www.gitorious.org Git - gnutls:gnutls.git/blob - lib/accelerated/x86/elf/aesni-x86

3 #

4 # Redistribution and use in source and binary forms, with or without

5 # modification, are permitted provided that the following conditions

6 # are met:

7 #

8 # * Redistributions of source code must retain copyright notices,

9 # this list of conditions and the following disclaimer.

10 #

11 # * Redistributions in binary form must reproduce the above

12 # copyright notice, this list of conditions and the following

13 # disclaimer in the documentation and/or other materials

14 # provided with the distribution.

15 #

16 # * Neither the name of the Andy Polyakov nor the names of its

17 # copyright holder and contributors may be used to endorse or

18 # promote products derived from this software without specific

19 # prior written permission.

20 #

21 # ALTERNATIVELY, provided that this notice is retained in full, this

22 # product may be distributed under the terms of the GNU General Public

23 # License (GPL), in which case the provisions of the GPL apply INSTEAD OF

24 # those given above.

25 #

26 # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER AND CONTRIBUTORS

27 # "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT

28 # LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR

29 # A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT

30 # OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,

31 # SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT

32 # LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,

33 # DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY

34 # THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT

35 # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE

36 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

37 #

38 # *** This file is auto-generated ***

39 #

40 .text

41 .globl aesni_encrypt

42 .type aesni_encrypt,@function

43 .align 16

44 aesni_encrypt:

45 movups (%rdi),%xmm2

46 movl 240(%rdx),%eax

47 movups (%rdx),%xmm0

48 movups 16(%rdx),%xmm1

49 leaq 32(%rdx),%rdx

50 xorps %xmm0,%xmm2

51 .Loop_enc1_1:

52 .byte 102,15,56,220,209

53 decl %eax

54 movups (%rdx),%xmm1

55 leaq 16(%rdx),%rdx

56 jnz .Loop_enc1_1

57 .byte 102,15,56,221,209

58 movups %xmm2,(%rsi)

59 .byte 0xf3,0xc3

60 .size aesni_encrypt,.-aesni_encrypt

62 .globl aesni_decrypt

63 .type aesni_decrypt,@function

64 .align 16

65 aesni_decrypt:

66 movups (%rdi),%xmm2

67 movl 240(%rdx),%eax

68 movups (%rdx),%xmm0

69 movups 16(%rdx),%xmm1

70 leaq 32(%rdx),%rdx

71 xorps %xmm0,%xmm2

72 .Loop_dec1_2:

73 .byte 102,15,56,222,209

74 decl %eax

75 movups (%rdx),%xmm1

76 leaq 16(%rdx),%rdx

77 jnz .Loop_dec1_2

78 .byte 102,15,56,223,209

79 movups %xmm2,(%rsi)

80 .byte 0xf3,0xc3

81 .size aesni_decrypt, .-aesni_decrypt

82 .type _aesni_encrypt3,@function

83 .align 16

84 _aesni_encrypt3:

85 movups (%rcx),%xmm0

86 shrl $1,%eax

87 movups 16(%rcx),%xmm1

88 leaq 32(%rcx),%rcx

89 xorps %xmm0,%xmm2

90 xorps %xmm0,%xmm3

91 xorps %xmm0,%xmm4

92 movups (%rcx),%xmm0

94 .Lenc_loop3:

95 .byte 102,15,56,220,209

96 .byte 102,15,56,220,217

97 decl %eax

98 .byte 102,15,56,220,225

99 movups 16(%rcx),%xmm1

100 .byte 102,15,56,220,208

101 .byte 102,15,56,220,216

102 leaq 32(%rcx),%rcx

103 .byte 102,15,56,220,224

104 movups (%rcx),%xmm0

105 jnz .Lenc_loop3

106

107 .byte 102,15,56,220,209

108 .byte 102,15,56,220,217

109 .byte 102,15,56,220,225

110 .byte 102,15,56,221,208

111 .byte 102,15,56,221,216

112 .byte 102,15,56,221,224

113 .byte 0xf3,0xc3

114 .size _aesni_encrypt3,.-_aesni_encrypt3

115 .type _aesni_decrypt3,@function

116 .align 16

117 _aesni_decrypt3:

118 movups (%rcx),%xmm0

119 shrl $1,%eax

120 movups 16(%rcx),%xmm1

121 leaq 32(%rcx),%rcx

122 xorps %xmm0,%xmm2

123 xorps %xmm0,%xmm3

124 xorps %xmm0,%xmm4

125 movups (%rcx),%xmm0

126

127 .Ldec_loop3:

128 .byte 102,15,56,222,209

129 .byte 102,15,56,222,217

130 decl %eax

131 .byte 102,15,56,222,225

132 movups 16(%rcx),%xmm1

133 .byte 102,15,56,222,208

134 .byte 102,15,56,222,216

135 leaq 32(%rcx),%rcx

136 .byte 102,15,56,222,224

137 movups (%rcx),%xmm0

138 jnz .Ldec_loop3

139

140 .byte 102,15,56,222,209

141 .byte 102,15,56,222,217

142 .byte 102,15,56,222,225

143 .byte 102,15,56,223,208

144 .byte 102,15,56,223,216

145 .byte 102,15,56,223,224

146 .byte 0xf3,0xc3

147 .size _aesni_decrypt3,.-_aesni_decrypt3

148 .type _aesni_encrypt4,@function

149 .align 16

150 _aesni_encrypt4:

151 movups (%rcx),%xmm0

152 shrl $1,%eax

153 movups 16(%rcx),%xmm1

154 leaq 32(%rcx),%rcx

155 xorps %xmm0,%xmm2

156 xorps %xmm0,%xmm3

157 xorps %xmm0,%xmm4

158 xorps %xmm0,%xmm5

159 movups (%rcx),%xmm0

160

161 .Lenc_loop4:

162 .byte 102,15,56,220,209

163 .byte 102,15,56,220,217

164 decl %eax

165 .byte 102,15,56,220,225

166 .byte 102,15,56,220,233

167 movups 16(%rcx),%xmm1

168 .byte 102,15,56,220,208

169 .byte 102,15,56,220,216

170 leaq 32(%rcx),%rcx

171 .byte 102,15,56,220,224

172 .byte 102,15,56,220,232

173 movups (%rcx),%xmm0

174 jnz .Lenc_loop4

175

176 .byte 102,15,56,220,209

177 .byte 102,15,56,220,217

178 .byte 102,15,56,220,225

179 .byte 102,15,56,220,233

180 .byte 102,15,56,221,208

181 .byte 102,15,56,221,216

182 .byte 102,15,56,221,224

183 .byte 102,15,56,221,232

184 .byte 0xf3,0xc3

185 .size _aesni_encrypt4,.-_aesni_encrypt4

186 .type _aesni_decrypt4,@function

187 .align 16

188 _aesni_decrypt4:

189 movups (%rcx),%xmm0

190 shrl $1,%eax

191 movups 16(%rcx),%xmm1

192 leaq 32(%rcx),%rcx

193 xorps %xmm0,%xmm2

194 xorps %xmm0,%xmm3

195 xorps %xmm0,%xmm4

196 xorps %xmm0,%xmm5

197 movups (%rcx),%xmm0

198

199 .Ldec_loop4:

200 .byte 102,15,56,222,209

201 .byte 102,15,56,222,217

202 decl %eax

203 .byte 102,15,56,222,225

204 .byte 102,15,56,222,233

205 movups 16(%rcx),%xmm1

206 .byte 102,15,56,222,208

207 .byte 102,15,56,222,216

208 leaq 32(%rcx),%rcx

209 .byte 102,15,56,222,224

210 .byte 102,15,56,222,232

211 movups (%rcx),%xmm0

212 jnz .Ldec_loop4

213

214 .byte 102,15,56,222,209

215 .byte 102,15,56,222,217

216 .byte 102,15,56,222,225

217 .byte 102,15,56,222,233

218 .byte 102,15,56,223,208

219 .byte 102,15,56,223,216

220 .byte 102,15,56,223,224

221 .byte 102,15,56,223,232

222 .byte 0xf3,0xc3

223 .size _aesni_decrypt4,.-_aesni_decrypt4

224 .type _aesni_encrypt6,@function

225 .align 16

226 _aesni_encrypt6:

227 movups (%rcx),%xmm0

228 shrl $1,%eax

229 movups 16(%rcx),%xmm1

230 leaq 32(%rcx),%rcx

231 xorps %xmm0,%xmm2

232 pxor %xmm0,%xmm3

233 .byte 102,15,56,220,209

234 pxor %xmm0,%xmm4

235 .byte 102,15,56,220,217

236 pxor %xmm0,%xmm5

237 .byte 102,15,56,220,225

238 pxor %xmm0,%xmm6

239 .byte 102,15,56,220,233

240 pxor %xmm0,%xmm7

241 decl %eax

242 .byte 102,15,56,220,241

243 movups (%rcx),%xmm0

244 .byte 102,15,56,220,249

245 jmp .Lenc_loop6_enter

246 .align 16

247 .Lenc_loop6:

248 .byte 102,15,56,220,209

249 .byte 102,15,56,220,217

250 decl %eax

251 .byte 102,15,56,220,225

252 .byte 102,15,56,220,233

253 .byte 102,15,56,220,241

254 .byte 102,15,56,220,249

255 .Lenc_loop6_enter:

256 movups 16(%rcx),%xmm1

257 .byte 102,15,56,220,208

258 .byte 102,15,56,220,216

259 leaq 32(%rcx),%rcx

260 .byte 102,15,56,220,224

261 .byte 102,15,56,220,232

262 .byte 102,15,56,220,240

263 .byte 102,15,56,220,248

264 movups (%rcx),%xmm0

265 jnz .Lenc_loop6

266

267 .byte 102,15,56,220,209

268 .byte 102,15,56,220,217

269 .byte 102,15,56,220,225

270 .byte 102,15,56,220,233

271 .byte 102,15,56,220,241

272 .byte 102,15,56,220,249

273 .byte 102,15,56,221,208

274 .byte 102,15,56,221,216

275 .byte 102,15,56,221,224

276 .byte 102,15,56,221,232

277 .byte 102,15,56,221,240

278 .byte 102,15,56,221,248

279 .byte 0xf3,0xc3

280 .size _aesni_encrypt6,.-_aesni_encrypt6

281 .type _aesni_decrypt6,@function

282 .align 16

283 _aesni_decrypt6:

284 movups (%rcx),%xmm0

285 shrl $1,%eax

286 movups 16(%rcx),%xmm1

287 leaq 32(%rcx),%rcx

288 xorps %xmm0,%xmm2

289 pxor %xmm0,%xmm3

290 .byte 102,15,56,222,209

291 pxor %xmm0,%xmm4

292 .byte 102,15,56,222,217

293 pxor %xmm0,%xmm5

294 .byte 102,15,56,222,225

295 pxor %xmm0,%xmm6

296 .byte 102,15,56,222,233

297 pxor %xmm0,%xmm7

298 decl %eax

299 .byte 102,15,56,222,241

300 movups (%rcx),%xmm0

301 .byte 102,15,56,222,249

302 jmp .Ldec_loop6_enter

303 .align 16

304 .Ldec_loop6:

305 .byte 102,15,56,222,209

306 .byte 102,15,56,222,217

307 decl %eax

308 .byte 102,15,56,222,225

309 .byte 102,15,56,222,233

310 .byte 102,15,56,222,241

311 .byte 102,15,56,222,249

312 .Ldec_loop6_enter:

313 movups 16(%rcx),%xmm1

314 .byte 102,15,56,222,208

315 .byte 102,15,56,222,216

316 leaq 32(%rcx),%rcx

317 .byte 102,15,56,222,224

318 .byte 102,15,56,222,232

319 .byte 102,15,56,222,240

320 .byte 102,15,56,222,248

321 movups (%rcx),%xmm0

322 jnz .Ldec_loop6

323

324 .byte 102,15,56,222,209

325 .byte 102,15,56,222,217

326 .byte 102,15,56,222,225

327 .byte 102,15,56,222,233

328 .byte 102,15,56,222,241

329 .byte 102,15,56,222,249

330 .byte 102,15,56,223,208

331 .byte 102,15,56,223,216

332 .byte 102,15,56,223,224

333 .byte 102,15,56,223,232

334 .byte 102,15,56,223,240

335 .byte 102,15,56,223,248

336 .byte 0xf3,0xc3

337 .size _aesni_decrypt6,.-_aesni_decrypt6

338 .type _aesni_encrypt8,@function

339 .align 16

340 _aesni_encrypt8:

341 movups (%rcx),%xmm0

342 shrl $1,%eax

343 movups 16(%rcx),%xmm1

344 leaq 32(%rcx),%rcx

345 xorps %xmm0,%xmm2

346 xorps %xmm0,%xmm3

347 .byte 102,15,56,220,209

348 pxor %xmm0,%xmm4

349 .byte 102,15,56,220,217

350 pxor %xmm0,%xmm5

351 .byte 102,15,56,220,225

352 pxor %xmm0,%xmm6

353 .byte 102,15,56,220,233

354 pxor %xmm0,%xmm7

355 decl %eax

356 .byte 102,15,56,220,241

357 pxor %xmm0,%xmm8

358 .byte 102,15,56,220,249

359 pxor %xmm0,%xmm9

360 movups (%rcx),%xmm0

361 .byte 102,68,15,56,220,193

362 .byte 102,68,15,56,220,201

363 movups 16(%rcx),%xmm1

364 jmp .Lenc_loop8_enter

365 .align 16

366 .Lenc_loop8:

367 .byte 102,15,56,220,209

368 .byte 102,15,56,220,217

369 decl %eax

370 .byte 102,15,56,220,225

371 .byte 102,15,56,220,233

372 .byte 102,15,56,220,241

373 .byte 102,15,56,220,249

374 .byte 102,68,15,56,220,193

375 .byte 102,68,15,56,220,201

376 movups 16(%rcx),%xmm1

377 .Lenc_loop8_enter:

