TLA Line data Source code
1 : /*
2 : * brin_minmax_multi.c
3 : * Implementation of Multi Min/Max opclass for BRIN
4 : *
5 : * Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
6 : * Portions Copyright (c) 1994, Regents of the University of California
7 : *
8 : *
9 : * Implements a variant of minmax opclass, where the summary is composed of
10 : * multiple smaller intervals. This allows us to handle outliers, which
11 : * usually make the simple minmax opclass inefficient.
12 : *
13 : * Consider for example page range with simple minmax interval [1000,2000],
14 : * and assume a new row gets inserted into the range with value 1000000.
15 : * Due to that the interval gets [1000,1000000]. I.e. the minmax interval
16 : * got 1000x wider and won't be useful to eliminate scan keys between 2001
17 : * and 1000000.
18 : *
19 : * With minmax-multi opclass, we may have [1000,2000] interval initially,
20 : * but after adding the new row we start tracking it as two interval:
21 : *
22 : * [1000,2000] and [1000000,1000000]
23 : *
24 : * This allows us to still eliminate the page range when the scan keys hit
25 : * the gap between 2000 and 1000000, making it useful in cases when the
26 : * simple minmax opclass gets inefficient.
27 : *
28 : * The number of intervals tracked per page range is somewhat flexible.
29 : * What is restricted is the number of values per page range, and the limit
30 : * is currently 32 (see values_per_range reloption). Collapsed intervals
31 : * (with equal minimum and maximum value) are stored as a single value,
32 : * while regular intervals require two values.
33 : *
34 : * When the number of values gets too high (by adding new values to the
35 : * summary), we merge some of the intervals to free space for more values.
36 : * This is done in a greedy way - we simply pick the two closest intervals,
37 : * merge them, and repeat this until the number of values to store gets
38 : * sufficiently low (below 50% of maximum values), but that is mostly
39 : * arbitrary threshold and may be changed easily).
40 : *
41 : * To pick the closest intervals we use the "distance" support procedure,
42 : * which measures space between two ranges (i.e. the length of an interval).
43 : * The computed value may be an approximation - in the worst case we will
44 : * merge two ranges that are slightly less optimal at that step, but the
45 : * index should still produce correct results.
46 : *
47 : * The compactions (reducing the number of values) is fairly expensive, as
48 : * it requires calling the distance functions, sorting etc. So when building
49 : * the summary, we use a significantly larger buffer, and only enforce the
50 : * exact limit at the very end. This improves performance, and it also helps
51 : * with building better ranges (due to the greedy approach).
52 : *
53 : *
54 : * IDENTIFICATION
55 : * src/backend/access/brin/brin_minmax_multi.c
56 : */
57 : #include "postgres.h"
58 :
59 : /* needed for PGSQL_AF_INET */
60 : #include <sys/socket.h>
61 :
62 : #include "access/genam.h"
63 : #include "access/brin.h"
64 : #include "access/brin_internal.h"
65 : #include "access/brin_tuple.h"
66 : #include "access/reloptions.h"
67 : #include "access/stratnum.h"
68 : #include "access/htup_details.h"
69 : #include "catalog/pg_type.h"
70 : #include "catalog/pg_am.h"
71 : #include "catalog/pg_amop.h"
72 : #include "utils/array.h"
73 : #include "utils/builtins.h"
74 : #include "utils/date.h"
75 : #include "utils/datum.h"
76 : #include "utils/float.h"
77 : #include "utils/inet.h"
78 : #include "utils/lsyscache.h"
79 : #include "utils/memutils.h"
80 : #include "utils/numeric.h"
81 : #include "utils/pg_lsn.h"
82 : #include "utils/rel.h"
83 : #include "utils/syscache.h"
84 : #include "utils/timestamp.h"
85 : #include "utils/uuid.h"
86 :
87 : /*
88 : * Additional SQL level support functions
89 : *
90 : * Procedure numbers must not use values reserved for BRIN itself; see
91 : * brin_internal.h.
92 : */
93 : #define MINMAX_MAX_PROCNUMS 1 /* maximum support procs we need */
94 : #define PROCNUM_DISTANCE 11 /* required, distance between values */
95 :
96 : /*
97 : * Subtract this from procnum to obtain index in MinmaxMultiOpaque arrays
98 : * (Must be equal to minimum of private procnums).
99 : */
100 : #define PROCNUM_BASE 11
101 :
102 : /*
103 : * Sizing the insert buffer - we use 10x the number of values specified
104 : * in the reloption, but we cap it to 8192 not to get too large. When
105 : * the buffer gets full, we reduce the number of values by half.
106 : */
107 : #define MINMAX_BUFFER_FACTOR 10
108 : #define MINMAX_BUFFER_MIN 256
109 : #define MINMAX_BUFFER_MAX 8192
110 : #define MINMAX_BUFFER_LOAD_FACTOR 0.5
111 :
112 : typedef struct MinmaxMultiOpaque
113 : {
114 : FmgrInfo extra_procinfos[MINMAX_MAX_PROCNUMS];
115 : bool extra_proc_missing[MINMAX_MAX_PROCNUMS];
116 : Oid cached_subtype;
117 : FmgrInfo strategy_procinfos[BTMaxStrategyNumber];
118 : } MinmaxMultiOpaque;
119 :
120 : /*
121 : * Storage type for BRIN's minmax reloptions
122 : */
123 : typedef struct MinMaxMultiOptions
124 : {
125 : int32 vl_len_; /* varlena header (do not touch directly!) */
126 : int valuesPerRange; /* number of values per range */
127 : } MinMaxMultiOptions;
128 :
129 : #define MINMAX_MULTI_DEFAULT_VALUES_PER_PAGE 32
130 :
131 : #define MinMaxMultiGetValuesPerRange(opts) \
132 : ((opts) && (((MinMaxMultiOptions *) (opts))->valuesPerRange != 0) ? \
133 : ((MinMaxMultiOptions *) (opts))->valuesPerRange : \
134 : MINMAX_MULTI_DEFAULT_VALUES_PER_PAGE)
135 :
136 : #define SAMESIGN(a,b) (((a) < 0) == ((b) < 0))
137 :
138 : /*
139 : * The summary of minmax-multi indexes has two representations - Ranges for
140 : * convenient processing, and SerializedRanges for storage in bytea value.
141 : *
142 : * The Ranges struct stores the boundary values in a single array, but we
143 : * treat regular and single-point ranges differently to save space. For
144 : * regular ranges (with different boundary values) we have to store both
145 : * the lower and upper bound of the range, while for "single-point ranges"
146 : * we only need to store a single value.
147 : *
148 : * The 'values' array stores boundary values for regular ranges first (there
149 : * are 2*nranges values to store), and then the nvalues boundary values for
150 : * single-point ranges. That is, we have (2*nranges + nvalues) boundary
151 : * values in the array.
152 : *
153 : * +-------------------------+----------------------------------+
154 : * | ranges (2 * nranges of) | single point values (nvalues of) |
155 : * +-------------------------+----------------------------------+
156 : *
157 : * This allows us to quickly add new values, and store outliers without
158 : * having to widen any of the existing range values.
159 : *
160 : * 'nsorted' denotes how many of 'nvalues' in the values[] array are sorted.
161 : * When nsorted == nvalues, all single point values are sorted.
162 : *
163 : * We never store more than maxvalues values (as set by values_per_range
164 : * reloption). If needed we merge some of the ranges.
165 : *
166 : * To minimize palloc overhead, we always allocate the full array with
167 : * space for maxvalues elements. This should be fine as long as the
168 : * maxvalues is reasonably small (64 seems fine), which is the case
169 : * thanks to values_per_range reloption being limited to 256.
170 : */
171 : typedef struct Ranges
172 : {
173 : /* Cache information that we need quite often. */
174 : Oid typid;
175 : Oid colloid;
176 : AttrNumber attno;
177 : FmgrInfo *cmp;
178 :
179 : /* (2*nranges + nvalues) <= maxvalues */
180 : int nranges; /* number of ranges in the values[] array */
181 : int nsorted; /* number of nvalues which are sorted */
182 : int nvalues; /* number of point values in values[] array */
183 : int maxvalues; /* number of elements in the values[] array */
184 :
185 : /*
186 : * We simply add the values into a large buffer, without any expensive
187 : * steps (sorting, deduplication, ...). The buffer is a multiple of the
188 : * target number of values, so the compaction happens less often,
189 : * amortizing the costs. We keep the actual target and compact to the
190 : * requested number of values at the very end, before serializing to
191 : * on-disk representation.
192 : */
193 : /* requested number of values */
194 : int target_maxvalues;
195 :
196 : /* values stored for this range - either raw values, or ranges */
197 : Datum values[FLEXIBLE_ARRAY_MEMBER];
198 : } Ranges;
199 :
200 : /*
201 : * On-disk the summary is stored as a bytea value, with a simple header
202 : * with basic metadata, followed by the boundary values. It has a varlena
203 : * header, so can be treated as varlena directly.
204 : *
205 : * See brin_range_serialize/brin_range_deserialize for serialization details.
206 : */
207 : typedef struct SerializedRanges
208 : {
209 : /* varlena header (do not touch directly!) */
210 : int32 vl_len_;
211 :
212 : /* type of values stored in the data array */
213 : Oid typid;
214 :
215 : /* (2*nranges + nvalues) <= maxvalues */
216 : int nranges; /* number of ranges in the array (stored) */
217 : int nvalues; /* number of values in the data array (all) */
218 : int maxvalues; /* maximum number of values (reloption) */
219 :
220 : /* contains the actual data */
221 : char data[FLEXIBLE_ARRAY_MEMBER];
222 : } SerializedRanges;
223 :
224 : static SerializedRanges *brin_range_serialize(Ranges *range);
225 :
226 : static Ranges *brin_range_deserialize(int maxvalues,
227 : SerializedRanges *serialized);
228 :
229 :
230 : /*
231 : * Used to represent ranges expanded to make merging and combining easier.
232 : *
233 : * Each expanded range is essentially an interval, represented by min/max
234 : * values, along with a flag whether it's a collapsed range (in which case
235 : * the min and max values are equal). We have the flag to handle by-ref
236 : * data types - we can't simply compare the datums, and this saves some
237 : * calls to the type-specific comparator function.
238 : */
239 : typedef struct ExpandedRange
240 : {
241 : Datum minval; /* lower boundary */
242 : Datum maxval; /* upper boundary */
243 : bool collapsed; /* true if minval==maxval */
244 : } ExpandedRange;
245 :
246 : /*
247 : * Represents a distance between two ranges (identified by index into
248 : * an array of extended ranges).
249 : */
250 : typedef struct DistanceValue
251 : {
252 : int index;
253 : double value;
254 : } DistanceValue;
255 :
256 :
257 : /* Cache for support and strategy procedures. */
258 :
259 : static FmgrInfo *minmax_multi_get_procinfo(BrinDesc *bdesc, uint16 attno,
260 : uint16 procnum);
261 :
262 : static FmgrInfo *minmax_multi_get_strategy_procinfo(BrinDesc *bdesc,
263 : uint16 attno, Oid subtype,
264 : uint16 strategynum);
265 :
266 : typedef struct compare_context
267 : {
268 : FmgrInfo *cmpFn;
269 : Oid colloid;
270 : } compare_context;
271 :
272 : static int compare_values(const void *a, const void *b, void *arg);
273 :
274 :
275 : #ifdef USE_ASSERT_CHECKING
276 : /*
277 : * Check that the order of the array values is correct, using the cmp
278 : * function (which should be BTLessStrategyNumber).
279 : */
280 : static void
281 GIC 189966 : AssertArrayOrder(FmgrInfo *cmp, Oid colloid, Datum *values, int nvalues)
282 ECB : {
283 : int i;
284 : Datum lt;
285 :
286 GIC 4848756 : for (i = 0; i < (nvalues - 1); i++)
287 ECB : {
288 GIC 4658790 : lt = FunctionCall2Coll(cmp, colloid, values[i], values[i + 1]);
289 CBC 4658790 : Assert(DatumGetBool(lt));
290 ECB : }
291 GIC 189966 : }
292 ECB : #endif
293 :
294 : /*
295 : * Comprehensive check of the Ranges structure.
296 : */
297 : static void
298 GIC 94983 : AssertCheckRanges(Ranges *ranges, FmgrInfo *cmpFn, Oid colloid)
299 ECB : {
300 : #ifdef USE_ASSERT_CHECKING
301 : int i;
302 :
303 : /* some basic sanity checks */
304 GIC 94983 : Assert(ranges->nranges >= 0);
305 CBC 94983 : Assert(ranges->nsorted >= 0);
306 94983 : Assert(ranges->nvalues >= ranges->nsorted);
307 94983 : Assert(ranges->maxvalues >= 2 * ranges->nranges + ranges->nvalues);
308 94983 : Assert(ranges->typid != InvalidOid);
309 ECB :
310 : /*
311 : * First the ranges - there are 2*nranges boundary values, and the values
312 : * have to be strictly ordered (equal values would mean the range is
313 : * collapsed, and should be stored as a point). This also guarantees that
314 : * the ranges do not overlap.
315 : */
316 GIC 94983 : AssertArrayOrder(cmpFn, colloid, ranges->values, 2 * ranges->nranges);
317 ECB :
318 : /* then the single-point ranges (with nvalues boundary values ) */
319 GIC 94983 : AssertArrayOrder(cmpFn, colloid, &ranges->values[2 * ranges->nranges],
320 ECB : ranges->nsorted);
321 :
322 : /*
323 : * Check that none of the values are not covered by ranges (both sorted
324 : * and unsorted)
325 : */
326 GIC 94983 : if (ranges->nranges > 0)
327 ECB : {
328 GIC 11620965 : for (i = 0; i < ranges->nvalues; i++)
329 ECB : {
330 : Datum compar;
331 : int start,
332 : end;
333 GIC 11562582 : Datum minvalue = ranges->values[0];
334 CBC 11562582 : Datum maxvalue = ranges->values[2 * ranges->nranges - 1];
335 11562582 : Datum value = ranges->values[2 * ranges->nranges + i];
336 ECB :
337 GIC 11562582 : compar = FunctionCall2Coll(cmpFn, colloid, value, minvalue);
338 ECB :
339 : /*
340 : * If the value is smaller than the lower bound in the first range
341 : * then it cannot possibly be in any of the ranges.
342 : */
343 GIC 11562582 : if (DatumGetBool(compar))
344 CBC 4607328 : continue;
345 ECB :
346 GIC 6955254 : compar = FunctionCall2Coll(cmpFn, colloid, maxvalue, value);
347 ECB :
348 : /*
349 : * Likewise, if the value is larger than the upper bound of the
350 : * final range, then it cannot possibly be inside any of the
351 : * ranges.
352 : */
353 GIC 6955254 : if (DatumGetBool(compar))
354 CBC 6955155 : continue;
355 ECB :
356 : /* bsearch the ranges to see if 'value' fits within any of them */
357 GIC 99 : start = 0; /* first range */
358 CBC 99 : end = ranges->nranges - 1; /* last range */
359 ECB : while (true)
360 GIC 357 : {
361 CBC 456 : int midpoint = (start + end) / 2;
362 ECB :
363 : /* this means we ran out of ranges in the last step */
364 GIC 456 : if (start > end)
365 CBC 99 : break;
366 ECB :
367 : /* copy the min/max values from the ranges */
368 GIC 357 : minvalue = ranges->values[2 * midpoint];
369 CBC 357 : maxvalue = ranges->values[2 * midpoint + 1];
370 ECB :
371 : /*
372 : * Is the value smaller than the minval? If yes, we'll recurse
373 : * to the left side of range array.
374 : */
375 GIC 357 : compar = FunctionCall2Coll(cmpFn, colloid, value, minvalue);
376 ECB :
377 : /* smaller than the smallest value in this range */
378 GIC 357 : if (DatumGetBool(compar))
379 ECB : {
380 GIC 135 : end = (midpoint - 1);
381 CBC 135 : continue;
382 ECB : }
383 :
384 : /*
385 : * Is the value greater than the minval? If yes, we'll recurse
386 : * to the right side of range array.
