| 1 |
// Copyright (c) Meta Platforms, Inc. and affiliates. |
| 2 |
// SPDX-License-Identifier: LGPL-2.1-or-later |
| 3 |
|
| 4 |
/** |
| 5 |
* @file |
| 6 |
* |
| 7 |
* High performance generic hash tables. |
| 8 |
* |
| 9 |
* See @ref HashTables. |
| 10 |
*/ |
| 11 |
|
| 12 |
#ifndef DRGN_HASH_TABLE_H |
| 13 |
#define DRGN_HASH_TABLE_H |
| 14 |
|
| 15 |
#ifdef __SSE2__ |
| 16 |
#include <emmintrin.h> // IWYU pragma: keep |
| 17 |
#endif |
| 18 |
#ifdef __SSE4_2__ |
| 19 |
#include <nmmintrin.h> |
| 20 |
#endif |
| 21 |
#ifdef __BMI2__ |
| 22 |
#include <immintrin.h> |
| 23 |
#endif |
| 24 |
#include <stdalign.h> |
| 25 |
#include <stdbool.h> |
| 26 |
#include <stddef.h> |
| 27 |
#include <stdint.h> |
| 28 |
#include <string.h> |
| 29 |
|
| 30 |
#include "bitops.h" |
| 31 |
#include "cityhash.h" |
| 32 |
#include "generics.h" |
| 33 |
#include "minmax.h" |
| 34 |
#include "nstring.h" // IWYU pragma: export |
| 35 |
#include "util.h" |
| 36 |
|
| 37 |
/** |
| 38 |
* @ingroup Internals |
| 39 |
* |
| 40 |
* @defgroup HashTables Hash tables |
| 41 |
* |
| 42 |
* High performance generic hash tables. |
| 43 |
* |
| 44 |
* This is an implementation of Facebook's <a |
| 45 |
* href="https://github.com/facebook/folly/blob/master/folly/container/F14.md"> |
| 46 |
* F14</a>, which provides both high performance and good memory efficiency by |
| 47 |
* using SIMD instructions to allow for a high load factor. |
| 48 |
* |
| 49 |
* These hash tables are generic, strongly typed (i.e., keys and values have |
| 50 |
* static types rather than <tt>void *</tt>), and don't have any function |
| 51 |
* pointer overhead. |
| 52 |
* |
| 53 |
* On non-x86 platforms, this falls back to a slower implementation that doesn't |
| 54 |
* use SIMD. |
| 55 |
* |
| 56 |
* Abstractly, a hash table stores @em entries which can be looked up by @em |
| 57 |
* key. A hash table is defined with @ref DEFINE_HASH_TABLE() (or the |
| 58 |
* higher-level wrappers, @ref DEFINE_HASH_MAP() and @ref DEFINE_HASH_SET()). |
| 59 |
* Each generated hash table interface is prefixed with a given name; the |
| 60 |
* interface documented here uses the example name @c hash_table, which could be |
| 61 |
* generated with this example code: |
| 62 |
* |
| 63 |
* @code{.c} |
| 64 |
* key_type entry_to_key(const entry_type *entry); |
| 65 |
* struct hash_pair hash_func(const key_type *key); |
| 66 |
* bool eq_func(const key_type *a, const key_type *b); |
| 67 |
* DEFINE_HASH_TABLE(hash_table, entry_type, entry_to_key, hash_func, eq_func); |
| 68 |
* @endcode |
| 69 |
* |
| 70 |
* @sa BinarySearchTrees |
| 71 |
* |
| 72 |
* @{ |
| 73 |
*/ |
| 74 |
|
| 75 |
/** |
| 76 |
* Double hash. |
| 77 |
* |
| 78 |
* @sa HashTableHelpers |
| 79 |
*/ |
| 80 |
struct hash_pair { |
| 81 |
/** |
| 82 |
* First hash. |
| 83 |
* |
| 84 |
* F14 uses this to select the chunk. |
| 85 |
*/ |
| 86 |
size_t first; |
| 87 |
/** |
| 88 |
* Second hash. |
| 89 |
* |
| 90 |
* F14 uses this as the tag within the chunk and as the probe stride |
| 91 |
* when a chunk overflows. |
| 92 |
* |
| 93 |
* Only the 8 least-significant bits of this are used; the rest are zero |
| 94 |
* (the folly implementation insists that storing this as @c size_t |
| 95 |
* generates better code). The 8th bit is always set. This is derived |
| 96 |
* from @ref hash_pair::first; see @ref |
| 97 |
* hash_pair_from_avalanching_hash() and @ref |
| 98 |
* hash_pair_from_non_avalanching_hash(). |
| 99 |
*/ |
| 100 |
size_t second; |
| 101 |
}; |
| 102 |
|
| 103 |
#ifdef DOXYGEN |
| 104 |
/** |
| 105 |
* @struct hash_table |
| 106 |
* |
| 107 |
* Hash table instance. |
| 108 |
* |
| 109 |
* There are no requirements on how this is allocated; it may be global, on the |
| 110 |
* stack, allocated by @c malloc(), embedded in another structure, etc. |
| 111 |
*/ |
| 112 |
struct hash_table; |
| 113 |
|
| 114 |
/** |
| 115 |
* Hash table iterator. |
| 116 |
* |
| 117 |
* Several functions return an iterator or take one as an argument. This |
| 118 |
* iterator has a reference to an entry, which can be @c NULL to indicate that |
| 119 |
* there is no such entry. It also contains private bookkeeping which should not |
| 120 |
* be used. |
| 121 |
* |
| 122 |
* An iterator remains valid until the table is rehashed or the entry or one |
| 123 |
* before it is deleted. |
| 124 |
*/ |
| 125 |
struct hash_table_iterator { |
| 126 |
/** Pointer to the entry in the hash table. */ |
| 127 |
entry_type *entry; |
| 128 |
}; |
| 129 |
|
| 130 |
/** |
| 131 |
* Compute the hash for a given key. |
| 132 |
* |
| 133 |
* Note that this function is simply a wrapper around the hash function that was |
| 134 |
* passed when defining the hash table. It is provided for convenience. |
| 135 |
*/ |
| 136 |
struct hash_pair hash_table_hash(const key_type *key); |
| 137 |
|
| 138 |
/** |
| 139 |
* Initialize a @ref hash_table. |
| 140 |
* |
| 141 |
* The new hash table is empty. It must be deinitialized with @ref |
| 142 |
* hash_table_deinit(). |
| 143 |
* |
| 144 |
* @sa HASH_TABLE_INIT |
| 145 |
*/ |
| 146 |
void hash_table_init(struct hash_table *table); |
| 147 |
|
| 148 |
/** |
| 149 |
* Free memory allocated by a @ref hash_table. |
| 150 |
* |
| 151 |
* After this is called, the hash table must not be used unless it is |
| 152 |
* reinitialized with @ref hash_table_init(). Note that this only frees memory |
| 153 |
* allocated by the hash table implementation; if the keys, values, or the hash |
| 154 |
* table structure itself are dynamically allocated, those must be freed |
| 155 |
* separately. |
| 156 |
*/ |
| 157 |
void hash_table_deinit(struct hash_table *table); |
| 158 |
|
| 159 |
/** |
| 160 |
* Return whether a @ref hash_table has no entries. |
| 161 |
* |
| 162 |
* This is O(1). |
| 163 |
*/ |
| 164 |
bool hash_table_empty(struct hash_table *table); |
| 165 |
|
| 166 |
/** |
| 167 |
* Return the number of entries in a @ref hash_table. |
| 168 |
* |
| 169 |
* This is O(1). |
| 170 |
*/ |
| 171 |
size_t hash_table_size(struct hash_table *table); |
| 172 |
|
| 173 |
/** |
| 174 |
* Maximum possible number of entries in a @ref hash_table. |
| 175 |
* |
| 176 |
* Attempts to increase the size or capacity beyond this will fail. |
| 177 |
*/ |
| 178 |
const size_t hash_table_max_size; |
| 179 |
|
| 180 |
/** |
| 181 |
* Delete all entries in a @ref hash_table. |
| 182 |
* |
| 183 |
* This does not necessarily free memory used by the hash table. |
| 184 |
*/ |
| 185 |
void hash_table_clear(struct hash_table *table); |
| 186 |
|
| 187 |
/** |
| 188 |
* Reserve entries in a @ref hash_table. |
| 189 |
* |
| 190 |
* This allocates space up front and rehashes the table to ensure that it will |
| 191 |
* not be rehashed until it contains the given number of entries. |
| 192 |
* |
| 193 |
* @return @c true on success, @c false on failure. |
| 194 |
*/ |
| 195 |
bool hash_table_reserve(struct hash_table *table, size_t capacity); |
| 196 |
|
| 197 |
/** |
| 198 |
* Insert an entry in a @ref hash_table. |
| 199 |
* |
| 200 |
* If an entry with the same key is already in the hash table, the entry is @em |
| 201 |
* not inserted. |
| 202 |
* |
| 203 |
* @param[out] it_ret If not @c NULL, a returned iterator pointing to the newly |
| 204 |
* inserted entry or the existing entry with the same key. |
| 205 |
* @return 1 if the entry was inserted, 0 if the key already existed, -1 if |
| 206 |
* allocating memory for a rehash failed. |
| 207 |
*/ |
| 208 |
int hash_table_insert(struct hash_table *table, const entry_type *entry, |
| 209 |
struct hash_table_iterator *it_ret); |
| 210 |
|
| 211 |
/** |
| 212 |
* Insert an entry in a @ref hash_table with a precomputed hash. |
| 213 |
* |
| 214 |
* Like @ref hash_table_insert(), but the hash was already computed. This saves |
| 215 |
* recomputing the hash when doing multiple operations with the same key. |
| 216 |
*/ |
| 217 |
int hash_table_insert_hashed(struct hash_table *table, const entry_type *entry, |
| 218 |
struct hash_pair hp, |
| 219 |
struct hash_table_iterator *it_ret); |
| 220 |
|
| 221 |
/** |
| 222 |
* Insert an entry in a @ref hash_table which is not in the table. |
| 223 |
* |
| 224 |
* Like @ref hash_table_insert_hashed(), but a search was previously done and |
| 225 |
* the key is not already in the table. This saves doing a redundant search in |
| 226 |
* that case but is unsafe otherwise. |
| 227 |
*/ |
| 228 |
int hash_table_insert_searched(struct hash_table *table, |
| 229 |
const entry_type *entry, struct hash_pair hp, |
| 230 |
struct hash_table_iterator *it_ret); |
| 231 |
|
| 232 |
/** |
| 233 |
* Search for an entry in a @ref hash_table. |
| 234 |
* |
| 235 |
* @return An iterator pointing to the entry with the given key, or an iterator |
| 236 |
* with <tt>entry == NULL</tt> if the key was not found. |
| 237 |
*/ |
| 238 |
struct hash_table_iterator hash_table_search(struct hash_table *table, |
| 239 |
const key_type *key); |
| 240 |
|
| 241 |
/** |
| 242 |
* Search for an entry in a @ref hash_table with a precomputed hash. |
| 243 |
* |
| 244 |
* Like @ref hash_table_search(), but the hash was already computed. This saves |
| 245 |
* recomputing the hash when doing multiple operations with the same key. |
| 246 |
*/ |
| 247 |
struct hash_table_iterator hash_table_search_hashed(struct hash_table *table, |
| 248 |
const key_type *key, |
| 249 |
struct hash_pair hp); |
| 250 |
|
| 251 |
/** |
| 252 |
* Delete an entry in a @ref hash_table. |
| 253 |
* |
| 254 |
* This deletes the entry with the given key. It will never rehash the table. |
| 255 |
* |
| 256 |
