1 /*
2 * Compute the CRC32 using a parallelized folding approach with the PCLMULQDQ
3 * instruction.
4 *
5 * A white paper describing this algorithm can be found at:
6 * https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
7 *
8 * Copyright (C) 2013 Intel Corporation. All rights reserved.
9 * Copyright (C) 2016 Marian Beermann (support for initial value)
10 * Authors:
11 * Wajdi Feghali <[email protected]>
12 * Jim Guilford <[email protected]>
13 * Vinodh Gopal <[email protected]>
14 * Erdinc Ozturk <[email protected]>
15 * Jim Kukunas <[email protected]>
16 *
17 * For conditions of distribution and use, see copyright notice in zlib.h
18 */
19
20 #ifdef X86_PCLMULQDQ_CRC
21 #include "../../zbuild.h"
22
23 #include <immintrin.h>
24 #include <wmmintrin.h>
25 #include <smmintrin.h> // _mm_extract_epi32
26
27 #include "x86_features.h"
28 #include "cpu_features.h"
29
30 #include "../../crc32_fold.h"
31 #include "../../crc32_braid_p.h"
32 #include <assert.h>
33
34 #ifdef X86_VPCLMULQDQ_CRC
35 extern size_t fold_16_vpclmulqdq(__m128i *xmm_crc0, __m128i *xmm_crc1,
36 __m128i *xmm_crc2, __m128i *xmm_crc3, uint8_t *dst, const uint8_t *src, size_t len);
37 extern size_t fold_16_vpclmulqdq_nocp(__m128i *xmm_crc0, __m128i *xmm_crc1,
38 __m128i *xmm_crc2, __m128i *xmm_crc3, const uint8_t *src, size_t len, __m128i init_crc,
39 int32_t first);
40 #endif
41
fold_1(__m128i * xmm_crc0,__m128i * xmm_crc1,__m128i * xmm_crc2,__m128i * xmm_crc3)42 static void fold_1(__m128i *xmm_crc0, __m128i *xmm_crc1, __m128i *xmm_crc2, __m128i *xmm_crc3) {
43 const __m128i xmm_fold4 = _mm_set_epi32( 0x00000001, 0x54442bd4,
44 0x00000001, 0xc6e41596);
45 __m128i x_tmp3;
46 __m128 ps_crc0, ps_crc3, ps_res;
47
48 x_tmp3 = *xmm_crc3;
49
50 *xmm_crc3 = *xmm_crc0;
51 *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01);
52 *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x10);
53 ps_crc0 = _mm_castsi128_ps(*xmm_crc0);
54 ps_crc3 = _mm_castsi128_ps(*xmm_crc3);
55 ps_res = _mm_xor_ps(ps_crc0, ps_crc3);
56
57 *xmm_crc0 = *xmm_crc1;
58 *xmm_crc1 = *xmm_crc2;
59 *xmm_crc2 = x_tmp3;
60 *xmm_crc3 = _mm_castps_si128(ps_res);
61 }
62
fold_2(__m128i * xmm_crc0,__m128i * xmm_crc1,__m128i * xmm_crc2,__m128i * xmm_crc3)63 static void fold_2(__m128i *xmm_crc0, __m128i *xmm_crc1, __m128i *xmm_crc2, __m128i *xmm_crc3) {
64 const __m128i xmm_fold4 = _mm_set_epi32( 0x00000001, 0x54442bd4,
65 0x00000001, 0xc6e41596);
66 __m128i x_tmp3, x_tmp2;
67 __m128 ps_crc0, ps_crc1, ps_crc2, ps_crc3, ps_res31, ps_res20;
68
69 x_tmp3 = *xmm_crc3;
70 x_tmp2 = *xmm_crc2;
71
72 *xmm_crc3 = *xmm_crc1;
73 *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x01);
74 *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x10);
75 ps_crc3 = _mm_castsi128_ps(*xmm_crc3);
76 ps_crc1 = _mm_castsi128_ps(*xmm_crc1);
77 ps_res31 = _mm_xor_ps(ps_crc3, ps_crc1);
78
79 *xmm_crc2 = *xmm_crc0;
80 *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01);
81 *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x10);
82 ps_crc0 = _mm_castsi128_ps(*xmm_crc0);
83 ps_crc2 = _mm_castsi128_ps(*xmm_crc2);
84 ps_res20 = _mm_xor_ps(ps_crc0, ps_crc2);
85
86 *xmm_crc0 = x_tmp2;
87 *xmm_crc1 = x_tmp3;
88 *xmm_crc2 = _mm_castps_si128(ps_res20);
89 *xmm_crc3 = _mm_castps_si128(ps_res31);
