OLD | NEW |
(Empty) | |
| 1 .file "crypto/aes/asm/aesni-x86.s" |
| 2 .text |
| 3 .globl _aesni_encrypt |
| 4 .align 4 |
| 5 _aesni_encrypt: |
| 6 L_aesni_encrypt_begin: |
| 7 movl 4(%esp),%eax |
| 8 movl 12(%esp),%edx |
| 9 movups (%eax),%xmm2 |
| 10 movl 240(%edx),%ecx |
| 11 movl 8(%esp),%eax |
| 12 movups (%edx),%xmm0 |
| 13 movups 16(%edx),%xmm1 |
| 14 leal 32(%edx),%edx |
| 15 xorps %xmm0,%xmm2 |
| 16 L000enc1_loop_1: |
| 17 .byte 102,15,56,220,209 |
| 18 decl %ecx |
| 19 movups (%edx),%xmm1 |
| 20 leal 16(%edx),%edx |
| 21 jnz L000enc1_loop_1 |
| 22 .byte 102,15,56,221,209 |
| 23 movups %xmm2,(%eax) |
| 24 ret |
| 25 .globl _aesni_decrypt |
| 26 .align 4 |
| 27 _aesni_decrypt: |
| 28 L_aesni_decrypt_begin: |
| 29 movl 4(%esp),%eax |
| 30 movl 12(%esp),%edx |
| 31 movups (%eax),%xmm2 |
| 32 movl 240(%edx),%ecx |
| 33 movl 8(%esp),%eax |
| 34 movups (%edx),%xmm0 |
| 35 movups 16(%edx),%xmm1 |
| 36 leal 32(%edx),%edx |
| 37 xorps %xmm0,%xmm2 |
| 38 L001dec1_loop_2: |
| 39 .byte 102,15,56,222,209 |
| 40 decl %ecx |
| 41 movups (%edx),%xmm1 |
| 42 leal 16(%edx),%edx |
| 43 jnz L001dec1_loop_2 |
| 44 .byte 102,15,56,223,209 |
| 45 movups %xmm2,(%eax) |
| 46 ret |
| 47 .align 4 |
| 48 __aesni_encrypt3: |
| 49 movups (%edx),%xmm0 |
| 50 shrl $1,%ecx |
| 51 movups 16(%edx),%xmm1 |
| 52 leal 32(%edx),%edx |
| 53 xorps %xmm0,%xmm2 |
| 54 pxor %xmm0,%xmm3 |
| 55 pxor %xmm0,%xmm4 |
| 56 movups (%edx),%xmm0 |
| 57 L002enc3_loop: |
| 58 .byte 102,15,56,220,209 |
| 59 .byte 102,15,56,220,217 |
| 60 decl %ecx |
| 61 .byte 102,15,56,220,225 |
| 62 movups 16(%edx),%xmm1 |
| 63 .byte 102,15,56,220,208 |
| 64 .byte 102,15,56,220,216 |
| 65 leal 32(%edx),%edx |
| 66 .byte 102,15,56,220,224 |
| 67 movups (%edx),%xmm0 |
| 68 jnz L002enc3_loop |
| 69 .byte 102,15,56,220,209 |
| 70 .byte 102,15,56,220,217 |
| 71 .byte 102,15,56,220,225 |
| 72 .byte 102,15,56,221,208 |
| 73 .byte 102,15,56,221,216 |
| 74 .byte 102,15,56,221,224 |
| 75 ret |
| 76 .align 4 |
| 77 __aesni_decrypt3: |
| 78 movups (%edx),%xmm0 |
| 79 shrl $1,%ecx |
| 80 movups 16(%edx),%xmm1 |
| 81 leal 32(%edx),%edx |
| 82 xorps %xmm0,%xmm2 |
| 83 pxor %xmm0,%xmm3 |
| 84 pxor %xmm0,%xmm4 |
| 85 movups (%edx),%xmm0 |
| 86 L003dec3_loop: |
| 87 .byte 102,15,56,222,209 |
| 88 .byte 102,15,56,222,217 |
| 89 decl %ecx |
| 90 .byte 102,15,56,222,225 |
| 91 movups 16(%edx),%xmm1 |
| 92 .byte 102,15,56,222,208 |
| 93 .byte 102,15,56,222,216 |
| 94 leal 32(%edx),%edx |
| 95 .byte 102,15,56,222,224 |
| 96 movups (%edx),%xmm0 |
| 97 jnz L003dec3_loop |
| 98 .byte 102,15,56,222,209 |
| 99 .byte 102,15,56,222,217 |
| 100 .byte 102,15,56,222,225 |
| 101 .byte 102,15,56,223,208 |
| 102 .byte 102,15,56,223,216 |
| 103 .byte 102,15,56,223,224 |
