| 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 |
spg_text_config(PG_FUNCTION_ARGS) |
| 97 |
{ |
| 98 |
/* spgConfigIn *cfgin = (spgConfigIn *) PG_GETARG_POINTER(0); */ |
| 99 |
spgConfigOut *cfg = (spgConfigOut *) PG_GETARG_POINTER(1); |
| 100 |
|
| 101 |
cfg->prefixType = TEXTOID; |
| 102 |
cfg->labelType = INT2OID; |
| 103 |
cfg->canReturnData = true; |
| 104 |
cfg->longValuesOK = true; /* suffixing will shorten long values */ |
| 105 |
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 |
formTextDatum(const char *data, int datalen) |
| 114 |
{ |
| 115 |
char *p; |
| 116 |
|
| 117 |
p = (char *) palloc(datalen + VARHDRSZ); |
| 118 |
|
| 119 |
if (datalen + VARHDRSZ_SHORT <= VARATT_SHORT_MAX) |
| 120 |
{ |
| 121 |
SET_VARSIZE_SHORT(p, datalen + VARHDRSZ_SHORT); |
| 122 |
if (datalen) |
| 123 |
memcpy(p + VARHDRSZ_SHORT, data, datalen); |
| 124 |
} |
| 125 |
else |
| 126 |
{ |
| 127 |
SET_VARSIZE(p, datalen + VARHDRSZ); |
| 128 |
memcpy(p + VARHDRSZ, data, datalen); |
| 129 |
} |
| 130 |
|
| 131 |
return PointerGetDatum(p); |
| 132 |
} |
| 133 |
|
| 134 |
/* |
| 135 |
* Find the length of the common prefix of a and b |
| 136 |
*/ |
| 137 |
static int |
| 138 |
commonPrefix(const char *a, const char *b, int lena, int lenb) |
| 139 |
{ |
| 140 |
int i = 0; |
| 141 |
|
| 142 |
while (i < lena && i < lenb && *a == *b) |
| 143 |
{ |
| 144 |
a++; |
| 145 |
b++; |
| 146 |
i++; |
| 147 |
} |
| 148 |
|
| 149 |
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 |
searchChar(Datum *nodeLabels, int nNodes, int16 c, int *i) |
| 159 |
{ |
| 160 |
int StopLow = 0, |
| 161 |
StopHigh = nNodes; |
| 162 |
|
| 163 |
while (StopLow < StopHigh) |
| 164 |
{ |
| 165 |
int StopMiddle = (StopLow + StopHigh) >> 1; |
| 166 |
int16 middle = DatumGetInt16(nodeLabels[StopMiddle]); |
| 167 |
|
| 168 |
if (c < middle) |
| 169 |
StopHigh = StopMiddle; |
| 170 |
else if (c > middle) |
| 171 |
StopLow = StopMiddle + 1; |
| 172 |
else |
| 173 |
{ |
| 174 |
*i = StopMiddle; |
| 175 |
return true; |
| 176 |
} |
| 177 |
} |
| 178 |
|
| 179 |
*i = StopHigh; |
| 180 |
return false; |
| 181 |
} |
| 182 |
|
| 183 |
Datum |
| 184 |
spg_text_choose(PG_FUNCTION_ARGS) |
| 185 |
{ |
| 186 |
spgChooseIn *in = (spgChooseIn *) PG_GETARG_POINTER(0); |
| 187 |
spgChooseOut *out = (spgChooseOut *) PG_GETARG_POINTER(1); |
| 188 |
text *inText = DatumGetTextPP(in->datum); |
| 189 |
char *inStr = VARDATA_ANY(inText); |
| 190 |
int inSize = VARSIZE_ANY_EXHDR(inText); |
| 191 |
char *prefixStr = NULL; |
| 192 |
int prefixSize = 0; |
| 193 |
int commonLen = 0; |
| 194 |
int16 nodeChar = 0; |
| 195 |
int i = 0; |
| 196 |
|
| 197 |
/* Check for prefix match, set nodeChar to first byte after prefix */ |
| 198 |
if (in->hasPrefix) |
| 199 |
{ |
| 200 |
text *prefixText = DatumGetTextPP(in->prefixDatum); |
| 201 |
|
| 202 |
prefixStr = VARDATA_ANY(prefixText); |
| 203 |
prefixSize = VARSIZE_ANY_EXHDR(prefixText); |
| 204 |
|
