinQualCounts *counts)
8578 : : {
8579 : 1704 : Oid clause_op = clause->opno;
8580 : 1704 : Node *operand = (Node *) lsecond(clause->args);
8581 : :
8582 : : /* aggressively reduce to a constant, and look through relabeling */
8583 : 1704 : operand = estimate_expression_value(root, operand);
8584 : :
8585 [ - + ]: 1704 : if (IsA(operand, RelabelType))
8586 : 0 : operand = (Node *) ((RelabelType *) operand)->arg;
8587 : :
8588 : : /*
8589 : : * It's impossible to call extractQuery method for unknown operand. So
8590 : : * unless operand is a Const we can't do much; just assume there will be
8591 : : * one ordinary search entry from the operand at runtime.
8592 : : */
8593 [ - + ]: 1704 : if (!IsA(operand, Const))
8594 : : {
8595 : 0 : counts->exactEntries++;
8596 : 0 : counts->searchEntries++;
8597 : 0 : return true;
8598 : : }
8599 : :
8600 : : /* If Const is null, there can be no matches */
8601 [ - + ]: 1704 : if (((Const *) operand)->constisnull)
8602 : 0 : return false;
8603 : :
8604 : : /* Otherwise, apply extractQuery and get the actual term counts */
8605 : 1704 : return gincost_pattern(index, indexcol, clause_op,
8606 : : ((Const *) operand)->constvalue,
8607 : : counts);
8608 : : }
8609 : :
8610 : : /*
8611 : : * Estimate the number of index terms that need to be searched for while
8612 : : * testing the given GIN index clause, and increment the counts in *counts
8613 : : * appropriately. If the query is unsatisfiable, return false.
8614 : : *
8615 : : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
8616 : : * each of which involves one value from the RHS array, plus all the
8617 : : * non-array quals (if any). To model this, we average the counts across
8618 : : * the RHS elements, and add the averages to the counts in *counts (which
8619 : : * correspond to per-indexscan costs). We also multiply counts->arrayScans
8620 : : * by N, causing gincostestimate to scale up its estimates accordingly.
8621 : : */
8622 : : static bool
8623 : 5 : gincost_scalararrayopexpr(PlannerInfo *root,
8624 : : IndexOptInfo *index,
8625 : : int indexcol,
8626 : : ScalarArrayOpExpr *clause,
8627 : : double numIndexEntries,
8628 : : GinQualCounts *counts)
8629 : : {
8630 : 5 : Oid clause_op = clause->opno;
8631 : 5 : Node *rightop = (Node *) lsecond(clause->args);
8632 : : ArrayType *arrayval;
8633 : : int16 elmlen;
8634 : : bool elmbyval;
8635 : : char elmalign;
8636 : : int numElems;
8637 : : Datum *elemValues;
8638 : : bool *elemNulls;
8639 : : GinQualCounts arraycounts;
8640 : 5 : int numPossible = 0;
8641 : : int i;
8642 : :
8643 : : Assert(clause->useOr);
8644 : :
8645 : : /* aggressively reduce to a constant, and look through relabeling */
8646 : 5 : rightop = estimate_expression_value(root, rightop);
8647 : :
8648 [ - + ]: 5 : if (IsA(rightop, RelabelType))
8649 : 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
8650 : :
8651 : : /*
8652 : : * It's impossible to call extractQuery method for unknown operand. So
8653 : : * unless operand is a Const we can't do much; just assume there will be
8654 : : * one ordinary search entry from each array entry at runtime, and fall
8655 : : * back on a probably-bad estimate of the number of array entries.