| 104 ret |
| 105 .align 4 |
| 106 __aesni_encrypt4: |
| 107 movups (%edx),%xmm0 |
| 108 movups 16(%edx),%xmm1 |
| 109 shrl $1,%ecx |
| 110 leal 32(%edx),%edx |
| 111 xorps %xmm0,%xmm2 |
| 112 pxor %xmm0,%xmm3 |
| 113 pxor %xmm0,%xmm4 |
| 114 pxor %xmm0,%xmm5 |
| 115 movups (%edx),%xmm0 |
| 116 L004enc4_loop: |
| 117 .byte 102,15,56,220,209 |
| 118 .byte 102,15,56,220,217 |
| 119 decl %ecx |
| 120 .byte 102,15,56,220,225 |
| 121 .byte 102,15,56,220,233 |
| 122 movups 16(%edx),%xmm1 |
| 123 .byte 102,15,56,220,208 |
| 124 .byte 102,15,56,220,216 |
| 125 leal 32(%edx),%edx |
| 126 .byte 102,15,56,220,224 |
| 127 .byte 102,15,56,220,232 |
| 128 movups (%edx),%xmm0 |
| 129 jnz L004enc4_loop |
| 130 .byte 102,15,56,220,209 |
| 131 .byte 102,15,56,220,217 |
| 132 .byte 102,15,56,220,225 |
| 133 .byte 102,15,56,220,233 |
| 134 .byte 102,15,56,221,208 |
| 135 .byte 102,15,56,221,216 |
| 136 .byte 102,15,56,221,224 |
| 137 .byte 102,15,56,221,232 |
| 138 ret |
| 139 .align 4 |
| 140 __aesni_decrypt4: |
| 141 movups (%edx),%xmm0 |
| 142 movups 16(%edx),%xmm1 |
| 143 shrl $1,%ecx |
| 144 leal 32(%edx),%edx |
| 145 xorps %xmm0,%xmm2 |
| 146 pxor %xmm0,%xmm3 |
| 147 pxor %xmm0,%xmm4 |
| 148 pxor %xmm0,%xmm5 |
| 149 movups (%edx),%xmm0 |
| 150 L005dec4_loop: |
| 151 .byte 102,15,56,222,209 |
| 152 .byte 102,15,56,222,217 |
| 153 decl %ecx |
| 154 .byte 102,15,56,222,225 |
| 155 .byte 102,15,56,222,233 |
| 156 movups 16(%edx),%xmm1 |
| 157 .byte 102,15,56,222,208 |
| 158 .byte 102,15,56,222,216 |
| 159 leal 32(%edx),%edx |
| 160 .byte 102,15,56,222,224 |
| 161 .byte 102,15,56,222,232 |
| 162 movups (%edx),%xmm0 |
| 163 jnz L005dec4_loop |
| 164 .byte 102,15,56,222,209 |
| 165 .byte 102,15,56,222,217 |
| 166 .byte 102,15,56,222,225 |
| 167 .byte 102,15,56,222,233 |
| 168 .byte 102,15,56,223,208 |
| 169 .byte 102,15,56,223,216 |
| 170 .byte 102,15,56,223,224 |
| 171 .byte 102,15,56,223,232 |
| 172 ret |
| 173 .align 4 |
| 174 __aesni_encrypt6: |
| 175 movups (%edx),%xmm0 |
| 176 shrl $1,%ecx |
| 177 movups 16(%edx),%xmm1 |
| 178 leal 32(%edx),%edx |
| 179 xorps %xmm0,%xmm2 |
| 180 pxor %xmm0,%xmm3 |
| 181 .byte 102,15,56,220,209 |
| 182 pxor %xmm0,%xmm4 |
| 183 .byte 102,15,56,220,217 |
| 184 pxor %xmm0,%xmm5 |
| 185 decl %ecx |
| 186 .byte 102,15,56,220,225 |
| 187 pxor %xmm0,%xmm6 |
| 188 .byte 102,15,56,220,233 |
| 189 pxor %xmm0,%xmm7 |
| 190 .byte 102,15,56,220,241 |
| 191 movups (%edx),%xmm0 |
| 192 .byte 102,15,56,220,249 |
| 193 jmp L_aesni_encrypt6_enter |
| 194 .align 4,0x90 |
| 195 L006enc6_loop: |
| 196 .byte 102,15,56,220,209 |
| 197 .byte 102,15,56,220,217 |
| 198 decl %ecx |
| 199 .byte 102,15,56,220,225 |
| 200 .byte 102,15,56,220,233 |
| 201 .byte 102,15,56,220,241 |
