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) 1996-2024, PostgreSQL Global Development Group
14 : : * Portions Copyright (c) 1994, Regents of the University of California
15 : : *
16 : : *
17 : : * IDENTIFICATION
18 : : * src/backend/utils/adt/selfuncs.c
19 : : *
20 : : *-------------------------------------------------------------------------
21 : : */
22 : :
23 : : /*----------
24 : : * Operator selectivity estimation functions are called to estimate the
25 : : * selectivity of WHERE clauses whose top-level operator is their operator.
26 : : * We divide the problem into two cases:
27 : : * Restriction clause estimation: the clause involves vars of just
28 : : * one relation.
29 : : * Join clause estimation: the clause involves vars of multiple rels.
30 : : * Join selectivity estimation is far more difficult and usually less accurate
31 : : * than restriction estimation.
32 : : *
33 : : * When dealing with the inner scan of a nestloop join, we consider the
34 : : * join's joinclauses as restriction clauses for the inner relation, and
35 : : * treat vars of the outer relation as parameters (a/k/a constants of unknown
36 : : * values). So, restriction estimators need to be able to accept an argument
37 : : * telling which relation is to be treated as the variable.
38 : : *
39 : : * The call convention for a restriction estimator (oprrest function) is
40 : : *
41 : : * Selectivity oprrest (PlannerInfo *root,
42 : : * Oid operator,
43 : : * List *args,
44 : : * int varRelid);
45 : : *
46 : : * root: general information about the query (rtable and RelOptInfo lists
47 : : * are particularly important for the estimator).
48 : : * operator: OID of the specific operator in question.
49 : : * args: argument list from the operator clause.
50 : : * varRelid: if not zero, the relid (rtable index) of the relation to
51 : : * be treated as the variable relation. May be zero if the args list
52 : : * is known to contain vars of only one relation.
53 : : *
54 : : * This is represented at the SQL level (in pg_proc) as
55 : : *
56 : : * float8 oprrest (internal, oid, internal, int4);
57 : : *
58 : : * The result is a selectivity, that is, a fraction (0 to 1) of the rows
59 : : * of the relation that are expected to produce a TRUE result for the
60 : : * given operator.
61 : : *
62 : : * The call convention for a join estimator (oprjoin function) is similar
63 : : * except that varRelid is not needed, and instead join information is
64 : : * supplied:
65 : : *
66 : : * Selectivity oprjoin (PlannerInfo *root,
67 : : * Oid operator,
68 : : * List *args,
69 : : * JoinType jointype,
70 : : * SpecialJoinInfo *sjinfo);
71 : : *
72 : : * float8 oprjoin (internal, oid, internal, int2, internal);
73 : : *
74 : : * (Before Postgres 8.4, join estimators had only the first four of these
75 : : * parameters. That signature is still allowed, but deprecated.) The
76 : : * relationship between jointype and sjinfo is explained in the comments for
77 : : * clause_selectivity() --- the short version is that jointype is usually
78 : : * best ignored in favor of examining sjinfo.
79 : : *
80 : : * Join selectivity for regular inner and outer joins is defined as the
81 : : * fraction (0 to 1) of the cross product of the relations that is expected
82 : : * to produce a TRUE result for the given operator. For both semi and anti
83 : : * joins, however, the selectivity is defined as the fraction of the left-hand
84 : : * side relation's rows that are expected to have a match (ie, at least one
85 : : * row with a TRUE result) in the right-hand side.
86 : : *
87 : : * For both oprrest and oprjoin functions, the operator's input collation OID
88 : : * (if any) is passed using the standard fmgr mechanism, so that the estimator
89 : : * function can fetch it with PG_GET_COLLATION(). Note, however, that all
90 : : * statistics in pg_statistic are currently built using the relevant column's
91 : : * collation.
92 : : *----------
93 : : */
94 : :
95 : : #include "postgres.h"
96 : :
97 : : #include <ctype.h>
98 : : #include <math.h>
99 : :
100 : : #include "access/brin.h"
101 : : #include "access/brin_page.h"
102 : : #include "access/gin.h"
103 : : #include "access/table.h"
104 : : #include "access/tableam.h"
105 : : #include "access/visibilitymap.h"
106 : : #include "catalog/pg_am.h"
107 : : #include "catalog/pg_collation.h"
108 : : #include "catalog/pg_operator.h"
109 : : #include "catalog/pg_statistic.h"
110 : : #include "catalog/pg_statistic_ext.h"
111 : : #include "executor/nodeAgg.h"
112 : : #include "miscadmin.h"
113 : : #include "nodes/makefuncs.h"
114 : : #include "nodes/nodeFuncs.h"
115 : : #include "optimizer/clauses.h"
116 : : #include "optimizer/cost.h"
117 : : #include "optimizer/optimizer.h"
118 : : #include "optimizer/pathnode.h"
119 : : #include "optimizer/paths.h"
120 : : #include "optimizer/plancat.h"
121 : : #include "parser/parse_clause.h"
122 : : #include "parser/parse_relation.h"
123 : : #include "parser/parsetree.h"
124 : : #include "statistics/statistics.h"
125 : : #include "storage/bufmgr.h"
126 : : #include "utils/acl.h"
127 : : #include "utils/array.h"
128 : : #include "utils/builtins.h"
129 : : #include "utils/date.h"
130 : : #include "utils/datum.h"
131 : : #include "utils/fmgroids.h"
132 : : #include "utils/index_selfuncs.h"
133 : : #include "utils/lsyscache.h"
134 : : #include "utils/memutils.h"
135 : : #include "utils/pg_locale.h"
136 : : #include "utils/rel.h"
137 : : #include "utils/selfuncs.h"
138 : : #include "utils/snapmgr.h"
139 : : #include "utils/spccache.h"
140 : : #include "utils/syscache.h"
141 : : #include "utils/timestamp.h"
142 : : #include "utils/typcache.h"
143 : :
144 : : #define DEFAULT_PAGE_CPU_MULTIPLIER 50.0
145 : :
146 : : /* Hooks for plugins to get control when we ask for stats */
147 : : get_relation_stats_hook_type get_relation_stats_hook = NULL;
148 : : get_index_stats_hook_type get_index_stats_hook = NULL;
149 : :
150 : : static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
151 : : static double eqjoinsel_inner(Oid opfuncoid, Oid collation,
152 : : VariableStatData *vardata1, VariableStatData *vardata2,
153 : : double nd1, double nd2,
154 : : bool isdefault1, bool isdefault2,
155 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
156 : : Form_pg_statistic stats1, Form_pg_statistic stats2,
157 : : bool have_mcvs1, bool have_mcvs2);
158 : : static double eqjoinsel_semi(Oid opfuncoid, Oid collation,
159 : : VariableStatData *vardata1, VariableStatData *vardata2,
160 : : double nd1, double nd2,
161 : : bool isdefault1, bool isdefault2,
162 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
163 : : Form_pg_statistic stats1, Form_pg_statistic stats2,
164 : : bool have_mcvs1, bool have_mcvs2,
165 : : RelOptInfo *inner_rel);
166 : : static bool estimate_multivariate_ndistinct(PlannerInfo *root,
167 : : RelOptInfo *rel, List **varinfos, double *ndistinct);
168 : : static bool convert_to_scalar(Datum value, Oid valuetypid, Oid collid,
169 : : double *scaledvalue,
170 : : Datum lobound, Datum hibound, Oid boundstypid,
171 : : double *scaledlobound, double *scaledhibound);
172 : : static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure);
173 : : static void convert_string_to_scalar(char *value,
174 : : double *scaledvalue,
175 : : char *lobound,
176 : : double *scaledlobound,
177 : : char *hibound,
178 : : double *scaledhibound);
179 : : static void convert_bytea_to_scalar(Datum value,
180 : : double *scaledvalue,
181 : : Datum lobound,
182 : : double *scaledlobound,
183 : : Datum hibound,
184 : : double *scaledhibound);
185 : : static double convert_one_string_to_scalar(char *value,
186 : : int rangelo, int rangehi);
187 : : static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
188 : : int rangelo, int rangehi);
189 : : static char *convert_string_datum(Datum value, Oid typid, Oid collid,
190 : : bool *failure);
191 : : static double convert_timevalue_to_scalar(Datum value, Oid typid,
192 : : bool *failure);
193 : : static void examine_simple_variable(PlannerInfo *root, Var *var,
194 : : VariableStatData *vardata);
195 : : static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata,
196 : : Oid sortop, Oid collation,
197 : : Datum *min, Datum *max);
198 : : static void get_stats_slot_range(AttStatsSlot *sslot,
199 : : Oid opfuncoid, FmgrInfo *opproc,
200 : : Oid collation, int16 typLen, bool typByVal,
201 : : Datum *min, Datum *max, bool *p_have_data);
202 : : static bool get_actual_variable_range(PlannerInfo *root,
203 : : VariableStatData *vardata,
204 : : Oid sortop, Oid collation,
205 : : Datum *min, Datum *max);
206 : : static bool get_actual_variable_endpoint(Relation heapRel,
207 : : Relation indexRel,
208 : : ScanDirection indexscandir,
209 : : ScanKey scankeys,
210 : : int16 typLen,
211 : : bool typByVal,
212 : : TupleTableSlot *tableslot,
213 : : MemoryContext outercontext,
214 : : Datum *endpointDatum);
215 : : static RelOptInfo *find_join_input_rel(PlannerInfo *root, Relids relids);
216 : :
217 : :
218 : : /*
219 : : * eqsel - Selectivity of "=" for any data types.
220 : : *
221 : : * Note: this routine is also used to estimate selectivity for some
222 : : * operators that are not "=" but have comparable selectivity behavior,
223 : : * such as "~=" (geometric approximate-match). Even for "=", we must
224 : : * keep in mind that the left and right datatypes may differ.
225 : : */
226 : : Datum
8714 tgl@sss.pgh.pa.us 227 :CBC 262472 : eqsel(PG_FUNCTION_ARGS)
228 : : {
2507 229 : 262472 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
230 : : }
231 : :
232 : : /*
233 : : * Common code for eqsel() and neqsel()
234 : : */
235 : : static double
236 : 282343 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
237 : : {
6888 238 : 282343 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
8365 239 : 282343 : Oid operator = PG_GETARG_OID(1);
240 : 282343 : List *args = (List *) PG_GETARG_POINTER(2);
241 : 282343 : int varRelid = PG_GETARG_INT32(3);
1409 242 : 282343 : Oid collation = PG_GET_COLLATION();
243 : : VariableStatData vardata;
244 : : Node *other;
245 : : bool varonleft;
246 : : double selec;
247 : :
248 : : /*
249 : : * When asked about <>, we do the estimation using the corresponding =
250 : : * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
251 : : */
2507 252 [ + + ]: 282343 : if (negate)
253 : : {
254 : 19871 : operator = get_negator(operator);
255 [ - + ]: 19871 : if (!OidIsValid(operator))
256 : : {
257 : : /* Use default selectivity (should we raise an error instead?) */
2507 tgl@sss.pgh.pa.us 258 :UBC 0 : return 1.0 - DEFAULT_EQ_SEL;
259 : : }
260 : : }
261 : :
262 : : /*
263 : : * If expression is not variable = something or something = variable, then
264 : : * punt and return a default estimate.
265 : : */
7362 tgl@sss.pgh.pa.us 266 [ + + ]:CBC 282343 : if (!get_restriction_variable(root, args, varRelid,
267 : : &vardata, &other, &varonleft))
2507 268 [ + + ]: 1851 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
269 : :
270 : : /*
271 : : * We can do a lot better if the something is a constant. (Note: the
272 : : * Const might result from estimation rather than being a simple constant
273 : : * in the query.)
274 : : */
5881 275 [ + + ]: 280492 : if (IsA(other, Const))
1409 276 : 123238 : selec = var_eq_const(&vardata, operator, collation,
5881 277 : 123238 : ((Const *) other)->constvalue,
278 : 123238 : ((Const *) other)->constisnull,
279 : : varonleft, negate);
280 : : else
1409 281 : 157254 : selec = var_eq_non_const(&vardata, operator, collation, other,
282 : : varonleft, negate);
283 : :
5881 284 [ + + ]: 280492 : ReleaseVariableStats(vardata);
285 : :
2507 286 : 280492 : return selec;
287 : : }
288 : :
289 : : /*
290 : : * var_eq_const --- eqsel for var = const case
291 : : *
292 : : * This is exported so that some other estimation functions can use it.
293 : : */
294 : : double
572 pg@bowt.ie 295 : 144181 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
296 : : Datum constval, bool constisnull,
297 : : bool varonleft, bool negate)
298 : : {
299 : : double selec;
2507 tgl@sss.pgh.pa.us 300 : 144181 : double nullfrac = 0.0;
301 : : bool isdefault;
302 : : Oid opfuncoid;
303 : :
304 : : /*
305 : : * If the constant is NULL, assume operator is strict and return zero, ie,
306 : : * operator will never return TRUE. (It's zero even for a negator op.)
307 : : */
5881 308 [ + + ]: 144181 : if (constisnull)
309 : 160 : return 0.0;
310 : :
311 : : /*
312 : : * Grab the nullfrac for use below. Note we allow use of nullfrac
313 : : * regardless of security check.
314 : : */
2507 315 [ + + ]: 144021 : if (HeapTupleIsValid(vardata->statsTuple))
316 : : {
317 : : Form_pg_statistic stats;
318 : :
319 : 106979 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
320 : 106979 : nullfrac = stats->stanullfrac;
321 : : }
322 : :
323 : : /*
324 : : * If we matched the var to a unique index or DISTINCT clause, assume
325 : : * there is exactly one match regardless of anything else. (This is
326 : : * slightly bogus, since the index or clause's equality operator might be
327 : : * different from ours, but it's much more likely to be right than
328 : : * ignoring the information.)
329 : : */
5537 330 [ + + + - : 144021 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
+ + ]
331 : : {
2507 332 : 38145 : selec = 1.0 / vardata->rel->tuples;
333 : : }
334 [ + + + - ]: 183169 : else if (HeapTupleIsValid(vardata->statsTuple) &&
335 : 77293 : statistic_proc_security_check(vardata,
572 pg@bowt.ie 336 : 77293 : (opfuncoid = get_opcode(oproid))))
9023 tgl@sss.pgh.pa.us 337 : 77293 : {
338 : : AttStatsSlot sslot;
5881 339 : 77293 : bool match = false;
340 : : int i;
341 : :
342 : : /*
343 : : * Is the constant "=" to any of the column's most common values?
344 : : * (Although the given operator may not really be "=", we will assume
345 : : * that seeing whether it returns TRUE is an appropriate test. If you
346 : : * don't like this, maybe you shouldn't be using eqsel for your
347 : : * operator...)
348 : : */
2528 349 [ + + ]: 77293 : if (get_attstatsslot(&sslot, vardata->statsTuple,
350 : : STATISTIC_KIND_MCV, InvalidOid,
351 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
352 : : {
1454 353 : 70362 : LOCAL_FCINFO(fcinfo, 2);
354 : : FmgrInfo eqproc;
355 : :
2536 peter_e@gmx.net 356 : 70362 : fmgr_info(opfuncoid, &eqproc);
357 : :
358 : : /*
359 : : * Save a few cycles by setting up the fcinfo struct just once.
360 : : * Using FunctionCallInvoke directly also avoids failure if the
361 : : * eqproc returns NULL, though really equality functions should
362 : : * never do that.
363 : : */
1409 tgl@sss.pgh.pa.us 364 : 70362 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
365 : : NULL, NULL);
1454 366 : 70362 : fcinfo->args[0].isnull = false;
367 : 70362 : fcinfo->args[1].isnull = false;
368 : : /* be careful to apply operator right way 'round */
369 [ + + ]: 70362 : if (varonleft)
370 : 70346 : fcinfo->args[1].value = constval;
371 : : else
372 : 16 : fcinfo->args[0].value = constval;
373 : :
2528 374 [ + + ]: 1063952 : for (i = 0; i < sslot.nvalues; i++)
375 : : {
376 : : Datum fresult;
377 : :
5881 378 [ + + ]: 1031053 : if (varonleft)
1454 379 : 1031025 : fcinfo->args[0].value = sslot.values[i];
380 : : else
381 : 28 : fcinfo->args[1].value = sslot.values[i];
382 : 1031053 : fcinfo->isnull = false;
383 : 1031053 : fresult = FunctionCallInvoke(fcinfo);
384 [ + - + + ]: 1031053 : if (!fcinfo->isnull && DatumGetBool(fresult))
385 : : {
386 : 37463 : match = true;
5881 387 : 37463 : break;
388 : : }
389 : : }
390 : : }
391 : : else
392 : : {
393 : : /* no most-common-value info available */
2528 394 : 6931 : i = 0; /* keep compiler quiet */
395 : : }
396 : :
5881 397 [ + + ]: 77293 : if (match)
398 : : {
399 : : /*
400 : : * Constant is "=" to this common value. We know selectivity
401 : : * exactly (or as exactly as ANALYZE could calculate it, anyway).
402 : : */
2528 403 : 37463 : selec = sslot.numbers[i];
404 : : }
405 : : else
406 : : {
407 : : /*
408 : : * Comparison is against a constant that is neither NULL nor any
409 : : * of the common values. Its selectivity cannot be more than
410 : : * this:
411 : : */
5881 412 : 39830 : double sumcommon = 0.0;
413 : : double otherdistinct;
414 : :
2528 415 [ + + ]: 934551 : for (i = 0; i < sslot.nnumbers; i++)
416 : 894721 : sumcommon += sslot.numbers[i];
2507 417 : 39830 : selec = 1.0 - sumcommon - nullfrac;
5881 418 [ + + - + ]: 39830 : CLAMP_PROBABILITY(selec);
419 : :
420 : : /*
421 : : * and in fact it's probably a good deal less. We approximate that
422 : : * all the not-common values share this remaining fraction
423 : : * equally, so we divide by the number of other distinct values.
424 : : */
2528 425 : 39830 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
426 : 39830 : sslot.nnumbers;
5881 427 [ + + ]: 39830 : if (otherdistinct > 1)
428 : 19925 : selec /= otherdistinct;
429 : :
430 : : /*
431 : : * Another cross-check: selectivity shouldn't be estimated as more
432 : : * than the least common "most common value".
433 : : */
2528 434 [ + + - + ]: 39830 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
2528 tgl@sss.pgh.pa.us 435 :UBC 0 : selec = sslot.numbers[sslot.nnumbers - 1];
436 : : }
437 : :
2528 tgl@sss.pgh.pa.us 438 :CBC 77293 : free_attstatsslot(&sslot);
439 : : }
440 : : else
441 : : {
442 : : /*
443 : : * No ANALYZE stats available, so make a guess using estimated number
444 : : * of distinct values and assuming they are equally common. (The guess
445 : : * is unlikely to be very good, but we do know a few special cases.)
446 : : */
4606 447 : 28583 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
448 : : }
449 : :
450 : : /* now adjust if we wanted <> rather than = */
2507 451 [ + + ]: 144021 : if (negate)
452 : 16135 : selec = 1.0 - selec - nullfrac;
453 : :
454 : : /* result should be in range, but make sure... */
5881 455 [ - + - + ]: 144021 : CLAMP_PROBABILITY(selec);
456 : :
457 : 144021 : return selec;
458 : : }
459 : :
460 : : /*
461 : : * var_eq_non_const --- eqsel for var = something-other-than-const case
462 : : *
463 : : * This is exported so that some other estimation functions can use it.
464 : : */
465 : : double
572 pg@bowt.ie 466 : 157254 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
467 : : Node *other,
468 : : bool varonleft, bool negate)
469 : : {
470 : : double selec;
2507 tgl@sss.pgh.pa.us 471 : 157254 : double nullfrac = 0.0;
472 : : bool isdefault;
473 : :
474 : : /*
475 : : * Grab the nullfrac for use below.
476 : : */
477 [ + + ]: 157254 : if (HeapTupleIsValid(vardata->statsTuple))
478 : : {
479 : : Form_pg_statistic stats;
480 : :
481 : 115519 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
482 : 115519 : nullfrac = stats->stanullfrac;
483 : : }
484 : :
485 : : /*
486 : : * If we matched the var to a unique index or DISTINCT clause, assume
487 : : * there is exactly one match regardless of anything else. (This is
488 : : * slightly bogus, since the index or clause's equality operator might be
489 : : * different from ours, but it's much more likely to be right than
490 : : * ignoring the information.)
491 : : */
5537 492 [ + + + - : 157254 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
+ + ]
493 : : {
2507 494 : 62151 : selec = 1.0 / vardata->rel->tuples;
495 : : }
496 [ + + ]: 95103 : else if (HeapTupleIsValid(vardata->statsTuple))
497 : : {
498 : : double ndistinct;
499 : : AttStatsSlot sslot;
500 : :
501 : : /*
502 : : * Search is for a value that we do not know a priori, but we will
503 : : * assume it is not NULL. Estimate the selectivity as non-null
504 : : * fraction divided by number of distinct values, so that we get a
505 : : * result averaged over all possible values whether common or
506 : : * uncommon. (Essentially, we are assuming that the not-yet-known
507 : : * comparison value is equally likely to be any of the possible
508 : : * values, regardless of their frequency in the table. Is that a good
509 : : * idea?)
510 : : */
511 : 59413 : selec = 1.0 - nullfrac;
4606 512 : 59413 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
5881 513 [ + + ]: 59413 : if (ndistinct > 1)
514 : 57357 : selec /= ndistinct;
515 : :
516 : : /*
517 : : * Cross-check: selectivity should never be estimated as more than the
518 : : * most common value's.
519 : : */
2528 520 [ + + ]: 59413 : if (get_attstatsslot(&sslot, vardata->statsTuple,
521 : : STATISTIC_KIND_MCV, InvalidOid,
522 : : ATTSTATSSLOT_NUMBERS))
523 : : {
524 [ + - + + ]: 50383 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
525 : 282 : selec = sslot.numbers[0];
526 : 50383 : free_attstatsslot(&sslot);
527 : : }
528 : : }
529 : : else
530 : : {
531 : : /*
532 : : * No ANALYZE stats available, so make a guess using estimated number
533 : : * of distinct values and assuming they are equally common. (The guess
534 : : * is unlikely to be very good, but we do know a few special cases.)
535 : : */
4606 536 : 35690 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
537 : : }
538 : :
539 : : /* now adjust if we wanted <> rather than = */
2507 540 [ + + ]: 157254 : if (negate)
541 : 2807 : selec = 1.0 - selec - nullfrac;
542 : :
543 : : /* result should be in range, but make sure... */
8137 544 [ - + - + ]: 157254 : CLAMP_PROBABILITY(selec);
545 : :
5881 546 : 157254 : return selec;
547 : : }
548 : :
549 : : /*
550 : : * neqsel - Selectivity of "!=" for any data types.
551 : : *
552 : : * This routine is also used for some operators that are not "!="
553 : : * but have comparable selectivity behavior. See above comments
554 : : * for eqsel().
555 : : */
556 : : Datum
8714 557 : 19871 : neqsel(PG_FUNCTION_ARGS)
558 : : {
2507 559 : 19871 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
560 : : }
561 : :
562 : : /*
563 : : * scalarineqsel - Selectivity of "<", "<=", ">", ">=" for scalars.
564 : : *
565 : : * This is the guts of scalarltsel/scalarlesel/scalargtsel/scalargesel.
566 : : * The isgt and iseq flags distinguish which of the four cases apply.
567 : : *
568 : : * The caller has commuted the clause, if necessary, so that we can treat
569 : : * the variable as being on the left. The caller must also make sure that
570 : : * the other side of the clause is a non-null Const, and dissect that into
571 : : * a value and datatype. (This definition simplifies some callers that
572 : : * want to estimate against a computed value instead of a Const node.)
573 : : *
574 : : * This routine works for any datatype (or pair of datatypes) known to
575 : : * convert_to_scalar(). If it is applied to some other datatype,
576 : : * it will return an approximate estimate based on assuming that the constant
577 : : * value falls in the middle of the bin identified by binary search.
578 : : */
579 : : static double
2405 580 : 136505 : scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
581 : : Oid collation,
582 : : VariableStatData *vardata, Datum constval, Oid consttype)
583 : : {
584 : : Form_pg_statistic stats;
585 : : FmgrInfo opproc;
586 : : double mcv_selec,
587 : : hist_selec,
588 : : sumcommon;
589 : : double selec;
590 : :
7362 591 [ + + ]: 136505 : if (!HeapTupleIsValid(vardata->statsTuple))
592 : : {
593 : : /*
594 : : * No stats are available. Typically this means we have to fall back
595 : : * on the default estimate; but if the variable is CTID then we can
596 : : * make an estimate based on comparing the constant to the table size.
597 : : */
1847 598 [ + - + + ]: 9657 : if (vardata->var && IsA(vardata->var, Var) &&
599 [ + + ]: 7531 : ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
600 : : {
601 : : ItemPointer itemptr;
602 : : double block;
603 : : double density;
604 : :
605 : : /*
606 : : * If the relation's empty, we're going to include all of it.
607 : : * (This is mostly to avoid divide-by-zero below.)
608 : : */
609 [ - + ]: 107 : if (vardata->rel->pages == 0)
1847 tgl@sss.pgh.pa.us 610 :UBC 0 : return 1.0;
611 : :
1847 tgl@sss.pgh.pa.us 612 :CBC 107 : itemptr = (ItemPointer) DatumGetPointer(constval);
613 : 107 : block = ItemPointerGetBlockNumberNoCheck(itemptr);
614 : :
615 : : /*
616 : : * Determine the average number of tuples per page (density).
617 : : *
618 : : * Since the last page will, on average, be only half full, we can
619 : : * estimate it to have half as many tuples as earlier pages. So
620 : : * give it half the weight of a regular page.
621 : : */
622 : 107 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
623 : :
624 : : /* If target is the last page, use half the density. */
625 [ + + ]: 107 : if (block >= vardata->rel->pages - 1)
626 : 15 : density *= 0.5;
627 : :
628 : : /*
629 : : * Using the average tuples per page, calculate how far into the
630 : : * page the itemptr is likely to be and adjust block accordingly,
631 : : * by adding that fraction of a whole block (but never more than a
632 : : * whole block, no matter how high the itemptr's offset is). Here
633 : : * we are ignoring the possibility of dead-tuple line pointers,
634 : : * which is fairly bogus, but we lack the info to do better.
635 : : */
636 [ + - ]: 107 : if (density > 0.0)
637 : : {
638 : 107 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
639 : :
640 [ + + ]: 107 : block += Min(offset / density, 1.0);
641 : : }
642 : :
643 : : /*
644 : : * Convert relative block number to selectivity. Again, the last
645 : : * page has only half weight.
646 : : */
647 : 107 : selec = block / (vardata->rel->pages - 0.5);
648 : :
649 : : /*
650 : : * The calculation so far gave us a selectivity for the "<=" case.
651 : : * We'll have one fewer tuple for "<" and one additional tuple for
652 : : * ">=", the latter of which we'll reverse the selectivity for
653 : : * below, so we can simply subtract one tuple for both cases. The
654 : : * cases that need this adjustment can be identified by iseq being
655 : : * equal to isgt.
656 : : */
657 [ + + + - ]: 107 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
658 : 51 : selec -= (1.0 / vardata->rel->tuples);
659 : :
660 : : /* Finally, reverse the selectivity for the ">", ">=" cases. */
661 [ + + ]: 107 : if (isgt)
662 : 50 : selec = 1.0 - selec;
663 : :
664 [ + + - + ]: 107 : CLAMP_PROBABILITY(selec);
665 : 107 : return selec;
666 : : }
667 : :
668 : : /* no stats available, so default result */
8365 669 : 9550 : return DEFAULT_INEQ_SEL;
670 : : }
7362 671 : 126848 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
672 : :
8365 673 : 126848 : fmgr_info(get_opcode(operator), &opproc);
674 : :
675 : : /*
676 : : * If we have most-common-values info, add up the fractions of the MCV
677 : : * entries that satisfy MCV OP CONST. These fractions contribute directly
678 : : * to the result selectivity. Also add up the total fraction represented
679 : : * by MCV entries.
680 : : */
1409 681 : 126848 : mcv_selec = mcv_selectivity(vardata, &opproc, collation, constval, true,
682 : : &sumcommon);
683 : :
684 : : /*
685 : : * If there is a histogram, determine which bin the constant falls in, and
686 : : * compute the resulting contribution to selectivity.
687 : : */
2405 688 : 126848 : hist_selec = ineq_histogram_selectivity(root, vardata,
689 : : operator, &opproc, isgt, iseq,
690 : : collation,
691 : : constval, consttype);
692 : :
693 : : /*
694 : : * Now merge the results from the MCV and histogram calculations,
695 : : * realizing that the histogram covers only the non-null values that are
696 : : * not listed in MCV.
697 : : */
6669 698 : 126848 : selec = 1.0 - stats->stanullfrac - sumcommon;
699 : :
5214 700 [ + + ]: 126848 : if (hist_selec >= 0.0)
6669 701 : 101098 : selec *= hist_selec;
702 : : else
703 : : {
704 : : /*
705 : : * If no histogram but there are values not accounted for by MCV,
706 : : * arbitrarily assume half of them will match.
707 : : */
708 : 25750 : selec *= 0.5;
709 : : }
710 : :
711 : 126848 : selec += mcv_selec;
712 : :
713 : : /* result should be in range, but make sure... */
714 [ + + + + ]: 126848 : CLAMP_PROBABILITY(selec);
715 : :
716 : 126848 : return selec;
717 : : }
718 : :
719 : : /*
720 : : * mcv_selectivity - Examine the MCV list for selectivity estimates
721 : : *
722 : : * Determine the fraction of the variable's MCV population that satisfies
723 : : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft. Also
724 : : * compute the fraction of the total column population represented by the MCV
725 : : * list. This code will work for any boolean-returning predicate operator.
726 : : *
727 : : * The function result is the MCV selectivity, and the fraction of the
728 : : * total population is returned into *sumcommonp. Zeroes are returned
729 : : * if there is no MCV list.
730 : : */
731 : : double
1409 732 : 129346 : mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation,
733 : : Datum constval, bool varonleft,
734 : : double *sumcommonp)
735 : : {
736 : : double mcv_selec,
737 : : sumcommon;
738 : : AttStatsSlot sslot;
739 : : int i;
740 : :
8378 741 : 129346 : mcv_selec = 0.0;
742 : 129346 : sumcommon = 0.0;
743 : :
6669 744 [ + + + + ]: 257558 : if (HeapTupleIsValid(vardata->statsTuple) &&
2536 peter_e@gmx.net 745 [ + + ]: 256403 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
2528 tgl@sss.pgh.pa.us 746 : 128191 : get_attstatsslot(&sslot, vardata->statsTuple,
747 : : STATISTIC_KIND_MCV, InvalidOid,
748 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
749 : : {
1454 750 : 58121 : LOCAL_FCINFO(fcinfo, 2);
751 : :
752 : : /*
753 : : * We invoke the opproc "by hand" so that we won't fail on NULL
754 : : * results. Such cases won't arise for normal comparison functions,
755 : : * but generic_restriction_selectivity could perhaps be used with
756 : : * operators that can return NULL. A small side benefit is to not
757 : : * need to re-initialize the fcinfo struct from scratch each time.
758 : : */
1409 759 : 58121 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
760 : : NULL, NULL);
1454 761 : 58121 : fcinfo->args[0].isnull = false;
762 : 58121 : fcinfo->args[1].isnull = false;
763 : : /* be careful to apply operator right way 'round */
764 [ + - ]: 58121 : if (varonleft)
765 : 58121 : fcinfo->args[1].value = constval;
766 : : else
1454 tgl@sss.pgh.pa.us 767 :UBC 0 : fcinfo->args[0].value = constval;
768 : :
2528 tgl@sss.pgh.pa.us 769 [ + + ]:CBC 1688628 : for (i = 0; i < sslot.nvalues; i++)
770 : : {
771 : : Datum fresult;
772 : :
1454 773 [ + - ]: 1630507 : if (varonleft)
774 : 1630507 : fcinfo->args[0].value = sslot.values[i];
775 : : else
1454 tgl@sss.pgh.pa.us 776 :UBC 0 : fcinfo->args[1].value = sslot.values[i];
1454 tgl@sss.pgh.pa.us 777 :CBC 1630507 : fcinfo->isnull = false;
778 : 1630507 : fresult = FunctionCallInvoke(fcinfo);
779 [ + - + + ]: 1630507 : if (!fcinfo->isnull && DatumGetBool(fresult))
2528 780 : 696010 : mcv_selec += sslot.numbers[i];
781 : 1630507 : sumcommon += sslot.numbers[i];
782 : : }
783 : 58121 : free_attstatsslot(&sslot);
784 : : }
785 : :
6669 786 : 129346 : *sumcommonp = sumcommon;
787 : 129346 : return mcv_selec;
788 : : }
789 : :
790 : : /*
791 : : * histogram_selectivity - Examine the histogram for selectivity estimates
792 : : *
793 : : * Determine the fraction of the variable's histogram entries that satisfy
794 : : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft.
795 : : *
796 : : * This code will work for any boolean-returning predicate operator, whether
797 : : * or not it has anything to do with the histogram sort operator. We are
798 : : * essentially using the histogram just as a representative sample. However,
799 : : * small histograms are unlikely to be all that representative, so the caller
800 : : * should be prepared to fall back on some other estimation approach when the
801 : : * histogram is missing or very small. It may also be prudent to combine this
802 : : * approach with another one when the histogram is small.
803 : : *
804 : : * If the actual histogram size is not at least min_hist_size, we won't bother
805 : : * to do the calculation at all. Also, if the n_skip parameter is > 0, we
806 : : * ignore the first and last n_skip histogram elements, on the grounds that
807 : : * they are outliers and hence not very representative. Typical values for
808 : : * these parameters are 10 and 1.
809 : : *
810 : : * The function result is the selectivity, or -1 if there is no histogram
811 : : * or it's smaller than min_hist_size.
812 : : *
813 : : * The output parameter *hist_size receives the actual histogram size,
814 : : * or zero if no histogram. Callers may use this number to decide how
815 : : * much faith to put in the function result.
816 : : *
817 : : * Note that the result disregards both the most-common-values (if any) and
818 : : * null entries. The caller is expected to combine this result with
819 : : * statistics for those portions of the column population. It may also be
820 : : * prudent to clamp the result range, ie, disbelieve exact 0 or 1 outputs.
821 : : */
822 : : double
1409 823 : 2498 : histogram_selectivity(VariableStatData *vardata,
824 : : FmgrInfo *opproc, Oid collation,
825 : : Datum constval, bool varonleft,
826 : : int min_hist_size, int n_skip,
827 : : int *hist_size)
828 : : {
829 : : double result;
830 : : AttStatsSlot sslot;
831 : :
832 : : /* check sanity of parameters */
6416 833 [ - + ]: 2498 : Assert(n_skip >= 0);
834 [ - + ]: 2498 : Assert(min_hist_size > 2 * n_skip);
835 : :
836 [ + + + + ]: 3862 : if (HeapTupleIsValid(vardata->statsTuple) &&
2536 peter_e@gmx.net 837 [ + + ]: 2725 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
2528 tgl@sss.pgh.pa.us 838 : 1361 : get_attstatsslot(&sslot, vardata->statsTuple,
839 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
840 : : ATTSTATSSLOT_VALUES))
841 : : {
842 : 1314 : *hist_size = sslot.nvalues;
843 [ + + ]: 1314 : if (sslot.nvalues >= min_hist_size)
844 : : {
1454 845 : 814 : LOCAL_FCINFO(fcinfo, 2);
6416 846 : 814 : int nmatch = 0;
847 : : int i;
848 : :
849 : : /*
850 : : * We invoke the opproc "by hand" so that we won't fail on NULL
851 : : * results. Such cases won't arise for normal comparison
852 : : * functions, but generic_restriction_selectivity could perhaps be
853 : : * used with operators that can return NULL. A small side benefit
854 : : * is to not need to re-initialize the fcinfo struct from scratch
855 : : * each time.
856 : : */
1409 857 : 814 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
858 : : NULL, NULL);
1454 859 : 814 : fcinfo->args[0].isnull = false;
860 : 814 : fcinfo->args[1].isnull = false;
861 : : /* be careful to apply operator right way 'round */
862 [ + - ]: 814 : if (varonleft)
863 : 814 : fcinfo->args[1].value = constval;
864 : : else
1454 tgl@sss.pgh.pa.us 865 :UBC 0 : fcinfo->args[0].value = constval;
866 : :
2528 tgl@sss.pgh.pa.us 867 [ + + ]:CBC 70800 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
868 : : {
869 : : Datum fresult;
870 : :
1454 871 [ + - ]: 69986 : if (varonleft)
872 : 69986 : fcinfo->args[0].value = sslot.values[i];
873 : : else
1454 tgl@sss.pgh.pa.us 874 :UBC 0 : fcinfo->args[1].value = sslot.values[i];
1454 tgl@sss.pgh.pa.us 875 :CBC 69986 : fcinfo->isnull = false;
876 : 69986 : fresult = FunctionCallInvoke(fcinfo);
877 [ + - + + ]: 69986 : if (!fcinfo->isnull && DatumGetBool(fresult))
6416 878 : 4469 : nmatch++;
879 : : }
2528 880 : 814 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
881 : : }
882 : : else
6416 883 : 500 : result = -1;
2528 884 : 1314 : free_attstatsslot(&sslot);
885 : : }
886 : : else
887 : : {
5880 888 : 1184 : *hist_size = 0;
6416 889 : 1184 : result = -1;
890 : : }
891 : :
892 : 2498 : return result;
893 : : }
894 : :
895 : : /*
896 : : * generic_restriction_selectivity - Selectivity for almost anything
897 : : *
898 : : * This function estimates selectivity for operators that we don't have any
899 : : * special knowledge about, but are on data types that we collect standard
900 : : * MCV and/or histogram statistics for. (Additional assumptions are that
901 : : * the operator is strict and immutable, or at least stable.)
902 : : *
903 : : * If we have "VAR OP CONST" or "CONST OP VAR", selectivity is estimated by
904 : : * applying the operator to each element of the column's MCV and/or histogram
905 : : * stats, and merging the results using the assumption that the histogram is
906 : : * a reasonable random sample of the column's non-MCV population. Note that
907 : : * if the operator's semantics are related to the histogram ordering, this
908 : : * might not be such a great assumption; other functions such as
909 : : * scalarineqsel() are probably a better match in such cases.
910 : : *
911 : : * Otherwise, fall back to the default selectivity provided by the caller.
912 : : */
913 : : double
1409 914 : 553 : generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation,
915 : : List *args, int varRelid,
916 : : double default_selectivity)
917 : : {
918 : : double selec;
919 : : VariableStatData vardata;
920 : : Node *other;
921 : : bool varonleft;
922 : :
923 : : /*
924 : : * If expression is not variable OP something or something OP variable,
925 : : * then punt and return the default estimate.
926 : : */
1474 927 [ - + ]: 553 : if (!get_restriction_variable(root, args, varRelid,
928 : : &vardata, &other, &varonleft))
1474 tgl@sss.pgh.pa.us 929 :UBC 0 : return default_selectivity;
930 : :
931 : : /*
932 : : * If the something is a NULL constant, assume operator is strict and
933 : : * return zero, ie, operator will never return TRUE.
934 : : */
1474 tgl@sss.pgh.pa.us 935 [ + - ]:CBC 553 : if (IsA(other, Const) &&
936 [ - + ]: 553 : ((Const *) other)->constisnull)
937 : : {
1474 tgl@sss.pgh.pa.us 938 [ # # ]:UBC 0 : ReleaseVariableStats(vardata);
939 : 0 : return 0.0;
940 : : }
941 : :
1474 tgl@sss.pgh.pa.us 942 [ + - ]:CBC 553 : if (IsA(other, Const))
943 : : {
944 : : /* Variable is being compared to a known non-null constant */
945 : 553 : Datum constval = ((Const *) other)->constvalue;
946 : : FmgrInfo opproc;
947 : : double mcvsum;
948 : : double mcvsel;
949 : : double nullfrac;
950 : : int hist_size;
951 : :
1454 952 : 553 : fmgr_info(get_opcode(oproid), &opproc);
953 : :
954 : : /*
955 : : * Calculate the selectivity for the column's most common values.
956 : : */
1409 957 : 553 : mcvsel = mcv_selectivity(&vardata, &opproc, collation,
958 : : constval, varonleft,
959 : : &mcvsum);
960 : :
961 : : /*
962 : : * If the histogram is large enough, see what fraction of it matches
963 : : * the query, and assume that's representative of the non-MCV
964 : : * population. Otherwise use the default selectivity for the non-MCV
965 : : * population.
966 : : */
967 : 553 : selec = histogram_selectivity(&vardata, &opproc, collation,
968 : : constval, varonleft,
969 : : 10, 1, &hist_size);
1474 970 [ + - ]: 553 : if (selec < 0)
971 : : {
972 : : /* Nope, fall back on default */
973 : 553 : selec = default_selectivity;
974 : : }
1474 tgl@sss.pgh.pa.us 975 [ # # ]:UBC 0 : else if (hist_size < 100)
976 : : {
977 : : /*
978 : : * For histogram sizes from 10 to 100, we combine the histogram
979 : : * and default selectivities, putting increasingly more trust in
980 : : * the histogram for larger sizes.
981 : : */
982 : 0 : double hist_weight = hist_size / 100.0;
983 : :
984 : 0 : selec = selec * hist_weight +
985 : 0 : default_selectivity * (1.0 - hist_weight);
986 : : }
987 : :
988 : : /* In any case, don't believe extremely small or large estimates. */
1474 tgl@sss.pgh.pa.us 989 [ - + ]:CBC 553 : if (selec < 0.0001)
1474 tgl@sss.pgh.pa.us 990 :UBC 0 : selec = 0.0001;
1474 tgl@sss.pgh.pa.us 991 [ - + ]:CBC 553 : else if (selec > 0.9999)
1474 tgl@sss.pgh.pa.us 992 :UBC 0 : selec = 0.9999;
993 : :
994 : : /* Don't forget to account for nulls. */
1474 tgl@sss.pgh.pa.us 995 [ + + ]:CBC 553 : if (HeapTupleIsValid(vardata.statsTuple))
996 : 42 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
997 : : else
998 : 511 : nullfrac = 0.0;
999 : :
1000 : : /*
1001 : : * Now merge the results from the MCV and histogram calculations,
1002 : : * realizing that the histogram covers only the non-null values that
1003 : : * are not listed in MCV.
1004 : : */
1005 : 553 : selec *= 1.0 - nullfrac - mcvsum;
1006 : 553 : selec += mcvsel;
1007 : : }
1008 : : else
1009 : : {
1010 : : /* Comparison value is not constant, so we can't do anything */
1474 tgl@sss.pgh.pa.us 1011 :UBC 0 : selec = default_selectivity;
1012 : : }
1013 : :
1474 tgl@sss.pgh.pa.us 1014 [ + + ]:CBC 553 : ReleaseVariableStats(vardata);
1015 : :
1016 : : /* result should be in range, but make sure... */
1017 [ - + - + ]: 553 : CLAMP_PROBABILITY(selec);
1018 : :
1019 : 553 : return selec;
1020 : : }
1021 : :
1022 : : /*
1023 : : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
1024 : : *
1025 : : * Determine the fraction of the variable's histogram population that
1026 : : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
1027 : : * The isgt and iseq flags distinguish which of the four cases apply.
1028 : : *
1029 : : * While opproc could be looked up from the operator OID, common callers
1030 : : * also need to call it separately, so we make the caller pass both.
1031 : : *
1032 : : * Returns -1 if there is no histogram (valid results will always be >= 0).
1033 : : *
1034 : : * Note that the result disregards both the most-common-values (if any) and
1035 : : * null entries. The caller is expected to combine this result with
1036 : : * statistics for those portions of the column population.
1037 : : *
1038 : : * This is exported so that some other estimation functions can use it.
1039 : : */
1040 : : double
5214 1041 : 128369 : ineq_histogram_selectivity(PlannerInfo *root,
1042 : : VariableStatData *vardata,
1043 : : Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq,
1044 : : Oid collation,
1045 : : Datum constval, Oid consttype)
1046 : : {
1047 : : double hist_selec;
1048 : : AttStatsSlot sslot;
1049 : :
1050 : 128369 : hist_selec = -1.0;
1051 : :
1052 : : /*
1053 : : * Someday, ANALYZE might store more than one histogram per rel/att,
1054 : : * corresponding to more than one possible sort ordering defined for the
1055 : : * column type. Right now, we know there is only one, so just grab it and
1056 : : * see if it matches the query.
1057 : : *
1058 : : * Note that we can't use opoid as search argument; the staop appearing in
1059 : : * pg_statistic will be for the relevant '<' operator, but what we have
1060 : : * might be some other inequality operator such as '>='. (Even if opoid
1061 : : * is a '<' operator, it could be cross-type.) Hence we must use
1062 : : * comparison_ops_are_compatible() to see if the operators match.
1063 : : */
6669 1064 [ + + + + ]: 256458 : if (HeapTupleIsValid(vardata->statsTuple) &&
2536 peter_e@gmx.net 1065 [ + + ]: 256160 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
2528 tgl@sss.pgh.pa.us 1066 : 128071 : get_attstatsslot(&sslot, vardata->statsTuple,
1067 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1068 : : ATTSTATSSLOT_VALUES))
1069 : : {
1409 1070 [ + - ]: 102338 : if (sslot.nvalues > 1 &&
1071 [ + + + + ]: 204638 : sslot.stacoll == collation &&
1072 : 102300 : comparison_ops_are_compatible(sslot.staop, opoid))
9022 1073 : 102246 : {
1074 : : /*
1075 : : * Use binary search to find the desired location, namely the
1076 : : * right end of the histogram bin containing the comparison value,
1077 : : * which is the leftmost entry for which the comparison operator
1078 : : * succeeds (if isgt) or fails (if !isgt).
1079 : : *
1080 : : * In this loop, we pay no attention to whether the operator iseq
1081 : : * or not; that detail will be mopped up below. (We cannot tell,
1082 : : * anyway, whether the operator thinks the values are equal.)