378 .byte 102,15,56,220,208

379 .byte 102,15,56,220,216

380 leaq 32(%rcx),%rcx

381 .byte 102,15,56,220,224

382 .byte 102,15,56,220,232

383 .byte 102,15,56,220,240

384 .byte 102,15,56,220,248

385 .byte 102,68,15,56,220,192

386 .byte 102,68,15,56,220,200

387 movups (%rcx),%xmm0

388 jnz .Lenc_loop8

389

390 .byte 102,15,56,220,209

391 .byte 102,15,56,220,217

392 .byte 102,15,56,220,225

393 .byte 102,15,56,220,233

394 .byte 102,15,56,220,241

395 .byte 102,15,56,220,249

396 .byte 102,68,15,56,220,193

397 .byte 102,68,15,56,220,201

398 .byte 102,15,56,221,208

399 .byte 102,15,56,221,216

400 .byte 102,15,56,221,224

401 .byte 102,15,56,221,232

402 .byte 102,15,56,221,240

403 .byte 102,15,56,221,248

404 .byte 102,68,15,56,221,192

405 .byte 102,68,15,56,221,200

406 .byte 0xf3,0xc3

407 .size _aesni_encrypt8,.-_aesni_encrypt8

408 .type _aesni_decrypt8,@function

409 .align 16

410 _aesni_decrypt8:

411 movups (%rcx),%xmm0

412 shrl $1,%eax

413 movups 16(%rcx),%xmm1

414 leaq 32(%rcx),%rcx

415 xorps %xmm0,%xmm2

416 xorps %xmm0,%xmm3

417 .byte 102,15,56,222,209

418 pxor %xmm0,%xmm4

419 .byte 102,15,56,222,217

420 pxor %xmm0,%xmm5

421 .byte 102,15,56,222,225

422 pxor %xmm0,%xmm6

423 .byte 102,15,56,222,233

424 pxor %xmm0,%xmm7

425 decl %eax

426 .byte 102,15,56,222,241

427 pxor %xmm0,%xmm8

428 .byte 102,15,56,222,249

429 pxor %xmm0,%xmm9

430 movups (%rcx),%xmm0

431 .byte 102,68,15,56,222,193

432 .byte 102,68,15,56,222,201

433 movups 16(%rcx),%xmm1

434 jmp .Ldec_loop8_enter

435 .align 16

436 .Ldec_loop8:

437 .byte 102,15,56,222,209

438 .byte 102,15,56,222,217

439 decl %eax

440 .byte 102,15,56,222,225

441 .byte 102,15,56,222,233

442 .byte 102,15,56,222,241

443 .byte 102,15,56,222,249

444 .byte 102,68,15,56,222,193

445 .byte 102,68,15,56,222,201

446 movups 16(%rcx),%xmm1

447 .Ldec_loop8_enter:

448 .byte 102,15,56,222,208

449 .byte 102,15,56,222,216

450 leaq 32(%rcx),%rcx

451 .byte 102,15,56,222,224

452 .byte 102,15,56,222,232

453 .byte 102,15,56,222,240

454 .byte 102,15,56,222,248

455 .byte 102,68,15,56,222,192

456 .byte 102,68,15,56,222,200

457 movups (%rcx),%xmm0

458 jnz .Ldec_loop8

459

460 .byte 102,15,56,222,209

461 .byte 102,15,56,222,217

462 .byte 102,15,56,222,225

463 .byte 102,15,56,222,233

464 .byte 102,15,56,222,241

465 .byte 102,15,56,222,249

466 .byte 102,68,15,56,222,193

467 .byte 102,68,15,56,222,201

468 .byte 102,15,56,223,208

469 .byte 102,15,56,223,216

470 .byte 102,15,56,223,224

471 .byte 102,15,56,223,232

472 .byte 102,15,56,223,240

473 .byte 102,15,56,223,248

474 .byte 102,68,15,56,223,192

475 .byte 102,68,15,56,223,200

476 .byte 0xf3,0xc3

477 .size _aesni_decrypt8,.-_aesni_decrypt8

478 .globl aesni_ecb_encrypt

479 .type aesni_ecb_encrypt,@function

480 .align 16

481 aesni_ecb_encrypt:

482 andq $-16,%rdx

483 jz .Lecb_ret

484

485 movl 240(%rcx),%eax

486 movups (%rcx),%xmm0

487 movq %rcx,%r11

488 movl %eax,%r10d

489 testl %r8d,%r8d

490 jz .Lecb_decrypt

491

492 cmpq $128,%rdx

493 jb .Lecb_enc_tail

494

495 movdqu (%rdi),%xmm2

496 movdqu 16(%rdi),%xmm3

497 movdqu 32(%rdi),%xmm4

498 movdqu 48(%rdi),%xmm5

499 movdqu 64(%rdi),%xmm6

500 movdqu 80(%rdi),%xmm7

501 movdqu 96(%rdi),%xmm8

502 movdqu 112(%rdi),%xmm9

503 leaq 128(%rdi),%rdi

504 subq $128,%rdx

505 jmp .Lecb_enc_loop8_enter

506 .align 16

507 .Lecb_enc_loop8:

508 movups %xmm2,(%rsi)

509 movq %r11,%rcx

510 movdqu (%rdi),%xmm2

511 movl %r10d,%eax

512 movups %xmm3,16(%rsi)

513 movdqu 16(%rdi),%xmm3

514 movups %xmm4,32(%rsi)

515 movdqu 32(%rdi),%xmm4

516 movups %xmm5,48(%rsi)

517 movdqu 48(%rdi),%xmm5

518 movups %xmm6,64(%rsi)

519 movdqu 64(%rdi),%xmm6

520 movups %xmm7,80(%rsi)

521 movdqu 80(%rdi),%xmm7

522 movups %xmm8,96(%rsi)

523 movdqu 96(%rdi),%xmm8

524 movups %xmm9,112(%rsi)

525 leaq 128(%rsi),%rsi

526 movdqu 112(%rdi),%xmm9

527 leaq 128(%rdi),%rdi

528 .Lecb_enc_loop8_enter:

529

530 call _aesni_encrypt8

531

532 subq $128,%rdx

533 jnc .Lecb_enc_loop8

534

535 movups %xmm2,(%rsi)

536 movq %r11,%rcx

537 movups %xmm3,16(%rsi)

538 movl %r10d,%eax

539 movups %xmm4,32(%rsi)

540 movups %xmm5,48(%rsi)

541 movups %xmm6,64(%rsi)

542 movups %xmm7,80(%rsi)

543 movups %xmm8,96(%rsi)

544 movups %xmm9,112(%rsi)

545 leaq 128(%rsi),%rsi

546 addq $128,%rdx

547 jz .Lecb_ret

548

549 .Lecb_enc_tail:

550 movups (%rdi),%xmm2

551 cmpq $32,%rdx

552 jb .Lecb_enc_one

553 movups 16(%rdi),%xmm3

554 je .Lecb_enc_two

555 movups 32(%rdi),%xmm4

556 cmpq $64,%rdx

557 jb .Lecb_enc_three

558 movups 48(%rdi),%xmm5

559 je .Lecb_enc_four

560 movups 64(%rdi),%xmm6

561 cmpq $96,%rdx

562 jb .Lecb_enc_five

563 movups 80(%rdi),%xmm7

564 je .Lecb_enc_six

565 movdqu 96(%rdi),%xmm8

566 call _aesni_encrypt8

567 movups %xmm2,(%rsi)

568 movups %xmm3,16(%rsi)

569 movups %xmm4,32(%rsi)

570 movups %xmm5,48(%rsi)

571 movups %xmm6,64(%rsi)

572 movups %xmm7,80(%rsi)

573 movups %xmm8,96(%rsi)

574 jmp .Lecb_ret

575 .align 16

576 .Lecb_enc_one:

577 movups (%rcx),%xmm0

578 movups 16(%rcx),%xmm1

579 leaq 32(%rcx),%rcx

580 xorps %xmm0,%xmm2

581 .Loop_enc1_3:

582 .byte 102,15,56,220,209

583 decl %eax

584 movups (%rcx),%xmm1

585 leaq 16(%rcx),%rcx

586 jnz .Loop_enc1_3

587 .byte 102,15,56,221,209

588 movups %xmm2,(%rsi)

589 jmp .Lecb_ret

590 .align 16

591 .Lecb_enc_two:

592 xorps %xmm4,%xmm4

593 call _aesni_encrypt3

594 movups %xmm2,(%rsi)

595 movups %xmm3,16(%rsi)

596 jmp .Lecb_ret

597 .align 16

598 .Lecb_enc_three:

599 call _aesni_encrypt3

600 movups %xmm2,(%rsi)

601 movups %xmm3,16(%rsi)

602 movups %xmm4,32(%rsi)

603 jmp .Lecb_ret

604 .align 16

605 .Lecb_enc_four:

606 call _aesni_encrypt4

607 movups %xmm2,(%rsi)

608 movups %xmm3,16(%rsi)

609 movups %xmm4,32(%rsi)

610 movups %xmm5,48(%rsi)

611 jmp .Lecb_ret

612 .align 16

613 .Lecb_enc_five:

614 xorps %xmm7,%xmm7

615 call _aesni_encrypt6

616 movups %xmm2,(%rsi)

617 movups %xmm3,16(%rsi)

618 movups %xmm4,32(%rsi)

619 movups %xmm5,48(%rsi)

620 movups %xmm6,64(%rsi)

621 jmp .Lecb_ret

622 .align 16

623 .Lecb_enc_six:

624 call _aesni_encrypt6

625 movups %xmm2,(%rsi)

626 movups %xmm3,16(%rsi)

627 movups %xmm4,32(%rsi)

628 movups %xmm5,48(%rsi)

629 movups %xmm6,64(%rsi)

630 movups %xmm7,80(%rsi)

631 jmp .Lecb_ret

632

633 .align 16

634 .Lecb_decrypt:

635 cmpq $128,%rdx

636 jb .Lecb_dec_tail

637

638 movdqu (%rdi),%xmm2

639 movdqu 16(%rdi),%xmm3

640 movdqu 32(%rdi),%xmm4

641 movdqu 48(%rdi),%xmm5

642 movdqu 64(%rdi),%xmm6

643 movdqu 80(%rdi),%xmm7

644 movdqu 96(%rdi),%xmm8

645 movdqu 112(%rdi),%xmm9

646 leaq 128(%rdi),%rdi

647 subq $128,%rdx

648 jmp .Lecb_dec_loop8_enter

649 .align 16

650 .Lecb_dec_loop8:

651 movups %xmm2,(%rsi)

652 movq %r11,%rcx

653 movdqu (%rdi),%xmm2

654 movl %r10d,%eax

655 movups %xmm3,16(%rsi)

656 movdqu 16(%rdi),%xmm3

657 movups %xmm4,32(%rsi)

658 movdqu 32(%rdi),%xmm4

659 movups %xmm5,48(%rsi)

660 movdqu 48(%rdi),%xmm5

661 movups %xmm6,64(%rsi)

662 movdqu 64(%rdi),%xmm6

663 movups %xmm7,80(%rsi)

664 movdqu 80(%rdi),%xmm7

665 movups %xmm8,96(%rsi)

666 movdqu 96(%rdi),%xmm8

667 movups %xmm9,112(%rsi)

668 leaq 128(%rsi),%rsi

669 movdqu 112(%rdi),%xmm9

670 leaq 128(%rdi),%rdi

671 .Lecb_dec_loop8_enter:

672

673 call _aesni_decrypt8

674

675 movups (%r11),%xmm0

676 subq $128,%rdx

677 jnc .Lecb_dec_loop8

678

679 movups %xmm2,(%rsi)

680 movq %r11,%rcx

681 movups %xmm3,16(%rsi)

682 movl %r10d,%eax

683 movups %xmm4,32(%rsi)

684 movups %xmm5,48(%rsi)

685 movups %xmm6,64(%rsi)

686 movups %xmm7,80(%rsi)

687 movups %xmm8,96(%rsi)

688 movups %xmm9,112(%rsi)

689 leaq 128(%rsi),%rsi

690 addq $128,%rdx

691 jz .Lecb_ret

692

693 .Lecb_dec_tail:

694 movups (%rdi),%xmm2

695 cmpq $32,%rdx

696 jb .Lecb_dec_one

697 movups 16(%rdi),%xmm3

698 je .Lecb_dec_two

699 movups 32(%rdi),%xmm4

700 cmpq $64,%rdx

701 jb .Lecb_dec_three

702 movups 48(%rdi),%xmm5

703 je .Lecb_dec_four

704 movups 64(%rdi),%xmm6

705 cmpq $96,%rdx

706 jb .Lecb_dec_five

707 movups 80(%rdi),%xmm7

708 je .Lecb_dec_six

709 movups 96(%rdi),%xmm8

710 movups (%rcx),%xmm0

711 call _aesni_decrypt8

712 movups %xmm2,(%rsi)

713 movups %xmm3,16(%rsi)

714 movups %xmm4,32(%rsi)

715 movups %xmm5,48(%rsi)

716 movups %xmm6,64(%rsi)

717 movups %xmm7,80(%rsi)

718 movups %xmm8,96(%rsi)

719 jmp .Lecb_ret

720 .align 16

721 .Lecb_dec_one:

722 movups (%rcx),%xmm0

723 movups 16(%rcx),%xmm1

724 leaq 32(%rcx),%rcx

725 xorps %xmm0,%xmm2

726 .Loop_dec1_4:

727 .byte 102,15,56,222,209

728 decl %eax

729 movups (%rcx),%xmm1

730 leaq 16(%rcx),%rcx

731 jnz .Loop_dec1_4

732 .byte 102,15,56,223,209

733 movups %xmm2,(%rsi)

734 jmp .Lecb_ret

735 .align 16

736 .Lecb_dec_two:

737 xorps %xmm4,%xmm4

738 call _aesni_decrypt3

739 movups %xmm2,(%rsi)

740 movups %xmm3,16(%rsi)

741 jmp .Lecb_ret

742 .align 16

743 .Lecb_dec_three:

744 call _aesni_decrypt3

745 movups %xmm2,(%rsi)

746 movups %xmm3,16(%rsi)

747 movups %xmm4,32(%rsi)

748 jmp .Lecb_ret

749 .align 16

750 .Lecb_dec_four:

751 call _aesni_decrypt4

752 movups %xmm2,(%rsi)

753 movups %xmm3,16(%rsi)

754 movups %xmm4,32(%rsi)

755 movups %xmm5,48(%rsi)

756 jmp .Lecb_ret

757 .align 16

758 .Lecb_dec_five:

759 xorps %xmm7,%xmm7

760 call _aesni_decrypt6

761 movups %xmm2,(%rsi)

762 movups %xmm3,16(%rsi)

763 movups %xmm4,32(%rsi)

764 movups %xmm5,48(%rsi)

765 movups %xmm6,64(%rsi)