| 202 .byte 102,15,56,220,249 |
| 203 .align 4,0x90 |
| 204 L_aesni_encrypt6_enter: |
| 205 movups 16(%edx),%xmm1 |
| 206 .byte 102,15,56,220,208 |
| 207 .byte 102,15,56,220,216 |
| 208 leal 32(%edx),%edx |
| 209 .byte 102,15,56,220,224 |
| 210 .byte 102,15,56,220,232 |
| 211 .byte 102,15,56,220,240 |
| 212 .byte 102,15,56,220,248 |
| 213 movups (%edx),%xmm0 |
| 214 jnz L006enc6_loop |
| 215 .byte 102,15,56,220,209 |
| 216 .byte 102,15,56,220,217 |
| 217 .byte 102,15,56,220,225 |
| 218 .byte 102,15,56,220,233 |
| 219 .byte 102,15,56,220,241 |
| 220 .byte 102,15,56,220,249 |
| 221 .byte 102,15,56,221,208 |
| 222 .byte 102,15,56,221,216 |
| 223 .byte 102,15,56,221,224 |
| 224 .byte 102,15,56,221,232 |
| 225 .byte 102,15,56,221,240 |
| 226 .byte 102,15,56,221,248 |
| 227 ret |
| 228 .align 4 |
| 229 __aesni_decrypt6: |
| 230 movups (%edx),%xmm0 |
| 231 shrl $1,%ecx |
| 232 movups 16(%edx),%xmm1 |
| 233 leal 32(%edx),%edx |
| 234 xorps %xmm0,%xmm2 |
| 235 pxor %xmm0,%xmm3 |
| 236 .byte 102,15,56,222,209 |
| 237 pxor %xmm0,%xmm4 |
| 238 .byte 102,15,56,222,217 |
| 239 pxor %xmm0,%xmm5 |
| 240 decl %ecx |
| 241 .byte 102,15,56,222,225 |
| 242 pxor %xmm0,%xmm6 |
| 243 .byte 102,15,56,222,233 |
| 244 pxor %xmm0,%xmm7 |
| 245 .byte 102,15,56,222,241 |
| 246 movups (%edx),%xmm0 |
| 247 .byte 102,15,56,222,249 |
| 248 jmp L_aesni_decrypt6_enter |
| 249 .align 4,0x90 |
| 250 L007dec6_loop: |
| 251 .byte 102,15,56,222,209 |
| 252 .byte 102,15,56,222,217 |
| 253 decl %ecx |
| 254 .byte 102,15,56,222,225 |
| 255 .byte 102,15,56,222,233 |
| 256 .byte 102,15,56,222,241 |
| 257 .byte 102,15,56,222,249 |
| 258 .align 4,0x90 |
| 259 L_aesni_decrypt6_enter: |
| 260 movups 16(%edx),%xmm1 |
| 261 .byte 102,15,56,222,208 |
| 262 .byte 102,15,56,222,216 |
| 263 leal 32(%edx),%edx |
| 264 .byte 102,15,56,222,224 |
| 265 .byte 102,15,56,222,232 |
| 266 .byte 102,15,56,222,240 |
| 267 .byte 102,15,56,222,248 |
| 268 movups (%edx),%xmm0 |
| 269 jnz L007dec6_loop |
| 270 .byte 102,15,56,222,209 |
| 271 .byte 102,15,56,222,217 |
| 272 .byte 102,15,56,222,225 |
| 273 .byte 102,15,56,222,233 |
| 274 .byte 102,15,56,222,241 |
| 275 .byte 102,15,56,222,249 |
| 276 .byte 102,15,56,223,208 |
| 277 .byte 102,15,56,223,216 |
| 278 .byte 102,15,56,223,224 |
| 279 .byte 102,15,56,223,232 |
| 280 .byte 102,15,56,223,240 |
| 281 .byte 102,15,56,223,248 |
| 282 ret |
| 283 .globl _aesni_ecb_encrypt |
| 284 .align 4 |
| 285 _aesni_ecb_encrypt: |
| 286 L_aesni_ecb_encrypt_begin: |
| 287 pushl %ebp |
| 288 pushl %ebx |
| 289 pushl %esi |
| 290 pushl %edi |
| 291 movl 20(%esp),%esi |
| 292 movl 24(%esp),%edi |
| 293 movl 28(%esp),%eax |
| 294 movl 32(%esp),%edx |
| 295 movl 36(%esp),%ebx |
| 296 andl $-16,%eax |
| 297 jz L008ecb_ret |
