]> granicus.if.org Git - postgresql/blob - src/backend/utils/adt/array_selfuncs.c
Fix initialization of fake LSN for unlogged relations
[postgresql] / src / backend / utils / adt / array_selfuncs.c
1 /*-------------------------------------------------------------------------
2  *
3  * array_selfuncs.c
4  *        Functions for selectivity estimation of array operators
5  *
6  * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
7  * Portions Copyright (c) 1994, Regents of the University of California
8  *
9  *
10  * IDENTIFICATION
11  *        src/backend/utils/adt/array_selfuncs.c
12  *
13  *-------------------------------------------------------------------------
14  */
15 #include "postgres.h"
16
17 #include <math.h>
18
19 #include "access/htup_details.h"
20 #include "catalog/pg_collation.h"
21 #include "catalog/pg_operator.h"
22 #include "catalog/pg_statistic.h"
23 #include "utils/array.h"
24 #include "utils/builtins.h"
25 #include "utils/lsyscache.h"
26 #include "utils/selfuncs.h"
27 #include "utils/typcache.h"
28
29
30 /* Default selectivity constant for "@>" and "<@" operators */
31 #define DEFAULT_CONTAIN_SEL 0.005
32
33 /* Default selectivity constant for "&&" operator */
34 #define DEFAULT_OVERLAP_SEL 0.01
35
36 /* Default selectivity for given operator */
37 #define DEFAULT_SEL(operator) \
38         ((operator) == OID_ARRAY_OVERLAP_OP ? \
39                 DEFAULT_OVERLAP_SEL : DEFAULT_CONTAIN_SEL)
40
41 static Selectivity calc_arraycontsel(VariableStatData *vardata, Datum constval,
42                                                                          Oid elemtype, Oid operator);
43 static Selectivity mcelem_array_selec(ArrayType *array,
44                                                                           TypeCacheEntry *typentry,
45                                                                           Datum *mcelem, int nmcelem,
46                                                                           float4 *numbers, int nnumbers,
47                                                                           float4 *hist, int nhist,
48                                                                           Oid operator);
49 static Selectivity mcelem_array_contain_overlap_selec(Datum *mcelem, int nmcelem,
50                                                                                                           float4 *numbers, int nnumbers,
51                                                                                                           Datum *array_data, int nitems,
52                                                                                                           Oid operator, TypeCacheEntry *typentry);
53 static Selectivity mcelem_array_contained_selec(Datum *mcelem, int nmcelem,
54                                                                                                 float4 *numbers, int nnumbers,
55                                                                                                 Datum *array_data, int nitems,
56                                                                                                 float4 *hist, int nhist,
57                                                                                                 Oid operator, TypeCacheEntry *typentry);
58 static float *calc_hist(const float4 *hist, int nhist, int n);
59 static float *calc_distr(const float *p, int n, int m, float rest);
60 static int      floor_log2(uint32 n);
61 static bool find_next_mcelem(Datum *mcelem, int nmcelem, Datum value,
62                                                          int *index, TypeCacheEntry *typentry);
63 static int      element_compare(const void *key1, const void *key2, void *arg);
64 static int      float_compare_desc(const void *key1, const void *key2);
65
66
67 /*
68  * scalararraysel_containment
69  *              Estimate selectivity of ScalarArrayOpExpr via array containment.
70  *
71  * If we have const =/<> ANY/ALL (array_var) then we can estimate the
72  * selectivity as though this were an array containment operator,
73  * array_var op ARRAY[const].
74  *
75  * scalararraysel() has already verified that the ScalarArrayOpExpr's operator
76  * is the array element type's default equality or inequality operator, and
77  * has aggressively simplified both inputs to constants.
78  *
79  * Returns selectivity (0..1), or -1 if we fail to estimate selectivity.
80  */
81 Selectivity
82 scalararraysel_containment(PlannerInfo *root,
83                                                    Node *leftop, Node *rightop,
84                                                    Oid elemtype, bool isEquality, bool useOr,
85                                                    int varRelid)
86 {
87         Selectivity selec;
88         VariableStatData vardata;
89         Datum           constval;
90         TypeCacheEntry *typentry;
91         FmgrInfo   *cmpfunc;
92
93         /*
94          * rightop must be a variable, else punt.
95          */
96         examine_variable(root, rightop, varRelid, &vardata);
97         if (!vardata.rel)
98         {
99                 ReleaseVariableStats(vardata);
100                 return -1.0;
101         }
102
103         /*
104          * leftop must be a constant, else punt.
105          */
106         if (!IsA(leftop, Const))
107         {
108                 ReleaseVariableStats(vardata);
109                 return -1.0;
110         }
111         if (((Const *) leftop)->constisnull)
112         {
113                 /* qual can't succeed if null on left */
114                 ReleaseVariableStats(vardata);
115                 return (Selectivity) 0.0;
116         }
117         constval = ((Const *) leftop)->constvalue;
118
119         /* Get element type's default comparison function */
120         typentry = lookup_type_cache(elemtype, TYPECACHE_CMP_PROC_FINFO);
121         if (!OidIsValid(typentry->cmp_proc_finfo.fn_oid))
122         {
123                 ReleaseVariableStats(vardata);
124                 return -1.0;
125         }
126         cmpfunc = &typentry->cmp_proc_finfo;
127
128         /*
129          * If the operator is <>, swap ANY/ALL, then invert the result later.
