Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * extended_stats.c
4 : * POSTGRES extended statistics
5 : *
6 : * Generic code supporting statistics objects created via CREATE STATISTICS.
7 : *
8 : *
9 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
10 : * Portions Copyright (c) 1994, Regents of the University of California
11 : *
12 : * IDENTIFICATION
13 : * src/backend/statistics/extended_stats.c
14 : *
15 : *-------------------------------------------------------------------------
16 : */
17 : #include "postgres.h"
18 :
19 : #include "access/detoast.h"
20 : #include "access/genam.h"
21 : #include "access/htup_details.h"
22 : #include "access/table.h"
23 : #include "catalog/indexing.h"
24 : #include "catalog/pg_statistic_ext.h"
25 : #include "catalog/pg_statistic_ext_data.h"
26 : #include "commands/defrem.h"
27 : #include "commands/progress.h"
28 : #include "executor/executor.h"
29 : #include "miscadmin.h"
30 : #include "nodes/nodeFuncs.h"
31 : #include "optimizer/optimizer.h"
32 : #include "parser/parsetree.h"
33 : #include "pgstat.h"
34 : #include "postmaster/autovacuum.h"
35 : #include "statistics/extended_stats_internal.h"
36 : #include "statistics/statistics.h"
37 : #include "utils/acl.h"
38 : #include "utils/array.h"
39 : #include "utils/attoptcache.h"
40 : #include "utils/builtins.h"
41 : #include "utils/datum.h"
42 : #include "utils/fmgroids.h"
43 : #include "utils/lsyscache.h"
44 : #include "utils/memutils.h"
45 : #include "utils/rel.h"
46 : #include "utils/selfuncs.h"
47 : #include "utils/syscache.h"
48 :
49 : /*
50 : * To avoid consuming too much memory during analysis and/or too much space
51 : * in the resulting pg_statistic rows, we ignore varlena datums that are wider
52 : * than WIDTH_THRESHOLD (after detoasting!). This is legitimate for MCV
53 : * and distinct-value calculations since a wide value is unlikely to be
54 : * duplicated at all, much less be a most-common value. For the same reason,
55 : * ignoring wide values will not affect our estimates of histogram bin
56 : * boundaries very much.
57 : */
58 : #define WIDTH_THRESHOLD 1024
59 :
60 : /*
61 : * Used internally to refer to an individual statistics object, i.e.,
62 : * a pg_statistic_ext entry.
63 : */
64 : typedef struct StatExtEntry
65 : {
66 : Oid statOid; /* OID of pg_statistic_ext entry */
67 : char *schema; /* statistics object's schema */
68 : char *name; /* statistics object's name */
69 : Bitmapset *columns; /* attribute numbers covered by the object */
70 : List *types; /* 'char' list of enabled statistics kinds */
71 : int stattarget; /* statistics target (-1 for default) */
72 : List *exprs; /* expressions */
73 : } StatExtEntry;
74 :
75 :
76 : static List *fetch_statentries_for_relation(Relation pg_statext, Oid relid);
77 : static VacAttrStats **lookup_var_attr_stats(Bitmapset *attrs, List *exprs,
78 : int nvacatts, VacAttrStats **vacatts);
79 : static void statext_store(Oid statOid, bool inh,
80 : MVNDistinct *ndistinct, MVDependencies *dependencies,
81 : MCVList *mcv, Datum exprs, VacAttrStats **stats);
82 : static int statext_compute_stattarget(int stattarget,
83 : int nattrs, VacAttrStats **stats);
84 :
85 : /* Information needed to analyze a single simple expression. */
86 : typedef struct AnlExprData
87 : {
88 : Node *expr; /* expression to analyze */
89 : VacAttrStats *vacattrstat; /* statistics attrs to analyze */
90 : } AnlExprData;
91 :
92 : static void compute_expr_stats(Relation onerel, AnlExprData *exprdata,
93 : int nexprs, HeapTuple *rows, int numrows);
94 : static Datum serialize_expr_stats(AnlExprData *exprdata, int nexprs);
95 : static Datum expr_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
96 : static AnlExprData *build_expr_data(List *exprs, int stattarget);
97 :
98 : static StatsBuildData *make_build_data(Relation rel, StatExtEntry *stat,
99 : int numrows, HeapTuple *rows,
100 : VacAttrStats **stats, int stattarget);
101 :
102 :
103 : /*
104 : * Compute requested extended stats, using the rows sampled for the plain
105 : * (single-column) stats.
106 : *
107 : * This fetches a list of stats types from pg_statistic_ext, computes the
108 : * requested stats, and serializes them back into the catalog.
109 : */
110 : void
111 5736 : BuildRelationExtStatistics(Relation onerel, bool inh, double totalrows,
112 : int numrows, HeapTuple *rows,
113 : int natts, VacAttrStats **vacattrstats)
114 : {
115 : Relation pg_stext;
116 : ListCell *lc;
117 : List *statslist;
118 : MemoryContext cxt;
119 : MemoryContext oldcxt;
120 : int64 ext_cnt;
121 :
122 : /* Do nothing if there are no columns to analyze. */
123 5736 : if (!natts)
124 9 : return;
125 :
126 : /* the list of stats has to be allocated outside the memory context */
127 5727 : pg_stext = table_open(StatisticExtRelationId, RowExclusiveLock);
128 5727 : statslist = fetch_statentries_for_relation(pg_stext, RelationGetRelid(onerel));
129 :
130 : /* memory context for building each statistics object */
131 5727 : cxt = AllocSetContextCreate(CurrentMemoryContext,
132 : "BuildRelationExtStatistics",
133 : ALLOCSET_DEFAULT_SIZES);
134 5727 : oldcxt = MemoryContextSwitchTo(cxt);
135 :
136 : /* report this phase */
137 5727 : if (statslist != NIL)
138 : {
139 173 : const int index[] = {
140 : PROGRESS_ANALYZE_PHASE,
141 : PROGRESS_ANALYZE_EXT_STATS_TOTAL
142 : };
143 346 : const int64 val[] = {
144 : PROGRESS_ANALYZE_PHASE_COMPUTE_EXT_STATS,
145 173 : list_length(statslist)
146 : };
147 :
148 173 : pgstat_progress_update_multi_param(2, index, val);
149 : }
150 :
151 5727 : ext_cnt = 0;
152 5993 : foreach(lc, statslist)
153 : {
154 266 : StatExtEntry *stat = (StatExtEntry *) lfirst(lc);
155 266 : MVNDistinct *ndistinct = NULL;
156 266 : MVDependencies *dependencies = NULL;
157 266 : MCVList *mcv = NULL;
158 266 : Datum exprstats = (Datum) 0;
159 : VacAttrStats **stats;
160 : ListCell *lc2;
161 : int stattarget;
162 : StatsBuildData *data;
163 :
164 : /*
165 : * Check if we can build these stats based on the column analyzed. If
166 : * not, report this fact (except in autovacuum) and move on.
167 : */
168 266 : stats = lookup_var_attr_stats(stat->columns, stat->exprs,
169 : natts, vacattrstats);
170 266 : if (!stats)
171 : {
172 6 : if (!AmAutoVacuumWorkerProcess())
173 6 : ereport(WARNING,
174 : (errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
175 : errmsg("statistics object \"%s.%s\" could not be computed for relation \"%s.%s\"",
176 : stat->schema, stat->name,
177 : get_namespace_name(onerel->rd_rel->relnamespace),
178 : RelationGetRelationName(onerel)),
179 : errtable(onerel)));
180 6 : continue;
181 : }
182 :
183 : /* compute statistics target for this statistics object */
184 260 : stattarget = statext_compute_stattarget(stat->stattarget,
185 260 : bms_num_members(stat->columns),
186 : stats);
187 :
188 : /*
189 : * Don't rebuild statistics objects with statistics target set to 0
190 : * (we just leave the existing values around, just like we do for
191 : * regular per-column statistics).
192 : */
193 260 : if (stattarget == 0)
194 3 : continue;
195 :
196 : /* evaluate expressions (if the statistics object has any) */
197 257 : data = make_build_data(onerel, stat, numrows, rows, stats, stattarget);
198 :
199 : /* compute statistic of each requested type */
200 730 : foreach(lc2, stat->types)
201 : {
202 473 : char t = (char) lfirst_int(lc2);
203 :
204 473 : if (t == STATS_EXT_NDISTINCT)
205 125 : ndistinct = statext_ndistinct_build(totalrows, data);
206 348 : else if (t == STATS_EXT_DEPENDENCIES)
207 98 : dependencies = statext_dependencies_build(data);
208 250 : else if (t == STATS_EXT_MCV)
209 137 : mcv = statext_mcv_build(data, totalrows, stattarget);
210 113 : else if (t == STATS_EXT_EXPRESSIONS)
211 : {
212 : AnlExprData *exprdata;
213 : int nexprs;
214 :
215 : /* should not happen, thanks to checks when defining stats */
216 113 : if (!stat->exprs)
217 0 : elog(ERROR, "requested expression stats, but there are no expressions");
218 :
219 113 : exprdata = build_expr_data(stat->exprs, stattarget);
220 113 : nexprs = list_length(stat->exprs);
221 :
222 113 : compute_expr_stats(onerel, exprdata, nexprs, rows, numrows);
223 :
224 113 : exprstats = serialize_expr_stats(exprdata, nexprs);
225 : }
226 : }
227 :
228 : /* store the statistics in the catalog */
229 257 : statext_store(stat->statOid, inh,
230 : ndistinct, dependencies, mcv, exprstats, stats);
231 :
232 : /* for reporting progress */
233 257 : pgstat_progress_update_param(PROGRESS_ANALYZE_EXT_STATS_COMPUTED,
234 : ++ext_cnt);
235 :
236 : /* free the data used for building this statistics object */
237 257 : MemoryContextReset(cxt);
238 : }
239 :
240 5727 : MemoryContextSwitchTo(oldcxt);
241 5727 : MemoryContextDelete(cxt);
242 :
243 5727 : list_free(statslist);
244 :
245 5727 : table_close(pg_stext, RowExclusiveLock);
246 : }
247 :
248 : /*
249 : * ComputeExtStatisticsRows
250 : * Compute number of rows required by extended statistics on a table.
251 : *
252 : * Computes number of rows we need to sample to build extended statistics on a
253 : * table. This only looks at statistics we can actually build - for example
254 : * when analyzing only some of the columns, this will skip statistics objects
255 : * that would require additional columns.
256 : *
257 : * See statext_compute_stattarget for details about how we compute the
258 : * statistics target for a statistics object (from the object target,
259 : * attribute targets and default statistics target).
260 : */
261 : int
262 8460 : ComputeExtStatisticsRows(Relation onerel,
263 : int natts, VacAttrStats **vacattrstats)
264 : {
265 : Relation pg_stext;
266 : ListCell *lc;
267 : List *lstats;
268 : MemoryContext cxt;
269 : MemoryContext oldcxt;
270 8460 : int result = 0;
271 :
272 : /* If there are no columns to analyze, just return 0. */
273 8460 : if (!natts)
274 36 : return 0;
275 :
276 8424 : cxt = AllocSetContextCreate(CurrentMemoryContext,
277 : "ComputeExtStatisticsRows",
278 : ALLOCSET_DEFAULT_SIZES);
279 8424 : oldcxt = MemoryContextSwitchTo(cxt);
280 :
281 8424 : pg_stext = table_open(StatisticExtRelationId, RowExclusiveLock);
282 8424 : lstats = fetch_statentries_for_relation(pg_stext, RelationGetRelid(onerel));
283 :
284 8690 : foreach(lc, lstats)
285 : {
286 266 : StatExtEntry *stat = (StatExtEntry *) lfirst(lc);
287 : int stattarget;
288 : VacAttrStats **stats;
289 266 : int nattrs = bms_num_members(stat->columns);
290 :
291 : /*
292 : * Check if we can build this statistics object based on the columns
293 : * analyzed. If not, ignore it (don't report anything, we'll do that
294 : * during the actual build BuildRelationExtStatistics).
295 : */
296 266 : stats = lookup_var_attr_stats(stat->columns, stat->exprs,
297 : natts, vacattrstats);
298 :
299 266 : if (!stats)
300 6 : continue;
301 :
302 : /*
303 : * Compute statistics target, based on what's set for the statistic
304 : * object itself, and for its attributes.
