Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * spgtextproc.c
4 : * implementation of radix tree (compressed trie) over text
5 : *
6 : * In a text_ops SPGiST index, inner tuples can have a prefix which is the
7 : * common prefix of all strings indexed under that tuple. The node labels
8 : * represent the next byte of the string(s) after the prefix. Assuming we
9 : * always use the longest possible prefix, we will get more than one node
10 : * label unless the prefix length is restricted by SPGIST_MAX_PREFIX_LENGTH.
11 : *
12 : * To reconstruct the indexed string for any index entry, concatenate the
13 : * inner-tuple prefixes and node labels starting at the root and working
14 : * down to the leaf entry, then append the datum in the leaf entry.
15 : * (While descending the tree, "level" is the number of bytes reconstructed
16 : * so far.)
17 : *
18 : * However, there are two special cases for node labels: -1 indicates that
19 : * there are no more bytes after the prefix-so-far, and -2 indicates that we
20 : * had to split an existing allTheSame tuple (in such a case we have to create
21 : * a node label that doesn't correspond to any string byte). In either case,
22 : * the node label does not contribute anything to the reconstructed string.
23 : *
24 : * Previously, we used a node label of zero for both special cases, but
25 : * this was problematic because one can't tell whether a string ending at
26 : * the current level can be pushed down into such a child node. For
27 : * backwards compatibility, we still support such node labels for reading;
28 : * but no new entries will ever be pushed down into a zero-labeled child.
29 : * No new entries ever get pushed into a -2-labeled child, either.
30 : *
31 : *
32 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
33 : * Portions Copyright (c) 1994, Regents of the University of California
34 : *
35 : * IDENTIFICATION
36 : * src/backend/access/spgist/spgtextproc.c
37 : *
38 : *-------------------------------------------------------------------------
39 : */
40 : #include "postgres.h"
41 :
42 : #include "access/spgist.h"
43 : #include "catalog/pg_type.h"
44 : #include "common/int.h"
45 : #include "mb/pg_wchar.h"
46 : #include "utils/datum.h"
47 : #include "utils/fmgrprotos.h"
48 : #include "utils/pg_locale.h"
49 : #include "utils/varlena.h"
50 : #include "varatt.h"
51 :
52 :
53 : /*
54 : * In the worst case, an inner tuple in a text radix tree could have as many
55 : * as 258 nodes (one for each possible byte value, plus the two special
56 : * cases). Each node can take 16 bytes on MAXALIGN=8 machines. The inner
57 : * tuple must fit on an index page of size BLCKSZ. Rather than assuming we
58 : * know the exact amount of overhead imposed by page headers, tuple headers,
59 : * etc, we leave 100 bytes for that (the actual overhead should be no more
60 : * than 56 bytes at this writing, so there is slop in this number).
61 : * So we can safely create prefixes up to BLCKSZ - 258 * 16 - 100 bytes long.
62 : * Unfortunately, because 258 * 16 is over 4K, there is no safe prefix length
63 : * when BLCKSZ is less than 8K; it is always possible to get "SPGiST inner
64 : * tuple size exceeds maximum" if there are too many distinct next-byte values
65 : * at a given place in the tree. Since use of nonstandard block sizes appears
66 : * to be negligible in the field, we just live with that fact for now,
67 : * choosing a max prefix size of 32 bytes when BLCKSZ is configured smaller
68 : * than default.
69 : */
70 : #define SPGIST_MAX_PREFIX_LENGTH Max((int) (BLCKSZ - 258 * 16 - 100), 32)
71 :
72 : /*
73 : * Strategy for collation aware operator on text is equal to btree strategy
74 : * plus value of 10.
