]>
Commit | Line | Data |
---|---|---|
d2912cb1 | 1 | /* SPDX-License-Identifier: GPL-2.0-only */ |
6ecc9d9f OM |
2 | /* |
3 | * AVX2 implementation of MORUS-1280 | |
4 | * | |
5 | * Copyright (c) 2017-2018 Ondrej Mosnacek <omosnacek@gmail.com> | |
6 | * Copyright (C) 2017-2018 Red Hat, Inc. All rights reserved. | |
6ecc9d9f OM |
7 | */ |
8 | ||
9 | #include <linux/linkage.h> | |
10 | #include <asm/frame.h> | |
11 | ||
12 | #define SHUFFLE_MASK(i0, i1, i2, i3) \ | |
13 | (i0 | (i1 << 2) | (i2 << 4) | (i3 << 6)) | |
14 | ||
15 | #define MASK1 SHUFFLE_MASK(3, 0, 1, 2) | |
16 | #define MASK2 SHUFFLE_MASK(2, 3, 0, 1) | |
17 | #define MASK3 SHUFFLE_MASK(1, 2, 3, 0) | |
18 | ||
19 | #define STATE0 %ymm0 | |
20 | #define STATE0_LOW %xmm0 | |
21 | #define STATE1 %ymm1 | |
22 | #define STATE2 %ymm2 | |
23 | #define STATE3 %ymm3 | |
24 | #define STATE4 %ymm4 | |
25 | #define KEY %ymm5 | |
26 | #define MSG %ymm5 | |
27 | #define MSG_LOW %xmm5 | |
28 | #define T0 %ymm6 | |
29 | #define T0_LOW %xmm6 | |
30 | #define T1 %ymm7 | |
31 | ||
32 | .section .rodata.cst32.morus1280_const, "aM", @progbits, 32 | |
33 | .align 32 | |
34 | .Lmorus1280_const: | |
35 | .byte 0x00, 0x01, 0x01, 0x02, 0x03, 0x05, 0x08, 0x0d | |
36 | .byte 0x15, 0x22, 0x37, 0x59, 0x90, 0xe9, 0x79, 0x62 | |
37 | .byte 0xdb, 0x3d, 0x18, 0x55, 0x6d, 0xc2, 0x2f, 0xf1 | |
38 | .byte 0x20, 0x11, 0x31, 0x42, 0x73, 0xb5, 0x28, 0xdd | |
39 | ||
40 | .section .rodata.cst32.morus1280_counter, "aM", @progbits, 32 | |
41 | .align 32 | |
42 | .Lmorus1280_counter: | |
43 | .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 | |
44 | .byte 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f | |
45 | .byte 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17 | |
46 | .byte 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f | |
47 | ||
48 | .text | |
49 | ||
50 | .macro morus1280_round s0, s1, s2, s3, s4, b, w | |
51 | vpand \s1, \s2, T0 | |
52 | vpxor T0, \s0, \s0 | |
53 | vpxor \s3, \s0, \s0 | |
54 | vpsllq $\b, \s0, T0 | |
55 | vpsrlq $(64 - \b), \s0, \s0 | |
56 | vpxor T0, \s0, \s0 | |
57 | vpermq $\w, \s3, \s3 | |
58 | .endm | |
59 | ||
60 | /* | |
61 | * __morus1280_update: internal ABI | |
62 | * input: | |
63 | * STATE[0-4] - input state | |
64 | * MSG - message block | |
65 | * output: | |
66 | * STATE[0-4] - output state | |
67 | * changed: | |
68 | * T0 | |
69 | */ | |
70 | __morus1280_update: | |
71 | morus1280_round STATE0, STATE1, STATE2, STATE3, STATE4, 13, MASK1 | |
72 | vpxor MSG, STATE1, STATE1 | |