130          */
131         if (!isEquality)
132                 useOr = !useOr;
133
134         /* Get array element stats for var, if available */
135         if (HeapTupleIsValid(vardata.statsTuple) &&
136                 statistic_proc_security_check(&vardata, cmpfunc->fn_oid))
137         {
138                 Form_pg_statistic stats;
139                 AttStatsSlot sslot;
140                 AttStatsSlot hslot;
141
142                 stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
143
144                 /* MCELEM will be an array of same type as element */
145                 if (get_attstatsslot(&sslot, vardata.statsTuple,
146                                                          STATISTIC_KIND_MCELEM, InvalidOid,
147                                                          ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
148                 {
149                         /* For ALL case, also get histogram of distinct-element counts */
150                         if (useOr ||
151                                 !get_attstatsslot(&hslot, vardata.statsTuple,
152                                                                   STATISTIC_KIND_DECHIST, InvalidOid,
153                                                                   ATTSTATSSLOT_NUMBERS))
154                                 memset(&hslot, 0, sizeof(hslot));
155
156                         /*
157                          * For = ANY, estimate as var @> ARRAY[const].
158                          *
159                          * For = ALL, estimate as var <@ ARRAY[const].
160                          */
161                         if (useOr)
162                                 selec = mcelem_array_contain_overlap_selec(sslot.values,
163                                                                                                                    sslot.nvalues,
164                                                                                                                    sslot.numbers,
165                                                                                                                    sslot.nnumbers,
166                                                                                                                    &constval, 1,
167                                                                                                                    OID_ARRAY_CONTAINS_OP,
168                                                                                                                    typentry);
169                         else
170                                 selec = mcelem_array_contained_selec(sslot.values,
171                                                                                                          sslot.nvalues,
172                                                                                                          sslot.numbers,
173                                                                                                          sslot.nnumbers,
174                                                                                                          &constval, 1,
175                                                                                                          hslot.numbers,
176                                                                                                          hslot.nnumbers,
177                                                                                                          OID_ARRAY_CONTAINED_OP,
178                                                                                                          typentry);
179
180                         free_attstatsslot(&hslot);
181                         free_attstatsslot(&sslot);
182                 }
183                 else
184                 {
185                         /* No most-common-elements info, so do without */
186                         if (useOr)
187                                 selec = mcelem_array_contain_overlap_selec(NULL, 0,
188                                                                                                                    NULL, 0,
189                                                                                                                    &constval, 1,
190                                                                                                                    OID_ARRAY_CONTAINS_OP,
191                                                                                                                    typentry);
192                         else
193                                 selec = mcelem_array_contained_selec(NULL, 0,
194                                                                                                          NULL, 0,
195                                                                                                          &constval, 1,
196                                                                                                          NULL, 0,
197                                                                                                          OID_ARRAY_CONTAINED_OP,
198                                                                                                          typentry);
199                 }
200
201                 /*
202                  * MCE stats count only non-null rows, so adjust for null rows.
203                  */
204                 selec *= (1.0 - stats->stanullfrac);
205         }
206         else
207         {
208                 /* No stats at all, so do without */
209                 if (useOr)
210                         selec = mcelem_array_contain_overlap_selec(NULL, 0,
211                                                                                                            NULL, 0,
212                                                                                                            &constval, 1,
213                                                                                                            OID_ARRAY_CONTAINS_OP,
214                                                                                                            typentry);
215                 else
216                         selec = mcelem_array_contained_selec(NULL, 0,
217                                                                                                  NULL, 0,
218                                                                                                  &constval, 1,
219                                                                                                  NULL, 0,
220                                                                                                  OID_ARRAY_CONTAINED_OP,
221                                                                                                  typentry);
222                 /* we assume no nulls here, so no stanullfrac correction */
223         }
224
225         ReleaseVariableStats(vardata);
226
227         /*
228          * If the operator is <>, invert the results.
229          */
230         if (!isEquality)
231                 selec = 1.0 - selec;
232
233         CLAMP_PROBABILITY(selec);
234
235         return selec;
236 }
237
238 /*
239  * arraycontsel -- restriction selectivity for array @>, &&, <@ operators
240  */
241 Datum
242 arraycontsel(PG_FUNCTION_ARGS)
243 {
244         PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
245         Oid                     operator = PG_GETARG_OID(1);
246         List       *args = (List *) PG_GETARG_POINTER(2);
247         int                     varRelid = PG_GETARG_INT32(3);
248         VariableStatData vardata;
249         Node       *other;
250         bool            varonleft;
251         Selectivity selec;
252         Oid                     element_typeid;
253
254         /*
255          * If expression is not (variable op something) or (something op
256          * variable), then punt and return a default estimate.
257          */
258         if (!get_restriction_variable(root, args, varRelid,
259                                                                   &vardata, &other, &varonleft))
260                 PG_RETURN_FLOAT8(DEFAULT_SEL(operator));
261
262         /*
263          * Can't do anything useful if the something is not a constant, either.
264          */
265         if (!IsA(other, Const))
266         {
267                 ReleaseVariableStats(vardata);
268                 PG_RETURN_FLOAT8(DEFAULT_SEL(operator));
269         }
270
271         /*
272          * The "&&", "@>" and "<@" operators are strict, so we can cope with a
273          * NULL constant right away.