| 298 movl 240(%edx),%ecx |
| 299 testl %ebx,%ebx |
| 300 jz L009ecb_decrypt |
| 301 movl %edx,%ebp |
| 302 movl %ecx,%ebx |
| 303 cmpl $96,%eax |
| 304 jb L010ecb_enc_tail |
| 305 movdqu (%esi),%xmm2 |
| 306 movdqu 16(%esi),%xmm3 |
| 307 movdqu 32(%esi),%xmm4 |
| 308 movdqu 48(%esi),%xmm5 |
| 309 movdqu 64(%esi),%xmm6 |
| 310 movdqu 80(%esi),%xmm7 |
| 311 leal 96(%esi),%esi |
| 312 subl $96,%eax |
| 313 jmp L011ecb_enc_loop6_enter |
| 314 .align 4,0x90 |
| 315 L012ecb_enc_loop6: |
| 316 movups %xmm2,(%edi) |
| 317 movdqu (%esi),%xmm2 |
| 318 movups %xmm3,16(%edi) |
| 319 movdqu 16(%esi),%xmm3 |
| 320 movups %xmm4,32(%edi) |
| 321 movdqu 32(%esi),%xmm4 |
| 322 movups %xmm5,48(%edi) |
| 323 movdqu 48(%esi),%xmm5 |
| 324 movups %xmm6,64(%edi) |
| 325 movdqu 64(%esi),%xmm6 |
| 326 movups %xmm7,80(%edi) |
| 327 leal 96(%edi),%edi |
| 328 movdqu 80(%esi),%xmm7 |
| 329 leal 96(%esi),%esi |
| 330 L011ecb_enc_loop6_enter: |
| 331 call __aesni_encrypt6 |
| 332 movl %ebp,%edx |
| 333 movl %ebx,%ecx |
| 334 subl $96,%eax |
| 335 jnc L012ecb_enc_loop6 |
| 336 movups %xmm2,(%edi) |
| 337 movups %xmm3,16(%edi) |
| 338 movups %xmm4,32(%edi) |
| 339 movups %xmm5,48(%edi) |
| 340 movups %xmm6,64(%edi) |
| 341 movups %xmm7,80(%edi) |
| 342 leal 96(%edi),%edi |
| 343 addl $96,%eax |
| 344 jz L008ecb_ret |
| 345 L010ecb_enc_tail: |
| 346 movups (%esi),%xmm2 |
| 347 cmpl $32,%eax |
| 348 jb L013ecb_enc_one |
| 349 movups 16(%esi),%xmm3 |
| 350 je L014ecb_enc_two |
| 351 movups 32(%esi),%xmm4 |
| 352 cmpl $64,%eax |
| 353 jb L015ecb_enc_three |
| 354 movups 48(%esi),%xmm5 |
| 355 je L016ecb_enc_four |
| 356 movups 64(%esi),%xmm6 |
| 357 xorps %xmm7,%xmm7 |
| 358 call __aesni_encrypt6 |
| 359 movups %xmm2,(%edi) |
| 360 movups %xmm3,16(%edi) |
| 361 movups %xmm4,32(%edi) |
| 362 movups %xmm5,48(%edi) |
| 363 movups %xmm6,64(%edi) |
| 364 jmp L008ecb_ret |
| 365 .align 4,0x90 |
| 366 L013ecb_enc_one: |
| 367 movups (%edx),%xmm0 |
| 368 movups 16(%edx),%xmm1 |
| 369 leal 32(%edx),%edx |
| 370 xorps %xmm0,%xmm2 |
| 371 L017enc1_loop_3: |
| 372 .byte 102,15,56,220,209 |
| 373 decl %ecx |
| 374 movups (%edx),%xmm1 |
| 375 leal 16(%edx),%edx |
| 376 jnz L017enc1_loop_3 |
| 377 .byte 102,15,56,221,209 |
| 378 movups %xmm2,(%edi) |
| 379 jmp L008ecb_ret |
| 380 .align 4,0x90 |
| 381 L014ecb_enc_two: |
| 382 xorps %xmm4,%xmm4 |
| 383 call __aesni_encrypt3 |
| 384 movups %xmm2,(%edi) |
| 385 movups %xmm3,16(%edi) |
| 386 jmp L008ecb_ret |
| 387 .align 4,0x90 |
| 388 L015ecb_enc_three: |
| 389 call __aesni_encrypt3 |
| 390 movups %xmm2,(%edi) |
| 391 movups %xmm3,16(%edi) |
| 392 movups %xmm4,32(%edi) |
| 393 jmp L008ecb_ret |
| 394 .align 4,0x90 |
| 395 L016ecb_enc_four: |
| 396 call __aesni_encrypt4 |