73 | morus1280_round STATE1, STATE2, STATE3, STATE4, STATE0, 46, MASK2 | |
74 | vpxor MSG, STATE2, STATE2 | |
75 | morus1280_round STATE2, STATE3, STATE4, STATE0, STATE1, 38, MASK3 | |
76 | vpxor MSG, STATE3, STATE3 | |
77 | morus1280_round STATE3, STATE4, STATE0, STATE1, STATE2, 7, MASK2 | |
78 | vpxor MSG, STATE4, STATE4 | |
79 | morus1280_round STATE4, STATE0, STATE1, STATE2, STATE3, 4, MASK1 | |
80 | ret | |
81 | ENDPROC(__morus1280_update) | |
82 | ||
83 | /* | |
84 | * __morus1280_update_zero: internal ABI | |
85 | * input: | |
86 | * STATE[0-4] - input state | |
87 | * output: | |
88 | * STATE[0-4] - output state | |
89 | * changed: | |
90 | * T0 | |
91 | */ | |
92 | __morus1280_update_zero: | |
93 | morus1280_round STATE0, STATE1, STATE2, STATE3, STATE4, 13, MASK1 | |
94 | morus1280_round STATE1, STATE2, STATE3, STATE4, STATE0, 46, MASK2 | |
95 | morus1280_round STATE2, STATE3, STATE4, STATE0, STATE1, 38, MASK3 | |
96 | morus1280_round STATE3, STATE4, STATE0, STATE1, STATE2, 7, MASK2 | |
97 | morus1280_round STATE4, STATE0, STATE1, STATE2, STATE3, 4, MASK1 | |
98 | ret | |
99 | ENDPROC(__morus1280_update_zero) | |
100 | ||
101 | /* | |
102 | * __load_partial: internal ABI | |
103 | * input: | |
104 | * %rsi - src | |
105 | * %rcx - bytes | |
106 | * output: | |
107 | * MSG - message block | |
108 | * changed: | |
109 | * %r8 | |
110 | * %r9 | |
111 | */ | |
112 | __load_partial: | |
a7bea830 | 113 | xor %r9d, %r9d |
6ecc9d9f OM |
114 | vpxor MSG, MSG, MSG |
115 | ||
116 | mov %rcx, %r8 | |
117 | and $0x1, %r8 | |
118 | jz .Lld_partial_1 | |
119 | ||
120 | mov %rcx, %r8 | |
121 | and $0x1E, %r8 | |
122 | add %rsi, %r8 | |
123 | mov (%r8), %r9b | |
124 | ||
125 | .Lld_partial_1: | |
126 | mov %rcx, %r8 | |
127 | and $0x2, %r8 | |
128 | jz .Lld_partial_2 | |
129 | ||
130 | mov %rcx, %r8 | |
131 | and $0x1C, %r8 | |
132 | add %rsi, %r8 | |
133 | shl $16, %r9 | |
134 | mov (%r8), %r9w | |
135 | ||
136 | .Lld_partial_2: | |
137 | mov %rcx, %r8 | |
138 | and $0x4, %r8 | |
139 | jz .Lld_partial_4 | |
140 | ||
141 | mov %rcx, %r8 | |
142 | and $0x18, %r8 | |
143 | add %rsi, %r8 | |
144 | shl $32, %r9 | |
145 | mov (%r8), %r8d | |
146 | xor %r8, %r9 | |
147 | ||
148 | .Lld_partial_4: | |
149 | movq %r9, MSG_LOW | |
150 | ||
151 | mov %rcx, %r8 | |
152 | and $0x8, %r8 | |
153 | jz .Lld_partial_8 | |
154 | ||
155 | mov %rcx, %r8 | |
156 | and $0x10, %r8 | |
157 | add %rsi, %r8 | |
158 | pshufd $MASK2, MSG_LOW, MSG_LOW | |
159 | pinsrq $0, (%r8), MSG_LOW | |
160 | ||
161 | .Lld_partial_8: | |
162 | mov %rcx, %r8 | |