| 205 |
commonLen = commonPrefix(inStr + in->level, |
| 206 |
prefixStr, |
| 207 |
inSize - in->level, |
| 208 |
prefixSize); |
| 209 |
|
| 210 |
if (commonLen == prefixSize) |
| 211 |
{ |
| 212 |
if (inSize - in->level > commonLen) |
| 213 |
nodeChar = *(unsigned char *) (inStr + in->level + commonLen); |
| 214 |
else |
| 215 |
nodeChar = -1; |
| 216 |
} |
| 217 |
else |
| 218 |
{ |
| 219 |
/* Must split tuple because incoming value doesn't match prefix */ |
| 220 |
out->resultType = spgSplitTuple; |
| 221 |
|
| 222 |
if (commonLen == 0) |
| 223 |
{ |
| 224 |
out->result.splitTuple.prefixHasPrefix = false; |
| 225 |
} |
| 226 |
else |
| 227 |
{ |
| 228 |
out->result.splitTuple.prefixHasPrefix = true; |
| 229 |
out->result.splitTuple.prefixPrefixDatum = |
| 230 |
formTextDatum(prefixStr, commonLen); |
| 231 |
} |
| 232 |
out->result.splitTuple.prefixNNodes = 1; |
| 233 |
out->result.splitTuple.prefixNodeLabels = |
| 234 |
(Datum *) palloc(sizeof(Datum)); |
| 235 |
out->result.splitTuple.prefixNodeLabels[0] = |
| 236 |
Int16GetDatum(*(unsigned char *) (prefixStr + commonLen)); |
| 237 |
|
| 238 |
out->result.splitTuple.childNodeN = 0; |
| 239 |
|
| 240 |
if (prefixSize - commonLen == 1) |
| 241 |
{ |
| 242 |
out->result.splitTuple.postfixHasPrefix = false; |
| 243 |
} |
| 244 |
else |
| 245 |
{ |
| 246 |
out->result.splitTuple.postfixHasPrefix = true; |
| 247 |
out->result.splitTuple.postfixPrefixDatum = |
| 248 |
formTextDatum(prefixStr + commonLen + 1, |
| 249 |
prefixSize - commonLen - 1); |
| 250 |
} |
| 251 |
|
| 252 |
PG_RETURN_VOID(); |
| 253 |
} |
| 254 |
} |
| 255 |
else if (inSize > in->level) |
| 256 |
{ |
| 257 |
nodeChar = *(unsigned char *) (inStr + in->level); |
| 258 |
} |
| 259 |
else |
| 260 |
{ |
| 261 |
nodeChar = -1; |
| 262 |
} |
| 263 |
|
| 264 |
/* Look up nodeChar in the node label array */ |
| 265 |
if (searchChar(in->nodeLabels, in->nNodes, nodeChar, &i)) |
| 266 |
{ |
| 267 |
/* |
| 268 |
* Descend to existing node. (If in->allTheSame, the core code will |
| 269 |
* ignore our nodeN specification here, but that's OK. We still have |
| 270 |
* to provide the correct levelAdd and restDatum values, and those are |
| 271 |
* the same regardless of which node gets chosen by core.) |
| 272 |
*/ |
| 273 |
int levelAdd; |
| 274 |
|
| 275 |
out->resultType = spgMatchNode; |
| 276 |
out->result.matchNode.nodeN = i; |
| 277 |
levelAdd = commonLen; |
| 278 |
if (nodeChar >= 0) |
| 279 |
levelAdd++; |
| 280 |
out->result.matchNode.levelAdd = levelAdd; |
| 281 |
if (inSize - in->level - levelAdd > 0) |
| 282 |
out->result.matchNode.restDatum = |
| 283 |
formTextDatum(inStr + in->level + levelAdd, |
| 284 |
inSize - in->level - levelAdd); |
| 285 |
else |
| 286 |
out->result.matchNode.restDatum = |
| 287 |
formTextDatum(NULL, 0); |
| 288 |
} |
| 289 |
else if (in->allTheSame) |
| 290 |
{ |
| 291 |
/* |
| 292 |
* Can't use AddNode action, so split the tuple. The upper tuple has |
