Age Owner Branch data TLA 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 examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
251 : : int indexcol, VariableStatData *vardata);
252 : : static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata,
253 : : Oid sortop, Oid collation,
254 : : Datum *min, Datum *max);
255 : : static void get_stats_slot_range(AttStatsSlot *sslot,
256 : : Oid opfuncoid, FmgrInfo *opproc,
257 : : Oid collation, int16 typLen, bool typByVal,
258 : : Datum *min, Datum *max, bool *p_have_data);
259 : : static bool get_actual_variable_range(PlannerInfo *root,
260 : : VariableStatData *vardata,
261 : : Oid sortop, Oid collation,
262 : : Datum *min, Datum *max);
263 : : static bool get_actual_variable_endpoint(Relation heapRel,
264 : : Relation indexRel,
265 : : ScanDirection indexscandir,
266 : : ScanKey scankeys,
267 : : int16 typLen,
268 : : bool typByVal,
269 : : TupleTableSlot *tableslot,
270 : : MemoryContext outercontext,
271 : : Datum *endpointDatum);
272 : : static RelOptInfo *find_join_input_rel(PlannerInfo *root, Relids relids);
273 : : static double btcost_correlation(IndexOptInfo *index,
274 : : VariableStatData *vardata);
275 : :
276 : : /* Define support routines for MCV hash tables */
277 : : #define SH_PREFIX MCVHashTable
278 : : #define SH_ELEMENT_TYPE MCVHashEntry
279 : : #define SH_KEY_TYPE Datum
280 : : #define SH_KEY value
281 : : #define SH_HASH_KEY(tab,key) hash_mcv(tab, key)
282 : : #define SH_EQUAL(tab,key0,key1) mcvs_equal(tab, key0, key1)
283 : : #define SH_SCOPE static inline
284 : : #define SH_STORE_HASH
285 : : #define SH_GET_HASH(tab,ent) (ent)->hash
286 : : #define SH_DEFINE
287 : : #define SH_DECLARE
288 : : #include "lib/simplehash.h"
289 : :
290 : :
291 : : /*
292 : : * eqsel - Selectivity of "=" for any data types.
293 : : *
294 : : * Note: this routine is also used to estimate selectivity for some
295 : : * operators that are not "=" but have comparable selectivity behavior,
296 : : * such as "~=" (geometric approximate-match). Even for "=", we must
297 : : * keep in mind that the left and right datatypes may differ.
298 : : */
299 : : Datum
9465 tgl@sss.pgh.pa.us 300 :CBC 563147 : eqsel(PG_FUNCTION_ARGS)
301 : : {
3258 302 : 563147 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
303 : : }
304 : :
305 : : /*
306 : : * Common code for eqsel() and neqsel()
307 : : */
308 : : static double
309 : 595595 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
310 : : {
7639 311 : 595595 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
9116 312 : 595595 : Oid operator = PG_GETARG_OID(1);
313 : 595595 : List *args = (List *) PG_GETARG_POINTER(2);
314 : 595595 : int varRelid = PG_GETARG_INT32(3);
2160 315 : 595595 : Oid collation = PG_GET_COLLATION();
316 : : VariableStatData vardata;
317 : : Node *other;
318 : : bool varonleft;
319 : : double selec;
320 : :
321 : : /*
322 : : * When asked about <>, we do the estimation using the corresponding =
323 : : * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
324 : : */
3258 325 [ + + ]: 595595 : if (negate)
326 : : {
327 : 32448 : operator = get_negator(operator);
328 [ - + ]: 32448 : if (!OidIsValid(operator))
329 : : {
330 : : /* Use default selectivity (should we raise an error instead?) */
3258 tgl@sss.pgh.pa.us 331 :UBC 0 : return 1.0 - DEFAULT_EQ_SEL;
332 : : }
333 : : }
334 : :
335 : : /*
336 : : * If expression is not variable = something or something = variable, then
337 : : * punt and return a default estimate.
338 : : */
8113 tgl@sss.pgh.pa.us 339 [ + + ]:CBC 595595 : if (!get_restriction_variable(root, args, varRelid,
340 : : &vardata, &other, &varonleft))
3258 341 [ + + ]: 3626 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
342 : :
343 : : /*
344 : : * We can do a lot better if the something is a constant. (Note: the
345 : : * Const might result from estimation rather than being a simple constant
346 : : * in the query.)
347 : : */
6632 348 [ + + ]: 591965 : if (IsA(other, Const))
2160 349 : 224405 : selec = var_eq_const(&vardata, operator, collation,
6632 350 : 224405 : ((Const *) other)->constvalue,
351 : 224405 : ((Const *) other)->constisnull,
352 : : varonleft, negate);
353 : : else
2160 354 : 367560 : selec = var_eq_non_const(&vardata, operator, collation, other,
355 : : varonleft, negate);
356 : :
6632 357 [ + + ]: 591965 : ReleaseVariableStats(vardata);
358 : :
3258 359 : 591965 : return selec;
360 : : }
361 : :
362 : : /*
363 : : * var_eq_const --- eqsel for var = const case
364 : : *
365 : : * This is exported so that some other estimation functions can use it.
366 : : */
367 : : double
1323 pg@bowt.ie 368 : 255523 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
369 : : Datum constval, bool constisnull,
370 : : bool varonleft, bool negate)
371 : : {
372 : : double selec;
3258 tgl@sss.pgh.pa.us 373 : 255523 : double nullfrac = 0.0;
374 : : bool isdefault;
375 : : Oid opfuncoid;
376 : :
377 : : /*
378 : : * If the constant is NULL, assume operator is strict and return zero, ie,
379 : : * operator will never return TRUE. (It's zero even for a negator op.)
380 : : */
6632 381 [ + + ]: 255523 : if (constisnull)
382 : 270 : return 0.0;
383 : :
384 : : /*
385 : : * Grab the nullfrac for use below. Note we allow use of nullfrac
386 : : * regardless of security check.
387 : : */
3258 388 [ + + ]: 255253 : if (HeapTupleIsValid(vardata->statsTuple))
389 : : {
390 : : Form_pg_statistic stats;
391 : :
392 : 189036 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
393 : 189036 : nullfrac = stats->stanullfrac;
394 : : }
395 : :
396 : : /*
397 : : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
398 : : * assume there is exactly one match regardless of anything else. (This
399 : : * is slightly bogus, since the index or clause's equality operator might
400 : : * be different from ours, but it's much more likely to be right than
401 : : * ignoring the information.)
402 : : */
6288 403 [ + + + - : 255253 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
+ + ]
404 : : {
3258 405 : 52471 : selec = 1.0 / vardata->rel->tuples;
406 : : }
407 [ + + + - ]: 348301 : else if (HeapTupleIsValid(vardata->statsTuple) &&
408 : 145519 : statistic_proc_security_check(vardata,
1323 pg@bowt.ie 409 : 145519 : (opfuncoid = get_opcode(oproid))))
9774 tgl@sss.pgh.pa.us 410 : 145519 : {
411 : : AttStatsSlot sslot;
6632 412 : 145519 : bool match = false;
413 : : int i;
414 : :
415 : : /*
416 : : * Is the constant "=" to any of the column's most common values?
417 : : * (Although the given operator may not really be "=", we will assume
418 : : * that seeing whether it returns TRUE is an appropriate test. If you
419 : : * don't like this, maybe you shouldn't be using eqsel for your
420 : : * operator...)
421 : : */
3279 422 [ + + ]: 145519 : if (get_attstatsslot(&sslot, vardata->statsTuple,
423 : : STATISTIC_KIND_MCV, InvalidOid,
424 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
425 : : {
2205 426 : 134321 : LOCAL_FCINFO(fcinfo, 2);
427 : : FmgrInfo eqproc;
428 : :
3287 peter_e@gmx.net 429 : 134321 : fmgr_info(opfuncoid, &eqproc);
430 : :
431 : : /*
432 : : * Save a few cycles by setting up the fcinfo struct just once.
433 : : * Using FunctionCallInvoke directly also avoids failure if the
434 : : * eqproc returns NULL, though really equality functions should
435 : : * never do that.
436 : : */
2160 tgl@sss.pgh.pa.us 437 : 134321 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
438 : : NULL, NULL);
2205 439 : 134321 : fcinfo->args[0].isnull = false;
440 : 134321 : fcinfo->args[1].isnull = false;
441 : : /* be careful to apply operator right way 'round */
442 [ + + ]: 134321 : if (varonleft)
443 : 134294 : fcinfo->args[1].value = constval;
444 : : else
445 : 27 : fcinfo->args[0].value = constval;
446 : :
3279 447 [ + + ]: 2339349 : for (i = 0; i < sslot.nvalues; i++)
448 : : {
449 : : Datum fresult;
450 : :
6632 451 [ + + ]: 2276146 : if (varonleft)
2205 452 : 2276075 : fcinfo->args[0].value = sslot.values[i];
453 : : else
454 : 71 : fcinfo->args[1].value = sslot.values[i];
455 : 2276146 : fcinfo->isnull = false;
456 : 2276146 : fresult = FunctionCallInvoke(fcinfo);
457 [ + - + + ]: 2276146 : if (!fcinfo->isnull && DatumGetBool(fresult))
458 : : {
459 : 71118 : match = true;
6632 460 : 71118 : break;
461 : : }
462 : : }
463 : : }
464 : : else
465 : : {
466 : : /* no most-common-value info available */
3279 467 : 11198 : i = 0; /* keep compiler quiet */
468 : : }
469 : :
6632 470 [ + + ]: 145519 : if (match)
471 : : {
472 : : /*
473 : : * Constant is "=" to this common value. We know selectivity
474 : : * exactly (or as exactly as ANALYZE could calculate it, anyway).
475 : : */
3279 476 : 71118 : selec = sslot.numbers[i];
477 : : }
478 : : else
479 : : {
480 : : /*
481 : : * Comparison is against a constant that is neither NULL nor any
482 : : * of the common values. Its selectivity cannot be more than
483 : : * this:
484 : : */
6632 485 : 74401 : double sumcommon = 0.0;
486 : : double otherdistinct;
487 : :
3279 488 [ + + ]: 1987733 : for (i = 0; i < sslot.nnumbers; i++)
489 : 1913332 : sumcommon += sslot.numbers[i];
3258 490 : 74401 : selec = 1.0 - sumcommon - nullfrac;
6632 491 [ + + - + ]: 74401 : CLAMP_PROBABILITY(selec);
492 : :
493 : : /*
494 : : * and in fact it's probably a good deal less. We approximate that
495 : : * all the not-common values share this remaining fraction
496 : : * equally, so we divide by the number of other distinct values.
497 : : */
3279 498 : 74401 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
499 : 74401 : sslot.nnumbers;
6632 500 [ + + ]: 74401 : if (otherdistinct > 1)
501 : 36663 : selec /= otherdistinct;
502 : :
503 : : /*
504 : : * Another cross-check: selectivity shouldn't be estimated as more
505 : : * than the least common "most common value".
506 : : */
3279 507 [ + + - + ]: 74401 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
3279 tgl@sss.pgh.pa.us 508 :UBC 0 : selec = sslot.numbers[sslot.nnumbers - 1];
509 : : }
510 : :
3279 tgl@sss.pgh.pa.us 511 :CBC 145519 : free_attstatsslot(&sslot);
512 : : }
513 : : else
514 : : {
515 : : /*
516 : : * No ANALYZE stats available, so make a guess using estimated number
517 : : * of distinct values and assuming they are equally common. (The guess
518 : : * is unlikely to be very good, but we do know a few special cases.)
519 : : */
5357 520 : 57263 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
521 : : }
522 : :
523 : : /* now adjust if we wanted <> rather than = */
3258 524 [ + + ]: 255253 : if (negate)
525 : 25695 : selec = 1.0 - selec - nullfrac;
526 : :
527 : : /* result should be in range, but make sure... */
6632 528 [ - + - + ]: 255253 : CLAMP_PROBABILITY(selec);
529 : :
530 : 255253 : return selec;
531 : : }
532 : :
533 : : /*
534 : : * var_eq_non_const --- eqsel for var = something-other-than-const case
535 : : *
536 : : * This is exported so that some other estimation functions can use it.
537 : : */
538 : : double
1323 pg@bowt.ie 539 : 367560 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
540 : : Node *other,
541 : : bool varonleft, bool negate)
542 : : {
543 : : double selec;
3258 tgl@sss.pgh.pa.us 544 : 367560 : double nullfrac = 0.0;
545 : : bool isdefault;
546 : :
547 : : /*
548 : : * Grab the nullfrac for use below.
549 : : */
550 [ + + ]: 367560 : if (HeapTupleIsValid(vardata->statsTuple))
551 : : {
552 : : Form_pg_statistic stats;
553 : :
554 : 242335 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
555 : 242335 : nullfrac = stats->stanullfrac;
556 : : }
557 : :
558 : : /*
559 : : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
560 : : * assume there is exactly one match regardless of anything else. (This
561 : : * is slightly bogus, since the index or clause's equality operator might
562 : : * be different from ours, but it's much more likely to be right than
563 : : * ignoring the information.)
564 : : */
6288 565 [ + + + - : 367560 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
+ + ]
566 : : {
3258 567 : 129629 : selec = 1.0 / vardata->rel->tuples;
568 : : }
569 [ + + ]: 237931 : else if (HeapTupleIsValid(vardata->statsTuple))
570 : : {
571 : : double ndistinct;
572 : : AttStatsSlot sslot;
573 : :
574 : : /*
575 : : * Search is for a value that we do not know a priori, but we will
576 : : * assume it is not NULL. Estimate the selectivity as non-null
577 : : * fraction divided by number of distinct values, so that we get a
578 : : * result averaged over all possible values whether common or
579 : : * uncommon. (Essentially, we are assuming that the not-yet-known
580 : : * comparison value is equally likely to be any of the possible
581 : : * values, regardless of their frequency in the table. Is that a good
582 : : * idea?)
583 : : */
584 : 130518 : selec = 1.0 - nullfrac;
5357 585 : 130518 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
6632 586 [ + + ]: 130518 : if (ndistinct > 1)
587 : 127906 : selec /= ndistinct;
588 : :
589 : : /*
590 : : * Cross-check: selectivity should never be estimated as more than the
591 : : * most common value's.
592 : : */
3279 593 [ + + ]: 130518 : if (get_attstatsslot(&sslot, vardata->statsTuple,
594 : : STATISTIC_KIND_MCV, InvalidOid,
595 : : ATTSTATSSLOT_NUMBERS))
596 : : {
597 [ + - + + ]: 114707 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
598 : 483 : selec = sslot.numbers[0];
599 : 114707 : free_attstatsslot(&sslot);
600 : : }
601 : : }
602 : : else
603 : : {
604 : : /*
605 : : * No ANALYZE stats available, so make a guess using estimated number
606 : : * of distinct values and assuming they are equally common. (The guess
607 : : * is unlikely to be very good, but we do know a few special cases.)
608 : : */
5357 609 : 107413 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
610 : : }
611 : :
612 : : /* now adjust if we wanted <> rather than = */
3258 613 [ + + ]: 367560 : if (negate)
614 : 5160 : selec = 1.0 - selec - nullfrac;
615 : :
616 : : /* result should be in range, but make sure... */
8888 617 [ - + - + ]: 367560 : CLAMP_PROBABILITY(selec);
618 : :
6632 619 : 367560 : return selec;
620 : : }
621 : :
622 : : /*
623 : : * neqsel - Selectivity of "!=" for any data types.
624 : : *
625 : : * This routine is also used for some operators that are not "!="
626 : : * but have comparable selectivity behavior. See above comments
627 : : * for eqsel().
628 : : */
629 : : Datum
9465 630 : 32448 : neqsel(PG_FUNCTION_ARGS)
631 : : {
3258 632 : 32448 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
633 : : }
634 : :
635 : : /*
636 : : * scalarineqsel - Selectivity of "<", "<=", ">", ">=" for scalars.
637 : : *
638 : : * This is the guts of scalarltsel/scalarlesel/scalargtsel/scalargesel.
639 : : * The isgt and iseq flags distinguish which of the four cases apply.
640 : : *
641 : : * The caller has commuted the clause, if necessary, so that we can treat
642 : : * the variable as being on the left. The caller must also make sure that
643 : : * the other side of the clause is a non-null Const, and dissect that into
644 : : * a value and datatype. (This definition simplifies some callers that
645 : : * want to estimate against a computed value instead of a Const node.)
646 : : *
647 : : * This routine works for any datatype (or pair of datatypes) known to
648 : : * convert_to_scalar(). If it is applied to some other datatype,
649 : : * it will return an approximate estimate based on assuming that the constant
650 : : * value falls in the middle of the bin identified by binary search.
651 : : */
652 : : static double
3156 653 : 252684 : scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
654 : : Oid collation,
655 : : VariableStatData *vardata, Datum constval, Oid consttype)
656 : : {
657 : : Form_pg_statistic stats;
658 : : FmgrInfo opproc;
659 : : double mcv_selec,
660 : : hist_selec,
661 : : sumcommon;
662 : : double selec;
663 : :
8113 664 [ + + ]: 252684 : if (!HeapTupleIsValid(vardata->statsTuple))
665 : : {
666 : : /*
667 : : * No stats are available. Typically this means we have to fall back
668 : : * on the default estimate; but if the variable is CTID then we can
669 : : * make an estimate based on comparing the constant to the table size.
670 : : */
2598 671 [ + - + + ]: 21439 : if (vardata->var && IsA(vardata->var, Var) &&
672 [ + + ]: 17403 : ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
673 : : {
674 : : ItemPointer itemptr;
675 : : double block;
676 : : double density;
677 : :
678 : : /*
679 : : * If the relation's empty, we're going to include all of it.
680 : : * (This is mostly to avoid divide-by-zero below.)
681 : : */
682 [ - + ]: 1680 : if (vardata->rel->pages == 0)
2598 tgl@sss.pgh.pa.us 683 :UBC 0 : return 1.0;
684 : :
2598 tgl@sss.pgh.pa.us 685 :CBC 1680 : itemptr = (ItemPointer) DatumGetPointer(constval);
686 : 1680 : block = ItemPointerGetBlockNumberNoCheck(itemptr);
687 : :
688 : : /*
689 : : * Determine the average number of tuples per page (density).
690 : : *
691 : : * Since the last page will, on average, be only half full, we can
692 : : * estimate it to have half as many tuples as earlier pages. So
693 : : * give it half the weight of a regular page.
694 : : */
695 : 1680 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
696 : :
697 : : /* If target is the last page, use half the density. */
698 [ + + ]: 1680 : if (block >= vardata->rel->pages - 1)
699 : 25 : density *= 0.5;
700 : :
701 : : /*
702 : : * Using the average tuples per page, calculate how far into the
703 : : * page the itemptr is likely to be and adjust block accordingly,
704 : : * by adding that fraction of a whole block (but never more than a
705 : : * whole block, no matter how high the itemptr's offset is). Here
706 : : * we are ignoring the possibility of dead-tuple line pointers,
707 : : * which is fairly bogus, but we lack the info to do better.
708 : : */
709 [ + - ]: 1680 : if (density > 0.0)
710 : : {
711 : 1680 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
712 : :
713 [ + + ]: 1680 : block += Min(offset / density, 1.0);
714 : : }
715 : :
716 : : /*
717 : : * Convert relative block number to selectivity. Again, the last
718 : : * page has only half weight.
719 : : */
720 : 1680 : selec = block / (vardata->rel->pages - 0.5);
721 : :
722 : : /*
723 : : * The calculation so far gave us a selectivity for the "<=" case.
724 : : * We'll have one fewer tuple for "<" and one additional tuple for
725 : : * ">=", the latter of which we'll reverse the selectivity for
726 : : * below, so we can simply subtract one tuple for both cases. The
727 : : * cases that need this adjustment can be identified by iseq being
728 : : * equal to isgt.
729 : : */
730 [ + + + - ]: 1680 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
731 : 1563 : selec -= (1.0 / vardata->rel->tuples);
732 : :
733 : : /* Finally, reverse the selectivity for the ">", ">=" cases. */
734 [ + + ]: 1680 : if (isgt)
735 : 1547 : selec = 1.0 - selec;
736 : :
737 [ + + - + ]: 1680 : CLAMP_PROBABILITY(selec);
738 : 1680 : return selec;
739 : : }
740 : :
741 : : /* no stats available, so default result */
9116 742 : 19759 : return DEFAULT_INEQ_SEL;
743 : : }
8113 744 : 231245 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
745 : :
9116 746 : 231245 : fmgr_info(get_opcode(operator), &opproc);
747 : :
748 : : /*
749 : : * If we have most-common-values info, add up the fractions of the MCV
750 : : * entries that satisfy MCV OP CONST. These fractions contribute directly
751 : : * to the result selectivity. Also add up the total fraction represented
752 : : * by MCV entries.
753 : : */
2160 754 : 231245 : mcv_selec = mcv_selectivity(vardata, &opproc, collation, constval, true,
755 : : &sumcommon);
756 : :
757 : : /*
758 : : * If there is a histogram, determine which bin the constant falls in, and
759 : : * compute the resulting contribution to selectivity.
760 : : */
3156 761 : 231245 : hist_selec = ineq_histogram_selectivity(root, vardata,
762 : : operator, &opproc, isgt, iseq,
763 : : collation,
764 : : constval, consttype);
765 : :
766 : : /*
767 : : * Now merge the results from the MCV and histogram calculations,
768 : : * realizing that the histogram covers only the non-null values that are
769 : : * not listed in MCV.
770 : : */
7420 771 : 231245 : selec = 1.0 - stats->stanullfrac - sumcommon;
772 : :
5965 773 [ + + ]: 231245 : if (hist_selec >= 0.0)
7420 774 : 146207 : selec *= hist_selec;
775 : : else
776 : : {
777 : : /*
778 : : * If no histogram but there are values not accounted for by MCV,
779 : : * arbitrarily assume half of them will match.
780 : : */
781 : 85038 : selec *= 0.5;
782 : : }
783 : :
784 : 231245 : selec += mcv_selec;
785 : :
786 : : /* result should be in range, but make sure... */
787 [ + + + + ]: 231245 : CLAMP_PROBABILITY(selec);
788 : :
789 : 231245 : return selec;
790 : : }
791 : :
792 : : /*
793 : : * mcv_selectivity - Examine the MCV list for selectivity estimates
794 : : *
795 : : * Determine the fraction of the variable's MCV population that satisfies
796 : : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft. Also
797 : : * compute the fraction of the total column population represented by the MCV
798 : : * list. This code will work for any boolean-returning predicate operator.
799 : : *
800 : : * The function result is the MCV selectivity, and the fraction of the
801 : : * total population is returned into *sumcommonp. Zeroes are returned
802 : : * if there is no MCV list.
803 : : */
804 : : double
2160 805 : 235647 : mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation,
806 : : Datum constval, bool varonleft,
807 : : double *sumcommonp)
808 : : {
809 : : double mcv_selec,
810 : : sumcommon;
811 : : AttStatsSlot sslot;
812 : : int i;
813 : :
9129 814 : 235647 : mcv_selec = 0.0;
815 : 235647 : sumcommon = 0.0;
816 : :
7420 817 [ + + + + ]: 469499 : if (HeapTupleIsValid(vardata->statsTuple) &&
3287 peter_e@gmx.net 818 [ + + ]: 467429 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
3279 tgl@sss.pgh.pa.us 819 : 233577 : get_attstatsslot(&sslot, vardata->statsTuple,
820 : : STATISTIC_KIND_MCV, InvalidOid,
821 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
822 : : {
2205 823 : 132053 : LOCAL_FCINFO(fcinfo, 2);
824 : :
825 : : /*
826 : : * We invoke the opproc "by hand" so that we won't fail on NULL
827 : : * results. Such cases won't arise for normal comparison functions,
828 : : * but generic_restriction_selectivity could perhaps be used with
829 : : * operators that can return NULL. A small side benefit is to not
830 : : * need to re-initialize the fcinfo struct from scratch each time.
831 : : */
2160 832 : 132053 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
833 : : NULL, NULL);
2205 834 : 132053 : fcinfo->args[0].isnull = false;
835 : 132053 : fcinfo->args[1].isnull = false;
836 : : /* be careful to apply operator right way 'round */
837 [ + - ]: 132053 : if (varonleft)
838 : 132053 : fcinfo->args[1].value = constval;
839 : : else
2205 tgl@sss.pgh.pa.us 840 :UBC 0 : fcinfo->args[0].value = constval;
841 : :
3279 tgl@sss.pgh.pa.us 842 [ + + ]:CBC 3297594 : for (i = 0; i < sslot.nvalues; i++)
843 : : {
844 : : Datum fresult;
845 : :
2205 846 [ + - ]: 3165541 : if (varonleft)
847 : 3165541 : fcinfo->args[0].value = sslot.values[i];
848 : : else
2205 tgl@sss.pgh.pa.us 849 :UBC 0 : fcinfo->args[1].value = sslot.values[i];
2205 tgl@sss.pgh.pa.us 850 :CBC 3165541 : fcinfo->isnull = false;
851 : 3165541 : fresult = FunctionCallInvoke(fcinfo);
852 [ + - + + ]: 3165541 : if (!fcinfo->isnull && DatumGetBool(fresult))
3279 853 : 1329110 : mcv_selec += sslot.numbers[i];
854 : 3165541 : sumcommon += sslot.numbers[i];
855 : : }
856 : 132053 : free_attstatsslot(&sslot);
857 : : }
858 : :
7420 859 : 235647 : *sumcommonp = sumcommon;
860 : 235647 : return mcv_selec;
861 : : }
862 : :
863 : : /*
864 : : * histogram_selectivity - Examine the histogram for selectivity estimates
865 : : *
866 : : * Determine the fraction of the variable's histogram entries that satisfy
867 : : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft.
868 : : *
869 : : * This code will work for any boolean-returning predicate operator, whether
870 : : * or not it has anything to do with the histogram sort operator. We are
871 : : * essentially using the histogram just as a representative sample. However,
872 : : * small histograms are unlikely to be all that representative, so the caller
873 : : * should be prepared to fall back on some other estimation approach when the
874 : : * histogram is missing or very small. It may also be prudent to combine this
875 : : * approach with another one when the histogram is small.
876 : : *
877 : : * If the actual histogram size is not at least min_hist_size, we won't bother
878 : : * to do the calculation at all. Also, if the n_skip parameter is > 0, we
879 : : * ignore the first and last n_skip histogram elements, on the grounds that
880 : : * they are outliers and hence not very representative. Typical values for
881 : : * these parameters are 10 and 1.
882 : : *
883 : : * The function result is the selectivity, or -1 if there is no histogram
884 : : * or it's smaller than min_hist_size.
885 : : *
886 : : * The output parameter *hist_size receives the actual histogram size,
887 : : * or zero if no histogram. Callers may use this number to decide how
888 : : * much faith to put in the function result.
889 : : *
890 : : * Note that the result disregards both the most-common-values (if any) and
891 : : * null entries. The caller is expected to combine this result with
892 : : * statistics for those portions of the column population. It may also be
893 : : * prudent to clamp the result range, ie, disbelieve exact 0 or 1 outputs.
894 : : */
895 : : double
2160 896 : 4402 : histogram_selectivity(VariableStatData *vardata,
897 : : FmgrInfo *opproc, Oid collation,
898 : : Datum constval, bool varonleft,
899 : : int min_hist_size, int n_skip,
900 : : int *hist_size)
901 : : {
902 : : double result;
903 : : AttStatsSlot sslot;
904 : :
905 : : /* check sanity of parameters */
7167 906 [ - + ]: 4402 : Assert(n_skip >= 0);
907 [ - + ]: 4402 : Assert(min_hist_size > 2 * n_skip);
908 : :
909 [ + + + + ]: 7009 : if (HeapTupleIsValid(vardata->statsTuple) &&
3287 peter_e@gmx.net 910 [ + + ]: 5209 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
3279 tgl@sss.pgh.pa.us 911 : 2602 : get_attstatsslot(&sslot, vardata->statsTuple,
912 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
913 : : ATTSTATSSLOT_VALUES))
914 : : {
915 : 2524 : *hist_size = sslot.nvalues;
916 [ + + ]: 2524 : if (sslot.nvalues >= min_hist_size)
917 : : {
2205 918 : 1475 : LOCAL_FCINFO(fcinfo, 2);
7167 919 : 1475 : int nmatch = 0;
920 : : int i;
921 : :
922 : : /*
923 : : * We invoke the opproc "by hand" so that we won't fail on NULL
924 : : * results. Such cases won't arise for normal comparison
925 : : * functions, but generic_restriction_selectivity could perhaps be
926 : : * used with operators that can return NULL. A small side benefit
927 : : * is to not need to re-initialize the fcinfo struct from scratch
928 : : * each time.
929 : : */
2160 930 : 1475 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
931 : : NULL, NULL);
2205 932 : 1475 : fcinfo->args[0].isnull = false;
933 : 1475 : fcinfo->args[1].isnull = false;
934 : : /* be careful to apply operator right way 'round */
935 [ + - ]: 1475 : if (varonleft)
936 : 1475 : fcinfo->args[1].value = constval;
937 : : else
2205 tgl@sss.pgh.pa.us 938 :UBC 0 : fcinfo->args[0].value = constval;
939 : :
3279 tgl@sss.pgh.pa.us 940 [ + + ]:CBC 124350 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
941 : : {
942 : : Datum fresult;
943 : :
2205 944 [ + - ]: 122875 : if (varonleft)
945 : 122875 : fcinfo->args[0].value = sslot.values[i];
946 : : else
2205 tgl@sss.pgh.pa.us 947 :UBC 0 : fcinfo->args[1].value = sslot.values[i];
2205 tgl@sss.pgh.pa.us 948 :CBC 122875 : fcinfo->isnull = false;
949 : 122875 : fresult = FunctionCallInvoke(fcinfo);
950 [ + - + + ]: 122875 : if (!fcinfo->isnull && DatumGetBool(fresult))
7167 951 : 8126 : nmatch++;
952 : : }
3279 953 : 1475 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
954 : : }
955 : : else
7167 956 : 1049 : result = -1;
3279 957 : 2524 : free_attstatsslot(&sslot);
958 : : }
959 : : else
960 : : {
6631 961 : 1878 : *hist_size = 0;
7167 962 : 1878 : result = -1;
963 : : }
964 : :
965 : 4402 : return result;
966 : : }
967 : :
968 : : /*
969 : : * generic_restriction_selectivity - Selectivity for almost anything
970 : : *
971 : : * This function estimates selectivity for operators that we don't have any
972 : : * special knowledge about, but are on data types that we collect standard
973 : : * MCV and/or histogram statistics for. (Additional assumptions are that
974 : : * the operator is strict and immutable, or at least stable.)
975 : : *
976 : : * If we have "VAR OP CONST" or "CONST OP VAR", selectivity is estimated by
977 : : * applying the operator to each element of the column's MCV and/or histogram
978 : : * stats, and merging the results using the assumption that the histogram is
979 : : * a reasonable random sample of the column's non-MCV population. Note that
980 : : * if the operator's semantics are related to the histogram ordering, this
981 : : * might not be such a great assumption; other functions such as
982 : : * scalarineqsel() are probably a better match in such cases.
983 : : *
984 : : * Otherwise, fall back to the default selectivity provided by the caller.
985 : : */
986 : : double
2160 987 : 845 : generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation,
988 : : List *args, int varRelid,
989 : : double default_selectivity)
990 : : {
991 : : double selec;
992 : : VariableStatData vardata;
993 : : Node *other;
994 : : bool varonleft;
995 : :
996 : : /*
997 : : * If expression is not variable OP something or something OP variable,
998 : : * then punt and return the default estimate.
999 : : */
2225 1000 [ - + ]: 845 : if (!get_restriction_variable(root, args, varRelid,
1001 : : &vardata, &other, &varonleft))
2225 tgl@sss.pgh.pa.us 1002 :UBC 0 : return default_selectivity;
1003 : :
1004 : : /*
1005 : : * If the something is a NULL constant, assume operator is strict and
1006 : : * return zero, ie, operator will never return TRUE.
1007 : : */
2225 tgl@sss.pgh.pa.us 1008 [ + - ]:CBC 845 : if (IsA(other, Const) &&
1009 [ - + ]: 845 : ((Const *) other)->constisnull)
1010 : : {
2225 tgl@sss.pgh.pa.us 1011 [ # # ]:UBC 0 : ReleaseVariableStats(vardata);
1012 : 0 : return 0.0;
1013 : : }
1014 : :
2225 tgl@sss.pgh.pa.us 1015 [ + - ]:CBC 845 : if (IsA(other, Const))
1016 : : {
1017 : : /* Variable is being compared to a known non-null constant */
1018 : 845 : Datum constval = ((Const *) other)->constvalue;
1019 : : FmgrInfo opproc;
1020 : : double mcvsum;
1021 : : double mcvsel;
1022 : : double nullfrac;
1023 : : int hist_size;
1024 : :
2205 1025 : 845 : fmgr_info(get_opcode(oproid), &opproc);
1026 : :
1027 : : /*
1028 : : * Calculate the selectivity for the column's most common values.
1029 : : */
2160 1030 : 845 : mcvsel = mcv_selectivity(&vardata, &opproc, collation,
1031 : : constval, varonleft,
1032 : : &mcvsum);
1033 : :
1034 : : /*
1035 : : * If the histogram is large enough, see what fraction of it matches
1036 : : * the query, and assume that's representative of the non-MCV
1037 : : * population. Otherwise use the default selectivity for the non-MCV
1038 : : * population.
1039 : : */
1040 : 845 : selec = histogram_selectivity(&vardata, &opproc, collation,
1041 : : constval, varonleft,
1042 : : 10, 1, &hist_size);
2225 1043 [ + - ]: 845 : if (selec < 0)
1044 : : {
1045 : : /* Nope, fall back on default */
1046 : 845 : selec = default_selectivity;
1047 : : }
2225 tgl@sss.pgh.pa.us 1048 [ # # ]:UBC 0 : else if (hist_size < 100)
1049 : : {
1050 : : /*
1051 : : * For histogram sizes from 10 to 100, we combine the histogram
1052 : : * and default selectivities, putting increasingly more trust in
1053 : : * the histogram for larger sizes.
1054 : : */
1055 : 0 : double hist_weight = hist_size / 100.0;
1056 : :
1057 : 0 : selec = selec * hist_weight +
1058 : 0 : default_selectivity * (1.0 - hist_weight);
1059 : : }
1060 : :
1061 : : /* In any case, don't believe extremely small or large estimates. */
2225 tgl@sss.pgh.pa.us 1062 [ - + ]:CBC 845 : if (selec < 0.0001)
2225 tgl@sss.pgh.pa.us 1063 :UBC 0 : selec = 0.0001;
2225 tgl@sss.pgh.pa.us 1064 [ - + ]:CBC 845 : else if (selec > 0.9999)
2225 tgl@sss.pgh.pa.us 1065 :UBC 0 : selec = 0.9999;
1066 : :
1067 : : /* Don't forget to account for nulls. */
2225 tgl@sss.pgh.pa.us 1068 [ + + ]:CBC 845 : if (HeapTupleIsValid(vardata.statsTuple))
1069 : 70 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
1070 : : else
1071 : 775 : nullfrac = 0.0;
1072 : :
1073 : : /*
1074 : : * Now merge the results from the MCV and histogram calculations,
1075 : : * realizing that the histogram covers only the non-null values that
1076 : : * are not listed in MCV.
1077 : : */
1078 : 845 : selec *= 1.0 - nullfrac - mcvsum;
1079 : 845 : selec += mcvsel;
1080 : : }
1081 : : else
1082 : : {
1083 : : /* Comparison value is not constant, so we can't do anything */
2225 tgl@sss.pgh.pa.us 1084 :UBC 0 : selec = default_selectivity;
1085 : : }
1086 : :
2225 tgl@sss.pgh.pa.us 1087 [ + + ]:CBC 845 : ReleaseVariableStats(vardata);
1088 : :
1089 : : /* result should be in range, but make sure... */
1090 [ - + - + ]: 845 : CLAMP_PROBABILITY(selec);
1091 : :
1092 : 845 : return selec;
1093 : : }
1094 : :
1095 : : /*
1096 : : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
1097 : : *
1098 : : * Determine the fraction of the variable's histogram population that
1099 : : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
1100 : : * The isgt and iseq flags distinguish which of the four cases apply.
1101 : : *
1102 : : * While opproc could be looked up from the operator OID, common callers
1103 : : * also need to call it separately, so we make the caller pass both.
1104 : : *
1105 : : * Returns -1 if there is no histogram (valid results will always be >= 0).
1106 : : *
1107 : : * Note that the result disregards both the most-common-values (if any) and
1108 : : * null entries. The caller is expected to combine this result with
1109 : : * statistics for those portions of the column population.
1110 : : *
1111 : : * This is exported so that some other estimation functions can use it.
1112 : : */
1113 : : double
5965 1114 : 234454 : ineq_histogram_selectivity(PlannerInfo *root,
1115 : : VariableStatData *vardata,
1116 : : Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq,
1117 : : Oid collation,
1118 : : Datum constval, Oid consttype)
1119 : : {
1120 : : double hist_selec;
1121 : : AttStatsSlot sslot;
1122 : :
1123 : 234454 : hist_selec = -1.0;
1124 : :
1125 : : /*
1126 : : * Someday, ANALYZE might store more than one histogram per rel/att,
1127 : : * corresponding to more than one possible sort ordering defined for the
1128 : : * column type. Right now, we know there is only one, so just grab it and
1129 : : * see if it matches the query.
1130 : : *
1131 : : * Note that we can't use opoid as search argument; the staop appearing in
1132 : : * pg_statistic will be for the relevant '<' operator, but what we have
1133 : : * might be some other inequality operator such as '>='. (Even if opoid
1134 : : * is a '<' operator, it could be cross-type.) Hence we must use
1135 : : * comparison_ops_are_compatible() to see if the operators match.
1136 : : */
7420 1137 [ + + + + ]: 468401 : if (HeapTupleIsValid(vardata->statsTuple) &&
3287 peter_e@gmx.net 1138 [ + + ]: 467624 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
3279 tgl@sss.pgh.pa.us 1139 : 233677 : get_attstatsslot(&sslot, vardata->statsTuple,
1140 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1141 : : ATTSTATSSLOT_VALUES))
1142 : : {
2160 1143 [ + - ]: 148907 : if (sslot.nvalues > 1 &&
1144 [ + + + + ]: 297746 : sslot.stacoll == collation &&
1145 : 148839 : comparison_ops_are_compatible(sslot.staop, opoid))
9773 1146 : 148749 : {
1147 : : /*
1148 : : * Use binary search to find the desired location, namely the
1149 : : * right end of the histogram bin containing the comparison value,
1150 : : * which is the leftmost entry for which the comparison operator
1151 : : * succeeds (if isgt) or fails (if !isgt).
1152 : : *
1153 : : * In this loop, we pay no attention to whether the operator iseq
1154 : : * or not; that detail will be mopped up below. (We cannot tell,
1155 : : * anyway, whether the operator thinks the values are equal.)
1156 : : *
1157 : : * If the binary search accesses the first or last histogram
1158 : : * entry, we try to replace that endpoint with the true column min
1159 : : * or max as found by get_actual_variable_range(). This
1160 : : * ameliorates misestimates when the min or max is moving as a
1161 : : * result of changes since the last ANALYZE. Note that this could
1162 : : * result in effectively including MCVs into the histogram that
1163 : : * weren't there before, but we don't try to correct for that.
1164 : : */
1165 : : double histfrac;
7153 bruce@momjian.us 1166 : 148749 : int lobound = 0; /* first possible slot to search */
3240 tgl@sss.pgh.pa.us 1167 : 148749 : int hibound = sslot.nvalues; /* last+1 slot to search */
5965 1168 : 148749 : bool have_end = false;
1169 : :
1170 : : /*
1171 : : * If there are only two histogram entries, we'll want up-to-date
1172 : : * values for both. (If there are more than two, we need at most
1173 : : * one of them to be updated, so we deal with that within the
1174 : : * loop.)
1175 : : */
3279 1176 [ + + ]: 148749 : if (sslot.nvalues == 2)
5965 1177 : 4839 : have_end = get_actual_variable_range(root,
1178 : : vardata,
1179 : : sslot.staop,
1180 : : collation,
1181 : : &sslot.values[0],
3279 1182 : 4839 : &sslot.values[1]);
1183 : :
7167 1184 [ + + ]: 996310 : while (lobound < hibound)
1185 : : {
7153 bruce@momjian.us 1186 : 847561 : int probe = (lobound + hibound) / 2;
1187 : : bool ltcmp;
1188 : :
1189 : : /*
1190 : : * If we find ourselves about to compare to the first or last
1191 : : * histogram entry, first try to replace it with the actual
1192 : : * current min or max (unless we already did so above).
1193 : : */
3279 tgl@sss.pgh.pa.us 1194 [ + + + + ]: 847561 : if (probe == 0 && sslot.nvalues > 2)
5965 1195 : 70012 : have_end = get_actual_variable_range(root,
1196 : : vardata,
1197 : : sslot.staop,
1198 : : collation,
1199 : : &sslot.values[0],
1200 : : NULL);
3279 1201 [ + + + + ]: 777549 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
5965 1202 : 50290 : have_end = get_actual_variable_range(root,
1203 : : vardata,
1204 : : sslot.staop,
1205 : : collation,
1206 : : NULL,
3240 1207 : 50290 : &sslot.values[probe]);
1208 : :
5502 1209 : 847561 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1210 : : collation,
3279 1211 : 847561 : sslot.values[probe],
1212 : : constval));
7167 1213 [ + + ]: 847561 : if (isgt)
1214 : 51947 : ltcmp = !ltcmp;
1215 [ + + ]: 847561 : if (ltcmp)
1216 : 326404 : lobound = probe + 1;
1217 : : else
1218 : 521157 : hibound = probe;
1219 : : }
1220 : :
1221 [ + + ]: 148749 : if (lobound <= 0)
1222 : : {
1223 : : /*
1224 : : * Constant is below lower histogram boundary. More
1225 : : * precisely, we have found that no entry in the histogram
1226 : : * satisfies the inequality clause (if !isgt) or they all do
1227 : : * (if isgt). We estimate that that's true of the entire
1228 : : * table, so set histfrac to 0.0 (which we'll flip to 1.0
1229 : : * below, if isgt).
1230 : : */
9129 1231 : 62050 : histfrac = 0.0;
1232 : : }
3279 1233 [ + + ]: 86699 : else if (lobound >= sslot.nvalues)
1234 : : {
1235 : : /*
1236 : : * Inverse case: constant is above upper histogram boundary.
1237 : : */
7167 1238 : 25849 : histfrac = 1.0;
1239 : : }
1240 : : else
1241 : : {
1242 : : /* We have values[i-1] <= constant <= values[i]. */
1243 : 60850 : int i = lobound;
3156 1244 : 60850 : double eq_selec = 0;
1245 : : double val,
1246 : : high,
1247 : : low;
1248 : : double binfrac;
1249 : :
1250 : : /*
1251 : : * In the cases where we'll need it below, obtain an estimate
1252 : : * of the selectivity of "x = constval". We use a calculation
1253 : : * similar to what var_eq_const() does for a non-MCV constant,
1254 : : * ie, estimate that all distinct non-MCV values occur equally
1255 : : * often. But multiplication by "1.0 - sumcommon - nullfrac"
1256 : : * will be done by our caller, so we shouldn't do that here.
1257 : : * Therefore we can't try to clamp the estimate by reference
1258 : : * to the least common MCV; the result would be too small.
1259 : : *
1260 : : * Note: since this is effectively assuming that constval
1261 : : * isn't an MCV, it's logically dubious if constval in fact is
1262 : : * one. But we have to apply *some* correction for equality,
1263 : : * and anyway we cannot tell if constval is an MCV, since we
1264 : : * don't have a suitable equality operator at hand.
1265 : : */
1266 [ + + + + ]: 60850 : if (i == 1 || isgt == iseq)
1267 : : {
1268 : : double otherdistinct;
1269 : : bool isdefault;
1270 : : AttStatsSlot mcvslot;
1271 : :
1272 : : /* Get estimated number of distinct values */
1273 : 26122 : otherdistinct = get_variable_numdistinct(vardata,
1274 : : &isdefault);
1275 : :
1276 : : /* Subtract off the number of known MCVs */
1277 [ + + ]: 26122 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1278 : : STATISTIC_KIND_MCV, InvalidOid,
1279 : : ATTSTATSSLOT_NUMBERS))
1280 : : {
1281 : 3804 : otherdistinct -= mcvslot.nnumbers;
1282 : 3804 : free_attstatsslot(&mcvslot);
1283 : : }
1284 : :
1285 : : /* If result doesn't seem sane, leave eq_selec at 0 */
1286 [ + + ]: 26122 : if (otherdistinct > 1)
1287 : 26101 : eq_selec = 1.0 / otherdistinct;
1288 : : }
1289 : :
1290 : : /*
1291 : : * Convert the constant and the two nearest bin boundary
1292 : : * values to a uniform comparison scale, and do a linear
1293 : : * interpolation within this bin.
1294 : : */
2160 1295 [ + - ]: 60850 : if (convert_to_scalar(constval, consttype, collation,
1296 : : &val,
3279 1297 : 60850 : sslot.values[i - 1], sslot.values[i],
1298 : : vardata->vartype,
1299 : : &low, &high))
1300 : : {
7167 1301 [ - + ]: 60850 : if (high <= low)
1302 : : {
1303 : : /* cope if bin boundaries appear identical */
7167 tgl@sss.pgh.pa.us 1304 :UBC 0 : binfrac = 0.5;
1305 : : }
7167 tgl@sss.pgh.pa.us 1306 [ + + ]:CBC 60850 : else if (val <= low)
1307 : 10734 : binfrac = 0.0;
1308 [ + + ]: 50116 : else if (val >= high)
1309 : 1998 : binfrac = 1.0;
1310 : : else
1311 : : {
1312 : 48118 : binfrac = (val - low) / (high - low);
1313 : :
1314 : : /*
1315 : : * Watch out for the possibility that we got a NaN or
1316 : : * Infinity from the division. This can happen
1317 : : * despite the previous checks, if for example "low"
1318 : : * is -Infinity.
1319 : : */
1320 [ + - + - ]: 48118 : if (isnan(binfrac) ||
1321 [ - + ]: 48118 : binfrac < 0.0 || binfrac > 1.0)
7167 tgl@sss.pgh.pa.us 1322 :UBC 0 : binfrac = 0.5;
1323 : : }
1324 : : }
1325 : : else
1326 : : {
1327 : : /*
1328 : : * Ideally we'd produce an error here, on the grounds that
1329 : : * the given operator shouldn't have scalarXXsel
1330 : : * registered as its selectivity func unless we can deal
1331 : : * with its operand types. But currently, all manner of
1332 : : * stuff is invoking scalarXXsel, so give a default
1333 : : * estimate until that can be fixed.
1334 : : */
1335 : 0 : binfrac = 0.5;
1336 : : }
1337 : :
1338 : : /*
1339 : : * Now, compute the overall selectivity across the values
1340 : : * represented by the histogram. We have i-1 full bins and
1341 : : * binfrac partial bin below the constant.
1342 : : */
7167 tgl@sss.pgh.pa.us 1343 :CBC 60850 : histfrac = (double) (i - 1) + binfrac;
3279 1344 : 60850 : histfrac /= (double) (sslot.nvalues - 1);
1345 : :
1346 : : /*
1347 : : * At this point, histfrac is an estimate of the fraction of
1348 : : * the population represented by the histogram that satisfies
1349 : : * "x <= constval". Somewhat remarkably, this statement is
1350 : : * true regardless of which operator we were doing the probes
1351 : : * with, so long as convert_to_scalar() delivers reasonable
1352 : : * results. If the probe constant is equal to some histogram
1353 : : * entry, we would have considered the bin to the left of that
1354 : : * entry if probing with "<" or ">=", or the bin to the right
1355 : : * if probing with "<=" or ">"; but binfrac would have come
1356 : : * out as 1.0 in the first case and 0.0 in the second, leading
1357 : : * to the same histfrac in either case. For probe constants
1358 : : * between histogram entries, we find the same bin and get the
1359 : : * same estimate with any operator.
1360 : : *
1361 : : * The fact that the estimate corresponds to "x <= constval"
1362 : : * and not "x < constval" is because of the way that ANALYZE
1363 : : * constructs the histogram: each entry is, effectively, the
1364 : : * rightmost value in its sample bucket. So selectivity
1365 : : * values that are exact multiples of 1/(histogram_size-1)
1366 : : * should be understood as estimates including a histogram
1367 : : * entry plus everything to its left.
1368 : : *
1369 : : * However, that breaks down for the first histogram entry,
1370 : : * which necessarily is the leftmost value in its sample
1371 : : * bucket. That means the first histogram bin is slightly
1372 : : * narrower than the rest, by an amount equal to eq_selec.
1373 : : * Another way to say that is that we want "x <= leftmost" to
1374 : : * be estimated as eq_selec not zero. So, if we're dealing
1375 : : * with the first bin (i==1), rescale to make that true while
1376 : : * adjusting the rest of that bin linearly.
1377 : : */
3156 1378 [ + + ]: 60850 : if (i == 1)
1379 : 10774 : histfrac += eq_selec * (1.0 - binfrac);
1380 : :
1381 : : /*
1382 : : * "x <= constval" is good if we want an estimate for "<=" or
1383 : : * ">", but if we are estimating for "<" or ">=", we now need
1384 : : * to decrease the estimate by eq_selec.
1385 : : */
1386 [ + + ]: 60850 : if (isgt == iseq)
1387 : 21025 : histfrac -= eq_selec;
1388 : : }
1389 : :
1390 : : /*
1391 : : * Now the estimate is finished for "<" and "<=" cases. If we are
1392 : : * estimating for ">" or ">=", flip it.
1393 : : */
9129 1394 [ + + ]: 148749 : hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1395 : :
1396 : : /*
1397 : : * The histogram boundaries are only approximate to begin with,
1398 : : * and may well be out of date anyway. Therefore, don't believe
1399 : : * extremely small or large selectivity estimates --- unless we
1400 : : * got actual current endpoint values from the table, in which
1401 : : * case just do the usual sanity clamp. Somewhat arbitrarily, we
1402 : : * set the cutoff for other cases at a hundredth of the histogram
1403 : : * resolution.
1404 : : */
5965 1405 [ + + ]: 148749 : if (have_end)
1406 [ + + - + ]: 82863 : CLAMP_PROBABILITY(hist_selec);
1407 : : else
1408 : : {
3156 1409 : 65886 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1410 : :
1411 [ + + ]: 65886 : if (hist_selec < cutoff)
1412 : 21504 : hist_selec = cutoff;
1413 [ + + ]: 44382 : else if (hist_selec > 1.0 - cutoff)
1414 : 16491 : hist_selec = 1.0 - cutoff;
1415 : : }
1416 : : }
2160 1417 [ + - ]: 158 : else if (sslot.nvalues > 1)
1418 : : {
1419 : : /*
1420 : : * If we get here, we have a histogram but it's not sorted the way
1421 : : * we want. Do a brute-force search to see how many of the
1422 : : * entries satisfy the comparison condition, and take that
1423 : : * fraction as our estimate. (This is identical to the inner loop
1424 : : * of histogram_selectivity; maybe share code?)