8656 : : */
8657 [ - + ]: 5 : if (!IsA(rightop, Const))
8658 : : {
8659 : 0 : counts->exactEntries++;
8660 : 0 : counts->searchEntries++;
8661 : 0 : counts->arrayScans *= estimate_array_length(root, rightop);
8662 : 0 : return true;
8663 : : }
8664 : :
8665 : : /* If Const is null, there can be no matches */
8666 [ - + ]: 5 : if (((Const *) rightop)->constisnull)
8667 : 0 : return false;
8668 : :
8669 : : /* Otherwise, extract the array elements and iterate over them */
8670 : 5 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
8671 : 5 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
8672 : : &elmlen, &elmbyval, &elmalign);
8673 : 5 : deconstruct_array(arrayval,
8674 : : ARR_ELEMTYPE(arrayval),
8675 : : elmlen, elmbyval, elmalign,
8676 : : &elemValues, &elemNulls, &numElems);
8677 : :
8678 : 5 : memset(&arraycounts, 0, sizeof(arraycounts));
8679 : :
8680 [ + + ]: 15 : for (i = 0; i < numElems; i++)
8681 : : {
8682 : : GinQualCounts elemcounts;
8683 : :
8684 : : /* NULL can't match anything, so ignore, as the executor will */
8685 [ - + ]: 10 : if (elemNulls[i])
8686 : 0 : continue;
8687 : :
8688 : : /* Otherwise, apply extractQuery and get the actual term counts */
8689 : 10 : memset(&elemcounts, 0, sizeof(elemcounts));
8690 : :
8691 [ + - ]: 10 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
8692 : : &elemcounts))
8693 : : {
8694 : : /* We ignore array elements that are unsatisfiable patterns */
8695 : 10 : numPossible++;
8696 : :
8697 [ - + ]: 10 : if (elemcounts.attHasFullScan[indexcol] &&
8698 [ # # ]: 0 : !elemcounts.attHasNormalScan[indexcol])
8699 : : {
8700 : : /*
8701 : : * Full index scan will be required. We treat this as if
8702 : : * every key in the index had been listed in the query; is
8703 : : * that reasonable?
8704 : : */
8705 : 0 : elemcounts.partialEntries = 0;
8706 : 0 : elemcounts.exactEntries = numIndexEntries;
8707 : 0 : elemcounts.searchEntries = numIndexEntries;
8708 : : }
8709 : 10 : arraycounts.partialEntries += elemcounts.partialEntries;
8710 : 10 : arraycounts.exactEntries += elemcounts.exactEntries;
8711 : 10 : arraycounts.searchEntries += elemcounts.searchEntries;
8712 : : }
8713 : : }
8714 : :
8715 [ - + ]: 5 : if (numPossible == 0)
8716 : : {
8717 : : /* No satisfiable patterns in the array */
8718 : 0 : return false;
8719 : : }
8720 : :
8721 : : /*
8722 : : * Now add the averages to the global counts. This will give us an
8723 : : * estimate of the average number of terms searched for in each indexscan,
8724 : : * including contributions from both array and non-array quals.
8725 : : */
8726 : 5 : counts->partialEntries += arraycounts.partialEntries / numPossible;
8727 : 5 : counts->exactEntries += arraycounts.exactEntries / numPossible;
8728 : 5 : counts->searchEntries += arraycounts.searchEntries / numPossible;
8729 : :
8730 : 5 : counts->arrayScans *= numPossible;
8731 : :
8732 : 5 : return true;
8733 : : }
8734 : :
8735 : : /*
8736 : : * GIN has search behavior completely different from other index types
8737 : : */
8738 : : void
8739 : 1546 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8740 : : Cost *indexStartupCost, Cost *indexTotalCost,
8741 : : Selectivity *indexSelectivity, double *indexCorrelation,
8742 : : double *indexPages)
8743 : : {
8744 : 1546 : IndexOptInfo *index = path->indexinfo;
8745 : 1546 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8746 : : List *selectivityQuals;
8747 : 1546 : double numPages = index->pages,
8748 : 1546 : numTuples = index->tuples;
8749 : : double numEntryPages,
8750 : : numDataPages,
8751 : : numPendingPages,
8752 : : numEntries;
8753 : : GinQualCounts counts;
8754 : : bool matchPossible;
8755 : : bool fullIndexScan;
8756 : : double partialScale;
8757 : : double entryPagesFetched,
8758 : : dataPagesFetched,
8759 : : dataPagesFetchedBySel;
8760 : : double qual_op_cost,
8761 : : qual_arg_cost,
8762 : : spc_random_page_cost,
8763 : : outer_scans;
8764 : : Cost descentCost;
8765 : : Relation indexRel;
8766 : : GinStatsData ginStats;
8767 : : ListCell *lc;
8768 : : int i;
8769 : :
8770 : : /*
8771 : : * Obtain statistical information from the meta page, if possible. Else
8772 : : * set ginStats to zeroes, and we'll cope below.