| 293 |
* the same prefix as before and uses a dummy node label -2 for the |
| 294 |
* lower tuple. The lower tuple has no prefix and the same node |
| 295 |
* labels as the original tuple. |
| 296 |
* |
| 297 |
* Note: it might seem tempting to shorten the upper tuple's prefix, |
| 298 |
* if it has one, then use its last byte as label for the lower tuple. |
| 299 |
* But that doesn't win since we know the incoming value matches the |
| 300 |
* whole prefix: we'd just end up splitting the lower tuple again. |
| 301 |
*/ |
| 302 |
out->resultType = spgSplitTuple; |
| 303 |
out->result.splitTuple.prefixHasPrefix = in->hasPrefix; |
| 304 |
out->result.splitTuple.prefixPrefixDatum = in->prefixDatum; |
| 305 |
out->result.splitTuple.prefixNNodes = 1; |
| 306 |
out->result.splitTuple.prefixNodeLabels = (Datum *) palloc(sizeof(Datum)); |
| 307 |
out->result.splitTuple.prefixNodeLabels[0] = Int16GetDatum(-2); |
| 308 |
out->result.splitTuple.childNodeN = 0; |
| 309 |
out->result.splitTuple.postfixHasPrefix = false; |
| 310 |
} |
| 311 |
else |
| 312 |
{ |
| 313 |
/* Add a node for the not-previously-seen nodeChar value */ |
| 314 |
out->resultType = spgAddNode; |
| 315 |
out->result.addNode.nodeLabel = Int16GetDatum(nodeChar); |
| 316 |
out->result.addNode.nodeN = i; |
| 317 |
} |
| 318 |
|
| 319 |
PG_RETURN_VOID(); |
| 320 |
} |
| 321 |
|
| 322 |
/* qsort comparator to sort spgNodePtr structs by "c" */ |
| 323 |
static int |
| 324 |
cmpNodePtr(const void *a, const void *b) |
| 325 |
{ |
| 326 |
const spgNodePtr *aa = (const spgNodePtr *) a; |
| 327 |
const spgNodePtr *bb = (const spgNodePtr *) b; |
| 328 |
|
| 329 |
return pg_cmp_s16(aa->c, bb->c); |
| 330 |
} |
| 331 |
|
| 332 |
Datum |
| 333 |
spg_text_picksplit(PG_FUNCTION_ARGS) |
| 334 |
{ |
| 335 |
spgPickSplitIn *in = (spgPickSplitIn *) PG_GETARG_POINTER(0); |
| 336 |
spgPickSplitOut *out = (spgPickSplitOut *) PG_GETARG_POINTER(1); |
| 337 |
text *text0 = DatumGetTextPP(in->datums[0]); |
| 338 |
int i, |
| 339 |
commonLen; |
| 340 |
spgNodePtr *nodes; |
| 341 |
|
| 342 |
/* Identify longest common prefix, if any */ |
| 343 |
commonLen = VARSIZE_ANY_EXHDR(text0); |
| 344 |
for (i = 1; i < in->nTuples && commonLen > 0; i++) |
| 345 |
{ |
| 346 |
text *texti = DatumGetTextPP(in->datums[i]); |
| 347 |
int tmp = commonPrefix(VARDATA_ANY(text0), |
| 348 |
VARDATA_ANY(texti), |
| 349 |
VARSIZE_ANY_EXHDR(text0), |
| 350 |
VARSIZE_ANY_EXHDR(texti)); |
| 351 |
|
| 352 |
if (tmp < commonLen) |
| 353 |
commonLen = tmp; |
| 354 |
} |
| 355 |
|
| 356 |
/* |
| 357 |
* Limit the prefix length, if necessary, to ensure that the resulting |
| 358 |
* inner tuple will fit on a page. |
| 359 |
*/ |
| 360 |
commonLen = Min(commonLen, SPGIST_MAX_PREFIX_LENGTH); |
| 361 |
|
| 362 |
/* Set node prefix to be that string, if it's not empty */ |
| 363 |
if (commonLen == 0) |
| 364 |
{ |
| 365 |
out->hasPrefix = false; |
| 366 |
} |
| 367 |