* @return @c true if the entry was found and deleted, @c false if not. |
| 257 |
*/ |
| 258 |
bool hash_table_delete(struct hash_table *table, const key_type *key); |
| 259 |
|
| 260 |
/** |
| 261 |
* Delete an entry in a @ref hash_table with a precomputed hash. |
| 262 |
* |
| 263 |
* Like @ref hash_table_delete(), but the hash was already computed. This saves |
| 264 |
* recomputing the hash when doing multiple operations with the same key. |
| 265 |
*/ |
| 266 |
bool hash_table_delete_hashed(struct hash_table *table, struct hash_pair hp); |
| 267 |
|
| 268 |
/** |
| 269 |
* Delete an entry given by an iterator in a @ref hash_table. |
| 270 |
* |
| 271 |
* This deletes the entry pointed to by the iterator. It will never rehash the |
| 272 |
* table. |
| 273 |
* |
| 274 |
* @return An iterator pointing to the next entry in the table. See @ref |
| 275 |
* hash_table_next(). |
| 276 |
*/ |
| 277 |
struct hash_table_iterator |
| 278 |
hash_table_delete_iterator(struct hash_table *table, |
| 279 |
struct hash_table_iterator it); |
| 280 |
|
| 281 |
/** |
| 282 |
* Delete an entry given by an iterator in a @ref hash_table with a precomputed |
| 283 |
* hash. |
| 284 |
* |
| 285 |
* Like @ref hash_table_delete_iterator(), but the hash was already computed. |
| 286 |
* This saves recomputing the hash when doing multiple operations with the same |
| 287 |
* key. |
| 288 |
*/ |
| 289 |
struct hash_table_iterator |
| 290 |
hash_table_delete_iterator_hashed(struct hash_table *table, |
| 291 |
struct hash_table_iterator it, |
| 292 |
struct hash_pair hp); |
| 293 |
|
| 294 |
/** |
| 295 |
* Delete an entry in a @ref hash_table. |
| 296 |
* |
| 297 |
* @return An iterator pointing to the next entry in the table. See @ref |
| 298 |
* hash_table_next(). |
| 299 |
*/ |
| 300 |
bool hash_table_delete_entry(struct hash_table *table, const entry_type *entry); |
| 301 |
|
| 302 |
/** |
| 303 |
* Get an iterator pointing to the first entry in a @ref hash_table. |
| 304 |
* |
| 305 |
* The first entry is arbitrary. |
| 306 |
* |
| 307 |
* @return An iterator pointing to the first entry, or an iterator with |
| 308 |
* <tt>entry == NULL</tt> if the table is empty. |
| 309 |
*/ |
| 310 |
struct hash_table_iterator hash_table_first(struct hash_table *table); |
| 311 |
|
| 312 |
/** |
| 313 |
* Get an iterator pointing to the next entry in a @ref hash_table. |
| 314 |
* |
| 315 |
* The order of entries is arbitrary. |
| 316 |
* |
| 317 |
* @return An iterator pointing to the next entry, or an iterator with <tt>entry |
| 318 |
* == NULL</tt> if there are no more entries. |
| 319 |
*/ |
| 320 |
struct hash_table_iterator hash_table_next(struct hash_table_iterator it); |
| 321 |
#endif |
| 322 |
|
| 323 |
enum { hash_table_chunk_alignment = max_iconst(alignof(max_align_t), 16) }; |
| 324 |
|
| 325 |
static inline size_t hash_table_probe_delta(struct hash_pair hp) |
| 326 |
{ |
| 327 |
return 2 * hp.second + 1; |
| 328 |
} |
| 329 |
|
| 330 |
static const uint8_t hosted_overflow_count_inc = 0x10; |
| 331 |
static const uint8_t hosted_overflow_count_dec = -0x10; |
| 332 |
|
| 333 |
#if SIZE_MAX == 0xffffffffffffffff |
| 334 |
_Static_assert(sizeof(size_t) == sizeof(uint64_t), |
| 335 |
"size_t/SIZE_MAX doesn't make sense"); |
| 336 |
|
| 337 |
struct hash_table_size_and_chunk_shift { |
| 338 |
uint64_t packed; |
| 339 |
}; |
| 340 |
|
| 341 |
static inline void |
| 342 |
hash_table_size_and_chunk_shift_init(struct hash_table_size_and_chunk_shift *scs) |
| 343 |
{ |
| 344 |
scs->packed = 0; |
| 345 |
} |
| 346 |
|
| 347 |
enum { hash_table_size_shift = 8 }; |
| 348 |
static const size_t hash_table_max_size = SIZE_MAX >> hash_table_size_shift; |
| 349 |
|
| 350 |
static inline size_t |
| 351 |
hash_table_size(struct hash_table_size_and_chunk_shift *scs) |
| 352 |
{ |
| 353 |
return scs->packed >> hash_table_size_shift; |
| 354 |
} |
| 355 |
|
| 356 |
static inline uint8_t |
| 357 |
hash_table_chunk_shift(struct hash_table_size_and_chunk_shift *scs) |
| 358 |
{ |
| 359 |
return scs->packed; |
| 360 |
} |
| 361 |
|
| 362 |
static inline void |
| 363 |
hash_table_set_size(struct hash_table_size_and_chunk_shift *scs, size_t size) |
| 364 |
{ |
| 365 |
scs->packed = (size << hash_table_size_shift) |
| 366 |
| hash_table_chunk_shift(scs); |
| 367 |
} |
| 368 |
|
| 369 |
static inline void |
| 370 |
hash_table_increment_size(struct hash_table_size_and_chunk_shift *scs) |
| 371 |
{ |
| 372 |
scs->packed += UINT64_C(1) << hash_table_size_shift; |
| 373 |
} |
| 374 |
|
| 375 |
static inline void |
| 376 |
hash_table_decrement_size(struct hash_table_size_and_chunk_shift *scs) |
| 377 |
{ |
| 378 |
scs->packed -= UINT64_C(1) << hash_table_size_shift; |
| 379 |
} |
| 380 |
|
| 381 |
static inline void |
| 382 |
hash_table_set_chunk_count(struct hash_table_size_and_chunk_shift *scs, |
| 383 |
size_t chunk_count) |
| 384 |
{ |
| 385 |
scs->packed = (hash_table_size(scs) << hash_table_size_shift) |
| 386 |
| ilog2(chunk_count); |
| 387 |
} |
| 388 |
#elif SIZE_MAX == 0xffffffff |
| 389 |
_Static_assert(sizeof(size_t) == sizeof(uint32_t), |
| 390 |
"size_t/SIZE_MAX doesn't make sense"); |
| 391 |
|
| 392 |
struct hash_table_size_and_chunk_shift { |
| 393 |
size_t size; |
| 394 |
uint32_t chunk_shift; |
| 395 |
}; |
| 396 |
|
| 397 |
static inline void |
| 398 |
hash_table_size_and_chunk_shift_init(struct hash_table_size_and_chunk_shift *scs) |
| 399 |
{ |
| 400 |
scs->size = 0; |
| 401 |
scs->chunk_shift = 0; |
| 402 |
} |
| 403 |
|
| 404 |
static const size_t hash_table_max_size = SIZE_MAX; |
| 405 |
|
| 406 |
static inline size_t |
| 407 |
hash_table_size(struct hash_table_size_and_chunk_shift *scs) |
| 408 |
{ |
| 409 |
return scs->size; |
| 410 |
} |
| 411 |
|
| 412 |
static inline uint8_t |
| 413 |
hash_table_chunk_shift(struct hash_table_size_and_chunk_shift *scs) |
| 414 |
{ |
| 415 |
return scs->chunk_shift; |
| 416 |
} |
| 417 |
|
| 418 |
static inline void |
| 419 |
hash_table_set_size(struct hash_table_size_and_chunk_shift *scs, size_t size) |
| 420 |
{ |
| 421 |
scs->size = size; |
| 422 |
} |
| 423 |
|
| 424 |
static inline void |
| 425 |
hash_table_increment_size(struct hash_table_size_and_chunk_shift *scs) |
| 426 |
{ |
| 427 |
scs->size++; |
| 428 |
} |
| 429 |
|
| 430 |
static inline void |
| 431 |
hash_table_decrement_size(struct hash_table_size_and_chunk_shift *scs) |
| 432 |
{ |
| 433 |
scs->size--; |
| 434 |
} |
| 435 |
|
| 436 |
static inline void |
| 437 |
hash_table_set_chunk_count(struct hash_table_size_and_chunk_shift *scs, |
| 438 |
size_t chunk_count) |
| 439 |
{ |
| 440 |
scs->chunk_shift = ilog2(chunk_count); |
| 441 |
} |
| 442 |
#else |
| 443 |
#error "unsupported SIZE_MAX" |
| 444 |
#endif |
| 445 |
|
| 446 |
static inline size_t |
| 447 |
hash_table_chunk_count(struct hash_table_size_and_chunk_shift *scs) |
| 448 |
{ |
| 449 |
return (size_t)1 << hash_table_chunk_shift(scs); |
| 450 |
} |
| 451 |
|
| 452 |
static inline size_t |
| 453 |
hash_table_modulo_by_chunk_count(struct hash_table_size_and_chunk_shift *scs, |
| 454 |
size_t index) |
| 455 |
{ |
| 456 |
// The folly implementation resorts to an instrinsic because BZHI wasn't |
| 457 |
// generated reliably by the compiler. |
| 458 |
#ifdef __BMI2__ |
| 459 |
return _bzhi_u64(index, hash_table_chunk_shift(scs)); |
| 460 |
#else |
| 461 |
return index & (((size_t)1 << hash_table_chunk_shift(scs)) - 1); |
| 462 |
#endif |
| 463 |
} |
| 464 |
|
| 465 |
/* |
| 466 |
* We could represent an empty hash table with chunks set to NULL. However, then |
| 467 |
* we would need a branch to check for this in insert, search, and delete. We |
| 468 |
* could avoid this by allocating an empty chunk, but that is wasteful since it |
| 469 |
* will never actually be used. Instead, we have a special empty chunk which is |
| 470 |
* used by all tables. |
| 471 |
*/ |
| 472 |
extern const uint8_t hash_table_empty_chunk_header[]; |
| 473 |
#define hash_table_empty_chunk (void *)hash_table_empty_chunk_header |
| 474 |
|
| 475 |
#ifdef __SSE2__ |
| 476 |
#define HASH_TABLE_CHUNK_MATCH(table) \ |
| 477 |
static inline unsigned int table##_chunk_match(struct table##_chunk *chunk, \ |
| 478 |
size_t needle) \ |
| 479 |
{ \ |
| 480 |
__m128i tag_vec = _mm_load_si128((__m128i *)chunk); \ |
| 481 |
__m128i needle_vec = _mm_set1_epi8((uint8_t)needle); \ |
| 482 |
__m128i eq_vec = _mm_cmpeq_epi8(tag_vec, needle_vec); \ |
| 483 |
return _mm_movemask_epi8(eq_vec) & table##_chunk_full_mask; \ |
| 484 |
} |
| 485 |
|
| 486 |
#define HASH_TABLE_CHUNK_OCCUPIED(table) \ |
| 487 |
static inline unsigned int table##_chunk_occupied(struct table##_chunk *chunk) \ |
| 488 |
{ \ |
| 489 |
__m128i tag_vec = _mm_load_si128((__m128i *)chunk); \ |
| 490 |
return _mm_movemask_epi8(tag_vec) & table##_chunk_full_mask; \ |
| 491 |
} |
| 492 |
#else |
| 493 |
#define HASH_TABLE_CHUNK_MATCH(table) \ |
| 494 |
static inline unsigned int table##_chunk_match(struct table##_chunk *chunk, \ |
| 495 |
size_t needle) \ |
| 496 |
{ \ |
| 497 |
unsigned int mask, i; \ |
| 498 |
for (mask = 0, i = 0; i < table##_chunk_capacity; i++) { \ |
| 499 |
if (chunk->tags[i] == needle) \ |
| 500 |
mask |= 1U << i; \ |
| 501 |
} \ |
| 502 |
return mask; \ |
| 503 |
} |
| 504 |
|
| 505 |
#define HASH_TABLE_CHUNK_OCCUPIED(table) \ |
| 506 |
static inline unsigned int table##_chunk_occupied(struct table##_chunk *chunk) \ |
| 507 |
{ \ |
| 508 |
unsigned int mask, i; \ |
| 509 |
for (mask = 0, i = 0; i < table##_chunk_capacity; i++) { \ |
| 510 |
if (chunk->tags[i]) \ |
| 511 |
mask |= 1U << i; \ |
| 512 |
} \ |
| 513 |
return mask; \ |
| 514 |
} |