1083 : : *
1084 : : * If the binary search accesses the first or last histogram
1085 : : * entry, we try to replace that endpoint with the true column min
1086 : : * or max as found by get_actual_variable_range(). This
1087 : : * ameliorates misestimates when the min or max is moving as a
1088 : : * result of changes since the last ANALYZE. Note that this could
1089 : : * result in effectively including MCVs into the histogram that
1090 : : * weren't there before, but we don't try to correct for that.
1091 : : */
1092 : : double histfrac;
6402 bruce@momjian.us 1093 : 102246 : int lobound = 0; /* first possible slot to search */
2489 tgl@sss.pgh.pa.us 1094 : 102246 : int hibound = sslot.nvalues; /* last+1 slot to search */
5214 1095 : 102246 : bool have_end = false;
1096 : :
1097 : : /*
1098 : : * If there are only two histogram entries, we'll want up-to-date
1099 : : * values for both. (If there are more than two, we need at most
1100 : : * one of them to be updated, so we deal with that within the
1101 : : * loop.)
1102 : : */
2528 1103 [ + + ]: 102246 : if (sslot.nvalues == 2)
5214 1104 : 775 : have_end = get_actual_variable_range(root,
1105 : : vardata,
1106 : : sslot.staop,
1107 : : collation,
1108 : : &sslot.values[0],
2528 1109 : 775 : &sslot.values[1]);
1110 : :
6416 1111 [ + + ]: 691764 : while (lobound < hibound)
1112 : : {
6402 bruce@momjian.us 1113 : 589518 : int probe = (lobound + hibound) / 2;
1114 : : bool ltcmp;
1115 : :
1116 : : /*
1117 : : * If we find ourselves about to compare to the first or last
1118 : : * histogram entry, first try to replace it with the actual
1119 : : * current min or max (unless we already did so above).
1120 : : */
2528 tgl@sss.pgh.pa.us 1121 [ + + + + ]: 589518 : if (probe == 0 && sslot.nvalues > 2)
5214 1122 : 49997 : have_end = get_actual_variable_range(root,
1123 : : vardata,
1124 : : sslot.staop,
1125 : : collation,
1126 : : &sslot.values[0],
1127 : : NULL);
2528 1128 [ + + + + ]: 539521 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
5214 1129 : 34202 : have_end = get_actual_variable_range(root,
1130 : : vardata,
1131 : : sslot.staop,
1132 : : collation,
1133 : : NULL,
2489 1134 : 34202 : &sslot.values[probe]);
1135 : :
4751 1136 : 589518 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1137 : : collation,
2528 1138 : 589518 : sslot.values[probe],
1139 : : constval));
6416 1140 [ + + ]: 589518 : if (isgt)
1141 : 28818 : ltcmp = !ltcmp;
1142 [ + + ]: 589518 : if (ltcmp)
1143 : 229429 : lobound = probe + 1;
1144 : : else
1145 : 360089 : hibound = probe;
1146 : : }
1147 : :
1148 [ + + ]: 102246 : if (lobound <= 0)
1149 : : {
1150 : : /*
1151 : : * Constant is below lower histogram boundary. More
1152 : : * precisely, we have found that no entry in the histogram
1153 : : * satisfies the inequality clause (if !isgt) or they all do
1154 : : * (if isgt). We estimate that that's true of the entire
1155 : : * table, so set histfrac to 0.0 (which we'll flip to 1.0
1156 : : * below, if isgt).
1157 : : */
8378 1158 : 44273 : histfrac = 0.0;
1159 : : }
2528 1160 [ + + ]: 57973 : else if (lobound >= sslot.nvalues)
1161 : : {
1162 : : /*
1163 : : * Inverse case: constant is above upper histogram boundary.
1164 : : */
6416 1165 : 16224 : histfrac = 1.0;
1166 : : }
1167 : : else
1168 : : {
1169 : : /* We have values[i-1] <= constant <= values[i]. */
1170 : 41749 : int i = lobound;
2405 1171 : 41749 : double eq_selec = 0;
1172 : : double val,
1173 : : high,
1174 : : low;
1175 : : double binfrac;
1176 : :
1177 : : /*
1178 : : * In the cases where we'll need it below, obtain an estimate
1179 : : * of the selectivity of "x = constval". We use a calculation
1180 : : * similar to what var_eq_const() does for a non-MCV constant,
1181 : : * ie, estimate that all distinct non-MCV values occur equally
1182 : : * often. But multiplication by "1.0 - sumcommon - nullfrac"
1183 : : * will be done by our caller, so we shouldn't do that here.
1184 : : * Therefore we can't try to clamp the estimate by reference
1185 : : * to the least common MCV; the result would be too small.
1186 : : *
1187 : : * Note: since this is effectively assuming that constval
1188 : : * isn't an MCV, it's logically dubious if constval in fact is
1189 : : * one. But we have to apply *some* correction for equality,
1190 : : * and anyway we cannot tell if constval is an MCV, since we
1191 : : * don't have a suitable equality operator at hand.
1192 : : */
1193 [ + + + + ]: 41749 : if (i == 1 || isgt == iseq)
1194 : : {
1195 : : double otherdistinct;
1196 : : bool isdefault;
1197 : : AttStatsSlot mcvslot;
1198 : :
1199 : : /* Get estimated number of distinct values */
1200 : 14019 : otherdistinct = get_variable_numdistinct(vardata,
1201 : : &isdefault);
1202 : :
1203 : : /* Subtract off the number of known MCVs */
1204 [ + + ]: 14019 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1205 : : STATISTIC_KIND_MCV, InvalidOid,
1206 : : ATTSTATSSLOT_NUMBERS))
1207 : : {
1208 : 2306 : otherdistinct -= mcvslot.nnumbers;
1209 : 2306 : free_attstatsslot(&mcvslot);
1210 : : }
1211 : :
1212 : : /* If result doesn't seem sane, leave eq_selec at 0 */
1213 [ + + ]: 14019 : if (otherdistinct > 1)
1214 : 13998 : eq_selec = 1.0 / otherdistinct;
1215 : : }
1216 : :
1217 : : /*
1218 : : * Convert the constant and the two nearest bin boundary
1219 : : * values to a uniform comparison scale, and do a linear
1220 : : * interpolation within this bin.
1221 : : */
1409 1222 [ + - ]: 41749 : if (convert_to_scalar(constval, consttype, collation,
1223 : : &val,
2528 1224 : 41749 : sslot.values[i - 1], sslot.values[i],
1225 : : vardata->vartype,
1226 : : &low, &high))
1227 : : {
6416 1228 [ - + ]: 41749 : if (high <= low)
1229 : : {
1230 : : /* cope if bin boundaries appear identical */
6416 tgl@sss.pgh.pa.us 1231 :UBC 0 : binfrac = 0.5;
1232 : : }
6416 tgl@sss.pgh.pa.us 1233 [ + + ]:CBC 41749 : else if (val <= low)
1234 : 7252 : binfrac = 0.0;
1235 [ + + ]: 34497 : else if (val >= high)
1236 : 1758 : binfrac = 1.0;
1237 : : else
1238 : : {
1239 : 32739 : binfrac = (val - low) / (high - low);
1240 : :
1241 : : /*
1242 : : * Watch out for the possibility that we got a NaN or
1243 : : * Infinity from the division. This can happen
1244 : : * despite the previous checks, if for example "low"
1245 : : * is -Infinity.
1246 : : */
1247 [ + - + - ]: 32739 : if (isnan(binfrac) ||
1248 [ - + ]: 32739 : binfrac < 0.0 || binfrac > 1.0)
6416 tgl@sss.pgh.pa.us 1249 :UBC 0 : binfrac = 0.5;
1250 : : }
1251 : : }
1252 : : else
1253 : : {
1254 : : /*
1255 : : * Ideally we'd produce an error here, on the grounds that
1256 : : * the given operator shouldn't have scalarXXsel
1257 : : * registered as its selectivity func unless we can deal
1258 : : * with its operand types. But currently, all manner of
1259 : : * stuff is invoking scalarXXsel, so give a default
1260 : : * estimate until that can be fixed.
1261 : : */
1262 : 0 : binfrac = 0.5;
1263 : : }
1264 : :
1265 : : /*
1266 : : * Now, compute the overall selectivity across the values
1267 : : * represented by the histogram. We have i-1 full bins and
1268 : : * binfrac partial bin below the constant.
1269 : : */
6416 tgl@sss.pgh.pa.us 1270 :CBC 41749 : histfrac = (double) (i - 1) + binfrac;
2528 1271 : 41749 : histfrac /= (double) (sslot.nvalues - 1);
1272 : :
1273 : : /*
1274 : : * At this point, histfrac is an estimate of the fraction of
1275 : : * the population represented by the histogram that satisfies
1276 : : * "x <= constval". Somewhat remarkably, this statement is
1277 : : * true regardless of which operator we were doing the probes
1278 : : * with, so long as convert_to_scalar() delivers reasonable
1279 : : * results. If the probe constant is equal to some histogram
1280 : : * entry, we would have considered the bin to the left of that
1281 : : * entry if probing with "<" or ">=", or the bin to the right
1282 : : * if probing with "<=" or ">"; but binfrac would have come
1283 : : * out as 1.0 in the first case and 0.0 in the second, leading
1284 : : * to the same histfrac in either case. For probe constants
1285 : : * between histogram entries, we find the same bin and get the
1286 : : * same estimate with any operator.
1287 : : *
1288 : : * The fact that the estimate corresponds to "x <= constval"
1289 : : * and not "x < constval" is because of the way that ANALYZE
1290 : : * constructs the histogram: each entry is, effectively, the
1291 : : * rightmost value in its sample bucket. So selectivity
1292 : : * values that are exact multiples of 1/(histogram_size-1)
1293 : : * should be understood as estimates including a histogram
1294 : : * entry plus everything to its left.
1295 : : *
1296 : : * However, that breaks down for the first histogram entry,
1297 : : * which necessarily is the leftmost value in its sample
1298 : : * bucket. That means the first histogram bin is slightly
1299 : : * narrower than the rest, by an amount equal to eq_selec.
1300 : : * Another way to say that is that we want "x <= leftmost" to
1301 : : * be estimated as eq_selec not zero. So, if we're dealing
1302 : : * with the first bin (i==1), rescale to make that true while
1303 : : * adjusting the rest of that bin linearly.
1304 : : */
2405 1305 [ + + ]: 41749 : if (i == 1)
1306 : 6259 : histfrac += eq_selec * (1.0 - binfrac);
1307 : :
1308 : : /*
1309 : : * "x <= constval" is good if we want an estimate for "<=" or
1310 : : * ">", but if we are estimating for "<" or ">=", we now need
1311 : : * to decrease the estimate by eq_selec.
1312 : : */
1313 [ + + ]: 41749 : if (isgt == iseq)
1314 : 11343 : histfrac -= eq_selec;
1315 : : }
1316 : :
1317 : : /*
1318 : : * Now the estimate is finished for "<" and "<=" cases. If we are
1319 : : * estimating for ">" or ">=", flip it.
1320 : : */
8378 1321 [ + + ]: 102246 : hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1322 : :
1323 : : /*
1324 : : * The histogram boundaries are only approximate to begin with,
1325 : : * and may well be out of date anyway. Therefore, don't believe
1326 : : * extremely small or large selectivity estimates --- unless we
1327 : : * got actual current endpoint values from the table, in which
1328 : : * case just do the usual sanity clamp. Somewhat arbitrarily, we
1329 : : * set the cutoff for other cases at a hundredth of the histogram
1330 : : * resolution.
1331 : : */
5214 1332 [ + + ]: 102246 : if (have_end)
1333 [ + + + + ]: 54701 : CLAMP_PROBABILITY(hist_selec);
1334 : : else
1335 : : {
2405 1336 : 47545 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1337 : :
1338 [ + + ]: 47545 : if (hist_selec < cutoff)
1339 : 16104 : hist_selec = cutoff;
1340 [ + + ]: 31441 : else if (hist_selec > 1.0 - cutoff)
1341 : 11594 : hist_selec = 1.0 - cutoff;
1342 : : }
1343 : : }
1409 1344 [ + - ]: 92 : else if (sslot.nvalues > 1)
1345 : : {
1346 : : /*
1347 : : * If we get here, we have a histogram but it's not sorted the way
1348 : : * we want. Do a brute-force search to see how many of the
1349 : : * entries satisfy the comparison condition, and take that
1350 : : * fraction as our estimate. (This is identical to the inner loop
1351 : : * of histogram_selectivity; maybe share code?)
1352 : : */
1353 : 92 : LOCAL_FCINFO(fcinfo, 2);
1354 : 92 : int nmatch = 0;
1355 : :
1356 : 92 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1357 : : NULL, NULL);
1358 : 92 : fcinfo->args[0].isnull = false;
1359 : 92 : fcinfo->args[1].isnull = false;
1360 : 92 : fcinfo->args[1].value = constval;
1361 [ + + ]: 481042 : for (int i = 0; i < sslot.nvalues; i++)
1362 : : {
1363 : : Datum fresult;
1364 : :
1365 : 480950 : fcinfo->args[0].value = sslot.values[i];
1366 : 480950 : fcinfo->isnull = false;
1367 : 480950 : fresult = FunctionCallInvoke(fcinfo);
1368 [ + - + + ]: 480950 : if (!fcinfo->isnull && DatumGetBool(fresult))
1369 : 964 : nmatch++;
1370 : : }
1371 : 92 : hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1372 : :
1373 : : /*
1374 : : * As above, clamp to a hundredth of the histogram resolution.
1375 : : * This case is surely even less trustworthy than the normal one,
1376 : : * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1377 : : * clamp should be more restrictive in this case?)
1378 : : */
1379 : : {
1380 : 92 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1381 : :
1382 [ + + ]: 92 : if (hist_selec < cutoff)
1409 tgl@sss.pgh.pa.us 1383 :GBC 6 : hist_selec = cutoff;
1409 tgl@sss.pgh.pa.us 1384 [ + + ]:CBC 86 : else if (hist_selec > 1.0 - cutoff)
1409 tgl@sss.pgh.pa.us 1385 :GBC 6 : hist_selec = 1.0 - cutoff;
1386 : : }
1387 : : }
1388 : :
2528 tgl@sss.pgh.pa.us 1389 :CBC 102338 : free_attstatsslot(&sslot);
1390 : : }
1391 : :
6669 1392 : 128369 : return hist_selec;
1393 : : }
1394 : :
1395 : : /*
1396 : : * Common wrapper function for the selectivity estimators that simply
1397 : : * invoke scalarineqsel().
1398 : : */
1399 : : static Datum
2405 1400 : 19623 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1401 : : {
6888 1402 : 19623 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
8365 1403 : 19623 : Oid operator = PG_GETARG_OID(1);
1404 : 19623 : List *args = (List *) PG_GETARG_POINTER(2);
1405 : 19623 : int varRelid = PG_GETARG_INT32(3);
1409 1406 : 19623 : Oid collation = PG_GET_COLLATION();
1407 : : VariableStatData vardata;
1408 : : Node *other;
1409 : : bool varonleft;
1410 : : Datum constval;
1411 : : Oid consttype;
1412 : : double selec;
1413 : :
1414 : : /*
1415 : : * If expression is not variable op something or something op variable,
1416 : : * then punt and return a default estimate.
1417 : : */
7362 1418 [ + + ]: 19623 : if (!get_restriction_variable(root, args, varRelid,
1419 : : &vardata, &other, &varonleft))
8365 1420 : 445 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1421 : :
1422 : : /*
1423 : : * Can't do anything useful if the something is not a constant, either.
1424 : : */
8080 1425 [ + + ]: 19178 : if (!IsA(other, Const))
1426 : : {
7362 1427 [ + + ]: 1318 : ReleaseVariableStats(vardata);
8080 1428 : 1318 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1429 : : }
1430 : :
1431 : : /*
1432 : : * If the constant is NULL, assume operator is strict and return zero, ie,
1433 : : * operator will never return TRUE.
1434 : : */
1435 [ + + ]: 17860 : if (((Const *) other)->constisnull)
1436 : : {
7362 1437 [ + + ]: 27 : ReleaseVariableStats(vardata);
8080 1438 : 27 : PG_RETURN_FLOAT8(0.0);
1439 : : }
1440 : 17833 : constval = ((Const *) other)->constvalue;
1441 : 17833 : consttype = ((Const *) other)->consttype;
1442 : :
1443 : : /*
1444 : : * Force the var to be on the left to simplify logic in scalarineqsel.
1445 : : */
2405 1446 [ + + ]: 17833 : if (!varonleft)
1447 : : {
8365 1448 : 168 : operator = get_commutator(operator);
1449 [ - + ]: 168 : if (!operator)
1450 : : {
1451 : : /* Use default selectivity (should we raise an error instead?) */
7362 tgl@sss.pgh.pa.us 1452 [ # # ]:UBC 0 : ReleaseVariableStats(vardata);
8365 1453 : 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1454 : : }
2405 tgl@sss.pgh.pa.us 1455 :CBC 168 : isgt = !isgt;
1456 : : }
1457 : :
1458 : : /* The rest of the work is done by scalarineqsel(). */
1409 1459 : 17833 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1460 : : &vardata, constval, consttype);
1461 : :
7362 1462 [ + + ]: 17833 : ReleaseVariableStats(vardata);
1463 : :
8378 1464 : 17833 : PG_RETURN_FLOAT8((float8) selec);
1465 : : }
1466 : :
1467 : : /*
1468 : : * scalarltsel - Selectivity of "<" for scalars.
1469 : : */
1470 : : Datum
2405 1471 : 6793 : scalarltsel(PG_FUNCTION_ARGS)
1472 : : {
1473 : 6793 : return scalarineqsel_wrapper(fcinfo, false, false);
1474 : : }
1475 : :
1476 : : /*
1477 : : * scalarlesel - Selectivity of "<=" for scalars.
1478 : : */
1479 : : Datum
1480 : 2151 : scalarlesel(PG_FUNCTION_ARGS)
1481 : : {
1482 : 2151 : return scalarineqsel_wrapper(fcinfo, false, true);
1483 : : }
1484 : :
1485 : : /*
1486 : : * scalargtsel - Selectivity of ">" for scalars.
1487 : : */
1488 : : Datum
1489 : 7010 : scalargtsel(PG_FUNCTION_ARGS)
1490 : : {
1491 : 7010 : return scalarineqsel_wrapper(fcinfo, true, false);
1492 : : }
1493 : :
1494 : : /*
1495 : : * scalargesel - Selectivity of ">=" for scalars.
1496 : : */
1497 : : Datum
1498 : 3669 : scalargesel(PG_FUNCTION_ARGS)
1499 : : {
1500 : 3669 : return scalarineqsel_wrapper(fcinfo, true, true);
1501 : : }
1502 : :
1503 : : /*
1504 : : * boolvarsel - Selectivity of Boolean variable.
1505 : : *
1506 : : * This can actually be called on any boolean-valued expression. If it
1507 : : * involves only Vars of the specified relation, and if there are statistics
1508 : : * about the Var or expression (the latter is possible if it's indexed) then
1509 : : * we'll produce a real estimate; otherwise it's just a default.
1510 : : */
1511 : : Selectivity
3125 1512 : 20765 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1513 : : {
1514 : : VariableStatData vardata;
1515 : : double selec;
1516 : :
1517 : 20765 : examine_variable(root, arg, varRelid, &vardata);
1518 [ + + ]: 20765 : if (HeapTupleIsValid(vardata.statsTuple))
1519 : : {
1520 : : /*
1521 : : * A boolean variable V is equivalent to the clause V = 't', so we
1522 : : * compute the selectivity as if that is what we have.
1523 : : */
1409 1524 : 17304 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1525 : : BoolGetDatum(true), false, true, false);
1526 : : }
1527 : : else
1528 : : {
1529 : : /* Otherwise, the default estimate is 0.5 */
3125 1530 : 3461 : selec = 0.5;
1531 : : }
1532 [ + + ]: 20765 : ReleaseVariableStats(vardata);
1533 : 20765 : return selec;
1534 : : }
1535 : :
1536 : : /*
1537 : : * booltestsel - Selectivity of BooleanTest Node.
1538 : : */
1539 : : Selectivity
6888 1540 : 425 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1541 : : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1542 : : {
1543 : : VariableStatData vardata;
1544 : : double selec;
1545 : :
7362 1546 : 425 : examine_variable(root, arg, varRelid, &vardata);
1547 : :
1548 [ - + ]: 425 : if (HeapTupleIsValid(vardata.statsTuple))
1549 : : {
1550 : : Form_pg_statistic stats;
1551 : : double freq_null;
1552 : : AttStatsSlot sslot;
1553 : :
7362 tgl@sss.pgh.pa.us 1554 :UBC 0 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
8329 1555 : 0 : freq_null = stats->stanullfrac;
1556 : :
2528 1557 [ # # ]: 0 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1558 : : STATISTIC_KIND_MCV, InvalidOid,
1559 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1560 [ # # ]: 0 : && sslot.nnumbers > 0)
8329 1561 : 0 : {
1562 : : double freq_true;
1563 : : double freq_false;
1564 : :
1565 : : /*
1566 : : * Get first MCV frequency and derive frequency for true.
1567 : : */
2528 1568 [ # # ]: 0 : if (DatumGetBool(sslot.values[0]))
1569 : 0 : freq_true = sslot.numbers[0];
1570 : : else
1571 : 0 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1572 : :
1573 : : /*
1574 : : * Next derive frequency for false. Then use these as appropriate
1575 : : * to derive frequency for each case.
1576 : : */
8329 1577 : 0 : freq_false = 1.0 - freq_true - freq_null;
1578 : :
7848 1579 [ # # # # : 0 : switch (booltesttype)
# # # ]
1580 : : {
8207 bruce@momjian.us 1581 : 0 : case IS_UNKNOWN:
1582 : : /* select only NULL values */
8329 tgl@sss.pgh.pa.us 1583 : 0 : selec = freq_null;
1584 : 0 : break;
8207 bruce@momjian.us 1585 : 0 : case IS_NOT_UNKNOWN:
1586 : : /* select non-NULL values */
8329 tgl@sss.pgh.pa.us 1587 : 0 : selec = 1.0 - freq_null;
1588 : 0 : break;
8207 bruce@momjian.us 1589 : 0 : case IS_TRUE:
1590 : : /* select only TRUE values */
8329 tgl@sss.pgh.pa.us 1591 : 0 : selec = freq_true;
1592 : 0 : break;
8207 bruce@momjian.us 1593 : 0 : case IS_NOT_TRUE:
1594 : : /* select non-TRUE values */
8329 tgl@sss.pgh.pa.us 1595 : 0 : selec = 1.0 - freq_true;
1596 : 0 : break;
8207 bruce@momjian.us 1597 : 0 : case IS_FALSE:
1598 : : /* select only FALSE values */
8329 tgl@sss.pgh.pa.us 1599 : 0 : selec = freq_false;
1600 : 0 : break;
8207 bruce@momjian.us 1601 : 0 : case IS_NOT_FALSE:
1602 : : /* select non-FALSE values */
8329 tgl@sss.pgh.pa.us 1603 : 0 : selec = 1.0 - freq_false;
1604 : 0 : break;
8207 bruce@momjian.us 1605 : 0 : default:
7567 tgl@sss.pgh.pa.us 1606 [ # # ]: 0 : elog(ERROR, "unrecognized booltesttype: %d",
1607 : : (int) booltesttype);
1608 : : selec = 0.0; /* Keep compiler quiet */
1609 : : break;
1610 : : }
1611 : :
2528 1612 : 0 : free_attstatsslot(&sslot);
1613 : : }
1614 : : else
1615 : : {
1616 : : /*
1617 : : * No most-common-value info available. Still have null fraction
1618 : : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1619 : : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1620 : : */
7848 1621 [ # # # # : 0 : switch (booltesttype)
# ]
1622 : : {
8207 bruce@momjian.us 1623 : 0 : case IS_UNKNOWN:
1624 : : /* select only NULL values */
8329 tgl@sss.pgh.pa.us 1625 : 0 : selec = freq_null;
1626 : 0 : break;
8207 bruce@momjian.us 1627 : 0 : case IS_NOT_UNKNOWN:
1628 : : /* select non-NULL values */
8329 tgl@sss.pgh.pa.us 1629 : 0 : selec = 1.0 - freq_null;
1630 : 0 : break;
8207 bruce@momjian.us 1631 : 0 : case IS_TRUE:
1632 : : case IS_FALSE:
1633 : : /* Assume we select half of the non-NULL values */
8329 tgl@sss.pgh.pa.us 1634 : 0 : selec = (1.0 - freq_null) / 2.0;
1635 : 0 : break;
3917 1636 : 0 : case IS_NOT_TRUE:
1637 : : case IS_NOT_FALSE:
1638 : : /* Assume we select NULLs plus half of the non-NULLs */
1639 : : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1640 : 0 : selec = (freq_null + 1.0) / 2.0;
1641 : 0 : break;
8207 bruce@momjian.us 1642 : 0 : default:
7567 tgl@sss.pgh.pa.us 1643 [ # # ]: 0 : elog(ERROR, "unrecognized booltesttype: %d",
1644 : : (int) booltesttype);
1645 : : selec = 0.0; /* Keep compiler quiet */
1646 : : break;
1647 : : }
1648 : : }
1649 : : }
1650 : : else
1651 : : {
1652 : : /*
1653 : : * If we can't get variable statistics for the argument, perhaps
1654 : : * clause_selectivity can do something with it. We ignore the
1655 : : * possibility of a NULL value when using clause_selectivity, and just
1656 : : * assume the value is either TRUE or FALSE.
1657 : : */
7848 tgl@sss.pgh.pa.us 1658 [ + + + + :CBC 425 : switch (booltesttype)
- ]
1659 : : {
8329 1660 : 24 : case IS_UNKNOWN:
1661 : 24 : selec = DEFAULT_UNK_SEL;
1662 : 24 : break;
1663 : 54 : case IS_NOT_UNKNOWN:
1664 : 54 : selec = DEFAULT_NOT_UNK_SEL;
1665 : 54 : break;
1666 : 126 : case IS_TRUE:
1667 : : case IS_NOT_FALSE:
7362 1668 : 126 : selec = (double) clause_selectivity(root, arg,
1669 : : varRelid,
1670 : : jointype, sjinfo);
1671 : 126 : break;
1672 : 221 : case IS_FALSE:
1673 : : case IS_NOT_TRUE:
1674 : 221 : selec = 1.0 - (double) clause_selectivity(root, arg,
1675 : : varRelid,
1676 : : jointype, sjinfo);
8329 1677 : 221 : break;
8329 tgl@sss.pgh.pa.us 1678 :UBC 0 : default:
7567 1679 [ # # ]: 0 : elog(ERROR, "unrecognized booltesttype: %d",
1680 : : (int) booltesttype);
1681 : : selec = 0.0; /* Keep compiler quiet */
1682 : : break;
1683 : : }
1684 : : }
1685 : :
7362 tgl@sss.pgh.pa.us 1686 [ - + ]:CBC 425 : ReleaseVariableStats(vardata);
1687 : :
1688 : : /* result should be in range, but make sure... */
8137 1689 [ - + - + ]: 425 : CLAMP_PROBABILITY(selec);
1690 : :
8329 1691 : 425 : return (Selectivity) selec;
1692 : : }
1693 : :
1694 : : /*
1695 : : * nulltestsel - Selectivity of NullTest Node.
1696 : : */
1697 : : Selectivity
5722 1698 : 8750 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1699 : : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1700 : : {
1701 : : VariableStatData vardata;
1702 : : double selec;
1703 : :
7362 1704 : 8750 : examine_variable(root, arg, varRelid, &vardata);
1705 : :
1706 [ + + ]: 8750 : if (HeapTupleIsValid(vardata.statsTuple))
1707 : : {
1708 : : Form_pg_statistic stats;
1709 : : double freq_null;
1710 : :
1711 : 4777 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
8329 1712 : 4777 : freq_null = stats->stanullfrac;
1713 : :
7848 1714 [ + + - ]: 4777 : switch (nulltesttype)
1715 : : {
8207 bruce@momjian.us 1716 : 3337 : case IS_NULL:
1717 : :
1718 : : /*
1719 : : * Use freq_null directly.
1720 : : */
8329 tgl@sss.pgh.pa.us 1721 : 3337 : selec = freq_null;
1722 : 3337 : break;
8207 bruce@momjian.us 1723 : 1440 : case IS_NOT_NULL:
1724 : :
1725 : : /*
1726 : : * Select not unknown (not null) values. Calculate from
1727 : : * freq_null.
1728 : : */
8329 tgl@sss.pgh.pa.us 1729 : 1440 : selec = 1.0 - freq_null;
1730 : 1440 : break;
8207 bruce@momjian.us 1731 :UBC 0 : default:
7567 tgl@sss.pgh.pa.us 1732 [ # # ]: 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1733 : : (int) nulltesttype);
1734 : : return (Selectivity) 0; /* keep compiler quiet */
1735 : : }
1736 : : }
1906 tgl@sss.pgh.pa.us 1737 [ + - + + ]:CBC 3973 : else if (vardata.var && IsA(vardata.var, Var) &&
1738 [ + + ]: 3640 : ((Var *) vardata.var)->varattno < 0)
1739 : : {
1740 : : /*
1741 : : * There are no stats for system columns, but we know they are never
1742 : : * NULL.
1743 : : */
1744 [ + - ]: 30 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1745 : : }
1746 : : else
1747 : : {
1748 : : /*
1749 : : * No ANALYZE stats available, so make a guess
1750 : : */
7362 1751 [ + + - ]: 3943 : switch (nulltesttype)
1752 : : {
1753 : 1004 : case IS_NULL:
1754 : 1004 : selec = DEFAULT_UNK_SEL;
1755 : 1004 : break;
1756 : 2939 : case IS_NOT_NULL:
1757 : 2939 : selec = DEFAULT_NOT_UNK_SEL;
1758 : 2939 : break;
7362 tgl@sss.pgh.pa.us 1759 :UBC 0 : default:
1760 [ # # ]: 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1761 : : (int) nulltesttype);
1762 : : return (Selectivity) 0; /* keep compiler quiet */
1763 : : }
1764 : : }
1765 : :
7362 tgl@sss.pgh.pa.us 1766 [ + + ]:CBC 8750 : ReleaseVariableStats(vardata);
1767 : :
1768 : : /* result should be in range, but make sure... */
8137 1769 [ - + - + ]: 8750 : CLAMP_PROBABILITY(selec);
1770 : :
8329 1771 : 8750 : return (Selectivity) selec;
1772 : : }
1773 : :
1774 : : /*
1775 : : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1776 : : *
1777 : : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1778 : : * but it seems possible that RelabelType might show up. Also, the planner
1779 : : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1780 : : * so we need to be ready to deal with more than one level.
1781 : : */
1782 : : static Node *
6286 1783 : 51272 : strip_array_coercion(Node *node)
1784 : : {
1785 : : for (;;)
1786 : : {
2388 1787 [ + - + + ]: 51290 : if (node && IsA(node, ArrayCoerceExpr))
6286 1788 : 18 : {
2388 1789 : 1332 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1790 : :
1791 : : /*
1792 : : * If the per-element expression is just a RelabelType on top of
1793 : : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1794 : : */
1795 [ + + ]: 1332 : if (IsA(acoerce->elemexpr, RelabelType) &&
1796 [ + - ]: 18 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1797 : 18 : node = (Node *) acoerce->arg;
1798 : : else
1799 : : break;
1800 : : }
6228 1801 [ + - - + ]: 49958 : else if (node && IsA(node, RelabelType))
1802 : : {
1803 : : /* We don't really expect this case, but may as well cope */
6228 tgl@sss.pgh.pa.us 1804 :UBC 0 : node = (Node *) ((RelabelType *) node)->arg;
1805 : : }
1806 : : else
1807 : : break;
1808 : : }
6286 tgl@sss.pgh.pa.us 1809 :CBC 51272 : return node;
1810 : : }
1811 : :
1812 : : /*
1813 : : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1814 : : */
1815 : : Selectivity
6715 1816 : 8283 : scalararraysel(PlannerInfo *root,
1817 : : ScalarArrayOpExpr *clause,
1818 : : bool is_join_clause,
1819 : : int varRelid,
1820 : : JoinType jointype,
1821 : : SpecialJoinInfo *sjinfo)
1822 : : {
1823 : 8283 : Oid operator = clause->opno;
1824 : 8283 : bool useOr = clause->useOr;
4425 1825 : 8283 : bool isEquality = false;
1826 : 8283 : bool isInequality = false;
1827 : : Node *leftop;
1828 : : Node *rightop;
1829 : : Oid nominal_element_type;
1830 : : Oid nominal_element_collation;
1831 : : TypeCacheEntry *typentry;
1832 : : RegProcedure oprsel;
1833 : : FmgrInfo oprselproc;
1834 : : Selectivity s1;
1835 : : Selectivity s1disjoint;
1836 : :
1837 : : /* First, deconstruct the expression */
6286 1838 [ - + ]: 8283 : Assert(list_length(clause->args) == 2);
1839 : 8283 : leftop = (Node *) linitial(clause->args);
1840 : 8283 : rightop = (Node *) lsecond(clause->args);
1841 : :
1842 : : /* aggressively reduce both sides to constants */
3705 1843 : 8283 : leftop = estimate_expression_value(root, leftop);
1844 : 8283 : rightop = estimate_expression_value(root, rightop);
1845 : :
1846 : : /* get nominal (after relabeling) element type of rightop */
4924 1847 : 8283 : nominal_element_type = get_base_element_type(exprType(rightop));
6286 1848 [ - + ]: 8283 : if (!OidIsValid(nominal_element_type))
2489 tgl@sss.pgh.pa.us 1849 :UBC 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1850 : : /* get nominal collation, too, for generating constants */
4769 tgl@sss.pgh.pa.us 1851 :CBC 8283 : nominal_element_collation = exprCollation(rightop);
1852 : :
1853 : : /* look through any binary-compatible relabeling of rightop */
6286 1854 : 8283 : rightop = strip_array_coercion(rightop);
1855 : :
1856 : : /*
1857 : : * Detect whether the operator is the default equality or inequality
1858 : : * operator of the array element type.
1859 : : */
4425 1860 : 8283 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1861 [ + - ]: 8283 : if (OidIsValid(typentry->eq_opr))
1862 : : {
1863 [ + + ]: 8283 : if (operator == typentry->eq_opr)
1864 : 7052 : isEquality = true;
1865 [ + + ]: 1231 : else if (get_negator(operator) == typentry->eq_opr)
1866 : 985 : isInequality = true;
1867 : : }
1868 : :
1869 : : /*
1870 : : * If it is equality or inequality, we might be able to estimate this as a
1871 : : * form of array containment; for instance "const = ANY(column)" can be
1872 : : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1873 : : * that, and returns the selectivity estimate if successful, or -1 if not.
1874 : : */
1875 [ + + + + : 8283 : if ((isEquality || isInequality) && !is_join_clause)
+ - ]
1876 : : {
1877 : 8037 : s1 = scalararraysel_containment(root, leftop, rightop,
1878 : : nominal_element_type,
1879 : : isEquality, useOr, varRelid);
1880 [ + + ]: 8037 : if (s1 >= 0.0)
1881 : 59 : return s1;
1882 : : }
1883 : :
1884 : : /*
1885 : : * Look up the underlying operator's selectivity estimator. Punt if it
1886 : : * hasn't got one.
1887 : : */
1888 [ - + ]: 8224 : if (is_join_clause)
4425 tgl@sss.pgh.pa.us 1889 :UBC 0 : oprsel = get_oprjoin(operator);
1890 : : else
4425 tgl@sss.pgh.pa.us 1891 :CBC 8224 : oprsel = get_oprrest(operator);
1892 [ - + ]: 8224 : if (!oprsel)
4425 tgl@sss.pgh.pa.us 1893 :UBC 0 : return (Selectivity) 0.5;
4425 tgl@sss.pgh.pa.us 1894 :CBC 8224 : fmgr_info(oprsel, &oprselproc);
1895 : :
1896 : : /*
1897 : : * In the array-containment check above, we must only believe that an
1898 : : * operator is equality or inequality if it is the default btree equality
1899 : : * operator (or its negator) for the element type, since those are the
1900 : : * operators that array containment will use. But in what follows, we can
1901 : : * be a little laxer, and also believe that any operators using eqsel() or
1902 : : * neqsel() as selectivity estimator act like equality or inequality.
1903 : : */
4421 1904 [ + + - + ]: 8224 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1905 : 7081 : isEquality = true;
1906 [ + + - + ]: 1143 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1907 : 930 : isInequality = true;
1908 : :
1909 : : /*
1910 : : * We consider three cases:
1911 : : *
1912 : : * 1. rightop is an Array constant: deconstruct the array, apply the
1913 : : * operator's selectivity function for each array element, and merge the
1914 : : * results in the same way that clausesel.c does for AND/OR combinations.
1915 : : *
1916 : : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
1917 : : * function for each element of the ARRAY[] construct, and merge.
1918 : : *
1919 : : * 3. otherwise, make a guess ...
1920 : : */
6715 1921 [ + - + + ]: 8224 : if (rightop && IsA(rightop, Const))
1922 : 6492 : {
1923 : 6501 : Datum arraydatum = ((Const *) rightop)->constvalue;
1924 : 6501 : bool arrayisnull = ((Const *) rightop)->constisnull;
1925 : : ArrayType *arrayval;
1926 : : int16 elmlen;
1927 : : bool elmbyval;
1928 : : char elmalign;
1929 : : int num_elems;
1930 : : Datum *elem_values;
1931 : : bool *elem_nulls;
1932 : : int i;
1933 : :
1934 [ + + ]: 6501 : if (arrayisnull) /* qual can't succeed if null array */
1935 : 9 : return (Selectivity) 0.0;
1936 : 6492 : arrayval = DatumGetArrayTypeP(arraydatum);
1937 : 6492 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
1938 : : &elmlen, &elmbyval, &elmalign);
1939 : 6492 : deconstruct_array(arrayval,
1940 : : ARR_ELEMTYPE(arrayval),
1941 : : elmlen, elmbyval, elmalign,
1942 : : &elem_values, &elem_nulls, &num_elems);
1943 : :
1944 : : /*
1945 : : * For generic operators, we assume the probability of success is
1946 : : * independent for each array element. But for "= ANY" or "<> ALL",
1947 : : * if the array elements are distinct (which'd typically be the case)
1948 : : * then the probabilities are disjoint, and we should just sum them.
1949 : : *
1950 : : * If we were being really tense we would try to confirm that the
1951 : : * elements are all distinct, but that would be expensive and it
1952 : : * doesn't seem to be worth the cycles; it would amount to penalizing
1953 : : * well-written queries in favor of poorly-written ones. However, we
1954 : : * do protect ourselves a little bit by checking whether the
1955 : : * disjointness assumption leads to an impossible (out of range)
1956 : : * probability; if so, we fall back to the normal calculation.
1957 : : */
4421 1958 [ + + ]: 6492 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
1959 : :
6715 1960 [ + + ]: 26398 : for (i = 0; i < num_elems; i++)
1961 : : {
1962 : : List *args;
1963 : : Selectivity s2;
1964 : :
1965 : 19906 : args = list_make2(leftop,
1966 : : makeConst(nominal_element_type,
1967 : : -1,
1968 : : nominal_element_collation,
1969 : : elmlen,
1970 : : elem_values[i],
1971 : : elem_nulls[i],
1972 : : elmbyval));
5720 1973 [ - + ]: 19906 : if (is_join_clause)
4298 tgl@sss.pgh.pa.us 1974 :UBC 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
1975 : : clause->inputcollid,
1976 : : PointerGetDatum(root),
1977 : : ObjectIdGetDatum(operator),
1978 : : PointerGetDatum(args),
1979 : : Int16GetDatum(jointype),
1980 : : PointerGetDatum(sjinfo)));
1981 : : else
4298 tgl@sss.pgh.pa.us 1982 :CBC 19906 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
1983 : : clause->inputcollid,
1984 : : PointerGetDatum(root),
1985 : : ObjectIdGetDatum(operator),
1986 : : PointerGetDatum(args),
1987 : : Int32GetDatum(varRelid)));
1988 : :
6715 1989 [ + + ]: 19906 : if (useOr)
1990 : : {
1991 : 16918 : s1 = s1 + s2 - s1 * s2;
4421 1992 [ + + ]: 16918 : if (isEquality)
1993 : 16486 : s1disjoint += s2;
1994 : : }
1995 : : else
1996 : : {
6715 1997 : 2988 : s1 = s1 * s2;
4421 1998 [ + + ]: 2988 : if (isInequality)
1999 : 2832 : s1disjoint += s2 - 1.0;
2000 : : }
2001 : : }
2002 : :
2003 : : /* accept disjoint-probability estimate if in range */
2004 [ + + + + : 6492 : if ((useOr ? isEquality : isInequality) &&
+ + ]
2005 [ + + ]: 6207 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2006 : 6192 : s1 = s1disjoint;
2007 : : }
6715 2008 [ + - + + ]: 1723 : else if (rightop && IsA(rightop, ArrayExpr) &&
2009 [ + - ]: 52 : !((ArrayExpr *) rightop)->multidims)
2010 : 52 : {
2011 : 52 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2012 : : int16 elmlen;
2013 : : bool elmbyval;
2014 : : ListCell *l;
2015 : :
2016 : 52 : get_typlenbyval(arrayexpr->element_typeid,
2017 : : &elmlen, &elmbyval);
2018 : :
2019 : : /*
2020 : : * We use the assumption of disjoint probabilities here too, although
2021 : : * the odds of equal array elements are rather higher if the elements
2022 : : * are not all constants (which they won't be, else constant folding
2023 : : * would have reduced the ArrayExpr to a Const). In this path it's
2024 : : * critical to have the sanity check on the s1disjoint estimate.
2025 : : */
4421 2026 [ + - ]: 52 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2027 : :
6715 2028 [ + - + + : 184 : foreach(l, arrayexpr->elements)
+ + ]
2029 : : {
6286 2030 : 132 : Node *elem = (Node *) lfirst(l);
2031 : : List *args;
2032 : : Selectivity s2;
2033 : :
2034 : : /*
2035 : : * Theoretically, if elem isn't of nominal_element_type we should
2036 : : * insert a RelabelType, but it seems unlikely that any operator
2037 : : * estimation function would really care ...
2038 : : */
2039 : 132 : args = list_make2(leftop, elem);
5720 2040 [ - + ]: 132 : if (is_join_clause)
4298 tgl@sss.pgh.pa.us 2041 :UBC 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2042 : : clause->inputcollid,
2043 : : PointerGetDatum(root),
2044 : : ObjectIdGetDatum(operator),
2045 : : PointerGetDatum(args),
2046 : : Int16GetDatum(jointype),
2047 : : PointerGetDatum(sjinfo)));
2048 : : else
4298 tgl@sss.pgh.pa.us 2049 :CBC 132 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2050 : : clause->inputcollid,
2051 : : PointerGetDatum(root),
2052 : : ObjectIdGetDatum(operator),
2053 : : PointerGetDatum(args),
2054 : : Int32GetDatum(varRelid)));
2055 : :
6715 2056 [ + - ]: 132 : if (useOr)
2057 : : {
2058 : 132 : s1 = s1 + s2 - s1 * s2;
4421 2059 [ + - ]: 132 : if (isEquality)
2060 : 132 : s1disjoint += s2;
2061 : : }
2062 : : else
2063 : : {
6715 tgl@sss.pgh.pa.us 2064 :UBC 0 : s1 = s1 * s2;
4421 2065 [ # # ]: 0 : if (isInequality)
2066 : 0 : s1disjoint += s2 - 1.0;
2067 : : }
2068 : : }
2069 : :
2070 : : /* accept disjoint-probability estimate if in range */
4421 tgl@sss.pgh.pa.us 2071 [ + - + - :CBC 52 : if ((useOr ? isEquality : isInequality) &&
+ - ]
2072 [ + - ]: 52 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2073 : 52 : s1 = s1disjoint;
2074 : : }
2075 : : else
2076 : : {
2077 : : CaseTestExpr *dummyexpr;
2078 : : List *args;
2079 : : Selectivity s2;
2080 : : int i;
2081 : :
2082 : : /*
2083 : : * We need a dummy rightop to pass to the operator selectivity
2084 : : * routine. It can be pretty much anything that doesn't look like a
2085 : : * constant; CaseTestExpr is a convenient choice.
2086 : : */
6715 2087 : 1671 : dummyexpr = makeNode(CaseTestExpr);
6286 2088 : 1671 : dummyexpr->typeId = nominal_element_type;
6715 2089 : 1671 : dummyexpr->typeMod = -1;
4775 2090 : 1671 : dummyexpr->collation = clause->inputcollid;
6715 2091 : 1671 : args = list_make2(leftop, dummyexpr);
5720 2092 [ - + ]: 1671 : if (is_join_clause)
4298 tgl@sss.pgh.pa.us 2093 :UBC 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2094 : : clause->inputcollid,
2095 : : PointerGetDatum(root),
2096 : : ObjectIdGetDatum(operator),
2097 : : PointerGetDatum(args),
2098 : : Int16GetDatum(jointype),
2099 : : PointerGetDatum(sjinfo)));
2100 : : else
4298 tgl@sss.pgh.pa.us 2101 :CBC 1671 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2102 : : clause->inputcollid,
2103 : : PointerGetDatum(root),
2104 : : ObjectIdGetDatum(operator),
2105 : : PointerGetDatum(args),
2106 : : Int32GetDatum(varRelid)));
6715 2107 [ + - ]: 1671 : s1 = useOr ? 0.0 : 1.0;
2108 : :
2109 : : /*
2110 : : * Arbitrarily assume 10 elements in the eventual array value (see
2111 : : * also estimate_array_length). We don't risk an assumption of
2112 : : * disjoint probabilities here.
2113 : : */
2114 [ + + ]: 18381 : for (i = 0; i < 10; i++)
2115 : : {
2116 [ + - ]: 16710 : if (useOr)
2117 : 16710 : s1 = s1 + s2 - s1 * s2;
2118 : : else
6715 tgl@sss.pgh.pa.us 2119 :UBC 0 : s1 = s1 * s2;
2120 : : }
2121 : : }
2122 : :
2123 : : /* result should be in range, but make sure... */
6715 tgl@sss.pgh.pa.us 2124 [ - + - + ]:CBC 8215 : CLAMP_PROBABILITY(s1);
2125 : :
2126 : 8215 : return s1;
2127 : : }
2128 : :
2129 : : /*
2130 : : * Estimate number of elements in the array yielded by an expression.