305 : */
306 260 : stattarget = statext_compute_stattarget(stat->stattarget,
307 : nattrs, stats);
308 :
309 : /* Use the largest value for all statistics objects. */
310 260 : if (stattarget > result)
311 164 : result = stattarget;
312 : }
313 :
314 8424 : table_close(pg_stext, RowExclusiveLock);
315 :
316 8424 : MemoryContextSwitchTo(oldcxt);
317 8424 : MemoryContextDelete(cxt);
318 :
319 : /* compute sample size based on the statistics target */
320 8424 : return (300 * result);
321 : }
322 :
323 : /*
324 : * statext_compute_stattarget
325 : * compute statistics target for an extended statistic
326 : *
327 : * When computing target for extended statistics objects, we consider three
328 : * places where the target may be set - the statistics object itself,
329 : * attributes the statistics object is defined on, and then the default
330 : * statistics target.
331 : *
332 : * First we look at what's set for the statistics object itself, using the
333 : * ALTER STATISTICS ... SET STATISTICS command. If we find a valid value
334 : * there (i.e. not -1) we're done. Otherwise we look at targets set for any
335 : * of the attributes the statistic is defined on, and if there are columns
336 : * with defined target, we use the maximum value. We do this mostly for
337 : * backwards compatibility, because this is what we did before having
338 : * statistics target for extended statistics.
339 : *
340 : * And finally, if we still don't have a statistics target, we use the value
341 : * set in default_statistics_target.
342 : */
343 : static int
344 520 : statext_compute_stattarget(int stattarget, int nattrs, VacAttrStats **stats)
345 : {
346 : int i;
347 :
348 : /*
349 : * If there's statistics target set for the statistics object, use it. It
350 : * may be set to 0 which disables building of that statistic.
351 : */
352 520 : if (stattarget >= 0)
353 6 : return stattarget;
354 :
355 : /*
356 : * The target for the statistics object is set to -1, in which case we
357 : * look at the maximum target set for any of the attributes the object is
358 : * defined on.
359 : */
360 1358 : for (i = 0; i < nattrs; i++)
361 : {
362 : /* keep the maximum statistics target */
363 844 : if (stats[i]->attstattarget > stattarget)
364 382 : stattarget = stats[i]->attstattarget;
365 : }
366 :
367 : /*
368 : * If the value is still negative (so neither the statistics object nor
369 : * any of the columns have custom statistics target set), use the global
370 : * default target.
371 : */
372 514 : if (stattarget < 0)
373 132 : stattarget = default_statistics_target;
374 :
375 : /* As this point we should have a valid statistics target. */
376 : Assert((stattarget >= 0) && (stattarget <= MAX_STATISTICS_TARGET));
377 :
378 514 : return stattarget;
379 : }
380 :
381 : /*
382 : * statext_is_kind_built
383 : * Is this stat kind built in the given pg_statistic_ext_data tuple?
384 : */
385 : bool
386 4016 : statext_is_kind_built(HeapTuple htup, char type)
387 : {
388 : AttrNumber attnum;
389 :
390 4016 : switch (type)
391 : {
392 1004 : case STATS_EXT_NDISTINCT:
393 1004 : attnum = Anum_pg_statistic_ext_data_stxdndistinct;
394 1004 : break;
395 :
396 1004 : case STATS_EXT_DEPENDENCIES:
397 1004 : attnum = Anum_pg_statistic_ext_data_stxddependencies;
398 1004 : break;
399 :
400 1004 : case STATS_EXT_MCV:
401 1004 : attnum = Anum_pg_statistic_ext_data_stxdmcv;
402 1004 : break;
403 :
404 1004 : case STATS_EXT_EXPRESSIONS:
405 1004 : attnum = Anum_pg_statistic_ext_data_stxdexpr;
406 1004 : break;
407 :
408 0 : default:
409 0 : elog(ERROR, "unexpected statistics type requested: %d", type);
410 : }
411 :
412 4016 : return !heap_attisnull(htup, attnum, NULL);
413 : }
414 :
415 : /*
416 : * Return a list (of StatExtEntry) of statistics objects for the given relation.
417 : */
418 : static List *
419 14151 : fetch_statentries_for_relation(Relation pg_statext, Oid relid)
420 : {
421 : SysScanDesc scan;
422 : ScanKeyData skey;
423 : HeapTuple htup;
424 14151 : List *result = NIL;
425 :
426 : /*
427 : * Prepare to scan pg_statistic_ext for entries having stxrelid = this
428 : * rel.
429 : */
430 14151 : ScanKeyInit(&skey,
431 : Anum_pg_statistic_ext_stxrelid,
432 : BTEqualStrategyNumber, F_OIDEQ,
433 : ObjectIdGetDatum(relid));
434 :
435 14151 : scan = systable_beginscan(pg_statext, StatisticExtRelidIndexId, true,
436 : NULL, 1, &skey);
437 :
438 14683 : while (HeapTupleIsValid(htup = systable_getnext(scan)))
439 : {
440 : StatExtEntry *entry;
441 : Datum datum;
442 : bool isnull;
443 : int i;
444 : ArrayType *arr;
445 : char *enabled;
446 : Form_pg_statistic_ext staForm;
447 532 : List *exprs = NIL;
448 :
449 532 : entry = palloc0_object(StatExtEntry);
450 532 : staForm = (Form_pg_statistic_ext) GETSTRUCT(htup);
451 532 : entry->statOid = staForm->oid;
452 532 : entry->schema = get_namespace_name(staForm->stxnamespace);
453 532 : entry->name = pstrdup(NameStr(staForm->stxname));
454 1412 : for (i = 0; i < staForm->stxkeys.dim1; i++)
455 : {
456 880 : entry->columns = bms_add_member(entry->columns,
457 880 : staForm->stxkeys.values[i]);
458 : }
459 :
460 532 : datum = SysCacheGetAttr(STATEXTOID, htup, Anum_pg_statistic_ext_stxstattarget, &isnull);
461 532 : entry->stattarget = isnull ? -1 : DatumGetInt16(datum);
462 :
463 : /* decode the stxkind char array into a list of chars */
464 532 : datum = SysCacheGetAttrNotNull(STATEXTOID, htup,
465 : Anum_pg_statistic_ext_stxkind);
466 532 : arr = DatumGetArrayTypeP(datum);
467 532 : if (ARR_NDIM(arr) != 1 ||
468 532 : ARR_HASNULL(arr) ||
469 532 : ARR_ELEMTYPE(arr) != CHAROID)
470 0 : elog(ERROR, "stxkind is not a 1-D char array");
471 532 : enabled = (char *) ARR_DATA_PTR(arr);
472 1532 : for (i = 0; i < ARR_DIMS(arr)[0]; i++)
473 : {
474 : Assert((enabled[i] == STATS_EXT_NDISTINCT) ||
475 : (enabled[i] == STATS_EXT_DEPENDENCIES) ||
476 : (enabled[i] == STATS_EXT_MCV) ||
477 : (enabled[i] == STATS_EXT_EXPRESSIONS));
478 1000 : entry->types = lappend_int(entry->types, (int) enabled[i]);
479 : }
480 :
481 : /* decode expression (if any) */
482 532 : datum = SysCacheGetAttr(STATEXTOID, htup,
483 : Anum_pg_statistic_ext_stxexprs, &isnull);
484 :
485 532 : if (!isnull)
486 : {
487 : char *exprsString;
488 :
489 226 : exprsString = TextDatumGetCString(datum);
490 226 : exprs = (List *) stringToNode(exprsString);
491 :
492 226 : pfree(exprsString);
493 :
494 : /*
495 : * Run the expressions through eval_const_expressions. This is not
496 : * just an optimization, but is necessary, because the planner
497 : * will be comparing them to similarly-processed qual clauses, and
498 : * may fail to detect valid matches without this. We must not use
499 : * canonicalize_qual, however, since these aren't qual
500 : * expressions.
501 : */
502 226 : exprs = (List *) eval_const_expressions(NULL, (Node *) exprs);
503 :
504 : /* May as well fix opfuncids too */
505 226 : fix_opfuncids((Node *) exprs);
506 : }
507 :
508 532 : entry->exprs = exprs;
509 :
510 532 : result = lappend(result, entry);
511 : }
512 :
513 14151 : systable_endscan(scan);
514 :
515 14151 : return result;
516 : }
517 :
518 : /*
519 : * examine_attribute -- pre-analysis of a single column
520 : *
521 : * Determine whether the column is analyzable; if so, create and initialize
522 : * a VacAttrStats struct for it. If not, return NULL.
523 : */
524 : static VacAttrStats *
525 418 : examine_attribute(Node *expr)
526 : {
527 : HeapTuple typtuple;
528 : VacAttrStats *stats;
529 : int i;
530 : bool ok;
531 :
532 : /*
533 : * Create the VacAttrStats struct.
534 : */
535 418 : stats = palloc0_object(VacAttrStats);
536 418 : stats->attstattarget = -1;
537 :
538 : /*
539 : * When analyzing an expression, believe the expression tree's type not
540 : * the column datatype --- the latter might be the opckeytype storage type
541 : * of the opclass, which is not interesting for our purposes. (Note: if
542 : * we did anything with non-expression statistics columns, we'd need to
543 : * figure out where to get the correct type info from, but for now that's
544 : * not a problem.) It's not clear whether anyone will care about the
545 : * typmod, but we store that too just in case.
546 : */
547 418 : stats->attrtypid = exprType(expr);
548 418 : stats->attrtypmod = exprTypmod(expr);
549 418 : stats->attrcollid = exprCollation(expr);
550 :
551 418 : typtuple = SearchSysCacheCopy1(TYPEOID,
552 : ObjectIdGetDatum(stats->attrtypid));
553 418 : if (!HeapTupleIsValid(typtuple))
554 0 : elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
555 418 : stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
556 :
557 : /*
558 : * We don't actually analyze individual attributes, so no need to set the
559 : * memory context.
560 : */
561 418 : stats->anl_context = NULL;
562 418 : stats->tupattnum = InvalidAttrNumber;
563 :
564 : /*
565 : * The fields describing the stats->stavalues[n] element types default to
566 : * the type of the data being analyzed, but the type-specific typanalyze
567 : * function can change them if it wants to store something else.
568 : */
569 2508 : for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
570 : {
571 2090 : stats->statypid[i] = stats->attrtypid;
572 2090 : stats->statyplen[i] = stats->attrtype->typlen;
573 2090 : stats->statypbyval[i] = stats->attrtype->typbyval;
574 2090 : stats->statypalign[i] = stats->attrtype->typalign;
575 : }
576 :
577 : /*
578 : * Call the type-specific typanalyze function. If none is specified, use
579 : * std_typanalyze().
580 : */
581 418 : if (OidIsValid(stats->attrtype->typanalyze))
582 6 : ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
583 : PointerGetDatum(stats)));
584 : else
585 412 : ok = std_typanalyze(stats);
586 :
587 418 : if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
588 : {
589 0 : heap_freetuple(typtuple);
590 0 : pfree(stats);
591 0 : return NULL;
592 : }
593 :
594 418 : return stats;
595 : }
596 :
597 : /*
598 : * examine_expression -- pre-analysis of a single expression
599 : *
600 : * Determine whether the expression is analyzable; if so, create and initialize
601 : * a VacAttrStats struct for it. If not, return NULL.
602 : */
603 : static VacAttrStats *
604 418 : examine_expression(Node *expr, int stattarget)
605 : {
606 : HeapTuple typtuple;
607 : VacAttrStats *stats;
608 : int i;
609 : bool ok;
610 :
611 : Assert(expr != NULL);
612 :
613 : /*
614 : * Create the VacAttrStats struct.
615 : */
616 418 : stats = palloc0_object(VacAttrStats);
617 :
618 : /*
619 : * We can't have statistics target specified for the expression, so we
620 : * could use either the default_statistics_target, or the target computed
621 : * for the extended statistics. The second option seems more reasonable.
622 : */
623 418 : stats->attstattarget = stattarget;
624 :
625 : /*
626 : * When analyzing an expression, believe the expression tree's type.
627 : */
628 418 : stats->attrtypid = exprType(expr);
629 418 : stats->attrtypmod = exprTypmod(expr);
630 :
631 : /*
632 : * We don't allow collation to be specified in CREATE STATISTICS, so we
633 : * have to use the collation specified for the expression. It's possible
634 : * to specify the collation in the expression "(col COLLATE "en_US")" in
635 : * which case exprCollation() does the right thing.
636 : */
637 418 : stats->attrcollid = exprCollation(expr);
638 :
639 418 : typtuple = SearchSysCacheCopy1(TYPEOID,
640 : ObjectIdGetDatum(stats->attrtypid));
641 418 : if (!HeapTupleIsValid(typtuple))
642 0 : elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
643 :
644 418 : stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
645 418 : stats->anl_context = CurrentMemoryContext; /* XXX should be using
646 : * something else? */
647 418 : stats->tupattnum = InvalidAttrNumber;
648 :
649 : /*
650 : * The fields describing the stats->stavalues[n] element types default to
651 : * the type of the data being analyzed, but the type-specific typanalyze
652 : * function can change them if it wants to store something else.
