| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* hashfunc.c |
| 4 |
* Support functions for hash access method. |
| 5 |
* |
| 6 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 7 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 8 |
* |
| 9 |
* |
| 10 |
* IDENTIFICATION |
| 11 |
* src/backend/access/hash/hashfunc.c |
| 12 |
* |
| 13 |
* NOTES |
| 14 |
* These functions are stored in pg_amproc. For each operator class |
| 15 |
* defined for hash indexes, they compute the hash value of the argument. |
| 16 |
* |
| 17 |
* Additional hash functions appear in /utils/adt/ files for various |
| 18 |
* specialized datatypes. |
| 19 |
* |
| 20 |
* It is expected that every bit of a hash function's 32-bit result is |
| 21 |
* as random as every other; failure to ensure this is likely to lead |
| 22 |
* to poor performance of hash joins, for example. In most cases a hash |
| 23 |
* function should use hash_any() or its variant hash_uint32(). |
| 24 |
*------------------------------------------------------------------------- |
| 25 |
*/ |
| 26 |
|
| 27 |
#include "postgres.h" |
| 28 |
|
| 29 |
#include "common/hashfn.h" |
| 30 |
#include "utils/builtins.h" |
| 31 |
#include "utils/float.h" |
| 32 |
#include "utils/fmgrprotos.h" |
| 33 |
#include "utils/pg_locale.h" |
| 34 |
#include "varatt.h" |
| 35 |
|
| 36 |
/* |
| 37 |
* Datatype-specific hash functions. |
| 38 |
* |
| 39 |
* These support both hash indexes and hash joins. |
| 40 |
* |
| 41 |
* NOTE: some of these are also used by catcache operations, without |
| 42 |
* any direct connection to hash indexes. Also, the common hash_any |
| 43 |
* routine is also used by dynahash tables. |
| 44 |
*/ |
| 45 |
|
| 46 |
/* Note: this is used for both "char" and boolean datatypes */ |
| 47 |
Datum |
| 48 |
hashchar(PG_FUNCTION_ARGS) |
| 49 |
{ |
| 50 |
return hash_uint32((int32) PG_GETARG_CHAR(0)); |
| 51 |
} |
| 52 |
|
| 53 |
Datum |
| 54 |
hashcharextended(PG_FUNCTION_ARGS) |
| 55 |
{ |
| 56 |
return hash_uint32_extended((int32) PG_GETARG_CHAR(0), PG_GETARG_INT64(1)); |
| 57 |
} |
| 58 |
|
| 59 |
Datum |
| 60 |
hashint2(PG_FUNCTION_ARGS) |
| 61 |
{ |
| 62 |
return hash_uint32((int32) PG_GETARG_INT16(0)); |
| 63 |
} |
| 64 |
|
| 65 |
Datum |
| 66 |
hashint2extended(PG_FUNCTION_ARGS) |
| 67 |
{ |
| 68 |
return hash_uint32_extended((int32) PG_GETARG_INT16(0), PG_GETARG_INT64(1)); |
| 69 |
} |
| 70 |
|
| 71 |
Datum |
| 72 |
hashint4(PG_FUNCTION_ARGS) |
| 73 |
{ |
| 74 |
return hash_uint32(PG_GETARG_INT32(0)); |
| 75 |
} |
| 76 |
|
| 77 |
Datum |
| 78 |
hashint4extended(PG_FUNCTION_ARGS) |
| 79 |
{ |
| 80 |
return hash_uint32_extended(PG_GETARG_INT32(0), PG_GETARG_INT64(1)); |
| 81 |
} |
| 82 |
|
| 83 |
Datum |
| 84 |
hashint8(PG_FUNCTION_ARGS) |
| 85 |
{ |
| 86 |
/* |
| 87 |
* The idea here is to produce a hash value compatible with the values |