163 | and $0x10, %r8 | |
164 | jz .Lld_partial_16 | |
165 | ||
166 | vpermq $MASK2, MSG, MSG | |
167 | movdqu (%rsi), MSG_LOW | |
168 | ||
169 | .Lld_partial_16: | |
170 | ret | |
171 | ENDPROC(__load_partial) | |
172 | ||
173 | /* | |
174 | * __store_partial: internal ABI | |
175 | * input: | |
176 | * %rdx - dst | |
177 | * %rcx - bytes | |
178 | * output: | |
179 | * T0 - message block | |
180 | * changed: | |
181 | * %r8 | |
182 | * %r9 | |
183 | * %r10 | |
184 | */ | |
185 | __store_partial: | |
186 | mov %rcx, %r8 | |
187 | mov %rdx, %r9 | |
188 | ||
189 | cmp $16, %r8 | |
190 | jl .Lst_partial_16 | |
191 | ||
192 | movdqu T0_LOW, (%r9) | |
193 | vpermq $MASK2, T0, T0 | |
194 | ||
195 | sub $16, %r8 | |
196 | add $16, %r9 | |
197 | ||
198 | .Lst_partial_16: | |
199 | movq T0_LOW, %r10 | |
200 | ||
201 | cmp $8, %r8 | |
202 | jl .Lst_partial_8 | |
203 | ||
204 | mov %r10, (%r9) | |
205 | pextrq $1, T0_LOW, %r10 | |
206 | ||
207 | sub $8, %r8 | |
208 | add $8, %r9 | |
209 | ||
210 | .Lst_partial_8: | |
211 | cmp $4, %r8 | |
212 | jl .Lst_partial_4 | |
213 | ||
214 | mov %r10d, (%r9) | |
215 | shr $32, %r10 | |
216 | ||
217 | sub $4, %r8 | |
218 | add $4, %r9 | |
219 | ||
220 | .Lst_partial_4: | |
221 | cmp $2, %r8 | |
222 | jl .Lst_partial_2 | |
223 | ||
224 | mov %r10w, (%r9) | |
225 | shr $16, %r10 | |
226 | ||
227 | sub $2, %r8 | |
228 | add $2, %r9 | |
229 | ||
230 | .Lst_partial_2: | |
231 | cmp $1, %r8 | |
232 | jl .Lst_partial_1 | |
233 | ||
234 | mov %r10b, (%r9) | |
235 | ||
236 | .Lst_partial_1: | |
237 | ret | |
238 | ENDPROC(__store_partial) | |
239 | ||
240 | /* | |
241 | * void crypto_morus1280_avx2_init(void *state, const void *key, | |
242 | * const void *iv); | |
243 | */ | |
244 | ENTRY(crypto_morus1280_avx2_init) | |
245 | FRAME_BEGIN | |
246 | ||
247 | /* load IV: */ | |
248 | vpxor STATE0, STATE0, STATE0 | |
249 | movdqu (%rdx), STATE0_LOW | |
250 | /* load key: */ | |
251 | vmovdqu (%rsi), KEY | |
252 | vmovdqa KEY, STATE1 | |
253 | /* load all ones: */ | |
254 | vpcmpeqd STATE2, STATE2, STATE2 | |
255 | /* load all zeros: */ | |
256 | vpxor STATE3, STATE3, STATE3 | |
257 | /* load the constant: */ | |
258 | vmovdqa .Lmorus1280_const, STATE4 | |
259 | ||
260 | /* update 16 times with zero: */ | |
261 | call __morus1280_update_zero | |
262 | call __morus1280_update_zero | |
263 | call __morus1280_update_zero | |
264 | call __morus1280_update_zero | |
265 | call __morus1280_update_zero | |
266 | call __morus1280_update_zero | |
267 | call __morus1280_update_zero | |
