Line data Source code
1 : /*
2 : * hashfn_unstable.h
3 : *
4 : * Building blocks for creating fast inlineable hash functions. The
5 : * functions in this file are not guaranteed to be stable between versions,
6 : * and may differ by hardware platform. Hence they must not be used in
7 : * indexes or other on-disk structures. See hashfn.h if you need stability.
8 : *
9 : *
10 : * Portions Copyright (c) 2024, PostgreSQL Global Development Group
11 : *
12 : * src/include/common/hashfn_unstable.h
13 : */
14 : #ifndef HASHFN_UNSTABLE_H
15 : #define HASHFN_UNSTABLE_H
16 :
17 : #include "port/pg_bitutils.h"
18 : #include "port/pg_bswap.h"
19 :
20 : /*
21 : * fasthash is a modification of code taken from
22 : * https://code.google.com/archive/p/fast-hash/source/default/source
23 : * under the terms of the MIT license. The original copyright
24 : * notice follows:
25 : */
26 :
27 : /* The MIT License
28 :
29 : Copyright (C) 2012 Zilong Tan (eric.zltan@gmail.com)
30 :
31 : Permission is hereby granted, free of charge, to any person
32 : obtaining a copy of this software and associated documentation
33 : files (the "Software"), to deal in the Software without
34 : restriction, including without limitation the rights to use, copy,
35 : modify, merge, publish, distribute, sublicense, and/or sell copies
36 : of the Software, and to permit persons to whom the Software is
37 : furnished to do so, subject to the following conditions:
38 :
39 : The above copyright notice and this permission notice shall be
40 : included in all copies or substantial portions of the Software.
41 :
42 : THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
43 : EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
44 : MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
45 : NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
46 : BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
47 : ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
48 : CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
49 : SOFTWARE.
50 : */
51 :
52 : /*
53 : * fasthash as implemented here has two interfaces:
54 : *
55 : * 1) Standalone functions, e.g. fasthash32() for a single value with a
56 : * known length. These return the same hash code as the original, at
57 : * least on little-endian machines.
58 : *
59 : * 2) Incremental interface. This can used for incorporating multiple
60 : * inputs. First, initialize the hash state (here with a zero seed):
61 : *
62 : * fasthash_state hs;
63 : * fasthash_init(&hs, 0);
64 : *
65 : * If the inputs are of types that can be trivially cast to uint64, it's
66 : * sufficient to do:
67 : *
68 : * hs.accum = value1;
69 : * fasthash_combine(&hs);
70 : * hs.accum = value2;
71 : * fasthash_combine(&hs);
72 : * ...
73 : *
74 : * For longer or variable-length input, fasthash_accum() is a more
75 : * flexible, but more verbose method. The standalone functions use this
76 : * internally, so see fasthash64() for an example of this.
77 : *
78 : * After all inputs have been mixed in, finalize the hash:
79 : *
80 : * hashcode = fasthash_final32(&hs, 0);
81 : *
82 : * The incremental interface allows an optimization for NUL-terminated
83 : * C strings:
84 : *
85 : * len = fasthash_accum_cstring(&hs, str);
86 : * hashcode = fasthash_final32(&hs, len);
87 : *
88 : * By handling the terminator on-the-fly, we can avoid needing a strlen()
89 : * call to tell us how many bytes to hash. Experimentation has found that
90 : * SMHasher fails unless we incorporate the length, so it is passed to
91 : * the finalizer as a tweak.
92 : */
93 :
94 :
95 : typedef struct fasthash_state
96 : {
97 : /* staging area for chunks of input */
98 : uint64 accum;
99 :
100 : uint64 hash;
101 : } fasthash_state;
102 :
103 : #define FH_SIZEOF_ACCUM sizeof(uint64)
104 :
105 :
106 : /*
107 : * Initialize the hash state.
108 : *
109 : * 'seed' can be zero.