653 : */
654 2508 : for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
655 : {
656 2090 : stats->statypid[i] = stats->attrtypid;
657 2090 : stats->statyplen[i] = stats->attrtype->typlen;
658 2090 : stats->statypbyval[i] = stats->attrtype->typbyval;
659 2090 : stats->statypalign[i] = stats->attrtype->typalign;
660 : }
661 :
662 : /*
663 : * Call the type-specific typanalyze function. If none is specified, use
664 : * std_typanalyze().
665 : */
666 418 : if (OidIsValid(stats->attrtype->typanalyze))
667 6 : ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
668 : PointerGetDatum(stats)));
669 : else
670 412 : ok = std_typanalyze(stats);
671 :
672 418 : if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
673 : {
674 0 : heap_freetuple(typtuple);
675 0 : pfree(stats);
676 0 : return NULL;
677 : }
678 :
679 418 : return stats;
680 : }
681 :
682 : /*
683 : * Using 'vacatts' of size 'nvacatts' as input data, return a newly-built
684 : * VacAttrStats array which includes only the items corresponding to
685 : * attributes indicated by 'attrs'. If we don't have all of the per-column
686 : * stats available to compute the extended stats, then we return NULL to
687 : * indicate to the caller that the stats should not be built.
688 : */
689 : static VacAttrStats **
690 532 : lookup_var_attr_stats(Bitmapset *attrs, List *exprs,
691 : int nvacatts, VacAttrStats **vacatts)
692 : {
693 532 : int i = 0;
694 532 : int x = -1;
695 : int natts;
696 : VacAttrStats **stats;
697 : ListCell *lc;
698 :
699 532 : natts = bms_num_members(attrs) + list_length(exprs);
700 :
701 532 : stats = (VacAttrStats **) palloc(natts * sizeof(VacAttrStats *));
702 :
703 : /* lookup VacAttrStats info for the requested columns (same attnum) */
704 1394 : while ((x = bms_next_member(attrs, x)) >= 0)
705 : {
706 : int j;
707 :
708 874 : stats[i] = NULL;
709 2512 : for (j = 0; j < nvacatts; j++)
710 : {
711 2500 : if (x == vacatts[j]->tupattnum)
712 : {
713 862 : stats[i] = vacatts[j];
714 862 : break;
715 : }
716 : }
717 :
718 874 : if (!stats[i])
719 : {
720 : /*
721 : * Looks like stats were not gathered for one of the columns
722 : * required. We'll be unable to build the extended stats without
723 : * this column.
724 : */
725 12 : pfree(stats);
726 12 : return NULL;
727 : }
728 :
729 862 : i++;
730 : }
731 :
732 : /* also add info for expressions */
733 938 : foreach(lc, exprs)
734 : {
735 418 : Node *expr = (Node *) lfirst(lc);
736 :
737 418 : stats[i] = examine_attribute(expr);
738 :
739 : /*
740 : * XXX We need tuple descriptor later, and we just grab it from
741 : * stats[0]->tupDesc (see e.g. statext_mcv_build). But as coded
742 : * examine_attribute does not set that, so just grab it from the first
743 : * vacatts element.
744 : */
745 418 : stats[i]->tupDesc = vacatts[0]->tupDesc;
746 :
747 418 : i++;
748 : }
749 :
750 520 : return stats;
751 : }
752 :
753 : /*
754 : * statext_store
755 : * Serializes the statistics and stores them into the pg_statistic_ext_data
756 : * tuple.
757 : */
758 : static void
759 257 : statext_store(Oid statOid, bool inh,
760 : MVNDistinct *ndistinct, MVDependencies *dependencies,
761 : MCVList *mcv, Datum exprs, VacAttrStats **stats)
762 : {
763 : Relation pg_stextdata;
764 : HeapTuple stup;
765 : Datum values[Natts_pg_statistic_ext_data];
766 : bool nulls[Natts_pg_statistic_ext_data];
767 :
768 257 : pg_stextdata = table_open(StatisticExtDataRelationId, RowExclusiveLock);
769 :
770 257 : memset(nulls, true, sizeof(nulls));
771 257 : memset(values, 0, sizeof(values));
772 :
773 : /* basic info */
774 257 : values[Anum_pg_statistic_ext_data_stxoid - 1] = ObjectIdGetDatum(statOid);
775 257 : nulls[Anum_pg_statistic_ext_data_stxoid - 1] = false;
776 :
777 257 : values[Anum_pg_statistic_ext_data_stxdinherit - 1] = BoolGetDatum(inh);
778 257 : nulls[Anum_pg_statistic_ext_data_stxdinherit - 1] = false;
779 :
780 : /*
781 : * Construct a new pg_statistic_ext_data tuple, replacing the calculated
782 : * stats.
783 : */
784 257 : if (ndistinct != NULL)
785 : {
786 125 : bytea *data = statext_ndistinct_serialize(ndistinct);
787 :
788 125 : nulls[Anum_pg_statistic_ext_data_stxdndistinct - 1] = (data == NULL);
789 125 : values[Anum_pg_statistic_ext_data_stxdndistinct - 1] = PointerGetDatum(data);
790 : }
791 :
792 257 : if (dependencies != NULL)
793 : {
794 89 : bytea *data = statext_dependencies_serialize(dependencies);
795 :
796 89 : nulls[Anum_pg_statistic_ext_data_stxddependencies - 1] = (data == NULL);
797 89 : values[Anum_pg_statistic_ext_data_stxddependencies - 1] = PointerGetDatum(data);
798 : }
799 257 : if (mcv != NULL)
800 : {
801 137 : bytea *data = statext_mcv_serialize(mcv, stats);
802 :
803 137 : nulls[Anum_pg_statistic_ext_data_stxdmcv - 1] = (data == NULL);
804 137 : values[Anum_pg_statistic_ext_data_stxdmcv - 1] = PointerGetDatum(data);
805 : }
806 257 : if (exprs != (Datum) 0)
807 : {
808 113 : nulls[Anum_pg_statistic_ext_data_stxdexpr - 1] = false;
809 113 : values[Anum_pg_statistic_ext_data_stxdexpr - 1] = exprs;
810 : }
811 :
812 : /*
813 : * Delete the old tuple if it exists, and insert a new one. It's easier
814 : * than trying to update or insert, based on various conditions.
815 : */
816 257 : RemoveStatisticsDataById(statOid, inh);
817 :
818 : /* form and insert a new tuple */
819 257 : stup = heap_form_tuple(RelationGetDescr(pg_stextdata), values, nulls);
820 257 : CatalogTupleInsert(pg_stextdata, stup);
821 :
822 257 : heap_freetuple(stup);
823 :
824 257 : table_close(pg_stextdata, RowExclusiveLock);
825 257 : }
826 :
827 : /* initialize multi-dimensional sort */
828 : MultiSortSupport
829 1034 : multi_sort_init(int ndims)
830 : {
831 : MultiSortSupport mss;
832 :
833 : Assert(ndims >= 2);
834 :
835 1034 : mss = (MultiSortSupport) palloc0(offsetof(MultiSortSupportData, ssup)
836 1034 : + sizeof(SortSupportData) * ndims);
837 :
838 1034 : mss->ndims = ndims;
839 :
840 1034 : return mss;
841 : }
842 :
843 : /*
844 : * Prepare sort support info using the given sort operator and collation
845 : * at the position 'sortdim'
846 : */
847 : void
848 2479 : multi_sort_add_dimension(MultiSortSupport mss, int sortdim,
849 : Oid oper, Oid collation)
850 : {
851 2479 : SortSupport ssup = &mss->ssup[sortdim];
852 :
853 2479 : ssup->ssup_cxt = CurrentMemoryContext;
854 2479 : ssup->ssup_collation = collation;
855 2479 : ssup->ssup_nulls_first = false;
856 :
857 2479 : PrepareSortSupportFromOrderingOp(oper, ssup);
858 2479 : }
859 :
860 : /* compare all the dimensions in the selected order */
861 : int
862 11080822 : multi_sort_compare(const void *a, const void *b, void *arg)
863 : {
864 11080822 : MultiSortSupport mss = (MultiSortSupport) arg;
865 11080822 : const SortItem *ia = a;
866 11080822 : const SortItem *ib = b;
867 : int i;
868 :
869 19620991 : for (i = 0; i < mss->ndims; i++)
870 : {
871 : int compare;
872 :
873 17546035 : compare = ApplySortComparator(ia->values[i], ia->isnull[i],
874 17546035 : ib->values[i], ib->isnull[i],
875 17546035 : &mss->ssup[i]);
876 :
877 17546035 : if (compare != 0)
878 9005866 : return compare;
879 : }
880 :
881 : /* equal by default */
882 2074956 : return 0;
883 : }
884 :
885 : /* compare selected dimension */
886 : int
887 736575 : multi_sort_compare_dim(int dim, const SortItem *a, const SortItem *b,
888 : MultiSortSupport mss)
889 : {
890 1473150 : return ApplySortComparator(a->values[dim], a->isnull[dim],
891 736575 : b->values[dim], b->isnull[dim],
892 736575 : &mss->ssup[dim]);
893 : }
894 :
895 : int
896 753329 : multi_sort_compare_dims(int start, int end,
897 : const SortItem *a, const SortItem *b,
898 : MultiSortSupport mss)
899 : {
900 : int dim;
901 :
902 1703183 : for (dim = start; dim <= end; dim++)
903 : {
904 966608 : int r = ApplySortComparator(a->values[dim], a->isnull[dim],
905 966608 : b->values[dim], b->isnull[dim],
906 966608 : &mss->ssup[dim]);
907 :
908 966608 : if (r != 0)
909 16754 : return r;
910 : }
911 :
912 736575 : return 0;
913 : }
914 :
915 : int
916 104566 : compare_scalars_simple(const void *a, const void *b, void *arg)
917 : {
918 104566 : return compare_datums_simple(*(const Datum *) a,
919 : *(const Datum *) b,
920 : (SortSupport) arg);
921 : }
922 :
923 : int
924 118570 : compare_datums_simple(Datum a, Datum b, SortSupport ssup)
925 : {
926 118570 : return ApplySortComparator(a, false, b, false, ssup);
927 : }
928 :
929 : /*
930 : * build_attnums_array
931 : * Transforms a bitmap into an array of AttrNumber values.
932 : *
933 : * This is used for extended statistics only, so all the attributes must be
934 : * user-defined. That means offsetting by FirstLowInvalidHeapAttributeNumber
935 : * is not necessary here (and when querying the bitmap).
936 : */
937 : AttrNumber *
938 0 : build_attnums_array(Bitmapset *attrs, int nexprs, int *numattrs)
939 : {
940 : int i,
941 : j;
942 : AttrNumber *attnums;
943 0 : int num = bms_num_members(attrs);
944 :
945 0 : if (numattrs)
946 0 : *numattrs = num;
947 :
948 : /* build attnums from the bitmapset */
949 0 : attnums = palloc_array(AttrNumber, num);
950 0 : i = 0;
951 0 : j = -1;
952 0 : while ((j = bms_next_member(attrs, j)) >= 0)
953 : {
954 0 : int attnum = (j - nexprs);
955 :
956 : /*
957 : * Make sure the bitmap contains only user-defined attributes. As
958 : * bitmaps can't contain negative values, this can be violated in two
959 : * ways. Firstly, the bitmap might contain 0 as a member, and secondly
960 : * the integer value might be larger than MaxAttrNumber.
961 : */
962 : Assert(AttributeNumberIsValid(attnum));
963 : Assert(attnum <= MaxAttrNumber);
964 : Assert(attnum >= (-nexprs));
965 :
966 0 : attnums[i++] = (AttrNumber) attnum;
967 :
968 : /* protect against overflows */
969 : Assert(i <= num);
970 : }
971 :
972 0 : return attnums;
973 : }
974 :
975 : /*
976 : * build_sorted_items
977 : * build a sorted array of SortItem with values from rows
978 : *
979 : * Note: All the memory is allocated in a single chunk, so that the caller
980 : * can simply pfree the return value to release all of it.