387 : */
388 GIC 222 : compar = FunctionCall2Coll(cmpFn, colloid, maxvalue, value);
389 ECB :
390 : /* larger than the largest value in this range */
391 GIC 222 : if (DatumGetBool(compar))
392 ECB : {
393 GIC 222 : start = (midpoint + 1);
394 CBC 222 : continue;
395 ECB : }
396 :
397 : /* hey, we found a matching range */
398 UIC 0 : Assert(false);
399 EUB : }
400 : }
401 : }
402 :
403 : /* and values in the unsorted part must not be in the sorted part */
404 GIC 94983 : if (ranges->nsorted > 0)
405 ECB : {
406 : compare_context cxt;
407 :
408 GIC 92739 : cxt.colloid = ranges->colloid;
409 CBC 92739 : cxt.cmpFn = ranges->cmp;
410 ECB :
411 GIC 7954167 : for (i = ranges->nsorted; i < ranges->nvalues; i++)
412 ECB : {
413 GIC 7861428 : Datum value = ranges->values[2 * ranges->nranges + i];
414 ECB :
415 GIC 7861428 : Assert(bsearch_arg(&value, &ranges->values[2 * ranges->nranges],
416 ECB : ranges->nsorted, sizeof(Datum),
417 : compare_values, (void *) &cxt) == NULL);
418 : }
419 : }
420 : #endif
421 GIC 94983 : }
422 ECB :
423 : /*
424 : * Check that the expanded ranges (built when reducing the number of ranges
425 : * by combining some of them) are correctly sorted and do not overlap.
426 : */
427 : static void
428 UIC 0 : AssertCheckExpandedRanges(BrinDesc *bdesc, Oid colloid, AttrNumber attno,
429 EUB : Form_pg_attribute attr, ExpandedRange *ranges,
430 : int nranges)
431 : {
432 : #ifdef USE_ASSERT_CHECKING
433 : int i;
434 : FmgrInfo *eq;
435 : FmgrInfo *lt;
436 :
437 UIC 0 : eq = minmax_multi_get_strategy_procinfo(bdesc, attno, attr->atttypid,
438 EUB : BTEqualStrategyNumber);
439 :
440 UIC 0 : lt = minmax_multi_get_strategy_procinfo(bdesc, attno, attr->atttypid,
441 EUB : BTLessStrategyNumber);
442 :
443 : /*
444 : * Each range independently should be valid, i.e. that for the boundary
445 : * values (lower <= upper).
446 : */
447 UIC 0 : for (i = 0; i < nranges; i++)
448 EUB : {
449 : Datum r;
450 UIC 0 : Datum minval = ranges[i].minval;
451 UBC 0 : Datum maxval = ranges[i].maxval;
452 EUB :
453 UIC 0 : if (ranges[i].collapsed) /* collapsed: minval == maxval */
454 UBC 0 : r = FunctionCall2Coll(eq, colloid, minval, maxval);
455 EUB : else /* non-collapsed: minval < maxval */
456 UIC 0 : r = FunctionCall2Coll(lt, colloid, minval, maxval);
457 EUB :
458 UIC 0 : Assert(DatumGetBool(r));
459 EUB : }
460 :
461 : /*
462 : * And the ranges should be ordered and must not overlap, i.e. upper <
463 : * lower for boundaries of consecutive ranges.
464 : */
465 UIC 0 : for (i = 0; i < nranges - 1; i++)
466 EUB : {
467 : Datum r;
468 UIC 0 : Datum maxval = ranges[i].maxval;
469 UBC 0 : Datum minval = ranges[i + 1].minval;
470 EUB :
471 UIC 0 : r = FunctionCall2Coll(lt, colloid, maxval, minval);
472 EUB :
473 UIC 0 : Assert(DatumGetBool(r));
474 EUB : }
475 : #endif
476 UIC 0 : }
477 EUB :
478 :
479 : /*
480 : * minmax_multi_init
481 : * Initialize the deserialized range list, allocate all the memory.
482 : *
483 : * This is only in-memory representation of the ranges, so we allocate
484 : * enough space for the maximum number of values (so as not to have to do
485 : * repallocs as the ranges grow).
486 : */
487 : static Ranges *
488 GIC 23187 : minmax_multi_init(int maxvalues)
489 ECB : {
490 : Size len;
491 : Ranges *ranges;
492 :
493 GIC 23187 : Assert(maxvalues > 0);
494 ECB :
495 GIC 23187 : len = offsetof(Ranges, values); /* fixed header */
496 CBC 23187 : len += maxvalues * sizeof(Datum); /* Datum values */
497 ECB :
498 GIC 23187 : ranges = (Ranges *) palloc0(len);
499 ECB :
500 GIC 23187 : ranges->maxvalues = maxvalues;
501 ECB :
502 GIC 23187 : return ranges;
503 ECB : }
504 :
505 :
506 : /*
507 : * range_deduplicate_values
508 : * Deduplicate the part with values in the simple points.
509 : *
510 : * This is meant to be a cheaper way of reducing the size of the ranges. It
511 : * does not touch the ranges, and only sorts the other values - it does not
512 : * call the distance functions, which may be quite expensive, etc.
513 : *
514 : * We do know the values are not duplicate with the ranges, because we check
515 : * that before adding a new value. Same for the sorted part of values.
516 : */
517 : static void
518 GIC 8658 : range_deduplicate_values(Ranges *range)
519 ECB : {
520 : int i,
521 : n;
522 : int start;
523 : compare_context cxt;
524 :
525 : /*
526 : * If there are no unsorted values, we're done (this probably can't
527 : * happen, as we're adding values to unsorted part).
528 : */
529 GIC 8658 : if (range->nsorted == range->nvalues)
530 CBC 8535 : return;
531 ECB :
532 : /* sort the values */
533 GIC 123 : cxt.colloid = range->colloid;
534 CBC 123 : cxt.cmpFn = range->cmp;
535 ECB :
536 : /* the values start right after the ranges (which are always sorted) */
537 GIC 123 : start = 2 * range->nranges;
538 ECB :
539 : /*
540 : * XXX This might do a merge sort, to leverage that the first part of the
541 : * array is already sorted. If the sorted part is large, it might be quite
542 : * a bit faster.
543 : */
544 GIC 123 : qsort_arg(&range->values[start],
545 CBC 123 : range->nvalues, sizeof(Datum),
546 : compare_values, &cxt);
547 :
548 GIC 123 : n = 1;
549 CBC 39120 : for (i = 1; i < range->nvalues; i++)
550 ECB : {
551 : /* same as preceding value, so store it */
552 GIC 38997 : if (compare_values(&range->values[start + i - 1],
553 CBC 38997 : &range->values[start + i],
554 ECB : (void *) &cxt) == 0)
555 UIC 0 : continue;
556 EUB :
557 GIC 38997 : range->values[start + n] = range->values[start + i];
558 ECB :
559 GIC 38997 : n++;
560 ECB : }
561 :
562 : /* now all the values are sorted */
563 GIC 123 : range->nvalues = n;
564 CBC 123 : range->nsorted = n;
565 ECB :
566 GIC 123 : AssertCheckRanges(range, range->cmp, range->colloid);
567 ECB : }
568 :
569 :
570 : /*
571 : * brin_range_serialize
572 : * Serialize the in-memory representation into a compact varlena value.
573 : *
574 : * Simply copy the header and then also the individual values, as stored
575 : * in the in-memory value array.
576 : */
577 : static SerializedRanges *
578 GIC 8535 : brin_range_serialize(Ranges *range)
579 ECB : {
580 : Size len;
581 : int nvalues;
582 : SerializedRanges *serialized;
583 : Oid typid;
584 : int typlen;
585 : bool typbyval;
586 :
587 : char *ptr;
588 :
589 : /* simple sanity checks */
590 CBC 8535 : Assert(range->nranges >= 0);
591 8535 : Assert(range->nsorted >= 0);
592 8535 : Assert(range->nvalues >= 0);
593 8535 : Assert(range->maxvalues > 0);
594 8535 : Assert(range->target_maxvalues > 0);
595 :
596 : /* at this point the range should be compacted to the target size */
597 8535 : Assert(2 * range->nranges + range->nvalues <= range->target_maxvalues);
598 :
599 8535 : Assert(range->target_maxvalues <= range->maxvalues);
600 :
601 : /* range boundaries are always sorted */
602 8535 : Assert(range->nvalues >= range->nsorted);
603 :
604 : /* deduplicate values, if there's unsorted part */
605 8535 : range_deduplicate_values(range);
606 :
607 : /* see how many Datum values we actually have */
608 8535 : nvalues = 2 * range->nranges + range->nvalues;
609 :
610 8535 : typid = range->typid;
611 8535 : typbyval = get_typbyval(typid);
612 8535 : typlen = get_typlen(typid);
613 :
614 : /* header is always needed */
615 8535 : len = offsetof(SerializedRanges, data);
616 :
617 : /*
618 : * The space needed depends on data type - for fixed-length data types
619 : * (by-value and some by-reference) it's pretty simple, just multiply
620 : * (attlen * nvalues) and we're done. For variable-length by-reference
621 : * types we need to actually walk all the values and sum the lengths.
622 : */
623 8535 : if (typlen == -1) /* varlena */
624 : {
625 : int i;
626 :
627 5928 : for (i = 0; i < nvalues; i++)
628 : {
629 4644 : len += VARSIZE_ANY(range->values[i]);
630 : }
631 : }
632 7251 : else if (typlen == -2) /* cstring */
633 : {
634 : int i;
635 :
636 UBC 0 : for (i = 0; i < nvalues; i++)
637 : {
638 : /* don't forget to include the null terminator ;-) */
639 0 : len += strlen(DatumGetCString(range->values[i])) + 1;
640 : }
641 : }
642 : else /* fixed-length types (even by-reference) */
643 : {
644 CBC 7251 : Assert(typlen > 0);
645 7251 : len += nvalues * typlen;
646 : }
647 :
648 : /*
649 : * Allocate the serialized object, copy the basic information. The
650 : * serialized object is a varlena, so update the header.
651 : */
652 8535 : serialized = (SerializedRanges *) palloc0(len);
653 8535 : SET_VARSIZE(serialized, len);
654 :
655 8535 : serialized->typid = typid;
656 8535 : serialized->nranges = range->nranges;
657 8535 : serialized->nvalues = range->nvalues;
658 8535 : serialized->maxvalues = range->target_maxvalues;
659 :
660 : /*
661 : * And now copy also the boundary values (like the length calculation this
662 : * depends on the particular data type).
663 : */
664 8535 : ptr = serialized->data; /* start of the serialized data */
665 :
666 GNC 37056 : for (int i = 0; i < nvalues; i++)
667 : {
668 CBC 28521 : if (typbyval) /* simple by-value data types */
669 : {
670 : Datum tmp;
671 :
672 : /*
673 : * For byval types, we need to copy just the significant bytes -
674 : * we can't use memcpy directly, as that assumes little-endian
675 : * behavior. store_att_byval does almost what we need, but it
676 : * requires a properly aligned buffer - the output buffer does not
677 : * guarantee that. So we simply use a local Datum variable (which
678 : * guarantees proper alignment), and then copy the value from it.
679 : */
680 15693 : store_att_byval(&tmp, range->values[i], typlen);
681 :
682 15693 : memcpy(ptr, &tmp, typlen);
683 15693 : ptr += typlen;
684 : }
685 12828 : else if (typlen > 0) /* fixed-length by-ref types */
686 : {
687 8184 : memcpy(ptr, DatumGetPointer(range->values[i]), typlen);
688 8184 : ptr += typlen;
689 : }
690 4644 : else if (typlen == -1) /* varlena */
691 : {
692 4644 : int tmp = VARSIZE_ANY(DatumGetPointer(range->values[i]));
693 :
694 4644 : memcpy(ptr, DatumGetPointer(range->values[i]), tmp);
695 4644 : ptr += tmp;
696 : }
697 UBC 0 : else if (typlen == -2) /* cstring */
698 : {
699 0 : int tmp = strlen(DatumGetCString(range->values[i])) + 1;
700 :
701 0 : memcpy(ptr, DatumGetCString(range->values[i]), tmp);
702 0 : ptr += tmp;
703 : }
704 :
705 : /* make sure we haven't overflown the buffer end */
706 CBC 28521 : Assert(ptr <= ((char *) serialized + len));
707 : }
708 :
709 : /* exact size */
710 8535 : Assert(ptr == ((char *) serialized + len));
711 :
712 8535 : return serialized;
713 : }
714 :
715 : /*
716 : * brin_range_deserialize
717 : * Serialize the in-memory representation into a compact varlena value.
718 : *
719 : * Simply copy the header and then also the individual values, as stored
720 : * in the in-memory value array.
721 : */
722 : static Ranges *
723 20961 : brin_range_deserialize(int maxvalues, SerializedRanges *serialized)
724 : {
725 : int i,
726 : nvalues;
727 : char *ptr,
728 : *dataptr;
729 : bool typbyval;
730 : int typlen;
731 : Size datalen;
732 :
733 : Ranges *range;
734 :
735 20961 : Assert(serialized->nranges >= 0);
736 20961 : Assert(serialized->nvalues >= 0);
737 20961 : Assert(serialized->maxvalues > 0);
738 :
739 20961 : nvalues = 2 * serialized->nranges + serialized->nvalues;
740 :
741 20961 : Assert(nvalues <= serialized->maxvalues);
742 20961 : Assert(serialized->maxvalues <= maxvalues);
743 :
744 20961 : range = minmax_multi_init(maxvalues);
745 :
746 : /* copy the header info */
747 20961 : range->nranges = serialized->nranges;
748 20961 : range->nvalues = serialized->nvalues;
749 20961 : range->nsorted = serialized->nvalues;
750 20961 : range->maxvalues = maxvalues;
751 20961 : range->target_maxvalues = serialized->maxvalues;
752 :
753 20961 : range->typid = serialized->typid;
754 :
755 20961 : typbyval = get_typbyval(serialized->typid);
756 20961 : typlen = get_typlen(serialized->typid);
757 :
758 : /*
759 : * And now deconstruct the values into Datum array. We have to copy the
760 : * data because the serialized representation ignores alignment, and we
761 : * don't want to rely on it being kept around anyway.
762 : */
763 20961 : ptr = serialized->data;
764 :
765 : /*
766 : * We don't want to allocate many pieces, so we just allocate everything
767 : * in one chunk. How much space will we need?
768 : *
769 : * XXX We don't need to copy simple by-value data types.
770 : */
771 20961 : datalen = 0;
772 20961 : dataptr = NULL;
773 47481 : for (i = 0; (i < nvalues) && (!typbyval); i++)
774 : {
775 26520 : if (typlen > 0) /* fixed-length by-ref types */
776 14706 : datalen += MAXALIGN(typlen);
777 11814 : else if (typlen == -1) /* varlena */
778 : {
779 GNC 11814 : datalen += MAXALIGN(VARSIZE_ANY(ptr));
780 11814 : ptr += VARSIZE_ANY(ptr);
781 : }
782 UBC 0 : else if (typlen == -2) /* cstring */
783 : {
784 UNC 0 : Size slen = strlen(ptr) + 1;
785 :
786 UBC 0 : datalen += MAXALIGN(slen);
787 0 : ptr += slen;
788 : }
789 : }
790 :
791 CBC 20961 : if (datalen > 0)
792 8184 : dataptr = palloc(datalen);
793 :
794 : /*
795 : * Restore the source pointer (might have been modified when calculating
796 : * the space we need to allocate).
797 : */
798 20961 : ptr = serialized->data;
799 :
800 86412 : for (i = 0; i < nvalues; i++)
801 : {
802 65451 : if (typbyval) /* simple by-value data types */
803 : {
804 38931 : Datum v = 0;
805 :
806 38931 : memcpy(&v, ptr, typlen);
807 :
808 38931 : range->values[i] = fetch_att(&v, true, typlen);
809 38931 : ptr += typlen;
810 : }
811 26520 : else if (typlen > 0) /* fixed-length by-ref types */
812 : {
813 14706 : range->values[i] = PointerGetDatum(dataptr);
814 :
815 14706 : memcpy(dataptr, ptr, typlen);
816 14706 : dataptr += MAXALIGN(typlen);
817 :
818 14706 : ptr += typlen;
819 : }
820 11814 : else if (typlen == -1) /* varlena */
821 : {
822 11814 : range->values[i] = PointerGetDatum(dataptr);
823 :
824 11814 : memcpy(dataptr, ptr, VARSIZE_ANY(ptr));
825 11814 : dataptr += MAXALIGN(VARSIZE_ANY(ptr));
826 11814 : ptr += VARSIZE_ANY(ptr);
827 : }
828 UBC 0 : else if (typlen == -2) /* cstring */
829 : {
830 0 : Size slen = strlen(ptr) + 1;
831 :
832 0 : range->values[i] = PointerGetDatum(dataptr);
833 :
834 0 : memcpy(dataptr, ptr, slen);
835 0 : dataptr += MAXALIGN(slen);
836 0 : ptr += slen;
837 : }
838 :
839 : /* make sure we haven't overflown the buffer end */
840 CBC 65451 : Assert(ptr <= ((char *) serialized + VARSIZE_ANY(serialized)));
841 : }
842 :
843 : /* should have consumed the whole input value exactly */
844 20961 : Assert(ptr == ((char *) serialized + VARSIZE_ANY(serialized)));
845 :
846 : /* return the deserialized value */
847 20961 : return range;
848 : }
849 :
850 : /*
851 : * compare_expanded_ranges
852 : * Compare the expanded ranges - first by minimum, then by maximum.