| 88 |
* produced by hashint4 and hashint2 for logically equal inputs; this is |
| 89 |
* necessary to support cross-type hash joins across these input types. |
| 90 |
* Since all three types are signed, we can xor the high half of the int8 |
| 91 |
* value if the sign is positive, or the complement of the high half when |
| 92 |
* the sign is negative. |
| 93 |
*/ |
| 94 |
int64 val = PG_GETARG_INT64(0); |
| 95 |
uint32 lohalf = (uint32) val; |
| 96 |
uint32 hihalf = (uint32) (val >> 32); |
| 97 |
|
| 98 |
lohalf ^= (val >= 0) ? hihalf : ~hihalf; |
| 99 |
|
| 100 |
return hash_uint32(lohalf); |
| 101 |
} |
| 102 |
|
| 103 |
Datum |
| 104 |
hashint8extended(PG_FUNCTION_ARGS) |
| 105 |
{ |
| 106 |
/* Same approach as hashint8 */ |
| 107 |
int64 val = PG_GETARG_INT64(0); |
| 108 |
uint32 lohalf = (uint32) val; |
| 109 |
uint32 hihalf = (uint32) (val >> 32); |
| 110 |
|
| 111 |
lohalf ^= (val >= 0) ? hihalf : ~hihalf; |
| 112 |
|
| 113 |
return hash_uint32_extended(lohalf, PG_GETARG_INT64(1)); |
| 114 |
} |
| 115 |
|
| 116 |
Datum |
| 117 |
hashoid(PG_FUNCTION_ARGS) |
| 118 |
{ |
| 119 |
return hash_uint32((uint32) PG_GETARG_OID(0)); |
| 120 |
} |
| 121 |
|
| 122 |
Datum |
| 123 |
hashoidextended(PG_FUNCTION_ARGS) |
| 124 |
{ |
| 125 |
return hash_uint32_extended((uint32) PG_GETARG_OID(0), PG_GETARG_INT64(1)); |
| 126 |
} |
| 127 |
|
| 128 |
Datum |
| 129 |
hashenum(PG_FUNCTION_ARGS) |
| 130 |
{ |
| 131 |
return hash_uint32((uint32) PG_GETARG_OID(0)); |
| 132 |
} |
| 133 |
|
| 134 |
Datum |
| 135 |
hashenumextended(PG_FUNCTION_ARGS) |
| 136 |
{ |
| 137 |
return hash_uint32_extended((uint32) PG_GETARG_OID(0), PG_GETARG_INT64(1)); |
| 138 |
} |
| 139 |
|
| 140 |
Datum |
| 141 |
hashfloat4(PG_FUNCTION_ARGS) |
| 142 |
{ |
| 143 |
float4 key = PG_GETARG_FLOAT4(0); |
| 144 |
float8 key8; |
| 145 |
|
| 146 |
/* |
| 147 |
* On IEEE-float machines, minus zero and zero have different bit patterns |
| 148 |
* but should compare as equal. We must ensure that they have the same |
| 149 |
* hash value, which is most reliably done this way: |
| 150 |
*/ |
| 151 |
if (key == (float4) 0) |
| 152 |
PG_RETURN_UINT32(0); |
| 153 |
|
| 154 |
/* |
| 155 |
* To support cross-type hashing of float8 and float4, we want to return |
| 156 |
* the same hash value hashfloat8 would produce for an equal float8 value. |
| 157 |
* So, widen the value to float8 and hash that. (We must do this rather |
| 158 |
* than have hashfloat8 try to narrow its value to float4; that could fail |
| 159 |
* on overflow.) |
| 160 |
*/ |
| 161 |
key8 = key; |
| 162 |
|
| 163 |
/* |
| 164 |
* Similarly, NaNs can have different bit patterns but they should all |
| 165 |
* compare as equal. For backwards-compatibility reasons we force them to |
| 166 |