75 : *
76 : * Current collation aware strategies and their corresponding btree strategies:
77 : * 11 BTLessStrategyNumber
78 : * 12 BTLessEqualStrategyNumber
79 : * 14 BTGreaterEqualStrategyNumber
80 : * 15 BTGreaterStrategyNumber
81 : */
82 : #define SPG_STRATEGY_ADDITION (10)
83 : #define SPG_IS_COLLATION_AWARE_STRATEGY(s) ((s) > SPG_STRATEGY_ADDITION \
84 : && (s) != RTPrefixStrategyNumber)
85 :
86 : /* Struct for sorting values in picksplit */
87 : typedef struct spgNodePtr
88 : {
89 : Datum d;
90 : int i;
91 : int16 c;
92 : } spgNodePtr;
93 :
94 :
95 : Datum
96 82 : spg_text_config(PG_FUNCTION_ARGS)
97 : {
98 : /* spgConfigIn *cfgin = (spgConfigIn *) PG_GETARG_POINTER(0); */
99 82 : spgConfigOut *cfg = (spgConfigOut *) PG_GETARG_POINTER(1);
100 :
101 82 : cfg->prefixType = TEXTOID;
102 82 : cfg->labelType = INT2OID;
103 82 : cfg->canReturnData = true;
104 82 : cfg->longValuesOK = true; /* suffixing will shorten long values */
105 82 : PG_RETURN_VOID();
106 : }
107 :
108 : /*
109 : * Form a text datum from the given not-necessarily-null-terminated string,
110 : * using short varlena header format if possible
111 : */
112 : static Datum
113 257944 : formTextDatum(const char *data, int datalen)
114 : {
115 : char *p;
116 :
117 257944 : p = (char *) palloc(datalen + VARHDRSZ);
118 :
119 257944 : if (datalen + VARHDRSZ_SHORT <= VARATT_SHORT_MAX)
120 : {
121 257944 : SET_VARSIZE_SHORT(p, datalen + VARHDRSZ_SHORT);
122 257944 : if (datalen)
123 242812 : memcpy(p + VARHDRSZ_SHORT, data, datalen);
124 : }
125 : else
126 : {
127 0 : SET_VARSIZE(p, datalen + VARHDRSZ);
128 0 : memcpy(p + VARHDRSZ, data, datalen);
129 : }
130 :
131 257944 : return PointerGetDatum(p);
132 : }
133 :
134 : /*
135 : * Find the length of the common prefix of a and b
136 : */
137 : static int
138 95118 : commonPrefix(const char *a, const char *b, int lena, int lenb)
139 : {
140 95118 : int i = 0;
141 :
142 6638562 : while (i < lena && i < lenb && *a == *b)
143 : {
144 6543444 : a++;
145 6543444 : b++;
146 6543444 : i++;
147 : }
148 :
149 95118 : return i;
150 : }
151 :
152 : /*
153 : * Binary search an array of int16 datums for a match to c
154 : *
155 : * On success, *i gets the match location; on failure, it gets where to insert
156 : */
157 : static bool
158 208724 : searchChar(const Datum *nodeLabels, int nNodes, int16 c, int *i)
159 : {
160 208724 : int StopLow = 0,
161 208724 : StopHigh = nNodes;
162 :
163 573066 : while (StopLow < StopHigh)
164 : {
165 571700 : int StopMiddle = (StopLow + StopHigh) >> 1;
166 571700 : int16 middle = DatumGetInt16(nodeLabels[StopMiddle]);
167 :
168 571700 : if (c < middle)
169 180908 : StopHigh = StopMiddle;
170 390792 : else if (c > middle)
171 183434 : StopLow = StopMiddle + 1;
172 : else
173 : {
174 207358 : *i = StopMiddle;
175 207358 : return true;
176 : }
177 : }
178 :
179 1366 : *i = StopHigh;
180 1366 : return false;
181 : }
182 :
183 : Datum
184 209352 : spg_text_choose(PG_FUNCTION_ARGS)
185 : {
186 209352 : spgChooseIn *in = (spgChooseIn *) PG_GETARG_POINTER(0);
187 209352 : spgChooseOut *out = (spgChooseOut *) PG_GETARG_POINTER(1);
188 209352 : text *inText = DatumGetTextPP(in->datum);
189 209352 : char *inStr = VARDATA_ANY(inText);
190 209352 : int inSize = VARSIZE_ANY_EXHDR(inText);
191 209352 : char *prefixStr = NULL;
192 209352 : int prefixSize = 0;
193 209352 : int commonLen = 0;
194 209352 : int16 nodeChar = 0;
195 209352 : int i = 0;
196 :
197 : /* Check for prefix match, set nodeChar to first byte after prefix */