| 515 |
#endif |
| 516 |
|
| 517 |
/** |
| 518 |
* Define a hash table type without defining its functions. |
| 519 |
* |
| 520 |
* This is useful when the hash table type must be defined in one place (e.g., a |
| 521 |
* header) but the interface is defined elsewhere (e.g., a source file) with |
| 522 |
* @ref DEFINE_HASH_TABLE_FUNCTIONS(). Otherwise, just use @ref |
| 523 |
* DEFINE_HASH_TABLE(). |
| 524 |
* |
| 525 |
* @sa DEFINE_HASH_TABLE() |
| 526 |
*/ |
| 527 |
#define DEFINE_HASH_TABLE_TYPE(table, entry_type) \ |
| 528 |
typedef typeof(entry_type) table##_entry_type; \ |
| 529 |
\ |
| 530 |
enum { \ |
| 531 |
/* \ |
| 532 |
* Whether this table uses the vector storage policy. \ |
| 533 |
* \ |
| 534 |
* The vector policy provides the best performance and memory \ |
| 535 |
* efficiency for medium and large entries. \ |
| 536 |
*/ \ |
| 537 |
table##_vector_policy = sizeof(table##_entry_type) >= 24, \ |
| 538 |
}; \ |
| 539 |
\ |
| 540 |
/* \ |
| 541 |
* The vector storage policy stores 32-bit indices, so it only needs 32-bit \ |
| 542 |
* sizes. \ |
| 543 |
*/ \ |
| 544 |
typedef_if(table##_size_type, table##_vector_policy, uint32_t, size_t); \ |
| 545 |
\ |
| 546 |
struct table { \ |
| 547 |
struct table##_chunk *chunks; \ |
| 548 |
struct hash_table_size_and_chunk_shift size_and_chunk_shift; \ |
| 549 |
union { \ |
| 550 |
/* Allocated together with chunks. */ \ |
| 551 |
table##_entry_type *vector; \ |
| 552 |
uintptr_t first_packed; \ |
| 553 |
}; \ |
| 554 |
}; \ |
| 555 |
struct DEFINE_HASH_TABLE_needs_semicolon |
| 556 |
|
| 557 |
/* |
| 558 |
* Common search function implementation returning an item iterator. This is |
| 559 |
* shared by key lookups and index lookups. |
| 560 |
*/ |
| 561 |
#define HASH_TABLE_SEARCH_IMPL(table, func, key_type, item_to_key, eq_func) \ |
| 562 |
static struct table##_iterator table##_##func(struct table *table, \ |
| 563 |
const key_type *key, \ |
| 564 |
struct hash_pair hp) \ |
| 565 |
{ \ |
| 566 |
const size_t delta = hash_table_probe_delta(hp); \ |
| 567 |
size_t index = hp.first; \ |
| 568 |
for (size_t tries = 0; tries >> table##_chunk_shift(table) == 0; \ |
| 569 |
tries++) { \ |
| 570 |
struct table##_chunk *chunk = \ |
| 571 |
&table->chunks[table##_modulo_by_chunk_count(table, index)]; \ |
| 572 |
if (sizeof(*chunk) > 64) \ |
| 573 |
__builtin_prefetch(&chunk->items[8]); \ |
| 574 |
unsigned int mask = table##_chunk_match(chunk, hp.second), i; \ |
| 575 |
for_each_bit(i, mask) { \ |
| 576 |
table##_item_type *item = &chunk->items[i]; \ |
| 577 |
key_type item_key = item_to_key(table, item); \ |
| 578 |
if (likely(eq_func(key, &item_key))) { \ |
| 579 |
return (struct table##_iterator){ \ |
| 580 |
.item = item, \ |
| 581 |
.index = i, \ |
| 582 |
}; \ |
| 583 |
} \ |
| 584 |
} \ |
| 585 |
if (likely(chunk->outbound_overflow_count == 0)) \ |
| 586 |
break; \ |
| 587 |
index += delta; \ |
| 588 |
} \ |
| 589 |
return (struct table##_iterator){}; \ |
| 590 |
} |
| 591 |
|
| 592 |
#define HASH_TABLE_SEARCH_BY_INDEX_ITEM_TO_KEY(table, item) (*(uint32_t *)item) |
| 593 |
|
| 594 |
/** |
| 595 |
* Define the functions for a hash table. |
| 596 |
* |
| 597 |
* The hash table type must have already been defined with @ref |
| 598 |
* DEFINE_HASH_TABLE_TYPE(). |
| 599 |
* |
| 600 |
* Unless the type and function definitions must be in separate places, use @ref |
| 601 |
* DEFINE_HASH_TABLE() instead. |
| 602 |
*/ |
| 603 |
#define DEFINE_HASH_TABLE_FUNCTIONS(table, entry_to_key, hash_func, eq_func) \ |
| 604 |
typedef typeof(entry_to_key((table##_entry_type *)0)) table##_key_type; \ |
| 605 |
\ |
| 606 |
static inline table##_key_type \ |
| 607 |
table##_entry_to_key(const table##_entry_type *entry) \ |
| 608 |
{ \ |
| 609 |
return entry_to_key(entry); \ |
| 610 |
} \ |
| 611 |
\ |
| 612 |
/* \ |
| 613 |
* Item stored in a chunk. \ |
| 614 |
* \ |
| 615 |
* When using the basic policy, the entry is stored directly in the item. When \ |
| 616 |
* using the vector policy, the item is an index to an out-of-band vector of \ |
| 617 |
* entries. \ |
| 618 |
*/ \ |
| 619 |
typedef_if(table##_item_type, table##_vector_policy, uint32_t, \ |
| 620 |
table##_entry_type); \ |
| 621 |
\ |
| 622 |
static inline uint32_t *table##_vector_item(table##_item_type *item) \ |
| 623 |
{ \ |
| 624 |
return (uint32_t *)item; \ |
| 625 |
} \ |
| 626 |
\ |
| 627 |
enum { \ |
| 628 |
/* \ |
| 629 |
* The number of items per chunk. 14 is the most space efficient, but \ |
| 630 |
* if an item is 4 bytes, 12 items makes a chunk exactly one cache \ |
| 631 |
* line. \ |
| 632 |
*/ \ |
| 633 |
table##_chunk_capacity = sizeof(table##_item_type) == 4 ? 12 : 14, \ |
| 634 |
/* The maximum load factor in terms of items per chunk. */ \ |
| 635 |
table##_chunk_desired_capacity = table##_chunk_capacity - 2, \ |
| 636 |
/* \ |
| 637 |
* If an item is 16 bytes, add an extra 16 bytes of padding to make a \ |
| 638 |
* chunk exactly four cache lines. \ |
| 639 |
*/ \ |
| 640 |
table##_chunk_allocated_capacity = \ |
| 641 |
(table##_chunk_capacity + \ |
| 642 |
(sizeof(table##_item_type) == 16 ? 1 : 0)), \ |
| 643 |
/* \ |
| 644 |
* If the chunk capacity is 12, we can use tags 12 and 13 for 16 bits. \ |
| 645 |
* Otherwise, we only get 4 from control. \ |
| 646 |
*/ \ |
| 647 |
table##_capacity_scale_bits = table##_chunk_capacity == 12 ? 16 : 4, \ |
| 648 |
table##_capacity_scale_shift = table##_capacity_scale_bits - 4, \ |
| 649 |
table##_chunk_full_mask = (1 << table##_chunk_capacity) - 1, \ |
| 650 |
}; \ |
| 651 |
\ |
| 652 |
struct table##_chunk { \ |
| 653 |
uint8_t tags[14]; \ |
| 654 |
/* \ |
| 655 |
* The lower 4 bits are capacity_scale: for the first chunk, this is \ |
| 656 |
* the scaling factor between the chunk count and the capacity; for \ |
| 657 |
* other chunks, this is zero. \ |
| 658 |
* \ |
| 659 |
* The upper 4 bits are hosted_overflow_count: the number of entries in \ |
| 660 |
* this chunk that overflowed their desired chunk. \ |
| 661 |
*/ \ |
| 662 |
uint8_t control; \ |
| 663 |
/* \ |
| 664 |
* The number of entries that would have been in this chunk if it were \ |
| 665 |
* not full. This value saturates if it hits 255, after which it will \ |
| 666 |
* not be updated. \ |
| 667 |
*/ \ |
| 668 |
uint8_t outbound_overflow_count; \ |
| 669 |
table##_item_type items[table##_chunk_allocated_capacity]; \ |
| 670 |
} __attribute__((__aligned__(hash_table_chunk_alignment))); \ |
| 671 |
\ |
| 672 |
/* \ |
| 673 |
* This may be a "public iterator" (used by the public interface to refer to an \ |
| 674 |
* entry) or an "item iterator" (used by certain internal functions to refer to \ |
| 675 |
* an item regardless of the storage policy). \ |
| 676 |
*/ \ |
| 677 |
struct table##_iterator { \ |
| 678 |
union { \ |
| 679 |
/* Entry if public iterator. */ \ |
| 680 |
table##_entry_type *entry; \ |
| 681 |
/* \ |
| 682 |
* Item if item iterator. Interchangable with entry when using \ |
| 683 |
* the basic storage policy. \ |
| 684 |
*/ \ |
| 685 |
table##_item_type *item; \ |
| 686 |
}; \ |
| 687 |
union { \ |
| 688 |
/* \ |
| 689 |
* Lowest entry if public iterator and using the vector storage \ |
| 690 |
* policy (i.e., table->vector). \ |
| 691 |
*/ \ |
| 692 |
table##_entry_type *lowest; \ |
| 693 |
/* \ |
| 694 |
* Index of item in its containing chunk if item iterator or \ |
| 695 |
* using the basic storage policy. \ |
| 696 |
*/ \ |
| 697 |
size_t index; \ |
| 698 |
}; \ |
| 699 |
}; \ |
| 700 |
\ |
| 701 |
static inline struct hash_pair table##_hash(const table##_key_type *key) \ |
| 702 |
{ \ |
| 703 |
return hash_func(key); \ |
| 704 |
} \ |
| 705 |
\ |
| 706 |
static inline table##_entry_type * \ |
| 707 |
table##_item_to_entry(struct table *table, table##_item_type *item) \ |
| 708 |
{ \ |
| 709 |
if (table##_vector_policy) \ |
| 710 |
return &table->vector[*table##_vector_item(item)]; \ |
| 711 |
else \ |
| 712 |
return (table##_entry_type *)item; \ |
| 713 |
} \ |
| 714 |
\ |
| 715 |
static inline table##_key_type \ |
| 716 |
table##_item_to_key(struct table *table, table##_item_type *item) \ |
| 717 |
{ \ |
| 718 |
return table##_entry_to_key(table##_item_to_entry(table, item)); \ |
| 719 |
} \ |
| 720 |
\ |
| 721 |
/* \ |
| 722 |
* We cache the first position in the table as a tagged pointer: we steal the \ |
| 723 |
* bottom bits of the chunk pointer for the entry index. We can do this because \ |
| 724 |
* chunks are aligned to 16 bytes and the index is always less than 16. \ |
| 725 |
* \ |
| 726 |
* The folly implementation mentions this strategy but uses a more complicated \ |
| 727 |
* scheme in order to avoid computing the chunk pointer from an entry pointer. \ |
| 728 |
* We always have the chunk pointer readily available when we want to pack an \ |
| 729 |
* entry, so we can use this much simpler scheme. \ |
| 730 |
*/ \ |
| 731 |
static inline uintptr_t table##_pack_iterator(struct table##_chunk *chunk, \ |
| 732 |
size_t index) \ |
| 733 |
{ \ |
| 734 |
return (uintptr_t)chunk | (uintptr_t)index; \ |
| 735 |
} \ |
| 736 |
\ |
| 737 |
static inline struct table##_chunk *table##_unpack_chunk(uintptr_t packed) \ |
| 738 |
{ \ |
| 739 |
return (struct table##_chunk *)(packed & ~(uintptr_t)0xf); \ |
| 740 |
} \ |
| 741 |
\ |
| 742 |
static inline size_t table##_unpack_index(uintptr_t packed) \ |
| 743 |
{ \ |
| 744 |
return packed & 0xf; \ |
| 745 |
} \ |
| 746 |
\ |
| 747 |
static inline struct table##_iterator table##_unpack_iterator(uintptr_t packed) \ |
| 748 |
{ \ |
| 749 |
struct table##_chunk *chunk = table##_unpack_chunk(packed); \ |
| 750 |