274          */
275         if (((Const *) other)->constisnull)
276         {
277                 ReleaseVariableStats(vardata);
278                 PG_RETURN_FLOAT8(0.0);
279         }
280
281         /*
282          * If var is on the right, commute the operator, so that we can assume the
283          * var is on the left in what follows.
284          */
285         if (!varonleft)
286         {
287                 if (operator == OID_ARRAY_CONTAINS_OP)
288                         operator = OID_ARRAY_CONTAINED_OP;
289                 else if (operator == OID_ARRAY_CONTAINED_OP)
290                         operator = OID_ARRAY_CONTAINS_OP;
291         }
292
293         /*
294          * OK, there's a Var and a Const we're dealing with here.  We need the
295          * Const to be an array with same element type as column, else we can't do
296          * anything useful.  (Such cases will likely fail at runtime, but here
297          * we'd rather just return a default estimate.)
298          */
299         element_typeid = get_base_element_type(((Const *) other)->consttype);
300         if (element_typeid != InvalidOid &&
301                 element_typeid == get_base_element_type(vardata.vartype))
302         {
303                 selec = calc_arraycontsel(&vardata, ((Const *) other)->constvalue,
304                                                                   element_typeid, operator);
305         }
306         else
307         {
308                 selec = DEFAULT_SEL(operator);
309         }
310
311         ReleaseVariableStats(vardata);
312
313         CLAMP_PROBABILITY(selec);
314
315         PG_RETURN_FLOAT8((float8) selec);
316 }
317
318 /*
319  * arraycontjoinsel -- join selectivity for array @>, &&, <@ operators
320  */
321 Datum
322 arraycontjoinsel(PG_FUNCTION_ARGS)
323 {
324         /* For the moment this is just a stub */
325         Oid                     operator = PG_GETARG_OID(1);
326
327         PG_RETURN_FLOAT8(DEFAULT_SEL(operator));
328 }
329
330 /*
331  * Calculate selectivity for "arraycolumn @> const", "arraycolumn && const"
332  * or "arraycolumn <@ const" based on the statistics
333  *
334  * This function is mainly responsible for extracting the pg_statistic data
335  * to be used; we then pass the problem on to mcelem_array_selec().
336  */
337 static Selectivity
338 calc_arraycontsel(VariableStatData *vardata, Datum constval,
339                                   Oid elemtype, Oid operator)
340 {
341         Selectivity selec;
342         TypeCacheEntry *typentry;
343         FmgrInfo   *cmpfunc;
344         ArrayType  *array;
345
346         /* Get element type's default comparison function */
347         typentry = lookup_type_cache(elemtype, TYPECACHE_CMP_PROC_FINFO);
348         if (!OidIsValid(typentry->cmp_proc_finfo.fn_oid))
349                 return DEFAULT_SEL(operator);
350         cmpfunc = &typentry->cmp_proc_finfo;
351
352         /*
353          * The caller made sure the const is an array with same element type, so
354          * get it now
355          */
356         array = DatumGetArrayTypeP(constval);
357
358         if (HeapTupleIsValid(vardata->statsTuple) &&
359                 statistic_proc_security_check(vardata, cmpfunc->fn_oid))
360         {
361                 Form_pg_statistic stats;
362                 AttStatsSlot sslot;
363                 AttStatsSlot hslot;
364
365                 stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
366
367                 /* MCELEM will be an array of same type as column */
368                 if (get_attstatsslot(&sslot, vardata->statsTuple,
369                                                          STATISTIC_KIND_MCELEM, InvalidOid,
370                                                          ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
371                 {
372                         /*
373                          * For "array <@ const" case we also need histogram of distinct
374                          * element counts.
375                          */
376                         if (operator != OID_ARRAY_CONTAINED_OP ||
377                                 !get_attstatsslot(&hslot, vardata->statsTuple,
378                                                                   STATISTIC_KIND_DECHIST, InvalidOid,
379                                                                   ATTSTATSSLOT_NUMBERS))
380                                 memset(&hslot, 0, sizeof(hslot));
381
382                         /* Use the most-common-elements slot for the array Var. */
383                         selec = mcelem_array_selec(array, typentry,
384                                                                            sslot.values, sslot.nvalues,
385                                                                            sslot.numbers, sslot.nnumbers,
386                                                                            hslot.numbers, hslot.nnumbers,
387                                                                            operator);
388
389                         free_attstatsslot(&hslot);
390                         free_attstatsslot(&sslot);
391                 }
392                 else
393                 {
394                         /* No most-common-elements info, so do without */
395                         selec = mcelem_array_selec(array, typentry,
396                                                                            NULL, 0, NULL, 0, NULL, 0,
397                                                                            operator);
398                 }
399
400                 /*
401                  * MCE stats count only non-null rows, so adjust for null rows.
402                  */
403                 selec *= (1.0 - stats->stanullfrac);
404         }
405         else
406         {
407                 /* No stats at all, so do without */
408                 selec = mcelem_array_selec(array, typentry,
409                                                                    NULL, 0, NULL, 0, NULL, 0,
410                                                                    operator);
411                 /* we assume no nulls here, so no stanullfrac correction */
412         }
413
414         /* If constant was toasted, release the copy we made */
415         if (PointerGetDatum(array) != constval)
416                 pfree(array);
417
418         return selec;
419 }
420
421 /*
422  * Array selectivity estimation based on most common elements statistics
423  *
424  * This function just deconstructs and sorts the array constant's contents,
425  * and then passes the problem on to mcelem_array_contain_overlap_selec or
426  * mcelem_array_contained_selec depending on the operator.