766 jmp .Lecb_ret

767 .align 16

768 .Lecb_dec_six:

769 call _aesni_decrypt6

770 movups %xmm2,(%rsi)

771 movups %xmm3,16(%rsi)

772 movups %xmm4,32(%rsi)

773 movups %xmm5,48(%rsi)

774 movups %xmm6,64(%rsi)

775 movups %xmm7,80(%rsi)

776

777 .Lecb_ret:

778 .byte 0xf3,0xc3

779 .size aesni_ecb_encrypt,.-aesni_ecb_encrypt

780 .globl aesni_ccm64_encrypt_blocks

781 .type aesni_ccm64_encrypt_blocks,@function

782 .align 16

783 aesni_ccm64_encrypt_blocks:

784 movl 240(%rcx),%eax

785 movdqu (%r8),%xmm9

786 movdqa .Lincrement64(%rip),%xmm6

787 movdqa .Lbswap_mask(%rip),%xmm7

788

789 shrl $1,%eax

790 leaq 0(%rcx),%r11

791 movdqu (%r9),%xmm3

792 movdqa %xmm9,%xmm2

793 movl %eax,%r10d

794 .byte 102,68,15,56,0,207

795 jmp .Lccm64_enc_outer

796 .align 16

797 .Lccm64_enc_outer:

798 movups (%r11),%xmm0

799 movl %r10d,%eax

800 movups (%rdi),%xmm8

801

802 xorps %xmm0,%xmm2

803 movups 16(%r11),%xmm1

804 xorps %xmm8,%xmm0

805 leaq 32(%r11),%rcx

806 xorps %xmm0,%xmm3

807 movups (%rcx),%xmm0

808

809 .Lccm64_enc2_loop:

810 .byte 102,15,56,220,209

811 decl %eax

812 .byte 102,15,56,220,217

813 movups 16(%rcx),%xmm1

814 .byte 102,15,56,220,208

815 leaq 32(%rcx),%rcx

816 .byte 102,15,56,220,216

817 movups 0(%rcx),%xmm0

818 jnz .Lccm64_enc2_loop

819 .byte 102,15,56,220,209

820 .byte 102,15,56,220,217

821 paddq %xmm6,%xmm9

822 .byte 102,15,56,221,208

823 .byte 102,15,56,221,216

824

825 decq %rdx

826 leaq 16(%rdi),%rdi

827 xorps %xmm2,%xmm8

828 movdqa %xmm9,%xmm2

829 movups %xmm8,(%rsi)

830 leaq 16(%rsi),%rsi

831 .byte 102,15,56,0,215

832 jnz .Lccm64_enc_outer

833

834 movups %xmm3,(%r9)

835 .byte 0xf3,0xc3

836 .size aesni_ccm64_encrypt_blocks,.-aesni_ccm64_encrypt_blocks

837 .globl aesni_ccm64_decrypt_blocks

838 .type aesni_ccm64_decrypt_blocks,@function

839 .align 16

840 aesni_ccm64_decrypt_blocks:

841 movl 240(%rcx),%eax

842 movups (%r8),%xmm9

843 movdqu (%r9),%xmm3

844 movdqa .Lincrement64(%rip),%xmm6

845 movdqa .Lbswap_mask(%rip),%xmm7

846

847 movaps %xmm9,%xmm2

848 movl %eax,%r10d

849 movq %rcx,%r11

850 .byte 102,68,15,56,0,207

851 movups (%rcx),%xmm0

852 movups 16(%rcx),%xmm1

853 leaq 32(%rcx),%rcx

854 xorps %xmm0,%xmm2

855 .Loop_enc1_5:

856 .byte 102,15,56,220,209

857 decl %eax

858 movups (%rcx),%xmm1

859 leaq 16(%rcx),%rcx

860 jnz .Loop_enc1_5

861 .byte 102,15,56,221,209

862 movups (%rdi),%xmm8

863 paddq %xmm6,%xmm9

864 leaq 16(%rdi),%rdi

865 jmp .Lccm64_dec_outer

866 .align 16

867 .Lccm64_dec_outer:

868 xorps %xmm2,%xmm8

869 movdqa %xmm9,%xmm2

870 movl %r10d,%eax

871 movups %xmm8,(%rsi)

872 leaq 16(%rsi),%rsi

873 .byte 102,15,56,0,215

874

875 subq $1,%rdx

876 jz .Lccm64_dec_break

877

878 movups (%r11),%xmm0

879 shrl $1,%eax

880 movups 16(%r11),%xmm1

881 xorps %xmm0,%xmm8

882 leaq 32(%r11),%rcx

883 xorps %xmm0,%xmm2

884 xorps %xmm8,%xmm3

885 movups (%rcx),%xmm0

886

887 .Lccm64_dec2_loop:

888 .byte 102,15,56,220,209

889 decl %eax

890 .byte 102,15,56,220,217

891 movups 16(%rcx),%xmm1

892 .byte 102,15,56,220,208

893 leaq 32(%rcx),%rcx

894 .byte 102,15,56,220,216

895 movups 0(%rcx),%xmm0

896 jnz .Lccm64_dec2_loop

897 movups (%rdi),%xmm8

898 paddq %xmm6,%xmm9

899 .byte 102,15,56,220,209

900 .byte 102,15,56,220,217

901 leaq 16(%rdi),%rdi

902 .byte 102,15,56,221,208

903 .byte 102,15,56,221,216

904 jmp .Lccm64_dec_outer

905

906 .align 16

907 .Lccm64_dec_break:

908

909 movups (%r11),%xmm0

910 movups 16(%r11),%xmm1

911 xorps %xmm0,%xmm8

912 leaq 32(%r11),%r11

913 xorps %xmm8,%xmm3

914 .Loop_enc1_6:

915 .byte 102,15,56,220,217

916 decl %eax

917 movups (%r11),%xmm1

918 leaq 16(%r11),%r11

919 jnz .Loop_enc1_6

920 .byte 102,15,56,221,217

921 movups %xmm3,(%r9)

922 .byte 0xf3,0xc3

923 .size aesni_ccm64_decrypt_blocks,.-aesni_ccm64_decrypt_blocks

924 .globl aesni_ctr32_encrypt_blocks

925 .type aesni_ctr32_encrypt_blocks,@function

926 .align 16

927 aesni_ctr32_encrypt_blocks:

928 leaq (%rsp),%rax

929 pushq %rbp

930 subq $128,%rsp

931 andq $-16,%rsp

932 leaq -8(%rax),%rbp

933

934 cmpq $1,%rdx

935 je .Lctr32_one_shortcut

936

937 movdqu (%r8),%xmm2

938 movdqu (%rcx),%xmm0

939 movl 12(%r8),%r8d

940 pxor %xmm0,%xmm2

941 movl 12(%rcx),%r11d

942 movdqa %xmm2,0(%rsp)

943 bswapl %r8d

944 movdqa %xmm2,%xmm3

945 movdqa %xmm2,%xmm4

946 movdqa %xmm2,%xmm5

947 movdqa %xmm2,64(%rsp)

948 movdqa %xmm2,80(%rsp)

949 movdqa %xmm2,96(%rsp)

950 movdqa %xmm2,112(%rsp)

951

952 movl 240(%rcx),%eax

953

954 leaq 1(%r8),%r9

955 leaq 2(%r8),%r10

956 bswapl %r9d

957 bswapl %r10d

958 xorl %r11d,%r9d

959 xorl %r11d,%r10d

960 .byte 102,65,15,58,34,217,3

961 leaq 3(%r8),%r9

962 movdqa %xmm3,16(%rsp)

963 .byte 102,65,15,58,34,226,3

964 bswapl %r9d

965 leaq 4(%r8),%r10

966 movdqa %xmm4,32(%rsp)

967 xorl %r11d,%r9d

968 bswapl %r10d

969 .byte 102,65,15,58,34,233,3

970 xorl %r11d,%r10d

971 movdqa %xmm5,48(%rsp)

972 leaq 5(%r8),%r9

973 movl %r10d,64+12(%rsp)

974 bswapl %r9d

975 leaq 6(%r8),%r10

976 xorl %r11d,%r9d

977 bswapl %r10d

978 movl %r9d,80+12(%rsp)

979 xorl %r11d,%r10d

980 leaq 7(%r8),%r9

981 movl %r10d,96+12(%rsp)

982 bswapl %r9d

983 xorl %r11d,%r9d

984 movl %r9d,112+12(%rsp)

985

986 movups 16(%rcx),%xmm1

987

988 movdqa 64(%rsp),%xmm6

989 movdqa 80(%rsp),%xmm7

990

991 cmpq $8,%rdx

992 jb .Lctr32_tail

993

994 leaq 128(%rcx),%rcx

995 subq $8,%rdx

996 jmp .Lctr32_loop8

997

998 .align 32

999 .Lctr32_loop8:

1000 addl $8,%r8d

1001 movdqa 96(%rsp),%xmm8

1002 .byte 102,15,56,220,209

1003 movl %r8d,%r9d

1004 movdqa 112(%rsp),%xmm9

1005 .byte 102,15,56,220,217

1006 bswapl %r9d

1007 movups 32-128(%rcx),%xmm0

1008 .byte 102,15,56,220,225

1009 xorl %r11d,%r9d

1010 .byte 102,15,56,220,233

1011 movl %r9d,0+12(%rsp)

1012 leaq 1(%r8),%r9

1013 .byte 102,15,56,220,241

1014 .byte 102,15,56,220,249

1015 .byte 102,68,15,56,220,193

1016 .byte 102,68,15,56,220,201

1017 movups 48-128(%rcx),%xmm1

1018 .byte 102,15,56,220,208

1019 .byte 102,15,56,220,216

1020 bswapl %r9d

1021 .byte 102,15,56,220,224

1022 xorl %r11d,%r9d

1023 .byte 102,15,56,220,232

1024 movl %r9d,16+12(%rsp)

1025 leaq 2(%r8),%r9

1026 .byte 102,15,56,220,240

1027 .byte 102,15,56,220,248

1028 .byte 102,68,15,56,220,192

1029 .byte 102,68,15,56,220,200

1030 movups 64-128(%rcx),%xmm0

1031 .byte 102,15,56,220,209

1032 .byte 102,15,56,220,217

1033 bswapl %r9d

1034 .byte 102,15,56,220,225

1035 xorl %r11d,%r9d

1036 .byte 102,15,56,220,233

1037 movl %r9d,32+12(%rsp)

1038 leaq 3(%r8),%r9

1039 .byte 102,15,56,220,241

1040 .byte 102,15,56,220,249

1041 .byte 102,68,15,56,220,193

1042 .byte 102,68,15,56,220,201

1043 movups 80-128(%rcx),%xmm1

1044 .byte 102,15,56,220,208

1045 .byte 102,15,56,220,216

1046 bswapl %r9d

1047 .byte 102,15,56,220,224

1048 xorl %r11d,%r9d

1049 .byte 102,15,56,220,232

1050 movl %r9d,48+12(%rsp)

1051 leaq 4(%r8),%r9

1052 .byte 102,15,56,220,240

1053 .byte 102,15,56,220,248

1054 .byte 102,68,15,56,220,192

1055 .byte 102,68,15,56,220,200

1056 movups 96-128(%rcx),%xmm0

1057 .byte 102,15,56,220,209

1058 .byte 102,15,56,220,217

1059 bswapl %r9d

1060 .byte 102,15,56,220,225

1061 xorl %r11d,%r9d

1062 .byte 102,15,56,220,233

1063 movl %r9d,64+12(%rsp)

1064 leaq 5(%r8),%r9

1065 .byte 102,15,56,220,241

1066 .byte 102,15,56,220,249

1067 .byte 102,68,15,56,220,193

1068 .byte 102,68,15,56,220,201

1069 movups 112-128(%rcx),%xmm1

1070 .byte 102,15,56,220,208

1071 .byte 102,15,56,220,216

1072 bswapl %r9d

1073 .byte 102,15,56,220,224

1074 xorl %r11d,%r9d

1075 .byte 102,15,56,220,232

1076 movl %r9d,80+12(%rsp)

1077 leaq 6(%r8),%r9

1078 .byte 102,15,56,220,240

1079 .byte 102,15,56,220,248

1080 .byte 102,68,15,56,220,192

1081 .byte 102,68,15,56,220,200

1082 movups 128-128(%rcx),%xmm0

1083 .byte 102,15,56,220,209

1084 .byte 102,15,56,220,217

1085 bswapl %r9d

1086 .byte 102,15,56,220,225

1087 xorl %r11d,%r9d

1088 .byte 102,15,56,220,233

1089 movl %r9d,96+12(%rsp)

1090 leaq 7(%r8),%r9

1091 .byte 102,15,56,220,241

1092 .byte 102,15,56,220,249

1093 .byte 102,68,15,56,220,193

1094 .byte 102,68,15,56,220,201

1095 movups 144-128(%rcx),%xmm1

1096 .byte 102,15,56,220,208

1097 .byte 102,15,56,220,216

1098 bswapl %r9d

1099 .byte 102,15,56,220,224

1100 xorl %r11d,%r9d

1101 .byte 102,15,56,220,232

1102 movl %r9d,112+12(%rsp)

1103 .byte 102,15,56,220,240

1104 .byte 102,15,56,220,248

1105 .byte 102,68,15,56,220,192

1106 movdqu 0(%rdi),%xmm10

1107 .byte 102,68,15,56,220,200

1108 movups 160-128(%rcx),%xmm0

1109

1110 cmpl $11,%eax

1111 jb .Lctr32_enc_done

1112

1113 .byte 102,15,56,220,209

1114 .byte 102,15,56,220,217

1115 .byte 102,15,56,220,225

1116 .byte 102,15,56,220,233

1117 .byte 102,15,56,220,241

1118 .byte 102,15,56,220,249

1119 .byte 102,68,15,56,220,193

1120 .byte 102,68,15,56,220,201

1121 movups 176-128(%rcx),%xmm1

1122

1123 .byte 102,15,56,220,208

1124 .byte 102,15,56,220,216

1125 .byte 102,15,56,220,224

1126 .byte 102,15,56,220,232

1127 .byte 102,15,56,220,240

1128 .byte 102,15,56,220,248

1129 .byte 102,68,15,56,220,192

1130 .byte 102,68,15,56,220,200

1131 movups 192-128(%rcx),%xmm0

1132 je .Lctr32_enc_done

1133

1134 .byte 102,15,56,220,209

1135 .byte 102,15,56,220,217

1136 .byte 102,15,56,220,225

1137 .byte 102,15,56,220,233

1138 .byte 102,15,56,220,241

1139 .byte 102,15,56,220,249

1140 .byte 102,68,15,56,220,193

1141 .byte 102,68,15,56,220,201

1142 movups 208-128(%rcx),%xmm1

1143

1144 .byte 102,15,56,220,208

1145 .byte 102,15,56,220,216

1146 .byte 102,15,56,220,224

1147 .byte 102,15,56,220,232

1148 .byte 102,15,56,220,240

1149 .byte 102,15,56,220,248

1150 .byte 102,68,15,56,220,192

1151 .byte 102,68,15,56,220,200

1152 movups 224-128(%rcx),%xmm0

1153

1154 .Lctr32_enc_done:

1155 movdqu 16(%rdi),%xmm11

1156 pxor %xmm0,%xmm10

1157 movdqu 32(%rdi),%xmm12

1158 pxor %xmm0,%xmm11

1159 movdqu 48(%rdi),%xmm13

1160 pxor %xmm0,%xmm12

1161 movdqu 64(%rdi),%xmm14

1162 pxor %xmm0,%xmm13

1163 movdqu 80(%rdi),%xmm15

1164 pxor %xmm0,%xmm14

1165 .byte 102,15,56,220,209

1166 pxor %xmm0,%xmm15

1167 .byte 102,15,56,220,217

1168 .byte 102,15,56,220,225

1169 .byte 102,15,56,220,233

1170 .byte 102,15,56,220,241

1171 .byte 102,15,56,220,249

1172 .byte 102,68,15,56,220,193

1173 .byte 102,68,15,56,220,201

1174 movdqu 96(%rdi),%xmm1

1175

1176 .byte 102,65,15,56,221,210

1177 pxor %xmm0,%xmm1

1178 movdqu 112(%rdi),%xmm10

1179 leaq 128(%rdi),%rdi

1180 .byte 102,65,15,56,221,219

1181 pxor %xmm0,%xmm10

1182 movdqa 0(%rsp),%xmm11

1183 .byte 102,65,15,56,221,228

1184 movdqa 16(%rsp),%xmm12

1185 .byte 102,65,15,56,221,237

1186 movdqa 32(%rsp),%xmm13

1187 .byte 102,65,15,56,221,246

1188 movdqa 48(%rsp),%xmm14

1189 .byte 102,65,15,56,221,255

1190 movdqa 64(%rsp),%xmm15

1191 .byte 102,68,15,56,221,193

1192 movdqa 80(%rsp),%xmm0

1193 .byte 102,69,15,56,221,202

1194 movups 16-128(%rcx),%xmm1

1195

1196 movups %xmm2,(%rsi)

1197 movdqa %xmm11,%xmm2

1198 movups %xmm3,16(%rsi)

1199 movdqa %xmm12,%xmm3

1200 movups %xmm4,32(%rsi)

1201 movdqa %xmm13,%xmm4

1202 movups %xmm5,48(%rsi)

1203 movdqa %xmm14,%xmm5

1204 movups %xmm6,64(%rsi)

1205 movdqa %xmm15,%xmm6

1206 movups %xmm7,80(%rsi)

1207 movdqa %xmm0,%xmm7

1208 movups %xmm8,96(%rsi)

1209 movups %xmm9,112(%rsi)

1210 leaq 128(%rsi),%rsi

1211

1212 subq $8,%rdx

1213 jnc .Lctr32_loop8

1214

1215 addq $8,%rdx

1216 jz .Lctr32_done

1217 leaq -128(%rcx),%rcx

1218

1219 .Lctr32_tail:

1220 leaq 16(%rcx),%rcx

1221 cmpq $4,%rdx

1222 jb .Lctr32_loop3

1223 je .Lctr32_loop4

1224

1225 movdqa 96(%rsp),%xmm8

1226 pxor %xmm9,%xmm9

1227

1228 movups 16(%rcx),%xmm0

1229 .byte 102,15,56,220,209

1230 leaq 16(%rcx),%rcx

1231 .byte 102,15,56,220,217

1232 shrl $1,%eax

1233 .byte 102,15,56,220,225

1234 decl %eax

1235 .byte 102,15,56,220,233

1236 movups (%rdi),%xmm10

1237 .byte 102,15,56,220,241

1238 movups 16(%rdi),%xmm11

1239 .byte 102,15,56,220,249

1240 movups 32(%rdi),%xmm12

1241 .byte 102,68,15,56,220,193

1242 movups 16(%rcx),%xmm1

1243

1244 call .Lenc_loop8_enter

1245

1246 movdqu 48(%rdi),%xmm13

1247 pxor %xmm10,%xmm2

1248 movdqu 64(%rdi),%xmm10

1249 pxor %xmm11,%xmm3

1250 movdqu %xmm2,(%rsi)

1251 pxor %xmm12,%xmm4

1252 movdqu %xmm3,16(%rsi)

1253 pxor %xmm13,%xmm5

1254 movdqu %xmm4,32(%rsi)

1255 pxor %xmm10,%xmm6

1256 movdqu %xmm5,48(%rsi)

1257 movdqu %xmm6,64(%rsi)

1258 cmpq $6,%rdx

1259 jb .Lctr32_done

1260

1261 movups 80(%rdi),%xmm11

1262 xorps %xmm11,%xmm7

1263 movups %xmm7,80(%rsi)

1264 je .Lctr32_done

1265

1266 movups 96(%rdi),%xmm12

1267 xorps %xmm12,%xmm8

1268 movups %xmm8,96(%rsi)

1269 jmp .Lctr32_done

1270

1271 .align 32

1272 .Lctr32_loop4:

1273 .byte 102,15,56,220,209

1274 leaq 16(%rcx),%rcx

1275 .byte 102,15,56,220,217

1276 .byte 102,15,56,220,225

1277 .byte 102,15,56,220,233

1278 movups (%rcx),%xmm1

1279 decl %eax

1280 jnz .Lctr32_loop4

1281 .byte 102,15,56,221,209

1282 movups (%rdi),%xmm10

1283 .byte 102,15,56,221,217

1284 movups 16(%rdi),%xmm11

1285 .byte 102,15,56,221,225

1286 movups 32(%rdi),%xmm12

1287 .byte 102,15,56,221,233

1288 movups 48(%rdi),%xmm13

1289

1290 xorps %xmm10,%xmm2

1291 movups %xmm2,(%rsi)

1292 xorps %xmm11,%xmm3

1293 movups %xmm3,16(%rsi)

1294 pxor %xmm12,%xmm4

1295 movdqu %xmm4,32(%rsi)

1296 pxor %xmm13,%xmm5

1297 movdqu %xmm5,48(%rsi)

1298 jmp .Lctr32_done

1299

1300 .align 32

1301 .Lctr32_loop3:

1302 .byte 102,15,56,220,209

1303 leaq 16(%rcx),%rcx

1304 .byte 102,15,56,220,217

1305 .byte 102,15,56,220,225

1306 movups (%rcx),%xmm1

1307 decl %eax

1308 jnz .Lctr32_loop3

1309 .byte 102,15,56,221,209

1310 .byte 102,15,56,221,217

1311 .byte 102,15,56,221,225

1312

1313 movups (%rdi),%xmm10

1314 xorps %xmm10,%xmm2

1315 movups %xmm2,(%rsi)

1316 cmpq $2,%rdx

1317 jb .Lctr32_done

1318

1319 movups 16(%rdi),%xmm11

1320 xorps %xmm11,%xmm3

1321 movups %xmm3,16(%rsi)

1322 je .Lctr32_done

1323

1324 movups 32(%rdi),%xmm12

1325 xorps %xmm12,%xmm4

1326 movups %xmm4,32(%rsi)

1327 jmp .Lctr32_done

1328

1329 .align 16

1330 .Lctr32_one_shortcut:

1331 movups (%r8),%xmm2

1332 movups (%rdi),%xmm10

1333 movl 240(%rcx),%eax

1334 movups (%rcx),%xmm0

1335 movups 16(%rcx),%xmm1

1336 leaq 32(%rcx),%rcx

1337 xorps %xmm0,%xmm2

1338 .Loop_enc1_7:

1339 .byte 102,15,56,220,209

1340 decl %eax

1341 movups (%rcx),%xmm1

1342 leaq 16(%rcx),%rcx

1343 jnz .Loop_enc1_7

1344 .byte 102,15,56,221,209

1345 xorps %xmm10,%xmm2

1346 movups %xmm2,(%rsi)

1347 jmp .Lctr32_done

1348

1349 .align 16

1350 .Lctr32_done:

1351 leaq (%rbp),%rsp

1352 popq %rbp

1353 .Lctr32_epilogue:

1354 .byte 0xf3,0xc3

1355 .size aesni_ctr32_encrypt_blocks,.-aesni_ctr32_encrypt_blocks

1356 .globl aesni_xts_encrypt

1357 .type aesni_xts_encrypt,@function

1358 .align 16

1359 aesni_xts_encrypt:

1360 leaq (%rsp),%rax

1361 pushq %rbp

1362 subq $112,%rsp

1363 andq $-16,%rsp

1364 leaq -8(%rax),%rbp

1365 movups (%r9),%xmm15

1366 movl 240(%r8),%eax

1367 movl 240(%rcx),%r10d

1368 movups (%r8),%xmm0

1369 movups 16(%r8),%xmm1

1370 leaq 32(%r8),%r8

1371 xorps %xmm0,%xmm15

1372 .Loop_enc1_8:

1373 .byte 102,68,15,56,220,249

1374 decl %eax

1375 movups (%r8),%xmm1

1376 leaq 16(%r8),%r8

1377 jnz .Loop_enc1_8

1378 .byte 102,68,15,56,221,249

1379 movups (%rcx),%xmm0

1380 movq %rcx,%r11

1381 movl %r10d,%eax

1382 shll $4,%r10d

1383 movq %rdx,%r9

1384 andq $-16,%rdx

1385

1386 movups 16(%rcx,%r10,1),%xmm1

1387 movl %eax,%r10d

1388

1389 movdqa .Lxts_magic(%rip),%xmm8

1390 pshufd $95,%xmm15,%xmm9

1391 pxor %xmm0,%xmm1

1392 movdqa %xmm9,%xmm14

1393 paddd %xmm9,%xmm9

1394 movdqa %xmm15,%xmm10

1395 psrad $31,%xmm14

1396 paddq %xmm15,%xmm15

1397 pand %xmm8,%xmm14

1398 pxor %xmm0,%xmm10

1399 pxor %xmm14,%xmm15

1400 movdqa %xmm9,%xmm14

1401 paddd %xmm9,%xmm9

1402 movdqa %xmm15,%xmm11

1403 psrad $31,%xmm14

1404 paddq %xmm15,%xmm15

1405 pand %xmm8,%xmm14

1406 pxor %xmm0,%xmm11

1407 pxor %xmm14,%xmm15

1408 movdqa %xmm9,%xmm14

1409 paddd %xmm9,%xmm9

1410 movdqa %xmm15,%xmm12

1411 psrad $31,%xmm14

1412 paddq %xmm15,%xmm15

1413 pand %xmm8,%xmm14

1414 pxor %xmm0,%xmm12

1415 pxor %xmm14,%xmm15

1416 movdqa %xmm9,%xmm14

1417 paddd %xmm9,%xmm9

1418 movdqa %xmm15,%xmm13

1419 psrad $31,%xmm14

1420 paddq %xmm15,%xmm15

1421 pand %xmm8,%xmm14

1422 pxor %xmm0,%xmm13

1423 pxor %xmm14,%xmm15

1424 movdqa %xmm15,%xmm14

1425 psrad $31,%xmm9

1426 paddq %xmm15,%xmm15

1427 pand %xmm8,%xmm9

1428 pxor %xmm0,%xmm14

1429 pxor %xmm9,%xmm15

1430 movaps %xmm1,96(%rsp)

1431

1432 subq $96,%rdx

1433 jc .Lxts_enc_short

1434

1435 shrl $1,%eax

1436 subl $3,%eax

1437 movups 16(%r11),%xmm1

1438 movl %eax,%r10d

1439 leaq .Lxts_magic(%rip),%r8

1440 jmp .Lxts_enc_grandloop

1441

1442 .align 32

1443 .Lxts_enc_grandloop:

1444 movdqu 0(%rdi),%xmm2

1445 movdqa %xmm0,%xmm8

1446 movdqu 16(%rdi),%xmm3

1447 pxor %xmm10,%xmm2

1448 movdqu 32(%rdi),%xmm4

1449 pxor %xmm11,%xmm3

1450 .byte 102,15,56,220,209

1451 movdqu 48(%rdi),%xmm5

1452 pxor %xmm12,%xmm4

1453 .byte 102,15,56,220,217

1454 movdqu 64(%rdi),%xmm6

1455 pxor %xmm13,%xmm5

1456 .byte 102,15,56,220,225

1457 movdqu 80(%rdi),%xmm7

1458 pxor %xmm15,%xmm8

1459 movdqa 96(%rsp),%xmm9

1460 pxor %xmm14,%xmm6

1461 .byte 102,15,56,220,233

1462 movups 32(%r11),%xmm0

1463 leaq 96(%rdi),%rdi

1464 pxor %xmm8,%xmm7

1465

1466 pxor %xmm9,%xmm10

1467 .byte 102,15,56,220,241

1468 pxor %xmm9,%xmm11

1469 movdqa %xmm10,0(%rsp)

1470 .byte 102,15,56,220,249

1471 movups 48(%r11),%xmm1

1472

1473 .byte 102,15,56,220,208

1474 pxor %xmm9,%xmm12

1475 movdqa %xmm11,16(%rsp)

1476 .byte 102,15,56,220,216

1477 pxor %xmm9,%xmm13

1478 movdqa %xmm12,32(%rsp)

1479 .byte 102,15,56,220,224

1480 pxor %xmm9,%xmm14

1481 .byte 102,15,56,220,232

1482 pxor %xmm9,%xmm8

1483 movdqa %xmm14,64(%rsp)

1484 .byte 102,15,56,220,240

1485 movdqa %xmm8,80(%rsp)

1486 .byte 102,15,56,220,248

1487 movups 64(%r11),%xmm0

1488 leaq 64(%r11),%rcx

1489 pshufd $95,%xmm15,%xmm9

1490 jmp .Lxts_enc_loop6

1491 .align 32

1492 .Lxts_enc_loop6:

1493 .byte 102,15,56,220,209

1494 .byte 102,15,56,220,217

1495 .byte 102,15,56,220,225

1496 .byte 102,15,56,220,233

1497 .byte 102,15,56,220,241

1498 .byte 102,15,56,220,249

1499 movups 16(%rcx),%xmm1

1500 leaq 32(%rcx),%rcx

1501

1502 .byte 102,15,56,220,208

1503 .byte 102,15,56,220,216

1504 .byte 102,15,56,220,224

1505 .byte 102,15,56,220,232

1506 .byte 102,15,56,220,240

1507 .byte 102,15,56,220,248

1508 movups (%rcx),%xmm0

1509 decl %eax

1510 jnz .Lxts_enc_loop6

1511

1512 movdqa (%r8),%xmm8

1513 movdqa %xmm9,%xmm14

1514 paddd %xmm9,%xmm9

1515 .byte 102,15,56,220,209

1516 paddq %xmm15,%xmm15

1517 psrad $31,%xmm14

1518 .byte 102,15,56,220,217

1519 pand %xmm8,%xmm14

1520 movups (%r11),%xmm10

1521 .byte 102,15,56,220,225

1522 .byte 102,15,56,220,233

1523 pxor %xmm14,%xmm15

1524 .byte 102,15,56,220,241

1525 movaps %xmm10,%xmm11

1526 .byte 102,15,56,220,249

1527 movups 16(%rcx),%xmm1

1528

1529 movdqa %xmm9,%xmm14

1530 paddd %xmm9,%xmm9

1531 .byte 102,15,56,220,208

1532 pxor %xmm15,%xmm10

1533 psrad $31,%xmm14

1534 .byte 102,15,56,220,216

1535 paddq %xmm15,%xmm15

1536 pand %xmm8,%xmm14

1537 .byte 102,15,56,220,224

1538 .byte 102,15,56,220,232

1539 pxor %xmm14,%xmm15

1540 .byte 102,15,56,220,240

1541 movaps %xmm11,%xmm12

1542 .byte 102,15,56,220,248

1543 movups 32(%rcx),%xmm0

1544

1545 movdqa %xmm9,%xmm14

1546 paddd %xmm9,%xmm9

1547 .byte 102,15,56,220,209

1548 pxor %xmm15,%xmm11

1549 psrad $31,%xmm14

1550 .byte 102,15,56,220,217

1551 paddq %xmm15,%xmm15

1552 pand %xmm8,%xmm14

1553 .byte 102,15,56,220,225

1554 movdqa %xmm13,48(%rsp)

1555 .byte 102,15,56,220,233

1556 pxor %xmm14,%xmm15

1557 .byte 102,15,56,220,241

1558 movaps %xmm12,%xmm13

1559 .byte 102,15,56,220,249

1560 movups 48(%rcx),%xmm1

1561

1562 movdqa %xmm9,%xmm14

1563 paddd %xmm9,%xmm9

1564 .byte 102,15,56,220,208

1565 pxor %xmm15,%xmm12

1566 psrad $31,%xmm14

1567 .byte 102,15,56,220,216

1568 paddq %xmm15,%xmm15

1569 pand %xmm8,%xmm14

1570 .byte 102,15,56,220,224

1571 .byte 102,15,56,220,232

1572 pxor %xmm14,%xmm15

1573 .byte 102,15,56,220,240

1574 movaps %xmm13,%xmm14

1575 .byte 102,15,56,220,248

1576

1577 movdqa %xmm9,%xmm0

1578 paddd %xmm9,%xmm9

1579 .byte 102,15,56,220,209

1580 pxor %xmm15,%xmm13

1581 psrad $31,%xmm0

1582 .byte 102,15,56,220,217

1583 paddq %xmm15,%xmm15

1584 pand %xmm8,%xmm0

1585 .byte 102,15,56,220,225

1586 .byte 102,15,56,220,233

1587 pxor %xmm0,%xmm15

1588 movups (%r11),%xmm0

1589 .byte 102,15,56,220,241

1590 .byte 102,15,56,220,249

1591 movups 16(%r11),%xmm1

1592

1593 pxor %xmm15,%xmm14

1594 psrad $31,%xmm9

1595 .byte 102,15,56,221,84,36,0

1596 paddq %xmm15,%xmm15

1597 pand %xmm8,%xmm9

1598 .byte 102,15,56,221,92,36,16

1599 .byte 102,15,56,221,100,36,32

1600 pxor %xmm9,%xmm15

1601 .byte 102,15,56,221,108,36,48

1602 .byte 102,15,56,221,116,36,64

1603 .byte 102,15,56,221,124,36,80

1604 movl %r10d,%eax

1605

1606 leaq 96(%rsi),%rsi

1607 movups %xmm2,-96(%rsi)

1608 movups %xmm3,-80(%rsi)

1609 movups %xmm4,-64(%rsi)

1610 movups %xmm5,-48(%rsi)

1611 movups %xmm6,-32(%rsi)

1612 movups %xmm7,-16(%rsi)

1613 subq $96,%rdx

1614 jnc .Lxts_enc_grandloop

1615

1616 leal 7(%rax,%rax,1),%eax

1617 movq %r11,%rcx

1618 movl %eax,%r10d

1619

1620 .Lxts_enc_short:

1621 pxor %xmm0,%xmm10

1622 addq $96,%rdx

1623 jz .Lxts_enc_done

1624

1625 pxor %xmm0,%xmm11

1626 cmpq $32,%rdx

1627 jb .Lxts_enc_one

1628 pxor %xmm0,%xmm12

1629 je .Lxts_enc_two

1630

1631 pxor %xmm0,%xmm13

1632 cmpq $64,%rdx

1633 jb .Lxts_enc_three

1634 pxor %xmm0,%xmm14

1635 je .Lxts_enc_four

1636

1637 movdqu (%rdi),%xmm2

1638 movdqu 16(%rdi),%xmm3

1639 movdqu 32(%rdi),%xmm4

1640 pxor %xmm10,%xmm2

1641 movdqu 48(%rdi),%xmm5

1642 pxor %xmm11,%xmm3

1643 movdqu 64(%rdi),%xmm6

1644 leaq 80(%rdi),%rdi

1645 pxor %xmm12,%xmm4

1646 pxor %xmm13,%xmm5

1647 pxor %xmm14,%xmm6

1648

1649 call _aesni_encrypt6

1650

1651 xorps %xmm10,%xmm2

1652 movdqa %xmm15,%xmm10

1653 xorps %xmm11,%xmm3

1654 xorps %xmm12,%xmm4

1655 movdqu %xmm2,(%rsi)

1656 xorps %xmm13,%xmm5

1657 movdqu %xmm3,16(%rsi)

1658 xorps %xmm14,%xmm6

1659 movdqu %xmm4,32(%rsi)

1660 movdqu %xmm5,48(%rsi)

1661 movdqu %xmm6,64(%rsi)

1662 leaq 80(%rsi),%rsi

1663 jmp .Lxts_enc_done

1664

1665 .align 16

1666 .Lxts_enc_one:

1667 movups (%rdi),%xmm2

1668 leaq 16(%rdi),%rdi

1669 xorps %xmm10,%xmm2

1670 movups (%rcx),%xmm0

1671 movups 16(%rcx),%xmm1

1672 leaq 32(%rcx),%rcx

1673 xorps %xmm0,%xmm2

1674 .Loop_enc1_9:

1675 .byte 102,15,56,220,209

1676 decl %eax

1677 movups (%rcx),%xmm1

1678 leaq 16(%rcx),%rcx

1679 jnz .Loop_enc1_9

1680 .byte 102,15,56,221,209

1681 xorps %xmm10,%xmm2

1682 movdqa %xmm11,%xmm10

1683 movups %xmm2,(%rsi)

1684 leaq 16(%rsi),%rsi

1685 jmp .Lxts_enc_done

1686

1687 .align 16

1688 .Lxts_enc_two:

1689 movups (%rdi),%xmm2

1690 movups 16(%rdi),%xmm3

1691 leaq 32(%rdi),%rdi

1692 xorps %xmm10,%xmm2

1693 xorps %xmm11,%xmm3

1694

1695 call _aesni_encrypt3

1696

1697 xorps %xmm10,%xmm2

1698 movdqa %xmm12,%xmm10

1699 xorps %xmm11,%xmm3

1700 movups %xmm2,(%rsi)

1701 movups %xmm3,16(%rsi)

1702 leaq 32(%rsi),%rsi

1703 jmp .Lxts_enc_done

1704

1705 .align 16

1706 .Lxts_enc_three:

1707 movups (%rdi),%xmm2

1708 movups 16(%rdi),%xmm3

1709 movups 32(%rdi),%xmm4

1710 leaq 48(%rdi),%rdi

1711 xorps %xmm10,%xmm2

1712 xorps %xmm11,%xmm3

1713 xorps %xmm12,%xmm4

1714

1715 call _aesni_encrypt3

1716

1717 xorps %xmm10,%xmm2

1718 movdqa %xmm13,%xmm10

1719 xorps %xmm11,%xmm3

1720 xorps %xmm12,%xmm4

1721 movups %xmm2,(%rsi)

1722 movups %xmm3,16(%rsi)

1723 movups %xmm4,32(%rsi)

1724 leaq 48(%rsi),%rsi

1725 jmp .Lxts_enc_done

1726

1727 .align 16

1728 .Lxts_enc_four:

1729 movups (%rdi),%xmm2

1730 movups 16(%rdi),%xmm3

1731 movups 32(%rdi),%xmm4

1732 xorps %xmm10,%xmm2

1733 movups 48(%rdi),%xmm5

1734 leaq 64(%rdi),%rdi

1735 xorps %xmm11,%xmm3

1736 xorps %xmm12,%xmm4

1737 xorps %xmm13,%xmm5

1738

1739 call _aesni_encrypt4

1740

1741 pxor %xmm10,%xmm2

1742 movdqa %xmm14,%xmm10

1743 pxor %xmm11,%xmm3

1744 pxor %xmm12,%xmm4

1745 movdqu %xmm2,(%rsi)

1746 pxor %xmm13,%xmm5

1747 movdqu %xmm3,16(%rsi)

1748 movdqu %xmm4,32(%rsi)

1749 movdqu %xmm5,48(%rsi)

1750 leaq 64(%rsi),%rsi

1751 jmp .Lxts_enc_done

1752

1753 .align 16

1754 .Lxts_enc_done:

1755 andq $15,%r9

1756 jz .Lxts_enc_ret

1757 movq %r9,%rdx

1758

1759 .Lxts_enc_steal:

1760 movzbl (%rdi),%eax

1761 movzbl -16(%rsi),%ecx

1762 leaq 1(%rdi),%rdi

1763 movb %al,-16(%rsi)

1764 movb %cl,0(%rsi)

1765 leaq 1(%rsi),%rsi

1766 subq $1,%rdx

1767 jnz .Lxts_enc_steal

1768

1769 subq %r9,%rsi

1770 movq %r11,%rcx

1771 movl %r10d,%eax

1772

1773 movups -16(%rsi),%xmm2

1774 xorps %xmm10,%xmm2

1775 movups (%rcx),%xmm0

1776 movups 16(%rcx),%xmm1

1777 leaq 32(%rcx),%rcx

1778 xorps %xmm0,%xmm2

1779 .Loop_enc1_10:

1780 .byte 102,15,56,220,209

1781 decl %eax

1782 movups (%rcx),%xmm1

1783 leaq 16(%rcx),%rcx

1784 jnz .Loop_enc1_10

1785 .byte 102,15,56,221,209

1786 xorps %xmm10,%xmm2

1787 movups %xmm2,-16(%rsi)

1788

1789 .Lxts_enc_ret:

1790 leaq (%rbp),%rsp

1791 popq %rbp

1792 .Lxts_enc_epilogue:

1793 .byte 0xf3,0xc3

1794 .size aesni_xts_encrypt,.-aesni_xts_encrypt

1795 .globl aesni_xts_decrypt

1796 .type aesni_xts_decrypt,@function

1797 .align 16

1798 aesni_xts_decrypt:

1799 leaq (%rsp),%rax

1800 pushq %rbp

1801 subq $112,%rsp

1802 andq $-16,%rsp

1803 leaq -8(%rax),%rbp

1804 movups (%r9),%xmm15

1805 movl 240(%r8),%eax

1806 movl 240(%rcx),%r10d

1807 movups (%r8),%xmm0

1808 movups 16(%r8),%xmm1

1809 leaq 32(%r8),%r8

1810 xorps %xmm0,%xmm15

1811 .Loop_enc1_11:

1812 .byte 102,68,15,56,220,249

1813 decl %eax

1814 movups (%r8),%xmm1

1815 leaq 16(%r8),%r8

1816 jnz .Loop_enc1_11

1817 .byte 102,68,15,56,221,249

1818 xorl %eax,%eax

1819 testq $15,%rdx

1820 setnz %al

1821 shlq $4,%rax

1822 subq %rax,%rdx

1823

1824 movups (%rcx),%xmm0

1825 movq %rcx,%r11

1826 movl %r10d,%eax

1827 shll $4,%r10d

1828 movq %rdx,%r9

1829 andq $-16,%rdx

1830

1831 movups 16(%rcx,%r10,1),%xmm1

1832 movl %eax,%r10d

1833

1834 movdqa .Lxts_magic(%rip),%xmm8

1835 pshufd $95,%xmm15,%xmm9

1836 pxor %xmm0,%xmm1

1837 movdqa %xmm9,%xmm14

1838 paddd %xmm9,%xmm9

1839 movdqa %xmm15,%xmm10

1840 psrad $31,%xmm14

1841 paddq %xmm15,%xmm15

1842 pand %xmm8,%xmm14

1843 pxor %xmm0,%xmm10

1844 pxor %xmm14,%xmm15

1845 movdqa %xmm9,%xmm14

1846 paddd %xmm9,%xmm9

1847 movdqa %xmm15,%xmm11

1848 psrad $31,%xmm14

1849 paddq %xmm15,%xmm15

1850 pand %xmm8,%xmm14

1851 pxor %xmm0,%xmm11

1852 pxor %xmm14,%xmm15

1853 movdqa %xmm9,%xmm14

1854 paddd %xmm9,%xmm9

1855 movdqa %xmm15,%xmm12

1856 psrad $31,%xmm14

1857 paddq %xmm15,%xmm15

1858 pand %xmm8,%xmm14

1859 pxor %xmm0,%xmm12

1860 pxor %xmm14,%xmm15

1861 movdqa %xmm9,%xmm14

1862 paddd %xmm9,%xmm9

1863 movdqa %xmm15,%xmm13

1864 psrad $31,%xmm14

1865 paddq %xmm15,%xmm15

1866 pand %xmm8,%xmm14

1867 pxor %xmm0,%xmm13

1868 pxor %xmm14,%xmm15

1869 movdqa %xmm15,%xmm14

1870 psrad $31,%xmm9

1871 paddq %xmm15,%xmm15

1872 pand %xmm8,%xmm9

1873 pxor %xmm0,%xmm14

1874 pxor %xmm9,%xmm15

1875 movaps %xmm1,96(%rsp)