853 : *
854 : * We do guarantee that ranges in a single Ranges object do not overlap, so it
855 : * may seem strange that we don't order just by minimum. But when merging two
856 : * Ranges (which happens in the union function), the ranges may in fact
857 : * overlap. So we do compare both.
858 : */
859 : static int
860 310020 : compare_expanded_ranges(const void *a, const void *b, void *arg)
861 : {
862 310020 : ExpandedRange *ra = (ExpandedRange *) a;
863 310020 : ExpandedRange *rb = (ExpandedRange *) b;
864 : Datum r;
865 :
866 310020 : compare_context *cxt = (compare_context *) arg;
867 :
868 : /* first compare minvals */
869 310020 : r = FunctionCall2Coll(cxt->cmpFn, cxt->colloid, ra->minval, rb->minval);
870 :
871 310020 : if (DatumGetBool(r))
872 227331 : return -1;
873 :
874 82689 : r = FunctionCall2Coll(cxt->cmpFn, cxt->colloid, rb->minval, ra->minval);
875 :
876 82689 : if (DatumGetBool(r))
877 81315 : return 1;
878 :
879 : /* then compare maxvals */
880 1374 : r = FunctionCall2Coll(cxt->cmpFn, cxt->colloid, ra->maxval, rb->maxval);
881 :
882 1374 : if (DatumGetBool(r))
883 UBC 0 : return -1;
884 :
885 CBC 1374 : r = FunctionCall2Coll(cxt->cmpFn, cxt->colloid, rb->maxval, ra->maxval);
886 :
887 1374 : if (DatumGetBool(r))
888 UBC 0 : return 1;
889 :
890 CBC 1374 : return 0;
891 : }
892 :
893 : /*
894 : * compare_values
895 : * Compare the values.
896 : */
897 : static int
898 43198377 : compare_values(const void *a, const void *b, void *arg)
899 : {
900 43198377 : Datum *da = (Datum *) a;
901 43198377 : Datum *db = (Datum *) b;
902 : Datum r;
903 :
904 43198377 : compare_context *cxt = (compare_context *) arg;
905 :
906 43198377 : r = FunctionCall2Coll(cxt->cmpFn, cxt->colloid, *da, *db);
907 :
908 43198377 : if (DatumGetBool(r))
909 674631 : return -1;
910 :
911 42523746 : r = FunctionCall2Coll(cxt->cmpFn, cxt->colloid, *db, *da);
912 :
913 42523746 : if (DatumGetBool(r))
914 42495945 : return 1;
915 :
916 27801 : return 0;
917 : }
918 :
919 : /*
920 : * Check if the new value matches one of the existing ranges.
921 : */
922 : static bool
923 92193 : has_matching_range(BrinDesc *bdesc, Oid colloid, Ranges *ranges,
924 : Datum newval, AttrNumber attno, Oid typid)
925 : {
926 : Datum compar;
927 :
928 : Datum minvalue;
929 : Datum maxvalue;
930 :
931 : FmgrInfo *cmpLessFn;
932 : FmgrInfo *cmpGreaterFn;
933 :
934 : /* binary search on ranges */
935 : int start,
936 : end;
937 :
938 92193 : if (ranges->nranges == 0)
939 34113 : return false;
940 :
941 58080 : minvalue = ranges->values[0];
942 58080 : maxvalue = ranges->values[2 * ranges->nranges - 1];
943 :
944 : /*
945 : * Otherwise, need to compare the new value with boundaries of all the
946 : * ranges. First check if it's less than the absolute minimum, which is
947 : * the first value in the array.
948 : */
949 58080 : cmpLessFn = minmax_multi_get_strategy_procinfo(bdesc, attno, typid,
950 : BTLessStrategyNumber);
951 58080 : compar = FunctionCall2Coll(cmpLessFn, colloid, newval, minvalue);
952 :
953 : /* smaller than the smallest value in the range list */
954 58080 : if (DatumGetBool(compar))
955 UBC 0 : return false;
956 :
957 : /*
958 : * And now compare it to the existing maximum (last value in the data
959 : * array). But only if we haven't already ruled out a possible match in
960 : * the minvalue check.
961 : */
962 CBC 58080 : cmpGreaterFn = minmax_multi_get_strategy_procinfo(bdesc, attno, typid,
963 : BTGreaterStrategyNumber);
964 58080 : compar = FunctionCall2Coll(cmpGreaterFn, colloid, newval, maxvalue);
965 :
966 58080 : if (DatumGetBool(compar))
967 58080 : return false;
968 :
969 : /*
970 : * So we know it's in the general min/max, the question is whether it
971 : * falls in one of the ranges or gaps. We'll do a binary search on
972 : * individual ranges - for each range we check equality (value falls into
973 : * the range), and then check ranges either above or below the current
974 : * range.
975 : */
976 UBC 0 : start = 0; /* first range */
977 0 : end = (ranges->nranges - 1); /* last range */
978 : while (true)
979 0 : {
980 0 : int midpoint = (start + end) / 2;
981 :
982 : /* this means we ran out of ranges in the last step */
983 0 : if (start > end)
984 0 : return false;
985 :
986 : /* copy the min/max values from the ranges */
987 0 : minvalue = ranges->values[2 * midpoint];
988 0 : maxvalue = ranges->values[2 * midpoint + 1];
989 :
990 : /*
991 : * Is the value smaller than the minval? If yes, we'll recurse to the
992 : * left side of range array.
993 : */
994 0 : compar = FunctionCall2Coll(cmpLessFn, colloid, newval, minvalue);
995 :
996 : /* smaller than the smallest value in this range */
997 0 : if (DatumGetBool(compar))
998 : {
999 0 : end = (midpoint - 1);
1000 0 : continue;
1001 : }
1002 :
1003 : /*
1004 : * Is the value greater than the minval? If yes, we'll recurse to the
1005 : * right side of range array.
1006 : */
1007 0 : compar = FunctionCall2Coll(cmpGreaterFn, colloid, newval, maxvalue);
1008 :
1009 : /* larger than the largest value in this range */
1010 0 : if (DatumGetBool(compar))
1011 : {
1012 0 : start = (midpoint + 1);
1013 0 : continue;
1014 : }
1015 :
1016 : /* hey, we found a matching range */
1017 0 : return true;
1018 : }
1019 :
1020 : return false;
1021 : }
1022 :
1023 :
1024 : /*
1025 : * range_contains_value
1026 : * See if the new value is already contained in the range list.
1027 : *
1028 : * We first inspect the list of intervals. We use a small trick - we check
1029 : * the value against min/max of the whole range (min of the first interval,
1030 : * max of the last one) first, and only inspect the individual intervals if
1031 : * this passes.
1032 : *
1033 : * If the value matches none of the intervals, we check the exact values.
1034 : * We simply loop through them and invoke equality operator on them.
1035 : *
1036 : * The last parameter (full) determines whether we need to search all the
1037 : * values, including the unsorted part. With full=false, the unsorted part
1038 : * is not searched, which may produce false negatives and duplicate values
1039 : * (in the unsorted part only), but when we're building the range that's
1040 : * fine - we'll deduplicate before serialization, and it can only happen
1041 : * if there already are unsorted values (so it was already modified).
1042 : *
1043 : * Serialized ranges don't have any unsorted values, so this can't cause
1044 : * false negatives during querying.
1045 : */
1046 : static bool
1047 CBC 92193 : range_contains_value(BrinDesc *bdesc, Oid colloid,
1048 : AttrNumber attno, Form_pg_attribute attr,
1049 : Ranges *ranges, Datum newval, bool full)
1050 : {
1051 : int i;
1052 : FmgrInfo *cmpEqualFn;
1053 92193 : Oid typid = attr->atttypid;
1054 :
1055 : /*
1056 : * First inspect the ranges, if there are any. We first check the whole
1057 : * range, and only when there's still a chance of getting a match we
1058 : * inspect the individual ranges.
1059 : */
1060 92193 : if (has_matching_range(bdesc, colloid, ranges, newval, attno, typid))
1061 UBC 0 : return true;
1062 :
1063 CBC 92193 : cmpEqualFn = minmax_multi_get_strategy_procinfo(bdesc, attno, typid,
1064 : BTEqualStrategyNumber);
1065 :
1066 : /*
1067 : * There is no matching range, so let's inspect the sorted values.
1068 : *
1069 : * We do a sequential search for small numbers of values, and binary
1070 : * search once we have more than 16 values. This threshold is somewhat
1071 : * arbitrary, as it depends on how expensive the comparison function is.
1072 : *
1073 : * XXX If we use the threshold here, maybe we should do the same thing in
1074 : * has_matching_range? Or maybe we should do the bin search all the time?
1075 : *
1076 : * XXX We could use the same optimization as for ranges, to check if the
1077 : * value is between min/max, to maybe rule out all sorted values without
1078 : * having to inspect all of them.
1079 : */
1080 92193 : if (ranges->nsorted >= 16)
1081 : {
1082 : compare_context cxt;
1083 :
1084 58080 : cxt.colloid = ranges->colloid;
1085 58080 : cxt.cmpFn = ranges->cmp;
1086 :
1087 58080 : if (bsearch_arg(&newval, &ranges->values[2 * ranges->nranges],
1088 58080 : ranges->nsorted, sizeof(Datum),
1089 : compare_values, (void *) &cxt) != NULL)
1090 UBC 0 : return true;
1091 : }
1092 : else
1093 : {
1094 CBC 65706 : for (i = 2 * ranges->nranges; i < 2 * ranges->nranges + ranges->nsorted; i++)
1095 : {
1096 : Datum compar;
1097 :
1098 39834 : compar = FunctionCall2Coll(cmpEqualFn, colloid, newval, ranges->values[i]);
1099 :
1100 : /* found an exact match */
1101 39834 : if (DatumGetBool(compar))
1102 8241 : return true;
1103 : }
1104 : }
1105 :
1106 : /* If not asked to inspect the unsorted part, we're done. */
1107 83952 : if (!full)
1108 43089 : return false;
1109 :
1110 : /* Inspect the unsorted part. */
1111 3909768 : for (i = 2 * ranges->nranges + ranges->nsorted; i < 2 * ranges->nranges + ranges->nvalues; i++)
1112 : {
1113 : Datum compar;
1114 :
1115 3909768 : compar = FunctionCall2Coll(cmpEqualFn, colloid, newval, ranges->values[i]);
1116 :
1117 : /* found an exact match */
1118 3909768 : if (DatumGetBool(compar))
1119 40863 : return true;
1120 : }
1121 :
1122 : /* the value is not covered by this BRIN tuple */
1123 UBC 0 : return false;
1124 : }
1125 :
1126 : /*
1127 : * Expand ranges from Ranges into ExpandedRange array. This expects the
1128 : * eranges to be pre-allocated and with the correct size - there needs to be
1129 : * (nranges + nvalues) elements.
1130 : *
1131 : * The order of expanded ranges is arbitrary. We do expand the ranges first,
1132 : * and this part is sorted. But then we expand the values, and this part may
1133 : * be unsorted.
1134 : */
1135 : static void
1136 CBC 2667 : fill_expanded_ranges(ExpandedRange *eranges, int neranges, Ranges *ranges)
1137 : {
1138 : int idx;
1139 : int i;
1140 :
1141 : /* Check that the output array has the right size. */
1142 2667 : Assert(neranges == (ranges->nranges + ranges->nvalues));
1143 :
1144 2667 : idx = 0;
1145 2790 : for (i = 0; i < ranges->nranges; i++)
1146 : {
1147 123 : eranges[idx].minval = ranges->values[2 * i];
1148 123 : eranges[idx].maxval = ranges->values[2 * i + 1];
1149 123 : eranges[idx].collapsed = false;
1150 123 : idx++;
1151 :
1152 123 : Assert(idx <= neranges);
1153 : }
1154 :
1155 56757 : for (i = 0; i < ranges->nvalues; i++)
1156 : {
1157 54090 : eranges[idx].minval = ranges->values[2 * ranges->nranges + i];
1158 54090 : eranges[idx].maxval = ranges->values[2 * ranges->nranges + i];
1159 54090 : eranges[idx].collapsed = true;
1160 54090 : idx++;
1161 :
1162 54090 : Assert(idx <= neranges);
1163 : }
1164 :
1165 : /* Did we produce the expected number of elements? */
1166 2667 : Assert(idx == neranges);
1167 :
1168 2667 : return;
1169 : }
1170 :
1171 : /*
1172 : * Sort and deduplicate expanded ranges.
1173 : *
1174 : * The ranges may be deduplicated - we're simply appending values, without
1175 : * checking for duplicates etc. So maybe the deduplication will reduce the
1176 : * number of ranges enough, and we won't have to compute the distances etc.
1177 : *
1178 : * Returns the number of expanded ranges.
1179 : */
1180 : static int
1181 2667 : sort_expanded_ranges(FmgrInfo *cmp, Oid colloid,
1182 : ExpandedRange *eranges, int neranges)
1183 : {
1184 : int n;
1185 : int i;
1186 : compare_context cxt;
1187 :
1188 2667 : Assert(neranges > 0);
1189 :
1190 : /* sort the values */
1191 2667 : cxt.colloid = colloid;
1192 2667 : cxt.cmpFn = cmp;
1193 :
1194 : /*
1195 : * XXX We do qsort on all the values, but we could also leverage the fact
1196 : * that some of the input data is already sorted (all the ranges and maybe
1197 : * some of the points) and do merge sort.
1198 : */
1199 2667 : qsort_arg(eranges, neranges, sizeof(ExpandedRange),
1200 : compare_expanded_ranges, &cxt);
1201 :
1202 : /*
1203 : * Deduplicate the ranges - simply compare each range to the preceding
1204 : * one, and skip the duplicate ones.
1205 : */
1206 2667 : n = 1;
1207 54213 : for (i = 1; i < neranges; i++)
1208 : {
1209 : /* if the current range is equal to the preceding one, do nothing */
1210 51546 : if (!compare_expanded_ranges(&eranges[i - 1], &eranges[i], (void *) &cxt))
1211 609 : continue;
1212 :
1213 : /* otherwise, copy it to n-th place (if not already there) */
1214 50937 : if (i != n)
1215 1464 : memcpy(&eranges[n], &eranges[i], sizeof(ExpandedRange));
1216 :
1217 50937 : n++;
1218 : }
1219 :
1220 2667 : Assert((n > 0) && (n <= neranges));
1221 :
1222 2667 : return n;
1223 : }
1224 :
1225 : /*
1226 : * When combining multiple Range values (in union function), some of the
1227 : * ranges may overlap. We simply merge the overlapping ranges to fix that.
1228 : *
1229 : * XXX This assumes the expanded ranges were previously sorted (by minval
1230 : * and then maxval). We leverage this when detecting overlap.
1231 : */
1232 : static int
1233 UBC 0 : merge_overlapping_ranges(FmgrInfo *cmp, Oid colloid,
1234 : ExpandedRange *eranges, int neranges)
1235 : {
1236 : int idx;
1237 :
1238 : /* Merge ranges (idx) and (idx+1) if they overlap. */
1239 0 : idx = 0;
1240 0 : while (idx < (neranges - 1))
1241 : {
1242 : Datum r;
1243 :
1244 : /*
1245 : * comparing [?,maxval] vs. [minval,?] - the ranges overlap if (minval
1246 : * < maxval)
1247 : */
1248 0 : r = FunctionCall2Coll(cmp, colloid,
1249 0 : eranges[idx].maxval,
1250 0 : eranges[idx + 1].minval);
1251 :
1252 : /*
1253 : * Nope, maxval < minval, so no overlap. And we know the ranges are
1254 : * ordered, so there are no more overlaps, because all the remaining
1255 : * ranges have greater or equal minval.