427  */
428 static Selectivity
429 mcelem_array_selec(ArrayType *array, TypeCacheEntry *typentry,
430                                    Datum *mcelem, int nmcelem,
431                                    float4 *numbers, int nnumbers,
432                                    float4 *hist, int nhist,
433                                    Oid operator)
434 {
435         Selectivity selec;
436         int                     num_elems;
437         Datum      *elem_values;
438         bool       *elem_nulls;
439         bool            null_present;
440         int                     nonnull_nitems;
441         int                     i;
442
443         /*
444          * Prepare constant array data for sorting.  Sorting lets us find unique
445          * elements and efficiently merge with the MCELEM array.
446          */
447         deconstruct_array(array,
448                                           typentry->type_id,
449                                           typentry->typlen,
450                                           typentry->typbyval,
451                                           typentry->typalign,
452                                           &elem_values, &elem_nulls, &num_elems);
453
454         /* Collapse out any null elements */
455         nonnull_nitems = 0;
456         null_present = false;
457         for (i = 0; i < num_elems; i++)
458         {
459                 if (elem_nulls[i])
460                         null_present = true;
461                 else
462                         elem_values[nonnull_nitems++] = elem_values[i];
463         }
464
465         /*
466          * Query "column @> '{anything, null}'" matches nothing.  For the other
467          * two operators, presence of a null in the constant can be ignored.
468          */
469         if (null_present && operator == OID_ARRAY_CONTAINS_OP)
470         {
471                 pfree(elem_values);
472                 pfree(elem_nulls);
473                 return (Selectivity) 0.0;
474         }
475
476         /* Sort extracted elements using their default comparison function. */
477         qsort_arg(elem_values, nonnull_nitems, sizeof(Datum),
478                           element_compare, typentry);
479
480         /* Separate cases according to operator */
481         if (operator == OID_ARRAY_CONTAINS_OP || operator == OID_ARRAY_OVERLAP_OP)
482                 selec = mcelem_array_contain_overlap_selec(mcelem, nmcelem,
483                                                                                                    numbers, nnumbers,
484                                                                                                    elem_values, nonnull_nitems,
485                                                                                                    operator, typentry);
486         else if (operator == OID_ARRAY_CONTAINED_OP)
487                 selec = mcelem_array_contained_selec(mcelem, nmcelem,
488                                                                                          numbers, nnumbers,
489                                                                                          elem_values, nonnull_nitems,
490                                                                                          hist, nhist,
491                                                                                          operator, typentry);
492         else
493         {
494                 elog(ERROR, "arraycontsel called for unrecognized operator %u",
495                          operator);
496                 selec = 0.0;                    /* keep compiler quiet */
497         }
498
499         pfree(elem_values);
500         pfree(elem_nulls);
501         return selec;
502 }
503
504 /*
505  * Estimate selectivity of "column @> const" and "column && const" based on
506  * most common element statistics.  This estimation assumes element
507  * occurrences are independent.
508  *
509  * mcelem (of length nmcelem) and numbers (of length nnumbers) are from
510  * the array column's MCELEM statistics slot, or are NULL/0 if stats are
511  * not available.  array_data (of length nitems) is the constant's elements.
512  *
513  * Both the mcelem and array_data arrays are assumed presorted according
514  * to the element type's cmpfunc.  Null elements are not present.
515  *
516  * TODO: this estimate probably could be improved by using the distinct
517  * elements count histogram.  For example, excepting the special case of
518  * "column @> '{}'", we can multiply the calculated selectivity by the
519  * fraction of nonempty arrays in the column.
520  */
521 static Selectivity
522 mcelem_array_contain_overlap_selec(Datum *mcelem, int nmcelem,
523                                                                    float4 *numbers, int nnumbers,
524                                                                    Datum *array_data, int nitems,
525                                                                    Oid operator, TypeCacheEntry *typentry)
526 {
527         Selectivity selec,
528                                 elem_selec;
529         int                     mcelem_index,
530                                 i;
531         bool            use_bsearch;
532         float4          minfreq;
533
534         /*
535          * There should be three more Numbers than Values, because the last three
536          * cells should hold minimal and maximal frequency among the non-null
537          * elements, and then the frequency of null elements.  Ignore the Numbers
538          * if not right.
539          */
540         if (nnumbers != nmcelem + 3)
541         {
542                 numbers = NULL;
543                 nnumbers = 0;
544         }
545
546         if (numbers)
547         {
548                 /* Grab the lowest observed frequency */
549                 minfreq = numbers[nmcelem];
550         }
551         else
552         {
553                 /* Without statistics make some default assumptions */
554                 minfreq = 2 * (float4) DEFAULT_CONTAIN_SEL;
555         }
556
557         /* Decide whether it is faster to use binary search or not. */
558         if (nitems * floor_log2((uint32) nmcelem) < nmcelem + nitems)
559                 use_bsearch = true;
560         else
561                 use_bsearch = false;
562
563         if (operator == OID_ARRAY_CONTAINS_OP)
564         {
565                 /*
566                  * Initial selectivity for "column @> const" query is 1.0, and it will
567                  * be decreased with each element of constant array.
568                  */
569                 selec = 1.0;
570         }
571         else
572         {
573                 /*
574                  * Initial selectivity for "column && const" query is 0.0, and it will
575                  * be increased with each element of constant array.