1876

1877 subq $96,%rdx

1878 jc .Lxts_dec_short

1879

1880 shrl $1,%eax

1881 subl $3,%eax

1882 movups 16(%r11),%xmm1

1883 movl %eax,%r10d

1884 leaq .Lxts_magic(%rip),%r8

1885 jmp .Lxts_dec_grandloop

1886

1887 .align 32

1888 .Lxts_dec_grandloop:

1889 movdqu 0(%rdi),%xmm2

1890 movdqa %xmm0,%xmm8

1891 movdqu 16(%rdi),%xmm3

1892 pxor %xmm10,%xmm2

1893 movdqu 32(%rdi),%xmm4

1894 pxor %xmm11,%xmm3

1895 .byte 102,15,56,222,209

1896 movdqu 48(%rdi),%xmm5

1897 pxor %xmm12,%xmm4

1898 .byte 102,15,56,222,217

1899 movdqu 64(%rdi),%xmm6

1900 pxor %xmm13,%xmm5

1901 .byte 102,15,56,222,225

1902 movdqu 80(%rdi),%xmm7

1903 pxor %xmm15,%xmm8

1904 movdqa 96(%rsp),%xmm9

1905 pxor %xmm14,%xmm6

1906 .byte 102,15,56,222,233

1907 movups 32(%r11),%xmm0

1908 leaq 96(%rdi),%rdi

1909 pxor %xmm8,%xmm7

1910

1911 pxor %xmm9,%xmm10

1912 .byte 102,15,56,222,241

1913 pxor %xmm9,%xmm11

1914 movdqa %xmm10,0(%rsp)

1915 .byte 102,15,56,222,249

1916 movups 48(%r11),%xmm1

1917

1918 .byte 102,15,56,222,208

1919 pxor %xmm9,%xmm12

1920 movdqa %xmm11,16(%rsp)

1921 .byte 102,15,56,222,216

1922 pxor %xmm9,%xmm13

1923 movdqa %xmm12,32(%rsp)

1924 .byte 102,15,56,222,224

1925 pxor %xmm9,%xmm14

1926 .byte 102,15,56,222,232

1927 pxor %xmm9,%xmm8

1928 movdqa %xmm14,64(%rsp)

1929 .byte 102,15,56,222,240

1930 movdqa %xmm8,80(%rsp)

1931 .byte 102,15,56,222,248

1932 movups 64(%r11),%xmm0

1933 leaq 64(%r11),%rcx

1934 pshufd $95,%xmm15,%xmm9

1935 jmp .Lxts_dec_loop6

1936 .align 32

1937 .Lxts_dec_loop6:

1938 .byte 102,15,56,222,209

1939 .byte 102,15,56,222,217

1940 .byte 102,15,56,222,225

1941 .byte 102,15,56,222,233

1942 .byte 102,15,56,222,241

1943 .byte 102,15,56,222,249

1944 movups 16(%rcx),%xmm1

1945 leaq 32(%rcx),%rcx

1946

1947 .byte 102,15,56,222,208

1948 .byte 102,15,56,222,216

1949 .byte 102,15,56,222,224

1950 .byte 102,15,56,222,232

1951 .byte 102,15,56,222,240

1952 .byte 102,15,56,222,248

1953 movups (%rcx),%xmm0

1954 decl %eax

1955 jnz .Lxts_dec_loop6

1956

1957 movdqa (%r8),%xmm8

1958 movdqa %xmm9,%xmm14

1959 paddd %xmm9,%xmm9

1960 .byte 102,15,56,222,209

1961 paddq %xmm15,%xmm15

1962 psrad $31,%xmm14

1963 .byte 102,15,56,222,217

1964 pand %xmm8,%xmm14

1965 movups (%r11),%xmm10

1966 .byte 102,15,56,222,225

1967 .byte 102,15,56,222,233

1968 pxor %xmm14,%xmm15

1969 .byte 102,15,56,222,241

1970 movaps %xmm10,%xmm11

1971 .byte 102,15,56,222,249

1972 movups 16(%rcx),%xmm1

1973

1974 movdqa %xmm9,%xmm14

1975 paddd %xmm9,%xmm9

1976 .byte 102,15,56,222,208

1977 pxor %xmm15,%xmm10

1978 psrad $31,%xmm14

1979 .byte 102,15,56,222,216

1980 paddq %xmm15,%xmm15

1981 pand %xmm8,%xmm14

1982 .byte 102,15,56,222,224

1983 .byte 102,15,56,222,232

1984 pxor %xmm14,%xmm15

1985 .byte 102,15,56,222,240

1986 movaps %xmm11,%xmm12

1987 .byte 102,15,56,222,248

1988 movups 32(%rcx),%xmm0

1989

1990 movdqa %xmm9,%xmm14

1991 paddd %xmm9,%xmm9

1992 .byte 102,15,56,222,209

1993 pxor %xmm15,%xmm11

1994 psrad $31,%xmm14

1995 .byte 102,15,56,222,217

1996 paddq %xmm15,%xmm15

1997 pand %xmm8,%xmm14

1998 .byte 102,15,56,222,225

1999 movdqa %xmm13,48(%rsp)

2000 .byte 102,15,56,222,233

2001 pxor %xmm14,%xmm15

2002 .byte 102,15,56,222,241

2003 movaps %xmm12,%xmm13

2004 .byte 102,15,56,222,249

2005 movups 48(%rcx),%xmm1

2006

2007 movdqa %xmm9,%xmm14

2008 paddd %xmm9,%xmm9

2009 .byte 102,15,56,222,208

2010 pxor %xmm15,%xmm12

2011 psrad $31,%xmm14

2012 .byte 102,15,56,222,216

2013 paddq %xmm15,%xmm15

2014 pand %xmm8,%xmm14

2015 .byte 102,15,56,222,224

2016 .byte 102,15,56,222,232

2017 pxor %xmm14,%xmm15

2018 .byte 102,15,56,222,240

2019 movaps %xmm13,%xmm14

2020 .byte 102,15,56,222,248

2021

2022 movdqa %xmm9,%xmm0

2023 paddd %xmm9,%xmm9

2024 .byte 102,15,56,222,209

2025 pxor %xmm15,%xmm13

2026 psrad $31,%xmm0

2027 .byte 102,15,56,222,217

2028 paddq %xmm15,%xmm15

2029 pand %xmm8,%xmm0

2030 .byte 102,15,56,222,225

2031 .byte 102,15,56,222,233

2032 pxor %xmm0,%xmm15

2033 movups (%r11),%xmm0

2034 .byte 102,15,56,222,241

2035 .byte 102,15,56,222,249

2036 movups 16(%r11),%xmm1

2037

2038 pxor %xmm15,%xmm14

2039 psrad $31,%xmm9

2040 .byte 102,15,56,223,84,36,0

2041 paddq %xmm15,%xmm15

2042 pand %xmm8,%xmm9

2043 .byte 102,15,56,223,92,36,16

2044 .byte 102,15,56,223,100,36,32

2045 pxor %xmm9,%xmm15

2046 .byte 102,15,56,223,108,36,48

2047 .byte 102,15,56,223,116,36,64

2048 .byte 102,15,56,223,124,36,80

2049 movl %r10d,%eax

2050

2051 leaq 96(%rsi),%rsi

2052 movups %xmm2,-96(%rsi)

2053 movups %xmm3,-80(%rsi)

2054 movups %xmm4,-64(%rsi)

2055 movups %xmm5,-48(%rsi)

2056 movups %xmm6,-32(%rsi)

2057 movups %xmm7,-16(%rsi)

2058 subq $96,%rdx

2059 jnc .Lxts_dec_grandloop

2060

2061 leal 7(%rax,%rax,1),%eax

2062 movq %r11,%rcx

2063 movl %eax,%r10d

2064

2065 .Lxts_dec_short:

2066 pxor %xmm0,%xmm10

2067 pxor %xmm0,%xmm11

2068 addq $96,%rdx

2069 jz .Lxts_dec_done

2070

2071 pxor %xmm0,%xmm12

2072 cmpq $32,%rdx

2073 jb .Lxts_dec_one

2074 pxor %xmm0,%xmm13

2075 je .Lxts_dec_two

2076

2077 pxor %xmm0,%xmm14

2078 cmpq $64,%rdx

2079 jb .Lxts_dec_three

2080 je .Lxts_dec_four

2081

2082 movdqu (%rdi),%xmm2

2083 movdqu 16(%rdi),%xmm3

2084 movdqu 32(%rdi),%xmm4

2085 pxor %xmm10,%xmm2

2086 movdqu 48(%rdi),%xmm5

2087 pxor %xmm11,%xmm3

2088 movdqu 64(%rdi),%xmm6

2089 leaq 80(%rdi),%rdi

2090 pxor %xmm12,%xmm4

2091 pxor %xmm13,%xmm5

2092 pxor %xmm14,%xmm6

2093

2094 call _aesni_decrypt6

2095

2096 xorps %xmm10,%xmm2

2097 xorps %xmm11,%xmm3

2098 xorps %xmm12,%xmm4

2099 movdqu %xmm2,(%rsi)

2100 xorps %xmm13,%xmm5

2101 movdqu %xmm3,16(%rsi)

2102 xorps %xmm14,%xmm6

2103 movdqu %xmm4,32(%rsi)

2104 pxor %xmm14,%xmm14

2105 movdqu %xmm5,48(%rsi)

2106 pcmpgtd %xmm15,%xmm14

2107 movdqu %xmm6,64(%rsi)

2108 leaq 80(%rsi),%rsi

2109 pshufd $19,%xmm14,%xmm11

2110 andq $15,%r9

2111 jz .Lxts_dec_ret

2112

2113 movdqa %xmm15,%xmm10

2114 paddq %xmm15,%xmm15

2115 pand %xmm8,%xmm11

2116 pxor %xmm15,%xmm11

2117 jmp .Lxts_dec_done2

2118

2119 .align 16

2120 .Lxts_dec_one:

2121 movups (%rdi),%xmm2

2122 leaq 16(%rdi),%rdi

2123 xorps %xmm10,%xmm2

2124 movups (%rcx),%xmm0

2125 movups 16(%rcx),%xmm1

2126 leaq 32(%rcx),%rcx

2127 xorps %xmm0,%xmm2

2128 .Loop_dec1_12:

2129 .byte 102,15,56,222,209

2130 decl %eax

2131 movups (%rcx),%xmm1

2132 leaq 16(%rcx),%rcx

2133 jnz .Loop_dec1_12

2134 .byte 102,15,56,223,209

2135 xorps %xmm10,%xmm2

2136 movdqa %xmm11,%xmm10

2137 movups %xmm2,(%rsi)

2138 movdqa %xmm12,%xmm11

2139 leaq 16(%rsi),%rsi

2140 jmp .Lxts_dec_done

2141

2142 .align 16

2143 .Lxts_dec_two:

2144 movups (%rdi),%xmm2

2145 movups 16(%rdi),%xmm3

2146 leaq 32(%rdi),%rdi

2147 xorps %xmm10,%xmm2

2148 xorps %xmm11,%xmm3

2149

2150 call _aesni_decrypt3

2151

2152 xorps %xmm10,%xmm2

2153 movdqa %xmm12,%xmm10

2154 xorps %xmm11,%xmm3

2155 movdqa %xmm13,%xmm11

2156 movups %xmm2,(%rsi)

2157 movups %xmm3,16(%rsi)

2158 leaq 32(%rsi),%rsi

2159 jmp .Lxts_dec_done

2160

2161 .align 16

2162 .Lxts_dec_three:

2163 movups (%rdi),%xmm2

2164 movups 16(%rdi),%xmm3

2165 movups 32(%rdi),%xmm4

2166 leaq 48(%rdi),%rdi

2167 xorps %xmm10,%xmm2

2168 xorps %xmm11,%xmm3

2169 xorps %xmm12,%xmm4

2170

2171 call _aesni_decrypt3

2172

2173 xorps %xmm10,%xmm2

2174 movdqa %xmm13,%xmm10

2175 xorps %xmm11,%xmm3

2176 movdqa %xmm14,%xmm11

2177 xorps %xmm12,%xmm4

2178 movups %xmm2,(%rsi)

2179 movups %xmm3,16(%rsi)

2180 movups %xmm4,32(%rsi)

2181 leaq 48(%rsi),%rsi

2182 jmp .Lxts_dec_done

2183

2184 .align 16

2185 .Lxts_dec_four:

2186 movups (%rdi),%xmm2

2187 movups 16(%rdi),%xmm3

2188 movups 32(%rdi),%xmm4

2189 xorps %xmm10,%xmm2

2190 movups 48(%rdi),%xmm5

2191 leaq 64(%rdi),%rdi

2192 xorps %xmm11,%xmm3

2193 xorps %xmm12,%xmm4

2194 xorps %xmm13,%xmm5

2195

2196 call _aesni_decrypt4

2197

2198 pxor %xmm10,%xmm2

2199 movdqa %xmm14,%xmm10

2200 pxor %xmm11,%xmm3

2201 movdqa %xmm15,%xmm11

2202 pxor %xmm12,%xmm4

2203 movdqu %xmm2,(%rsi)

2204 pxor %xmm13,%xmm5

2205 movdqu %xmm3,16(%rsi)

2206 movdqu %xmm4,32(%rsi)

2207 movdqu %xmm5,48(%rsi)

2208 leaq 64(%rsi),%rsi

2209 jmp .Lxts_dec_done

2210

2211 .align 16

2212 .Lxts_dec_done:

2213 andq $15,%r9

2214 jz .Lxts_dec_ret

2215 .Lxts_dec_done2:

2216 movq %r9,%rdx

2217 movq %r11,%rcx

2218 movl %r10d,%eax

2219

2220 movups (%rdi),%xmm2

2221 xorps %xmm11,%xmm2

2222 movups (%rcx),%xmm0

2223 movups 16(%rcx),%xmm1

2224 leaq 32(%rcx),%rcx

2225 xorps %xmm0,%xmm2

2226 .Loop_dec1_13:

2227 .byte 102,15,56,222,209

2228 decl %eax

2229 movups (%rcx),%xmm1

2230 leaq 16(%rcx),%rcx

2231 jnz .Loop_dec1_13

2232 .byte 102,15,56,223,209

2233 xorps %xmm11,%xmm2

2234 movups %xmm2,(%rsi)

2235

2236 .Lxts_dec_steal:

2237 movzbl 16(%rdi),%eax

2238 movzbl (%rsi),%ecx

2239 leaq 1(%rdi),%rdi

2240 movb %al,(%rsi)

2241 movb %cl,16(%rsi)

2242 leaq 1(%rsi),%rsi

2243 subq $1,%rdx

2244 jnz .Lxts_dec_steal

2245

2246 subq %r9,%rsi

2247 movq %r11,%rcx

2248 movl %r10d,%eax

2249

2250 movups (%rsi),%xmm2

2251 xorps %xmm10,%xmm2

2252 movups (%rcx),%xmm0

2253 movups 16(%rcx),%xmm1

2254 leaq 32(%rcx),%rcx

2255 xorps %xmm0,%xmm2

2256 .Loop_dec1_14:

2257 .byte 102,15,56,222,209

2258 decl %eax

2259 movups (%rcx),%xmm1

2260 leaq 16(%rcx),%rcx

2261 jnz .Loop_dec1_14

2262 .byte 102,15,56,223,209

2263 xorps %xmm10,%xmm2

2264 movups %xmm2,(%rsi)

2265

2266 .Lxts_dec_ret:

2267 leaq (%rbp),%rsp

2268 popq %rbp

2269 .Lxts_dec_epilogue:

2270 .byte 0xf3,0xc3

2271 .size aesni_xts_decrypt,.-aesni_xts_decrypt

2272 .globl aesni_cbc_encrypt

2273 .type aesni_cbc_encrypt,@function

2274 .align 16

2275 aesni_cbc_encrypt:

2276 testq %rdx,%rdx

2277 jz .Lcbc_ret

2278

2279 movl 240(%rcx),%r10d

2280 movq %rcx,%r11

2281 testl %r9d,%r9d

2282 jz .Lcbc_decrypt

2283

2284 movups (%r8),%xmm2

2285 movl %r10d,%eax

2286 cmpq $16,%rdx

2287 jb .Lcbc_enc_tail

2288 subq $16,%rdx

2289 jmp .Lcbc_enc_loop

2290 .align 16

2291 .Lcbc_enc_loop:

2292 movups (%rdi),%xmm3

2293 leaq 16(%rdi),%rdi

2294

2295 movups (%rcx),%xmm0

2296 movups 16(%rcx),%xmm1

2297 xorps %xmm0,%xmm3

2298 leaq 32(%rcx),%rcx

2299 xorps %xmm3,%xmm2

2300 .Loop_enc1_15:

2301 .byte 102,15,56,220,209

2302 decl %eax

2303 movups (%rcx),%xmm1

2304 leaq 16(%rcx),%rcx

2305 jnz .Loop_enc1_15

2306 .byte 102,15,56,221,209

2307 movl %r10d,%eax

2308 movq %r11,%rcx

2309 movups %xmm2,0(%rsi)

2310 leaq 16(%rsi),%rsi

2311 subq $16,%rdx

2312 jnc .Lcbc_enc_loop

2313 addq $16,%rdx

2314 jnz .Lcbc_enc_tail

2315 movups %xmm2,(%r8)

2316 jmp .Lcbc_ret

2317

2318 .Lcbc_enc_tail:

2319 movq %rdx,%rcx

2320 xchgq %rdi,%rsi

2321 .long 0x9066A4F3

2322 movl $16,%ecx

2323 subq %rdx,%rcx

2324 xorl %eax,%eax

2325 .long 0x9066AAF3

2326 leaq -16(%rdi),%rdi

2327 movl %r10d,%eax

2328 movq %rdi,%rsi

2329 movq %r11,%rcx

2330 xorq %rdx,%rdx

2331 jmp .Lcbc_enc_loop

2332

2333 .align 16

2334 .Lcbc_decrypt:

2335 leaq (%rsp),%rax

2336 pushq %rbp

2337 subq $16,%rsp

2338 andq $-16,%rsp

2339 leaq -8(%rax),%rbp

2340 movups (%r8),%xmm10

2341 movl %r10d,%eax

2342 cmpq $80,%rdx

2343 jbe .Lcbc_dec_tail

2344

2345 movups (%rcx),%xmm0

2346 movdqu 0(%rdi),%xmm2

2347 movdqu 16(%rdi),%xmm3

2348 movdqa %xmm2,%xmm11

2349 movdqu 32(%rdi),%xmm4

2350 movdqa %xmm3,%xmm12

2351 movdqu 48(%rdi),%xmm5

2352 movdqa %xmm4,%xmm13

2353 movdqu 64(%rdi),%xmm6

2354 movdqa %xmm5,%xmm14

2355 movdqu 80(%rdi),%xmm7

2356 movdqa %xmm6,%xmm15

2357 cmpq $112,%rdx

2358 jbe .Lcbc_dec_six_or_seven

2359

2360 subq $112,%rdx

2361 leaq 112(%rcx),%rcx

2362 jmp .Lcbc_dec_loop8_enter

2363 .align 16

2364 .Lcbc_dec_loop8:

2365 movups %xmm9,(%rsi)

2366 leaq 16(%rsi),%rsi

2367 .Lcbc_dec_loop8_enter:

2368 movdqu 96(%rdi),%xmm8

2369 pxor %xmm0,%xmm2

2370 movdqu 112(%rdi),%xmm9

2371 pxor %xmm0,%xmm3

2372 movups 16-112(%rcx),%xmm1

2373 pxor %xmm0,%xmm4

2374 xorq %r11,%r11

2375 cmpq $112,%rdx

2376 pxor %xmm0,%xmm5

2377 pxor %xmm0,%xmm6

2378 pxor %xmm0,%xmm7

2379 pxor %xmm0,%xmm8

2380

2381 .byte 102,15,56,222,209

2382 pxor %xmm0,%xmm9

2383 movups 32-112(%rcx),%xmm0

2384 .byte 102,15,56,222,217

2385 .byte 102,15,56,222,225

2386 .byte 102,15,56,222,233

2387 .byte 102,15,56,222,241

2388 .byte 102,15,56,222,249

2389 setnc %r11b

2390 .byte 102,68,15,56,222,193

2391 shlq $7,%r11

2392 .byte 102,68,15,56,222,201

2393 addq %rdi,%r11

2394 movups 48-112(%rcx),%xmm1

2395 .byte 102,15,56,222,208

2396 .byte 102,15,56,222,216

2397 .byte 102,15,56,222,224

2398 .byte 102,15,56,222,232

2399 .byte 102,15,56,222,240

2400 .byte 102,15,56,222,248

2401 .byte 102,68,15,56,222,192

2402 .byte 102,68,15,56,222,200

2403 movups 64-112(%rcx),%xmm0

2404 .byte 102,15,56,222,209

2405 .byte 102,15,56,222,217

2406 .byte 102,15,56,222,225

2407 .byte 102,15,56,222,233

2408 .byte 102,15,56,222,241

2409 .byte 102,15,56,222,249

2410 .byte 102,68,15,56,222,193

2411 .byte 102,68,15,56,222,201

2412 movups 80-112(%rcx),%xmm1

2413 .byte 102,15,56,222,208

2414 .byte 102,15,56,222,216

2415 .byte 102,15,56,222,224

2416 .byte 102,15,56,222,232

2417 .byte 102,15,56,222,240

2418 .byte 102,15,56,222,248

2419 .byte 102,68,15,56,222,192

2420 .byte 102,68,15,56,222,200

2421 movups 96-112(%rcx),%xmm0

2422 .byte 102,15,56,222,209

2423 .byte 102,15,56,222,217

2424 .byte 102,15,56,222,225

2425 .byte 102,15,56,222,233

2426 .byte 102,15,56,222,241

2427 .byte 102,15,56,222,249

2428 .byte 102,68,15,56,222,193

2429 .byte 102,68,15,56,222,201

2430 movups 112-112(%rcx),%xmm1

2431 .byte 102,15,56,222,208

2432 .byte 102,15,56,222,216

2433 .byte 102,15,56,222,224

2434 .byte 102,15,56,222,232

2435 .byte 102,15,56,222,240

2436 .byte 102,15,56,222,248

2437 .byte 102,68,15,56,222,192

2438 .byte 102,68,15,56,222,200

2439 movups 128-112(%rcx),%xmm0

2440 .byte 102,15,56,222,209

2441 .byte 102,15,56,222,217

2442 .byte 102,15,56,222,225

2443 .byte 102,15,56,222,233

2444 .byte 102,15,56,222,241

2445 .byte 102,15,56,222,249

2446 .byte 102,68,15,56,222,193

2447 .byte 102,68,15,56,222,201

2448 movups 144-112(%rcx),%xmm1

2449 .byte 102,15,56,222,208

2450 .byte 102,15,56,222,216

2451 .byte 102,15,56,222,224

2452 .byte 102,15,56,222,232

2453 .byte 102,15,56,222,240

2454 .byte 102,15,56,222,248

2455 .byte 102,68,15,56,222,192

2456 .byte 102,68,15,56,222,200

2457 movups 160-112(%rcx),%xmm0

2458 cmpl $11,%eax

2459 jb .Lcbc_dec_done

2460 .byte 102,15,56,222,209

2461 .byte 102,15,56,222,217

2462 .byte 102,15,56,222,225

2463 .byte 102,15,56,222,233

2464 .byte 102,15,56,222,241

2465 .byte 102,15,56,222,249

2466 .byte 102,68,15,56,222,193

2467 .byte 102,68,15,56,222,201

2468 movups 176-112(%rcx),%xmm1

2469 .byte 102,15,56,222,208

2470 .byte 102,15,56,222,216

2471 .byte 102,15,56,222,224

2472 .byte 102,15,56,222,232

2473 .byte 102,15,56,222,240

2474 .byte 102,15,56,222,248

2475 .byte 102,68,15,56,222,192

2476 .byte 102,68,15,56,222,200

2477 movups 192-112(%rcx),%xmm0

2478 je .Lcbc_dec_done

2479 .byte 102,15,56,222,209

2480 .byte 102,15,56,222,217

2481 .byte 102,15,56,222,225

2482 .byte 102,15,56,222,233

2483 .byte 102,15,56,222,241

2484 .byte 102,15,56,222,249

2485 .byte 102,68,15,56,222,193

2486 .byte 102,68,15,56,222,201

2487 movups 208-112(%rcx),%xmm1

2488 .byte 102,15,56,222,208

2489 .byte 102,15,56,222,216

2490 .byte 102,15,56,222,224

2491 .byte 102,15,56,222,232

2492 .byte 102,15,56,222,240

2493 .byte 102,15,56,222,248

2494 .byte 102,68,15,56,222,192

2495 .byte 102,68,15,56,222,200

2496 movups 224-112(%rcx),%xmm0

2497 .Lcbc_dec_done:

2498 .byte 102,15,56,222,209

2499 pxor %xmm0,%xmm10

2500 .byte 102,15,56,222,217

2501 pxor %xmm0,%xmm11

2502 .byte 102,15,56,222,225

2503 pxor %xmm0,%xmm12

2504 .byte 102,15,56,222,233

2505 pxor %xmm0,%xmm13

2506 .byte 102,15,56,222,241

2507 pxor %xmm0,%xmm14

2508 .byte 102,15,56,222,249

2509 pxor %xmm0,%xmm15

2510 .byte 102,68,15,56,222,193

2511 .byte 102,68,15,56,222,201

2512 movdqu 80(%rdi),%xmm1

2513

2514 .byte 102,65,15,56,223,210

2515 movdqu 96(%rdi),%xmm10

2516 pxor %xmm0,%xmm1

2517 .byte 102,65,15,56,223,219

2518 pxor %xmm0,%xmm10

2519 movdqu 112(%rdi),%xmm0

2520 leaq 128(%rdi),%rdi

2521 .byte 102,65,15,56,223,228

2522 movdqu 0(%r11),%xmm11

2523 .byte 102,65,15,56,223,237

2524 movdqu 16(%r11),%xmm12

2525 .byte 102,65,15,56,223,246

2526 movdqu 32(%r11),%xmm13

2527 .byte 102,65,15,56,223,255

2528 movdqu 48(%r11),%xmm14

2529 .byte 102,68,15,56,223,193

2530 movdqu 64(%r11),%xmm15

2531 .byte 102,69,15,56,223,202

2532 movdqa %xmm0,%xmm10

2533 movdqu 80(%r11),%xmm1

2534 movups -112(%rcx),%xmm0

2535

2536 movups %xmm2,(%rsi)

2537 movdqa %xmm11,%xmm2

2538 movups %xmm3,16(%rsi)

2539 movdqa %xmm12,%xmm3

2540 movups %xmm4,32(%rsi)

2541 movdqa %xmm13,%xmm4

2542 movups %xmm5,48(%rsi)

2543 movdqa %xmm14,%xmm5

2544 movups %xmm6,64(%rsi)

2545 movdqa %xmm15,%xmm6

2546 movups %xmm7,80(%rsi)

2547 movdqa %xmm1,%xmm7

2548 movups %xmm8,96(%rsi)

2549 leaq 112(%rsi),%rsi

2550

2551 subq $128,%rdx

2552 ja .Lcbc_dec_loop8

2553

2554 movaps %xmm9,%xmm2

2555 leaq -112(%rcx),%rcx

2556 addq $112,%rdx

2557 jle .Lcbc_dec_tail_collected

2558 movups %xmm9,(%rsi)

2559 leaq 16(%rsi),%rsi

2560 cmpq $80,%rdx

2561 jbe .Lcbc_dec_tail

2562

2563 movaps %xmm11,%xmm2

2564 .Lcbc_dec_six_or_seven:

2565 cmpq $96,%rdx

2566 ja .Lcbc_dec_seven

2567

2568 movaps %xmm7,%xmm8

2569 call _aesni_decrypt6

2570 pxor %xmm10,%xmm2

2571 movaps %xmm8,%xmm10

2572 pxor %xmm11,%xmm3

2573 movdqu %xmm2,(%rsi)