2131 : : *
2132 : : * Note: the result is integral, but we use "double" to avoid overflow
2133 : : * concerns. Most callers will use it in double-type expressions anyway.
2134 : : */
2135 : : double
101 tgl@sss.pgh.pa.us 2136 :GNC 42989 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2137 : : {
2138 : : /* look through any binary-compatible relabeling of arrayexpr */
6286 tgl@sss.pgh.pa.us 2139 :CBC 42989 : arrayexpr = strip_array_coercion(arrayexpr);
2140 : :
6497 2141 [ + - + + ]: 42989 : if (arrayexpr && IsA(arrayexpr, Const))
2142 : : {
2143 : 18004 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2144 : 18004 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2145 : : ArrayType *arrayval;
2146 : :
2147 [ + + ]: 18004 : if (arrayisnull)
2148 : 18 : return 0;
2149 : 17986 : arrayval = DatumGetArrayTypeP(arraydatum);
2150 : 17986 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2151 : : }
2152 [ + - + + ]: 24985 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2153 [ + - ]: 200 : !((ArrayExpr *) arrayexpr)->multidims)
2154 : : {
2155 : 200 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2156 : : }
101 tgl@sss.pgh.pa.us 2157 [ + - ]:GNC 24785 : else if (arrayexpr)
2158 : : {
2159 : : /* See if we can find any statistics about it */
2160 : : VariableStatData vardata;
2161 : : AttStatsSlot sslot;
2162 : 24785 : double nelem = 0;
2163 : :
2164 : 24785 : examine_variable(root, arrayexpr, 0, &vardata);
2165 [ + + ]: 24785 : if (HeapTupleIsValid(vardata.statsTuple))
2166 : : {
2167 : : /*
2168 : : * Found stats, so use the average element count, which is stored
2169 : : * in the last stanumbers element of the DECHIST statistics.
2170 : : * Actually that is the average count of *distinct* elements;
2171 : : * perhaps we should scale it up somewhat?
2172 : : */
2173 [ + + ]: 6328 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2174 : : STATISTIC_KIND_DECHIST, InvalidOid,
2175 : : ATTSTATSSLOT_NUMBERS))
2176 : : {
2177 [ + - ]: 6271 : if (sslot.nnumbers > 0)
2178 : 6271 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2179 : 6271 : free_attstatsslot(&sslot);
2180 : : }
2181 : : }
2182 [ + + ]: 24785 : ReleaseVariableStats(vardata);
2183 : :
2184 [ + + ]: 24785 : if (nelem > 0)
2185 : 6271 : return nelem;
2186 : : }
2187 : :
2188 : : /* Else use a default guess --- this should match scalararraysel */
2189 : 18514 : return 10;
2190 : : }
2191 : :
2192 : : /*
2193 : : * rowcomparesel - Selectivity of RowCompareExpr Node.
2194 : : *
2195 : : * We estimate RowCompare selectivity by considering just the first (high
2196 : : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2197 : : * this estimate could be refined by considering additional columns, it
2198 : : * seems unlikely that we could do a lot better without multi-column
2199 : : * statistics.
2200 : : */
2201 : : Selectivity
6665 tgl@sss.pgh.pa.us 2202 :CBC 78 : rowcomparesel(PlannerInfo *root,
2203 : : RowCompareExpr *clause,
2204 : : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2205 : : {
2206 : : Selectivity s1;
2207 : 78 : Oid opno = linitial_oid(clause->opnos);
4298 2208 : 78 : Oid inputcollid = linitial_oid(clause->inputcollids);
2209 : : List *opargs;
2210 : : bool is_join_clause;
2211 : :
2212 : : /* Build equivalent arg list for single operator */
6665 2213 : 78 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2214 : :
2215 : : /*
2216 : : * Decide if it's a join clause. This should match clausesel.c's
2217 : : * treat_as_join_clause(), except that we intentionally consider only the
2218 : : * leading columns and not the rest of the clause.
2219 : : */
2220 [ + + ]: 78 : if (varRelid != 0)
2221 : : {
2222 : : /*
2223 : : * Caller is forcing restriction mode (eg, because we are examining an
2224 : : * inner indexscan qual).
2225 : : */
5720 2226 : 27 : is_join_clause = false;
2227 : : }
2228 [ + + ]: 51 : else if (sjinfo == NULL)
2229 : : {
2230 : : /*
2231 : : * It must be a restriction clause, since it's being evaluated at a
2232 : : * scan node.
2233 : : */
6665 2234 : 45 : is_join_clause = false;
2235 : : }
2236 : : else
2237 : : {
2238 : : /*
2239 : : * Otherwise, it's a join if there's more than one base relation used.
2240 : : */
1179 2241 : 6 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2242 : : }
2243 : :
6665 2244 [ + + ]: 78 : if (is_join_clause)
2245 : : {
2246 : : /* Estimate selectivity for a join clause. */
2247 : 6 : s1 = join_selectivity(root, opno,
2248 : : opargs,
2249 : : inputcollid,
2250 : : jointype,
2251 : : sjinfo);
2252 : : }
2253 : : else
2254 : : {
2255 : : /* Estimate selectivity for a restriction clause. */
2256 : 72 : s1 = restriction_selectivity(root, opno,
2257 : : opargs,
2258 : : inputcollid,
2259 : : varRelid);
2260 : : }
2261 : :
2262 : 78 : return s1;
2263 : : }
2264 : :
2265 : : /*
2266 : : * eqjoinsel - Join selectivity of "="
2267 : : */
2268 : : Datum
8714 2269 : 95525 : eqjoinsel(PG_FUNCTION_ARGS)
2270 : : {
6888 2271 : 95525 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
8365 2272 : 95525 : Oid operator = PG_GETARG_OID(1);
2273 : 95525 : List *args = (List *) PG_GETARG_POINTER(2);
2274 : :
2275 : : #ifdef NOT_USED
2276 : : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2277 : : #endif
5720 2278 : 95525 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
1409 2279 : 95525 : Oid collation = PG_GET_COLLATION();
2280 : : double selec;
2281 : : double selec_inner;
2282 : : VariableStatData vardata1;
2283 : : VariableStatData vardata2;
2284 : : double nd1;
2285 : : double nd2;
2286 : : bool isdefault1;
2287 : : bool isdefault2;
2288 : : Oid opfuncoid;
2289 : : AttStatsSlot sslot1;
2290 : : AttStatsSlot sslot2;
1969 2291 : 95525 : Form_pg_statistic stats1 = NULL;
2292 : 95525 : Form_pg_statistic stats2 = NULL;
2293 : 95525 : bool have_mcvs1 = false;
2294 : 95525 : bool have_mcvs2 = false;
2295 : : bool get_mcv_stats;
2296 : : bool join_is_reversed;
2297 : : RelOptInfo *inner_rel;
2298 : :
5720 2299 : 95525 : get_join_variables(root, args, sjinfo,
2300 : : &vardata1, &vardata2, &join_is_reversed);
2301 : :
1969 2302 : 95525 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2303 : 95525 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2304 : :
2305 : 95525 : opfuncoid = get_opcode(operator);
2306 : :
2307 : 95525 : memset(&sslot1, 0, sizeof(sslot1));
2308 : 95525 : memset(&sslot2, 0, sizeof(sslot2));
2309 : :
2310 : : /*
2311 : : * There is no use in fetching one side's MCVs if we lack MCVs for the
2312 : : * other side, so do a quick check to verify that both stats exist.
2313 : : */
513 2314 : 268299 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2315 [ + + + + ]: 139113 : HeapTupleIsValid(vardata2.statsTuple) &&
2316 : 61864 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2317 : : STATISTIC_KIND_MCV, InvalidOid,
2318 [ + + + + ]: 172774 : 0) &&
2319 : 27731 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2320 : : STATISTIC_KIND_MCV, InvalidOid,
2321 : : 0));
2322 : :
1969 2323 [ + + ]: 95525 : if (HeapTupleIsValid(vardata1.statsTuple))
2324 : : {
2325 : : /* note we allow use of nullfrac regardless of security check */
2326 : 77249 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
513 2327 [ + + + - ]: 85644 : if (get_mcv_stats &&
2328 : 8395 : statistic_proc_security_check(&vardata1, opfuncoid))
1969 2329 : 8395 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2330 : : STATISTIC_KIND_MCV, InvalidOid,
2331 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2332 : : }
2333 : :
2334 [ + + ]: 95525 : if (HeapTupleIsValid(vardata2.statsTuple))
2335 : : {
2336 : : /* note we allow use of nullfrac regardless of security check */
2337 : 66236 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
513 2338 [ + + + - ]: 74631 : if (get_mcv_stats &&
2339 : 8395 : statistic_proc_security_check(&vardata2, opfuncoid))
1969 2340 : 8395 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2341 : : STATISTIC_KIND_MCV, InvalidOid,
2342 : : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2343 : : }
2344 : :
2345 : : /* We need to compute the inner-join selectivity in all cases */
1409 2346 : 95525 : selec_inner = eqjoinsel_inner(opfuncoid, collation,
2347 : : &vardata1, &vardata2,
2348 : : nd1, nd2,
2349 : : isdefault1, isdefault2,
2350 : : &sslot1, &sslot2,
2351 : : stats1, stats2,
2352 : : have_mcvs1, have_mcvs2);
2353 : :
5720 2354 [ + + - ]: 95525 : switch (sjinfo->jointype)
2355 : : {
2356 : 92001 : case JOIN_INNER:
2357 : : case JOIN_LEFT:
2358 : : case JOIN_FULL:
1969 2359 : 92001 : selec = selec_inner;
5720 2360 : 92001 : break;
2361 : 3524 : case JOIN_SEMI:
2362 : : case JOIN_ANTI:
2363 : :
2364 : : /*
2365 : : * Look up the join's inner relation. min_righthand is sufficient
2366 : : * information because neither SEMI nor ANTI joins permit any
2367 : : * reassociation into or out of their RHS, so the righthand will
2368 : : * always be exactly that set of rels.
2369 : : */
4609 2370 : 3524 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2371 : :
5720 2372 [ + + ]: 3524 : if (!join_is_reversed)
1409 2373 : 1718 : selec = eqjoinsel_semi(opfuncoid, collation,
2374 : : &vardata1, &vardata2,
2375 : : nd1, nd2,
2376 : : isdefault1, isdefault2,
2377 : : &sslot1, &sslot2,
2378 : : stats1, stats2,
2379 : : have_mcvs1, have_mcvs2,
2380 : : inner_rel);
2381 : : else
2382 : : {
1969 2383 : 1806 : Oid commop = get_commutator(operator);
2384 [ + - ]: 1806 : Oid commopfuncoid = OidIsValid(commop) ? get_opcode(commop) : InvalidOid;
2385 : :
1409 2386 : 1806 : selec = eqjoinsel_semi(commopfuncoid, collation,
2387 : : &vardata2, &vardata1,
2388 : : nd2, nd1,
2389 : : isdefault2, isdefault1,
2390 : : &sslot2, &sslot1,
2391 : : stats2, stats1,
2392 : : have_mcvs2, have_mcvs1,
2393 : : inner_rel);
2394 : : }
2395 : :
2396 : : /*
2397 : : * We should never estimate the output of a semijoin to be more
2398 : : * rows than we estimate for an inner join with the same input
2399 : : * rels and join condition; it's obviously impossible for that to
2400 : : * happen. The former estimate is N1 * Ssemi while the latter is
2401 : : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2402 : : * this is worthwhile because of the shakier estimation rules we
2403 : : * use in eqjoinsel_semi, particularly in cases where it has to
2404 : : * punt entirely.
2405 : : */
1969 2406 [ + + ]: 3524 : selec = Min(selec, inner_rel->rows * selec_inner);
5720 2407 : 3524 : break;
5720 tgl@sss.pgh.pa.us 2408 :UBC 0 : default:
2409 : : /* other values not expected here */
2410 [ # # ]: 0 : elog(ERROR, "unrecognized join type: %d",
2411 : : (int) sjinfo->jointype);
2412 : : selec = 0; /* keep compiler quiet */
2413 : : break;
2414 : : }
2415 : :
1969 tgl@sss.pgh.pa.us 2416 :CBC 95525 : free_attstatsslot(&sslot1);
2417 : 95525 : free_attstatsslot(&sslot2);
2418 : :
5720 2419 [ + + ]: 95525 : ReleaseVariableStats(vardata1);
2420 [ + + ]: 95525 : ReleaseVariableStats(vardata2);
2421 : :
2422 [ - + - + ]: 95525 : CLAMP_PROBABILITY(selec);
2423 : :
2424 : 95525 : PG_RETURN_FLOAT8((float8) selec);
2425 : : }
2426 : :
2427 : : /*
2428 : : * eqjoinsel_inner --- eqjoinsel for normal inner join
2429 : : *
2430 : : * We also use this for LEFT/FULL outer joins; it's not presently clear
2431 : : * that it's worth trying to distinguish them here.
2432 : : */
2433 : : static double
1409 2434 : 95525 : eqjoinsel_inner(Oid opfuncoid, Oid collation,
2435 : : VariableStatData *vardata1, VariableStatData *vardata2,
2436 : : double nd1, double nd2,
2437 : : bool isdefault1, bool isdefault2,
2438 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2439 : : Form_pg_statistic stats1, Form_pg_statistic stats2,
2440 : : bool have_mcvs1, bool have_mcvs2)
2441 : : {
2442 : : double selec;
2443 : :
7362 2444 [ + + + - ]: 95525 : if (have_mcvs1 && have_mcvs2)
9716 bruce@momjian.us 2445 : 8395 : {
2446 : : /*
2447 : : * We have most-common-value lists for both relations. Run through
2448 : : * the lists to see which MCVs actually join to each other with the
2449 : : * given operator. This allows us to determine the exact join
2450 : : * selectivity for the portion of the relations represented by the MCV
2451 : : * lists. We still have to estimate for the remaining population, but
2452 : : * in a skewed distribution this gives us a big leg up in accuracy.
2453 : : * For motivation see the analysis in Y. Ioannidis and S.
2454 : : * Christodoulakis, "On the propagation of errors in the size of join
2455 : : * results", Technical Report 1018, Computer Science Dept., University
2456 : : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2457 : : */
1454 tgl@sss.pgh.pa.us 2458 : 8395 : LOCAL_FCINFO(fcinfo, 2);
2459 : : FmgrInfo eqproc;
2460 : : bool *hasmatch1;
2461 : : bool *hasmatch2;
7362 2462 : 8395 : double nullfrac1 = stats1->stanullfrac;
2463 : 8395 : double nullfrac2 = stats2->stanullfrac;
2464 : : double matchprodfreq,
2465 : : matchfreq1,
2466 : : matchfreq2,
2467 : : unmatchfreq1,
2468 : : unmatchfreq2,
2469 : : otherfreq1,
2470 : : otherfreq2,
2471 : : totalsel1,
2472 : : totalsel2;
2473 : : int i,
2474 : : nmatches;
2475 : :
2536 peter_e@gmx.net 2476 : 8395 : fmgr_info(opfuncoid, &eqproc);
2477 : :
2478 : : /*
2479 : : * Save a few cycles by setting up the fcinfo struct just once. Using
2480 : : * FunctionCallInvoke directly also avoids failure if the eqproc
2481 : : * returns NULL, though really equality functions should never do
2482 : : * that.
2483 : : */
1409 tgl@sss.pgh.pa.us 2484 : 8395 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2485 : : NULL, NULL);
1454 2486 : 8395 : fcinfo->args[0].isnull = false;
2487 : 8395 : fcinfo->args[1].isnull = false;
2488 : :
1969 2489 : 8395 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
2490 : 8395 : hasmatch2 = (bool *) palloc0(sslot2->nvalues * sizeof(bool));
2491 : :
2492 : : /*
2493 : : * Note we assume that each MCV will match at most one member of the
2494 : : * other MCV list. If the operator isn't really equality, there could
2495 : : * be multiple matches --- but we don't look for them, both for speed
2496 : : * and because the math wouldn't add up...
2497 : : */
7362 2498 : 8395 : matchprodfreq = 0.0;
2499 : 8395 : nmatches = 0;
1969 2500 [ + + ]: 268770 : for (i = 0; i < sslot1->nvalues; i++)
2501 : : {
2502 : : int j;
2503 : :
1454 2504 : 260375 : fcinfo->args[0].value = sslot1->values[i];
2505 : :
1969 2506 [ + + ]: 9861176 : for (j = 0; j < sslot2->nvalues; j++)
2507 : : {
2508 : : Datum fresult;
2509 : :
7362 2510 [ + + ]: 9682956 : if (hasmatch2[j])
2511 : 2593297 : continue;
1454 2512 : 7089659 : fcinfo->args[1].value = sslot2->values[j];
2513 : 7089659 : fcinfo->isnull = false;
2514 : 7089659 : fresult = FunctionCallInvoke(fcinfo);
2515 [ + - + + ]: 7089659 : if (!fcinfo->isnull && DatumGetBool(fresult))
2516 : : {
7362 2517 : 82155 : hasmatch1[i] = hasmatch2[j] = true;
1969 2518 : 82155 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
7362 2519 : 82155 : nmatches++;
2520 : 82155 : break;
2521 : : }
2522 : : }
2523 : : }
2524 [ - + - + ]: 8395 : CLAMP_PROBABILITY(matchprodfreq);
2525 : : /* Sum up frequencies of matched and unmatched MCVs */
2526 : 8395 : matchfreq1 = unmatchfreq1 = 0.0;
1969 2527 [ + + ]: 268770 : for (i = 0; i < sslot1->nvalues; i++)
2528 : : {
7362 2529 [ + + ]: 260375 : if (hasmatch1[i])
1969 2530 : 82155 : matchfreq1 += sslot1->numbers[i];
2531 : : else
2532 : 178220 : unmatchfreq1 += sslot1->numbers[i];
2533 : : }
7362 2534 [ - + + + ]: 8395 : CLAMP_PROBABILITY(matchfreq1);
2535 [ - + - + ]: 8395 : CLAMP_PROBABILITY(unmatchfreq1);
2536 : 8395 : matchfreq2 = unmatchfreq2 = 0.0;
1969 2537 [ + + ]: 323752 : for (i = 0; i < sslot2->nvalues; i++)
2538 : : {
7362 2539 [ + + ]: 315357 : if (hasmatch2[i])
1969 2540 : 82155 : matchfreq2 += sslot2->numbers[i];
2541 : : else
2542 : 233202 : unmatchfreq2 += sslot2->numbers[i];
2543 : : }
7362 2544 [ - + + + ]: 8395 : CLAMP_PROBABILITY(matchfreq2);
2545 [ - + - + ]: 8395 : CLAMP_PROBABILITY(unmatchfreq2);
2546 : 8395 : pfree(hasmatch1);
2547 : 8395 : pfree(hasmatch2);
2548 : :
2549 : : /*
2550 : : * Compute total frequency of non-null values that are not in the MCV
2551 : : * lists.
2552 : : */
2553 : 8395 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2554 : 8395 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2555 [ + + - + ]: 8395 : CLAMP_PROBABILITY(otherfreq1);
2556 [ + + - + ]: 8395 : CLAMP_PROBABILITY(otherfreq2);
2557 : :
2558 : : /*
2559 : : * We can estimate the total selectivity from the point of view of
2560 : : * relation 1 as: the known selectivity for matched MCVs, plus
2561 : : * unmatched MCVs that are assumed to match against random members of
2562 : : * relation 2's non-MCV population, plus non-MCV values that are
2563 : : * assumed to match against random members of relation 2's unmatched
2564 : : * MCVs plus non-MCV values.
2565 : : */
2566 : 8395 : totalsel1 = matchprodfreq;
1969 2567 [ + + ]: 8395 : if (nd2 > sslot2->nvalues)
2568 : 5274 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
7362 2569 [ + + ]: 8395 : if (nd2 > nmatches)
2570 : 7025 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2571 : 7025 : (nd2 - nmatches);
2572 : : /* Same estimate from the point of view of relation 2. */
2573 : 8395 : totalsel2 = matchprodfreq;
1969 2574 [ + + ]: 8395 : if (nd1 > sslot1->nvalues)
2575 : 5309 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
7362 2576 [ + + ]: 8395 : if (nd1 > nmatches)
2577 : 6509 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2578 : 6509 : (nd1 - nmatches);
2579 : :
2580 : : /*
2581 : : * Use the smaller of the two estimates. This can be justified in
2582 : : * essentially the same terms as given below for the no-stats case: to
2583 : : * a first approximation, we are estimating from the point of view of
2584 : : * the relation with smaller nd.
2585 : : */
2586 [ + + ]: 8395 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2587 : : }
2588 : : else
2589 : : {
2590 : : /*
2591 : : * We do not have MCV lists for both sides. Estimate the join
2592 : : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2593 : : * is plausible if we assume that the join operator is strict and the
2594 : : * non-null values are about equally distributed: a given non-null
2595 : : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2596 : : * of rel2, so total join rows are at most
2597 : : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2598 : : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2599 : : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2600 : : * with MIN() is an upper bound. Using the MIN() means we estimate
2601 : : * from the point of view of the relation with smaller nd (since the
2602 : : * larger nd is determining the MIN). It is reasonable to assume that
2603 : : * most tuples in this rel will have join partners, so the bound is
2604 : : * probably reasonably tight and should be taken as-is.
2605 : : *
2606 : : * XXX Can we be smarter if we have an MCV list for just one side? It
2607 : : * seems that if we assume equal distribution for the other side, we
2608 : : * end up with the same answer anyway.
2609 : : */
2610 [ + + ]: 87130 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2611 [ + + ]: 87130 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2612 : :
2613 : 87130 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2614 [ + + ]: 87130 : if (nd1 > nd2)
2615 : 41601 : selec /= nd1;
2616 : : else
2617 : 45529 : selec /= nd2;
2618 : : }
2619 : :
5720 2620 : 95525 : return selec;
2621 : : }
2622 : :
2623 : : /*
2624 : : * eqjoinsel_semi --- eqjoinsel for semi join
2625 : : *
2626 : : * (Also used for anti join, which we are supposed to estimate the same way.)
2627 : : * Caller has ensured that vardata1 is the LHS variable.
2628 : : * Unlike eqjoinsel_inner, we have to cope with opfuncoid being InvalidOid.
2629 : : */
2630 : : static double
1409 2631 : 3524 : eqjoinsel_semi(Oid opfuncoid, Oid collation,
2632 : : VariableStatData *vardata1, VariableStatData *vardata2,
2633 : : double nd1, double nd2,
2634 : : bool isdefault1, bool isdefault2,
2635 : : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2636 : : Form_pg_statistic stats1, Form_pg_statistic stats2,
2637 : : bool have_mcvs1, bool have_mcvs2,
2638 : : RelOptInfo *inner_rel)
2639 : : {
2640 : : double selec;
2641 : :
2642 : : /*
2643 : : * We clamp nd2 to be not more than what we estimate the inner relation's
2644 : : * size to be. This is intuitively somewhat reasonable since obviously
2645 : : * there can't be more than that many distinct values coming from the
2646 : : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2647 : : * likewise) is that this is the only pathway by which restriction clauses
2648 : : * applied to the inner rel will affect the join result size estimate,
2649 : : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2650 : : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2651 : : * the selectivity of outer-rel restrictions.
2652 : : *
2653 : : * We can apply this clamping both with respect to the base relation from
2654 : : * which the join variable comes (if there is just one), and to the
2655 : : * immediate inner input relation of the current join.
2656 : : *
2657 : : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2658 : : * great, maybe, but it didn't come out of nowhere either. This is most
2659 : : * helpful when the inner relation is empty and consequently has no stats.
2660 : : */
4609 2661 [ + + ]: 3524 : if (vardata2->rel)
2662 : : {
2693 2663 [ + + ]: 3521 : if (nd2 >= vardata2->rel->rows)
2664 : : {
2665 : 2703 : nd2 = vardata2->rel->rows;
2666 : 2703 : isdefault2 = false;
2667 : : }
2668 : : }
2669 [ + + ]: 3524 : if (nd2 >= inner_rel->rows)
2670 : : {
2671 : 2676 : nd2 = inner_rel->rows;
2672 : 2676 : isdefault2 = false;
2673 : : }
2674 : :
1969 2675 [ + + + - : 3524 : if (have_mcvs1 && have_mcvs2 && OidIsValid(opfuncoid))
+ - ]
5720 2676 : 264 : {
2677 : : /*
2678 : : * We have most-common-value lists for both relations. Run through
2679 : : * the lists to see which MCVs actually join to each other with the
2680 : : * given operator. This allows us to determine the exact join
2681 : : * selectivity for the portion of the relations represented by the MCV
2682 : : * lists. We still have to estimate for the remaining population, but
2683 : : * in a skewed distribution this gives us a big leg up in accuracy.
2684 : : */
1454 2685 : 264 : LOCAL_FCINFO(fcinfo, 2);
2686 : : FmgrInfo eqproc;
2687 : : bool *hasmatch1;
2688 : : bool *hasmatch2;
5720 2689 : 264 : double nullfrac1 = stats1->stanullfrac;
2690 : : double matchfreq1,
2691 : : uncertainfrac,
2692 : : uncertain;
2693 : : int i,
2694 : : nmatches,
2695 : : clamped_nvalues2;
2696 : :
2697 : : /*
2698 : : * The clamping above could have resulted in nd2 being less than
2699 : : * sslot2->nvalues; in which case, we assume that precisely the nd2
2700 : : * most common values in the relation will appear in the join input,
2701 : : * and so compare to only the first nd2 members of the MCV list. Of
2702 : : * course this is frequently wrong, but it's the best bet we can make.
2703 : : */
1969 2704 [ + + ]: 264 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2705 : :
2536 peter_e@gmx.net 2706 : 264 : fmgr_info(opfuncoid, &eqproc);
2707 : :
2708 : : /*
2709 : : * Save a few cycles by setting up the fcinfo struct just once. Using
2710 : : * FunctionCallInvoke directly also avoids failure if the eqproc
2711 : : * returns NULL, though really equality functions should never do
2712 : : * that.
2713 : : */
1409 tgl@sss.pgh.pa.us 2714 : 264 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
2715 : : NULL, NULL);
1454 2716 : 264 : fcinfo->args[0].isnull = false;
2717 : 264 : fcinfo->args[1].isnull = false;
2718 : :
1969 2719 : 264 : hasmatch1 = (bool *) palloc0(sslot1->nvalues * sizeof(bool));
4609 2720 : 264 : hasmatch2 = (bool *) palloc0(clamped_nvalues2 * sizeof(bool));
2721 : :
2722 : : /*
2723 : : * Note we assume that each MCV will match at most one member of the
2724 : : * other MCV list. If the operator isn't really equality, there could
2725 : : * be multiple matches --- but we don't look for them, both for speed
2726 : : * and because the math wouldn't add up...
2727 : : */
5720 2728 : 264 : nmatches = 0;
1969 2729 [ + + ]: 6275 : for (i = 0; i < sslot1->nvalues; i++)
2730 : : {
2731 : : int j;
2732 : :
1454 2733 : 6011 : fcinfo->args[0].value = sslot1->values[i];
2734 : :
4609 2735 [ + + ]: 254239 : for (j = 0; j < clamped_nvalues2; j++)
2736 : : {
2737 : : Datum fresult;
2738 : :
5720 2739 [ + + ]: 253295 : if (hasmatch2[j])
2740 : 192596 : continue;
1454 2741 : 60699 : fcinfo->args[1].value = sslot2->values[j];
2742 : 60699 : fcinfo->isnull = false;
2743 : 60699 : fresult = FunctionCallInvoke(fcinfo);
2744 [ + - + + ]: 60699 : if (!fcinfo->isnull && DatumGetBool(fresult))
2745 : : {
5720 2746 : 5067 : hasmatch1[i] = hasmatch2[j] = true;
2747 : 5067 : nmatches++;
2748 : 5067 : break;
2749 : : }
2750 : : }
2751 : : }
2752 : : /* Sum up frequencies of matched MCVs */
2753 : 264 : matchfreq1 = 0.0;
1969 2754 [ + + ]: 6275 : for (i = 0; i < sslot1->nvalues; i++)
2755 : : {
5720 2756 [ + + ]: 6011 : if (hasmatch1[i])
1969 2757 : 5067 : matchfreq1 += sslot1->numbers[i];
2758 : : }
5720 2759 [ - + + + ]: 264 : CLAMP_PROBABILITY(matchfreq1);
2760 : 264 : pfree(hasmatch1);
2761 : 264 : pfree(hasmatch2);
2762 : :
2763 : : /*
2764 : : * Now we need to estimate the fraction of relation 1 that has at
2765 : : * least one join partner. We know for certain that the matched MCVs
2766 : : * do, so that gives us a lower bound, but we're really in the dark
2767 : : * about everything else. Our crude approach is: if nd1 <= nd2 then
2768 : : * assume all non-null rel1 rows have join partners, else assume for
2769 : : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2770 : : * can discount the known-matched MCVs from the distinct-values counts
2771 : : * before doing the division.
2772 : : *
2773 : : * Crude as the above is, it's completely useless if we don't have
2774 : : * reliable ndistinct values for both sides. Hence, if either nd1 or
2775 : : * nd2 is default, punt and assume half of the uncertain rows have
2776 : : * join partners.
2777 : : */
4606 2778 [ + - + - ]: 264 : if (!isdefault1 && !isdefault2)
2779 : : {
4751 2780 : 264 : nd1 -= nmatches;
2781 : 264 : nd2 -= nmatches;
4609 2782 [ + + - + ]: 264 : if (nd1 <= nd2 || nd2 < 0)
4751 2783 : 249 : uncertainfrac = 1.0;
2784 : : else
2785 : 15 : uncertainfrac = nd2 / nd1;
2786 : : }
2787 : : else
4751 tgl@sss.pgh.pa.us 2788 :UBC 0 : uncertainfrac = 0.5;
4751 tgl@sss.pgh.pa.us 2789 :CBC 264 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2790 [ - + - + ]: 264 : CLAMP_PROBABILITY(uncertain);
2791 : 264 : selec = matchfreq1 + uncertainfrac * uncertain;
2792 : : }
2793 : : else
2794 : : {
2795 : : /*
2796 : : * Without MCV lists for both sides, we can only use the heuristic
2797 : : * about nd1 vs nd2.
2798 : : */
5720 2799 [ + + ]: 3260 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2800 : :
4606 2801 [ + + + + ]: 3260 : if (!isdefault1 && !isdefault2)
2802 : : {
4609 2803 [ + + - + ]: 2324 : if (nd1 <= nd2 || nd2 < 0)
4751 2804 : 1212 : selec = 1.0 - nullfrac1;
2805 : : else
2806 : 1112 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2807 : : }
2808 : : else
2809 : 936 : selec = 0.5 * (1.0 - nullfrac1);
2810 : : }
2811 : :
5720 2812 : 3524 : return selec;
2813 : : }
2814 : :
2815 : : /*
2816 : : * neqjoinsel - Join selectivity of "!="
2817 : : */
2818 : : Datum
8714 2819 : 1660 : neqjoinsel(PG_FUNCTION_ARGS)
2820 : : {
6888 2821 : 1660 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
8358 2822 : 1660 : Oid operator = PG_GETARG_OID(1);
2823 : 1660 : List *args = (List *) PG_GETARG_POINTER(2);
7747 2824 : 1660 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
5720 2825 : 1660 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
1363 2826 : 1660 : Oid collation = PG_GET_COLLATION();
2827 : : float8 result;
2828 : :
2328 2829 [ + + - + ]: 1660 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
8358 2830 : 549 : {
2831 : : /*
2832 : : * For semi-joins, if there is more than one distinct value in the RHS
2833 : : * relation then every non-null LHS row must find a row to join since
2834 : : * it can only be equal to one of them. We'll assume that there is
2835 : : * always more than one distinct RHS value for the sake of stability,
2836 : : * though in theory we could have special cases for empty RHS
2837 : : * (selectivity = 0) and single-distinct-value RHS (selectivity =
2838 : : * fraction of LHS that has the same value as the single RHS value).
2839 : : *
2840 : : * For anti-joins, if we use the same assumption that there is more
2841 : : * than one distinct key in the RHS relation, then every non-null LHS
2842 : : * row must be suppressed by the anti-join.
2843 : : *
2844 : : * So either way, the selectivity estimate should be 1 - nullfrac.
2845 : : */
2846 : : VariableStatData leftvar;
2847 : : VariableStatData rightvar;
2848 : : bool reversed;
2849 : : HeapTuple statsTuple;
2850 : : double nullfrac;
2851 : :
2328 2852 : 549 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
2853 [ + + ]: 549 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
2854 [ + + ]: 549 : if (HeapTupleIsValid(statsTuple))
2855 : 452 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
2856 : : else
2857 : 97 : nullfrac = 0.0;
2858 [ + + ]: 549 : ReleaseVariableStats(leftvar);
2859 [ + + ]: 549 : ReleaseVariableStats(rightvar);
2860 : :
2861 : 549 : result = 1.0 - nullfrac;
2862 : : }
2863 : : else
2864 : : {
2865 : : /*
2866 : : * We want 1 - eqjoinsel() where the equality operator is the one
2867 : : * associated with this != operator, that is, its negator.
2868 : : */
2869 : 1111 : Oid eqop = get_negator(operator);
2870 : :
2871 [ + - ]: 1111 : if (eqop)
2872 : : {
2873 : : result =
1363 2874 : 1111 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
2875 : : collation,
2876 : : PointerGetDatum(root),
2877 : : ObjectIdGetDatum(eqop),
2878 : : PointerGetDatum(args),
2879 : : Int16GetDatum(jointype),
2880 : : PointerGetDatum(sjinfo)));
2881 : : }
2882 : : else
2883 : : {
2884 : : /* Use default selectivity (should we raise an error instead?) */
2328 tgl@sss.pgh.pa.us 2885 :UBC 0 : result = DEFAULT_EQ_SEL;
2886 : : }
2328 tgl@sss.pgh.pa.us 2887 :CBC 1111 : result = 1.0 - result;
2888 : : }
2889 : :
8714 2890 : 1660 : PG_RETURN_FLOAT8(result);
2891 : : }
2892 : :
2893 : : /*
2894 : : * scalarltjoinsel - Join selectivity of "<" for scalars
2895 : : */
2896 : : Datum
2897 : 159 : scalarltjoinsel(PG_FUNCTION_ARGS)
2898 : : {
2899 : 159 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2900 : : }
2901 : :
2902 : : /*
2903 : : * scalarlejoinsel - Join selectivity of "<=" for scalars
2904 : : */
2905 : : Datum
2405 2906 : 138 : scalarlejoinsel(PG_FUNCTION_ARGS)
2907 : : {
2908 : 138 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2909 : : }
2910 : :
2911 : : /*
2912 : : * scalargtjoinsel - Join selectivity of ">" for scalars
2913 : : */
2914 : : Datum
8714 2915 : 114 : scalargtjoinsel(PG_FUNCTION_ARGS)
2916 : : {
2917 : 114 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2918 : : }
2919 : :
2920 : : /*
2921 : : * scalargejoinsel - Join selectivity of ">=" for scalars
2922 : : */
2923 : : Datum
2405 2924 : 92 : scalargejoinsel(PG_FUNCTION_ARGS)
2925 : : {
2926 : 92 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
2927 : : }
2928 : :
2929 : :
2930 : : /*
2931 : : * mergejoinscansel - Scan selectivity of merge join.
2932 : : *
2933 : : * A merge join will stop as soon as it exhausts either input stream.
2934 : : * Therefore, if we can estimate the ranges of both input variables,
2935 : : * we can estimate how much of the input will actually be read. This
2936 : : * can have a considerable impact on the cost when using indexscans.
2937 : : *
2938 : : * Also, we can estimate how much of each input has to be read before the
2939 : : * first join pair is found, which will affect the join's startup time.
2940 : : *
2941 : : * clause should be a clause already known to be mergejoinable. opfamily,
2942 : : * strategy, and nulls_first specify the sort ordering being used.
2943 : : *
2944 : : * The outputs are:
2945 : : * *leftstart is set to the fraction of the left-hand variable expected
2946 : : * to be scanned before the first join pair is found (0 to 1).
2947 : : * *leftend is set to the fraction of the left-hand variable expected
2948 : : * to be scanned before the join terminates (0 to 1).
2949 : : * *rightstart, *rightend similarly for the right-hand variable.
2950 : : */
2951 : : void
1886 2952 : 48325 : mergejoinscansel(PlannerInfo *root, Node *clause,
2953 : : Oid opfamily, int strategy, bool nulls_first,
2954 : : Selectivity *leftstart, Selectivity *leftend,
2955 : : Selectivity *rightstart, Selectivity *rightend)
2956 : : {
2957 : : Node *left,
2958 : : *right;
2959 : : VariableStatData leftvar,
2960 : : rightvar;
2961 : : int op_strategy;
2962 : : Oid op_lefttype;
2963 : : Oid op_righttype;
2964 : : Oid opno,
2965 : : collation,
2966 : : lsortop,
2967 : : rsortop,
2968 : : lstatop,
2969 : : rstatop,
2970 : : ltop,
2971 : : leop,
2972 : : revltop,
2973 : : revleop;
2974 : : bool isgt;
2975 : : Datum leftmin,
2976 : : leftmax,
2977 : : rightmin,
2978 : : rightmax;
2979 : : double selec;
2980 : :
2981 : : /* Set default results if we can't figure anything out. */
2982 : : /* XXX should default "start" fraction be a bit more than 0? */
5972 2983 : 48325 : *leftstart = *rightstart = 0.0;
2984 : 48325 : *leftend = *rightend = 1.0;
2985 : :
2986 : : /* Deconstruct the merge clause */
8080 2987 [ - + ]: 48325 : if (!is_opclause(clause))
8080 tgl@sss.pgh.pa.us 2988 :UBC 0 : return; /* shouldn't happen */
7794 tgl@sss.pgh.pa.us 2989 :CBC 48325 : opno = ((OpExpr *) clause)->opno;
1409 2990 : 48325 : collation = ((OpExpr *) clause)->inputcollid;
7362 2991 : 48325 : left = get_leftop((Expr *) clause);
2992 : 48325 : right = get_rightop((Expr *) clause);
8080 2993 [ - + ]: 48325 : if (!right)
8080 tgl@sss.pgh.pa.us 2994 :UBC 0 : return; /* shouldn't happen */
2995 : :
2996 : : /* Look for stats for the inputs */
7362 tgl@sss.pgh.pa.us 2997 :CBC 48325 : examine_variable(root, left, 0, &leftvar);
2998 : 48325 : examine_variable(root, right, 0, &rightvar);
2999 : :
3000 : : /* Extract the operator's declared left/right datatypes */
4882 3001 : 48325 : get_op_opfamily_properties(opno, opfamily, false,
3002 : : &op_strategy,
3003 : : &op_lefttype,
3004 : : &op_righttype);
6322 3005 [ - + ]: 48325 : Assert(op_strategy == BTEqualStrategyNumber);
3006 : :
3007 : : /*
3008 : : * Look up the various operators we need. If we don't find them all, it
3009 : : * probably means the opfamily is broken, but we just fail silently.
3010 : : *
3011 : : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3012 : : * operator, regardless of which sort direction we are considering.
3013 : : */
3014 [ + + - ]: 48325 : switch (strategy)
3015 : : {
3016 : 48298 : case BTLessStrategyNumber:
5972 3017 : 48298 : isgt = false;
3018 [ + + ]: 48298 : if (op_lefttype == op_righttype)
3019 : : {
3020 : : /* easy case */
3021 : 47661 : ltop = get_opfamily_member(opfamily,
3022 : : op_lefttype, op_righttype,
3023 : : BTLessStrategyNumber);
3024 : 47661 : leop = get_opfamily_member(opfamily,
3025 : : op_lefttype, op_righttype,
3026 : : BTLessEqualStrategyNumber);
3027 : 47661 : lsortop = ltop;
3028 : 47661 : rsortop = ltop;
3029 : 47661 : lstatop = lsortop;
3030 : 47661 : rstatop = rsortop;
3031 : 47661 : revltop = ltop;
3032 : 47661 : revleop = leop;
3033 : : }
3034 : : else
3035 : : {
3036 : 637 : ltop = get_opfamily_member(opfamily,
3037 : : op_lefttype, op_righttype,
3038 : : BTLessStrategyNumber);
3039 : 637 : leop = get_opfamily_member(opfamily,
3040 : : op_lefttype, op_righttype,
3041 : : BTLessEqualStrategyNumber);
3042 : 637 : lsortop = get_opfamily_member(opfamily,
3043 : : op_lefttype, op_lefttype,
3044 : : BTLessStrategyNumber);
3045 : 637 : rsortop = get_opfamily_member(opfamily,
3046 : : op_righttype, op_righttype,
3047 : : BTLessStrategyNumber);
3048 : 637 : lstatop = lsortop;
3049 : 637 : rstatop = rsortop;
3050 : 637 : revltop = get_opfamily_member(opfamily,
3051 : : op_righttype, op_lefttype,
3052 : : BTLessStrategyNumber);
3053 : 637 : revleop = get_opfamily_member(opfamily,
3054 : : op_righttype, op_lefttype,
3055 : : BTLessEqualStrategyNumber);
3056 : : }
6322 3057 : 48298 : break;
3058 : 27 : case BTGreaterStrategyNumber:
3059 : : /* descending-order case */
5972 3060 : 27 : isgt = true;
3061 [ + - ]: 27 : if (op_lefttype == op_righttype)
3062 : : {
3063 : : /* easy case */
3064 : 27 : ltop = get_opfamily_member(opfamily,
3065 : : op_lefttype, op_righttype,
3066 : : BTGreaterStrategyNumber);
3067 : 27 : leop = get_opfamily_member(opfamily,
3068 : : op_lefttype, op_righttype,
3069 : : BTGreaterEqualStrategyNumber);
3070 : 27 : lsortop = ltop;
3071 : 27 : rsortop = ltop;
3072 : 27 : lstatop = get_opfamily_member(opfamily,
3073 : : op_lefttype, op_lefttype,
3074 : : BTLessStrategyNumber);
3075 : 27 : rstatop = lstatop;
3076 : 27 : revltop = ltop;
3077 : 27 : revleop = leop;
3078 : : }
3079 : : else
3080 : : {
5972 tgl@sss.pgh.pa.us 3081 :UBC 0 : ltop = get_opfamily_member(opfamily,
3082 : : op_lefttype, op_righttype,
3083 : : BTGreaterStrategyNumber);
3084 : 0 : leop = get_opfamily_member(opfamily,
3085 : : op_lefttype, op_righttype,
3086 : : BTGreaterEqualStrategyNumber);
3087 : 0 : lsortop = get_opfamily_member(opfamily,
3088 : : op_lefttype, op_lefttype,
3089 : : BTGreaterStrategyNumber);
3090 : 0 : rsortop = get_opfamily_member(opfamily,
3091 : : op_righttype, op_righttype,
3092 : : BTGreaterStrategyNumber);
3093 : 0 : lstatop = get_opfamily_member(opfamily,
3094 : : op_lefttype, op_lefttype,
3095 : : BTLessStrategyNumber);
3096 : 0 : rstatop = get_opfamily_member(opfamily,
3097 : : op_righttype, op_righttype,
3098 : : BTLessStrategyNumber);
3099 : 0 : revltop = get_opfamily_member(opfamily,
3100 : : op_righttype, op_lefttype,
3101 : : BTGreaterStrategyNumber);
3102 : 0 : revleop = get_opfamily_member(opfamily,
3103 : : op_righttype, op_lefttype,
3104 : : BTGreaterEqualStrategyNumber);
3105 : : }
6322 tgl@sss.pgh.pa.us 3106 :CBC 27 : break;
6322 tgl@sss.pgh.pa.us 3107 :UBC 0 : default:
3108 : 0 : goto fail; /* shouldn't get here */
3109 : : }
3110 : :
6322 tgl@sss.pgh.pa.us 3111 [ + - + - ]:CBC 48325 : if (!OidIsValid(lsortop) ||
3112 [ + - ]: 48325 : !OidIsValid(rsortop) ||
5972 3113 [ + - ]: 48325 : !OidIsValid(lstatop) ||
3114 [ + + ]: 48325 : !OidIsValid(rstatop) ||
3115 [ + - ]: 48319 : !OidIsValid(ltop) ||
6322 3116 [ + - ]: 48319 : !OidIsValid(leop) ||
5972 3117 [ - + ]: 48319 : !OidIsValid(revltop) ||
3118 : : !OidIsValid(revleop))
6322 3119 : 6 : goto fail; /* insufficient info in catalogs */
3120 : :
3121 : : /* Try to get ranges of both inputs */
5972 3122 [ + + ]: 48319 : if (!isgt)
3123 : : {
1409 3124 [ + + ]: 48292 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3125 : : &leftmin, &leftmax))
5972 3126 : 9935 : goto fail; /* no range available from stats */
1409 3127 [ + + ]: 38357 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3128 : : &rightmin, &rightmax))
5972 3129 : 8701 : goto fail; /* no range available from stats */
3130 : : }
3131 : : else
3132 : : {
3133 : : /* need to swap the max and min */
1409 3134 [ + + ]: 27 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3135 : : &leftmax, &leftmin))
5972 3136 : 15 : goto fail; /* no range available from stats */
1409 3137 [ - + ]: 12 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3138 : : &rightmax, &rightmin))
5972 tgl@sss.pgh.pa.us 3139 :UBC 0 : goto fail; /* no range available from stats */
3140 : : }
3141 : :
3142 : : /*
3143 : : * Now, the fraction of the left variable that will be scanned is the
3144 : : * fraction that's <= the right-side maximum value. But only believe
3145 : : * non-default estimates, else stick with our 1.0.