size_t index = table##_unpack_index(packed); \ |
| 751 |
return (struct table##_iterator) { \ |
| 752 |
.item = chunk ? &chunk->items[index] : NULL, \ |
| 753 |
.index = index, \ |
| 754 |
}; \ |
| 755 |
} \ |
| 756 |
\ |
| 757 |
static inline struct table##_chunk * \ |
| 758 |
table##_iterator_chunk(struct table##_iterator it) \ |
| 759 |
{ \ |
| 760 |
return container_of(it.item - it.index, struct table##_chunk, items[0]);\ |
| 761 |
} \ |
| 762 |
\ |
| 763 |
HASH_TABLE_CHUNK_MATCH(table) \ |
| 764 |
HASH_TABLE_CHUNK_OCCUPIED(table) \ |
| 765 |
\ |
| 766 |
static inline unsigned int \ |
| 767 |
table##_chunk_first_empty(struct table##_chunk *chunk) \ |
| 768 |
{ \ |
| 769 |
unsigned int mask = \ |
| 770 |
table##_chunk_occupied(chunk) ^ table##_chunk_full_mask; \ |
| 771 |
return mask ? ctz(mask) : (unsigned int)-1; \ |
| 772 |
} \ |
| 773 |
\ |
| 774 |
static inline unsigned int \ |
| 775 |
table##_chunk_last_occupied(struct table##_chunk *chunk) \ |
| 776 |
{ \ |
| 777 |
unsigned int mask = table##_chunk_occupied(chunk); \ |
| 778 |
return mask ? fls(mask) - 1 : (unsigned int)-1; \ |
| 779 |
} \ |
| 780 |
\ |
| 781 |
static inline size_t \ |
| 782 |
table##_chunk_hosted_overflow_count(struct table##_chunk *chunk) \ |
| 783 |
{ \ |
| 784 |
return chunk->control >> 4; \ |
| 785 |
} \ |
| 786 |
\ |
| 787 |
static inline void \ |
| 788 |
table##_chunk_adjust_hosted_overflow_count(struct table##_chunk *chunk, \ |
| 789 |
size_t op) \ |
| 790 |
{ \ |
| 791 |
chunk->control += op; \ |
| 792 |
} \ |
| 793 |
\ |
| 794 |
static inline size_t table##_chunk_capacity_scale(struct table##_chunk *chunk) \ |
| 795 |
{ \ |
| 796 |
if (table##_capacity_scale_bits == 4) { \ |
| 797 |
return chunk->control & 0xf; \ |
| 798 |
} else { \ |
| 799 |
uint16_t val; \ |
| 800 |
memcpy(&val, &chunk->tags[12], 2); \ |
| 801 |
return val; \ |
| 802 |
} \ |
| 803 |
} \ |
| 804 |
\ |
| 805 |
static inline bool table##_chunk_eof(struct table##_chunk *chunk) \ |
| 806 |
{ \ |
| 807 |
return table##_chunk_capacity_scale(chunk) != 0; \ |
| 808 |
} \ |
| 809 |
\ |
| 810 |
static inline void table##_chunk_mark_eof(struct table##_chunk *chunk, \ |
| 811 |
size_t capacity_scale) \ |
| 812 |
{ \ |
| 813 |
if (table##_capacity_scale_bits == 4) { \ |
| 814 |
chunk->control = capacity_scale; \ |
| 815 |
} else { \ |
| 816 |
uint16_t val = capacity_scale; \ |
| 817 |
memcpy(&chunk->tags[12], &val, 2); \ |
| 818 |
} \ |
| 819 |
} \ |
| 820 |
\ |
| 821 |
static inline void \ |
| 822 |
table##_chunk_inc_outbound_overflow_count(struct table##_chunk *chunk) \ |
| 823 |
{ \ |
| 824 |
if (chunk->outbound_overflow_count != UINT8_MAX) \ |
| 825 |
chunk->outbound_overflow_count++; \ |
| 826 |
} \ |
| 827 |
\ |
| 828 |
static inline void \ |
| 829 |
table##_chunk_dec_outbound_overflow_count(struct table##_chunk *chunk) \ |
| 830 |
{ \ |
| 831 |
if (chunk->outbound_overflow_count != UINT8_MAX) \ |
| 832 |
chunk->outbound_overflow_count--; \ |
| 833 |
} \ |
| 834 |
\ |
| 835 |
__attribute__((__unused__)) \ |
| 836 |
static void table##_init(struct table *table) \ |
| 837 |
{ \ |
| 838 |
table->chunks = hash_table_empty_chunk; \ |
| 839 |
hash_table_size_and_chunk_shift_init(&table->size_and_chunk_shift); \ |
| 840 |
if (table##_vector_policy) \ |
| 841 |
table->vector = NULL; \ |
| 842 |
else \ |
| 843 |
table->first_packed = 0; \ |
| 844 |
} \ |
| 845 |
\ |
| 846 |
__attribute__((__unused__)) \ |
| 847 |
static void table##_deinit(struct table *table) \ |
| 848 |
{ \ |
| 849 |
if (table->chunks != hash_table_empty_chunk) \ |
| 850 |
free(table->chunks); \ |
| 851 |
} \ |
| 852 |
\ |
| 853 |
static inline size_t table##_size(struct table *table) \ |
| 854 |
{ \ |
| 855 |
return hash_table_size(&table->size_and_chunk_shift); \ |
| 856 |
} \ |
| 857 |
\ |
| 858 |
static inline uint8_t table##_chunk_shift(struct table *table) \ |
| 859 |
{ \ |
| 860 |
return hash_table_chunk_shift(&table->size_and_chunk_shift); \ |
| 861 |
} \ |
| 862 |
\ |
| 863 |
static inline size_t table##_chunk_count(struct table *table) \ |
| 864 |
{ \ |
| 865 |
return hash_table_chunk_count(&table->size_and_chunk_shift); \ |
| 866 |
} \ |
| 867 |
\ |
| 868 |
static inline size_t table##_modulo_by_chunk_count(struct table *table, \ |
| 869 |
size_t i) \ |
| 870 |
{ \ |
| 871 |
return hash_table_modulo_by_chunk_count(&table->size_and_chunk_shift, \ |
| 872 |
i); \ |
| 873 |
} \ |
| 874 |
\ |
| 875 |
static inline void table##_set_size(struct table *table, size_t size) \ |
| 876 |
{ \ |
| 877 |
hash_table_set_size(&table->size_and_chunk_shift, size); \ |
| 878 |
} \ |
| 879 |
\ |
| 880 |
static inline void table##_increment_size(struct table *table) \ |
| 881 |
{ \ |
| 882 |
hash_table_increment_size(&table->size_and_chunk_shift); \ |
| 883 |
} \ |
| 884 |
\ |
| 885 |
static inline void table##_decrement_size(struct table *table) \ |
| 886 |
{ \ |
| 887 |
hash_table_decrement_size(&table->size_and_chunk_shift); \ |
| 888 |
} \ |
| 889 |
\ |
| 890 |
static inline void table##_set_chunk_count(struct table *table, \ |
| 891 |
size_t chunk_count) \ |
| 892 |
{ \ |
| 893 |
hash_table_set_chunk_count(&table->size_and_chunk_shift, chunk_count); \ |
| 894 |
} \ |
| 895 |
\ |
| 896 |
__attribute__((__unused__)) \ |
| 897 |
static inline bool table##_empty(struct table *table) \ |
| 898 |
{ \ |
| 899 |
return table##_size(table) == 0; \ |
| 900 |
} \ |
| 901 |
\ |
| 902 |
static const size_t table##_max_size = \ |
| 903 |
min_iconst(min_iconst(PTRDIFF_MAX / sizeof(table##_entry_type), \ |
| 904 |
table##_vector_policy ? UINT32_MAX : SIZE_MAX), \ |
| 905 |
(table##_size_type)-1); \ |
| 906 |
\ |
| 907 |
static table##_item_type *table##_allocate_tag(struct table *table, \ |
| 908 |
uint8_t *fullness, \ |
| 909 |
struct hash_pair hp) \ |
| 910 |
{ \ |
| 911 |
const size_t delta = hash_table_probe_delta(hp); \ |
| 912 |
size_t index = hp.first; \ |
| 913 |
struct table##_chunk *chunk; \ |
| 914 |
uint8_t hosted_op = 0; \ |
| 915 |
for (;;) { \ |
| 916 |
index = table##_modulo_by_chunk_count(table, index); \ |
| 917 |
chunk = &table->chunks[index]; \ |
| 918 |
if (likely(fullness[index] < table##_chunk_capacity)) \ |
| 919 |
break; \ |
| 920 |
table##_chunk_inc_outbound_overflow_count(chunk); \ |
| 921 |
hosted_op = hosted_overflow_count_inc; \ |
| 922 |
index += delta; \ |
| 923 |
} \ |
| 924 |
size_t item_index = fullness[index]++; \ |
| 925 |
chunk->tags[item_index] = hp.second; \ |
| 926 |
table##_chunk_adjust_hosted_overflow_count(chunk, hosted_op); \ |
| 927 |
return &chunk->items[item_index]; \ |
| 928 |
} \ |
| 929 |
\ |
| 930 |
static size_t table##_compute_capacity(size_t chunk_count, size_t scale) \ |
| 931 |
{ \ |
| 932 |
return (((chunk_count - 1) >> table##_capacity_scale_shift) + 1) * scale;\ |
| 933 |
} \ |
| 934 |
\ |
| 935 |
static bool \ |
| 936 |
table##_compute_chunk_count_and_scale(size_t capacity, \ |
| 937 |
bool continuous_single_chunk_capacity, \ |
| 938 |
bool continuous_multi_chunk_capacity, \ |
| 939 |
size_t *chunk_count_ret, \ |
| 940 |
size_t *scale_ret) \ |
| 941 |
{ \ |
| 942 |
if (capacity <= table##_chunk_capacity) { \ |
| 943 |
if (!continuous_single_chunk_capacity) { \ |
| 944 |
if (capacity <= 2) \ |
| 945 |
capacity = 2; \ |
| 946 |
else if (capacity <= 6) \ |
| 947 |
capacity = 6; \ |
| 948 |
else \ |
| 949 |
capacity = table##_chunk_capacity; \ |
| 950 |
} \ |
| 951 |
*chunk_count_ret = 1; \ |
| 952 |
*scale_ret = capacity; \ |
| 953 |
} else { \ |
| 954 |
size_t min_chunks = \ |
| 955 |
(capacity - 1) / table##_chunk_desired_capacity + 1; \ |
| 956 |
size_t chunk_pow = fls(min_chunks - 1); \ |
| 957 |
if (chunk_pow == 8 * sizeof(size_t)) \ |
| 958 |
return false; \ |
| 959 |
size_t chunk_count = (size_t)1 << chunk_pow; \ |
| 960 |
size_t ss = (chunk_pow >= table##_capacity_scale_shift ? \ |
| 961 |
chunk_pow - table##_capacity_scale_shift : 0); \ |
| 962 |
size_t scale = \ |
| 963 |
continuous_multi_chunk_capacity ? \ |
| 964 |
((capacity - 1) >> ss) + 1 : \ |
| 965 |
table##_chunk_desired_capacity << (chunk_pow - ss); \ |
| 966 |
if (table##_compute_capacity(chunk_count, scale) \ |
| 967 |
> table##_max_size) \ |
| 968 |
return false; \ |
| 969 |
*chunk_count_ret = chunk_count; \ |
| 970 |
*scale_ret = scale; \ |
| 971 |
} \ |
| 972 |
return true; \ |
| 973 |
} \ |
| 974 |
\ |
| 975 |
static inline size_t table##_chunk_alloc_size(size_t chunk_count, \ |
| 976 |
size_t capacity_scale) \ |
| 977 |
{ \ |
| 978 |
/* \ |
| 979 |
* Small hash tables are common, so for capacities of less than a full \ |
| 980 |
* chunk, we only allocate the required items. \ |
| 981 |
*/ \ |
| 982 |
if (chunk_count == 1) { \ |
| 983 |
return (offsetof(struct table##_chunk, items) + \ |
| 984 |
table##_compute_capacity(1, capacity_scale) * \ |
| 985 |
sizeof(table##_item_type)); \ |
| 986 |
} else { \ |
| 987 |
return chunk_count * sizeof(struct table##_chunk); \ |
| 988 |
} \ |
| 989 |
} \ |
| 990 |
\ |
| 991 |
static bool table##_rehash(struct table *table, size_t orig_chunk_count, \ |
| 992 |
size_t orig_capacity_scale, size_t new_chunk_count, \ |
| 993 |
size_t new_capacity_scale) \ |
| 994 |
{ \ |
| 995 |
size_t chunk_alloc_size = table##_chunk_alloc_size(new_chunk_count, \ |
| 996 |
new_capacity_scale); \ |
| 997 |
size_t alloc_size, entries_offset; \ |
| 998 |
if (table##_vector_policy) { \ |
| 999 |
entries_offset = chunk_alloc_size; \ |
| 1000 |
if (alignof(table##_entry_type) > alignof(table##_item_type)) { \ |
| 1001 |
entries_offset = -(-entries_offset & \ |
| 1002 |
~(alignof(table##_entry_type) - 1)); \ |
| 1003 |
} \ |
| 1004 |
size_t new_capacity = \ |
| 1005 |