1256 : */
1257 0 : if (DatumGetBool(r))
1258 : {
1259 : /* proceed to the next range */
1260 0 : idx += 1;
1261 0 : continue;
1262 : }
1263 :
1264 : /*
1265 : * So ranges 'idx' and 'idx+1' do overlap, but we don't know if
1266 : * 'idx+1' is contained in 'idx', or if they overlap only partially.
1267 : * So compare the upper bounds and keep the larger one.
1268 : */
1269 0 : r = FunctionCall2Coll(cmp, colloid,
1270 0 : eranges[idx].maxval,
1271 0 : eranges[idx + 1].maxval);
1272 :
1273 0 : if (DatumGetBool(r))
1274 0 : eranges[idx].maxval = eranges[idx + 1].maxval;
1275 :
1276 : /*
1277 : * The range certainly is no longer collapsed (irrespectively of the
1278 : * previous state).
1279 : */
1280 0 : eranges[idx].collapsed = false;
1281 :
1282 : /*
1283 : * Now get rid of the (idx+1) range entirely by shifting the remaining
1284 : * ranges by 1. There are neranges elements, and we need to move
1285 : * elements from (idx+2). That means the number of elements to move is
1286 : * [ncranges - (idx+2)].
1287 : */
1288 0 : memmove(&eranges[idx + 1], &eranges[idx + 2],
1289 0 : (neranges - (idx + 2)) * sizeof(ExpandedRange));
1290 :
1291 : /*
1292 : * Decrease the number of ranges, and repeat (with the same range, as
1293 : * it might overlap with additional ranges thanks to the merge).
1294 : */
1295 0 : neranges--;
1296 : }
1297 :
1298 0 : return neranges;
1299 : }
1300 :
1301 : /*
1302 : * Simple comparator for distance values, comparing the double value.
1303 : * This is intentionally sorting the distances in descending order, i.e.
1304 : * the longer gaps will be at the front.
1305 : */
1306 : static int
1307 CBC 72960 : compare_distances(const void *a, const void *b)
1308 : {
1309 72960 : DistanceValue *da = (DistanceValue *) a;
1310 72960 : DistanceValue *db = (DistanceValue *) b;
1311 :
1312 72960 : if (da->value < db->value)
1313 18168 : return 1;
1314 54792 : else if (da->value > db->value)
1315 12696 : return -1;
1316 :
1317 42096 : return 0;
1318 : }
1319 :
1320 : /*
1321 : * Given an array of expanded ranges, compute size of the gaps between each
1322 : * range. For neranges there are (neranges-1) gaps.
1323 : *
1324 : * We simply call the "distance" function to compute the (max-min) for pairs
1325 : * of consecutive ranges. The function may be fairly expensive, so we do that
1326 : * just once (and then use it to pick as many ranges to merge as possible).
1327 : *
1328 : * See reduce_expanded_ranges for details.
1329 : */
1330 : static DistanceValue *
1331 2667 : build_distances(FmgrInfo *distanceFn, Oid colloid,
1332 : ExpandedRange *eranges, int neranges)
1333 : {
1334 : int i;
1335 : int ndistances;
1336 : DistanceValue *distances;
1337 :
1338 2667 : Assert(neranges > 0);
1339 :
1340 : /* If there's only a single range, there's no distance to calculate. */
1341 2667 : if (neranges == 1)
1342 UBC 0 : return NULL;
1343 :
1344 CBC 2667 : ndistances = (neranges - 1);
1345 2667 : distances = (DistanceValue *) palloc0(sizeof(DistanceValue) * ndistances);
1346 :
1347 : /*
1348 : * Walk through the ranges once and compute the distance between the
1349 : * ranges so that we can sort them once.
1350 : */
1351 53604 : for (i = 0; i < ndistances; i++)
1352 : {
1353 : Datum a1,
1354 : a2,
1355 : r;
1356 :
1357 50937 : a1 = eranges[i].maxval;
1358 50937 : a2 = eranges[i + 1].minval;
1359 :
1360 : /* compute length of the gap (between max/min) */
1361 50937 : r = FunctionCall2Coll(distanceFn, colloid, a1, a2);
1362 :
1363 : /* remember the index of the gap the distance is for */
1364 50937 : distances[i].index = i;
1365 50937 : distances[i].value = DatumGetFloat8(r);
1366 : }
1367 :
1368 : /*
1369 : * Sort the distances in descending order, so that the longest gaps are at
1370 : * the front.
1371 : */
1372 2667 : pg_qsort(distances, ndistances, sizeof(DistanceValue), compare_distances);
1373 :
1374 2667 : return distances;
1375 : }
1376 :
1377 : /*
1378 : * Builds expanded ranges for the existing ranges (and single-point ranges),
1379 : * and also the new value (which did not fit into the array). This expanded
1380 : * representation makes the processing a bit easier, as it allows handling
1381 : * ranges and points the same way.
1382 : *
1383 : * We sort and deduplicate the expanded ranges - this is necessary, because
1384 : * the points may be unsorted. And moreover the two parts (ranges and
1385 : * points) are sorted on their own.
1386 : */
1387 : static ExpandedRange *
1388 2667 : build_expanded_ranges(FmgrInfo *cmp, Oid colloid, Ranges *ranges,
1389 : int *nranges)
1390 : {
1391 : int neranges;
1392 : ExpandedRange *eranges;
1393 :
1394 : /* both ranges and points are expanded into a separate element */
1395 2667 : neranges = ranges->nranges + ranges->nvalues;
1396 :
1397 2667 : eranges = (ExpandedRange *) palloc0(neranges * sizeof(ExpandedRange));
1398 :
1399 : /* fill the expanded ranges */
1400 2667 : fill_expanded_ranges(eranges, neranges, ranges);
1401 :
1402 : /* sort and deduplicate the expanded ranges */
1403 2667 : neranges = sort_expanded_ranges(cmp, colloid, eranges, neranges);
1404 :
1405 : /* remember how many ranges we built */
1406 2667 : *nranges = neranges;
1407 :
1408 2667 : return eranges;
1409 : }
1410 :
1411 : #ifdef USE_ASSERT_CHECKING
1412 : /*
1413 : * Counts boundary values needed to store the ranges. Each single-point
1414 : * range is stored using a single value, each regular range needs two.
1415 : */
1416 : static int
1417 5334 : count_values(ExpandedRange *cranges, int ncranges)
1418 : {
1419 : int i;
1420 : int count;
1421 :
1422 5334 : count = 0;
1423 43500 : for (i = 0; i < ncranges; i++)
1424 : {
1425 38166 : if (cranges[i].collapsed)
1426 36348 : count += 1;
1427 : else
1428 1818 : count += 2;
1429 : }
1430 :
1431 5334 : return count;
1432 : }
1433 : #endif
1434 :
1435 : /*
1436 : * reduce_expanded_ranges
1437 : * reduce the ranges until the number of values is low enough
1438 : *
1439 : * Combines ranges until the number of boundary values drops below the
1440 : * threshold specified by max_values. This happens by merging enough
1441 : * ranges by the distance between them.
1442 : *
1443 : * Returns the number of result ranges.
1444 : *
1445 : * We simply use the global min/max and then add boundaries for enough
1446 : * largest gaps. Each gap adds 2 values, so we simply use (target/2-1)
1447 : * distances. Then we simply sort all the values - each two values are
1448 : * a boundary of a range (possibly collapsed).
1449 : *
1450 : * XXX Some of the ranges may be collapsed (i.e. the min/max values are
1451 : * equal), but we ignore that for now. We could repeat the process,
1452 : * adding a couple more gaps recursively.
1453 : *
1454 : * XXX The ranges to merge are selected solely using the distance. But
1455 : * that may not be the best strategy, for example when multiple gaps
1456 : * are of equal (or very similar) length.
1457 : *
1458 : * Consider for example points 1, 2, 3, .., 64, which have gaps of the
1459 : * same length 1 of course. In that case, we tend to pick the first
1460 : * gap of that length, which leads to this:
1461 : *
1462 : * step 1: [1, 2], 3, 4, 5, .., 64
1463 : * step 2: [1, 3], 4, 5, .., 64
1464 : * step 3: [1, 4], 5, .., 64
1465 : * ...
1466 : *
1467 : * So in the end we'll have one "large" range and multiple small points.
1468 : * That may be fine, but it seems a bit strange and non-optimal. Maybe
1469 : * we should consider other things when picking ranges to merge - e.g.
1470 : * length of the ranges? Or perhaps randomize the choice of ranges, with
1471 : * probability inversely proportional to the distance (the gap lengths
1472 : * may be very close, but not exactly the same).
1473 : *
1474 : * XXX Or maybe we could just handle this by using random value as a
1475 : * tie-break, or by adding random noise to the actual distance.
1476 : */
1477 : static int
1478 2667 : reduce_expanded_ranges(ExpandedRange *eranges, int neranges,
1479 : DistanceValue *distances, int max_values,
1480 : FmgrInfo *cmp, Oid colloid)
1481 : {
1482 : int i;
1483 : int nvalues;
1484 : Datum *values;
1485 :
1486 : compare_context cxt;
1487 :
1488 : /* total number of gaps between ranges */
1489 2667 : int ndistances = (neranges - 1);
1490 :
1491 : /* number of gaps to keep */
1492 2667 : int keep = (max_values / 2 - 1);
1493 :
1494 : /*
1495 : * Maybe we have a sufficiently low number of ranges already?
1496 : *
1497 : * XXX This should happen before we actually do the expensive stuff like
1498 : * sorting, so maybe this should be just an assert.
1499 : */
1500 2667 : if (keep >= ndistances)
1501 2481 : return neranges;
1502 :
1503 : /* sort the values */
1504 186 : cxt.colloid = colloid;
1505 186 : cxt.cmpFn = cmp;
1506 :
1507 : /* allocate space for the boundary values */
1508 186 : nvalues = 0;
1509 186 : values = (Datum *) palloc(sizeof(Datum) * max_values);
1510 :
1511 : /* add the global min/max values, from the first/last range */
1512 186 : values[nvalues++] = eranges[0].minval;
1513 186 : values[nvalues++] = eranges[neranges - 1].maxval;
1514 :
1515 : /* add boundary values for enough gaps */
1516 10848 : for (i = 0; i < keep; i++)
1517 : {
1518 : /* index of the gap between (index) and (index+1) ranges */
1519 10662 : int index = distances[i].index;
1520 :
1521 10662 : Assert((index >= 0) && ((index + 1) < neranges));
1522 :
1523 : /* add max from the preceding range, minval from the next one */
1524 10662 : values[nvalues++] = eranges[index].maxval;
1525 10662 : values[nvalues++] = eranges[index + 1].minval;
1526 :
1527 10662 : Assert(nvalues <= max_values);
1528 : }
1529 :
1530 : /* We should have an even number of range values. */
1531 186 : Assert(nvalues % 2 == 0);
1532 :
1533 : /*
1534 : * Sort the values using the comparator function, and form ranges from the
1535 : * sorted result.
1536 : */
1537 186 : qsort_arg(values, nvalues, sizeof(Datum),
1538 : compare_values, &cxt);
1539 :
1540 : /* We have nvalues boundary values, which means nvalues/2 ranges. */
1541 11034 : for (i = 0; i < (nvalues / 2); i++)
1542 : {
1543 10848 : eranges[i].minval = values[2 * i];
1544 10848 : eranges[i].maxval = values[2 * i + 1];
1545 :
1546 : /* if the boundary values are the same, it's a collapsed range */
1547 21696 : eranges[i].collapsed = (compare_values(&values[2 * i],
1548 10848 : &values[2 * i + 1],
1549 10848 : &cxt) == 0);
1550 : }
1551 :
1552 186 : return (nvalues / 2);
1553 : }
1554 :
1555 : /*
1556 : * Store the boundary values from ExpandedRanges back into 'ranges' (using
1557 : * only the minimal number of values needed).
1558 : */
1559 : static void
1560 2667 : store_expanded_ranges(Ranges *ranges, ExpandedRange *eranges, int neranges)
1561 : {
1562 : int i;
1563 2667 : int idx = 0;
1564 :
1565 : /* first copy in the regular ranges */
1566 2667 : ranges->nranges = 0;
1567 21750 : for (i = 0; i < neranges; i++)
1568 : {
1569 19083 : if (!eranges[i].collapsed)
1570 : {
1571 909 : ranges->values[idx++] = eranges[i].minval;
1572 909 : ranges->values[idx++] = eranges[i].maxval;
1573 909 : ranges->nranges++;
1574 : }
1575 : }
1576 :
1577 : /* now copy in the collapsed ones */
1578 2667 : ranges->nvalues = 0;
1579 21750 : for (i = 0; i < neranges; i++)
1580 : {
1581 19083 : if (eranges[i].collapsed)
1582 : {
1583 18174 : ranges->values[idx++] = eranges[i].minval;
1584 18174 : ranges->nvalues++;
1585 : }
1586 : }
1587 :
1588 : /* all the values are sorted */
1589 2667 : ranges->nsorted = ranges->nvalues;
1590 :
1591 2667 : Assert(count_values(eranges, neranges) == 2 * ranges->nranges + ranges->nvalues);
1592 2667 : Assert(2 * ranges->nranges + ranges->nvalues <= ranges->maxvalues);
1593 2667 : }
1594 :
1595 :
1596 : /*
1597 : * Consider freeing space in the ranges. Checks if there's space for at least
1598 : * one new value, and performs compaction if needed.
1599 : *
1600 : * Returns true if the value was actually modified.
1601 : */
1602 : static bool
1603 49104 : ensure_free_space_in_buffer(BrinDesc *bdesc, Oid colloid,
1604 : AttrNumber attno, Form_pg_attribute attr,
1605 : Ranges *range)
1606 : {
1607 : MemoryContext ctx;
1608 : MemoryContext oldctx;
1609 :
1610 : FmgrInfo *cmpFn,
1611 : *distanceFn;
1612 :
1613 : /* expanded ranges */
1614 : ExpandedRange *eranges;
1615 : int neranges;
1616 : DistanceValue *distances;
1617 :
1618 : /*
1619 : * If there is free space in the buffer, we're done without having to
1620 : * modify anything.
1621 : */
1622 49104 : if (2 * range->nranges + range->nvalues < range->maxvalues)
1623 48981 : return false;
1624 :
1625 : /* we'll certainly need the comparator, so just look it up now */
1626 123 : cmpFn = minmax_multi_get_strategy_procinfo(bdesc, attno, attr->atttypid,
1627 : BTLessStrategyNumber);
1628 :
1629 : /* deduplicate values, if there's an unsorted part */
1630 123 : range_deduplicate_values(range);
1631 :
1632 : /*
1633 : * Did we reduce enough free space by just the deduplication?
1634 : *
1635 : * We don't simply check against range->maxvalues again. The deduplication
1636 : * might have freed very little space (e.g. just one value), forcing us to
1637 : * do deduplication very often. In that case, it's better to do the
1638 : * compaction and reduce more space.
1639 : */
1640 123 : if (2 * range->nranges + range->nvalues <= range->maxvalues * MINMAX_BUFFER_LOAD_FACTOR)
1641 UBC 0 : return true;
1642 :
1643 : /*
1644 : * We need to combine some of the existing ranges, to reduce the number of
1645 : * values we have to store.
1646 : *
1647 : * The distanceFn calls (which may internally call e.g. numeric_le) may
1648 : * allocate quite a bit of memory, and we must not leak it (we might have
1649 : * to do this repeatedly, even for a single BRIN page range). Otherwise
1650 : * we'd have problems e.g. when building new indexes. So we use a memory
1651 : * context and make sure we free the memory at the end (so if we call the
1652 : * distance function many times, it might be an issue, but meh).