else |
| 368 |
{ |
| 369 |
out->hasPrefix = true; |
| 370 |
out->prefixDatum = formTextDatum(VARDATA_ANY(text0), commonLen); |
| 371 |
} |
| 372 |
|
| 373 |
/* Extract the node label (first non-common byte) from each value */ |
| 374 |
nodes = (spgNodePtr *) palloc(sizeof(spgNodePtr) * in->nTuples); |
| 375 |
|
| 376 |
for (i = 0; i < in->nTuples; i++) |
| 377 |
{ |
| 378 |
text *texti = DatumGetTextPP(in->datums[i]); |
| 379 |
|
| 380 |
if (commonLen < VARSIZE_ANY_EXHDR(texti)) |
| 381 |
nodes[i].c = *(unsigned char *) (VARDATA_ANY(texti) + commonLen); |
| 382 |
else |
| 383 |
nodes[i].c = -1; /* use -1 if string is all common */ |
| 384 |
nodes[i].i = i; |
| 385 |
nodes[i].d = in->datums[i]; |
| 386 |
} |
| 387 |
|
| 388 |
/* |
| 389 |
* Sort by label values so that we can group the values into nodes. This |
| 390 |
* also ensures that the nodes are ordered by label value, allowing the |
| 391 |
* use of binary search in searchChar. |
| 392 |
*/ |
| 393 |
qsort(nodes, in->nTuples, sizeof(*nodes), cmpNodePtr); |
| 394 |
|
| 395 |
/* And emit results */ |
| 396 |
out->nNodes = 0; |
| 397 |
out->nodeLabels = (Datum *) palloc(sizeof(Datum) * in->nTuples); |
| 398 |
out->mapTuplesToNodes = (int *) palloc(sizeof(int) * in->nTuples); |
| 399 |
out->leafTupleDatums = (Datum *) palloc(sizeof(Datum) * in->nTuples); |
| 400 |
|
| 401 |
for (i = 0; i < in->nTuples; i++) |
| 402 |
{ |
| 403 |
text *texti = DatumGetTextPP(nodes[i].d); |
| 404 |
Datum leafD; |
| 405 |
|
| 406 |
if (i == 0 || nodes[i].c != nodes[i - 1].c) |
| 407 |
{ |
| 408 |
out->nodeLabels[out->nNodes] = Int16GetDatum(nodes[i].c); |
| 409 |
out->nNodes++; |
| 410 |
} |
| 411 |
|
| 412 |
if (commonLen < VARSIZE_ANY_EXHDR(texti)) |
| 413 |
leafD = formTextDatum(VARDATA_ANY(texti) + commonLen + 1, |
| 414 |
VARSIZE_ANY_EXHDR(texti) - commonLen - 1); |
| 415 |
else |
| 416 |
leafD = formTextDatum(NULL, 0); |
| 417 |
|
| 418 |
out->leafTupleDatums[nodes[i].i] = leafD; |
| 419 |
out->mapTuplesToNodes[nodes[i].i] = out->nNodes - 1; |
| 420 |
} |
| 421 |
|
| 422 |
PG_RETURN_VOID(); |
| 423 |
} |
| 424 |
|
| 425 |
Datum |
| 426 |
spg_text_inner_consistent(PG_FUNCTION_ARGS) |
| 427 |
{ |
| 428 |
spgInnerConsistentIn *in = (spgInnerConsistentIn *) PG_GETARG_POINTER(0); |
| 429 |
spgInnerConsistentOut *out = (spgInnerConsistentOut *) PG_GETARG_POINTER(1); |
| 430 |
bool collate_is_c = pg_newlocale_from_collation(PG_GET_COLLATION())->collate_is_c; |
| 431 |
text *reconstructedValue; |
| 432 |
text *reconstrText; |
| 433 |
int maxReconstrLen; |
| 434 |
text *prefixText = NULL; |
| 435 |
int prefixSize = 0; |
| 436 |
int i; |
| 437 |
|
| 438 |
/* |
| 439 |
* Reconstruct values represented at this tuple, including parent data, |
| 440 |
* prefix of this tuple if any, and the node label if it's non-dummy. |
| 441 |
* in->level should be the length of the previously reconstructed value, |
| 442 |
* and the number of bytes added here is prefixSize or prefixSize + 1. |