table##_compute_capacity(new_chunk_count, \ |
| 1006 |
new_capacity_scale); \ |
| 1007 |
alloc_size = (entries_offset + \ |
| 1008 |
new_capacity * sizeof(table##_entry_type)); \ |
| 1009 |
} else { \ |
| 1010 |
alloc_size = chunk_alloc_size; \ |
| 1011 |
} \ |
| 1012 |
\ |
| 1013 |
void *new_chunks; \ |
| 1014 |
if (posix_memalign(&new_chunks, hash_table_chunk_alignment, alloc_size))\ |
| 1015 |
return false; \ |
| 1016 |
\ |
| 1017 |
struct table##_chunk *orig_chunks = table->chunks; \ |
| 1018 |
table->chunks = new_chunks; \ |
| 1019 |
table##_entry_type *orig_entries; \ |
| 1020 |
if (table##_vector_policy) { \ |
| 1021 |
orig_entries = table->vector; \ |
| 1022 |
table->vector = (void *)((char *)new_chunks + entries_offset); \ |
| 1023 |
if (table##_size(table) > 0) { \ |
| 1024 |
memcpy(table->vector, orig_entries, \ |
| 1025 |
table##_size(table) * \ |
| 1026 |
sizeof(table##_entry_type)); \ |
| 1027 |
} \ |
| 1028 |
} \ |
| 1029 |
\ |
| 1030 |
memset(table->chunks, 0, chunk_alloc_size); \ |
| 1031 |
table##_chunk_mark_eof(table->chunks, new_capacity_scale); \ |
| 1032 |
table##_set_chunk_count(table, new_chunk_count); \ |
| 1033 |
\ |
| 1034 |
if (table##_size(table) == 0) { \ |
| 1035 |
/* Nothing to do. */ \ |
| 1036 |
} else if (orig_chunk_count == 1 && new_chunk_count == 1) { \ |
| 1037 |
struct table##_chunk *src = orig_chunks; \ |
| 1038 |
struct table##_chunk *dst = table->chunks; \ |
| 1039 |
size_t src_i = 0, dst_i = 0; \ |
| 1040 |
while (dst_i < table##_size(table)) { \ |
| 1041 |
if (likely(src->tags[src_i])) { \ |
| 1042 |
dst->tags[dst_i] = src->tags[src_i]; \ |
| 1043 |
memcpy(&dst->items[dst_i], &src->items[src_i], \ |
| 1044 |
sizeof(dst->items[dst_i])); \ |
| 1045 |
dst_i++; \ |
| 1046 |
} \ |
| 1047 |
src_i++; \ |
| 1048 |
} \ |
| 1049 |
if (!table##_vector_policy) { \ |
| 1050 |
table->first_packed = \ |
| 1051 |
table##_pack_iterator(dst, dst_i - 1); \ |
| 1052 |
} \ |
| 1053 |
} else { \ |
| 1054 |
uint8_t stack_fullness[256]; \ |
| 1055 |
uint8_t *fullness; \ |
| 1056 |
if (new_chunk_count <= sizeof(stack_fullness)) { \ |
| 1057 |
memset(stack_fullness, 0, sizeof(stack_fullness)); \ |
| 1058 |
fullness = stack_fullness; \ |
| 1059 |
} else { \ |
| 1060 |
fullness = calloc(new_chunk_count, 1); \ |
| 1061 |
if (!fullness) \ |
| 1062 |
goto err; \ |
| 1063 |
} \ |
| 1064 |
\ |
| 1065 |
struct table##_chunk *src = &orig_chunks[orig_chunk_count - 1]; \ |
| 1066 |
size_t remaining = table##_size(table); \ |
| 1067 |
while (remaining) { \ |
| 1068 |
unsigned int mask = table##_chunk_occupied(src), i; \ |
| 1069 |
if (table##_vector_policy) { \ |
| 1070 |
unsigned int pmask = mask; \ |
| 1071 |
for_each_bit(i, pmask) { \ |
| 1072 |
table##_item_type *item = \ |
| 1073 |
&src->items[i]; \ |
| 1074 |
table##_entry_type *entry = \ |
| 1075 |
table##_item_to_entry(table, \ |
| 1076 |
item); \ |
| 1077 |
__builtin_prefetch(entry); \ |
| 1078 |
} \ |
| 1079 |
} \ |
| 1080 |
for_each_bit(i, mask) { \ |
| 1081 |
remaining--; \ |
| 1082 |
\ |
| 1083 |
table##_item_type *src_item = &src->items[i]; \ |
| 1084 |
table##_key_type key = \ |
| 1085 |
table##_item_to_key(table, src_item); \ |
| 1086 |
struct hash_pair hp = table##_hash(&key); \ |
| 1087 |
table##_item_type *dst_item = \ |
| 1088 |
table##_allocate_tag(table, fullness, \ |
| 1089 |
hp); \ |
| 1090 |
memcpy(dst_item, src_item, sizeof(*dst_item)); \ |
| 1091 |
} \ |
| 1092 |
src--; \ |
| 1093 |
} \ |
| 1094 |
\ |
| 1095 |
if (!table##_vector_policy) { \ |
| 1096 |
size_t i = table##_chunk_count(table) - 1; \ |
| 1097 |
while (fullness[i] == 0) \ |
| 1098 |
i--; \ |
| 1099 |
table->first_packed = \ |
| 1100 |
table##_pack_iterator(&table->chunks[i], \ |
| 1101 |
fullness[i] - 1); \ |
| 1102 |
} \ |
| 1103 |
\ |
| 1104 |
if (fullness != stack_fullness) \ |
| 1105 |
free(fullness); \ |
| 1106 |
} \ |
| 1107 |
\ |
| 1108 |
if (orig_chunks != hash_table_empty_chunk) \ |
| 1109 |
free(orig_chunks); \ |
| 1110 |
return true; \ |
| 1111 |
\ |
| 1112 |
err: \ |
| 1113 |
free(table->chunks); \ |
| 1114 |
table->chunks = orig_chunks; \ |
| 1115 |
table##_set_chunk_count(table, orig_chunk_count); \ |
| 1116 |
if (table##_vector_policy) \ |
| 1117 |
table->vector = orig_entries; \ |
| 1118 |
return false; \ |
| 1119 |
} \ |
| 1120 |
\ |
| 1121 |
static void table##_do_clear(struct table *table, bool reset) \ |
| 1122 |
{ \ |
| 1123 |
if (table->chunks == hash_table_empty_chunk) \ |
| 1124 |
return; \ |
| 1125 |
\ |
| 1126 |
size_t chunk_count = table##_chunk_count(table); \ |
| 1127 |
/* Always reset large tables. */ \ |
| 1128 |
if (chunk_count >= 16) \ |
| 1129 |
reset = true; \ |
| 1130 |
if (!table##_empty(table)) { \ |
| 1131 |
if (!reset) { \ |
| 1132 |
size_t capacity_scale = \ |
| 1133 |
table##_chunk_capacity_scale(table->chunks); \ |
| 1134 |
memset(table->chunks, 0, \ |
| 1135 |
table##_chunk_alloc_size(chunk_count, \ |
| 1136 |
capacity_scale)); \ |
| 1137 |
table##_chunk_mark_eof(table->chunks, capacity_scale); \ |
| 1138 |
} \ |
| 1139 |
if (!table##_vector_policy) \ |
| 1140 |
table->first_packed = 0; \ |
| 1141 |
table##_set_size(table, 0); \ |
| 1142 |
} \ |
| 1143 |
if (reset) { \ |
| 1144 |
free(table->chunks); \ |
| 1145 |
table->chunks = hash_table_empty_chunk; \ |
| 1146 |
table##_set_chunk_count(table, 1); \ |
| 1147 |
if (table##_vector_policy) \ |
| 1148 |
table->vector = NULL; \ |
| 1149 |
} \ |
| 1150 |
} \ |
| 1151 |
\ |
| 1152 |
__attribute__((__unused__)) \ |
| 1153 |
static bool table##_reserve(struct table *table, size_t capacity) \ |
| 1154 |
{ \ |
| 1155 |
capacity = max(capacity, table##_size(table)); \ |
| 1156 |
if (!capacity) { \ |
| 1157 |
table##_do_clear(table, true); \ |
| 1158 |
return true; \ |
| 1159 |
} \ |
| 1160 |
\ |
| 1161 |
size_t orig_chunk_count = table##_chunk_count(table); \ |
| 1162 |
size_t orig_capacity_scale = table##_chunk_capacity_scale(table->chunks);\ |
| 1163 |
size_t orig_capacity = table##_compute_capacity(orig_chunk_count, \ |
| 1164 |
orig_capacity_scale); \ |
| 1165 |
\ |
| 1166 |
/* \ |
| 1167 |
* To avoid pathological behavior, ignore decreases that aren't at \ |
| 1168 |
* least a 1/8 decrease, and double for increases that aren't at least \ |
| 1169 |
* a 1/8 increase. \ |
| 1170 |
*/ \ |
| 1171 |
if (capacity <= orig_capacity && \ |
| 1172 |
capacity >= orig_capacity - orig_capacity / 8) \ |
| 1173 |
return true; \ |
| 1174 |
bool attempt_exact = !(capacity > orig_capacity && \ |
| 1175 |
capacity < orig_capacity + orig_capacity / 8); \ |
| 1176 |
\ |
| 1177 |
size_t new_chunk_count; \ |
| 1178 |
size_t new_capacity_scale; \ |
| 1179 |
if (!table##_compute_chunk_count_and_scale(capacity, attempt_exact, \ |
| 1180 |
table##_vector_policy && \ |
| 1181 |
attempt_exact, \ |
| 1182 |
&new_chunk_count, \ |
| 1183 |
&new_capacity_scale)) \ |
| 1184 |
return false; \ |
| 1185 |
size_t new_capacity = table##_compute_capacity(new_chunk_count, \ |
| 1186 |
new_capacity_scale); \ |
| 1187 |
if (new_capacity == orig_capacity) \ |
| 1188 |
return true; \ |
| 1189 |
return table##_rehash(table, orig_chunk_count, orig_capacity_scale, \ |
| 1190 |
new_chunk_count, new_capacity_scale); \ |
| 1191 |
} \ |
| 1192 |
\ |
| 1193 |
__attribute__((__unused__)) \ |
| 1194 |
static void table##_clear(struct table *table) \ |
| 1195 |
{ \ |
| 1196 |
table##_do_clear(table, false); \ |
| 1197 |
} \ |
| 1198 |
\ |
| 1199 |
\ |
| 1200 |
HASH_TABLE_SEARCH_IMPL(table, search_by_key, table##_key_type, \ |
| 1201 |
table##_item_to_key, eq_func) \ |
| 1202 |
HASH_TABLE_SEARCH_IMPL(table, search_by_index, uint32_t, \ |
| 1203 |
HASH_TABLE_SEARCH_BY_INDEX_ITEM_TO_KEY, scalar_key_eq) \ |
| 1204 |
\ |
| 1205 |
\ |
| 1206 |
static struct table##_iterator \ |
| 1207 |
table##_search_hashed(struct table *table, const table##_key_type *key, \ |
| 1208 |
struct hash_pair hp) \ |
| 1209 |
{ \ |
| 1210 |
struct table##_iterator it = table##_search_by_key(table, key, hp); \ |
| 1211 |
/* Convert the item iterator to a public iterator. */ \ |
| 1212 |
if (table##_vector_policy && it.item) { \ |
| 1213 |
it.entry = table##_item_to_entry(table, it.item); \ |
| 1214 |
it.lowest = table->vector; \ |
| 1215 |
} \ |
| 1216 |
return it; \ |
| 1217 |
} \ |
| 1218 |
\ |
| 1219 |
__attribute__((__unused__)) \ |
| 1220 |
static struct table##_iterator \ |
| 1221 |
table##_search(struct table *table, const table##_key_type *key) \ |
| 1222 |
{ \ |
| 1223 |
return table##_search_hashed(table, key, table##_hash(key)); \ |
| 1224 |
} \ |
| 1225 |
\ |
| 1226 |
static bool table##_reserve_for_insert(struct table *table) \ |
| 1227 |
{ \ |
| 1228 |
size_t orig_chunk_count = table##_chunk_count(table); \ |
| 1229 |
size_t orig_capacity_scale = table##_chunk_capacity_scale(table->chunks);\ |
| 1230 |
size_t orig_capacity = table##_compute_capacity(orig_chunk_count, \ |
| 1231 |
orig_capacity_scale); \ |
| 1232 |
size_t capacity = table##_size(table) + 1; \ |
| 1233 |
if (capacity <= orig_capacity) \ |
| 1234 |
return true; \ |
| 1235 |
/* Grow by at least orig_capacity * 2^0.5. */ \ |
| 1236 |
size_t min_growth = (orig_capacity + \ |
| 1237 |
(orig_capacity >> 2) + \ |
| 1238 |
(orig_capacity >> 3) + \ |
| 1239 |
(orig_capacity >> 5)); \ |
| 1240 |
capacity = max(capacity, min_growth); \ |
| 1241 |
size_t new_chunk_count, new_capacity_scale; \ |
| 1242 |
if (!table##_compute_chunk_count_and_scale(capacity, false, false, \ |
| 1243 |
&new_chunk_count, \ |
| 1244 |
&new_capacity_scale)) \ |
| 1245 |
return false; \ |
| 1246 |
return table##_rehash(table, orig_chunk_count, orig_capacity_scale, \ |
| 1247 |
new_chunk_count, new_capacity_scale); \ |