90 }
91
fold_3(__m128i * xmm_crc0,__m128i * xmm_crc1,__m128i * xmm_crc2,__m128i * xmm_crc3)92 static void fold_3(__m128i *xmm_crc0, __m128i *xmm_crc1, __m128i *xmm_crc2, __m128i *xmm_crc3) {
93 const __m128i xmm_fold4 = _mm_set_epi32( 0x00000001, 0x54442bd4,
94 0x00000001, 0xc6e41596);
95 __m128i x_tmp3;
96 __m128 ps_crc0, ps_crc1, ps_crc2, ps_crc3, ps_res32, ps_res21, ps_res10;
97
98 x_tmp3 = *xmm_crc3;
99
100 *xmm_crc3 = *xmm_crc2;
101 *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x01);
102 *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x10);
103 ps_crc2 = _mm_castsi128_ps(*xmm_crc2);
104 ps_crc3 = _mm_castsi128_ps(*xmm_crc3);
105 ps_res32 = _mm_xor_ps(ps_crc2, ps_crc3);
106
107 *xmm_crc2 = *xmm_crc1;
108 *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x01);
109 *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x10);
110 ps_crc1 = _mm_castsi128_ps(*xmm_crc1);
111 ps_crc2 = _mm_castsi128_ps(*xmm_crc2);
112 ps_res21 = _mm_xor_ps(ps_crc1, ps_crc2);
113
114 *xmm_crc1 = *xmm_crc0;
115 *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01);
116 *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x10);
117 ps_crc0 = _mm_castsi128_ps(*xmm_crc0);
118 ps_crc1 = _mm_castsi128_ps(*xmm_crc1);
119 ps_res10 = _mm_xor_ps(ps_crc0, ps_crc1);
120
121 *xmm_crc0 = x_tmp3;
122 *xmm_crc1 = _mm_castps_si128(ps_res10);
123 *xmm_crc2 = _mm_castps_si128(ps_res21);
124 *xmm_crc3 = _mm_castps_si128(ps_res32);
125 }
126
fold_4(__m128i * xmm_crc0,__m128i * xmm_crc1,__m128i * xmm_crc2,__m128i * xmm_crc3)127 static void fold_4(__m128i *xmm_crc0, __m128i *xmm_crc1, __m128i *xmm_crc2, __m128i *xmm_crc3) {
128 const __m128i xmm_fold4 = _mm_set_epi32( 0x00000001, 0x54442bd4,
129 0x00000001, 0xc6e41596);
130 __m128i x_tmp0, x_tmp1, x_tmp2, x_tmp3;
131 __m128 ps_crc0, ps_crc1, ps_crc2, ps_crc3;
132 __m128 ps_t0, ps_t1, ps_t2, ps_t3;
133 __m128 ps_res0, ps_res1, ps_res2, ps_res3;
134
135 x_tmp0 = *xmm_crc0;
136 x_tmp1 = *xmm_crc1;
137 x_tmp2 = *xmm_crc2;
138 x_tmp3 = *xmm_crc3;
139
140 *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01);
141 x_tmp0 = _mm_clmulepi64_si128(x_tmp0, xmm_fold4, 0x10);
142 ps_crc0 = _mm_castsi128_ps(*xmm_crc0);
143 ps_t0 = _mm_castsi128_ps(x_tmp0);
144 ps_res0 = _mm_xor_ps(ps_crc0, ps_t0);
145
146 *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x01);
147 x_tmp1 = _mm_clmulepi64_si128(x_tmp1, xmm_fold4, 0x10);
148 ps_crc1 = _mm_castsi128_ps(*xmm_crc1);
149 ps_t1 = _mm_castsi128_ps(x_tmp1);
150 ps_res1 = _mm_xor_ps(ps_crc1, ps_t1);
151
152 *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x01);
153 x_tmp2 = _mm_clmulepi64_si128(x_tmp2, xmm_fold4, 0x10);
154 ps_crc2 = _mm_castsi128_ps(*xmm_crc2);
155 ps_t2 = _mm_castsi128_ps(x_tmp2);
156 ps_res2 = _mm_xor_ps(ps_crc2, ps_t2);
157
158 *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x01);
159 x_tmp3 = _mm_clmulepi64_si128(x_tmp3, xmm_fold4, 0x10);
160 ps_crc3 = _mm_castsi128_ps(*xmm_crc3);
161 ps_t3 = _mm_castsi128_ps(x_tmp3);
162 ps_res3 = _mm_xor_ps(ps_crc3, ps_t3);
163
164 *xmm_crc0 = _mm_castps_si128(ps_res0);
165 *xmm_crc1 = _mm_castps_si128(ps_res1);
166 *xmm_crc2 = _mm_castps_si128(ps_res2);
167 *xmm_crc3 = _mm_castps_si128(ps_res3);
168 }
169