1425 : : */
1426 : 158 : LOCAL_FCINFO(fcinfo, 2);
1427 : 158 : int nmatch = 0;
1428 : :
1429 : 158 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1430 : : NULL, NULL);
1431 : 158 : fcinfo->args[0].isnull = false;
1432 : 158 : fcinfo->args[1].isnull = false;
1433 : 158 : fcinfo->args[1].value = constval;
1434 [ + + ]: 802994 : for (int i = 0; i < sslot.nvalues; i++)
1435 : : {
1436 : : Datum fresult;
1437 : :
1438 : 802836 : fcinfo->args[0].value = sslot.values[i];
1439 : 802836 : fcinfo->isnull = false;
1440 : 802836 : fresult = FunctionCallInvoke(fcinfo);
1441 [ + - + + ]: 802836 : if (!fcinfo->isnull && DatumGetBool(fresult))
1442 : 2507 : nmatch++;
1443 : : }
1444 : 158 : hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1445 : :
1446 : : /*
1447 : : * As above, clamp to a hundredth of the histogram resolution.
1448 : : * This case is surely even less trustworthy than the normal one,
1449 : : * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1450 : : * clamp should be more restrictive in this case?)
1451 : : */
1452 : : {
1453 : 158 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1454 : :
1455 [ + + ]: 158 : if (hist_selec < cutoff)
2160 tgl@sss.pgh.pa.us 1456 :GBC 3 : hist_selec = cutoff;
2160 tgl@sss.pgh.pa.us 1457 [ + + ]:CBC 155 : else if (hist_selec > 1.0 - cutoff)
2160 tgl@sss.pgh.pa.us 1458 :GBC 3 : hist_selec = 1.0 - cutoff;
1459 : : }
1460 : : }
1461 : :
3279 tgl@sss.pgh.pa.us 1462 :CBC 148907 : free_attstatsslot(&sslot);
1463 : : }
1464 : :
7420 1465 : 234454 : return hist_selec;
1466 : : }
1467 : :
1468 : : /*
1469 : : * Common wrapper function for the selectivity estimators that simply
1470 : : * invoke scalarineqsel().
1471 : : */
1472 : : static Datum
3156 1473 : 39042 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1474 : : {
7639 1475 : 39042 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
9116 1476 : 39042 : Oid operator = PG_GETARG_OID(1);
1477 : 39042 : List *args = (List *) PG_GETARG_POINTER(2);
1478 : 39042 : int varRelid = PG_GETARG_INT32(3);
2160 1479 : 39042 : Oid collation = PG_GET_COLLATION();
1480 : : VariableStatData vardata;
1481 : : Node *other;
1482 : : bool varonleft;
1483 : : Datum constval;
1484 : : Oid consttype;
1485 : : double selec;
1486 : :
1487 : : /*
1488 : : * If expression is not variable op something or something op variable,
1489 : : * then punt and return a default estimate.
1490 : : */
8113 1491 [ + + ]: 39042 : if (!get_restriction_variable(root, args, varRelid,
1492 : : &vardata, &other, &varonleft))
9116 1493 : 462 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1494 : :
1495 : : /*
1496 : : * Can't do anything useful if the something is not a constant, either.
1497 : : */
8831 1498 [ + + ]: 38580 : if (!IsA(other, Const))
1499 : : {
8113 1500 [ + + ]: 2365 : ReleaseVariableStats(vardata);
8831 1501 : 2365 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1502 : : }
1503 : :
1504 : : /*
1505 : : * If the constant is NULL, assume operator is strict and return zero, ie,
1506 : : * operator will never return TRUE.
1507 : : */
1508 [ + + ]: 36215 : if (((Const *) other)->constisnull)
1509 : : {
8113 1510 [ + + ]: 55 : ReleaseVariableStats(vardata);
8831 1511 : 55 : PG_RETURN_FLOAT8(0.0);
1512 : : }
1513 : 36160 : constval = ((Const *) other)->constvalue;
1514 : 36160 : consttype = ((Const *) other)->consttype;
1515 : :
1516 : : /*
1517 : : * Force the var to be on the left to simplify logic in scalarineqsel.
1518 : : */
3156 1519 [ + + ]: 36160 : if (!varonleft)
1520 : : {
9116 1521 : 318 : operator = get_commutator(operator);
1522 [ - + ]: 318 : if (!operator)
1523 : : {
1524 : : /* Use default selectivity (should we raise an error instead?) */
8113 tgl@sss.pgh.pa.us 1525 [ # # ]:UBC 0 : ReleaseVariableStats(vardata);
9116 1526 : 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1527 : : }
3156 tgl@sss.pgh.pa.us 1528 :CBC 318 : isgt = !isgt;
1529 : : }
1530 : :
1531 : : /* The rest of the work is done by scalarineqsel(). */
2160 1532 : 36160 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1533 : : &vardata, constval, consttype);
1534 : :
8113 1535 [ + + ]: 36160 : ReleaseVariableStats(vardata);
1536 : :
9129 1537 : 36160 : PG_RETURN_FLOAT8((float8) selec);
1538 : : }
1539 : :
1540 : : /*
1541 : : * scalarltsel - Selectivity of "<" for scalars.
1542 : : */
1543 : : Datum
3156 1544 : 12240 : scalarltsel(PG_FUNCTION_ARGS)
1545 : : {
1546 : 12240 : return scalarineqsel_wrapper(fcinfo, false, false);
1547 : : }
1548 : :
1549 : : /*
1550 : : * scalarlesel - Selectivity of "<=" for scalars.
1551 : : */
1552 : : Datum
1553 : 3816 : scalarlesel(PG_FUNCTION_ARGS)
1554 : : {
1555 : 3816 : return scalarineqsel_wrapper(fcinfo, false, true);
1556 : : }
1557 : :
1558 : : /*
1559 : : * scalargtsel - Selectivity of ">" for scalars.
1560 : : */
1561 : : Datum
1562 : 12400 : scalargtsel(PG_FUNCTION_ARGS)
1563 : : {
1564 : 12400 : return scalarineqsel_wrapper(fcinfo, true, false);
1565 : : }
1566 : :
1567 : : /*
1568 : : * scalargesel - Selectivity of ">=" for scalars.
1569 : : */
1570 : : Datum
1571 : 10586 : scalargesel(PG_FUNCTION_ARGS)
1572 : : {
1573 : 10586 : return scalarineqsel_wrapper(fcinfo, true, true);
1574 : : }
1575 : :
1576 : : /*
1577 : : * boolvarsel - Selectivity of Boolean variable.
1578 : : *
1579 : : * This can actually be called on any boolean-valued expression. If it
1580 : : * involves only Vars of the specified relation, and if there are statistics
1581 : : * about the Var or expression (the latter is possible if it's indexed) then
1582 : : * we'll produce a real estimate; otherwise it's just a default.
1583 : : */
1584 : : Selectivity
3876 1585 : 46971 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1586 : : {
1587 : : VariableStatData vardata;
1588 : : double selec;
1589 : :
1590 : 46971 : examine_variable(root, arg, varRelid, &vardata);
1591 [ + + ]: 46971 : if (HeapTupleIsValid(vardata.statsTuple))
1592 : : {
1593 : : /*
1594 : : * A boolean variable V is equivalent to the clause V = 't', so we
1595 : : * compute the selectivity as if that is what we have.
1596 : : */
2160 1597 : 24080 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1598 : : BoolGetDatum(true), false, true, false);
1599 : : }
227 tgl@sss.pgh.pa.us 1600 [ + + ]:GNC 22891 : else if (is_funcclause(arg))
1601 : : {
1602 : : /*
1603 : : * If we have no stats and it's a function call, estimate 0.3333333.
1604 : : * This seems a pretty unprincipled choice, but Postgres has been
1605 : : * using that estimate for function calls since 1992. The hoariness
1606 : : * of this behavior suggests that we should not be in too much hurry
1607 : : * to use another value.
1608 : : */
1609 : 11379 : selec = 0.3333333;
1610 : : }
1611 : : else
1612 : : {
1613 : : /* Otherwise, the default estimate is 0.5 */
3876 tgl@sss.pgh.pa.us 1614 :CBC 11512 : selec = 0.5;
1615 : : }
1616 [ + + ]: 46971 : ReleaseVariableStats(vardata);
1617 : 46971 : return selec;
1618 : : }
1619 : :
1620 : : /*
1621 : : * booltestsel - Selectivity of BooleanTest Node.
1622 : : */
1623 : : Selectivity
7639 1624 : 791 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1625 : : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1626 : : {
1627 : : VariableStatData vardata;
1628 : : double selec;
1629 : :
8113 1630 : 791 : examine_variable(root, arg, varRelid, &vardata);
1631 : :
1632 [ + + ]: 791 : if (HeapTupleIsValid(vardata.statsTuple))
1633 : : {
1634 : : Form_pg_statistic stats;
1635 : : double freq_null;
1636 : : AttStatsSlot sslot;
1637 : :
8113 tgl@sss.pgh.pa.us 1638 :GBC 20 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
9080 1639 : 20 : freq_null = stats->stanullfrac;
1640 : :
3279 1641 [ + + ]: 20 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1642 : : STATISTIC_KIND_MCV, InvalidOid,
1643 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1644 [ + - ]: 10 : && sslot.nnumbers > 0)
9080 1645 : 10 : {
1646 : : double freq_true;
1647 : : double freq_false;
1648 : :
1649 : : /*
1650 : : * Get first MCV frequency and derive frequency for true.
1651 : : */
3279 1652 [ - + ]: 10 : if (DatumGetBool(sslot.values[0]))
3279 tgl@sss.pgh.pa.us 1653 :UBC 0 : freq_true = sslot.numbers[0];
1654 : : else
3279 tgl@sss.pgh.pa.us 1655 :GBC 10 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1656 : :
1657 : : /*
1658 : : * Next derive frequency for false. Then use these as appropriate
1659 : : * to derive frequency for each case.
1660 : : */
9080 1661 : 10 : freq_false = 1.0 - freq_true - freq_null;
1662 : :
8599 1663 [ - - + - : 10 : switch (booltesttype)
- - - ]
1664 : : {
8958 bruce@momjian.us 1665 :UBC 0 : case IS_UNKNOWN:
1666 : : /* select only NULL values */
9080 tgl@sss.pgh.pa.us 1667 : 0 : selec = freq_null;
1668 : 0 : break;
8958 bruce@momjian.us 1669 : 0 : case IS_NOT_UNKNOWN:
1670 : : /* select non-NULL values */
9080 tgl@sss.pgh.pa.us 1671 : 0 : selec = 1.0 - freq_null;
1672 : 0 : break;
8958 bruce@momjian.us 1673 :GBC 10 : case IS_TRUE:
1674 : : /* select only TRUE values */
9080 tgl@sss.pgh.pa.us 1675 : 10 : selec = freq_true;
1676 : 10 : break;
8958 bruce@momjian.us 1677 :UBC 0 : case IS_NOT_TRUE:
1678 : : /* select non-TRUE values */
9080 tgl@sss.pgh.pa.us 1679 : 0 : selec = 1.0 - freq_true;
1680 : 0 : break;
8958 bruce@momjian.us 1681 : 0 : case IS_FALSE:
1682 : : /* select only FALSE values */
9080 tgl@sss.pgh.pa.us 1683 : 0 : selec = freq_false;
1684 : 0 : break;
8958 bruce@momjian.us 1685 : 0 : case IS_NOT_FALSE:
1686 : : /* select non-FALSE values */
9080 tgl@sss.pgh.pa.us 1687 : 0 : selec = 1.0 - freq_false;
1688 : 0 : break;
8958 bruce@momjian.us 1689 : 0 : default:
8318 tgl@sss.pgh.pa.us 1690 [ # # ]: 0 : elog(ERROR, "unrecognized booltesttype: %d",
1691 : : (int) booltesttype);
1692 : : selec = 0.0; /* Keep compiler quiet */
1693 : : break;
1694 : : }
1695 : :
3279 tgl@sss.pgh.pa.us 1696 :GBC 10 : free_attstatsslot(&sslot);
1697 : : }
1698 : : else
1699 : : {
1700 : : /*
1701 : : * No most-common-value info available. Still have null fraction
1702 : : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1703 : : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1704 : : */
8599 1705 [ + - - - : 10 : switch (booltesttype)
- ]
1706 : : {
8958 bruce@momjian.us 1707 : 10 : case IS_UNKNOWN:
1708 : : /* select only NULL values */
9080 tgl@sss.pgh.pa.us 1709 : 10 : selec = freq_null;
1710 : 10 : break;
8958 bruce@momjian.us 1711 :UBC 0 : case IS_NOT_UNKNOWN:
1712 : : /* select non-NULL values */
9080 tgl@sss.pgh.pa.us 1713 : 0 : selec = 1.0 - freq_null;
1714 : 0 : break;
8958 bruce@momjian.us 1715 : 0 : case IS_TRUE:
1716 : : case IS_FALSE:
1717 : : /* Assume we select half of the non-NULL values */
9080 tgl@sss.pgh.pa.us 1718 : 0 : selec = (1.0 - freq_null) / 2.0;
1719 : 0 : break;
4668 1720 : 0 : case IS_NOT_TRUE:
1721 : : case IS_NOT_FALSE:
1722 : : /* Assume we select NULLs plus half of the non-NULLs */
1723 : : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1724 : 0 : selec = (freq_null + 1.0) / 2.0;
1725 : 0 : break;
8958 bruce@momjian.us 1726 : 0 : default:
8318 tgl@sss.pgh.pa.us 1727 [ # # ]: 0 : elog(ERROR, "unrecognized booltesttype: %d",
1728 : : (int) booltesttype);
1729 : : selec = 0.0; /* Keep compiler quiet */
1730 : : break;
1731 : : }
1732 : : }
1733 : : }
1734 : : else
1735 : : {
1736 : : /*
1737 : : * If we can't get variable statistics for the argument, perhaps
1738 : : * clause_selectivity can do something with it. We ignore the
1739 : : * possibility of a NULL value when using clause_selectivity, and just
1740 : : * assume the value is either TRUE or FALSE.
1741 : : */
8599 tgl@sss.pgh.pa.us 1742 [ + + + + :CBC 771 : switch (booltesttype)
- ]
1743 : : {
9080 1744 : 40 : case IS_UNKNOWN:
1745 : 40 : selec = DEFAULT_UNK_SEL;
1746 : 40 : break;
1747 : 90 : case IS_NOT_UNKNOWN:
1748 : 90 : selec = DEFAULT_NOT_UNK_SEL;
1749 : 90 : break;
1750 : 220 : case IS_TRUE:
1751 : : case IS_NOT_FALSE:
8113 1752 : 220 : selec = (double) clause_selectivity(root, arg,
1753 : : varRelid,
1754 : : jointype, sjinfo);
1755 : 220 : break;
1756 : 421 : case IS_FALSE:
1757 : : case IS_NOT_TRUE:
1758 : 421 : selec = 1.0 - (double) clause_selectivity(root, arg,
1759 : : varRelid,
1760 : : jointype, sjinfo);
9080 1761 : 421 : break;
9080 tgl@sss.pgh.pa.us 1762 :UBC 0 : default:
8318 1763 [ # # ]: 0 : elog(ERROR, "unrecognized booltesttype: %d",
1764 : : (int) booltesttype);
1765 : : selec = 0.0; /* Keep compiler quiet */
1766 : : break;
1767 : : }
1768 : : }
1769 : :
8113 tgl@sss.pgh.pa.us 1770 [ + + ]:CBC 791 : ReleaseVariableStats(vardata);
1771 : :
1772 : : /* result should be in range, but make sure... */
8888 1773 [ - + - + ]: 791 : CLAMP_PROBABILITY(selec);
1774 : :
9080 1775 : 791 : return (Selectivity) selec;
1776 : : }
1777 : :
1778 : : /*
1779 : : * nulltestsel - Selectivity of NullTest Node.
1780 : : */
1781 : : Selectivity
6473 1782 : 14453 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1783 : : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1784 : : {
1785 : : VariableStatData vardata;
1786 : : double selec;
1787 : :
8113 1788 : 14453 : examine_variable(root, arg, varRelid, &vardata);
1789 : :
1790 [ + + ]: 14453 : if (HeapTupleIsValid(vardata.statsTuple))
1791 : : {
1792 : : Form_pg_statistic stats;
1793 : : double freq_null;
1794 : :
1795 : 7786 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
9080 1796 : 7786 : freq_null = stats->stanullfrac;
1797 : :
8599 1798 [ + + - ]: 7786 : switch (nulltesttype)
1799 : : {
8958 bruce@momjian.us 1800 : 5976 : case IS_NULL:
1801 : :
1802 : : /*
1803 : : * Use freq_null directly.
1804 : : */
9080 tgl@sss.pgh.pa.us 1805 : 5976 : selec = freq_null;
1806 : 5976 : break;
8958 bruce@momjian.us 1807 : 1810 : case IS_NOT_NULL:
1808 : :
1809 : : /*
1810 : : * Select not unknown (not null) values. Calculate from
1811 : : * freq_null.
1812 : : */
9080 tgl@sss.pgh.pa.us 1813 : 1810 : selec = 1.0 - freq_null;
1814 : 1810 : break;
8958 bruce@momjian.us 1815 :UBC 0 : default:
8318 tgl@sss.pgh.pa.us 1816 [ # # ]: 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1817 : : (int) nulltesttype);
1818 : : return (Selectivity) 0; /* keep compiler quiet */
1819 : : }
1820 : : }
2657 tgl@sss.pgh.pa.us 1821 [ + - + + ]:CBC 6667 : else if (vardata.var && IsA(vardata.var, Var) &&
1822 [ + + ]: 6137 : ((Var *) vardata.var)->varattno < 0)
1823 : : {
1824 : : /*
1825 : : * There are no stats for system columns, but we know they are never
1826 : : * NULL.
1827 : : */
1828 [ + - ]: 89 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1829 : : }
1830 : : else
1831 : : {
1832 : : /*
1833 : : * No ANALYZE stats available, so make a guess
1834 : : */
8113 1835 [ + + - ]: 6578 : switch (nulltesttype)
1836 : : {
1837 : 1648 : case IS_NULL:
1838 : 1648 : selec = DEFAULT_UNK_SEL;
1839 : 1648 : break;
1840 : 4930 : case IS_NOT_NULL:
1841 : 4930 : selec = DEFAULT_NOT_UNK_SEL;
1842 : 4930 : break;
8113 tgl@sss.pgh.pa.us 1843 :UBC 0 : default:
1844 [ # # ]: 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1845 : : (int) nulltesttype);
1846 : : return (Selectivity) 0; /* keep compiler quiet */
1847 : : }
1848 : : }
1849 : :
8113 tgl@sss.pgh.pa.us 1850 [ + + ]:CBC 14453 : ReleaseVariableStats(vardata);
1851 : :
1852 : : /* result should be in range, but make sure... */
8888 1853 [ - + - + ]: 14453 : CLAMP_PROBABILITY(selec);
1854 : :
9080 1855 : 14453 : return (Selectivity) selec;
1856 : : }
1857 : :
1858 : : /*
1859 : : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1860 : : *
1861 : : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1862 : : * but it seems possible that RelabelType might show up. Also, the planner
1863 : : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1864 : : * so we need to be ready to deal with more than one level.
1865 : : */
1866 : : static Node *
7037 1867 : 106231 : strip_array_coercion(Node *node)
1868 : : {
1869 : : for (;;)
1870 : : {
3139 1871 [ + - + + ]: 106319 : if (node && IsA(node, ArrayCoerceExpr))
7037 1872 : 88 : {
3139 1873 : 1968 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1874 : :
1875 : : /*
1876 : : * If the per-element expression is just a RelabelType on top of
1877 : : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1878 : : */
1879 [ + + ]: 1968 : if (IsA(acoerce->elemexpr, RelabelType) &&
1880 [ + - ]: 88 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1881 : 88 : node = (Node *) acoerce->arg;
1882 : : else
1883 : : break;
1884 : : }
6979 1885 [ + - - + ]: 104351 : else if (node && IsA(node, RelabelType))
1886 : : {
1887 : : /* We don't really expect this case, but may as well cope */
6979 tgl@sss.pgh.pa.us 1888 :UBC 0 : node = (Node *) ((RelabelType *) node)->arg;
1889 : : }
1890 : : else
1891 : : break;
1892 : : }
7037 tgl@sss.pgh.pa.us 1893 :CBC 106231 : return node;
1894 : : }
1895 : :
1896 : : /*
1897 : : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1898 : : */
1899 : : Selectivity
7466 1900 : 18468 : scalararraysel(PlannerInfo *root,
1901 : : ScalarArrayOpExpr *clause,
1902 : : bool is_join_clause,
1903 : : int varRelid,
1904 : : JoinType jointype,
1905 : : SpecialJoinInfo *sjinfo)
1906 : : {
1907 : 18468 : Oid operator = clause->opno;
1908 : 18468 : bool useOr = clause->useOr;
5176 1909 : 18468 : bool isEquality = false;
1910 : 18468 : bool isInequality = false;
1911 : : Node *leftop;
1912 : : Node *rightop;
1913 : : Oid nominal_element_type;
1914 : : Oid nominal_element_collation;
1915 : : TypeCacheEntry *typentry;
1916 : : RegProcedure oprsel;
1917 : : FmgrInfo oprselproc;
1918 : : Selectivity s1;
1919 : : Selectivity s1disjoint;
1920 : :
1921 : : /* First, deconstruct the expression */
7037 1922 [ - + ]: 18468 : Assert(list_length(clause->args) == 2);
1923 : 18468 : leftop = (Node *) linitial(clause->args);
1924 : 18468 : rightop = (Node *) lsecond(clause->args);
1925 : :
1926 : : /* aggressively reduce both sides to constants */
4456 1927 : 18468 : leftop = estimate_expression_value(root, leftop);
1928 : 18468 : rightop = estimate_expression_value(root, rightop);
1929 : :
1930 : : /* get nominal (after relabeling) element type of rightop */
5675 1931 : 18468 : nominal_element_type = get_base_element_type(exprType(rightop));
7037 1932 [ - + ]: 18468 : if (!OidIsValid(nominal_element_type))
3240 tgl@sss.pgh.pa.us 1933 :UBC 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1934 : : /* get nominal collation, too, for generating constants */
5520 tgl@sss.pgh.pa.us 1935 :CBC 18468 : nominal_element_collation = exprCollation(rightop);
1936 : :
1937 : : /* look through any binary-compatible relabeling of rightop */
7037 1938 : 18468 : rightop = strip_array_coercion(rightop);
1939 : :
1940 : : /*
1941 : : * Detect whether the operator is the default equality or inequality
1942 : : * operator of the array element type.
1943 : : */
5176 1944 : 18468 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1945 [ + + ]: 18468 : if (OidIsValid(typentry->eq_opr))
1946 : : {
1947 [ + + ]: 18466 : if (operator == typentry->eq_opr)
1948 : 16000 : isEquality = true;
1949 [ + + ]: 2466 : else if (get_negator(operator) == typentry->eq_opr)
1950 : 1988 : isInequality = true;
1951 : : }
1952 : :
1953 : : /*
1954 : : * If it is equality or inequality, we might be able to estimate this as a
1955 : : * form of array containment; for instance "const = ANY(column)" can be
1956 : : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1957 : : * that, and returns the selectivity estimate if successful, or -1 if not.
1958 : : */
1959 [ + + + + : 18468 : if ((isEquality || isInequality) && !is_join_clause)
+ + ]
1960 : : {
1961 : 17987 : s1 = scalararraysel_containment(root, leftop, rightop,
1962 : : nominal_element_type,
1963 : : isEquality, useOr, varRelid);
1964 [ + + ]: 17987 : if (s1 >= 0.0)
1965 : 94 : return s1;
1966 : : }
1967 : :
1968 : : /*
1969 : : * Look up the underlying operator's selectivity estimator. Punt if it
1970 : : * hasn't got one.
1971 : : */
1972 [ + + ]: 18374 : if (is_join_clause)
1973 : 1 : oprsel = get_oprjoin(operator);
1974 : : else
1975 : 18373 : oprsel = get_oprrest(operator);
1976 [ + + ]: 18374 : if (!oprsel)
1977 : 2 : return (Selectivity) 0.5;
1978 : 18372 : fmgr_info(oprsel, &oprselproc);
1979 : :
1980 : : /*
1981 : : * In the array-containment check above, we must only believe that an
1982 : : * operator is equality or inequality if it is the default btree equality
1983 : : * operator (or its negator) for the element type, since those are the
1984 : : * operators that array containment will use. But in what follows, we can
1985 : : * be a little laxer, and also believe that any operators using eqsel() or
1986 : : * neqsel() as selectivity estimator act like equality or inequality.
1987 : : */
5172 1988 [ + + + + ]: 18372 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1989 : 16055 : isEquality = true;
1990 [ + + - + ]: 2317 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1991 : 1917 : isInequality = true;
1992 : :
1993 : : /*
1994 : : * We consider three cases:
1995 : : *
1996 : : * 1. rightop is an Array constant: deconstruct the array, apply the
1997 : : * operator's selectivity function for each array element, and merge the
1998 : : * results in the same way that clausesel.c does for AND/OR combinations.
1999 : : *
2000 : : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
2001 : : * function for each element of the ARRAY[] construct, and merge.
2002 : : *
2003 : : * 3. otherwise, make a guess ...
2004 : : */
7466 2005 [ + - + + ]: 18372 : if (rightop && IsA(rightop, Const))
2006 : 14825 : {
2007 : 14855 : Datum arraydatum = ((Const *) rightop)->constvalue;
2008 : 14855 : bool arrayisnull = ((Const *) rightop)->constisnull;
2009 : : ArrayType *arrayval;
2010 : : int16 elmlen;
2011 : : bool elmbyval;
2012 : : char elmalign;
2013 : : int num_elems;
2014 : : Datum *elem_values;
2015 : : bool *elem_nulls;
2016 : : int i;
2017 : :
2018 [ + + ]: 14855 : if (arrayisnull) /* qual can't succeed if null array */
2019 : 30 : return (Selectivity) 0.0;
2020 : 14830 : arrayval = DatumGetArrayTypeP(arraydatum);
2021 : :
2022 : : /*
2023 : : * When the array contains a NULL constant, same as var_eq_const, we
2024 : : * assume the operator is strict and nothing will match, thus return
2025 : : * 0.0.
2026 : : */
47 drowley@postgresql.o 2027 [ + + + + ]:GNC 14830 : if (!useOr && array_contains_nulls(arrayval))
2028 : 5 : return (Selectivity) 0.0;
2029 : :
7466 tgl@sss.pgh.pa.us 2030 :CBC 14825 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
2031 : : &elmlen, &elmbyval, &elmalign);
2032 : 14825 : deconstruct_array(arrayval,
2033 : : ARR_ELEMTYPE(arrayval),
2034 : : elmlen, elmbyval, elmalign,
2035 : : &elem_values, &elem_nulls, &num_elems);
2036 : :
2037 : : /*
2038 : : * For generic operators, we assume the probability of success is
2039 : : * independent for each array element. But for "= ANY" or "<> ALL",
2040 : : * if the array elements are distinct (which'd typically be the case)
2041 : : * then the probabilities are disjoint, and we should just sum them.
2042 : : *
2043 : : * If we were being really tense we would try to confirm that the
2044 : : * elements are all distinct, but that would be expensive and it
2045 : : * doesn't seem to be worth the cycles; it would amount to penalizing
2046 : : * well-written queries in favor of poorly-written ones. However, we
2047 : : * do protect ourselves a little bit by checking whether the
2048 : : * disjointness assumption leads to an impossible (out of range)
2049 : : * probability; if so, we fall back to the normal calculation.
2050 : : */
5172 2051 [ + + ]: 14825 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2052 : :
7466 2053 [ + + ]: 57605 : for (i = 0; i < num_elems; i++)
2054 : : {
2055 : : List *args;
2056 : : Selectivity s2;
2057 : :
2058 : 42780 : args = list_make2(leftop,
2059 : : makeConst(nominal_element_type,
2060 : : -1,
2061 : : nominal_element_collation,
2062 : : elmlen,
2063 : : elem_values[i],
2064 : : elem_nulls[i],
2065 : : elmbyval));
6471 2066 [ - + ]: 42780 : if (is_join_clause)
5049 tgl@sss.pgh.pa.us 2067 :UBC 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2068 : : clause->inputcollid,
2069 : : PointerGetDatum(root),
2070 : : ObjectIdGetDatum(operator),
2071 : : PointerGetDatum(args),
2072 : : Int16GetDatum(jointype),
2073 : : PointerGetDatum(sjinfo)));
2074 : : else
5049 tgl@sss.pgh.pa.us 2075 :CBC 42780 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2076 : : clause->inputcollid,
2077 : : PointerGetDatum(root),
2078 : : ObjectIdGetDatum(operator),
2079 : : PointerGetDatum(args),
2080 : : Int32GetDatum(varRelid)));
2081 : :
7466 2082 [ + + ]: 42780 : if (useOr)
2083 : : {
2084 : 36744 : s1 = s1 + s2 - s1 * s2;
5172 2085 [ + + ]: 36744 : if (isEquality)
2086 : 35874 : s1disjoint += s2;
2087 : : }
2088 : : else
2089 : : {
7466 2090 : 6036 : s1 = s1 * s2;
5172 2091 [ + + ]: 6036 : if (isInequality)
2092 : 5776 : s1disjoint += s2 - 1.0;
2093 : : }
2094 : : }
2095 : :
2096 : : /* accept disjoint-probability estimate if in range */
2097 [ + + + + : 14825 : if ((useOr ? isEquality : isInequality) &&
+ + ]
2098 [ + + ]: 14319 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2099 : 14292 : s1 = s1disjoint;
2100 : : }
7466 2101 [ + - + + ]: 3517 : else if (rightop && IsA(rightop, ArrayExpr) &&
2102 [ + - ]: 314 : !((ArrayExpr *) rightop)->multidims)
2103 : 309 : {
2104 : 314 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2105 : : int16 elmlen;
2106 : : bool elmbyval;
2107 : : ListCell *l;
2108 : :
2109 : 314 : get_typlenbyval(arrayexpr->element_typeid,
2110 : : &elmlen, &elmbyval);
2111 : :
2112 : : /*
2113 : : * We use the assumption of disjoint probabilities here too, although
2114 : : * the odds of equal array elements are rather higher if the elements
2115 : : * are not all constants (which they won't be, else constant folding
2116 : : * would have reduced the ArrayExpr to a Const). In this path it's
2117 : : * critical to have the sanity check on the s1disjoint estimate.
2118 : : */
5172 2119 [ + + ]: 314 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2120 : :
7466 2121 [ + - + + : 1147 : foreach(l, arrayexpr->elements)
+ + ]
2122 : : {
7037 2123 : 838 : Node *elem = (Node *) lfirst(l);
2124 : : List *args;
2125 : : Selectivity s2;
2126 : :
2127 : : /*
2128 : : * When the array contains a NULL constant, same as var_eq_const,
2129 : : * we assume the operator is strict and nothing will match, thus
2130 : : * return 0.0.
2131 : : */
47 drowley@postgresql.o 2132 [ + + + + :GNC 838 : if (!useOr && IsA(elem, Const) && ((Const *) elem)->constisnull)
+ + ]
2133 : 5 : return (Selectivity) 0.0;
2134 : :
2135 : : /*
2136 : : * Theoretically, if elem isn't of nominal_element_type we should
2137 : : * insert a RelabelType, but it seems unlikely that any operator
2138 : : * estimation function would really care ...
2139 : : */
7037 tgl@sss.pgh.pa.us 2140 :CBC 833 : args = list_make2(leftop, elem);
6471 2141 [ + + ]: 833 : if (is_join_clause)
5049 2142 : 3 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2143 : : clause->inputcollid,
2144 : : PointerGetDatum(root),
2145 : : ObjectIdGetDatum(operator),
2146 : : PointerGetDatum(args),
2147 : : Int16GetDatum(jointype),
2148 : : PointerGetDatum(sjinfo)));
2149 : : else
2150 : 830 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2151 : : clause->inputcollid,
2152 : : PointerGetDatum(root),
2153 : : ObjectIdGetDatum(operator),
2154 : : PointerGetDatum(args),
2155 : : Int32GetDatum(varRelid)));
2156 : :
7466 2157 [ + + ]: 833 : if (useOr)
2158 : : {
2159 : 813 : s1 = s1 + s2 - s1 * s2;
5172 2160 [ + - ]: 813 : if (isEquality)
2161 : 813 : s1disjoint += s2;
2162 : : }
2163 : : else
2164 : : {
7466 tgl@sss.pgh.pa.us 2165 :GBC 20 : s1 = s1 * s2;
5172 2166 [ + - ]: 20 : if (isInequality)
2167 : 20 : s1disjoint += s2 - 1.0;
2168 : : }
2169 : : }
2170 : :
2171 : : /* accept disjoint-probability estimate if in range */
5172 tgl@sss.pgh.pa.us 2172 [ + - + - :CBC 309 : if ((useOr ? isEquality : isInequality) &&
+ - ]
2173 [ + - ]: 309 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2174 : 309 : s1 = s1disjoint;
2175 : : }
2176 : : else
2177 : : {
2178 : : CaseTestExpr *dummyexpr;
2179 : : List *args;
2180 : : Selectivity s2;
2181 : : int i;
2182 : :
2183 : : /*
2184 : : * We need a dummy rightop to pass to the operator selectivity
2185 : : * routine. It can be pretty much anything that doesn't look like a
2186 : : * constant; CaseTestExpr is a convenient choice.
2187 : : */
7466 2188 : 3203 : dummyexpr = makeNode(CaseTestExpr);
7037 2189 : 3203 : dummyexpr->typeId = nominal_element_type;
7466 2190 : 3203 : dummyexpr->typeMod = -1;
5526 2191 : 3203 : dummyexpr->collation = clause->inputcollid;
7466 2192 : 3203 : args = list_make2(leftop, dummyexpr);
6471 2193 [ - + ]: 3203 : if (is_join_clause)
5049 tgl@sss.pgh.pa.us 2194 :UBC 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2195 : : clause->inputcollid,
2196 : : PointerGetDatum(root),
2197 : : ObjectIdGetDatum(operator),
2198 : : PointerGetDatum(args),
2199 : : Int16GetDatum(jointype),
2200 : : PointerGetDatum(sjinfo)));
2201 : : else
5049 tgl@sss.pgh.pa.us 2202 :CBC 3203 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2203 : : clause->inputcollid,
2204 : : PointerGetDatum(root),
2205 : : ObjectIdGetDatum(operator),
2206 : : PointerGetDatum(args),
2207 : : Int32GetDatum(varRelid)));
7466 2208 [ + - ]: 3203 : s1 = useOr ? 0.0 : 1.0;
2209 : :
2210 : : /*
2211 : : * Arbitrarily assume 10 elements in the eventual array value (see
2212 : : * also estimate_array_length). We don't risk an assumption of
2213 : : * disjoint probabilities here.
2214 : : */
2215 [ + + ]: 35233 : for (i = 0; i < 10; i++)
2216 : : {
2217 [ + - ]: 32030 : if (useOr)
2218 : 32030 : s1 = s1 + s2 - s1 * s2;
2219 : : else
7466 tgl@sss.pgh.pa.us 2220 :UBC 0 : s1 = s1 * s2;
2221 : : }
2222 : : }
2223 : :
2224 : : /* result should be in range, but make sure... */
7466 tgl@sss.pgh.pa.us 2225 [ - + - + ]:CBC 18337 : CLAMP_PROBABILITY(s1);
2226 : :
2227 : 18337 : return s1;
2228 : : }
2229 : :
2230 : : /*
2231 : : * Estimate number of elements in the array yielded by an expression.
2232 : : *
2233 : : * Note: the result is integral, but we use "double" to avoid overflow
2234 : : * concerns. Most callers will use it in double-type expressions anyway.
2235 : : *
2236 : : * Note: in some code paths root can be passed as NULL, resulting in
2237 : : * slightly worse estimates.
2238 : : */
2239 : : double
852 2240 : 87763 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2241 : : {
2242 : : /* look through any binary-compatible relabeling of arrayexpr */
7037 2243 : 87763 : arrayexpr = strip_array_coercion(arrayexpr);
2244 : :
7248 2245 [ + - + + ]: 87763 : if (arrayexpr && IsA(arrayexpr, Const))
2246 : : {
2247 : 38740 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2248 : 38740 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2249 : : ArrayType *arrayval;
2250 : :
2251 [ + + ]: 38740 : if (arrayisnull)
2252 : 75 : return 0;
2253 : 38665 : arrayval = DatumGetArrayTypeP(arraydatum);
2254 : 38665 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2255 : : }
2256 [ + - + + ]: 49023 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2257 [ + - ]: 548 : !((ArrayExpr *) arrayexpr)->multidims)
2258 : : {
2259 : 548 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2260 : : }
573 2261 [ + - + + ]: 48475 : else if (arrayexpr && root)
2262 : : {
2263 : : /* See if we can find any statistics about it */
2264 : : VariableStatData vardata;
2265 : : AttStatsSlot sslot;
852 2266 : 48455 : double nelem = 0;
2267 : :
2268 : : /*
2269 : : * Skip calling examine_variable for Var with varno 0, which has no
2270 : : * valid relation entry and would error in find_base_rel. Such a Var
2271 : : * can appear when a nested set operation's output type doesn't match
2272 : : * the parent's expected type, because recurse_set_operations builds a
2273 : : * projection target list using generate_setop_tlist with varno 0, and
2274 : : * if the required type coercion involves an ArrayCoerceExpr, we can
2275 : : * be called on that Var.
2276 : : */
24 rguo@postgresql.org 2277 [ + + + + ]: 48455 : if (IsA(arrayexpr, Var) && ((Var *) arrayexpr)->varno == 0)
2278 : 11693 : return 10; /* default guess, should match scalararraysel */
2279 : :
852 tgl@sss.pgh.pa.us 2280 : 48450 : examine_variable(root, arrayexpr, 0, &vardata);
2281 [ + + ]: 48450 : if (HeapTupleIsValid(vardata.statsTuple))
2282 : : {
2283 : : /*
2284 : : * Found stats, so use the average element count, which is stored
2285 : : * in the last stanumbers element of the DECHIST statistics.
2286 : : * Actually that is the average count of *distinct* elements;
2287 : : * perhaps we should scale it up somewhat?
2288 : : */
2289 [ + + ]: 11783 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2290 : : STATISTIC_KIND_DECHIST, InvalidOid,
2291 : : ATTSTATSSLOT_NUMBERS))
2292 : : {
2293 [ + - ]: 11688 : if (sslot.nnumbers > 0)
2294 : 11688 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2295 : 11688 : free_attstatsslot(&sslot);
2296 : : }
2297 : : }
2298 [ + + ]: 48450 : ReleaseVariableStats(vardata);
2299 : :
2300 [ + + ]: 48450 : if (nelem > 0)
2301 : 11688 : return nelem;
2302 : : }
2303 : :
2304 : : /* Else use a default guess --- this should match scalararraysel */
2305 : 36782 : return 10;
2306 : : }
2307 : :
2308 : : /*
2309 : : * rowcomparesel - Selectivity of RowCompareExpr Node.
2310 : : *
2311 : : * We estimate RowCompare selectivity by considering just the first (high
2312 : : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2313 : : * this estimate could be refined by considering additional columns, it
2314 : : * seems unlikely that we could do a lot better without multi-column
2315 : : * statistics.
2316 : : */
2317 : : Selectivity
7416 2318 : 240 : rowcomparesel(PlannerInfo *root,
2319 : : RowCompareExpr *clause,
2320 : : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2321 : : {
2322 : : Selectivity s1;
2323 : 240 : Oid opno = linitial_oid(clause->opnos);
5049 2324 : 240 : Oid inputcollid = linitial_oid(clause->inputcollids);
2325 : : List *opargs;
2326 : : bool is_join_clause;
2327 : :
2328 : : /* Build equivalent arg list for single operator */
7416 2329 : 240 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2330 : :
2331 : : /*
2332 : : * Decide if it's a join clause. This should match clausesel.c's
2333 : : * treat_as_join_clause(), except that we intentionally consider only the
2334 : : * leading columns and not the rest of the clause.
2335 : : */
2336 [ + + ]: 240 : if (varRelid != 0)
2337 : : {
2338 : : /*
2339 : : * Caller is forcing restriction mode (eg, because we are examining an
2340 : : * inner indexscan qual).
2341 : : */
6471 2342 : 45 : is_join_clause = false;
2343 : : }
2344 [ + + ]: 195 : else if (sjinfo == NULL)
2345 : : {
2346 : : /*
2347 : : * It must be a restriction clause, since it's being evaluated at a
2348 : : * scan node.
2349 : : */
7416 2350 : 175 : is_join_clause = false;
2351 : : }
2352 : : else
2353 : : {
2354 : : /*
2355 : : * Otherwise, it's a join if there's more than one base relation used.
2356 : : */
1930 2357 : 20 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2358 : : }
2359 : :
7416 2360 [ + + ]: 240 : if (is_join_clause)
2361 : : {
2362 : : /* Estimate selectivity for a join clause. */
2363 : 20 : s1 = join_selectivity(root, opno,
2364 : : opargs,
2365 : : inputcollid,
2366 : : jointype,
2367 : : sjinfo);
2368 : : }
2369 : : else
2370 : : {
2371 : : /* Estimate selectivity for a restriction clause. */
2372 : 220 : s1 = restriction_selectivity(root, opno,
2373 : : opargs,
2374 : : inputcollid,
2375 : : varRelid);
2376 : : }
2377 : :
2378 : 240 : return s1;
2379 : : }
2380 : :
2381 : : /*
2382 : : * eqjoinsel - Join selectivity of "="
2383 : : */
2384 : : Datum
9465 2385 : 218813 : eqjoinsel(PG_FUNCTION_ARGS)
2386 : : {
7639 2387 : 218813 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
9116 2388 : 218813 : Oid operator = PG_GETARG_OID(1);
2389 : 218813 : List *args = (List *) PG_GETARG_POINTER(2);
2390 : :
2391 : : #ifdef NOT_USED
2392 : : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2393 : : #endif
6471 2394 : 218813 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2160 2395 : 218813 : Oid collation = PG_GET_COLLATION();
2396 : : double selec;
2397 : : double selec_inner;
2398 : : VariableStatData vardata1;
2399 : : VariableStatData vardata2;
2400 : : double nd1;
2401 : : double nd2;
2402 : : bool isdefault1;
2403 : : bool isdefault2;
2404 : : Oid opfuncoid;
2405 : : FmgrInfo eqproc;
167 tgl@sss.pgh.pa.us 2406 :GNC 218813 : Oid hashLeft = InvalidOid;
2407 : 218813 : Oid hashRight = InvalidOid;
2408 : : AttStatsSlot sslot1;
2409 : : AttStatsSlot sslot2;
2720 tgl@sss.pgh.pa.us 2410 :CBC 218813 : Form_pg_statistic stats1 = NULL;
2411 : 218813 : Form_pg_statistic stats2 = NULL;
2412 : 218813 : bool have_mcvs1 = false;
2413 : 218813 : bool have_mcvs2 = false;
167 tgl@sss.pgh.pa.us 2414 :GNC 218813 : bool *hasmatch1 = NULL;
2415 : 218813 : bool *hasmatch2 = NULL;
2416 : 218813 : int nmatches = 0;
2417 : : bool get_mcv_stats;
2418 : : bool join_is_reversed;
2419 : : RelOptInfo *inner_rel;
2420 : :
6471 tgl@sss.pgh.pa.us 2421 :CBC 218813 : get_join_variables(root, args, sjinfo,
2422 : : &vardata1, &vardata2, &join_is_reversed);
2423 : :
2720 2424 : 218813 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2425 : 218813 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2426 : :
2427 : 218813 : opfuncoid = get_opcode(operator);
2428 : :
2429 : 218813 : memset(&sslot1, 0, sizeof(sslot1));
2430 : 218813 : memset(&sslot2, 0, sizeof(sslot2));
2431 : :
2432 : : /*
2433 : : * There is no use in fetching one side's MCVs if we lack MCVs for the
2434 : : * other side, so do a quick check to verify that both stats exist.
2435 : : */
1264 2436 : 593211 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2437 [ + + + + ]: 277414 : HeapTupleIsValid(vardata2.statsTuple) &&
2438 : 121829 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2439 : : STATISTIC_KIND_MCV, InvalidOid,
2440 [ + + + + ]: 374398 : 0) &&
2441 : 58359 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2442 : : STATISTIC_KIND_MCV, InvalidOid,
2443 : : 0));
2444 : :
2720 2445 [ + + ]: 218813 : if (HeapTupleIsValid(vardata1.statsTuple))
2446 : : {
2447 : : /* note we allow use of nullfrac regardless of security check */
2448 : 155585 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
1264 2449 [ + + + - ]: 179173 : if (get_mcv_stats &&
2450 : 23588 : statistic_proc_security_check(&vardata1, opfuncoid))
2720 2451 : 23588 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2452 : : STATISTIC_KIND_MCV, InvalidOid,
2453 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2454 : : }
2455 : :
2456 [ + + ]: 218813 : if (HeapTupleIsValid(vardata2.statsTuple))
2457 : : {
2458 : : /* note we allow use of nullfrac regardless of security check */
2459 : 138089 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
1264 2460 [ + + + - ]: 161677 : if (get_mcv_stats &&
2461 : 23588 : statistic_proc_security_check(&vardata2, opfuncoid))
2720 2462 : 23588 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2463 : : STATISTIC_KIND_MCV, InvalidOid,
2464 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2465 : : }
2466 : :
2467 : : /* Prepare info usable by both eqjoinsel_inner and eqjoinsel_semi */
167 tgl@sss.pgh.pa.us 2468 [ + + + - ]:GNC 218813 : if (have_mcvs1 && have_mcvs2)
2469 : : {
2470 : 23588 : fmgr_info(opfuncoid, &eqproc);
2471 : 23588 : hasmatch1 = (bool *) palloc0(sslot1.nvalues * sizeof(bool));
2472 : 23588 : hasmatch2 = (bool *) palloc0(sslot2.nvalues * sizeof(bool));
2473 : :
2474 : : /*
2475 : : * If the MCV lists are long enough to justify hashing, try to look up
2476 : : * hash functions for the join operator.
2477 : : */
2478 [ + + ]: 23588 : if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
2479 : 9737 : (void) get_op_hash_functions(operator, &hashLeft, &hashRight);
2480 : : }
2481 : : else
2482 : 195225 : memset(&eqproc, 0, sizeof(eqproc)); /* silence uninit-var warnings */
2483 : :
2484 : : /* We need to compute the inner-join selectivity in all cases */
2485 : 218813 : selec_inner = eqjoinsel_inner(&eqproc, collation,
2486 : : hashLeft, hashRight,
2487 : : &vardata1, &vardata2,
2488 : : nd1, nd2,
2489 : : isdefault1, isdefault2,
2490 : : &sslot1, &sslot2,
2491 : : stats1, stats2,
2492 : : have_mcvs1, have_mcvs2,
2493 : : hasmatch1, hasmatch2,
2494 : : &nmatches);
2495 : :
6471 tgl@sss.pgh.pa.us 2496 [ + + - ]:CBC 218813 : switch (sjinfo->jointype)
2497 : : {
2498 : 207419 : case JOIN_INNER:
2499 : : case JOIN_LEFT:
2500 : : case JOIN_FULL:
2720 2501 : 207419 : selec = selec_inner;
6471 2502 : 207419 : break;
2503 : 11394 : case JOIN_SEMI:
2504 : : case JOIN_ANTI:
2505 : :
2506 : : /*
2507 : : * Look up the join's inner relation. min_righthand is sufficient
2508 : : * information because neither SEMI nor ANTI joins permit any
2509 : : * reassociation into or out of their RHS, so the righthand will
2510 : : * always be exactly that set of rels.
2511 : : */
5360 2512 : 11394 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2513 : :
6471 2514 [ + + ]: 11394 : if (!join_is_reversed)
167 tgl@sss.pgh.pa.us 2515 :GNC 5413 : selec = eqjoinsel_semi(&eqproc, collation,
2516 : : hashLeft, hashRight,
2517 : : false,
2518 : : &vardata1, &vardata2,
2519 : : nd1, nd2,
2520 : : isdefault1, isdefault2,
2521 : : &sslot1, &sslot2,
2522 : : stats1, stats2,
2523 : : have_mcvs1, have_mcvs2,
2524 : : hasmatch1, hasmatch2,
2525 : : &nmatches,
2526 : : inner_rel);
2527 : : else
2528 : 5981 : selec = eqjoinsel_semi(&eqproc, collation,
2529 : : hashLeft, hashRight,
2530 : : true,
2531 : : &vardata2, &vardata1,
2532 : : nd2, nd1,
2533 : : isdefault2, isdefault1,
2534 : : &sslot2, &sslot1,
2535 : : stats2, stats1,
2536 : : have_mcvs2, have_mcvs1,
2537 : : hasmatch2, hasmatch1,
2538 : : &nmatches,
2539 : : inner_rel);
2540 : :
2541 : : /*
2542 : : * We should never estimate the output of a semijoin to be more
2543 : : * rows than we estimate for an inner join with the same input
2544 : : * rels and join condition; it's obviously impossible for that to
2545 : : * happen. The former estimate is N1 * Ssemi while the latter is
2546 : : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2547 : : * this is worthwhile because of the shakier estimation rules we
2548 : : * use in eqjoinsel_semi, particularly in cases where it has to
2549 : : * punt entirely.
2550 : : */
2720 tgl@sss.pgh.pa.us 2551 [ + + ]:CBC 11394 : selec = Min(selec, inner_rel->rows * selec_inner);
6471 2552 : 11394 : break;
6471 tgl@sss.pgh.pa.us 2553 :UBC 0 : default:
2554 : : /* other values not expected here */
2555 [ # # ]: 0 : elog(ERROR, "unrecognized join type: %d",
2556 : : (int) sjinfo->jointype);
2557 : : selec = 0; /* keep compiler quiet */
2558 : : break;
2559 : : }
2560 : :
2720 tgl@sss.pgh.pa.us 2561 :CBC 218813 : free_attstatsslot(&sslot1);
2562 : 218813 : free_attstatsslot(&sslot2);
2563 : :
6471 2564 [ + + ]: 218813 : ReleaseVariableStats(vardata1);
2565 [ + + ]: 218813 : ReleaseVariableStats(vardata2);
2566 : :
167 tgl@sss.pgh.pa.us 2567 [ + + ]:GNC 218813 : if (hasmatch1)
2568 : 23588 : pfree(hasmatch1);
2569 [ + + ]: 218813 : if (hasmatch2)
2570 : 23588 : pfree(hasmatch2);
2571 : :
6471 tgl@sss.pgh.pa.us 2572 [ - + - + ]:CBC 218813 : CLAMP_PROBABILITY(selec);
2573 : :
2574 : 218813 : PG_RETURN_FLOAT8((float8) selec);
2575 : : }
2576 : :
2577 : : /*
2578 : : * eqjoinsel_inner --- eqjoinsel for normal inner join
2579 : : *
2580 : : * In addition to computing the selectivity estimate, this will fill
2581 : : * hasmatch1[], hasmatch2[], and *p_nmatches (if have_mcvs1 && have_mcvs2).