981 : */
982 : SortItem *
983 657 : build_sorted_items(StatsBuildData *data, int *nitems,
984 : MultiSortSupport mss,
985 : int numattrs, AttrNumber *attnums)
986 : {
987 : int i,
988 : j,
989 : nrows;
990 657 : int nvalues = data->numrows * numattrs;
991 : Size len;
992 : SortItem *items;
993 : Datum *values;
994 : bool *isnull;
995 : char *ptr;
996 : int *typlen;
997 :
998 : /* Compute the total amount of memory we need (both items and values). */
999 657 : len = MAXALIGN(data->numrows * sizeof(SortItem)) +
1000 657 : nvalues * (sizeof(Datum) + sizeof(bool));
1001 :
1002 : /* Allocate the memory and split it into the pieces. */
1003 657 : ptr = palloc0(len);
1004 :
1005 : /* items to sort */
1006 657 : items = (SortItem *) ptr;
1007 : /* MAXALIGN ensures that the following Datums are suitably aligned */
1008 657 : ptr += MAXALIGN(data->numrows * sizeof(SortItem));
1009 :
1010 : /* values and null flags */
1011 657 : values = (Datum *) ptr;
1012 657 : ptr += nvalues * sizeof(Datum);
1013 :
1014 657 : isnull = (bool *) ptr;
1015 657 : ptr += nvalues * sizeof(bool);
1016 :
1017 : /* make sure we consumed the whole buffer exactly */
1018 : Assert((ptr - (char *) items) == len);
1019 :
1020 : /* fix the pointers to Datum and bool arrays */
1021 657 : nrows = 0;
1022 996438 : for (i = 0; i < data->numrows; i++)
1023 : {
1024 995781 : items[nrows].values = &values[nrows * numattrs];
1025 995781 : items[nrows].isnull = &isnull[nrows * numattrs];
1026 :
1027 995781 : nrows++;
1028 : }
1029 :
1030 : /* build a local cache of typlen for all attributes */
1031 657 : typlen = palloc_array(int, data->nattnums);
1032 2610 : for (i = 0; i < data->nattnums; i++)
1033 1953 : typlen[i] = get_typlen(data->stats[i]->attrtypid);
1034 :
1035 657 : nrows = 0;
1036 996438 : for (i = 0; i < data->numrows; i++)
1037 : {
1038 995781 : bool toowide = false;
1039 :
1040 : /* load the values/null flags from sample rows */
1041 3426159 : for (j = 0; j < numattrs; j++)
1042 : {
1043 : Datum value;
1044 : bool isnull;
1045 : int attlen;
1046 2430378 : AttrNumber attnum = attnums[j];
1047 :
1048 : int idx;
1049 :
1050 : /* match attnum to the pre-calculated data */
1051 4790901 : for (idx = 0; idx < data->nattnums; idx++)
1052 : {
1053 4790901 : if (attnum == data->attnums[idx])
1054 2430378 : break;
1055 : }
1056 :
1057 : Assert(idx < data->nattnums);
1058 :
1059 2430378 : value = data->values[idx][i];
1060 2430378 : isnull = data->nulls[idx][i];
1061 2430378 : attlen = typlen[idx];
1062 :
1063 : /*
1064 : * If this is a varlena value, check if it's too wide and if yes
1065 : * then skip the whole item. Otherwise detoast the value.
1066 : *
1067 : * XXX It may happen that we've already detoasted some preceding
1068 : * values for the current item. We don't bother to cleanup those
1069 : * on the assumption that those are small (below WIDTH_THRESHOLD)
1070 : * and will be discarded at the end of analyze.
1071 : */
1072 2430378 : if ((!isnull) && (attlen == -1))
1073 : {
1074 741282 : if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
1075 : {
1076 0 : toowide = true;
1077 0 : break;
1078 : }
1079 :
1080 741282 : value = PointerGetDatum(PG_DETOAST_DATUM(value));
1081 : }
1082 :
1083 2430378 : items[nrows].values[j] = value;
1084 2430378 : items[nrows].isnull[j] = isnull;
1085 : }
1086 :
1087 995781 : if (toowide)
1088 0 : continue;
1089 :
1090 995781 : nrows++;
1091 : }
1092 :
1093 : /* store the actual number of items (ignoring the too-wide ones) */
1094 657 : *nitems = nrows;
1095 :
1096 : /* all items were too wide */
1097 657 : if (nrows == 0)
1098 : {
1099 : /* everything is allocated as a single chunk */
1100 0 : pfree(items);
1101 0 : return NULL;
1102 : }
1103 :
1104 : /* do the sort, using the multi-sort */
1105 657 : qsort_interruptible(items, nrows, sizeof(SortItem),
1106 : multi_sort_compare, mss);
1107 :
1108 657 : return items;
1109 : }
1110 :
1111 : /*
1112 : * has_stats_of_kind
1113 : * Check whether the list contains statistic of a given kind
1114 : */
1115 : bool
1116 2514 : has_stats_of_kind(List *stats, char requiredkind)
1117 : {
1118 : ListCell *l;
1119 :
1120 4143 : foreach(l, stats)
1121 : {
1122 2895 : StatisticExtInfo *stat = (StatisticExtInfo *) lfirst(l);
1123 :
1124 2895 : if (stat->kind == requiredkind)
1125 1266 : return true;
1126 : }
1127 :
1128 1248 : return false;
1129 : }
1130 :
1131 : /*
1132 : * stat_find_expression
1133 : * Search for an expression in statistics object's list of expressions.
1134 : *
1135 : * Returns the index of the expression in the statistics object's list of
1136 : * expressions, or -1 if not found.
1137 : */
1138 : static int
1139 297 : stat_find_expression(StatisticExtInfo *stat, Node *expr)
1140 : {
1141 : ListCell *lc;
1142 : int idx;
1143 :
1144 297 : idx = 0;
1145 537 : foreach(lc, stat->exprs)
1146 : {
1147 486 : Node *stat_expr = (Node *) lfirst(lc);
1148 :
1149 486 : if (equal(stat_expr, expr))
1150 246 : return idx;
1151 240 : idx++;
1152 : }
1153 :
1154 : /* Expression not found */
1155 51 : return -1;
1156 : }
1157 :
1158 : /*
1159 : * stat_covers_expressions
1160 : * Test whether a statistics object covers all expressions in a list.
1161 : *
1162 : * Returns true if all expressions are covered. If expr_idxs is non-NULL, it
1163 : * is populated with the indexes of the expressions found.
1164 : */
1165 : static bool
1166 1557 : stat_covers_expressions(StatisticExtInfo *stat, List *exprs,
1167 : Bitmapset **expr_idxs)
1168 : {
1169 : ListCell *lc;
1170 :
1171 1803 : foreach(lc, exprs)
1172 : {
1173 297 : Node *expr = (Node *) lfirst(lc);
1174 : int expr_idx;
1175 :
1176 297 : expr_idx = stat_find_expression(stat, expr);
1177 297 : if (expr_idx == -1)
1178 51 : return false;
1179 :
1180 246 : if (expr_idxs != NULL)
1181 123 : *expr_idxs = bms_add_member(*expr_idxs, expr_idx);
1182 : }
1183 :
1184 : /* If we reach here, all expressions are covered */
1185 1506 : return true;
1186 : }
1187 :
1188 : /*
1189 : * choose_best_statistics
1190 : * Look for and return statistics with the specified 'requiredkind' which
1191 : * have keys that match at least two of the given attnums. Return NULL if
1192 : * there's no match.
1193 : *
1194 : * The current selection criteria is very simple - we choose the statistics
1195 : * object referencing the most attributes in covered (and still unestimated
1196 : * clauses), breaking ties in favor of objects with fewer keys overall.
1197 : *
1198 : * The clause_attnums is an array of bitmaps, storing attnums for individual
1199 : * clauses. A NULL element means the clause is either incompatible or already
1200 : * estimated.
1201 : *
1202 : * XXX If multiple statistics objects tie on both criteria, then which object
1203 : * is chosen depends on the order that they appear in the stats list. Perhaps
1204 : * further tiebreakers are needed.
1205 : */
1206 : StatisticExtInfo *
1207 675 : choose_best_statistics(List *stats, char requiredkind, bool inh,
1208 : Bitmapset **clause_attnums, List **clause_exprs,
1209 : int nclauses)
1210 : {
1211 : ListCell *lc;
1212 675 : StatisticExtInfo *best_match = NULL;
1213 675 : int best_num_matched = 2; /* goal #1: maximize */
1214 675 : int best_match_keys = (STATS_MAX_DIMENSIONS + 1); /* goal #2: minimize */
1215 :
1216 1773 : foreach(lc, stats)
1217 : {
1218 : int i;
1219 1098 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
1220 1098 : Bitmapset *matched_attnums = NULL;
1221 1098 : Bitmapset *matched_exprs = NULL;
1222 : int num_matched;
1223 : int numkeys;
1224 :
1225 : /* skip statistics that are not of the correct type */
1226 1098 : if (info->kind != requiredkind)
1227 234 : continue;
1228 :
1229 : /* skip statistics with mismatching inheritance flag */
1230 864 : if (info->inherit != inh)
1231 12 : continue;
1232 :
1233 : /*
1234 : * Collect attributes and expressions in remaining (unestimated)
1235 : * clauses fully covered by this statistic object.
1236 : *
1237 : * We know already estimated clauses have both clause_attnums and
1238 : * clause_exprs set to NULL. We leave the pointers NULL if already
1239 : * estimated, or we reset them to NULL after estimating the clause.
1240 : */
1241 2898 : for (i = 0; i < nclauses; i++)
1242 : {
1243 2046 : Bitmapset *expr_idxs = NULL;
1244 :
1245 : /* ignore incompatible/estimated clauses */
1246 2046 : if (!clause_attnums[i] && !clause_exprs[i])
1247 1242 : continue;
1248 :
1249 : /* ignore clauses that are not covered by this object */
1250 1017 : if (!bms_is_subset(clause_attnums[i], info->keys) ||
1251 852 : !stat_covers_expressions(info, clause_exprs[i], &expr_idxs))
1252 213 : continue;
1253 :
1254 : /* record attnums and indexes of expressions covered */
1255 804 : matched_attnums = bms_add_members(matched_attnums, clause_attnums[i]);
1256 804 : matched_exprs = bms_add_members(matched_exprs, expr_idxs);
1257 : }
1258 :
1259 852 : num_matched = bms_num_members(matched_attnums) + bms_num_members(matched_exprs);
1260 :
1261 852 : bms_free(matched_attnums);
1262 852 : bms_free(matched_exprs);
1263 :
1264 : /*
1265 : * save the actual number of keys in the stats so that we can choose
1266 : * the narrowest stats with the most matching keys.
1267 : */
1268 852 : numkeys = bms_num_members(info->keys) + list_length(info->exprs);
1269 :
1270 : /*
1271 : * Use this object when it increases the number of matched attributes
1272 : * and expressions or when it matches the same number of attributes
1273 : * and expressions but these stats have fewer keys than any previous
1274 : * match.
1275 : */
1276 852 : if (num_matched > best_num_matched ||
1277 207 : (num_matched == best_num_matched && numkeys < best_match_keys))
1278 : {
1279 303 : best_match = info;
1280 303 : best_num_matched = num_matched;
1281 303 : best_match_keys = numkeys;
1282 : }
1283 : }
1284 :
1285 675 : return best_match;
1286 : }
1287 :
1288 : /*
1289 : * statext_is_compatible_clause_internal
1290 : * Determines if the clause is compatible with MCV lists.
1291 : *
1292 : * To be compatible, the given clause must be a combination of supported
1293 : * clauses built from Vars or sub-expressions (where a sub-expression is
1294 : * something that exactly matches an expression found in statistics objects).
1295 : * This function recursively examines the clause and extracts any
1296 : * sub-expressions that will need to be matched against statistics.
1297 : *
1298 : * Currently, we only support the following types of clauses:
1299 : *
1300 : * (a) OpExprs of the form (Var/Expr op Const), or (Const op Var/Expr), where
1301 : * the op is one of ("=", "<", ">", ">=", "<=")
1302 : *
1303 : * (b) (Var/Expr IS [NOT] NULL)
1304 : *
1305 : * (c) combinations using AND/OR/NOT
1306 : *
1307 : * (d) ScalarArrayOpExprs of the form (Var/Expr op ANY (Const)) or
1308 : * (Var/Expr op ALL (Const))
1309 : *
1310 : * In the future, the range of supported clauses may be expanded to more
1311 : * complex cases, for example (Var op Var).
1312 : *
1313 : * Arguments:
1314 : * clause: (sub)clause to be inspected (bare clause, not a RestrictInfo)
1315 : * relid: rel that all Vars in clause must belong to
1316 : * *attnums: input/output parameter collecting attribute numbers of all
1317 : * mentioned Vars. Note that we do not offset the attribute numbers,
1318 : * so we can't cope with system columns.
1319 : * *exprs: input/output parameter collecting primitive subclauses within
1320 : * the clause tree
1321 : * *leakproof: input/output parameter recording the leakproofness of the
1322 : * clause tree. This should be true initially, and will be set to false
1323 : * if any operator function used in an OpExpr is not leakproof.
1324 : *
1325 : * Returns false if there is something we definitively can't handle.
1326 : * On true return, we can proceed to match the *exprs against statistics.