1653 : */
1654 CBC 123 : ctx = AllocSetContextCreate(CurrentMemoryContext,
1655 : "minmax-multi context",
1656 : ALLOCSET_DEFAULT_SIZES);
1657 :
1658 123 : oldctx = MemoryContextSwitchTo(ctx);
1659 :
1660 : /* build the expanded ranges */
1661 123 : eranges = build_expanded_ranges(cmpFn, colloid, range, &neranges);
1662 :
1663 : /* and we'll also need the 'distance' procedure */
1664 123 : distanceFn = minmax_multi_get_procinfo(bdesc, attno, PROCNUM_DISTANCE);
1665 :
1666 : /* build array of gap distances and sort them in ascending order */
1667 123 : distances = build_distances(distanceFn, colloid, eranges, neranges);
1668 :
1669 : /*
1670 : * Combine ranges until we release at least 50% of the space. This
1671 : * threshold is somewhat arbitrary, perhaps needs tuning. We must not use
1672 : * too low or high value.
1673 : */
1674 246 : neranges = reduce_expanded_ranges(eranges, neranges, distances,
1675 123 : range->maxvalues * MINMAX_BUFFER_LOAD_FACTOR,
1676 : cmpFn, colloid);
1677 :
1678 : /* Make sure we've sufficiently reduced the number of ranges. */
1679 123 : Assert(count_values(eranges, neranges) <= range->maxvalues * MINMAX_BUFFER_LOAD_FACTOR);
1680 :
1681 : /* decompose the expanded ranges into regular ranges and single values */
1682 123 : store_expanded_ranges(range, eranges, neranges);
1683 :
1684 123 : MemoryContextSwitchTo(oldctx);
1685 123 : MemoryContextDelete(ctx);
1686 :
1687 : /* Did we break the ranges somehow? */
1688 123 : AssertCheckRanges(range, cmpFn, colloid);
1689 :
1690 123 : return true;
1691 : }
1692 :
1693 : /*
1694 : * range_add_value
1695 : * Add the new value to the minmax-multi range.
1696 : */
1697 : static bool
1698 49104 : range_add_value(BrinDesc *bdesc, Oid colloid,
1699 : AttrNumber attno, Form_pg_attribute attr,
1700 : Ranges *ranges, Datum newval)
1701 : {
1702 : FmgrInfo *cmpFn;
1703 49104 : bool modified = false;
1704 :
1705 : /* we'll certainly need the comparator, so just look it up now */
1706 49104 : cmpFn = minmax_multi_get_strategy_procinfo(bdesc, attno, attr->atttypid,
1707 : BTLessStrategyNumber);
1708 :
1709 : /* comprehensive checks of the input ranges */
1710 49104 : AssertCheckRanges(ranges, cmpFn, colloid);
1711 :
1712 : /*
1713 : * Make sure there's enough free space in the buffer. We only trigger this
1714 : * when the buffer is full, which means it had to be modified as we size
1715 : * it to be larger than what is stored on disk.
1716 : *
1717 : * This needs to happen before we check if the value is contained in the
1718 : * range, because the value might be in the unsorted part, and we don't
1719 : * check that in range_contains_value. The deduplication would then move
1720 : * it to the sorted part, and we'd add the value too, which violates the
1721 : * rule that we never have duplicates with the ranges or sorted values.
1722 : *
1723 : * We might also deduplicate and recheck if the value is contained, but
1724 : * that seems like overkill. We'd need to deduplicate anyway, so why not
1725 : * do it now.
1726 : */
1727 49104 : modified = ensure_free_space_in_buffer(bdesc, colloid,
1728 : attno, attr, ranges);
1729 :
1730 : /*
1731 : * Bail out if the value already is covered by the range.
1732 : *
1733 : * We could also add values until we hit values_per_range, and then do the
1734 : * deduplication in a batch, hoping for better efficiency. But that would
1735 : * mean we actually modify the range every time, which means having to
1736 : * serialize the value, which does palloc, walks the values, copies them,
1737 : * etc. Not exactly cheap.
1738 : *
1739 : * So instead we do the check, which should be fairly cheap - assuming the
1740 : * comparator function is not very expensive.
1741 : *
1742 : * This also implies the values array can't contain duplicate values.
1743 : */
1744 49104 : if (range_contains_value(bdesc, colloid, attno, attr, ranges, newval, false))
1745 6015 : return modified;
1746 :
1747 : /* Make a copy of the value, if needed. */
1748 43089 : newval = datumCopy(newval, attr->attbyval, attr->attlen);
1749 :
1750 : /*
1751 : * If there's space in the values array, copy it in and we're done.
1752 : *
1753 : * We do want to keep the values sorted (to speed up searches), so we do a
1754 : * simple insertion sort. We could do something more elaborate, e.g. by
1755 : * sorting the values only now and then, but for small counts (e.g. when
1756 : * maxvalues is 64) this should be fine.
1757 : */
1758 43089 : ranges->values[2 * ranges->nranges + ranges->nvalues] = newval;
1759 43089 : ranges->nvalues++;
1760 :
1761 : /* If we added the first value, we can consider it as sorted. */
1762 43089 : if (ranges->nvalues == 1)
1763 2226 : ranges->nsorted = 1;
1764 :
1765 : /*
1766 : * Check we haven't broken the ordering of boundary values (checks both
1767 : * parts, but that doesn't hurt).
1768 : */
1769 43089 : AssertCheckRanges(ranges, cmpFn, colloid);
1770 :
1771 : /* Check the range contains the value we just added. */
1772 43089 : Assert(range_contains_value(bdesc, colloid, attno, attr, ranges, newval, true));
1773 :
1774 : /* yep, we've modified the range */
1775 43089 : return true;
1776 : }
1777 :
1778 : /*
1779 : * Generate range representation of data collected during "batch mode".
1780 : * This is similar to reduce_expanded_ranges, except that we can't assume
1781 : * the values are sorted and there may be duplicate values.
1782 : */
1783 : static void
1784 8535 : compactify_ranges(BrinDesc *bdesc, Ranges *ranges, int max_values)
1785 : {
1786 : FmgrInfo *cmpFn,
1787 : *distanceFn;
1788 :
1789 : /* expanded ranges */
1790 : ExpandedRange *eranges;
1791 : int neranges;
1792 : DistanceValue *distances;
1793 :
1794 : MemoryContext ctx;
1795 : MemoryContext oldctx;
1796 :
1797 : /*
1798 : * Do we need to actually compactify anything?
1799 : *
1800 : * There are two reasons why compaction may be needed - firstly, there may
1801 : * be too many values, or some of the values may be unsorted.
1802 : */
1803 8535 : if ((ranges->nranges * 2 + ranges->nvalues <= max_values) &&
1804 8472 : (ranges->nsorted == ranges->nvalues))
1805 5991 : return;
1806 :
1807 : /* we'll certainly need the comparator, so just look it up now */
1808 2544 : cmpFn = minmax_multi_get_strategy_procinfo(bdesc, ranges->attno, ranges->typid,
1809 : BTLessStrategyNumber);
1810 :
1811 : /* and we'll also need the 'distance' procedure */
1812 2544 : distanceFn = minmax_multi_get_procinfo(bdesc, ranges->attno, PROCNUM_DISTANCE);
1813 :
1814 : /*
1815 : * The distanceFn calls (which may internally call e.g. numeric_le) may
1816 : * allocate quite a bit of memory, and we must not leak it. Otherwise,
1817 : * we'd have problems e.g. when building indexes. So we create a local
1818 : * memory context and make sure we free the memory before leaving this
1819 : * function (not after every call).
1820 : */
1821 2544 : ctx = AllocSetContextCreate(CurrentMemoryContext,
1822 : "minmax-multi context",
1823 : ALLOCSET_DEFAULT_SIZES);
1824 :
1825 2544 : oldctx = MemoryContextSwitchTo(ctx);
1826 :
1827 : /* build the expanded ranges */
1828 2544 : eranges = build_expanded_ranges(cmpFn, ranges->colloid, ranges, &neranges);
1829 :
1830 : /* build array of gap distances and sort them in ascending order */
1831 2544 : distances = build_distances(distanceFn, ranges->colloid,
1832 : eranges, neranges);
1833 :
1834 : /*
1835 : * Combine ranges until we get below max_values. We don't use any scale
1836 : * factor, because this is used during serialization, and we don't expect
1837 : * more tuples to be inserted anytime soon.
1838 : */
1839 2544 : neranges = reduce_expanded_ranges(eranges, neranges, distances,
1840 : max_values, cmpFn, ranges->colloid);
1841 :
1842 2544 : Assert(count_values(eranges, neranges) <= max_values);
1843 :
1844 : /* transform back into regular ranges and single values */
1845 2544 : store_expanded_ranges(ranges, eranges, neranges);
1846 :
1847 : /* check all the range invariants */
1848 2544 : AssertCheckRanges(ranges, cmpFn, ranges->colloid);
1849 :
1850 2544 : MemoryContextSwitchTo(oldctx);
1851 2544 : MemoryContextDelete(ctx);
1852 : }
1853 :
1854 : Datum
1855 9392 : brin_minmax_multi_opcinfo(PG_FUNCTION_ARGS)
1856 : {
1857 : BrinOpcInfo *result;
1858 :
1859 : /*
1860 : * opaque->strategy_procinfos is initialized lazily; here it is set to
1861 : * all-uninitialized by palloc0 which sets fn_oid to InvalidOid.
1862 : */
1863 :
1864 9392 : result = palloc0(MAXALIGN(SizeofBrinOpcInfo(1)) +
1865 : sizeof(MinmaxMultiOpaque));
1866 9392 : result->oi_nstored = 1;
1867 9392 : result->oi_regular_nulls = true;
1868 9392 : result->oi_opaque = (MinmaxMultiOpaque *)
1869 9392 : MAXALIGN((char *) result + SizeofBrinOpcInfo(1));
1870 9392 : result->oi_typcache[0] = lookup_type_cache(PG_BRIN_MINMAX_MULTI_SUMMARYOID, 0);
1871 :
1872 9392 : PG_RETURN_POINTER(result);
1873 : }
1874 :
1875 : /*
1876 : * Compute the distance between two float4 values (plain subtraction).
1877 : */
1878 : Datum
1879 348 : brin_minmax_multi_distance_float4(PG_FUNCTION_ARGS)
1880 : {
1881 348 : float a1 = PG_GETARG_FLOAT4(0);
1882 348 : float a2 = PG_GETARG_FLOAT4(1);
1883 :
1884 : /* if both values are NaN, then we consider them the same */
1885 348 : if (isnan(a1) && isnan(a2))
1886 UBC 0 : PG_RETURN_FLOAT8(0.0);
1887 :
1888 : /* if one value is NaN, use infinite distance */
1889 CBC 348 : if (isnan(a1) || isnan(a2))
1890 3 : PG_RETURN_FLOAT8(get_float8_infinity());
1891 :
1892 : /*
1893 : * We know the values are range boundaries, but the range may be collapsed
1894 : * (i.e. single points), with equal values.
1895 : */
1896 345 : Assert(a1 <= a2);
1897 :
1898 345 : PG_RETURN_FLOAT8((double) a2 - (double) a1);
1899 : }
1900 :
1901 : /*
1902 : * Compute the distance between two float8 values (plain subtraction).
1903 : */
1904 : Datum
1905 522 : brin_minmax_multi_distance_float8(PG_FUNCTION_ARGS)
1906 : {
1907 522 : double a1 = PG_GETARG_FLOAT8(0);
1908 522 : double a2 = PG_GETARG_FLOAT8(1);
1909 :
1910 : /* if both values are NaN, then we consider them the same */
1911 522 : if (isnan(a1) && isnan(a2))
1912 UBC 0 : PG_RETURN_FLOAT8(0.0);
1913 :
1914 : /* if one value is NaN, use infinite distance */
1915 CBC 522 : if (isnan(a1) || isnan(a2))
1916 3 : PG_RETURN_FLOAT8(get_float8_infinity());
1917 :
1918 : /*
1919 : * We know the values are range boundaries, but the range may be collapsed
1920 : * (i.e. single points), with equal values.
1921 : */
1922 519 : Assert(a1 <= a2);
1923 :
1924 519 : PG_RETURN_FLOAT8(a2 - a1);
1925 : }
1926 :
1927 : /*
1928 : * Compute the distance between two int2 values (plain subtraction).
1929 : */
1930 : Datum
1931 507 : brin_minmax_multi_distance_int2(PG_FUNCTION_ARGS)
1932 : {
1933 507 : int16 a1 = PG_GETARG_INT16(0);
1934 507 : int16 a2 = PG_GETARG_INT16(1);
1935 :
1936 : /*
1937 : * We know the values are range boundaries, but the range may be collapsed
1938 : * (i.e. single points), with equal values.
1939 : */
1940 507 : Assert(a1 <= a2);
1941 :
1942 507 : PG_RETURN_FLOAT8((double) a2 - (double) a1);
1943 : }
1944 :
1945 : /*
1946 : * Compute the distance between two int4 values (plain subtraction).
1947 : */
1948 : Datum
1949 40737 : brin_minmax_multi_distance_int4(PG_FUNCTION_ARGS)
1950 : {
1951 40737 : int32 a1 = PG_GETARG_INT32(0);
1952 40737 : int32 a2 = PG_GETARG_INT32(1);
1953 :
1954 : /*
1955 : * We know the values are range boundaries, but the range may be collapsed
1956 : * (i.e. single points), with equal values.
1957 : */
1958 40737 : Assert(a1 <= a2);
1959 :
1960 40737 : PG_RETURN_FLOAT8((double) a2 - (double) a1);
1961 : }
1962 :
1963 : /*
1964 : * Compute the distance between two int8 values (plain subtraction).
1965 : */
1966 : Datum
1967 1782 : brin_minmax_multi_distance_int8(PG_FUNCTION_ARGS)
1968 : {
1969 1782 : int64 a1 = PG_GETARG_INT64(0);
1970 1782 : int64 a2 = PG_GETARG_INT64(1);
1971 :
1972 : /*
1973 : * We know the values are range boundaries, but the range may be collapsed
1974 : * (i.e. single points), with equal values.
1975 : */
1976 1782 : Assert(a1 <= a2);
1977 :
1978 1782 : PG_RETURN_FLOAT8((double) a2 - (double) a1);
1979 : }
1980 :
1981 : /*
1982 : * Compute the distance between two tid values (by mapping them to float8 and
1983 : * then subtracting them).
1984 : */
1985 : Datum
1986 513 : brin_minmax_multi_distance_tid(PG_FUNCTION_ARGS)
1987 : {
1988 : double da1,
1989 : da2;
1990 :
1991 513 : ItemPointer pa1 = (ItemPointer) PG_GETARG_DATUM(0);
1992 513 : ItemPointer pa2 = (ItemPointer) PG_GETARG_DATUM(1);
1993 :
1994 : /*
1995 : * We know the values are range boundaries, but the range may be collapsed
1996 : * (i.e. single points), with equal values.
1997 : */
1998 513 : Assert(ItemPointerCompare(pa1, pa2) <= 0);
1999 :
2000 : /*
2001 : * We use the no-check variants here, because user-supplied values may
2002 : * have (ip_posid == 0). See ItemPointerCompare.
2003 : */
2004 513 : da1 = ItemPointerGetBlockNumberNoCheck(pa1) * MaxHeapTuplesPerPage +
2005 513 : ItemPointerGetOffsetNumberNoCheck(pa1);
2006 :
2007 513 : da2 = ItemPointerGetBlockNumberNoCheck(pa2) * MaxHeapTuplesPerPage +
2008 513 : ItemPointerGetOffsetNumberNoCheck(pa2);
2009 :
2010 513 : PG_RETURN_FLOAT8(da2 - da1);
2011 : }
2012 :
2013 : /*
2014 : * Compute the distance between two numeric values (plain subtraction).
2015 : */
2016 : Datum
2017 513 : brin_minmax_multi_distance_numeric(PG_FUNCTION_ARGS)
2018 : {
2019 : Datum d;
2020 513 : Datum a1 = PG_GETARG_DATUM(0);
2021 513 : Datum a2 = PG_GETARG_DATUM(1);
2022 :
2023 : /*
2024 : * We know the values are range boundaries, but the range may be collapsed
2025 : * (i.e. single points), with equal values.
2026 : */
2027 513 : Assert(DatumGetBool(DirectFunctionCall2(numeric_le, a1, a2)));
2028 :
2029 513 : d = DirectFunctionCall2(numeric_sub, a2, a1); /* a2 - a1 */
2030 :
2031 513 : PG_RETURN_FLOAT8(DirectFunctionCall1(numeric_float8, d));
2032 : }
2033 :
2034 : /*
2035 : * Compute the approximate distance between two UUID values.