| 443 |
* |
| 444 |
* Note: we assume that in->reconstructedValue isn't toasted and doesn't |
| 445 |
* have a short varlena header. This is okay because it must have been |
| 446 |
* created by a previous invocation of this routine, and we always emit |
| 447 |
* long-format reconstructed values. |
| 448 |
*/ |
| 449 |
reconstructedValue = (text *) DatumGetPointer(in->reconstructedValue); |
| 450 |
Assert(reconstructedValue == NULL ? in->level == 0 : |
| 451 |
VARSIZE_ANY_EXHDR(reconstructedValue) == in->level); |
| 452 |
|
| 453 |
maxReconstrLen = in->level + 1; |
| 454 |
if (in->hasPrefix) |
| 455 |
{ |
| 456 |
prefixText = DatumGetTextPP(in->prefixDatum); |
| 457 |
prefixSize = VARSIZE_ANY_EXHDR(prefixText); |
| 458 |
maxReconstrLen += prefixSize; |
| 459 |
} |
| 460 |
|
| 461 |
reconstrText = palloc(VARHDRSZ + maxReconstrLen); |
| 462 |
SET_VARSIZE(reconstrText, VARHDRSZ + maxReconstrLen); |
| 463 |
|
| 464 |
if (in->level) |
| 465 |
memcpy(VARDATA(reconstrText), |
| 466 |
VARDATA(reconstructedValue), |
| 467 |
in->level); |
| 468 |
if (prefixSize) |
| 469 |
memcpy(((char *) VARDATA(reconstrText)) + in->level, |
| 470 |
VARDATA_ANY(prefixText), |
| 471 |
prefixSize); |
| 472 |
/* last byte of reconstrText will be filled in below */ |
| 473 |
|
| 474 |
/* |
| 475 |
* Scan the child nodes. For each one, complete the reconstructed value |
| 476 |
* and see if it's consistent with the query. If so, emit an entry into |
| 477 |
* the output arrays. |
| 478 |
*/ |
| 479 |
out->nodeNumbers = (int *) palloc(sizeof(int) * in->nNodes); |
| 480 |
out->levelAdds = (int *) palloc(sizeof(int) * in->nNodes); |
| 481 |
out->reconstructedValues = (Datum *) palloc(sizeof(Datum) * in->nNodes); |
| 482 |
out->nNodes = 0; |
| 483 |
|
| 484 |
for (i = 0; i < in->nNodes; i++) |
| 485 |
{ |
| 486 |
int16 nodeChar = DatumGetInt16(in->nodeLabels[i]); |
| 487 |
int thisLen; |
| 488 |
bool res = true; |
| 489 |
int j; |
| 490 |
|
| 491 |
/* If nodeChar is a dummy value, don't include it in data */ |
| 492 |
if (nodeChar <= 0) |
| 493 |
thisLen = maxReconstrLen - 1; |
| 494 |
else |
| 495 |
{ |
| 496 |
((unsigned char *) VARDATA(reconstrText))[maxReconstrLen - 1] = nodeChar; |
| 497 |
thisLen = maxReconstrLen; |
| 498 |
} |
| 499 |
|
| 500 |
for (j = 0; j < in->nkeys; j++) |
| 501 |
{ |
| 502 |
StrategyNumber strategy = in->scankeys[j].sk_strategy; |
| 503 |
text *inText; |
| 504 |
int inSize; |
| 505 |
int r; |
| 506 |
|
| 507 |
/* |
| 508 |
* If it's a collation-aware operator, but the collation is C, we |
| 509 |
* can treat it as non-collation-aware. With non-C collation we |
| 510 |
* need to traverse whole tree :-( so there's no point in making |
| 511 |
* any check here. (Note also that our reconstructed value may |
| 512 |
* well end with a partial multibyte character, so that applying |
| 513 |
* any encoding-sensitive test to it would be risky anyhow.) |
| 514 |