1327 : */
1328 : static bool
1329 1731 : statext_is_compatible_clause_internal(PlannerInfo *root, Node *clause,
1330 : Index relid, Bitmapset **attnums,
1331 : List **exprs, bool *leakproof)
1332 : {
1333 : /* Look inside any binary-compatible relabeling (as in examine_variable) */
1334 1731 : if (IsA(clause, RelabelType))
1335 0 : clause = (Node *) ((RelabelType *) clause)->arg;
1336 :
1337 : /* plain Var references (boolean Vars or recursive checks) */
1338 1731 : if (IsA(clause, Var))
1339 : {
1340 774 : Var *var = (Var *) clause;
1341 :
1342 : /* Ensure var is from the correct relation */
1343 774 : if (var->varno != relid)
1344 0 : return false;
1345 :
1346 : /* we also better ensure the Var is from the current level */
1347 774 : if (var->varlevelsup > 0)
1348 0 : return false;
1349 :
1350 : /*
1351 : * Also reject system attributes and whole-row Vars (we don't allow
1352 : * stats on those).
1353 : */
1354 774 : if (!AttrNumberIsForUserDefinedAttr(var->varattno))
1355 0 : return false;
1356 :
1357 : /* OK, record the attnum for later permissions checks. */
1358 774 : *attnums = bms_add_member(*attnums, var->varattno);
1359 :
1360 774 : return true;
1361 : }
1362 :
1363 : /* (Var/Expr op Const) or (Const op Var/Expr) */
1364 957 : if (is_opclause(clause))
1365 : {
1366 696 : OpExpr *expr = (OpExpr *) clause;
1367 : Node *clause_expr;
1368 :
1369 : /* Only expressions with two arguments are considered compatible. */
1370 696 : if (list_length(expr->args) != 2)
1371 0 : return false;
1372 :
1373 : /* Check if the expression has the right shape */
1374 696 : if (!examine_opclause_args(expr->args, &clause_expr, NULL, NULL))
1375 0 : return false;
1376 :
1377 : /*
1378 : * If it's not one of the supported operators ("=", "<", ">", etc.),
1379 : * just ignore the clause, as it's not compatible with MCV lists.
1380 : *
1381 : * This uses the function for estimating selectivity, not the operator
1382 : * directly (a bit awkward, but well ...).
1383 : */
1384 696 : switch (get_oprrest(expr->opno))
1385 : {
1386 696 : case F_EQSEL:
1387 : case F_NEQSEL:
1388 : case F_SCALARLTSEL:
1389 : case F_SCALARLESEL:
1390 : case F_SCALARGTSEL:
1391 : case F_SCALARGESEL:
1392 : /* supported, will continue with inspection of the Var/Expr */
1393 696 : break;
1394 :
1395 0 : default:
1396 : /* other estimators are considered unknown/unsupported */
1397 0 : return false;
1398 : }
1399 :
1400 : /* Check if the operator is leakproof */
1401 696 : if (*leakproof)
1402 690 : *leakproof = get_func_leakproof(get_opcode(expr->opno));
1403 :
1404 : /* Check (Var op Const) or (Const op Var) clauses by recursing. */
1405 696 : if (IsA(clause_expr, Var))
1406 576 : return statext_is_compatible_clause_internal(root, clause_expr,
1407 : relid, attnums,
1408 : exprs, leakproof);
1409 :
1410 : /* Otherwise we have (Expr op Const) or (Const op Expr). */
1411 120 : *exprs = lappend(*exprs, clause_expr);
1412 120 : return true;
1413 : }
1414 :
1415 : /* Var/Expr IN Array */
1416 261 : if (IsA(clause, ScalarArrayOpExpr))
1417 : {
1418 144 : ScalarArrayOpExpr *expr = (ScalarArrayOpExpr *) clause;
1419 : Node *clause_expr;
1420 : bool expronleft;
1421 :
1422 : /* Only expressions with two arguments are considered compatible. */
1423 144 : if (list_length(expr->args) != 2)
1424 0 : return false;
1425 :
1426 : /* Check if the expression has the right shape (one Var, one Const) */
1427 144 : if (!examine_opclause_args(expr->args, &clause_expr, NULL, &expronleft))
1428 0 : return false;
1429 :
1430 : /* We only support Var on left, Const on right */
1431 144 : if (!expronleft)
1432 3 : return false;
1433 :
1434 : /*
1435 : * If it's not one of the supported operators ("=", "<", ">", etc.),
1436 : * just ignore the clause, as it's not compatible with MCV lists.
1437 : *
1438 : * This uses the function for estimating selectivity, not the operator
1439 : * directly (a bit awkward, but well ...).
1440 : */
1441 141 : switch (get_oprrest(expr->opno))
1442 : {
1443 141 : case F_EQSEL:
1444 : case F_NEQSEL:
1445 : case F_SCALARLTSEL:
1446 : case F_SCALARLESEL:
1447 : case F_SCALARGTSEL:
1448 : case F_SCALARGESEL:
1449 : /* supported, will continue with inspection of the Var/Expr */
1450 141 : break;
1451 :
1452 0 : default:
1453 : /* other estimators are considered unknown/unsupported */
1454 0 : return false;
1455 : }
1456 :
1457 : /* Check if the operator is leakproof */
1458 141 : if (*leakproof)
1459 141 : *leakproof = get_func_leakproof(get_opcode(expr->opno));
1460 :
1461 : /* Check Var IN Array clauses by recursing. */
1462 141 : if (IsA(clause_expr, Var))
1463 114 : return statext_is_compatible_clause_internal(root, clause_expr,
1464 : relid, attnums,
1465 : exprs, leakproof);
1466 :
1467 : /* Otherwise we have Expr IN Array. */
1468 27 : *exprs = lappend(*exprs, clause_expr);
1469 27 : return true;
1470 : }
1471 :
1472 : /* AND/OR/NOT clause */
1473 234 : if (is_andclause(clause) ||
1474 207 : is_orclause(clause) ||
1475 90 : is_notclause(clause))
1476 : {
1477 : /*
1478 : * AND/OR/NOT-clauses are supported if all sub-clauses are supported
1479 : *
1480 : * Perhaps we could improve this by handling mixed cases, when some of
1481 : * the clauses are supported and some are not. Selectivity for the
1482 : * supported subclauses would be computed using extended statistics,
1483 : * and the remaining clauses would be estimated using the traditional
1484 : * algorithm (product of selectivities).
1485 : *
1486 : * It however seems overly complex, and in a way we already do that
1487 : * because if we reject the whole clause as unsupported here, it will
1488 : * be eventually passed to clauselist_selectivity() which does exactly
1489 : * this (split into supported/unsupported clauses etc).
1490 : */
1491 42 : BoolExpr *expr = (BoolExpr *) clause;
1492 : ListCell *lc;
1493 :
1494 111 : foreach(lc, expr->args)
1495 : {
1496 : /*
1497 : * If we find an incompatible clause in the arguments, treat the
1498 : * whole clause as incompatible.
1499 : */
1500 69 : if (!statext_is_compatible_clause_internal(root,
1501 69 : (Node *) lfirst(lc),
1502 : relid, attnums, exprs,
1503 : leakproof))
1504 0 : return false;
1505 : }
1506 :
1507 42 : return true;
1508 : }
1509 :
1510 : /* Var/Expr IS NULL */
1511 75 : if (IsA(clause, NullTest))
1512 : {
1513 72 : NullTest *nt = (NullTest *) clause;
1514 :
1515 : /* Check Var IS NULL clauses by recursing. */
1516 72 : if (IsA(nt->arg, Var))
1517 45 : return statext_is_compatible_clause_internal(root,
1518 45 : (Node *) (nt->arg),
1519 : relid, attnums,
1520 : exprs, leakproof);
1521 :
1522 : /* Otherwise we have Expr IS NULL. */
1523 27 : *exprs = lappend(*exprs, nt->arg);
1524 27 : return true;
1525 : }
1526 :
1527 : /*
1528 : * Treat any other expressions as bare expressions to be matched against
1529 : * expressions in statistics objects.
1530 : */
1531 3 : *exprs = lappend(*exprs, clause);
1532 3 : return true;
1533 : }
1534 :
1535 : /*
1536 : * statext_is_compatible_clause
1537 : * Determines if the clause is compatible with MCV lists.
1538 : *
1539 : * See statext_is_compatible_clause_internal, above, for the basic rules.
1540 : * This layer deals with RestrictInfo superstructure and applies permissions
1541 : * checks to verify that it's okay to examine all mentioned Vars.
1542 : *
1543 : * Arguments:
1544 : * clause: clause to be inspected (in RestrictInfo form)
1545 : * relid: rel that all Vars in clause must belong to
1546 : * *attnums: input/output parameter collecting attribute numbers of all
1547 : * mentioned Vars. Note that we do not offset the attribute numbers,
1548 : * so we can't cope with system columns.
1549 : * *exprs: input/output parameter collecting primitive subclauses within
1550 : * the clause tree
1551 : *
1552 : * Returns false if there is something we definitively can't handle.
1553 : * On true return, we can proceed to match the *exprs against statistics.
1554 : */
1555 : static bool
1556 954 : statext_is_compatible_clause(PlannerInfo *root, Node *clause, Index relid,
1557 : Bitmapset **attnums, List **exprs)
1558 : {
1559 : RestrictInfo *rinfo;
1560 : int clause_relid;
1561 : bool leakproof;
1562 :
1563 : /*
1564 : * Special-case handling for bare BoolExpr AND clauses, because the
1565 : * restrictinfo machinery doesn't build RestrictInfos on top of AND
1566 : * clauses.
1567 : */
1568 954 : if (is_andclause(clause))
1569 : {
1570 24 : BoolExpr *expr = (BoolExpr *) clause;
1571 : ListCell *lc;
1572 :
1573 : /*
1574 : * Check that each sub-clause is compatible. We expect these to be
1575 : * RestrictInfos.
1576 : */
1577 81 : foreach(lc, expr->args)
1578 : {
1579 57 : if (!statext_is_compatible_clause(root, (Node *) lfirst(lc),
1580 : relid, attnums, exprs))
1581 0 : return false;
1582 : }
1583 :
1584 24 : return true;
1585 : }
1586 :
1587 : /* Otherwise it must be a RestrictInfo. */
1588 930 : if (!IsA(clause, RestrictInfo))
1589 0 : return false;
1590 930 : rinfo = (RestrictInfo *) clause;
1591 :
1592 : /* Pseudoconstants are not really interesting here. */
1593 930 : if (rinfo->pseudoconstant)
1594 3 : return false;
1595 :
1596 : /* Clauses referencing other varnos are incompatible. */
1597 927 : if (!bms_get_singleton_member(rinfo->clause_relids, &clause_relid) ||
1598 927 : clause_relid != relid)
1599 0 : return false;
1600 :
1601 : /*
1602 : * Check the clause, determine what attributes it references, and whether
1603 : * it includes any non-leakproof operators.
1604 : */
1605 927 : leakproof = true;
1606 927 : if (!statext_is_compatible_clause_internal(root, (Node *) rinfo->clause,
1607 : relid, attnums, exprs,
1608 : &leakproof))
1609 3 : return false;
1610 :
1611 : /*
1612 : * If the clause includes any non-leakproof operators, check that the user
1613 : * has permission to read all required attributes, otherwise the operators
1614 : * might reveal values from the MCV list that the user doesn't have
1615 : * permission to see. We require all rows to be selectable --- there must
1616 : * be no securityQuals from security barrier views or RLS policies. See
1617 : * similar code in examine_variable(), examine_simple_variable(), and
1618 : * statistic_proc_security_check().
1619 : *
1620 : * Note that for an inheritance child, the permission checks are performed
1621 : * on the inheritance root parent, and whole-table select privilege on the
1622 : * parent doesn't guarantee that the user could read all columns of the
1623 : * child. Therefore we must check all referenced columns.
1624 : */
1625 924 : if (!leakproof)
1626 : {
1627 126 : Bitmapset *clause_attnums = NULL;
1628 126 : int attnum = -1;
1629 :
1630 : /*
1631 : * We have to check per-column privileges. *attnums has the attnums
1632 : * for individual Vars we saw, but there may also be Vars within
1633 : * subexpressions in *exprs. We can use pull_varattnos() to extract
1634 : * those, but there's an impedance mismatch: attnums returned by
1635 : * pull_varattnos() are offset by FirstLowInvalidHeapAttributeNumber,
1636 : * while attnums within *attnums aren't. Convert *attnums to the
1637 : * offset style so we can combine the results.