170 static const unsigned ALIGNED_(32) pshufb_shf_table[60] = {
171 0x84838281, 0x88878685, 0x8c8b8a89, 0x008f8e8d, /* shl 15 (16 - 1)/shr1 */
172 0x85848382, 0x89888786, 0x8d8c8b8a, 0x01008f8e, /* shl 14 (16 - 3)/shr2 */
173 0x86858483, 0x8a898887, 0x8e8d8c8b, 0x0201008f, /* shl 13 (16 - 4)/shr3 */
174 0x87868584, 0x8b8a8988, 0x8f8e8d8c, 0x03020100, /* shl 12 (16 - 4)/shr4 */
175 0x88878685, 0x8c8b8a89, 0x008f8e8d, 0x04030201, /* shl 11 (16 - 5)/shr5 */
176 0x89888786, 0x8d8c8b8a, 0x01008f8e, 0x05040302, /* shl 10 (16 - 6)/shr6 */
177 0x8a898887, 0x8e8d8c8b, 0x0201008f, 0x06050403, /* shl 9 (16 - 7)/shr7 */
178 0x8b8a8988, 0x8f8e8d8c, 0x03020100, 0x07060504, /* shl 8 (16 - 8)/shr8 */
179 0x8c8b8a89, 0x008f8e8d, 0x04030201, 0x08070605, /* shl 7 (16 - 9)/shr9 */
180 0x8d8c8b8a, 0x01008f8e, 0x05040302, 0x09080706, /* shl 6 (16 -10)/shr10*/
181 0x8e8d8c8b, 0x0201008f, 0x06050403, 0x0a090807, /* shl 5 (16 -11)/shr11*/
182 0x8f8e8d8c, 0x03020100, 0x07060504, 0x0b0a0908, /* shl 4 (16 -12)/shr12*/
183 0x008f8e8d, 0x04030201, 0x08070605, 0x0c0b0a09, /* shl 3 (16 -13)/shr13*/
184 0x01008f8e, 0x05040302, 0x09080706, 0x0d0c0b0a, /* shl 2 (16 -14)/shr14*/
185 0x0201008f, 0x06050403, 0x0a090807, 0x0e0d0c0b /* shl 1 (16 -15)/shr15*/
186 };
187
partial_fold(const size_t len,__m128i * xmm_crc0,__m128i * xmm_crc1,__m128i * xmm_crc2,__m128i * xmm_crc3,__m128i * xmm_crc_part)188 static void partial_fold(const size_t len, __m128i *xmm_crc0, __m128i *xmm_crc1, __m128i *xmm_crc2,
189 __m128i *xmm_crc3, __m128i *xmm_crc_part) {
190
191 const __m128i xmm_fold4 = _mm_set_epi32( 0x00000001, 0x54442bd4,
192 0x00000001, 0xc6e41596);
193 const __m128i xmm_mask3 = _mm_set1_epi32((int32_t)0x80808080);
194
195 __m128i xmm_shl, xmm_shr, xmm_tmp1, xmm_tmp2, xmm_tmp3;
196 __m128i xmm_a0_0, xmm_a0_1;
197 __m128 ps_crc3, psa0_0, psa0_1, ps_res;
198
199 xmm_shl = _mm_load_si128((__m128i *)pshufb_shf_table + (len - 1));
200 xmm_shr = xmm_shl;
201 xmm_shr = _mm_xor_si128(xmm_shr, xmm_mask3);
202
203 xmm_a0_0 = _mm_shuffle_epi8(*xmm_crc0, xmm_shl);
204
205 *xmm_crc0 = _mm_shuffle_epi8(*xmm_crc0, xmm_shr);
206 xmm_tmp1 = _mm_shuffle_epi8(*xmm_crc1, xmm_shl);
207 *xmm_crc0 = _mm_or_si128(*xmm_crc0, xmm_tmp1);
208
209 *xmm_crc1 = _mm_shuffle_epi8(*xmm_crc1, xmm_shr);
210 xmm_tmp2 = _mm_shuffle_epi8(*xmm_crc2, xmm_shl);
211 *xmm_crc1 = _mm_or_si128(*xmm_crc1, xmm_tmp2);
212
213 *xmm_crc2 = _mm_shuffle_epi8(*xmm_crc2, xmm_shr);
214 xmm_tmp3 = _mm_shuffle_epi8(*xmm_crc3, xmm_shl);
215 *xmm_crc2 = _mm_or_si128(*xmm_crc2, xmm_tmp3);
216
217 *xmm_crc3 = _mm_shuffle_epi8(*xmm_crc3, xmm_shr);
218 *xmm_crc_part = _mm_shuffle_epi8(*xmm_crc_part, xmm_shl);
219 *xmm_crc3 = _mm_or_si128(*xmm_crc3, *xmm_crc_part);
220
221 xmm_a0_1 = _mm_clmulepi64_si128(xmm_a0_0, xmm_fold4, 0x10);
222 xmm_a0_0 = _mm_clmulepi64_si128(xmm_a0_0, xmm_fold4, 0x01);
223
224 ps_crc3 = _mm_castsi128_ps(*xmm_crc3);
225 psa0_0 = _mm_castsi128_ps(xmm_a0_0);
226 psa0_1 = _mm_castsi128_ps(xmm_a0_1);
227
228 ps_res = _mm_xor_ps(ps_crc3, psa0_0);
229 ps_res = _mm_xor_ps(ps_res, psa0_1);
230
231 *xmm_crc3 = _mm_castps_si128(ps_res);
232 }
233
crc32_fold_load(__m128i * fold,__m128i * fold0,__m128i * fold1,__m128i * fold2,__m128i * fold3)234 static inline void crc32_fold_load(__m128i *fold, __m128i *fold0, __m128i *fold1, __m128i *fold2, __m128i *fold3) {