8773 : : */
8774 [ + - ]: 1546 : if (!index->hypothetical)
8775 : : {
8776 : : /* Lock should have already been obtained in plancat.c */
8777 : 1546 : indexRel = index_open(index->indexoid, NoLock);
8778 : 1546 : ginGetStats(indexRel, &ginStats);
8779 : 1546 : index_close(indexRel, NoLock);
8780 : : }
8781 : : else
8782 : : {
8783 : 0 : memset(&ginStats, 0, sizeof(ginStats));
8784 : : }
8785 : :
8786 : : /*
8787 : : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
8788 : : * trusted, but the other fields are data as of the last VACUUM. We can
8789 : : * scale them up to account for growth since then, but that method only
8790 : : * goes so far; in the worst case, the stats might be for a completely
8791 : : * empty index, and scaling them will produce pretty bogus numbers.
8792 : : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
8793 : : * it's grown more than that, fall back to estimating things only from the
8794 : : * assumed-accurate index size. But we'll trust nPendingPages in any case
8795 : : * so long as it's not clearly insane, ie, more than the index size.
8796 : : */
8797 [ + - ]: 1546 : if (ginStats.nPendingPages < numPages)
8798 : 1546 : numPendingPages = ginStats.nPendingPages;
8799 : : else
8800 : 0 : numPendingPages = 0;
8801 : :
8802 [ + - + - ]: 1546 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
8803 [ + + ]: 1546 : ginStats.nTotalPages > numPages / 4 &&
8804 [ + - + + ]: 1506 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
8805 : 1292 : {
8806 : : /*
8807 : : * OK, the stats seem close enough to sane to be trusted. But we
8808 : : * still need to scale them by the ratio numPages / nTotalPages to
8809 : : * account for growth since the last VACUUM.
8810 : : */
8811 : 1292 : double scale = numPages / ginStats.nTotalPages;
8812 : :
8813 : 1292 : numEntryPages = ceil(ginStats.nEntryPages * scale);
8814 : 1292 : numDataPages = ceil(ginStats.nDataPages * scale);
8815 : 1292 : numEntries = ceil(ginStats.nEntries * scale);
8816 : : /* ensure we didn't round up too much */
8817 [ + + ]: 1292 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
8818 [ + + ]: 1292 : numDataPages = Min(numDataPages,
8819 : : numPages - numPendingPages - numEntryPages);
8820 : : }
8821 : : else
8822 : : {
8823 : : /*
8824 : : * We might get here because it's a hypothetical index, or an index
8825 : : * created pre-9.1 and never vacuumed since upgrading (in which case
8826 : : * its stats would read as zeroes), or just because it's grown too
8827 : : * much since the last VACUUM for us to put our faith in scaling.
8828 : : *
8829 : : * Invent some plausible internal statistics based on the index page
8830 : : * count (and clamp that to at least 10 pages, just in case). We
8831 : : * estimate that 90% of the index is entry pages, and the rest is data
8832 : : * pages. Estimate 100 entries per entry page; this is rather bogus
8833 : : * since it'll depend on the size of the keys, but it's more robust
8834 : : * than trying to predict the number of entries per heap tuple.
8835 : : */
8836 [ + + ]: 254 : numPages = Max(numPages, 10);
8837 : 254 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
8838 : 254 : numDataPages = numPages - numPendingPages - numEntryPages;
8839 : 254 : numEntries = floor(numEntryPages * 100);
8840 : : }
8841 : :
8842 : : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
8843 [ - + ]: 1546 : if (numEntries < 1)
8844 : 0 : numEntries = 1;
8845 : :
8846 : : /*
8847 : : * If the index is partial, AND the index predicate with the index-bound
8848 : : * quals to produce a more accurate idea of the number of rows covered by
8849 : : * the bound conditions.