3146 : : */
1409 tgl@sss.pgh.pa.us 3147 :CBC 29668 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3148 : : rightmax, op_righttype);
8080 3149 [ + + ]: 29668 : if (selec != DEFAULT_INEQ_SEL)
5972 3150 : 29665 : *leftend = selec;
3151 : :
3152 : : /* And similarly for the right variable. */
1409 3153 : 29668 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3154 : : leftmax, op_lefttype);
8080 3155 [ + - ]: 29668 : if (selec != DEFAULT_INEQ_SEL)
5972 3156 : 29668 : *rightend = selec;
3157 : :
3158 : : /*
3159 : : * Only one of the two "end" fractions can really be less than 1.0;
3160 : : * believe the smaller estimate and reset the other one to exactly 1.0. If
3161 : : * we get exactly equal estimates (as can easily happen with self-joins),
3162 : : * believe neither.
3163 : : */
3164 [ + + ]: 29668 : if (*leftend > *rightend)
3165 : 12972 : *leftend = 1.0;
3166 [ + + ]: 16696 : else if (*leftend < *rightend)
3167 : 13015 : *rightend = 1.0;
3168 : : else
3169 : 3681 : *leftend = *rightend = 1.0;
3170 : :
3171 : : /*
3172 : : * Also, the fraction of the left variable that will be scanned before the
3173 : : * first join pair is found is the fraction that's < the right-side
3174 : : * minimum value. But only believe non-default estimates, else stick with
3175 : : * our own default.
3176 : : */
1409 3177 : 29668 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3178 : : rightmin, op_righttype);
5972 3179 [ + - ]: 29668 : if (selec != DEFAULT_INEQ_SEL)
3180 : 29668 : *leftstart = selec;
3181 : :
3182 : : /* And similarly for the right variable. */
1409 3183 : 29668 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3184 : : leftmin, op_lefttype);
5972 3185 [ + - ]: 29668 : if (selec != DEFAULT_INEQ_SEL)
3186 : 29668 : *rightstart = selec;
3187 : :
3188 : : /*
3189 : : * Only one of the two "start" fractions can really be more than zero;
3190 : : * believe the larger estimate and reset the other one to exactly 0.0. If
3191 : : * we get exactly equal estimates (as can easily happen with self-joins),
3192 : : * believe neither.
3193 : : */
3194 [ + + ]: 29668 : if (*leftstart < *rightstart)
3195 : 8429 : *leftstart = 0.0;
3196 [ + + ]: 21239 : else if (*leftstart > *rightstart)
3197 : 13784 : *rightstart = 0.0;
3198 : : else
3199 : 7455 : *leftstart = *rightstart = 0.0;
3200 : :
3201 : : /*
3202 : : * If the sort order is nulls-first, we're going to have to skip over any
3203 : : * nulls too. These would not have been counted by scalarineqsel, and we
3204 : : * can safely add in this fraction regardless of whether we believe
3205 : : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3206 : : */
3207 [ + + ]: 29668 : if (nulls_first)
3208 : : {
3209 : : Form_pg_statistic stats;
3210 : :
3211 [ + - ]: 12 : if (HeapTupleIsValid(leftvar.statsTuple))
3212 : : {
3213 : 12 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3214 : 12 : *leftstart += stats->stanullfrac;
3215 [ - + - + ]: 12 : CLAMP_PROBABILITY(*leftstart);
3216 : 12 : *leftend += stats->stanullfrac;
3217 [ - + - + ]: 12 : CLAMP_PROBABILITY(*leftend);
3218 : : }
3219 [ + - ]: 12 : if (HeapTupleIsValid(rightvar.statsTuple))
3220 : : {
6292 3221 : 12 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
5972 3222 : 12 : *rightstart += stats->stanullfrac;
3223 [ - + - + ]: 12 : CLAMP_PROBABILITY(*rightstart);
3224 : 12 : *rightend += stats->stanullfrac;
3225 [ - + - + ]: 12 : CLAMP_PROBABILITY(*rightend);
3226 : : }
3227 : : }
3228 : :
3229 : : /* Disbelieve start >= end, just in case that can happen */
3230 [ + + ]: 29668 : if (*leftstart >= *leftend)
3231 : : {
3232 : 80 : *leftstart = 0.0;
3233 : 80 : *leftend = 1.0;
3234 : : }
3235 [ + + ]: 29668 : if (*rightstart >= *rightend)
3236 : : {
3237 : 495 : *rightstart = 0.0;
3238 : 495 : *rightend = 1.0;
3239 : : }
3240 : :
7362 3241 : 29173 : fail:
3242 [ + + ]: 48325 : ReleaseVariableStats(leftvar);
3243 [ + + ]: 48325 : ReleaseVariableStats(rightvar);
3244 : : }
3245 : :
3246 : :
3247 : : /*
3248 : : * matchingsel -- generic matching-operator selectivity support
3249 : : *
3250 : : * Use these for any operators that (a) are on data types for which we collect
3251 : : * standard statistics, and (b) have behavior for which the default estimate
3252 : : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3253 : : * operators.
3254 : : */
3255 : :
3256 : : Datum
1474 3257 : 553 : matchingsel(PG_FUNCTION_ARGS)
3258 : : {
3259 : 553 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3260 : 553 : Oid operator = PG_GETARG_OID(1);
3261 : 553 : List *args = (List *) PG_GETARG_POINTER(2);
3262 : 553 : int varRelid = PG_GETARG_INT32(3);
1409 3263 : 553 : Oid collation = PG_GET_COLLATION();
3264 : : double selec;
3265 : :
3266 : : /* Use generic restriction selectivity logic. */
3267 : 553 : selec = generic_restriction_selectivity(root, operator, collation,
3268 : : args, varRelid,
3269 : : DEFAULT_MATCHING_SEL);
3270 : :
1474 3271 : 553 : PG_RETURN_FLOAT8((float8) selec);
3272 : : }
3273 : :
3274 : : Datum
3275 : 3 : matchingjoinsel(PG_FUNCTION_ARGS)
3276 : : {
3277 : : /* Just punt, for the moment. */
3278 : 3 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3279 : : }
3280 : :
3281 : :
3282 : : /*
3283 : : * Helper routine for estimate_num_groups: add an item to a list of
3284 : : * GroupVarInfos, but only if it's not known equal to any of the existing
3285 : : * entries.
3286 : : */
3287 : : typedef struct
3288 : : {
3289 : : Node *var; /* might be an expression, not just a Var */
3290 : : RelOptInfo *rel; /* relation it belongs to */
3291 : : double ndistinct; /* # distinct values */
3292 : : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3293 : : } GroupVarInfo;
3294 : :
3295 : : static List *
6888 3296 : 143013 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3297 : : Node *var, VariableStatData *vardata)
3298 : : {
3299 : : GroupVarInfo *varinfo;
3300 : : double ndistinct;
3301 : : bool isdefault;
3302 : : ListCell *lc;
3303 : :
4606 3304 : 143013 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3305 : :
1735 3306 [ + + + + : 164761 : foreach(lc, varinfos)
+ + ]
3307 : : {
7148 3308 : 22123 : varinfo = (GroupVarInfo *) lfirst(lc);
3309 : :
3310 : : /* Drop exact duplicates */
3311 [ + + ]: 22123 : if (equal(var, varinfo->var))
3312 : 375 : return varinfos;
3313 : :
3314 : : /*
3315 : : * Drop known-equal vars, but only if they belong to different
3316 : : * relations (see comments for estimate_num_groups)
3317 : : */
3318 [ + + + + ]: 23427 : if (vardata->rel != varinfo->rel &&
3319 : 1619 : exprs_known_equal(root, var, varinfo->var))
3320 : : {
3321 [ + - ]: 60 : if (varinfo->ndistinct <= ndistinct)
3322 : : {
3323 : : /* Keep older item, forget new one */
3324 : 60 : return varinfos;
3325 : : }
3326 : : else
3327 : : {
3328 : : /* Delete the older item */
1735 tgl@sss.pgh.pa.us 3329 :UBC 0 : varinfos = foreach_delete_current(varinfos, lc);
3330 : : }
3331 : : }
3332 : : }
3333 : :
7148 tgl@sss.pgh.pa.us 3334 :CBC 142638 : varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
3335 : :
3336 : 142638 : varinfo->var = var;
3337 : 142638 : varinfo->rel = vardata->rel;
3338 : 142638 : varinfo->ndistinct = ndistinct;
1111 drowley@postgresql.o 3339 : 142638 : varinfo->isdefault = isdefault;
7148 tgl@sss.pgh.pa.us 3340 : 142638 : varinfos = lappend(varinfos, varinfo);
3341 : 142638 : return varinfos;
3342 : : }
3343 : :
3344 : : /*
3345 : : * estimate_num_groups - Estimate number of groups in a grouped query
3346 : : *
3347 : : * Given a query having a GROUP BY clause, estimate how many groups there
3348 : : * will be --- ie, the number of distinct combinations of the GROUP BY
3349 : : * expressions.
3350 : : *
3351 : : * This routine is also used to estimate the number of rows emitted by
3352 : : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3353 : : * actually, we only use it for DISTINCT when there's no grouping or
3354 : : * aggregation ahead of the DISTINCT.)
3355 : : *
3356 : : * Inputs:
3357 : : * root - the query
3358 : : * groupExprs - list of expressions being grouped by
3359 : : * input_rows - number of rows estimated to arrive at the group/unique
3360 : : * filter step
3361 : : * pgset - NULL, or a List** pointing to a grouping set to filter the
3362 : : * groupExprs against
3363 : : *
3364 : : * Outputs:
3365 : : * estinfo - When passed as non-NULL, the function will set bits in the
3366 : : * "flags" field in order to provide callers with additional information
3367 : : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3368 : : * bit if we used any default values in the estimation.
3369 : : *
3370 : : * Given the lack of any cross-correlation statistics in the system, it's
3371 : : * impossible to do anything really trustworthy with GROUP BY conditions
3372 : : * involving multiple Vars. We should however avoid assuming the worst
3373 : : * case (all possible cross-product terms actually appear as groups) since
3374 : : * very often the grouped-by Vars are highly correlated. Our current approach
3375 : : * is as follows:
3376 : : * 1. Expressions yielding boolean are assumed to contribute two groups,
3377 : : * independently of their content, and are ignored in the subsequent
3378 : : * steps. This is mainly because tests like "col IS NULL" break the
3379 : : * heuristic used in step 2 especially badly.
3380 : : * 2. Reduce the given expressions to a list of unique Vars used. For
3381 : : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3382 : : * It is clearly correct not to count the same Var more than once.
3383 : : * It is also reasonable to treat f(x) the same as x: f() cannot
3384 : : * increase the number of distinct values (unless it is volatile,
3385 : : * which we consider unlikely for grouping), but it probably won't
3386 : : * reduce the number of distinct values much either.
3387 : : * As a special case, if a GROUP BY expression can be matched to an
3388 : : * expressional index for which we have statistics, then we treat the
3389 : : * whole expression as though it were just a Var.
3390 : : * 3. If the list contains Vars of different relations that are known equal
3391 : : * due to equivalence classes, then drop all but one of the Vars from each
3392 : : * known-equal set, keeping the one with smallest estimated # of values
3393 : : * (since the extra values of the others can't appear in joined rows).
3394 : : * Note the reason we only consider Vars of different relations is that
3395 : : * if we considered ones of the same rel, we'd be double-counting the
3396 : : * restriction selectivity of the equality in the next step.
3397 : : * 4. For Vars within a single source rel, we multiply together the numbers
3398 : : * of values, clamp to the number of rows in the rel (divided by 10 if
3399 : : * more than one Var), and then multiply by a factor based on the
3400 : : * selectivity of the restriction clauses for that rel. When there's
3401 : : * more than one Var, the initial product is probably too high (it's the
3402 : : * worst case) but clamping to a fraction of the rel's rows seems to be a
3403 : : * helpful heuristic for not letting the estimate get out of hand. (The
3404 : : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3405 : : * we multiply by to adjust for the restriction selectivity assumes that
3406 : : * the restriction clauses are independent of the grouping, which may not
3407 : : * be a valid assumption, but it's hard to do better.
3408 : : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3409 : : * rel, and multiply the results together.
3410 : : * Note that rels not containing grouped Vars are ignored completely, as are
3411 : : * join clauses. Such rels cannot increase the number of groups, and we
3412 : : * assume such clauses do not reduce the number either (somewhat bogus,
3413 : : * but we don't have the info to do better).
3414 : : */
3415 : : double
3256 andres@anarazel.de 3416 : 125992 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3417 : : List **pgset, EstimationInfo *estinfo)
3418 : : {
559 tgl@sss.pgh.pa.us 3419 : 125992 : List *varinfos = NIL;
2332 3420 : 125992 : double srf_multiplier = 1.0;
3421 : : double numdistinct;
3422 : : ListCell *l;
3423 : : int i;
3424 : :
3425 : : /* Zero the estinfo output parameter, if non-NULL */
1111 drowley@postgresql.o 3426 [ + + ]: 125992 : if (estinfo != NULL)
3427 : 112612 : memset(estinfo, 0, sizeof(EstimationInfo));
3428 : :
3429 : : /*
3430 : : * We don't ever want to return an estimate of zero groups, as that tends
3431 : : * to lead to division-by-zero and other unpleasantness. The input_rows
3432 : : * estimate is usually already at least 1, but clamp it just in case it
3433 : : * isn't.
3434 : : */
3992 tgl@sss.pgh.pa.us 3435 : 125992 : input_rows = clamp_row_est(input_rows);
3436 : :
3437 : : /*
3438 : : * If no grouping columns, there's exactly one group. (This can't happen
3439 : : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3440 : : * corner cases with set operations.)
3441 : : */
606 3442 [ + + + + : 125992 : if (groupExprs == NIL || (pgset && *pgset == NIL))
+ + ]
4388 3443 : 478 : return 1.0;
3444 : :
3445 : : /*
3446 : : * Count groups derived from boolean grouping expressions. For other
3447 : : * expressions, find the unique Vars used, treating an expression as a Var
3448 : : * if we can find stats for it. For each one, record the statistical
3449 : : * estimate of number of distinct values (total in its table, without
3450 : : * regard for filtering).
3451 : : */
5760 3452 : 125514 : numdistinct = 1.0;
3453 : :
559 3454 : 125514 : i = 0;
7755 3455 [ + - + + : 267689 : foreach(l, groupExprs)
+ + ]
3456 : : {
3457 : 142199 : Node *groupexpr = (Node *) lfirst(l);
3458 : : double this_srf_multiplier;
3459 : : VariableStatData vardata;
3460 : : List *varshere;
3461 : : ListCell *l2;
3462 : :
3463 : : /* is expression in this grouping set? */
3256 andres@anarazel.de 3464 [ + + + + ]: 142199 : if (pgset && !list_member_int(*pgset, i++))
3465 : 119758 : continue;
3466 : :
3467 : : /*
3468 : : * Set-returning functions in grouping columns are a bit problematic.
3469 : : * The code below will effectively ignore their SRF nature and come up
3470 : : * with a numdistinct estimate as though they were scalar functions.
3471 : : * We compensate by scaling up the end result by the largest SRF
3472 : : * rowcount estimate. (This will be an overestimate if the SRF
3473 : : * produces multiple copies of any output value, but it seems best to
3474 : : * assume the SRF's outputs are distinct. In any case, it's probably
3475 : : * pointless to worry too much about this without much better
3476 : : * estimates for SRF output rowcounts than we have today.)
3477 : : */
1891 tgl@sss.pgh.pa.us 3478 : 141835 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
2332 3479 [ + + ]: 141835 : if (srf_multiplier < this_srf_multiplier)
3480 : 66 : srf_multiplier = this_srf_multiplier;
3481 : :
3482 : : /* Short-circuit for expressions returning boolean */
5760 3483 [ + + ]: 141835 : if (exprType(groupexpr) == BOOLOID)
3484 : : {
3485 : 57 : numdistinct *= 2.0;
3486 : 57 : continue;
3487 : : }
3488 : :
3489 : : /*
3490 : : * If examine_variable is able to deduce anything about the GROUP BY
3491 : : * expression, treat it as a single variable even if it's really more
3492 : : * complicated.
3493 : : *
3494 : : * XXX This has the consequence that if there's a statistics object on
3495 : : * the expression, we don't split it into individual Vars. This
3496 : : * affects our selection of statistics in
3497 : : * estimate_multivariate_ndistinct, because it's probably better to
3498 : : * use more accurate estimate for each expression and treat them as
3499 : : * independent, than to combine estimates for the extracted variables
3500 : : * when we don't know how that relates to the expressions.
3501 : : */
7148 3502 : 141778 : examine_variable(root, groupexpr, 0, &vardata);
5677 3503 [ + + + + ]: 141778 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3504 : : {
7148 3505 : 119079 : varinfos = add_unique_group_var(root, varinfos,
3506 : : groupexpr, &vardata);
3507 [ + + ]: 119079 : ReleaseVariableStats(vardata);
3508 : 119079 : continue;
3509 : : }
3510 [ - + ]: 22699 : ReleaseVariableStats(vardata);
3511 : :
3512 : : /*
3513 : : * Else pull out the component Vars. Handle PlaceHolderVars by
3514 : : * recursing into their arguments (effectively assuming that the
3515 : : * PlaceHolderVar doesn't change the number of groups, which boils
3516 : : * down to ignoring the possible addition of nulls to the result set).
3517 : : */
4660 3518 : 22699 : varshere = pull_var_clause(groupexpr,
3519 : : PVC_RECURSE_AGGREGATES |
3520 : : PVC_RECURSE_WINDOWFUNCS |
3521 : : PVC_RECURSE_PLACEHOLDERS);
3522 : :
3523 : : /*
3524 : : * If we find any variable-free GROUP BY item, then either it is a
3525 : : * constant (and we can ignore it) or it contains a volatile function;
3526 : : * in the latter case we punt and assume that each input row will
3527 : : * yield a distinct group.
3528 : : */
7817 3529 [ + + ]: 22699 : if (varshere == NIL)
3530 : : {
3531 [ + + ]: 282 : if (contain_volatile_functions(groupexpr))
3532 : 24 : return input_rows;
3533 : 258 : continue;
3534 : : }
3535 : :
3536 : : /*
3537 : : * Else add variables to varinfos list
3538 : : */
7148 3539 [ + - + + : 46351 : foreach(l2, varshere)
+ + ]
3540 : : {
3541 : 23934 : Node *var = (Node *) lfirst(l2);
3542 : :
3543 : 23934 : examine_variable(root, var, 0, &vardata);
3544 : 23934 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3545 [ + + ]: 23934 : ReleaseVariableStats(vardata);
3546 : : }
3547 : : }
3548 : :
3549 : : /*
3550 : : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3551 : : * list.
3552 : : */
3553 [ + + ]: 125490 : if (varinfos == NIL)
3554 : : {
3555 : : /* Apply SRF multiplier as we would do in the long path */
2332 3556 : 146 : numdistinct *= srf_multiplier;
3557 : : /* Round off */
3558 : 146 : numdistinct = ceil(numdistinct);
3559 : : /* Guard against out-of-range answers */
5760 3560 [ + + ]: 146 : if (numdistinct > input_rows)
5760 tgl@sss.pgh.pa.us 3561 :GBC 7 : numdistinct = input_rows;
2332 tgl@sss.pgh.pa.us 3562 [ - + ]:CBC 146 : if (numdistinct < 1.0)
2332 tgl@sss.pgh.pa.us 3563 :UBC 0 : numdistinct = 1.0;
5760 tgl@sss.pgh.pa.us 3564 :CBC 146 : return numdistinct;
3565 : : }
3566 : :
3567 : : /*
3568 : : * Group Vars by relation and estimate total numdistinct.
3569 : : *
3570 : : * For each iteration of the outer loop, we process the frontmost Var in
3571 : : * varinfos, plus all other Vars in the same relation. We remove these
3572 : : * Vars from the newvarinfos list for the next iteration. This is the
3573 : : * easiest way to group Vars of same rel together.
3574 : : */
3575 : : do
3576 : : {
7148 3577 : 126223 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3578 : 126223 : RelOptInfo *rel = varinfo1->rel;
2578 alvherre@alvh.no-ip. 3579 : 126223 : double reldistinct = 1;
7012 tgl@sss.pgh.pa.us 3580 : 126223 : double relmaxndistinct = reldistinct;
2575 alvherre@alvh.no-ip. 3581 : 126223 : int relvarcount = 0;
7559 bruce@momjian.us 3582 : 126223 : List *newvarinfos = NIL;
2578 alvherre@alvh.no-ip. 3583 : 126223 : List *relvarinfos = NIL;
3584 : :
3585 : : /*
3586 : : * Split the list of varinfos in two - one for the current rel, one
3587 : : * for remaining Vars on other rels.
3588 : : */
1733 tgl@sss.pgh.pa.us 3589 : 126223 : relvarinfos = lappend(relvarinfos, varinfo1);
1294 3590 [ + - + + : 143840 : for_each_from(l, varinfos, 1)
+ + ]
3591 : : {
7148 3592 : 17617 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3593 : :
3594 [ + + ]: 17617 : if (varinfo2->rel == varinfo1->rel)
3595 : : {
3596 : : /* varinfos on current rel */
1733 3597 : 16415 : relvarinfos = lappend(relvarinfos, varinfo2);
3598 : : }
3599 : : else
3600 : : {
3601 : : /* not time to process varinfo2 yet */
3602 : 1202 : newvarinfos = lappend(newvarinfos, varinfo2);
3603 : : }
3604 : : }
3605 : :
3606 : : /*
3607 : : * Get the numdistinct estimate for the Vars of this rel. We
3608 : : * iteratively search for multivariate n-distinct with maximum number
3609 : : * of vars; assuming that each var group is independent of the others,
3610 : : * we multiply them together. Any remaining relvarinfos after no more
3611 : : * multivariate matches are found are assumed independent too, so
3612 : : * their individual ndistinct estimates are multiplied also.
3613 : : *
3614 : : * While iterating, count how many separate numdistinct values we
3615 : : * apply. We apply a fudge factor below, but only if we multiplied
3616 : : * more than one such values.
3617 : : */
2578 alvherre@alvh.no-ip. 3618 [ + + ]: 252509 : while (relvarinfos)
3619 : : {
3620 : : double mvndistinct;
3621 : :
3622 [ + + ]: 126286 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3623 : : &mvndistinct))
3624 : : {
3625 : 201 : reldistinct *= mvndistinct;
3626 [ + + ]: 201 : if (relmaxndistinct < mvndistinct)
3627 : 195 : relmaxndistinct = mvndistinct;
2575 3628 : 201 : relvarcount++;
3629 : : }
3630 : : else
3631 : : {
2524 bruce@momjian.us 3632 [ + - + + : 268297 : foreach(l, relvarinfos)
+ + ]
3633 : : {
2578 alvherre@alvh.no-ip. 3634 : 142212 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3635 : :
3636 : 142212 : reldistinct *= varinfo2->ndistinct;
3637 [ + + ]: 142212 : if (relmaxndistinct < varinfo2->ndistinct)
3638 : 128632 : relmaxndistinct = varinfo2->ndistinct;
3639 : 142212 : relvarcount++;
3640 : :
3641 : : /*
3642 : : * When varinfo2's isdefault is set then we'd better set
3643 : : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3644 : : */
1111 drowley@postgresql.o 3645 [ + + + + ]: 142212 : if (estinfo != NULL && varinfo2->isdefault)
3646 : 5410 : estinfo->flags |= SELFLAG_USED_DEFAULT;
3647 : : }
3648 : :
3649 : : /* we're done with this relation */
2578 alvherre@alvh.no-ip. 3650 : 126085 : relvarinfos = NIL;
3651 : : }
3652 : : }
3653 : :
3654 : : /*
3655 : : * Sanity check --- don't divide by zero if empty relation.
3656 : : */
2568 rhaas@postgresql.org 3657 [ + + - + ]: 126223 : Assert(IS_SIMPLE_REL(rel));
7486 tgl@sss.pgh.pa.us 3658 [ + + ]: 126223 : if (rel->tuples > 0)
3659 : : {
3660 : : /*
3661 : : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
3662 : : * fudge factor is because the Vars are probably correlated but we
3663 : : * don't know by how much. We should never clamp to less than the
3664 : : * largest ndistinct value for any of the Vars, though, since
3665 : : * there will surely be at least that many groups.
3666 : : */
7016 3667 : 125530 : double clamp = rel->tuples;
3668 : :
3669 [ + + ]: 125530 : if (relvarcount > 1)
3670 : : {
3671 : 14944 : clamp *= 0.1;
7012 3672 [ + + ]: 14944 : if (clamp < relmaxndistinct)
3673 : : {
3674 : 14207 : clamp = relmaxndistinct;
3675 : : /* for sanity in case some ndistinct is too large: */
3676 [ + + ]: 14207 : if (clamp > rel->tuples)
3677 : 39 : clamp = rel->tuples;
3678 : : }
3679 : : }
7016 3680 [ + + ]: 125530 : if (reldistinct > clamp)
3681 : 14064 : reldistinct = clamp;
3682 : :
3683 : : /*
3684 : : * Update the estimate based on the restriction selectivity,
3685 : : * guarding against division by zero when reldistinct is zero.
3686 : : * Also skip this if we know that we are returning all rows.
3687 : : */
2932 dean.a.rasheed@gmail 3688 [ + - + + ]: 125530 : if (reldistinct > 0 && rel->rows < rel->tuples)
3689 : : {
3690 : : /*
3691 : : * Given a table containing N rows with n distinct values in a
3692 : : * uniform distribution, if we select p rows at random then
3693 : : * the expected number of distinct values selected is
3694 : : *
3695 : : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
3696 : : *
3697 : : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
3698 : : *
3699 : : * See "Approximating block accesses in database
3700 : : * organizations", S. B. Yao, Communications of the ACM,
3701 : : * Volume 20 Issue 4, April 1977 Pages 260-261.
3702 : : *
3703 : : * Alternatively, re-arranging the terms from the factorials,
3704 : : * this may be written as
3705 : : *
3706 : : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
3707 : : *
3708 : : * This form of the formula is more efficient to compute in
3709 : : * the common case where p is larger than N/n. Additionally,
3710 : : * as pointed out by Dell'Era, if i << N for all terms in the
3711 : : * product, it can be approximated by
3712 : : *
3713 : : * n * (1 - ((N-p)/N)^(N/n))
3714 : : *
3715 : : * See "Expected distinct values when selecting from a bag
3716 : : * without replacement", Alberto Dell'Era,
3717 : : * http://www.adellera.it/investigations/distinct_balls/.
3718 : : *
3719 : : * The condition i << N is equivalent to n >> 1, so this is a
3720 : : * good approximation when the number of distinct values in
3721 : : * the table is large. It turns out that this formula also
3722 : : * works well even when n is small.
3723 : : */
3724 : 37024 : reldistinct *=
3725 : 37024 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
3726 : 37024 : rel->tuples / reldistinct));
3727 : : }
3728 : 125530 : reldistinct = clamp_row_est(reldistinct);
3729 : :
3730 : : /*
3731 : : * Update estimate of total distinct groups.
3732 : : */
7486 tgl@sss.pgh.pa.us 3733 : 125530 : numdistinct *= reldistinct;
3734 : : }
3735 : :
7817 3736 : 126223 : varinfos = newvarinfos;
3737 [ + + ]: 126223 : } while (varinfos != NIL);
3738 : :
3739 : : /* Now we can account for the effects of any SRFs */
2332 3740 : 125344 : numdistinct *= srf_multiplier;
3741 : :
3742 : : /* Round off */
7748 3743 : 125344 : numdistinct = ceil(numdistinct);
3744 : :
3745 : : /* Guard against out-of-range answers */
7817 3746 [ + + ]: 125344 : if (numdistinct > input_rows)
3747 : 28687 : numdistinct = input_rows;
3748 [ - + ]: 125344 : if (numdistinct < 1.0)
7817 tgl@sss.pgh.pa.us 3749 :UBC 0 : numdistinct = 1.0;
3750 : :
7817 tgl@sss.pgh.pa.us 3751 :CBC 125344 : return numdistinct;
3752 : : }
3753 : :
3754 : : /*
3755 : : * Estimate hash bucket statistics when the specified expression is used
3756 : : * as a hash key for the given number of buckets.
3757 : : *
3758 : : * This attempts to determine two values:
3759 : : *
3760 : : * 1. The frequency of the most common value of the expression (returns
3761 : : * zero into *mcv_freq if we can't get that).
3762 : : *
3763 : : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
3764 : : * divided by total tuples in relation.
3765 : : *
3766 : : * XXX This is really pretty bogus since we're effectively assuming that the
3767 : : * distribution of hash keys will be the same after applying restriction
3768 : : * clauses as it was in the underlying relation. However, we are not nearly
3769 : : * smart enough to figure out how the restrict clauses might change the
3770 : : * distribution, so this will have to do for now.
3771 : : *
3772 : : * We are passed the number of buckets the executor will use for the given
3773 : : * input relation. If the data were perfectly distributed, with the same
3774 : : * number of tuples going into each available bucket, then the bucketsize
3775 : : * fraction would be 1/nbuckets. But this happy state of affairs will occur
3776 : : * only if (a) there are at least nbuckets distinct data values, and (b)
3777 : : * we have a not-too-skewed data distribution. Otherwise the buckets will
3778 : : * be nonuniformly occupied. If the other relation in the join has a key
3779 : : * distribution similar to this one's, then the most-loaded buckets are
3780 : : * exactly those that will be probed most often. Therefore, the "average"
3781 : : * bucket size for costing purposes should really be taken as something close
3782 : : * to the "worst case" bucket size. We try to estimate this by adjusting the
3783 : : * fraction if there are too few distinct data values, and then scaling up
3784 : : * by the ratio of the most common value's frequency to the average frequency.
3785 : : *
3786 : : * If no statistics are available, use a default estimate of 0.1. This will
3787 : : * discourage use of a hash rather strongly if the inner relation is large,
3788 : : * which is what we want. We do not want to hash unless we know that the
3789 : : * inner rel is well-dispersed (or the alternatives seem much worse).
3790 : : *
3791 : : * The caller should also check that the mcv_freq is not so large that the
3792 : : * most common value would by itself require an impractically large bucket.
3793 : : * In a hash join, the executor can split buckets if they get too big, but
3794 : : * obviously that doesn't help for a bucket that contains many duplicates of
3795 : : * the same value.
3796 : : */
3797 : : void
2434 3798 : 65928 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
3799 : : Selectivity *mcv_freq,
3800 : : Selectivity *bucketsize_frac)
3801 : : {
3802 : : VariableStatData vardata;
3803 : : double estfract,
3804 : : ndistinct,
3805 : : stanullfrac,
3806 : : avgfreq;
3807 : : bool isdefault;
3808 : : AttStatsSlot sslot;
3809 : :
7362 3810 : 65928 : examine_variable(root, hashkey, 0, &vardata);
3811 : :
3812 : : /* Look up the frequency of the most common value, if available */
2434 3813 : 65928 : *mcv_freq = 0.0;
3814 : :
3815 [ + + ]: 65928 : if (HeapTupleIsValid(vardata.statsTuple))
3816 : : {
3817 [ + + ]: 46097 : if (get_attstatsslot(&sslot, vardata.statsTuple,
3818 : : STATISTIC_KIND_MCV, InvalidOid,
3819 : : ATTSTATSSLOT_NUMBERS))
3820 : : {
3821 : : /*
3822 : : * The first MCV stat is for the most common value.
3823 : : */
3824 [ + - ]: 21308 : if (sslot.nnumbers > 0)
3825 : 21308 : *mcv_freq = sslot.numbers[0];
3826 : 21308 : free_attstatsslot(&sslot);
3827 : : }
3828 : : }
3829 : :
3830 : : /* Get number of distinct values */
4606 3831 : 65928 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
3832 : :
3833 : : /*
3834 : : * If ndistinct isn't real, punt. We normally return 0.1, but if the
3835 : : * mcv_freq is known to be even higher than that, use it instead.
3836 : : */
3837 [ + + ]: 65928 : if (isdefault)
3838 : : {
2434 3839 [ + - ]: 9644 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4606 3840 [ + + ]: 9644 : ReleaseVariableStats(vardata);
2434 3841 : 9644 : return;
3842 : : }
3843 : :
3844 : : /* Get fraction that are null */
7362 3845 [ + + ]: 56284 : if (HeapTupleIsValid(vardata.statsTuple))
3846 : : {
3847 : : Form_pg_statistic stats;
3848 : :
3849 : 46088 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
3850 : 46088 : stanullfrac = stats->stanullfrac;
3851 : : }
3852 : : else
3853 : 10196 : stanullfrac = 0.0;
3854 : :
3855 : : /* Compute avg freq of all distinct data values in raw relation */
3856 : 56284 : avgfreq = (1.0 - stanullfrac) / ndistinct;
3857 : :
3858 : : /*
3859 : : * Adjust ndistinct to account for restriction clauses. Observe we are
3860 : : * assuming that the data distribution is affected uniformly by the
3861 : : * restriction clauses!
3862 : : *
3863 : : * XXX Possibly better way, but much more expensive: multiply by
3864 : : * selectivity of rel's restriction clauses that mention the target Var.
3865 : : */
2940 3866 [ + - + + ]: 56284 : if (vardata.rel && vardata.rel->tuples > 0)
3867 : : {
7362 3868 : 56277 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
2940 3869 : 56277 : ndistinct = clamp_row_est(ndistinct);
3870 : : }
3871 : :
3872 : : /*
3873 : : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
3874 : : * number of buckets is less than the expected number of distinct values;
3875 : : * otherwise it is 1/ndistinct.
3876 : : */
6979 3877 [ + + ]: 56284 : if (ndistinct > nbuckets)
3878 : 48 : estfract = 1.0 / nbuckets;
3879 : : else
7362 3880 : 56236 : estfract = 1.0 / ndistinct;
3881 : :
3882 : : /*
3883 : : * Adjust estimated bucketsize upward to account for skewed distribution.
3884 : : */
2434 3885 [ + + + + ]: 56284 : if (avgfreq > 0.0 && *mcv_freq > avgfreq)
3886 : 19067 : estfract *= *mcv_freq / avgfreq;
3887 : :
3888 : : /*
3889 : : * Clamp bucketsize to sane range (the above adjustment could easily
3890 : : * produce an out-of-range result). We set the lower bound a little above
3891 : : * zero, since zero isn't a very sane result.
3892 : : */
7362 3893 [ - + ]: 56284 : if (estfract < 1.0e-6)
7362 tgl@sss.pgh.pa.us 3894 :UBC 0 : estfract = 1.0e-6;
7362 tgl@sss.pgh.pa.us 3895 [ + + ]:CBC 56284 : else if (estfract > 1.0)
3896 : 13493 : estfract = 1.0;
3897 : :
2434 3898 : 56284 : *bucketsize_frac = (Selectivity) estfract;
3899 : :
3900 [ + + ]: 56284 : ReleaseVariableStats(vardata);
3901 : : }
3902 : :
3903 : : /*
3904 : : * estimate_hashagg_tablesize
3905 : : * estimate the number of bytes that a hash aggregate hashtable will
3906 : : * require based on the agg_costs, path width and number of groups.
3907 : : *
3908 : : * We return the result as "double" to forestall any possible overflow
3909 : : * problem in the multiplication by dNumGroups.
3910 : : *
3911 : : * XXX this may be over-estimating the size now that hashagg knows to omit
3912 : : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
3913 : : * grouping columns not in the hashed set are counted here even though hashagg
3914 : : * won't store them. Is this a problem?
3915 : : */
3916 : : double
1237 heikki.linnakangas@i 3917 : 989 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
3918 : : const AggClauseCosts *agg_costs, double dNumGroups)
3919 : : {
3920 : : Size hashentrysize;
3921 : :
3922 : 989 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
3923 : 989 : path->pathtarget->width,
3924 : 989 : agg_costs->transitionSpace);
3925 : :
3926 : : /*
3927 : : * Note that this disregards the effect of fill-factor and growth policy
3928 : : * of the hash table. That's probably ok, given that the default
3929 : : * fill-factor is relatively high. It'd be hard to meaningfully factor in
3930 : : * "double-in-size" growth policies here.
3931 : : */
1879 tgl@sss.pgh.pa.us 3932 : 989 : return hashentrysize * dNumGroups;
3933 : : }
3934 : :
3935 : :
3936 : : /*-------------------------------------------------------------------------
3937 : : *
3938 : : * Support routines
3939 : : *
3940 : : *-------------------------------------------------------------------------
3941 : : */
3942 : :
3943 : : /*
3944 : : * Find applicable ndistinct statistics for the given list of VarInfos (which
3945 : : * must all belong to the given rel), and update *ndistinct to the estimate of
3946 : : * the MVNDistinctItem that best matches. If a match it found, *varinfos is
3947 : : * updated to remove the list of matched varinfos.
3948 : : *
3949 : : * Varinfos that aren't for simple Vars are ignored.
3950 : : *
3951 : : * Return true if we're able to find a match, false otherwise.
3952 : : */
3953 : : static bool
2578 alvherre@alvh.no-ip. 3954 : 126286 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
3955 : : List **varinfos, double *ndistinct)
3956 : : {
3957 : : ListCell *lc;
3958 : : int nmatches_vars;
3959 : : int nmatches_exprs;
3960 : 126286 : Oid statOid = InvalidOid;
3961 : : MVNDistinct *stats;
1115 tomas.vondra@postgre 3962 : 126286 : StatisticExtInfo *matched_info = NULL;
530 tgl@sss.pgh.pa.us 3963 [ + - ]: 126286 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
3964 : :
3965 : : /* bail out immediately if the table has no extended statistics */
2578 alvherre@alvh.no-ip. 3966 [ + + ]: 126286 : if (!rel->statlist)
3967 : 126025 : return false;
3968 : :
3969 : : /* look for the ndistinct statistics object matching the most vars */
1115 tomas.vondra@postgre 3970 : 261 : nmatches_vars = 0; /* we require at least two matches */
3971 : 261 : nmatches_exprs = 0;
2578 alvherre@alvh.no-ip. 3972 [ + - + + : 1056 : foreach(lc, rel->statlist)
+ + ]
3973 : : {
3974 : : ListCell *lc2;
3975 : 795 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
1115 tomas.vondra@postgre 3976 : 795 : int nshared_vars = 0;
3977 : 795 : int nshared_exprs = 0;
3978 : :
3979 : : /* skip statistics of other kinds */
2578 alvherre@alvh.no-ip. 3980 [ + + ]: 795 : if (info->kind != STATS_EXT_NDISTINCT)
3981 : 375 : continue;
3982 : :
3983 : : /* skip statistics with mismatching stxdinherit value */
530 tgl@sss.pgh.pa.us 3984 [ + + ]: 420 : if (info->inherit != rte->inh)
3985 : 12 : continue;
3986 : :
3987 : : /*
3988 : : * Determine how many expressions (and variables in non-matched
3989 : : * expressions) match. We'll then use these numbers to pick the
3990 : : * statistics object that best matches the clauses.
3991 : : */
1115 tomas.vondra@postgre 3992 [ + - + + : 1308 : foreach(lc2, *varinfos)
+ + ]
3993 : : {
3994 : : ListCell *lc3;
3995 : 900 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
3996 : : AttrNumber attnum;
3997 : :
3998 [ - + ]: 900 : Assert(varinfo->rel == rel);
3999 : :
4000 : : /* simple Var, search in statistics keys directly */
4001 [ + + ]: 900 : if (IsA(varinfo->var, Var))
4002 : : {
4003 : 717 : attnum = ((Var *) varinfo->var)->varattno;
4004 : :
4005 : : /*
4006 : : * Ignore system attributes - we don't support statistics on
4007 : : * them, so can't match them (and it'd fail as the values are
4008 : : * negative).
4009 : : */
4010 [ + + ]: 717 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4011 : 6 : continue;
4012 : :
4013 [ + + ]: 711 : if (bms_is_member(attnum, info->keys))
4014 : 408 : nshared_vars++;
4015 : :
4016 : 711 : continue;
4017 : : }
4018 : :
4019 : : /* expression - see if it's in the statistics object */
4020 [ + + + + : 330 : foreach(lc3, info->exprs)
+ + ]
4021 : : {
4022 : 264 : Node *expr = (Node *) lfirst(lc3);
4023 : :
4024 [ + + ]: 264 : if (equal(varinfo->var, expr))
4025 : : {
4026 : 117 : nshared_exprs++;
4027 : 117 : break;
4028 : : }
4029 : : }
4030 : : }
4031 : :
4032 [ + + ]: 408 : if (nshared_vars + nshared_exprs < 2)
4033 : 189 : continue;
4034 : :
4035 : : /*
4036 : : * Does this statistics object match more columns than the currently
4037 : : * best object? If so, use this one instead.
4038 : : *
4039 : : * XXX This should break ties using name of the object, or something
4040 : : * like that, to make the outcome stable.
4041 : : */
4042 [ + + + - ]: 219 : if ((nshared_exprs > nmatches_exprs) ||
4043 [ + + ]: 165 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4044 : : {
2578 alvherre@alvh.no-ip. 4045 : 207 : statOid = info->statOid;
1115 tomas.vondra@postgre 4046 : 207 : nmatches_vars = nshared_vars;
4047 : 207 : nmatches_exprs = nshared_exprs;
4048 : 207 : matched_info = info;
4049 : : }
4050 : : }
4051 : :
4052 : : /* No match? */
2578 alvherre@alvh.no-ip. 4053 [ + + ]: 261 : if (statOid == InvalidOid)
4054 : 60 : return false;
4055 : :
1115 tomas.vondra@postgre 4056 [ - + ]: 201 : Assert(nmatches_vars + nmatches_exprs > 1);
4057 : :
819 4058 : 201 : stats = statext_ndistinct_load(statOid, rte->inh);
4059 : :
4060 : : /*
4061 : : * If we have a match, search it for the specific item that matches (there
4062 : : * must be one), and construct the output values.
4063 : : */
2578 alvherre@alvh.no-ip. 4064 [ + - ]: 201 : if (stats)
4065 : : {
4066 : : int i;
2524 bruce@momjian.us 4067 : 201 : List *newlist = NIL;
2578 alvherre@alvh.no-ip. 4068 : 201 : MVNDistinctItem *item = NULL;
4069 : : ListCell *lc2;
1115 tomas.vondra@postgre 4070 : 201 : Bitmapset *matched = NULL;
4071 : : AttrNumber attnum_offset;
4072 : :
4073 : : /*
4074 : : * How much we need to offset the attnums? If there are no
4075 : : * expressions, no offset is needed. Otherwise offset enough to move
4076 : : * the lowest one (which is equal to number of expressions) to 1.
4077 : : */
4078 [ + + ]: 201 : if (matched_info->exprs)
4079 : 72 : attnum_offset = (list_length(matched_info->exprs) + 1);
4080 : : else
4081 : 129 : attnum_offset = 0;
4082 : :
4083 : : /* see what actually matched */
4084 [ + - + + : 705 : foreach(lc2, *varinfos)
+ + ]
4085 : : {
4086 : : ListCell *lc3;
4087 : : int idx;
4088 : 504 : bool found = false;
4089 : :
4090 : 504 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4091 : :
4092 : : /*
4093 : : * Process a simple Var expression, by matching it to keys
4094 : : * directly. If there's a matching expression, we'll try matching
4095 : : * it later.
4096 : : */
4097 [ + + ]: 504 : if (IsA(varinfo->var, Var))
4098 : : {
4099 : 411 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4100 : :
4101 : : /*
4102 : : * Ignore expressions on system attributes. Can't rely on the
4103 : : * bms check for negative values.
4104 : : */
4105 [ + + ]: 411 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4106 : 3 : continue;
4107 : :
4108 : : /* Is the variable covered by the statistics object? */
4109 [ + + ]: 408 : if (!bms_is_member(attnum, matched_info->keys))
4110 : 60 : continue;
4111 : :
4112 : 348 : attnum = attnum + attnum_offset;
4113 : :
4114 : : /* ensure sufficient offset */
4115 [ - + ]: 348 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4116 : :
4117 : 348 : matched = bms_add_member(matched, attnum);
4118 : :
4119 : 348 : found = true;
4120 : : }
4121 : :
4122 : : /*
4123 : : * XXX Maybe we should allow searching the expressions even if we
4124 : : * found an attribute matching the expression? That would handle
4125 : : * trivial expressions like "(a)" but it seems fairly useless.
4126 : : */
4127 [ + + ]: 441 : if (found)
4128 : 348 : continue;
4129 : :
4130 : : /* expression - see if it's in the statistics object */
4131 : 93 : idx = 0;
4132 [ + + + + : 153 : foreach(lc3, matched_info->exprs)
+ + ]
4133 : : {
4134 : 138 : Node *expr = (Node *) lfirst(lc3);
4135 : :
4136 [ + + ]: 138 : if (equal(varinfo->var, expr))
4137 : : {
4138 : 78 : AttrNumber attnum = -(idx + 1);
4139 : :
4140 : 78 : attnum = attnum + attnum_offset;
4141 : :
4142 : : /* ensure sufficient offset */
4143 [ - + ]: 78 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4144 : :
4145 : 78 : matched = bms_add_member(matched, attnum);
4146 : :
4147 : : /* there should be just one matching expression */
4148 : 78 : break;
4149 : : }
4150 : :
4151 : 60 : idx++;
4152 : : }
4153 : : }
4154 : :
4155 : : /* Find the specific item that exactly matches the combination */
2578 alvherre@alvh.no-ip. 4156 [ + - ]: 411 : for (i = 0; i < stats->nitems; i++)
4157 : : {
4158 : : int j;
4159 : 411 : MVNDistinctItem *tmpitem = &stats->items[i];
4160 : :
1115 tomas.vondra@postgre 4161 [ + + ]: 411 : if (tmpitem->nattributes != bms_num_members(matched))
4162 : 72 : continue;
4163 : :
4164 : : /* assume it's the right item */
4165 : 339 : item = tmpitem;
4166 : :
4167 : : /* check that all item attributes/expressions fit the match */
4168 [ + + ]: 807 : for (j = 0; j < tmpitem->nattributes; j++)
4169 : : {
4170 : 606 : AttrNumber attnum = tmpitem->attributes[j];
4171 : :
4172 : : /*
4173 : : * Thanks to how we constructed the matched bitmap above, we
4174 : : * can just offset all attnums the same way.