235 *fold0 = _mm_load_si128(fold + 0);
236 *fold1 = _mm_load_si128(fold + 1);
237 *fold2 = _mm_load_si128(fold + 2);
238 *fold3 = _mm_load_si128(fold + 3);
239 }
240
crc32_fold_save(__m128i * fold,__m128i fold0,__m128i fold1,__m128i fold2,__m128i fold3)241 static inline void crc32_fold_save(__m128i *fold, __m128i fold0, __m128i fold1, __m128i fold2, __m128i fold3) {
242 _mm_storeu_si128(fold + 0, fold0);
243 _mm_storeu_si128(fold + 1, fold1);
244 _mm_storeu_si128(fold + 2, fold2);
245 _mm_storeu_si128(fold + 3, fold3);
246 }
247
crc32_fold_save_partial(__m128i * fold,__m128i foldp)248 static inline void crc32_fold_save_partial(__m128i *fold, __m128i foldp) {
249 _mm_store_si128(fold + 4, foldp);
250 }
251
crc32_fold_reset_pclmulqdq(crc32_fold * crc)252 Z_INTERNAL uint32_t crc32_fold_reset_pclmulqdq(crc32_fold *crc) {
253 __m128i xmm_crc0 = _mm_cvtsi32_si128(0x9db42487);
254 __m128i xmm_zero = _mm_setzero_si128();
255 crc32_fold_save((__m128i *)crc->fold, xmm_crc0, xmm_zero, xmm_zero, xmm_zero);
256 return 0;
257 }
258
crc32_fold_copy_pclmulqdq(crc32_fold * crc,uint8_t * dst,const uint8_t * src,size_t len)259 Z_INTERNAL void crc32_fold_copy_pclmulqdq(crc32_fold *crc, uint8_t *dst, const uint8_t *src, size_t len) {
260 unsigned long algn_diff;
261 __m128i xmm_t0, xmm_t1, xmm_t2, xmm_t3;
262 __m128i xmm_crc0, xmm_crc1, xmm_crc2, xmm_crc3, xmm_crc_part;
263 char ALIGNED_(16) partial_buf[16] = { 0 };
264
265 crc32_fold_load((__m128i *)crc->fold, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
266
267 if (len < 16) {
268 if (len == 0)
269 return;
270
271 memcpy(partial_buf, src, len);
272 xmm_crc_part = _mm_load_si128((const __m128i *)partial_buf);
273 memcpy(dst, partial_buf, len);
274 goto partial;
275 }
276
277 algn_diff = ((uintptr_t)16 - ((uintptr_t)src & 0xF)) & 0xF;
278 if (algn_diff) {
279 xmm_crc_part = _mm_loadu_si128((__m128i *)src);
280 _mm_storeu_si128((__m128i *)dst, xmm_crc_part);
281
282 dst += algn_diff;
283 src += algn_diff;
284 len -= algn_diff;
285
286 partial_fold(algn_diff, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3, &xmm_crc_part);
287 } else {
288 xmm_crc_part = _mm_setzero_si128();
289 }
290
291 #ifdef X86_VPCLMULQDQ_CRC
292 if (x86_cpu_has_vpclmulqdq && x86_cpu_has_avx512 && (len >= 256)) {
293 size_t n = fold_16_vpclmulqdq(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3, dst, src, len);
294
295 len -= n;
296 src += n;
297 dst += n;
298 }
299 #endif
300
301 while (len >= 64) {
302 crc32_fold_load((__m128i *)src, &xmm_t0, &xmm_t1, &xmm_t2, &xmm_t3);
303
304 fold_4(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
305
306 crc32_fold_save((__m128i *)dst, xmm_t0, xmm_t1, xmm_t2, xmm_t3);
307
308 xmm_crc0 = _mm_xor_si128(xmm_crc0, xmm_t0);
309 xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_t1);
310 xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t2);
311 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t3);
312
313 src += 64;
314 dst += 64;
315 len -= 64;
316 }
317
318 /*
319 * len = num bytes left - 64
320 */
321 if (len >= 48) {
322 xmm_t0 = _mm_load_si128((__m128i *)src);
323 xmm_t1 = _mm_load_si128((__m128i *)src + 1);
324 xmm_t2 = _mm_load_si128((__m128i *)src + 2);