8850 : : */
8851 : 1546 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
8852 : :
8853 : : /* Estimate the fraction of main-table tuples that will be visited */
8854 : 3092 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
8855 : 1546 : index->rel->relid,
8856 : : JOIN_INNER,
8857 : : NULL);
8858 : :
8859 : : /* fetch estimated page cost for tablespace containing index */
8860 : 1546 : get_tablespace_page_costs(index->reltablespace,
8861 : : &spc_random_page_cost,
8862 : : NULL);
8863 : :
8864 : : /*
8865 : : * Generic assumption about index correlation: there isn't any.
8866 : : */
8867 : 1546 : *indexCorrelation = 0.0;
8868 : :
8869 : : /*
8870 : : * Examine quals to estimate number of search entries & partial matches
8871 : : */
8872 : 1546 : memset(&counts, 0, sizeof(counts));
8873 : 1546 : counts.arrayScans = 1;
8874 : 1546 : matchPossible = true;
8875 : :
8876 [ + - + + : 3255 : foreach(lc, path->indexclauses)
+ + ]
8877 : : {
8878 : 1709 : IndexClause *iclause = lfirst_node(IndexClause, lc);
8879 : : ListCell *lc2;
8880 : :
8881 [ + - + + : 3408 : foreach(lc2, iclause->indexquals)
+ + ]
8882 : : {
8883 : 1709 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8884 : 1709 : Expr *clause = rinfo->clause;
8885 : :
8886 [ + + ]: 1709 : if (IsA(clause, OpExpr))
8887 : : {
8888 : 1704 : matchPossible = gincost_opexpr(root,
8889 : : index,
8890 : 1704 : iclause->indexcol,
8891 : : (OpExpr *) clause,
8892 : : &counts);
8893 [ + + ]: 1704 : if (!matchPossible)
8894 : 10 : break;
8895 : : }
8896 [ + - ]: 5 : else if (IsA(clause, ScalarArrayOpExpr))
8897 : : {
8898 : 5 : matchPossible = gincost_scalararrayopexpr(root,
8899 : : index,
8900 : 5 : iclause->indexcol,
8901 : : (ScalarArrayOpExpr *) clause,
8902 : : numEntries,
8903 : : &counts);
8904 [ - + ]: 5 : if (!matchPossible)
8905 : 0 : break;
8906 : : }
8907 : : else
8908 : : {
8909 : : /* shouldn't be anything else for a GIN index */
8910 [ # # ]: 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
8911 : : (int) nodeTag(clause));
8912 : : }
8913 : : }
8914 : : }
8915 : :
8916 : : /* Fall out if there were any provably-unsatisfiable quals */
8917 [ + + ]: 1546 : if (!matchPossible)
8918 : : {
8919 : 10 : *indexStartupCost = 0;
8920 : 10 : *indexTotalCost = 0;
8921 : 10 : *indexSelectivity = 0;
8922 : 10 : return;
8923 : : }
8924 : :
8925 : : /*
8926 : : * If attribute has a full scan and at the same time doesn't have normal
8927 : : * scan, then we'll have to scan all non-null entries of that attribute.
8928 : : * Currently, we don't have per-attribute statistics for GIN. Thus, we
8929 : : * must assume the whole GIN index has to be scanned in this case.
8930 : : */
8931 : 1536 : fullIndexScan = false;
8932 [ + + ]: 2989 : for (i = 0; i < index->nkeycolumns; i++)
8933 : : {
8934 [ + + + + ]: 1726 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
8935 : : {
8936 : 273 : fullIndexScan = true;
8937 : 273 : break;
8938 : : }
8939 : : }
8940 : :
8941 [ + + - + ]: 1536 : if (fullIndexScan || indexQuals == NIL)
8942 : : {
8943 : : /*
8944 : : * Full index scan will be required. We treat this as if every key in
8945 : : * the index had been listed in the query; is that reasonable?
8946 : : */
8947 : 273 : counts.partialEntries = 0;
8948 : 273 : counts.exactEntries = numEntries;
8949 : 273 : counts.searchEntries = numEntries;
8950 : : }
8951 : :
8952 : : /* Will we have more than one iteration of a nestloop scan? */
8953 : 1536 : outer_scans = loop_count;
8954 : :
8955 : : /*
8956 : : * Compute cost to begin scan, first of all, pay attention to pending
8957 : : * list.