| 397 movups %xmm2,(%edi) |
| 398 movups %xmm3,16(%edi) |
| 399 movups %xmm4,32(%edi) |
| 400 movups %xmm5,48(%edi) |
| 401 jmp L008ecb_ret |
| 402 .align 4,0x90 |
| 403 L009ecb_decrypt: |
| 404 movl %edx,%ebp |
| 405 movl %ecx,%ebx |
| 406 cmpl $96,%eax |
| 407 jb L018ecb_dec_tail |
| 408 movdqu (%esi),%xmm2 |
| 409 movdqu 16(%esi),%xmm3 |
| 410 movdqu 32(%esi),%xmm4 |
| 411 movdqu 48(%esi),%xmm5 |
| 412 movdqu 64(%esi),%xmm6 |
| 413 movdqu 80(%esi),%xmm7 |
| 414 leal 96(%esi),%esi |
| 415 subl $96,%eax |
| 416 jmp L019ecb_dec_loop6_enter |
| 417 .align 4,0x90 |
| 418 L020ecb_dec_loop6: |
| 419 movups %xmm2,(%edi) |
| 420 movdqu (%esi),%xmm2 |
| 421 movups %xmm3,16(%edi) |
| 422 movdqu 16(%esi),%xmm3 |
| 423 movups %xmm4,32(%edi) |
| 424 movdqu 32(%esi),%xmm4 |
| 425 movups %xmm5,48(%edi) |
| 426 movdqu 48(%esi),%xmm5 |
| 427 movups %xmm6,64(%edi) |
| 428 movdqu 64(%esi),%xmm6 |
| 429 movups %xmm7,80(%edi) |
| 430 leal 96(%edi),%edi |
| 431 movdqu 80(%esi),%xmm7 |
| 432 leal 96(%esi),%esi |
| 433 L019ecb_dec_loop6_enter: |
| 434 call __aesni_decrypt6 |
| 435 movl %ebp,%edx |
| 436 movl %ebx,%ecx |
| 437 subl $96,%eax |
| 438 jnc L020ecb_dec_loop6 |
| 439 movups %xmm2,(%edi) |
| 440 movups %xmm3,16(%edi) |
| 441 movups %xmm4,32(%edi) |
| 442 movups %xmm5,48(%edi) |
| 443 movups %xmm6,64(%edi) |
| 444 movups %xmm7,80(%edi) |
| 445 leal 96(%edi),%edi |
| 446 addl $96,%eax |
| 447 jz L008ecb_ret |
| 448 L018ecb_dec_tail: |
| 449 movups (%esi),%xmm2 |
| 450 cmpl $32,%eax |
| 451 jb L021ecb_dec_one |
| 452 movups 16(%esi),%xmm3 |
| 453 je L022ecb_dec_two |
| 454 movups 32(%esi),%xmm4 |
| 455 cmpl $64,%eax |
| 456 jb L023ecb_dec_three |
| 457 movups 48(%esi),%xmm5 |
| 458 je L024ecb_dec_four |
| 459 movups 64(%esi),%xmm6 |
| 460 xorps %xmm7,%xmm7 |
| 461 call __aesni_decrypt6 |
| 462 movups %xmm2,(%edi) |
| 463 movups %xmm3,16(%edi) |
| 464 movups %xmm4,32(%edi) |
| 465 movups %xmm5,48(%edi) |
| 466 movups %xmm6,64(%edi) |
| 467 jmp L008ecb_ret |
| 468 .align 4,0x90 |
| 469 L021ecb_dec_one: |
| 470 movups (%edx),%xmm0 |
| 471 movups 16(%edx),%xmm1 |
| 472 leal 32(%edx),%edx |
| 473 xorps %xmm0,%xmm2 |
| 474 L025dec1_loop_4: |
| 475 .byte 102,15,56,222,209 |
| 476 decl %ecx |
| 477 movups (%edx),%xmm1 |
| 478 leal 16(%edx),%edx |
| 479 jnz L025dec1_loop_4 |
| 480 .byte 102,15,56,223,209 |
| 481 movups %xmm2,(%edi) |
| 482 jmp L008ecb_ret |
| 483 .align 4,0x90 |
| 484 L022ecb_dec_two: |
| 485 xorps %xmm4,%xmm4 |
| 486 call __aesni_decrypt3 |
| 487 movups %xmm2,(%edi) |
| 488 movups %xmm3,16(%edi) |
| 489 jmp L008ecb_ret |
| 490 .align 4,0x90 |
| 491 L023ecb_dec_three: |
| 492 call __aesni_decrypt3 |
| 493 movups %xmm2,(%edi) |
| 494 movups %xmm3,16(%edi) |
| 495 movups %xmm4,32(%edi) |