325
326 fold_3(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
327
328 _mm_storeu_si128((__m128i *)dst, xmm_t0);
329 _mm_storeu_si128((__m128i *)dst + 1, xmm_t1);
330 _mm_storeu_si128((__m128i *)dst + 2, xmm_t2);
331
332 xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_t0);
333 xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t1);
334 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t2);
335 len -= 48;
336 if (len == 0)
337 goto done;
338
339 dst += 48;
340 memcpy(&xmm_crc_part, (__m128i *)src + 3, len);
341 } else if (len >= 32) {
342 xmm_t0 = _mm_load_si128((__m128i *)src);
343 xmm_t1 = _mm_load_si128((__m128i *)src + 1);
344
345 fold_2(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
346
347 _mm_storeu_si128((__m128i *)dst, xmm_t0);
348 _mm_storeu_si128((__m128i *)dst + 1, xmm_t1);
349
350 xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t0);
351 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t1);
352
353 len -= 32;
354 if (len == 0)
355 goto done;
356
357 dst += 32;
358 memcpy(&xmm_crc_part, (__m128i *)src + 2, len);
359 } else if (len >= 16) {
360 xmm_t0 = _mm_load_si128((__m128i *)src);
361
362 fold_1(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
363
364 _mm_storeu_si128((__m128i *)dst, xmm_t0);
365
366 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t0);
367
368 len -= 16;
369 if (len == 0)
370 goto done;
371
372 dst += 16;
373 memcpy(&xmm_crc_part, (__m128i *)src + 1, len);
374 } else {
375 if (len == 0)
376 goto done;
377 memcpy(&xmm_crc_part, src, len);
378 }
379
380 _mm_storeu_si128((__m128i *)partial_buf, xmm_crc_part);
381 memcpy(dst, partial_buf, len);
382
383 partial:
384 partial_fold((size_t)len, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3, &xmm_crc_part);
385 done:
386 crc32_fold_save((__m128i *)crc->fold, xmm_crc0, xmm_crc1, xmm_crc2, xmm_crc3);
387 crc32_fold_save_partial((__m128i *)crc->fold, xmm_crc_part);
388 }
389
390 #define ONCE(op) if (first) { \
391 first = 0; \
392 (op); \
393 }
394 #define XOR_INITIAL(where) ONCE(where = _mm_xor_si128(where, xmm_initial))
395
crc32_fold_pclmulqdq(crc32_fold * crc,const uint8_t * src,size_t len,uint32_t init_crc)396 Z_INTERNAL void crc32_fold_pclmulqdq(crc32_fold *crc, const uint8_t *src, size_t len, uint32_t init_crc) {
397 unsigned long algn_diff;
398 __m128i xmm_t0, xmm_t1, xmm_t2, xmm_t3;
399 __m128i xmm_crc0, xmm_crc1, xmm_crc2, xmm_crc3, xmm_crc_part;
400 __m128i xmm_initial = _mm_cvtsi32_si128(init_crc);
401 xmm_crc_part = _mm_setzero_si128();
402 int32_t first = init_crc != 0;
403
404 /* Technically the CRC functions don't even call this for input < 64, but a bare minimum of 31
405 * bytes of input is needed for the aligning load that occurs. If there's an initial CRC, to
406 * carry it forward through the folded CRC there must be 16 - src % 16 + 16 bytes available, which
407 * by definition can be up to 15 bytes + one full vector load. */
408 assert(len >= 31 || first == 0);
409 crc32_fold_load((__m128i *)crc->fold, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
410
411 if (len < 16) {
412 goto partial_nocpy;
413 }
414
415 algn_diff = ((uintptr_t)16 - ((uintptr_t)src & 0xF)) & 0xF;
416 if (algn_diff) {
417 if (algn_diff >= 4 || init_crc == 0) {