576                  */
577                 selec = 0.0;
578         }
579
580         /* Scan mcelem and array in parallel. */
581         mcelem_index = 0;
582         for (i = 0; i < nitems; i++)
583         {
584                 bool            match = false;
585
586                 /* Ignore any duplicates in the array data. */
587                 if (i > 0 &&
588                         element_compare(&array_data[i - 1], &array_data[i], typentry) == 0)
589                         continue;
590
591                 /* Find the smallest MCELEM >= this array item. */
592                 if (use_bsearch)
593                 {
594                         match = find_next_mcelem(mcelem, nmcelem, array_data[i],
595                                                                          &mcelem_index, typentry);
596                 }
597                 else
598                 {
599                         while (mcelem_index < nmcelem)
600                         {
601                                 int                     cmp = element_compare(&mcelem[mcelem_index],
602                                                                                                   &array_data[i],
603                                                                                                   typentry);
604
605                                 if (cmp < 0)
606                                         mcelem_index++;
607                                 else
608                                 {
609                                         if (cmp == 0)
610                                                 match = true;   /* mcelem is found */
611                                         break;
612                                 }
613                         }
614                 }
615
616                 if (match && numbers)
617                 {
618                         /* MCELEM matches the array item; use its frequency. */
619                         elem_selec = numbers[mcelem_index];
620                         mcelem_index++;
621                 }
622                 else
623                 {
624                         /*
625                          * The element is not in MCELEM.  Punt, but assume that the
626                          * selectivity cannot be more than minfreq / 2.
627                          */
628                         elem_selec = Min(DEFAULT_CONTAIN_SEL, minfreq / 2);
629                 }
630
631                 /*
632                  * Update overall selectivity using the current element's selectivity
633                  * and an assumption of element occurrence independence.
634                  */
635                 if (operator == OID_ARRAY_CONTAINS_OP)
636                         selec *= elem_selec;
637                 else
638                         selec = selec + elem_selec - selec * elem_selec;
639
640                 /* Clamp intermediate results to stay sane despite roundoff error */
641                 CLAMP_PROBABILITY(selec);
642         }
643
644         return selec;
645 }
646
647 /*
648  * Estimate selectivity of "column <@ const" based on most common element
649  * statistics.
650  *
651  * mcelem (of length nmcelem) and numbers (of length nnumbers) are from
652  * the array column's MCELEM statistics slot, or are NULL/0 if stats are
653  * not available.  array_data (of length nitems) is the constant's elements.
654  * hist (of length nhist) is from the array column's DECHIST statistics slot,
655  * or is NULL/0 if those stats are not available.
656  *
657  * Both the mcelem and array_data arrays are assumed presorted according
658  * to the element type's cmpfunc.  Null elements are not present.
659  *
660  * Independent element occurrence would imply a particular distribution of
661  * distinct element counts among matching rows.  Real data usually falsifies
662  * that assumption.  For example, in a set of 11-element integer arrays having
663  * elements in the range [0..10], element occurrences are typically not
664  * independent.  If they were, a sufficiently-large set would include all
665  * distinct element counts 0 through 11.  We correct for this using the
666  * histogram of distinct element counts.
667  *
668  * In the "column @> const" and "column && const" cases, we usually have a
669  * "const" with low number of elements (otherwise we have selectivity close
670  * to 0 or 1 respectively).  That's why the effect of dependence related
671  * to distinct element count distribution is negligible there.  In the
672  * "column <@ const" case, number of elements is usually high (otherwise we
673  * have selectivity close to 0).  That's why we should do a correction with
674  * the array distinct element count distribution here.
675  *
676  * Using the histogram of distinct element counts produces a different
677  * distribution law than independent occurrences of elements.  This
678  * distribution law can be described as follows:
679  *
680  * P(o1, o2, ..., on) = f1^o1 * (1 - f1)^(1 - o1) * f2^o2 *
681  *        (1 - f2)^(1 - o2) * ... * fn^on * (1 - fn)^(1 - on) * hist[m] / ind[m]
682  *
683  * where:
684  * o1, o2, ..., on - occurrences of elements 1, 2, ..., n
685  *              (1 - occurrence, 0 - no occurrence) in row
686  * f1, f2, ..., fn - frequencies of elements 1, 2, ..., n
687  *              (scalar values in [0..1]) according to collected statistics
688  * m = o1 + o2 + ... + on = total number of distinct elements in row
689  * hist[m] - histogram data for occurrence of m elements.
690  * ind[m] - probability of m occurrences from n events assuming their
691  *        probabilities to be equal to frequencies of array elements.
692  *
693  * ind[m] = sum(f1^o1 * (1 - f1)^(1 - o1) * f2^o2 * (1 - f2)^(1 - o2) *
694  * ... * fn^on * (1 - fn)^(1 - on), o1, o2, ..., on) | o1 + o2 + .. on = m
695  */
696 static Selectivity
697 mcelem_array_contained_selec(Datum *mcelem, int nmcelem,
698                                                          float4 *numbers, int nnumbers,
699                                                          Datum *array_data, int nitems,
700                                                          float4 *hist, int nhist,
701                                                          Oid operator, TypeCacheEntry *typentry)
702 {
703         int                     mcelem_index,
704                                 i,
705                                 unique_nitems = 0;
706         float           selec,
707                                 minfreq,
708                                 nullelem_freq;
709         float      *dist,
710                            *mcelem_dist,
711                            *hist_part;
712         float           avg_count,
713                                 mult,
714                                 rest;
715         float      *elem_selec;
716
717         /*
718          * There should be three more Numbers than Values in the MCELEM slot,
719          * because the last three cells should hold minimal and maximal frequency
720          * among the non-null elements, and then the frequency of null elements.