198 209352 : if (in->hasPrefix)
199 : {
200 82838 : text *prefixText = DatumGetTextPP(in->prefixDatum);
201 :
202 82838 : prefixStr = VARDATA_ANY(prefixText);
203 82838 : prefixSize = VARSIZE_ANY_EXHDR(prefixText);
204 :
205 82838 : commonLen = commonPrefix(inStr + in->level,
206 : prefixStr,
207 82838 : inSize - in->level,
208 : prefixSize);
209 :
210 82838 : if (commonLen == prefixSize)
211 : {
212 82210 : if (inSize - in->level > commonLen)
213 76054 : nodeChar = *(unsigned char *) (inStr + in->level + commonLen);
214 : else
215 6156 : nodeChar = -1;
216 : }
217 : else
218 : {
219 : /* Must split tuple because incoming value doesn't match prefix */
220 628 : out->resultType = spgSplitTuple;
221 :
222 628 : if (commonLen == 0)
223 : {
224 22 : out->result.splitTuple.prefixHasPrefix = false;
225 : }
226 : else
227 : {
228 606 : out->result.splitTuple.prefixHasPrefix = true;
229 606 : out->result.splitTuple.prefixPrefixDatum =
230 606 : formTextDatum(prefixStr, commonLen);
231 : }
232 628 : out->result.splitTuple.prefixNNodes = 1;
233 628 : out->result.splitTuple.prefixNodeLabels = palloc_object(Datum);
234 1256 : out->result.splitTuple.prefixNodeLabels[0] =
235 628 : Int16GetDatum(*(unsigned char *) (prefixStr + commonLen));
236 :
237 628 : out->result.splitTuple.childNodeN = 0;
238 :
239 628 : if (prefixSize - commonLen == 1)
240 : {
241 616 : out->result.splitTuple.postfixHasPrefix = false;
242 : }
243 : else
244 : {
245 12 : out->result.splitTuple.postfixHasPrefix = true;
246 12 : out->result.splitTuple.postfixPrefixDatum =
247 12 : formTextDatum(prefixStr + commonLen + 1,
248 12 : prefixSize - commonLen - 1);
249 : }
250 :
251 628 : PG_RETURN_VOID();
252 : }
253 : }
254 126514 : else if (inSize > in->level)
255 : {
256 125528 : nodeChar = *(unsigned char *) (inStr + in->level);
257 : }
258 : else
259 : {
260 986 : nodeChar = -1;
261 : }
262 :
263 : /* Look up nodeChar in the node label array */
264 208724 : if (searchChar(in->nodeLabels, in->nNodes, nodeChar, &i))
265 : {
266 : /*
267 : * Descend to existing node. (If in->allTheSame, the core code will
268 : * ignore our nodeN specification here, but that's OK. We still have
269 : * to provide the correct levelAdd and restDatum values, and those are
270 : * the same regardless of which node gets chosen by core.)
271 : */
272 : int levelAdd;
273 :
274 207358 : out->resultType = spgMatchNode;
275 207358 : out->result.matchNode.nodeN = i;
276 207358 : levelAdd = commonLen;
277 207358 : if (nodeChar >= 0)
278 200222 : levelAdd++;
279 207358 : out->result.matchNode.levelAdd = levelAdd;
280 207358 : if (inSize - in->level - levelAdd > 0)
281 200216 : out->result.matchNode.restDatum =
282 200216 : formTextDatum(inStr + in->level + levelAdd,
283 200216 : inSize - in->level - levelAdd);
284 : else
285 7142 : out->result.matchNode.restDatum =
286 7142 : formTextDatum(NULL, 0);
287 : }
288 1366 : else if (in->allTheSame)
289 : {
290 : /*
291 : * Can't use AddNode action, so split the tuple. The upper tuple has
292 : * the same prefix as before and uses a dummy node label -2 for the
293 : * lower tuple. The lower tuple has no prefix and the same node
294 : * labels as the original tuple.
295 : *
296 : * Note: it might seem tempting to shorten the upper tuple's prefix,
297 : * if it has one, then use its last byte as label for the lower tuple.
298 : * But that doesn't win since we know the incoming value matches the
299 : * whole prefix: we'd just end up splitting the lower tuple again.