2582 : : * We may be able to re-use that data in eqjoinsel_semi.
2583 : : *
2584 : : * We also use this for LEFT/FULL outer joins; it's not presently clear
2585 : : * that it's worth trying to distinguish them here.
2586 : : */
2587 : : static double
167 tgl@sss.pgh.pa.us 2588 :GNC 218813 : eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
2589 : : Oid hashLeft, Oid hashRight,
2590 : : VariableStatData *vardata1, VariableStatData *vardata2,
2591 : : double nd1, double nd2,
2592 : : bool isdefault1, bool isdefault2,
2593 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2594 : : Form_pg_statistic stats1, Form_pg_statistic stats2,
2595 : : bool have_mcvs1, bool have_mcvs2,
2596 : : bool *hasmatch1, bool *hasmatch2,
2597 : : int *p_nmatches)
2598 : : {
2599 : : double selec;
2600 : :
8113 tgl@sss.pgh.pa.us 2601 [ + + + - ]:CBC 218813 : if (have_mcvs1 && have_mcvs2)
10467 bruce@momjian.us 2602 : 23588 : {
2603 : : /*
2604 : : * We have most-common-value lists for both relations. Run through
2605 : : * the lists to see which MCVs actually join to each other with the
2606 : : * given operator. This allows us to determine the exact join
2607 : : * selectivity for the portion of the relations represented by the MCV
2608 : : * lists. We still have to estimate for the remaining population, but
2609 : : * in a skewed distribution this gives us a big leg up in accuracy.
2610 : : * For motivation see the analysis in Y. Ioannidis and S.
2611 : : * Christodoulakis, "On the propagation of errors in the size of join
2612 : : * results", Technical Report 1018, Computer Science Dept., University
2613 : : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2614 : : */
8113 tgl@sss.pgh.pa.us 2615 : 23588 : double nullfrac1 = stats1->stanullfrac;
2616 : 23588 : double nullfrac2 = stats2->stanullfrac;
2617 : : double matchprodfreq,
2618 : : matchfreq1,
2619 : : matchfreq2,
2620 : : unmatchfreq1,
2621 : : unmatchfreq2,
2622 : : otherfreq1,
2623 : : otherfreq2,
2624 : : totalsel1,
2625 : : totalsel2;
2626 : : int i,
2627 : : nmatches;
2628 : :
2629 : : /* Fill the match arrays */
167 tgl@sss.pgh.pa.us 2630 :GNC 23588 : eqjoinsel_find_matches(eqproc, collation,
2631 : : hashLeft, hashRight,
2632 : : false,
2633 : : sslot1, sslot2,
2634 : : sslot1->nvalues, sslot2->nvalues,
2635 : : hasmatch1, hasmatch2,
2636 : : p_nmatches, &matchprodfreq);
2637 : 23588 : nmatches = *p_nmatches;
8113 tgl@sss.pgh.pa.us 2638 [ - + - + ]:CBC 23588 : CLAMP_PROBABILITY(matchprodfreq);
2639 : :
2640 : : /* Sum up frequencies of matched and unmatched MCVs */
2641 : 23588 : matchfreq1 = unmatchfreq1 = 0.0;
2720 2642 [ + + ]: 537528 : for (i = 0; i < sslot1->nvalues; i++)
2643 : : {
8113 2644 [ + + ]: 513940 : if (hasmatch1[i])
2720 2645 : 232059 : matchfreq1 += sslot1->numbers[i];
2646 : : else
2647 : 281881 : unmatchfreq1 += sslot1->numbers[i];
2648 : : }
8113 2649 [ - + + + ]: 23588 : CLAMP_PROBABILITY(matchfreq1);
2650 [ - + - + ]: 23588 : CLAMP_PROBABILITY(unmatchfreq1);
2651 : 23588 : matchfreq2 = unmatchfreq2 = 0.0;
2720 2652 [ + + ]: 436124 : for (i = 0; i < sslot2->nvalues; i++)
2653 : : {
8113 2654 [ + + ]: 412536 : if (hasmatch2[i])
2720 2655 : 232059 : matchfreq2 += sslot2->numbers[i];
2656 : : else
2657 : 180477 : unmatchfreq2 += sslot2->numbers[i];
2658 : : }
8113 2659 [ - + + + ]: 23588 : CLAMP_PROBABILITY(matchfreq2);
2660 [ - + - + ]: 23588 : CLAMP_PROBABILITY(unmatchfreq2);
2661 : :
2662 : : /*
2663 : : * Compute total frequency of non-null values that are not in the MCV
2664 : : * lists.
2665 : : */
2666 : 23588 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2667 : 23588 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2668 [ + + - + ]: 23588 : CLAMP_PROBABILITY(otherfreq1);
2669 [ + + - + ]: 23588 : CLAMP_PROBABILITY(otherfreq2);
2670 : :
2671 : : /*
2672 : : * We can estimate the total selectivity from the point of view of
2673 : : * relation 1 as: the known selectivity for matched MCVs, plus
2674 : : * unmatched MCVs that are assumed to match against random members of
2675 : : * relation 2's non-MCV population, plus non-MCV values that are
2676 : : * assumed to match against random members of relation 2's unmatched
2677 : : * MCVs plus non-MCV values.
2678 : : */
2679 : 23588 : totalsel1 = matchprodfreq;
2720 2680 [ + + ]: 23588 : if (nd2 > sslot2->nvalues)
2681 : 5330 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
8113 2682 [ + + ]: 23588 : if (nd2 > nmatches)
2683 : 9265 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2684 : 9265 : (nd2 - nmatches);
2685 : : /* Same estimate from the point of view of relation 2. */
2686 : 23588 : totalsel2 = matchprodfreq;
2720 2687 [ + + ]: 23588 : if (nd1 > sslot1->nvalues)
2688 : 5092 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
8113 2689 [ + + ]: 23588 : if (nd1 > nmatches)
2690 : 8249 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2691 : 8249 : (nd1 - nmatches);
2692 : :
2693 : : /*
2694 : : * Use the smaller of the two estimates. This can be justified in
2695 : : * essentially the same terms as given below for the no-stats case: to
2696 : : * a first approximation, we are estimating from the point of view of
2697 : : * the relation with smaller nd.
2698 : : */
2699 [ + + ]: 23588 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2700 : : }
2701 : : else
2702 : : {
2703 : : /*
2704 : : * We do not have MCV lists for both sides. Estimate the join
2705 : : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2706 : : * is plausible if we assume that the join operator is strict and the
2707 : : * non-null values are about equally distributed: a given non-null
2708 : : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2709 : : * of rel2, so total join rows are at most
2710 : : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2711 : : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2712 : : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2713 : : * with MIN() is an upper bound. Using the MIN() means we estimate
2714 : : * from the point of view of the relation with smaller nd (since the
2715 : : * larger nd is determining the MIN). It is reasonable to assume that
2716 : : * most tuples in this rel will have join partners, so the bound is
2717 : : * probably reasonably tight and should be taken as-is.
2718 : : *
2719 : : * XXX Can we be smarter if we have an MCV list for just one side? It
2720 : : * seems that if we assume equal distribution for the other side, we
2721 : : * end up with the same answer anyway.
2722 : : */
2723 [ + + ]: 195225 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2724 [ + + ]: 195225 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2725 : :
2726 : 195225 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2727 [ + + ]: 195225 : if (nd1 > nd2)
2728 : 103422 : selec /= nd1;
2729 : : else
2730 : 91803 : selec /= nd2;
2731 : : }
2732 : :
6471 2733 : 218813 : return selec;
2734 : : }
2735 : :
2736 : : /*
2737 : : * eqjoinsel_semi --- eqjoinsel for semi join
2738 : : *
2739 : : * (Also used for anti join, which we are supposed to estimate the same way.)
2740 : : * Caller has ensured that vardata1 is the LHS variable; however, eqproc
2741 : : * is for the original join operator, which might now need to have the inputs
2742 : : * swapped in order to apply correctly. Also, if have_mcvs1 && have_mcvs2
2743 : : * then hasmatch1[], hasmatch2[], and *p_nmatches were filled by
2744 : : * eqjoinsel_inner.
2745 : : */
2746 : : static double
167 tgl@sss.pgh.pa.us 2747 :GNC 11394 : eqjoinsel_semi(FmgrInfo *eqproc, Oid collation,
2748 : : Oid hashLeft, Oid hashRight,
2749 : : bool op_is_reversed,
2750 : : VariableStatData *vardata1, VariableStatData *vardata2,
2751 : : double nd1, double nd2,
2752 : : bool isdefault1, bool isdefault2,
2753 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2754 : : Form_pg_statistic stats1, Form_pg_statistic stats2,
2755 : : bool have_mcvs1, bool have_mcvs2,
2756 : : bool *hasmatch1, bool *hasmatch2,
2757 : : int *p_nmatches,
2758 : : RelOptInfo *inner_rel)
2759 : : {
2760 : : double selec;
2761 : :
2762 : : /*
2763 : : * We clamp nd2 to be not more than what we estimate the inner relation's
2764 : : * size to be. This is intuitively somewhat reasonable since obviously
2765 : : * there can't be more than that many distinct values coming from the
2766 : : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2767 : : * likewise) is that this is the only pathway by which restriction clauses
2768 : : * applied to the inner rel will affect the join result size estimate,
2769 : : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2770 : : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2771 : : * the selectivity of outer-rel restrictions.
2772 : : *
2773 : : * We can apply this clamping both with respect to the base relation from
2774 : : * which the join variable comes (if there is just one), and to the
2775 : : * immediate inner input relation of the current join.
2776 : : *
2777 : : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2778 : : * great, maybe, but it didn't come out of nowhere either. This is most
2779 : : * helpful when the inner relation is empty and consequently has no stats.
2780 : : */
5360 tgl@sss.pgh.pa.us 2781 [ + + ]:CBC 11394 : if (vardata2->rel)
2782 : : {
3444 2783 [ + + ]: 11389 : if (nd2 >= vardata2->rel->rows)
2784 : : {
2785 : 9637 : nd2 = vardata2->rel->rows;
2786 : 9637 : isdefault2 = false;
2787 : : }
2788 : : }
2789 [ + + ]: 11394 : if (nd2 >= inner_rel->rows)
2790 : : {
2791 : 9591 : nd2 = inner_rel->rows;
2792 : 9591 : isdefault2 = false;
2793 : : }
2794 : :
167 tgl@sss.pgh.pa.us 2795 [ + + + - ]:GNC 11394 : if (have_mcvs1 && have_mcvs2)
6471 tgl@sss.pgh.pa.us 2796 :CBC 525 : {
2797 : : /*
2798 : : * We have most-common-value lists for both relations. Run through
2799 : : * the lists to see which MCVs actually join to each other with the
2800 : : * given operator. This allows us to determine the exact join
2801 : : * selectivity for the portion of the relations represented by the MCV
2802 : : * lists. We still have to estimate for the remaining population, but
2803 : : * in a skewed distribution this gives us a big leg up in accuracy.
2804 : : */
2805 : 525 : double nullfrac1 = stats1->stanullfrac;
2806 : : double matchprodfreq,
2807 : : matchfreq1,
2808 : : uncertainfrac,
2809 : : uncertain;
2810 : : int i,
2811 : : nmatches,
2812 : : clamped_nvalues2;
2813 : :
2814 : : /*
2815 : : * The clamping above could have resulted in nd2 being less than
2816 : : * sslot2->nvalues; in which case, we assume that precisely the nd2
2817 : : * most common values in the relation will appear in the join input,
2818 : : * and so compare to only the first nd2 members of the MCV list. Of
2819 : : * course this is frequently wrong, but it's the best bet we can make.
2820 : : */
2720 2821 [ + + ]: 525 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2822 : :
2823 : : /*
2824 : : * If we did not set clamped_nvalues2 to less than sslot2->nvalues,
2825 : : * then the hasmatch1[] and hasmatch2[] match flags computed by
2826 : : * eqjoinsel_inner are still perfectly applicable, so we need not
2827 : : * re-do the matching work. Note that it does not matter if
2828 : : * op_is_reversed: we'd get the same answers.
2829 : : *
2830 : : * If we did clamp, then a different set of sslot2 values is to be
2831 : : * compared, so we have to re-do the matching.
2832 : : */
167 tgl@sss.pgh.pa.us 2833 [ - + ]:GNC 525 : if (clamped_nvalues2 != sslot2->nvalues)
2834 : : {
2835 : : /* Must re-zero the arrays */
167 tgl@sss.pgh.pa.us 2836 :UNC 0 : memset(hasmatch1, 0, sslot1->nvalues * sizeof(bool));
2837 : 0 : memset(hasmatch2, 0, clamped_nvalues2 * sizeof(bool));
2838 : : /* Re-fill the match arrays */
2839 : 0 : eqjoinsel_find_matches(eqproc, collation,
2840 : : hashLeft, hashRight,
2841 : : op_is_reversed,
2842 : : sslot1, sslot2,
2843 : : sslot1->nvalues, clamped_nvalues2,
2844 : : hasmatch1, hasmatch2,
2845 : : p_nmatches, &matchprodfreq);
2846 : : }
167 tgl@sss.pgh.pa.us 2847 :GNC 525 : nmatches = *p_nmatches;
2848 : :
2849 : : /* Sum up frequencies of matched MCVs */
6471 tgl@sss.pgh.pa.us 2850 :CBC 525 : matchfreq1 = 0.0;
2720 2851 [ + + ]: 11907 : for (i = 0; i < sslot1->nvalues; i++)
2852 : : {
6471 2853 [ + + ]: 11382 : if (hasmatch1[i])
2720 2854 : 9490 : matchfreq1 += sslot1->numbers[i];
2855 : : }
6471 2856 [ - + + + ]: 525 : CLAMP_PROBABILITY(matchfreq1);
2857 : :
2858 : : /*
2859 : : * Now we need to estimate the fraction of relation 1 that has at
2860 : : * least one join partner. We know for certain that the matched MCVs
2861 : : * do, so that gives us a lower bound, but we're really in the dark
2862 : : * about everything else. Our crude approach is: if nd1 <= nd2 then
2863 : : * assume all non-null rel1 rows have join partners, else assume for
2864 : : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2865 : : * can discount the known-matched MCVs from the distinct-values counts
2866 : : * before doing the division.
2867 : : *
2868 : : * Crude as the above is, it's completely useless if we don't have
2869 : : * reliable ndistinct values for both sides. Hence, if either nd1 or
2870 : : * nd2 is default, punt and assume half of the uncertain rows have
2871 : : * join partners.
2872 : : */
5357 2873 [ + - + - ]: 525 : if (!isdefault1 && !isdefault2)
2874 : : {
5502 2875 : 525 : nd1 -= nmatches;
2876 : 525 : nd2 -= nmatches;
5360 2877 [ + + - + ]: 525 : if (nd1 <= nd2 || nd2 < 0)
5502 2878 : 495 : uncertainfrac = 1.0;
2879 : : else
2880 : 30 : uncertainfrac = nd2 / nd1;
2881 : : }
2882 : : else
5502 tgl@sss.pgh.pa.us 2883 :UBC 0 : uncertainfrac = 0.5;
5502 tgl@sss.pgh.pa.us 2884 :CBC 525 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2885 [ - + - + ]: 525 : CLAMP_PROBABILITY(uncertain);
2886 : 525 : selec = matchfreq1 + uncertainfrac * uncertain;
2887 : : }
2888 : : else
2889 : : {
2890 : : /*
2891 : : * Without MCV lists for both sides, we can only use the heuristic
2892 : : * about nd1 vs nd2.
2893 : : */
6471 2894 [ + + ]: 10869 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2895 : :
5357 2896 [ + + + + ]: 10869 : if (!isdefault1 && !isdefault2)
2897 : : {
5360 2898 [ + + - + ]: 8068 : if (nd1 <= nd2 || nd2 < 0)
5502 2899 : 3728 : selec = 1.0 - nullfrac1;
2900 : : else
2901 : 4340 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2902 : : }
2903 : : else
2904 : 2801 : selec = 0.5 * (1.0 - nullfrac1);
2905 : : }
2906 : :
6471 2907 : 11394 : return selec;
2908 : : }
2909 : :
2910 : : /*
2911 : : * Identify matching MCVs for eqjoinsel_inner or eqjoinsel_semi.
2912 : : *
2913 : : * Inputs:
2914 : : * eqproc: FmgrInfo for equality function to use (might be reversed)
2915 : : * collation: OID of collation to use
2916 : : * hashLeft, hashRight: OIDs of hash functions associated with equality op,
2917 : : * or InvalidOid if we're not to use hashing
2918 : : * op_is_reversed: indicates that eqproc compares right type to left type
2919 : : * sslot1, sslot2: MCV values for the lefthand and righthand inputs
2920 : : * nvalues1, nvalues2: number of values to be considered (can be less than
2921 : : * sslotN->nvalues, but not more)
2922 : : * Outputs:
2923 : : * hasmatch1[], hasmatch2[]: pre-zeroed arrays of lengths nvalues1, nvalues2;
2924 : : * entries are set to true if that MCV has a match on the other side
2925 : : * *p_nmatches: receives number of MCV pairs that match
2926 : : * *p_matchprodfreq: receives sum(sslot1->numbers[i] * sslot2->numbers[j])
2927 : : * for matching MCVs
2928 : : *
2929 : : * Note that hashLeft is for the eqproc's left-hand input type, hashRight
2930 : : * for its right, regardless of op_is_reversed.
2931 : : *
2932 : : * Note we assume that each MCV will match at most one member of the other
2933 : : * MCV list. If the operator isn't really equality, there could be multiple
2934 : : * matches --- but we don't look for them, both for speed and because the
2935 : : * math wouldn't add up...
2936 : : */
2937 : : static void
167 tgl@sss.pgh.pa.us 2938 :GNC 23588 : eqjoinsel_find_matches(FmgrInfo *eqproc, Oid collation,
2939 : : Oid hashLeft, Oid hashRight,
2940 : : bool op_is_reversed,
2941 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2942 : : int nvalues1, int nvalues2,
2943 : : bool *hasmatch1, bool *hasmatch2,
2944 : : int *p_nmatches, double *p_matchprodfreq)
2945 : : {
2946 : 23588 : LOCAL_FCINFO(fcinfo, 2);
2947 : 23588 : double matchprodfreq = 0.0;
2948 : 23588 : int nmatches = 0;
2949 : :
2950 : : /*
2951 : : * Save a few cycles by setting up the fcinfo struct just once. Using
2952 : : * FunctionCallInvoke directly also avoids failure if the eqproc returns
2953 : : * NULL, though really equality functions should never do that.
2954 : : */
2955 : 23588 : InitFunctionCallInfoData(*fcinfo, eqproc, 2, collation,
2956 : : NULL, NULL);
2957 : 23588 : fcinfo->args[0].isnull = false;
2958 : 23588 : fcinfo->args[1].isnull = false;
2959 : :
2960 [ + + + - ]: 23588 : if (OidIsValid(hashLeft) && OidIsValid(hashRight))
2961 : 9737 : {
2962 : : /* Use a hash table to speed up the matching */
2963 : 9737 : LOCAL_FCINFO(hash_fcinfo, 1);
2964 : : FmgrInfo hash_proc;
2965 : : MCVHashContext hashContext;
2966 : : MCVHashTable_hash *hashTable;
2967 : : AttStatsSlot *statsProbe;
2968 : : AttStatsSlot *statsHash;
2969 : : bool *hasMatchProbe;
2970 : : bool *hasMatchHash;
2971 : : int nvaluesProbe;
2972 : : int nvaluesHash;
2973 : :
2974 : : /* Make sure we build the hash table on the smaller array. */
2975 [ + + ]: 9737 : if (sslot1->nvalues >= sslot2->nvalues)
2976 : : {
2977 : 8224 : statsProbe = sslot1;
2978 : 8224 : statsHash = sslot2;
2979 : 8224 : hasMatchProbe = hasmatch1;
2980 : 8224 : hasMatchHash = hasmatch2;
2981 : 8224 : nvaluesProbe = nvalues1;
2982 : 8224 : nvaluesHash = nvalues2;
2983 : : }
2984 : : else
2985 : : {
2986 : : /* We'll have to reverse the direction of use of the operator. */
2987 : 1513 : op_is_reversed = !op_is_reversed;
2988 : 1513 : statsProbe = sslot2;
2989 : 1513 : statsHash = sslot1;
2990 : 1513 : hasMatchProbe = hasmatch2;
2991 : 1513 : hasMatchHash = hasmatch1;
2992 : 1513 : nvaluesProbe = nvalues2;
2993 : 1513 : nvaluesHash = nvalues1;
2994 : : }
2995 : :
2996 : : /*
2997 : : * Build the hash table on the smaller array, using the appropriate
2998 : : * hash function for its data type.
2999 : : */
3000 [ + + ]: 9737 : fmgr_info(op_is_reversed ? hashLeft : hashRight, &hash_proc);
3001 : 9737 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
3002 : : NULL, NULL);
3003 : 9737 : hash_fcinfo->args[0].isnull = false;
3004 : :
3005 : 9737 : hashContext.equal_fcinfo = fcinfo;
3006 : 9737 : hashContext.hash_fcinfo = hash_fcinfo;
3007 : 9737 : hashContext.op_is_reversed = op_is_reversed;
3008 : 9737 : hashContext.insert_mode = true;
3009 : 9737 : get_typlenbyval(statsHash->valuetype,
3010 : : &hashContext.hash_typlen,
3011 : : &hashContext.hash_typbyval);
3012 : :
3013 : 9737 : hashTable = MCVHashTable_create(CurrentMemoryContext,
3014 : : nvaluesHash,
3015 : : &hashContext);
3016 : :
3017 [ + + ]: 313126 : for (int i = 0; i < nvaluesHash; i++)
3018 : : {
3019 : 303389 : bool found = false;
3020 : 303389 : MCVHashEntry *entry = MCVHashTable_insert(hashTable,
3021 : 303389 : statsHash->values[i],
3022 : : &found);
3023 : :
3024 : : /*
3025 : : * MCVHashTable_insert will only report "found" if the new value
3026 : : * is equal to some previous one per datum_image_eq(). That
3027 : : * probably shouldn't happen, since we're not expecting duplicates
3028 : : * in the MCV list. If we do find a dup, just ignore it, leaving
3029 : : * the hash entry's index pointing at the first occurrence. That
3030 : : * matches the behavior that the non-hashed code path would have.
3031 : : */
3032 [ + - ]: 303389 : if (likely(!found))
3033 : 303389 : entry->index = i;
3034 : : }
3035 : :
3036 : : /*
3037 : : * Prepare to probe the hash table. If the probe values are of a
3038 : : * different data type, then we need to change hash functions. (This
3039 : : * code relies on the assumption that since we defined SH_STORE_HASH,
3040 : : * simplehash.h will never need to compute hash values for existing
3041 : : * hash table entries.)
3042 : : */
3043 : 9737 : hashContext.insert_mode = false;
3044 [ + + ]: 9737 : if (hashLeft != hashRight)
3045 : : {
3046 [ + + ]: 1264 : fmgr_info(op_is_reversed ? hashRight : hashLeft, &hash_proc);
3047 : : /* Resetting hash_fcinfo is probably unnecessary, but be safe */
3048 : 1264 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
3049 : : NULL, NULL);
3050 : 1264 : hash_fcinfo->args[0].isnull = false;
3051 : : }
3052 : :
3053 : : /* Look up each probe value in turn. */
3054 [ + + ]: 501477 : for (int i = 0; i < nvaluesProbe; i++)
3055 : : {
3056 : 491740 : MCVHashEntry *entry = MCVHashTable_lookup(hashTable,
3057 : 491740 : statsProbe->values[i]);
3058 : :
3059 : : /* As in the other code path, skip already-matched hash entries */
3060 [ + + + - ]: 491740 : if (entry != NULL && !hasMatchHash[entry->index])
3061 : : {
3062 : 179954 : hasMatchHash[entry->index] = hasMatchProbe[i] = true;
3063 : 179954 : nmatches++;
3064 : 179954 : matchprodfreq += statsHash->numbers[entry->index] * statsProbe->numbers[i];
3065 : : }
3066 : : }
3067 : :
3068 : 9737 : MCVHashTable_destroy(hashTable);
3069 : : }
3070 : : else
3071 : : {
3072 : : /* We're not to use hashing, so do it the O(N^2) way */
3073 : : int index1,
3074 : : index2;
3075 : :
3076 : : /* Set up to supply the values in the order the operator expects */
3077 [ - + ]: 13851 : if (op_is_reversed)
3078 : : {
167 tgl@sss.pgh.pa.us 3079 :UNC 0 : index1 = 1;
3080 : 0 : index2 = 0;
3081 : : }
3082 : : else
3083 : : {
167 tgl@sss.pgh.pa.us 3084 :GNC 13851 : index1 = 0;
3085 : 13851 : index2 = 1;
3086 : : }
3087 : :
3088 [ + + ]: 70243 : for (int i = 0; i < nvalues1; i++)
3089 : : {
3090 : 56392 : fcinfo->args[index1].value = sslot1->values[i];
3091 : :
3092 [ + + ]: 200926 : for (int j = 0; j < nvalues2; j++)
3093 : : {
3094 : : Datum fresult;
3095 : :
3096 [ + + ]: 196639 : if (hasmatch2[j])
3097 : 95118 : continue;
3098 : 101521 : fcinfo->args[index2].value = sslot2->values[j];
3099 : 101521 : fcinfo->isnull = false;
3100 : 101521 : fresult = FunctionCallInvoke(fcinfo);
3101 [ + - + + ]: 101521 : if (!fcinfo->isnull && DatumGetBool(fresult))
3102 : : {
3103 : 52105 : hasmatch1[i] = hasmatch2[j] = true;
3104 : 52105 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
3105 : 52105 : nmatches++;
3106 : 52105 : break;
3107 : : }
3108 : : }
3109 : : }
3110 : : }
3111 : :
3112 : 23588 : *p_nmatches = nmatches;
3113 : 23588 : *p_matchprodfreq = matchprodfreq;
3114 : 23588 : }
3115 : :
3116 : : /*
3117 : : * Support functions for the hash tables used by eqjoinsel_find_matches
3118 : : */
3119 : : static uint32
3120 : 795129 : hash_mcv(MCVHashTable_hash *tab, Datum key)
3121 : : {
3122 : 795129 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3123 : 795129 : FunctionCallInfo fcinfo = context->hash_fcinfo;
3124 : : Datum fresult;
3125 : :
3126 : 795129 : fcinfo->args[0].value = key;
3127 : 795129 : fcinfo->isnull = false;
3128 : 795129 : fresult = FunctionCallInvoke(fcinfo);
3129 [ - + ]: 795129 : Assert(!fcinfo->isnull);
3130 : 795129 : return DatumGetUInt32(fresult);
3131 : : }
3132 : :
3133 : : static bool
3134 : 179954 : mcvs_equal(MCVHashTable_hash *tab, Datum key0, Datum key1)
3135 : : {
3136 : 179954 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3137 : :
3138 [ - + ]: 179954 : if (context->insert_mode)
3139 : : {
3140 : : /*
3141 : : * During the insertion step, any comparisons will be between two
3142 : : * Datums of the hash table's data type, so if the given operator is
3143 : : * cross-type it will be the wrong thing to use. Fortunately, we can
3144 : : * use datum_image_eq instead. The MCV values should all be distinct
3145 : : * anyway, so it's mostly pro-forma to compare them at all.
3146 : : */
167 tgl@sss.pgh.pa.us 3147 :UNC 0 : return datum_image_eq(key0, key1,
3148 : 0 : context->hash_typbyval, context->hash_typlen);
3149 : : }
3150 : : else
3151 : : {
167 tgl@sss.pgh.pa.us 3152 :GNC 179954 : FunctionCallInfo fcinfo = context->equal_fcinfo;
3153 : : Datum fresult;
3154 : :
3155 : : /*
3156 : : * Apply the operator the correct way around. Although simplehash.h
3157 : : * doesn't document this explicitly, during lookups key0 is from the
3158 : : * hash table while key1 is the probe value, so we should compare them
3159 : : * in that order only if op_is_reversed.
3160 : : */
3161 [ + + ]: 179954 : if (context->op_is_reversed)
3162 : : {
3163 : 45959 : fcinfo->args[0].value = key0;
3164 : 45959 : fcinfo->args[1].value = key1;
3165 : : }
3166 : : else
3167 : : {
3168 : 133995 : fcinfo->args[0].value = key1;
3169 : 133995 : fcinfo->args[1].value = key0;
3170 : : }
3171 : 179954 : fcinfo->isnull = false;
3172 : 179954 : fresult = FunctionCallInvoke(fcinfo);
3173 [ + - + - ]: 179954 : return (!fcinfo->isnull && DatumGetBool(fresult));
3174 : : }
3175 : : }
3176 : :
3177 : : /*
3178 : : * neqjoinsel - Join selectivity of "!="
3179 : : */
3180 : : Datum
9465 tgl@sss.pgh.pa.us 3181 :CBC 2952 : neqjoinsel(PG_FUNCTION_ARGS)
3182 : : {
7639 3183 : 2952 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
9109 3184 : 2952 : Oid operator = PG_GETARG_OID(1);
3185 : 2952 : List *args = (List *) PG_GETARG_POINTER(2);
8498 3186 : 2952 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
6471 3187 : 2952 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2114 3188 : 2952 : Oid collation = PG_GET_COLLATION();
3189 : : float8 result;
3190 : :
3079 3191 [ + + - + ]: 2952 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
9109 3192 : 1019 : {
3193 : : /*
3194 : : * For semi-joins, if there is more than one distinct value in the RHS
3195 : : * relation then every non-null LHS row must find a row to join since
3196 : : * it can only be equal to one of them. We'll assume that there is
3197 : : * always more than one distinct RHS value for the sake of stability,
3198 : : * though in theory we could have special cases for empty RHS
3199 : : * (selectivity = 0) and single-distinct-value RHS (selectivity =
3200 : : * fraction of LHS that has the same value as the single RHS value).
3201 : : *
3202 : : * For anti-joins, if we use the same assumption that there is more
3203 : : * than one distinct key in the RHS relation, then every non-null LHS
3204 : : * row must be suppressed by the anti-join.
3205 : : *
3206 : : * So either way, the selectivity estimate should be 1 - nullfrac.
3207 : : */
3208 : : VariableStatData leftvar;
3209 : : VariableStatData rightvar;
3210 : : bool reversed;
3211 : : HeapTuple statsTuple;
3212 : : double nullfrac;
3213 : :
3079 3214 : 1019 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
3215 [ + + ]: 1019 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
3216 [ + + ]: 1019 : if (HeapTupleIsValid(statsTuple))
3217 : 842 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
3218 : : else
3219 : 177 : nullfrac = 0.0;
3220 [ + + ]: 1019 : ReleaseVariableStats(leftvar);
3221 [ + + ]: 1019 : ReleaseVariableStats(rightvar);
3222 : :
3223 : 1019 : result = 1.0 - nullfrac;
3224 : : }
3225 : : else
3226 : : {
3227 : : /*
3228 : : * We want 1 - eqjoinsel() where the equality operator is the one
3229 : : * associated with this != operator, that is, its negator.
3230 : : */
3231 : 1933 : Oid eqop = get_negator(operator);
3232 : :
3233 [ + - ]: 1933 : if (eqop)
3234 : : {
3235 : : result =
2114 3236 : 1933 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
3237 : : collation,
3238 : : PointerGetDatum(root),
3239 : : ObjectIdGetDatum(eqop),
3240 : : PointerGetDatum(args),
3241 : : Int16GetDatum(jointype),
3242 : : PointerGetDatum(sjinfo)));
3243 : : }
3244 : : else
3245 : : {
3246 : : /* Use default selectivity (should we raise an error instead?) */
3079 tgl@sss.pgh.pa.us 3247 :UBC 0 : result = DEFAULT_EQ_SEL;
3248 : : }
3079 tgl@sss.pgh.pa.us 3249 :CBC 1933 : result = 1.0 - result;
3250 : : }
3251 : :
9465 3252 : 2952 : PG_RETURN_FLOAT8(result);
3253 : : }
3254 : :
3255 : : /*
3256 : : * scalarltjoinsel - Join selectivity of "<" for scalars
3257 : : */
3258 : : Datum
3259 : 270 : scalarltjoinsel(PG_FUNCTION_ARGS)
3260 : : {
3261 : 270 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3262 : : }
3263 : :
3264 : : /*
3265 : : * scalarlejoinsel - Join selectivity of "<=" for scalars
3266 : : */
3267 : : Datum
3156 3268 : 198 : scalarlejoinsel(PG_FUNCTION_ARGS)
3269 : : {
3270 : 198 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3271 : : }
3272 : :
3273 : : /*
3274 : : * scalargtjoinsel - Join selectivity of ">" for scalars
3275 : : */
3276 : : Datum
9465 3277 : 240 : scalargtjoinsel(PG_FUNCTION_ARGS)
3278 : : {
3279 : 240 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3280 : : }
3281 : :
3282 : : /*
3283 : : * scalargejoinsel - Join selectivity of ">=" for scalars
3284 : : */
3285 : : Datum
3156 3286 : 152 : scalargejoinsel(PG_FUNCTION_ARGS)
3287 : : {
3288 : 152 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3289 : : }
3290 : :
3291 : :
3292 : : /*
3293 : : * mergejoinscansel - Scan selectivity of merge join.
3294 : : *
3295 : : * A merge join will stop as soon as it exhausts either input stream.
3296 : : * Therefore, if we can estimate the ranges of both input variables,
3297 : : * we can estimate how much of the input will actually be read. This
3298 : : * can have a considerable impact on the cost when using indexscans.
3299 : : *
3300 : : * Also, we can estimate how much of each input has to be read before the
3301 : : * first join pair is found, which will affect the join's startup time.
3302 : : *
3303 : : * clause should be a clause already known to be mergejoinable. opfamily,
3304 : : * cmptype, and nulls_first specify the sort ordering being used.
3305 : : *
3306 : : * The outputs are:
3307 : : * *leftstart is set to the fraction of the left-hand variable expected
3308 : : * to be scanned before the first join pair is found (0 to 1).
3309 : : * *leftend is set to the fraction of the left-hand variable expected
3310 : : * to be scanned before the join terminates (0 to 1).
3311 : : * *rightstart, *rightend similarly for the right-hand variable.
3312 : : */
3313 : : void
2637 3314 : 102901 : mergejoinscansel(PlannerInfo *root, Node *clause,
3315 : : Oid opfamily, CompareType cmptype, bool nulls_first,
3316 : : Selectivity *leftstart, Selectivity *leftend,
3317 : : Selectivity *rightstart, Selectivity *rightend)
3318 : : {
3319 : : Node *left,
3320 : : *right;
3321 : : VariableStatData leftvar,
3322 : : rightvar;
3323 : : Oid opmethod;
3324 : : int op_strategy;
3325 : : Oid op_lefttype;
3326 : : Oid op_righttype;
3327 : : Oid opno,
3328 : : collation,
3329 : : lsortop,
3330 : : rsortop,
3331 : : lstatop,
3332 : : rstatop,
3333 : : ltop,
3334 : : leop,
3335 : : revltop,
3336 : : revleop;
3337 : : StrategyNumber ltstrat,
3338 : : lestrat,
3339 : : gtstrat,
3340 : : gestrat;
3341 : : bool isgt;
3342 : : Datum leftmin,
3343 : : leftmax,
3344 : : rightmin,
3345 : : rightmax;
3346 : : double selec;
3347 : :
3348 : : /* Set default results if we can't figure anything out. */
3349 : : /* XXX should default "start" fraction be a bit more than 0? */
6723 3350 : 102901 : *leftstart = *rightstart = 0.0;
3351 : 102901 : *leftend = *rightend = 1.0;
3352 : :
3353 : : /* Deconstruct the merge clause */
8831 3354 [ - + ]: 102901 : if (!is_opclause(clause))
8831 tgl@sss.pgh.pa.us 3355 :UBC 0 : return; /* shouldn't happen */
8545 tgl@sss.pgh.pa.us 3356 :CBC 102901 : opno = ((OpExpr *) clause)->opno;
2160 3357 : 102901 : collation = ((OpExpr *) clause)->inputcollid;
8113 3358 : 102901 : left = get_leftop((Expr *) clause);
3359 : 102901 : right = get_rightop((Expr *) clause);
8831 3360 [ - + ]: 102901 : if (!right)
8831 tgl@sss.pgh.pa.us 3361 :UBC 0 : return; /* shouldn't happen */
3362 : :
3363 : : /* Look for stats for the inputs */
8113 tgl@sss.pgh.pa.us 3364 :CBC 102901 : examine_variable(root, left, 0, &leftvar);
3365 : 102901 : examine_variable(root, right, 0, &rightvar);
3366 : :
394 peter@eisentraut.org 3367 : 102901 : opmethod = get_opfamily_method(opfamily);
3368 : :
3369 : : /* Extract the operator's declared left/right datatypes */
5633 tgl@sss.pgh.pa.us 3370 : 102901 : get_op_opfamily_properties(opno, opfamily, false,
3371 : : &op_strategy,
3372 : : &op_lefttype,
3373 : : &op_righttype);
394 peter@eisentraut.org 3374 [ - + ]: 102901 : Assert(IndexAmTranslateStrategy(op_strategy, opmethod, opfamily, true) == COMPARE_EQ);
3375 : :
3376 : : /*
3377 : : * Look up the various operators we need. If we don't find them all, it
3378 : : * probably means the opfamily is broken, but we just fail silently.
3379 : : *
3380 : : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3381 : : * operator, regardless of which sort direction we are considering.
3382 : : */
3383 [ + + - ]: 102901 : switch (cmptype)
3384 : : {
3385 : 102872 : case COMPARE_LT:
6723 tgl@sss.pgh.pa.us 3386 : 102872 : isgt = false;
394 peter@eisentraut.org 3387 : 102872 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3388 : 102872 : lestrat = IndexAmTranslateCompareType(COMPARE_LE, opmethod, opfamily, true);
6723 tgl@sss.pgh.pa.us 3389 [ + + ]: 102872 : if (op_lefttype == op_righttype)
3390 : : {
3391 : : /* easy case */
3392 : 101697 : ltop = get_opfamily_member(opfamily,
3393 : : op_lefttype, op_righttype,
3394 : : ltstrat);
3395 : 101697 : leop = get_opfamily_member(opfamily,
3396 : : op_lefttype, op_righttype,
3397 : : lestrat);
3398 : 101697 : lsortop = ltop;
3399 : 101697 : rsortop = ltop;
3400 : 101697 : lstatop = lsortop;
3401 : 101697 : rstatop = rsortop;
3402 : 101697 : revltop = ltop;
3403 : 101697 : revleop = leop;
3404 : : }
3405 : : else
3406 : : {
3407 : 1175 : ltop = get_opfamily_member(opfamily,
3408 : : op_lefttype, op_righttype,
3409 : : ltstrat);
3410 : 1175 : leop = get_opfamily_member(opfamily,
3411 : : op_lefttype, op_righttype,
3412 : : lestrat);
3413 : 1175 : lsortop = get_opfamily_member(opfamily,
3414 : : op_lefttype, op_lefttype,
3415 : : ltstrat);
3416 : 1175 : rsortop = get_opfamily_member(opfamily,
3417 : : op_righttype, op_righttype,
3418 : : ltstrat);
3419 : 1175 : lstatop = lsortop;
3420 : 1175 : rstatop = rsortop;
3421 : 1175 : revltop = get_opfamily_member(opfamily,
3422 : : op_righttype, op_lefttype,
3423 : : ltstrat);
3424 : 1175 : revleop = get_opfamily_member(opfamily,
3425 : : op_righttype, op_lefttype,
3426 : : lestrat);
3427 : : }
7073 3428 : 102872 : break;
394 peter@eisentraut.org 3429 : 29 : case COMPARE_GT:
3430 : : /* descending-order case */
6723 tgl@sss.pgh.pa.us 3431 : 29 : isgt = true;
394 peter@eisentraut.org 3432 : 29 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3433 : 29 : gtstrat = IndexAmTranslateCompareType(COMPARE_GT, opmethod, opfamily, true);
3434 : 29 : gestrat = IndexAmTranslateCompareType(COMPARE_GE, opmethod, opfamily, true);
6723 tgl@sss.pgh.pa.us 3435 [ + - ]: 29 : if (op_lefttype == op_righttype)
3436 : : {
3437 : : /* easy case */
3438 : 29 : ltop = get_opfamily_member(opfamily,
3439 : : op_lefttype, op_righttype,
3440 : : gtstrat);
3441 : 29 : leop = get_opfamily_member(opfamily,
3442 : : op_lefttype, op_righttype,
3443 : : gestrat);
3444 : 29 : lsortop = ltop;
3445 : 29 : rsortop = ltop;
3446 : 29 : lstatop = get_opfamily_member(opfamily,
3447 : : op_lefttype, op_lefttype,
3448 : : ltstrat);
3449 : 29 : rstatop = lstatop;
3450 : 29 : revltop = ltop;
3451 : 29 : revleop = leop;
3452 : : }
3453 : : else
3454 : : {
6723 tgl@sss.pgh.pa.us 3455 :UBC 0 : ltop = get_opfamily_member(opfamily,
3456 : : op_lefttype, op_righttype,
3457 : : gtstrat);
3458 : 0 : leop = get_opfamily_member(opfamily,
3459 : : op_lefttype, op_righttype,
3460 : : gestrat);
3461 : 0 : lsortop = get_opfamily_member(opfamily,
3462 : : op_lefttype, op_lefttype,
3463 : : gtstrat);
3464 : 0 : rsortop = get_opfamily_member(opfamily,
3465 : : op_righttype, op_righttype,
3466 : : gtstrat);
3467 : 0 : lstatop = get_opfamily_member(opfamily,
3468 : : op_lefttype, op_lefttype,
3469 : : ltstrat);
3470 : 0 : rstatop = get_opfamily_member(opfamily,
3471 : : op_righttype, op_righttype,
3472 : : ltstrat);
3473 : 0 : revltop = get_opfamily_member(opfamily,
3474 : : op_righttype, op_lefttype,
3475 : : gtstrat);
3476 : 0 : revleop = get_opfamily_member(opfamily,
3477 : : op_righttype, op_lefttype,
3478 : : gestrat);
3479 : : }
7073 tgl@sss.pgh.pa.us 3480 :CBC 29 : break;
7073 tgl@sss.pgh.pa.us 3481 :UBC 0 : default:
3482 : 0 : goto fail; /* shouldn't get here */
3483 : : }
3484 : :
7073 tgl@sss.pgh.pa.us 3485 [ + - + - ]:CBC 102901 : if (!OidIsValid(lsortop) ||
3486 [ + - ]: 102901 : !OidIsValid(rsortop) ||
6723 3487 [ + - ]: 102901 : !OidIsValid(lstatop) ||
3488 [ + + ]: 102901 : !OidIsValid(rstatop) ||
3489 [ + - ]: 102891 : !OidIsValid(ltop) ||
7073 3490 [ + - ]: 102891 : !OidIsValid(leop) ||
6723 3491 [ - + ]: 102891 : !OidIsValid(revltop) ||
3492 : : !OidIsValid(revleop))
7073 3493 : 10 : goto fail; /* insufficient info in catalogs */
3494 : :
3495 : : /* Try to get ranges of both inputs */
6723 3496 [ + + ]: 102891 : if (!isgt)
3497 : : {
2160 3498 [ + + ]: 102862 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3499 : : &leftmin, &leftmax))
6723 3500 : 30876 : goto fail; /* no range available from stats */
2160 3501 [ + + ]: 71986 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3502 : : &rightmin, &rightmax))
6723 3503 : 17860 : goto fail; /* no range available from stats */
3504 : : }
3505 : : else
3506 : : {
3507 : : /* need to swap the max and min */
2160 3508 [ + + ]: 29 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3509 : : &leftmax, &leftmin))
6723 3510 : 24 : goto fail; /* no range available from stats */
2160 3511 [ - + ]: 5 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3512 : : &rightmax, &rightmin))
6723 tgl@sss.pgh.pa.us 3513 :UBC 0 : goto fail; /* no range available from stats */
3514 : : }
3515 : :
3516 : : /*
3517 : : * Now, the fraction of the left variable that will be scanned is the
3518 : : * fraction that's <= the right-side maximum value. But only believe
3519 : : * non-default estimates, else stick with our 1.0.
3520 : : */
2160 tgl@sss.pgh.pa.us 3521 :CBC 54131 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3522 : : rightmax, op_righttype);
8831 3523 [ + + ]: 54131 : if (selec != DEFAULT_INEQ_SEL)
6723 3524 : 54127 : *leftend = selec;
3525 : :
3526 : : /* And similarly for the right variable. */
2160 3527 : 54131 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3528 : : leftmax, op_lefttype);
8831 3529 [ + - ]: 54131 : if (selec != DEFAULT_INEQ_SEL)
6723 3530 : 54131 : *rightend = selec;
3531 : :
3532 : : /*
3533 : : * Only one of the two "end" fractions can really be less than 1.0;
3534 : : * believe the smaller estimate and reset the other one to exactly 1.0. If
3535 : : * we get exactly equal estimates (as can easily happen with self-joins),
3536 : : * believe neither.
3537 : : */
3538 [ + + ]: 54131 : if (*leftend > *rightend)
3539 : 16567 : *leftend = 1.0;
3540 [ + + ]: 37564 : else if (*leftend < *rightend)
3541 : 20597 : *rightend = 1.0;
3542 : : else
3543 : 16967 : *leftend = *rightend = 1.0;
3544 : :
3545 : : /*
3546 : : * Also, the fraction of the left variable that will be scanned before the
3547 : : * first join pair is found is the fraction that's < the right-side
3548 : : * minimum value. But only believe non-default estimates, else stick with
3549 : : * our own default.
3550 : : */
2160 3551 : 54131 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3552 : : rightmin, op_righttype);
6723 3553 [ + - ]: 54131 : if (selec != DEFAULT_INEQ_SEL)
3554 : 54131 : *leftstart = selec;
3555 : :
3556 : : /* And similarly for the right variable. */
2160 3557 : 54131 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3558 : : leftmin, op_lefttype);
6723 3559 [ + - ]: 54131 : if (selec != DEFAULT_INEQ_SEL)
3560 : 54131 : *rightstart = selec;
3561 : :
3562 : : /*
3563 : : * Only one of the two "start" fractions can really be more than zero;
3564 : : * believe the larger estimate and reset the other one to exactly 0.0. If
3565 : : * we get exactly equal estimates (as can easily happen with self-joins),
3566 : : * believe neither.
3567 : : */
3568 [ + + ]: 54131 : if (*leftstart < *rightstart)
3569 : 11573 : *leftstart = 0.0;
3570 [ + + ]: 42558 : else if (*leftstart > *rightstart)
3571 : 17473 : *rightstart = 0.0;
3572 : : else
3573 : 25085 : *leftstart = *rightstart = 0.0;
3574 : :
3575 : : /*
3576 : : * If the sort order is nulls-first, we're going to have to skip over any
3577 : : * nulls too. These would not have been counted by scalarineqsel, and we
3578 : : * can safely add in this fraction regardless of whether we believe
3579 : : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3580 : : */
3581 [ + + ]: 54131 : if (nulls_first)
3582 : : {
3583 : : Form_pg_statistic stats;
3584 : :
3585 [ + - ]: 5 : if (HeapTupleIsValid(leftvar.statsTuple))
3586 : : {
3587 : 5 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3588 : 5 : *leftstart += stats->stanullfrac;
3589 [ - + - + ]: 5 : CLAMP_PROBABILITY(*leftstart);
3590 : 5 : *leftend += stats->stanullfrac;
3591 [ - + - + ]: 5 : CLAMP_PROBABILITY(*leftend);
3592 : : }
3593 [ + - ]: 5 : if (HeapTupleIsValid(rightvar.statsTuple))
3594 : : {
7043 3595 : 5 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
6723 3596 : 5 : *rightstart += stats->stanullfrac;
3597 [ - + - + ]: 5 : CLAMP_PROBABILITY(*rightstart);
3598 : 5 : *rightend += stats->stanullfrac;
3599 [ - + - + ]: 5 : CLAMP_PROBABILITY(*rightend);
3600 : : }
3601 : : }
3602 : :
3603 : : /* Disbelieve start >= end, just in case that can happen */
3604 [ + + ]: 54131 : if (*leftstart >= *leftend)
3605 : : {
3606 : 97 : *leftstart = 0.0;
3607 : 97 : *leftend = 1.0;
3608 : : }
3609 [ + + ]: 54131 : if (*rightstart >= *rightend)
3610 : : {
3611 : 770 : *rightstart = 0.0;
3612 : 770 : *rightend = 1.0;
3613 : : }
3614 : :
8113 3615 : 53361 : fail:
3616 [ + + ]: 102901 : ReleaseVariableStats(leftvar);
3617 [ + + ]: 102901 : ReleaseVariableStats(rightvar);
3618 : : }
3619 : :
3620 : :
3621 : : /*
3622 : : * matchingsel -- generic matching-operator selectivity support
3623 : : *
3624 : : * Use these for any operators that (a) are on data types for which we collect
3625 : : * standard statistics, and (b) have behavior for which the default estimate
3626 : : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3627 : : * operators.
3628 : : */
3629 : :
3630 : : Datum
2225 3631 : 845 : matchingsel(PG_FUNCTION_ARGS)
3632 : : {
3633 : 845 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3634 : 845 : Oid operator = PG_GETARG_OID(1);
3635 : 845 : List *args = (List *) PG_GETARG_POINTER(2);
3636 : 845 : int varRelid = PG_GETARG_INT32(3);
2160 3637 : 845 : Oid collation = PG_GET_COLLATION();
3638 : : double selec;
3639 : :
3640 : : /* Use generic restriction selectivity logic. */
3641 : 845 : selec = generic_restriction_selectivity(root, operator, collation,
3642 : : args, varRelid,
3643 : : DEFAULT_MATCHING_SEL);
3644 : :
2225 3645 : 845 : PG_RETURN_FLOAT8((float8) selec);
3646 : : }
3647 : :
3648 : : Datum
3649 : 5 : matchingjoinsel(PG_FUNCTION_ARGS)
3650 : : {
3651 : : /* Just punt, for the moment. */
3652 : 5 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3653 : : }
3654 : :
3655 : :
3656 : : /*
3657 : : * Helper routine for estimate_num_groups: add an item to a list of
3658 : : * GroupVarInfos, but only if it's not known equal to any of the existing
3659 : : * entries.