721          * Punt if not right, because we can't do much without the element freqs.
722          */
723         if (numbers == NULL || nnumbers != nmcelem + 3)
724                 return DEFAULT_CONTAIN_SEL;
725
726         /* Can't do much without a count histogram, either */
727         if (hist == NULL || nhist < 3)
728                 return DEFAULT_CONTAIN_SEL;
729
730         /*
731          * Grab some of the summary statistics that compute_array_stats() stores:
732          * lowest frequency, frequency of null elements, and average distinct
733          * element count.
734          */
735         minfreq = numbers[nmcelem];
736         nullelem_freq = numbers[nmcelem + 2];
737         avg_count = hist[nhist - 1];
738
739         /*
740          * "rest" will be the sum of the frequencies of all elements not
741          * represented in MCELEM.  The average distinct element count is the sum
742          * of the frequencies of *all* elements.  Begin with that; we will proceed
743          * to subtract the MCELEM frequencies.
744          */
745         rest = avg_count;
746
747         /*
748          * mult is a multiplier representing estimate of probability that each
749          * mcelem that is not present in constant doesn't occur.
750          */
751         mult = 1.0f;
752
753         /*
754          * elem_selec is array of estimated frequencies for elements in the
755          * constant.
756          */
757         elem_selec = (float *) palloc(sizeof(float) * nitems);
758
759         /* Scan mcelem and array in parallel. */
760         mcelem_index = 0;
761         for (i = 0; i < nitems; i++)
762         {
763                 bool            match = false;
764
765                 /* Ignore any duplicates in the array data. */
766                 if (i > 0 &&
767                         element_compare(&array_data[i - 1], &array_data[i], typentry) == 0)
768                         continue;
769
770                 /*
771                  * Iterate over MCELEM until we find an entry greater than or equal to
772                  * this element of the constant.  Update "rest" and "mult" for mcelem
773                  * entries skipped over.
774                  */
775                 while (mcelem_index < nmcelem)
776                 {
777                         int                     cmp = element_compare(&mcelem[mcelem_index],
778                                                                                           &array_data[i],
779                                                                                           typentry);
780
781                         if (cmp < 0)
782                         {
783                                 mult *= (1.0f - numbers[mcelem_index]);
784                                 rest -= numbers[mcelem_index];
785                                 mcelem_index++;
786                         }
787                         else
788                         {
789                                 if (cmp == 0)
790                                         match = true;   /* mcelem is found */
791                                 break;
792                         }
793                 }
794
795                 if (match)
796                 {
797                         /* MCELEM matches the array item. */
798                         elem_selec[unique_nitems] = numbers[mcelem_index];
799                         /* "rest" is decremented for all mcelems, matched or not */
800                         rest -= numbers[mcelem_index];
801                         mcelem_index++;
802                 }
803                 else
804                 {
805                         /*
806                          * The element is not in MCELEM.  Punt, but assume that the
807                          * selectivity cannot be more than minfreq / 2.
808                          */
809                         elem_selec[unique_nitems] = Min(DEFAULT_CONTAIN_SEL,
810                                                                                         minfreq / 2);
811                 }
812
813                 unique_nitems++;
814         }
815
816         /*
817          * If we handled all constant elements without exhausting the MCELEM
818          * array, finish walking it to complete calculation of "rest" and "mult".
819          */
820         while (mcelem_index < nmcelem)
821         {
822                 mult *= (1.0f - numbers[mcelem_index]);
823                 rest -= numbers[mcelem_index];
824                 mcelem_index++;
825         }
826
827         /*
828          * The presence of many distinct rare elements materially decreases
829          * selectivity.  Use the Poisson distribution to estimate the probability
830          * of a column value having zero occurrences of such elements.  See above
831          * for the definition of "rest".
832          */
833         mult *= exp(-rest);
834
835         /*----------
836          * Using the distinct element count histogram requires
837          *              O(unique_nitems * (nmcelem + unique_nitems))
838          * operations.  Beyond a certain computational cost threshold, it's
839          * reasonable to sacrifice accuracy for decreased planning time.  We limit
840          * the number of operations to EFFORT * nmcelem; since nmcelem is limited
841          * by the column's statistics target, the work done is user-controllable.
842          *
843          * If the number of operations would be too large, we can reduce it
844          * without losing all accuracy by reducing unique_nitems and considering
845          * only the most-common elements of the constant array.  To make the
846          * results exactly match what we would have gotten with only those
847          * elements to start with, we'd have to remove any discarded elements'
848          * frequencies from "mult", but since this is only an approximation
849          * anyway, we don't bother with that.  Therefore it's sufficient to qsort
850          * elem_selec[] and take the largest elements.  (They will no longer match
851          * up with the elements of array_data[], but we don't care.)
852          *----------
853          */
854 #define EFFORT 100
855
856         if ((nmcelem + unique_nitems) > 0 &&
857                 unique_nitems > EFFORT * nmcelem / (nmcelem + unique_nitems))
858         {
859                 /*
860                  * Use the quadratic formula to solve for largest allowable N.  We
861                  * have A = 1, B = nmcelem, C = - EFFORT * nmcelem.