*/ |
| 515 |
if (SPG_IS_COLLATION_AWARE_STRATEGY(strategy)) |
| 516 |
{ |
| 517 |
if (collate_is_c) |
| 518 |
strategy -= SPG_STRATEGY_ADDITION; |
| 519 |
else |
| 520 |
continue; |
| 521 |
} |
| 522 |
|
| 523 |
inText = DatumGetTextPP(in->scankeys[j].sk_argument); |
| 524 |
inSize = VARSIZE_ANY_EXHDR(inText); |
| 525 |
|
| 526 |
r = memcmp(VARDATA(reconstrText), VARDATA_ANY(inText), |
| 527 |
Min(inSize, thisLen)); |
| 528 |
|
| 529 |
switch (strategy) |
| 530 |
{ |
| 531 |
case BTLessStrategyNumber: |
| 532 |
case BTLessEqualStrategyNumber: |
| 533 |
if (r > 0) |
| 534 |
res = false; |
| 535 |
break; |
| 536 |
case BTEqualStrategyNumber: |
| 537 |
if (r != 0 || inSize < thisLen) |
| 538 |
res = false; |
| 539 |
break; |
| 540 |
case BTGreaterEqualStrategyNumber: |
| 541 |
case BTGreaterStrategyNumber: |
| 542 |
if (r < 0) |
| 543 |
res = false; |
| 544 |
break; |
| 545 |
case RTPrefixStrategyNumber: |
| 546 |
if (r != 0) |
| 547 |
res = false; |
| 548 |
break; |
| 549 |
default: |
| 550 |
elog(ERROR, "unrecognized strategy number: %d", |
| 551 |
in->scankeys[j].sk_strategy); |
| 552 |
break; |
| 553 |
} |
| 554 |
|
| 555 |
if (!res) |
| 556 |
break; /* no need to consider remaining conditions */ |
| 557 |
} |
| 558 |
|
| 559 |
if (res) |
| 560 |
{ |
| 561 |
out->nodeNumbers[out->nNodes] = i; |
| 562 |
out->levelAdds[out->nNodes] = thisLen - in->level; |
| 563 |
SET_VARSIZE(reconstrText, VARHDRSZ + thisLen); |
| 564 |
out->reconstructedValues[out->nNodes] = |
| 565 |
datumCopy(PointerGetDatum(reconstrText), false, -1); |
| 566 |
out->nNodes++; |
| 567 |
} |
| 568 |
} |
| 569 |
|
| 570 |
PG_RETURN_VOID(); |
| 571 |
} |
| 572 |
|
| 573 |
Datum |
| 574 |
spg_text_leaf_consistent(PG_FUNCTION_ARGS) |
| 575 |
{ |
| 576 |
spgLeafConsistentIn *in = (spgLeafConsistentIn *) PG_GETARG_POINTER(0); |
| 577 |
spgLeafConsistentOut *out = (spgLeafConsistentOut *) PG_GETARG_POINTER(1); |
| 578 |
int level = in->level; |
| 579 |
text *leafValue, |
| 580 |
*reconstrValue = NULL; |
| 581 |
char *fullValue; |
| 582 |
int fullLen; |
| 583 |
bool res; |
| 584 |
int j; |
| 585 |
|
| 586 |
/* all tests are exact */ |
| 587 |
out->recheck = false; |
| 588 |
|
| 589 |
leafValue = DatumGetTextPP(in->leafDatum); |
| 590 |
|
| 591 |
/* As above, in->reconstructedValue isn't toasted or short. */ |
| 592 |
if (DatumGetPointer(in->reconstructedValue)) |
| 593 |
reconstrValue = (text *) DatumGetPointer(in->reconstructedValue); |
| 594 |
|
| 595 |
Assert(reconstrValue == NULL ? level == 0 : |
| 596 |
VARSIZE_ANY_EXHDR(reconstrValue) == level); |
| 597 |
|
| 598 |
/* Reconstruct the full string represented by this leaf tuple */ |
| 599 |
fullLen = level + VARSIZE_ANY_EXHDR(leafValue); |
| 600 |
if (VARSIZE_ANY_EXHDR(leafValue) == 0 && level > 0) |
| 601 |
{ |
| 602 |
fullValue = VARDATA(reconstrValue); |
| 603 |
out->leafValue = PointerGetDatum(reconstrValue); |
| 604 |
} |
| 605 |
else |
| 606 |
{ |
| 607 |