300 : */
301 6 : out->resultType = spgSplitTuple;
302 6 : out->result.splitTuple.prefixHasPrefix = in->hasPrefix;
303 6 : out->result.splitTuple.prefixPrefixDatum = in->prefixDatum;
304 6 : out->result.splitTuple.prefixNNodes = 1;
305 6 : out->result.splitTuple.prefixNodeLabels = palloc_object(Datum);
306 6 : out->result.splitTuple.prefixNodeLabels[0] = Int16GetDatum(-2);
307 6 : out->result.splitTuple.childNodeN = 0;
308 6 : out->result.splitTuple.postfixHasPrefix = false;
309 : }
310 : else
311 : {
312 : /* Add a node for the not-previously-seen nodeChar value */
313 1360 : out->resultType = spgAddNode;
314 1360 : out->result.addNode.nodeLabel = Int16GetDatum(nodeChar);
315 1360 : out->result.addNode.nodeN = i;
316 : }
317 :
318 208724 : PG_RETURN_VOID();
319 : }
320 :
321 : /* qsort comparator to sort spgNodePtr structs by "c" */
322 : static int
323 114432 : cmpNodePtr(const void *a, const void *b)
324 : {
325 114432 : const spgNodePtr *aa = (const spgNodePtr *) a;
326 114432 : const spgNodePtr *bb = (const spgNodePtr *) b;
327 :
328 114432 : return pg_cmp_s16(aa->c, bb->c);
329 : }
330 :
331 : Datum
332 514 : spg_text_picksplit(PG_FUNCTION_ARGS)
333 : {
334 514 : spgPickSplitIn *in = (spgPickSplitIn *) PG_GETARG_POINTER(0);
335 514 : spgPickSplitOut *out = (spgPickSplitOut *) PG_GETARG_POINTER(1);
336 514 : text *text0 = DatumGetTextPP(in->datums[0]);
337 : int i,
338 : commonLen;
339 : spgNodePtr *nodes;
340 :
341 : /* Identify longest common prefix, if any */
342 514 : commonLen = VARSIZE_ANY_EXHDR(text0);
343 12794 : for (i = 1; i < in->nTuples && commonLen > 0; i++)
344 : {
345 12280 : text *texti = DatumGetTextPP(in->datums[i]);
346 12280 : int tmp = commonPrefix(VARDATA_ANY(text0),
347 12280 : VARDATA_ANY(texti),
348 12280 : VARSIZE_ANY_EXHDR(text0),
349 12280 : VARSIZE_ANY_EXHDR(texti));
350 :
351 12280 : if (tmp < commonLen)
352 416 : commonLen = tmp;
353 : }
354 :
355 : /*
356 : * Limit the prefix length, if necessary, to ensure that the resulting
357 : * inner tuple will fit on a page.
358 : */
359 514 : commonLen = Min(commonLen, SPGIST_MAX_PREFIX_LENGTH);
360 :
361 : /* Set node prefix to be that string, if it's not empty */
362 514 : if (commonLen == 0)
363 : {
364 424 : out->hasPrefix = false;
365 : }
366 : else
367 : {
368 90 : out->hasPrefix = true;
369 90 : out->prefixDatum = formTextDatum(VARDATA_ANY(text0), commonLen);
370 : }
371 :
372 : /* Extract the node label (first non-common byte) from each value */
373 514 : nodes = palloc_array(spgNodePtr, in->nTuples);
374 :
375 50392 : for (i = 0; i < in->nTuples; i++)
376 : {
377 49878 : text *texti = DatumGetTextPP(in->datums[i]);
378 :
379 49878 : if (commonLen < VARSIZE_ANY_EXHDR(texti))
380 43624 : nodes[i].c = *(unsigned char *) (VARDATA_ANY(texti) + commonLen);
381 : else
382 6254 : nodes[i].c = -1; /* use -1 if string is all common */
383 49878 : nodes[i].i = i;
384 49878 : nodes[i].d = in->datums[i];
385 : }
386 :
387 : /*
388 : * Sort by label values so that we can group the values into nodes. This
389 : * also ensures that the nodes are ordered by label value, allowing the
390 : * use of binary search in searchChar.