862                  */
863                 double          b = (double) nmcelem;
864                 int                     n;
865
866                 n = (int) ((sqrt(b * b + 4 * EFFORT * b) - b) / 2);
867
868                 /* Sort, then take just the first n elements */
869                 qsort(elem_selec, unique_nitems, sizeof(float),
870                           float_compare_desc);
871                 unique_nitems = n;
872         }
873
874         /*
875          * Calculate probabilities of each distinct element count for both mcelems
876          * and constant elements.  At this point, assume independent element
877          * occurrence.
878          */
879         dist = calc_distr(elem_selec, unique_nitems, unique_nitems, 0.0f);
880         mcelem_dist = calc_distr(numbers, nmcelem, unique_nitems, rest);
881
882         /* ignore hist[nhist-1], which is the average not a histogram member */
883         hist_part = calc_hist(hist, nhist - 1, unique_nitems);
884
885         selec = 0.0f;
886         for (i = 0; i <= unique_nitems; i++)
887         {
888                 /*
889                  * mult * dist[i] / mcelem_dist[i] gives us probability of qual
890                  * matching from assumption of independent element occurrence with the
891                  * condition that distinct element count = i.
892                  */
893                 if (mcelem_dist[i] > 0)
894                         selec += hist_part[i] * mult * dist[i] / mcelem_dist[i];
895         }
896
897         pfree(dist);
898         pfree(mcelem_dist);
899         pfree(hist_part);
900         pfree(elem_selec);
901
902         /* Take into account occurrence of NULL element. */
903         selec *= (1.0f - nullelem_freq);
904
905         CLAMP_PROBABILITY(selec);
906
907         return selec;
908 }
909
910 /*
911  * Calculate the first n distinct element count probabilities from a
912  * histogram of distinct element counts.
913  *
914  * Returns a palloc'd array of n+1 entries, with array[k] being the
915  * probability of element count k, k in [0..n].
916  *
917  * We assume that a histogram box with bounds a and b gives 1 / ((b - a + 1) *
918  * (nhist - 1)) probability to each value in (a,b) and an additional half of
919  * that to a and b themselves.
920  */
921 static float *
922 calc_hist(const float4 *hist, int nhist, int n)
923 {
924         float      *hist_part;
925         int                     k,
926                                 i = 0;
927         float           prev_interval = 0,
928                                 next_interval;
929         float           frac;
930
931         hist_part = (float *) palloc((n + 1) * sizeof(float));
932
933         /*
934          * frac is a probability contribution for each interval between histogram
935          * values.  We have nhist - 1 intervals, so contribution of each one will
936          * be 1 / (nhist - 1).
937          */
938         frac = 1.0f / ((float) (nhist - 1));
939
940         for (k = 0; k <= n; k++)
941         {
942                 int                     count = 0;
943
944                 /*
945                  * Count the histogram boundaries equal to k.  (Although the histogram
946                  * should theoretically contain only exact integers, entries are
947                  * floats so there could be roundoff error in large values.  Treat any
948                  * fractional value as equal to the next larger k.)
949                  */
950                 while (i < nhist && hist[i] <= k)
951                 {
952                         count++;
953                         i++;
954                 }
955
956                 if (count > 0)
957                 {
958                         /* k is an exact bound for at least one histogram box. */
959                         float           val;
960
961                         /* Find length between current histogram value and the next one */
962                         if (i < nhist)
963                                 next_interval = hist[i] - hist[i - 1];
964                         else
965                                 next_interval = 0;
966
967                         /*
968                          * count - 1 histogram boxes contain k exclusively.  They
969                          * contribute a total of (count - 1) * frac probability.  Also
970                          * factor in the partial histogram boxes on either side.
971                          */
972                         val = (float) (count - 1);
973                         if (next_interval > 0)
974                                 val += 0.5f / next_interval;
975                         if (prev_interval > 0)
976                                 val += 0.5f / prev_interval;
977                         hist_part[k] = frac * val;
978
979                         prev_interval = next_interval;
980                 }
981                 else
982                 {
983                         /* k does not appear as an exact histogram bound. */
984                         if (prev_interval > 0)
985                                 hist_part[k] = frac / prev_interval;
986                         else
987                                 hist_part[k] = 0.0f;
988                 }
989         }
990
991         return hist_part;
992 }
993
994 /*
995  * Consider n independent events with probabilities p[].  This function
996  * calculates probabilities of exact k of events occurrence for k in [0..m].
997  * Returns a palloc'd array of size m+1.
998  *
999  * "rest" is the sum of the probabilities of all low-probability events not
1000  * included in p.
1001  *
1002  * Imagine matrix M of size (n + 1) x (m + 1).  Element M[i,j] denotes the
1003  * probability that exactly j of first i events occur.  Obviously M[0,0] = 1.
1004  * For any constant j, each increment of i increases the probability iff the
1005  * event occurs.  So, by the law of total probability:
1006  *      M[i,j] = M[i - 1, j] * (1 - p[i]) + M[i - 1, j - 1] * p[i]
1007  *              for i > 0, j > 0.
1008  *      M[i,0] = M[i - 1, 0] * (1 - p[i]) for i > 0.