| 496 jmp L008ecb_ret |
| 497 .align 4,0x90 |
| 498 L024ecb_dec_four: |
| 499 call __aesni_decrypt4 |
| 500 movups %xmm2,(%edi) |
| 501 movups %xmm3,16(%edi) |
| 502 movups %xmm4,32(%edi) |
| 503 movups %xmm5,48(%edi) |
| 504 L008ecb_ret: |
| 505 popl %edi |
| 506 popl %esi |
| 507 popl %ebx |
| 508 popl %ebp |
| 509 ret |
| 510 .globl _aesni_ccm64_encrypt_blocks |
| 511 .align 4 |
| 512 _aesni_ccm64_encrypt_blocks: |
| 513 L_aesni_ccm64_encrypt_blocks_begin: |
| 514 pushl %ebp |
| 515 pushl %ebx |
| 516 pushl %esi |
| 517 pushl %edi |
| 518 movl 20(%esp),%esi |
| 519 movl 24(%esp),%edi |
| 520 movl 28(%esp),%eax |
| 521 movl 32(%esp),%edx |
| 522 movl 36(%esp),%ebx |
| 523 movl 40(%esp),%ecx |
| 524 movl %esp,%ebp |
| 525 subl $60,%esp |
| 526 andl $-16,%esp |
| 527 movl %ebp,48(%esp) |
| 528 movdqu (%ebx),%xmm7 |
| 529 movdqu (%ecx),%xmm3 |
| 530 movl 240(%edx),%ecx |
| 531 movl $202182159,(%esp) |
| 532 movl $134810123,4(%esp) |
| 533 movl $67438087,8(%esp) |
| 534 movl $66051,12(%esp) |
| 535 movl $1,%ebx |
| 536 xorl %ebp,%ebp |
| 537 movl %ebx,16(%esp) |
| 538 movl %ebp,20(%esp) |
| 539 movl %ebp,24(%esp) |
| 540 movl %ebp,28(%esp) |
| 541 shrl $1,%ecx |
| 542 leal (%edx),%ebp |
| 543 movdqa (%esp),%xmm5 |
| 544 movdqa %xmm7,%xmm2 |
| 545 movl %ecx,%ebx |
| 546 .byte 102,15,56,0,253 |
| 547 L026ccm64_enc_outer: |
| 548 movups (%ebp),%xmm0 |
| 549 movl %ebx,%ecx |
| 550 movups (%esi),%xmm6 |
| 551 xorps %xmm0,%xmm2 |
| 552 movups 16(%ebp),%xmm1 |
| 553 xorps %xmm6,%xmm0 |
| 554 leal 32(%ebp),%edx |
| 555 xorps %xmm0,%xmm3 |
| 556 movups (%edx),%xmm0 |
| 557 L027ccm64_enc2_loop: |
| 558 .byte 102,15,56,220,209 |
| 559 decl %ecx |
| 560 .byte 102,15,56,220,217 |
| 561 movups 16(%edx),%xmm1 |
| 562 .byte 102,15,56,220,208 |
| 563 leal 32(%edx),%edx |
| 564 .byte 102,15,56,220,216 |
| 565 movups (%edx),%xmm0 |
| 566 jnz L027ccm64_enc2_loop |
| 567 .byte 102,15,56,220,209 |
| 568 .byte 102,15,56,220,217 |
| 569 paddq 16(%esp),%xmm7 |
| 570 .byte 102,15,56,221,208 |
| 571 .byte 102,15,56,221,216 |
| 572 decl %eax |
| 573 leal 16(%esi),%esi |
| 574 xorps %xmm2,%xmm6 |
| 575 movdqa %xmm7,%xmm2 |
| 576 movups %xmm6,(%edi) |
| 577 leal 16(%edi),%edi |
| 578 .byte 102,15,56,0,213 |
| 579 jnz L026ccm64_enc_outer |
| 580 movl 48(%esp),%esp |
| 581 movl 40(%esp),%edi |
| 582 movups %xmm3,(%edi) |
| 583 popl %edi |
| 584 popl %esi |
| 585 popl %ebx |
| 586 popl %ebp |
| 587 ret |
| 588 .globl _aesni_ccm64_decrypt_blocks |
| 589 .align 4 |
| 590 _aesni_ccm64_decrypt_blocks: |
| 591 L_aesni_ccm64_decrypt_blocks_begin: |
| 592 pushl %ebp |
| 593 pushl %ebx |
| 594 pushl %esi |
| 595 pushl %edi |
| 596 movl 20(%esp),%esi |
| 597 movl 24(%esp),%edi |
| 598 movl 28(%esp),%eax |
| 599 movl 32(%esp),%edx |