4175 : : */
4176 : 606 : attnum = attnum + attnum_offset;
4177 : :
4178 [ + + ]: 606 : if (!bms_is_member(attnum, matched))
4179 : : {
4180 : : /* nah, it's not this item */
4181 : 138 : item = NULL;
4182 : 138 : break;
4183 : : }
4184 : : }
4185 : :
4186 : : /*
4187 : : * If the item has all the matched attributes, we know it's the
4188 : : * right one - there can't be a better one. matching more.
4189 : : */
4190 [ + + ]: 339 : if (item)
4191 : 201 : break;
4192 : : }
4193 : :
4194 : : /*
4195 : : * Make sure we found an item. There has to be one, because ndistinct
4196 : : * statistics includes all combinations of attributes.
4197 : : */
2578 alvherre@alvh.no-ip. 4198 [ - + ]: 201 : if (!item)
2578 alvherre@alvh.no-ip. 4199 [ # # ]:UBC 0 : elog(ERROR, "corrupt MVNDistinct entry");
4200 : :
4201 : : /* Form the output varinfo list, keeping only unmatched ones */
2578 alvherre@alvh.no-ip. 4202 [ + - + + :CBC 705 : foreach(lc, *varinfos)
+ + ]
4203 : : {
4204 : 504 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4205 : : ListCell *lc3;
1115 tomas.vondra@postgre 4206 : 504 : bool found = false;
4207 : :
4208 : : /*
4209 : : * Let's look at plain variables first, because it's the most
4210 : : * common case and the check is quite cheap. We can simply get the
4211 : : * attnum and check (with an offset) matched bitmap.
4212 : : */
4213 [ + + ]: 504 : if (IsA(varinfo->var, Var))
2578 alvherre@alvh.no-ip. 4214 : 408 : {
1115 tomas.vondra@postgre 4215 : 411 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4216 : :
4217 : : /*
4218 : : * If it's a system attribute, we're done. We don't support
4219 : : * extended statistics on system attributes, so it's clearly
4220 : : * not matched. Just keep the expression and continue.
4221 : : */
4222 [ + + ]: 411 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4223 : : {
4224 : 3 : newlist = lappend(newlist, varinfo);
4225 : 3 : continue;
4226 : : }
4227 : :
4228 : : /* apply the same offset as above */
4229 : 408 : attnum += attnum_offset;
4230 : :
4231 : : /* if it's not matched, keep the varinfo */
4232 [ + + ]: 408 : if (!bms_is_member(attnum, matched))
4233 : 60 : newlist = lappend(newlist, varinfo);
4234 : :
4235 : : /* The rest of the loop deals with complex expressions. */
2578 alvherre@alvh.no-ip. 4236 : 408 : continue;
4237 : : }
4238 : :
4239 : : /*
4240 : : * Process complex expressions, not just simple Vars.
4241 : : *
4242 : : * First, we search for an exact match of an expression. If we
4243 : : * find one, we can just discard the whole GroupVarInfo, with all
4244 : : * the variables we extracted from it.
4245 : : *
4246 : : * Otherwise we inspect the individual vars, and try matching it
4247 : : * to variables in the item.
4248 : : */
1115 tomas.vondra@postgre 4249 [ + + + + : 153 : foreach(lc3, matched_info->exprs)
+ + ]
4250 : : {
4251 : 138 : Node *expr = (Node *) lfirst(lc3);
4252 : :
4253 [ + + ]: 138 : if (equal(varinfo->var, expr))
4254 : : {
4255 : 78 : found = true;
4256 : 78 : break;
4257 : : }
4258 : : }
4259 : :
4260 : : /* found exact match, skip */
4261 [ + + ]: 93 : if (found)
1611 4262 : 78 : continue;
4263 : :
1115 4264 : 15 : newlist = lappend(newlist, varinfo);
4265 : : }
4266 : :
2578 alvherre@alvh.no-ip. 4267 : 201 : *varinfos = newlist;
4268 : 201 : *ndistinct = item->ndistinct;
4269 : 201 : return true;
4270 : : }
4271 : :
2578 alvherre@alvh.no-ip. 4272 :UBC 0 : return false;
4273 : : }
4274 : :
4275 : : /*
4276 : : * convert_to_scalar
4277 : : * Convert non-NULL values of the indicated types to the comparison
4278 : : * scale needed by scalarineqsel().
4279 : : * Returns "true" if successful.
4280 : : *
4281 : : * XXX this routine is a hack: ideally we should look up the conversion
4282 : : * subroutines in pg_type.
4283 : : *
4284 : : * All numeric datatypes are simply converted to their equivalent
4285 : : * "double" values. (NUMERIC values that are outside the range of "double"
4286 : : * are clamped to +/- HUGE_VAL.)
4287 : : *
4288 : : * String datatypes are converted by convert_string_to_scalar(),
4289 : : * which is explained below. The reason why this routine deals with
4290 : : * three values at a time, not just one, is that we need it for strings.
4291 : : *
4292 : : * The bytea datatype is just enough different from strings that it has
4293 : : * to be treated separately.
4294 : : *
4295 : : * The several datatypes representing absolute times are all converted
4296 : : * to Timestamp, which is actually an int64, and then we promote that to
4297 : : * a double. Note this will give correct results even for the "special"
4298 : : * values of Timestamp, since those are chosen to compare correctly;
4299 : : * see timestamp_cmp.
4300 : : *
4301 : : * The several datatypes representing relative times (intervals) are all
4302 : : * converted to measurements expressed in seconds.
4303 : : */
4304 : : static bool
1948 tgl@sss.pgh.pa.us 4305 :CBC 41749 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4306 : : Datum lobound, Datum hibound, Oid boundstypid,
4307 : : double *scaledlobound, double *scaledhibound)
4308 : : {
2234 4309 : 41749 : bool failure = false;
4310 : :
4311 : : /*
4312 : : * Both the valuetypid and the boundstypid should exactly match the
4313 : : * declared input type(s) of the operator we are invoked for. However,
4314 : : * extensions might try to use scalarineqsel as estimator for operators
4315 : : * with input type(s) we don't handle here; in such cases, we want to
4316 : : * return false, not fail. In any case, we mustn't assume that valuetypid
4317 : : * and boundstypid are identical.
4318 : : *
4319 : : * XXX The histogram we are interpolating between points of could belong
4320 : : * to a column that's only binary-compatible with the declared type. In
4321 : : * essence we are assuming that the semantics of binary-compatible types
4322 : : * are enough alike that we can use a histogram generated with one type's
4323 : : * operators to estimate selectivity for the other's. This is outright
4324 : : * wrong in some cases --- in particular signed versus unsigned
4325 : : * interpretation could trip us up. But it's useful enough in the
4326 : : * majority of cases that we do it anyway. Should think about more
4327 : : * rigorous ways to do it.
4328 : : */
8764 4329 [ + + - - : 41749 : switch (valuetypid)
- - ]
4330 : : {
4331 : : /*
4332 : : * Built-in numeric types
4333 : : */
8365 4334 : 39350 : case BOOLOID:
4335 : : case INT2OID:
4336 : : case INT4OID:
4337 : : case INT8OID:
4338 : : case FLOAT4OID:
4339 : : case FLOAT8OID:
4340 : : case NUMERICOID:
4341 : : case OIDOID:
4342 : : case REGPROCOID:
4343 : : case REGPROCEDUREOID:
4344 : : case REGOPEROID:
4345 : : case REGOPERATOROID:
4346 : : case REGCLASSOID:
4347 : : case REGTYPEOID:
4348 : : case REGCOLLATIONOID:
4349 : : case REGCONFIGOID:
4350 : : case REGDICTIONARYOID:
4351 : : case REGROLEOID:
4352 : : case REGNAMESPACEOID:
2234 4353 : 39350 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4354 : : &failure);
4355 : 39350 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4356 : : &failure);
4357 : 39350 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4358 : : &failure);
4359 : 39350 : return !failure;
4360 : :
4361 : : /*
4362 : : * Built-in string types
4363 : : */
8847 4364 : 2399 : case CHAROID:
4365 : : case BPCHAROID:
4366 : : case VARCHAROID:
4367 : : case TEXTOID:
4368 : : case NAMEOID:
4369 : : {
2234 4370 : 2399 : char *valstr = convert_string_datum(value, valuetypid,
4371 : : collid, &failure);
4372 : 2399 : char *lostr = convert_string_datum(lobound, boundstypid,
4373 : : collid, &failure);
4374 : 2399 : char *histr = convert_string_datum(hibound, boundstypid,
4375 : : collid, &failure);
4376 : :
4377 : : /*
4378 : : * Bail out if any of the values is not of string type. We
4379 : : * might leak converted strings for the other value(s), but
4380 : : * that's not worth troubling over.
4381 : : */
4382 [ - + ]: 2399 : if (failure)
2234 tgl@sss.pgh.pa.us 4383 :UBC 0 : return false;
4384 : :
8424 bruce@momjian.us 4385 :CBC 2399 : convert_string_to_scalar(valstr, scaledvalue,
4386 : : lostr, scaledlobound,
4387 : : histr, scaledhibound);
4388 : 2399 : pfree(valstr);
4389 : 2399 : pfree(lostr);
4390 : 2399 : pfree(histr);
4391 : 2399 : return true;
4392 : : }
4393 : :
4394 : : /*
4395 : : * Built-in bytea type
4396 : : */
8280 tgl@sss.pgh.pa.us 4397 :UBC 0 : case BYTEAOID:
4398 : : {
4399 : : /* We only support bytea vs bytea comparison */
2234 4400 [ # # ]: 0 : if (boundstypid != BYTEAOID)
4401 : 0 : return false;
8280 4402 : 0 : convert_bytea_to_scalar(value, scaledvalue,
4403 : : lobound, scaledlobound,
4404 : : hibound, scaledhibound);
4405 : 0 : return true;
4406 : : }
4407 : :
4408 : : /*
4409 : : * Built-in time types
4410 : : */
8814 4411 : 0 : case TIMESTAMPOID:
4412 : : case TIMESTAMPTZOID:
4413 : : case DATEOID:
4414 : : case INTERVALOID:
4415 : : case TIMEOID:
4416 : : case TIMETZOID:
2234 4417 : 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
4418 : : &failure);
4419 : 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
4420 : : &failure);
4421 : 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
4422 : : &failure);
4423 : 0 : return !failure;
4424 : :
4425 : : /*
4426 : : * Built-in network types
4427 : : */
8345 4428 : 0 : case INETOID:
4429 : : case CIDROID:
4430 : : case MACADDROID:
4431 : : case MACADDR8OID:
2234 4432 : 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
4433 : : &failure);
4434 : 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
4435 : : &failure);
4436 : 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
4437 : : &failure);
4438 : 0 : return !failure;
4439 : : }
4440 : : /* Don't know how to convert */
6777 4441 : 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
9022 4442 : 0 : return false;
4443 : : }
4444 : :
4445 : : /*
4446 : : * Do convert_to_scalar()'s work for any numeric data type.
4447 : : *
4448 : : * On failure (e.g., unsupported typid), set *failure to true;
4449 : : * otherwise, that variable is not changed.
4450 : : */
4451 : : static double
2234 tgl@sss.pgh.pa.us 4452 :CBC 118050 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
4453 : : {
8764 4454 [ - + + - : 118050 : switch (typid)
- + - +
- ]
4455 : : {
8378 tgl@sss.pgh.pa.us 4456 :UBC 0 : case BOOLOID:
8722 4457 : 0 : return (double) DatumGetBool(value);
8764 tgl@sss.pgh.pa.us 4458 :CBC 6 : case INT2OID:
4459 : 6 : return (double) DatumGetInt16(value);
4460 : 14619 : case INT4OID:
4461 : 14619 : return (double) DatumGetInt32(value);
8764 tgl@sss.pgh.pa.us 4462 :UBC 0 : case INT8OID:
8722 4463 : 0 : return (double) DatumGetInt64(value);
8764 4464 : 0 : case FLOAT4OID:
8722 4465 : 0 : return (double) DatumGetFloat4(value);
8764 tgl@sss.pgh.pa.us 4466 :CBC 18 : case FLOAT8OID:
8722 4467 : 18 : return (double) DatumGetFloat8(value);
8764 tgl@sss.pgh.pa.us 4468 :UBC 0 : case NUMERICOID:
4469 : : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
8219 4470 : 0 : return (double)
4471 : 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
4472 : : value));
8764 tgl@sss.pgh.pa.us 4473 :CBC 103407 : case OIDOID:
4474 : : case REGPROCOID:
4475 : : case REGPROCEDUREOID:
4476 : : case REGOPEROID:
4477 : : case REGOPERATOROID:
4478 : : case REGCLASSOID:
4479 : : case REGTYPEOID:
4480 : : case REGCOLLATIONOID:
4481 : : case REGCONFIGOID:
4482 : : case REGDICTIONARYOID:
4483 : : case REGROLEOID:
4484 : : case REGNAMESPACEOID:
4485 : : /* we can treat OIDs as integers... */
4486 : 103407 : return (double) DatumGetObjectId(value);
4487 : : }
4488 : :
2234 tgl@sss.pgh.pa.us 4489 :UBC 0 : *failure = true;
8764 4490 : 0 : return 0;
4491 : : }
4492 : :
4493 : : /*
4494 : : * Do convert_to_scalar()'s work for any character-string data type.
4495 : : *
4496 : : * String datatypes are converted to a scale that ranges from 0 to 1,
4497 : : * where we visualize the bytes of the string as fractional digits.
4498 : : *
4499 : : * We do not want the base to be 256, however, since that tends to
4500 : : * generate inflated selectivity estimates; few databases will have
4501 : : * occurrences of all 256 possible byte values at each position.
4502 : : * Instead, use the smallest and largest byte values seen in the bounds
4503 : : * as the estimated range for each byte, after some fudging to deal with
4504 : : * the fact that we probably aren't going to see the full range that way.
4505 : : *
4506 : : * An additional refinement is that we discard any common prefix of the
4507 : : * three strings before computing the scaled values. This allows us to
4508 : : * "zoom in" when we encounter a narrow data range. An example is a phone
4509 : : * number database where all the values begin with the same area code.
4510 : : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
4511 : : * so this is more likely to happen than you might think.)
4512 : : */
4513 : : static void
6777 tgl@sss.pgh.pa.us 4514 :CBC 2399 : convert_string_to_scalar(char *value,
4515 : : double *scaledvalue,
4516 : : char *lobound,
4517 : : double *scaledlobound,
4518 : : char *hibound,
4519 : : double *scaledhibound)
4520 : : {
4521 : : int rangelo,
4522 : : rangehi;
4523 : : char *sptr;
4524 : :
4525 : 2399 : rangelo = rangehi = (unsigned char) hibound[0];
8764 4526 [ + + ]: 30083 : for (sptr = lobound; *sptr; sptr++)
4527 : : {
6777 4528 [ + + ]: 27684 : if (rangelo > (unsigned char) *sptr)
4529 : 5964 : rangelo = (unsigned char) *sptr;
4530 [ + + ]: 27684 : if (rangehi < (unsigned char) *sptr)
4531 : 3103 : rangehi = (unsigned char) *sptr;
4532 : : }
8764 4533 [ + + ]: 29892 : for (sptr = hibound; *sptr; sptr++)
4534 : : {
6777 4535 [ + + ]: 27493 : if (rangelo > (unsigned char) *sptr)
4536 : 538 : rangelo = (unsigned char) *sptr;
4537 [ + + ]: 27493 : if (rangehi < (unsigned char) *sptr)
4538 : 1328 : rangehi = (unsigned char) *sptr;
4539 : : }
4540 : : /* If range includes any upper-case ASCII chars, make it include all */
8764 4541 [ + + + + ]: 2399 : if (rangelo <= 'Z' && rangehi >= 'A')
4542 : : {
4543 [ + + ]: 685 : if (rangelo > 'A')
4544 : 45 : rangelo = 'A';
4545 [ + + ]: 685 : if (rangehi < 'Z')
4546 : 240 : rangehi = 'Z';
4547 : : }
4548 : : /* Ditto lower-case */
4549 [ + - + + ]: 2399 : if (rangelo <= 'z' && rangehi >= 'a')
4550 : : {
4551 [ + + ]: 2148 : if (rangelo > 'a')
4552 : 9 : rangelo = 'a';
4553 [ + + ]: 2148 : if (rangehi < 'z')
4554 : 2105 : rangehi = 'z';
4555 : : }
4556 : : /* Ditto digits */
4557 [ + + + - ]: 2399 : if (rangelo <= '9' && rangehi >= '0')
4558 : : {
4559 [ + + ]: 411 : if (rangelo > '0')
4560 : 364 : rangelo = '0';
4561 [ + + ]: 411 : if (rangehi < '9')
4562 : 7 : rangehi = '9';
4563 : : }
4564 : :
4565 : : /*
4566 : : * If range includes less than 10 chars, assume we have not got enough
4567 : : * data, and make it include regular ASCII set.
4568 : : */
4569 [ - + ]: 2399 : if (rangehi - rangelo < 9)
4570 : : {
8764 tgl@sss.pgh.pa.us 4571 :UBC 0 : rangelo = ' ';
4572 : 0 : rangehi = 127;
4573 : : }
4574 : :
4575 : : /*
4576 : : * Now strip any common prefix of the three strings.
4577 : : */
8764 tgl@sss.pgh.pa.us 4578 [ + - ]:CBC 4378 : while (*lobound)
4579 : : {
4580 [ + + + - ]: 4378 : if (*lobound != *hibound || *lobound != *value)
4581 : : break;
4582 : 1979 : lobound++, hibound++, value++;
4583 : : }
4584 : :
4585 : : /*
4586 : : * Now we can do the conversions.
4587 : : */
4588 : 2399 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
4589 : 2399 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
4590 : 2399 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
4591 : 2399 : }
4592 : :
4593 : : static double
6777 4594 : 7197 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
4595 : : {
4596 : 7197 : int slen = strlen(value);
4597 : : double num,
4598 : : denom,
4599 : : base;
4600 : :
8764 4601 [ - + ]: 7197 : if (slen <= 0)
8764 tgl@sss.pgh.pa.us 4602 :LBC (13) : return 0.0; /* empty string has scalar value 0 */
4603 : :
4604 : : /*
4605 : : * There seems little point in considering more than a dozen bytes from
4606 : : * the string. Since base is at least 10, that will give us nominal
4607 : : * resolution of at least 12 decimal digits, which is surely far more
4608 : : * precision than this estimation technique has got anyway (especially in
4609 : : * non-C locales). Also, even with the maximum possible base of 256, this
4610 : : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
4611 : : * overflow on any known machine.
4612 : : */
3157 tgl@sss.pgh.pa.us 4613 [ + + ]:CBC 7197 : if (slen > 12)
4614 : 1900 : slen = 12;
4615 : :
4616 : : /* Convert initial characters to fraction */
8764 4617 : 7197 : base = rangehi - rangelo + 1;
4618 : 7197 : num = 0.0;
4619 : 7197 : denom = base;
4620 [ + + ]: 60171 : while (slen-- > 0)
4621 : : {
6777 4622 : 52974 : int ch = (unsigned char) *value++;
4623 : :
8764 4624 [ + + ]: 52974 : if (ch < rangelo)
8424 bruce@momjian.us 4625 : 68 : ch = rangelo - 1;
8764 tgl@sss.pgh.pa.us 4626 [ - + ]: 52906 : else if (ch > rangehi)
8424 bruce@momjian.us 4627 :UBC 0 : ch = rangehi + 1;
8764 tgl@sss.pgh.pa.us 4628 :CBC 52974 : num += ((double) (ch - rangelo)) / denom;
4629 : 52974 : denom *= base;
4630 : : }
4631 : :
4632 : 7197 : return num;
4633 : : }
4634 : :
4635 : : /*
4636 : : * Convert a string-type Datum into a palloc'd, null-terminated string.
4637 : : *
4638 : : * On failure (e.g., unsupported typid), set *failure to true;
4639 : : * otherwise, that variable is not changed. (We'll return NULL on failure.)
4640 : : *
4641 : : * When using a non-C locale, we must pass the string through strxfrm()
4642 : : * before continuing, so as to generate correct locale-specific results.
4643 : : */
4644 : : static char *
1948 4645 : 7197 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
4646 : : {
4647 : : char *val;
4648 : :
8764 4649 [ - + + - ]: 7197 : switch (typid)
4650 : : {
8764 tgl@sss.pgh.pa.us 4651 :UBC 0 : case CHAROID:
4652 : 0 : val = (char *) palloc(2);
4653 : 0 : val[0] = DatumGetChar(value);
4654 : 0 : val[1] = '\0';
4655 : 0 : break;
8764 tgl@sss.pgh.pa.us 4656 :CBC 2197 : case BPCHAROID:
4657 : : case VARCHAROID:
4658 : : case TEXTOID:
5864 4659 : 2197 : val = TextDatumGetCString(value);
4660 : 2197 : break;
8764 4661 : 5000 : case NAMEOID:
4662 : : {
8424 bruce@momjian.us 4663 : 5000 : NameData *nm = (NameData *) DatumGetPointer(value);
4664 : :
4665 : 5000 : val = pstrdup(NameStr(*nm));
4666 : 5000 : break;
4667 : : }
8764 tgl@sss.pgh.pa.us 4668 :UBC 0 : default:
2234 4669 : 0 : *failure = true;
8764 4670 : 0 : return NULL;
4671 : : }
4672 : :
1948 tgl@sss.pgh.pa.us 4673 [ + + ]:CBC 7197 : if (!lc_collate_is_c(collid))
4674 : : {
4675 : : char *xfrmstr;
4676 : : size_t xfrmlen;
4677 : : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
4678 : :
4679 : : /*
4680 : : * XXX: We could guess at a suitable output buffer size and only call
4681 : : * strxfrm twice if our guess is too small.
4682 : : *
4683 : : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
4684 : : * bogus data or set an error. This is not really a problem unless it
4685 : : * crashes since it will only give an estimation error and nothing
4686 : : * fatal.
4687 : : */
6402 bruce@momjian.us 4688 : 36 : xfrmlen = strxfrm(NULL, val, 0);
4689 : : #ifdef WIN32
4690 : :
4691 : : /*
4692 : : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
4693 : : * of trying to allocate this much memory (and fail), just return the
4694 : : * original string unmodified as if we were in the C locale.
4695 : : */
4696 : : if (xfrmlen == INT_MAX)
4697 : : return val;
4698 : : #endif
7577 tgl@sss.pgh.pa.us 4699 : 36 : xfrmstr = (char *) palloc(xfrmlen + 1);
4700 : 36 : xfrmlen2 = strxfrm(xfrmstr, val, xfrmlen + 1);
4701 : :
4702 : : /*
4703 : : * Some systems (e.g., glibc) can return a smaller value from the
4704 : : * second call than the first; thus the Assert must be <= not ==.
4705 : : */
4706 [ - + ]: 36 : Assert(xfrmlen2 <= xfrmlen);
8047 peter_e@gmx.net 4707 : 36 : pfree(val);
4708 : 36 : val = xfrmstr;
4709 : : }
4710 : :
6777 tgl@sss.pgh.pa.us 4711 : 7197 : return val;
4712 : : }
4713 : :
4714 : : /*
4715 : : * Do convert_to_scalar()'s work for any bytea data type.
4716 : : *
4717 : : * Very similar to convert_string_to_scalar except we can't assume
4718 : : * null-termination and therefore pass explicit lengths around.
4719 : : *
4720 : : * Also, assumptions about likely "normal" ranges of characters have been
4721 : : * removed - a data range of 0..255 is always used, for now. (Perhaps
4722 : : * someday we will add information about actual byte data range to
4723 : : * pg_statistic.)
4724 : : */
4725 : : static void
8280 tgl@sss.pgh.pa.us 4726 :UBC 0 : convert_bytea_to_scalar(Datum value,
4727 : : double *scaledvalue,
4728 : : Datum lobound,
4729 : : double *scaledlobound,
4730 : : Datum hibound,
4731 : : double *scaledhibound)
4732 : : {
2234 4733 : 0 : bytea *valuep = DatumGetByteaPP(value);
4734 : 0 : bytea *loboundp = DatumGetByteaPP(lobound);
4735 : 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
4736 : : int rangelo,
4737 : : rangehi,
4738 [ # # # # : 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
# # # # #
# ]
4739 [ # # # # : 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
# # # # #
# ]
4740 [ # # # # : 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
# # # # #
# ]
4741 : : i,
4742 : : minlen;
4743 [ # # ]: 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
4744 [ # # ]: 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
4745 [ # # ]: 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
4746 : :
4747 : : /*
4748 : : * Assume bytea data is uniformly distributed across all byte values.
4749 : : */
8280 4750 : 0 : rangelo = 0;
4751 : 0 : rangehi = 255;
4752 : :
4753 : : /*
4754 : : * Now strip any common prefix of the three strings.
4755 : : */
4756 : 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
4757 [ # # ]: 0 : for (i = 0; i < minlen; i++)
4758 : : {
4759 [ # # # # ]: 0 : if (*lostr != *histr || *lostr != *valstr)
4760 : : break;
4761 : 0 : lostr++, histr++, valstr++;
4762 : 0 : loboundlen--, hiboundlen--, valuelen--;
4763 : : }
4764 : :
4765 : : /*
4766 : : * Now we can do the conversions.
4767 : : */
4768 : 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
4769 : 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
4770 : 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
4771 : 0 : }
4772 : :
4773 : : static double
4774 : 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
4775 : : int rangelo, int rangehi)
4776 : : {
4777 : : double num,
4778 : : denom,
4779 : : base;
4780 : :
4781 [ # # ]: 0 : if (valuelen <= 0)
4782 : 0 : return 0.0; /* empty string has scalar value 0 */
4783 : :
4784 : : /*
4785 : : * Since base is 256, need not consider more than about 10 chars (even
4786 : : * this many seems like overkill)
4787 : : */
4788 [ # # ]: 0 : if (valuelen > 10)
4789 : 0 : valuelen = 10;
4790 : :
4791 : : /* Convert initial characters to fraction */
4792 : 0 : base = rangehi - rangelo + 1;
4793 : 0 : num = 0.0;
4794 : 0 : denom = base;
4795 [ # # ]: 0 : while (valuelen-- > 0)
4796 : : {
4797 : 0 : int ch = *value++;
4798 : :
4799 [ # # ]: 0 : if (ch < rangelo)
4800 : 0 : ch = rangelo - 1;
4801 [ # # ]: 0 : else if (ch > rangehi)
4802 : 0 : ch = rangehi + 1;
4803 : 0 : num += ((double) (ch - rangelo)) / denom;
4804 : 0 : denom *= base;
4805 : : }
4806 : :
4807 : 0 : return num;
4808 : : }
4809 : :
4810 : : /*
4811 : : * Do convert_to_scalar()'s work for any timevalue data type.
4812 : : *
4813 : : * On failure (e.g., unsupported typid), set *failure to true;
4814 : : * otherwise, that variable is not changed.
4815 : : */
4816 : : static double
2234 4817 : 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
4818 : : {
8764 4819 [ # # # # : 0 : switch (typid)
# # # ]
4820 : : {
8365 4821 : 0 : case TIMESTAMPOID:
8710 4822 : 0 : return DatumGetTimestamp(value);
8229 4823 : 0 : case TIMESTAMPTZOID:
4824 : 0 : return DatumGetTimestampTz(value);
8764 4825 : 0 : case DATEOID:
4856 4826 : 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
8764 4827 : 0 : case INTERVALOID:
4828 : : {
8424 bruce@momjian.us 4829 : 0 : Interval *interval = DatumGetIntervalP(value);
4830 : :
4831 : : /*
4832 : : * Convert the month part of Interval to days using assumed
4833 : : * average month length of 365.25/12.0 days. Not too
4834 : : * accurate, but plenty good enough for our purposes.
4835 : : *
4836 : : * This also works for infinite intervals, which just have all
4837 : : * fields set to INT_MIN/INT_MAX, and so will produce a result
4838 : : * smaller/larger than any finite interval.
4839 : : */
6756 4840 : 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
4841 : 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
4842 : : }
8764 tgl@sss.pgh.pa.us 4843 : 0 : case TIMEOID:
8710 4844 : 0 : return DatumGetTimeADT(value);
8345 4845 : 0 : case TIMETZOID:
4846 : : {
4847 : 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
4848 : :
4849 : : /* use GMT-equivalent time */
8029 lockhart@fourpalms.o 4850 : 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
4851 : : }
4852 : : }
4853 : :
2234 tgl@sss.pgh.pa.us 4854 : 0 : *failure = true;
8764 4855 : 0 : return 0;
4856 : : }
4857 : :
4858 : :
4859 : : /*
4860 : : * get_restriction_variable
4861 : : * Examine the args of a restriction clause to see if it's of the
4862 : : * form (variable op pseudoconstant) or (pseudoconstant op variable),
4863 : : * where "variable" could be either a Var or an expression in vars of a
4864 : : * single relation. If so, extract information about the variable,
4865 : : * and also indicate which side it was on and the other argument.
4866 : : *
4867 : : * Inputs:
4868 : : * root: the planner info
4869 : : * args: clause argument list
4870 : : * varRelid: see specs for restriction selectivity functions
4871 : : *
4872 : : * Outputs: (these are valid only if true is returned)
4873 : : * *vardata: gets information about variable (see examine_variable)
4874 : : * *other: gets other clause argument, aggressively reduced to a constant
4875 : : * *varonleft: set true if variable is on the left, false if on the right
4876 : : *
4877 : : * Returns true if a variable is identified, otherwise false.
4878 : : *
4879 : : * Note: if there are Vars on both sides of the clause, we must fail, because
4880 : : * callers are expecting that the other side will act like a pseudoconstant.
4881 : : */
4882 : : bool
6888 tgl@sss.pgh.pa.us 4883 :CBC 310557 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
4884 : : VariableStatData *vardata, Node **other,
4885 : : bool *varonleft)
4886 : : {
4887 : : Node *left,
4888 : : *right;
4889 : : VariableStatData rdata;
4890 : :
4891 : : /* Fail if not a binary opclause (probably shouldn't happen) */
7259 neilc@samurai.com 4892 [ - + ]: 310557 : if (list_length(args) != 2)
7362 tgl@sss.pgh.pa.us 4893 :UBC 0 : return false;
4894 : :
7263 neilc@samurai.com 4895 :CBC 310557 : left = (Node *) linitial(args);
7362 tgl@sss.pgh.pa.us 4896 : 310557 : right = (Node *) lsecond(args);
4897 : :
4898 : : /*
4899 : : * Examine both sides. Note that when varRelid is nonzero, Vars of other
4900 : : * relations will be treated as pseudoconstants.
4901 : : */
4902 : 310557 : examine_variable(root, left, varRelid, vardata);
4903 : 310557 : examine_variable(root, right, varRelid, &rdata);
4904 : :
4905 : : /*
4906 : : * If one side is a variable and the other not, we win.
4907 : : */
4908 [ + + + + ]: 310557 : if (vardata->rel && rdata.rel == NULL)
4909 : : {
4910 : 276664 : *varonleft = true;
6264 4911 : 276664 : *other = estimate_expression_value(root, rdata.var);
4912 : : /* Assume we need no ReleaseVariableStats(rdata) here */
7362 4913 : 276664 : return true;
4914 : : }
4915 : :
4916 [ + + + + ]: 33893 : if (vardata->rel == NULL && rdata.rel)
4917 : : {
4918 : 31469 : *varonleft = false;
6264 4919 : 31469 : *other = estimate_expression_value(root, vardata->var);
4920 : : /* Assume we need no ReleaseVariableStats(*vardata) here */
7362 4921 : 31469 : *vardata = rdata;
4922 : 31469 : return true;
4923 : : }
4924 : :
4925 : : /* Oops, clause has wrong structure (probably var op var) */
4926 [ + + ]: 2424 : ReleaseVariableStats(*vardata);
4927 [ + + ]: 2424 : ReleaseVariableStats(rdata);
4928 : :
4929 : 2424 : return false;
4930 : : }
4931 : :
4932 : : /*
4933 : : * get_join_variables
4934 : : * Apply examine_variable() to each side of a join clause.
4935 : : * Also, attempt to identify whether the join clause has the same
4936 : : * or reversed sense compared to the SpecialJoinInfo.
4937 : : *
4938 : : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
4939 : : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
4940 : : * where we can't tell for sure, we default to assuming it's normal.
4941 : : */
4942 : : void
5720 4943 : 96074 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
4944 : : VariableStatData *vardata1, VariableStatData *vardata2,
4945 : : bool *join_is_reversed)
4946 : : {
4947 : : Node *left,
4948 : : *right;
4949 : :
7259 neilc@samurai.com 4950 [ - + ]: 96074 : if (list_length(args) != 2)
7362 tgl@sss.pgh.pa.us 4951 [ # # ]:UBC 0 : elog(ERROR, "join operator should take two arguments");
4952 : :
7263 neilc@samurai.com 4953 :CBC 96074 : left = (Node *) linitial(args);
8365 tgl@sss.pgh.pa.us 4954 : 96074 : right = (Node *) lsecond(args);
4955 : :
7362 4956 : 96074 : examine_variable(root, left, 0, vardata1);
4957 : 96074 : examine_variable(root, right, 0, vardata2);
4958 : :
5720 4959 [ + + + + ]: 192063 : if (vardata1->rel &&
4960 : 95989 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
2489 4961 : 30853 : *join_is_reversed = true; /* var1 is on RHS */
5720 4962 [ + + + + ]: 130357 : else if (vardata2->rel &&
4963 : 65136 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
2489 4964 : 70 : *join_is_reversed = true; /* var2 is on LHS */
4965 : : else
5720 4966 : 65151 : *join_is_reversed = false;
7362 4967 : 96074 : }
4968 : :
4969 : : /* statext_expressions_load copies the tuple, so just pfree it. */
4970 : : static void
1115 tomas.vondra@postgre 4971 : 822 : ReleaseDummy(HeapTuple tuple)
4972 : : {
4973 : 822 : pfree(tuple);
4974 : 822 : }
4975 : :
4976 : : /*
4977 : : * examine_variable
4978 : : * Try to look up statistical data about an expression.
4979 : : * Fill in a VariableStatData struct to describe the expression.
4980 : : *
4981 : : * Inputs:
4982 : : * root: the planner info
4983 : : * node: the expression tree to examine
4984 : : * varRelid: see specs for restriction selectivity functions
4985 : : *
4986 : : * Outputs: *vardata is filled as follows:
4987 : : * var: the input expression (with any binary relabeling stripped, if
4988 : : * it is or contains a variable; but otherwise the type is preserved)
4989 : : * rel: RelOptInfo for relation containing variable; NULL if expression
4990 : : * contains no Vars (NOTE this could point to a RelOptInfo of a
4991 : : * subquery, not one in the current query).
4992 : : * statsTuple: the pg_statistic entry for the variable, if one exists;
4993 : : * otherwise NULL.
4994 : : * freefunc: pointer to a function to release statsTuple with.
4995 : : * vartype: exposed type of the expression; this should always match
4996 : : * the declared input type of the operator we are estimating for.
4997 : : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
4998 : : * commonly the same as the exposed type of the variable argument,
4999 : : * but can be different in binary-compatible-type cases.
5000 : : * isunique: true if we were able to match the var to a unique index or a
5001 : : * single-column DISTINCT clause, implying its values are unique for
5002 : : * this query. (Caution: this should be trusted for statistical
5003 : : * purposes only, since we do not check indimmediate nor verify that
5004 : : * the exact same definition of equality applies.)
5005 : : * acl_ok: true if current user has permission to read the column(s)
5006 : : * underlying the pg_statistic entry. This is consulted by
5007 : : * statistic_proc_security_check().
5008 : : *
5009 : : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5010 : : */
5011 : : void
6888 tgl@sss.pgh.pa.us 5012 : 1204314 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5013 : : VariableStatData *vardata)
5014 : : {
5015 : : Node *basenode;
5016 : : Relids varnos;
5017 : : RelOptInfo *onerel;
5018 : :
5019 : : /* Make sure we don't return dangling pointers in vardata */
7362 5020 [ + - + - : 8430198 : MemSet(vardata, 0, sizeof(VariableStatData));
+ - + - +
+ ]
5021 : :
5022 : : /* Save the exposed type of the expression */
6953 5023 : 1204314 : vardata->vartype = exprType(node);
5024 : :
5025 : : /* Look inside any binary-compatible relabeling */
5026 : :
7362 5027 [ + + ]: 1204314 : if (IsA(node, RelabelType))
6959 5028 : 11375 : basenode = (Node *) ((RelabelType *) node)->arg;
5029 : : else
5030 : 1192939 : basenode = node;
5031 : :
5032 : : /* Fast path for a simple Var */
5033 : :
5034 [ + + + + ]: 1204314 : if (IsA(basenode, Var) &&
5035 [ + + ]: 281236 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5036 : : {
5037 : 851135 : Var *var = (Var *) basenode;
5038 : :
5039 : : /* Set up result fields other than the stats tuple */
5040 : 851135 : vardata->var = basenode; /* return Var without relabeling */
7362 5041 : 851135 : vardata->rel = find_base_rel(root, var->varno);
5042 : 851135 : vardata->atttype = var->vartype;
5043 : 851135 : vardata->atttypmod = var->vartypmod;
5537 5044 : 851135 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5045 : :
5046 : : /* Try to locate some stats */
4606 5047 : 851135 : examine_simple_variable(root, var, vardata);
5048 : :
7362 5049 : 851135 : return;
5050 : : }
5051 : :
5052 : : /*
5053 : : * Okay, it's a more complicated expression. Determine variable
5054 : : * membership. Note that when varRelid isn't zero, only vars of that
5055 : : * relation are considered "real" vars.
5056 : : */
1179 5057 : 353179 : varnos = pull_varnos(root, basenode);
5058 : :
7362 5059 : 353179 : onerel = NULL;
5060 : :
138 drowley@postgresql.o 5061 [ + + ]:GNC 353179 : if (bms_is_empty(varnos))
5062 : : {
5063 : : /* No Vars at all ... must be pseudo-constant clause */
5064 : : }
5065 : : else
5066 : : {
5067 : : int relid;
5068 : :
5069 [ + + ]: 168360 : if (bms_get_singleton_member(varnos, &relid))
5070 : : {
5071 [ + + + + ]: 165621 : if (varRelid == 0 || varRelid == relid)
5072 : : {
5073 : 25937 : onerel = find_base_rel(root, relid);
7362 tgl@sss.pgh.pa.us 5074 :CBC 25937 : vardata->rel = onerel;
6756 bruce@momjian.us 5075 : 25937 : node = basenode; /* strip any relabeling */
5076 : : }
5077 : : /* else treat it as a constant */
5078 : : }
5079 : : else
5080 : : {
5081 : : /* varnos has multiple relids */
7362 tgl@sss.pgh.pa.us 5082 [ + + ]: 2739 : if (varRelid == 0)
5083 : : {
5084 : : /* treat it as a variable of a join relation */
5085 : 2488 : vardata->rel = find_join_rel(root, varnos);
6756 bruce@momjian.us 5086 : 2488 : node = basenode; /* strip any relabeling */
5087 : : }
7362 tgl@sss.pgh.pa.us 5088 [ + + ]: 251 : else if (bms_is_member(varRelid, varnos))
5089 : : {
5090 : : /* ignore the vars belonging to other relations */
5091 : 46 : vardata->rel = find_base_rel(root, varRelid);
6756 bruce@momjian.us 5092 : 46 : node = basenode; /* strip any relabeling */
5093 : : /* note: no point in expressional-index search here */
5094 : : }
5095 : : /* else treat it as a constant */
5096 : : }
5097 : : }
5098 : :
7362 tgl@sss.pgh.pa.us 5099 : 353179 : bms_free(varnos);
5100 : :
6959 5101 : 353179 : vardata->var = node;
7362 5102 : 353179 : vardata->atttype = exprType(node);
5103 : 353179 : vardata->atttypmod = exprTypmod(node);
5104 : :
5105 [ + + ]: 353179 : if (onerel)
5106 : : {
5107 : : /*
5108 : : * We have an expression in vars of a single relation. Try to match
5109 : : * it to expressional index columns, in hopes of finding some
5110 : : * statistics.
5111 : : *
5112 : : * Note that we consider all index columns including INCLUDE columns,
5113 : : * since there could be stats for such columns. But the test for
5114 : : * uniqueness needs to be warier.
5115 : : *
5116 : : * XXX it's conceivable that there are multiple matches with different
5117 : : * index opfamilies; if so, we need to pick one that matches the
5118 : : * operator we are estimating for. FIXME later.
5119 : : */
5120 : : ListCell *ilist;
5121 : : ListCell *slist;
5122 : : Oid userid;
5123 : :
5124 : : /*
5125 : : * Determine the user ID to use for privilege checks: either
5126 : : * onerel->userid if it's set (e.g., in case we're accessing the table
5127 : : * via a view), or the current user otherwise.
5128 : : *
5129 : : * If we drill down to child relations, we keep using the same userid:
5130 : : * it's going to be the same anyway, due to how we set up the relation
5131 : : * tree (q.v. build_simple_rel).
5132 : : */
451 alvherre@alvh.no-ip. 5133 [ + + ]: 25937 : userid = OidIsValid(onerel->userid) ? onerel->userid : GetUserId();
5134 : :
7362 tgl@sss.pgh.pa.us 5135 [ + + + + : 50885 : foreach(ilist, onerel->indexlist)
+ + ]
5136 : : {
5137 : 26337 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5138 : : ListCell *indexpr_item;
5139 : : int pos;
5140 : :
7263 neilc@samurai.com 5141 : 26337 : indexpr_item = list_head(index->indexprs);
5142 [ + + ]: 26337 : if (indexpr_item == NULL)
7362 tgl@sss.pgh.pa.us 5143 : 23964 : continue; /* no expressions here... */
5144 : :
5145 [ + + ]: 3393 : for (pos = 0; pos < index->ncolumns; pos++)
5146 : : {
5147 [ + + ]: 2409 : if (index->indexkeys[pos] == 0)
5148 : : {
5149 : : Node *indexkey;
5150 : :
7263 neilc@samurai.com 5151 [ - + ]: 2373 : if (indexpr_item == NULL)
7362 tgl@sss.pgh.pa.us 5152 [ # # ]:UBC 0 : elog(ERROR, "too few entries in indexprs list");
7263 neilc@samurai.com 5153 :CBC 2373 : indexkey = (Node *) lfirst(indexpr_item);
7362 tgl@sss.pgh.pa.us 5154 [ + - - + ]: 2373 : if (indexkey && IsA(indexkey, RelabelType))
7362 tgl@sss.pgh.pa.us 5155 :UBC 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
7362 tgl@sss.pgh.pa.us 5156 [ + + ]:CBC 2373 : if (equal(node, indexkey))
5157 : : {
5158 : : /*
5159 : : * Found a match ... is it a unique index? Tests here
5160 : : * should match has_unique_index().
5161 : : */
5162 [ + + ]: 1743 : if (index->unique &&
2199 teodor@sigaev.ru 5163 [ + - + - ]: 255 : index->nkeycolumns == 1 &&
1888 tgl@sss.pgh.pa.us 5164 : 255 : pos == 0 &&
5537 5165 [ - + - - ]: 255 : (index->indpred == NIL || index->predOK))
7362 5166 : 255 : vardata->isunique = true;
5167 : :
5168 : : /*
5169 : : * Has it got stats? We only consider stats for
5170 : : * non-partial indexes, since partial indexes probably
5171 : : * don't reflect whole-relation statistics; the above
5172 : : * check for uniqueness is the only info we take from
5173 : : * a partial index.
5174 : : *
5175 : : * An index stats hook, however, must make its own
5176 : : * decisions about what to do with partial indexes.
5177 : : */
5677 5178 [ - + - - ]: 1743 : if (get_index_stats_hook &&
5677 tgl@sss.pgh.pa.us 5179 :UBC 0 : (*get_index_stats_hook) (root, index->indexoid,
5180 : 0 : pos + 1, vardata))
5181 : : {
5182 : : /*
5183 : : * The hook took control of acquiring a stats
5184 : : * tuple. If it did supply a tuple, it'd better
5185 : : * have supplied a freefunc.
5186 : : */
5187 [ # # ]: 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5188 [ # # ]: 0 : !vardata->freefunc)
5189 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
5190 : : }
5537 tgl@sss.pgh.pa.us 5191 [ + - ]:CBC 1743 : else if (index->indpred == NIL)
5192 : : {
5677 5193 : 1743 : vardata->statsTuple =
5173 rhaas@postgresql.org 5194 : 3486 : SearchSysCache3(STATRELATTINH,
5195 : : ObjectIdGetDatum(index->indexoid),
5161 bruce@momjian.us 5196 : 1743 : Int16GetDatum(pos + 1),
5197 : : BoolGetDatum(false));
5677 tgl@sss.pgh.pa.us 5198 : 1743 : vardata->freefunc = ReleaseSysCache;
5199 : :
2536 peter_e@gmx.net 5200 [ + + ]: 1743 : if (HeapTupleIsValid(vardata->statsTuple))
5201 : : {
5202 : : /* Get index's table for permission check */
5203 : : RangeTblEntry *rte;
5204 : :
5205 [ + - ]: 1389 : rte = planner_rt_fetch(index->rel->relid, root);
5206 [ - + ]: 1389 : Assert(rte->rtekind == RTE_RELATION);
5207 : :
5208 : : /*
5209 : : * For simplicity, we insist on the whole
5210 : : * table being selectable, rather than trying
5211 : : * to identify which column(s) the index
5212 : : * depends on. Also require all rows to be
5213 : : * selectable --- there must be no
5214 : : * securityQuals from security barrier views
5215 : : * or RLS policies.
5216 : : */
5217 : 1389 : vardata->acl_ok =
1805 dean.a.rasheed@gmail 5218 [ + - + + ]: 2778 : rte->securityQuals == NIL &&
5219 : 1389 : (pg_class_aclcheck(rte->relid, userid,
2489 tgl@sss.pgh.pa.us 5220 :ECB (1383) : ACL_SELECT) == ACLCHECK_OK);
5221 : :
5222 : : /*
5223 : : * If the user doesn't have permissions to
5224 : : * access an inheritance child relation, check
5225 : : * the permissions of the table actually
5226 : : * mentioned in the query, since most likely
5227 : : * the user does have that permission. Note
5228 : : * that whole-table select privilege on the
5229 : : * parent doesn't quite guarantee that the
5230 : : * user could read all columns of the child.