391 : */
392 514 : qsort(nodes, in->nTuples, sizeof(*nodes), cmpNodePtr);
393 :
394 : /* And emit results */
395 514 : out->nNodes = 0;
396 514 : out->nodeLabels = palloc_array(Datum, in->nTuples);
397 514 : out->mapTuplesToNodes = palloc_array(int, in->nTuples);
398 514 : out->leafTupleDatums = palloc_array(Datum, in->nTuples);
399 :
400 50392 : for (i = 0; i < in->nTuples; i++)
401 : {
402 49878 : text *texti = DatumGetTextPP(nodes[i].d);
403 : Datum leafD;
404 :
405 49878 : if (i == 0 || nodes[i].c != nodes[i - 1].c)
406 : {
407 3190 : out->nodeLabels[out->nNodes] = Int16GetDatum(nodes[i].c);
408 3190 : out->nNodes++;
409 : }
410 :
411 49878 : if (commonLen < VARSIZE_ANY_EXHDR(texti))
412 43624 : leafD = formTextDatum(VARDATA_ANY(texti) + commonLen + 1,
413 43624 : VARSIZE_ANY_EXHDR(texti) - commonLen - 1);
414 : else
415 6254 : leafD = formTextDatum(NULL, 0);
416 :
417 49878 : out->leafTupleDatums[nodes[i].i] = leafD;
418 49878 : out->mapTuplesToNodes[nodes[i].i] = out->nNodes - 1;
419 : }
420 :
421 514 : PG_RETURN_VOID();
422 : }
423 :
424 : Datum
425 1732 : spg_text_inner_consistent(PG_FUNCTION_ARGS)
426 : {
427 1732 : spgInnerConsistentIn *in = (spgInnerConsistentIn *) PG_GETARG_POINTER(0);
428 1732 : spgInnerConsistentOut *out = (spgInnerConsistentOut *) PG_GETARG_POINTER(1);
429 1732 : bool collate_is_c = pg_newlocale_from_collation(PG_GET_COLLATION())->collate_is_c;
430 : text *reconstructedValue;
431 : text *reconstrText;
432 : int maxReconstrLen;
433 1732 : text *prefixText = NULL;
434 1732 : int prefixSize = 0;
435 : int i;
436 :
437 : /*
438 : * Reconstruct values represented at this tuple, including parent data,
439 : * prefix of this tuple if any, and the node label if it's non-dummy.
440 : * in->level should be the length of the previously reconstructed value,
441 : * and the number of bytes added here is prefixSize or prefixSize + 1.
442 : *
443 : * Note: we assume that in->reconstructedValue isn't toasted and doesn't
444 : * have a short varlena header. This is okay because it must have been
445 : * created by a previous invocation of this routine, and we always emit
446 : * long-format reconstructed values.
447 : */
448 1732 : reconstructedValue = (text *) DatumGetPointer(in->reconstructedValue);
449 : Assert(reconstructedValue == NULL ? in->level == 0 :
450 : VARSIZE_ANY_EXHDR(reconstructedValue) == in->level);
451 :
452 1732 : maxReconstrLen = in->level + 1;
453 1732 : if (in->hasPrefix)
454 : {
455 324 : prefixText = DatumGetTextPP(in->prefixDatum);
456 324 : prefixSize = VARSIZE_ANY_EXHDR(prefixText);
457 324 : maxReconstrLen += prefixSize;
458 : }
459 :
460 1732 : reconstrText = palloc(VARHDRSZ + maxReconstrLen);
461 1732 : SET_VARSIZE(reconstrText, VARHDRSZ + maxReconstrLen);
462 :
463 1732 : if (in->level)
464 1552 : memcpy(VARDATA(reconstrText),
465 1552 : VARDATA(reconstructedValue),
466 1552 : in->level);
467 1732 : if (prefixSize)
468 324 : memcpy(((char *) VARDATA(reconstrText)) + in->level,
469 324 : VARDATA_ANY(prefixText),
470 : prefixSize);
471 : /* last byte of reconstrText will be filled in below */
472 :
473 : /*
474 : * Scan the child nodes. For each one, complete the reconstructed value
475 : * and see if it's consistent with the query. If so, emit an entry into
476 : * the output arrays.