8958 : : */
8959 : 1536 : entryPagesFetched = numPendingPages;
8960 : :
8961 : : /*
8962 : : * Estimate number of entry pages read. We need to do
8963 : : * counts.searchEntries searches. Use a power function as it should be,
8964 : : * but tuples on leaf pages usually is much greater. Here we include all
8965 : : * searches in entry tree, including search of first entry in partial
8966 : : * match algorithm
8967 : : */
8968 : 1536 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8969 : :
8970 : : /*
8971 : : * Add an estimate of entry pages read by partial match algorithm. It's a
8972 : : * scan over leaf pages in entry tree. We haven't any useful stats here,
8973 : : * so estimate it as proportion. Because counts.partialEntries is really
8974 : : * pretty bogus (see code above), it's possible that it is more than
8975 : : * numEntries; clamp the proportion to ensure sanity.
8976 : : */
8977 : 1536 : partialScale = counts.partialEntries / numEntries;
8978 [ + + ]: 1536 : partialScale = Min(partialScale, 1.0);
8979 : :
8980 : 1536 : entryPagesFetched += ceil(numEntryPages * partialScale);
8981 : :
8982 : : /*
8983 : : * Partial match algorithm reads all data pages before doing actual scan,
8984 : : * so it's a startup cost. Again, we haven't any useful stats here, so
8985 : : * estimate it as proportion.
8986 : : */
8987 : 1536 : dataPagesFetched = ceil(numDataPages * partialScale);
8988 : :
8989 : 1536 : *indexStartupCost = 0;
8990 : 1536 : *indexTotalCost = 0;
8991 : :
8992 : : /*
8993 : : * Add a CPU-cost component to represent the costs of initial entry btree
8994 : : * descent. We don't charge any I/O cost for touching upper btree levels,
8995 : : * since they tend to stay in cache, but we still have to do about log2(N)
8996 : : * comparisons to descend a btree of N leaf tuples. We charge one
8997 : : * cpu_operator_cost per comparison.
8998 : : *
8999 : : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
9000 : : * ones after the first one are not startup cost so far as the overall
9001 : : * plan is concerned, so add them only to "total" cost.
9002 : : */
9003 [ + - ]: 1536 : if (numEntries > 1) /* avoid computing log(0) */
9004 : : {
9005 : 1536 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
9006 : 1536 : *indexStartupCost += descentCost * counts.searchEntries;
9007 : 1536 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
9008 : : }
9009 : :
9010 : : /*
9011 : : * Add a cpu cost per entry-page fetched. This is not amortized over a
9012 : : * loop.
9013 : : */
9014 : 1536 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
9015 : 1536 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
9016 : :
9017 : : /*
9018 : : * Add a cpu cost per data-page fetched. This is also not amortized over a
9019 : : * loop. Since those are the data pages from the partial match algorithm,
9020 : : * charge them as startup cost.
9021 : : */
9022 : 1536 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
9023 : :
9024 : : /*
9025 : : * Since we add the startup cost to the total cost later on, remove the
9026 : : * initial arrayscan from the total.
9027 : : */
9028 : 1536 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
9029 : :
9030 : : /*
9031 : : * Calculate cache effects if more than one scan due to nestloops or array
9032 : : * quals. The result is pro-rated per nestloop scan, but the array qual
9033 : : * factor shouldn't be pro-rated (compare genericcostestimate).
9034 : : */
9035 [ + - + + ]: 1536 : if (outer_scans > 1 || counts.arrayScans > 1)
9036 : : {
9037 : 5 : entryPagesFetched *= outer_scans * counts.arrayScans;
9038 : 5 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
9039 : : (BlockNumber) numEntryPages,
9040 : : numEntryPages, root);
9041 : 5 : entryPagesFetched /= outer_scans;
9042 : 5 : dataPagesFetched *= outer_scans * counts.arrayScans;
9043 : 5 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
9044 : : (BlockNumber) numDataPages,
9045 : : numDataPages, root);
9046 : 5 : dataPagesFetched /= outer_scans;
9047 : : }
9048 : :
9049 : : /*
9050 : : * Here we use random page cost because logically-close pages could be far
9051 : : * apart on disk.