1009  */
1010 static float *
1011 calc_distr(const float *p, int n, int m, float rest)
1012 {
1013         float      *row,
1014                            *prev_row,
1015                            *tmp;
1016         int                     i,
1017                                 j;
1018
1019         /*
1020          * Since we return only the last row of the matrix and need only the
1021          * current and previous row for calculations, allocate two rows.
1022          */
1023         row = (float *) palloc((m + 1) * sizeof(float));
1024         prev_row = (float *) palloc((m + 1) * sizeof(float));
1025
1026         /* M[0,0] = 1 */
1027         row[0] = 1.0f;
1028         for (i = 1; i <= n; i++)
1029         {
1030                 float           t = p[i - 1];
1031
1032                 /* Swap rows */
1033                 tmp = row;
1034                 row = prev_row;
1035                 prev_row = tmp;
1036
1037                 /* Calculate next row */
1038                 for (j = 0; j <= i && j <= m; j++)
1039                 {
1040                         float           val = 0.0f;
1041
1042                         if (j < i)
1043                                 val += prev_row[j] * (1.0f - t);
1044                         if (j > 0)
1045                                 val += prev_row[j - 1] * t;
1046                         row[j] = val;
1047                 }
1048         }
1049
1050         /*
1051          * The presence of many distinct rare (not in "p") elements materially
1052          * decreases selectivity.  Model their collective occurrence with the
1053          * Poisson distribution.
1054          */
1055         if (rest > DEFAULT_CONTAIN_SEL)
1056         {
1057                 float           t;
1058
1059                 /* Swap rows */
1060                 tmp = row;
1061                 row = prev_row;
1062                 prev_row = tmp;
1063
1064                 for (i = 0; i <= m; i++)
1065                         row[i] = 0.0f;
1066
1067                 /* Value of Poisson distribution for 0 occurrences */
1068                 t = exp(-rest);
1069
1070                 /*
1071                  * Calculate convolution of previously computed distribution and the
1072                  * Poisson distribution.
1073                  */
1074                 for (i = 0; i <= m; i++)
1075                 {
1076                         for (j = 0; j <= m - i; j++)
1077                                 row[j + i] += prev_row[j] * t;
1078
1079                         /* Get Poisson distribution value for (i + 1) occurrences */
1080                         t *= rest / (float) (i + 1);
1081                 }
1082         }
1083
1084         pfree(prev_row);
1085         return row;
1086 }
1087
1088 /* Fast function for floor value of 2 based logarithm calculation. */
1089 static int
1090 floor_log2(uint32 n)
1091 {
1092         int                     logval = 0;
1093
1094         if (n == 0)
1095                 return -1;
1096         if (n >= (1 << 16))
1097         {
1098                 n >>= 16;
1099                 logval += 16;
1100         }
1101         if (n >= (1 << 8))
1102         {
1103                 n >>= 8;
1104                 logval += 8;
1105         }
1106         if (n >= (1 << 4))
1107         {
1108                 n >>= 4;
1109                 logval += 4;
1110         }
1111         if (n >= (1 << 2))
1112         {
1113                 n >>= 2;
1114                 logval += 2;
1115         }
1116         if (n >= (1 << 1))
1117         {
1118                 logval += 1;
1119         }
1120         return logval;
1121 }
1122
1123 /*
1124  * find_next_mcelem binary-searches a most common elements array, starting
1125  * from *index, for the first member >= value.  It saves the position of the
1126  * match into *index and returns true if it's an exact match.  (Note: we
1127  * assume the mcelem elements are distinct so there can't be more than one
1128  * exact match.)
1129  */
1130 static bool
1131 find_next_mcelem(Datum *mcelem, int nmcelem, Datum value, int *index,
1132                                  TypeCacheEntry *typentry)
1133 {
1134         int                     l = *index,
1135                                 r = nmcelem - 1,
1136                                 i,
1137                                 res;
1138
1139         while (l <= r)
1140         {
1141                 i = (l + r) / 2;
1142                 res = element_compare(&mcelem[i], &value, typentry);
1143                 if (res == 0)
1144                 {
1145                         *index = i;
1146                         return true;
1147                 }
1148                 else if (res < 0)
1149                         l = i + 1;
1150                 else
1151                         r = i - 1;
1152         }
1153         *index = l;
1154         return false;
1155 }
1156
1157 /*
1158  * Comparison function for elements.
1159  *
1160  * We use the element type's default btree opclass, and its default collation
1161  * if the type is collation-sensitive.
1162  *
1163  * XXX consider using SortSupport infrastructure
1164  */
1165 static int
1166 element_compare(const void *key1, const void *key2, void *arg)
1167 {
1168         Datum           d1 = *((const Datum *) key1);
1169         Datum           d2 = *((const Datum *) key2);
1170         TypeCacheEntry *typentry = (TypeCacheEntry *) arg;
1171         FmgrInfo   *cmpfunc = &typentry->cmp_proc_finfo;
1172         Datum           c;
1173
1174         c = FunctionCall2Coll(cmpfunc, typentry->typcollation, d1, d2);
1175         return DatumGetInt32(c);
1176 }
1177
1178 /*
1179  * Comparison function for sorting floats into descending order.
1180  */
1181 static int
1182 float_compare_desc(const void *key1, const void *key2)
1183 {
1184         float           d1 = *((const float *) key1);
1185         float           d2 = *((const float *) key2);
1186
1187         if (d1 > d2)
1188                 return -1;
1189         else if (d1 < d2)
1190                 return 1;
1191         else
1192                 return 0;
1193 }