477 : */
478 1732 : out->nodeNumbers = palloc_array(int, in->nNodes);
479 1732 : out->levelAdds = palloc_array(int, in->nNodes);
480 1732 : out->reconstructedValues = palloc_array(Datum, in->nNodes);
481 1732 : out->nNodes = 0;
482 :
483 18304 : for (i = 0; i < in->nNodes; i++)
484 : {
485 16572 : int16 nodeChar = DatumGetInt16(in->nodeLabels[i]);
486 : int thisLen;
487 16572 : bool res = true;
488 : int j;
489 :
490 : /* If nodeChar is a dummy value, don't include it in data */
491 16572 : if (nodeChar <= 0)
492 3324 : thisLen = maxReconstrLen - 1;
493 : else
494 : {
495 13248 : ((unsigned char *) VARDATA(reconstrText))[maxReconstrLen - 1] = nodeChar;
496 13248 : thisLen = maxReconstrLen;
497 : }
498 :
499 28644 : for (j = 0; j < in->nkeys; j++)
500 : {
501 16572 : StrategyNumber strategy = in->scankeys[j].sk_strategy;
502 : text *inText;
503 : int inSize;
504 : int r;
505 :
506 : /*
507 : * If it's a collation-aware operator, but the collation is C, we
508 : * can treat it as non-collation-aware. With non-C collation we
509 : * need to traverse whole tree :-( so there's no point in making
510 : * any check here. (Note also that our reconstructed value may
511 : * well end with a partial multibyte character, so that applying
512 : * any encoding-sensitive test to it would be risky anyhow.)
513 : */
514 16572 : if (SPG_IS_COLLATION_AWARE_STRATEGY(strategy))
515 : {
516 10416 : if (collate_is_c)
517 624 : strategy -= SPG_STRATEGY_ADDITION;
518 : else
519 9792 : continue;
520 : }
521 :
522 6780 : inText = DatumGetTextPP(in->scankeys[j].sk_argument);
523 6780 : inSize = VARSIZE_ANY_EXHDR(inText);
524 :
525 6780 : r = memcmp(VARDATA(reconstrText), VARDATA_ANY(inText),
526 6780 : Min(inSize, thisLen));
527 :
528 6780 : switch (strategy)
529 : {
530 1408 : case BTLessStrategyNumber:
531 : case BTLessEqualStrategyNumber:
532 1408 : if (r > 0)
533 800 : res = false;
534 1408 : break;
535 3468 : case BTEqualStrategyNumber:
536 3468 : if (r != 0 || inSize < thisLen)
537 2100 : res = false;
538 3468 : break;
539 1088 : case BTGreaterEqualStrategyNumber:
540 : case BTGreaterStrategyNumber:
541 1088 : if (r < 0)
542 832 : res = false;
543 1088 : break;
544 816 : case RTPrefixStrategyNumber:
545 816 : if (r != 0)
546 768 : res = false;
547 816 : break;
548 0 : default:
549 0 : elog(ERROR, "unrecognized strategy number: %d",
550 : in->scankeys[j].sk_strategy);
551 : break;
552 : }
553 :
554 6780 : if (!res)
555 4500 : break; /* no need to consider remaining conditions */
556 : }
557 :
558 16572 : if (res)
559 : {
560 12072 : out->nodeNumbers[out->nNodes] = i;
561 12072 : out->levelAdds[out->nNodes] = thisLen - in->level;
562 12072 : SET_VARSIZE(reconstrText, VARHDRSZ + thisLen);
563 24144 : out->reconstructedValues[out->nNodes] =
564 12072 : datumCopy(PointerGetDatum(reconstrText), false, -1);
565 12072 : out->nNodes++;
566 : }
567 : }
568 :
569 1732 : PG_RETURN_VOID();
570 : }
571 :
572 : Datum
573 235500 : spg_text_leaf_consistent(PG_FUNCTION_ARGS)
574 : {
575 235500 : spgLeafConsistentIn *in = (spgLeafConsistentIn *) PG_GETARG_POINTER(0);
576 235500 : spgLeafConsistentOut *out = (spgLeafConsistentOut *) PG_GETARG_POINTER(1);
577 235500 : int level = in->level;
578 : text *leafValue,
579 235500 : *reconstrValue = NULL;
580 : char *fullValue;
581 : int fullLen;
582 : bool res;
583 : int j;
584 :
585 : /* all tests are exact */
586 235500 : out->recheck = false;
587 :
588 235500 : leafValue = DatumGetTextPP(in->leafDatum);
589 :