2036 : *
2037 : * XXX We do not need a perfectly accurate value, so we approximate the
2038 : * deltas (which would have to be 128-bit integers) with a 64-bit float.
2039 : * The small inaccuracies do not matter in practice, in the worst case
2040 : * we'll decide to merge ranges that are not the closest ones.
2041 : */
2042 : Datum
2043 513 : brin_minmax_multi_distance_uuid(PG_FUNCTION_ARGS)
2044 : {
2045 : int i;
2046 513 : float8 delta = 0;
2047 :
2048 513 : Datum a1 = PG_GETARG_DATUM(0);
2049 513 : Datum a2 = PG_GETARG_DATUM(1);
2050 :
2051 513 : pg_uuid_t *u1 = DatumGetUUIDP(a1);
2052 513 : pg_uuid_t *u2 = DatumGetUUIDP(a2);
2053 :
2054 : /*
2055 : * We know the values are range boundaries, but the range may be collapsed
2056 : * (i.e. single points), with equal values.
2057 : */
2058 513 : Assert(DatumGetBool(DirectFunctionCall2(uuid_le, a1, a2)));
2059 :
2060 : /* compute approximate delta as a double precision value */
2061 8721 : for (i = UUID_LEN - 1; i >= 0; i--)
2062 : {
2063 8208 : delta += (int) u2->data[i] - (int) u1->data[i];
2064 8208 : delta /= 256;
2065 : }
2066 :
2067 513 : Assert(delta >= 0);
2068 :
2069 513 : PG_RETURN_FLOAT8(delta);
2070 : }
2071 :
2072 : /*
2073 : * Compute the approximate distance between two dates.
2074 : */
2075 : Datum
2076 513 : brin_minmax_multi_distance_date(PG_FUNCTION_ARGS)
2077 : {
2078 513 : DateADT dateVal1 = PG_GETARG_DATEADT(0);
2079 513 : DateADT dateVal2 = PG_GETARG_DATEADT(1);
2080 :
2081 513 : if (DATE_NOT_FINITE(dateVal1) || DATE_NOT_FINITE(dateVal2))
2082 UBC 0 : PG_RETURN_FLOAT8(0);
2083 :
2084 CBC 513 : PG_RETURN_FLOAT8(dateVal1 - dateVal2);
2085 : }
2086 :
2087 : /*
2088 : * Compute the approximate distance between two time (without tz) values.
2089 : *
2090 : * TimeADT is just an int64, so we simply subtract the values directly.
2091 : */
2092 : Datum
2093 507 : brin_minmax_multi_distance_time(PG_FUNCTION_ARGS)
2094 : {
2095 507 : float8 delta = 0;
2096 :
2097 507 : TimeADT ta = PG_GETARG_TIMEADT(0);
2098 507 : TimeADT tb = PG_GETARG_TIMEADT(1);
2099 :
2100 507 : delta = (tb - ta);
2101 :
2102 507 : Assert(delta >= 0);
2103 :
2104 507 : PG_RETURN_FLOAT8(delta);
2105 : }
2106 :
2107 : /*
2108 : * Compute the approximate distance between two timetz values.
2109 : *
2110 : * Simply subtracts the TimeADT (int64) values embedded in TimeTzADT.
2111 : */
2112 : Datum
2113 393 : brin_minmax_multi_distance_timetz(PG_FUNCTION_ARGS)
2114 : {
2115 393 : float8 delta = 0;
2116 :
2117 393 : TimeTzADT *ta = PG_GETARG_TIMETZADT_P(0);
2118 393 : TimeTzADT *tb = PG_GETARG_TIMETZADT_P(1);
2119 :
2120 393 : delta = (tb->time - ta->time) + (tb->zone - ta->zone) * USECS_PER_SEC;
2121 :
2122 393 : Assert(delta >= 0);
2123 :
2124 393 : PG_RETURN_FLOAT8(delta);
2125 : }
2126 :
2127 : /*
2128 : * Compute the distance between two timestamp values.
2129 : */
2130 : Datum
2131 1020 : brin_minmax_multi_distance_timestamp(PG_FUNCTION_ARGS)
2132 : {
2133 1020 : float8 delta = 0;
2134 :
2135 1020 : Timestamp dt1 = PG_GETARG_TIMESTAMP(0);
2136 1020 : Timestamp dt2 = PG_GETARG_TIMESTAMP(1);
2137 :
2138 1020 : if (TIMESTAMP_NOT_FINITE(dt1) || TIMESTAMP_NOT_FINITE(dt2))
2139 UBC 0 : PG_RETURN_FLOAT8(0);
2140 :
2141 CBC 1020 : delta = dt2 - dt1;
2142 :
2143 1020 : Assert(delta >= 0);
2144 :
2145 1020 : PG_RETURN_FLOAT8(delta);
2146 : }
2147 :
2148 : /*
2149 : * Compute the distance between two interval values.
2150 : */
2151 : Datum
2152 513 : brin_minmax_multi_distance_interval(PG_FUNCTION_ARGS)
2153 : {
2154 513 : float8 delta = 0;
2155 :
2156 513 : Interval *ia = PG_GETARG_INTERVAL_P(0);
2157 513 : Interval *ib = PG_GETARG_INTERVAL_P(1);
2158 : Interval *result;
2159 :
2160 : int64 dayfraction;
2161 : int64 days;
2162 :
2163 513 : result = (Interval *) palloc(sizeof(Interval));
2164 :
2165 513 : result->month = ib->month - ia->month;
2166 : /* overflow check copied from int4mi */
2167 513 : if (!SAMESIGN(ib->month, ia->month) &&
2168 UBC 0 : !SAMESIGN(result->month, ib->month))
2169 0 : ereport(ERROR,
2170 : (errcode(ERRCODE_DATETIME_VALUE_OUT_OF_RANGE),
2171 : errmsg("interval out of range")));
2172 :
2173 CBC 513 : result->day = ib->day - ia->day;
2174 513 : if (!SAMESIGN(ib->day, ia->day) &&
2175 UBC 0 : !SAMESIGN(result->day, ib->day))
2176 0 : ereport(ERROR,
2177 : (errcode(ERRCODE_DATETIME_VALUE_OUT_OF_RANGE),
2178 : errmsg("interval out of range")));
2179 :
2180 CBC 513 : result->time = ib->time - ia->time;
2181 513 : if (!SAMESIGN(ib->time, ia->time) &&
2182 UBC 0 : !SAMESIGN(result->time, ib->time))
2183 0 : ereport(ERROR,
2184 : (errcode(ERRCODE_DATETIME_VALUE_OUT_OF_RANGE),
2185 : errmsg("interval out of range")));
2186 :
2187 : /*
2188 : * Delta is (fractional) number of days between the intervals. Assume
2189 : * months have 30 days for consistency with interval_cmp_internal. We
2190 : * don't need to be exact, in the worst case we'll build a bit less
2191 : * efficient ranges. But we should not contradict interval_cmp.
2192 : */
2193 CBC 513 : dayfraction = result->time % USECS_PER_DAY;
2194 513 : days = result->time / USECS_PER_DAY;
2195 513 : days += result->month * INT64CONST(30);
2196 513 : days += result->day;
2197 :
2198 : /* convert to double precision */
2199 513 : delta = (double) days + dayfraction / (double) USECS_PER_DAY;
2200 :
2201 513 : Assert(delta >= 0);
2202 :
2203 513 : PG_RETURN_FLOAT8(delta);
2204 : }
2205 :
2206 : /*
2207 : * Compute the distance between two pg_lsn values.
2208 : *
2209 : * LSN is just an int64 encoding position in the stream, so just subtract
2210 : * those int64 values directly.
2211 : */
2212 : Datum
2213 513 : brin_minmax_multi_distance_pg_lsn(PG_FUNCTION_ARGS)
2214 : {
2215 513 : float8 delta = 0;
2216 :
2217 513 : XLogRecPtr lsna = PG_GETARG_LSN(0);
2218 513 : XLogRecPtr lsnb = PG_GETARG_LSN(1);
2219 :
2220 513 : delta = (lsnb - lsna);
2221 :
2222 513 : Assert(delta >= 0);
2223 :
2224 513 : PG_RETURN_FLOAT8(delta);
2225 : }
2226 :
2227 : /*
2228 : * Compute the distance between two macaddr values.
2229 : *
2230 : * mac addresses are treated as 6 unsigned chars, so do the same thing we
2231 : * already do for UUID values.
2232 : */
2233 : Datum
2234 393 : brin_minmax_multi_distance_macaddr(PG_FUNCTION_ARGS)
2235 : {
2236 : float8 delta;
2237 :
2238 393 : macaddr *a = PG_GETARG_MACADDR_P(0);
2239 393 : macaddr *b = PG_GETARG_MACADDR_P(1);
2240 :
2241 393 : delta = ((float8) b->f - (float8) a->f);
2242 393 : delta /= 256;
2243 :
2244 393 : delta += ((float8) b->e - (float8) a->e);
2245 393 : delta /= 256;
2246 :
2247 393 : delta += ((float8) b->d - (float8) a->d);
2248 393 : delta /= 256;
2249 :
2250 393 : delta += ((float8) b->c - (float8) a->c);
2251 393 : delta /= 256;
2252 :
2253 393 : delta += ((float8) b->b - (float8) a->b);
2254 393 : delta /= 256;
2255 :
2256 393 : delta += ((float8) b->a - (float8) a->a);
2257 393 : delta /= 256;
2258 :
2259 393 : Assert(delta >= 0);
2260 :
2261 393 : PG_RETURN_FLOAT8(delta);
2262 : }
2263 :
2264 : /*
2265 : * Compute the distance between two macaddr8 values.
2266 : *
2267 : * macaddr8 addresses are 8 unsigned chars, so do the same thing we
2268 : * already do for UUID values.
2269 : */
2270 : Datum
2271 513 : brin_minmax_multi_distance_macaddr8(PG_FUNCTION_ARGS)
2272 : {
2273 : float8 delta;
2274 :
2275 513 : macaddr8 *a = PG_GETARG_MACADDR8_P(0);
2276 513 : macaddr8 *b = PG_GETARG_MACADDR8_P(1);
2277 :
2278 513 : delta = ((float8) b->h - (float8) a->h);
2279 513 : delta /= 256;
2280 :
2281 513 : delta += ((float8) b->g - (float8) a->g);
2282 513 : delta /= 256;
2283 :
2284 513 : delta += ((float8) b->f - (float8) a->f);
2285 513 : delta /= 256;
2286 :
2287 513 : delta += ((float8) b->e - (float8) a->e);
2288 513 : delta /= 256;
2289 :
2290 513 : delta += ((float8) b->d - (float8) a->d);
2291 513 : delta /= 256;
2292 :
2293 513 : delta += ((float8) b->c - (float8) a->c);
2294 513 : delta /= 256;
2295 :
2296 513 : delta += ((float8) b->b - (float8) a->b);
2297 513 : delta /= 256;
2298 :
2299 513 : delta += ((float8) b->a - (float8) a->a);
2300 513 : delta /= 256;
2301 :
2302 513 : Assert(delta >= 0);
2303 :
2304 513 : PG_RETURN_FLOAT8(delta);
2305 : }
2306 :
2307 : /*
2308 : * Compute the distance between two inet values.
2309 : *
2310 : * The distance is defined as the difference between 32-bit/128-bit values,
2311 : * depending on the IP version. The distance is computed by subtracting
2312 : * the bytes and normalizing it to [0,1] range for each IP family.
2313 : * Addresses from different families are considered to be in maximum
2314 : * distance, which is 1.0.
2315 : *
2316 : * XXX Does this need to consider the mask (bits)? For now, it's ignored.
2317 : */
2318 : Datum
2319 1137 : brin_minmax_multi_distance_inet(PG_FUNCTION_ARGS)
2320 : {
2321 : float8 delta;
2322 : int i;
2323 : int len;
2324 : unsigned char *addra,
2325 : *addrb;
2326 :
2327 1137 : inet *ipa = PG_GETARG_INET_PP(0);
2328 1137 : inet *ipb = PG_GETARG_INET_PP(1);
2329 :
2330 : int lena,
2331 : lenb;
2332 :
2333 : /*
2334 : * If the addresses are from different families, consider them to be in
2335 : * maximal possible distance (which is 1.0).
2336 : */
2337 1137 : if (ip_family(ipa) != ip_family(ipb))
2338 90 : PG_RETURN_FLOAT8(1.0);
2339 :
2340 1047 : addra = (unsigned char *) palloc(ip_addrsize(ipa));
2341 1047 : memcpy(addra, ip_addr(ipa), ip_addrsize(ipa));
2342 :
2343 1047 : addrb = (unsigned char *) palloc(ip_addrsize(ipb));
2344 1047 : memcpy(addrb, ip_addr(ipb), ip_addrsize(ipb));
2345 :
2346 : /*
2347 : * The length is calculated from the mask length, because we sort the
2348 : * addresses by first address in the range, so A.B.C.D/24 < A.B.C.1 (the
2349 : * first range starts at A.B.C.0, which is before A.B.C.1). We don't want
2350 : * to produce a negative delta in this case, so we just cut the extra
2351 : * bytes.
2352 : *
2353 : * XXX Maybe this should be a bit more careful and cut the bits, not just
2354 : * whole bytes.
2355 : */
2356 1047 : lena = ip_bits(ipa);
2357 1047 : lenb = ip_bits(ipb);
2358 :
2359 1047 : len = ip_addrsize(ipa);
2360 :
2361 : /* apply the network mask to both addresses */
2362 7899 : for (i = 0; i < len; i++)
2363 : {
2364 : unsigned char mask;
2365 : int nbits;
2366 :
2367 6852 : nbits = Max(0, lena - (i * 8));
2368 6852 : if (nbits < 8)
2369 : {
2370 825 : mask = (0xFF << (8 - nbits));
2371 825 : addra[i] = (addra[i] & mask);
2372 : }
2373 :
2374 6852 : nbits = Max(0, lenb - (i * 8));
2375 6852 : if (nbits < 8)
2376 : {
2377 822 : mask = (0xFF << (8 - nbits));
2378 822 : addrb[i] = (addrb[i] & mask);
2379 : }
2380 : }
2381 :
2382 : /* Calculate the difference between the addresses. */
2383 1047 : delta = 0;
2384 7899 : for (i = len - 1; i >= 0; i--)
2385 : {
2386 6852 : unsigned char a = addra[i];
2387 6852 : unsigned char b = addrb[i];
2388 :
2389 6852 : delta += (float8) b - (float8) a;
2390 6852 : delta /= 256;
2391 : }
2392 :
2393 1047 : Assert((delta >= 0) && (delta <= 1));
2394 :
2395 1047 : pfree(addra);
2396 1047 : pfree(addrb);
2397 :
2398 1047 : PG_RETURN_FLOAT8(delta);
2399 : }
2400 :
2401 : static void
2402 8535 : brin_minmax_multi_serialize(BrinDesc *bdesc, Datum src, Datum *dst)
2403 : {
2404 8535 : Ranges *ranges = (Ranges *) DatumGetPointer(src);
2405 : SerializedRanges *s;
2406 :
2407 : /*
2408 : * In batch mode, we need to compress the accumulated values to the
2409 : * actually requested number of values/ranges.
2410 : */
2411 8535 : compactify_ranges(bdesc, ranges, ranges->target_maxvalues);
2412 :
2413 : /* At this point everything has to be fully sorted. */
2414 8535 : Assert(ranges->nsorted == ranges->nvalues);
2415 :
2416 8535 : s = brin_range_serialize(ranges);
2417 8535 : dst[0] = PointerGetDatum(s);
2418 8535 : }
2419 :
2420 : static int
2421 2226 : brin_minmax_multi_get_values(BrinDesc *bdesc, MinMaxMultiOptions *opts)
2422 : {
2423 2226 : return MinMaxMultiGetValuesPerRange(opts);
2424 : }
2425 :
2426 : /*
2427 : * Examine the given index tuple (which contains the partial status of a
2428 : * certain page range) by comparing it to the given value that comes from
2429 : * another heap tuple. If the new value is outside the min/max range
2430 : * specified by the existing tuple values, update the index tuple and return
2431 : * true. Otherwise, return false and do not modify in this case.