9052 : : */
9053 : 1536 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
9054 : :
9055 : : /*
9056 : : * Now compute the number of data pages fetched during the scan.
9057 : : *
9058 : : * We assume every entry to have the same number of items, and that there
9059 : : * is no overlap between them. (XXX: tsvector and array opclasses collect
9060 : : * statistics on the frequency of individual keys; it would be nice to use
9061 : : * those here.)
9062 : : */
9063 : 1536 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
9064 : :
9065 : : /*
9066 : : * If there is a lot of overlap among the entries, in particular if one of
9067 : : * the entries is very frequent, the above calculation can grossly
9068 : : * under-estimate. As a simple cross-check, calculate a lower bound based
9069 : : * on the overall selectivity of the quals. At a minimum, we must read
9070 : : * one item pointer for each matching entry.
9071 : : *
9072 : : * The width of each item pointer varies, based on the level of
9073 : : * compression. We don't have statistics on that, but an average of
9074 : : * around 3 bytes per item is fairly typical.
9075 : : */
9076 : 1536 : dataPagesFetchedBySel = ceil(*indexSelectivity *
9077 : 1536 : (numTuples / (BLCKSZ / 3)));
9078 [ + + ]: 1536 : if (dataPagesFetchedBySel > dataPagesFetched)
9079 : 1243 : dataPagesFetched = dataPagesFetchedBySel;
9080 : :
9081 : : /* Add one page cpu-cost to the startup cost */
9082 : 1536 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
9083 : :
9084 : : /*
9085 : : * Add once again a CPU-cost for those data pages, before amortizing for
9086 : : * cache.
9087 : : */
9088 : 1536 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
9089 : :
9090 : : /* Account for cache effects, the same as above */
9091 [ + - + + ]: 1536 : if (outer_scans > 1 || counts.arrayScans > 1)
9092 : : {
9093 : 5 : dataPagesFetched *= outer_scans * counts.arrayScans;
9094 : 5 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
9095 : : (BlockNumber) numDataPages,
9096 : : numDataPages, root);
9097 : 5 : dataPagesFetched /= outer_scans;
9098 : : }
9099 : :
9100 : : /* And apply random_page_cost as the cost per page */
9101 : 1536 : *indexTotalCost += *indexStartupCost +
9102 : 1536 : dataPagesFetched * spc_random_page_cost;
9103 : :
9104 : : /*
9105 : : * Add on index qual eval costs, much as in genericcostestimate. We charge
9106 : : * cpu but we can disregard indexorderbys, since GIN doesn't support
9107 : : * those.
9108 : : */
9109 : 1536 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
9110 : 1536 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
9111 : :
9112 : 1536 : *indexStartupCost += qual_arg_cost;
9113 : 1536 : *indexTotalCost += qual_arg_cost;
9114 : :
9115 : : /*
9116 : : * Add a cpu cost per search entry, corresponding to the actual visited
9117 : : * entries.
9118 : : */
9119 : 1536 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
9120 : : /* Now add a cpu cost per tuple in the posting lists / trees */
9121 : 1536 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
9122 : 1536 : *indexPages = dataPagesFetched;
9123 : : }
9124 : :
9125 : : /*
9126 : : * BRIN has search behavior completely different from other index types
9127 : : */
9128 : : void
9129 : 8944 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
9130 : : Cost *indexStartupCost, Cost *indexTotalCost,
9131 : : Selectivity *indexSelectivity, double *indexCorrelation,
9132 : : double *indexPages)
9133 : : {
9134 : 8944 : IndexOptInfo *index = path->indexinfo;
9135 : 8944 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
9136 : 8944 : double numPages = index->pages;
9137 : 8944 : RelOptInfo *baserel = index->rel;
9138 [ + - ]: 8944 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
9139 : : Cost spc_seq_page_cost;
9140 : : Cost spc_random_page_cost;
9141 : : double qual_arg_cost;
9142 : : double qualSelectivity;
9143 : : BrinStatsData statsData;
9144 : : double indexRanges;
9145 : : double minimalRanges;
9146 : : double estimatedRanges;
9147 : : double selec;
9148 : : Relation indexRel;
9149 : : ListCell *l;
9150 : : VariableStatData vardata;
9151 : :
9152 : : Assert(rte->rtekind == RTE_RELATION);
9153 : :
9154 : : /* fetch estimated page cost for the tablespace containing the index */
9155 : 8944 : get_tablespace_page_costs(index->reltablespace,
9156 : : &spc_random_page_cost,
9157 : : &spc_seq_page_cost);
9158 : :
9159 : : /*
9160 : : * Obtain some data from the index itself, if possible. Otherwise invent
9161 : : * some plausible internal statistics based on the relation page count.