* have the hash value of a standard float8 NaN. (You'd think we could |
| 167 |
* replace key with a float4 NaN and then widen it; but on some old |
| 168 |
* platforms, that way produces a different bit pattern.) |
| 169 |
*/ |
| 170 |
if (isnan(key8)) |
| 171 |
key8 = get_float8_nan(); |
| 172 |
|
| 173 |
return hash_any((unsigned char *) &key8, sizeof(key8)); |
| 174 |
} |
| 175 |
|
| 176 |
Datum |
| 177 |
hashfloat4extended(PG_FUNCTION_ARGS) |
| 178 |
{ |
| 179 |
float4 key = PG_GETARG_FLOAT4(0); |
| 180 |
uint64 seed = PG_GETARG_INT64(1); |
| 181 |
float8 key8; |
| 182 |
|
| 183 |
/* Same approach as hashfloat4 */ |
| 184 |
if (key == (float4) 0) |
| 185 |
PG_RETURN_UINT64(seed); |
| 186 |
key8 = key; |
| 187 |
if (isnan(key8)) |
| 188 |
key8 = get_float8_nan(); |
| 189 |
|
| 190 |
return hash_any_extended((unsigned char *) &key8, sizeof(key8), seed); |
| 191 |
} |
| 192 |
|
| 193 |
Datum |
| 194 |
hashfloat8(PG_FUNCTION_ARGS) |
| 195 |
{ |
| 196 |
float8 key = PG_GETARG_FLOAT8(0); |
| 197 |
|
| 198 |
/* |
| 199 |
* On IEEE-float machines, minus zero and zero have different bit patterns |
| 200 |
* but should compare as equal. We must ensure that they have the same |
| 201 |
* hash value, which is most reliably done this way: |
| 202 |
*/ |
| 203 |
if (key == (float8) 0) |
| 204 |
PG_RETURN_UINT32(0); |
| 205 |
|
| 206 |
/* |
| 207 |
* Similarly, NaNs can have different bit patterns but they should all |
| 208 |
* compare as equal. For backwards-compatibility reasons we force them to |
| 209 |
* have the hash value of a standard NaN. |
| 210 |
*/ |
| 211 |
if (isnan(key)) |
| 212 |
key = get_float8_nan(); |
| 213 |
|
| 214 |
return hash_any((unsigned char *) &key, sizeof(key)); |
| 215 |
} |
| 216 |
|
| 217 |
Datum |
| 218 |
hashfloat8extended(PG_FUNCTION_ARGS) |
| 219 |
{ |
| 220 |
float8 key = PG_GETARG_FLOAT8(0); |
| 221 |
uint64 seed = PG_GETARG_INT64(1); |
| 222 |
|
| 223 |
/* Same approach as hashfloat8 */ |
| 224 |
if (key == (float8) 0) |
| 225 |
PG_RETURN_UINT64(seed); |
| 226 |
if (isnan(key)) |
| 227 |
key = get_float8_nan(); |
| 228 |
|
| 229 |
return hash_any_extended((unsigned char *) &key, sizeof(key), seed); |
| 230 |
} |
| 231 |
|
| 232 |
Datum |
| 233 |
hashoidvector(PG_FUNCTION_ARGS) |
| 234 |
{ |
| 235 |
oidvector *key = (oidvector *) PG_GETARG_POINTER(0); |
| 236 |
|
| 237 |
check_valid_oidvector(key); |
| 238 |
return hash_any((unsigned char *) key->values, key->dim1 * sizeof(Oid)); |
| 239 |
} |
| 240 |
|
| 241 |
Datum |
| 242 |
hashoidvectorextended(PG_FUNCTION_ARGS) |
| 243 |
{ |
| 244 |
oidvector *key = (oidvector *) PG_GETARG_POINTER(0); |
| 245 |
|
| 246 |
check_valid_oidvector(key); |
| 247 |
return hash_any_extended((unsigned char *) key->values, |
| 248 |
key->dim1 * sizeof(Oid), |
| 249 |
PG_GETARG_INT64(1)); |
| 250 |
} |
| 251 |
|
| 252 |
Datum |
| 253 |