| 1248 |
} \ |
| 1249 |
\ |
| 1250 |
static void \ |
| 1251 |
table##_adjust_size_and_first_after_insert(struct table *table, \ |
| 1252 |
struct table##_chunk *chunk, \ |
| 1253 |
size_t index) \ |
| 1254 |
{ \ |
| 1255 |
if (!table##_vector_policy) { \ |
| 1256 |
uintptr_t first_packed = table##_pack_iterator(chunk, index); \ |
| 1257 |
if (first_packed > table->first_packed) \ |
| 1258 |
table->first_packed = first_packed; \ |
| 1259 |
} \ |
| 1260 |
table##_increment_size(table); \ |
| 1261 |
} \ |
| 1262 |
\ |
| 1263 |
static int table##_insert_searched(struct table *table, \ |
| 1264 |
const table##_entry_type *entry, \ |
| 1265 |
struct hash_pair hp, \ |
| 1266 |
struct table##_iterator *it_ret) \ |
| 1267 |
{ \ |
| 1268 |
if (!table##_reserve_for_insert(table)) \ |
| 1269 |
return -1; \ |
| 1270 |
\ |
| 1271 |
size_t index = hp.first; \ |
| 1272 |
struct table##_chunk *chunk = \ |
| 1273 |
&table->chunks[table##_modulo_by_chunk_count(table, index)]; \ |
| 1274 |
unsigned int first_empty = table##_chunk_first_empty(chunk); \ |
| 1275 |
if (first_empty == (unsigned int)-1) { \ |
| 1276 |
size_t delta = hash_table_probe_delta(hp); \ |
| 1277 |
do { \ |
| 1278 |
table##_chunk_inc_outbound_overflow_count(chunk); \ |
| 1279 |
index += delta; \ |
| 1280 |
chunk = &table->chunks[table##_modulo_by_chunk_count(table, index)];\ |
| 1281 |
first_empty = table##_chunk_first_empty(chunk); \ |
| 1282 |
} while (first_empty == (unsigned int)-1); \ |
| 1283 |
table##_chunk_adjust_hosted_overflow_count(chunk, \ |
| 1284 |
hosted_overflow_count_inc);\ |
| 1285 |
} \ |
| 1286 |
chunk->tags[first_empty] = hp.second; \ |
| 1287 |
if (table##_vector_policy) { \ |
| 1288 |
*table##_vector_item(&chunk->items[first_empty]) = \ |
| 1289 |
table##_size(table); \ |
| 1290 |
memcpy(&table->vector[table##_size(table)], entry, \ |
| 1291 |
sizeof(*entry)); \ |
| 1292 |
} else { \ |
| 1293 |
memcpy(&chunk->items[first_empty], entry, sizeof(*entry)); \ |
| 1294 |
} \ |
| 1295 |
table##_adjust_size_and_first_after_insert(table, chunk, first_empty); \ |
| 1296 |
if (it_ret) { \ |
| 1297 |
if (table##_vector_policy) { \ |
| 1298 |
it_ret->entry = \ |
| 1299 |
&table->vector[table##_size(table) - 1]; \ |
| 1300 |
it_ret->lowest = table->vector; \ |
| 1301 |
} else { \ |
| 1302 |
it_ret->item = &chunk->items[first_empty]; \ |
| 1303 |
it_ret->index = first_empty; \ |
| 1304 |
} \ |
| 1305 |
} \ |
| 1306 |
return 1; \ |
| 1307 |
} \ |
| 1308 |
\ |
| 1309 |
static int table##_insert_hashed(struct table *table, \ |
| 1310 |
const table##_entry_type *entry, \ |
| 1311 |
struct hash_pair hp, \ |
| 1312 |
struct table##_iterator *it_ret) \ |
| 1313 |
{ \ |
| 1314 |
table##_key_type key = table##_entry_to_key(entry); \ |
| 1315 |
struct table##_iterator it = table##_search_hashed(table, &key, hp); \ |
| 1316 |
if (it.entry) { \ |
| 1317 |
if (it_ret) \ |
| 1318 |
*it_ret = it; \ |
| 1319 |
return 0; \ |
| 1320 |
} else { \ |
| 1321 |
return table##_insert_searched(table, entry, hp, it_ret); \ |
| 1322 |
} \ |
| 1323 |
} \ |
| 1324 |
\ |
| 1325 |
__attribute__((__unused__)) \ |
| 1326 |
static int table##_insert(struct table *table, \ |
| 1327 |
const table##_entry_type *entry, \ |
| 1328 |
struct table##_iterator *it_ret) \ |
| 1329 |
{ \ |
| 1330 |
table##_key_type key = table##_entry_to_key(entry); \ |
| 1331 |
return table##_insert_hashed(table, entry, table##_hash(&key), it_ret); \ |
| 1332 |
} \ |
| 1333 |
\ |
| 1334 |
/* Similar to table##_next_impl() but for the cached first position. */ \ |
| 1335 |
static void table##_advance_first_packed(struct table *table) \ |
| 1336 |
{ \ |
| 1337 |
uintptr_t packed = table->first_packed; \ |
| 1338 |
struct table##_chunk *chunk = table##_unpack_chunk(packed); \ |
| 1339 |
size_t index = table##_unpack_index(packed); \ |
| 1340 |
while (index > 0) { \ |
| 1341 |
index--; \ |
| 1342 |
if (chunk->tags[index]) { \ |
| 1343 |
table->first_packed = \ |
| 1344 |
table##_pack_iterator(chunk, index); \ |
| 1345 |
return; \ |
| 1346 |
} \ |
| 1347 |
} \ |
| 1348 |
\ |
| 1349 |
/* \ |
| 1350 |
* This is only called when there is another entry in the table, so we \ |
| 1351 |
* don't need to check if we hit the end. \ |
| 1352 |
*/ \ |
| 1353 |
for (;;) { \ |
| 1354 |
chunk--; \ |
| 1355 |
unsigned int last = table##_chunk_last_occupied(chunk); \ |
| 1356 |
if (last != (unsigned int)-1) { \ |
| 1357 |
table->first_packed = \ |
| 1358 |
table##_pack_iterator(chunk, last); \ |
| 1359 |
return; \ |
| 1360 |
} \ |
| 1361 |
} \ |
| 1362 |
} \ |
| 1363 |
\ |
| 1364 |
static void \ |
| 1365 |
table##_adjust_size_and_first_before_delete(struct table *table, \ |
| 1366 |
struct table##_chunk *chunk, \ |
| 1367 |
size_t index) \ |
| 1368 |
{ \ |
| 1369 |
table##_decrement_size(table); \ |
| 1370 |
if (!table##_vector_policy && \ |
| 1371 |
table##_pack_iterator(chunk, index) == table->first_packed) { \ |
| 1372 |
if (table##_empty(table)) \ |
| 1373 |
table->first_packed = 0; \ |
| 1374 |
else \ |
| 1375 |
table##_advance_first_packed(table); \ |
| 1376 |
} \ |
| 1377 |
} \ |
| 1378 |
\ |
| 1379 |
/* \ |
| 1380 |
* We want this inlined so that the whole function call can be optimized away \ |
| 1381 |
* in the likely_dead case, and so that the counter can be optimized away in \ |
| 1382 |
* the not likely_dead case. \ |
| 1383 |
*/ \ |
| 1384 |
__attribute__((__always_inline__)) \ |
| 1385 |
static inline struct table##_iterator \ |
| 1386 |
table##_next_impl(struct table##_iterator it, bool likely_dead) \ |
| 1387 |
{ \ |
| 1388 |
struct table##_chunk *chunk = table##_iterator_chunk(it); \ |
| 1389 |
while (it.index > 0) { \ |
| 1390 |
it.index--; \ |
| 1391 |
it.entry--; \ |
| 1392 |
if (likely(chunk->tags[it.index])) \ |
| 1393 |
return it; \ |
| 1394 |
} \ |
| 1395 |
\ |
| 1396 |
/* \ |
| 1397 |
* This hack is copied from the folly implementation: this is dead code \ |
| 1398 |
* if the return value is not used (e.g., the return value of \ |
| 1399 |
* table##_delete_iterator() is often ignored), but the compiler needs \ |
| 1400 |
* some help proving that the following loop terminates. \ |
| 1401 |
*/ \ |
| 1402 |
for (size_t i = 1; !likely_dead || i != 0; i++) { \ |
| 1403 |
if (unlikely(table##_chunk_eof(chunk))) \ |
| 1404 |
break; \ |
| 1405 |
\ |
| 1406 |
chunk--; \ |
| 1407 |
unsigned int last = table##_chunk_last_occupied(chunk); \ |
| 1408 |
if (!likely_dead) \ |
| 1409 |
__builtin_prefetch(chunk - 1); \ |
| 1410 |
if (likely(last != (unsigned int)-1)) { \ |
| 1411 |
it.index = last; \ |
| 1412 |
it.item = &chunk->items[last]; \ |
| 1413 |
return it; \ |
| 1414 |
} \ |
| 1415 |
} \ |
| 1416 |
return (struct table##_iterator){}; \ |
| 1417 |
} \ |
| 1418 |
\ |
| 1419 |
static void table##_delete_impl(struct table *table, \ |
| 1420 |
struct table##_iterator item_it, \ |
| 1421 |
struct hash_pair hp) \ |
| 1422 |
{ \ |
| 1423 |
struct table##_chunk *it_chunk = table##_iterator_chunk(item_it); \ |
| 1424 |
it_chunk->tags[item_it.index] = 0; \ |
| 1425 |
\ |
| 1426 |
table##_adjust_size_and_first_before_delete(table, it_chunk, \ |
| 1427 |
item_it.index); \ |
| 1428 |
\ |
| 1429 |
if (table##_chunk_hosted_overflow_count(it_chunk)) { \ |
| 1430 |
const size_t delta = hash_table_probe_delta(hp); \ |
| 1431 |
size_t index = hp.first; \ |
| 1432 |
uint8_t hosted_op = 0; \ |
| 1433 |
for (;;) { \ |
| 1434 |
struct table##_chunk *chunk = \ |
| 1435 |
&table->chunks[table##_modulo_by_chunk_count(table, index)];\ |
| 1436 |
if (chunk == it_chunk) { \ |
| 1437 |
table##_chunk_adjust_hosted_overflow_count(chunk,\ |
| 1438 |
hosted_op);\ |
| 1439 |
break; \ |
| 1440 |
} \ |
| 1441 |
table##_chunk_dec_outbound_overflow_count(chunk); \ |
| 1442 |
hosted_op = hosted_overflow_count_dec; \ |
| 1443 |
index += delta; \ |
| 1444 |
} \ |
| 1445 |
} \ |
| 1446 |
} \ |
| 1447 |
\ |
| 1448 |
static void table##_vector_delete_impl(struct table *table, \ |
| 1449 |
struct table##_iterator item_it, \ |
| 1450 |
struct hash_pair hp) \ |
| 1451 |
{ \ |
| 1452 |
/* Delete the index from the table. */ \ |
| 1453 |
uint32_t index = *table##_vector_item(item_it.item); \ |
| 1454 |
table##_delete_impl(table, item_it, hp); \ |
| 1455 |
\ |
| 1456 |
/* Replace it with the last entry and update its index in the table. */ \ |
| 1457 |
uint32_t tail_index = table##_size(table); \ |
| 1458 |
if (tail_index != index) { \ |
| 1459 |
table##_entry_type *tail = \ |
| 1460 |
&table->vector[tail_index]; \ |
| 1461 |
table##_key_type tail_key = table##_entry_to_key(tail); \ |
| 1462 |
item_it = table##_search_by_index(table, &tail_index, \ |
| 1463 |
table##_hash(&tail_key)); \ |
| 1464 |
*table##_vector_item(item_it.item) = index; \ |
| 1465 |
memcpy(&table->vector[index], tail, sizeof(*tail)); \ |
| 1466 |
} \ |
| 1467 |
} \ |
| 1468 |
\ |
| 1469 |
/* \ |
| 1470 |
* We want this inlined so that the call to table##_next_impl() can be \ |
| 1471 |
* optimized away. \ |
| 1472 |
*/ \ |
| 1473 |
__attribute__((__always_inline__)) \ |
| 1474 |
static inline struct table##_iterator \ |
| 1475 |
table##_delete_iterator_hashed(struct table *table, struct table##_iterator it, \ |
| 1476 |
struct hash_pair hp) \ |
| 1477 |
{ \ |
| 1478 |
if (table##_vector_policy) { \ |
| 1479 |
uint32_t index = it.entry - it.lowest; \ |
| 1480 |
struct table##_iterator item_it = \ |
| 1481 |
table##_search_by_index(table, &index, hp); \ |
| 1482 |
table##_vector_delete_impl(table, item_it, hp); \ |
| 1483 |
if (index == 0) { \ |
| 1484 |
return (struct table##_iterator){}; \ |
| 1485 |
} else { \ |
| 1486 |
it.entry--; \ |
| 1487 |
return it; \ |
| 1488 |
} \ |