text *fullText = palloc(VARHDRSZ + fullLen); |
| 608 |
|
| 609 |
SET_VARSIZE(fullText, VARHDRSZ + fullLen); |
| 610 |
fullValue = VARDATA(fullText); |
| 611 |
if (level) |
| 612 |
memcpy(fullValue, VARDATA(reconstrValue), level); |
| 613 |
if (VARSIZE_ANY_EXHDR(leafValue) > 0) |
| 614 |
memcpy(fullValue + level, VARDATA_ANY(leafValue), |
| 615 |
VARSIZE_ANY_EXHDR(leafValue)); |
| 616 |
out->leafValue = PointerGetDatum(fullText); |
| 617 |
} |
| 618 |
|
| 619 |
/* Perform the required comparison(s) */ |
| 620 |
res = true; |
| 621 |
for (j = 0; j < in->nkeys; j++) |
| 622 |
{ |
| 623 |
StrategyNumber strategy = in->scankeys[j].sk_strategy; |
| 624 |
text *query = DatumGetTextPP(in->scankeys[j].sk_argument); |
| 625 |
int queryLen = VARSIZE_ANY_EXHDR(query); |
| 626 |
int r; |
| 627 |
|
| 628 |
if (strategy == RTPrefixStrategyNumber) |
| 629 |
{ |
| 630 |
/* |
| 631 |
* if level >= length of query then reconstrValue must begin with |
| 632 |
* query (prefix) string, so we don't need to check it again. |
| 633 |
*/ |
| 634 |
res = (level >= queryLen) || |
| 635 |
DatumGetBool(DirectFunctionCall2Coll(text_starts_with, |
| 636 |
PG_GET_COLLATION(), |
| 637 |
out->leafValue, |
| 638 |
PointerGetDatum(query))); |
| 639 |
|
| 640 |
if (!res) /* no need to consider remaining conditions */ |
| 641 |
break; |
| 642 |
|
| 643 |
continue; |
| 644 |
} |
| 645 |
|
| 646 |
if (SPG_IS_COLLATION_AWARE_STRATEGY(strategy)) |
| 647 |
{ |
| 648 |
/* Collation-aware comparison */ |
| 649 |
strategy -= SPG_STRATEGY_ADDITION; |
| 650 |
|
| 651 |
/* If asserts enabled, verify encoding of reconstructed string */ |
| 652 |
Assert(pg_verifymbstr(fullValue, fullLen, false)); |
| 653 |
|
| 654 |
r = varstr_cmp(fullValue, fullLen, |
| 655 |
VARDATA_ANY(query), queryLen, |
| 656 |
PG_GET_COLLATION()); |
| 657 |
} |
| 658 |
else |
| 659 |
{ |
| 660 |
/* Non-collation-aware comparison */ |
| 661 |
r = memcmp(fullValue, VARDATA_ANY(query), Min(queryLen, fullLen)); |
| 662 |
|
| 663 |
if (r == 0) |
| 664 |
{ |
| 665 |
if (queryLen > fullLen) |
| 666 |
r = -1; |
| 667 |
else if (queryLen < fullLen) |
| 668 |
r = 1; |
| 669 |
} |
| 670 |
} |
| 671 |
|
| 672 |
switch (strategy) |
| 673 |
{ |
| 674 |
case BTLessStrategyNumber: |
| 675 |
res = (r < 0); |
| 676 |
break; |
| 677 |
case BTLessEqualStrategyNumber: |
| 678 |
res = (r <= 0); |
| 679 |
break; |
| 680 |
case BTEqualStrategyNumber: |
| 681 |
res = (r == 0); |
| 682 |
break; |
| 683 |
case BTGreaterEqualStrategyNumber: |
| 684 |
res = (r >= 0); |
| 685 |
break; |
| 686 |
case BTGreaterStrategyNumber: |
| 687 |
res = (r > 0); |
| 688 |
break; |
| 689 |
default: |
| 690 |
elog(ERROR, "unrecognized strategy number: %d", |
| 691 |
in->scankeys[j].sk_strategy); |
| 692 |
res = false; |
| 693 |
break; |
| 694 |
} |
| 695 |
|
| 696 |
if (!res) |
| 697 |
break; /* no need to consider remaining conditions */ |
| 698 |
} |
| 699 |
|
| 700 |
PG_RETURN_BOOL(res); |
| 701 |
} |