| 600 movl 36(%esp),%ebx |
| 601 movl 40(%esp),%ecx |
| 602 movl %esp,%ebp |
| 603 subl $60,%esp |
| 604 andl $-16,%esp |
| 605 movl %ebp,48(%esp) |
| 606 movdqu (%ebx),%xmm7 |
| 607 movdqu (%ecx),%xmm3 |
| 608 movl 240(%edx),%ecx |
| 609 movl $202182159,(%esp) |
| 610 movl $134810123,4(%esp) |
| 611 movl $67438087,8(%esp) |
| 612 movl $66051,12(%esp) |
| 613 movl $1,%ebx |
| 614 xorl %ebp,%ebp |
| 615 movl %ebx,16(%esp) |
| 616 movl %ebp,20(%esp) |
| 617 movl %ebp,24(%esp) |
| 618 movl %ebp,28(%esp) |
| 619 movdqa (%esp),%xmm5 |
| 620 movdqa %xmm7,%xmm2 |
| 621 movl %edx,%ebp |
| 622 movl %ecx,%ebx |
| 623 .byte 102,15,56,0,253 |
| 624 movups (%edx),%xmm0 |
| 625 movups 16(%edx),%xmm1 |
| 626 leal 32(%edx),%edx |
| 627 xorps %xmm0,%xmm2 |
| 628 L028enc1_loop_5: |
| 629 .byte 102,15,56,220,209 |
| 630 decl %ecx |
| 631 movups (%edx),%xmm1 |
| 632 leal 16(%edx),%edx |
| 633 jnz L028enc1_loop_5 |
| 634 .byte 102,15,56,221,209 |
| 635 movups (%esi),%xmm6 |
| 636 paddq 16(%esp),%xmm7 |
| 637 leal 16(%esi),%esi |
| 638 jmp L029ccm64_dec_outer |
| 639 .align 4,0x90 |
| 640 L029ccm64_dec_outer: |
| 641 xorps %xmm2,%xmm6 |
| 642 movdqa %xmm7,%xmm2 |
| 643 movl %ebx,%ecx |
| 644 movups %xmm6,(%edi) |
| 645 leal 16(%edi),%edi |
| 646 .byte 102,15,56,0,213 |
| 647 subl $1,%eax |
| 648 jz L030ccm64_dec_break |
| 649 movups (%ebp),%xmm0 |
| 650 shrl $1,%ecx |
| 651 movups 16(%ebp),%xmm1 |
| 652 xorps %xmm0,%xmm6 |
| 653 leal 32(%ebp),%edx |
| 654 xorps %xmm0,%xmm2 |
| 655 xorps %xmm6,%xmm3 |
| 656 movups (%edx),%xmm0 |
| 657 L031ccm64_dec2_loop: |
| 658 .byte 102,15,56,220,209 |
| 659 decl %ecx |
| 660 .byte 102,15,56,220,217 |
| 661 movups 16(%edx),%xmm1 |
| 662 .byte 102,15,56,220,208 |
| 663 leal 32(%edx),%edx |
| 664 .byte 102,15,56,220,216 |
| 665 movups (%edx),%xmm0 |
| 666 jnz L031ccm64_dec2_loop |
| 667 movups (%esi),%xmm6 |
| 668 paddq 16(%esp),%xmm7 |
| 669 .byte 102,15,56,220,209 |
| 670 .byte 102,15,56,220,217 |
| 671 leal 16(%esi),%esi |
| 672 .byte 102,15,56,221,208 |
| 673 .byte 102,15,56,221,216 |
| 674 jmp L029ccm64_dec_outer |
| 675 .align 4,0x90 |
| 676 L030ccm64_dec_break: |
| 677 movl %ebp,%edx |
| 678 movups (%edx),%xmm0 |
| 679 movups 16(%edx),%xmm1 |
| 680 xorps %xmm0,%xmm6 |
| 681 leal 32(%edx),%edx |
| 682 xorps %xmm6,%xmm3 |
| 683 L032enc1_loop_6: |
| 684 .byte 102,15,56,220,217 |
| 685 decl %ecx |
| 686 movups (%edx),%xmm1 |
| 687 leal 16(%edx),%edx |
| 688 jnz L032enc1_loop_6 |
| 689 .byte 102,15,56,221,217 |
| 690 movl 48(%esp),%esp |
| 691 movl 40(%esp),%edi |
| 692 movups %xmm3,(%edi) |
| 693 popl %edi |
| 694 popl %esi |
| 695 popl %ebx |
| 696 popl %ebp |
| 697 ret |
| 698 .globl _aesni_ctr32_encrypt_blocks |
| 699 .align 4 |