1638 : */
1639 246 : while ((attnum = bms_next_member(*attnums, attnum)) >= 0)
1640 : {
1641 120 : clause_attnums =
1642 120 : bms_add_member(clause_attnums,
1643 : attnum - FirstLowInvalidHeapAttributeNumber);
1644 : }
1645 :
1646 : /* Now merge attnums from *exprs into clause_attnums */
1647 126 : if (*exprs != NIL)
1648 24 : pull_varattnos((Node *) *exprs, relid, &clause_attnums);
1649 :
1650 : /* Must have permission to read all rows from these columns */
1651 126 : if (!all_rows_selectable(root, relid, clause_attnums))
1652 114 : return false;
1653 : }
1654 :
1655 : /* If we reach here, the clause is OK */
1656 810 : return true;
1657 : }
1658 :
1659 : /*
1660 : * statext_mcv_clauselist_selectivity
1661 : * Estimate clauses using the best multi-column statistics.
1662 : *
1663 : * Applies available extended (multi-column) statistics on a table. There may
1664 : * be multiple applicable statistics (with respect to the clauses), in which
1665 : * case we use greedy approach. In each round we select the best statistic on
1666 : * a table (measured by the number of attributes extracted from the clauses
1667 : * and covered by it), and compute the selectivity for the supplied clauses.
1668 : * We repeat this process with the remaining clauses (if any), until none of
1669 : * the available statistics can be used.
1670 : *
1671 : * One of the main challenges with using MCV lists is how to extrapolate the
1672 : * estimate to the data not covered by the MCV list. To do that, we compute
1673 : * not only the "MCV selectivity" (selectivities for MCV items matching the
1674 : * supplied clauses), but also the following related selectivities:
1675 : *
1676 : * - simple selectivity: Computed without extended statistics, i.e. as if the
1677 : * columns/clauses were independent.
1678 : *
1679 : * - base selectivity: Similar to simple selectivity, but is computed using
1680 : * the extended statistic by adding up the base frequencies (that we compute
1681 : * and store for each MCV item) of matching MCV items.
1682 : *
1683 : * - total selectivity: Selectivity covered by the whole MCV list.
1684 : *
1685 : * These are passed to mcv_combine_selectivities() which combines them to
1686 : * produce a selectivity estimate that makes use of both per-column statistics
1687 : * and the multi-column MCV statistics.
1688 : *
1689 : * 'estimatedclauses' is an input/output parameter. We set bits for the
1690 : * 0-based 'clauses' indexes we estimate for and also skip clause items that
1691 : * already have a bit set.
1692 : */
1693 : static Selectivity
1694 1296 : statext_mcv_clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid,
1695 : JoinType jointype, SpecialJoinInfo *sjinfo,
1696 : RelOptInfo *rel, Bitmapset **estimatedclauses,
1697 : bool is_or)
1698 : {
1699 : ListCell *l;
1700 : Bitmapset **list_attnums; /* attnums extracted from the clause */
1701 : List **list_exprs; /* expressions matched to any statistic */
1702 : int listidx;
1703 1296 : Selectivity sel = (is_or) ? 0.0 : 1.0;
1704 1296 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
1705 :
1706 : /* check if there's any stats that might be useful for us. */
1707 1296 : if (!has_stats_of_kind(rel->statlist, STATS_EXT_MCV))
1708 924 : return sel;
1709 :
1710 372 : list_attnums = palloc_array(Bitmapset *, list_length(clauses));
1711 :
1712 : /* expressions extracted from complex expressions */
1713 372 : list_exprs = palloc_array(List *, list_length(clauses));
1714 :
1715 : /*
1716 : * Pre-process the clauses list to extract the attnums and expressions
1717 : * seen in each item. We need to determine if there are any clauses which
1718 : * will be useful for selectivity estimations with extended stats. Along
1719 : * the way we'll record all of the attnums and expressions for each clause
1720 : * in lists which we'll reference later so we don't need to repeat the
1721 : * same work again.
1722 : *
1723 : * We also skip clauses that we already estimated using different types of
1724 : * statistics (we treat them as incompatible).
1725 : */
1726 372 : listidx = 0;
1727 1269 : foreach(l, clauses)
1728 : {
1729 897 : Node *clause = (Node *) lfirst(l);
1730 897 : Bitmapset *attnums = NULL;
1731 897 : List *exprs = NIL;
1732 :
1733 1794 : if (!bms_is_member(listidx, *estimatedclauses) &&
1734 897 : statext_is_compatible_clause(root, clause, rel->relid, &attnums, &exprs))
1735 : {
1736 777 : list_attnums[listidx] = attnums;
1737 777 : list_exprs[listidx] = exprs;
1738 : }
1739 : else
1740 : {
1741 120 : list_attnums[listidx] = NULL;
1742 120 : list_exprs[listidx] = NIL;
1743 : }
1744 :
1745 897 : listidx++;
1746 : }
1747 :
1748 : /* apply as many extended statistics as possible */
1749 : while (true)
1750 303 : {
1751 : StatisticExtInfo *stat;
1752 : List *stat_clauses;
1753 : Bitmapset *simple_clauses;
1754 :
1755 : /* find the best suited statistics object for these attnums */
1756 675 : stat = choose_best_statistics(rel->statlist, STATS_EXT_MCV, rte->inh,
1757 : list_attnums, list_exprs,
1758 : list_length(clauses));
1759 :
1760 : /*
1761 : * if no (additional) matching stats could be found then we've nothing
1762 : * to do
1763 : */
1764 675 : if (!stat)
1765 372 : break;
1766 :
1767 : /* Ensure choose_best_statistics produced an expected stats type. */
1768 : Assert(stat->kind == STATS_EXT_MCV);
1769 :
1770 : /* now filter the clauses to be estimated using the selected MCV */
1771 303 : stat_clauses = NIL;
1772 :
1773 : /* record which clauses are simple (single column or expression) */
1774 303 : simple_clauses = NULL;
1775 :
1776 303 : listidx = -1;
1777 1056 : foreach(l, clauses)
1778 : {
1779 : /* Increment the index before we decide if to skip the clause. */
1780 753 : listidx++;
1781 :
1782 : /*
1783 : * Ignore clauses from which we did not extract any attnums or
1784 : * expressions (this needs to be consistent with what we do in
1785 : * choose_best_statistics).
1786 : *
1787 : * This also eliminates already estimated clauses - both those
1788 : * estimated before and during applying extended statistics.
1789 : *
1790 : * XXX This check is needed because both bms_is_subset and
1791 : * stat_covers_expressions return true for empty attnums and
1792 : * expressions.
1793 : */
1794 753 : if (!list_attnums[listidx] && !list_exprs[listidx])
1795 18 : continue;
1796 :
1797 : /*
1798 : * The clause was not estimated yet, and we've extracted either
1799 : * attnums or expressions from it. Ignore it if it's not fully
1800 : * covered by the chosen statistics object.
1801 : *
1802 : * We need to check both attributes and expressions, and reject if
1803 : * either is not covered.
1804 : */
1805 735 : if (!bms_is_subset(list_attnums[listidx], stat->keys) ||
1806 705 : !stat_covers_expressions(stat, list_exprs[listidx], NULL))
1807 33 : continue;
1808 :
1809 : /*
1810 : * Now we know the clause is compatible (we have either attnums or
1811 : * expressions extracted from it), and was not estimated yet.
1812 : */
1813 :
1814 : /* record simple clauses (single column or expression) */
1815 825 : if ((list_attnums[listidx] == NULL &&
1816 123 : list_length(list_exprs[listidx]) == 1) ||
1817 1158 : (list_exprs[listidx] == NIL &&
1818 579 : bms_membership(list_attnums[listidx]) == BMS_SINGLETON))
1819 672 : simple_clauses = bms_add_member(simple_clauses,
1820 : list_length(stat_clauses));
1821 :
1822 : /* add clause to list and mark it as estimated */
1823 702 : stat_clauses = lappend(stat_clauses, (Node *) lfirst(l));
1824 702 : *estimatedclauses = bms_add_member(*estimatedclauses, listidx);
1825 :
1826 : /*
1827 : * Reset the pointers, so that choose_best_statistics knows this
1828 : * clause was estimated and does not consider it again.
1829 : */
1830 702 : bms_free(list_attnums[listidx]);
1831 702 : list_attnums[listidx] = NULL;
1832 :
1833 702 : list_free(list_exprs[listidx]);
1834 702 : list_exprs[listidx] = NULL;
1835 : }
1836 :
1837 303 : if (is_or)
1838 : {
1839 48 : bool *or_matches = NULL;
1840 48 : Selectivity simple_or_sel = 0.0,
1841 48 : stat_sel = 0.0;
1842 : MCVList *mcv_list;
1843 :
1844 : /* Load the MCV list stored in the statistics object */
1845 48 : mcv_list = statext_mcv_load(stat->statOid, rte->inh);
1846 :
1847 : /*
1848 : * Compute the selectivity of the ORed list of clauses covered by
1849 : * this statistics object by estimating each in turn and combining
1850 : * them using the formula P(A OR B) = P(A) + P(B) - P(A AND B).
1851 : * This allows us to use the multivariate MCV stats to better
1852 : * estimate the individual terms and their overlap.
1853 : *
1854 : * Each time we iterate this formula, the clause "A" above is
1855 : * equal to all the clauses processed so far, combined with "OR".
1856 : */
1857 48 : listidx = 0;
1858 168 : foreach(l, stat_clauses)
1859 : {
1860 120 : Node *clause = (Node *) lfirst(l);
1861 : Selectivity simple_sel,
1862 : overlap_simple_sel,
1863 : mcv_sel,
1864 : mcv_basesel,
1865 : overlap_mcvsel,
1866 : overlap_basesel,
1867 : mcv_totalsel,
1868 : clause_sel,
1869 : overlap_sel;
1870 :
1871 : /*
1872 : * "Simple" selectivity of the next clause and its overlap
1873 : * with any of the previous clauses. These are our initial
1874 : * estimates of P(B) and P(A AND B), assuming independence of
1875 : * columns/clauses.
1876 : */
1877 120 : simple_sel = clause_selectivity_ext(root, clause, varRelid,
1878 : jointype, sjinfo, false);
1879 :
1880 120 : overlap_simple_sel = simple_or_sel * simple_sel;
1881 :
1882 : /*
1883 : * New "simple" selectivity of all clauses seen so far,
1884 : * assuming independence.
1885 : */
1886 120 : simple_or_sel += simple_sel - overlap_simple_sel;
1887 120 : CLAMP_PROBABILITY(simple_or_sel);
1888 :
1889 : /*
1890 : * Multi-column estimate of this clause using MCV statistics,
1891 : * along with base and total selectivities, and corresponding
1892 : * selectivities for the overlap term P(A AND B).
1893 : */
1894 120 : mcv_sel = mcv_clause_selectivity_or(root, stat, mcv_list,
1895 : clause, &or_matches,
1896 : &mcv_basesel,
1897 : &overlap_mcvsel,
1898 : &overlap_basesel,
1899 : &mcv_totalsel);
1900 :
1901 : /*
1902 : * Combine the simple and multi-column estimates.
1903 : *
1904 : * If this clause is a simple single-column clause, then we
1905 : * just use the simple selectivity estimate for it, since the
1906 : * multi-column statistics are unlikely to improve on that
1907 : * (and in fact could make it worse). For the overlap, we
1908 : * always make use of the multi-column statistics.
1909 : */
1910 120 : if (bms_is_member(listidx, simple_clauses))
1911 96 : clause_sel = simple_sel;
1912 : else
1913 24 : clause_sel = mcv_combine_selectivities(simple_sel,
1914 : mcv_sel,
1915 : mcv_basesel,
1916 : mcv_totalsel);
1917 :
1918 120 : overlap_sel = mcv_combine_selectivities(overlap_simple_sel,
1919 : overlap_mcvsel,
1920 : overlap_basesel,
1921 : mcv_totalsel);
1922 :
1923 : /* Factor these into the result for this statistics object */
1924 120 : stat_sel += clause_sel - overlap_sel;
1925 120 : CLAMP_PROBABILITY(stat_sel);
1926 :
1927 120 : listidx++;
1928 : }
1929 :
1930 : /*
1931 : * Factor the result for this statistics object into the overall
1932 : * result. We treat the results from each separate statistics
1933 : * object as independent of one another.
1934 : */
1935 48 : sel = sel + stat_sel - sel * stat_sel;
1936 : }
1937 : else /* Implicitly-ANDed list of clauses */
1938 : {
1939 : Selectivity simple_sel,
1940 : mcv_sel,
1941 : mcv_basesel,
1942 : mcv_totalsel,
1943 : stat_sel;
1944 :
1945 : /*
1946 : * "Simple" selectivity, i.e. without any extended statistics,
1947 : * essentially assuming independence of the columns/clauses.
1948 : */
1949 255 : simple_sel = clauselist_selectivity_ext(root, stat_clauses,
1950 : varRelid, jointype,
1951 : sjinfo, false);
1952 :
1953 : /*
1954 : * Multi-column estimate using MCV statistics, along with base and
1955 : * total selectivities.