110 : */
111 : static inline void
112 10788106 : fasthash_init(fasthash_state *hs, uint64 seed)
113 : {
114 10788106 : memset(hs, 0, sizeof(fasthash_state));
115 10788106 : hs->hash = seed ^ 0x880355f21e6d1965;
116 10788106 : }
117 :
118 : /* both the finalizer and part of the combining step */
119 : static inline uint64
120 32614776 : fasthash_mix(uint64 h, uint64 tweak)
121 : {
122 32614776 : h ^= (h >> 23) + tweak;
123 32614776 : h *= 0x2127599bf4325c37;
124 32614776 : h ^= h >> 47;
125 32614776 : return h;
126 : }
127 :
128 : /* combine one chunk of input into the hash */
129 : static inline void
130 21826670 : fasthash_combine(fasthash_state *hs)
131 : {
132 21826670 : hs->hash ^= fasthash_mix(hs->accum, 0);
133 21826670 : hs->hash *= 0x880355f21e6d1965;
134 21826670 : }
135 :
136 : /* accumulate up to 8 bytes of input and combine it into the hash */
137 : static inline void
138 29700978 : fasthash_accum(fasthash_state *hs, const char *k, size_t len)
139 : {
140 : uint32 lower_four;
141 :
142 : Assert(len <= FH_SIZEOF_ACCUM);
143 29700978 : hs->accum = 0;
144 :
145 : /*
146 : * For consistency, bytewise loads must match the platform's endianness.
147 : */
148 : #ifdef WORDS_BIGENDIAN
149 : switch (len)
150 : {
151 : case 8:
152 : memcpy(&hs->accum, k, 8);
153 : break;
154 : case 7:
155 : hs->accum |= (uint64) k[6] << 8;
156 : /* FALLTHROUGH */
157 : case 6:
158 : hs->accum |= (uint64) k[5] << 16;
159 : /* FALLTHROUGH */
160 : case 5:
161 : hs->accum |= (uint64) k[4] << 24;
162 : /* FALLTHROUGH */
163 : case 4:
164 : memcpy(&lower_four, k, sizeof(lower_four));
165 : hs->accum |= (uint64) lower_four << 32;
166 : break;
167 : case 3:
168 : hs->accum |= (uint64) k[2] << 40;
169 : /* FALLTHROUGH */
170 : case 2:
171 : hs->accum |= (uint64) k[1] << 48;
172 : /* FALLTHROUGH */
173 : case 1:
174 : hs->accum |= (uint64) k[0] << 56;
175 : break;
176 : case 0:
177 : return;
178 : }
179 : #else
180 29700978 : switch (len)
181 : {
182 19800652 : case 8:
183 19800652 : memcpy(&hs->accum, k, 8);
184 19800652 : break;
185 0 : case 7:
186 0 : hs->accum |= (uint64) k[6] << 48;
187 : /* FALLTHROUGH */
188 0 : case 6:
189 0 : hs->accum |= (uint64) k[5] << 40;
190 : /* FALLTHROUGH */
191 0 : case 5:
192 0 : hs->accum |= (uint64) k[4] << 32;
193 : /* FALLTHROUGH */
194 0 : case 4:
195 0 : memcpy(&lower_four, k, sizeof(lower_four));
196 0 : hs->accum |= lower_four;
197 0 : break;
198 0 : case 3:
199 0 : hs->accum |= (uint64) k[2] << 16;
200 : /* FALLTHROUGH */
201 0 : case 2:
202 0 : hs->accum |= (uint64) k[1] << 8;
203 : /* FALLTHROUGH */
204 0 : case 1:
205 0 : hs->accum |= (uint64) k[0];
206 0 : break;
207 9900326 : case 0:
208 9900326 : return;
209 : }
210 : #endif
211 :
212 19800652 : fasthash_combine(hs);
213 : }
214 :
215 : /*
216 : * Set high bit in lowest byte where the input is zero, from:
217 : * https://graphics.stanford.edu/~seander/bithacks.html#ZeroInWord
218 : */
219 : #define haszero64(v) \
220 : (((v) - 0x0101010101010101) & ~(v) & 0x8080808080808080)
221 :
222 : /* get first byte in memory order */
223 : #ifdef WORDS_BIGENDIAN
224 : #define firstbyte64(v) ((v) >> 56)
225 : #else
226 : #define firstbyte64(v) ((v) & 0xFF)
227 : #endif
228 :
229 : /*
230 : * all-purpose workhorse for fasthash_accum_cstring
231 : */
232 : static inline size_t
233 0 : fasthash_accum_cstring_unaligned(fasthash_state *hs, const char *str)
234 : {
235 0 : const char *const start = str;
236 :
237 0 : while (*str)
238 : {
239 0 : size_t chunk_len = 0;
240 :
241 0 : while (chunk_len < FH_SIZEOF_ACCUM && str[chunk_len] != '\0')
242 0 : chunk_len++;
243 :
244 0 : fasthash_accum(hs, str, chunk_len);
245 0 : str += chunk_len;
246 : }
247 :
248 0 : return str - start;
249 : }
250 :
251 : /*
252 : * specialized workhorse for fasthash_accum_cstring
253 : *
254 : * With an aligned pointer, we consume the string a word at a time.