| 1489 |
} else { \ |
| 1490 |
table##_delete_impl(table, it, hp); \ |
| 1491 |
return table##_next_impl(it, true); \ |
| 1492 |
} \ |
| 1493 |
} \ |
| 1494 |
\ |
| 1495 |
__attribute__((__always_inline__, __unused__)) \ |
| 1496 |
static inline struct table##_iterator \ |
| 1497 |
table##_delete_iterator(struct table *table, struct table##_iterator it) \ |
| 1498 |
{ \ |
| 1499 |
struct hash_pair hp = {}; \ |
| 1500 |
/* \ |
| 1501 |
* The basic policy only needs the hash if the chunk hosts an \ |
| 1502 |
* overflowed entry. \ |
| 1503 |
*/ \ |
| 1504 |
if (table##_vector_policy || \ |
| 1505 |
table##_chunk_hosted_overflow_count(table##_iterator_chunk(it))) { \ |
| 1506 |
table##_key_type key = table##_entry_to_key(it.entry); \ |
| 1507 |
hp = table##_hash(&key); \ |
| 1508 |
} \ |
| 1509 |
return table##_delete_iterator_hashed(table, it, hp); \ |
| 1510 |
} \ |
| 1511 |
\ |
| 1512 |
static bool table##_delete_hashed(struct table *table, \ |
| 1513 |
const table##_key_type *key, \ |
| 1514 |
struct hash_pair hp) \ |
| 1515 |
{ \ |
| 1516 |
struct table##_iterator item_it = table##_search_by_key(table, key, hp);\ |
| 1517 |
if (!item_it.item) \ |
| 1518 |
return false; \ |
| 1519 |
if (table##_vector_policy) \ |
| 1520 |
table##_vector_delete_impl(table, item_it, hp); \ |
| 1521 |
else \ |
| 1522 |
table##_delete_impl(table, item_it, hp); \ |
| 1523 |
return true; \ |
| 1524 |
} \ |
| 1525 |
\ |
| 1526 |
static bool table##_delete(struct table *table, const table##_key_type *key) \ |
| 1527 |
{ \ |
| 1528 |
return table##_delete_hashed(table, key, table##_hash(key)); \ |
| 1529 |
} \ |
| 1530 |
\ |
| 1531 |
__attribute__((__unused__)) \ |
| 1532 |
static inline bool table##_delete_entry(struct table *table, \ |
| 1533 |
const table##_entry_type *entry) \ |
| 1534 |
{ \ |
| 1535 |
const table##_key_type key = table##_entry_to_key(entry); \ |
| 1536 |
return table##_delete(table, &key); \ |
| 1537 |
} \ |
| 1538 |
\ |
| 1539 |
__attribute__((__unused__)) \ |
| 1540 |
static struct table##_iterator table##_first(struct table *table) \ |
| 1541 |
{ \ |
| 1542 |
if (table##_vector_policy) { \ |
| 1543 |
table##_entry_type *entry; \ |
| 1544 |
if (table##_empty(table)) \ |
| 1545 |
entry = NULL; \ |
| 1546 |
else \ |
| 1547 |
entry = &table->vector[table##_size(table) - 1]; \ |
| 1548 |
return (struct table##_iterator){ \ |
| 1549 |
.entry = entry, \ |
| 1550 |
.lowest = table->vector, \ |
| 1551 |
}; \ |
| 1552 |
} else { \ |
| 1553 |
return table##_unpack_iterator(table->first_packed); \ |
| 1554 |
} \ |
| 1555 |
} \ |
| 1556 |
\ |
| 1557 |
__attribute__((__unused__)) \ |
| 1558 |
static struct table##_iterator table##_next(struct table##_iterator it) \ |
| 1559 |
{ \ |
| 1560 |
if (table##_vector_policy) { \ |
| 1561 |
if (it.entry == it.lowest) { \ |
| 1562 |
return (struct table##_iterator){}; \ |
| 1563 |
} else { \ |
| 1564 |
it.entry--; \ |
| 1565 |
return it; \ |
| 1566 |
} \ |
| 1567 |
} else { \ |
| 1568 |
return table##_next_impl(it, false); \ |
| 1569 |
} \ |
| 1570 |
} \ |
| 1571 |
struct DEFINE_HASH_TABLE_needs_semicolon |
| 1572 |
|
| 1573 |
/** |
| 1574 |
* Define a hash table interface. |
| 1575 |
* |
| 1576 |
* This macro defines a hash table type along with its functions. |
| 1577 |
* |
| 1578 |
* @param[in] table Name of the type to define. This is prefixed to all of the |
| 1579 |
* types and functions defined for that type. |
| 1580 |
* @param[in] entry_type Type of entries in the table. |
| 1581 |
* @param[in] entry_to_key Name of function or macro which is passed a <tt>const |
| 1582 |
* entry_type *</tt> and returns the key for that entry. The return type is the |
| 1583 |
* @c key_type of the hash table. The passed entry is never @c NULL. |
| 1584 |
* @param[in] hash_func Hash function which takes a <tt>const key_type *</tt> |
| 1585 |
* and returns a @ref hash_pair. |
| 1586 |
* @param[in] eq_func Comparison function which takes two <tt>const key_type |
| 1587 |
* *</tt> and returns a @c bool. |
| 1588 |
*/ |
| 1589 |
#define DEFINE_HASH_TABLE(table, entry_type, entry_to_key, hash_func, eq_func) \ |
| 1590 |
DEFINE_HASH_TABLE_TYPE(table, entry_type); \ |
| 1591 |
DEFINE_HASH_TABLE_FUNCTIONS(table, entry_to_key, hash_func, eq_func) |
| 1592 |
|
| 1593 |
/** |
| 1594 |
* Define a hash map type without defining its functions. |
| 1595 |
* |
| 1596 |
* The functions are defined with @ref DEFINE_HASH_MAP_FUNCTIONS(). |
| 1597 |
* |
| 1598 |
* @sa DEFINE_HASH_MAP(), DEFINE_HASH_TABLE_TYPE() |
| 1599 |
*/ |
| 1600 |
#define DEFINE_HASH_MAP_TYPE(table, key_type, value_type) \ |
| 1601 |
struct table##_entry { \ |
| 1602 |
typeof(key_type) key; \ |
| 1603 |
typeof(value_type) value; \ |
| 1604 |
}; \ |
| 1605 |
DEFINE_HASH_TABLE_TYPE(table, struct table##_entry) |
| 1606 |
|
| 1607 |
#define HASH_MAP_ENTRY_TO_KEY(entry) ((entry)->key) |
| 1608 |
|
| 1609 |
/** |
| 1610 |
* Define the functions for a hash map. |
| 1611 |
* |
| 1612 |
* The hash map type must have already been defined with @ref |
| 1613 |
* DEFINE_HASH_MAP_TYPE(). |
| 1614 |
* |
| 1615 |
* Unless the type and function definitions must be in separate places, use @ref |
| 1616 |
* DEFINE_HASH_MAP() instead. |
| 1617 |
* |
| 1618 |
* @sa DEFINE_HASH_TABLE_FUNCTIONS |
| 1619 |
*/ |
| 1620 |
#define DEFINE_HASH_MAP_FUNCTIONS(table, hash_func, eq_func) \ |
| 1621 |
DEFINE_HASH_TABLE_FUNCTIONS(table, HASH_MAP_ENTRY_TO_KEY, hash_func, eq_func) |
| 1622 |
|
| 1623 |
/** |
| 1624 |
* Define a hash map interface. |
| 1625 |
* |
| 1626 |
* This is a higher-level wrapper for @ref DEFINE_HASH_TABLE() with entries of |
| 1627 |
* the following type (with the example name @c hash_map): |
| 1628 |
* |
| 1629 |
* @code{.c} |
| 1630 |
* struct hash_map_entry { |
| 1631 |
* key_type key; |
| 1632 |
* value_type value; |
| 1633 |
* }; |
| 1634 |
* @endcode |
| 1635 |
* |
| 1636 |
* @param[in] table Name of the map type to define. This is prefixed to all of |
| 1637 |
* the types and functions defined for that type. |
| 1638 |
* @param[in] key_type Type of keys in the map. |
| 1639 |
* @param[in] value_type Type of values in the map. |
| 1640 |
* @param[in] hash_func See @ref DEFINE_HASH_TABLE(). |
| 1641 |
* @param[in] eq_func See @ref DEFINE_HASH_TABLE(). |
| 1642 |
*/ |
| 1643 |
#define DEFINE_HASH_MAP(table, key_type, value_type, hash_func, eq_func) \ |
| 1644 |
DEFINE_HASH_MAP_TYPE(table, key_type, value_type); \ |
| 1645 |
DEFINE_HASH_MAP_FUNCTIONS(table, hash_func, eq_func) |
| 1646 |
|
| 1647 |
/** |
| 1648 |
* Define a hash set type without defining its functions. |
| 1649 |
* |
| 1650 |
* The functions are defined with @ref DEFINE_HASH_SET_FUNCTIONS(). |
| 1651 |
* |
| 1652 |
* @sa DEFINE_HASH_SET(), DEFINE_HASH_TABLE_TYPE() |
| 1653 |
*/ |
| 1654 |
#define DEFINE_HASH_SET_TYPE DEFINE_HASH_TABLE_TYPE |
| 1655 |
|
| 1656 |
#define HASH_SET_ENTRY_TO_KEY(entry) (*(entry)) |
| 1657 |
|
| 1658 |
/** |
| 1659 |
* Define the functions for a hash set. |
| 1660 |
* |
| 1661 |
* The hash set type must have already been defined with @ref |
| 1662 |
* DEFINE_HASH_SET_TYPE(). |
| 1663 |
* |
| 1664 |
* Unless the type and function definitions must be in separate places, use @ref |
| 1665 |
* DEFINE_HASH_SET() instead. |
| 1666 |
* |
| 1667 |
* @sa DEFINE_HASH_TABLE_FUNCTIONS |
| 1668 |
*/ |
| 1669 |
#define DEFINE_HASH_SET_FUNCTIONS(table, hash_func, eq_func) \ |
| 1670 |
DEFINE_HASH_TABLE_FUNCTIONS(table, HASH_SET_ENTRY_TO_KEY, hash_func, eq_func) |
| 1671 |
|
| 1672 |
/** |
| 1673 |
* Define a hash set interface. |
| 1674 |
* |
| 1675 |
* This is a higher-level wrapper for @ref DEFINE_HASH_TABLE() where @p |
| 1676 |
* entry_type is the same as @p key_type. |
| 1677 |
* |
| 1678 |
* @param[in] table Name of the set type to define. This is prefixed to all of |
| 1679 |
* the types and functions defined for that type. |
| 1680 |
* @param[in] key_type Type of keys in the set. |
| 1681 |
* @param[in] hash_func See @ref DEFINE_HASH_TABLE(). |
| 1682 |
* @param[in] eq_func See @ref DEFINE_HASH_TABLE(). |
| 1683 |
*/ |
| 1684 |
#define DEFINE_HASH_SET(table, key_type, hash_func, eq_func) \ |
| 1685 |
DEFINE_HASH_SET_TYPE(table, key_type); \ |
| 1686 |
DEFINE_HASH_SET_FUNCTIONS(table, hash_func, eq_func) |
| 1687 |
|
| 1688 |
/** |
| 1689 |
* Empty hash table initializer. |
| 1690 |
* |
| 1691 |
* This can be used to initialize a hash table when declaring it. |
| 1692 |
* |
| 1693 |
* @sa hash_table_init() |
| 1694 |
*/ |
| 1695 |
#define HASH_TABLE_INIT { hash_table_empty_chunk } |
| 1696 |
|
| 1697 |
/** |
| 1698 |
* Define and initialize an empty @ref hash_table of type @p table_type named @p |
| 1699 |
* table that is automatically deinitialized when it goes out of scope. |
| 1700 |
*/ |
| 1701 |
#define HASH_TABLE(table_type, table) \ |
| 1702 |
__attribute__((__cleanup__(table_type##_deinit))) \ |
| 1703 |
struct table_type table = HASH_TABLE_INIT |
| 1704 |
|
| 1705 |
/** |
| 1706 |
* Iterate over every entry in a @ref hash_table. |
| 1707 |
* |
| 1708 |
* @param[in] table_type Name of hash table type. |
| 1709 |
* @param[out] it Name of iterator variable. |
| 1710 |
* @param[in] table Hash table to iterate over. |
| 1711 |
*/ |
| 1712 |
#define hash_table_for_each(table_type, it, table) \ |
| 1713 |
for (struct table_type##_iterator it = table_type##_first(table); \ |
| 1714 |
it.entry; it = table_type##_next(it)) |
| 1715 |
|
| 1716 |
/** |
| 1717 |
* @defgroup HashTableHelpers Hash table helpers |
| 1718 |
* |
| 1719 |
* Hash functions and comparators for use with @ref HashTables. |
| 1720 |
* |
| 1721 |