hashname(PG_FUNCTION_ARGS) |
| 254 |
{ |
| 255 |
char *key = NameStr(*PG_GETARG_NAME(0)); |
| 256 |
|
| 257 |
return hash_any((unsigned char *) key, strlen(key)); |
| 258 |
} |
| 259 |
|
| 260 |
Datum |
| 261 |
hashnameextended(PG_FUNCTION_ARGS) |
| 262 |
{ |
| 263 |
char *key = NameStr(*PG_GETARG_NAME(0)); |
| 264 |
|
| 265 |
return hash_any_extended((unsigned char *) key, strlen(key), |
| 266 |
PG_GETARG_INT64(1)); |
| 267 |
} |
| 268 |
|
| 269 |
Datum |
| 270 |
hashtext(PG_FUNCTION_ARGS) |
| 271 |
{ |
| 272 |
text *key = PG_GETARG_TEXT_PP(0); |
| 273 |
Oid collid = PG_GET_COLLATION(); |
| 274 |
pg_locale_t mylocale; |
| 275 |
Datum result; |
| 276 |
|
| 277 |
if (!collid) |
| 278 |
ereport(ERROR, |
| 279 |
(errcode(ERRCODE_INDETERMINATE_COLLATION), |
| 280 |
errmsg("could not determine which collation to use for string hashing"), |
| 281 |
errhint("Use the COLLATE clause to set the collation explicitly."))); |
| 282 |
|
| 283 |
mylocale = pg_newlocale_from_collation(collid); |
| 284 |
|
| 285 |
if (mylocale->deterministic) |
| 286 |
{ |
| 287 |
result = hash_any((unsigned char *) VARDATA_ANY(key), |
| 288 |
VARSIZE_ANY_EXHDR(key)); |
| 289 |
} |
| 290 |
else |
| 291 |
{ |
| 292 |
Size bsize, |
| 293 |
rsize; |
| 294 |
char *buf; |
| 295 |
const char *keydata = VARDATA_ANY(key); |
| 296 |
size_t keylen = VARSIZE_ANY_EXHDR(key); |
| 297 |
|
| 298 |
|
| 299 |
bsize = pg_strnxfrm(NULL, 0, keydata, keylen, mylocale); |
| 300 |
buf = palloc(bsize + 1); |
| 301 |
|
| 302 |
rsize = pg_strnxfrm(buf, bsize + 1, keydata, keylen, mylocale); |
| 303 |
|
| 304 |
/* the second call may return a smaller value than the first */ |
| 305 |
if (rsize > bsize) |
| 306 |
elog(ERROR, "pg_strnxfrm() returned unexpected result"); |
| 307 |
|
| 308 |
/* |
| 309 |
* In principle, there's no reason to include the terminating NUL |
| 310 |
* character in the hash, but it was done before and the behavior must |
| 311 |
* be preserved. |
| 312 |
*/ |
| 313 |
result = hash_any((uint8_t *) buf, bsize + 1); |
| 314 |
|
| 315 |
pfree(buf); |
| 316 |
} |
| 317 |
|
| 318 |
/* Avoid leaking memory for toasted inputs */ |
| 319 |
PG_FREE_IF_COPY(key, 0); |
| 320 |
|
| 321 |
return result; |
| 322 |
} |
| 323 |
|
| 324 |
Datum |
| 325 |
hashtextextended(PG_FUNCTION_ARGS) |
| 326 |
{ |
| 327 |
text *key = PG_GETARG_TEXT_PP(0); |
| 328 |
Oid collid = PG_GET_COLLATION(); |
| 329 |
pg_locale_t mylocale; |
| 330 |
Datum result; |
| 331 |
|
| 332 |
if (!collid) |
| 333 |
ereport(ERROR, |
| 334 |
(errcode(ERRCODE_INDETERMINATE_COLLATION), |
| 335 |
errmsg("could not determine which collation to use for string hashing"), |
| 336 |
errhint("Use the COLLATE clause to set the collation explicitly."))); |
| 337 |
|
| 338 |
mylocale = pg_newlocale_from_collation(collid); |
| 339 |
|
| 340 |
if (mylocale->deterministic) |