268 | call __morus1280_update_zero | |
269 | call __morus1280_update_zero | |
270 | call __morus1280_update_zero | |
271 | call __morus1280_update_zero | |
272 | call __morus1280_update_zero | |
273 | call __morus1280_update_zero | |
274 | call __morus1280_update_zero | |
275 | call __morus1280_update_zero | |
276 | call __morus1280_update_zero | |
277 | ||
278 | /* xor-in the key again after updates: */ | |
279 | vpxor KEY, STATE1, STATE1 | |
280 | ||
281 | /* store the state: */ | |
282 | vmovdqu STATE0, (0 * 32)(%rdi) | |
283 | vmovdqu STATE1, (1 * 32)(%rdi) | |
284 | vmovdqu STATE2, (2 * 32)(%rdi) | |
285 | vmovdqu STATE3, (3 * 32)(%rdi) | |
286 | vmovdqu STATE4, (4 * 32)(%rdi) | |
287 | ||
288 | FRAME_END | |
289 | ret | |
290 | ENDPROC(crypto_morus1280_avx2_init) | |
291 | ||
292 | /* | |
293 | * void crypto_morus1280_avx2_ad(void *state, const void *data, | |
294 | * unsigned int length); | |
295 | */ | |
296 | ENTRY(crypto_morus1280_avx2_ad) | |
297 | FRAME_BEGIN | |
298 | ||
299 | cmp $32, %rdx | |
300 | jb .Lad_out | |
301 | ||
302 | /* load the state: */ | |
303 | vmovdqu (0 * 32)(%rdi), STATE0 | |
304 | vmovdqu (1 * 32)(%rdi), STATE1 | |
305 | vmovdqu (2 * 32)(%rdi), STATE2 | |
306 | vmovdqu (3 * 32)(%rdi), STATE3 | |
307 | vmovdqu (4 * 32)(%rdi), STATE4 | |
308 | ||
309 | mov %rsi, %r8 | |
310 | and $0x1F, %r8 | |
311 | jnz .Lad_u_loop | |
312 | ||
313 | .align 4 | |
314 | .Lad_a_loop: | |
315 | vmovdqa (%rsi), MSG | |
316 | call __morus1280_update | |
317 | sub $32, %rdx | |
318 | add $32, %rsi | |
319 | cmp $32, %rdx | |
320 | jge .Lad_a_loop | |
321 | ||
322 | jmp .Lad_cont | |
323 | .align 4 | |
324 | .Lad_u_loop: | |
325 | vmovdqu (%rsi), MSG | |
326 | call __morus1280_update | |
327 | sub $32, %rdx | |
328 | add $32, %rsi | |
329 | cmp $32, %rdx | |
330 | jge .Lad_u_loop | |
331 | ||
332 | .Lad_cont: | |
333 | /* store the state: */ | |
334 | vmovdqu STATE0, (0 * 32)(%rdi) | |
335 | vmovdqu STATE1, (1 * 32)(%rdi) | |
336 | vmovdqu STATE2, (2 * 32)(%rdi) | |
337 | vmovdqu STATE3, (3 * 32)(%rdi) | |
338 | vmovdqu STATE4, (4 * 32)(%rdi) | |
339 | ||
340 | .Lad_out: | |
341 | FRAME_END | |
342 | ret | |
343 | ENDPROC(crypto_morus1280_avx2_ad) | |
344 | ||
345 | /* | |
346 | * void crypto_morus1280_avx2_enc(void *state, const void *src, void *dst, | |
347 | * unsigned int length); | |
348 | */ | |
349 | ENTRY(crypto_morus1280_avx2_enc) | |
350 | FRAME_BEGIN | |
351 | ||
352 | cmp $32, %rcx | |
353 | jb .Lenc_out | |
354 | ||
355 | /* load the state: */ | |
356 | vmovdqu (0 * 32)(%rdi), STATE0 | |