9162 : : */
9163 [ + - ]: 8944 : if (!index->hypothetical)
9164 : : {
9165 : : /*
9166 : : * A lock should have already been obtained on the index in plancat.c.
9167 : : */
9168 : 8944 : indexRel = index_open(index->indexoid, NoLock);
9169 : 8944 : brinGetStats(indexRel, &statsData);
9170 : 8944 : index_close(indexRel, NoLock);
9171 : :
9172 : : /* work out the actual number of ranges in the index */
9173 [ + + ]: 8944 : indexRanges = Max(ceil((double) baserel->pages /
9174 : : statsData.pagesPerRange), 1.0);
9175 : : }
9176 : : else
9177 : : {
9178 : : /*
9179 : : * Assume default number of pages per range, and estimate the number
9180 : : * of ranges based on that.
9181 : : */
9182 [ # # ]: 0 : indexRanges = Max(ceil((double) baserel->pages /
9183 : : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
9184 : :
9185 : 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
9186 : 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
9187 : : }
9188 : :
9189 : : /*
9190 : : * Compute index correlation
9191 : : *
9192 : : * Because we can use all index quals equally when scanning, we can use
9193 : : * the largest correlation (in absolute value) among columns used by the
9194 : : * query. Start at zero, the worst possible case. If we cannot find any
9195 : : * correlation statistics, we will keep it as 0.
9196 : : */
9197 : 8944 : *indexCorrelation = 0;
9198 : :
9199 [ + - + + : 17889 : foreach(l, path->indexclauses)
+ + ]
9200 : : {
9201 : 8945 : IndexClause *iclause = lfirst_node(IndexClause, l);
9202 : 8945 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
9203 : :
9204 : : /* attempt to lookup stats in relation for this index column */
9205 [ + - ]: 8945 : if (attnum != 0)
9206 : : {
9207 : : /* Simple variable -- look to stats for the underlying table */
9208 [ - + - - ]: 8945 : if (get_relation_stats_hook &&
9209 : 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
9210 : : {
9211 : : /*
9212 : : * The hook took control of acquiring a stats tuple. If it
9213 : : * did supply a tuple, it'd better have supplied a freefunc.
9214 : : */
9215 [ # # # # ]: 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
9216 [ # # ]: 0 : elog(ERROR,
9217 : : "no function provided to release variable stats with");
9218 : : }
9219 : : else
9220 : : {
9221 : 8945 : vardata.statsTuple =
9222 : 8945 : SearchSysCache3(STATRELATTINH,
9223 : : ObjectIdGetDatum(rte->relid),
9224 : : Int16GetDatum(attnum),
9225 : : BoolGetDatum(false));
9226 : 8945 : vardata.freefunc = ReleaseSysCache;
9227 : : }
9228 : : }
9229 : : else
9230 : : {
9231 : : /*
9232 : : * Looks like we've found an expression column in the index. Let's
9233 : : * see if there's any stats for it.
9234 : : */
9235 : :
9236 : : /* get the attnum from the 0-based index. */
9237 : 0 : attnum = iclause->indexcol + 1;
9238 : :
9239 [ # # # # ]: 0 : if (get_index_stats_hook &&
9240 : 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
9241 : : {
9242 : : /*
9243 : : * The hook took control of acquiring a stats tuple. If it
9244 : : * did supply a tuple, it'd better have supplied a freefunc.