2432 : */
2433 : Datum
2434 49104 : brin_minmax_multi_add_value(PG_FUNCTION_ARGS)
2435 : {
2436 49104 : BrinDesc *bdesc = (BrinDesc *) PG_GETARG_POINTER(0);
2437 49104 : BrinValues *column = (BrinValues *) PG_GETARG_POINTER(1);
2438 49104 : Datum newval = PG_GETARG_DATUM(2);
2439 49104 : bool isnull PG_USED_FOR_ASSERTS_ONLY = PG_GETARG_DATUM(3);
2440 49104 : MinMaxMultiOptions *opts = (MinMaxMultiOptions *) PG_GET_OPCLASS_OPTIONS();
2441 49104 : Oid colloid = PG_GET_COLLATION();
2442 49104 : bool modified = false;
2443 : Form_pg_attribute attr;
2444 : AttrNumber attno;
2445 : Ranges *ranges;
2446 49104 : SerializedRanges *serialized = NULL;
2447 :
2448 49104 : Assert(!isnull);
2449 :
2450 49104 : attno = column->bv_attno;
2451 49104 : attr = TupleDescAttr(bdesc->bd_tupdesc, attno - 1);
2452 :
2453 : /* use the already deserialized value, if possible */
2454 49104 : ranges = (Ranges *) DatumGetPointer(column->bv_mem_value);
2455 :
2456 : /*
2457 : * If this is the first non-null value, we need to initialize the range
2458 : * list. Otherwise, just extract the existing range list from BrinValues.
2459 : *
2460 : * When starting with an empty range, we assume this is a batch mode and
2461 : * we use a larger buffer. The buffer size is derived from the BRIN range
2462 : * size, number of rows per page, with some sensible min/max values. A
2463 : * small buffer would be bad for performance, but a large buffer might
2464 : * require a lot of memory (because of keeping all the values).
2465 : */
2466 49104 : if (column->bv_allnulls)
2467 : {
2468 : MemoryContext oldctx;
2469 :
2470 : int target_maxvalues;
2471 : int maxvalues;
2472 2226 : BlockNumber pagesPerRange = BrinGetPagesPerRange(bdesc->bd_index);
2473 :
2474 : /* what was specified as a reloption? */
2475 2226 : target_maxvalues = brin_minmax_multi_get_values(bdesc, opts);
2476 :
2477 : /*
2478 : * Determine the insert buffer size - we use 10x the target, capped to
2479 : * the maximum number of values in the heap range. This is more than
2480 : * enough, considering the actual number of rows per page is likely
2481 : * much lower, but meh.
2482 : */
2483 2226 : maxvalues = Min(target_maxvalues * MINMAX_BUFFER_FACTOR,
2484 : MaxHeapTuplesPerPage * pagesPerRange);
2485 :
2486 : /* but always at least the original value */
2487 2226 : maxvalues = Max(maxvalues, target_maxvalues);
2488 :
2489 : /* always cap by MIN/MAX */
2490 2226 : maxvalues = Max(maxvalues, MINMAX_BUFFER_MIN);
2491 2226 : maxvalues = Min(maxvalues, MINMAX_BUFFER_MAX);
2492 :
2493 2226 : oldctx = MemoryContextSwitchTo(column->bv_context);
2494 2226 : ranges = minmax_multi_init(maxvalues);
2495 2226 : ranges->attno = attno;
2496 2226 : ranges->colloid = colloid;
2497 2226 : ranges->typid = attr->atttypid;
2498 2226 : ranges->target_maxvalues = target_maxvalues;
2499 :
2500 : /* we'll certainly need the comparator, so just look it up now */
2501 2226 : ranges->cmp = minmax_multi_get_strategy_procinfo(bdesc, attno, attr->atttypid,
2502 : BTLessStrategyNumber);
2503 :
2504 2226 : MemoryContextSwitchTo(oldctx);
2505 :
2506 2226 : column->bv_allnulls = false;
2507 2226 : modified = true;
2508 :
2509 2226 : column->bv_mem_value = PointerGetDatum(ranges);
2510 2226 : column->bv_serialize = brin_minmax_multi_serialize;
2511 : }
2512 46878 : else if (!ranges)
2513 : {
2514 : MemoryContext oldctx;
2515 :
2516 : int maxvalues;
2517 6309 : BlockNumber pagesPerRange = BrinGetPagesPerRange(bdesc->bd_index);
2518 :
2519 6309 : oldctx = MemoryContextSwitchTo(column->bv_context);
2520 :
2521 6309 : serialized = (SerializedRanges *) PG_DETOAST_DATUM(column->bv_values[0]);
2522 :
2523 : /*
2524 : * Determine the insert buffer size - we use 10x the target, capped to
2525 : * the maximum number of values in the heap range. This is more than
2526 : * enough, considering the actual number of rows per page is likely
2527 : * much lower, but meh.
2528 : */
2529 6309 : maxvalues = Min(serialized->maxvalues * MINMAX_BUFFER_FACTOR,
2530 : MaxHeapTuplesPerPage * pagesPerRange);
2531 :
2532 : /* but always at least the original value */
2533 6309 : maxvalues = Max(maxvalues, serialized->maxvalues);
2534 :
2535 : /* always cap by MIN/MAX */
2536 6309 : maxvalues = Max(maxvalues, MINMAX_BUFFER_MIN);
2537 6309 : maxvalues = Min(maxvalues, MINMAX_BUFFER_MAX);
2538 :
2539 6309 : ranges = brin_range_deserialize(maxvalues, serialized);
2540 :
2541 6309 : ranges->attno = attno;
2542 6309 : ranges->colloid = colloid;
2543 6309 : ranges->typid = attr->atttypid;
2544 :
2545 : /* we'll certainly need the comparator, so just look it up now */
2546 6309 : ranges->cmp = minmax_multi_get_strategy_procinfo(bdesc, attno, attr->atttypid,
2547 : BTLessStrategyNumber);
2548 :
2549 6309 : column->bv_mem_value = PointerGetDatum(ranges);
2550 6309 : column->bv_serialize = brin_minmax_multi_serialize;
2551 :
2552 6309 : MemoryContextSwitchTo(oldctx);
2553 : }
2554 :
2555 : /*
2556 : * Try to add the new value to the range. We need to update the modified
2557 : * flag, so that we serialize the updated summary later.
2558 : */
2559 49104 : modified |= range_add_value(bdesc, colloid, attno, attr, ranges, newval);
2560 :
2561 :
2562 49104 : PG_RETURN_BOOL(modified);
2563 : }
2564 :
2565 : /*
2566 : * Given an index tuple corresponding to a certain page range and a scan key,
2567 : * return whether the scan key is consistent with the index tuple's min/max
2568 : * values. Return true if so, false otherwise.
2569 : */
2570 : Datum
2571 14652 : brin_minmax_multi_consistent(PG_FUNCTION_ARGS)
2572 : {
2573 14652 : BrinDesc *bdesc = (BrinDesc *) PG_GETARG_POINTER(0);
2574 14652 : BrinValues *column = (BrinValues *) PG_GETARG_POINTER(1);
2575 14652 : ScanKey *keys = (ScanKey *) PG_GETARG_POINTER(2);
2576 14652 : int nkeys = PG_GETARG_INT32(3);
2577 :
2578 14652 : Oid colloid = PG_GET_COLLATION(),
2579 : subtype;
2580 : AttrNumber attno;
2581 : Datum value;
2582 : FmgrInfo *finfo;
2583 : SerializedRanges *serialized;
2584 : Ranges *ranges;
2585 : int keyno;
2586 : int rangeno;
2587 : int i;
2588 :
2589 14652 : attno = column->bv_attno;
2590 :
2591 14652 : serialized = (SerializedRanges *) PG_DETOAST_DATUM(column->bv_values[0]);
2592 14652 : ranges = brin_range_deserialize(serialized->maxvalues, serialized);
2593 :
2594 : /* inspect the ranges, and for each one evaluate the scan keys */
2595 14652 : for (rangeno = 0; rangeno < ranges->nranges; rangeno++)
2596 : {
2597 UBC 0 : Datum minval = ranges->values[2 * rangeno];
2598 0 : Datum maxval = ranges->values[2 * rangeno + 1];
2599 :
2600 : /* assume the range is matching, and we'll try to prove otherwise */
2601 0 : bool matching = true;
2602 :
2603 0 : for (keyno = 0; keyno < nkeys; keyno++)
2604 : {
2605 : Datum matches;
2606 0 : ScanKey key = keys[keyno];
2607 :
2608 : /* NULL keys are handled and filtered-out in bringetbitmap */
2609 0 : Assert(!(key->sk_flags & SK_ISNULL));
2610 :
2611 0 : attno = key->sk_attno;
2612 0 : subtype = key->sk_subtype;
2613 0 : value = key->sk_argument;
2614 0 : switch (key->sk_strategy)
2615 : {
2616 0 : case BTLessStrategyNumber:
2617 : case BTLessEqualStrategyNumber:
2618 0 : finfo = minmax_multi_get_strategy_procinfo(bdesc, attno, subtype,
2619 0 : key->sk_strategy);
2620 : /* first value from the array */
2621 0 : matches = FunctionCall2Coll(finfo, colloid, minval, value);
2622 0 : break;
2623 :
2624 0 : case BTEqualStrategyNumber:
2625 : {
2626 : Datum compar;
2627 : FmgrInfo *cmpFn;
2628 :
2629 : /* by default this range does not match */
2630 0 : matches = false;
2631 :
2632 : /*
2633 : * Otherwise, need to compare the new value with
2634 : * boundaries of all the ranges. First check if it's
2635 : * less than the absolute minimum, which is the first
2636 : * value in the array.
2637 : */
2638 0 : cmpFn = minmax_multi_get_strategy_procinfo(bdesc, attno, subtype,
2639 : BTGreaterStrategyNumber);
2640 0 : compar = FunctionCall2Coll(cmpFn, colloid, minval, value);
2641 :
2642 : /* smaller than the smallest value in this range */
2643 0 : if (DatumGetBool(compar))
2644 0 : break;
2645 :
2646 0 : cmpFn = minmax_multi_get_strategy_procinfo(bdesc, attno, subtype,
2647 : BTLessStrategyNumber);
2648 0 : compar = FunctionCall2Coll(cmpFn, colloid, maxval, value);
2649 :
2650 : /* larger than the largest value in this range */
2651 0 : if (DatumGetBool(compar))
2652 0 : break;
2653 :
2654 : /*
2655 : * We haven't managed to eliminate this range, so
2656 : * consider it matching.
2657 : */
2658 0 : matches = true;
2659 :
2660 0 : break;
2661 : }
2662 0 : case BTGreaterEqualStrategyNumber:
2663 : case BTGreaterStrategyNumber:
2664 0 : finfo = minmax_multi_get_strategy_procinfo(bdesc, attno, subtype,
2665 0 : key->sk_strategy);
2666 : /* last value from the array */
2667 0 : matches = FunctionCall2Coll(finfo, colloid, maxval, value);
2668 0 : break;
2669 :
2670 0 : default:
2671 : /* shouldn't happen */
2672 0 : elog(ERROR, "invalid strategy number %d", key->sk_strategy);
2673 : matches = 0;
2674 : break;
2675 : }
2676 :
2677 : /* the range has to match all the scan keys */
2678 0 : matching &= DatumGetBool(matches);
2679 :
2680 : /* once we find a non-matching key, we're done */
2681 0 : if (!matching)
2682 0 : break;
2683 : }
2684 :
2685 : /*
2686 : * have we found a range matching all scan keys? if yes, we're done
2687 : */
2688 0 : if (matching)
2689 0 : PG_RETURN_DATUM(BoolGetDatum(true));
2690 : }
2691 :
2692 : /*
2693 : * And now inspect the values. We don't bother with doing a binary search
2694 : * here, because we're dealing with serialized / fully compacted ranges,
2695 : * so there should be only very few values.
2696 : */
2697 CBC 23937 : for (i = 0; i < ranges->nvalues; i++)
2698 : {
2699 21033 : Datum val = ranges->values[2 * ranges->nranges + i];
2700 :
2701 : /* assume the range is matching, and we'll try to prove otherwise */
2702 21033 : bool matching = true;
2703 :
2704 32781 : for (keyno = 0; keyno < nkeys; keyno++)
2705 : {
2706 : Datum matches;
2707 21033 : ScanKey key = keys[keyno];
2708 :
2709 : /* we've already dealt with NULL keys at the beginning */
2710 21033 : if (key->sk_flags & SK_ISNULL)
2711 UBC 0 : continue;
2712 :
2713 CBC 21033 : attno = key->sk_attno;
2714 21033 : subtype = key->sk_subtype;
2715 21033 : value = key->sk_argument;
2716 21033 : switch (key->sk_strategy)
2717 : {
2718 21033 : case BTLessStrategyNumber:
2719 : case BTLessEqualStrategyNumber:
2720 : case BTEqualStrategyNumber:
2721 : case BTGreaterEqualStrategyNumber:
2722 : case BTGreaterStrategyNumber:
2723 :
2724 21033 : finfo = minmax_multi_get_strategy_procinfo(bdesc, attno, subtype,
2725 21033 : key->sk_strategy);
2726 21033 : matches = FunctionCall2Coll(finfo, colloid, val, value);
2727 21033 : break;
2728 :
2729 UBC 0 : default:
2730 : /* shouldn't happen */
2731 0 : elog(ERROR, "invalid strategy number %d", key->sk_strategy);
2732 : matches = 0;
2733 : break;
2734 : }
2735 :
2736 : /* the range has to match all the scan keys */
2737 CBC 21033 : matching &= DatumGetBool(matches);
2738 :
2739 : /* once we find a non-matching key, we're done */
2740 21033 : if (!matching)
2741 9285 : break;
2742 : }
2743 :
2744 : /* have we found a range matching all scan keys? if yes, we're done */
2745 21033 : if (matching)
2746 11748 : PG_RETURN_DATUM(BoolGetDatum(true));
2747 : }
2748 :
2749 2904 : PG_RETURN_DATUM(BoolGetDatum(false));
2750 : }
2751 :
2752 : /*
2753 : * Given two BrinValues, update the first of them as a union of the summary
2754 : * values contained in both. The second one is untouched.
2755 : */
2756 : Datum
2757 UBC 0 : brin_minmax_multi_union(PG_FUNCTION_ARGS)
2758 : {
2759 0 : BrinDesc *bdesc = (BrinDesc *) PG_GETARG_POINTER(0);
2760 0 : BrinValues *col_a = (BrinValues *) PG_GETARG_POINTER(1);
2761 0 : BrinValues *col_b = (BrinValues *) PG_GETARG_POINTER(2);
2762 :
2763 0 : Oid colloid = PG_GET_COLLATION();
2764 : SerializedRanges *serialized_a;
2765 : SerializedRanges *serialized_b;
2766 : Ranges *ranges_a;
2767 : Ranges *ranges_b;
2768 : AttrNumber attno;
2769 : Form_pg_attribute attr;
2770 : ExpandedRange *eranges;
2771 : int neranges;
2772 : FmgrInfo *cmpFn,
2773 : *distanceFn;
2774 : DistanceValue *distances;
2775 : MemoryContext ctx;
2776 : MemoryContext oldctx;
2777 :
2778 0 : Assert(col_a->bv_attno == col_b->bv_attno);
2779 0 : Assert(!col_a->bv_allnulls && !col_b->bv_allnulls);
2780 :
2781 0 : attno = col_a->bv_attno;
2782 0 : attr = TupleDescAttr(bdesc->bd_tupdesc, attno - 1);
2783 :
2784 0 : serialized_a = (SerializedRanges *) PG_DETOAST_DATUM(col_a->bv_values[0]);
2785 0 : serialized_b = (SerializedRanges *) PG_DETOAST_DATUM(col_b->bv_values[0]);
2786 :
2787 0 : ranges_a = brin_range_deserialize(serialized_a->maxvalues, serialized_a);
2788 0 : ranges_b = brin_range_deserialize(serialized_b->maxvalues, serialized_b);
2789 :
2790 : /* make sure neither of the ranges is NULL */
2791 0 : Assert(ranges_a && ranges_b);
2792 :
2793 0 : neranges = (ranges_a->nranges + ranges_a->nvalues) +
2794 0 : (ranges_b->nranges + ranges_b->nvalues);
2795 :
2796 : /*
2797 : * The distanceFn calls (which may internally call e.g. numeric_le) may
2798 : * allocate quite a bit of memory, and we must not leak it. Otherwise,
2799 : * we'd have problems e.g. when building indexes. So we create a local
2800 : * memory context and make sure we free the memory before leaving this
2801 : * function (not after every call).