3660 : : */
3661 : : typedef struct
3662 : : {
3663 : : Node *var; /* might be an expression, not just a Var */
3664 : : RelOptInfo *rel; /* relation it belongs to */
3665 : : double ndistinct; /* # distinct values */
3666 : : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3667 : : } GroupVarInfo;
3668 : :
3669 : : static List *
7639 3670 : 281067 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3671 : : Node *var, VariableStatData *vardata)
3672 : : {
3673 : : GroupVarInfo *varinfo;
3674 : : double ndistinct;
3675 : : bool isdefault;
3676 : : ListCell *lc;
3677 : :
5357 3678 : 281067 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3679 : :
3680 : : /*
3681 : : * The nullingrels bits within the var could cause the same var to be
3682 : : * counted multiple times if it's marked with different nullingrels. They
3683 : : * could also prevent us from matching the var to the expressions in
3684 : : * extended statistics (see estimate_multivariate_ndistinct). So strip
3685 : : * them out first.
3686 : : */
488 rguo@postgresql.org 3687 : 281067 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3688 : :
2486 tgl@sss.pgh.pa.us 3689 [ + + + + : 347056 : foreach(lc, varinfos)
+ + ]
3690 : : {
7899 3691 : 66765 : varinfo = (GroupVarInfo *) lfirst(lc);
3692 : :
3693 : : /* Drop exact duplicates */
3694 [ + + ]: 66765 : if (equal(var, varinfo->var))
3695 : 776 : return varinfos;
3696 : :
3697 : : /*
3698 : : * Drop known-equal vars, but only if they belong to different
3699 : : * relations (see comments for estimate_num_groups). We aren't too
3700 : : * fussy about the semantics of "equal" here.
3701 : : */
3702 [ + + + + ]: 70307 : if (vardata->rel != varinfo->rel &&
644 rguo@postgresql.org 3703 : 4124 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3704 : : {
7899 tgl@sss.pgh.pa.us 3705 [ + + ]: 210 : if (varinfo->ndistinct <= ndistinct)
3706 : : {
3707 : : /* Keep older item, forget new one */
3708 : 194 : return varinfos;
3709 : : }
3710 : : else
3711 : : {
3712 : : /* Delete the older item */
2486 3713 : 16 : varinfos = foreach_delete_current(varinfos, lc);
3714 : : }
3715 : : }
3716 : : }
3717 : :
146 michael@paquier.xyz 3718 :GNC 280291 : varinfo = palloc_object(GroupVarInfo);
3719 : :
7899 tgl@sss.pgh.pa.us 3720 :CBC 280291 : varinfo->var = var;
3721 : 280291 : varinfo->rel = vardata->rel;
3722 : 280291 : varinfo->ndistinct = ndistinct;
1862 drowley@postgresql.o 3723 : 280291 : varinfo->isdefault = isdefault;
7899 tgl@sss.pgh.pa.us 3724 : 280291 : varinfos = lappend(varinfos, varinfo);
3725 : 280291 : return varinfos;
3726 : : }
3727 : :
3728 : : /*
3729 : : * estimate_num_groups - Estimate number of groups in a grouped query
3730 : : *
3731 : : * Given a query having a GROUP BY clause, estimate how many groups there
3732 : : * will be --- ie, the number of distinct combinations of the GROUP BY
3733 : : * expressions.
3734 : : *
3735 : : * This routine is also used to estimate the number of rows emitted by
3736 : : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3737 : : * actually, we only use it for DISTINCT when there's no grouping or
3738 : : * aggregation ahead of the DISTINCT.)
3739 : : *
3740 : : * Inputs:
3741 : : * root - the query
3742 : : * groupExprs - list of expressions being grouped by
3743 : : * input_rows - number of rows estimated to arrive at the group/unique
3744 : : * filter step
3745 : : * pgset - NULL, or a List** pointing to a grouping set to filter the
3746 : : * groupExprs against
3747 : : *
3748 : : * Outputs:
3749 : : * estinfo - When passed as non-NULL, the function will set bits in the
3750 : : * "flags" field in order to provide callers with additional information
3751 : : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3752 : : * bit if we used any default values in the estimation.
3753 : : *
3754 : : * Given the lack of any cross-correlation statistics in the system, it's
3755 : : * impossible to do anything really trustworthy with GROUP BY conditions
3756 : : * involving multiple Vars. We should however avoid assuming the worst
3757 : : * case (all possible cross-product terms actually appear as groups) since
3758 : : * very often the grouped-by Vars are highly correlated. Our current approach
3759 : : * is as follows:
3760 : : * 1. Expressions yielding boolean are assumed to contribute two groups,
3761 : : * independently of their content, and are ignored in the subsequent
3762 : : * steps. This is mainly because tests like "col IS NULL" break the
3763 : : * heuristic used in step 2 especially badly.
3764 : : * 2. Reduce the given expressions to a list of unique Vars used. For
3765 : : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3766 : : * It is clearly correct not to count the same Var more than once.
3767 : : * It is also reasonable to treat f(x) the same as x: f() cannot
3768 : : * increase the number of distinct values (unless it is volatile,
3769 : : * which we consider unlikely for grouping), but it probably won't
3770 : : * reduce the number of distinct values much either.
3771 : : * As a special case, if a GROUP BY expression can be matched to an
3772 : : * expressional index for which we have statistics, then we treat the
3773 : : * whole expression as though it were just a Var.
3774 : : * 3. If the list contains Vars of different relations that are known equal
3775 : : * due to equivalence classes, then drop all but one of the Vars from each
3776 : : * known-equal set, keeping the one with smallest estimated # of values
3777 : : * (since the extra values of the others can't appear in joined rows).
3778 : : * Note the reason we only consider Vars of different relations is that
3779 : : * if we considered ones of the same rel, we'd be double-counting the
3780 : : * restriction selectivity of the equality in the next step.
3781 : : * 4. For Vars within a single source rel, we multiply together the numbers
3782 : : * of values, clamp to the number of rows in the rel (divided by 10 if
3783 : : * more than one Var), and then multiply by a factor based on the
3784 : : * selectivity of the restriction clauses for that rel. When there's
3785 : : * more than one Var, the initial product is probably too high (it's the
3786 : : * worst case) but clamping to a fraction of the rel's rows seems to be a
3787 : : * helpful heuristic for not letting the estimate get out of hand. (The
3788 : : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3789 : : * we multiply by to adjust for the restriction selectivity assumes that
3790 : : * the restriction clauses are independent of the grouping, which may not
3791 : : * be a valid assumption, but it's hard to do better.
3792 : : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3793 : : * rel, and multiply the results together.
3794 : : * Note that rels not containing grouped Vars are ignored completely, as are
3795 : : * join clauses. Such rels cannot increase the number of groups, and we
3796 : : * assume such clauses do not reduce the number either (somewhat bogus,
3797 : : * but we don't have the info to do better).
3798 : : */
3799 : : double
4007 andres@anarazel.de 3800 : 243848 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3801 : : List **pgset, EstimationInfo *estinfo)
3802 : : {
1310 tgl@sss.pgh.pa.us 3803 : 243848 : List *varinfos = NIL;
3083 3804 : 243848 : double srf_multiplier = 1.0;
3805 : : double numdistinct;
3806 : : ListCell *l;
3807 : : int i;
3808 : :
3809 : : /* Zero the estinfo output parameter, if non-NULL */
1862 drowley@postgresql.o 3810 [ + + ]: 243848 : if (estinfo != NULL)
3811 : 203204 : memset(estinfo, 0, sizeof(EstimationInfo));
3812 : :
3813 : : /*
3814 : : * We don't ever want to return an estimate of zero groups, as that tends
3815 : : * to lead to division-by-zero and other unpleasantness. The input_rows
3816 : : * estimate is usually already at least 1, but clamp it just in case it
3817 : : * isn't.
3818 : : */
4743 tgl@sss.pgh.pa.us 3819 : 243848 : input_rows = clamp_row_est(input_rows);
3820 : :
3821 : : /*
3822 : : * If no grouping columns, there's exactly one group. (This can't happen
3823 : : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3824 : : * corner cases with set operations.)
3825 : : */
1357 3826 [ + + + + : 243848 : if (groupExprs == NIL || (pgset && *pgset == NIL))
+ + ]
5139 3827 : 923 : return 1.0;
3828 : :
3829 : : /*
3830 : : * Count groups derived from boolean grouping expressions. For other
3831 : : * expressions, find the unique Vars used, treating an expression as a Var
3832 : : * if we can find stats for it. For each one, record the statistical
3833 : : * estimate of number of distinct values (total in its table, without
3834 : : * regard for filtering).
3835 : : */
6511 3836 : 242925 : numdistinct = 1.0;
3837 : :
1310 3838 : 242925 : i = 0;
8506 3839 [ + - + + : 523375 : foreach(l, groupExprs)
+ + ]
3840 : : {
3841 : 280490 : Node *groupexpr = (Node *) lfirst(l);
3842 : : double this_srf_multiplier;
3843 : : VariableStatData vardata;
3844 : : List *varshere;
3845 : : ListCell *l2;
3846 : :
3847 : : /* is expression in this grouping set? */
4007 andres@anarazel.de 3848 [ + + + + ]: 280490 : if (pgset && !list_member_int(*pgset, i++))
3849 : 222919 : continue;
3850 : :
3851 : : /*
3852 : : * Set-returning functions in grouping columns are a bit problematic.
3853 : : * The code below will effectively ignore their SRF nature and come up
3854 : : * with a numdistinct estimate as though they were scalar functions.
3855 : : * We compensate by scaling up the end result by the largest SRF
3856 : : * rowcount estimate. (This will be an overestimate if the SRF
3857 : : * produces multiple copies of any output value, but it seems best to
3858 : : * assume the SRF's outputs are distinct. In any case, it's probably
3859 : : * pointless to worry too much about this without much better
3860 : : * estimates for SRF output rowcounts than we have today.)
3861 : : */
2642 tgl@sss.pgh.pa.us 3862 : 279822 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3083 3863 [ + + ]: 279822 : if (srf_multiplier < this_srf_multiplier)
3864 : 130 : srf_multiplier = this_srf_multiplier;
3865 : :
3866 : : /* Short-circuit for expressions returning boolean */
6511 3867 [ + + ]: 279822 : if (exprType(groupexpr) == BOOLOID)
3868 : : {
3869 : 292 : numdistinct *= 2.0;
3870 : 292 : continue;
3871 : : }
3872 : :
3873 : : /*
3874 : : * If examine_variable is able to deduce anything about the GROUP BY
3875 : : * expression, treat it as a single variable even if it's really more
3876 : : * complicated.
3877 : : *
3878 : : * XXX This has the consequence that if there's a statistics object on
3879 : : * the expression, we don't split it into individual Vars. This
3880 : : * affects our selection of statistics in
3881 : : * estimate_multivariate_ndistinct, because it's probably better to
3882 : : * use more accurate estimate for each expression and treat them as
3883 : : * independent, than to combine estimates for the extracted variables
3884 : : * when we don't know how that relates to the expressions.
3885 : : */
7899 3886 : 279530 : examine_variable(root, groupexpr, 0, &vardata);
6428 3887 [ + + + + ]: 279530 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3888 : : {
7899 3889 : 221461 : varinfos = add_unique_group_var(root, varinfos,
3890 : : groupexpr, &vardata);
3891 [ + + ]: 221461 : ReleaseVariableStats(vardata);
3892 : 221461 : continue;
3893 : : }
3894 [ - + ]: 58069 : ReleaseVariableStats(vardata);
3895 : :
3896 : : /*
3897 : : * Else pull out the component Vars. Handle PlaceHolderVars by
3898 : : * recursing into their arguments (effectively assuming that the
3899 : : * PlaceHolderVar doesn't change the number of groups, which boils
3900 : : * down to ignoring the possible addition of nulls to the result set).
3901 : : */
5411 3902 : 58069 : varshere = pull_var_clause(groupexpr,
3903 : : PVC_RECURSE_AGGREGATES |
3904 : : PVC_RECURSE_WINDOWFUNCS |
3905 : : PVC_RECURSE_PLACEHOLDERS);
3906 : :
3907 : : /*
3908 : : * If we find any variable-free GROUP BY item, then either it is a
3909 : : * constant (and we can ignore it) or it contains a volatile function;
3910 : : * in the latter case we punt and assume that each input row will
3911 : : * yield a distinct group.
3912 : : */
8568 3913 [ + + ]: 58069 : if (varshere == NIL)
3914 : : {
3915 [ + + ]: 538 : if (contain_volatile_functions(groupexpr))
3916 : 40 : return input_rows;
3917 : 498 : continue;
3918 : : }
3919 : :
3920 : : /*
3921 : : * Else add variables to varinfos list
3922 : : */
7899 3923 [ + - + + : 117137 : foreach(l2, varshere)
+ + ]
3924 : : {
3925 : 59606 : Node *var = (Node *) lfirst(l2);
3926 : :
3927 : 59606 : examine_variable(root, var, 0, &vardata);
3928 : 59606 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3929 [ + + ]: 59606 : ReleaseVariableStats(vardata);
3930 : : }
3931 : : }
3932 : :
3933 : : /*
3934 : : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3935 : : * list.
3936 : : */
3937 [ + + ]: 242885 : if (varinfos == NIL)
3938 : : {
3939 : : /* Apply SRF multiplier as we would do in the long path */
3083 3940 : 385 : numdistinct *= srf_multiplier;
3941 : : /* Round off */
3942 : 385 : numdistinct = ceil(numdistinct);
3943 : : /* Guard against out-of-range answers */
6511 3944 [ + + ]: 385 : if (numdistinct > input_rows)
3945 : 71 : numdistinct = input_rows;
3083 3946 [ - + ]: 385 : if (numdistinct < 1.0)
3083 tgl@sss.pgh.pa.us 3947 :UBC 0 : numdistinct = 1.0;
6511 tgl@sss.pgh.pa.us 3948 :CBC 385 : return numdistinct;
3949 : : }
3950 : :
3951 : : /*
3952 : : * Group Vars by relation and estimate total numdistinct.
3953 : : *
3954 : : * For each iteration of the outer loop, we process the frontmost Var in
3955 : : * varinfos, plus all other Vars in the same relation. We remove these
3956 : : * Vars from the newvarinfos list for the next iteration. This is the
3957 : : * easiest way to group Vars of same rel together.
3958 : : */
3959 : : do
3960 : : {
7899 3961 : 244665 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3962 : 244665 : RelOptInfo *rel = varinfo1->rel;
3329 alvherre@alvh.no-ip. 3963 : 244665 : double reldistinct = 1;
7763 tgl@sss.pgh.pa.us 3964 : 244665 : double relmaxndistinct = reldistinct;
3326 alvherre@alvh.no-ip. 3965 : 244665 : int relvarcount = 0;
8310 bruce@momjian.us 3966 : 244665 : List *newvarinfos = NIL;
3329 alvherre@alvh.no-ip. 3967 : 244665 : List *relvarinfos = NIL;
3968 : :
3969 : : /*
3970 : : * Split the list of varinfos in two - one for the current rel, one
3971 : : * for remaining Vars on other rels.
3972 : : */
2484 tgl@sss.pgh.pa.us 3973 : 244665 : relvarinfos = lappend(relvarinfos, varinfo1);
2045 3974 [ + - + + : 283418 : for_each_from(l, varinfos, 1)
+ + ]
3975 : : {
7899 3976 : 38753 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3977 : :
3978 [ + + ]: 38753 : if (varinfo2->rel == varinfo1->rel)
3979 : : {
3980 : : /* varinfos on current rel */
2484 3981 : 35610 : relvarinfos = lappend(relvarinfos, varinfo2);
3982 : : }
3983 : : else
3984 : : {
3985 : : /* not time to process varinfo2 yet */
3986 : 3143 : newvarinfos = lappend(newvarinfos, varinfo2);
3987 : : }
3988 : : }
3989 : :
3990 : : /*
3991 : : * Get the numdistinct estimate for the Vars of this rel. We
3992 : : * iteratively search for multivariate n-distinct with maximum number
3993 : : * of vars; assuming that each var group is independent of the others,
3994 : : * we multiply them together. Any remaining relvarinfos after no more
3995 : : * multivariate matches are found are assumed independent too, so
3996 : : * their individual ndistinct estimates are multiplied also.
3997 : : *
3998 : : * While iterating, count how many separate numdistinct values we
3999 : : * apply. We apply a fudge factor below, but only if we multiplied
4000 : : * more than one such values.
4001 : : */
3329 alvherre@alvh.no-ip. 4002 [ + + ]: 489435 : while (relvarinfos)
4003 : : {
4004 : : double mvndistinct;
4005 : :
4006 [ + + ]: 244770 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
4007 : : &mvndistinct))
4008 : : {
4009 : 345 : reldistinct *= mvndistinct;
4010 [ + + ]: 345 : if (relmaxndistinct < mvndistinct)
4011 : 335 : relmaxndistinct = mvndistinct;
3326 4012 : 345 : relvarcount++;
4013 : : }
4014 : : else
4015 : : {
3275 bruce@momjian.us 4016 [ + - + + : 523970 : foreach(l, relvarinfos)
+ + ]
4017 : : {
3329 alvherre@alvh.no-ip. 4018 : 279545 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
4019 : :
4020 : 279545 : reldistinct *= varinfo2->ndistinct;
4021 [ + + ]: 279545 : if (relmaxndistinct < varinfo2->ndistinct)
4022 : 248326 : relmaxndistinct = varinfo2->ndistinct;
4023 : 279545 : relvarcount++;
4024 : :
4025 : : /*
4026 : : * When varinfo2's isdefault is set then we'd better set
4027 : : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
4028 : : */
1862 drowley@postgresql.o 4029 [ + + + + ]: 279545 : if (estinfo != NULL && varinfo2->isdefault)
4030 : 16908 : estinfo->flags |= SELFLAG_USED_DEFAULT;
4031 : : }
4032 : :
4033 : : /* we're done with this relation */
3329 alvherre@alvh.no-ip. 4034 : 244425 : relvarinfos = NIL;
4035 : : }
4036 : : }
4037 : :
4038 : : /*
4039 : : * Sanity check --- don't divide by zero if empty relation.
4040 : : */
3319 rhaas@postgresql.org 4041 [ + + - + ]: 244665 : Assert(IS_SIMPLE_REL(rel));
8237 tgl@sss.pgh.pa.us 4042 [ + + ]: 244665 : if (rel->tuples > 0)
4043 : : {
4044 : : /*
4045 : : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
4046 : : * fudge factor is because the Vars are probably correlated but we
4047 : : * don't know by how much. We should never clamp to less than the
4048 : : * largest ndistinct value for any of the Vars, though, since
4049 : : * there will surely be at least that many groups.
4050 : : */
7767 4051 : 243800 : double clamp = rel->tuples;
4052 : :
4053 [ + + ]: 243800 : if (relvarcount > 1)
4054 : : {
4055 : 31151 : clamp *= 0.1;
7763 4056 [ + + ]: 31151 : if (clamp < relmaxndistinct)
4057 : : {
4058 : 29009 : clamp = relmaxndistinct;
4059 : : /* for sanity in case some ndistinct is too large: */
4060 [ + + ]: 29009 : if (clamp > rel->tuples)
4061 : 65 : clamp = rel->tuples;
4062 : : }
4063 : : }
7767 4064 [ + + ]: 243800 : if (reldistinct > clamp)
4065 : 25438 : reldistinct = clamp;
4066 : :
4067 : : /*
4068 : : * Update the estimate based on the restriction selectivity,
4069 : : * guarding against division by zero when reldistinct is zero.
4070 : : * Also skip this if we know that we are returning all rows.
4071 : : */
3683 dean.a.rasheed@gmail 4072 [ + - + + ]: 243800 : if (reldistinct > 0 && rel->rows < rel->tuples)
4073 : : {
4074 : : /*
4075 : : * Given a table containing N rows with n distinct values in a
4076 : : * uniform distribution, if we select p rows at random then
4077 : : * the expected number of distinct values selected is
4078 : : *
4079 : : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
4080 : : *
4081 : : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
4082 : : *
4083 : : * See "Approximating block accesses in database
4084 : : * organizations", S. B. Yao, Communications of the ACM,
4085 : : * Volume 20 Issue 4, April 1977 Pages 260-261.
4086 : : *
4087 : : * Alternatively, re-arranging the terms from the factorials,
4088 : : * this may be written as
4089 : : *
4090 : : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
4091 : : *
4092 : : * This form of the formula is more efficient to compute in
4093 : : * the common case where p is larger than N/n. Additionally,
4094 : : * as pointed out by Dell'Era, if i << N for all terms in the
4095 : : * product, it can be approximated by
4096 : : *
4097 : : * n * (1 - ((N-p)/N)^(N/n))
4098 : : *
4099 : : * See "Expected distinct values when selecting from a bag
4100 : : * without replacement", Alberto Dell'Era,
4101 : : * http://www.adellera.it/investigations/distinct_balls/.
4102 : : *
4103 : : * The condition i << N is equivalent to n >> 1, so this is a
4104 : : * good approximation when the number of distinct values in
4105 : : * the table is large. It turns out that this formula also
4106 : : * works well even when n is small.
4107 : : */
4108 : 75282 : reldistinct *=
4109 : 75282 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
4110 : 75282 : rel->tuples / reldistinct));
4111 : : }
4112 : 243800 : reldistinct = clamp_row_est(reldistinct);
4113 : :
4114 : : /*
4115 : : * Update estimate of total distinct groups.
4116 : : */
8237 tgl@sss.pgh.pa.us 4117 : 243800 : numdistinct *= reldistinct;
4118 : : }
4119 : :
8568 4120 : 244665 : varinfos = newvarinfos;
4121 [ + + ]: 244665 : } while (varinfos != NIL);
4122 : :
4123 : : /* Now we can account for the effects of any SRFs */
3083 4124 : 242500 : numdistinct *= srf_multiplier;
4125 : :
4126 : : /* Round off */
8499 4127 : 242500 : numdistinct = ceil(numdistinct);
4128 : :
4129 : : /* Guard against out-of-range answers */
8568 4130 [ + + ]: 242500 : if (numdistinct > input_rows)
4131 : 52691 : numdistinct = input_rows;
4132 [ - + ]: 242500 : if (numdistinct < 1.0)
8568 tgl@sss.pgh.pa.us 4133 :UBC 0 : numdistinct = 1.0;
4134 : :
8568 tgl@sss.pgh.pa.us 4135 :CBC 242500 : return numdistinct;
4136 : : }
4137 : :
4138 : : /*
4139 : : * Try to estimate the bucket size of the hash join inner side when the join
4140 : : * condition contains two or more clauses by employing extended statistics.
4141 : : *
4142 : : * The main idea of this approach is that the distinct value generated by
4143 : : * multivariate estimation on two or more columns would provide less bucket size
4144 : : * than estimation on one separate column.
4145 : : *
4146 : : * IMPORTANT: It is crucial to synchronize the approach of combining different
4147 : : * estimations with the caller's method.
4148 : : *
4149 : : * Return a list of clauses that didn't fetch any extended statistics.
4150 : : */
4151 : : List *
421 akorotkov@postgresql 4152 : 348681 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
4153 : : List *hashclauses,
4154 : : Selectivity *innerbucketsize)
4155 : : {
4156 : : List *clauses;
4157 : : List *otherclauses;
4158 : : double ndistinct;
4159 : :
4160 [ + + ]: 348681 : if (list_length(hashclauses) <= 1)
4161 : : {
4162 : : /*
4163 : : * Nothing to do for a single clause. Could we employ univariate
4164 : : * extended stat here?
4165 : : */
4166 : 318297 : return hashclauses;
4167 : : }
4168 : :
4169 : : /* "clauses" is the list of hashclauses we've not dealt with yet */
290 tgl@sss.pgh.pa.us 4170 :GNC 30384 : clauses = list_copy(hashclauses);
4171 : : /* "otherclauses" holds clauses we are going to return to caller */
4172 : 30384 : otherclauses = NIL;
4173 : : /* current estimate of ndistinct */
4174 : 30384 : ndistinct = 1.0;
421 akorotkov@postgresql 4175 [ + + ]:CBC 60778 : while (clauses != NIL)
4176 : : {
4177 : : ListCell *lc;
4178 : 30394 : int relid = -1;
4179 : 30394 : List *varinfos = NIL;
4180 : 30394 : List *origin_rinfos = NIL;
4181 : : double mvndistinct;
4182 : : List *origin_varinfos;
4183 : 30394 : int group_relid = -1;
4184 : 30394 : RelOptInfo *group_rel = NULL;
4185 : : ListCell *lc1,
4186 : : *lc2;
4187 : :
4188 : : /*
4189 : : * Find clauses, referencing the same single base relation and try to
4190 : : * estimate such a group with extended statistics. Create varinfo for
4191 : : * an approved clause, push it to otherclauses, if it can't be
4192 : : * estimated here or ignore to process at the next iteration.
4193 : : */
4194 [ + + + + : 93209 : foreach(lc, clauses)
+ + ]
4195 : : {
4196 : 62815 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
4197 : : Node *expr;
4198 : : Relids relids;
4199 : : GroupVarInfo *varinfo;
4200 : :
4201 : : /*
4202 : : * Find the inner side of the join, which we need to estimate the
4203 : : * number of buckets. Use outer_is_left because the
4204 : : * clause_sides_match_join routine has called on hash clauses.
4205 : : */
4206 : 125630 : relids = rinfo->outer_is_left ?
4207 [ + + ]: 62815 : rinfo->right_relids : rinfo->left_relids;
4208 : 125630 : expr = rinfo->outer_is_left ?
4209 [ + + ]: 62815 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
4210 : :
4211 [ + + ]: 62815 : if (bms_get_singleton_member(relids, &relid) &&
4212 [ + + ]: 61195 : root->simple_rel_array[relid]->statlist != NIL)
4213 : 40 : {
377 4214 : 50 : bool is_duplicate = false;
4215 : :
4216 : : /*
4217 : : * This inner-side expression references only one relation.
4218 : : * Extended statistics on this clause can exist.
4219 : : */
421 4220 [ + + ]: 50 : if (group_relid < 0)
4221 : : {
4222 : 25 : RangeTblEntry *rte = root->simple_rte_array[relid];
4223 : :
4224 [ + - - + ]: 25 : if (!rte || (rte->relkind != RELKIND_RELATION &&
421 akorotkov@postgresql 4225 [ # # ]:UBC 0 : rte->relkind != RELKIND_MATVIEW &&
4226 [ # # ]: 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
4227 [ # # ]: 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
4228 : : {
4229 : : /* Extended statistics can't exist in principle */
4230 : 0 : otherclauses = lappend(otherclauses, rinfo);
4231 : 0 : clauses = foreach_delete_current(clauses, lc);
4232 : 0 : continue;
4233 : : }
4234 : :
421 akorotkov@postgresql 4235 :CBC 25 : group_relid = relid;
4236 : 25 : group_rel = root->simple_rel_array[relid];
4237 : : }
4238 [ - + ]: 25 : else if (group_relid != relid)
4239 : : {
4240 : : /*
4241 : : * Being in the group forming state we don't need other
4242 : : * clauses.
4243 : : */
421 akorotkov@postgresql 4244 :UBC 0 : continue;
4245 : : }
4246 : :
4247 : : /*
4248 : : * We're going to add the new clause to the varinfos list. We
4249 : : * might re-use add_unique_group_var(), but we don't do so for
4250 : : * two reasons.
4251 : : *
4252 : : * 1) We must keep the origin_rinfos list ordered exactly the
4253 : : * same way as varinfos.
4254 : : *
4255 : : * 2) add_unique_group_var() is designed for
4256 : : * estimate_num_groups(), where a larger number of groups is
4257 : : * worse. While estimating the number of hash buckets, we
4258 : : * have the opposite: a lesser number of groups is worse.
4259 : : * Therefore, we don't have to remove "known equal" vars: the
4260 : : * removed var may valuably contribute to the multivariate
4261 : : * statistics to grow the number of groups.
4262 : : */
4263 : :
4264 : : /*
4265 : : * Clear nullingrels to correctly match hash keys. See
4266 : : * add_unique_group_var()'s comment for details.
4267 : : */
377 akorotkov@postgresql 4268 :CBC 50 : expr = remove_nulling_relids(expr, root->outer_join_rels, NULL);
4269 : :
4270 : : /*
4271 : : * Detect and exclude exact duplicates from the list of hash
4272 : : * keys (like add_unique_group_var does).
4273 : : */
4274 [ + + + + : 70 : foreach(lc1, varinfos)
+ + ]
4275 : : {
4276 : 30 : varinfo = (GroupVarInfo *) lfirst(lc1);
4277 : :
4278 [ + + ]: 30 : if (!equal(expr, varinfo->var))
4279 : 20 : continue;
4280 : :
4281 : 10 : is_duplicate = true;
4282 : 10 : break;
4283 : : }
4284 : :
4285 [ + + ]: 50 : if (is_duplicate)
4286 : : {
4287 : : /*
4288 : : * Skip exact duplicates. Adding them to the otherclauses
4289 : : * list also doesn't make sense.
4290 : : */
4291 : 10 : continue;
4292 : : }
4293 : :
4294 : : /*
4295 : : * Initialize GroupVarInfo. We only use it to call
4296 : : * estimate_multivariate_ndistinct(), which doesn't care about
4297 : : * ndistinct and isdefault fields. Thus, skip these fields.
4298 : : */
146 michael@paquier.xyz 4299 :GNC 40 : varinfo = palloc0_object(GroupVarInfo);
421 akorotkov@postgresql 4300 :CBC 40 : varinfo->var = expr;
4301 : 40 : varinfo->rel = root->simple_rel_array[relid];
4302 : 40 : varinfos = lappend(varinfos, varinfo);
4303 : :
4304 : : /*
4305 : : * Remember the link to RestrictInfo for the case the clause
4306 : : * is failed to be estimated.
4307 : : */
4308 : 40 : origin_rinfos = lappend(origin_rinfos, rinfo);
4309 : : }
4310 : : else
4311 : : {
4312 : : /* This clause can't be estimated with extended statistics */
4313 : 62765 : otherclauses = lappend(otherclauses, rinfo);
4314 : : }
4315 : :
4316 : 62805 : clauses = foreach_delete_current(clauses, lc);
4317 : : }
4318 : :
4319 [ + + ]: 30394 : if (list_length(varinfos) < 2)
4320 : : {
4321 : : /*
4322 : : * Multivariate statistics doesn't apply to single columns except
4323 : : * for expressions, but it has not been implemented yet.
4324 : : */
4325 : 30384 : otherclauses = list_concat(otherclauses, origin_rinfos);
4326 : 30384 : list_free_deep(varinfos);
4327 : 30384 : list_free(origin_rinfos);
4328 : 30384 : continue;
4329 : : }
4330 : :
4331 [ - + ]: 10 : Assert(group_rel != NULL);
4332 : :
4333 : : /* Employ the extended statistics. */
4334 : 10 : origin_varinfos = varinfos;
4335 : : for (;;)
4336 : 10 : {
4337 : 20 : bool estimated = estimate_multivariate_ndistinct(root,
4338 : : group_rel,
4339 : : &varinfos,
4340 : : &mvndistinct);
4341 : :
4342 [ + + ]: 20 : if (!estimated)
4343 : 10 : break;
4344 : :
4345 : : /*
4346 : : * We've got an estimation. Use ndistinct value in a consistent
4347 : : * way - according to the caller's logic (see
4348 : : * final_cost_hashjoin).
4349 : : */
4350 [ + - ]: 10 : if (ndistinct < mvndistinct)
4351 : 10 : ndistinct = mvndistinct;
4352 [ - + ]: 10 : Assert(ndistinct >= 1.0);
4353 : : }
4354 : :
4355 [ - + ]: 10 : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
4356 : :
4357 : : /* Collect unmatched clauses as otherclauses. */
4358 [ + - + + : 35 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
+ - + + +
+ + - +
+ ]
4359 : : {
4360 : 25 : GroupVarInfo *vinfo = lfirst(lc1);
4361 : :
4362 [ + - ]: 25 : if (!list_member_ptr(varinfos, vinfo))
4363 : : /* Already estimated */
4364 : 25 : continue;
4365 : :
4366 : : /* Can't be estimated here - push to the returning list */
421 akorotkov@postgresql 4367 :UBC 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
4368 : : }
4369 : : }
4370 : :
421 akorotkov@postgresql 4371 :CBC 30384 : *innerbucketsize = 1.0 / ndistinct;
4372 : 30384 : return otherclauses;
4373 : : }
4374 : :
4375 : : /*
4376 : : * Estimate hash bucket statistics when the specified expression is used
4377 : : * as a hash key for the given number of buckets.
4378 : : *
4379 : : * This attempts to determine two values:
4380 : : *
4381 : : * 1. The frequency of the most common value of the expression (returns
4382 : : * zero into *mcv_freq if we can't get that). This will be frequency
4383 : : * relative to the entire underlying table.
4384 : : *
4385 : : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
4386 : : * divided by total number of tuples to be hashed.
4387 : : *
4388 : : * XXX This is really pretty bogus since we're effectively assuming that the
4389 : : * distribution of hash keys will be the same after applying restriction
4390 : : * clauses as it was in the underlying relation. However, we are not nearly
4391 : : * smart enough to figure out how the restrict clauses might change the
4392 : : * distribution, so this will have to do for now.
4393 : : *
4394 : : * We are passed the number of buckets the executor will use for the given
4395 : : * input relation. If the data were perfectly distributed, with the same
4396 : : * number of tuples going into each available bucket, then the bucketsize
4397 : : * fraction would be 1/nbuckets. But this happy state of affairs will occur
4398 : : * only if (a) there are at least nbuckets distinct data values, and (b)
4399 : : * we have a not-too-skewed data distribution. Otherwise the buckets will
4400 : : * be nonuniformly occupied. If the other relation in the join has a key
4401 : : * distribution similar to this one's, then the most-loaded buckets are
4402 : : * exactly those that will be probed most often. Therefore, the "average"
4403 : : * bucket size for costing purposes should really be taken as something close
4404 : : * to the "worst case" bucket size. We try to estimate this by adjusting the
4405 : : * fraction if there are too few distinct data values, and then clamping to
4406 : : * at least the bucket size implied by the most common value's frequency.
4407 : : *
4408 : : * If no statistics are available, use a default estimate of 0.1. This will
4409 : : * discourage use of a hash rather strongly if the inner relation is large,
4410 : : * which is what we want. We do not want to hash unless we know that the
4411 : : * inner rel is well-dispersed (or the alternatives seem much worse).
4412 : : *
4413 : : * The caller should also check that the mcv_freq is not so large that the
4414 : : * most common value would by itself require an impractically large bucket.
4415 : : * In a hash join, the executor can split buckets if they get too big, but
4416 : : * obviously that doesn't help for a bucket that contains many duplicates of
4417 : : * the same value.
4418 : : */
4419 : : void
3185 tgl@sss.pgh.pa.us 4420 : 155604 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
4421 : : Selectivity *mcv_freq,
4422 : : Selectivity *bucketsize_frac)
4423 : : {
4424 : : VariableStatData vardata;
4425 : : double estfract,
4426 : : ndistinct;
4427 : : bool isdefault;
4428 : : AttStatsSlot sslot;
4429 : :
8113 4430 : 155604 : examine_variable(root, hashkey, 0, &vardata);
4431 : :
4432 : : /* Initialize *mcv_freq to "unknown" */
3185 4433 : 155604 : *mcv_freq = 0.0;
4434 : :
4435 : : /* Look up the frequency of the most common value, if available */
4436 [ + + ]: 155604 : if (HeapTupleIsValid(vardata.statsTuple))
4437 : : {
4438 [ + + ]: 105850 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4439 : : STATISTIC_KIND_MCV, InvalidOid,
4440 : : ATTSTATSSLOT_NUMBERS))
4441 : : {
4442 : : /*
4443 : : * The first MCV stat is for the most common value.
4444 : : */
4445 [ + - ]: 63469 : if (sslot.nnumbers > 0)
4446 : 63469 : *mcv_freq = sslot.numbers[0];
4447 : 63469 : free_attstatsslot(&sslot);
4448 : : }
127 tgl@sss.pgh.pa.us 4449 [ + + ]:GNC 42381 : else if (get_attstatsslot(&sslot, vardata.statsTuple,
4450 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
4451 : : 0))
4452 : : {
4453 : : /*
4454 : : * If there are no recorded MCVs, but we do have a histogram, then
4455 : : * assume that ANALYZE determined that the column is unique.
4456 : : */
65 4457 [ + - + + ]: 40886 : if (vardata.rel && vardata.rel->tuples > 0)
4458 : 40871 : *mcv_freq = 1.0 / vardata.rel->tuples;
4459 : : }
4460 : : }
4461 : :
4462 : : /* Get number of distinct values */
5357 tgl@sss.pgh.pa.us 4463 :CBC 155604 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4464 : :
4465 : : /*
4466 : : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4467 : : * mcv_freq is known to be even higher than that, use it instead.
4468 : : */
4469 [ + + ]: 155604 : if (isdefault)
4470 : : {
3185 4471 [ + - ]: 24409 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
5357 4472 [ + + ]: 24409 : ReleaseVariableStats(vardata);
3185 4473 : 24409 : return;
4474 : : }
4475 : :
4476 : : /*
4477 : : * Adjust ndistinct to account for restriction clauses. Observe we are
4478 : : * assuming that the data distribution is affected uniformly by the
4479 : : * restriction clauses!
4480 : : *
4481 : : * XXX Possibly better way, but much more expensive: multiply by
4482 : : * selectivity of rel's restriction clauses that mention the target Var.
4483 : : */
3691 4484 [ + - + + ]: 131195 : if (vardata.rel && vardata.rel->tuples > 0)
4485 : : {
8113 4486 : 131165 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
3691 4487 : 131165 : ndistinct = clamp_row_est(ndistinct);
4488 : : }
4489 : :
4490 : : /*
4491 : : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
4492 : : * number of buckets is less than the expected number of distinct values;
4493 : : * otherwise it is 1/ndistinct.
4494 : : */
7730 4495 [ + + ]: 131195 : if (ndistinct > nbuckets)
4496 : 75 : estfract = 1.0 / nbuckets;
4497 : : else
8113 4498 : 131120 : estfract = 1.0 / ndistinct;
4499 : :
4500 : : /*
4501 : : * Clamp the bucketsize fraction to be not less than the MCV frequency,
4502 : : * since whichever bucket the MCV values end up in will have at least that
4503 : : * size. This has no effect if *mcv_freq is still zero.
4504 : : */
62 tgl@sss.pgh.pa.us 4505 [ + + ]:GNC 131195 : estfract = Max(estfract, *mcv_freq);
4506 : :
3185 tgl@sss.pgh.pa.us 4507 :CBC 131195 : *bucketsize_frac = (Selectivity) estfract;
4508 : :
4509 [ + + ]: 131195 : ReleaseVariableStats(vardata);
4510 : : }
4511 : :
4512 : : /*
4513 : : * estimate_hashagg_tablesize
4514 : : * estimate the number of bytes that a hash aggregate hashtable will
4515 : : * require based on the agg_costs, path width and number of groups.
4516 : : *
4517 : : * We return the result as "double" to forestall any possible overflow
4518 : : * problem in the multiplication by dNumGroups.
4519 : : *
4520 : : * XXX this may be over-estimating the size now that hashagg knows to omit
4521 : : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4522 : : * grouping columns not in the hashed set are counted here even though hashagg
4523 : : * won't store them. Is this a problem?
4524 : : */
4525 : : double
1988 heikki.linnakangas@i 4526 : 2424 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4527 : : const AggClauseCosts *agg_costs, double dNumGroups)
4528 : : {
4529 : : Size hashentrysize;
4530 : :
4531 : 2424 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4532 : 2424 : path->pathtarget->width,
4533 : 2424 : agg_costs->transitionSpace);
4534 : :
4535 : : /*
4536 : : * Note that this disregards the effect of fill-factor and growth policy
4537 : : * of the hash table. That's probably ok, given that the default
4538 : : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4539 : : * "double-in-size" growth policies here.
4540 : : */
2630 tgl@sss.pgh.pa.us 4541 : 2424 : return hashentrysize * dNumGroups;
4542 : : }
4543 : :
4544 : :
4545 : : /*-------------------------------------------------------------------------
4546 : : *
4547 : : * Support routines
4548 : : *
4549 : : *-------------------------------------------------------------------------
4550 : : */
4551 : :
4552 : : /*
4553 : : * Find the best matching ndistinct extended statistics for the given list of
4554 : : * GroupVarInfos.
4555 : : *
4556 : : * Callers must ensure that the given GroupVarInfos all belong to 'rel' and
4557 : : * the GroupVarInfos list does not contain any duplicate Vars or expressions.
4558 : : *
4559 : : * When statistics are found that match > 1 of the given GroupVarInfo, the
4560 : : * *ndistinct parameter is set according to the ndistinct estimate and a new
4561 : : * list is built with the matching GroupVarInfos removed, which is output via
4562 : : * the *varinfos parameter before returning true. When no matching stats are
4563 : : * found, false is returned and the *varinfos and *ndistinct parameters are
4564 : : * left untouched.
4565 : : */
4566 : : static bool
3329 alvherre@alvh.no-ip. 4567 : 244790 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4568 : : List **varinfos, double *ndistinct)
4569 : : {
4570 : : ListCell *lc;
4571 : : int nmatches_vars;
4572 : : int nmatches_exprs;
4573 : 244790 : Oid statOid = InvalidOid;
4574 : : MVNDistinct *stats;
1866 tomas.vondra@postgre 4575 : 244790 : StatisticExtInfo *matched_info = NULL;
1281 tgl@sss.pgh.pa.us 4576 [ + - ]: 244790 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4577 : :
4578 : : /* bail out immediately if the table has no extended statistics */
3329 alvherre@alvh.no-ip. 4579 [ + + ]: 244790 : if (!rel->statlist)
4580 : 244319 : return false;
4581 : :
4582 : : /* look for the ndistinct statistics object matching the most vars */
1866 tomas.vondra@postgre 4583 : 471 : nmatches_vars = 0; /* we require at least two matches */
4584 : 471 : nmatches_exprs = 0;
3329 alvherre@alvh.no-ip. 4585 [ + - + + : 1875 : foreach(lc, rel->statlist)
+ + ]
4586 : : {
4587 : : ListCell *lc2;
4588 : 1404 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
1866 tomas.vondra@postgre 4589 : 1404 : int nshared_vars = 0;
4590 : 1404 : int nshared_exprs = 0;
4591 : :
4592 : : /* skip statistics of other kinds */
3329 alvherre@alvh.no-ip. 4593 [ + + ]: 1404 : if (info->kind != STATS_EXT_NDISTINCT)
4594 : 663 : continue;
4595 : :
4596 : : /* skip statistics with mismatching stxdinherit value */
1281 tgl@sss.pgh.pa.us 4597 [ + + ]: 741 : if (info->inherit != rte->inh)
4598 : 25 : continue;
4599 : :
4600 : : /*
4601 : : * Determine how many expressions (and variables in non-matched
4602 : : * expressions) match. We'll then use these numbers to pick the
4603 : : * statistics object that best matches the clauses.
4604 : : */
1866 tomas.vondra@postgre 4605 [ + + + + : 2267 : foreach(lc2, *varinfos)
+ + ]
4606 : : {
4607 : : ListCell *lc3;
4608 : 1551 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4609 : : AttrNumber attnum;
4610 : :
4611 [ - + ]: 1551 : Assert(varinfo->rel == rel);
4612 : :
4613 : : /* simple Var, search in statistics keys directly */
4614 [ + + ]: 1551 : if (IsA(varinfo->var, Var))
4615 : : {
4616 : 1246 : attnum = ((Var *) varinfo->var)->varattno;
4617 : :
4618 : : /*
4619 : : * Ignore system attributes - we don't support statistics on
4620 : : * them, so can't match them (and it'd fail as the values are
4621 : : * negative).
4622 : : */
4623 [ + + ]: 1246 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4624 : 10 : continue;
4625 : :
4626 [ + + ]: 1236 : if (bms_is_member(attnum, info->keys))
4627 : 730 : nshared_vars++;
4628 : :
4629 : 1236 : continue;
4630 : : }
4631 : :
4632 : : /* expression - see if it's in the statistics object */
4633 [ + + + + : 550 : foreach(lc3, info->exprs)
+ + ]
4634 : : {
4635 : 440 : Node *expr = (Node *) lfirst(lc3);
4636 : :
4637 [ + + ]: 440 : if (equal(varinfo->var, expr))
4638 : : {
4639 : 195 : nshared_exprs++;
4640 : 195 : break;
4641 : : }
4642 : : }
4643 : : }
4644 : :
4645 : : /*
4646 : : * The ndistinct extended statistics contain estimates for a minimum
4647 : : * of pairs of columns which the statistics are defined on and
4648 : : * certainly not single columns. Here we skip unless we managed to
4649 : : * match to at least two columns.
4650 : : */
4651 [ + + ]: 716 : if (nshared_vars + nshared_exprs < 2)
4652 : 331 : continue;
4653 : :
4654 : : /*
4655 : : * Check if these statistics are a better match than the previous best
4656 : : * match and if so, take note of the StatisticExtInfo.
4657 : : *
4658 : : * The statslist is sorted by statOid, so the StatisticExtInfo we
4659 : : * select as the best match is deterministic even when multiple sets
4660 : : * of statistics match equally as well.
4661 : : */
4662 [ + + + - ]: 385 : if ((nshared_exprs > nmatches_exprs) ||
4663 [ + + ]: 295 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4664 : : {
3329 alvherre@alvh.no-ip. 4665 : 365 : statOid = info->statOid;
1866 tomas.vondra@postgre 4666 : 365 : nmatches_vars = nshared_vars;
4667 : 365 : nmatches_exprs = nshared_exprs;
4668 : 365 : matched_info = info;
4669 : : }
4670 : : }
4671 : :
4672 : : /* No match? */
3329 alvherre@alvh.no-ip. 4673 [ + + ]: 471 : if (statOid == InvalidOid)
4674 : 116 : return false;
4675 : :
1866 tomas.vondra@postgre 4676 [ - + ]: 355 : Assert(nmatches_vars + nmatches_exprs > 1);
4677 : :
1570 4678 : 355 : stats = statext_ndistinct_load(statOid, rte->inh);
4679 : :
4680 : : /*
4681 : : * If we have a match, search it for the specific item that matches (there
4682 : : * must be one), and construct the output values.
4683 : : */
3329 alvherre@alvh.no-ip. 4684 [ + - ]: 355 : if (stats)
4685 : : {
4686 : : int i;
3275 bruce@momjian.us 4687 : 355 : List *newlist = NIL;
3329 alvherre@alvh.no-ip. 4688 : 355 : MVNDistinctItem *item = NULL;
4689 : : ListCell *lc2;
1866 tomas.vondra@postgre 4690 : 355 : Bitmapset *matched = NULL;
4691 : : AttrNumber attnum_offset;
4692 : :
4693 : : /*
4694 : : * How much we need to offset the attnums? If there are no
4695 : : * expressions, no offset is needed. Otherwise offset enough to move
4696 : : * the lowest one (which is equal to number of expressions) to 1.
4697 : : */
4698 [ + + ]: 355 : if (matched_info->exprs)
4699 : 125 : attnum_offset = (list_length(matched_info->exprs) + 1);
4700 : : else
4701 : 230 : attnum_offset = 0;
4702 : :
4703 : : /* see what actually matched */
4704 [ + - + + : 1240 : foreach(lc2, *varinfos)
+ + ]
4705 : : {
4706 : : ListCell *lc3;
4707 : : int idx;
4708 : 885 : bool found = false;
4709 : :
4710 : 885 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4711 : :
4712 : : /*
4713 : : * Process a simple Var expression, by matching it to keys
4714 : : * directly. If there's a matching expression, we'll try matching
4715 : : * it later.
4716 : : */
4717 [ + + ]: 885 : if (IsA(varinfo->var, Var))
4718 : : {
4719 : 730 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4720 : :
4721 : : /*
4722 : : * Ignore expressions on system attributes. Can't rely on the
4723 : : * bms check for negative values.
4724 : : */
4725 [ + + ]: 730 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4726 : 5 : continue;
4727 : :
4728 : : /* Is the variable covered by the statistics object? */
4729 [ + + ]: 725 : if (!bms_is_member(attnum, matched_info->keys))
4730 : 100 : continue;
4731 : :
4732 : 625 : attnum = attnum + attnum_offset;
4733 : :
4734 : : /* ensure sufficient offset */
4735 [ - + ]: 625 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4736 : :
4737 : 625 : matched = bms_add_member(matched, attnum);
4738 : :
4739 : 625 : found = true;
4740 : : }
4741 : :
4742 : : /*
4743 : : * XXX Maybe we should allow searching the expressions even if we
4744 : : * found an attribute matching the expression? That would handle
4745 : : * trivial expressions like "(a)" but it seems fairly useless.
4746 : : */
4747 [ + + ]: 780 : if (found)
4748 : 625 : continue;
4749 : :
4750 : : /* expression - see if it's in the statistics object */
4751 : 155 : idx = 0;
4752 [ + + + + : 255 : foreach(lc3, matched_info->exprs)
+ + ]
4753 : : {
4754 : 230 : Node *expr = (Node *) lfirst(lc3);
4755 : :
4756 [ + + ]: 230 : if (equal(varinfo->var, expr))
4757 : : {
4758 : 130 : AttrNumber attnum = -(idx + 1);
4759 : :
4760 : 130 : attnum = attnum + attnum_offset;
4761 : :
4762 : : /* ensure sufficient offset */
4763 [ - + ]: 130 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4764 : :
4765 : 130 : matched = bms_add_member(matched, attnum);
4766 : :
4767 : : /* there should be just one matching expression */
4768 : 130 : break;
4769 : : }
4770 : :
4771 : 100 : idx++;
4772 : : }
4773 : : }
4774 : :
4775 : : /* Find the specific item that exactly matches the combination */
3329 alvherre@alvh.no-ip. 4776 [ + - ]: 720 : for (i = 0; i < stats->nitems; i++)
4777 : : {
4778 : : int j;
4779 : 720 : MVNDistinctItem *tmpitem = &stats->items[i];
4780 : :
1866 tomas.vondra@postgre 4781 [ + + ]: 720 : if (tmpitem->nattributes != bms_num_members(matched))
4782 : 135 : continue;
4783 : :
4784 : : /* assume it's the right item */
4785 : 585 : item = tmpitem;
4786 : :
4787 : : /* check that all item attributes/expressions fit the match */
4788 [ + + ]: 1410 : for (j = 0; j < tmpitem->nattributes; j++)
4789 : : {
4790 : 1055 : AttrNumber attnum = tmpitem->attributes[j];
4791 : :
4792 : : /*
4793 : : * Thanks to how we constructed the matched bitmap above, we
4794 : : * can just offset all attnums the same way.