5231 : : * But in practice it's unlikely that any
5232 : : * interesting security violation could result
5233 : : * from allowing access to the expression
5234 : : * index's stats, so we allow it anyway. See
5235 : : * similar code in examine_simple_variable()
5236 : : * for additional comments.
5237 : : */
1601 tgl@sss.pgh.pa.us 5238 [ + + ]:CBC 1389 : if (!vardata->acl_ok &&
5239 [ + + ]: 9 : root->append_rel_array != NULL)
5240 : : {
5241 : : AppendRelInfo *appinfo;
5242 : 6 : Index varno = index->rel->relid;
5243 : :
5244 : 6 : appinfo = root->append_rel_array[varno];
5245 [ + + ]: 18 : while (appinfo &&
5246 [ + - ]: 12 : planner_rt_fetch(appinfo->parent_relid,
5247 [ + - ]: 12 : root)->rtekind == RTE_RELATION)
5248 : : {
5249 : 12 : varno = appinfo->parent_relid;
5250 : 12 : appinfo = root->append_rel_array[varno];
5251 : : }
5252 [ + - ]: 6 : if (varno != index->rel->relid)
5253 : : {
5254 : : /* Repeat access check on this rel */
5255 [ + - ]: 6 : rte = planner_rt_fetch(varno, root);
5256 [ - + ]: 6 : Assert(rte->rtekind == RTE_RELATION);
5257 : :
5258 : 6 : vardata->acl_ok =
5259 [ + - + + ]: 12 : rte->securityQuals == NIL &&
5260 : 6 : (pg_class_aclcheck(rte->relid,
5261 : : userid,
1601 tgl@sss.pgh.pa.us 5262 :ECB (6) : ACL_SELECT) == ACLCHECK_OK);
5263 : : }
5264 : : }
5265 : : }
5266 : : else
5267 : : {
5268 : : /* suppress leakproofness checks later */
2536 peter_e@gmx.net 5269 :CBC 354 : vardata->acl_ok = true;
5270 : : }
5271 : : }
7362 tgl@sss.pgh.pa.us 5272 [ + + ]: 1743 : if (vardata->statsTuple)
5273 : 1389 : break;
5274 : : }
1735 5275 : 984 : indexpr_item = lnext(index->indexprs, indexpr_item);
5276 : : }
5277 : : }
7362 5278 [ + + ]: 2373 : if (vardata->statsTuple)
5279 : 1389 : break;
5280 : : }
5281 : :
5282 : : /*
5283 : : * Search extended statistics for one with a matching expression.
5284 : : * There might be multiple ones, so just grab the first one. In the
5285 : : * future, we might consider the statistics target (and pick the most
5286 : : * accurate statistics) and maybe some other parameters.
5287 : : */
1115 tomas.vondra@postgre 5288 [ + + + + : 27947 : foreach(slist, onerel->statlist)
+ + ]
5289 : : {
5290 : 2154 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
703 tgl@sss.pgh.pa.us 5291 [ + - ]: 2154 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5292 : : ListCell *expr_item;
5293 : : int pos;
5294 : :
5295 : : /*
5296 : : * Stop once we've found statistics for the expression (either
5297 : : * from extended stats, or for an index in the preceding loop).
5298 : : */
1115 tomas.vondra@postgre 5299 [ + + ]: 2154 : if (vardata->statsTuple)
5300 : 144 : break;
5301 : :
5302 : : /* skip stats without per-expression stats */
5303 [ + + ]: 2010 : if (info->kind != STATS_EXT_EXPRESSIONS)
5304 : 1008 : continue;
5305 : :
5306 : : /* skip stats with mismatching stxdinherit value */
530 tgl@sss.pgh.pa.us 5307 [ + + ]: 1002 : if (info->inherit != rte->inh)
5308 : 3 : continue;
5309 : :
1115 tomas.vondra@postgre 5310 : 999 : pos = 0;
5311 [ + - + + : 1650 : foreach(expr_item, info->exprs)
+ + ]
5312 : : {
5313 : 1473 : Node *expr = (Node *) lfirst(expr_item);
5314 : :
5315 [ - + ]: 1473 : Assert(expr);
5316 : :
5317 : : /* strip RelabelType before comparing it */
5318 [ + - - + ]: 1473 : if (expr && IsA(expr, RelabelType))
1115 tomas.vondra@postgre 5319 :UBC 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5320 : :
5321 : : /* found a match, see if we can extract pg_statistic row */
1115 tomas.vondra@postgre 5322 [ + + ]:CBC 1473 : if (equal(node, expr))
5323 : : {
5324 : : /*
5325 : : * XXX Not sure if we should cache the tuple somewhere.
5326 : : * Now we just create a new copy every time.
5327 : : */
819 5328 : 822 : vardata->statsTuple =
5329 : 822 : statext_expressions_load(info->statOid, rte->inh, pos);
5330 : :
5331 : 822 : vardata->freefunc = ReleaseDummy;
5332 : :
5333 : : /*
5334 : : * For simplicity, we insist on the whole table being
5335 : : * selectable, rather than trying to identify which
5336 : : * column(s) the statistics object depends on. Also
5337 : : * require all rows to be selectable --- there must be no
5338 : : * securityQuals from security barrier views or RLS
5339 : : * policies.
5340 : : */
1115 5341 : 822 : vardata->acl_ok =
5342 [ + - + - ]: 1644 : rte->securityQuals == NIL &&
5343 : 822 : (pg_class_aclcheck(rte->relid, userid,
1115 tomas.vondra@postgre 5344 :ECB (822) : ACL_SELECT) == ACLCHECK_OK);
5345 : :
5346 : : /*
5347 : : * If the user doesn't have permissions to access an
5348 : : * inheritance child relation, check the permissions of
5349 : : * the table actually mentioned in the query, since most
5350 : : * likely the user does have that permission. Note that
5351 : : * whole-table select privilege on the parent doesn't
5352 : : * quite guarantee that the user could read all columns of
5353 : : * the child. But in practice it's unlikely that any
5354 : : * interesting security violation could result from
5355 : : * allowing access to the expression stats, so we allow it
5356 : : * anyway. See similar code in examine_simple_variable()
5357 : : * for additional comments.
5358 : : */
1115 tomas.vondra@postgre 5359 [ - + ]:CBC 822 : if (!vardata->acl_ok &&
1115 tomas.vondra@postgre 5360 [ # # ]:UBC 0 : root->append_rel_array != NULL)
5361 : : {
5362 : : AppendRelInfo *appinfo;
5363 : 0 : Index varno = onerel->relid;
5364 : :
5365 : 0 : appinfo = root->append_rel_array[varno];
5366 [ # # ]: 0 : while (appinfo &&
5367 [ # # ]: 0 : planner_rt_fetch(appinfo->parent_relid,
5368 [ # # ]: 0 : root)->rtekind == RTE_RELATION)
5369 : : {
5370 : 0 : varno = appinfo->parent_relid;
5371 : 0 : appinfo = root->append_rel_array[varno];
5372 : : }
5373 [ # # ]: 0 : if (varno != onerel->relid)
5374 : : {
5375 : : /* Repeat access check on this rel */
5376 [ # # ]: 0 : rte = planner_rt_fetch(varno, root);
5377 [ # # ]: 0 : Assert(rte->rtekind == RTE_RELATION);
5378 : :
5379 : 0 : vardata->acl_ok =
5380 [ # # # # ]: 0 : rte->securityQuals == NIL &&
5381 : 0 : (pg_class_aclcheck(rte->relid,
5382 : : userid,
1115 tomas.vondra@postgre 5383 :EUB : ACL_SELECT) == ACLCHECK_OK);
5384 : : }
5385 : : }
5386 : :
1115 tomas.vondra@postgre 5387 :CBC 822 : break;
5388 : : }
5389 : :
5390 : 651 : pos++;
5391 : : }
5392 : : }
5393 : : }
5394 : : }
5395 : :
5396 : : /*
5397 : : * examine_simple_variable
5398 : : * Handle a simple Var for examine_variable
5399 : : *
5400 : : * This is split out as a subroutine so that we can recurse to deal with
5401 : : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
5402 : : *
5403 : : * We already filled in all the fields of *vardata except for the stats tuple.
5404 : : */
5405 : : static void
4606 tgl@sss.pgh.pa.us 5406 : 853214 : examine_simple_variable(PlannerInfo *root, Var *var,
5407 : : VariableStatData *vardata)
5408 : : {
5409 : 853214 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
5410 : :
5411 [ - + ]: 853214 : Assert(IsA(rte, RangeTblEntry));
5412 : :
5413 [ - + - - ]: 853214 : if (get_relation_stats_hook &&
4606 tgl@sss.pgh.pa.us 5414 :UBC 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
5415 : : {
5416 : : /*
5417 : : * The hook took control of acquiring a stats tuple. If it did supply
5418 : : * a tuple, it'd better have supplied a freefunc.
5419 : : */
5420 [ # # ]: 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5421 [ # # ]: 0 : !vardata->freefunc)
5422 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
5423 : : }
4606 tgl@sss.pgh.pa.us 5424 [ + + ]:CBC 853214 : else if (rte->rtekind == RTE_RELATION)
5425 : : {
5426 : : /*
5427 : : * Plain table or parent of an inheritance appendrel, so look up the
5428 : : * column in pg_statistic
5429 : : */
5430 : 814037 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
5431 : : ObjectIdGetDatum(rte->relid),
5432 : 814037 : Int16GetDatum(var->varattno),
5433 : 814037 : BoolGetDatum(rte->inh));
5434 : 814037 : vardata->freefunc = ReleaseSysCache;
5435 : :
2536 peter_e@gmx.net 5436 [ + + ]: 814037 : if (HeapTupleIsValid(vardata->statsTuple))
5437 : : {
97 tgl@sss.pgh.pa.us 5438 :GNC 621737 : RelOptInfo *onerel = find_base_rel_noerr(root, var->varno);
5439 : : Oid userid;
5440 : :
5441 : : /*
5442 : : * Check if user has permission to read this column. We require
5443 : : * all rows to be accessible, so there must be no securityQuals
5444 : : * from security barrier views or RLS policies.
5445 : : *
5446 : : * Normally the Var will have an associated RelOptInfo from which
5447 : : * we can find out which userid to do the check as; but it might
5448 : : * not if it's a RETURNING Var for an INSERT target relation. In
5449 : : * that case use the RTEPermissionInfo associated with the RTE.
5450 : : */
5451 [ + + ]: 621737 : if (onerel)
5452 : 621716 : userid = onerel->userid;
5453 : : else
5454 : : {
5455 : : RTEPermissionInfo *perminfo;
5456 : :
5457 : 21 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
5458 : 21 : userid = perminfo->checkAsUser;
5459 : : }
5460 [ + + ]: 621737 : if (!OidIsValid(userid))
5461 : 585249 : userid = GetUserId();
5462 : :
2536 peter_e@gmx.net 5463 :CBC 621737 : vardata->acl_ok =
1805 dean.a.rasheed@gmail 5464 [ + - + + ]: 1243598 : rte->securityQuals == NIL &&
5465 : 621737 : ((pg_class_aclcheck(rte->relid, userid,
5466 [ + + ]: 124 : ACL_SELECT) == ACLCHECK_OK) ||
5467 : 124 : (pg_attribute_aclcheck(rte->relid, var->varattno, userid,
1805 dean.a.rasheed@gmail 5468 :ECB (529928) : ACL_SELECT) == ACLCHECK_OK));
5469 : :
5470 : : /*
5471 : : * If the user doesn't have permissions to access an inheritance
5472 : : * child relation or specifically this attribute, check the
5473 : : * permissions of the table/column actually mentioned in the
5474 : : * query, since most likely the user does have that permission
5475 : : * (else the query will fail at runtime), and if the user can read
5476 : : * the column there then he can get the values of the child table
5477 : : * too. To do that, we must find out which of the root parent's
5478 : : * attributes the child relation's attribute corresponds to.
5479 : : */
1601 tgl@sss.pgh.pa.us 5480 [ + + + - ]:CBC 621737 : if (!vardata->acl_ok && var->varattno > 0 &&
5481 [ + + ]: 24 : root->append_rel_array != NULL)
5482 : : {
5483 : : AppendRelInfo *appinfo;
5484 : 6 : Index varno = var->varno;
5485 : 6 : int varattno = var->varattno;
5486 : 6 : bool found = false;
5487 : :
5488 : 6 : appinfo = root->append_rel_array[varno];
5489 : :
5490 : : /*
5491 : : * Partitions are mapped to their immediate parent, not the
5492 : : * root parent, so must be ready to walk up multiple
5493 : : * AppendRelInfos. But stop if we hit a parent that is not
5494 : : * RTE_RELATION --- that's a flattened UNION ALL subquery, not
5495 : : * an inheritance parent.
5496 : : */
5497 [ + + ]: 18 : while (appinfo &&
5498 [ + - ]: 12 : planner_rt_fetch(appinfo->parent_relid,
5499 [ + - ]: 12 : root)->rtekind == RTE_RELATION)
5500 : : {
5501 : : int parent_varattno;
5502 : :
5503 : 12 : found = false;
1595 5504 [ + - + - ]: 12 : if (varattno <= 0 || varattno > appinfo->num_child_cols)
5505 : : break; /* safety check */
5506 : 12 : parent_varattno = appinfo->parent_colnos[varattno - 1];
5507 [ - + ]: 12 : if (parent_varattno == 0)
1595 tgl@sss.pgh.pa.us 5508 :UBC 0 : break; /* Var is local to child */
5509 : :
1601 tgl@sss.pgh.pa.us 5510 :CBC 12 : varno = appinfo->parent_relid;
5511 : 12 : varattno = parent_varattno;
1595 5512 : 12 : found = true;
5513 : :
5514 : : /* If the parent is itself a child, continue up. */
1601 5515 : 12 : appinfo = root->append_rel_array[varno];
5516 : : }
5517 : :
5518 : : /*
5519 : : * In rare cases, the Var may be local to the child table, in
5520 : : * which case, we've got to live with having no access to this
5521 : : * column's stats.
5522 : : */
5523 [ - + ]: 6 : if (!found)
1601 tgl@sss.pgh.pa.us 5524 :UBC 0 : return;
5525 : :
5526 : : /* Repeat the access check on this parent rel & column */
1601 tgl@sss.pgh.pa.us 5527 [ + - ]:CBC 6 : rte = planner_rt_fetch(varno, root);
5528 [ - + ]: 6 : Assert(rte->rtekind == RTE_RELATION);
5529 : :
5530 : : /*
5531 : : * Fine to use the same userid as it's the same in all
5532 : : * relations of a given inheritance tree.
5533 : : */
5534 : 6 : vardata->acl_ok =
5535 [ + - + + ]: 15 : rte->securityQuals == NIL &&
5536 : 6 : ((pg_class_aclcheck(rte->relid, userid,
5537 [ + - ]: 3 : ACL_SELECT) == ACLCHECK_OK) ||
5538 : 3 : (pg_attribute_aclcheck(rte->relid, varattno, userid,
1601 tgl@sss.pgh.pa.us 5539 :ECB (6) : ACL_SELECT) == ACLCHECK_OK));
5540 : : }
5541 : : }
5542 : : else
5543 : : {
5544 : : /* suppress any possible leakproofness checks later */
2536 peter_e@gmx.net 5545 :CBC 192300 : vardata->acl_ok = true;
5546 : : }
5547 : : }
149 tgl@sss.pgh.pa.us 5548 [ + + + + ]:GNC 39177 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
5549 [ + + + + ]: 35561 : (rte->rtekind == RTE_CTE && !rte->self_reference))
5550 : : {
5551 : : /*
5552 : : * Plain subquery (not one that was converted to an appendrel) or
5553 : : * non-recursive CTE. In either case, we can try to find out what the
5554 : : * Var refers to within the subquery. We skip this for appendrel and
5555 : : * recursive-CTE cases because any column stats we did find would
5556 : : * likely not be very relevant.
5557 : : */
5558 : : PlannerInfo *subroot;
5559 : : Query *subquery;
5560 : : List *subtlist;
5561 : : TargetEntry *ste;
5562 : :
5563 : : /*
5564 : : * Punt if it's a whole-row var rather than a plain column reference.
5565 : : */
3807 tgl@sss.pgh.pa.us 5566 [ - + ]:CBC 6060 : if (var->varattno == InvalidAttrNumber)
3807 tgl@sss.pgh.pa.us 5567 :UBC 0 : return;
5568 : :
5569 : : /*
5570 : : * Otherwise, find the subquery's planner subroot.
5571 : : */
149 tgl@sss.pgh.pa.us 5572 [ + + ]:GNC 6060 : if (rte->rtekind == RTE_SUBQUERY)
5573 : : {
5574 : : RelOptInfo *rel;
5575 : :
5576 : : /*
5577 : : * Fetch RelOptInfo for subquery. Note that we don't change the
5578 : : * rel returned in vardata, since caller expects it to be a rel of
5579 : : * the caller's query level. Because we might already be
5580 : : * recursing, we can't use that rel pointer either, but have to
5581 : : * look up the Var's rel afresh.
5582 : : */
5583 : 3616 : rel = find_base_rel(root, var->varno);
5584 : :
5585 : 3616 : subroot = rel->subroot;
5586 : : }
5587 : : else
5588 : : {
5589 : : /* CTE case is more difficult */
5590 : : PlannerInfo *cteroot;
5591 : : Index levelsup;
5592 : : int ndx;
5593 : : int plan_id;
5594 : : ListCell *lc;
5595 : :
5596 : : /*
5597 : : * Find the referenced CTE, and locate the subroot previously made
5598 : : * for it.
5599 : : */
5600 : 2444 : levelsup = rte->ctelevelsup;
5601 : 2444 : cteroot = root;
5602 [ + + ]: 4725 : while (levelsup-- > 0)
5603 : : {
5604 : 2281 : cteroot = cteroot->parent_root;
5605 [ - + ]: 2281 : if (!cteroot) /* shouldn't happen */
149 tgl@sss.pgh.pa.us 5606 [ # # ]:UNC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
5607 : : }
5608 : :
5609 : : /*
5610 : : * Note: cte_plan_ids can be shorter than cteList, if we are still
5611 : : * working on planning the CTEs (ie, this is a side-reference from
5612 : : * another CTE). So we mustn't use forboth here.
5613 : : */
149 tgl@sss.pgh.pa.us 5614 :GNC 2444 : ndx = 0;
5615 [ + - + - : 3563 : foreach(lc, cteroot->parse->cteList)
+ - ]
5616 : : {
5617 : 3563 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
5618 : :
5619 [ + + ]: 3563 : if (strcmp(cte->ctename, rte->ctename) == 0)
5620 : 2444 : break;
5621 : 1119 : ndx++;
5622 : : }
5623 [ - + ]: 2444 : if (lc == NULL) /* shouldn't happen */
149 tgl@sss.pgh.pa.us 5624 [ # # ]:UNC 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
149 tgl@sss.pgh.pa.us 5625 [ - + ]:GNC 2444 : if (ndx >= list_length(cteroot->cte_plan_ids))
149 tgl@sss.pgh.pa.us 5626 [ # # ]:UNC 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
149 tgl@sss.pgh.pa.us 5627 :GNC 2444 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
5628 [ - + ]: 2444 : if (plan_id <= 0)
149 tgl@sss.pgh.pa.us 5629 [ # # ]:UNC 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
149 tgl@sss.pgh.pa.us 5630 :GNC 2444 : subroot = list_nth(root->glob->subroots, plan_id - 1);
5631 : : }
5632 : :
5633 : : /* If the subquery hasn't been planned yet, we have to punt */
5634 [ - + ]: 6060 : if (subroot == NULL)
149 tgl@sss.pgh.pa.us 5635 :UNC 0 : return;
149 tgl@sss.pgh.pa.us 5636 [ - + ]:GNC 6060 : Assert(IsA(subroot, PlannerInfo));
5637 : :
5638 : : /*
5639 : : * We must use the subquery parsetree as mangled by the planner, not
5640 : : * the raw version from the RTE, because we need a Var that will refer
5641 : : * to the subroot's live RelOptInfos. For instance, if any subquery
5642 : : * pullup happened during planning, Vars in the targetlist might have
5643 : : * gotten replaced, and we need to see the replacement expressions.
5644 : : */
5645 : 6060 : subquery = subroot->parse;
5646 [ - + ]: 6060 : Assert(IsA(subquery, Query));
5647 : :
5648 : : /*
5649 : : * Punt if subquery uses set operations or GROUP BY, as these will
5650 : : * mash underlying columns' stats beyond recognition. (Set ops are
5651 : : * particularly nasty; if we forged ahead, we would return stats
5652 : : * relevant to only the leftmost subselect...) DISTINCT is also
5653 : : * problematic, but we check that later because there is a possibility
5654 : : * of learning something even with it.
5655 : : */
4441 tgl@sss.pgh.pa.us 5656 [ + + ]:CBC 6060 : if (subquery->setOperations ||
1048 5657 [ + + ]: 5420 : subquery->groupClause ||
5658 [ - + ]: 5092 : subquery->groupingSets)
4497 rhaas@postgresql.org 5659 : 968 : return;
5660 : :
5661 : : /* Get the subquery output expression referenced by the upper Var */
149 tgl@sss.pgh.pa.us 5662 [ + + ]:GNC 5092 : if (subquery->returningList)
5663 : 103 : subtlist = subquery->returningList;
5664 : : else
5665 : 4989 : subtlist = subquery->targetList;
5666 : 5092 : ste = get_tle_by_resno(subtlist, var->varattno);
4606 tgl@sss.pgh.pa.us 5667 [ + - - + ]:CBC 5092 : if (ste == NULL || ste->resjunk)
4606 tgl@sss.pgh.pa.us 5668 [ # # ]:UBC 0 : elog(ERROR, "subquery %s does not have attribute %d",
5669 : : rte->eref->aliasname, var->varattno);
4606 tgl@sss.pgh.pa.us 5670 :CBC 5092 : var = (Var *) ste->expr;
5671 : :
5672 : : /*
5673 : : * If subquery uses DISTINCT, we can't make use of any stats for the
5674 : : * variable ... but, if it's the only DISTINCT column, we are entitled
5675 : : * to consider it unique. We do the test this way so that it works
5676 : : * for cases involving DISTINCT ON.
5677 : : */
4441 5678 [ + + ]: 5092 : if (subquery->distinctClause)
5679 : : {
5680 [ + + + + ]: 853 : if (list_length(subquery->distinctClause) == 1 &&
5681 : 281 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
5682 : 132 : vardata->isunique = true;
5683 : : /* cannot go further */
5684 : 572 : return;
5685 : : }
5686 : :
5687 : : /*
5688 : : * If the sub-query originated from a view with the security_barrier
5689 : : * attribute, we must not look at the variable's statistics, though it
5690 : : * seems all right to notice the existence of a DISTINCT clause. So
5691 : : * stop here.
5692 : : *
5693 : : * This is probably a harsher restriction than necessary; it's
5694 : : * certainly OK for the selectivity estimator (which is a C function,
5695 : : * and therefore omnipotent anyway) to look at the statistics. But
5696 : : * many selectivity estimators will happily *invoke the operator
5697 : : * function* to try to work out a good estimate - and that's not OK.
5698 : : * So for now, don't dig down for stats.
5699 : : */
5700 [ + + ]: 4520 : if (rte->security_barrier)
5701 : 114 : return;
5702 : :
5703 : : /* Can only handle a simple Var of subquery's query level */
4606 5704 [ + - + + ]: 4406 : if (var && IsA(var, Var) &&
5705 [ + - ]: 2079 : var->varlevelsup == 0)
5706 : : {
5707 : : /*
5708 : : * OK, recurse into the subquery. Note that the original setting
5709 : : * of vardata->isunique (which will surely be false) is left
5710 : : * unchanged in this situation. That's what we want, since even
5711 : : * if the underlying column is unique, the subquery may have
5712 : : * joined to other tables in a way that creates duplicates.
5713 : : */
149 tgl@sss.pgh.pa.us 5714 :GNC 2079 : examine_simple_variable(subroot, var, vardata);
5715 : : }
5716 : : }
5717 : : else
5718 : : {
5719 : : /*
5720 : : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
5721 : : * see RTE_JOIN here because join alias Vars have already been
5722 : : * flattened.) There's not much we can do with function outputs, but
5723 : : * maybe someday try to be smarter about VALUES.
5724 : : */
5725 : : }
5726 : : }
5727 : :
5728 : : /*
5729 : : * Check whether it is permitted to call func_oid passing some of the
5730 : : * pg_statistic data in vardata. We allow this either if the user has SELECT
5731 : : * privileges on the table or column underlying the pg_statistic data or if
5732 : : * the function is marked leak-proof.
5733 : : */
5734 : : bool
2536 peter_e@gmx.net 5735 :CBC 423393 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
5736 : : {
5737 [ + + ]: 423393 : if (vardata->acl_ok)
5738 : 422930 : return true;
5739 : :
5740 [ - + ]: 463 : if (!OidIsValid(func_oid))
2536 peter_e@gmx.net 5741 :UBC 0 : return false;
5742 : :
2536 peter_e@gmx.net 5743 [ + + ]:CBC 463 : if (get_func_leakproof(func_oid))
5744 : 212 : return true;
5745 : :
5746 [ - + ]: 251 : ereport(DEBUG2,
5747 : : (errmsg_internal("not using statistics because function \"%s\" is not leak-proof",
5748 : : get_func_name(func_oid))));
5749 : 251 : return false;
5750 : : }
5751 : :
5752 : : /*
5753 : : * get_variable_numdistinct
5754 : : * Estimate the number of distinct values of a variable.
5755 : : *
5756 : : * vardata: results of examine_variable
5757 : : * *isdefault: set to true if the result is a default rather than based on
5758 : : * anything meaningful.
5759 : : *
5760 : : * NB: be careful to produce a positive integral result, since callers may
5761 : : * compare the result to exact integer counts, or might divide by it.
5762 : : */
5763 : : double
4606 tgl@sss.pgh.pa.us 5764 : 577526 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
5765 : : {
5766 : : double stadistinct;
2807 5767 : 577526 : double stanullfrac = 0.0;
5768 : : double ntuples;
5769 : :
4606 5770 : 577526 : *isdefault = false;
5771 : :
5772 : : /*
5773 : : * Determine the stadistinct value to use. There are cases where we can
5774 : : * get an estimate even without a pg_statistic entry, or can get a better
5775 : : * value than is in pg_statistic. Grab stanullfrac too if we can find it
5776 : : * (otherwise, assume no nulls, for lack of any better idea).
5777 : : */
7362 5778 [ + + ]: 577526 : if (HeapTupleIsValid(vardata->statsTuple))
5779 : : {
5780 : : /* Use the pg_statistic entry */
5781 : : Form_pg_statistic stats;
5782 : :
5783 : 420102 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
5784 : 420102 : stadistinct = stats->stadistinct;
2807 5785 : 420102 : stanullfrac = stats->stanullfrac;
5786 : : }
6953 5787 [ + + ]: 157424 : else if (vardata->vartype == BOOLOID)
5788 : : {
5789 : : /*
5790 : : * Special-case boolean columns: presumably, two distinct values.
5791 : : *
5792 : : * Are there any other datatypes we should wire in special estimates
5793 : : * for?
5794 : : */
7362 5795 : 162 : stadistinct = 2.0;
5796 : : }
2433 5797 [ + + + + ]: 157262 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
5798 : : {
5799 : : /*
5800 : : * If the Var represents a column of a VALUES RTE, assume it's unique.
5801 : : * This could of course be very wrong, but it should tend to be true
5802 : : * in well-written queries. We could consider examining the VALUES'
5803 : : * contents to get some real statistics; but that only works if the
5804 : : * entries are all constants, and it would be pretty expensive anyway.
5805 : : */
5806 : 1430 : stadistinct = -1.0; /* unique (and all non null) */
5807 : : }
5808 : : else
5809 : : {
5810 : : /*
5811 : : * We don't keep statistics for system columns, but in some cases we
5812 : : * can infer distinctness anyway.
5813 : : */
7362 5814 [ + - + + ]: 155832 : if (vardata->var && IsA(vardata->var, Var))
5815 : : {
5816 [ + + + ]: 145409 : switch (((Var *) vardata->var)->varattno)
5817 : : {
5818 : 495 : case SelfItemPointerAttributeNumber:
2807 5819 : 495 : stadistinct = -1.0; /* unique (and all non null) */
7362 5820 : 495 : break;
5821 : 1259 : case TableOidAttributeNumber:
7168 bruce@momjian.us 5822 : 1259 : stadistinct = 1.0; /* only 1 value */
7362 tgl@sss.pgh.pa.us 5823 : 1259 : break;
5824 : 143655 : default:
7168 bruce@momjian.us 5825 : 143655 : stadistinct = 0.0; /* means "unknown" */
7362 tgl@sss.pgh.pa.us 5826 : 143655 : break;
5827 : : }
5828 : : }
5829 : : else
7168 bruce@momjian.us 5830 : 10423 : stadistinct = 0.0; /* means "unknown" */
5831 : :
5832 : : /*
5833 : : * XXX consider using estimate_num_groups on expressions?
5834 : : */
5835 : : }
5836 : :
5837 : : /*
5838 : : * If there is a unique index or DISTINCT clause for the variable, assume
5839 : : * it is unique no matter what pg_statistic says; the statistics could be
5840 : : * out of date, or we might have found a partial unique index that proves
5841 : : * the var is unique for this query. However, we'd better still believe
5842 : : * the null-fraction statistic.
5843 : : */
5537 tgl@sss.pgh.pa.us 5844 [ + + ]: 577526 : if (vardata->isunique)
2807 5845 : 162694 : stadistinct = -1.0 * (1.0 - stanullfrac);
5846 : :
5847 : : /*
5848 : : * If we had an absolute estimate, use that.
5849 : : */
7362 5850 [ + + ]: 577526 : if (stadistinct > 0.0)
3181 5851 : 122259 : return clamp_row_est(stadistinct);
5852 : :
5853 : : /*
5854 : : * Otherwise we need to get the relation size; punt if not available.
5855 : : */
7362 5856 [ + + ]: 455267 : if (vardata->rel == NULL)
5857 : : {
4606 5858 : 200 : *isdefault = true;
7362 5859 : 200 : return DEFAULT_NUM_DISTINCT;
5860 : : }
5861 : 455067 : ntuples = vardata->rel->tuples;
5862 [ + + ]: 455067 : if (ntuples <= 0.0)
5863 : : {
4606 5864 : 28716 : *isdefault = true;
7362 5865 : 28716 : return DEFAULT_NUM_DISTINCT;
5866 : : }
5867 : :
5868 : : /*
5869 : : * If we had a relative estimate, use that.
5870 : : */
5871 [ + + ]: 426351 : if (stadistinct < 0.0)
3181 5872 : 314758 : return clamp_row_est(-stadistinct * ntuples);
5873 : :
5874 : : /*
5875 : : * With no data, estimate ndistinct = ntuples if the table is small, else
5876 : : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
5877 : : * that the behavior isn't discontinuous.
5878 : : */
7362 5879 [ + + ]: 111593 : if (ntuples < DEFAULT_NUM_DISTINCT)
3181 5880 : 52995 : return clamp_row_est(ntuples);
5881 : :
4606 5882 : 58598 : *isdefault = true;
7362 5883 : 58598 : return DEFAULT_NUM_DISTINCT;
5884 : : }
5885 : :
5886 : : /*
5887 : : * get_variable_range
5888 : : * Estimate the minimum and maximum value of the specified variable.
5889 : : * If successful, store values in *min and *max, and return true.
5890 : : * If no data available, return false.
5891 : : *
5892 : : * sortop is the "<" comparison operator to use. This should generally
5893 : : * be "<" not ">", as only the former is likely to be found in pg_statistic.
5894 : : * The collation must be specified too.
5895 : : */
5896 : : static bool
1409 5897 : 86688 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
5898 : : Oid sortop, Oid collation,
5899 : : Datum *min, Datum *max)
5900 : : {
5972 5901 : 86688 : Datum tmin = 0;
7362 5902 : 86688 : Datum tmax = 0;
5972 5903 : 86688 : bool have_data = false;
5904 : : int16 typLen;
5905 : : bool typByVal;
5906 : : Oid opfuncoid;
5907 : : FmgrInfo opproc;
5908 : : AttStatsSlot sslot;
5909 : :
5910 : : /*
5911 : : * XXX It's very tempting to try to use the actual column min and max, if
5912 : : * we can get them relatively-cheaply with an index probe. However, since
5913 : : * this function is called many times during join planning, that could
5914 : : * have unpleasant effects on planning speed. Need more investigation
5915 : : * before enabling this.
5916 : : */
5917 : : #ifdef NOT_USED
5918 : : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
5919 : : return true;
5920 : : #endif
5921 : :
7362 5922 [ + + ]: 86688 : if (!HeapTupleIsValid(vardata->statsTuple))
5923 : : {
5924 : : /* no stats available, so default result */
5925 : 17443 : return false;
5926 : : }
5927 : :
5928 : : /*
5929 : : * If we can't apply the sortop to the stats data, just fail. In
5930 : : * principle, if there's a histogram and no MCVs, we could return the
5931 : : * histogram endpoints without ever applying the sortop ... but it's
5932 : : * probably not worth trying, because whatever the caller wants to do with
5933 : : * the endpoints would likely fail the security check too.
5934 : : */
2536 peter_e@gmx.net 5935 [ - + ]: 69245 : if (!statistic_proc_security_check(vardata,
5936 : 69245 : (opfuncoid = get_opcode(sortop))))
2536 peter_e@gmx.net 5937 :UBC 0 : return false;
5938 : :
1409 tgl@sss.pgh.pa.us 5939 :CBC 69245 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
5940 : :
7362 5941 : 69245 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
5942 : :
5943 : : /*
5944 : : * If there is a histogram with the ordering we want, grab the first and
5945 : : * last values.
5946 : : */
2528 5947 [ + + ]: 69245 : if (get_attstatsslot(&sslot, vardata->statsTuple,
5948 : : STATISTIC_KIND_HISTOGRAM, sortop,
5949 : : ATTSTATSSLOT_VALUES))
5950 : : {
1409 5951 [ + - + - ]: 55424 : if (sslot.stacoll == collation && sslot.nvalues > 0)
5952 : : {
2528 5953 : 55424 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
5954 : 55424 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
5972 5955 : 55424 : have_data = true;
5956 : : }
2528 5957 : 55424 : free_attstatsslot(&sslot);
5958 : : }
5959 : :
5960 : : /*
5961 : : * Otherwise, if there is a histogram with some other ordering, scan it
5962 : : * and get the min and max values according to the ordering we want. This
5963 : : * of course may not find values that are really extremal according to our
5964 : : * ordering, but it beats ignoring available data.
5965 : : */
1409 5966 [ + + - + ]: 83066 : if (!have_data &&
5967 : 13821 : get_attstatsslot(&sslot, vardata->statsTuple,
5968 : : STATISTIC_KIND_HISTOGRAM, InvalidOid,
5969 : : ATTSTATSSLOT_VALUES))
5970 : : {
1409 tgl@sss.pgh.pa.us 5971 :UBC 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
5972 : : collation, typLen, typByVal,
5973 : : &tmin, &tmax, &have_data);
2528 5974 : 0 : free_attstatsslot(&sslot);
5975 : : }
5976 : :
5977 : : /*
5978 : : * If we have most-common-values info, look for extreme MCVs. This is
5979 : : * needed even if we also have a histogram, since the histogram excludes
5980 : : * the MCVs. However, if we *only* have MCVs and no histogram, we should
5981 : : * be pretty wary of deciding that that is a full representation of the
5982 : : * data. Proceed only if the MCVs represent the whole table (to within
5983 : : * roundoff error).
5984 : : */
2528 tgl@sss.pgh.pa.us 5985 [ + + ]:CBC 69245 : if (get_attstatsslot(&sslot, vardata->statsTuple,
5986 : : STATISTIC_KIND_MCV, InvalidOid,
926 5987 [ + + ]: 69245 : have_data ? ATTSTATSSLOT_VALUES :
5988 : : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
5989 : : {
5990 : 31944 : bool use_mcvs = have_data;
5991 : :
5992 [ + + ]: 31944 : if (!have_data)
5993 : : {
5994 : 13262 : double sumcommon = 0.0;
5995 : : double nullfrac;
5996 : : int i;
5997 : :
5998 [ + + ]: 73233 : for (i = 0; i < sslot.nnumbers; i++)
5999 : 59971 : sumcommon += sslot.numbers[i];
6000 : 13262 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6001 [ + + ]: 13262 : if (sumcommon + nullfrac > 0.99999)
6002 : 12613 : use_mcvs = true;
6003 : : }
6004 : :
6005 [ + + ]: 31944 : if (use_mcvs)
6006 : 31295 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6007 : : collation, typLen, typByVal,
6008 : : &tmin, &tmax, &have_data);
2528 6009 : 31944 : free_attstatsslot(&sslot);
6010 : : }
6011 : :
5972 6012 : 69245 : *min = tmin;
7362 6013 : 69245 : *max = tmax;
5972 6014 : 69245 : return have_data;
6015 : : }
6016 : :
6017 : : /*
6018 : : * get_stats_slot_range: scan sslot for min/max values
6019 : : *
6020 : : * Subroutine for get_variable_range: update min/max/have_data according
6021 : : * to what we find in the statistics array.
6022 : : */
6023 : : static void
1409 6024 : 31295 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6025 : : Oid collation, int16 typLen, bool typByVal,
6026 : : Datum *min, Datum *max, bool *p_have_data)
6027 : : {
6028 : 31295 : Datum tmin = *min;
6029 : 31295 : Datum tmax = *max;
6030 : 31295 : bool have_data = *p_have_data;
6031 : 31295 : bool found_tmin = false;
6032 : 31295 : bool found_tmax = false;
6033 : :
6034 : : /* Look up the comparison function, if we didn't already do so */
6035 [ + - ]: 31295 : if (opproc->fn_oid != opfuncoid)
6036 : 31295 : fmgr_info(opfuncoid, opproc);
6037 : :
6038 : : /* Scan all the slot's values */
6039 [ + + ]: 936169 : for (int i = 0; i < sslot->nvalues; i++)
6040 : : {
6041 [ + + ]: 904874 : if (!have_data)
6042 : : {
6043 : 12613 : tmin = tmax = sslot->values[i];
6044 : 12613 : found_tmin = found_tmax = true;
6045 : 12613 : *p_have_data = have_data = true;
6046 : 12613 : continue;
6047 : : }
6048 [ + + ]: 892261 : if (DatumGetBool(FunctionCall2Coll(opproc,
6049 : : collation,
6050 : 892261 : sslot->values[i], tmin)))
6051 : : {
6052 : 26808 : tmin = sslot->values[i];
6053 : 26808 : found_tmin = true;
6054 : : }
6055 [ + + ]: 892261 : if (DatumGetBool(FunctionCall2Coll(opproc,
6056 : : collation,
6057 : 892261 : tmax, sslot->values[i])))
6058 : : {
6059 : 38785 : tmax = sslot->values[i];
6060 : 38785 : found_tmax = true;
6061 : : }
6062 : : }
6063 : :
6064 : : /*
6065 : : * Copy the slot's values, if we found new extreme values.
6066 : : */
6067 [ + + ]: 31295 : if (found_tmin)
6068 : 26085 : *min = datumCopy(tmin, typByVal, typLen);
6069 [ + + ]: 31295 : if (found_tmax)
6070 : 14775 : *max = datumCopy(tmax, typByVal, typLen);
6071 : 31295 : }
6072 : :
6073 : :
6074 : : /*
6075 : : * get_actual_variable_range
6076 : : * Attempt to identify the current *actual* minimum and/or maximum
6077 : : * of the specified variable, by looking for a suitable btree index
6078 : : * and fetching its low and/or high values.
6079 : : * If successful, store values in *min and *max, and return true.
6080 : : * (Either pointer can be NULL if that endpoint isn't needed.)
6081 : : * If unsuccessful, return false.
6082 : : *
6083 : : * sortop is the "<" comparison operator to use.
6084 : : * collation is the required collation.
6085 : : */
6086 : : static bool
5214 6087 : 84974 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6088 : : Oid sortop, Oid collation,
6089 : : Datum *min, Datum *max)
6090 : : {
6091 : 84974 : bool have_data = false;
6092 : 84974 : RelOptInfo *rel = vardata->rel;
6093 : : RangeTblEntry *rte;
6094 : : ListCell *lc;
6095 : :
6096 : : /* No hope if no relation or it doesn't have indexes */
6097 [ + - + + ]: 84974 : if (rel == NULL || rel->indexlist == NIL)
6098 : 6397 : return false;
6099 : : /* If it has indexes it must be a plain relation */
6100 : 78577 : rte = root->simple_rte_array[rel->relid];
6101 [ - + ]: 78577 : Assert(rte->rtekind == RTE_RELATION);
6102 : :
6103 : : /* ignore partitioned tables. Any indexes here are not real indexes */
461 drowley@postgresql.o 6104 [ + + ]: 78577 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6105 : 414 : return false;
6106 : :
6107 : : /* Search through the indexes to see if any match our problem */
5214 tgl@sss.pgh.pa.us 6108 [ + - + + : 165901 : foreach(lc, rel->indexlist)
+ + ]
6109 : : {
6110 : 142439 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6111 : : ScanDirection indexscandir;
6112 : :
6113 : : /* Ignore non-btree indexes */
6114 [ - + ]: 142439 : if (index->relam != BTREE_AM_OID)
5214 tgl@sss.pgh.pa.us 6115 :UBC 0 : continue;
6116 : :
6117 : : /*
6118 : : * Ignore partial indexes --- we only want stats that cover the entire
6119 : : * relation.
6120 : : */
5214 tgl@sss.pgh.pa.us 6121 [ + + ]:CBC 142439 : if (index->indpred != NIL)
6122 : 144 : continue;
6123 : :
6124 : : /*
6125 : : * The index list might include hypothetical indexes inserted by a
6126 : : * get_relation_info hook --- don't try to access them.
6127 : : */
4806 6128 [ - + ]: 142295 : if (index->hypothetical)
5214 tgl@sss.pgh.pa.us 6129 :UBC 0 : continue;
6130 : :
6131 : : /*
6132 : : * The first index column must match the desired variable, sortop, and
6133 : : * collation --- but we can use a descending-order index.
6134 : : */
1409 tgl@sss.pgh.pa.us 6135 [ + + ]:CBC 142295 : if (collation != index->indexcollations[0])
6136 : 21623 : continue; /* test first 'cause it's cheapest */
5214 6137 [ + + ]: 120672 : if (!match_index_to_operand(vardata->var, 0, index))
6138 : 65971 : continue;
4885 6139 [ + - - ]: 54701 : switch (get_op_opfamily_strategy(sortop, index->sortopfamily[0]))
6140 : : {
6141 : 54701 : case BTLessStrategyNumber:
6142 [ - + ]: 54701 : if (index->reverse_sort[0])
4885 tgl@sss.pgh.pa.us 6143 :UBC 0 : indexscandir = BackwardScanDirection;
6144 : : else
4885 tgl@sss.pgh.pa.us 6145 :CBC 54701 : indexscandir = ForwardScanDirection;
6146 : 54701 : break;
4885 tgl@sss.pgh.pa.us 6147 :UBC 0 : case BTGreaterStrategyNumber:
6148 [ # # ]: 0 : if (index->reverse_sort[0])
6149 : 0 : indexscandir = ForwardScanDirection;
6150 : : else
6151 : 0 : indexscandir = BackwardScanDirection;
6152 : 0 : break;
6153 : 0 : default:
6154 : : /* index doesn't match the sortop */
6155 : 0 : continue;
6156 : : }
6157 : :
6158 : : /*
6159 : : * Found a suitable index to extract data from. Set up some data that
6160 : : * can be used by both invocations of get_actual_variable_endpoint.
6161 : : */
6162 : : {
6163 : : MemoryContext tmpcontext;
6164 : : MemoryContext oldcontext;
6165 : : Relation heapRel;
6166 : : Relation indexRel;
6167 : : TupleTableSlot *slot;
6168 : : int16 typLen;
6169 : : bool typByVal;
6170 : : ScanKeyData scankeys[1];
6171 : :
6172 : : /* Make sure any cruft gets recycled when we're done */
1738 tgl@sss.pgh.pa.us 6173 :CBC 54701 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
6174 : : "get_actual_variable_range workspace",
6175 : : ALLOCSET_DEFAULT_SIZES);
5214 6176 : 54701 : oldcontext = MemoryContextSwitchTo(tmpcontext);
6177 : :
6178 : : /*
6179 : : * Open the table and index so we can read from them. We should
6180 : : * already have some type of lock on each.