418 xmm_crc_part = _mm_loadu_si128((__m128i *)src);
419
420 src += algn_diff;
421 len -= algn_diff;
422
423 XOR_INITIAL(xmm_crc_part);
424 partial_fold(algn_diff, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3, &xmm_crc_part);
425 } else {
426 xmm_t0 = _mm_loadu_si128((__m128i*)src);
427 xmm_crc_part = _mm_loadu_si128((__m128i*)src + 1);
428 XOR_INITIAL(xmm_t0);
429 fold_1(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
430 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t0);
431 partial_fold(algn_diff, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3, &xmm_crc_part);
432
433 src += (algn_diff + 16);
434 len -= (algn_diff + 16);
435 }
436
437 xmm_crc_part = _mm_setzero_si128();
438 }
439
440 #ifdef X86_VPCLMULQDQ_CRC
441 if (x86_cpu_has_vpclmulqdq && x86_cpu_has_avx512 && (len >= 256)) {
442 size_t n = fold_16_vpclmulqdq_nocp(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3, src, len,
443 xmm_initial, first);
444 first = 0;
445
446 len -= n;
447 src += n;
448 }
449 #endif
450
451 while (len >= 64) {
452 crc32_fold_load((__m128i *)src, &xmm_t0, &xmm_t1, &xmm_t2, &xmm_t3);
453 XOR_INITIAL(xmm_t0);
454 fold_4(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
455
456 xmm_crc0 = _mm_xor_si128(xmm_crc0, xmm_t0);
457 xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_t1);
458 xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t2);
459 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t3);
460
461 src += 64;
462 len -= 64;
463 }
464
465 /*
466 * len = num bytes left - 64
467 */
468 if (len >= 48) {
469 xmm_t0 = _mm_load_si128((__m128i *)src);
470 xmm_t1 = _mm_load_si128((__m128i *)src + 1);
471 xmm_t2 = _mm_load_si128((__m128i *)src + 2);
472 XOR_INITIAL(xmm_t0);
473
474 fold_3(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
475
476 xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_t0);
477 xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t1);
478 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t2);
479 len -= 48;
480 src += 48;
481 } else if (len >= 32) {
482 xmm_t0 = _mm_load_si128((__m128i *)src);
483 xmm_t1 = _mm_load_si128((__m128i *)src + 1);
484 XOR_INITIAL(xmm_t0);
485
486 fold_2(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
487
488 xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t0);
489 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t1);
490
491 len -= 32;
492 src += 32;
493 } else if (len >= 16) {
494 xmm_t0 = _mm_load_si128((__m128i *)src);
495 XOR_INITIAL(xmm_t0);
496
497 fold_1(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
498
499 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t0);
500
501 len -= 16;
502 src += 16;
503 }
504
505 partial_nocpy:
506 if (len) {
507 memcpy(&xmm_crc_part, src, len);
508 partial_fold((size_t)len, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3, &xmm_crc_part);
509 }
510
511 crc32_fold_save((__m128i *)crc->fold, xmm_crc0, xmm_crc1, xmm_crc2, xmm_crc3);
512 }
513
514 static const unsigned ALIGNED_(16) crc_k[] = {
515 0xccaa009e, 0x00000000, /* rk1 */
516 0x751997d0, 0x00000001, /* rk2 */
517 0xccaa009e, 0x00000000, /* rk5 */
518 0x63cd6124, 0x00000001, /* rk6 */
519 0xf7011640, 0x00000001, /* rk7 */
520 0xdb710640, 0x00000001 /* rk8 */
