1 /*
2  * The following hash function is based on MurmurHash3, placed into the public
3  * domain by Austin Appleby.  See https://github.com/aappleby/smhasher for
4  * details.
5  */
6 /******************************************************************************/
7 #ifdef JEMALLOC_H_TYPES
8 
9 #endif /* JEMALLOC_H_TYPES */
10 /******************************************************************************/
11 #ifdef JEMALLOC_H_STRUCTS
12 
13 #endif /* JEMALLOC_H_STRUCTS */
14 /******************************************************************************/
15 #ifdef JEMALLOC_H_EXTERNS
16 
17 #endif /* JEMALLOC_H_EXTERNS */
18 /******************************************************************************/
19 #ifdef JEMALLOC_H_INLINES
20 
21 #ifndef JEMALLOC_ENABLE_INLINE
22 uint32_t	hash_x86_32(const void *key, int len, uint32_t seed);
23 void	hash_x86_128(const void *key, const int len, uint32_t seed,
24     uint64_t r_out[2]);
25 void	hash_x64_128(const void *key, const int len, const uint32_t seed,
26     uint64_t r_out[2]);
27 void	hash(const void *key, size_t len, const uint32_t seed,
28     size_t r_hash[2]);
29 #endif
30 
31 #if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_HASH_C_))
32 /******************************************************************************/
33 /* Internal implementation. */
34 JEMALLOC_INLINE uint32_t
35 hash_rotl_32(uint32_t x, int8_t r)
36 {
37 
38 	return ((x << r) | (x >> (32 - r)));
39 }
40 
41 JEMALLOC_INLINE uint64_t
42 hash_rotl_64(uint64_t x, int8_t r)
43 {
44 
45 	return ((x << r) | (x >> (64 - r)));
46 }
47 
48 JEMALLOC_INLINE uint32_t
49 hash_get_block_32(const uint32_t *p, int i)
50 {
51 
52 	/* Handle unaligned read. */
53 	if (unlikely((uintptr_t)p & (sizeof(uint32_t)-1)) != 0) {
54 		uint32_t ret;
55 
56 		memcpy(&ret, (uint8_t *)(p + i), sizeof(uint32_t));
57 		return (ret);
58 	}
59 
60 	return (p[i]);
61 }
62 
63 JEMALLOC_INLINE uint64_t
64 hash_get_block_64(const uint64_t *p, int i)
65 {
66 
67 	/* Handle unaligned read. */
68 	if (unlikely((uintptr_t)p & (sizeof(uint64_t)-1)) != 0) {
69 		uint64_t ret;
70 
71 		memcpy(&ret, (uint8_t *)(p + i), sizeof(uint64_t));
72 		return (ret);
73 	}
74 
75 	return (p[i]);
76 }
77 
78 JEMALLOC_INLINE uint32_t
79 hash_fmix_32(uint32_t h)
80 {
81 
82 	h ^= h >> 16;
83 	h *= 0x85ebca6b;
84 	h ^= h >> 13;
85 	h *= 0xc2b2ae35;
86 	h ^= h >> 16;
87 
88 	return (h);
89 }
90 
91 JEMALLOC_INLINE uint64_t
92 hash_fmix_64(uint64_t k)
93 {
94 
95 	k ^= k >> 33;
96 	k *= KQU(0xff51afd7ed558ccd);
97 	k ^= k >> 33;
98 	k *= KQU(0xc4ceb9fe1a85ec53);
99 	k ^= k >> 33;
100 
101 	return (k);
102 }
103 
104 JEMALLOC_INLINE uint32_t
105 hash_x86_32(const void *key, int len, uint32_t seed)
106 {
107 	const uint8_t *data = (const uint8_t *) key;
108 	const int nblocks = len / 4;
109 
110 	uint32_t h1 = seed;
111 
112 	const uint32_t c1 = 0xcc9e2d51;
113 	const uint32_t c2 = 0x1b873593;
114 
115 	/* body */
116 	{
117 		const uint32_t *blocks = (const uint32_t *) (data + nblocks*4);
118 		int i;
119 
120 		for (i = -nblocks; i; i++) {
121 			uint32_t k1 = hash_get_block_32(blocks, i);
122 
123 			k1 *= c1;
124 			k1 = hash_rotl_32(k1, 15);
125 			k1 *= c2;
126 
127 			h1 ^= k1;
128 			h1 = hash_rotl_32(h1, 13);