6181 : : */
1910 andres@anarazel.de 6182 : 54701 : heapRel = table_open(rte->relid, NoLock);
1837 tgl@sss.pgh.pa.us 6183 : 54701 : indexRel = index_open(index->indexoid, NoLock);
6184 : :
6185 : : /* build some stuff needed for indexscan execution */
1861 andres@anarazel.de 6186 : 54701 : slot = table_slot_create(heapRel, NULL);
5214 tgl@sss.pgh.pa.us 6187 : 54701 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6188 : :
6189 : : /* set up an IS NOT NULL scan key so that we ignore nulls */
6190 : 54701 : ScanKeyEntryInitialize(&scankeys[0],
6191 : : SK_ISNULL | SK_SEARCHNOTNULL,
6192 : : 1, /* index col to scan */
6193 : : InvalidStrategy, /* no strategy */
6194 : : InvalidOid, /* no strategy subtype */
6195 : : InvalidOid, /* no collation */
6196 : : InvalidOid, /* no reg proc for this */
6197 : : (Datum) 0); /* constant */
6198 : :
6199 : : /* If min is requested ... */
6200 [ + + ]: 54701 : if (min)
6201 : : {
1738 6202 : 31522 : have_data = get_actual_variable_endpoint(heapRel,
6203 : : indexRel,
6204 : : indexscandir,
6205 : : scankeys,
6206 : : typLen,
6207 : : typByVal,
6208 : : slot,
6209 : : oldcontext,
6210 : : min);
6211 : : }
6212 : : else
6213 : : {
6214 : : /* If min not requested, still want to fetch max */
6215 : 23179 : have_data = true;
6216 : : }
6217 : :
6218 : : /* If max is requested, and we didn't already fail ... */
5214 6219 [ + + + - ]: 54701 : if (max && have_data)
6220 : : {
6221 : : /* scan in the opposite direction; all else is the same */
1738 6222 : 23775 : have_data = get_actual_variable_endpoint(heapRel,
6223 : : indexRel,
6224 : 23775 : -indexscandir,
6225 : : scankeys,
6226 : : typLen,
6227 : : typByVal,
6228 : : slot,
6229 : : oldcontext,
6230 : : max);
6231 : : }
6232 : :
6233 : : /* Clean everything up */
5214 6234 : 54701 : ExecDropSingleTupleTableSlot(slot);
6235 : :
1837 6236 : 54701 : index_close(indexRel, NoLock);
1910 andres@anarazel.de 6237 : 54701 : table_close(heapRel, NoLock);
6238 : :
5214 tgl@sss.pgh.pa.us 6239 : 54701 : MemoryContextSwitchTo(oldcontext);
1738 6240 : 54701 : MemoryContextDelete(tmpcontext);
6241 : :
6242 : : /* And we're done */
5214 6243 : 54701 : break;
6244 : : }
6245 : : }
6246 : :
6247 : 78163 : return have_data;
6248 : : }
6249 : :
6250 : : /*
6251 : : * Get one endpoint datum (min or max depending on indexscandir) from the
6252 : : * specified index. Return true if successful, false if not.
6253 : : * On success, endpoint value is stored to *endpointDatum (and copied into
6254 : : * outercontext).
6255 : : *
6256 : : * scankeys is a 1-element scankey array set up to reject nulls.
6257 : : * typLen/typByVal describe the datatype of the index's first column.
6258 : : * tableslot is a slot suitable to hold table tuples, in case we need
6259 : : * to probe the heap.
6260 : : * (We could compute these values locally, but that would mean computing them
6261 : : * twice when get_actual_variable_range needs both the min and the max.)
6262 : : *
6263 : : * Failure occurs either when the index is empty, or we decide that it's
6264 : : * taking too long to find a suitable tuple.
6265 : : */
6266 : : static bool
1738 6267 : 55297 : get_actual_variable_endpoint(Relation heapRel,
6268 : : Relation indexRel,
6269 : : ScanDirection indexscandir,
6270 : : ScanKey scankeys,
6271 : : int16 typLen,
6272 : : bool typByVal,
6273 : : TupleTableSlot *tableslot,
6274 : : MemoryContext outercontext,
6275 : : Datum *endpointDatum)
6276 : : {
6277 : 55297 : bool have_data = false;
6278 : : SnapshotData SnapshotNonVacuumable;
6279 : : IndexScanDesc index_scan;
6280 : 55297 : Buffer vmbuffer = InvalidBuffer;
509 6281 : 55297 : BlockNumber last_heap_block = InvalidBlockNumber;
6282 : 55297 : int n_visited_heap_pages = 0;
6283 : : ItemPointer tid;
6284 : : Datum values[INDEX_MAX_KEYS];
6285 : : bool isnull[INDEX_MAX_KEYS];
6286 : : MemoryContext oldcontext;
6287 : :
6288 : : /*
6289 : : * We use the index-only-scan machinery for this. With mostly-static
6290 : : * tables that's a win because it avoids a heap visit. It's also a win
6291 : : * for dynamic data, but the reason is less obvious; read on for details.
6292 : : *
6293 : : * In principle, we should scan the index with our current active
6294 : : * snapshot, which is the best approximation we've got to what the query
6295 : : * will see when executed. But that won't be exact if a new snap is taken
6296 : : * before running the query, and it can be very expensive if a lot of
6297 : : * recently-dead or uncommitted rows exist at the beginning or end of the
6298 : : * index (because we'll laboriously fetch each one and reject it).
6299 : : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
6300 : : * and uncommitted rows as well as normal visible rows. On the other
6301 : : * hand, it will reject known-dead rows, and thus not give a bogus answer
6302 : : * when the extreme value has been deleted (unless the deletion was quite
6303 : : * recent); that case motivates not using SnapshotAny here.
6304 : : *
6305 : : * A crucial point here is that SnapshotNonVacuumable, with
6306 : : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
6307 : : * condition that the indexscan will use to decide that index entries are
6308 : : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
6309 : : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
6310 : : * have to continue scanning past it, we know that the indexscan will mark
6311 : : * that index entry killed. That means that the next
6312 : : * get_actual_variable_endpoint() call will not have to re-consider that
6313 : : * index entry. In this way we avoid repetitive work when this function
6314 : : * is used a lot during planning.
6315 : : *
6316 : : * But using SnapshotNonVacuumable creates a hazard of its own. In a
6317 : : * recently-created index, some index entries may point at "broken" HOT
6318 : : * chains in which not all the tuple versions contain data matching the
6319 : : * index entry. The live tuple version(s) certainly do match the index,
6320 : : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
6321 : : * don't match. Hence, if we took data from the selected heap tuple, we
6322 : : * might get a bogus answer that's not close to the index extremal value,
6323 : : * or could even be NULL. We avoid this hazard because we take the data
6324 : : * from the index entry not the heap.
6325 : : *
6326 : : * Despite all this care, there are situations where we might find many
6327 : : * non-visible tuples near the end of the index. We don't want to expend
6328 : : * a huge amount of time here, so we give up once we've read too many heap
6329 : : * pages. When we fail for that reason, the caller will end up using
6330 : : * whatever extremal value is recorded in pg_statistic.
6331 : : */
1341 andres@anarazel.de 6332 : 55297 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
6333 : : GlobalVisTestFor(heapRel));
6334 : :
1738 tgl@sss.pgh.pa.us 6335 : 55297 : index_scan = index_beginscan(heapRel, indexRel,
6336 : : &SnapshotNonVacuumable,
6337 : : 1, 0);
6338 : : /* Set it up for index-only scan */
6339 : 55297 : index_scan->xs_want_itup = true;
6340 : 55297 : index_rescan(index_scan, scankeys, 1, NULL, 0);
6341 : :
6342 : : /* Fetch first/next tuple in specified direction */
6343 [ + - ]: 65324 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
6344 : : {
509 6345 : 65324 : BlockNumber block = ItemPointerGetBlockNumber(tid);
6346 : :
1738 6347 [ + + ]: 65324 : if (!VM_ALL_VISIBLE(heapRel,
6348 : : block,
6349 : : &vmbuffer))
6350 : : {
6351 : : /* Rats, we have to visit the heap to check visibility */
6352 [ + + ]: 45321 : if (!index_fetch_heap(index_scan, tableslot))
6353 : : {
6354 : : /*
6355 : : * No visible tuple for this index entry, so we need to
6356 : : * advance to the next entry. Before doing so, count heap
6357 : : * page fetches and give up if we've done too many.
6358 : : *
6359 : : * We don't charge a page fetch if this is the same heap page
6360 : : * as the previous tuple. This is on the conservative side,
6361 : : * since other recently-accessed pages are probably still in
6362 : : * buffers too; but it's good enough for this heuristic.
6363 : : */
6364 : : #define VISITED_PAGES_LIMIT 100
6365 : :
509 6366 [ + + ]: 10027 : if (block != last_heap_block)
6367 : : {
6368 : 715 : last_heap_block = block;
6369 : 715 : n_visited_heap_pages++;
6370 [ - + ]: 715 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
509 tgl@sss.pgh.pa.us 6371 :UBC 0 : break;
6372 : : }
6373 : :
1738 tgl@sss.pgh.pa.us 6374 :CBC 10027 : continue; /* no visible tuple, try next index entry */
6375 : : }
6376 : :
6377 : : /* We don't actually need the heap tuple for anything */
6378 : 35294 : ExecClearTuple(tableslot);
6379 : :
6380 : : /*
6381 : : * We don't care whether there's more than one visible tuple in
6382 : : * the HOT chain; if any are visible, that's good enough.
6383 : : */
6384 : : }
6385 : :
6386 : : /*
6387 : : * We expect that btree will return data in IndexTuple not HeapTuple
6388 : : * format. It's not lossy either.
6389 : : */
6390 [ - + ]: 55297 : if (!index_scan->xs_itup)
1738 tgl@sss.pgh.pa.us 6391 [ # # ]:UBC 0 : elog(ERROR, "no data returned for index-only scan");
1738 tgl@sss.pgh.pa.us 6392 [ - + ]:CBC 55297 : if (index_scan->xs_recheck)
1738 tgl@sss.pgh.pa.us 6393 [ # # ]:UBC 0 : elog(ERROR, "unexpected recheck indication from btree");
6394 : :
6395 : : /* OK to deconstruct the index tuple */
1738 tgl@sss.pgh.pa.us 6396 :CBC 55297 : index_deform_tuple(index_scan->xs_itup,
6397 : : index_scan->xs_itupdesc,
6398 : : values, isnull);
6399 : :
6400 : : /* Shouldn't have got a null, but be careful */
6401 [ - + ]: 55297 : if (isnull[0])
1738 tgl@sss.pgh.pa.us 6402 [ # # ]:UBC 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
6403 : : RelationGetRelationName(indexRel));
6404 : :
6405 : : /* Copy the index column value out to caller's context */
1738 tgl@sss.pgh.pa.us 6406 :CBC 55297 : oldcontext = MemoryContextSwitchTo(outercontext);
6407 : 55297 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
6408 : 55297 : MemoryContextSwitchTo(oldcontext);
6409 : 55297 : have_data = true;
6410 : 55297 : break;
6411 : : }
6412 : :
6413 [ + + ]: 55297 : if (vmbuffer != InvalidBuffer)
6414 : 49585 : ReleaseBuffer(vmbuffer);
6415 : 55297 : index_endscan(index_scan);
6416 : :
6417 : 55297 : return have_data;
6418 : : }
6419 : :
6420 : : /*
6421 : : * find_join_input_rel
6422 : : * Look up the input relation for a join.
6423 : : *
6424 : : * We assume that the input relation's RelOptInfo must have been constructed
6425 : : * already.
6426 : : */
6427 : : static RelOptInfo *
4610 6428 : 3524 : find_join_input_rel(PlannerInfo *root, Relids relids)
6429 : : {
6430 : 3524 : RelOptInfo *rel = NULL;
6431 : :
138 drowley@postgresql.o 6432 [ + - ]:GNC 3524 : if (!bms_is_empty(relids))
6433 : : {
6434 : : int relid;
6435 : :
6436 [ + + ]: 3524 : if (bms_get_singleton_member(relids, &relid))
6437 : 3372 : rel = find_base_rel(root, relid);
6438 : : else
4610 tgl@sss.pgh.pa.us 6439 :CBC 152 : rel = find_join_rel(root, relids);
6440 : : }
6441 : :
6442 [ - + ]: 3524 : if (rel == NULL)
4610 tgl@sss.pgh.pa.us 6443 [ # # ]:UBC 0 : elog(ERROR, "could not find RelOptInfo for given relids");
6444 : :
4610 tgl@sss.pgh.pa.us 6445 :CBC 3524 : return rel;
6446 : : }
6447 : :
6448 : :
6449 : : /*-------------------------------------------------------------------------
6450 : : *
6451 : : * Index cost estimation functions
6452 : : *
6453 : : *-------------------------------------------------------------------------
6454 : : */
6455 : :
6456 : : /*
6457 : : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
6458 : : */
6459 : : List *
1885 6460 : 322598 : get_quals_from_indexclauses(List *indexclauses)
6461 : : {
1891 6462 : 322598 : List *result = NIL;
6463 : : ListCell *lc;
6464 : :
6465 [ + + + + : 567480 : foreach(lc, indexclauses)
+ + ]
6466 : : {
6467 : 244882 : IndexClause *iclause = lfirst_node(IndexClause, lc);
6468 : : ListCell *lc2;
6469 : :
1886 6470 [ + - + + : 491172 : foreach(lc2, iclause->indexquals)
+ + ]
6471 : : {
6472 : 246290 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6473 : :
6474 : 246290 : result = lappend(result, rinfo);
6475 : : }
6476 : : }
1891 6477 : 322598 : return result;
6478 : : }
6479 : :
6480 : : /*
6481 : : * Compute the total evaluation cost of the comparison operands in a list
6482 : : * of index qual expressions. Since we know these will be evaluated just
6483 : : * once per scan, there's no need to distinguish startup from per-row cost.
6484 : : *
6485 : : * This can be used either on the result of get_quals_from_indexclauses(),
6486 : : * or directly on an indexorderbys list. In both cases, we expect that the
6487 : : * index key expression is on the left side of binary clauses.
6488 : : */
6489 : : Cost
1885 6490 : 638891 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
6491 : : {
3330 6492 : 638891 : Cost qual_arg_cost = 0;
6493 : : ListCell *lc;
6494 : :
1885 6495 [ + + + + : 885412 : foreach(lc, indexquals)
+ + ]
6496 : : {
3330 6497 : 246521 : Expr *clause = (Expr *) lfirst(lc);
6498 : : Node *other_operand;
6499 : : QualCost index_qual_cost;
6500 : :
6501 : : /*
6502 : : * Index quals will have RestrictInfos, indexorderbys won't. Look
6503 : : * through RestrictInfo if present.
6504 : : */
1885 6505 [ + + ]: 246521 : if (IsA(clause, RestrictInfo))
6506 : 246284 : clause = ((RestrictInfo *) clause)->clause;
6507 : :
3330 6508 [ + + ]: 246521 : if (IsA(clause, OpExpr))
6509 : : {
1885 6510 : 241289 : OpExpr *op = (OpExpr *) clause;
6511 : :
6512 : 241289 : other_operand = (Node *) lsecond(op->args);
6513 : : }
6514 [ + + ]: 5232 : else if (IsA(clause, RowCompareExpr))
6515 : : {
6516 : 66 : RowCompareExpr *rc = (RowCompareExpr *) clause;
6517 : :
6518 : 66 : other_operand = (Node *) rc->rargs;
6519 : : }
6520 [ + + ]: 5166 : else if (IsA(clause, ScalarArrayOpExpr))
6521 : : {
6522 : 3763 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6523 : :
6524 : 3763 : other_operand = (Node *) lsecond(saop->args);
6525 : : }
6526 [ + - ]: 1403 : else if (IsA(clause, NullTest))
6527 : : {
6528 : 1403 : other_operand = NULL;
6529 : : }
6530 : : else
6531 : : {
1885 tgl@sss.pgh.pa.us 6532 [ # # ]:UBC 0 : elog(ERROR, "unsupported indexqual type: %d",
6533 : : (int) nodeTag(clause));
6534 : : other_operand = NULL; /* keep compiler quiet */
6535 : : }
6536 : :
3330 tgl@sss.pgh.pa.us 6537 :CBC 246521 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
6538 : 246521 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
6539 : : }
6540 : 638891 : return qual_arg_cost;
6541 : : }
6542 : :
6543 : : void
6888 6544 : 316299 : genericcostestimate(PlannerInfo *root,
6545 : : IndexPath *path,
6546 : : double loop_count,
6547 : : GenericCosts *costs)
6548 : : {
4495 6549 : 316299 : IndexOptInfo *index = path->indexinfo;
1885 6550 : 316299 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
4495 6551 : 316299 : List *indexOrderBys = path->indexorderbys;
6552 : : Cost indexStartupCost;
6553 : : Cost indexTotalCost;
6554 : : Selectivity indexSelectivity;
6555 : : double indexCorrelation;
6556 : : double numIndexPages;
6557 : : double numIndexTuples;
6558 : : double spc_random_page_cost;
6559 : : double num_sa_scans;
6560 : : double num_outer_scans;
6561 : : double num_scans;
6562 : : double qual_op_cost;
6563 : : double qual_arg_cost;
6564 : : List *selectivityQuals;
6565 : : ListCell *l;
6566 : :
6567 : : /*
6568 : : * If the index is partial, AND the index predicate with the explicitly
6569 : : * given indexquals to produce a more accurate idea of the index
6570 : : * selectivity.
6571 : : */
1885 6572 : 316299 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
6573 : :
6574 : : /*
6575 : : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
6576 : : * just assume that the number of index descents is the number of distinct
6577 : : * combinations of array elements from all of the scan's SAOP clauses.
6578 : : */
8 pg@bowt.ie 6579 :GNC 316299 : num_sa_scans = costs->num_sa_scans;
6580 [ + + ]: 316299 : if (num_sa_scans < 1)
6581 : : {
6582 : 3728 : num_sa_scans = 1;
6583 [ + + + + : 7770 : foreach(l, indexQuals)
+ + ]
6584 : : {
6585 : 4042 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
6586 : :
6587 [ + + ]: 4042 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
6588 : : {
6589 : 13 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
6590 : 13 : double alength = estimate_array_length(root, lsecond(saop->args));
6591 : :
6592 [ + - ]: 13 : if (alength > 1)
6593 : 13 : num_sa_scans *= alength;
6594 : : }
6595 : : }
6596 : : }
6597 : :
6598 : : /* Estimate the fraction of main-table tuples that will be visited */
4111 tgl@sss.pgh.pa.us 6599 :CBC 316299 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
6600 : 316299 : index->rel->relid,
6601 : : JOIN_INNER,
6602 : : NULL);
6603 : :
6604 : : /*
6605 : : * If caller didn't give us an estimate, estimate the number of index
6606 : : * tuples that will be visited. We do it in this rather peculiar-looking
6607 : : * way in order to get the right answer for partial indexes.
6608 : : */
6609 : 316299 : numIndexTuples = costs->numIndexTuples;
6880 6610 [ + + ]: 316299 : if (numIndexTuples <= 0.0)
6611 : : {
4111 6612 : 31108 : numIndexTuples = indexSelectivity * index->rel->tuples;
6613 : :
6614 : : /*
6615 : : * The above calculation counts all the tuples visited across all
6616 : : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
6617 : : * average per-indexscan number, so adjust. This is a handy place to
6618 : : * round to integer, too. (If caller supplied tuple estimate, it's
6619 : : * responsible for handling these considerations.)
6620 : : */
6330 6621 : 31108 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
6622 : : }
6623 : :
6624 : : /*
6625 : : * We can bound the number of tuples by the index size in any case. Also,
6626 : : * always estimate at least one tuple is touched, even when
6627 : : * indexSelectivity estimate is tiny.
6628 : : */
6880 6629 [ + + ]: 316299 : if (numIndexTuples > index->tuples)
6630 : 3317 : numIndexTuples = index->tuples;
8771 6631 [ + + ]: 316299 : if (numIndexTuples < 1.0)
6632 : 31596 : numIndexTuples = 1.0;
6633 : :
6634 : : /*
6635 : : * Estimate the number of index pages that will be retrieved.
6636 : : *
6637 : : * We use the simplistic method of taking a pro-rata fraction of the total
6638 : : * number of index pages. In effect, this counts only leaf pages and not
6639 : : * any overhead such as index metapage or upper tree levels.
6640 : : *
6641 : : * In practice access to upper index levels is often nearly free because
6642 : : * those tend to stay in cache under load; moreover, the cost involved is
6643 : : * highly dependent on index type. We therefore ignore such costs here
6644 : : * and leave it to the caller to add a suitable charge if needed.
6645 : : */
6522 6646 [ + + + + ]: 316299 : if (index->pages > 1 && index->tuples > 1)
6647 : 290780 : numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
6648 : : else
8771 6649 : 25519 : numIndexPages = 1.0;
6650 : :
6651 : : /* fetch estimated page cost for tablespace containing index */
5213 rhaas@postgresql.org 6652 : 316299 : get_tablespace_page_costs(index->reltablespace,
6653 : : &spc_random_page_cost,
6654 : : NULL);
6655 : :
6656 : : /*
6657 : : * Now compute the disk access costs.
6658 : : *
6659 : : * The above calculations are all per-index-scan. However, if we are in a
6660 : : * nestloop inner scan, we can expect the scan to be repeated (with
6661 : : * different search keys) for each row of the outer relation. Likewise,
6662 : : * ScalarArrayOpExpr quals result in multiple index scans. This creates
6663 : : * the potential for cache effects to reduce the number of disk page
6664 : : * fetches needed. We want to estimate the average per-scan I/O cost in
6665 : : * the presence of caching.
6666 : : *
6667 : : * We use the Mackert-Lohman formula (see costsize.c for details) to
6668 : : * estimate the total number of page fetches that occur. While this
6669 : : * wasn't what it was designed for, it seems a reasonable model anyway.
6670 : : * Note that we are counting pages not tuples anymore, so we take N = T =
6671 : : * index size, as if there were one "tuple" per page.
6672 : : */
4461 tgl@sss.pgh.pa.us 6673 : 316299 : num_outer_scans = loop_count;
6674 : 316299 : num_scans = num_sa_scans * num_outer_scans;
6675 : :
6497 6676 [ + + ]: 316299 : if (num_scans > 1)
6677 : : {
6678 : : double pages_fetched;
6679 : :
6680 : : /* total page fetches ignoring cache effects */
6522 6681 : 32232 : pages_fetched = numIndexPages * num_scans;
6682 : :
6683 : : /* use Mackert and Lohman formula to adjust for cache effects */
6684 : 32232 : pages_fetched = index_pages_fetched(pages_fetched,
6685 : : index->pages,
6417 6686 : 32232 : (double) index->pages,
6687 : : root);
6688 : :
6689 : : /*
6690 : : * Now compute the total disk access cost, and then report a pro-rated
6691 : : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
6692 : : * since that's internal to the indexscan.)
6693 : : */
4111 6694 : 32232 : indexTotalCost = (pages_fetched * spc_random_page_cost)
6695 : : / num_outer_scans;
6696 : : }
6697 : : else
6698 : : {
6699 : : /*
6700 : : * For a single index scan, we just charge spc_random_page_cost per
6701 : : * page touched.
6702 : : */
6703 : 284067 : indexTotalCost = numIndexPages * spc_random_page_cost;
6704 : : }
6705 : :
6706 : : /*
6707 : : * CPU cost: any complex expressions in the indexquals will need to be
6708 : : * evaluated once at the start of the scan to reduce them to runtime keys
6709 : : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
6710 : : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
6711 : : * indexqual operator. Because we have numIndexTuples as a per-scan
6712 : : * number, we have to multiply by num_sa_scans to get the correct result
6713 : : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
6714 : : * ORDER BY expressions.
6715 : : *
6716 : : * Note: this neglects the possible costs of rechecking lossy operators.
6717 : : * Detecting that that might be needed seems more expensive than it's
6718 : : * worth, though, considering all the other inaccuracies here ...
6719 : : */
1885 6720 : 316299 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
6721 : 316299 : index_other_operands_eval_cost(root, indexOrderBys);
4882 6722 : 316299 : qual_op_cost = cpu_operator_cost *
6723 : 316299 : (list_length(indexQuals) + list_length(indexOrderBys));
6724 : :
4111 6725 : 316299 : indexStartupCost = qual_arg_cost;
6726 : 316299 : indexTotalCost += qual_arg_cost;
6727 : 316299 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
6728 : :
6729 : : /*
6730 : : * Generic assumption about index correlation: there isn't any.
6731 : : */
6732 : 316299 : indexCorrelation = 0.0;
6733 : :
6734 : : /*
6735 : : * Return everything to caller.
6736 : : */
6737 : 316299 : costs->indexStartupCost = indexStartupCost;
6738 : 316299 : costs->indexTotalCost = indexTotalCost;
6739 : 316299 : costs->indexSelectivity = indexSelectivity;
6740 : 316299 : costs->indexCorrelation = indexCorrelation;
6741 : 316299 : costs->numIndexPages = numIndexPages;
6742 : 316299 : costs->numIndexTuples = numIndexTuples;
6743 : 316299 : costs->spc_random_page_cost = spc_random_page_cost;
6744 : 316299 : costs->num_sa_scans = num_sa_scans;
6745 : 316299 : }
6746 : :
6747 : : /*
6748 : : * If the index is partial, add its predicate to the given qual list.
6749 : : *
6750 : : * ANDing the index predicate with the explicitly given indexquals produces
6751 : : * a more accurate idea of the index's selectivity. However, we need to be
6752 : : * careful not to insert redundant clauses, because clauselist_selectivity()
6753 : : * is easily fooled into computing a too-low selectivity estimate. Our
6754 : : * approach is to add only the predicate clause(s) that cannot be proven to
6755 : : * be implied by the given indexquals. This successfully handles cases such
6756 : : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
6757 : : * There are many other cases where we won't detect redundancy, leading to a
6758 : : * too-low selectivity estimate, which will bias the system in favor of using
6759 : : * partial indexes where possible. That is not necessarily bad though.
6760 : : *
6761 : : * Note that indexQuals contains RestrictInfo nodes while the indpred
6762 : : * does not, so the output list will be mixed. This is OK for both
6763 : : * predicate_implied_by() and clauselist_selectivity(), but might be
6764 : : * problematic if the result were passed to other things.
6765 : : */
6766 : : List *
1885 6767 : 522044 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
6768 : : {
4111 6769 : 522044 : List *predExtraQuals = NIL;
6770 : : ListCell *lc;
6771 : :
6772 [ + + ]: 522044 : if (index->indpred == NIL)
6773 : 521038 : return indexQuals;
6774 : :
6775 [ + - + + : 2018 : foreach(lc, index->indpred)
+ + ]
6776 : : {
6777 : 1012 : Node *predQual = (Node *) lfirst(lc);
6778 : 1012 : List *oneQual = list_make1(predQual);
6779 : :
2496 rhaas@postgresql.org 6780 [ + + ]: 1012 : if (!predicate_implied_by(oneQual, indexQuals, false))
4111 tgl@sss.pgh.pa.us 6781 : 906 : predExtraQuals = list_concat(predExtraQuals, oneQual);
6782 : : }
6783 : 1006 : return list_concat(predExtraQuals, indexQuals);
6784 : : }
6785 : :
6786 : :
6787 : : void
3010 6788 : 312571 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
6789 : : Cost *indexStartupCost, Cost *indexTotalCost,
6790 : : Selectivity *indexSelectivity, double *indexCorrelation,
6791 : : double *indexPages)
6792 : : {
4495 6793 : 312571 : IndexOptInfo *index = path->indexinfo;
638 peter@eisentraut.org 6794 : 312571 : GenericCosts costs = {0};
6795 : : Oid relid;
6796 : : AttrNumber colnum;
6797 : 312571 : VariableStatData vardata = {0};
6798 : : double numIndexTuples;
6799 : : Cost descentCost;
6800 : : List *indexBoundQuals;
6801 : : int indexcol;
6802 : : bool eqQualHere;
6803 : : bool found_saop;
6804 : : bool found_is_null_op;
6805 : : double num_sa_scans;
6806 : : ListCell *lc;
6807 : :
6808 : : /*
6809 : : * For a btree scan, only leading '=' quals plus inequality quals for the
6810 : : * immediately next attribute contribute to index selectivity (these are
6811 : : * the "boundary quals" that determine the starting and stopping points of
6812 : : * the index scan). Additional quals can suppress visits to the heap, so
6813 : : * it's OK to count them in indexSelectivity, but they should not count
6814 : : * for estimating numIndexTuples. So we must examine the given indexquals
6815 : : * to find out which ones count as boundary quals. We rely on the
6816 : : * knowledge that they are given in index column order.
6817 : : *
6818 : : * For a RowCompareExpr, we consider only the first column, just as
6819 : : * rowcomparesel() does.
6820 : : *
6821 : : * If there's a ScalarArrayOpExpr in the quals, we'll actually perform up
6822 : : * to N index descents (not just one), but the ScalarArrayOpExpr's
6823 : : * operator can be considered to act the same as it normally does.
6824 : : */
6880 tgl@sss.pgh.pa.us 6825 : 312571 : indexBoundQuals = NIL;
4495 6826 : 312571 : indexcol = 0;
6880 6827 : 312571 : eqQualHere = false;
6715 6828 : 312571 : found_saop = false;
5217 6829 : 312571 : found_is_null_op = false;
6330 6830 : 312571 : num_sa_scans = 1;
1885 6831 [ + + + + : 535061 : foreach(lc, path->indexclauses)
+ + ]
6832 : : {
6833 : 233418 : IndexClause *iclause = lfirst_node(IndexClause, lc);
6834 : : ListCell *lc2;
6835 : :
6836 [ + + ]: 233418 : if (indexcol != iclause->indexcol)
6837 : : {
6838 : : /* Beginning of a new column's quals */
4495 6839 [ + + ]: 38774 : if (!eqQualHere)
6840 : 10347 : break; /* done if no '=' qual for indexcol */
6841 : 28427 : eqQualHere = false;
6842 : 28427 : indexcol++;
1885 6843 [ + + ]: 28427 : if (indexcol != iclause->indexcol)
4495 6844 : 581 : break; /* no quals at all for indexcol */
6845 : : }
6846 : :
6847 : : /* Examine each indexqual associated with this index clause */
1885 6848 [ + - + + : 446308 : foreach(lc2, iclause->indexquals)
+ + ]
6849 : : {
6850 : 223818 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
6851 : 223818 : Expr *clause = rinfo->clause;
6852 : 223818 : Oid clause_op = InvalidOid;
6853 : : int op_strategy;
6854 : :
6855 [ + + ]: 223818 : if (IsA(clause, OpExpr))
6856 : : {
6857 : 218998 : OpExpr *op = (OpExpr *) clause;
6858 : :
6859 : 218998 : clause_op = op->opno;
6860 : : }
6861 [ + + ]: 4820 : else if (IsA(clause, RowCompareExpr))
6862 : : {
6863 : 66 : RowCompareExpr *rc = (RowCompareExpr *) clause;
6864 : :
6865 : 66 : clause_op = linitial_oid(rc->opnos);
6866 : : }
6867 [ + + ]: 4754 : else if (IsA(clause, ScalarArrayOpExpr))
6868 : : {
6869 : 3648 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
6870 : 3648 : Node *other_operand = (Node *) lsecond(saop->args);
101 tgl@sss.pgh.pa.us 6871 :GNC 3648 : double alength = estimate_array_length(root, other_operand);
6872 : :
1885 tgl@sss.pgh.pa.us 6873 :CBC 3648 : clause_op = saop->opno;
6874 : 3648 : found_saop = true;
6875 : : /* estimate SA descents by indexBoundQuals only */
6876 [ + + ]: 3648 : if (alength > 1)
6877 : 3600 : num_sa_scans *= alength;
6878 : : }
6879 [ + - ]: 1106 : else if (IsA(clause, NullTest))
6880 : : {
6881 : 1106 : NullTest *nt = (NullTest *) clause;
6882 : :
6883 [ + + ]: 1106 : if (nt->nulltesttype == IS_NULL)
6884 : : {
6885 : 102 : found_is_null_op = true;
6886 : : /* IS NULL is like = for selectivity purposes */
6887 : 102 : eqQualHere = true;
6888 : : }
6889 : : }
6890 : : else
1885 tgl@sss.pgh.pa.us 6891 [ # # ]:UBC 0 : elog(ERROR, "unsupported indexqual type: %d",
6892 : : (int) nodeTag(clause));
6893 : :
6894 : : /* check for equality operator */
1885 tgl@sss.pgh.pa.us 6895 [ + + ]:CBC 223818 : if (OidIsValid(clause_op))
6896 : : {
6897 : 222712 : op_strategy = get_op_opfamily_strategy(clause_op,
6898 : 222712 : index->opfamily[indexcol]);
6899 [ - + ]: 222712 : Assert(op_strategy != 0); /* not a member of opfamily?? */
6900 [ + + ]: 222712 : if (op_strategy == BTEqualStrategyNumber)
6901 : 210936 : eqQualHere = true;
6902 : : }
6903 : :
6904 : 223818 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
6905 : : }
6906 : : }
6907 : :
6908 : : /*
6909 : : * If index is unique and we found an '=' clause for each column, we can
6910 : : * just assume numIndexTuples = 1 and skip the expensive
6911 : : * clauselist_selectivity calculations. However, a ScalarArrayOp or
6912 : : * NullTest invalidates that theory, even though it sets eqQualHere.
6913 : : */
6715 6914 [ + + ]: 312571 : if (index->unique &&
2199 teodor@sigaev.ru 6915 [ + + + + ]: 259078 : indexcol == index->nkeycolumns - 1 &&
6715 tgl@sss.pgh.pa.us 6916 : 109934 : eqQualHere &&
6218 6917 [ + + ]: 109934 : !found_saop &&
5217 6918 [ + + ]: 107793 : !found_is_null_op)
6880 6919 : 107760 : numIndexTuples = 1.0;
6920 : : else
6921 : : {
6922 : : List *selectivityQuals;
6923 : : Selectivity btreeSelectivity;
6924 : :
6925 : : /*
6926 : : * If the index is partial, AND the index predicate with the
6927 : : * index-bound quals to produce a more accurate idea of the number of
6928 : : * rows covered by the bound conditions.
6929 : : */
1885 6930 : 204811 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
6931 : :
4459 6932 : 204811 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
6880 6933 : 204811 : index->rel->relid,
6934 : : JOIN_INNER,
6935 : : NULL);
6936 : 204811 : numIndexTuples = btreeSelectivity * index->rel->tuples;
6937 : :
6938 : : /*
6939 : : * btree automatically combines individual ScalarArrayOpExpr primitive
6940 : : * index scans whenever the tuples covered by the next set of array
6941 : : * keys are close to tuples covered by the current set. That puts a
6942 : : * natural ceiling on the worst case number of descents -- there
6943 : : * cannot possibly be more than one descent per leaf page scanned.
6944 : : *
6945 : : * Clamp the number of descents to at most 1/3 the number of index
6946 : : * pages. This avoids implausibly high estimates with low selectivity
6947 : : * paths, where scans usually require only one or two descents. This
6948 : : * is most likely to help when there are several SAOP clauses, where
6949 : : * naively accepting the total number of distinct combinations of
6950 : : * array elements as the number of descents would frequently lead to
6951 : : * wild overestimates.
6952 : : *
6953 : : * We somewhat arbitrarily don't just make the cutoff the total number
6954 : : * of leaf pages (we make it 1/3 the total number of pages instead) to
6955 : : * give the btree code credit for its ability to continue on the leaf
6956 : : * level with low selectivity scans.
6957 : : */
8 pg@bowt.ie 6958 [ + + ]:GNC 204811 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
6959 [ + + ]: 204811 : num_sa_scans = Max(num_sa_scans, 1);
6960 : :
6961 : : /*
6962 : : * As in genericcostestimate(), we have to adjust for any
6963 : : * ScalarArrayOpExpr quals included in indexBoundQuals, and then round
6964 : : * to integer.
6965 : : *
6966 : : * It is tempting to make genericcostestimate behave as if SAOP
6967 : : * clauses work in almost the same way as scalar operators during
6968 : : * btree scans, making the top-level scan look like a continuous scan
6969 : : * (as opposed to num_sa_scans-many primitive index scans). After
6970 : : * all, btree scans mostly work like that at runtime. However, such a
6971 : : * scheme would badly bias genericcostestimate's simplistic appraoch
6972 : : * to calculating numIndexPages through prorating.
6973 : : *
6974 : : * Stick with the approach taken by non-native SAOP scans for now.
6975 : : * genericcostestimate will use the Mackert-Lohman formula to
6976 : : * compensate for repeat page fetches, even though that definitely
6977 : : * won't happen during btree scans (not for leaf pages, at least).
6978 : : * We're usually very pessimistic about the number of primitive index
6979 : : * scans that will be required, but it's not clear how to do better.
6980 : : */
6330 tgl@sss.pgh.pa.us 6981 :CBC 204811 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
6982 : : }
6983 : :
6984 : : /*
6985 : : * Now do generic index cost estimation.
6986 : : */
4111 6987 : 312571 : costs.numIndexTuples = numIndexTuples;
8 pg@bowt.ie 6988 :GNC 312571 : costs.num_sa_scans = num_sa_scans;
6989 : :
1885 tgl@sss.pgh.pa.us 6990 :CBC 312571 : genericcostestimate(root, path, loop_count, &costs);
6991 : :
6992 : : /*
6993 : : * Add a CPU-cost component to represent the costs of initial btree
6994 : : * descent. We don't charge any I/O cost for touching upper btree levels,
6995 : : * since they tend to stay in cache, but we still have to do about log2(N)
6996 : : * comparisons to descend a btree of N leaf tuples. We charge one
6997 : : * cpu_operator_cost per comparison.
6998 : : *
6999 : : * If there are ScalarArrayOpExprs, charge this once per estimated SA
7000 : : * index descent. The ones after the first one are not startup cost so
7001 : : * far as the overall plan goes, so just add them to "total" cost.
7002 : : */
4111 7003 [ + + ]: 312571 : if (index->tuples > 1) /* avoid computing log(0) */
7004 : : {
7005 : 291040 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
7006 : 291040 : costs.indexStartupCost += descentCost;
7007 : 291040 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7008 : : }
7009 : :
7010 : : /*
7011 : : * Even though we're not charging I/O cost for touching upper btree pages,
7012 : : * it's still reasonable to charge some CPU cost per page descended
7013 : : * through. Moreover, if we had no such charge at all, bloated indexes
7014 : : * would appear to have the same search cost as unbloated ones, at least
7015 : : * in cases where only a single leaf page is expected to be visited. This
7016 : : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
7017 : : * touched. The number of such pages is btree tree height plus one (ie,
7018 : : * we charge for the leaf page too). As above, charge once per estimated
7019 : : * SA index descent.
7020 : : */
462 akorotkov@postgresql 7021 : 312571 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
4111 tgl@sss.pgh.pa.us 7022 : 312571 : costs.indexStartupCost += descentCost;
7023 : 312571 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7024 : :
7025 : : /*
7026 : : * If we can get an estimate of the first column's ordering correlation C
7027 : : * from pg_statistic, estimate the index correlation as C for a
7028 : : * single-column index, or C * 0.75 for multiple columns. (The idea here
7029 : : * is that multiple columns dilute the importance of the first column's
7030 : : * ordering, but don't negate it entirely. Before 8.0 we divided the
7031 : : * correlation by the number of columns, but that seems too strong.)
7032 : : */
7627 7033 [ + + ]: 312571 : if (index->indexkeys[0] != 0)
7034 : : {
7035 : : /* Simple variable --- look to stats for the underlying table */
5677 7036 [ + - ]: 311488 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
7037 : :
7038 [ - + ]: 311488 : Assert(rte->rtekind == RTE_RELATION);
7039 : 311488 : relid = rte->relid;
8376 7040 [ - + ]: 311488 : Assert(relid != InvalidOid);
7362 7041 : 311488 : colnum = index->indexkeys[0];
7042 : :
5677 7043 [ - + - - ]: 311488 : if (get_relation_stats_hook &&
5677 tgl@sss.pgh.pa.us 7044 :UBC 0 : (*get_relation_stats_hook) (root, rte, colnum, &vardata))
7045 : : {
7046 : : /*
7047 : : * The hook took control of acquiring a stats tuple. If it did
7048 : : * supply a tuple, it'd better have supplied a freefunc.
7049 : : */
7050 [ # # ]: 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
7051 [ # # ]: 0 : !vardata.freefunc)
7052 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
7053 : : }
7054 : : else
7055 : : {
5173 rhaas@postgresql.org 7056 :CBC 311488 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
7057 : : ObjectIdGetDatum(relid),
7058 : : Int16GetDatum(colnum),
7059 : 311488 : BoolGetDatum(rte->inh));
5677 tgl@sss.pgh.pa.us 7060 : 311488 : vardata.freefunc = ReleaseSysCache;
7061 : : }
7062 : : }
7063 : : else
7064 : : {
7065 : : /* Expression --- maybe there are stats for the index itself */
7362 7066 : 1083 : relid = index->indexoid;
7067 : 1083 : colnum = 1;
7068 : :
5677 7069 [ - + - - ]: 1083 : if (get_index_stats_hook &&
5677 tgl@sss.pgh.pa.us 7070 :UBC 0 : (*get_index_stats_hook) (root, relid, colnum, &vardata))
7071 : : {
7072 : : /*
7073 : : * The hook took control of acquiring a stats tuple. If it did
7074 : : * supply a tuple, it'd better have supplied a freefunc.
7075 : : */
7076 [ # # ]: 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
7077 [ # # ]: 0 : !vardata.freefunc)
7078 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
7079 : : }
7080 : : else
7081 : : {
5173 rhaas@postgresql.org 7082 :CBC 1083 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
7083 : : ObjectIdGetDatum(relid),
7084 : : Int16GetDatum(colnum),
7085 : : BoolGetDatum(false));
5677 tgl@sss.pgh.pa.us 7086 : 1083 : vardata.freefunc = ReleaseSysCache;
7087 : : }
7088 : : }
7089 : :
7090 [ + + ]: 312571 : if (HeapTupleIsValid(vardata.statsTuple))
7091 : : {
7092 : : Oid sortop;
7093 : : AttStatsSlot sslot;
7094 : :
4885 7095 : 235869 : sortop = get_opfamily_member(index->opfamily[0],
7096 : 235869 : index->opcintype[0],
7097 : 235869 : index->opcintype[0],
7098 : : BTLessStrategyNumber);
7099 [ + - + + ]: 471738 : if (OidIsValid(sortop) &&
2528 7100 : 235869 : get_attstatsslot(&sslot, vardata.statsTuple,
7101 : : STATISTIC_KIND_CORRELATION, sortop,
7102 : : ATTSTATSSLOT_NUMBERS))
7103 : : {
7104 : : double varCorrelation;
7105 : :
7106 [ - + ]: 233418 : Assert(sslot.nnumbers == 1);
7107 : 233418 : varCorrelation = sslot.numbers[0];
7108 : :
4885 7109 [ - + ]: 233418 : if (index->reverse_sort[0])
4885 tgl@sss.pgh.pa.us 7110 :UBC 0 : varCorrelation = -varCorrelation;
7111 : :
1888 tgl@sss.pgh.pa.us 7112 [ + + ]:CBC 233418 : if (index->nkeycolumns > 1)
4111 7113 : 73596 : costs.indexCorrelation = varCorrelation * 0.75;
7114 : : else
7115 : 159822 : costs.indexCorrelation = varCorrelation;
7116 : :
2528 7117 : 233418 : free_attstatsslot(&sslot);
7118 : : }
7119 : : }
7120 : :
5677 7121 [ + + ]: 312571 : ReleaseVariableStats(vardata);
7122 : :
4111 7123 : 312571 : *indexStartupCost = costs.indexStartupCost;
7124 : 312571 : *indexTotalCost = costs.indexTotalCost;
7125 : 312571 : *indexSelectivity = costs.indexSelectivity;
7126 : 312571 : *indexCorrelation = costs.indexCorrelation;
2615 rhaas@postgresql.org 7127 : 312571 : *indexPages = costs.numIndexPages;
10141 scrappy@hub.org 7128 : 312571 : }
7129 : :
7130 : : void
3010 tgl@sss.pgh.pa.us 7131 : 199 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7132 : : Cost *indexStartupCost, Cost *indexTotalCost,
7133 : : Selectivity *indexSelectivity, double *indexCorrelation,
7134 : : double *indexPages)
7135 : : {
638 peter@eisentraut.org 7136 : 199 : GenericCosts costs = {0};
7137 : :
1885 tgl@sss.pgh.pa.us 7138 : 199 : genericcostestimate(root, path, loop_count, &costs);
7139 : :
7140 : : /*
7141 : : * A hash index has no descent costs as such, since the index AM can go
7142 : : * directly to the target bucket after computing the hash value. There
7143 : : * are a couple of other hash-specific costs that we could conceivably add
7144 : : * here, though:
7145 : : *
7146 : : * Ideally we'd charge spc_random_page_cost for each page in the target
7147 : : * bucket, not just the numIndexPages pages that genericcostestimate
7148 : : * thought we'd visit. However in most cases we don't know which bucket
7149 : : * that will be. There's no point in considering the average bucket size
7150 : : * because the hash AM makes sure that's always one page.
7151 : : *
7152 : : * Likewise, we could consider charging some CPU for each index tuple in
7153 : : * the bucket, if we knew how many there were. But the per-tuple cost is
7154 : : * just a hash value comparison, not a general datatype-dependent
7155 : : * comparison, so any such charge ought to be quite a bit less than
7156 : : * cpu_operator_cost; which makes it probably not worth worrying about.
7157 : : *
7158 : : * A bigger issue is that chance hash-value collisions will result in
7159 : : * wasted probes into the heap. We don't currently attempt to model this
7160 : : * cost on the grounds that it's rare, but maybe it's not rare enough.
7161 : : * (Any fix for this ought to consider the generic lossy-operator problem,
7162 : : * though; it's not entirely hash-specific.)
7163 : : */
7164 : :
4111 7165 : 199 : *indexStartupCost = costs.indexStartupCost;
7166 : 199 : *indexTotalCost = costs.indexTotalCost;
7167 : 199 : *indexSelectivity = costs.indexSelectivity;
7168 : 199 : *indexCorrelation = costs.indexCorrelation;
2615 rhaas@postgresql.org 7169 : 199 : *indexPages = costs.numIndexPages;
10093 scrappy@hub.org 7170 : 199 : }
7171 : :
7172 : : void
3010 tgl@sss.pgh.pa.us 7173 : 2231 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7174 : : Cost *indexStartupCost, Cost *indexTotalCost,
7175 : : Selectivity *indexSelectivity, double *indexCorrelation,
7176 : : double *indexPages)
7177 : : {
4111 7178 : 2231 : IndexOptInfo *index = path->indexinfo;
638 peter@eisentraut.org 7179 : 2231 : GenericCosts costs = {0};
7180 : : Cost descentCost;
7181 : :
1885 tgl@sss.pgh.pa.us 7182 : 2231 : genericcostestimate(root, path, loop_count, &costs);
7183 : :
7184 : : /*
7185 : : * We model index descent costs similarly to those for btree, but to do
7186 : : * that we first need an idea of the tree height. We somewhat arbitrarily
7187 : : * assume that the fanout is 100, meaning the tree height is at most
7188 : : * log100(index->pages).