1956 : */
1957 255 : mcv_sel = mcv_clauselist_selectivity(root, stat, stat_clauses,
1958 : varRelid, jointype, sjinfo,
1959 : rel, &mcv_basesel,
1960 : &mcv_totalsel);
1961 :
1962 : /* Combine the simple and multi-column estimates. */
1963 255 : stat_sel = mcv_combine_selectivities(simple_sel,
1964 : mcv_sel,
1965 : mcv_basesel,
1966 : mcv_totalsel);
1967 :
1968 : /* Factor this into the overall result */
1969 255 : sel *= stat_sel;
1970 : }
1971 : }
1972 :
1973 372 : return sel;
1974 : }
1975 :
1976 : /*
1977 : * statext_clauselist_selectivity
1978 : * Estimate clauses using the best multi-column statistics.
1979 : */
1980 : Selectivity
1981 1296 : statext_clauselist_selectivity(PlannerInfo *root, List *clauses, int varRelid,
1982 : JoinType jointype, SpecialJoinInfo *sjinfo,
1983 : RelOptInfo *rel, Bitmapset **estimatedclauses,
1984 : bool is_or)
1985 : {
1986 : Selectivity sel;
1987 :
1988 : /* First, try estimating clauses using a multivariate MCV list. */
1989 1296 : sel = statext_mcv_clauselist_selectivity(root, clauses, varRelid, jointype,
1990 : sjinfo, rel, estimatedclauses, is_or);
1991 :
1992 : /*
1993 : * Functional dependencies only work for clauses connected by AND, so for
1994 : * OR clauses we're done.
1995 : */
1996 1296 : if (is_or)
1997 78 : return sel;
1998 :
1999 : /*
2000 : * Then, apply functional dependencies on the remaining clauses by calling
2001 : * dependencies_clauselist_selectivity. Pass 'estimatedclauses' so the
2002 : * function can properly skip clauses already estimated above.
2003 : *
2004 : * The reasoning for applying dependencies last is that the more complex
2005 : * stats can track more complex correlations between the attributes, and
2006 : * so may be considered more reliable.
2007 : *
2008 : * For example, MCV list can give us an exact selectivity for values in
2009 : * two columns, while functional dependencies can only provide information
2010 : * about the overall strength of the dependency.
2011 : */
2012 1218 : sel *= dependencies_clauselist_selectivity(root, clauses, varRelid,
2013 : jointype, sjinfo, rel,
2014 : estimatedclauses);
2015 :
2016 1218 : return sel;
2017 : }
2018 :
2019 : /*
2020 : * examine_opclause_args
2021 : * Split an operator expression's arguments into Expr and Const parts.
2022 : *
2023 : * Attempts to match the arguments to either (Expr op Const) or (Const op
2024 : * Expr), possibly with a RelabelType on top. When the expression matches this
2025 : * form, returns true, otherwise returns false.
2026 : *
2027 : * Optionally returns pointers to the extracted Expr/Const nodes, when passed
2028 : * non-null pointers (exprp, cstp and expronleftp). The expronleftp flag
2029 : * specifies on which side of the operator we found the expression node.
2030 : */
2031 : bool
2032 1509 : examine_opclause_args(List *args, Node **exprp, Const **cstp,
2033 : bool *expronleftp)
2034 : {
2035 : Node *expr;
2036 : Const *cst;
2037 : bool expronleft;
2038 : Node *leftop,
2039 : *rightop;
2040 :
2041 : /* enforced by statext_is_compatible_clause_internal */
2042 : Assert(list_length(args) == 2);
2043 :
2044 1509 : leftop = linitial(args);
2045 1509 : rightop = lsecond(args);
2046 :
2047 : /* strip RelabelType from either side of the expression */
2048 1509 : if (IsA(leftop, RelabelType))
2049 162 : leftop = (Node *) ((RelabelType *) leftop)->arg;
2050 :
2051 1509 : if (IsA(rightop, RelabelType))
2052 30 : rightop = (Node *) ((RelabelType *) rightop)->arg;
2053 :
2054 1509 : if (IsA(rightop, Const))
2055 : {
2056 1428 : expr = leftop;
2057 1428 : cst = (Const *) rightop;
2058 1428 : expronleft = true;
2059 : }
2060 81 : else if (IsA(leftop, Const))
2061 : {
2062 81 : expr = rightop;
2063 81 : cst = (Const *) leftop;
2064 81 : expronleft = false;
2065 : }
2066 : else
2067 0 : return false;
2068 :
2069 : /* return pointers to the extracted parts if requested */
2070 1509 : if (exprp)
2071 1509 : *exprp = expr;
2072 :
2073 1509 : if (cstp)
2074 669 : *cstp = cst;
2075 :
2076 1509 : if (expronleftp)
2077 813 : *expronleftp = expronleft;
2078 :
2079 1509 : return true;
2080 : }
2081 :
2082 :
2083 : /*
2084 : * Compute statistics about expressions of a relation.
2085 : */
2086 : static void
2087 113 : compute_expr_stats(Relation onerel, AnlExprData *exprdata, int nexprs,
2088 : HeapTuple *rows, int numrows)
2089 : {
2090 : MemoryContext expr_context,
2091 : old_context;
2092 : int ind,
2093 : i;
2094 :
2095 113 : expr_context = AllocSetContextCreate(CurrentMemoryContext,
2096 : "Analyze Expression",
2097 : ALLOCSET_DEFAULT_SIZES);
2098 113 : old_context = MemoryContextSwitchTo(expr_context);
2099 :
2100 322 : for (ind = 0; ind < nexprs; ind++)
2101 : {
2102 209 : AnlExprData *thisdata = &exprdata[ind];
2103 209 : VacAttrStats *stats = thisdata->vacattrstat;
2104 209 : Node *expr = thisdata->expr;
2105 : TupleTableSlot *slot;
2106 : EState *estate;
2107 : ExprContext *econtext;
2108 : Datum *exprvals;
2109 : bool *exprnulls;
2110 : ExprState *exprstate;
2111 : int tcnt;
2112 :
2113 : /* Are we still in the main context? */
2114 : Assert(CurrentMemoryContext == expr_context);
2115 :
2116 : /*
2117 : * Need an EState for evaluation of expressions. Create it in the
2118 : * per-expression context to be sure it gets cleaned up at the bottom
2119 : * of the loop.
2120 : */
2121 209 : estate = CreateExecutorState();
2122 209 : econtext = GetPerTupleExprContext(estate);
2123 :
2124 : /* Set up expression evaluation state */
2125 209 : exprstate = ExecPrepareExpr((Expr *) expr, estate);
2126 :
2127 : /* Need a slot to hold the current heap tuple, too */
2128 209 : slot = MakeSingleTupleTableSlot(RelationGetDescr(onerel),
2129 : &TTSOpsHeapTuple);
2130 :
2131 : /* Arrange for econtext's scan tuple to be the tuple under test */
2132 209 : econtext->ecxt_scantuple = slot;
2133 :
2134 : /* Compute and save expression values */
2135 209 : exprvals = (Datum *) palloc(numrows * sizeof(Datum));
2136 209 : exprnulls = (bool *) palloc(numrows * sizeof(bool));
2137 :
2138 209 : tcnt = 0;
2139 230455 : for (i = 0; i < numrows; i++)
2140 : {
2141 : Datum datum;
2142 : bool isnull;
2143 :
2144 : /*
2145 : * Reset the per-tuple context each time, to reclaim any cruft
2146 : * left behind by evaluating the statistics expressions.
2147 : */
2148 230246 : ResetExprContext(econtext);
2149 :
2150 : /* Set up for expression evaluation */
2151 230246 : ExecStoreHeapTuple(rows[i], slot, false);
2152 :
2153 : /*
2154 : * Evaluate the expression. We do this in the per-tuple context so
2155 : * as not to leak memory, and then copy the result into the
2156 : * context created at the beginning of this function.
2157 : */
2158 230246 : datum = ExecEvalExprSwitchContext(exprstate,
2159 230246 : GetPerTupleExprContext(estate),
2160 : &isnull);
2161 230246 : if (isnull)
2162 : {
2163 6 : exprvals[tcnt] = (Datum) 0;
2164 6 : exprnulls[tcnt] = true;
2165 : }
2166 : else
2167 : {
2168 : /* Make sure we copy the data into the context. */
2169 : Assert(CurrentMemoryContext == expr_context);
2170 :
2171 460480 : exprvals[tcnt] = datumCopy(datum,
2172 230240 : stats->attrtype->typbyval,
2173 230240 : stats->attrtype->typlen);
2174 230240 : exprnulls[tcnt] = false;
2175 : }
2176 :
2177 230246 : tcnt++;
2178 : }
2179 :
2180 : /*
2181 : * Now we can compute the statistics for the expression columns.
2182 : *
2183 : * XXX Unlike compute_index_stats we don't need to switch and reset
2184 : * memory contexts here, because we're only computing stats for a
2185 : * single expression (and not iterating over many indexes), so we just
2186 : * do it in expr_context. Note that compute_stats copies the result
2187 : * into stats->anl_context, so it does not disappear.
2188 : */
2189 209 : if (tcnt > 0)
2190 : {
2191 : AttributeOpts *aopt =
2192 209 : get_attribute_options(onerel->rd_id, stats->tupattnum);
2193 :
2194 209 : stats->exprvals = exprvals;
2195 209 : stats->exprnulls = exprnulls;
2196 209 : stats->rowstride = 1;
2197 209 : stats->compute_stats(stats,
2198 : expr_fetch_func,
2199 : tcnt,
2200 : tcnt);
2201 :
2202 : /*
2203 : * If the n_distinct option is specified, it overrides the above
2204 : * computation.
2205 : */
2206 209 : if (aopt != NULL && aopt->n_distinct != 0.0)
2207 0 : stats->stadistinct = aopt->n_distinct;
2208 : }
2209 :
2210 : /* And clean up */
2211 209 : MemoryContextSwitchTo(expr_context);
2212 :
2213 209 : ExecDropSingleTupleTableSlot(slot);
2214 209 : FreeExecutorState(estate);
2215 209 : MemoryContextReset(expr_context);
2216 : }
2217 :
2218 113 : MemoryContextSwitchTo(old_context);
2219 113 : MemoryContextDelete(expr_context);
2220 113 : }
2221 :
2222 :
2223 : /*
2224 : * Fetch function for analyzing statistics object expressions.
2225 : *
2226 : * We have not bothered to construct tuples from the data, instead the data
2227 : * is just in Datum arrays.
2228 : */
2229 : static Datum
2230 230246 : expr_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
2231 : {
2232 : int i;
2233 :
2234 : /* exprvals and exprnulls are already offset for proper column */
2235 230246 : i = rownum * stats->rowstride;
2236 230246 : *isNull = stats->exprnulls[i];
2237 230246 : return stats->exprvals[i];
2238 : }
2239 :
2240 : /*
2241 : * Build analyze data for a list of expressions. As this is not tied
2242 : * directly to a relation (table or index), we have to fake some of
2243 : * the fields in examine_expression().