129 			h1 = h1*5 + 0xe6546b64;
130 		}
131 	}
132 
133 	/* tail */
134 	{
135 		const uint8_t *tail = (const uint8_t *) (data + nblocks*4);
136 
137 		uint32_t k1 = 0;
138 
139 		switch (len & 3) {
140 		case 3: k1 ^= tail[2] << 16;
141 		case 2: k1 ^= tail[1] << 8;
142 		case 1: k1 ^= tail[0]; k1 *= c1; k1 = hash_rotl_32(k1, 15);
143 			k1 *= c2; h1 ^= k1;
144 		}
145 	}
146 
147 	/* finalization */
148 	h1 ^= len;
149 
150 	h1 = hash_fmix_32(h1);
151 
152 	return (h1);
153 }
154 
155 UNUSED JEMALLOC_INLINE void
156 hash_x86_128(const void *key, const int len, uint32_t seed,
157     uint64_t r_out[2])
158 {
159 	const uint8_t * data = (const uint8_t *) key;
160 	const int nblocks = len / 16;
161 
162 	uint32_t h1 = seed;
163 	uint32_t h2 = seed;
164 	uint32_t h3 = seed;
165 	uint32_t h4 = seed;
166 
167 	const uint32_t c1 = 0x239b961b;
168 	const uint32_t c2 = 0xab0e9789;
169 	const uint32_t c3 = 0x38b34ae5;
170 	const uint32_t c4 = 0xa1e38b93;
171 
172 	/* body */
173 	{
174 		const uint32_t *blocks = (const uint32_t *) (data + nblocks*16);
175 		int i;
176 
177 		for (i = -nblocks; i; i++) {
178 			uint32_t k1 = hash_get_block_32(blocks, i*4 + 0);
179 			uint32_t k2 = hash_get_block_32(blocks, i*4 + 1);
180 			uint32_t k3 = hash_get_block_32(blocks, i*4 + 2);
181 			uint32_t k4 = hash_get_block_32(blocks, i*4 + 3);
182 
183 			k1 *= c1; k1 = hash_rotl_32(k1, 15); k1 *= c2; h1 ^= k1;
184 
185 			h1 = hash_rotl_32(h1, 19); h1 += h2;
186 			h1 = h1*5 + 0x561ccd1b;
187 
188 			k2 *= c2; k2 = hash_rotl_32(k2, 16); k2 *= c3; h2 ^= k2;
189 
190 			h2 = hash_rotl_32(h2, 17); h2 += h3;
191 			h2 = h2*5 + 0x0bcaa747;
192 
193 			k3 *= c3; k3 = hash_rotl_32(k3, 17); k3 *= c4; h3 ^= k3;
194 
195 			h3 = hash_rotl_32(h3, 15); h3 += h4;
196 			h3 = h3*5 + 0x96cd1c35;
197 
198 			k4 *= c4; k4 = hash_rotl_32(k4, 18); k4 *= c1; h4 ^= k4;
199 
200 			h4 = hash_rotl_32(h4, 13); h4 += h1;
201 			h4 = h4*5 + 0x32ac3b17;
202 		}
203 	}
204 
205 	/* tail */
206 	{
207 		const uint8_t *tail = (const uint8_t *) (data + nblocks*16);
208 		uint32_t k1 = 0;
209 		uint32_t k2 = 0;
210 		uint32_t k3 = 0;
211 		uint32_t k4 = 0;
212 
213 		switch (len & 15) {
214 		case 15: k4 ^= tail[14] << 16;
215 		case 14: k4 ^= tail[13] << 8;
216 		case 13: k4 ^= tail[12] << 0;
217 			k4 *= c4; k4 = hash_rotl_32(k4, 18); k4 *= c1; h4 ^= k4;
218 
219 		case 12: k3 ^= tail[11] << 24;
220 		case 11: k3 ^= tail[10] << 16;
221 		case 10: k3 ^= tail[ 9] << 8;
222 		case  9: k3 ^= tail[ 8] << 0;
223 		     k3 *= c3; k3 = hash_rotl_32(k3, 17); k3 *= c4; h3 ^= k3;
224 
225 		case  8: k2 ^= tail[ 7] << 24;
226 		case  7: k2 ^= tail[ 6] << 16;
227 		case  6: k2 ^= tail[ 5] << 8;
228 		case  5: k2 ^= tail[ 4] << 0;
229 			k2 *= c2; k2 = hash_rotl_32(k2, 16); k2 *= c3; h2 ^= k2;
230 
231 		case  4: k1 ^= tail[ 3] << 24;
232 		case  3: k1 ^= tail[ 2] << 16;
233 		case  2: k1 ^= tail[ 1] << 8;
234 		case  1: k1 ^= tail[ 0] << 0;
235 			k1 *= c1; k1 = hash_rotl_32(k1, 15); k1 *= c2; h1 ^= k1;
236 		}
237 	}
238 
239 	/* finalization */
240 	h1 ^= len; h2 ^= len; h3 ^= len; h4 ^= len;
241 
242 	h1 += h2; h1 += h3; h1 += h4;
243 	h2 += h1; h3 += h1; h4 += h1;
244 
245 	h1 = hash_fmix_32(h1);
246 	h2 = hash_fmix_32(h2);
247 	h3 = hash_fmix_32(h3);