2574 pxor %xmm12,%xmm4

2575 movdqu %xmm3,16(%rsi)

2576 pxor %xmm13,%xmm5

2577 movdqu %xmm4,32(%rsi)

2578 pxor %xmm14,%xmm6

2579 movdqu %xmm5,48(%rsi)

2580 pxor %xmm15,%xmm7

2581 movdqu %xmm6,64(%rsi)

2582 leaq 80(%rsi),%rsi

2583 movdqa %xmm7,%xmm2

2584 jmp .Lcbc_dec_tail_collected

2585

2586 .align 16

2587 .Lcbc_dec_seven:

2588 movups 96(%rdi),%xmm8

2589 xorps %xmm9,%xmm9

2590 call _aesni_decrypt8

2591 movups 80(%rdi),%xmm9

2592 pxor %xmm10,%xmm2

2593 movups 96(%rdi),%xmm10

2594 pxor %xmm11,%xmm3

2595 movdqu %xmm2,(%rsi)

2596 pxor %xmm12,%xmm4

2597 movdqu %xmm3,16(%rsi)

2598 pxor %xmm13,%xmm5

2599 movdqu %xmm4,32(%rsi)

2600 pxor %xmm14,%xmm6

2601 movdqu %xmm5,48(%rsi)

2602 pxor %xmm15,%xmm7

2603 movdqu %xmm6,64(%rsi)

2604 pxor %xmm9,%xmm8

2605 movdqu %xmm7,80(%rsi)

2606 leaq 96(%rsi),%rsi

2607 movdqa %xmm8,%xmm2

2608 jmp .Lcbc_dec_tail_collected

2609

2610 .Lcbc_dec_tail:

2611 movups (%rdi),%xmm2

2612 subq $16,%rdx

2613 jbe .Lcbc_dec_one

2614

2615 movups 16(%rdi),%xmm3

2616 movaps %xmm2,%xmm11

2617 subq $16,%rdx

2618 jbe .Lcbc_dec_two

2619

2620 movups 32(%rdi),%xmm4

2621 movaps %xmm3,%xmm12

2622 subq $16,%rdx

2623 jbe .Lcbc_dec_three

2624

2625 movups 48(%rdi),%xmm5

2626 movaps %xmm4,%xmm13

2627 subq $16,%rdx

2628 jbe .Lcbc_dec_four

2629

2630 movups 64(%rdi),%xmm6

2631 movaps %xmm5,%xmm14

2632 movaps %xmm6,%xmm15

2633 xorps %xmm7,%xmm7

2634 call _aesni_decrypt6

2635 pxor %xmm10,%xmm2

2636 movaps %xmm15,%xmm10

2637 pxor %xmm11,%xmm3

2638 movdqu %xmm2,(%rsi)

2639 pxor %xmm12,%xmm4

2640 movdqu %xmm3,16(%rsi)

2641 pxor %xmm13,%xmm5

2642 movdqu %xmm4,32(%rsi)

2643 pxor %xmm14,%xmm6

2644 movdqu %xmm5,48(%rsi)

2645 leaq 64(%rsi),%rsi

2646 movdqa %xmm6,%xmm2

2647 subq $16,%rdx

2648 jmp .Lcbc_dec_tail_collected

2649

2650 .align 16

2651 .Lcbc_dec_one:

2652 movaps %xmm2,%xmm11

2653 movups (%rcx),%xmm0

2654 movups 16(%rcx),%xmm1

2655 leaq 32(%rcx),%rcx

2656 xorps %xmm0,%xmm2

2657 .Loop_dec1_16:

2658 .byte 102,15,56,222,209

2659 decl %eax

2660 movups (%rcx),%xmm1

2661 leaq 16(%rcx),%rcx

2662 jnz .Loop_dec1_16

2663 .byte 102,15,56,223,209

2664 xorps %xmm10,%xmm2

2665 movaps %xmm11,%xmm10

2666 jmp .Lcbc_dec_tail_collected

2667 .align 16

2668 .Lcbc_dec_two:

2669 movaps %xmm3,%xmm12

2670 xorps %xmm4,%xmm4

2671 call _aesni_decrypt3

2672 pxor %xmm10,%xmm2

2673 movaps %xmm12,%xmm10

2674 pxor %xmm11,%xmm3

2675 movdqu %xmm2,(%rsi)

2676 movdqa %xmm3,%xmm2

2677 leaq 16(%rsi),%rsi

2678 jmp .Lcbc_dec_tail_collected

2679 .align 16

2680 .Lcbc_dec_three:

2681 movaps %xmm4,%xmm13

2682 call _aesni_decrypt3

2683 pxor %xmm10,%xmm2

2684 movaps %xmm13,%xmm10

2685 pxor %xmm11,%xmm3

2686 movdqu %xmm2,(%rsi)

2687 pxor %xmm12,%xmm4

2688 movdqu %xmm3,16(%rsi)

2689 movdqa %xmm4,%xmm2

2690 leaq 32(%rsi),%rsi

2691 jmp .Lcbc_dec_tail_collected

2692 .align 16

2693 .Lcbc_dec_four:

2694 movaps %xmm5,%xmm14

2695 call _aesni_decrypt4

2696 pxor %xmm10,%xmm2

2697 movaps %xmm14,%xmm10

2698 pxor %xmm11,%xmm3

2699 movdqu %xmm2,(%rsi)

2700 pxor %xmm12,%xmm4

2701 movdqu %xmm3,16(%rsi)

2702 pxor %xmm13,%xmm5

2703 movdqu %xmm4,32(%rsi)

2704 movdqa %xmm5,%xmm2

2705 leaq 48(%rsi),%rsi

2706 jmp .Lcbc_dec_tail_collected

2707

2708 .align 16

2709 .Lcbc_dec_tail_collected:

2710 movups %xmm10,(%r8)

2711 andq $15,%rdx

2712 jnz .Lcbc_dec_tail_partial

2713 movups %xmm2,(%rsi)

2714 jmp .Lcbc_dec_ret

2715 .align 16

2716 .Lcbc_dec_tail_partial:

2717 movaps %xmm2,(%rsp)

2718 movq $16,%rcx

2719 movq %rsi,%rdi

2720 subq %rdx,%rcx

2721 leaq (%rsp),%rsi

2722 .long 0x9066A4F3

2723

2724 .Lcbc_dec_ret:

2725 leaq (%rbp),%rsp

2726 popq %rbp

2727 .Lcbc_ret:

2728 .byte 0xf3,0xc3

2729 .size aesni_cbc_encrypt,.-aesni_cbc_encrypt

2730 .globl aesni_set_decrypt_key

2731 .type aesni_set_decrypt_key,@function

2732 .align 16

2733 aesni_set_decrypt_key:

2734 .byte 0x48,0x83,0xEC,0x08

2735 call __aesni_set_encrypt_key

2736 shll $4,%esi

2737 testl %eax,%eax

2738 jnz .Ldec_key_ret

2739 leaq 16(%rdx,%rsi,1),%rdi

2740

2741 movups (%rdx),%xmm0

2742 movups (%rdi),%xmm1

2743 movups %xmm0,(%rdi)

2744 movups %xmm1,(%rdx)

2745 leaq 16(%rdx),%rdx

2746 leaq -16(%rdi),%rdi

2747

2748 .Ldec_key_inverse:

2749 movups (%rdx),%xmm0

2750 movups (%rdi),%xmm1

2751 .byte 102,15,56,219,192

2752 .byte 102,15,56,219,201

2753 leaq 16(%rdx),%rdx

2754 leaq -16(%rdi),%rdi

2755 movups %xmm0,16(%rdi)

2756 movups %xmm1,-16(%rdx)

2757 cmpq %rdx,%rdi

2758 ja .Ldec_key_inverse

2759

2760 movups (%rdx),%xmm0

2761 .byte 102,15,56,219,192

2762 movups %xmm0,(%rdi)

2763 .Ldec_key_ret:

2764 addq $8,%rsp

2765 .byte 0xf3,0xc3

2766 .LSEH_end_set_decrypt_key:

2767 .size aesni_set_decrypt_key,.-aesni_set_decrypt_key

2768 .globl aesni_set_encrypt_key

2769 .type aesni_set_encrypt_key,@function

2770 .align 16

2771 aesni_set_encrypt_key:

2772 __aesni_set_encrypt_key:

2773 .byte 0x48,0x83,0xEC,0x08

2774 movq $-1,%rax

2775 testq %rdi,%rdi

2776 jz .Lenc_key_ret

2777 testq %rdx,%rdx

2778 jz .Lenc_key_ret

2779

2780 movups (%rdi),%xmm0

2781 xorps %xmm4,%xmm4

2782 leaq 16(%rdx),%rax

2783 cmpl $256,%esi

2784 je .L14rounds

2785 cmpl $192,%esi

2786 je .L12rounds

2787 cmpl $128,%esi

2788 jne .Lbad_keybits

2789

2790 .L10rounds:

2791 movl $9,%esi

2792 movups %xmm0,(%rdx)

2793 .byte 102,15,58,223,200,1

2794 call .Lkey_expansion_128_cold

2795 .byte 102,15,58,223,200,2

2796 call .Lkey_expansion_128

2797 .byte 102,15,58,223,200,4

2798 call .Lkey_expansion_128

2799 .byte 102,15,58,223,200,8

2800 call .Lkey_expansion_128

2801 .byte 102,15,58,223,200,16

2802 call .Lkey_expansion_128

2803 .byte 102,15,58,223,200,32

2804 call .Lkey_expansion_128

2805 .byte 102,15,58,223,200,64

2806 call .Lkey_expansion_128

2807 .byte 102,15,58,223,200,128

2808 call .Lkey_expansion_128

2809 .byte 102,15,58,223,200,27

2810 call .Lkey_expansion_128

2811 .byte 102,15,58,223,200,54

2812 call .Lkey_expansion_128

2813 movups %xmm0,(%rax)

2814 movl %esi,80(%rax)

2815 xorl %eax,%eax

2816 jmp .Lenc_key_ret

2817

2818 .align 16

2819 .L12rounds:

2820 movq 16(%rdi),%xmm2

2821 movl $11,%esi

2822 movups %xmm0,(%rdx)

2823 .byte 102,15,58,223,202,1

2824 call .Lkey_expansion_192a_cold

2825 .byte 102,15,58,223,202,2

2826 call .Lkey_expansion_192b

2827 .byte 102,15,58,223,202,4

2828 call .Lkey_expansion_192a

2829 .byte 102,15,58,223,202,8

2830 call .Lkey_expansion_192b

2831 .byte 102,15,58,223,202,16

2832 call .Lkey_expansion_192a

2833 .byte 102,15,58,223,202,32

2834 call .Lkey_expansion_192b

2835 .byte 102,15,58,223,202,64

2836 call .Lkey_expansion_192a

2837 .byte 102,15,58,223,202,128

2838 call .Lkey_expansion_192b

2839 movups %xmm0,(%rax)

2840 movl %esi,48(%rax)

2841 xorq %rax,%rax

2842 jmp .Lenc_key_ret

2843

2844 .align 16

2845 .L14rounds:

2846 movups 16(%rdi),%xmm2

2847 movl $13,%esi

2848 leaq 16(%rax),%rax

2849 movups %xmm0,(%rdx)

2850 movups %xmm2,16(%rdx)

2851 .byte 102,15,58,223,202,1

2852 call .Lkey_expansion_256a_cold

2853 .byte 102,15,58,223,200,1

2854 call .Lkey_expansion_256b

2855 .byte 102,15,58,223,202,2

2856 call .Lkey_expansion_256a

2857 .byte 102,15,58,223,200,2

2858 call .Lkey_expansion_256b

2859 .byte 102,15,58,223,202,4

2860 call .Lkey_expansion_256a

2861 .byte 102,15,58,223,200,4

2862 call .Lkey_expansion_256b

2863 .byte 102,15,58,223,202,8

2864 call .Lkey_expansion_256a

2865 .byte 102,15,58,223,200,8

2866 call .Lkey_expansion_256b

2867 .byte 102,15,58,223,202,16

2868 call .Lkey_expansion_256a

2869 .byte 102,15,58,223,200,16

2870 call .Lkey_expansion_256b

2871 .byte 102,15,58,223,202,32

2872 call .Lkey_expansion_256a

2873 .byte 102,15,58,223,200,32

2874 call .Lkey_expansion_256b

2875 .byte 102,15,58,223,202,64

2876 call .Lkey_expansion_256a

2877 movups %xmm0,(%rax)

2878 movl %esi,16(%rax)

2879 xorq %rax,%rax

2880 jmp .Lenc_key_ret

2881

2882 .align 16

2883 .Lbad_keybits:

2884 movq $-2,%rax

2885 .Lenc_key_ret:

2886 addq $8,%rsp

2887 .byte 0xf3,0xc3

2888 .LSEH_end_set_encrypt_key:

2889

2890 .align 16

2891 .Lkey_expansion_128:

2892 movups %xmm0,(%rax)

2893 leaq 16(%rax),%rax

2894 .Lkey_expansion_128_cold:

2895 shufps $16,%xmm0,%xmm4

2896 xorps %xmm4,%xmm0

2897 shufps $140,%xmm0,%xmm4

2898 xorps %xmm4,%xmm0

2899 shufps $255,%xmm1,%xmm1

2900 xorps %xmm1,%xmm0

2901 .byte 0xf3,0xc3

2902

2903 .align 16

2904 .Lkey_expansion_192a:

2905 movups %xmm0,(%rax)

2906 leaq 16(%rax),%rax

2907 .Lkey_expansion_192a_cold:

2908 movaps %xmm2,%xmm5

2909 .Lkey_expansion_192b_warm:

2910 shufps $16,%xmm0,%xmm4

2911 movdqa %xmm2,%xmm3

2912 xorps %xmm4,%xmm0

2913 shufps $140,%xmm0,%xmm4

2914 pslldq $4,%xmm3

2915 xorps %xmm4,%xmm0

2916 pshufd $85,%xmm1,%xmm1

2917 pxor %xmm3,%xmm2

2918 pxor %xmm1,%xmm0

2919 pshufd $255,%xmm0,%xmm3

2920 pxor %xmm3,%xmm2

2921 .byte 0xf3,0xc3

2922

2923 .align 16

2924 .Lkey_expansion_192b: