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