248 	h4 = hash_fmix_32(h4);
249 
250 	h1 += h2; h1 += h3; h1 += h4;
251 	h2 += h1; h3 += h1; h4 += h1;
252 
253 	r_out[0] = (((uint64_t) h2) << 32) | h1;
254 	r_out[1] = (((uint64_t) h4) << 32) | h3;
255 }
256 
257 UNUSED JEMALLOC_INLINE void
258 hash_x64_128(const void *key, const int len, const uint32_t seed,
259     uint64_t r_out[2])
260 {
261 	const uint8_t *data = (const uint8_t *) key;
262 	const int nblocks = len / 16;
263 
264 	uint64_t h1 = seed;
265 	uint64_t h2 = seed;
266 
267 	const uint64_t c1 = KQU(0x87c37b91114253d5);
268 	const uint64_t c2 = KQU(0x4cf5ad432745937f);
269 
270 	/* body */
271 	{
272 		const uint64_t *blocks = (const uint64_t *) (data);
273 		int i;
274 
275 		for (i = 0; i < nblocks; i++) {
276 			uint64_t k1 = hash_get_block_64(blocks, i*2 + 0);
277 			uint64_t k2 = hash_get_block_64(blocks, i*2 + 1);
278 
279 			k1 *= c1; k1 = hash_rotl_64(k1, 31); k1 *= c2; h1 ^= k1;
280 
281 			h1 = hash_rotl_64(h1, 27); h1 += h2;
282 			h1 = h1*5 + 0x52dce729;
283 
284 			k2 *= c2; k2 = hash_rotl_64(k2, 33); k2 *= c1; h2 ^= k2;
285 
286 			h2 = hash_rotl_64(h2, 31); h2 += h1;
287 			h2 = h2*5 + 0x38495ab5;
288 		}
289 	}
290 
291 	/* tail */
292 	{
293 		const uint8_t *tail = (const uint8_t*)(data + nblocks*16);
294 		uint64_t k1 = 0;
295 		uint64_t k2 = 0;
296 
297 		switch (len & 15) {
298 		case 15: k2 ^= ((uint64_t)(tail[14])) << 48;
299 		case 14: k2 ^= ((uint64_t)(tail[13])) << 40;
300 		case 13: k2 ^= ((uint64_t)(tail[12])) << 32;
301 		case 12: k2 ^= ((uint64_t)(tail[11])) << 24;
302 		case 11: k2 ^= ((uint64_t)(tail[10])) << 16;
303 		case 10: k2 ^= ((uint64_t)(tail[ 9])) << 8;
304 		case  9: k2 ^= ((uint64_t)(tail[ 8])) << 0;
305 			k2 *= c2; k2 = hash_rotl_64(k2, 33); k2 *= c1; h2 ^= k2;
306 
307 		case  8: k1 ^= ((uint64_t)(tail[ 7])) << 56;
308 		case  7: k1 ^= ((uint64_t)(tail[ 6])) << 48;
309 		case  6: k1 ^= ((uint64_t)(tail[ 5])) << 40;
310 		case  5: k1 ^= ((uint64_t)(tail[ 4])) << 32;
311 		case  4: k1 ^= ((uint64_t)(tail[ 3])) << 24;
312 		case  3: k1 ^= ((uint64_t)(tail[ 2])) << 16;
313 		case  2: k1 ^= ((uint64_t)(tail[ 1])) << 8;
314 		case  1: k1 ^= ((uint64_t)(tail[ 0])) << 0;
315 			k1 *= c1; k1 = hash_rotl_64(k1, 31); k1 *= c2; h1 ^= k1;
316 		}
317 	}
318 
319 	/* finalization */
320 	h1 ^= len; h2 ^= len;
321 
322 	h1 += h2;
323 	h2 += h1;
324 
325 	h1 = hash_fmix_64(h1);
326 	h2 = hash_fmix_64(h2);
327 
328 	h1 += h2;
329 	h2 += h1;
330 
331 	r_out[0] = h1;
332 	r_out[1] = h2;
333 }
334 
335 /******************************************************************************/
336 /* API. */
337 JEMALLOC_INLINE void
338 hash(const void *key, size_t len, const uint32_t seed, size_t r_hash[2])
339 {
340 
341 	assert(len <= INT_MAX); /* Unfortunate implementation limitation. */
342 
343 #if (LG_SIZEOF_PTR == 3 && !defined(JEMALLOC_BIG_ENDIAN))
344 	hash_x64_128(key, (int)len, seed, (uint64_t *)r_hash);
345 #else
346 	{
347 		uint64_t hashes[2];
348 		hash_x86_128(key, (int)len, seed, hashes);
349 		r_hash[0] = (size_t)hashes[0];
350 		r_hash[1] = (size_t)hashes[1];
351 	}
352 #endif
353 }
354 #endif
355 
356 #endif /* JEMALLOC_H_INLINES */
357 /******************************************************************************/
358