590 : /* As above, in->reconstructedValue isn't toasted or short. */
591 235500 : if (DatumGetPointer(in->reconstructedValue))
592 235476 : reconstrValue = (text *) DatumGetPointer(in->reconstructedValue);
593 :
594 : Assert(reconstrValue == NULL ? level == 0 :
595 : VARSIZE_ANY_EXHDR(reconstrValue) == level);
596 :
597 : /* Reconstruct the full string represented by this leaf tuple */
598 235500 : fullLen = level + VARSIZE_ANY_EXHDR(leafValue);
599 235500 : if (VARSIZE_ANY_EXHDR(leafValue) == 0 && level > 0)
600 : {
601 74352 : fullValue = VARDATA(reconstrValue);
602 74352 : out->leafValue = PointerGetDatum(reconstrValue);
603 : }
604 : else
605 : {
606 161148 : text *fullText = palloc(VARHDRSZ + fullLen);
607 :
608 161148 : SET_VARSIZE(fullText, VARHDRSZ + fullLen);
609 161148 : fullValue = VARDATA(fullText);
610 161148 : if (level)
611 161124 : memcpy(fullValue, VARDATA(reconstrValue), level);
612 161148 : if (VARSIZE_ANY_EXHDR(leafValue) > 0)
613 161148 : memcpy(fullValue + level, VARDATA_ANY(leafValue),
614 : VARSIZE_ANY_EXHDR(leafValue));
615 161148 : out->leafValue = PointerGetDatum(fullText);
616 : }
617 :
618 : /* Perform the required comparison(s) */
619 235500 : res = true;
620 263046 : for (j = 0; j < in->nkeys; j++)
621 : {
622 235500 : StrategyNumber strategy = in->scankeys[j].sk_strategy;
623 235500 : text *query = DatumGetTextPP(in->scankeys[j].sk_argument);
624 235500 : int queryLen = VARSIZE_ANY_EXHDR(query);
625 : int r;
626 :
627 235500 : if (strategy == RTPrefixStrategyNumber)
628 : {
629 : /*
630 : * if level >= length of query then reconstrValue must begin with
631 : * query (prefix) string, so we don't need to check it again.
632 : */
633 768 : res = (level >= queryLen) ||
634 384 : DatumGetBool(DirectFunctionCall2Coll(text_starts_with,
635 : PG_GET_COLLATION(),
636 : out->leafValue,
637 : PointerGetDatum(query)));
638 :
639 384 : if (!res) /* no need to consider remaining conditions */
640 336 : break;
641 :
642 48 : continue;
643 : }
644 :
645 235116 : if (SPG_IS_COLLATION_AWARE_STRATEGY(strategy))
646 : {
647 : /* Collation-aware comparison */
648 202728 : strategy -= SPG_STRATEGY_ADDITION;
649 :
650 : /* If asserts enabled, verify encoding of reconstructed string */
651 : Assert(pg_verifymbstr(fullValue, fullLen, false));
652 :
653 202728 : r = varstr_cmp(fullValue, fullLen,
654 202728 : VARDATA_ANY(query), queryLen,
655 : PG_GET_COLLATION());
656 : }
657 : else
658 : {
659 : /* Non-collation-aware comparison */
660 32388 : r = memcmp(fullValue, VARDATA_ANY(query), Min(queryLen, fullLen));
661 :
662 32388 : if (r == 0)
663 : {
664 24162 : if (queryLen > fullLen)
665 12024 : r = -1;
666 12138 : else if (queryLen < fullLen)
667 0 : r = 1;
668 : }
669 : }
670 :
671 235116 : switch (strategy)
672 : {
673 54376 : case BTLessStrategyNumber:
674 54376 : res = (r < 0);
675 54376 : break;
676 54376 : case BTLessEqualStrategyNumber:
677 54376 : res = (r <= 0);
678 54376 : break;
679 24300 : case BTEqualStrategyNumber:
680 24300 : res = (r == 0);
681 24300 : break;
682 51032 : case BTGreaterEqualStrategyNumber:
683 51032 : res = (r >= 0);
684 51032 : break;
685 51032 : case BTGreaterStrategyNumber:
686 51032 : res = (r > 0);
687 51032 : break;
688 0 : default:
689 0 : elog(ERROR, "unrecognized strategy number: %d",
690 : in->scankeys[j].sk_strategy);
691 : res = false;
692 : break;
693 : }
694 :
695 235116 : if (!res)
696 207618 : break; /* no need to consider remaining conditions */
697 : }
698 :
699 235500 : PG_RETURN_BOOL(res);
700 : }
|