* F14 resolves collisions by double hashing. Rather than using two independent |
| 1722 |
* hash functions, this provides two options for efficiently deriving a pair of |
| 1723 |
* hashes from a single input hash function depending on whether the hash |
| 1724 |
* function is _avalanching_. See @ref hash_pair_from_avalanching_hash() and |
| 1725 |
* @ref hash_pair_from_non_avalanching_hash(). |
| 1726 |
* |
| 1727 |
* This provides: |
| 1728 |
* * Functions for double hashing common key types: `*_hash_pair()`. |
| 1729 |
* * Primitives for double hashing more complicated key types. |
| 1730 |
* * Equality functions for common key types: `*_eq()`. |
| 1731 |
* |
| 1732 |
* @{ |
| 1733 |
*/ |
| 1734 |
|
| 1735 |
/** |
| 1736 |
* Split an avalanching hash into a @ref hash_pair. |
| 1737 |
* |
| 1738 |
* A hash function is avalanching if each bit of the hash value has a 50% chance |
| 1739 |
* of being the same for different inputs. This is true for cryptographic hash |
| 1740 |
* functions as well as certain non-cryptographic hash functions including |
| 1741 |
* CityHash, MurmurHash, SipHash, and xxHash. Simple hashes like DJBX33A, ad-hoc |
| 1742 |
* combinations like `53 * x + y`, and the identity function are not |
| 1743 |
* avalanching. |
| 1744 |
* |
| 1745 |
* We use the input hash value as the first hash and the upper bits of the input |
| 1746 |
* hash value as the second hash (which would otherwise be discarded when |
| 1747 |
* masking to select the bucket). |
| 1748 |
*/ |
| 1749 |
static inline struct hash_pair hash_pair_from_avalanching_hash(size_t hash) |
| 1750 |
{ |
| 1751 |
return (struct hash_pair){ |
| 1752 |
.first = hash, |
| 1753 |
.second = (hash >> (8 * sizeof(hash) - 8)) | 0x80, |
| 1754 |
}; |
| 1755 |
} |
| 1756 |
|
| 1757 |
/** |
| 1758 |
* Mix a non-avalanching hash and split it into a @ref hash_pair. |
| 1759 |
* |
| 1760 |
* This is architecture-dependent. |
| 1761 |
*/ |
| 1762 |
static inline struct hash_pair hash_pair_from_non_avalanching_hash(size_t hash) |
| 1763 |
{ |
| 1764 |
#if SIZE_MAX == 0xffffffffffffffff |
| 1765 |
#ifdef __SSE4_2__ |
| 1766 |
/* 64-bit with SSE4.2 uses CRC32 */ |
| 1767 |
size_t c = _mm_crc32_u64(0, hash); |
| 1768 |
return (struct hash_pair){ |
| 1769 |
.first = hash + c, |
| 1770 |
.second = (c >> 24) | 0x80, |
| 1771 |
}; |
| 1772 |
#else |
| 1773 |
/* 64-bit without SSE4.2 uses a 128-bit multiplication-based mixer */ |
| 1774 |
static const uint64_t multiplier = UINT64_C(0xc4ceb9fe1a85ec53); |
| 1775 |
uint64_t hi = ((unsigned __int128)hash * multiplier) >> 64; |
| 1776 |
uint64_t lo = hash * multiplier; |
| 1777 |
hash = hi ^ lo; |
| 1778 |
hash *= multiplier; |
| 1779 |
return (struct hash_pair){ |
| 1780 |
.first = hash >> 22, |
| 1781 |
.second = ((hash >> 15) & 0x7f) | 0x80, |
| 1782 |
}; |
| 1783 |
#endif |
| 1784 |
#else |
| 1785 |
#ifdef __SSE4_2__ |
| 1786 |
/* 32-bit with SSE4.2 uses CRC32 */ |
| 1787 |
size_t c = _mm_crc32_u32(0, hash); |
| 1788 |
return (struct hash_pair){ |
| 1789 |
.first = hash + c, |
| 1790 |
.second = (uint8_t)(~(c >> 25)), |
| 1791 |
}; |
| 1792 |
#else |
| 1793 |
/* 32-bit without SSE4.2 uses the 32-bit Murmur2 finalizer */ |
| 1794 |
hash ^= hash >> 13; |
| 1795 |
hash *= 0x5bd1e995; |
| 1796 |
hash ^= hash >> 15; |
| 1797 |
return (struct hash_pair){ |
| 1798 |
.first = hash, |
| 1799 |
.second = (uint8_t)(~(hash >> 25)), |
| 1800 |
}; |
| 1801 |
#endif |
| 1802 |
#endif |
| 1803 |
} |
| 1804 |
|
| 1805 |
#ifdef DOXYGEN |
| 1806 |
/** |
| 1807 |
* Double hash an integral key. |
| 1808 |
* |
| 1809 |
* This can be used for any integer key type. |
| 1810 |
*/ |
| 1811 |
struct hash_pair int_key_hash_pair(const T *key); |
| 1812 |
#else |
| 1813 |
#if SIZE_MAX == 0xffffffffffffffff |
| 1814 |
static inline uint64_t hash_128_to_64(unsigned __int128 hash) |
| 1815 |
{ |
| 1816 |
return cityhash_128_to_64(hash, hash >> 64); |
| 1817 |
} |
| 1818 |
|
| 1819 |
#define int_key_hash_pair(key) ({ \ |
| 1820 |
__auto_type _key = *(key); \ |
| 1821 |
_Static_assert(sizeof(_key) <= sizeof(unsigned __int128), \ |
| 1822 |
"unsupported integer size"); \ |
| 1823 |
sizeof(_key) > sizeof(size_t) ? \ |
| 1824 |
hash_pair_from_avalanching_hash(hash_128_to_64(_key)) : \ |
| 1825 |
hash_pair_from_non_avalanching_hash(_key); \ |
| 1826 |
}) |
| 1827 |
#else |
| 1828 |
/* Thomas Wang downscaling hash function. */ |
| 1829 |
static inline uint32_t hash_64_to_32(uint64_t hash) |
| 1830 |
{ |
| 1831 |
hash = (~hash) + (hash << 18); |
| 1832 |
hash = hash ^ (hash >> 31); |
| 1833 |
hash = hash * 21; |
| 1834 |
hash = hash ^ (hash >> 11); |
| 1835 |
hash = hash + (hash << 6); |
| 1836 |
hash = hash ^ (hash >> 22); |
| 1837 |
return hash; |
| 1838 |
} |
| 1839 |
|
| 1840 |
#define int_key_hash_pair(key) ({ \ |
| 1841 |
__auto_type _key = *(key); \ |
| 1842 |
_Static_assert(sizeof(_key) <= sizeof(uint64_t), \ |
| 1843 |
"unsupported integer size"); \ |
| 1844 |
sizeof(_key) > sizeof(size_t) ? \ |
| 1845 |
hash_pair_from_avalanching_hash(hash_64_to_32(_key)) : \ |
| 1846 |
hash_pair_from_non_avalanching_hash(_key); \ |
| 1847 |
}) |
| 1848 |
#endif |
| 1849 |
#endif |
| 1850 |
|
| 1851 |
#ifdef DOXYGEN |
| 1852 |
/** |
| 1853 |
* Double hash a pointer key. |
| 1854 |
* |
| 1855 |
* This can be used when the key is a pointer value (rather than the |
| 1856 |
* dereferenced value). |
| 1857 |
*/ |
| 1858 |
struct hash_pair ptr_key_hash_pair(T * const *key); |
| 1859 |
#else |
| 1860 |
#define ptr_key_hash_pair(key) ({ \ |
| 1861 |
uintptr_t _ptr = (uintptr_t)*(key); \ |
| 1862 |
int_key_hash_pair(&_ptr); \ |
| 1863 |
}) |
| 1864 |
#endif |
| 1865 |
|
| 1866 |
#ifdef DOXYGEN |
| 1867 |
/** |
| 1868 |
* Return whether two scalar keys are equal. |
| 1869 |
* |
| 1870 |
* This can be used as the key comparison function for any scalar key type |
| 1871 |
* (e.g., integers, floating-point numbers, pointers). |
| 1872 |
*/ |
| 1873 |
bool scalar_key_eq(const T *a, const T *b); |
| 1874 |
#else |
| 1875 |
#define scalar_key_eq(a, b) ((bool)(*(a) == *(b))) |
| 1876 |
#endif |
| 1877 |
|
| 1878 |
/** |
| 1879 |
* Hash two integers. |
| 1880 |
* |
| 1881 |
* This is an avalanching hash function. It can be used for any integer types. |
| 1882 |
* The two integers can have different types. |
| 1883 |
* |
| 1884 |
* This can be used to combine input hash functions in order to hash records |
| 1885 |
* with multiple fields (e.g., structures or arrays). For example: |
| 1886 |
* |
| 1887 |
* ``` |
| 1888 |
* struct point3d { |
| 1889 |
* int x, y, z; |
| 1890 |
* }; |
| 1891 |
* |
| 1892 |
* static struct hash_pair point3d_key_hash_pair(const struct point3d *key) |
| 1893 |
* { |
| 1894 |
* return hash_pair_from_avalanching_hash(hash_combine(hash_combine(key->x, key->y), key->z)); |
| 1895 |
* } |
| 1896 |
* ``` |
| 1897 |
* |
| 1898 |
* Note that the input hash functions need not be avalanching; the output will |
| 1899 |
* be avalanching regardless. |
| 1900 |
*/ |
| 1901 |
#ifdef DOXYGEN |
| 1902 |
size_t hash_combine(T1 a, T2 b); |
| 1903 |
#else |
| 1904 |
#if SIZE_MAX == 0xffffffffffffffff |
| 1905 |
#define hash_combine(a, b) ({ \ |
| 1906 |
_Static_assert(sizeof(a) <= sizeof(unsigned __int128) && \ |
| 1907 |
sizeof(b) <= sizeof(unsigned __int128), \ |
| 1908 |
"unsupported integer size"); \ |
| 1909 |
size_t _a = sizeof(a) > sizeof(size_t) ? hash_128_to_64(a) : (a); \ |
| 1910 |
size_t _b = sizeof(b) > sizeof(size_t) ? hash_128_to_64(b) : (b); \ |
| 1911 |
cityhash_128_to_64(_b, _a); \ |
| 1912 |
}) |
| 1913 |
#else |
| 1914 |
#define hash_combine(a, b) ({ \ |
| 1915 |
_Static_assert(sizeof(a) <= sizeof(uint64_t) && \ |
| 1916 |
sizeof(b) <= sizeof(uint64_t), \ |
| 1917 |
"unsupported integer size"); \ |
| 1918 |
size_t _a = sizeof(a) > sizeof(size_t) ? hash_64_to_32(a) : (a); \ |
| 1919 |
size_t _b = sizeof(b) > sizeof(size_t) ? hash_64_to_32(b) : (b); \ |
| 1920 |
hash_64_to_32(((uint64_t)_a << 32) | _b); \ |
| 1921 |
}) |
| 1922 |
#endif |
| 1923 |
#endif |
| 1924 |
|
| 1925 |
/** |
| 1926 |
* Hash a byte buffer. |
| 1927 |
* |
| 1928 |
* This is an avalanching hash function. |
| 1929 |
*/ |
| 1930 |
static inline size_t hash_bytes(const void *data, size_t len) |
| 1931 |
{ |
| 1932 |
return cityhash_size_t(data, len); |
| 1933 |
} |
| 1934 |
|
| 1935 |
/** |
| 1936 |
* Hash a null-terminated string. |
| 1937 |
* |
| 1938 |
* This is an avalanching hash function. |
| 1939 |
*/ |
| 1940 |
static inline size_t hash_c_string(const char *s) |
| 1941 |
{ |
| 1942 |
return hash_bytes(s, strlen(s)); |
| 1943 |
} |
| 1944 |
|
| 1945 |
#ifdef DOXYGEN |
| 1946 |
/** Double hash a null-terminated string key. */ |
| 1947 |
struct hash_pair c_string_key_hash_pair(const char * const *key); |
| 1948 |
#else |
| 1949 |
/* This is a macro so that it works with char * and const char * keys. */ |
| 1950 |
#define c_string_key_hash_pair(key) \ |
| 1951 |
hash_pair_from_avalanching_hash(hash_c_string(*(key))) |
| 1952 |
#endif |
| 1953 |
|
| 1954 |
#ifdef DOXYGEN |
| 1955 |
/** Compare two null-terminated string keys for equality. */ |
| 1956 |
bool c_string_key_eq(const char * const *a, const char * const *b); |
| 1957 |
#else |
| 1958 |
#define c_string_key_eq(a, b) ((bool)(strcmp(*(a), *(b)) == 0)) |
| 1959 |
#endif |
| 1960 |
|
| 1961 |
/** Double hash a @ref nstring. */ |
| 1962 |
static inline struct hash_pair nstring_hash_pair(const struct nstring *key) |
| 1963 |
{ |
| 1964 |
return hash_pair_from_avalanching_hash(hash_bytes(key->str, key->len)); |
| 1965 |
} |
| 1966 |
|
| 1967 |
/** @} */ |
| 1968 |
|
| 1969 |
/** @} */ |
| 1970 |
|
| 1971 |
#endif /* DRGN_HASH_TABLE_H */ |