| 341 |
{ |
| 342 |
result = hash_any_extended((unsigned char *) VARDATA_ANY(key), |
| 343 |
VARSIZE_ANY_EXHDR(key), |
| 344 |
PG_GETARG_INT64(1)); |
| 345 |
} |
| 346 |
else |
| 347 |
{ |
| 348 |
Size bsize, |
| 349 |
rsize; |
| 350 |
char *buf; |
| 351 |
const char *keydata = VARDATA_ANY(key); |
| 352 |
size_t keylen = VARSIZE_ANY_EXHDR(key); |
| 353 |
|
| 354 |
bsize = pg_strnxfrm(NULL, 0, keydata, keylen, mylocale); |
| 355 |
buf = palloc(bsize + 1); |
| 356 |
|
| 357 |
rsize = pg_strnxfrm(buf, bsize + 1, keydata, keylen, mylocale); |
| 358 |
|
| 359 |
/* the second call may return a smaller value than the first */ |
| 360 |
if (rsize > bsize) |
| 361 |
elog(ERROR, "pg_strnxfrm() returned unexpected result"); |
| 362 |
|
| 363 |
/* |
| 364 |
* In principle, there's no reason to include the terminating NUL |
| 365 |
* character in the hash, but it was done before and the behavior must |
| 366 |
* be preserved. |
| 367 |
*/ |
| 368 |
result = hash_any_extended((uint8_t *) buf, bsize + 1, |
| 369 |
PG_GETARG_INT64(1)); |
| 370 |
|
| 371 |
pfree(buf); |
| 372 |
} |
| 373 |
|
| 374 |
PG_FREE_IF_COPY(key, 0); |
| 375 |
|
| 376 |
return result; |
| 377 |
} |
| 378 |
|
| 379 |
/* |
| 380 |
* hashvarlena() can be used for any varlena datatype in which there are |
| 381 |
* no non-significant bits, ie, distinct bitpatterns never compare as equal. |
| 382 |
* |
| 383 |
* (However, you need to define an SQL-level wrapper function around it with |
| 384 |
* the concrete input data type; otherwise hashvalidate() won't accept it. |
| 385 |
* Moreover, at least for built-in types, a C-level wrapper function is also |
| 386 |
* recommended; otherwise, the opr_sanity test will get upset.) |
| 387 |
*/ |
| 388 |
Datum |
| 389 |
hashvarlena(PG_FUNCTION_ARGS) |
| 390 |
{ |
| 391 |
struct varlena *key = PG_GETARG_VARLENA_PP(0); |
| 392 |
Datum result; |
| 393 |
|
| 394 |
result = hash_any((unsigned char *) VARDATA_ANY(key), |
| 395 |
VARSIZE_ANY_EXHDR(key)); |
| 396 |
|
| 397 |
/* Avoid leaking memory for toasted inputs */ |
| 398 |
PG_FREE_IF_COPY(key, 0); |
| 399 |
|
| 400 |
return result; |
| 401 |
} |
| 402 |
|
| 403 |
Datum |
| 404 |
hashvarlenaextended(PG_FUNCTION_ARGS) |
| 405 |
{ |
| 406 |
struct varlena *key = PG_GETARG_VARLENA_PP(0); |
| 407 |
Datum result; |
| 408 |
|
| 409 |
result = hash_any_extended((unsigned char *) VARDATA_ANY(key), |
| 410 |
VARSIZE_ANY_EXHDR(key), |
| 411 |
PG_GETARG_INT64(1)); |
| 412 |
|
| 413 |
PG_FREE_IF_COPY(key, 0); |
| 414 |
|
| 415 |
return result; |
| 416 |
} |
| 417 |
|
| 418 |
Datum |
| 419 |
hashbytea(PG_FUNCTION_ARGS) |
| 420 |
{ |
| 421 |
return hashvarlena(fcinfo); |
| 422 |
} |
| 423 |
|
| 424 |
Datum |
| 425 |
hashbyteaextended(PG_FUNCTION_ARGS) |
| 426 |
{ |
| 427 |
return hashvarlenaextended(fcinfo); |
| 428 |
} |