357 | vmovdqu (1 * 32)(%rdi), STATE1 | |
358 | vmovdqu (2 * 32)(%rdi), STATE2 | |
359 | vmovdqu (3 * 32)(%rdi), STATE3 | |
360 | vmovdqu (4 * 32)(%rdi), STATE4 | |
361 | ||
362 | mov %rsi, %r8 | |
363 | or %rdx, %r8 | |
364 | and $0x1F, %r8 | |
365 | jnz .Lenc_u_loop | |
366 | ||
367 | .align 4 | |
368 | .Lenc_a_loop: | |
369 | vmovdqa (%rsi), MSG | |
370 | vmovdqa MSG, T0 | |
371 | vpxor STATE0, T0, T0 | |
372 | vpermq $MASK3, STATE1, T1 | |
373 | vpxor T1, T0, T0 | |
374 | vpand STATE2, STATE3, T1 | |
375 | vpxor T1, T0, T0 | |
376 | vmovdqa T0, (%rdx) | |
377 | ||
378 | call __morus1280_update | |
379 | sub $32, %rcx | |
380 | add $32, %rsi | |
381 | add $32, %rdx | |
382 | cmp $32, %rcx | |
383 | jge .Lenc_a_loop | |
384 | ||
385 | jmp .Lenc_cont | |
386 | .align 4 | |
387 | .Lenc_u_loop: | |
388 | vmovdqu (%rsi), MSG | |
389 | vmovdqa MSG, T0 | |
390 | vpxor STATE0, T0, T0 | |
391 | vpermq $MASK3, STATE1, T1 | |
392 | vpxor T1, T0, T0 | |
393 | vpand STATE2, STATE3, T1 | |
394 | vpxor T1, T0, T0 | |
395 | vmovdqu T0, (%rdx) | |
396 | ||
397 | call __morus1280_update | |
398 | sub $32, %rcx | |
399 | add $32, %rsi | |
400 | add $32, %rdx | |
401 | cmp $32, %rcx | |
402 | jge .Lenc_u_loop | |
403 | ||
404 | .Lenc_cont: | |
405 | /* store the state: */ | |
406 | vmovdqu STATE0, (0 * 32)(%rdi) | |
407 | vmovdqu STATE1, (1 * 32)(%rdi) | |
408 | vmovdqu STATE2, (2 * 32)(%rdi) | |
409 | vmovdqu STATE3, (3 * 32)(%rdi) | |
410 | vmovdqu STATE4, (4 * 32)(%rdi) | |
411 | ||
412 | .Lenc_out: | |
413 | FRAME_END | |
414 | ret | |
415 | ENDPROC(crypto_morus1280_avx2_enc) | |
416 | ||
417 | /* | |
418 | * void crypto_morus1280_avx2_enc_tail(void *state, const void *src, void *dst, | |
419 | * unsigned int length); | |
420 | */ | |
421 | ENTRY(crypto_morus1280_avx2_enc_tail) | |
422 | FRAME_BEGIN | |
423 | ||
424 | /* load the state: */ | |
425 | vmovdqu (0 * 32)(%rdi), STATE0 | |
426 | vmovdqu (1 * 32)(%rdi), STATE1 | |
427 | vmovdqu (2 * 32)(%rdi), STATE2 | |
428 | vmovdqu (3 * 32)(%rdi), STATE3 | |
429 | vmovdqu (4 * 32)(%rdi), STATE4 | |
430 | ||
431 | /* encrypt message: */ | |
432 | call __load_partial | |
433 | ||
434 | vmovdqa MSG, T0 | |
435 | vpxor STATE0, T0, T0 | |
436 | vpermq $MASK3, STATE1, T1 | |
437 | vpxor T1, T0, T0 | |
438 | vpand STATE2, STATE3, T1 | |
439 | vpxor T1, T0, T0 | |
440 | ||
441 | call __store_partial | |
442 | ||
443 | call __morus1280_update | |
444 | ||
445 | /* store the state: */ | |
446 | vmovdqu STATE0, (0 * 32)(%rdi) | |