4795 : : */
4796 : 1055 : attnum = attnum + attnum_offset;
4797 : :
4798 [ + + ]: 1055 : if (!bms_is_member(attnum, matched))
4799 : : {
4800 : : /* nah, it's not this item */
4801 : 230 : item = NULL;
4802 : 230 : break;
4803 : : }
4804 : : }
4805 : :
4806 : : /*
4807 : : * If the item has all the matched attributes, we know it's the
4808 : : * right one - there can't be a better one. matching more.
4809 : : */
4810 [ + + ]: 585 : if (item)
4811 : 355 : break;
4812 : : }
4813 : :
4814 : : /*
4815 : : * Make sure we found an item. There has to be one, because ndistinct
4816 : : * statistics includes all combinations of attributes.
4817 : : */
3329 alvherre@alvh.no-ip. 4818 [ - + ]: 355 : if (!item)
3329 alvherre@alvh.no-ip. 4819 [ # # ]:UBC 0 : elog(ERROR, "corrupt MVNDistinct entry");
4820 : :
4821 : : /* Form the output varinfo list, keeping only unmatched ones */
3329 alvherre@alvh.no-ip. 4822 [ + - + + :CBC 1240 : foreach(lc, *varinfos)
+ + ]
4823 : : {
4824 : 885 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4825 : : ListCell *lc3;
1866 tomas.vondra@postgre 4826 : 885 : bool found = false;
4827 : :
4828 : : /*
4829 : : * Let's look at plain variables first, because it's the most
4830 : : * common case and the check is quite cheap. We can simply get the
4831 : : * attnum and check (with an offset) matched bitmap.
4832 : : */
4833 [ + + ]: 885 : if (IsA(varinfo->var, Var))
3329 alvherre@alvh.no-ip. 4834 : 725 : {
1866 tomas.vondra@postgre 4835 : 730 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4836 : :
4837 : : /*
4838 : : * If it's a system attribute, we're done. We don't support
4839 : : * extended statistics on system attributes, so it's clearly
4840 : : * not matched. Just keep the expression and continue.
4841 : : */
4842 [ + + ]: 730 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4843 : : {
4844 : 5 : newlist = lappend(newlist, varinfo);
4845 : 5 : continue;
4846 : : }
4847 : :
4848 : : /* apply the same offset as above */
4849 : 725 : attnum += attnum_offset;
4850 : :
4851 : : /* if it's not matched, keep the varinfo */
4852 [ + + ]: 725 : if (!bms_is_member(attnum, matched))
4853 : 100 : newlist = lappend(newlist, varinfo);
4854 : :
4855 : : /* The rest of the loop deals with complex expressions. */
3329 alvherre@alvh.no-ip. 4856 : 725 : continue;
4857 : : }
4858 : :
4859 : : /*
4860 : : * Process complex expressions, not just simple Vars.
4861 : : *
4862 : : * First, we search for an exact match of an expression. If we
4863 : : * find one, we can just discard the whole GroupVarInfo, with all
4864 : : * the variables we extracted from it.
4865 : : *
4866 : : * Otherwise we inspect the individual vars, and try matching it
4867 : : * to variables in the item.
4868 : : */
1866 tomas.vondra@postgre 4869 [ + + + + : 255 : foreach(lc3, matched_info->exprs)
+ + ]
4870 : : {
4871 : 230 : Node *expr = (Node *) lfirst(lc3);
4872 : :
4873 [ + + ]: 230 : if (equal(varinfo->var, expr))
4874 : : {
4875 : 130 : found = true;
4876 : 130 : break;
4877 : : }
4878 : : }
4879 : :
4880 : : /* found exact match, skip */
4881 [ + + ]: 155 : if (found)
2362 4882 : 130 : continue;
4883 : :
1866 4884 : 25 : newlist = lappend(newlist, varinfo);
4885 : : }
4886 : :
3329 alvherre@alvh.no-ip. 4887 : 355 : *varinfos = newlist;
4888 : 355 : *ndistinct = item->ndistinct;
4889 : 355 : return true;
4890 : : }
4891 : :
3329 alvherre@alvh.no-ip. 4892 :UBC 0 : return false;
4893 : : }
4894 : :
4895 : : /*
4896 : : * convert_to_scalar
4897 : : * Convert non-NULL values of the indicated types to the comparison
4898 : : * scale needed by scalarineqsel().
4899 : : * Returns "true" if successful.
4900 : : *
4901 : : * XXX this routine is a hack: ideally we should look up the conversion
4902 : : * subroutines in pg_type.
4903 : : *
4904 : : * All numeric datatypes are simply converted to their equivalent
4905 : : * "double" values. (NUMERIC values that are outside the range of "double"
4906 : : * are clamped to +/- HUGE_VAL.)
4907 : : *
4908 : : * String datatypes are converted by convert_string_to_scalar(),
4909 : : * which is explained below. The reason why this routine deals with
4910 : : * three values at a time, not just one, is that we need it for strings.
4911 : : *
4912 : : * The bytea datatype is just enough different from strings that it has
4913 : : * to be treated separately.
4914 : : *
4915 : : * The several datatypes representing absolute times are all converted
4916 : : * to Timestamp, which is actually an int64, and then we promote that to
4917 : : * a double. Note this will give correct results even for the "special"
4918 : : * values of Timestamp, since those are chosen to compare correctly;
4919 : : * see timestamp_cmp.
4920 : : *
4921 : : * The several datatypes representing relative times (intervals) are all
4922 : : * converted to measurements expressed in seconds.
4923 : : */
4924 : : static bool
2699 tgl@sss.pgh.pa.us 4925 :CBC 60850 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4926 : : Datum lobound, Datum hibound, Oid boundstypid,
4927 : : double *scaledlobound, double *scaledhibound)
4928 : : {
2985 4929 : 60850 : bool failure = false;
4930 : :
4931 : : /*
4932 : : * Both the valuetypid and the boundstypid should exactly match the
4933 : : * declared input type(s) of the operator we are invoked for. However,
4934 : : * extensions might try to use scalarineqsel as estimator for operators
4935 : : * with input type(s) we don't handle here; in such cases, we want to
4936 : : * return false, not fail. In any case, we mustn't assume that valuetypid
4937 : : * and boundstypid are identical.
4938 : : *
4939 : : * XXX The histogram we are interpolating between points of could belong
4940 : : * to a column that's only binary-compatible with the declared type. In
4941 : : * essence we are assuming that the semantics of binary-compatible types
4942 : : * are enough alike that we can use a histogram generated with one type's
4943 : : * operators to estimate selectivity for the other's. This is outright
4944 : : * wrong in some cases --- in particular signed versus unsigned
4945 : : * interpretation could trip us up. But it's useful enough in the
4946 : : * majority of cases that we do it anyway. Should think about more
4947 : : * rigorous ways to do it.
4948 : : */
9515 4949 [ + + - - : 60850 : switch (valuetypid)
- - ]
4950 : : {
4951 : : /*
4952 : : * Built-in numeric types
4953 : : */
9116 4954 : 55965 : case BOOLOID:
4955 : : case INT2OID:
4956 : : case INT4OID:
4957 : : case INT8OID:
4958 : : case FLOAT4OID:
4959 : : case FLOAT8OID:
4960 : : case NUMERICOID:
4961 : : case OIDOID:
4962 : : case REGPROCOID:
4963 : : case REGPROCEDUREOID:
4964 : : case REGOPEROID:
4965 : : case REGOPERATOROID:
4966 : : case REGCLASSOID:
4967 : : case REGTYPEOID:
4968 : : case REGCOLLATIONOID:
4969 : : case REGCONFIGOID:
4970 : : case REGDICTIONARYOID:
4971 : : case REGROLEOID:
4972 : : case REGNAMESPACEOID:
4973 : : case REGDATABASEOID:
2985 4974 : 55965 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4975 : : &failure);
4976 : 55965 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4977 : : &failure);
4978 : 55965 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4979 : : &failure);
4980 : 55965 : return !failure;
4981 : :
4982 : : /*
4983 : : * Built-in string types
4984 : : */
9598 4985 : 4885 : case CHAROID:
4986 : : case BPCHAROID:
4987 : : case VARCHAROID:
4988 : : case TEXTOID:
4989 : : case NAMEOID:
4990 : : {
2985 4991 : 4885 : char *valstr = convert_string_datum(value, valuetypid,
4992 : : collid, &failure);
4993 : 4885 : char *lostr = convert_string_datum(lobound, boundstypid,
4994 : : collid, &failure);
4995 : 4885 : char *histr = convert_string_datum(hibound, boundstypid,
4996 : : collid, &failure);
4997 : :
4998 : : /*
4999 : : * Bail out if any of the values is not of string type. We
5000 : : * might leak converted strings for the other value(s), but
5001 : : * that's not worth troubling over.
5002 : : */
5003 [ - + ]: 4885 : if (failure)
2985 tgl@sss.pgh.pa.us 5004 :UBC 0 : return false;
5005 : :
9175 bruce@momjian.us 5006 :CBC 4885 : convert_string_to_scalar(valstr, scaledvalue,
5007 : : lostr, scaledlobound,
5008 : : histr, scaledhibound);
5009 : 4885 : pfree(valstr);
5010 : 4885 : pfree(lostr);
5011 : 4885 : pfree(histr);
5012 : 4885 : return true;
5013 : : }
5014 : :
5015 : : /*
5016 : : * Built-in bytea type
5017 : : */
9031 tgl@sss.pgh.pa.us 5018 :UBC 0 : case BYTEAOID:
5019 : : {
5020 : : /* We only support bytea vs bytea comparison */
2985 5021 [ # # ]: 0 : if (boundstypid != BYTEAOID)
5022 : 0 : return false;
9031 5023 : 0 : convert_bytea_to_scalar(value, scaledvalue,
5024 : : lobound, scaledlobound,
5025 : : hibound, scaledhibound);
5026 : 0 : return true;
5027 : : }
5028 : :
5029 : : /*
5030 : : * Built-in time types
5031 : : */
9565 5032 : 0 : case TIMESTAMPOID:
5033 : : case TIMESTAMPTZOID:
5034 : : case DATEOID:
5035 : : case INTERVALOID:
5036 : : case TIMEOID:
5037 : : case TIMETZOID:
2985 5038 : 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
5039 : : &failure);
5040 : 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
5041 : : &failure);
5042 : 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
5043 : : &failure);
5044 : 0 : return !failure;
5045 : :
5046 : : /*
5047 : : * Built-in network types
5048 : : */
9096 5049 : 0 : case INETOID:
5050 : : case CIDROID:
5051 : : case MACADDROID:
5052 : : case MACADDR8OID:
2985 5053 : 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
5054 : : &failure);
5055 : 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
5056 : : &failure);
5057 : 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
5058 : : &failure);
5059 : 0 : return !failure;
5060 : : }
5061 : : /* Don't know how to convert */
7528 5062 : 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
9773 5063 : 0 : return false;
5064 : : }
5065 : :
5066 : : /*
5067 : : * Do convert_to_scalar()'s work for any numeric data type.
5068 : : *
5069 : : * On failure (e.g., unsupported typid), set *failure to true;
5070 : : * otherwise, that variable is not changed.
5071 : : */
5072 : : static double
2985 tgl@sss.pgh.pa.us 5073 :CBC 167895 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
5074 : : {
9515 5075 [ - + + - : 167895 : switch (typid)
- + - +
- ]
5076 : : {
9129 tgl@sss.pgh.pa.us 5077 :UBC 0 : case BOOLOID:
9473 5078 : 0 : return (double) DatumGetBool(value);
9515 tgl@sss.pgh.pa.us 5079 :CBC 10 : case INT2OID:
5080 : 10 : return (double) DatumGetInt16(value);
5081 : 22121 : case INT4OID:
5082 : 22121 : return (double) DatumGetInt32(value);
9515 tgl@sss.pgh.pa.us 5083 :UBC 0 : case INT8OID:
9473 5084 : 0 : return (double) DatumGetInt64(value);
9515 5085 : 0 : case FLOAT4OID:
9473 5086 : 0 : return (double) DatumGetFloat4(value);
9515 tgl@sss.pgh.pa.us 5087 :CBC 45 : case FLOAT8OID:
9473 5088 : 45 : return (double) DatumGetFloat8(value);
9515 tgl@sss.pgh.pa.us 5089 :UBC 0 : case NUMERICOID:
5090 : : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
8970 5091 : 0 : return (double)
5092 : 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
5093 : : value));
9515 tgl@sss.pgh.pa.us 5094 :CBC 145719 : case OIDOID:
5095 : : case REGPROCOID:
5096 : : case REGPROCEDUREOID:
5097 : : case REGOPEROID:
5098 : : case REGOPERATOROID:
5099 : : case REGCLASSOID:
5100 : : case REGTYPEOID:
5101 : : case REGCOLLATIONOID:
5102 : : case REGCONFIGOID:
5103 : : case REGDICTIONARYOID:
5104 : : case REGROLEOID:
5105 : : case REGNAMESPACEOID:
5106 : : case REGDATABASEOID:
5107 : : /* we can treat OIDs as integers... */
5108 : 145719 : return (double) DatumGetObjectId(value);
5109 : : }
5110 : :
2985 tgl@sss.pgh.pa.us 5111 :UBC 0 : *failure = true;
9515 5112 : 0 : return 0;
5113 : : }
5114 : :
5115 : : /*
5116 : : * Do convert_to_scalar()'s work for any character-string data type.
5117 : : *
5118 : : * String datatypes are converted to a scale that ranges from 0 to 1,
5119 : : * where we visualize the bytes of the string as fractional digits.
5120 : : *
5121 : : * We do not want the base to be 256, however, since that tends to
5122 : : * generate inflated selectivity estimates; few databases will have
5123 : : * occurrences of all 256 possible byte values at each position.
5124 : : * Instead, use the smallest and largest byte values seen in the bounds
5125 : : * as the estimated range for each byte, after some fudging to deal with
5126 : : * the fact that we probably aren't going to see the full range that way.
5127 : : *
5128 : : * An additional refinement is that we discard any common prefix of the
5129 : : * three strings before computing the scaled values. This allows us to
5130 : : * "zoom in" when we encounter a narrow data range. An example is a phone
5131 : : * number database where all the values begin with the same area code.
5132 : : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
5133 : : * so this is more likely to happen than you might think.)
5134 : : */
5135 : : static void
7528 tgl@sss.pgh.pa.us 5136 :CBC 4885 : convert_string_to_scalar(char *value,
5137 : : double *scaledvalue,
5138 : : char *lobound,
5139 : : double *scaledlobound,
5140 : : char *hibound,
5141 : : double *scaledhibound)
5142 : : {
5143 : : int rangelo,
5144 : : rangehi;
5145 : : char *sptr;
5146 : :
5147 : 4885 : rangelo = rangehi = (unsigned char) hibound[0];
9515 5148 [ + + ]: 71302 : for (sptr = lobound; *sptr; sptr++)
5149 : : {
7528 5150 [ + + ]: 66417 : if (rangelo > (unsigned char) *sptr)
5151 : 12325 : rangelo = (unsigned char) *sptr;
5152 [ + + ]: 66417 : if (rangehi < (unsigned char) *sptr)
5153 : 6425 : rangehi = (unsigned char) *sptr;
5154 : : }
9515 5155 [ + + ]: 52064 : for (sptr = hibound; *sptr; sptr++)
5156 : : {
7528 5157 [ + + ]: 47179 : if (rangelo > (unsigned char) *sptr)
5158 : 471 : rangelo = (unsigned char) *sptr;
5159 [ + + ]: 47179 : if (rangehi < (unsigned char) *sptr)
5160 : 1926 : rangehi = (unsigned char) *sptr;
5161 : : }
5162 : : /* If range includes any upper-case ASCII chars, make it include all */
9515 5163 [ + + + + ]: 4885 : if (rangelo <= 'Z' && rangehi >= 'A')
5164 : : {
5165 [ + + ]: 1397 : if (rangelo > 'A')
5166 : 185 : rangelo = 'A';
5167 [ + + ]: 1397 : if (rangehi < 'Z')
5168 : 425 : rangehi = 'Z';
5169 : : }
5170 : : /* Ditto lower-case */
5171 [ + - + + ]: 4885 : if (rangelo <= 'z' && rangehi >= 'a')
5172 : : {
5173 [ + + ]: 4426 : if (rangelo > 'a')
5174 : 1 : rangelo = 'a';
5175 [ + + ]: 4426 : if (rangehi < 'z')
5176 : 4392 : rangehi = 'z';
5177 : : }
5178 : : /* Ditto digits */
5179 [ + + + - ]: 4885 : if (rangelo <= '9' && rangehi >= '0')
5180 : : {
5181 [ + + ]: 821 : if (rangelo > '0')
5182 : 748 : rangelo = '0';
5183 [ + + ]: 821 : if (rangehi < '9')
5184 : 24 : rangehi = '9';
5185 : : }
5186 : :
5187 : : /*
5188 : : * If range includes less than 10 chars, assume we have not got enough
5189 : : * data, and make it include regular ASCII set.
5190 : : */
5191 [ - + ]: 4885 : if (rangehi - rangelo < 9)
5192 : : {
9515 tgl@sss.pgh.pa.us 5193 :UBC 0 : rangelo = ' ';
5194 : 0 : rangehi = 127;
5195 : : }
5196 : :
5197 : : /*
5198 : : * Now strip any common prefix of the three strings.
5199 : : */
9515 tgl@sss.pgh.pa.us 5200 [ + - ]:CBC 9667 : while (*lobound)
5201 : : {
5202 [ + + + - ]: 9667 : if (*lobound != *hibound || *lobound != *value)
5203 : : break;
5204 : 4782 : lobound++, hibound++, value++;
5205 : : }
5206 : :
5207 : : /*
5208 : : * Now we can do the conversions.
5209 : : */
5210 : 4885 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
5211 : 4885 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
5212 : 4885 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
5213 : 4885 : }
5214 : :
5215 : : static double
7528 5216 : 14655 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
5217 : : {
5218 : 14655 : int slen = strlen(value);
5219 : : double num,
5220 : : denom,
5221 : : base;
5222 : :
9515 5223 [ - + ]: 14655 : if (slen <= 0)
9515 tgl@sss.pgh.pa.us 5224 :UBC 0 : return 0.0; /* empty string has scalar value 0 */
5225 : :
5226 : : /*
5227 : : * There seems little point in considering more than a dozen bytes from
5228 : : * the string. Since base is at least 10, that will give us nominal
5229 : : * resolution of at least 12 decimal digits, which is surely far more
5230 : : * precision than this estimation technique has got anyway (especially in
5231 : : * non-C locales). Also, even with the maximum possible base of 256, this
5232 : : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
5233 : : * overflow on any known machine.
5234 : : */
3908 tgl@sss.pgh.pa.us 5235 [ + + ]:CBC 14655 : if (slen > 12)
5236 : 3827 : slen = 12;
5237 : :
5238 : : /* Convert initial characters to fraction */
9515 5239 : 14655 : base = rangehi - rangelo + 1;
5240 : 14655 : num = 0.0;
5241 : 14655 : denom = base;
5242 [ + + ]: 121709 : while (slen-- > 0)
5243 : : {
7528 5244 : 107054 : int ch = (unsigned char) *value++;
5245 : :
9515 5246 [ + + ]: 107054 : if (ch < rangelo)
9175 bruce@momjian.us 5247 : 68 : ch = rangelo - 1;
9515 tgl@sss.pgh.pa.us 5248 [ - + ]: 106986 : else if (ch > rangehi)
9175 bruce@momjian.us 5249 :UBC 0 : ch = rangehi + 1;
9515 tgl@sss.pgh.pa.us 5250 :CBC 107054 : num += ((double) (ch - rangelo)) / denom;
5251 : 107054 : denom *= base;
5252 : : }
5253 : :
5254 : 14655 : return num;
5255 : : }
5256 : :
5257 : : /*
5258 : : * Convert a string-type Datum into a palloc'd, null-terminated string.
5259 : : *
5260 : : * On failure (e.g., unsupported typid), set *failure to true;
5261 : : * otherwise, that variable is not changed. (We'll return NULL on failure.)
5262 : : *
5263 : : * When using a non-C locale, we must pass the string through pg_strxfrm()
5264 : : * before continuing, so as to generate correct locale-specific results.
5265 : : */
5266 : : static char *
2699 5267 : 14655 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
5268 : : {
5269 : : char *val;
5270 : : pg_locale_t mylocale;
5271 : :
9515 5272 [ - + + - ]: 14655 : switch (typid)
5273 : : {
9515 tgl@sss.pgh.pa.us 5274 :UBC 0 : case CHAROID:
5275 : 0 : val = (char *) palloc(2);
5276 : 0 : val[0] = DatumGetChar(value);
5277 : 0 : val[1] = '\0';
5278 : 0 : break;
9515 tgl@sss.pgh.pa.us 5279 :CBC 4416 : case BPCHAROID:
5280 : : case VARCHAROID:
5281 : : case TEXTOID:
6615 5282 : 4416 : val = TextDatumGetCString(value);
5283 : 4416 : break;
9515 5284 : 10239 : case NAMEOID:
5285 : : {
9175 bruce@momjian.us 5286 : 10239 : NameData *nm = (NameData *) DatumGetPointer(value);
5287 : :
5288 : 10239 : val = pstrdup(NameStr(*nm));
5289 : 10239 : break;
5290 : : }
9515 tgl@sss.pgh.pa.us 5291 :UBC 0 : default:
2985 5292 : 0 : *failure = true;
9515 5293 : 0 : return NULL;
5294 : : }
5295 : :
608 jdavis@postgresql.or 5296 :CBC 14655 : mylocale = pg_newlocale_from_collation(collid);
5297 : :
5298 [ + + ]: 14655 : if (!mylocale->collate_is_c)
5299 : : {
5300 : : char *xfrmstr;
5301 : : size_t xfrmlen;
5302 : : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
5303 : :
5304 : : /*
5305 : : * XXX: We could guess at a suitable output buffer size and only call
5306 : : * pg_strxfrm() twice if our guess is too small.
5307 : : *
5308 : : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
5309 : : * bogus data or set an error. This is not really a problem unless it
5310 : : * crashes since it will only give an estimation error and nothing
5311 : : * fatal.
5312 : : *
5313 : : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
5314 : : * some cases, libc strxfrm() may return the wrong results, but that
5315 : : * will only lead to an estimation error.
5316 : : */
637 5317 : 66 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
5318 : : #ifdef WIN32
5319 : :
5320 : : /*
5321 : : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
5322 : : * of trying to allocate this much memory (and fail), just return the
5323 : : * original string unmodified as if we were in the C locale.
5324 : : */
5325 : : if (xfrmlen == INT_MAX)
5326 : : return val;
5327 : : #endif
8328 tgl@sss.pgh.pa.us 5328 : 66 : xfrmstr = (char *) palloc(xfrmlen + 1);
637 jdavis@postgresql.or 5329 : 66 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
5330 : :
5331 : : /*
5332 : : * Some systems (e.g., glibc) can return a smaller value from the
5333 : : * second call than the first; thus the Assert must be <= not ==.
5334 : : */
8328 tgl@sss.pgh.pa.us 5335 [ - + ]: 66 : Assert(xfrmlen2 <= xfrmlen);
8798 peter_e@gmx.net 5336 : 66 : pfree(val);
5337 : 66 : val = xfrmstr;
5338 : : }
5339 : :
7528 tgl@sss.pgh.pa.us 5340 : 14655 : return val;
5341 : : }
5342 : :
5343 : : /*
5344 : : * Do convert_to_scalar()'s work for any bytea data type.
5345 : : *
5346 : : * Very similar to convert_string_to_scalar except we can't assume
5347 : : * null-termination and therefore pass explicit lengths around.
5348 : : *
5349 : : * Also, assumptions about likely "normal" ranges of characters have been
5350 : : * removed - a data range of 0..255 is always used, for now. (Perhaps
5351 : : * someday we will add information about actual byte data range to
5352 : : * pg_statistic.)
5353 : : */
5354 : : static void
9031 tgl@sss.pgh.pa.us 5355 :UBC 0 : convert_bytea_to_scalar(Datum value,
5356 : : double *scaledvalue,
5357 : : Datum lobound,
5358 : : double *scaledlobound,
5359 : : Datum hibound,
5360 : : double *scaledhibound)
5361 : : {
2985 5362 : 0 : bytea *valuep = DatumGetByteaPP(value);
5363 : 0 : bytea *loboundp = DatumGetByteaPP(lobound);
5364 : 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
5365 : : int rangelo,
5366 : : rangehi,
5367 [ # # # # : 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
# # # # #
# ]
5368 [ # # # # : 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
# # # # #
# ]
5369 [ # # # # : 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
# # # # #
# ]
5370 : : i,
5371 : : minlen;
5372 [ # # ]: 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
5373 [ # # ]: 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
5374 [ # # ]: 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
5375 : :
5376 : : /*
5377 : : * Assume bytea data is uniformly distributed across all byte values.
5378 : : */
9031 5379 : 0 : rangelo = 0;
5380 : 0 : rangehi = 255;
5381 : :
5382 : : /*
5383 : : * Now strip any common prefix of the three strings.
5384 : : */
5385 : 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
5386 [ # # ]: 0 : for (i = 0; i < minlen; i++)
5387 : : {
5388 [ # # # # ]: 0 : if (*lostr != *histr || *lostr != *valstr)
5389 : : break;
5390 : 0 : lostr++, histr++, valstr++;
5391 : 0 : loboundlen--, hiboundlen--, valuelen--;
5392 : : }
5393 : :
5394 : : /*
5395 : : * Now we can do the conversions.
5396 : : */
5397 : 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
5398 : 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
5399 : 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
5400 : 0 : }
5401 : :
5402 : : static double
5403 : 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
5404 : : int rangelo, int rangehi)
5405 : : {
5406 : : double num,
5407 : : denom,
5408 : : base;
5409 : :
5410 [ # # ]: 0 : if (valuelen <= 0)
5411 : 0 : return 0.0; /* empty string has scalar value 0 */
5412 : :
5413 : : /*
5414 : : * Since base is 256, need not consider more than about 10 chars (even
5415 : : * this many seems like overkill)
5416 : : */
5417 [ # # ]: 0 : if (valuelen > 10)
5418 : 0 : valuelen = 10;
5419 : :
5420 : : /* Convert initial characters to fraction */
5421 : 0 : base = rangehi - rangelo + 1;
5422 : 0 : num = 0.0;
5423 : 0 : denom = base;
5424 [ # # ]: 0 : while (valuelen-- > 0)
5425 : : {
5426 : 0 : int ch = *value++;
5427 : :
5428 [ # # ]: 0 : if (ch < rangelo)
5429 : 0 : ch = rangelo - 1;
5430 [ # # ]: 0 : else if (ch > rangehi)
5431 : 0 : ch = rangehi + 1;
5432 : 0 : num += ((double) (ch - rangelo)) / denom;
5433 : 0 : denom *= base;
5434 : : }
5435 : :
5436 : 0 : return num;
5437 : : }
5438 : :
5439 : : /*
5440 : : * Do convert_to_scalar()'s work for any timevalue data type.
5441 : : *
5442 : : * On failure (e.g., unsupported typid), set *failure to true;
5443 : : * otherwise, that variable is not changed.
5444 : : */
5445 : : static double
2985 5446 : 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5447 : : {
9515 5448 [ # # # # : 0 : switch (typid)
# # # ]
5449 : : {
9116 5450 : 0 : case TIMESTAMPOID:
9461 5451 : 0 : return DatumGetTimestamp(value);
8980 5452 : 0 : case TIMESTAMPTZOID:
5453 : 0 : return DatumGetTimestampTz(value);
9515 5454 : 0 : case DATEOID:
5607 5455 : 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
9515 5456 : 0 : case INTERVALOID:
5457 : : {
9175 bruce@momjian.us 5458 : 0 : Interval *interval = DatumGetIntervalP(value);
5459 : :
5460 : : /*
5461 : : * Convert the month part of Interval to days using assumed
5462 : : * average month length of 365.25/12.0 days. Not too
5463 : : * accurate, but plenty good enough for our purposes.
5464 : : *
5465 : : * This also works for infinite intervals, which just have all
5466 : : * fields set to INT_MIN/INT_MAX, and so will produce a result
5467 : : * smaller/larger than any finite interval.
5468 : : */
7507 5469 : 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5470 : 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5471 : : }
9515 tgl@sss.pgh.pa.us 5472 : 0 : case TIMEOID:
9461 5473 : 0 : return DatumGetTimeADT(value);
9096 5474 : 0 : case TIMETZOID:
5475 : : {
5476 : 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5477 : :
5478 : : /* use GMT-equivalent time */
8780 lockhart@fourpalms.o 5479 : 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5480 : : }
5481 : : }
5482 : :
2985 tgl@sss.pgh.pa.us 5483 : 0 : *failure = true;
9515 5484 : 0 : return 0;
5485 : : }
5486 : :
5487 : :
5488 : : /*
5489 : : * get_restriction_variable
5490 : : * Examine the args of a restriction clause to see if it's of the
5491 : : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5492 : : * where "variable" could be either a Var or an expression in vars of a
5493 : : * single relation. If so, extract information about the variable,
5494 : : * and also indicate which side it was on and the other argument.
5495 : : *
5496 : : * Inputs:
5497 : : * root: the planner info
5498 : : * args: clause argument list
5499 : : * varRelid: see specs for restriction selectivity functions
5500 : : *
5501 : : * Outputs: (these are valid only if true is returned)
5502 : : * *vardata: gets information about variable (see examine_variable)
5503 : : * *other: gets other clause argument, aggressively reduced to a constant
5504 : : * *varonleft: set true if variable is on the left, false if on the right
5505 : : *
5506 : : * Returns true if a variable is identified, otherwise false.
5507 : : *
5508 : : * Note: if there are Vars on both sides of the clause, we must fail, because
5509 : : * callers are expecting that the other side will act like a pseudoconstant.
5510 : : */
5511 : : bool
7639 tgl@sss.pgh.pa.us 5512 :CBC 650958 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5513 : : VariableStatData *vardata, Node **other,
5514 : : bool *varonleft)
5515 : : {
5516 : : Node *left,
5517 : : *right;
5518 : : VariableStatData rdata;
5519 : :
5520 : : /* Fail if not a binary opclause (probably shouldn't happen) */
8010 neilc@samurai.com 5521 [ - + ]: 650958 : if (list_length(args) != 2)
8113 tgl@sss.pgh.pa.us 5522 :UBC 0 : return false;
5523 : :
8014 neilc@samurai.com 5524 :CBC 650958 : left = (Node *) linitial(args);
8113 tgl@sss.pgh.pa.us 5525 : 650958 : right = (Node *) lsecond(args);
5526 : :
5527 : : /*
5528 : : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5529 : : * relations will be treated as pseudoconstants.
5530 : : */
5531 : 650958 : examine_variable(root, left, varRelid, vardata);
5532 : 650958 : examine_variable(root, right, varRelid, &rdata);
5533 : :
5534 : : /*
5535 : : * If one side is a variable and the other not, we win.
5536 : : */
5537 [ + + + + ]: 650958 : if (vardata->rel && rdata.rel == NULL)
5538 : : {
5539 : 579501 : *varonleft = true;
7015 5540 : 579501 : *other = estimate_expression_value(root, rdata.var);
5541 : : /* Assume we need no ReleaseVariableStats(rdata) here */
8113 5542 : 579497 : return true;
5543 : : }
5544 : :
5545 [ + + + + ]: 71457 : if (vardata->rel == NULL && rdata.rel)
5546 : : {
5547 : 67131 : *varonleft = false;
7015 5548 : 67131 : *other = estimate_expression_value(root, vardata->var);
5549 : : /* Assume we need no ReleaseVariableStats(*vardata) here */
8113 5550 : 67131 : *vardata = rdata;
5551 : 67131 : return true;
5552 : : }
5553 : :
5554 : : /* Oops, clause has wrong structure (probably var op var) */
5555 [ + + ]: 4326 : ReleaseVariableStats(*vardata);
5556 [ + + ]: 4326 : ReleaseVariableStats(rdata);
5557 : :
5558 : 4326 : return false;
5559 : : }
5560 : :
5561 : : /*
5562 : : * get_join_variables
5563 : : * Apply examine_variable() to each side of a join clause.
5564 : : * Also, attempt to identify whether the join clause has the same
5565 : : * or reversed sense compared to the SpecialJoinInfo.
5566 : : *
5567 : : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5568 : : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5569 : : * where we can't tell for sure, we default to assuming it's normal.
5570 : : */
5571 : : void
6471 5572 : 219832 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5573 : : VariableStatData *vardata1, VariableStatData *vardata2,
5574 : : bool *join_is_reversed)
5575 : : {
5576 : : Node *left,
5577 : : *right;
5578 : :
8010 neilc@samurai.com 5579 [ - + ]: 219832 : if (list_length(args) != 2)
8113 tgl@sss.pgh.pa.us 5580 [ # # ]:UBC 0 : elog(ERROR, "join operator should take two arguments");
5581 : :
8014 neilc@samurai.com 5582 :CBC 219832 : left = (Node *) linitial(args);
9116 tgl@sss.pgh.pa.us 5583 : 219832 : right = (Node *) lsecond(args);
5584 : :
8113 5585 : 219832 : examine_variable(root, left, 0, vardata1);
5586 : 219832 : examine_variable(root, right, 0, vardata2);
5587 : :
6471 5588 [ + + + + ]: 439453 : if (vardata1->rel &&
5589 : 219621 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
3240 5590 : 71828 : *join_is_reversed = true; /* var1 is on RHS */
6471 5591 [ + + + + ]: 295797 : else if (vardata2->rel &&
5592 : 147793 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
3240 5593 : 304 : *join_is_reversed = true; /* var2 is on LHS */
5594 : : else
6471 5595 : 147700 : *join_is_reversed = false;
8113 5596 : 219832 : }
5597 : :
5598 : : /* statext_expressions_load copies the tuple, so just pfree it. */
5599 : : static void
1866 tomas.vondra@postgre 5600 : 1415 : ReleaseDummy(HeapTuple tuple)
5601 : : {
5602 : 1415 : pfree(tuple);
5603 : 1415 : }
5604 : :
5605 : : /*
5606 : : * examine_variable
5607 : : * Try to look up statistical data about an expression.
5608 : : * Fill in a VariableStatData struct to describe the expression.
5609 : : *
5610 : : * Inputs:
5611 : : * root: the planner info
5612 : : * node: the expression tree to examine
5613 : : * varRelid: see specs for restriction selectivity functions
5614 : : *
5615 : : * Outputs: *vardata is filled as follows:
5616 : : * var: the input expression (with any phvs or binary relabeling stripped,
5617 : : * if it is or contains a variable; but otherwise unchanged)
5618 : : * rel: RelOptInfo for relation containing variable; NULL if expression
5619 : : * contains no Vars (NOTE this could point to a RelOptInfo of a
5620 : : * subquery, not one in the current query).
5621 : : * statsTuple: the pg_statistic entry for the variable, if one exists;
5622 : : * otherwise NULL.
5623 : : * freefunc: pointer to a function to release statsTuple with.
5624 : : * vartype: exposed type of the expression; this should always match
5625 : : * the declared input type of the operator we are estimating for.
5626 : : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5627 : : * commonly the same as the exposed type of the variable argument,
5628 : : * but can be different in binary-compatible-type cases.
5629 : : * isunique: true if we were able to match the var to a unique index, a
5630 : : * single-column DISTINCT or GROUP-BY clause, implying its values are
5631 : : * unique for this query. (Caution: this should be trusted for
5632 : : * statistical purposes only, since we do not check indimmediate nor
5633 : : * verify that the exact same definition of equality applies.)
5634 : : * acl_ok: true if current user has permission to read all table rows from
5635 : : * the column(s) underlying the pg_statistic entry. This is consulted by
5636 : : * statistic_proc_security_check().
5637 : : *
5638 : : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5639 : : */
5640 : : void
7639 tgl@sss.pgh.pa.us 5641 : 2570774 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5642 : : VariableStatData *vardata)
5643 : : {
5644 : : Node *basenode;
5645 : : Relids varnos;
5646 : : Relids basevarnos;
5647 : : RelOptInfo *onerel;
5648 : :
5649 : : /* Make sure we don't return dangling pointers in vardata */
8113 5650 [ + - + - : 17995418 : MemSet(vardata, 0, sizeof(VariableStatData));
+ - + - +
+ ]
5651 : :
5652 : : /* Save the exposed type of the expression */
7704 5653 : 2570774 : vardata->vartype = exprType(node);
5654 : :
5655 : : /*
5656 : : * PlaceHolderVars are transparent for the purpose of statistics lookup;
5657 : : * they do not alter the value distribution of the underlying expression.
5658 : : * However, they can obscure the structure, preventing us from recognizing
5659 : : * matches to base columns, index expressions, or extended statistics. So
5660 : : * strip them out first.
5661 : : */
127 rguo@postgresql.org 5662 : 2570774 : basenode = strip_all_phvs_deep(root, node);
5663 : :
5664 : : /*
5665 : : * Look inside any binary-compatible relabeling. We need to handle nested
5666 : : * RelabelType nodes here, because the prior stripping of PlaceHolderVars
5667 : : * may have brought separate RelabelTypes into adjacency.
5668 : : */
5669 [ + + ]: 2611517 : while (IsA(basenode, RelabelType))
5670 : 40743 : basenode = (Node *) ((RelabelType *) basenode)->arg;
5671 : :
5672 : : /* Fast path for a simple Var */
7710 tgl@sss.pgh.pa.us 5673 [ + + + + ]: 2570774 : if (IsA(basenode, Var) &&
5674 [ + + ]: 675788 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5675 : : {
5676 : 1830049 : Var *var = (Var *) basenode;
5677 : :
5678 : : /* Set up result fields other than the stats tuple */
127 rguo@postgresql.org 5679 : 1830049 : vardata->var = basenode; /* return Var without phvs or relabeling */
8113 tgl@sss.pgh.pa.us 5680 : 1830049 : vardata->rel = find_base_rel(root, var->varno);
5681 : 1830049 : vardata->atttype = var->vartype;
5682 : 1830049 : vardata->atttypmod = var->vartypmod;
6288 5683 : 1830049 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5684 : :
5685 : : /* Try to locate some stats */
5357 5686 : 1830049 : examine_simple_variable(root, var, vardata);
5687 : :
8113 5688 : 1830049 : return;
5689 : : }
5690 : :
5691 : : /*
5692 : : * Okay, it's a more complicated expression. Determine variable
5693 : : * membership. Note that when varRelid isn't zero, only vars of that
5694 : : * relation are considered "real" vars.
5695 : : */
1930 5696 : 740725 : varnos = pull_varnos(root, basenode);
488 rguo@postgresql.org 5697 : 740725 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5698 : :
8113 tgl@sss.pgh.pa.us 5699 : 740725 : onerel = NULL;
5700 : :
488 rguo@postgresql.org 5701 [ + + ]: 740725 : if (bms_is_empty(basevarnos))
5702 : : {
5703 : : /* No Vars at all ... must be pseudo-constant clause */
5704 : : }
5705 : : else
5706 : : {
5707 : : int relid;
5708 : :
5709 : : /* Check if the expression is in vars of a single base relation */
5710 [ + + ]: 394675 : if (bms_get_singleton_member(basevarnos, &relid))
5711 : : {
889 drowley@postgresql.o 5712 [ + + + + ]: 391056 : if (varRelid == 0 || varRelid == relid)
5713 : : {
5714 : 54135 : onerel = find_base_rel(root, relid);
8113 tgl@sss.pgh.pa.us 5715 : 54135 : vardata->rel = onerel;
127 rguo@postgresql.org 5716 : 54135 : node = basenode; /* strip any phvs or relabeling */
5717 : : }
5718 : : /* else treat it as a constant */
5719 : : }
5720 : : else
5721 : : {
5722 : : /* varnos has multiple relids */
8113 tgl@sss.pgh.pa.us 5723 [ + + ]: 3619 : if (varRelid == 0)
5724 : : {
5725 : : /* treat it as a variable of a join relation */
5726 : 2659 : vardata->rel = find_join_rel(root, varnos);
127 rguo@postgresql.org 5727 : 2659 : node = basenode; /* strip any phvs or relabeling */
5728 : : }
8113 tgl@sss.pgh.pa.us 5729 [ + + ]: 960 : else if (bms_is_member(varRelid, varnos))
5730 : : {
5731 : : /* ignore the vars belonging to other relations */
5732 : 865 : vardata->rel = find_base_rel(root, varRelid);
127 rguo@postgresql.org 5733 : 865 : node = basenode; /* strip any phvs or relabeling */
5734 : : /* note: no point in expressional-index search here */
5735 : : }
5736 : : /* else treat it as a constant */
5737 : : }
5738 : : }
5739 : :
488 5740 : 740725 : bms_free(basevarnos);
5741 : :
7710 tgl@sss.pgh.pa.us 5742 : 740725 : vardata->var = node;
8113 5743 : 740725 : vardata->atttype = exprType(node);
5744 : 740725 : vardata->atttypmod = exprTypmod(node);
5745 : :
5746 [ + + ]: 740725 : if (onerel)
5747 : : {
5748 : : /*
5749 : : * We have an expression in vars of a single relation. Try to match
5750 : : * it to expressional index columns, in hopes of finding some
5751 : : * statistics.
5752 : : *
5753 : : * Note that we consider all index columns including INCLUDE columns,
5754 : : * since there could be stats for such columns. But the test for
5755 : : * uniqueness needs to be warier.
5756 : : *
5757 : : * XXX it's conceivable that there are multiple matches with different
5758 : : * index opfamilies; if so, we need to pick one that matches the
5759 : : * operator we are estimating for. FIXME later.
5760 : : */
5761 : : ListCell *ilist;
5762 : : ListCell *slist;
5763 : :
5764 : : /*
5765 : : * The nullingrels bits within the expression could prevent us from
5766 : : * matching it to expressional index columns or to the expressions in
5767 : : * extended statistics. So strip them out first.
5768 : : */
488 rguo@postgresql.org 5769 [ + + ]: 54135 : if (bms_overlap(varnos, root->outer_join_rels))
5770 : 1565 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5771 : :
8113 tgl@sss.pgh.pa.us 5772 [ + + + + : 119249 : foreach(ilist, onerel->indexlist)
+ + ]
5773 : : {
5774 : 67491 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5775 : : ListCell *indexpr_item;
5776 : : int pos;
5777 : :
8014 neilc@samurai.com 5778 : 67491 : indexpr_item = list_head(index->indexprs);
5779 [ + + ]: 67491 : if (indexpr_item == NULL)
8113 tgl@sss.pgh.pa.us 5780 : 63539 : continue; /* no expressions here... */
5781 : :
5782 [ + + ]: 5587 : for (pos = 0; pos < index->ncolumns; pos++)
5783 : : {
5784 [ + + ]: 4012 : if (index->indexkeys[pos] == 0)
5785 : : {
5786 : : Node *indexkey;
5787 : :
8014 neilc@samurai.com 5788 [ - + ]: 3952 : if (indexpr_item == NULL)
8113 tgl@sss.pgh.pa.us 5789 [ # # ]:UBC 0 : elog(ERROR, "too few entries in indexprs list");
8014 neilc@samurai.com 5790 :CBC 3952 : indexkey = (Node *) lfirst(indexpr_item);
8113 tgl@sss.pgh.pa.us 5791 [ + - - + ]: 3952 : if (indexkey && IsA(indexkey, RelabelType))
8113 tgl@sss.pgh.pa.us 5792 :UBC 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
8113 tgl@sss.pgh.pa.us 5793 [ + + ]:CBC 3952 : if (equal(node, indexkey))
5794 : : {
5795 : : /*
5796 : : * Found a match ... is it a unique index? Tests here
5797 : : * should match has_unique_index().
5798 : : */
5799 [ + + ]: 2907 : if (index->unique &&
2950 teodor@sigaev.ru 5800 [ + - + - ]: 365 : index->nkeycolumns == 1 &&
2639 tgl@sss.pgh.pa.us 5801 : 365 : pos == 0 &&
6288 5802 [ - + - - ]: 365 : (index->indpred == NIL || index->predOK))
8113 5803 : 365 : vardata->isunique = true;
5804 : :
5805 : : /*
5806 : : * Has it got stats? We only consider stats for
5807 : : * non-partial indexes, since partial indexes probably
5808 : : * don't reflect whole-relation statistics; the above
5809 : : * check for uniqueness is the only info we take from
5810 : : * a partial index.
5811 : : *
5812 : : * An index stats hook, however, must make its own
5813 : : * decisions about what to do with partial indexes.
5814 : : */
6428 5815 [ - + - - ]: 2907 : if (get_index_stats_hook &&
6428 tgl@sss.pgh.pa.us 5816 :UBC 0 : (*get_index_stats_hook) (root, index->indexoid,
5817 : 0 : pos + 1, vardata))
5818 : : {
5819 : : /*
5820 : : * The hook took control of acquiring a stats
5821 : : * tuple. If it did supply a tuple, it'd better
5822 : : * have supplied a freefunc.
5823 : : */
5824 [ # # ]: 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5825 [ # # ]: 0 : !vardata->freefunc)
5826 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
5827 : : }
6288 tgl@sss.pgh.pa.us 5828 [ + - ]:CBC 2907 : else if (index->indpred == NIL)
5829 : : {
6428 5830 : 2907 : vardata->statsTuple =
5924 rhaas@postgresql.org 5831 : 5814 : SearchSysCache3(STATRELATTINH,
5832 : : ObjectIdGetDatum(index->indexoid),
5912 bruce@momjian.us 5833 : 2907 : Int16GetDatum(pos + 1),
5834 : : BoolGetDatum(false));
6428 tgl@sss.pgh.pa.us 5835 : 2907 : vardata->freefunc = ReleaseSysCache;
5836 : :
3287 peter_e@gmx.net 5837 [ + + ]: 2907 : if (HeapTupleIsValid(vardata->statsTuple))
5838 : : {
5839 : : /*
5840 : : * Test if user has permission to access all
5841 : : * rows from the index's table.
5842 : : *
5843 : : * For simplicity, we insist on the whole
5844 : : * table being selectable, rather than trying
5845 : : * to identify which column(s) the index
5846 : : * depends on.
5847 : : *
5848 : : * Note that for an inheritance child,
5849 : : * permissions are checked on the inheritance
5850 : : * root parent, and whole-table select
5851 : : * privilege on the parent doesn't quite
5852 : : * guarantee that the user could read all
5853 : : * columns of the child. But in practice it's
5854 : : * unlikely that any interesting security
5855 : : * violation could result from allowing access
5856 : : * to the expression index's stats, so we
5857 : : * allow it anyway. See similar code in
5858 : : * examine_simple_variable() for additional
5859 : : * comments.
5860 : : */
5861 : 2377 : vardata->acl_ok =
267 dean.a.rasheed@gmail 5862 : 2377 : all_rows_selectable(root,
5863 : 2377 : index->rel->relid,
5864 : : NULL);
5865 : : }
5866 : : else
5867 : : {
5868 : : /* suppress leakproofness checks later */
3287 peter_e@gmx.net 5869 : 530 : vardata->acl_ok = true;
5870 : : }
5871 : : }
8113 tgl@sss.pgh.pa.us 5872 [ + + ]: 2907 : if (vardata->statsTuple)
5873 : 2377 : break;
5874 : : }
2486 5875 : 1575 : indexpr_item = lnext(index->indexprs, indexpr_item);
5876 : : }
5877 : : }
8113 5878 [ + + ]: 3952 : if (vardata->statsTuple)
5879 : 2377 : break;
5880 : : }
5881 : :
5882 : : /*
5883 : : * Search extended statistics for one with a matching expression.
5884 : : * There might be multiple ones, so just grab the first one. In the
5885 : : * future, we might consider the statistics target (and pick the most
5886 : : * accurate statistics) and maybe some other parameters.
5887 : : */
1866 tomas.vondra@postgre 5888 [ + + + + : 57664 : foreach(slist, onerel->statlist)
+ + ]
5889 : : {
5890 : 3774 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
1454 tgl@sss.pgh.pa.us 5891 [ + - ]: 3774 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5892 : : ListCell *expr_item;
5893 : : int pos;
5894 : :
5895 : : /*
5896 : : * Stop once we've found statistics for the expression (either
5897 : : * from extended stats, or for an index in the preceding loop).
5898 : : */
1866 tomas.vondra@postgre 5899 [ + + ]: 3774 : if (vardata->statsTuple)
5900 : 245 : break;
5901 : :
5902 : : /* skip stats without per-expression stats */
5903 [ + + ]: 3529 : if (info->kind != STATS_EXT_EXPRESSIONS)
5904 : 1808 : continue;
5905 : :
5906 : : /* skip stats with mismatching stxdinherit value */
1281 tgl@sss.pgh.pa.us 5907 [ + + ]: 1721 : if (info->inherit != rte->inh)
5908 : 5 : continue;
5909 : :
1866 tomas.vondra@postgre 5910 : 1716 : pos = 0;
5911 [ + - + + : 2836 : foreach(expr_item, info->exprs)
+ + ]
5912 : : {
5913 : 2536 : Node *expr = (Node *) lfirst(expr_item);
5914 : :
5915 [ - + ]: 2536 : Assert(expr);
5916 : :
5917 : : /* strip RelabelType before comparing it */
5918 [ + - - + ]: 2536 : if (expr && IsA(expr, RelabelType))
1866 tomas.vondra@postgre 5919 :UBC 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5920 : :
5921 : : /* found a match, see if we can extract pg_statistic row */
1866 tomas.vondra@postgre 5922 [ + + ]:CBC 2536 : if (equal(node, expr))
5923 : : {
5924 : : /*
5925 : : * XXX Not sure if we should cache the tuple somewhere.
5926 : : * Now we just create a new copy every time.
5927 : : */
1570 5928 : 1416 : vardata->statsTuple =
5929 : 1416 : statext_expressions_load(info->statOid, rte->inh, pos);
5930 : :
5931 : : /* Nothing to release if no data found */
64 michael@paquier.xyz 5932 [ + + ]: 1416 : if (vardata->statsTuple != NULL)
5933 : : {
5934 : 1415 : vardata->freefunc = ReleaseDummy;
5935 : : }
5936 : :
5937 : : /*
5938 : : * Test if user has permission to access all rows from the
5939 : : * table.
5940 : : *
5941 : : * For simplicity, we insist on the whole table being
5942 : : * selectable, rather than trying to identify which
5943 : : * column(s) the statistics object depends on.