521 };
522
523 static const unsigned ALIGNED_(16) crc_mask[4] = {
524 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000
525 };
526
527 static const unsigned ALIGNED_(16) crc_mask2[4] = {
528 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF
529 };
530
crc32_fold_final_pclmulqdq(crc32_fold * crc)531 Z_INTERNAL uint32_t crc32_fold_final_pclmulqdq(crc32_fold *crc) {
532 const __m128i xmm_mask = _mm_load_si128((__m128i *)crc_mask);
533 const __m128i xmm_mask2 = _mm_load_si128((__m128i *)crc_mask2);
534 __m128i xmm_crc0, xmm_crc1, xmm_crc2, xmm_crc3;
535 __m128i x_tmp0, x_tmp1, x_tmp2, crc_fold;
536
537 crc32_fold_load((__m128i *)crc->fold, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3);
538
539 /*
540 * k1
541 */
542 crc_fold = _mm_load_si128((__m128i *)crc_k);
543
544 x_tmp0 = _mm_clmulepi64_si128(xmm_crc0, crc_fold, 0x10);
545 xmm_crc0 = _mm_clmulepi64_si128(xmm_crc0, crc_fold, 0x01);
546 xmm_crc1 = _mm_xor_si128(xmm_crc1, x_tmp0);
547 xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_crc0);
548
549 x_tmp1 = _mm_clmulepi64_si128(xmm_crc1, crc_fold, 0x10);
550 xmm_crc1 = _mm_clmulepi64_si128(xmm_crc1, crc_fold, 0x01);
551 xmm_crc2 = _mm_xor_si128(xmm_crc2, x_tmp1);
552 xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_crc1);
553
554 x_tmp2 = _mm_clmulepi64_si128(xmm_crc2, crc_fold, 0x10);
555 xmm_crc2 = _mm_clmulepi64_si128(xmm_crc2, crc_fold, 0x01);
556 xmm_crc3 = _mm_xor_si128(xmm_crc3, x_tmp2);
557 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc2);
558
559 /*
560 * k5
561 */
562 crc_fold = _mm_load_si128((__m128i *)crc_k + 1);
563
564 xmm_crc0 = xmm_crc3;
565 xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0);
566 xmm_crc0 = _mm_srli_si128(xmm_crc0, 8);
567 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc0);
568
569 xmm_crc0 = xmm_crc3;
570 xmm_crc3 = _mm_slli_si128(xmm_crc3, 4);
571 xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0x10);
572 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc0);
573 xmm_crc3 = _mm_and_si128(xmm_crc3, xmm_mask2);
574
575 /*
576 * k7
577 */
578 xmm_crc1 = xmm_crc3;
579 xmm_crc2 = xmm_crc3;
580 crc_fold = _mm_load_si128((__m128i *)crc_k + 2);
581
582 xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0);
583 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc2);
584 xmm_crc3 = _mm_and_si128(xmm_crc3, xmm_mask);
585
586 xmm_crc2 = xmm_crc3;
587 xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0x10);
588 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc2);
589 xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc1);
590
591 crc->value = ~((uint32_t)_mm_extract_epi32(xmm_crc3, 2));
592
593 return crc->value;
594 }
595
crc32_pclmulqdq(uint32_t crc32,const unsigned char * buf,uint64_t len)596 uint32_t crc32_pclmulqdq(uint32_t crc32, const unsigned char* buf, uint64_t len) {
597 /* For lens < 64, crc32_braid method is faster. The CRC32 instruction for
598 * these short lengths might also prove to be effective */
599 if (len < 64)
600 return crc32_braid(crc32, buf, len);
601
602 crc32_fold ALIGNED_(16) crc_state;
603 crc32_fold_reset_pclmulqdq(&crc_state);
604 crc32_fold_pclmulqdq(&crc_state, buf, len, crc32);
605 return crc32_fold_final_pclmulqdq(&crc_state);
606 }
607
608 #endif
609