2802 : */
2803 0 : ctx = AllocSetContextCreate(CurrentMemoryContext,
2804 : "minmax-multi context",
2805 : ALLOCSET_DEFAULT_SIZES);
2806 :
2807 0 : oldctx = MemoryContextSwitchTo(ctx);
2808 :
2809 : /* allocate and fill */
2810 0 : eranges = (ExpandedRange *) palloc0(neranges * sizeof(ExpandedRange));
2811 :
2812 : /* fill the expanded ranges with entries for the first range */
2813 0 : fill_expanded_ranges(eranges, ranges_a->nranges + ranges_a->nvalues,
2814 : ranges_a);
2815 :
2816 : /* and now add combine ranges for the second range */
2817 0 : fill_expanded_ranges(&eranges[ranges_a->nranges + ranges_a->nvalues],
2818 0 : ranges_b->nranges + ranges_b->nvalues,
2819 : ranges_b);
2820 :
2821 0 : cmpFn = minmax_multi_get_strategy_procinfo(bdesc, attno, attr->atttypid,
2822 : BTLessStrategyNumber);
2823 :
2824 : /* sort the expanded ranges */
2825 0 : neranges = sort_expanded_ranges(cmpFn, colloid, eranges, neranges);
2826 :
2827 : /*
2828 : * We've loaded two different lists of expanded ranges, so some of them
2829 : * may be overlapping. So walk through them and merge them.
2830 : */
2831 0 : neranges = merge_overlapping_ranges(cmpFn, colloid, eranges, neranges);
2832 :
2833 : /* check that the combine ranges are correct (no overlaps, ordering) */
2834 0 : AssertCheckExpandedRanges(bdesc, colloid, attno, attr, eranges, neranges);
2835 :
2836 : /*
2837 : * If needed, reduce some of the ranges.
2838 : *
2839 : * XXX This may be fairly expensive, so maybe we should do it only when
2840 : * it's actually needed (when we have too many ranges).
2841 : */
2842 :
2843 : /* build array of gap distances and sort them in ascending order */
2844 0 : distanceFn = minmax_multi_get_procinfo(bdesc, attno, PROCNUM_DISTANCE);
2845 0 : distances = build_distances(distanceFn, colloid, eranges, neranges);
2846 :
2847 : /*
2848 : * See how many values would be needed to store the current ranges, and if
2849 : * needed combine as many of them to get below the threshold. The
2850 : * collapsed ranges will be stored as a single value.
2851 : *
2852 : * XXX This does not apply the load factor, as we don't expect to add more
2853 : * values to the range, so we prefer to keep as many ranges as possible.
2854 : *
2855 : * XXX Can the maxvalues be different in the two ranges? Perhaps we should
2856 : * use maximum of those?
2857 : */
2858 0 : neranges = reduce_expanded_ranges(eranges, neranges, distances,
2859 : ranges_a->maxvalues,
2860 : cmpFn, colloid);
2861 :
2862 : /* update the first range summary */
2863 0 : store_expanded_ranges(ranges_a, eranges, neranges);
2864 :
2865 0 : MemoryContextSwitchTo(oldctx);
2866 0 : MemoryContextDelete(ctx);
2867 :
2868 : /* cleanup and update the serialized value */
2869 0 : pfree(serialized_a);
2870 0 : col_a->bv_values[0] = PointerGetDatum(brin_range_serialize(ranges_a));
2871 :
2872 0 : PG_RETURN_VOID();
2873 : }
2874 :
2875 : /*
2876 : * Cache and return minmax multi opclass support procedure
2877 : *
2878 : * Return the procedure corresponding to the given function support number
2879 : * or null if it does not exist.
2880 : */
2881 : static FmgrInfo *
2882 CBC 2667 : minmax_multi_get_procinfo(BrinDesc *bdesc, uint16 attno, uint16 procnum)
2883 : {
2884 : MinmaxMultiOpaque *opaque;
2885 2667 : uint16 basenum = procnum - PROCNUM_BASE;
2886 :
2887 : /*
2888 : * We cache these in the opaque struct, to avoid repetitive syscache
2889 : * lookups.
2890 : */
2891 2667 : opaque = (MinmaxMultiOpaque *) bdesc->bd_info[attno - 1]->oi_opaque;
2892 :
2893 : /*
2894 : * If we already searched for this proc and didn't find it, don't bother
2895 : * searching again.
2896 : */
2897 2667 : if (opaque->extra_proc_missing[basenum])
2898 UBC 0 : return NULL;
2899 :
2900 CBC 2667 : if (opaque->extra_procinfos[basenum].fn_oid == InvalidOid)
2901 : {
2902 195 : if (RegProcedureIsValid(index_getprocid(bdesc->bd_index, attno,
2903 : procnum)))
2904 : {
2905 195 : fmgr_info_copy(&opaque->extra_procinfos[basenum],
2906 : index_getprocinfo(bdesc->bd_index, attno, procnum),
2907 : bdesc->bd_context);
2908 : }
2909 : else
2910 : {
2911 UBC 0 : opaque->extra_proc_missing[basenum] = true;
2912 0 : return NULL;
2913 : }
2914 : }
2915 :
2916 CBC 2667 : return &opaque->extra_procinfos[basenum];
2917 : }
2918 :
2919 : /*
2920 : * Cache and return the procedure for the given strategy.
2921 : *
2922 : * Note: this function mirrors minmax_multi_get_strategy_procinfo; see notes
2923 : * there. If changes are made here, see that function too.
2924 : */
2925 : static FmgrInfo *
2926 289692 : minmax_multi_get_strategy_procinfo(BrinDesc *bdesc, uint16 attno, Oid subtype,
2927 : uint16 strategynum)
2928 : {
2929 : MinmaxMultiOpaque *opaque;
2930 :
2931 289692 : Assert(strategynum >= 1 &&
2932 : strategynum <= BTMaxStrategyNumber);
2933 :
2934 289692 : opaque = (MinmaxMultiOpaque *) bdesc->bd_info[attno - 1]->oi_opaque;
2935 :
2936 : /*
2937 : * We cache the procedures for the previous subtype in the opaque struct,
2938 : * to avoid repetitive syscache lookups. If the subtype changed,
2939 : * invalidate all the cached entries.
2940 : */
2941 289692 : if (opaque->cached_subtype != subtype)
2942 : {
2943 : uint16 i;
2944 :
2945 4194 : for (i = 1; i <= BTMaxStrategyNumber; i++)
2946 3495 : opaque->strategy_procinfos[i - 1].fn_oid = InvalidOid;
2947 699 : opaque->cached_subtype = subtype;
2948 : }
2949 :
2950 289692 : if (opaque->strategy_procinfos[strategynum - 1].fn_oid == InvalidOid)
2951 : {
2952 : Form_pg_attribute attr;
2953 : HeapTuple tuple;
2954 : Oid opfamily,
2955 : oprid;
2956 ECB :
2957 CBC 957 : opfamily = bdesc->bd_index->rd_opfamily[attno - 1];
2958 957 : attr = TupleDescAttr(bdesc->bd_tupdesc, attno - 1);
2959 GIC 957 : tuple = SearchSysCache4(AMOPSTRATEGY, ObjectIdGetDatum(opfamily),
2960 : ObjectIdGetDatum(attr->atttypid),
2961 : ObjectIdGetDatum(subtype),
2962 ECB : Int16GetDatum(strategynum));
2963 GBC 957 : if (!HeapTupleIsValid(tuple))
2964 UIC 0 : elog(ERROR, "missing operator %d(%u,%u) in opfamily %u",
2965 : strategynum, attr->atttypid, subtype, opfamily);
2966 ECB :
2967 GNC 957 : oprid = DatumGetObjectId(SysCacheGetAttrNotNull(AMOPSTRATEGY, tuple,
2968 : Anum_pg_amop_amopopr));
2969 CBC 957 : ReleaseSysCache(tuple);
2970 GNC 957 : Assert(RegProcedureIsValid(oprid));
2971 ECB :
2972 CBC 957 : fmgr_info_cxt(get_opcode(oprid),
2973 GIC 957 : &opaque->strategy_procinfos[strategynum - 1],
2974 : bdesc->bd_context);
2975 : }
2976 ECB :
2977 GIC 289692 : return &opaque->strategy_procinfos[strategynum - 1];
2978 : }
2979 :
2980 ECB : Datum
2981 GIC 344 : brin_minmax_multi_options(PG_FUNCTION_ARGS)
2982 ECB : {
2983 GIC 344 : local_relopts *relopts = (local_relopts *) PG_GETARG_POINTER(0);
2984 ECB :
2985 GIC 344 : init_local_reloptions(relopts, sizeof(MinMaxMultiOptions));
2986 ECB :
2987 GIC 344 : add_local_int_reloption(relopts, "values_per_range", "desc",
2988 : MINMAX_MULTI_DEFAULT_VALUES_PER_PAGE, 8, 256,
2989 : offsetof(MinMaxMultiOptions, valuesPerRange));
2990 ECB :
2991 GIC 344 : PG_RETURN_VOID();
2992 : }
2993 :
2994 : /*
2995 : * brin_minmax_multi_summary_in
2996 : * - input routine for type brin_minmax_multi_summary.
2997 : *
2998 : * brin_minmax_multi_summary is only used internally to represent summaries
2999 : * in BRIN minmax-multi indexes, so it has no operations of its own, and we
3000 : * disallow input too.
3001 : */
3002 EUB : Datum
3003 UIC 0 : brin_minmax_multi_summary_in(PG_FUNCTION_ARGS)
3004 : {
3005 : /*
3006 : * brin_minmax_multi_summary stores the data in binary form and parsing
3007 : * text input is not needed, so disallow this.
3008 EUB : */
3009 UIC 0 : ereport(ERROR,
3010 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
3011 : errmsg("cannot accept a value of type %s", "brin_minmax_multi_summary")));
3012 :
3013 : PG_RETURN_VOID(); /* keep compiler quiet */
3014 : }
3015 :
3016 :
3017 : /*
3018 : * brin_minmax_multi_summary_out
3019 : * - output routine for type brin_minmax_multi_summary.
3020 : *
3021 : * BRIN minmax-multi summaries are serialized into a bytea value, but we
3022 : * want to output something nicer humans can understand.
3023 : */
3024 EUB : Datum
3025 UIC 0 : brin_minmax_multi_summary_out(PG_FUNCTION_ARGS)
3026 : {
3027 : int i;
3028 : int idx;
3029 : SerializedRanges *ranges;
3030 : Ranges *ranges_deserialized;
3031 : StringInfoData str;
3032 : bool isvarlena;
3033 : Oid outfunc;
3034 EUB : FmgrInfo fmgrinfo;
3035 UIC 0 : ArrayBuildState *astate_values = NULL;
3036 EUB :
3037 UBC 0 : initStringInfo(&str);
3038 UIC 0 : appendStringInfoChar(&str, '{');
3039 :
3040 : /*
3041 : * Detoast to get value with full 4B header (can't be stored in a toast
3042 : * table, but can use 1B header).
3043 EUB : */
3044 UNC 0 : ranges = (SerializedRanges *) PG_DETOAST_DATUM_PACKED(PG_GETARG_DATUM(0));
3045 :
3046 EUB : /* lookup output func for the type */
3047 UBC 0 : getTypeOutputInfo(ranges->typid, &outfunc, &isvarlena);
3048 UIC 0 : fmgr_info(outfunc, &fmgrinfo);
3049 :
3050 EUB : /* deserialize the range info easy-to-process pieces */
3051 UIC 0 : ranges_deserialized = brin_range_deserialize(ranges->maxvalues, ranges);
3052 EUB :
3053 UIC 0 : appendStringInfo(&str, "nranges: %d nvalues: %d maxvalues: %d",
3054 : ranges_deserialized->nranges,
3055 : ranges_deserialized->nvalues,
3056 : ranges_deserialized->maxvalues);
3057 :
3058 EUB : /* serialize ranges */
3059 UBC 0 : idx = 0;
3060 UIC 0 : for (i = 0; i < ranges_deserialized->nranges; i++)
3061 : {
3062 : char *a,
3063 : *b;
3064 : text *c;
3065 : StringInfoData buf;
3066 EUB :
3067 UNC 0 : initStringInfo(&buf);
3068 EUB :
3069 UBC 0 : a = OutputFunctionCall(&fmgrinfo, ranges_deserialized->values[idx++]);
3070 UIC 0 : b = OutputFunctionCall(&fmgrinfo, ranges_deserialized->values[idx++]);
3071 EUB :
3072 UNC 0 : appendStringInfo(&buf, "%s ... %s", a, b);
3073 EUB :
3074 UNC 0 : c = cstring_to_text_with_len(buf.data, buf.len);
3075 EUB :
3076 UIC 0 : astate_values = accumArrayResult(astate_values,
3077 : PointerGetDatum(c),
3078 : false,
3079 : TEXTOID,
3080 : CurrentMemoryContext);
3081 : }
3082 EUB :
3083 UIC 0 : if (ranges_deserialized->nranges > 0)
3084 : {
3085 : Oid typoutput;
3086 : bool typIsVarlena;
3087 : Datum val;
3088 : char *extval;
3089 EUB :
3090 UIC 0 : getTypeOutputInfo(ANYARRAYOID, &typoutput, &typIsVarlena);
3091 EUB :
3092 UNC 0 : val = makeArrayResult(astate_values, CurrentMemoryContext);
3093 EUB :
3094 UIC 0 : extval = OidOutputFunctionCall(typoutput, val);
3095 EUB :
3096 UIC 0 : appendStringInfo(&str, " ranges: %s", extval);
3097 : }
3098 :
3099 EUB : /* serialize individual values */
3100 UIC 0 : astate_values = NULL;
3101 EUB :
3102 UIC 0 : for (i = 0; i < ranges_deserialized->nvalues; i++)
3103 : {
3104 : Datum a;
3105 : text *b;
3106 EUB :
3107 UIC 0 : a = FunctionCall1(&fmgrinfo, ranges_deserialized->values[idx++]);
3108 UNC 0 : b = cstring_to_text(DatumGetCString(a));
3109 :
3110 UBC 0 : astate_values = accumArrayResult(astate_values,
3111 : PointerGetDatum(b),
3112 : false,
3113 : TEXTOID,
3114 : CurrentMemoryContext);
3115 : }
3116 :
3117 0 : if (ranges_deserialized->nvalues > 0)
3118 : {
3119 EUB : Oid typoutput;
3120 : bool typIsVarlena;
3121 : Datum val;
3122 : char *extval;
3123 :
3124 UIC 0 : getTypeOutputInfo(ANYARRAYOID, &typoutput, &typIsVarlena);
3125 :
3126 UNC 0 : val = makeArrayResult(astate_values, CurrentMemoryContext);
3127 EUB :
3128 UIC 0 : extval = OidOutputFunctionCall(typoutput, val);
3129 EUB :
3130 UIC 0 : appendStringInfo(&str, " values: %s", extval);
3131 : }
3132 :
3133 :
3134 0 : appendStringInfoChar(&str, '}');
3135 :
3136 0 : PG_RETURN_CSTRING(str.data);
3137 EUB : }
3138 :
3139 : /*
3140 : * brin_minmax_multi_summary_recv
3141 : * - binary input routine for type brin_minmax_multi_summary.
3142 : */
3143 : Datum
3144 UIC 0 : brin_minmax_multi_summary_recv(PG_FUNCTION_ARGS)
3145 : {
3146 0 : ereport(ERROR,
3147 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
3148 : errmsg("cannot accept a value of type %s", "brin_minmax_multi_summary")));
3149 :
3150 : PG_RETURN_VOID(); /* keep compiler quiet */
3151 : }
3152 :
3153 : /*
3154 EUB : * brin_minmax_multi_summary_send
3155 : * - binary output routine for type brin_minmax_multi_summary.
3156 : *
3157 : * BRIN minmax-multi summaries are serialized in a bytea value (although
3158 : * the type is named differently), so let's just send that.
3159 : */
3160 : Datum
3161 UIC 0 : brin_minmax_multi_summary_send(PG_FUNCTION_ARGS)
3162 : {
3163 0 : return byteasend(fcinfo);
3164 : }
|