5944 : : *
5945 : : * Note that for an inheritance child, permissions are
5946 : : * checked on the inheritance root parent, and whole-table
5947 : : * select privilege on the parent doesn't quite guarantee
5948 : : * that the user could read all columns of the child. But
5949 : : * in practice it's unlikely that any interesting security
5950 : : * violation could result from allowing access to the
5951 : : * expression stats, so we allow it anyway. See similar
5952 : : * code in examine_simple_variable() for additional
5953 : : * comments.
5954 : : */
267 dean.a.rasheed@gmail 5955 : 1416 : vardata->acl_ok = all_rows_selectable(root,
5956 : : onerel->relid,
5957 : : NULL);
5958 : :
1866 tomas.vondra@postgre 5959 : 1416 : break;
5960 : : }
5961 : :
5962 : 1120 : pos++;
5963 : : }
5964 : : }
5965 : : }
5966 : :
488 rguo@postgresql.org 5967 : 740725 : bms_free(varnos);
5968 : : }
5969 : :
5970 : : /*
5971 : : * strip_all_phvs_deep
5972 : : * Deeply strip all PlaceHolderVars in an expression.
5973 : :
5974 : : * As a performance optimization, we first use a lightweight walker to check
5975 : : * for the presence of any PlaceHolderVars. The expensive mutator is invoked
5976 : : * only if a PlaceHolderVar is found, avoiding unnecessary memory allocation
5977 : : * and tree copying in the common case where no PlaceHolderVars are present.
5978 : : */
5979 : : static Node *
127 5980 : 2570774 : strip_all_phvs_deep(PlannerInfo *root, Node *node)
5981 : : {
5982 : : /* If there are no PHVs anywhere, we needn't work hard */
5983 [ + + ]: 2570774 : if (root->glob->lastPHId == 0)
5984 : 2544750 : return node;
5985 : :
5986 [ + + ]: 26024 : if (!contain_placeholder_walker(node, NULL))
5987 : 22346 : return node;
5988 : 3678 : return strip_all_phvs_mutator(node, NULL);
5989 : : }
5990 : :
5991 : : /*
5992 : : * contain_placeholder_walker
5993 : : * Lightweight walker to check if an expression contains any
5994 : : * PlaceHolderVars
5995 : : */
5996 : : static bool
5997 : 29350 : contain_placeholder_walker(Node *node, void *context)
5998 : : {
5999 [ + + ]: 29350 : if (node == NULL)
6000 : 109 : return false;
6001 [ + + ]: 29241 : if (IsA(node, PlaceHolderVar))
6002 : 3678 : return true;
6003 : :
6004 : 25563 : return expression_tree_walker(node, contain_placeholder_walker, context);
6005 : : }
6006 : :
6007 : : /*
6008 : : * strip_all_phvs_mutator
6009 : : * Mutator to deeply strip all PlaceHolderVars
6010 : : */
6011 : : static Node *
6012 : 9809 : strip_all_phvs_mutator(Node *node, void *context)
6013 : : {
6014 [ + + ]: 9809 : if (node == NULL)
6015 : 34 : return NULL;
6016 [ + + ]: 9775 : if (IsA(node, PlaceHolderVar))
6017 : : {
6018 : : /* Strip it and recurse into its contained expression */
6019 : 3798 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
6020 : :
6021 : 3798 : return strip_all_phvs_mutator((Node *) phv->phexpr, context);
6022 : : }
6023 : :
6024 : 5977 : return expression_tree_mutator(node, strip_all_phvs_mutator, context);
6025 : : }
6026 : :
6027 : : /*
6028 : : * examine_simple_variable
6029 : : * Handle a simple Var for examine_variable
6030 : : *
6031 : : * This is split out as a subroutine so that we can recurse to deal with
6032 : : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
6033 : : *
6034 : : * We already filled in all the fields of *vardata except for the stats tuple.
6035 : : */
6036 : : static void
5357 tgl@sss.pgh.pa.us 6037 : 1838143 : examine_simple_variable(PlannerInfo *root, Var *var,
6038 : : VariableStatData *vardata)
6039 : : {
6040 : 1838143 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
6041 : :
6042 [ - + ]: 1838143 : Assert(IsA(rte, RangeTblEntry));
6043 : :
6044 [ - + - - ]: 1838143 : if (get_relation_stats_hook &&
5357 tgl@sss.pgh.pa.us 6045 :UBC 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
6046 : : {
6047 : : /*
6048 : : * The hook took control of acquiring a stats tuple. If it did supply
6049 : : * a tuple, it'd better have supplied a freefunc.
6050 : : */
6051 [ # # ]: 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6052 [ # # ]: 0 : !vardata->freefunc)
6053 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
6054 : : }
5357 tgl@sss.pgh.pa.us 6055 [ + + ]:CBC 1838143 : else if (rte->rtekind == RTE_RELATION)
6056 : : {
6057 : : /*
6058 : : * Plain table or parent of an inheritance appendrel, so look up the
6059 : : * column in pg_statistic
6060 : : */
6061 : 1745336 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6062 : : ObjectIdGetDatum(rte->relid),
6063 : 1745336 : Int16GetDatum(var->varattno),
6064 : 1745336 : BoolGetDatum(rte->inh));
6065 : 1745336 : vardata->freefunc = ReleaseSysCache;
6066 : :
3287 peter_e@gmx.net 6067 [ + + ]: 1745336 : if (HeapTupleIsValid(vardata->statsTuple))
6068 : : {
6069 : : /*
6070 : : * Test if user has permission to read all rows from this column.
6071 : : *
6072 : : * This requires that the user has the appropriate SELECT
6073 : : * privileges and that there are no securityQuals from security
6074 : : * barrier views or RLS policies. If that's not the case, then we
6075 : : * only permit leakproof functions to be passed pg_statistic data
6076 : : * in vardata, otherwise the functions might reveal data that the
6077 : : * user doesn't have permission to see --- see
6078 : : * statistic_proc_security_check().
6079 : : */
6080 : 1218783 : vardata->acl_ok =
267 dean.a.rasheed@gmail 6081 : 1218783 : all_rows_selectable(root, var->varno,
6082 : 1218783 : bms_make_singleton(var->varattno - FirstLowInvalidHeapAttributeNumber));
6083 : : }
6084 : : else
6085 : : {
6086 : : /* suppress any possible leakproofness checks later */
3287 peter_e@gmx.net 6087 : 526553 : vardata->acl_ok = true;
6088 : : }
6089 : : }
900 tgl@sss.pgh.pa.us 6090 [ + + + + ]: 92807 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
6091 [ + + + + ]: 85040 : (rte->rtekind == RTE_CTE && !rte->self_reference))
6092 : : {
6093 : : /*
6094 : : * Plain subquery (not one that was converted to an appendrel) or
6095 : : * non-recursive CTE. In either case, we can try to find out what the
6096 : : * Var refers to within the subquery. We skip this for appendrel and
6097 : : * recursive-CTE cases because any column stats we did find would
6098 : : * likely not be very relevant.
6099 : : */
6100 : : PlannerInfo *subroot;
6101 : : Query *subquery;
6102 : : List *subtlist;
6103 : : TargetEntry *ste;
6104 : :
6105 : : /*
6106 : : * Punt if it's a whole-row var rather than a plain column reference.
6107 : : */
4558 6108 [ - + ]: 15738 : if (var->varattno == InvalidAttrNumber)
4558 tgl@sss.pgh.pa.us 6109 :UBC 0 : return;
6110 : :
6111 : : /*
6112 : : * Otherwise, find the subquery's planner subroot.
6113 : : */
900 tgl@sss.pgh.pa.us 6114 [ + + ]:CBC 15738 : if (rte->rtekind == RTE_SUBQUERY)
6115 : : {
6116 : : RelOptInfo *rel;
6117 : :
6118 : : /*
6119 : : * Fetch RelOptInfo for subquery. Note that we don't change the
6120 : : * rel returned in vardata, since caller expects it to be a rel of
6121 : : * the caller's query level. Because we might already be
6122 : : * recursing, we can't use that rel pointer either, but have to
6123 : : * look up the Var's rel afresh.
6124 : : */
6125 : 7767 : rel = find_base_rel(root, var->varno);
6126 : :
6127 : 7767 : subroot = rel->subroot;
6128 : : }
6129 : : else
6130 : : {
6131 : : /* CTE case is more difficult */
6132 : : PlannerInfo *cteroot;
6133 : : Index levelsup;
6134 : : int ndx;
6135 : : int plan_id;
6136 : : ListCell *lc;
6137 : :
6138 : : /*
6139 : : * Find the referenced CTE, and locate the subroot previously made
6140 : : * for it.
6141 : : */
6142 : 7971 : levelsup = rte->ctelevelsup;
6143 : 7971 : cteroot = root;
6144 [ + + ]: 19011 : while (levelsup-- > 0)
6145 : : {
6146 : 11040 : cteroot = cteroot->parent_root;
6147 [ - + ]: 11040 : if (!cteroot) /* shouldn't happen */
900 tgl@sss.pgh.pa.us 6148 [ # # ]:UBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
6149 : : }
6150 : :
6151 : : /*
6152 : : * Note: cte_plan_ids can be shorter than cteList, if we are still
6153 : : * working on planning the CTEs (ie, this is a side-reference from
6154 : : * another CTE). So we mustn't use forboth here.
6155 : : */
900 tgl@sss.pgh.pa.us 6156 :CBC 7971 : ndx = 0;
6157 [ + - + - : 11017 : foreach(lc, cteroot->parse->cteList)
+ - ]
6158 : : {
6159 : 11017 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
6160 : :
6161 [ + + ]: 11017 : if (strcmp(cte->ctename, rte->ctename) == 0)
6162 : 7971 : break;
6163 : 3046 : ndx++;
6164 : : }
6165 [ - + ]: 7971 : if (lc == NULL) /* shouldn't happen */
900 tgl@sss.pgh.pa.us 6166 [ # # ]:UBC 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
900 tgl@sss.pgh.pa.us 6167 [ - + ]:CBC 7971 : if (ndx >= list_length(cteroot->cte_plan_ids))
900 tgl@sss.pgh.pa.us 6168 [ # # ]:UBC 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
900 tgl@sss.pgh.pa.us 6169 :CBC 7971 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
6170 [ - + ]: 7971 : if (plan_id <= 0)
900 tgl@sss.pgh.pa.us 6171 [ # # ]:UBC 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
900 tgl@sss.pgh.pa.us 6172 :CBC 7971 : subroot = list_nth(root->glob->subroots, plan_id - 1);
6173 : : }
6174 : :
6175 : : /* If the subquery hasn't been planned yet, we have to punt */
6176 [ - + ]: 15738 : if (subroot == NULL)
900 tgl@sss.pgh.pa.us 6177 :UBC 0 : return;
900 tgl@sss.pgh.pa.us 6178 [ - + ]:CBC 15738 : Assert(IsA(subroot, PlannerInfo));
6179 : :
6180 : : /*
6181 : : * We must use the subquery parsetree as mangled by the planner, not
6182 : : * the raw version from the RTE, because we need a Var that will refer
6183 : : * to the subroot's live RelOptInfos. For instance, if any subquery
6184 : : * pullup happened during planning, Vars in the targetlist might have
6185 : : * gotten replaced, and we need to see the replacement expressions.
6186 : : */
6187 : 15738 : subquery = subroot->parse;
6188 [ - + ]: 15738 : Assert(IsA(subquery, Query));
6189 : :
6190 : : /*
6191 : : * Punt if subquery uses set operations or grouping sets, as these
6192 : : * will mash underlying columns' stats beyond recognition. (Set ops
6193 : : * are particularly nasty; if we forged ahead, we would return stats
6194 : : * relevant to only the leftmost subselect...) DISTINCT is also
6195 : : * problematic, but we check that later because there is a possibility
6196 : : * of learning something even with it.
6197 : : */
5192 6198 [ + + ]: 15738 : if (subquery->setOperations ||
1799 6199 [ + + ]: 14328 : subquery->groupingSets)
5248 rhaas@postgresql.org 6200 : 1487 : return;
6201 : :
6202 : : /* Get the subquery output expression referenced by the upper Var */
900 tgl@sss.pgh.pa.us 6203 [ + + ]: 14251 : if (subquery->returningList)
6204 : 179 : subtlist = subquery->returningList;
6205 : : else
6206 : 14072 : subtlist = subquery->targetList;
6207 : 14251 : ste = get_tle_by_resno(subtlist, var->varattno);
5357 6208 [ + - - + ]: 14251 : if (ste == NULL || ste->resjunk)
5357 tgl@sss.pgh.pa.us 6209 [ # # ]:UBC 0 : elog(ERROR, "subquery %s does not have attribute %d",
6210 : : rte->eref->aliasname, var->varattno);
5357 tgl@sss.pgh.pa.us 6211 :CBC 14251 : var = (Var *) ste->expr;
6212 : :
6213 : : /*
6214 : : * If subquery uses DISTINCT, we can't make use of any stats for the
6215 : : * variable ... but, if it's the only DISTINCT column, we are entitled
6216 : : * to consider it unique. We do the test this way so that it works
6217 : : * for cases involving DISTINCT ON.
6218 : : */
5192 6219 [ + + ]: 14251 : if (subquery->distinctClause)
6220 : : {
6221 [ + + + + ]: 1338 : if (list_length(subquery->distinctClause) == 1 &&
6222 : 464 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
6223 : 310 : vardata->isunique = true;
6224 : : /* cannot go further */
6225 : 874 : return;
6226 : : }
6227 : :
6228 : : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
440 akorotkov@postgresql 6229 [ + + ]: 13377 : if (subquery->groupClause)
6230 : : {
6231 [ + + + + ]: 947 : if (list_length(subquery->groupClause) == 1 &&
6232 : 402 : targetIsInSortList(ste, InvalidOid, subquery->groupClause))
6233 : 322 : vardata->isunique = true;
6234 : : /* cannot go further */
6235 : 545 : return;
6236 : : }
6237 : :
6238 : : /*
6239 : : * If the sub-query originated from a view with the security_barrier
6240 : : * attribute, we must not look at the variable's statistics, though it
6241 : : * seems all right to notice the existence of a DISTINCT clause. So
6242 : : * stop here.
6243 : : *
6244 : : * This is probably a harsher restriction than necessary; it's
6245 : : * certainly OK for the selectivity estimator (which is a C function,
6246 : : * and therefore omnipotent anyway) to look at the statistics. But
6247 : : * many selectivity estimators will happily *invoke the operator
6248 : : * function* to try to work out a good estimate - and that's not OK.
6249 : : * So for now, don't dig down for stats.
6250 : : */
5192 tgl@sss.pgh.pa.us 6251 [ + + ]: 12832 : if (rte->security_barrier)
6252 : 480 : return;
6253 : :
6254 : : /* Can only handle a simple Var of subquery's query level */
5357 6255 [ + - + + ]: 12352 : if (var && IsA(var, Var) &&
6256 [ + - ]: 8094 : var->varlevelsup == 0)
6257 : : {
6258 : : /*
6259 : : * OK, recurse into the subquery. Note that the original setting
6260 : : * of vardata->isunique (which will surely be false) is left
6261 : : * unchanged in this situation. That's what we want, since even
6262 : : * if the underlying column is unique, the subquery may have
6263 : : * joined to other tables in a way that creates duplicates.
6264 : : */
900 6265 : 8094 : examine_simple_variable(subroot, var, vardata);
6266 : : }
6267 : : }
6268 : : else
6269 : : {
6270 : : /*
6271 : : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
6272 : : * see RTE_JOIN here because join alias Vars have already been
6273 : : * flattened.) There's not much we can do with function outputs, but
6274 : : * maybe someday try to be smarter about VALUES.
6275 : : */
6276 : : }
6277 : : }
6278 : :
6279 : : /*
6280 : : * all_rows_selectable
6281 : : * Test whether the user has permission to select all rows from a given
6282 : : * relation.
6283 : : *
6284 : : * Inputs:
6285 : : * root: the planner info
6286 : : * varno: the index of the relation (assumed to be an RTE_RELATION)
6287 : : * varattnos: the attributes for which permission is required, or NULL if
6288 : : * whole-table access is required
6289 : : *
6290 : : * Returns true if the user has the required select permissions, and there are
6291 : : * no securityQuals from security barrier views or RLS policies.
6292 : : *
6293 : : * Note that if the relation is an inheritance child relation, securityQuals
6294 : : * and access permissions are checked against the inheritance root parent (the
6295 : : * relation actually mentioned in the query) --- see the comments in
6296 : : * expand_single_inheritance_child() for an explanation of why it has to be
6297 : : * done this way.
6298 : : *
6299 : : * If varattnos is non-NULL, its attribute numbers should be offset by
6300 : : * FirstLowInvalidHeapAttributeNumber so that system attributes can be
6301 : : * checked. If varattnos is NULL, only table-level SELECT privileges are
6302 : : * checked, not any column-level privileges.
6303 : : *
6304 : : * Note: if the relation is accessed via a view, this function actually tests
6305 : : * whether the view owner has permission to select from the relation. To
6306 : : * ensure that the current user has permission, it is also necessary to check
6307 : : * that the current user has permission to select from the view, which we do
6308 : : * at planner-startup --- see subquery_planner().
6309 : : *
6310 : : * This is exported so that other estimation functions can use it.
6311 : : */
6312 : : bool
267 dean.a.rasheed@gmail 6313 : 1222786 : all_rows_selectable(PlannerInfo *root, Index varno, Bitmapset *varattnos)
6314 : : {
6315 : 1222786 : RelOptInfo *rel = find_base_rel_noerr(root, varno);
6316 [ + - ]: 1222786 : RangeTblEntry *rte = planner_rt_fetch(varno, root);
6317 : : Oid userid;
6318 : : int varattno;
6319 : :
6320 [ - + ]: 1222786 : Assert(rte->rtekind == RTE_RELATION);
6321 : :
6322 : : /*
6323 : : * Determine the user ID to use for privilege checks (either the current
6324 : : * user or the view owner, if we're accessing the table via a view).
6325 : : *
6326 : : * Normally the relation will have an associated RelOptInfo from which we
6327 : : * can find the userid, but it might not if it's a RETURNING Var for an
6328 : : * INSERT target relation. In that case use the RTEPermissionInfo
6329 : : * associated with the RTE.
6330 : : *
6331 : : * If we navigate up to a parent relation, we keep using the same userid,
6332 : : * since it's the same in all relations of a given inheritance tree.
6333 : : */
6334 [ + + ]: 1222786 : if (rel)
6335 : 1222753 : userid = rel->userid;
6336 : : else
6337 : : {
6338 : : RTEPermissionInfo *perminfo;
6339 : :
6340 : 33 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
6341 : 33 : userid = perminfo->checkAsUser;
6342 : : }
6343 [ + + ]: 1222786 : if (!OidIsValid(userid))
6344 : 1091226 : userid = GetUserId();
6345 : :
6346 : : /*
6347 : : * Permissions and securityQuals must be checked on the table actually
6348 : : * mentioned in the query, so if this is an inheritance child, navigate up
6349 : : * to the inheritance root parent. If the user can read the whole table
6350 : : * or the required columns there, then they can read from the child table
6351 : : * too. For per-column checks, we must find out which of the root
6352 : : * parent's attributes the child relation's attributes correspond to.
6353 : : */
6354 [ + + ]: 1222786 : if (root->append_rel_array != NULL)
6355 : : {
6356 : : AppendRelInfo *appinfo;
6357 : :
6358 : 177850 : appinfo = root->append_rel_array[varno];
6359 : :
6360 : : /*
6361 : : * Partitions are mapped to their immediate parent, not the root
6362 : : * parent, so must be ready to walk up multiple AppendRelInfos. But
6363 : : * stop if we hit a parent that is not RTE_RELATION --- that's a
6364 : : * flattened UNION ALL subquery, not an inheritance parent.
6365 : : */
6366 [ + + ]: 334192 : while (appinfo &&
6367 [ + - ]: 156727 : planner_rt_fetch(appinfo->parent_relid,
6368 [ + + ]: 156727 : root)->rtekind == RTE_RELATION)
6369 : : {
6370 : 156342 : Bitmapset *parent_varattnos = NULL;
6371 : :
6372 : : /*
6373 : : * For each child attribute, find the corresponding parent
6374 : : * attribute. In rare cases, the attribute may be local to the
6375 : : * child table, in which case, we've got to live with having no
6376 : : * access to this column.
6377 : : */
6378 : 156342 : varattno = -1;
6379 [ + + ]: 310417 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6380 : : {
6381 : : AttrNumber attno;
6382 : : AttrNumber parent_attno;
6383 : :
6384 : 154075 : attno = varattno + FirstLowInvalidHeapAttributeNumber;
6385 : :
6386 [ + + ]: 154075 : if (attno == InvalidAttrNumber)
6387 : : {
6388 : : /*
6389 : : * Whole-row reference, so must map each column of the
6390 : : * child to the parent table.
6391 : : */
6392 [ + + ]: 30 : for (attno = 1; attno <= appinfo->num_child_cols; attno++)
6393 : : {
6394 : 20 : parent_attno = appinfo->parent_colnos[attno - 1];
6395 [ - + ]: 20 : if (parent_attno == 0)
267 dean.a.rasheed@gmail 6396 :UBC 0 : return false; /* attr is local to child */
6397 : : parent_varattnos =
267 dean.a.rasheed@gmail 6398 :CBC 20 : bms_add_member(parent_varattnos,
6399 : : parent_attno - FirstLowInvalidHeapAttributeNumber);
6400 : : }
6401 : : }
6402 : : else
6403 : : {
6404 [ - + ]: 154065 : if (attno < 0)
6405 : : {
6406 : : /* System attnos are the same in all tables */
267 dean.a.rasheed@gmail 6407 :UBC 0 : parent_attno = attno;
6408 : : }
6409 : : else
6410 : : {
267 dean.a.rasheed@gmail 6411 [ - + ]:CBC 154065 : if (attno > appinfo->num_child_cols)
267 dean.a.rasheed@gmail 6412 :UBC 0 : return false; /* safety check */
267 dean.a.rasheed@gmail 6413 :CBC 154065 : parent_attno = appinfo->parent_colnos[attno - 1];
6414 [ - + ]: 154065 : if (parent_attno == 0)
267 dean.a.rasheed@gmail 6415 :UBC 0 : return false; /* attr is local to child */
6416 : : }
6417 : : parent_varattnos =
267 dean.a.rasheed@gmail 6418 :CBC 154065 : bms_add_member(parent_varattnos,
6419 : : parent_attno - FirstLowInvalidHeapAttributeNumber);
6420 : : }
6421 : : }
6422 : :
6423 : : /* If the parent is itself a child, continue up */
6424 : 156342 : varno = appinfo->parent_relid;
6425 : 156342 : varattnos = parent_varattnos;
6426 : 156342 : appinfo = root->append_rel_array[varno];
6427 : : }
6428 : :
6429 : : /* Perform the access check on this parent rel */
6430 [ + - ]: 177850 : rte = planner_rt_fetch(varno, root);
6431 [ - + ]: 177850 : Assert(rte->rtekind == RTE_RELATION);
6432 : : }
6433 : :
6434 : : /*
6435 : : * For all rows to be accessible, there must be no securityQuals from
6436 : : * security barrier views or RLS policies.
6437 : : */
6438 [ + + ]: 1222786 : if (rte->securityQuals != NIL)
6439 : 690 : return false;
6440 : :
6441 : : /*
6442 : : * Test for table-level SELECT privilege.
6443 : : *
6444 : : * If varattnos is non-NULL, this is sufficient to give access to all
6445 : : * requested attributes, even for a child table, since we have verified
6446 : : * that all required child columns have matching parent columns.
6447 : : *
6448 : : * If varattnos is NULL (whole-table access requested), this doesn't
6449 : : * necessarily guarantee that the user can read all columns of a child
6450 : : * table, but we allow it anyway (see comments in examine_variable()) and
6451 : : * don't bother checking any column privileges.
6452 : : */
6453 [ + + ]: 1222096 : if (pg_class_aclcheck(rte->relid, userid, ACL_SELECT) == ACLCHECK_OK)
6454 : 1221753 : return true;
6455 : :
6456 [ + + ]: 343 : if (varattnos == NULL)
6457 : 10 : return false; /* whole-table access requested */
6458 : :
6459 : : /*
6460 : : * Don't have table-level SELECT privilege, so check per-column
6461 : : * privileges.
6462 : : */
6463 : 333 : varattno = -1;
6464 [ + + ]: 471 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6465 : : {
6466 : 333 : AttrNumber attno = varattno + FirstLowInvalidHeapAttributeNumber;
6467 : :
6468 [ + + ]: 333 : if (attno == InvalidAttrNumber)
6469 : : {
6470 : : /* Whole-row reference, so must have access to all columns */
6471 [ + - ]: 5 : if (pg_attribute_aclcheck_all(rte->relid, userid, ACL_SELECT,
6472 : : ACLMASK_ALL) != ACLCHECK_OK)
6473 : 5 : return false;
6474 : : }
6475 : : else
6476 : : {
6477 [ + + ]: 328 : if (pg_attribute_aclcheck(rte->relid, attno, userid,
6478 : : ACL_SELECT) != ACLCHECK_OK)
6479 : 190 : return false;
6480 : : }
6481 : : }
6482 : :
6483 : : /* If we reach here, have all required column privileges */
6484 : 138 : return true;
6485 : : }
6486 : :
6487 : : /*
6488 : : * examine_indexcol_variable
6489 : : * Try to look up statistical data about an index column/expression.
6490 : : * Fill in a VariableStatData struct to describe the column.
6491 : : *
6492 : : * Inputs:
6493 : : * root: the planner info
6494 : : * index: the index whose column we're interested in
6495 : : * indexcol: 0-based index column number (subscripts index->indexkeys[])
6496 : : *
6497 : : * Outputs: *vardata is filled as follows:
6498 : : * var: the input expression (with any binary relabeling stripped, if
6499 : : * it is or contains a variable; but otherwise the type is preserved)
6500 : : * rel: RelOptInfo for table relation containing variable.
6501 : : * statsTuple: the pg_statistic entry for the variable, if one exists;
6502 : : * otherwise NULL.
6503 : : * freefunc: pointer to a function to release statsTuple with.
6504 : : *
6505 : : * Caller is responsible for doing ReleaseVariableStats() before exiting.
6506 : : */
6507 : : static void
396 pg@bowt.ie 6508 : 654424 : examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
6509 : : int indexcol, VariableStatData *vardata)
6510 : : {
6511 : : AttrNumber colnum;
6512 : : Oid relid;
6513 : :
6514 [ + + ]: 654424 : if (index->indexkeys[indexcol] != 0)
6515 : : {
6516 : : /* Simple variable --- look to stats for the underlying table */
6517 [ + - ]: 652591 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6518 : :
6519 [ - + ]: 652591 : Assert(rte->rtekind == RTE_RELATION);
6520 : 652591 : relid = rte->relid;
6521 [ - + ]: 652591 : Assert(relid != InvalidOid);
6522 : 652591 : colnum = index->indexkeys[indexcol];
6523 : 652591 : vardata->rel = index->rel;
6524 : :
6525 [ - + - - ]: 652591 : if (get_relation_stats_hook &&
396 pg@bowt.ie 6526 :UBC 0 : (*get_relation_stats_hook) (root, rte, colnum, vardata))
6527 : : {
6528 : : /*
6529 : : * The hook took control of acquiring a stats tuple. If it did
6530 : : * supply a tuple, it'd better have supplied a freefunc.
6531 : : */
6532 [ # # ]: 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6533 [ # # ]: 0 : !vardata->freefunc)
6534 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
6535 : : }
6536 : : else
6537 : : {
396 pg@bowt.ie 6538 :CBC 652591 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6539 : : ObjectIdGetDatum(relid),
6540 : : Int16GetDatum(colnum),
6541 : 652591 : BoolGetDatum(rte->inh));
6542 : 652591 : vardata->freefunc = ReleaseSysCache;
6543 : : }
6544 : : }
6545 : : else
6546 : : {
6547 : : /* Expression --- maybe there are stats for the index itself */
6548 : 1833 : relid = index->indexoid;
6549 : 1833 : colnum = indexcol + 1;
6550 : :
6551 [ - + - - ]: 1833 : if (get_index_stats_hook &&
396 pg@bowt.ie 6552 :UBC 0 : (*get_index_stats_hook) (root, relid, colnum, vardata))
6553 : : {
6554 : : /*
6555 : : * The hook took control of acquiring a stats tuple. If it did
6556 : : * supply a tuple, it'd better have supplied a freefunc.
6557 : : */
6558 [ # # ]: 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6559 [ # # ]: 0 : !vardata->freefunc)
6560 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
6561 : : }
6562 : : else
6563 : : {
396 pg@bowt.ie 6564 :CBC 1833 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6565 : : ObjectIdGetDatum(relid),
6566 : : Int16GetDatum(colnum),
6567 : : BoolGetDatum(false));
6568 : 1833 : vardata->freefunc = ReleaseSysCache;
6569 : : }
6570 : : }
6571 : 654424 : }
6572 : :
6573 : : /*
6574 : : * Check whether it is permitted to call func_oid passing some of the
6575 : : * pg_statistic data in vardata. We allow this if either of the following
6576 : : * conditions is met: (1) the user has SELECT privileges on the table or
6577 : : * column underlying the pg_statistic data and there are no securityQuals from
6578 : : * security barrier views or RLS policies, or (2) the function is marked
6579 : : * leakproof.
6580 : : */
6581 : : bool
3287 peter_e@gmx.net 6582 : 799564 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
6583 : : {
6584 [ + + ]: 799564 : if (vardata->acl_ok)
267 dean.a.rasheed@gmail 6585 : 798053 : return true; /* have SELECT privs and no securityQuals */
6586 : :
3287 peter_e@gmx.net 6587 [ - + ]: 1511 : if (!OidIsValid(func_oid))
3287 peter_e@gmx.net 6588 :UBC 0 : return false;
6589 : :
3287 peter_e@gmx.net 6590 [ + + ]:CBC 1511 : if (get_func_leakproof(func_oid))
6591 : 748 : return true;
6592 : :
6593 [ - + ]: 763 : ereport(DEBUG2,
6594 : : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
6595 : : get_func_name(func_oid))));
6596 : 763 : return false;
6597 : : }
6598 : :
6599 : : /*
6600 : : * get_variable_numdistinct
6601 : : * Estimate the number of distinct values of a variable.
6602 : : *
6603 : : * vardata: results of examine_variable
6604 : : * *isdefault: set to true if the result is a default rather than based on
6605 : : * anything meaningful.
6606 : : *
6607 : : * NB: be careful to produce a positive integral result, since callers may
6608 : : * compare the result to exact integer counts, or might divide by it.
6609 : : */
6610 : : double
5357 tgl@sss.pgh.pa.us 6611 : 1299400 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
6612 : : {
6613 : : double stadistinct;
3558 6614 : 1299400 : double stanullfrac = 0.0;
6615 : : double ntuples;
6616 : :
5357 6617 : 1299400 : *isdefault = false;
6618 : :
6619 : : /*
6620 : : * Determine the stadistinct value to use. There are cases where we can
6621 : : * get an estimate even without a pg_statistic entry, or can get a better
6622 : : * value than is in pg_statistic. Grab stanullfrac too if we can find it
6623 : : * (otherwise, assume no nulls, for lack of any better idea).
6624 : : */
8113 6625 [ + + ]: 1299400 : if (HeapTupleIsValid(vardata->statsTuple))
6626 : : {
6627 : : /* Use the pg_statistic entry */
6628 : : Form_pg_statistic stats;
6629 : :
6630 : 860477 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6631 : 860477 : stadistinct = stats->stadistinct;
3558 6632 : 860477 : stanullfrac = stats->stanullfrac;
6633 : : }
7704 6634 [ + + ]: 438923 : else if (vardata->vartype == BOOLOID)
6635 : : {
6636 : : /*
6637 : : * Special-case boolean columns: presumably, two distinct values.
6638 : : *
6639 : : * Are there any other datatypes we should wire in special estimates
6640 : : * for?
6641 : : */
8113 6642 : 665 : stadistinct = 2.0;
6643 : : }
3184 6644 [ + + + + ]: 438258 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6645 : : {
6646 : : /*
6647 : : * If the Var represents a column of a VALUES RTE, assume it's unique.
6648 : : * This could of course be very wrong, but it should tend to be true
6649 : : * in well-written queries. We could consider examining the VALUES'
6650 : : * contents to get some real statistics; but that only works if the
6651 : : * entries are all constants, and it would be pretty expensive anyway.
6652 : : */
6653 : 2999 : stadistinct = -1.0; /* unique (and all non null) */
6654 : : }
6655 : : else
6656 : : {
6657 : : /*
6658 : : * We don't keep statistics for system columns, but in some cases we
6659 : : * can infer distinctness anyway.
6660 : : */
8113 6661 [ + + + + ]: 435259 : if (vardata->var && IsA(vardata->var, Var))
6662 : : {
6663 [ + + + ]: 404861 : switch (((Var *) vardata->var)->varattno)
6664 : : {
6665 : 865 : case SelfItemPointerAttributeNumber:
3558 6666 : 865 : stadistinct = -1.0; /* unique (and all non null) */
8113 6667 : 865 : break;
6668 : 18378 : case TableOidAttributeNumber:
7919 bruce@momjian.us 6669 : 18378 : stadistinct = 1.0; /* only 1 value */
8113 tgl@sss.pgh.pa.us 6670 : 18378 : break;
6671 : 385618 : default:
7919 bruce@momjian.us 6672 : 385618 : stadistinct = 0.0; /* means "unknown" */
8113 tgl@sss.pgh.pa.us 6673 : 385618 : break;
6674 : : }
6675 : : }
6676 : : else
7919 bruce@momjian.us 6677 : 30398 : stadistinct = 0.0; /* means "unknown" */
6678 : :
6679 : : /*
6680 : : * XXX consider using estimate_num_groups on expressions?
6681 : : */
6682 : : }
6683 : :
6684 : : /*
6685 : : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6686 : : * variable, assume it is unique no matter what pg_statistic says; the
6687 : : * statistics could be out of date, or we might have found a partial
6688 : : * unique index that proves the var is unique for this query. However,
6689 : : * we'd better still believe the null-fraction statistic.
6690 : : */
6288 tgl@sss.pgh.pa.us 6691 [ + + ]: 1299400 : if (vardata->isunique)
3558 6692 : 312149 : stadistinct = -1.0 * (1.0 - stanullfrac);
6693 : :
6694 : : /*
6695 : : * If we had an absolute estimate, use that.
6696 : : */
8113 6697 [ + + ]: 1299400 : if (stadistinct > 0.0)
3932 6698 : 304817 : return clamp_row_est(stadistinct);
6699 : :
6700 : : /*
6701 : : * Otherwise we need to get the relation size; punt if not available.
6702 : : */
8113 6703 [ + + ]: 994583 : if (vardata->rel == NULL)
6704 : : {
5357 6705 : 591 : *isdefault = true;
8113 6706 : 591 : return DEFAULT_NUM_DISTINCT;
6707 : : }
6708 : 993992 : ntuples = vardata->rel->tuples;
6709 [ + + ]: 993992 : if (ntuples <= 0.0)
6710 : : {
5357 6711 : 110950 : *isdefault = true;
8113 6712 : 110950 : return DEFAULT_NUM_DISTINCT;
6713 : : }
6714 : :
6715 : : /*
6716 : : * If we had a relative estimate, use that.
6717 : : */
6718 [ + + ]: 883042 : if (stadistinct < 0.0)
3932 6719 : 619692 : return clamp_row_est(-stadistinct * ntuples);
6720 : :
6721 : : /*
6722 : : * With no data, estimate ndistinct = ntuples if the table is small, else
6723 : : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6724 : : * that the behavior isn't discontinuous.
6725 : : */
8113 6726 [ + + ]: 263350 : if (ntuples < DEFAULT_NUM_DISTINCT)
3932 6727 : 124223 : return clamp_row_est(ntuples);
6728 : :
5357 6729 : 139127 : *isdefault = true;
8113 6730 : 139127 : return DEFAULT_NUM_DISTINCT;
6731 : : }
6732 : :
6733 : : /*
6734 : : * get_variable_range
6735 : : * Estimate the minimum and maximum value of the specified variable.
6736 : : * If successful, store values in *min and *max, and return true.
6737 : : * If no data available, return false.
6738 : : *
6739 : : * sortop is the "<" comparison operator to use. This should generally
6740 : : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6741 : : * The collation must be specified too.
6742 : : */
6743 : : static bool
2160 6744 : 174882 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6745 : : Oid sortop, Oid collation,
6746 : : Datum *min, Datum *max)
6747 : : {
6723 6748 : 174882 : Datum tmin = 0;
8113 6749 : 174882 : Datum tmax = 0;
6723 6750 : 174882 : bool have_data = false;
6751 : : int16 typLen;
6752 : : bool typByVal;
6753 : : Oid opfuncoid;
6754 : : FmgrInfo opproc;
6755 : : AttStatsSlot sslot;
6756 : :
6757 : : /*
6758 : : * XXX It's very tempting to try to use the actual column min and max, if
6759 : : * we can get them relatively-cheaply with an index probe. However, since
6760 : : * this function is called many times during join planning, that could
6761 : : * have unpleasant effects on planning speed. Need more investigation
6762 : : * before enabling this.
6763 : : */
6764 : : #ifdef NOT_USED
6765 : : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6766 : : return true;
6767 : : #endif
6768 : :
8113 6769 [ + + ]: 174882 : if (!HeapTupleIsValid(vardata->statsTuple))
6770 : : {
6771 : : /* no stats available, so default result */
6772 : 44343 : return false;
6773 : : }
6774 : :
6775 : : /*
6776 : : * If we can't apply the sortop to the stats data, just fail. In
6777 : : * principle, if there's a histogram and no MCVs, we could return the
6778 : : * histogram endpoints without ever applying the sortop ... but it's
6779 : : * probably not worth trying, because whatever the caller wants to do with
6780 : : * the endpoints would likely fail the security check too.
6781 : : */
3287 peter_e@gmx.net 6782 [ - + ]: 130539 : if (!statistic_proc_security_check(vardata,
6783 : 130539 : (opfuncoid = get_opcode(sortop))))
3287 peter_e@gmx.net 6784 :UBC 0 : return false;
6785 : :
2160 tgl@sss.pgh.pa.us 6786 :CBC 130539 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6787 : :
8113 6788 : 130539 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6789 : :
6790 : : /*
6791 : : * If there is a histogram with the ordering we want, grab the first and
6792 : : * last values.
6793 : : */
3279 6794 [ + + ]: 130539 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6795 : : STATISTIC_KIND_HISTOGRAM, sortop,
6796 : : ATTSTATSSLOT_VALUES))
6797 : : {
2160 6798 [ + - + - ]: 82940 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6799 : : {
3279 6800 : 82940 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6801 : 82940 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6723 6802 : 82940 : have_data = true;
6803 : : }
3279 6804 : 82940 : free_attstatsslot(&sslot);
6805 : : }
6806 : :
6807 : : /*
6808 : : * Otherwise, if there is a histogram with some other ordering, scan it
6809 : : * and get the min and max values according to the ordering we want. This
6810 : : * of course may not find values that are really extremal according to our
6811 : : * ordering, but it beats ignoring available data.
6812 : : */
2160 6813 [ + + - + ]: 178138 : if (!have_data &&
6814 : 47599 : get_attstatsslot(&sslot, vardata->statsTuple,
6815 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6816 : : ATTSTATSSLOT_VALUES))
6817 : : {
2160 tgl@sss.pgh.pa.us 6818 :UBC 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6819 : : collation, typLen, typByVal,
6820 : : &tmin, &tmax, &have_data);
3279 6821 : 0 : free_attstatsslot(&sslot);
6822 : : }
6823 : :
6824 : : /*
6825 : : * If we have most-common-values info, look for extreme MCVs. This is
6826 : : * needed even if we also have a histogram, since the histogram excludes
6827 : : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6828 : : * be pretty wary of deciding that that is a full representation of the
6829 : : * data. Proceed only if the MCVs represent the whole table (to within
6830 : : * roundoff error).
6831 : : */
3279 tgl@sss.pgh.pa.us 6832 [ + + ]:CBC 130539 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6833 : : STATISTIC_KIND_MCV, InvalidOid,
1677 6834 [ + + ]: 130539 : have_data ? ATTSTATSSLOT_VALUES :
6835 : : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6836 : : {
6837 : 75093 : bool use_mcvs = have_data;
6838 : :
6839 [ + + ]: 75093 : if (!have_data)
6840 : : {
6841 : 46660 : double sumcommon = 0.0;
6842 : : double nullfrac;
6843 : : int i;
6844 : :
6845 [ + + ]: 370145 : for (i = 0; i < sslot.nnumbers; i++)
6846 : 323485 : sumcommon += sslot.numbers[i];
6847 : 46660 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6848 [ + + ]: 46660 : if (sumcommon + nullfrac > 0.99999)
6849 : 43182 : use_mcvs = true;
6850 : : }
6851 : :
6852 [ + + ]: 75093 : if (use_mcvs)
6853 : 71615 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6854 : : collation, typLen, typByVal,
6855 : : &tmin, &tmax, &have_data);
3279 6856 : 75093 : free_attstatsslot(&sslot);
6857 : : }
6858 : :
6723 6859 : 130539 : *min = tmin;
8113 6860 : 130539 : *max = tmax;
6723 6861 : 130539 : return have_data;
6862 : : }
6863 : :
6864 : : /*
6865 : : * get_stats_slot_range: scan sslot for min/max values
6866 : : *
6867 : : * Subroutine for get_variable_range: update min/max/have_data according
6868 : : * to what we find in the statistics array.
6869 : : */
6870 : : static void
2160 6871 : 71615 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6872 : : Oid collation, int16 typLen, bool typByVal,
6873 : : Datum *min, Datum *max, bool *p_have_data)
6874 : : {
6875 : 71615 : Datum tmin = *min;
6876 : 71615 : Datum tmax = *max;
6877 : 71615 : bool have_data = *p_have_data;
6878 : 71615 : bool found_tmin = false;
6879 : 71615 : bool found_tmax = false;
6880 : :
6881 : : /* Look up the comparison function, if we didn't already do so */
6882 [ + - ]: 71615 : if (opproc->fn_oid != opfuncoid)
6883 : 71615 : fmgr_info(opfuncoid, opproc);
6884 : :
6885 : : /* Scan all the slot's values */
6886 [ + + ]: 1829695 : for (int i = 0; i < sslot->nvalues; i++)
6887 : : {
6888 [ + + ]: 1758080 : if (!have_data)
6889 : : {
6890 : 43182 : tmin = tmax = sslot->values[i];
6891 : 43182 : found_tmin = found_tmax = true;
6892 : 43182 : *p_have_data = have_data = true;
6893 : 43182 : continue;
6894 : : }
6895 [ + + ]: 1714898 : if (DatumGetBool(FunctionCall2Coll(opproc,
6896 : : collation,
6897 : 1714898 : sslot->values[i], tmin)))
6898 : : {
6899 : 44829 : tmin = sslot->values[i];
6900 : 44829 : found_tmin = true;
6901 : : }
6902 [ + + ]: 1714898 : if (DatumGetBool(FunctionCall2Coll(opproc,
6903 : : collation,
6904 : 1714898 : tmax, sslot->values[i])))
6905 : : {
6906 : 195365 : tmax = sslot->values[i];
6907 : 195365 : found_tmax = true;
6908 : : }
6909 : : }
6910 : :
6911 : : /*
6912 : : * Copy the slot's values, if we found new extreme values.
6913 : : */
6914 [ + + ]: 71615 : if (found_tmin)
6915 : 61296 : *min = datumCopy(tmin, typByVal, typLen);
6916 [ + + ]: 71615 : if (found_tmax)
6917 : 47651 : *max = datumCopy(tmax, typByVal, typLen);
6918 : 71615 : }
6919 : :
6920 : :
6921 : : /*
6922 : : * get_actual_variable_range
6923 : : * Attempt to identify the current *actual* minimum and/or maximum
6924 : : * of the specified variable, by looking for a suitable btree index
6925 : : * and fetching its low and/or high values.
6926 : : * If successful, store values in *min and *max, and return true.
6927 : : * (Either pointer can be NULL if that endpoint isn't needed.)
6928 : : * If unsuccessful, return false.
6929 : : *
6930 : : * sortop is the "<" comparison operator to use.
6931 : : * collation is the required collation.
6932 : : */
6933 : : static bool
5965 6934 : 125141 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6935 : : Oid sortop, Oid collation,
6936 : : Datum *min, Datum *max)
6937 : : {
6938 : 125141 : bool have_data = false;
6939 : 125141 : RelOptInfo *rel = vardata->rel;
6940 : : RangeTblEntry *rte;
6941 : : ListCell *lc;
6942 : :
6943 : : /* No hope if no relation or it doesn't have indexes */
6944 [ + - + + ]: 125141 : if (rel == NULL || rel->indexlist == NIL)
6945 : 8975 : return false;
6946 : : /* If it has indexes it must be a plain relation */
6947 : 116166 : rte = root->simple_rte_array[rel->relid];
6948 [ - + ]: 116166 : Assert(rte->rtekind == RTE_RELATION);
6949 : :
6950 : : /* ignore partitioned tables. Any indexes here are not real indexes */
1212 drowley@postgresql.o 6951 [ + + ]: 116166 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6952 : 560 : return false;
6953 : :
6954 : : /* Search through the indexes to see if any match our problem */
5965 tgl@sss.pgh.pa.us 6955 [ + - + + : 234355 : foreach(lc, rel->indexlist)
+ + ]
6956 : : {
6957 : 201612 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6958 : : ScanDirection indexscandir;
6959 : : StrategyNumber strategy;
6960 : :
6961 : : /* Ignore non-ordering indexes */
396 peter@eisentraut.org 6962 [ + + ]: 201612 : if (index->sortopfamily == NULL)
396 peter@eisentraut.org 6963 :GBC 3 : continue;
6964 : :
6965 : : /*
6966 : : * Ignore partial indexes --- we only want stats that cover the entire
6967 : : * relation.
6968 : : */
5965 tgl@sss.pgh.pa.us 6969 [ + + ]:CBC 201609 : if (index->indpred != NIL)
6970 : 240 : continue;
6971 : :
6972 : : /*
6973 : : * The index list might include hypothetical indexes inserted by a
6974 : : * get_relation_info hook --- don't try to access them.
6975 : : */
5557 6976 [ - + ]: 201369 : if (index->hypothetical)
5965 tgl@sss.pgh.pa.us 6977 :UBC 0 : continue;
6978 : :
6979 : : /*
6980 : : * get_actual_variable_endpoint uses the index-only-scan machinery, so
6981 : : * ignore indexes that can't use it on their first column.
6982 : : */
189 peter@eisentraut.org 6983 [ - + ]:CBC 201369 : if (!index->canreturn[0])
189 peter@eisentraut.org 6984 :UBC 0 : continue;
6985 : :
6986 : : /*
6987 : : * The first index column must match the desired variable, sortop, and
6988 : : * collation --- but we can use a descending-order index.
6989 : : */
2160 tgl@sss.pgh.pa.us 6990 [ + + ]:CBC 201369 : if (collation != index->indexcollations[0])
6991 : 28984 : continue; /* test first 'cause it's cheapest */
5965 6992 [ + + ]: 172385 : if (!match_index_to_operand(vardata->var, 0, index))
6993 : 89522 : continue;
396 peter@eisentraut.org 6994 : 82863 : strategy = get_op_opfamily_strategy(sortop, index->sortopfamily[0]);
6995 [ + - - ]: 82863 : switch (IndexAmTranslateStrategy(strategy, index->relam, index->sortopfamily[0], true))
6996 : : {
6997 : 82863 : case COMPARE_LT:
5636 tgl@sss.pgh.pa.us 6998 [ - + ]: 82863 : if (index->reverse_sort[0])
5636 tgl@sss.pgh.pa.us 6999 :UBC 0 : indexscandir = BackwardScanDirection;
7000 : : else
5636 tgl@sss.pgh.pa.us 7001 :CBC 82863 : indexscandir = ForwardScanDirection;
7002 : 82863 : break;
396 peter@eisentraut.org 7003 :UBC 0 : case COMPARE_GT:
5636 tgl@sss.pgh.pa.us 7004 [ # # ]: 0 : if (index->reverse_sort[0])
7005 : 0 : indexscandir = ForwardScanDirection;
7006 : : else
7007 : 0 : indexscandir = BackwardScanDirection;
7008 : 0 : break;
7009 : 0 : default:
7010 : : /* index doesn't match the sortop */
7011 : 0 : continue;
7012 : : }
7013 : :
7014 : : /*
7015 : : * Found a suitable index to extract data from. Set up some data that
7016 : : * can be used by both invocations of get_actual_variable_endpoint.
7017 : : */
7018 : : {
7019 : : MemoryContext tmpcontext;
7020 : : MemoryContext oldcontext;
7021 : : Relation heapRel;
7022 : : Relation indexRel;
7023 : : TupleTableSlot *slot;
7024 : : int16 typLen;
7025 : : bool typByVal;
7026 : : ScanKeyData scankeys[1];
7027 : :
7028 : : /* Make sure any cruft gets recycled when we're done */
2489 tgl@sss.pgh.pa.us 7029 :CBC 82863 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
7030 : : "get_actual_variable_range workspace",
7031 : : ALLOCSET_DEFAULT_SIZES);
5965 7032 : 82863 : oldcontext = MemoryContextSwitchTo(tmpcontext);
7033 : :
7034 : : /*
7035 : : * Open the table and index so we can read from them. We should
7036 : : * already have some type of lock on each.