255 : * Loading the word containing the NUL terminator cannot segfault since
256 : * allocation boundaries are suitably aligned. To keep from setting
257 : * off alarms with address sanitizers, exclude this function from
258 : * such testing.
259 : */
260 : pg_attribute_no_sanitize_address()
261 : static inline size_t
262 887780 : fasthash_accum_cstring_aligned(fasthash_state *hs, const char *str)
263 : {
264 887780 : const char *const start = str;
265 : uint64 chunk;
266 : uint64 zero_byte_low;
267 :
268 : Assert(PointerIsAligned(start, uint64));
269 :
270 : /*
271 : * For every chunk of input, check for zero bytes before mixing into the
272 : * hash. The chunk with zeros must contain the NUL terminator. We arrange
273 : * so that zero_byte_low tells us not only that a zero exists, but also
274 : * where it is, so we can hash the remainder of the string.
275 : *
276 : * The haszero64 calculation will set bits corresponding to the lowest
277 : * byte where a zero exists, so that suffices for little-endian machines.
278 : * For big-endian machines, we would need bits set for the highest zero
279 : * byte in the chunk, since the trailing junk past the terminator could
280 : * contain additional zeros. haszero64 does not give us that, so we
281 : * byteswap the chunk first.
282 : */
283 : for (;;)
284 : {
285 1887674 : chunk = *(uint64 *) str;
286 :
287 : #ifdef WORDS_BIGENDIAN
288 : zero_byte_low = haszero64(pg_bswap64(chunk));
289 : #else
290 1887674 : zero_byte_low = haszero64(chunk);
291 : #endif
292 1887674 : if (zero_byte_low)
293 887780 : break;
294 :
295 999894 : hs->accum = chunk;
296 999894 : fasthash_combine(hs);
297 999894 : str += FH_SIZEOF_ACCUM;
298 : }
299 :
300 887780 : if (firstbyte64(chunk) != 0)
301 : {
302 : size_t remainder;
303 : uint64 mask;
304 :
305 : /*
306 : * The byte corresponding to the NUL will be 0x80, so the rightmost
307 : * bit position will be in the range 15, 23, ..., 63. Turn this into
308 : * byte position by dividing by 8.
309 : */
310 862724 : remainder = pg_rightmost_one_pos64(zero_byte_low) / BITS_PER_BYTE;
311 :
312 : /*
313 : * Create a mask for the remaining bytes so we can combine them into
314 : * the hash. This must have the same result as mixing the remaining
315 : * bytes with fasthash_accum().
316 : */
317 : #ifdef WORDS_BIGENDIAN
318 : mask = ~UINT64CONST(0) << BITS_PER_BYTE * (FH_SIZEOF_ACCUM - remainder);
319 : #else
320 862724 : mask = ~UINT64CONST(0) >> BITS_PER_BYTE * (FH_SIZEOF_ACCUM - remainder);
321 : #endif
322 862724 : hs->accum = chunk & mask;
323 862724 : fasthash_combine(hs);
324 :
325 862724 : str += remainder;
326 : }
327 :
328 887780 : return str - start;
329 : }
330 :
331 : /*
332 : * Mix 'str' into the hash state and return the length of the string.