7189 : : *
7190 : : * Although this computation isn't really expensive enough to require
7191 : : * caching, we might as well use index->tree_height to cache it.
7192 : : */
3973 bruce@momjian.us 7193 [ + + ]: 2231 : if (index->tree_height < 0) /* unknown? */
7194 : : {
4111 tgl@sss.pgh.pa.us 7195 [ + + ]: 2224 : if (index->pages > 1) /* avoid computing log(0) */
7196 : 1352 : index->tree_height = (int) (log(index->pages) / log(100.0));
7197 : : else
7198 : 872 : index->tree_height = 0;
7199 : : }
7200 : :
7201 : : /*
7202 : : * Add a CPU-cost component to represent the costs of initial descent. We
7203 : : * just use log(N) here not log2(N) since the branching factor isn't
7204 : : * necessarily two anyway. As for btree, charge once per SA scan.
7205 : : */
7206 [ + - ]: 2231 : if (index->tuples > 1) /* avoid computing log(0) */
7207 : : {
7208 : 2231 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7209 : 2231 : costs.indexStartupCost += descentCost;
7210 : 2231 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7211 : : }
7212 : :
7213 : : /*
7214 : : * Likewise add a per-page charge, calculated the same as for btrees.
7215 : : */
462 akorotkov@postgresql 7216 : 2231 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
4111 tgl@sss.pgh.pa.us 7217 : 2231 : costs.indexStartupCost += descentCost;
7218 : 2231 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7219 : :
7220 : 2231 : *indexStartupCost = costs.indexStartupCost;
7221 : 2231 : *indexTotalCost = costs.indexTotalCost;
7222 : 2231 : *indexSelectivity = costs.indexSelectivity;
7223 : 2231 : *indexCorrelation = costs.indexCorrelation;
2615 rhaas@postgresql.org 7224 : 2231 : *indexPages = costs.numIndexPages;
10093 scrappy@hub.org 7225 : 2231 : }
7226 : :
7227 : : void
3010 tgl@sss.pgh.pa.us 7228 : 892 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7229 : : Cost *indexStartupCost, Cost *indexTotalCost,
7230 : : Selectivity *indexSelectivity, double *indexCorrelation,
7231 : : double *indexPages)
7232 : : {
4111 7233 : 892 : IndexOptInfo *index = path->indexinfo;
638 peter@eisentraut.org 7234 : 892 : GenericCosts costs = {0};
7235 : : Cost descentCost;
7236 : :
1885 tgl@sss.pgh.pa.us 7237 : 892 : genericcostestimate(root, path, loop_count, &costs);
7238 : :
7239 : : /*
7240 : : * We model index descent costs similarly to those for btree, but to do
7241 : : * that we first need an idea of the tree height. We somewhat arbitrarily
7242 : : * assume that the fanout is 100, meaning the tree height is at most
7243 : : * log100(index->pages).
7244 : : *
7245 : : * Although this computation isn't really expensive enough to require
7246 : : * caching, we might as well use index->tree_height to cache it.
7247 : : */
3973 bruce@momjian.us 7248 [ + + ]: 892 : if (index->tree_height < 0) /* unknown? */
7249 : : {
4111 tgl@sss.pgh.pa.us 7250 [ + - ]: 889 : if (index->pages > 1) /* avoid computing log(0) */
7251 : 889 : index->tree_height = (int) (log(index->pages) / log(100.0));
7252 : : else
4111 tgl@sss.pgh.pa.us 7253 :UBC 0 : index->tree_height = 0;
7254 : : }
7255 : :
7256 : : /*
7257 : : * Add a CPU-cost component to represent the costs of initial descent. We
7258 : : * just use log(N) here not log2(N) since the branching factor isn't
7259 : : * necessarily two anyway. As for btree, charge once per SA scan.
7260 : : */
4111 tgl@sss.pgh.pa.us 7261 [ + - ]:CBC 892 : if (index->tuples > 1) /* avoid computing log(0) */
7262 : : {
7263 : 892 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
7264 : 892 : costs.indexStartupCost += descentCost;
7265 : 892 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7266 : : }
7267 : :
7268 : : /*
7269 : : * Likewise add a per-page charge, calculated the same as for btrees.
7270 : : */
462 akorotkov@postgresql 7271 : 892 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
4111 tgl@sss.pgh.pa.us 7272 : 892 : costs.indexStartupCost += descentCost;
7273 : 892 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
7274 : :
7275 : 892 : *indexStartupCost = costs.indexStartupCost;
7276 : 892 : *indexTotalCost = costs.indexTotalCost;
7277 : 892 : *indexSelectivity = costs.indexSelectivity;
7278 : 892 : *indexCorrelation = costs.indexCorrelation;
2615 rhaas@postgresql.org 7279 : 892 : *indexPages = costs.numIndexPages;
4502 tgl@sss.pgh.pa.us 7280 : 892 : }
7281 : :
7282 : :
7283 : : /*
7284 : : * Support routines for gincostestimate
7285 : : */
7286 : :
7287 : : typedef struct
7288 : : {
7289 : : bool attHasFullScan[INDEX_MAX_KEYS];
7290 : : bool attHasNormalScan[INDEX_MAX_KEYS];
7291 : : double partialEntries;
7292 : : double exactEntries;
7293 : : double searchEntries;
7294 : : double arrayScans;
7295 : : } GinQualCounts;
7296 : :
7297 : : /*
7298 : : * Estimate the number of index terms that need to be searched for while
7299 : : * testing the given GIN query, and increment the counts in *counts
7300 : : * appropriately. If the query is unsatisfiable, return false.
7301 : : */
7302 : : static bool
4499 7303 : 1036 : gincost_pattern(IndexOptInfo *index, int indexcol,
7304 : : Oid clause_op, Datum query,
7305 : : GinQualCounts *counts)
7306 : : {
7307 : : FmgrInfo flinfo;
7308 : : Oid extractProcOid;
7309 : : Oid collation;
7310 : : int strategy_op;
7311 : : Oid lefttype,
7312 : : righttype;
7313 : 1036 : int32 nentries = 0;
7314 : 1036 : bool *partial_matches = NULL;
7315 : 1036 : Pointer *extra_data = NULL;
7316 : 1036 : bool *nullFlags = NULL;
7317 : 1036 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
7318 : : int32 i;
7319 : :
2194 teodor@sigaev.ru 7320 [ - + ]: 1036 : Assert(indexcol < index->nkeycolumns);
7321 : :
7322 : : /*
7323 : : * Get the operator's strategy number and declared input data types within
7324 : : * the index opfamily. (We don't need the latter, but we use
7325 : : * get_op_opfamily_properties because it will throw error if it fails to
7326 : : * find a matching pg_amop entry.)
7327 : : */
4499 tgl@sss.pgh.pa.us 7328 : 1036 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
7329 : : &strategy_op, &lefttype, &righttype);
7330 : :
7331 : : /*
7332 : : * GIN always uses the "default" support functions, which are those with
7333 : : * lefttype == righttype == the opclass' opcintype (see
7334 : : * IndexSupportInitialize in relcache.c).
7335 : : */
7336 : 1036 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
7337 : 1036 : index->opcintype[indexcol],
7338 : 1036 : index->opcintype[indexcol],
7339 : : GIN_EXTRACTQUERY_PROC);
7340 : :
7341 [ - + ]: 1036 : if (!OidIsValid(extractProcOid))
7342 : : {
7343 : : /* should not happen; throw same error as index_getprocinfo */
4499 tgl@sss.pgh.pa.us 7344 [ # # ]:UBC 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
7345 : : GIN_EXTRACTQUERY_PROC, indexcol + 1,
7346 : : get_rel_name(index->indexoid));
7347 : : }
7348 : :
7349 : : /*
7350 : : * Choose collation to pass to extractProc (should match initGinState).
7351 : : */
4023 tgl@sss.pgh.pa.us 7352 [ + + ]:CBC 1036 : if (OidIsValid(index->indexcollations[indexcol]))
7353 : 182 : collation = index->indexcollations[indexcol];
7354 : : else
7355 : 854 : collation = DEFAULT_COLLATION_OID;
7356 : :
1476 akorotkov@postgresql 7357 : 1036 : fmgr_info(extractProcOid, &flinfo);
7358 : :
7359 : 1036 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
7360 : :
7361 : 1036 : FunctionCall7Coll(&flinfo,
7362 : : collation,
7363 : : query,
7364 : : PointerGetDatum(&nentries),
7365 : : UInt16GetDatum(strategy_op),
7366 : : PointerGetDatum(&partial_matches),
7367 : : PointerGetDatum(&extra_data),
7368 : : PointerGetDatum(&nullFlags),
7369 : : PointerGetDatum(&searchMode));
7370 : :
4499 tgl@sss.pgh.pa.us 7371 [ + + + + ]: 1036 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
7372 : : {
7373 : : /* No match is possible */
7374 : 6 : return false;
7375 : : }
7376 : :
7377 [ + + ]: 4398 : for (i = 0; i < nentries; i++)
7378 : : {
7379 : : /*
7380 : : * For partial match we haven't any information to estimate number of
7381 : : * matched entries in index, so, we just estimate it as 100
7382 : : */
7383 [ + + + + ]: 3368 : if (partial_matches && partial_matches[i])
7384 : 158 : counts->partialEntries += 100;
7385 : : else
7386 : 3210 : counts->exactEntries++;
7387 : :
7388 : 3368 : counts->searchEntries++;
7389 : : }
7390 : :
1548 akorotkov@postgresql 7391 [ + + ]: 1030 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
7392 : : {
7393 : 794 : counts->attHasNormalScan[indexcol] = true;
7394 : : }
7395 [ + + ]: 236 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
7396 : : {
7397 : : /* Treat "include empty" like an exact-match item */
7398 : 22 : counts->attHasNormalScan[indexcol] = true;
4499 tgl@sss.pgh.pa.us 7399 : 22 : counts->exactEntries++;
7400 : 22 : counts->searchEntries++;
7401 : : }
7402 : : else
7403 : : {
7404 : : /* It's GIN_SEARCH_MODE_ALL */
1548 akorotkov@postgresql 7405 : 214 : counts->attHasFullScan[indexcol] = true;
7406 : : }
7407 : :
4499 tgl@sss.pgh.pa.us 7408 : 1030 : return true;
7409 : : }
7410 : :
7411 : : /*
7412 : : * Estimate the number of index terms that need to be searched for while
7413 : : * testing the given GIN index clause, and increment the counts in *counts
7414 : : * appropriately. If the query is unsatisfiable, return false.
7415 : : */
7416 : : static bool
3330 7417 : 1030 : gincost_opexpr(PlannerInfo *root,
7418 : : IndexOptInfo *index,
7419 : : int indexcol,
7420 : : OpExpr *clause,
7421 : : GinQualCounts *counts)
7422 : : {
1885 7423 : 1030 : Oid clause_op = clause->opno;
7424 : 1030 : Node *operand = (Node *) lsecond(clause->args);
7425 : :
7426 : : /* aggressively reduce to a constant, and look through relabeling */
3705 7427 : 1030 : operand = estimate_expression_value(root, operand);
7428 : :
4499 7429 [ - + ]: 1030 : if (IsA(operand, RelabelType))
4499 tgl@sss.pgh.pa.us 7430 :UBC 0 : operand = (Node *) ((RelabelType *) operand)->arg;
7431 : :
7432 : : /*
7433 : : * It's impossible to call extractQuery method for unknown operand. So
7434 : : * unless operand is a Const we can't do much; just assume there will be
7435 : : * one ordinary search entry from the operand at runtime.
7436 : : */
4499 tgl@sss.pgh.pa.us 7437 [ - + ]:CBC 1030 : if (!IsA(operand, Const))
7438 : : {
4499 tgl@sss.pgh.pa.us 7439 :UBC 0 : counts->exactEntries++;
7440 : 0 : counts->searchEntries++;
7441 : 0 : return true;
7442 : : }
7443 : :
7444 : : /* If Const is null, there can be no matches */
4499 tgl@sss.pgh.pa.us 7445 [ - + ]:CBC 1030 : if (((Const *) operand)->constisnull)
4499 tgl@sss.pgh.pa.us 7446 :UBC 0 : return false;
7447 : :
7448 : : /* Otherwise, apply extractQuery and get the actual term counts */
4499 tgl@sss.pgh.pa.us 7449 :CBC 1030 : return gincost_pattern(index, indexcol, clause_op,
7450 : : ((Const *) operand)->constvalue,
7451 : : counts);
7452 : : }
7453 : :
7454 : : /*
7455 : : * Estimate the number of index terms that need to be searched for while
7456 : : * testing the given GIN index clause, and increment the counts in *counts
7457 : : * appropriately. If the query is unsatisfiable, return false.
7458 : : *
7459 : : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
7460 : : * each of which involves one value from the RHS array, plus all the
7461 : : * non-array quals (if any). To model this, we average the counts across
7462 : : * the RHS elements, and add the averages to the counts in *counts (which
7463 : : * correspond to per-indexscan costs). We also multiply counts->arrayScans
7464 : : * by N, causing gincostestimate to scale up its estimates accordingly.
7465 : : */
7466 : : static bool
3705 7467 : 3 : gincost_scalararrayopexpr(PlannerInfo *root,
7468 : : IndexOptInfo *index,
7469 : : int indexcol,
7470 : : ScalarArrayOpExpr *clause,
7471 : : double numIndexEntries,
7472 : : GinQualCounts *counts)
7473 : : {
1885 7474 : 3 : Oid clause_op = clause->opno;
7475 : 3 : Node *rightop = (Node *) lsecond(clause->args);
7476 : : ArrayType *arrayval;
7477 : : int16 elmlen;
7478 : : bool elmbyval;
7479 : : char elmalign;
7480 : : int numElems;
7481 : : Datum *elemValues;
7482 : : bool *elemNulls;
7483 : : GinQualCounts arraycounts;
4499 7484 : 3 : int numPossible = 0;
7485 : : int i;
7486 : :
1885 7487 [ - + ]: 3 : Assert(clause->useOr);
7488 : :
7489 : : /* aggressively reduce to a constant, and look through relabeling */
3705 7490 : 3 : rightop = estimate_expression_value(root, rightop);
7491 : :
4499 7492 [ - + ]: 3 : if (IsA(rightop, RelabelType))
4499 tgl@sss.pgh.pa.us 7493 :UBC 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
7494 : :
7495 : : /*
7496 : : * It's impossible to call extractQuery method for unknown operand. So
7497 : : * unless operand is a Const we can't do much; just assume there will be
7498 : : * one ordinary search entry from each array entry at runtime, and fall
7499 : : * back on a probably-bad estimate of the number of array entries.
7500 : : */
4499 tgl@sss.pgh.pa.us 7501 [ - + ]:CBC 3 : if (!IsA(rightop, Const))
7502 : : {
4499 tgl@sss.pgh.pa.us 7503 :UBC 0 : counts->exactEntries++;
7504 : 0 : counts->searchEntries++;
101 tgl@sss.pgh.pa.us 7505 :UNC 0 : counts->arrayScans *= estimate_array_length(root, rightop);
4499 tgl@sss.pgh.pa.us 7506 :UBC 0 : return true;
7507 : : }
7508 : :
7509 : : /* If Const is null, there can be no matches */
4499 tgl@sss.pgh.pa.us 7510 [ - + ]:CBC 3 : if (((Const *) rightop)->constisnull)
4499 tgl@sss.pgh.pa.us 7511 :UBC 0 : return false;
7512 : :
7513 : : /* Otherwise, extract the array elements and iterate over them */
4499 tgl@sss.pgh.pa.us 7514 :CBC 3 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
7515 : 3 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
7516 : : &elmlen, &elmbyval, &elmalign);
7517 : 3 : deconstruct_array(arrayval,
7518 : : ARR_ELEMTYPE(arrayval),
7519 : : elmlen, elmbyval, elmalign,
7520 : : &elemValues, &elemNulls, &numElems);
7521 : :
7522 : 3 : memset(&arraycounts, 0, sizeof(arraycounts));
7523 : :
7524 [ + + ]: 9 : for (i = 0; i < numElems; i++)
7525 : : {
7526 : : GinQualCounts elemcounts;
7527 : :
7528 : : /* NULL can't match anything, so ignore, as the executor will */
7529 [ - + ]: 6 : if (elemNulls[i])
4499 tgl@sss.pgh.pa.us 7530 :UBC 0 : continue;
7531 : :
7532 : : /* Otherwise, apply extractQuery and get the actual term counts */
4499 tgl@sss.pgh.pa.us 7533 :CBC 6 : memset(&elemcounts, 0, sizeof(elemcounts));
7534 : :
7535 [ + - ]: 6 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
7536 : : &elemcounts))
7537 : : {
7538 : : /* We ignore array elements that are unsatisfiable patterns */
7539 : 6 : numPossible++;
7540 : :
1548 akorotkov@postgresql 7541 [ - + ]: 6 : if (elemcounts.attHasFullScan[indexcol] &&
1548 akorotkov@postgresql 7542 [ # # ]:UBC 0 : !elemcounts.attHasNormalScan[indexcol])
7543 : : {
7544 : : /*
7545 : : * Full index scan will be required. We treat this as if
7546 : : * every key in the index had been listed in the query; is
7547 : : * that reasonable?
7548 : : */
4499 tgl@sss.pgh.pa.us 7549 : 0 : elemcounts.partialEntries = 0;
7550 : 0 : elemcounts.exactEntries = numIndexEntries;
7551 : 0 : elemcounts.searchEntries = numIndexEntries;
7552 : : }
4499 tgl@sss.pgh.pa.us 7553 :CBC 6 : arraycounts.partialEntries += elemcounts.partialEntries;
7554 : 6 : arraycounts.exactEntries += elemcounts.exactEntries;
7555 : 6 : arraycounts.searchEntries += elemcounts.searchEntries;
7556 : : }
7557 : : }
7558 : :
7559 [ - + ]: 3 : if (numPossible == 0)
7560 : : {
7561 : : /* No satisfiable patterns in the array */
4499 tgl@sss.pgh.pa.us 7562 :UBC 0 : return false;
7563 : : }
7564 : :
7565 : : /*
7566 : : * Now add the averages to the global counts. This will give us an
7567 : : * estimate of the average number of terms searched for in each indexscan,
7568 : : * including contributions from both array and non-array quals.
7569 : : */
4499 tgl@sss.pgh.pa.us 7570 :CBC 3 : counts->partialEntries += arraycounts.partialEntries / numPossible;
7571 : 3 : counts->exactEntries += arraycounts.exactEntries / numPossible;
7572 : 3 : counts->searchEntries += arraycounts.searchEntries / numPossible;
7573 : :
7574 : 3 : counts->arrayScans *= numPossible;
7575 : :
7576 : 3 : return true;
7577 : : }
7578 : :
7579 : : /*
7580 : : * GIN has search behavior completely different from other index types
7581 : : */
7582 : : void
3010 7583 : 934 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7584 : : Cost *indexStartupCost, Cost *indexTotalCost,
7585 : : Selectivity *indexSelectivity, double *indexCorrelation,
7586 : : double *indexPages)
7587 : : {
4495 7588 : 934 : IndexOptInfo *index = path->indexinfo;
1885 7589 : 934 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7590 : : List *selectivityQuals;
4753 bruce@momjian.us 7591 : 934 : double numPages = index->pages,
7592 : 934 : numTuples = index->tuples;
7593 : : double numEntryPages,
7594 : : numDataPages,
7595 : : numPendingPages,
7596 : : numEntries;
7597 : : GinQualCounts counts;
7598 : : bool matchPossible;
7599 : : bool fullIndexScan;
7600 : : double partialScale;
7601 : : double entryPagesFetched,
7602 : : dataPagesFetched,
7603 : : dataPagesFetchedBySel;
7604 : : double qual_op_cost,
7605 : : qual_arg_cost,
7606 : : spc_random_page_cost,
7607 : : outer_scans;
7608 : : Cost descentCost;
7609 : : Relation indexRel;
7610 : : GinStatsData ginStats;
7611 : : ListCell *lc;
7612 : : int i;
7613 : :
7614 : : /*
7615 : : * Obtain statistical information from the meta page, if possible. Else
7616 : : * set ginStats to zeroes, and we'll cope below.
7617 : : */
3057 tgl@sss.pgh.pa.us 7618 [ + - ]: 934 : if (!index->hypothetical)
7619 : : {
7620 : : /* Lock should have already been obtained in plancat.c */
1837 7621 : 934 : indexRel = index_open(index->indexoid, NoLock);
3057 7622 : 934 : ginGetStats(indexRel, &ginStats);
1837 7623 : 934 : index_close(indexRel, NoLock);
7624 : : }
7625 : : else
7626 : : {
3057 tgl@sss.pgh.pa.us 7627 :UBC 0 : memset(&ginStats, 0, sizeof(ginStats));
7628 : : }
7629 : :
7630 : : /*
7631 : : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
7632 : : * trusted, but the other fields are data as of the last VACUUM. We can
7633 : : * scale them up to account for growth since then, but that method only
7634 : : * goes so far; in the worst case, the stats might be for a completely
7635 : : * empty index, and scaling them will produce pretty bogus numbers.
7636 : : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
7637 : : * it's grown more than that, fall back to estimating things only from the
7638 : : * assumed-accurate index size. But we'll trust nPendingPages in any case
7639 : : * so long as it's not clearly insane, ie, more than the index size.
7640 : : */
3026 tgl@sss.pgh.pa.us 7641 [ + - ]:CBC 934 : if (ginStats.nPendingPages < numPages)
7642 : 934 : numPendingPages = ginStats.nPendingPages;
7643 : : else
3026 tgl@sss.pgh.pa.us 7644 :UBC 0 : numPendingPages = 0;
7645 : :
3026 tgl@sss.pgh.pa.us 7646 [ + - + - ]:CBC 934 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
7647 [ + + ]: 934 : ginStats.nTotalPages > numPages / 4 &&
7648 [ + - + + ]: 908 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
3057 7649 : 779 : {
7650 : : /*
7651 : : * OK, the stats seem close enough to sane to be trusted. But we
7652 : : * still need to scale them by the ratio numPages / nTotalPages to
7653 : : * account for growth since the last VACUUM.
7654 : : */
4753 bruce@momjian.us 7655 : 779 : double scale = numPages / ginStats.nTotalPages;
7656 : :
3026 tgl@sss.pgh.pa.us 7657 : 779 : numEntryPages = ceil(ginStats.nEntryPages * scale);
7658 : 779 : numDataPages = ceil(ginStats.nDataPages * scale);
7659 : 779 : numEntries = ceil(ginStats.nEntries * scale);
7660 : : /* ensure we didn't round up too much */
7661 [ + + ]: 779 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
7662 [ + + ]: 779 : numDataPages = Min(numDataPages,
7663 : : numPages - numPendingPages - numEntryPages);
7664 : : }
7665 : : else
7666 : : {
7667 : : /*
7668 : : * We might get here because it's a hypothetical index, or an index
7669 : : * created pre-9.1 and never vacuumed since upgrading (in which case
7670 : : * its stats would read as zeroes), or just because it's grown too
7671 : : * much since the last VACUUM for us to put our faith in scaling.
7672 : : *
7673 : : * Invent some plausible internal statistics based on the index page
7674 : : * count (and clamp that to at least 10 pages, just in case). We
7675 : : * estimate that 90% of the index is entry pages, and the rest is data
7676 : : * pages. Estimate 100 entries per entry page; this is rather bogus
7677 : : * since it'll depend on the size of the keys, but it's more robust
7678 : : * than trying to predict the number of entries per heap tuple.
7679 : : */
3057 7680 [ + + ]: 155 : numPages = Max(numPages, 10);
3026 7681 : 155 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
7682 : 155 : numDataPages = numPages - numPendingPages - numEntryPages;
3057 7683 : 155 : numEntries = floor(numEntryPages * 100);
7684 : : }
7685 : :
7686 : : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
4742 7687 [ - + ]: 934 : if (numEntries < 1)
4742 tgl@sss.pgh.pa.us 7688 :UBC 0 : numEntries = 1;
7689 : :
7690 : : /*
7691 : : * If the index is partial, AND the index predicate with the index-bound
7692 : : * quals to produce a more accurate idea of the number of rows covered by
7693 : : * the bound conditions.
7694 : : */
1885 tgl@sss.pgh.pa.us 7695 :CBC 934 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7696 : :
7697 : : /* Estimate the fraction of main-table tuples that will be visited */
4928 7698 : 1868 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
4753 bruce@momjian.us 7699 : 934 : index->rel->relid,
7700 : : JOIN_INNER,
7701 : : NULL);
7702 : :
7703 : : /* fetch estimated page cost for tablespace containing index */
7704 : 934 : get_tablespace_page_costs(index->reltablespace,
7705 : : &spc_random_page_cost,
7706 : : NULL);
7707 : :
7708 : : /*
7709 : : * Generic assumption about index correlation: there isn't any.
7710 : : */
4928 tgl@sss.pgh.pa.us 7711 : 934 : *indexCorrelation = 0.0;
7712 : :
7713 : : /*
7714 : : * Examine quals to estimate number of search entries & partial matches
7715 : : */
4499 7716 : 934 : memset(&counts, 0, sizeof(counts));
7717 : 934 : counts.arrayScans = 1;
7718 : 934 : matchPossible = true;
7719 : :
1885 7720 [ + - + + : 1967 : foreach(lc, path->indexclauses)
+ + ]
7721 : : {
7722 : 1033 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7723 : : ListCell *lc2;
7724 : :
7725 [ + - + + : 2060 : foreach(lc2, iclause->indexquals)
+ + ]
7726 : : {
7727 : 1033 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7728 : 1033 : Expr *clause = rinfo->clause;
7729 : :
7730 [ + + ]: 1033 : if (IsA(clause, OpExpr))
7731 : : {
7732 : 1030 : matchPossible = gincost_opexpr(root,
7733 : : index,
7734 : 1030 : iclause->indexcol,
7735 : : (OpExpr *) clause,
7736 : : &counts);
7737 [ + + ]: 1030 : if (!matchPossible)
7738 : 6 : break;
7739 : : }
7740 [ + - ]: 3 : else if (IsA(clause, ScalarArrayOpExpr))
7741 : : {
7742 : 3 : matchPossible = gincost_scalararrayopexpr(root,
7743 : : index,
7744 : 3 : iclause->indexcol,
7745 : : (ScalarArrayOpExpr *) clause,
7746 : : numEntries,
7747 : : &counts);
7748 [ - + ]: 3 : if (!matchPossible)
1885 tgl@sss.pgh.pa.us 7749 :UBC 0 : break;
7750 : : }
7751 : : else
7752 : : {
7753 : : /* shouldn't be anything else for a GIN index */
7754 [ # # ]: 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
7755 : : (int) nodeTag(clause));
7756 : : }
7757 : : }
7758 : : }
7759 : :
7760 : : /* Fall out if there were any provably-unsatisfiable quals */
4499 tgl@sss.pgh.pa.us 7761 [ + + ]:CBC 934 : if (!matchPossible)
7762 : : {
7763 : 6 : *indexStartupCost = 0;
7764 : 6 : *indexTotalCost = 0;
7765 : 6 : *indexSelectivity = 0;
3010 7766 : 6 : return;
7767 : : }
7768 : :
7769 : : /*
7770 : : * If attribute has a full scan and at the same time doesn't have normal
7771 : : * scan, then we'll have to scan all non-null entries of that attribute.
7772 : : * Currently, we don't have per-attribute statistics for GIN. Thus, we
7773 : : * must assume the whole GIN index has to be scanned in this case.
7774 : : */
1548 akorotkov@postgresql 7775 : 928 : fullIndexScan = false;
7776 [ + + ]: 1801 : for (i = 0; i < index->nkeycolumns; i++)
7777 : : {
7778 [ + + + + ]: 1042 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
7779 : : {
7780 : 169 : fullIndexScan = true;
7781 : 169 : break;
7782 : : }
7783 : : }
7784 : :
7785 [ + + - + ]: 928 : if (fullIndexScan || indexQuals == NIL)
7786 : : {
7787 : : /*
7788 : : * Full index scan will be required. We treat this as if every key in
7789 : : * the index had been listed in the query; is that reasonable?
7790 : : */
4499 tgl@sss.pgh.pa.us 7791 : 169 : counts.partialEntries = 0;
7792 : 169 : counts.exactEntries = numEntries;
7793 : 169 : counts.searchEntries = numEntries;
7794 : : }
7795 : :
7796 : : /* Will we have more than one iteration of a nestloop scan? */
4461 7797 : 928 : outer_scans = loop_count;
7798 : :
7799 : : /*
7800 : : * Compute cost to begin scan, first of all, pay attention to pending
7801 : : * list.
7802 : : */
4928 7803 : 928 : entryPagesFetched = numPendingPages;
7804 : :
7805 : : /*
7806 : : * Estimate number of entry pages read. We need to do
7807 : : * counts.searchEntries searches. Use a power function as it should be,
7808 : : * but tuples on leaf pages usually is much greater. Here we include all
7809 : : * searches in entry tree, including search of first entry in partial
7810 : : * match algorithm
7811 : : */
4499 7812 : 928 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
7813 : :
7814 : : /*
7815 : : * Add an estimate of entry pages read by partial match algorithm. It's a
7816 : : * scan over leaf pages in entry tree. We haven't any useful stats here,
7817 : : * so estimate it as proportion. Because counts.partialEntries is really
7818 : : * pretty bogus (see code above), it's possible that it is more than
7819 : : * numEntries; clamp the proportion to ensure sanity.
7820 : : */
3026 7821 : 928 : partialScale = counts.partialEntries / numEntries;
7822 [ + + ]: 928 : partialScale = Min(partialScale, 1.0);
7823 : :
7824 : 928 : entryPagesFetched += ceil(numEntryPages * partialScale);
7825 : :
7826 : : /*
7827 : : * Partial match algorithm reads all data pages before doing actual scan,
7828 : : * so it's a startup cost. Again, we haven't any useful stats here, so
7829 : : * estimate it as proportion.
7830 : : */
7831 : 928 : dataPagesFetched = ceil(numDataPages * partialScale);
7832 : :
462 akorotkov@postgresql 7833 : 928 : *indexStartupCost = 0;
7834 : 928 : *indexTotalCost = 0;
7835 : :
7836 : : /*
7837 : : * Add a CPU-cost component to represent the costs of initial entry btree
7838 : : * descent. We don't charge any I/O cost for touching upper btree levels,
7839 : : * since they tend to stay in cache, but we still have to do about log2(N)
7840 : : * comparisons to descend a btree of N leaf tuples. We charge one
7841 : : * cpu_operator_cost per comparison.
7842 : : *
7843 : : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
7844 : : * ones after the first one are not startup cost so far as the overall
7845 : : * plan is concerned, so add them only to "total" cost.
7846 : : */
7847 [ + - ]: 928 : if (numEntries > 1) /* avoid computing log(0) */
7848 : : {
7849 : 928 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
7850 : 928 : *indexStartupCost += descentCost * counts.searchEntries;
7851 : 928 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
7852 : : }
7853 : :
7854 : : /*
7855 : : * Add a cpu cost per entry-page fetched. This is not amortized over a
7856 : : * loop.
7857 : : */
7858 : 928 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7859 : 928 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7860 : :
7861 : : /*
7862 : : * Add a cpu cost per data-page fetched. This is also not amortized over a
7863 : : * loop. Since those are the data pages from the partial match algorithm,
7864 : : * charge them as startup cost.
7865 : : */
7866 : 928 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
7867 : :
7868 : : /*
7869 : : * Since we add the startup cost to the total cost later on, remove the
7870 : : * initial arrayscan from the total.
7871 : : */
7872 : 928 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7873 : :
7874 : : /*
7875 : : * Calculate cache effects if more than one scan due to nestloops or array
7876 : : * quals. The result is pro-rated per nestloop scan, but the array qual
7877 : : * factor shouldn't be pro-rated (compare genericcostestimate).
7878 : : */
4499 tgl@sss.pgh.pa.us 7879 [ + - + + ]: 928 : if (outer_scans > 1 || counts.arrayScans > 1)
7880 : : {
7881 : 3 : entryPagesFetched *= outer_scans * counts.arrayScans;
4928 7882 : 3 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
7883 : : (BlockNumber) numEntryPages,
7884 : : numEntryPages, root);
4499 7885 : 3 : entryPagesFetched /= outer_scans;
7886 : 3 : dataPagesFetched *= outer_scans * counts.arrayScans;
4928 7887 : 3 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
7888 : : (BlockNumber) numDataPages,
7889 : : numDataPages, root);
4499 7890 : 3 : dataPagesFetched /= outer_scans;
7891 : : }
7892 : :
7893 : : /*
7894 : : * Here we use random page cost because logically-close pages could be far
7895 : : * apart on disk.
7896 : : */
462 akorotkov@postgresql 7897 : 928 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
7898 : :
7899 : : /*
7900 : : * Now compute the number of data pages fetched during the scan.
7901 : : *
7902 : : * We assume every entry to have the same number of items, and that there
7903 : : * is no overlap between them. (XXX: tsvector and array opclasses collect
7904 : : * statistics on the frequency of individual keys; it would be nice to use
7905 : : * those here.)
7906 : : */
4499 tgl@sss.pgh.pa.us 7907 : 928 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
7908 : :
7909 : : /*
7910 : : * If there is a lot of overlap among the entries, in particular if one of
7911 : : * the entries is very frequent, the above calculation can grossly
7912 : : * under-estimate. As a simple cross-check, calculate a lower bound based
7913 : : * on the overall selectivity of the quals. At a minimum, we must read
7914 : : * one item pointer for each matching entry.
7915 : : *
7916 : : * The width of each item pointer varies, based on the level of
7917 : : * compression. We don't have statistics on that, but an average of
7918 : : * around 3 bytes per item is fairly typical.
7919 : : */
4928 7920 : 928 : dataPagesFetchedBySel = ceil(*indexSelectivity *
3686 heikki.linnakangas@i 7921 : 928 : (numTuples / (BLCKSZ / 3)));
4928 tgl@sss.pgh.pa.us 7922 [ + + ]: 928 : if (dataPagesFetchedBySel > dataPagesFetched)
7923 : 738 : dataPagesFetched = dataPagesFetchedBySel;
7924 : :
7925 : : /* Add one page cpu-cost to the startup cost */
462 akorotkov@postgresql 7926 : 928 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
7927 : :
7928 : : /*
7929 : : * Add once again a CPU-cost for those data pages, before amortizing for
7930 : : * cache.
7931 : : */
7932 : 928 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
7933 : :
7934 : : /* Account for cache effects, the same as above */
4499 tgl@sss.pgh.pa.us 7935 [ + - + + ]: 928 : if (outer_scans > 1 || counts.arrayScans > 1)
7936 : : {
7937 : 3 : dataPagesFetched *= outer_scans * counts.arrayScans;
4928 7938 : 3 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
7939 : : (BlockNumber) numDataPages,
7940 : : numDataPages, root);
4499 7941 : 3 : dataPagesFetched /= outer_scans;
7942 : : }
7943 : :
7944 : : /* And apply random_page_cost as the cost per page */
462 akorotkov@postgresql 7945 : 928 : *indexTotalCost += *indexStartupCost +
4928 tgl@sss.pgh.pa.us 7946 : 928 : dataPagesFetched * spc_random_page_cost;
7947 : :
7948 : : /*
7949 : : * Add on index qual eval costs, much as in genericcostestimate. We charge
7950 : : * cpu but we can disregard indexorderbys, since GIN doesn't support
7951 : : * those.
7952 : : */
1885 7953 : 928 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
7954 : 928 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
7955 : :
4928 7956 : 928 : *indexStartupCost += qual_arg_cost;
7957 : 928 : *indexTotalCost += qual_arg_cost;
7958 : :
7959 : : /*
7960 : : * Add a cpu cost per search entry, corresponding to the actual visited
7961 : : * entries.
7962 : : */
462 akorotkov@postgresql 7963 : 928 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
7964 : : /* Now add a cpu cost per tuple in the posting lists / trees */
7965 : 928 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
2615 rhaas@postgresql.org 7966 : 928 : *indexPages = dataPagesFetched;
7967 : : }
7968 : :
7969 : : /*
7970 : : * BRIN has search behavior completely different from other index types
7971 : : */
7972 : : void
3010 tgl@sss.pgh.pa.us 7973 : 5365 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7974 : : Cost *indexStartupCost, Cost *indexTotalCost,
7975 : : Selectivity *indexSelectivity, double *indexCorrelation,
7976 : : double *indexPages)
7977 : : {
3446 alvherre@alvh.no-ip. 7978 : 5365 : IndexOptInfo *index = path->indexinfo;
1885 tgl@sss.pgh.pa.us 7979 : 5365 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
3446 alvherre@alvh.no-ip. 7980 : 5365 : double numPages = index->pages;
2565 7981 : 5365 : RelOptInfo *baserel = index->rel;
7982 [ + - ]: 5365 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
7983 : : Cost spc_seq_page_cost;
7984 : : Cost spc_random_page_cost;
7985 : : double qual_arg_cost;
7986 : : double qualSelectivity;
7987 : : BrinStatsData statsData;
7988 : : double indexRanges;
7989 : : double minimalRanges;
7990 : : double estimatedRanges;
7991 : : double selec;
7992 : : Relation indexRel;
7993 : : ListCell *l;
7994 : : VariableStatData vardata;
7995 : :
7996 [ - + ]: 5365 : Assert(rte->rtekind == RTE_RELATION);
7997 : :
7998 : : /* fetch estimated page cost for the tablespace containing the index */
3446 7999 : 5365 : get_tablespace_page_costs(index->reltablespace,
8000 : : &spc_random_page_cost,
8001 : : &spc_seq_page_cost);
8002 : :
8003 : : /*
8004 : : * Obtain some data from the index itself, if possible. Otherwise invent
8005 : : * some plausible internal statistics based on the relation page count.
8006 : : */
1606 michael@paquier.xyz 8007 [ + - ]: 5365 : if (!index->hypothetical)
8008 : : {
8009 : : /*
8010 : : * A lock should have already been obtained on the index in plancat.c.
8011 : : */
8012 : 5365 : indexRel = index_open(index->indexoid, NoLock);
8013 : 5365 : brinGetStats(indexRel, &statsData);
8014 : 5365 : index_close(indexRel, NoLock);
8015 : :
8016 : : /* work out the actual number of ranges in the index */
8017 [ + + ]: 5365 : indexRanges = Max(ceil((double) baserel->pages /
8018 : : statsData.pagesPerRange), 1.0);
8019 : : }
8020 : : else
8021 : : {
8022 : : /*
8023 : : * Assume default number of pages per range, and estimate the number
8024 : : * of ranges based on that.
8025 : : */
1606 michael@paquier.xyz 8026 [ # # ]:UBC 0 : indexRanges = Max(ceil((double) baserel->pages /
8027 : : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
8028 : :
8029 : 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
8030 : 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
8031 : : }
8032 : :
8033 : : /*
8034 : : * Compute index correlation
8035 : : *
8036 : : * Because we can use all index quals equally when scanning, we can use
8037 : : * the largest correlation (in absolute value) among columns used by the
8038 : : * query. Start at zero, the worst possible case. If we cannot find any
8039 : : * correlation statistics, we will keep it as 0.
8040 : : */
2565 alvherre@alvh.no-ip. 8041 :CBC 5365 : *indexCorrelation = 0;
8042 : :
1885 tgl@sss.pgh.pa.us 8043 [ + - + + : 10731 : foreach(l, path->indexclauses)
+ + ]
8044 : : {
8045 : 5366 : IndexClause *iclause = lfirst_node(IndexClause, l);
8046 : 5366 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
8047 : :
8048 : : /* attempt to lookup stats in relation for this index column */
2565 alvherre@alvh.no-ip. 8049 [ + - ]: 5366 : if (attnum != 0)
8050 : : {
8051 : : /* Simple variable -- look to stats for the underlying table */
8052 [ - + - - ]: 5366 : if (get_relation_stats_hook &&
2565 alvherre@alvh.no-ip. 8053 :UBC 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
8054 : : {
8055 : : /*
8056 : : * The hook took control of acquiring a stats tuple. If it
8057 : : * did supply a tuple, it'd better have supplied a freefunc.
8058 : : */
8059 [ # # # # ]: 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
8060 [ # # ]: 0 : elog(ERROR,
8061 : : "no function provided to release variable stats with");
8062 : : }
8063 : : else
8064 : : {
2565 alvherre@alvh.no-ip. 8065 :CBC 5366 : vardata.statsTuple =
8066 : 5366 : SearchSysCache3(STATRELATTINH,
8067 : : ObjectIdGetDatum(rte->relid),
8068 : : Int16GetDatum(attnum),
8069 : : BoolGetDatum(false));
8070 : 5366 : vardata.freefunc = ReleaseSysCache;
8071 : : }
8072 : : }
8073 : : else
8074 : : {
8075 : : /*
8076 : : * Looks like we've found an expression column in the index. Let's
8077 : : * see if there's any stats for it.
8078 : : */
8079 : :
8080 : : /* get the attnum from the 0-based index. */
1885 tgl@sss.pgh.pa.us 8081 :UBC 0 : attnum = iclause->indexcol + 1;
8082 : :
2565 alvherre@alvh.no-ip. 8083 [ # # # # ]: 0 : if (get_index_stats_hook &&
2489 tgl@sss.pgh.pa.us 8084 : 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
8085 : : {
8086 : : /*
8087 : : * The hook took control of acquiring a stats tuple. If it
8088 : : * did supply a tuple, it'd better have supplied a freefunc.
8089 : : */
2565 alvherre@alvh.no-ip. 8090 [ # # ]: 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
8091 [ # # ]: 0 : !vardata.freefunc)
8092 [ # # ]: 0 : elog(ERROR, "no function provided to release variable stats with");
8093 : : }
8094 : : else
8095 : : {
8096 : 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
8097 : : ObjectIdGetDatum(index->indexoid),
8098 : : Int16GetDatum(attnum),
8099 : : BoolGetDatum(false));
8100 : 0 : vardata.freefunc = ReleaseSysCache;
8101 : : }
8102 : : }
8103 : :
2565 alvherre@alvh.no-ip. 8104 [ + + ]:CBC 5366 : if (HeapTupleIsValid(vardata.statsTuple))
8105 : : {
8106 : : AttStatsSlot sslot;
8107 : :
2528 tgl@sss.pgh.pa.us 8108 [ + - ]: 18 : if (get_attstatsslot(&sslot, vardata.statsTuple,
8109 : : STATISTIC_KIND_CORRELATION, InvalidOid,
8110 : : ATTSTATSSLOT_NUMBERS))
8111 : : {
2565 alvherre@alvh.no-ip. 8112 : 18 : double varCorrelation = 0.0;
8113 : :
2528 tgl@sss.pgh.pa.us 8114 [ + - ]: 18 : if (sslot.nnumbers > 0)
555 peter@eisentraut.org 8115 : 18 : varCorrelation = fabs(sslot.numbers[0]);
8116 : :
2565 alvherre@alvh.no-ip. 8117 [ + - ]: 18 : if (varCorrelation > *indexCorrelation)
8118 : 18 : *indexCorrelation = varCorrelation;
8119 : :
2528 tgl@sss.pgh.pa.us 8120 : 18 : free_attstatsslot(&sslot);
8121 : : }
8122 : : }
8123 : :
2565 alvherre@alvh.no-ip. 8124 [ + + ]: 5366 : ReleaseVariableStats(vardata);
8125 : : }
8126 : :
8127 : 5365 : qualSelectivity = clauselist_selectivity(root, indexQuals,
8128 : 5365 : baserel->relid,
8129 : : JOIN_INNER, NULL);
8130 : :
8131 : : /*
8132 : : * Now calculate the minimum possible ranges we could match with if all of
8133 : : * the rows were in the perfect order in the table's heap.
8134 : : */
8135 : 5365 : minimalRanges = ceil(indexRanges * qualSelectivity);
8136 : :
8137 : : /*
8138 : : * Now estimate the number of ranges that we'll touch by using the
8139 : : * indexCorrelation from the stats. Careful not to divide by zero (note
8140 : : * we're using the absolute value of the correlation).
8141 : : */
8142 [ + + ]: 5365 : if (*indexCorrelation < 1.0e-10)
8143 : 5347 : estimatedRanges = indexRanges;
8144 : : else
8145 [ + + ]: 18 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
8146 : :
8147 : : /* we expect to visit this portion of the table */
8148 : 5365 : selec = estimatedRanges / indexRanges;
8149 : :
8150 [ - + - + ]: 5365 : CLAMP_PROBABILITY(selec);
8151 : :
8152 : 5365 : *indexSelectivity = selec;
8153 : :
8154 : : /*
8155 : : * Compute the index qual costs, much as in genericcostestimate, to add to
8156 : : * the index costs. We can disregard indexorderbys, since BRIN doesn't
8157 : : * support those.
8158 : : */
1885 tgl@sss.pgh.pa.us 8159 : 5365 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8160 : :
8161 : : /*
8162 : : * Compute the startup cost as the cost to read the whole revmap
8163 : : * sequentially, including the cost to execute the index quals.
8164 : : */
2565 alvherre@alvh.no-ip. 8165 : 5365 : *indexStartupCost =
8166 : 5365 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
3446 8167 : 5365 : *indexStartupCost += qual_arg_cost;
8168 : :
8169 : : /*
8170 : : * To read a BRIN index there might be a bit of back and forth over
8171 : : * regular pages, as revmap might point to them out of sequential order;
8172 : : * calculate the total cost as reading the whole index in random order.
8173 : : */
2565 8174 : 5365 : *indexTotalCost = *indexStartupCost +
8175 : 5365 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
8176 : :
8177 : : /*
8178 : : * Charge a small amount per range tuple which we expect to match to. This
8179 : : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
8180 : : * will set a bit for each page in the range when we find a matching
8181 : : * range, so we must multiply the charge by the number of pages in the
8182 : : * range.
8183 : : */
8184 : 5365 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
8185 : 5365 : statsData.pagesPerRange;
8186 : :
8187 : 5365 : *indexPages = index->pages;
3446 8188 : 5365 : }
|