7037 : : */
2661 andres@anarazel.de 7038 : 82863 : heapRel = table_open(rte->relid, NoLock);
2588 tgl@sss.pgh.pa.us 7039 : 82863 : indexRel = index_open(index->indexoid, NoLock);
7040 : :
7041 : : /* build some stuff needed for indexscan execution */
2612 andres@anarazel.de 7042 : 82863 : slot = table_slot_create(heapRel, NULL);
5965 tgl@sss.pgh.pa.us 7043 : 82863 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
7044 : :
7045 : : /* set up an IS NOT NULL scan key so that we ignore nulls */
7046 : 82863 : ScanKeyEntryInitialize(&scankeys[0],
7047 : : SK_ISNULL | SK_SEARCHNOTNULL,
7048 : : 1, /* index col to scan */
7049 : : InvalidStrategy, /* no strategy */
7050 : : InvalidOid, /* no strategy subtype */
7051 : : InvalidOid, /* no collation */
7052 : : InvalidOid, /* no reg proc for this */
7053 : : (Datum) 0); /* constant */
7054 : :
7055 : : /* If min is requested ... */
7056 [ + + ]: 82863 : if (min)
7057 : : {
2489 7058 : 45771 : have_data = get_actual_variable_endpoint(heapRel,
7059 : : indexRel,
7060 : : indexscandir,
7061 : : scankeys,
7062 : : typLen,
7063 : : typByVal,
7064 : : slot,
7065 : : oldcontext,
7066 : : min);
7067 : : }
7068 : : else
7069 : : {
7070 : : /* If min not requested, still want to fetch max */
7071 : 37092 : have_data = true;
7072 : : }
7073 : :
7074 : : /* If max is requested, and we didn't already fail ... */
5965 7075 [ + + + - ]: 82863 : if (max && have_data)
7076 : : {
7077 : : /* scan in the opposite direction; all else is the same */
2489 7078 : 38210 : have_data = get_actual_variable_endpoint(heapRel,
7079 : : indexRel,
7080 : 38210 : -indexscandir,
7081 : : scankeys,
7082 : : typLen,
7083 : : typByVal,
7084 : : slot,
7085 : : oldcontext,
7086 : : max);
7087 : : }
7088 : :
7089 : : /* Clean everything up */
5965 7090 : 82863 : ExecDropSingleTupleTableSlot(slot);
7091 : :
2588 7092 : 82863 : index_close(indexRel, NoLock);
2661 andres@anarazel.de 7093 : 82863 : table_close(heapRel, NoLock);
7094 : :
5965 tgl@sss.pgh.pa.us 7095 : 82863 : MemoryContextSwitchTo(oldcontext);
2489 7096 : 82863 : MemoryContextDelete(tmpcontext);
7097 : :
7098 : : /* And we're done */
5965 7099 : 82863 : break;
7100 : : }
7101 : : }
7102 : :
7103 : 115606 : return have_data;
7104 : : }
7105 : :
7106 : : /*
7107 : : * Get one endpoint datum (min or max depending on indexscandir) from the
7108 : : * specified index. Return true if successful, false if not.
7109 : : * On success, endpoint value is stored to *endpointDatum (and copied into
7110 : : * outercontext).
7111 : : *
7112 : : * scankeys is a 1-element scankey array set up to reject nulls.
7113 : : * typLen/typByVal describe the datatype of the index's first column.
7114 : : * tableslot is a slot suitable to hold table tuples, in case we need
7115 : : * to probe the heap.
7116 : : * (We could compute these values locally, but that would mean computing them
7117 : : * twice when get_actual_variable_range needs both the min and the max.)
7118 : : *
7119 : : * Failure occurs either when the index is empty, or we decide that it's
7120 : : * taking too long to find a suitable tuple.
7121 : : */
7122 : : static bool
2489 7123 : 83981 : get_actual_variable_endpoint(Relation heapRel,
7124 : : Relation indexRel,
7125 : : ScanDirection indexscandir,
7126 : : ScanKey scankeys,
7127 : : int16 typLen,
7128 : : bool typByVal,
7129 : : TupleTableSlot *tableslot,
7130 : : MemoryContext outercontext,
7131 : : Datum *endpointDatum)
7132 : : {
7133 : 83981 : bool have_data = false;
7134 : : SnapshotData SnapshotNonVacuumable;
7135 : : IndexScanDesc index_scan;
7136 : 83981 : Buffer vmbuffer = InvalidBuffer;
1260 7137 : 83981 : BlockNumber last_heap_block = InvalidBlockNumber;
7138 : 83981 : int n_visited_heap_pages = 0;
7139 : : ItemPointer tid;
7140 : : Datum values[INDEX_MAX_KEYS];
7141 : : bool isnull[INDEX_MAX_KEYS];
7142 : : MemoryContext oldcontext;
7143 : :
7144 : : /*
7145 : : * We use the index-only-scan machinery for this. With mostly-static
7146 : : * tables that's a win because it avoids a heap visit. It's also a win
7147 : : * for dynamic data, but the reason is less obvious; read on for details.
7148 : : *
7149 : : * In principle, we should scan the index with our current active
7150 : : * snapshot, which is the best approximation we've got to what the query
7151 : : * will see when executed. But that won't be exact if a new snap is taken
7152 : : * before running the query, and it can be very expensive if a lot of
7153 : : * recently-dead or uncommitted rows exist at the beginning or end of the
7154 : : * index (because we'll laboriously fetch each one and reject it).
7155 : : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
7156 : : * and uncommitted rows as well as normal visible rows. On the other
7157 : : * hand, it will reject known-dead rows, and thus not give a bogus answer
7158 : : * when the extreme value has been deleted (unless the deletion was quite
7159 : : * recent); that case motivates not using SnapshotAny here.
7160 : : *
7161 : : * A crucial point here is that SnapshotNonVacuumable, with
7162 : : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
7163 : : * condition that the indexscan will use to decide that index entries are
7164 : : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
7165 : : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
7166 : : * have to continue scanning past it, we know that the indexscan will mark
7167 : : * that index entry killed. That means that the next
7168 : : * get_actual_variable_endpoint() call will not have to re-consider that
7169 : : * index entry. In this way we avoid repetitive work when this function
7170 : : * is used a lot during planning.
7171 : : *
7172 : : * But using SnapshotNonVacuumable creates a hazard of its own. In a
7173 : : * recently-created index, some index entries may point at "broken" HOT
7174 : : * chains in which not all the tuple versions contain data matching the
7175 : : * index entry. The live tuple version(s) certainly do match the index,
7176 : : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
7177 : : * don't match. Hence, if we took data from the selected heap tuple, we
7178 : : * might get a bogus answer that's not close to the index extremal value,
7179 : : * or could even be NULL. We avoid this hazard because we take the data
7180 : : * from the index entry not the heap.
7181 : : *
7182 : : * Despite all this care, there are situations where we might find many
7183 : : * non-visible tuples near the end of the index. We don't want to expend
7184 : : * a huge amount of time here, so we give up once we've read too many heap
7185 : : * pages. When we fail for that reason, the caller will end up using
7186 : : * whatever extremal value is recorded in pg_statistic.
7187 : : */
2092 andres@anarazel.de 7188 : 83981 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
7189 : : GlobalVisTestFor(heapRel));
7190 : :
2489 tgl@sss.pgh.pa.us 7191 : 83981 : index_scan = index_beginscan(heapRel, indexRel,
7192 : : &SnapshotNonVacuumable, NULL,
7193 : : 1, 0,
7194 : : SO_NONE);
7195 : : /* Set it up for index-only scan */
7196 : 83981 : index_scan->xs_want_itup = true;
7197 : 83981 : index_rescan(index_scan, scankeys, 1, NULL, 0);
7198 : :
7199 : : /* Fetch first/next tuple in specified direction */
7200 [ + - ]: 102011 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
7201 : : {
1260 7202 : 102011 : BlockNumber block = ItemPointerGetBlockNumber(tid);
7203 : :
2489 7204 [ + + ]: 102011 : if (!VM_ALL_VISIBLE(heapRel,
7205 : : block,
7206 : : &vmbuffer))
7207 : : {
7208 : : /* Rats, we have to visit the heap to check visibility */
7209 [ + + ]: 69929 : if (!index_fetch_heap(index_scan, tableslot))
7210 : : {
7211 : : /*
7212 : : * No visible tuple for this index entry, so we need to
7213 : : * advance to the next entry. Before doing so, count heap
7214 : : * page fetches and give up if we've done too many.
7215 : : *
7216 : : * We don't charge a page fetch if this is the same heap page
7217 : : * as the previous tuple. This is on the conservative side,
7218 : : * since other recently-accessed pages are probably still in
7219 : : * buffers too; but it's good enough for this heuristic.
7220 : : */
7221 : : #define VISITED_PAGES_LIMIT 100
7222 : :
1260 7223 [ + + ]: 18030 : if (block != last_heap_block)
7224 : : {
7225 : 1825 : last_heap_block = block;
7226 : 1825 : n_visited_heap_pages++;
7227 [ - + ]: 1825 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
1260 tgl@sss.pgh.pa.us 7228 :UBC 0 : break;
7229 : : }
7230 : :
2489 tgl@sss.pgh.pa.us 7231 :CBC 18030 : continue; /* no visible tuple, try next index entry */
7232 : : }
7233 : :
7234 : : /* We don't actually need the heap tuple for anything */
7235 : 51899 : ExecClearTuple(tableslot);
7236 : :
7237 : : /*
7238 : : * We don't care whether there's more than one visible tuple in
7239 : : * the HOT chain; if any are visible, that's good enough.
7240 : : */
7241 : : }
7242 : :
7243 : : /*
7244 : : * We expect that the index will return data in IndexTuple not
7245 : : * HeapTuple format.
7246 : : */
7247 [ - + ]: 83981 : if (!index_scan->xs_itup)
2489 tgl@sss.pgh.pa.us 7248 [ # # ]:UBC 0 : elog(ERROR, "no data returned for index-only scan");
7249 : :
7250 : : /*
7251 : : * We do not yet support recheck here.
7252 : : */
2489 tgl@sss.pgh.pa.us 7253 [ - + ]:CBC 83981 : if (index_scan->xs_recheck)
396 peter@eisentraut.org 7254 :UBC 0 : break;
7255 : :
7256 : : /* OK to deconstruct the index tuple */
2489 tgl@sss.pgh.pa.us 7257 :CBC 83981 : index_deform_tuple(index_scan->xs_itup,
7258 : : index_scan->xs_itupdesc,
7259 : : values, isnull);
7260 : :
7261 : : /* Shouldn't have got a null, but be careful */
7262 [ - + ]: 83981 : if (isnull[0])
2489 tgl@sss.pgh.pa.us 7263 [ # # ]:UBC 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
7264 : : RelationGetRelationName(indexRel));
7265 : :
7266 : : /* Copy the index column value out to caller's context */
2489 tgl@sss.pgh.pa.us 7267 :CBC 83981 : oldcontext = MemoryContextSwitchTo(outercontext);
7268 : 83981 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
7269 : 83981 : MemoryContextSwitchTo(oldcontext);
7270 : 83981 : have_data = true;
7271 : 83981 : break;
7272 : : }
7273 : :
7274 [ + + ]: 83981 : if (vmbuffer != InvalidBuffer)
7275 : 77023 : ReleaseBuffer(vmbuffer);
7276 : 83981 : index_endscan(index_scan);
7277 : :
7278 : 83981 : return have_data;
7279 : : }
7280 : :
7281 : : /*
7282 : : * find_join_input_rel
7283 : : * Look up the input relation for a join.
7284 : : *
7285 : : * We assume that the input relation's RelOptInfo must have been constructed
7286 : : * already.
7287 : : */
7288 : : static RelOptInfo *
5361 7289 : 11394 : find_join_input_rel(PlannerInfo *root, Relids relids)
7290 : : {
7291 : 11394 : RelOptInfo *rel = NULL;
7292 : :
889 drowley@postgresql.o 7293 [ + - ]: 11394 : if (!bms_is_empty(relids))
7294 : : {
7295 : : int relid;
7296 : :
7297 [ + + ]: 11394 : if (bms_get_singleton_member(relids, &relid))
7298 : 11118 : rel = find_base_rel(root, relid);
7299 : : else
5361 tgl@sss.pgh.pa.us 7300 : 276 : rel = find_join_rel(root, relids);
7301 : : }
7302 : :
7303 [ - + ]: 11394 : if (rel == NULL)
5361 tgl@sss.pgh.pa.us 7304 [ # # ]:UBC 0 : elog(ERROR, "could not find RelOptInfo for given relids");
7305 : :
5361 tgl@sss.pgh.pa.us 7306 :CBC 11394 : return rel;
7307 : : }
7308 : :
7309 : :
7310 : : /*-------------------------------------------------------------------------
7311 : : *
7312 : : * Index cost estimation functions
7313 : : *
7314 : : *-------------------------------------------------------------------------
7315 : : */
7316 : :
7317 : : /*
7318 : : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
7319 : : */
7320 : : List *
2636 7321 : 669775 : get_quals_from_indexclauses(List *indexclauses)
7322 : : {
2642 7323 : 669775 : List *result = NIL;
7324 : : ListCell *lc;
7325 : :
7326 [ + + + + : 1163239 : foreach(lc, indexclauses)
+ + ]
7327 : : {
7328 : 493464 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7329 : : ListCell *lc2;
7330 : :
2637 7331 [ + - + + : 989351 : foreach(lc2, iclause->indexquals)
+ + ]
7332 : : {
7333 : 495887 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7334 : :
7335 : 495887 : result = lappend(result, rinfo);
7336 : : }
7337 : : }
2642 7338 : 669775 : return result;
7339 : : }
7340 : :
7341 : : /*
7342 : : * Compute the total evaluation cost of the comparison operands in a list
7343 : : * of index qual expressions. Since we know these will be evaluated just
7344 : : * once per scan, there's no need to distinguish startup from per-row cost.
7345 : : *
7346 : : * This can be used either on the result of get_quals_from_indexclauses(),
7347 : : * or directly on an indexorderbys list. In both cases, we expect that the
7348 : : * index key expression is on the left side of binary clauses.
7349 : : */
7350 : : Cost
2636 7351 : 1329050 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
7352 : : {
4081 7353 : 1329050 : Cost qual_arg_cost = 0;
7354 : : ListCell *lc;
7355 : :
2636 7356 [ + + + + : 1825286 : foreach(lc, indexquals)
+ + ]
7357 : : {
4081 7358 : 496236 : Expr *clause = (Expr *) lfirst(lc);
7359 : : Node *other_operand;
7360 : : QualCost index_qual_cost;
7361 : :
7362 : : /*
7363 : : * Index quals will have RestrictInfos, indexorderbys won't. Look
7364 : : * through RestrictInfo if present.
7365 : : */
2636 7366 [ + + ]: 496236 : if (IsA(clause, RestrictInfo))
7367 : 495877 : clause = ((RestrictInfo *) clause)->clause;
7368 : :
4081 7369 [ + + ]: 496236 : if (IsA(clause, OpExpr))
7370 : : {
2636 7371 : 480825 : OpExpr *op = (OpExpr *) clause;
7372 : :
7373 : 480825 : other_operand = (Node *) lsecond(op->args);
7374 : : }
7375 [ + + ]: 15411 : else if (IsA(clause, RowCompareExpr))
7376 : : {
7377 : 330 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7378 : :
7379 : 330 : other_operand = (Node *) rc->rargs;
7380 : : }
7381 [ + + ]: 15081 : else if (IsA(clause, ScalarArrayOpExpr))
7382 : : {
7383 : 12642 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7384 : :
7385 : 12642 : other_operand = (Node *) lsecond(saop->args);
7386 : : }
7387 [ + - ]: 2439 : else if (IsA(clause, NullTest))
7388 : : {
7389 : 2439 : other_operand = NULL;
7390 : : }
7391 : : else
7392 : : {
2636 tgl@sss.pgh.pa.us 7393 [ # # ]:UBC 0 : elog(ERROR, "unsupported indexqual type: %d",
7394 : : (int) nodeTag(clause));
7395 : : other_operand = NULL; /* keep compiler quiet */
7396 : : }
7397 : :
4081 tgl@sss.pgh.pa.us 7398 :CBC 496236 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
7399 : 496236 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
7400 : : }
7401 : 1329050 : return qual_arg_cost;
7402 : : }
7403 : :
7404 : : /*
7405 : : * Compute generic index access cost estimates.
7406 : : *
7407 : : * See struct GenericCosts in selfuncs.h for more info.
7408 : : */
7409 : : void
7639 7410 : 659285 : genericcostestimate(PlannerInfo *root,
7411 : : IndexPath *path,
7412 : : double loop_count,
7413 : : GenericCosts *costs)
7414 : : {
5246 7415 : 659285 : IndexOptInfo *index = path->indexinfo;
2636 7416 : 659285 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
5246 7417 : 659285 : List *indexOrderBys = path->indexorderbys;
7418 : : Cost indexStartupCost;
7419 : : Cost indexTotalCost;
7420 : : Selectivity indexSelectivity;
7421 : : double indexCorrelation;
7422 : : double numIndexPages;
7423 : : double numIndexTuples;
7424 : : double spc_random_page_cost;
7425 : : double num_sa_scans;
7426 : : double num_outer_scans;
7427 : : double num_scans;
7428 : : double qual_op_cost;
7429 : : double qual_arg_cost;
7430 : : List *selectivityQuals;
7431 : : ListCell *l;
7432 : :
7433 : : /*
7434 : : * If the index is partial, AND the index predicate with the explicitly
7435 : : * given indexquals to produce a more accurate idea of the index
7436 : : * selectivity.
7437 : : */
2636 7438 : 659285 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7439 : :
7440 : : /*
7441 : : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
7442 : : * just assume that the number of index descents is the number of distinct
7443 : : * combinations of array elements from all of the scan's SAOP clauses.
7444 : : */
759 pg@bowt.ie 7445 : 659285 : num_sa_scans = costs->num_sa_scans;
7446 [ + + ]: 659285 : if (num_sa_scans < 1)
7447 : : {
7448 : 6942 : num_sa_scans = 1;
7449 [ + + + + : 15192 : foreach(l, indexQuals)
+ + ]
7450 : : {
7451 : 8250 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
7452 : :
7453 [ + + ]: 8250 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
7454 : : {
7455 : 46 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
7456 : 46 : double alength = estimate_array_length(root, lsecond(saop->args));
7457 : :
7458 [ + - ]: 46 : if (alength > 1)
7459 : 46 : num_sa_scans *= alength;
7460 : : }
7461 : : }
7462 : : }
7463 : :
7464 : : /* Estimate the fraction of main-table tuples that will be visited */
4862 tgl@sss.pgh.pa.us 7465 : 659285 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7466 : 659285 : index->rel->relid,
7467 : : JOIN_INNER,
7468 : : NULL);
7469 : :
7470 : : /*
7471 : : * If caller didn't give us an estimate, estimate the number of index
7472 : : * tuples that will be visited. We do it in this rather peculiar-looking
7473 : : * way in order to get the right answer for partial indexes.
7474 : : */
7475 : 659285 : numIndexTuples = costs->numIndexTuples;
7631 7476 [ + + ]: 659285 : if (numIndexTuples <= 0.0)
7477 : : {
4862 7478 : 87238 : numIndexTuples = indexSelectivity * index->rel->tuples;
7479 : :
7480 : : /*
7481 : : * The above calculation counts all the tuples visited across all
7482 : : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
7483 : : * average per-indexscan number, so adjust. This is a handy place to
7484 : : * round to integer, too. (If caller supplied tuple estimate, it's
7485 : : * responsible for handling these considerations.)
7486 : : */
7081 7487 : 87238 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7488 : : }
7489 : :
7490 : : /*
7491 : : * We can bound the number of tuples by the index size in any case. Also,
7492 : : * always estimate at least one tuple is touched, even when
7493 : : * indexSelectivity estimate is tiny.
7494 : : */
7631 7495 [ + + ]: 659285 : if (numIndexTuples > index->tuples)
7496 : 8440 : numIndexTuples = index->tuples;
9522 7497 [ + + ]: 659285 : if (numIndexTuples < 1.0)
7498 : 90862 : numIndexTuples = 1.0;
7499 : :
7500 : : /*
7501 : : * Estimate the number of index pages that will be retrieved.
7502 : : *
7503 : : * We use the simplistic method of taking a pro-rata fraction of the total
7504 : : * number of index leaf pages. We disregard any overhead such as index
7505 : : * metapages or upper tree levels.
7506 : : *
7507 : : * In practice access to upper index levels is often nearly free because
7508 : : * those tend to stay in cache under load; moreover, the cost involved is
7509 : : * highly dependent on index type. We therefore ignore such costs here
7510 : : * and leave it to the caller to add a suitable charge if needed.
7511 : : */
46 tgl@sss.pgh.pa.us 7512 [ + + + + ]:GNC 659285 : if (index->pages > costs->numNonLeafPages && index->tuples > 1)
7513 : 592624 : numIndexPages =
7514 : 592624 : ceil(numIndexTuples * (index->pages - costs->numNonLeafPages)
7515 : 592624 : / index->tuples);
7516 : : else
9522 tgl@sss.pgh.pa.us 7517 :CBC 66661 : numIndexPages = 1.0;
7518 : :
7519 : : /* fetch estimated page cost for tablespace containing index */
5964 rhaas@postgresql.org 7520 : 659285 : get_tablespace_page_costs(index->reltablespace,
7521 : : &spc_random_page_cost,
7522 : : NULL);
7523 : :
7524 : : /*
7525 : : * Now compute the disk access costs.
7526 : : *
7527 : : * The above calculations are all per-index-scan. However, if we are in a
7528 : : * nestloop inner scan, we can expect the scan to be repeated (with
7529 : : * different search keys) for each row of the outer relation. Likewise,
7530 : : * ScalarArrayOpExpr quals result in multiple index scans. This creates
7531 : : * the potential for cache effects to reduce the number of disk page
7532 : : * fetches needed. We want to estimate the average per-scan I/O cost in
7533 : : * the presence of caching.
7534 : : *
7535 : : * We use the Mackert-Lohman formula (see costsize.c for details) to
7536 : : * estimate the total number of page fetches that occur. While this
7537 : : * wasn't what it was designed for, it seems a reasonable model anyway.
7538 : : * Note that we are counting pages not tuples anymore, so we take N = T =
7539 : : * index size, as if there were one "tuple" per page.
7540 : : */
5212 tgl@sss.pgh.pa.us 7541 : 659285 : num_outer_scans = loop_count;
7542 : 659285 : num_scans = num_sa_scans * num_outer_scans;
7543 : :
7248 7544 [ + + ]: 659285 : if (num_scans > 1)
7545 : : {
7546 : : double pages_fetched;
7547 : :
7548 : : /* total page fetches ignoring cache effects */
7273 7549 : 79869 : pages_fetched = numIndexPages * num_scans;
7550 : :
7551 : : /* use Mackert and Lohman formula to adjust for cache effects */
7552 : 79869 : pages_fetched = index_pages_fetched(pages_fetched,
7553 : : index->pages,
7168 7554 : 79869 : (double) index->pages,
7555 : : root);
7556 : :
7557 : : /*
7558 : : * Now compute the total disk access cost, and then report a pro-rated
7559 : : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
7560 : : * since that's internal to the indexscan.)
7561 : : */
4862 7562 : 79869 : indexTotalCost = (pages_fetched * spc_random_page_cost)
7563 : : / num_outer_scans;
7564 : : }
7565 : : else
7566 : : {
7567 : : /*
7568 : : * For a single index scan, we just charge spc_random_page_cost per
7569 : : * page touched.
7570 : : */
7571 : 579416 : indexTotalCost = numIndexPages * spc_random_page_cost;
7572 : : }
7573 : :
7574 : : /*
7575 : : * CPU cost: any complex expressions in the indexquals will need to be
7576 : : * evaluated once at the start of the scan to reduce them to runtime keys
7577 : : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
7578 : : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
7579 : : * indexqual operator. Because we have numIndexTuples as a per-scan
7580 : : * number, we have to multiply by num_sa_scans to get the correct result
7581 : : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
7582 : : * ORDER BY expressions.
7583 : : *
7584 : : * Note: this neglects the possible costs of rechecking lossy operators.
7585 : : * Detecting that that might be needed seems more expensive than it's
7586 : : * worth, though, considering all the other inaccuracies here ...
7587 : : */
2636 7588 : 659285 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
7589 : 659285 : index_other_operands_eval_cost(root, indexOrderBys);
5633 7590 : 659285 : qual_op_cost = cpu_operator_cost *
7591 : 659285 : (list_length(indexQuals) + list_length(indexOrderBys));
7592 : :
4862 7593 : 659285 : indexStartupCost = qual_arg_cost;
7594 : 659285 : indexTotalCost += qual_arg_cost;
7595 : 659285 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
7596 : :
7597 : : /*
7598 : : * Generic assumption about index correlation: there isn't any.
7599 : : */
7600 : 659285 : indexCorrelation = 0.0;
7601 : :
7602 : : /*
7603 : : * Return everything to caller.
7604 : : */
7605 : 659285 : costs->indexStartupCost = indexStartupCost;
7606 : 659285 : costs->indexTotalCost = indexTotalCost;
7607 : 659285 : costs->indexSelectivity = indexSelectivity;
7608 : 659285 : costs->indexCorrelation = indexCorrelation;
7609 : 659285 : costs->numIndexPages = numIndexPages;
7610 : 659285 : costs->numIndexTuples = numIndexTuples;
7611 : 659285 : costs->spc_random_page_cost = spc_random_page_cost;
7612 : 659285 : costs->num_sa_scans = num_sa_scans;
7613 : 659285 : }
7614 : :
7615 : : /*
7616 : : * If the index is partial, add its predicate to the given qual list.
7617 : : *
7618 : : * ANDing the index predicate with the explicitly given indexquals produces
7619 : : * a more accurate idea of the index's selectivity. However, we need to be
7620 : : * careful not to insert redundant clauses, because clauselist_selectivity()
7621 : : * is easily fooled into computing a too-low selectivity estimate. Our
7622 : : * approach is to add only the predicate clause(s) that cannot be proven to
7623 : : * be implied by the given indexquals. This successfully handles cases such
7624 : : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
7625 : : * There are many other cases where we won't detect redundancy, leading to a
7626 : : * too-low selectivity estimate, which will bias the system in favor of using
7627 : : * partial indexes where possible. That is not necessarily bad though.
7628 : : *
7629 : : * Note that indexQuals contains RestrictInfo nodes while the indpred
7630 : : * does not, so the output list will be mixed. This is OK for both
7631 : : * predicate_implied_by() and clauselist_selectivity(), but might be
7632 : : * problematic if the result were passed to other things.
7633 : : */
7634 : : List *
2636 7635 : 1130220 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
7636 : : {
4862 7637 : 1130220 : List *predExtraQuals = NIL;
7638 : : ListCell *lc;
7639 : :
7640 [ + + ]: 1130220 : if (index->indpred == NIL)
7641 : 1128683 : return indexQuals;
7642 : :
7643 [ + - + + : 3084 : foreach(lc, index->indpred)
+ + ]
7644 : : {
7645 : 1547 : Node *predQual = (Node *) lfirst(lc);
7646 : 1547 : List *oneQual = list_make1(predQual);
7647 : :
3247 rhaas@postgresql.org 7648 [ + + ]: 1547 : if (!predicate_implied_by(oneQual, indexQuals, false))
4862 tgl@sss.pgh.pa.us 7649 : 1370 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7650 : : }
7651 : 1537 : return list_concat(predExtraQuals, indexQuals);
7652 : : }
7653 : :
7654 : : /*
7655 : : * Estimate correlation of btree index's first column.
7656 : : *
7657 : : * If we can get an estimate of the first column's ordering correlation C
7658 : : * from pg_statistic, estimate the index correlation as C for a single-column
7659 : : * index, or C * 0.75 for multiple columns. The idea here is that multiple
7660 : : * columns dilute the importance of the first column's ordering, but don't
7661 : : * negate it entirely.
7662 : : *
7663 : : * We already filled in the stats tuple for *vardata when called.
7664 : : */
7665 : : static double
396 pg@bowt.ie 7666 : 464930 : btcost_correlation(IndexOptInfo *index, VariableStatData *vardata)
7667 : : {
7668 : : Oid sortop;
7669 : : AttStatsSlot sslot;
7670 : 464930 : double indexCorrelation = 0;
7671 : :
7672 [ - + ]: 464930 : Assert(HeapTupleIsValid(vardata->statsTuple));
7673 : :
7674 : 464930 : sortop = get_opfamily_member(index->opfamily[0],
7675 : 464930 : index->opcintype[0],
7676 : 464930 : index->opcintype[0],
7677 : : BTLessStrategyNumber);
7678 [ + - + + ]: 929860 : if (OidIsValid(sortop) &&
7679 : 464930 : get_attstatsslot(&sslot, vardata->statsTuple,
7680 : : STATISTIC_KIND_CORRELATION, sortop,
7681 : : ATTSTATSSLOT_NUMBERS))
7682 : : {
7683 : : double varCorrelation;
7684 : :
7685 [ - + ]: 459360 : Assert(sslot.nnumbers == 1);
7686 : 459360 : varCorrelation = sslot.numbers[0];
7687 : :
7688 [ - + ]: 459360 : if (index->reverse_sort[0])
396 pg@bowt.ie 7689 :UBC 0 : varCorrelation = -varCorrelation;
7690 : :
396 pg@bowt.ie 7691 [ + + ]:CBC 459360 : if (index->nkeycolumns > 1)
7692 : 161739 : indexCorrelation = varCorrelation * 0.75;
7693 : : else
7694 : 297621 : indexCorrelation = varCorrelation;
7695 : :
7696 : 459360 : free_attstatsslot(&sslot);
7697 : : }
7698 : :
7699 : 464930 : return indexCorrelation;
7700 : : }
7701 : :
7702 : : void
3761 tgl@sss.pgh.pa.us 7703 : 652343 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7704 : : Cost *indexStartupCost, Cost *indexTotalCost,
7705 : : Selectivity *indexSelectivity, double *indexCorrelation,
7706 : : double *indexPages)
7707 : : {
5246 7708 : 652343 : IndexOptInfo *index = path->indexinfo;
1389 peter@eisentraut.org 7709 : 652343 : GenericCosts costs = {0};
7710 : 652343 : VariableStatData vardata = {0};
7711 : : double numIndexTuples;
7712 : : Cost descentCost;
7713 : : List *indexBoundQuals;
7714 : : List *indexSkipQuals;
7715 : : int indexcol;
7716 : : bool eqQualHere;
7717 : : bool found_row_compare;
7718 : : bool found_array;
7719 : : bool found_is_null_op;
396 pg@bowt.ie 7720 : 652343 : bool have_correlation = false;
7721 : : double num_sa_scans;
7722 : 652343 : double correlation = 0.0;
7723 : : ListCell *lc;
7724 : :
7725 : : /*
7726 : : * For a btree scan, only leading '=' quals plus inequality quals for the
7727 : : * immediately next attribute contribute to index selectivity (these are
7728 : : * the "boundary quals" that determine the starting and stopping points of
7729 : : * the index scan). Additional quals can suppress visits to the heap, so
7730 : : * it's OK to count them in indexSelectivity, but they should not count
7731 : : * for estimating numIndexTuples. So we must examine the given indexquals
7732 : : * to find out which ones count as boundary quals. We rely on the
7733 : : * knowledge that they are given in index column order. Note that nbtree
7734 : : * preprocessing can add skip arrays that act as leading '=' quals in the
7735 : : * absence of ordinary input '=' quals, so in practice _most_ input quals
7736 : : * are able to act as index bound quals (which we take into account here).
7737 : : *
7738 : : * For a RowCompareExpr, we consider only the first column, just as
7739 : : * rowcomparesel() does.
7740 : : *
7741 : : * If there's a SAOP or skip array in the quals, we'll actually perform up
7742 : : * to N index descents (not just one), but the underlying array key's
7743 : : * operator can be considered to act the same as it normally does.
7744 : : */
7631 tgl@sss.pgh.pa.us 7745 : 652343 : indexBoundQuals = NIL;
396 pg@bowt.ie 7746 : 652343 : indexSkipQuals = NIL;
5246 tgl@sss.pgh.pa.us 7747 : 652343 : indexcol = 0;
7631 7748 : 652343 : eqQualHere = false;
396 pg@bowt.ie 7749 : 652343 : found_row_compare = false;
7750 : 652343 : found_array = false;
5968 tgl@sss.pgh.pa.us 7751 : 652343 : found_is_null_op = false;
7081 7752 : 652343 : num_sa_scans = 1;
2636 7753 [ + + + + : 1096915 : foreach(lc, path->indexclauses)
+ + ]
7754 : : {
7755 : 472343 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7756 : : ListCell *lc2;
7757 : :
396 pg@bowt.ie 7758 [ + + ]: 472343 : if (indexcol < iclause->indexcol)
7759 : : {
7760 : 93283 : double num_sa_scans_prev_cols = num_sa_scans;
7761 : :
7762 : : /*
7763 : : * Beginning of a new column's quals.
7764 : : *
7765 : : * Skip scans use skip arrays, which are ScalarArrayOp style
7766 : : * arrays that generate their elements procedurally and on demand.
7767 : : * Given a multi-column index on "(a, b)", and an SQL WHERE clause
7768 : : * "WHERE b = 42", a skip scan will effectively use an indexqual
7769 : : * "WHERE a = ANY('{every col a value}') AND b = 42". (Obviously,
7770 : : * the array on "a" must also return "IS NULL" matches, since our
7771 : : * WHERE clause used no strict operator on "a").
7772 : : *
7773 : : * Here we consider how nbtree will backfill skip arrays for any
7774 : : * index columns that lacked an '=' qual. This maintains our
7775 : : * num_sa_scans estimate, and determines if this new column (the
7776 : : * "iclause->indexcol" column, not the prior "indexcol" column)
7777 : : * can have its RestrictInfos/quals added to indexBoundQuals.
7778 : : *
7779 : : * We'll need to handle columns that have inequality quals, where
7780 : : * the skip array generates values from a range constrained by the
7781 : : * quals (not every possible value). We've been maintaining
7782 : : * indexSkipQuals to help with this; it will now contain all of
7783 : : * the prior column's quals (that is, indexcol's quals) when they
7784 : : * might be used for this.
7785 : : */
7786 [ + + ]: 93283 : if (found_row_compare)
7787 : : {
7788 : : /*
7789 : : * Skip arrays can't be added after a RowCompare input qual
7790 : : * due to limitations in nbtree
7791 : : */
7792 : 20 : break;
7793 : : }
7794 [ + + ]: 93263 : if (eqQualHere)
7795 : : {
7796 : : /*
7797 : : * Don't need to add a skip array for an indexcol that already
7798 : : * has an '=' qual/equality constraint
7799 : : */
7800 : 65948 : indexcol++;
7801 : 65948 : indexSkipQuals = NIL;
7802 : : }
5246 tgl@sss.pgh.pa.us 7803 : 93263 : eqQualHere = false;
7804 : :
396 pg@bowt.ie 7805 [ + + ]: 94898 : while (indexcol < iclause->indexcol)
7806 : : {
7807 : : double ndistinct;
7808 : 29386 : bool isdefault = true;
7809 : :
7810 : 29386 : found_array = true;
7811 : :
7812 : : /*
7813 : : * A skipped attribute's ndistinct forms the basis of our
7814 : : * estimate of the total number of "array elements" used by
7815 : : * its skip array at runtime. Look that up first.
7816 : : */
7817 : 29386 : examine_indexcol_variable(root, index, indexcol, &vardata);
7818 : 29386 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
7819 : :
7820 [ + + ]: 29386 : if (indexcol == 0)
7821 : : {
7822 : : /*
7823 : : * Get an estimate of the leading column's correlation in
7824 : : * passing (avoids rereading variable stats below)
7825 : : */
7826 [ + + ]: 27305 : if (HeapTupleIsValid(vardata.statsTuple))
7827 : 16657 : correlation = btcost_correlation(index, &vardata);
7828 : 27305 : have_correlation = true;
7829 : : }
7830 : :
7831 [ + + ]: 29386 : ReleaseVariableStats(vardata);
7832 : :
7833 : : /*
7834 : : * If ndistinct is a default estimate, conservatively assume
7835 : : * that no skipping will happen at runtime
7836 : : */
7837 [ + + ]: 29386 : if (isdefault)
7838 : : {
7839 : 8479 : num_sa_scans = num_sa_scans_prev_cols;
7840 : 27751 : break; /* done building indexBoundQuals */
7841 : : }
7842 : :
7843 : : /*
7844 : : * Apply indexcol's indexSkipQuals selectivity to ndistinct
7845 : : */
7846 [ + + ]: 20907 : if (indexSkipQuals != NIL)
7847 : : {
7848 : : List *partialSkipQuals;
7849 : : Selectivity ndistinctfrac;
7850 : :
7851 : : /*
7852 : : * If the index is partial, AND the index predicate with
7853 : : * the index-bound quals to produce a more accurate idea
7854 : : * of the number of distinct values for prior indexcol
7855 : : */
7856 : 552 : partialSkipQuals = add_predicate_to_index_quals(index,
7857 : : indexSkipQuals);
7858 : :
7859 : 552 : ndistinctfrac = clauselist_selectivity(root, partialSkipQuals,
7860 : 552 : index->rel->relid,
7861 : : JOIN_INNER,
7862 : : NULL);
7863 : :
7864 : : /*
7865 : : * If ndistinctfrac is selective (on its own), the scan is
7866 : : * unlikely to benefit from repositioning itself using
7867 : : * later quals. Do not allow iclause->indexcol's quals to
7868 : : * be added to indexBoundQuals (it would increase descent
7869 : : * costs, without lowering numIndexTuples costs by much).
7870 : : */
7871 [ + + ]: 552 : if (ndistinctfrac < DEFAULT_RANGE_INEQ_SEL)
7872 : : {
7873 : 311 : num_sa_scans = num_sa_scans_prev_cols;
7874 : 311 : break; /* done building indexBoundQuals */
7875 : : }
7876 : :
7877 : : /* Adjust ndistinct downward */
7878 : 241 : ndistinct = rint(ndistinct * ndistinctfrac);
7879 [ + - ]: 241 : ndistinct = Max(ndistinct, 1);
7880 : : }
7881 : :
7882 : : /*
7883 : : * When there's no inequality quals, account for the need to
7884 : : * find an initial value by counting -inf/+inf as a value.
7885 : : *
7886 : : * We don't charge anything extra for possible next/prior key
7887 : : * index probes, which are sometimes used to find the next
7888 : : * valid skip array element (ahead of using the located
7889 : : * element value to relocate the scan to the next position
7890 : : * that might contain matching tuples). It seems hard to do
7891 : : * better here. Use of the skip support infrastructure often
7892 : : * avoids most next/prior key probes. But even when it can't,
7893 : : * there's a decent chance that most individual next/prior key
7894 : : * probes will locate a leaf page whose key space overlaps all
7895 : : * of the scan's keys (even the lower-order keys) -- which
7896 : : * also avoids the need for a separate, extra index descent.
7897 : : * Note also that these probes are much cheaper than non-probe
7898 : : * primitive index scans: they're reliably very selective.
7899 : : */
7900 [ + + ]: 20596 : if (indexSkipQuals == NIL)
7901 : 20355 : ndistinct += 1;
7902 : :
7903 : : /*
7904 : : * Update num_sa_scans estimate by multiplying by ndistinct.
7905 : : *
7906 : : * We make the pessimistic assumption that there is no
7907 : : * naturally occurring cross-column correlation. This is
7908 : : * often wrong, but it seems best to err on the side of not
7909 : : * expecting skipping to be helpful...
7910 : : */
7911 : 20596 : num_sa_scans *= ndistinct;
7912 : :
7913 : : /*
7914 : : * ...but back out of adding this latest group of 1 or more
7915 : : * skip arrays when num_sa_scans exceeds the total number of
7916 : : * index pages (revert to num_sa_scans from before indexcol).
7917 : : * This causes a sharp discontinuity in cost (as a function of
7918 : : * the indexcol's ndistinct), but that is representative of
7919 : : * actual runtime costs.
7920 : : *
7921 : : * Note that skipping is helpful when each primitive index
7922 : : * scan only manages to skip over 1 or 2 irrelevant leaf pages
7923 : : * on average. Skip arrays bring savings in CPU costs due to
7924 : : * the scan not needing to evaluate indexquals against every
7925 : : * tuple, which can greatly exceed any savings in I/O costs.
7926 : : * This test is a test of whether num_sa_scans implies that
7927 : : * we're past the point where the ability to skip ceases to
7928 : : * lower the scan's costs (even qual evaluation CPU costs).
7929 : : */
7930 [ + + ]: 20596 : if (index->pages < num_sa_scans)
7931 : : {
7932 : 18961 : num_sa_scans = num_sa_scans_prev_cols;
7933 : 18961 : break; /* done building indexBoundQuals */
7934 : : }
7935 : :
7936 : 1635 : indexcol++;
7937 : 1635 : indexSkipQuals = NIL;
7938 : : }
7939 : :
7940 : : /*
7941 : : * Finished considering the need to add skip arrays to bridge an
7942 : : * initial eqQualHere gap between the old and new index columns
7943 : : * (or there was no initial eqQualHere gap in the first place).
7944 : : *
7945 : : * If an initial gap could not be bridged, then new column's quals
7946 : : * (i.e. iclause->indexcol's quals) won't go into indexBoundQuals,
7947 : : * and so won't affect our final numIndexTuples estimate.
7948 : : */
2636 tgl@sss.pgh.pa.us 7949 [ + + ]: 93263 : if (indexcol != iclause->indexcol)
396 pg@bowt.ie 7950 : 27751 : break; /* done building indexBoundQuals */
7951 : : }
7952 : :
7953 [ - + ]: 444572 : Assert(indexcol == iclause->indexcol);
7954 : :
7955 : : /* Examine each indexqual associated with this index clause */
2636 tgl@sss.pgh.pa.us 7956 [ + - + + : 891445 : foreach(lc2, iclause->indexquals)
+ + ]
7957 : : {
7958 : 446873 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7959 : 446873 : Expr *clause = rinfo->clause;
7960 : 446873 : Oid clause_op = InvalidOid;
7961 : : int op_strategy;
7962 : :
7963 [ + + ]: 446873 : if (IsA(clause, OpExpr))
7964 : : {
7965 : 432413 : OpExpr *op = (OpExpr *) clause;
7966 : :
7967 : 432413 : clause_op = op->opno;
7968 : : }
7969 [ + + ]: 14460 : else if (IsA(clause, RowCompareExpr))
7970 : : {
7971 : 330 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7972 : :
7973 : 330 : clause_op = linitial_oid(rc->opnos);
396 pg@bowt.ie 7974 : 330 : found_row_compare = true;
7975 : : }
2636 tgl@sss.pgh.pa.us 7976 [ + + ]: 14130 : else if (IsA(clause, ScalarArrayOpExpr))
7977 : : {
7978 : 12232 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7979 : 12232 : Node *other_operand = (Node *) lsecond(saop->args);
852 7980 : 12232 : double alength = estimate_array_length(root, other_operand);
7981 : :
2636 7982 : 12232 : clause_op = saop->opno;
396 pg@bowt.ie 7983 : 12232 : found_array = true;
7984 : : /* estimate SA descents by indexBoundQuals only */
2636 tgl@sss.pgh.pa.us 7985 [ + + ]: 12232 : if (alength > 1)
7986 : 12008 : num_sa_scans *= alength;
7987 : : }
7988 [ + - ]: 1898 : else if (IsA(clause, NullTest))
7989 : : {
7990 : 1898 : NullTest *nt = (NullTest *) clause;
7991 : :
7992 [ + + ]: 1898 : if (nt->nulltesttype == IS_NULL)
7993 : : {
7994 : 200 : found_is_null_op = true;
7995 : : /* IS NULL is like = for selectivity/skip scan purposes */
7996 : 200 : eqQualHere = true;
7997 : : }
7998 : : }
7999 : : else
2636 tgl@sss.pgh.pa.us 8000 [ # # ]:UBC 0 : elog(ERROR, "unsupported indexqual type: %d",
8001 : : (int) nodeTag(clause));
8002 : :
8003 : : /* check for equality operator */
2636 tgl@sss.pgh.pa.us 8004 [ + + ]:CBC 446873 : if (OidIsValid(clause_op))
8005 : : {
8006 : 444975 : op_strategy = get_op_opfamily_strategy(clause_op,
8007 : 444975 : index->opfamily[indexcol]);
8008 [ - + ]: 444975 : Assert(op_strategy != 0); /* not a member of opfamily?? */
8009 [ + + ]: 444975 : if (op_strategy == BTEqualStrategyNumber)
8010 : 421150 : eqQualHere = true;
8011 : : }
8012 : :
8013 : 446873 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
8014 : :
8015 : : /*
8016 : : * We apply inequality selectivities to estimate index descent
8017 : : * costs with scans that use skip arrays. Save this indexcol's
8018 : : * RestrictInfos if it looks like they'll be needed for that.
8019 : : */
396 pg@bowt.ie 8020 [ + + + + ]: 446873 : if (!eqQualHere && !found_row_compare &&
8021 [ + + ]: 24626 : indexcol < index->nkeycolumns - 1)
8022 : 4768 : indexSkipQuals = lappend(indexSkipQuals, rinfo);
8023 : : }
8024 : : }
8025 : :
8026 : : /*
8027 : : * If index is unique and we found an '=' clause for each column, we can
8028 : : * just assume numIndexTuples = 1 and skip the expensive
8029 : : * clauselist_selectivity calculations. However, an array or NullTest
8030 : : * always invalidates that theory (even when eqQualHere has been set).
8031 : : */
7466 tgl@sss.pgh.pa.us 8032 [ + + ]: 652343 : if (index->unique &&
2950 teodor@sigaev.ru 8033 [ + + + + ]: 521694 : indexcol == index->nkeycolumns - 1 &&
7466 tgl@sss.pgh.pa.us 8034 : 188987 : eqQualHere &&
396 pg@bowt.ie 8035 [ + + ]: 188987 : !found_array &&
5968 tgl@sss.pgh.pa.us 8036 [ + + ]: 183546 : !found_is_null_op)
7631 8037 : 183506 : numIndexTuples = 1.0;
8038 : : else
8039 : : {
8040 : : List *selectivityQuals;
8041 : : Selectivity btreeSelectivity;
8042 : :
8043 : : /*
8044 : : * If the index is partial, AND the index predicate with the
8045 : : * index-bound quals to produce a more accurate idea of the number of
8046 : : * rows covered by the bound conditions.
8047 : : */
2636 8048 : 468837 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
8049 : :
5210 8050 : 468837 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
7631 8051 : 468837 : index->rel->relid,
8052 : : JOIN_INNER,
8053 : : NULL);
8054 : 468837 : numIndexTuples = btreeSelectivity * index->rel->tuples;
8055 : :
8056 : : /*
8057 : : * btree automatically combines individual array element primitive
8058 : : * index scans whenever the tuples covered by the next set of array
8059 : : * keys are close to tuples covered by the current set. That puts a
8060 : : * natural ceiling on the worst case number of descents -- there
8061 : : * cannot possibly be more than one descent per leaf page scanned.
8062 : : *
8063 : : * Clamp the number of descents to at most 1/3 the number of index
8064 : : * pages. This avoids implausibly high estimates with low selectivity
8065 : : * paths, where scans usually require only one or two descents. This
8066 : : * is most likely to help when there are several SAOP clauses, where
8067 : : * naively accepting the total number of distinct combinations of
8068 : : * array elements as the number of descents would frequently lead to
8069 : : * wild overestimates.
8070 : : *
8071 : : * We somewhat arbitrarily don't just make the cutoff the total number
8072 : : * of leaf pages (we make it 1/3 the total number of pages instead) to
8073 : : * give the btree code credit for its ability to continue on the leaf
8074 : : * level with low selectivity scans.
8075 : : *
8076 : : * Note: num_sa_scans includes both ScalarArrayOp array elements and
8077 : : * skip array elements whose qual affects our numIndexTuples estimate.
8078 : : */
759 pg@bowt.ie 8079 [ + + ]: 468837 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
8080 [ + + ]: 468837 : num_sa_scans = Max(num_sa_scans, 1);
8081 : :
8082 : : /*
8083 : : * As in genericcostestimate(), we have to adjust for any array quals
8084 : : * included in indexBoundQuals, and then round to integer.
8085 : : *
8086 : : * It is tempting to make genericcostestimate behave as if array
8087 : : * clauses work in almost the same way as scalar operators during
8088 : : * btree scans, making the top-level scan look like a continuous scan
8089 : : * (as opposed to num_sa_scans-many primitive index scans). After
8090 : : * all, btree scans mostly work like that at runtime. However, such a
8091 : : * scheme would badly bias genericcostestimate's simplistic approach
8092 : : * to calculating numIndexPages through prorating.
8093 : : *
8094 : : * Stick with the approach taken by non-native SAOP scans for now.
8095 : : * genericcostestimate will use the Mackert-Lohman formula to
8096 : : * compensate for repeat page fetches, even though that definitely
8097 : : * won't happen during btree scans (not for leaf pages, at least).
8098 : : * We're usually very pessimistic about the number of primitive index
8099 : : * scans that will be required, but it's not clear how to do better.
8100 : : */
7081 tgl@sss.pgh.pa.us 8101 : 468837 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
8102 : : }
8103 : :
8104 : : /*
8105 : : * Now do generic index cost estimation.
8106 : : *
8107 : : * While we expended effort to make realistic estimates of numIndexTuples
8108 : : * and num_sa_scans, we are content to count only the btree metapage as
8109 : : * non-leaf. btree fanout is typically high enough that upper pages are
8110 : : * few relative to leaf pages, so accounting for them would move the
8111 : : * estimates at most a percent or two. Given the uncertainty in just how
8112 : : * many upper pages exist in a particular index, we'll skip trying to
8113 : : * handle that.
8114 : : */
4862 8115 : 652343 : costs.numIndexTuples = numIndexTuples;
759 pg@bowt.ie 8116 : 652343 : costs.num_sa_scans = num_sa_scans;
46 tgl@sss.pgh.pa.us 8117 :GNC 652343 : costs.numNonLeafPages = 1;
8118 : :
2636 tgl@sss.pgh.pa.us 8119 :CBC 652343 : genericcostestimate(root, path, loop_count, &costs);
8120 : :
8121 : : /*
8122 : : * Add a CPU-cost component to represent the costs of initial btree
8123 : : * descent. We don't charge any I/O cost for touching upper btree levels,
8124 : : * since they tend to stay in cache, but we still have to do about log2(N)
8125 : : * comparisons to descend a btree of N leaf tuples. We charge one
8126 : : * cpu_operator_cost per comparison.
8127 : : *
8128 : : * If there are SAOP or skip array keys, charge this once per estimated
8129 : : * index descent. The ones after the first one are not startup cost so
8130 : : * far as the overall plan goes, so just add them to "total" cost.
8131 : : */
4862 8132 [ + + ]: 652343 : if (index->tuples > 1) /* avoid computing log(0) */
8133 : : {
8134 : 591293 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
8135 : 591293 : costs.indexStartupCost += descentCost;
8136 : 591293 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8137 : : }
8138 : :
8139 : : /*
8140 : : * Even though we're not charging I/O cost for touching upper btree pages,
8141 : : * it's still reasonable to charge some CPU cost per page descended
8142 : : * through. Moreover, if we had no such charge at all, bloated indexes
8143 : : * would appear to have the same search cost as unbloated ones, at least
8144 : : * in cases where only a single leaf page is expected to be visited. This
8145 : : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
8146 : : * touched. The number of such pages is btree tree height plus one (ie,
8147 : : * we charge for the leaf page too). As above, charge once per estimated
8148 : : * SAOP/skip array descent.