447 | vmovdqu STATE1, (1 * 32)(%rdi) | |
448 | vmovdqu STATE2, (2 * 32)(%rdi) | |
449 | vmovdqu STATE3, (3 * 32)(%rdi) | |
450 | vmovdqu STATE4, (4 * 32)(%rdi) | |
451 | ||
452 | FRAME_END | |
221e00d1 | 453 | ret |
6ecc9d9f OM |
454 | ENDPROC(crypto_morus1280_avx2_enc_tail) |
455 | ||
456 | /* | |
457 | * void crypto_morus1280_avx2_dec(void *state, const void *src, void *dst, | |
458 | * unsigned int length); | |
459 | */ | |
460 | ENTRY(crypto_morus1280_avx2_dec) | |
461 | FRAME_BEGIN | |
462 | ||
463 | cmp $32, %rcx | |
464 | jb .Ldec_out | |
465 | ||
466 | /* load the state: */ | |
467 | vmovdqu (0 * 32)(%rdi), STATE0 | |
468 | vmovdqu (1 * 32)(%rdi), STATE1 | |
469 | vmovdqu (2 * 32)(%rdi), STATE2 | |
470 | vmovdqu (3 * 32)(%rdi), STATE3 | |
471 | vmovdqu (4 * 32)(%rdi), STATE4 | |
472 | ||
473 | mov %rsi, %r8 | |
474 | or %rdx, %r8 | |
475 | and $0x1F, %r8 | |
476 | jnz .Ldec_u_loop | |
477 | ||
478 | .align 4 | |
479 | .Ldec_a_loop: | |
480 | vmovdqa (%rsi), MSG | |
481 | vpxor STATE0, MSG, MSG | |
482 | vpermq $MASK3, STATE1, T0 | |
483 | vpxor T0, MSG, MSG | |
484 | vpand STATE2, STATE3, T0 | |
485 | vpxor T0, MSG, MSG | |
486 | vmovdqa MSG, (%rdx) | |
487 | ||
488 | call __morus1280_update | |
489 | sub $32, %rcx | |
490 | add $32, %rsi | |
491 | add $32, %rdx | |
492 | cmp $32, %rcx | |
493 | jge .Ldec_a_loop | |
494 | ||
495 | jmp .Ldec_cont | |
496 | .align 4 | |
497 | .Ldec_u_loop: | |
498 | vmovdqu (%rsi), MSG | |
499 | vpxor STATE0, MSG, MSG | |
500 | vpermq $MASK3, STATE1, T0 | |
501 | vpxor T0, MSG, MSG | |
502 | vpand STATE2, STATE3, T0 | |
503 | vpxor T0, MSG, MSG | |
504 | vmovdqu MSG, (%rdx) | |
505 | ||
506 | call __morus1280_update | |
507 | sub $32, %rcx | |
508 | add $32, %rsi | |
509 | add $32, %rdx | |
510 | cmp $32, %rcx | |
511 | jge .Ldec_u_loop | |
512 | ||
513 | .Ldec_cont: | |
514 | /* store the state: */ | |
515 | vmovdqu STATE0, (0 * 32)(%rdi) | |
516 | vmovdqu STATE1, (1 * 32)(%rdi) | |
517 | vmovdqu STATE2, (2 * 32)(%rdi) | |
518 | vmovdqu STATE3, (3 * 32)(%rdi) | |
519 | vmovdqu STATE4, (4 * 32)(%rdi) | |
520 | ||
521 | .Ldec_out: | |
522 | FRAME_END | |
523 | ret | |
524 | ENDPROC(crypto_morus1280_avx2_dec) | |
525 | ||
526 | /* | |
527 | * void crypto_morus1280_avx2_dec_tail(void *state, const void *src, void *dst, | |
528 | * unsigned int length); | |
529 | */ | |
530 | ENTRY(crypto_morus1280_avx2_dec_tail) | |
531 | FRAME_BEGIN | |
532 | ||
533 | /* load the state: */ | |
534 | vmovdqu (0 * 32)(%rdi), STATE0 | |
535 | vmovdqu (1 * 32)(%rdi), STATE1 | |