333 : */
334 : static inline size_t
335 887780 : fasthash_accum_cstring(fasthash_state *hs, const char *str)
336 : {
337 : #if SIZEOF_VOID_P >= 8
338 :
339 : size_t len;
340 : #ifdef USE_ASSERT_CHECKING
341 : size_t len_check;
342 : fasthash_state hs_check;
343 :
344 : memcpy(&hs_check, hs, sizeof(fasthash_state));
345 : len_check = fasthash_accum_cstring_unaligned(&hs_check, str);
346 : #endif
347 887780 : if (PointerIsAligned(str, uint64))
348 : {
349 887780 : len = fasthash_accum_cstring_aligned(hs, str);
350 : Assert(len_check == len);
351 : Assert(hs_check.hash == hs->hash);
352 887780 : return len;
353 : }
354 : #endif /* SIZEOF_VOID_P */
355 :
356 : /*
357 : * It's not worth it to try to make the word-at-a-time optimization work
358 : * on 32-bit platforms.
359 : */
360 0 : return fasthash_accum_cstring_unaligned(hs, str);
361 : }
362 :
363 : /*
364 : * The finalizer
365 : *
366 : * 'tweak' is intended to be the input length when the caller doesn't know
367 : * the length ahead of time, such as for NUL-terminated strings, otherwise
368 : * zero.
369 : */
370 : static inline uint64
371 10788106 : fasthash_final64(fasthash_state *hs, uint64 tweak)
372 : {
373 10788106 : return fasthash_mix(hs->hash, tweak);
374 : }
375 :
376 : /*
377 : * Reduce a 64-bit hash to a 32-bit hash.
378 : *
379 : * This optional step provides a bit more additional mixing compared to
380 : * just taking the lower 32-bits.
381 : */
382 : static inline uint32
383 10788106 : fasthash_reduce32(uint64 h)
384 : {
385 : /*
386 : * Convert the 64-bit hashcode to Fermat residue, which shall retain
387 : * information from both the higher and lower parts of hashcode.
388 : */
389 10788106 : return h - (h >> 32);
390 : }
391 :
392 : /* finalize and reduce */
393 : static inline uint32
394 887780 : fasthash_final32(fasthash_state *hs, uint64 tweak)
395 : {
396 887780 : return fasthash_reduce32(fasthash_final64(hs, tweak));
397 : }
398 :
399 : /*
400 : * The original fasthash64 function, re-implemented using the incremental
401 : * interface. Returns a 64-bit hashcode. 'len' controls not only how
402 : * many bytes to hash, but also modifies the internal seed.
403 : * 'seed' can be zero.
404 : */
405 : static inline uint64
406 9900326 : fasthash64(const char *k, size_t len, uint64 seed)
407 : {
408 : fasthash_state hs;
409 :
410 9900326 : fasthash_init(&hs, 0);
411 :
412 : /* re-initialize the seed according to input length */
413 9900326 : hs.hash = seed ^ (len * 0x880355f21e6d1965);
414 :
415 29700978 : while (len >= FH_SIZEOF_ACCUM)
416 : {
417 19800652 : fasthash_accum(&hs, k, FH_SIZEOF_ACCUM);
418 19800652 : k += FH_SIZEOF_ACCUM;
419 19800652 : len -= FH_SIZEOF_ACCUM;
420 : }
421 :
422 9900326 : fasthash_accum(&hs, k, len);
423 9900326 : return fasthash_final64(&hs, 0);
424 : }
425 :
426 : /* like fasthash64, but returns a 32-bit hashcode */
427 : static inline uint32
428 9900326 : fasthash32(const char *k, size_t len, uint64 seed)
429 : {
430 9900326 : return fasthash_reduce32(fasthash64(k, len, seed));
431 : }
432 :
433 : /*
434 : * Convenience function for hashing NUL-terminated strings
435 : */
436 : static inline uint32
437 724380 : hash_string(const char *s)
438 : {
439 : fasthash_state hs;
440 : size_t s_len;
441 :
442 724380 : fasthash_init(&hs, 0);
443 :
444 : /*
445 : * Combine string into the hash and save the length for tweaking the final
446 : * mix.
447 : */
448 724380 : s_len = fasthash_accum_cstring(&hs, s);
449 :
450 724380 : return fasthash_final32(&hs, s_len);
451 : }
452 :
453 : #endif /* HASHFN_UNSTABLE_H */
|