8149 : : */
1213 akorotkov@postgresql 8150 : 652343 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
4862 tgl@sss.pgh.pa.us 8151 : 652343 : costs.indexStartupCost += descentCost;
8152 : 652343 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8153 : :
396 pg@bowt.ie 8154 [ + + ]: 652343 : if (!have_correlation)
8155 : : {
8156 : 625038 : examine_indexcol_variable(root, index, 0, &vardata);
8157 [ + + ]: 625038 : if (HeapTupleIsValid(vardata.statsTuple))
8158 : 448273 : costs.indexCorrelation = btcost_correlation(index, &vardata);
8159 [ + + ]: 625038 : ReleaseVariableStats(vardata);
8160 : : }
8161 : : else
8162 : : {
8163 : : /* btcost_correlation already called earlier on */
8164 : 27305 : costs.indexCorrelation = correlation;
8165 : : }
8166 : :
4862 tgl@sss.pgh.pa.us 8167 : 652343 : *indexStartupCost = costs.indexStartupCost;
8168 : 652343 : *indexTotalCost = costs.indexTotalCost;
8169 : 652343 : *indexSelectivity = costs.indexSelectivity;
8170 : 652343 : *indexCorrelation = costs.indexCorrelation;
3366 rhaas@postgresql.org 8171 : 652343 : *indexPages = costs.numIndexPages;
10892 scrappy@hub.org 8172 : 652343 : }
8173 : :
8174 : : void
3761 tgl@sss.pgh.pa.us 8175 : 308 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8176 : : Cost *indexStartupCost, Cost *indexTotalCost,
8177 : : Selectivity *indexSelectivity, double *indexCorrelation,
8178 : : double *indexPages)
8179 : : {
1389 peter@eisentraut.org 8180 : 308 : GenericCosts costs = {0};
8181 : :
8182 : : /* As in btcostestimate, count only the metapage as non-leaf */
46 tgl@sss.pgh.pa.us 8183 :GNC 308 : costs.numNonLeafPages = 1;
8184 : :
2636 tgl@sss.pgh.pa.us 8185 :CBC 308 : genericcostestimate(root, path, loop_count, &costs);
8186 : :
8187 : : /*
8188 : : * A hash index has no descent costs as such, since the index AM can go
8189 : : * directly to the target bucket after computing the hash value. There
8190 : : * are a couple of other hash-specific costs that we could conceivably add
8191 : : * here, though:
8192 : : *
8193 : : * Ideally we'd charge spc_random_page_cost for each page in the target
8194 : : * bucket, not just the numIndexPages pages that genericcostestimate
8195 : : * thought we'd visit. However in most cases we don't know which bucket
8196 : : * that will be. There's no point in considering the average bucket size
8197 : : * because the hash AM makes sure that's always one page.
8198 : : *
8199 : : * Likewise, we could consider charging some CPU for each index tuple in
8200 : : * the bucket, if we knew how many there were. But the per-tuple cost is
8201 : : * just a hash value comparison, not a general datatype-dependent
8202 : : * comparison, so any such charge ought to be quite a bit less than
8203 : : * cpu_operator_cost; which makes it probably not worth worrying about.
8204 : : *
8205 : : * A bigger issue is that chance hash-value collisions will result in
8206 : : * wasted probes into the heap. We don't currently attempt to model this
8207 : : * cost on the grounds that it's rare, but maybe it's not rare enough.
8208 : : * (Any fix for this ought to consider the generic lossy-operator problem,
8209 : : * though; it's not entirely hash-specific.)
8210 : : */
8211 : :
4862 8212 : 308 : *indexStartupCost = costs.indexStartupCost;
8213 : 308 : *indexTotalCost = costs.indexTotalCost;
8214 : 308 : *indexSelectivity = costs.indexSelectivity;
8215 : 308 : *indexCorrelation = costs.indexCorrelation;
3366 rhaas@postgresql.org 8216 : 308 : *indexPages = costs.numIndexPages;
10844 scrappy@hub.org 8217 : 308 : }
8218 : :
8219 : : void
3761 tgl@sss.pgh.pa.us 8220 : 4746 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8221 : : Cost *indexStartupCost, Cost *indexTotalCost,
8222 : : Selectivity *indexSelectivity, double *indexCorrelation,
8223 : : double *indexPages)
8224 : : {
4862 8225 : 4746 : IndexOptInfo *index = path->indexinfo;
1389 peter@eisentraut.org 8226 : 4746 : GenericCosts costs = {0};
8227 : : Cost descentCost;
8228 : :
8229 : : /* GiST has no metapage, so we treat all pages as leaf pages */
8230 : :
2636 tgl@sss.pgh.pa.us 8231 : 4746 : genericcostestimate(root, path, loop_count, &costs);
8232 : :
8233 : : /*
8234 : : * We model index descent costs similarly to those for btree, but to do
8235 : : * that we first need an idea of the tree height. We somewhat arbitrarily
8236 : : * assume that the fanout is 100, meaning the tree height is at most
8237 : : * log100(index->pages).
8238 : : *
8239 : : * Although this computation isn't really expensive enough to require
8240 : : * caching, we might as well use index->tree_height to cache it.
8241 : : */
4724 bruce@momjian.us 8242 [ + + ]: 4746 : if (index->tree_height < 0) /* unknown? */
8243 : : {
4862 tgl@sss.pgh.pa.us 8244 [ + + ]: 4719 : if (index->pages > 1) /* avoid computing log(0) */
8245 : 1956 : index->tree_height = (int) (log(index->pages) / log(100.0));
8246 : : else
8247 : 2763 : index->tree_height = 0;
8248 : : }
8249 : :
8250 : : /*
8251 : : * Add a CPU-cost component to represent the costs of initial descent. We
8252 : : * just use log(N) here not log2(N) since the branching factor isn't
8253 : : * necessarily two anyway. As for btree, charge once per SA scan.
8254 : : */
8255 [ + - ]: 4746 : if (index->tuples > 1) /* avoid computing log(0) */
8256 : : {
8257 : 4746 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8258 : 4746 : costs.indexStartupCost += descentCost;
8259 : 4746 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8260 : : }
8261 : :
8262 : : /*
8263 : : * Likewise add a per-page charge, calculated the same as for btrees.
8264 : : */
1213 akorotkov@postgresql 8265 : 4746 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
4862 tgl@sss.pgh.pa.us 8266 : 4746 : costs.indexStartupCost += descentCost;
8267 : 4746 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8268 : :
8269 : 4746 : *indexStartupCost = costs.indexStartupCost;
8270 : 4746 : *indexTotalCost = costs.indexTotalCost;
8271 : 4746 : *indexSelectivity = costs.indexSelectivity;
8272 : 4746 : *indexCorrelation = costs.indexCorrelation;
3366 rhaas@postgresql.org 8273 : 4746 : *indexPages = costs.numIndexPages;
10844 scrappy@hub.org 8274 : 4746 : }
8275 : :
8276 : : void
3761 tgl@sss.pgh.pa.us 8277 : 1482 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8278 : : Cost *indexStartupCost, Cost *indexTotalCost,
8279 : : Selectivity *indexSelectivity, double *indexCorrelation,
8280 : : double *indexPages)
8281 : : {
4862 8282 : 1482 : IndexOptInfo *index = path->indexinfo;
1389 peter@eisentraut.org 8283 : 1482 : GenericCosts costs = {0};
8284 : : Cost descentCost;
8285 : :
8286 : : /* As in btcostestimate, count only the metapage as non-leaf */
46 tgl@sss.pgh.pa.us 8287 :GNC 1482 : costs.numNonLeafPages = 1;
8288 : :
2636 tgl@sss.pgh.pa.us 8289 :CBC 1482 : genericcostestimate(root, path, loop_count, &costs);
8290 : :
8291 : : /*
8292 : : * We model index descent costs similarly to those for btree, but to do
8293 : : * that we first need an idea of the tree height. We somewhat arbitrarily
8294 : : * assume that the fanout is 100, meaning the tree height is at most
8295 : : * log100(index->pages).
8296 : : *
8297 : : * Although this computation isn't really expensive enough to require
8298 : : * caching, we might as well use index->tree_height to cache it.
8299 : : */
4724 bruce@momjian.us 8300 [ + + ]: 1482 : if (index->tree_height < 0) /* unknown? */
8301 : : {
4862 tgl@sss.pgh.pa.us 8302 [ + - ]: 1477 : if (index->pages > 1) /* avoid computing log(0) */
8303 : 1477 : index->tree_height = (int) (log(index->pages) / log(100.0));
8304 : : else
4862 tgl@sss.pgh.pa.us 8305 :UBC 0 : index->tree_height = 0;
8306 : : }
8307 : :
8308 : : /*
8309 : : * Add a CPU-cost component to represent the costs of initial descent. We
8310 : : * just use log(N) here not log2(N) since the branching factor isn't
8311 : : * necessarily two anyway. As for btree, charge once per SA scan.
8312 : : */
4862 tgl@sss.pgh.pa.us 8313 [ + - ]:CBC 1482 : if (index->tuples > 1) /* avoid computing log(0) */
8314 : : {
8315 : 1482 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8316 : 1482 : costs.indexStartupCost += descentCost;
8317 : 1482 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8318 : : }
8319 : :
8320 : : /*
8321 : : * Likewise add a per-page charge, calculated the same as for btrees.
8322 : : */
1213 akorotkov@postgresql 8323 : 1482 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
4862 tgl@sss.pgh.pa.us 8324 : 1482 : costs.indexStartupCost += descentCost;
8325 : 1482 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8326 : :
8327 : 1482 : *indexStartupCost = costs.indexStartupCost;
8328 : 1482 : *indexTotalCost = costs.indexTotalCost;
8329 : 1482 : *indexSelectivity = costs.indexSelectivity;
8330 : 1482 : *indexCorrelation = costs.indexCorrelation;
3366 rhaas@postgresql.org 8331 : 1482 : *indexPages = costs.numIndexPages;
5253 tgl@sss.pgh.pa.us 8332 : 1482 : }
8333 : :
8334 : :
8335 : : /*
8336 : : * Support routines for gincostestimate
8337 : : */
8338 : :
8339 : : typedef struct
8340 : : {
8341 : : bool attHasFullScan[INDEX_MAX_KEYS];
8342 : : bool attHasNormalScan[INDEX_MAX_KEYS];
8343 : : double partialEntries;
8344 : : double exactEntries;
8345 : : double searchEntries;
8346 : : double arrayScans;
8347 : : } GinQualCounts;
8348 : :
8349 : : /*
8350 : : * Estimate the number of index terms that need to be searched for while
8351 : : * testing the given GIN query, and increment the counts in *counts
8352 : : * appropriately. If the query is unsatisfiable, return false.
8353 : : */
8354 : : static bool
5250 8355 : 1714 : gincost_pattern(IndexOptInfo *index, int indexcol,
8356 : : Oid clause_op, Datum query,
8357 : : GinQualCounts *counts)
8358 : : {
8359 : : FmgrInfo flinfo;
8360 : : Oid extractProcOid;
8361 : : Oid collation;
8362 : : int strategy_op;
8363 : : Oid lefttype,
8364 : : righttype;
8365 : 1714 : int32 nentries = 0;
8366 : 1714 : bool *partial_matches = NULL;
8367 : 1714 : Pointer *extra_data = NULL;
8368 : 1714 : bool *nullFlags = NULL;
8369 : 1714 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
8370 : : int32 i;
8371 : :
2945 teodor@sigaev.ru 8372 [ - + ]: 1714 : Assert(indexcol < index->nkeycolumns);
8373 : :
8374 : : /*
8375 : : * Get the operator's strategy number and declared input data types within
8376 : : * the index opfamily. (We don't need the latter, but we use
8377 : : * get_op_opfamily_properties because it will throw error if it fails to
8378 : : * find a matching pg_amop entry.)
8379 : : */
5250 tgl@sss.pgh.pa.us 8380 : 1714 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
8381 : : &strategy_op, &lefttype, &righttype);
8382 : :
8383 : : /*
8384 : : * GIN always uses the "default" support functions, which are those with
8385 : : * lefttype == righttype == the opclass' opcintype (see
8386 : : * IndexSupportInitialize in relcache.c).
8387 : : */
8388 : 1714 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
8389 : 1714 : index->opcintype[indexcol],
8390 : 1714 : index->opcintype[indexcol],
8391 : : GIN_EXTRACTQUERY_PROC);
8392 : :
8393 [ - + ]: 1714 : if (!OidIsValid(extractProcOid))
8394 : : {
8395 : : /* should not happen; throw same error as index_getprocinfo */
5250 tgl@sss.pgh.pa.us 8396 [ # # ]:UBC 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
8397 : : GIN_EXTRACTQUERY_PROC, indexcol + 1,
8398 : : get_rel_name(index->indexoid));
8399 : : }
8400 : :
8401 : : /*
8402 : : * Choose collation to pass to extractProc (should match initGinState).
8403 : : */
4774 tgl@sss.pgh.pa.us 8404 [ + + ]:CBC 1714 : if (OidIsValid(index->indexcollations[indexcol]))
8405 : 243 : collation = index->indexcollations[indexcol];
8406 : : else
8407 : 1471 : collation = DEFAULT_COLLATION_OID;
8408 : :
2227 akorotkov@postgresql 8409 : 1714 : fmgr_info(extractProcOid, &flinfo);
8410 : :
8411 : 1714 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
8412 : :
8413 : 1714 : FunctionCall7Coll(&flinfo,
8414 : : collation,
8415 : : query,
8416 : : PointerGetDatum(&nentries),
8417 : : UInt16GetDatum(strategy_op),
8418 : : PointerGetDatum(&partial_matches),
8419 : : PointerGetDatum(&extra_data),
8420 : : PointerGetDatum(&nullFlags),
8421 : : PointerGetDatum(&searchMode));
8422 : :
5250 tgl@sss.pgh.pa.us 8423 [ + + + + ]: 1714 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
8424 : : {
8425 : : /* No match is possible */
8426 : 10 : return false;
8427 : : }
8428 : :
8429 [ + + ]: 5854 : for (i = 0; i < nentries; i++)
8430 : : {
8431 : : /*
8432 : : * For partial match we haven't any information to estimate number of
8433 : : * matched entries in index, so, we just estimate it as 100
8434 : : */
8435 [ + + + + ]: 4150 : if (partial_matches && partial_matches[i])
8436 : 361 : counts->partialEntries += 100;
8437 : : else
8438 : 3789 : counts->exactEntries++;
8439 : :
8440 : 4150 : counts->searchEntries++;
8441 : : }
8442 : :
2299 akorotkov@postgresql 8443 [ + + ]: 1704 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
8444 : : {
8445 : 1322 : counts->attHasNormalScan[indexcol] = true;
8446 : : }
8447 [ + + ]: 382 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
8448 : : {
8449 : : /* Treat "include empty" like an exact-match item */
8450 : 36 : counts->attHasNormalScan[indexcol] = true;
5250 tgl@sss.pgh.pa.us 8451 : 36 : counts->exactEntries++;
8452 : 36 : counts->searchEntries++;
8453 : : }
8454 : : else
8455 : : {
8456 : : /* It's GIN_SEARCH_MODE_ALL */
2299 akorotkov@postgresql 8457 : 346 : counts->attHasFullScan[indexcol] = true;
8458 : : }
8459 : :
5250 tgl@sss.pgh.pa.us 8460 : 1704 : return true;
8461 : : }
8462 : :
8463 : : /*
8464 : : * Estimate the number of index terms that need to be searched for while
8465 : : * testing the given GIN index clause, and increment the counts in *counts
8466 : : * appropriately. If the query is unsatisfiable, return false.
8467 : : */
8468 : : static bool
4081 8469 : 1704 : gincost_opexpr(PlannerInfo *root,
8470 : : IndexOptInfo *index,
8471 : : int indexcol,
8472 : : OpExpr *clause,
8473 : : GinQualCounts *counts)
8474 : : {
2636 8475 : 1704 : Oid clause_op = clause->opno;
8476 : 1704 : Node *operand = (Node *) lsecond(clause->args);
8477 : :
8478 : : /* aggressively reduce to a constant, and look through relabeling */
4456 8479 : 1704 : operand = estimate_expression_value(root, operand);
8480 : :
5250 8481 [ - + ]: 1704 : if (IsA(operand, RelabelType))
5250 tgl@sss.pgh.pa.us 8482 :UBC 0 : operand = (Node *) ((RelabelType *) operand)->arg;
8483 : :
8484 : : /*
8485 : : * It's impossible to call extractQuery method for unknown operand. So
8486 : : * unless operand is a Const we can't do much; just assume there will be
8487 : : * one ordinary search entry from the operand at runtime.
8488 : : */
5250 tgl@sss.pgh.pa.us 8489 [ - + ]:CBC 1704 : if (!IsA(operand, Const))
8490 : : {
5250 tgl@sss.pgh.pa.us 8491 :UBC 0 : counts->exactEntries++;
8492 : 0 : counts->searchEntries++;
8493 : 0 : return true;
8494 : : }
8495 : :
8496 : : /* If Const is null, there can be no matches */
5250 tgl@sss.pgh.pa.us 8497 [ - + ]:CBC 1704 : if (((Const *) operand)->constisnull)
5250 tgl@sss.pgh.pa.us 8498 :UBC 0 : return false;
8499 : :
8500 : : /* Otherwise, apply extractQuery and get the actual term counts */
5250 tgl@sss.pgh.pa.us 8501 :CBC 1704 : return gincost_pattern(index, indexcol, clause_op,
8502 : : ((Const *) operand)->constvalue,
8503 : : counts);
8504 : : }
8505 : :
8506 : : /*
8507 : : * Estimate the number of index terms that need to be searched for while
8508 : : * testing the given GIN index clause, and increment the counts in *counts
8509 : : * appropriately. If the query is unsatisfiable, return false.
8510 : : *
8511 : : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
8512 : : * each of which involves one value from the RHS array, plus all the
8513 : : * non-array quals (if any). To model this, we average the counts across
8514 : : * the RHS elements, and add the averages to the counts in *counts (which
8515 : : * correspond to per-indexscan costs). We also multiply counts->arrayScans
8516 : : * by N, causing gincostestimate to scale up its estimates accordingly.
8517 : : */
8518 : : static bool
4456 8519 : 5 : gincost_scalararrayopexpr(PlannerInfo *root,
8520 : : IndexOptInfo *index,
8521 : : int indexcol,
8522 : : ScalarArrayOpExpr *clause,
8523 : : double numIndexEntries,
8524 : : GinQualCounts *counts)
8525 : : {
2636 8526 : 5 : Oid clause_op = clause->opno;
8527 : 5 : Node *rightop = (Node *) lsecond(clause->args);
8528 : : ArrayType *arrayval;
8529 : : int16 elmlen;
8530 : : bool elmbyval;
8531 : : char elmalign;
8532 : : int numElems;
8533 : : Datum *elemValues;
8534 : : bool *elemNulls;
8535 : : GinQualCounts arraycounts;
5250 8536 : 5 : int numPossible = 0;
8537 : : int i;
8538 : :
2636 8539 [ - + ]: 5 : Assert(clause->useOr);
8540 : :
8541 : : /* aggressively reduce to a constant, and look through relabeling */
4456 8542 : 5 : rightop = estimate_expression_value(root, rightop);
8543 : :
5250 8544 [ - + ]: 5 : if (IsA(rightop, RelabelType))
5250 tgl@sss.pgh.pa.us 8545 :UBC 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
8546 : :
8547 : : /*
8548 : : * It's impossible to call extractQuery method for unknown operand. So
8549 : : * unless operand is a Const we can't do much; just assume there will be
8550 : : * one ordinary search entry from each array entry at runtime, and fall
8551 : : * back on a probably-bad estimate of the number of array entries.
8552 : : */
5250 tgl@sss.pgh.pa.us 8553 [ - + ]:CBC 5 : if (!IsA(rightop, Const))
8554 : : {
5250 tgl@sss.pgh.pa.us 8555 :UBC 0 : counts->exactEntries++;
8556 : 0 : counts->searchEntries++;
852 8557 : 0 : counts->arrayScans *= estimate_array_length(root, rightop);
5250 8558 : 0 : return true;
8559 : : }
8560 : :
8561 : : /* If Const is null, there can be no matches */
5250 tgl@sss.pgh.pa.us 8562 [ - + ]:CBC 5 : if (((Const *) rightop)->constisnull)
5250 tgl@sss.pgh.pa.us 8563 :UBC 0 : return false;
8564 : :
8565 : : /* Otherwise, extract the array elements and iterate over them */
5250 tgl@sss.pgh.pa.us 8566 :CBC 5 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
8567 : 5 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
8568 : : &elmlen, &elmbyval, &elmalign);
8569 : 5 : deconstruct_array(arrayval,
8570 : : ARR_ELEMTYPE(arrayval),
8571 : : elmlen, elmbyval, elmalign,
8572 : : &elemValues, &elemNulls, &numElems);
8573 : :
8574 : 5 : memset(&arraycounts, 0, sizeof(arraycounts));
8575 : :
8576 [ + + ]: 15 : for (i = 0; i < numElems; i++)
8577 : : {
8578 : : GinQualCounts elemcounts;
8579 : :
8580 : : /* NULL can't match anything, so ignore, as the executor will */
8581 [ - + ]: 10 : if (elemNulls[i])
5250 tgl@sss.pgh.pa.us 8582 :UBC 0 : continue;
8583 : :
8584 : : /* Otherwise, apply extractQuery and get the actual term counts */
5250 tgl@sss.pgh.pa.us 8585 :CBC 10 : memset(&elemcounts, 0, sizeof(elemcounts));
8586 : :
8587 [ + - ]: 10 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
8588 : : &elemcounts))
8589 : : {
8590 : : /* We ignore array elements that are unsatisfiable patterns */
8591 : 10 : numPossible++;
8592 : :
2299 akorotkov@postgresql 8593 [ - + ]: 10 : if (elemcounts.attHasFullScan[indexcol] &&
2299 akorotkov@postgresql 8594 [ # # ]:UBC 0 : !elemcounts.attHasNormalScan[indexcol])
8595 : : {
8596 : : /*
8597 : : * Full index scan will be required. We treat this as if
8598 : : * every key in the index had been listed in the query; is
8599 : : * that reasonable?
8600 : : */
5250 tgl@sss.pgh.pa.us 8601 : 0 : elemcounts.partialEntries = 0;
8602 : 0 : elemcounts.exactEntries = numIndexEntries;
8603 : 0 : elemcounts.searchEntries = numIndexEntries;
8604 : : }
5250 tgl@sss.pgh.pa.us 8605 :CBC 10 : arraycounts.partialEntries += elemcounts.partialEntries;
8606 : 10 : arraycounts.exactEntries += elemcounts.exactEntries;
8607 : 10 : arraycounts.searchEntries += elemcounts.searchEntries;
8608 : : }
8609 : : }
8610 : :
8611 [ - + ]: 5 : if (numPossible == 0)
8612 : : {
8613 : : /* No satisfiable patterns in the array */
5250 tgl@sss.pgh.pa.us 8614 :UBC 0 : return false;
8615 : : }
8616 : :
8617 : : /*
8618 : : * Now add the averages to the global counts. This will give us an
8619 : : * estimate of the average number of terms searched for in each indexscan,
8620 : : * including contributions from both array and non-array quals.
8621 : : */
5250 tgl@sss.pgh.pa.us 8622 :CBC 5 : counts->partialEntries += arraycounts.partialEntries / numPossible;
8623 : 5 : counts->exactEntries += arraycounts.exactEntries / numPossible;
8624 : 5 : counts->searchEntries += arraycounts.searchEntries / numPossible;
8625 : :
8626 : 5 : counts->arrayScans *= numPossible;
8627 : :
8628 : 5 : return true;
8629 : : }
8630 : :
8631 : : /*
8632 : : * GIN has search behavior completely different from other index types
8633 : : */
8634 : : void
3761 8635 : 1546 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8636 : : Cost *indexStartupCost, Cost *indexTotalCost,
8637 : : Selectivity *indexSelectivity, double *indexCorrelation,
8638 : : double *indexPages)
8639 : : {
5246 8640 : 1546 : IndexOptInfo *index = path->indexinfo;
2636 8641 : 1546 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8642 : : List *selectivityQuals;
5504 bruce@momjian.us 8643 : 1546 : double numPages = index->pages,
8644 : 1546 : numTuples = index->tuples;
8645 : : double numEntryPages,
8646 : : numDataPages,
8647 : : numPendingPages,
8648 : : numEntries;
8649 : : GinQualCounts counts;
8650 : : bool matchPossible;
8651 : : bool fullIndexScan;
8652 : : double partialScale;
8653 : : double entryPagesFetched,
8654 : : dataPagesFetched,
8655 : : dataPagesFetchedBySel;
8656 : : double qual_op_cost,
8657 : : qual_arg_cost,
8658 : : spc_random_page_cost,
8659 : : outer_scans;
8660 : : Cost descentCost;
8661 : : Relation indexRel;
8662 : : GinStatsData ginStats;
8663 : : ListCell *lc;
8664 : : int i;
8665 : :
8666 : : /*
8667 : : * Obtain statistical information from the meta page, if possible. Else
8668 : : * set ginStats to zeroes, and we'll cope below.
8669 : : */
3808 tgl@sss.pgh.pa.us 8670 [ + - ]: 1546 : if (!index->hypothetical)
8671 : : {
8672 : : /* Lock should have already been obtained in plancat.c */
2588 8673 : 1546 : indexRel = index_open(index->indexoid, NoLock);
3808 8674 : 1546 : ginGetStats(indexRel, &ginStats);
2588 8675 : 1546 : index_close(indexRel, NoLock);
8676 : : }
8677 : : else
8678 : : {
3808 tgl@sss.pgh.pa.us 8679 :UBC 0 : memset(&ginStats, 0, sizeof(ginStats));
8680 : : }
8681 : :
8682 : : /*
8683 : : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
8684 : : * trusted, but the other fields are data as of the last VACUUM. We can
8685 : : * scale them up to account for growth since then, but that method only
8686 : : * goes so far; in the worst case, the stats might be for a completely
8687 : : * empty index, and scaling them will produce pretty bogus numbers.
8688 : : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
8689 : : * it's grown more than that, fall back to estimating things only from the
8690 : : * assumed-accurate index size. But we'll trust nPendingPages in any case
8691 : : * so long as it's not clearly insane, ie, more than the index size.
8692 : : */
3777 tgl@sss.pgh.pa.us 8693 [ + - ]:CBC 1546 : if (ginStats.nPendingPages < numPages)
8694 : 1546 : numPendingPages = ginStats.nPendingPages;
8695 : : else
3777 tgl@sss.pgh.pa.us 8696 :UBC 0 : numPendingPages = 0;
8697 : :
3777 tgl@sss.pgh.pa.us 8698 [ + - + - ]:CBC 1546 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
8699 [ + + ]: 1546 : ginStats.nTotalPages > numPages / 4 &&
8700 [ + - + + ]: 1506 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
3808 8701 : 1292 : {
8702 : : /*
8703 : : * OK, the stats seem close enough to sane to be trusted. But we
8704 : : * still need to scale them by the ratio numPages / nTotalPages to
8705 : : * account for growth since the last VACUUM.
8706 : : */
5504 bruce@momjian.us 8707 : 1292 : double scale = numPages / ginStats.nTotalPages;
8708 : :
3777 tgl@sss.pgh.pa.us 8709 : 1292 : numEntryPages = ceil(ginStats.nEntryPages * scale);
8710 : 1292 : numDataPages = ceil(ginStats.nDataPages * scale);
8711 : 1292 : numEntries = ceil(ginStats.nEntries * scale);
8712 : : /* ensure we didn't round up too much */
8713 [ + + ]: 1292 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
8714 [ + + ]: 1292 : numDataPages = Min(numDataPages,
8715 : : numPages - numPendingPages - numEntryPages);
8716 : : }
8717 : : else
8718 : : {
8719 : : /*
8720 : : * We might get here because it's a hypothetical index, or an index
8721 : : * created pre-9.1 and never vacuumed since upgrading (in which case
8722 : : * its stats would read as zeroes), or just because it's grown too
8723 : : * much since the last VACUUM for us to put our faith in scaling.
8724 : : *
8725 : : * Invent some plausible internal statistics based on the index page
8726 : : * count (and clamp that to at least 10 pages, just in case). We
8727 : : * estimate that 90% of the index is entry pages, and the rest is data
8728 : : * pages. Estimate 100 entries per entry page; this is rather bogus
8729 : : * since it'll depend on the size of the keys, but it's more robust
8730 : : * than trying to predict the number of entries per heap tuple.
8731 : : */
3808 8732 [ + + ]: 254 : numPages = Max(numPages, 10);
3777 8733 : 254 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
8734 : 254 : numDataPages = numPages - numPendingPages - numEntryPages;
3808 8735 : 254 : numEntries = floor(numEntryPages * 100);
8736 : : }
8737 : :
8738 : : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
5493 8739 [ - + ]: 1546 : if (numEntries < 1)
5493 tgl@sss.pgh.pa.us 8740 :UBC 0 : numEntries = 1;
8741 : :
8742 : : /*
8743 : : * If the index is partial, AND the index predicate with the index-bound
8744 : : * quals to produce a more accurate idea of the number of rows covered by
8745 : : * the bound conditions.
8746 : : */
2636 tgl@sss.pgh.pa.us 8747 :CBC 1546 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
8748 : :
8749 : : /* Estimate the fraction of main-table tuples that will be visited */
5679 8750 : 3092 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
5504 bruce@momjian.us 8751 : 1546 : index->rel->relid,
8752 : : JOIN_INNER,
8753 : : NULL);
8754 : :
8755 : : /* fetch estimated page cost for tablespace containing index */
8756 : 1546 : get_tablespace_page_costs(index->reltablespace,
8757 : : &spc_random_page_cost,
8758 : : NULL);
8759 : :
8760 : : /*
8761 : : * Generic assumption about index correlation: there isn't any.
8762 : : */
5679 tgl@sss.pgh.pa.us 8763 : 1546 : *indexCorrelation = 0.0;
8764 : :
8765 : : /*
8766 : : * Examine quals to estimate number of search entries & partial matches
8767 : : */
5250 8768 : 1546 : memset(&counts, 0, sizeof(counts));
8769 : 1546 : counts.arrayScans = 1;
8770 : 1546 : matchPossible = true;
8771 : :
2636 8772 [ + - + + : 3255 : foreach(lc, path->indexclauses)
+ + ]
8773 : : {
8774 : 1709 : IndexClause *iclause = lfirst_node(IndexClause, lc);
8775 : : ListCell *lc2;
8776 : :
8777 [ + - + + : 3408 : foreach(lc2, iclause->indexquals)
+ + ]
8778 : : {
8779 : 1709 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8780 : 1709 : Expr *clause = rinfo->clause;
8781 : :
8782 [ + + ]: 1709 : if (IsA(clause, OpExpr))
8783 : : {
8784 : 1704 : matchPossible = gincost_opexpr(root,
8785 : : index,
8786 : 1704 : iclause->indexcol,
8787 : : (OpExpr *) clause,
8788 : : &counts);
8789 [ + + ]: 1704 : if (!matchPossible)
8790 : 10 : break;
8791 : : }
8792 [ + - ]: 5 : else if (IsA(clause, ScalarArrayOpExpr))
8793 : : {
8794 : 5 : matchPossible = gincost_scalararrayopexpr(root,
8795 : : index,
8796 : 5 : iclause->indexcol,
8797 : : (ScalarArrayOpExpr *) clause,
8798 : : numEntries,
8799 : : &counts);
8800 [ - + ]: 5 : if (!matchPossible)
2636 tgl@sss.pgh.pa.us 8801 :UBC 0 : break;
8802 : : }
8803 : : else
8804 : : {
8805 : : /* shouldn't be anything else for a GIN index */
8806 [ # # ]: 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
8807 : : (int) nodeTag(clause));
8808 : : }
8809 : : }
8810 : : }
8811 : :
8812 : : /* Fall out if there were any provably-unsatisfiable quals */
5250 tgl@sss.pgh.pa.us 8813 [ + + ]:CBC 1546 : if (!matchPossible)
8814 : : {
8815 : 10 : *indexStartupCost = 0;
8816 : 10 : *indexTotalCost = 0;
8817 : 10 : *indexSelectivity = 0;
3761 8818 : 10 : return;
8819 : : }
8820 : :
8821 : : /*
8822 : : * If attribute has a full scan and at the same time doesn't have normal
8823 : : * scan, then we'll have to scan all non-null entries of that attribute.
8824 : : * Currently, we don't have per-attribute statistics for GIN. Thus, we
8825 : : * must assume the whole GIN index has to be scanned in this case.
8826 : : */
2299 akorotkov@postgresql 8827 : 1536 : fullIndexScan = false;
8828 [ + + ]: 2989 : for (i = 0; i < index->nkeycolumns; i++)
8829 : : {
8830 [ + + + + ]: 1726 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
8831 : : {
8832 : 273 : fullIndexScan = true;
8833 : 273 : break;
8834 : : }
8835 : : }
8836 : :
8837 [ + + - + ]: 1536 : if (fullIndexScan || indexQuals == NIL)
8838 : : {
8839 : : /*
8840 : : * Full index scan will be required. We treat this as if every key in
8841 : : * the index had been listed in the query; is that reasonable?
8842 : : */
5250 tgl@sss.pgh.pa.us 8843 : 273 : counts.partialEntries = 0;
8844 : 273 : counts.exactEntries = numEntries;
8845 : 273 : counts.searchEntries = numEntries;
8846 : : }
8847 : :
8848 : : /* Will we have more than one iteration of a nestloop scan? */
5212 8849 : 1536 : outer_scans = loop_count;
8850 : :
8851 : : /*
8852 : : * Compute cost to begin scan, first of all, pay attention to pending
8853 : : * list.
8854 : : */
5679 8855 : 1536 : entryPagesFetched = numPendingPages;
8856 : :
8857 : : /*
8858 : : * Estimate number of entry pages read. We need to do
8859 : : * counts.searchEntries searches. Use a power function as it should be,
8860 : : * but tuples on leaf pages usually is much greater. Here we include all
8861 : : * searches in entry tree, including search of first entry in partial
8862 : : * match algorithm
8863 : : */
5250 8864 : 1536 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8865 : :
8866 : : /*
8867 : : * Add an estimate of entry pages read by partial match algorithm. It's a
8868 : : * scan over leaf pages in entry tree. We haven't any useful stats here,
8869 : : * so estimate it as proportion. Because counts.partialEntries is really
8870 : : * pretty bogus (see code above), it's possible that it is more than
8871 : : * numEntries; clamp the proportion to ensure sanity.
8872 : : */
3777 8873 : 1536 : partialScale = counts.partialEntries / numEntries;
8874 [ + + ]: 1536 : partialScale = Min(partialScale, 1.0);
8875 : :
8876 : 1536 : entryPagesFetched += ceil(numEntryPages * partialScale);
8877 : :
8878 : : /*
8879 : : * Partial match algorithm reads all data pages before doing actual scan,
8880 : : * so it's a startup cost. Again, we haven't any useful stats here, so
8881 : : * estimate it as proportion.
8882 : : */
8883 : 1536 : dataPagesFetched = ceil(numDataPages * partialScale);
8884 : :
1213 akorotkov@postgresql 8885 : 1536 : *indexStartupCost = 0;
8886 : 1536 : *indexTotalCost = 0;
8887 : :
8888 : : /*
8889 : : * Add a CPU-cost component to represent the costs of initial entry btree
8890 : : * descent. We don't charge any I/O cost for touching upper btree levels,
8891 : : * since they tend to stay in cache, but we still have to do about log2(N)
8892 : : * comparisons to descend a btree of N leaf tuples. We charge one
8893 : : * cpu_operator_cost per comparison.
8894 : : *
8895 : : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
8896 : : * ones after the first one are not startup cost so far as the overall
8897 : : * plan is concerned, so add them only to "total" cost.
8898 : : */
8899 [ + - ]: 1536 : if (numEntries > 1) /* avoid computing log(0) */
8900 : : {
8901 : 1536 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
8902 : 1536 : *indexStartupCost += descentCost * counts.searchEntries;
8903 : 1536 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
8904 : : }
8905 : :
8906 : : /*
8907 : : * Add a cpu cost per entry-page fetched. This is not amortized over a
8908 : : * loop.
8909 : : */
8910 : 1536 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8911 : 1536 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8912 : :
8913 : : /*
8914 : : * Add a cpu cost per data-page fetched. This is also not amortized over a
8915 : : * loop. Since those are the data pages from the partial match algorithm,
8916 : : * charge them as startup cost.
8917 : : */
8918 : 1536 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
8919 : :
8920 : : /*
8921 : : * Since we add the startup cost to the total cost later on, remove the
8922 : : * initial arrayscan from the total.
8923 : : */
8924 : 1536 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8925 : :
8926 : : /*
8927 : : * Calculate cache effects if more than one scan due to nestloops or array
8928 : : * quals. The result is pro-rated per nestloop scan, but the array qual
8929 : : * factor shouldn't be pro-rated (compare genericcostestimate).
8930 : : */
5250 tgl@sss.pgh.pa.us 8931 [ + - + + ]: 1536 : if (outer_scans > 1 || counts.arrayScans > 1)
8932 : : {
8933 : 5 : entryPagesFetched *= outer_scans * counts.arrayScans;
5679 8934 : 5 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
8935 : : (BlockNumber) numEntryPages,
8936 : : numEntryPages, root);
5250 8937 : 5 : entryPagesFetched /= outer_scans;
8938 : 5 : dataPagesFetched *= outer_scans * counts.arrayScans;
5679 8939 : 5 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8940 : : (BlockNumber) numDataPages,
8941 : : numDataPages, root);
5250 8942 : 5 : dataPagesFetched /= outer_scans;
8943 : : }
8944 : :
8945 : : /*
8946 : : * Here we use random page cost because logically-close pages could be far
8947 : : * apart on disk.
8948 : : */
1213 akorotkov@postgresql 8949 : 1536 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
8950 : :
8951 : : /*
8952 : : * Now compute the number of data pages fetched during the scan.
8953 : : *
8954 : : * We assume every entry to have the same number of items, and that there
8955 : : * is no overlap between them. (XXX: tsvector and array opclasses collect
8956 : : * statistics on the frequency of individual keys; it would be nice to use
8957 : : * those here.)
8958 : : */
5250 tgl@sss.pgh.pa.us 8959 : 1536 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
8960 : :
8961 : : /*
8962 : : * If there is a lot of overlap among the entries, in particular if one of
8963 : : * the entries is very frequent, the above calculation can grossly
8964 : : * under-estimate. As a simple cross-check, calculate a lower bound based
8965 : : * on the overall selectivity of the quals. At a minimum, we must read
8966 : : * one item pointer for each matching entry.
8967 : : *
8968 : : * The width of each item pointer varies, based on the level of
8969 : : * compression. We don't have statistics on that, but an average of
8970 : : * around 3 bytes per item is fairly typical.
8971 : : */
5679 8972 : 1536 : dataPagesFetchedBySel = ceil(*indexSelectivity *
4437 heikki.linnakangas@i 8973 : 1536 : (numTuples / (BLCKSZ / 3)));
5679 tgl@sss.pgh.pa.us 8974 [ + + ]: 1536 : if (dataPagesFetchedBySel > dataPagesFetched)
8975 : 1243 : dataPagesFetched = dataPagesFetchedBySel;
8976 : :
8977 : : /* Add one page cpu-cost to the startup cost */
1213 akorotkov@postgresql 8978 : 1536 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
8979 : :
8980 : : /*
8981 : : * Add once again a CPU-cost for those data pages, before amortizing for
8982 : : * cache.
8983 : : */
8984 : 1536 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8985 : :
8986 : : /* Account for cache effects, the same as above */
5250 tgl@sss.pgh.pa.us 8987 [ + - + + ]: 1536 : if (outer_scans > 1 || counts.arrayScans > 1)
8988 : : {
8989 : 5 : dataPagesFetched *= outer_scans * counts.arrayScans;
5679 8990 : 5 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8991 : : (BlockNumber) numDataPages,
8992 : : numDataPages, root);
5250 8993 : 5 : dataPagesFetched /= outer_scans;
8994 : : }
8995 : :
8996 : : /* And apply random_page_cost as the cost per page */
1213 akorotkov@postgresql 8997 : 1536 : *indexTotalCost += *indexStartupCost +
5679 tgl@sss.pgh.pa.us 8998 : 1536 : dataPagesFetched * spc_random_page_cost;
8999 : :
9000 : : /*
9001 : : * Add on index qual eval costs, much as in genericcostestimate. We charge
9002 : : * cpu but we can disregard indexorderbys, since GIN doesn't support
9003 : : * those.
9004 : : */
2636 9005 : 1536 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
9006 : 1536 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
9007 : :
5679 9008 : 1536 : *indexStartupCost += qual_arg_cost;
9009 : 1536 : *indexTotalCost += qual_arg_cost;
9010 : :
9011 : : /*
9012 : : * Add a cpu cost per search entry, corresponding to the actual visited
9013 : : * entries.
9014 : : */
1213 akorotkov@postgresql 9015 : 1536 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
9016 : : /* Now add a cpu cost per tuple in the posting lists / trees */
9017 : 1536 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
3366 rhaas@postgresql.org 9018 : 1536 : *indexPages = dataPagesFetched;
9019 : : }
9020 : :
9021 : : /*
9022 : : * BRIN has search behavior completely different from other index types
9023 : : */
9024 : : void
3761 tgl@sss.pgh.pa.us 9025 : 8944 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
9026 : : Cost *indexStartupCost, Cost *indexTotalCost,
9027 : : Selectivity *indexSelectivity, double *indexCorrelation,
9028 : : double *indexPages)
9029 : : {
4197 alvherre@alvh.no-ip. 9030 : 8944 : IndexOptInfo *index = path->indexinfo;
2636 tgl@sss.pgh.pa.us 9031 : 8944 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
4197 alvherre@alvh.no-ip. 9032 : 8944 : double numPages = index->pages;
3316 9033 : 8944 : RelOptInfo *baserel = index->rel;
9034 [ + - ]: 8944 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
9035 : : Cost spc_seq_page_cost;
9036 : : Cost spc_random_page_cost;
9037 : : double qual_arg_cost;
9038 : : double qualSelectivity;
9039 : : BrinStatsData statsData;
9040 : : double indexRanges;
9041 : : double minimalRanges;
9042 : : double estimatedRanges;
9043 : : double selec;
9044 : : Relation indexRel;
9045 : : ListCell *l;
9046 : : VariableStatData vardata;
9047 : :
9048 [ - + ]: 8944 : Assert(rte->rtekind == RTE_RELATION);
9049 : :
9050 : : /* fetch estimated page cost for the tablespace containing the index */
4197 9051 : 8944 : get_tablespace_page_costs(index->reltablespace,
9052 : : &spc_random_page_cost,
9053 : : &spc_seq_page_cost);
9054 : :
9055 : : /*
9056 : : * Obtain some data from the index itself, if possible. Otherwise invent
9057 : : * some plausible internal statistics based on the relation page count.
9058 : : */
2357 michael@paquier.xyz 9059 [ + - ]: 8944 : if (!index->hypothetical)
9060 : : {
9061 : : /*
9062 : : * A lock should have already been obtained on the index in plancat.c.
9063 : : */
9064 : 8944 : indexRel = index_open(index->indexoid, NoLock);
9065 : 8944 : brinGetStats(indexRel, &statsData);
9066 : 8944 : index_close(indexRel, NoLock);
9067 : :
9068 : : /* work out the actual number of ranges in the index */
9069 [ + + ]: 8944 : indexRanges = Max(ceil((double) baserel->pages /
9070 : : statsData.pagesPerRange), 1.0);
9071 : : }
9072 : : else
9073 : : {
9074 : : /*
9075 : : * Assume default number of pages per range, and estimate the number
9076 : : * of ranges based on that.
9077 : : */
2357 michael@paquier.xyz 9078 [ # # ]:UBC 0 : indexRanges = Max(ceil((double) baserel->pages /
9079 : : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
9080 : :
9081 : 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
9082 : 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
9083 : : }
9084 : :
9085 : : /*
9086 : : * Compute index correlation
9087 : : *
9088 : : * Because we can use all index quals equally when scanning, we can use
9089 : : * the largest correlation (in absolute value) among columns used by the
9090 : : * query. Start at zero, the worst possible case. If we cannot find any
9091 : : * correlation statistics, we will keep it as 0.
9092 : : */
3316 alvherre@alvh.no-ip. 9093 :CBC 8944 : *indexCorrelation = 0;
9094 : :
2636 tgl@sss.pgh.pa.us 9095 [ + - + + : 17889 : foreach(l, path->indexclauses)
+ + ]
9096 : : {
9097 : 8945 : IndexClause *iclause = lfirst_node(IndexClause, l);
9098 : 8945 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
9099 : :
9100 : : /* attempt to lookup stats in relation for this index column */
3316 alvherre@alvh.no-ip. 9101 [ + - ]: 8945 : if (attnum != 0)
9102 : : {
9103 : : /* Simple variable -- look to stats for the underlying table */
9104 [ - + - - ]: 8945 : if (get_relation_stats_hook &&
3316 alvherre@alvh.no-ip. 9105 :UBC 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
9106 : : {
9107 : : /*
9108 : : * The hook took control of acquiring a stats tuple. If it
9109 : : * did supply a tuple, it'd better have supplied a freefunc.
9110 : : */
9111 [ # # # # ]: 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
9112 [ # # ]: 0 : elog(ERROR,
9113 : : "no function provided to release variable stats with");
9114 : : }
9115 : : else
9116 : : {
3316 alvherre@alvh.no-ip. 9117 :CBC 8945 : vardata.statsTuple =
9118 : 8945 : SearchSysCache3(STATRELATTINH,
9119 : : ObjectIdGetDatum(rte->relid),
9120 : : Int16GetDatum(attnum),
9121 : : BoolGetDatum(false));
9122 : 8945 : vardata.freefunc = ReleaseSysCache;
9123 : : }
9124 : : }
9125 : : else
9126 : : {
9127 : : /*
9128 : : * Looks like we've found an expression column in the index. Let's
9129 : : * see if there's any stats for it.
9130 : : */
9131 : :
9132 : : /* get the attnum from the 0-based index. */
2636 tgl@sss.pgh.pa.us 9133 :UBC 0 : attnum = iclause->indexcol + 1;
9134 : :
3316 alvherre@alvh.no-ip. 9135 [ # # # # ]: 0 : if (get_index_stats_hook &&
3240 tgl@sss.pgh.pa.us 9136 : 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
9137 : : {
9138 : : /*
9139 : : * The hook took control of acquiring a stats tuple. If it
9140 : : * did supply a tuple, it'd better have supplied a freefunc.
9141 : : */
3316 alvherre@alvh.no-ip. 9142 [ # # ]: 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
9143 [ # # ]: 0 : !vardata.freefunc)
9144 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
9145 : : }
9146 : : else
9147 : : {
9148 : 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
9149 : : ObjectIdGetDatum(index->indexoid),
9150 : : Int16GetDatum(attnum),
9151 : : BoolGetDatum(false));
9152 : 0 : vardata.freefunc = ReleaseSysCache;
9153 : : }
9154 : : }
9155 : :
3316 alvherre@alvh.no-ip. 9156 [ + + ]:CBC 8945 : if (HeapTupleIsValid(vardata.statsTuple))
9157 : : {
9158 : : AttStatsSlot sslot;
9159 : :
3279 tgl@sss.pgh.pa.us 9160 [ + - ]: 33 : if (get_attstatsslot(&sslot, vardata.statsTuple,
9161 : : STATISTIC_KIND_CORRELATION, InvalidOid,
9162 : : ATTSTATSSLOT_NUMBERS))
9163 : : {
3316 alvherre@alvh.no-ip. 9164 : 33 : double varCorrelation = 0.0;
9165 : :
3279 tgl@sss.pgh.pa.us 9166 [ + - ]: 33 : if (sslot.nnumbers > 0)
1306 peter@eisentraut.org 9167 : 33 : varCorrelation = fabs(sslot.numbers[0]);
9168 : :
3316 alvherre@alvh.no-ip. 9169 [ + - ]: 33 : if (varCorrelation > *indexCorrelation)
9170 : 33 : *indexCorrelation = varCorrelation;
9171 : :
3279 tgl@sss.pgh.pa.us 9172 : 33 : free_attstatsslot(&sslot);
9173 : : }
9174 : : }
9175 : :
3316 alvherre@alvh.no-ip. 9176 [ + + ]: 8945 : ReleaseVariableStats(vardata);
9177 : : }
9178 : :
9179 : 8944 : qualSelectivity = clauselist_selectivity(root, indexQuals,
9180 : 8944 : baserel->relid,
9181 : : JOIN_INNER, NULL);
9182 : :
9183 : : /*
9184 : : * Now calculate the minimum possible ranges we could match with if all of
9185 : : * the rows were in the perfect order in the table's heap.
9186 : : */
9187 : 8944 : minimalRanges = ceil(indexRanges * qualSelectivity);
9188 : :
9189 : : /*
9190 : : * Now estimate the number of ranges that we'll touch by using the
9191 : : * indexCorrelation from the stats. Careful not to divide by zero (note
9192 : : * we're using the absolute value of the correlation).
9193 : : */
9194 [ + + ]: 8944 : if (*indexCorrelation < 1.0e-10)
9195 : 8911 : estimatedRanges = indexRanges;
9196 : : else
9197 [ + + ]: 33 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
9198 : :
9199 : : /* we expect to visit this portion of the table */
9200 : 8944 : selec = estimatedRanges / indexRanges;
9201 : :
9202 [ - + - + ]: 8944 : CLAMP_PROBABILITY(selec);
9203 : :
9204 : 8944 : *indexSelectivity = selec;
9205 : :
9206 : : /*
9207 : : * Compute the index qual costs, much as in genericcostestimate, to add to
9208 : : * the index costs. We can disregard indexorderbys, since BRIN doesn't
9209 : : * support those.
9210 : : */
2636 tgl@sss.pgh.pa.us 9211 : 8944 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
9212 : :
9213 : : /*
9214 : : * Compute the startup cost as the cost to read the whole revmap
9215 : : * sequentially, including the cost to execute the index quals.
9216 : : */
3316 alvherre@alvh.no-ip. 9217 : 8944 : *indexStartupCost =
9218 : 8944 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
4197 9219 : 8944 : *indexStartupCost += qual_arg_cost;
9220 : :
9221 : : /*
9222 : : * To read a BRIN index there might be a bit of back and forth over
9223 : : * regular pages, as revmap might point to them out of sequential order;
9224 : : * calculate the total cost as reading the whole index in random order.
9225 : : */
3316 9226 : 8944 : *indexTotalCost = *indexStartupCost +
9227 : 8944 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
9228 : :
9229 : : /*
9230 : : * Charge a small amount per range tuple which we expect to match to. This
9231 : : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
9232 : : * will set a bit for each page in the range when we find a matching
9233 : : * range, so we must multiply the charge by the number of pages in the
9234 : : * range.
9235 : : */
9236 : 8944 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
9237 : 8944 : statsData.pagesPerRange;
9238 : :
9239 : 8944 : *indexPages = index->pages;
4197 9240 : 8944 : }
|