536 | vmovdqu (2 * 32)(%rdi), STATE2 | |
537 | vmovdqu (3 * 32)(%rdi), STATE3 | |
538 | vmovdqu (4 * 32)(%rdi), STATE4 | |
539 | ||
540 | /* decrypt message: */ | |
541 | call __load_partial | |
542 | ||
543 | vpxor STATE0, MSG, MSG | |
544 | vpermq $MASK3, STATE1, T0 | |
545 | vpxor T0, MSG, MSG | |
546 | vpand STATE2, STATE3, T0 | |
547 | vpxor T0, MSG, MSG | |
548 | vmovdqa MSG, T0 | |
549 | ||
550 | call __store_partial | |
551 | ||
552 | /* mask with byte count: */ | |
553 | movq %rcx, T0_LOW | |
554 | vpbroadcastb T0_LOW, T0 | |
555 | vmovdqa .Lmorus1280_counter, T1 | |
556 | vpcmpgtb T1, T0, T0 | |
557 | vpand T0, MSG, MSG | |
558 | ||
559 | call __morus1280_update | |
560 | ||
561 | /* store the state: */ | |
562 | vmovdqu STATE0, (0 * 32)(%rdi) | |
563 | vmovdqu STATE1, (1 * 32)(%rdi) | |
564 | vmovdqu STATE2, (2 * 32)(%rdi) | |
565 | vmovdqu STATE3, (3 * 32)(%rdi) | |
566 | vmovdqu STATE4, (4 * 32)(%rdi) | |
567 | ||
568 | FRAME_END | |
569 | ret | |
570 | ENDPROC(crypto_morus1280_avx2_dec_tail) | |
571 | ||
572 | /* | |
573 | * void crypto_morus1280_avx2_final(void *state, void *tag_xor, | |
574 | * u64 assoclen, u64 cryptlen); | |
575 | */ | |
576 | ENTRY(crypto_morus1280_avx2_final) | |
577 | FRAME_BEGIN | |
578 | ||
579 | /* load the state: */ | |
580 | vmovdqu (0 * 32)(%rdi), STATE0 | |
581 | vmovdqu (1 * 32)(%rdi), STATE1 | |
582 | vmovdqu (2 * 32)(%rdi), STATE2 | |
583 | vmovdqu (3 * 32)(%rdi), STATE3 | |
584 | vmovdqu (4 * 32)(%rdi), STATE4 | |
585 | ||
586 | /* xor state[0] into state[4]: */ | |
587 | vpxor STATE0, STATE4, STATE4 | |
588 | ||
589 | /* prepare length block: */ | |
590 | vpxor MSG, MSG, MSG | |
591 | vpinsrq $0, %rdx, MSG_LOW, MSG_LOW | |
592 | vpinsrq $1, %rcx, MSG_LOW, MSG_LOW | |
593 | vpsllq $3, MSG, MSG /* multiply by 8 (to get bit count) */ | |
594 | ||
595 | /* update state: */ | |
596 | call __morus1280_update | |
597 | call __morus1280_update | |
598 | call __morus1280_update | |
599 | call __morus1280_update | |
600 | call __morus1280_update | |
601 | call __morus1280_update | |
602 | call __morus1280_update | |
603 | call __morus1280_update | |
604 | call __morus1280_update | |
605 | call __morus1280_update | |
606 | ||
607 | /* xor tag: */ | |
608 | vmovdqu (%rsi), MSG | |
609 | ||
610 | vpxor STATE0, MSG, MSG | |
611 | vpermq $MASK3, STATE1, T0 | |
612 | vpxor T0, MSG, MSG | |
613 | vpand STATE2, STATE3, T0 | |
614 | vpxor T0, MSG, MSG | |
615 | vmovdqu MSG, (%rsi) | |
616 | ||
617 | FRAME_END | |
618 | ret | |
619 | ENDPROC(crypto_morus1280_avx2_final) |