9245 : : */
9246 [ # # ]: 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
9247 [ # # ]: 0 : !vardata.freefunc)
9248 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
9249 : : }
9250 : : else
9251 : : {
9252 : 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
9253 : : ObjectIdGetDatum(index->indexoid),
9254 : : Int16GetDatum(attnum),
9255 : : BoolGetDatum(false));
9256 : 0 : vardata.freefunc = ReleaseSysCache;
9257 : : }
9258 : : }
9259 : :
9260 [ + + ]: 8945 : if (HeapTupleIsValid(vardata.statsTuple))
9261 : : {
9262 : : AttStatsSlot sslot;
9263 : :
9264 [ + - ]: 33 : if (get_attstatsslot(&sslot, vardata.statsTuple,
9265 : : STATISTIC_KIND_CORRELATION, InvalidOid,
9266 : : ATTSTATSSLOT_NUMBERS))
9267 : : {
9268 : 33 : double varCorrelation = 0.0;
9269 : :
9270 [ + - ]: 33 : if (sslot.nnumbers > 0)
9271 : 33 : varCorrelation = fabs(sslot.numbers[0]);
9272 : :
9273 [ + - ]: 33 : if (varCorrelation > *indexCorrelation)
9274 : 33 : *indexCorrelation = varCorrelation;
9275 : :
9276 : 33 : free_attstatsslot(&sslot);
9277 : : }
9278 : : }
9279 : :
9280 [ + + ]: 8945 : ReleaseVariableStats(vardata);
9281 : : }
9282 : :
9283 : 8944 : qualSelectivity = clauselist_selectivity(root, indexQuals,
9284 : 8944 : baserel->relid,
9285 : : JOIN_INNER, NULL);
9286 : :
9287 : : /*
9288 : : * Now calculate the minimum possible ranges we could match with if all of
9289 : : * the rows were in the perfect order in the table's heap.
9290 : : */
9291 : 8944 : minimalRanges = ceil(indexRanges * qualSelectivity);
9292 : :
9293 : : /*
9294 : : * Now estimate the number of ranges that we'll touch by using the
9295 : : * indexCorrelation from the stats. Careful not to divide by zero (note
9296 : : * we're using the absolute value of the correlation).
9297 : : */
9298 [ + + ]: 8944 : if (*indexCorrelation < 1.0e-10)
9299 : 8911 : estimatedRanges = indexRanges;
9300 : : else
9301 [ + + ]: 33 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
9302 : :
9303 : : /* we expect to visit this portion of the table */
9304 : 8944 : selec = estimatedRanges / indexRanges;
9305 : :
9306 [ - + - + ]: 8944 : CLAMP_PROBABILITY(selec);
9307 : :
9308 : 8944 : *indexSelectivity = selec;
9309 : :
9310 : : /*
9311 : : * Compute the index qual costs, much as in genericcostestimate, to add to
9312 : : * the index costs. We can disregard indexorderbys, since BRIN doesn't
9313 : : * support those.
9314 : : */
9315 : 8944 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
9316 : :
9317 : : /*
9318 : : * Compute the startup cost as the cost to read the whole revmap
9319 : : * sequentially, including the cost to execute the index quals.
9320 : : */
9321 : 8944 : *indexStartupCost =
9322 : 8944 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
9323 : 8944 : *indexStartupCost += qual_arg_cost;
9324 : :
9325 : : /*
9326 : : * To read a BRIN index there might be a bit of back and forth over
9327 : : * regular pages, as revmap might point to them out of sequential order;
9328 : : * calculate the total cost as reading the whole index in random order.
9329 : : */
9330 : 8944 : *indexTotalCost = *indexStartupCost +
9331 : 8944 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
9332 : :
9333 : : /*
9334 : : * Charge a small amount per range tuple which we expect to match to. This
9335 : : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
9336 : : * will set a bit for each page in the range when we find a matching
9337 : : * range, so we must multiply the charge by the number of pages in the
9338 : : * range.
9339 : : */
9340 : 8944 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
9341 : 8944 : statsData.pagesPerRange;
9342 : :
9343 : 8944 : *indexPages = index->pages;
9344 : 8944 : }
|