2244 : */
2245 : static AnlExprData *
2246 113 : build_expr_data(List *exprs, int stattarget)
2247 : {
2248 : int idx;
2249 113 : int nexprs = list_length(exprs);
2250 : AnlExprData *exprdata;
2251 : ListCell *lc;
2252 :
2253 113 : exprdata = (AnlExprData *) palloc0(nexprs * sizeof(AnlExprData));
2254 :
2255 113 : idx = 0;
2256 322 : foreach(lc, exprs)
2257 : {
2258 209 : Node *expr = (Node *) lfirst(lc);
2259 209 : AnlExprData *thisdata = &exprdata[idx];
2260 :
2261 209 : thisdata->expr = expr;
2262 209 : thisdata->vacattrstat = examine_expression(expr, stattarget);
2263 209 : idx++;
2264 : }
2265 :
2266 113 : return exprdata;
2267 : }
2268 :
2269 : /* form an array of pg_statistic rows (per update_attstats) */
2270 : static Datum
2271 113 : serialize_expr_stats(AnlExprData *exprdata, int nexprs)
2272 : {
2273 : int exprno;
2274 : Oid typOid;
2275 : Relation sd;
2276 :
2277 113 : ArrayBuildState *astate = NULL;
2278 :
2279 113 : sd = table_open(StatisticRelationId, RowExclusiveLock);
2280 :
2281 : /* lookup OID of composite type for pg_statistic */
2282 113 : typOid = get_rel_type_id(StatisticRelationId);
2283 113 : if (!OidIsValid(typOid))
2284 0 : ereport(ERROR,
2285 : (errcode(ERRCODE_WRONG_OBJECT_TYPE),
2286 : errmsg("relation \"%s\" does not have a composite type",
2287 : "pg_statistic")));
2288 :
2289 322 : for (exprno = 0; exprno < nexprs; exprno++)
2290 : {
2291 : int i,
2292 : k;
2293 209 : VacAttrStats *stats = exprdata[exprno].vacattrstat;
2294 :
2295 : Datum values[Natts_pg_statistic];
2296 : bool nulls[Natts_pg_statistic];
2297 : HeapTuple stup;
2298 :
2299 209 : if (!stats->stats_valid)
2300 : {
2301 0 : astate = accumArrayResult(astate,
2302 : (Datum) 0,
2303 : true,
2304 : typOid,
2305 : CurrentMemoryContext);
2306 0 : continue;
2307 : }
2308 :
2309 : /*
2310 : * Construct a new pg_statistic tuple
2311 : */
2312 6688 : for (i = 0; i < Natts_pg_statistic; ++i)
2313 : {
2314 6479 : nulls[i] = false;
2315 : }
2316 :
2317 209 : values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(InvalidOid);
2318 209 : values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(InvalidAttrNumber);
2319 209 : values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(false);
2320 209 : values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
2321 209 : values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
2322 209 : values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
2323 209 : i = Anum_pg_statistic_stakind1 - 1;
2324 1254 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
2325 : {
2326 1045 : values[i++] = Int16GetDatum(stats->stakind[k]); /* stakindN */
2327 : }
2328 209 : i = Anum_pg_statistic_staop1 - 1;
2329 1254 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
2330 : {
2331 1045 : values[i++] = ObjectIdGetDatum(stats->staop[k]); /* staopN */
2332 : }
2333 209 : i = Anum_pg_statistic_stacoll1 - 1;
2334 1254 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
2335 : {
2336 1045 : values[i++] = ObjectIdGetDatum(stats->stacoll[k]); /* stacollN */
2337 : }
2338 209 : i = Anum_pg_statistic_stanumbers1 - 1;
2339 1254 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
2340 : {
2341 1045 : int nnum = stats->numnumbers[k];
2342 :
2343 1045 : if (nnum > 0)
2344 : {
2345 : int n;
2346 373 : Datum *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
2347 : ArrayType *arry;
2348 :
2349 2659 : for (n = 0; n < nnum; n++)
2350 2286 : numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
2351 373 : arry = construct_array_builtin(numdatums, nnum, FLOAT4OID);
2352 373 : values[i++] = PointerGetDatum(arry); /* stanumbersN */
2353 : }
2354 : else
2355 : {
2356 672 : nulls[i] = true;
2357 672 : values[i++] = (Datum) 0;
2358 : }
2359 : }
2360 209 : i = Anum_pg_statistic_stavalues1 - 1;
2361 1254 : for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
2362 : {
2363 1045 : if (stats->numvalues[k] > 0)
2364 : {
2365 : ArrayType *arry;
2366 :
2367 227 : arry = construct_array(stats->stavalues[k],
2368 : stats->numvalues[k],
2369 : stats->statypid[k],
2370 227 : stats->statyplen[k],
2371 227 : stats->statypbyval[k],
2372 227 : stats->statypalign[k]);
2373 227 : values[i++] = PointerGetDatum(arry); /* stavaluesN */
2374 : }
2375 : else
2376 : {
2377 818 : nulls[i] = true;
2378 818 : values[i++] = (Datum) 0;
2379 : }
2380 : }
2381 :
2382 209 : stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
2383 :
2384 209 : astate = accumArrayResult(astate,
2385 : heap_copy_tuple_as_datum(stup, RelationGetDescr(sd)),
2386 : false,
2387 : typOid,
2388 : CurrentMemoryContext);
2389 : }
2390 :
2391 113 : table_close(sd, RowExclusiveLock);
2392 :
2393 113 : return makeArrayResult(astate, CurrentMemoryContext);
2394 : }
2395 :
2396 : /*
2397 : * Loads pg_statistic record from expression statistics for expression
2398 : * identified by the supplied index.
2399 : */
2400 : HeapTuple
2401 825 : statext_expressions_load(Oid stxoid, bool inh, int idx)
2402 : {
2403 : bool isnull;
2404 : Datum value;
2405 : HeapTuple htup;
2406 : ExpandedArrayHeader *eah;
2407 : HeapTupleHeader td;
2408 : HeapTupleData tmptup;
2409 : HeapTuple tup;
2410 :
2411 825 : htup = SearchSysCache2(STATEXTDATASTXOID,
2412 : ObjectIdGetDatum(stxoid), BoolGetDatum(inh));
2413 825 : if (!HeapTupleIsValid(htup))
2414 0 : elog(ERROR, "cache lookup failed for statistics object %u", stxoid);
2415 :
2416 825 : value = SysCacheGetAttr(STATEXTDATASTXOID, htup,
2417 : Anum_pg_statistic_ext_data_stxdexpr, &isnull);
2418 825 : if (isnull)
2419 0 : elog(ERROR,
2420 : "requested statistics kind \"%c\" is not yet built for statistics object %u",
2421 : STATS_EXT_EXPRESSIONS, stxoid);
2422 :
2423 825 : eah = DatumGetExpandedArray(value);
2424 :
2425 825 : deconstruct_expanded_array(eah);
2426 :
2427 825 : td = DatumGetHeapTupleHeader(eah->dvalues[idx]);
2428 :
2429 : /* Build a temporary HeapTuple control structure */
2430 825 : tmptup.t_len = HeapTupleHeaderGetDatumLength(td);
2431 825 : ItemPointerSetInvalid(&(tmptup.t_self));
2432 825 : tmptup.t_tableOid = InvalidOid;
2433 825 : tmptup.t_data = td;
2434 :
2435 825 : tup = heap_copytuple(&tmptup);
2436 :
2437 825 : ReleaseSysCache(htup);
2438 :
2439 825 : return tup;
2440 : }
2441 :
2442 : /*
2443 : * Evaluate the expressions, so that we can use the results to build
2444 : * all the requested statistics types. This matters especially for
2445 : * expensive expressions, of course.
2446 : */
2447 : static StatsBuildData *
2448 257 : make_build_data(Relation rel, StatExtEntry *stat, int numrows, HeapTuple *rows,
2449 : VacAttrStats **stats, int stattarget)
2450 : {
2451 : /* evaluated expressions */
2452 : StatsBuildData *result;
2453 : char *ptr;
2454 : Size len;
2455 :
2456 : int i;
2457 : int k;
2458 : int idx;
2459 : TupleTableSlot *slot;
2460 : EState *estate;
2461 : ExprContext *econtext;
2462 257 : List *exprstates = NIL;
2463 257 : int nkeys = bms_num_members(stat->columns) + list_length(stat->exprs);
2464 : ListCell *lc;
2465 :
2466 : /* allocate everything as a single chunk, so we can free it easily */
2467 257 : len = MAXALIGN(sizeof(StatsBuildData));
2468 257 : len += MAXALIGN(sizeof(AttrNumber) * nkeys); /* attnums */
2469 257 : len += MAXALIGN(sizeof(VacAttrStats *) * nkeys); /* stats */
2470 :
2471 : /* values */
2472 257 : len += MAXALIGN(sizeof(Datum *) * nkeys);
2473 257 : len += nkeys * MAXALIGN(sizeof(Datum) * numrows);
2474 :
2475 : /* nulls */
2476 257 : len += MAXALIGN(sizeof(bool *) * nkeys);
2477 257 : len += nkeys * MAXALIGN(sizeof(bool) * numrows);
2478 :
2479 257 : ptr = palloc(len);
2480 :
2481 : /* set the pointers */
2482 257 : result = (StatsBuildData *) ptr;
2483 257 : ptr += MAXALIGN(sizeof(StatsBuildData));
2484 :
2485 : /* attnums */
2486 257 : result->attnums = (AttrNumber *) ptr;
2487 257 : ptr += MAXALIGN(sizeof(AttrNumber) * nkeys);
2488 :
2489 : /* stats */
2490 257 : result->stats = (VacAttrStats **) ptr;
2491 257 : ptr += MAXALIGN(sizeof(VacAttrStats *) * nkeys);
2492 :
2493 : /* values */
2494 257 : result->values = (Datum **) ptr;
2495 257 : ptr += MAXALIGN(sizeof(Datum *) * nkeys);
2496 :
2497 : /* nulls */
2498 257 : result->nulls = (bool **) ptr;
2499 257 : ptr += MAXALIGN(sizeof(bool *) * nkeys);
2500 :
2501 888 : for (i = 0; i < nkeys; i++)
2502 : {
2503 631 : result->values[i] = (Datum *) ptr;
2504 631 : ptr += MAXALIGN(sizeof(Datum) * numrows);
2505 :
2506 631 : result->nulls[i] = (bool *) ptr;
2507 631 : ptr += MAXALIGN(sizeof(bool) * numrows);
2508 : }
2509 :
2510 : Assert((ptr - (char *) result) == len);
2511 :
2512 : /* we have it allocated, so let's fill the values */
2513 257 : result->nattnums = nkeys;
2514 257 : result->numrows = numrows;
2515 :
2516 : /* fill the attribute info - first attributes, then expressions */
2517 257 : idx = 0;
2518 257 : k = -1;
2519 679 : while ((k = bms_next_member(stat->columns, k)) >= 0)
2520 : {
2521 422 : result->attnums[idx] = k;
2522 422 : result->stats[idx] = stats[idx];
2523 :
2524 422 : idx++;
2525 : }
2526 :
2527 257 : k = -1;
2528 466 : foreach(lc, stat->exprs)
2529 : {
2530 209 : Node *expr = (Node *) lfirst(lc);
2531 :
2532 209 : result->attnums[idx] = k;
2533 209 : result->stats[idx] = examine_expression(expr, stattarget);
2534 :
2535 209 : idx++;
2536 209 : k--;
2537 : }
2538 :
2539 : /* first extract values for all the regular attributes */
2540 472088 : for (i = 0; i < numrows; i++)
2541 : {
2542 471831 : idx = 0;
2543 471831 : k = -1;
2544 1523392 : while ((k = bms_next_member(stat->columns, k)) >= 0)
2545 : {
2546 2103122 : result->values[idx][i] = heap_getattr(rows[i], k,
2547 1051561 : result->stats[idx]->tupDesc,
2548 1051561 : &result->nulls[idx][i]);
2549 :
2550 1051561 : idx++;
2551 : }
2552 : }
2553 :
2554 : /* Need an EState for evaluation expressions. */
2555 257 : estate = CreateExecutorState();
2556 257 : econtext = GetPerTupleExprContext(estate);
2557 :
2558 : /* Need a slot to hold the current heap tuple, too */
2559 257 : slot = MakeSingleTupleTableSlot(RelationGetDescr(rel),
2560 : &TTSOpsHeapTuple);
2561 :
2562 : /* Arrange for econtext's scan tuple to be the tuple under test */
2563 257 : econtext->ecxt_scantuple = slot;
2564 :
2565 : /* Set up expression evaluation state */
2566 257 : exprstates = ExecPrepareExprList(stat->exprs, estate);
2567 :
2568 472088 : for (i = 0; i < numrows; i++)
2569 : {
2570 : /*
2571 : * Reset the per-tuple context each time, to reclaim any cruft left
2572 : * behind by evaluating the statistics object expressions.
2573 : */
2574 471831 : ResetExprContext(econtext);
2575 :
2576 : /* Set up for expression evaluation */
2577 471831 : ExecStoreHeapTuple(rows[i], slot, false);
2578 :
2579 471831 : idx = bms_num_members(stat->columns);
2580 702077 : foreach(lc, exprstates)
2581 : {
2582 : Datum datum;
2583 : bool isnull;
2584 230246 : ExprState *exprstate = (ExprState *) lfirst(lc);
2585 :
2586 : /*
2587 : * XXX This probably leaks memory. Maybe we should use
2588 : * ExecEvalExprSwitchContext but then we need to copy the result
2589 : * somewhere else.
2590 : */
2591 230246 : datum = ExecEvalExpr(exprstate,
2592 230246 : GetPerTupleExprContext(estate),
2593 : &isnull);
2594 230246 : if (isnull)
2595 : {
2596 6 : result->values[idx][i] = (Datum) 0;
2597 6 : result->nulls[idx][i] = true;
2598 : }
2599 : else
2600 : {
2601 230240 : result->values[idx][i] = datum;
2602 230240 : result->nulls[idx][i] = false;
2603 : }
2604 :
2605 230246 : idx++;
2606 : }
2607 : }
2608 :
2609 257 : ExecDropSingleTupleTableSlot(slot);
2610 257 : FreeExecutorState(estate);
2611 :
2612 257 : return result;
2613 : }
|