| 700 _aesni_ctr32_encrypt_blocks: |
| 701 L_aesni_ctr32_encrypt_blocks_begin: |
| 702 pushl %ebp |
| 703 pushl %ebx |
| 704 pushl %esi |
| 705 pushl %edi |
| 706 movl 20(%esp),%esi |
| 707 movl 24(%esp),%edi |
| 708 movl 28(%esp),%eax |
| 709 movl 32(%esp),%edx |
| 710 movl 36(%esp),%ebx |
| 711 movl %esp,%ebp |
| 712 subl $88,%esp |
| 713 andl $-16,%esp |
| 714 movl %ebp,80(%esp) |
| 715 cmpl $1,%eax |
| 716 je L033ctr32_one_shortcut |
| 717 movdqu (%ebx),%xmm7 |
| 718 movl $202182159,(%esp) |
| 719 movl $134810123,4(%esp) |
| 720 movl $67438087,8(%esp) |
| 721 movl $66051,12(%esp) |
| 722 movl $6,%ecx |
| 723 xorl %ebp,%ebp |
| 724 movl %ecx,16(%esp) |
| 725 movl %ecx,20(%esp) |
| 726 movl %ecx,24(%esp) |
| 727 movl %ebp,28(%esp) |
| 728 .byte 102,15,58,22,251,3 |
| 729 .byte 102,15,58,34,253,3 |
| 730 movl 240(%edx),%ecx |
| 731 bswap %ebx |
| 732 pxor %xmm1,%xmm1 |
| 733 pxor %xmm0,%xmm0 |
| 734 movdqa (%esp),%xmm2 |
| 735 .byte 102,15,58,34,203,0 |
| 736 leal 3(%ebx),%ebp |
| 737 .byte 102,15,58,34,197,0 |
| 738 incl %ebx |
| 739 .byte 102,15,58,34,203,1 |
| 740 incl %ebp |
| 741 .byte 102,15,58,34,197,1 |
| 742 incl %ebx |
| 743 .byte 102,15,58,34,203,2 |
| 744 incl %ebp |
| 745 .byte 102,15,58,34,197,2 |
| 746 movdqa %xmm1,48(%esp) |
| 747 .byte 102,15,56,0,202 |
| 748 movdqa %xmm0,64(%esp) |
| 749 .byte 102,15,56,0,194 |
| 750 pshufd $192,%xmm1,%xmm2 |
| 751 pshufd $128,%xmm1,%xmm3 |
| 752 cmpl $6,%eax |
| 753 jb L034ctr32_tail |
| 754 movdqa %xmm7,32(%esp) |
| 755 shrl $1,%ecx |
| 756 movl %edx,%ebp |
| 757 movl %ecx,%ebx |
| 758 subl $6,%eax |
| 759 jmp L035ctr32_loop6 |
| 760 .align 4,0x90 |
| 761 L035ctr32_loop6: |
| 762 pshufd $64,%xmm1,%xmm4 |
| 763 movdqa 32(%esp),%xmm1 |
| 764 pshufd $192,%xmm0,%xmm5 |
| 765 por %xmm1,%xmm2 |
| 766 pshufd $128,%xmm0,%xmm6 |
| 767 por %xmm1,%xmm3 |
| 768 pshufd $64,%xmm0,%xmm7 |
| 769 por %xmm1,%xmm4 |
| 770 por %xmm1,%xmm5 |
| 771 por %xmm1,%xmm6 |
| 772 por %xmm1,%xmm7 |
| 773 movups (%ebp),%xmm0 |
| 774 movups 16(%ebp),%xmm1 |
| 775 leal 32(%ebp),%edx |
| 776 decl %ecx |
| 777 pxor %xmm0,%xmm2 |
| 778 pxor %xmm0,%xmm3 |
| 779 .byte 102,15,56,220,209 |
| 780 pxor %xmm0,%xmm4 |
| 781 .byte 102,15,56,220,217 |
| 782 pxor %xmm0,%xmm5 |
| 783 .byte 102,15,56,220,225 |
| 784 pxor %xmm0,%xmm6 |
| 785 .byte 102,15,56,220,233 |
| 786 pxor %xmm0,%xmm7 |
| 787 .byte 102,15,56,220,241 |
| 788 movups (%edx),%xmm0 |
| 789 .byte 102,15,56,220,249 |
| 790 call L_aesni_encrypt6_enter |
| 791 movups (%esi),%xmm1 |
| 792 movups 16(%esi),%xmm0 |
| 793 xorps %xmm1,%xmm2 |
| 794 movups 32(%esi),%xmm1 |
| 795 xorps %xmm0,%xmm3 |
| 796 movups %xmm2,(%edi) |
| 797 movdqa 16(%esp),%xmm0 |
| 798 xorps %xmm1,%xmm4 |
| 799 movdqa 48(%esp),%xmm1 |
| 800 movups %xmm3,16(%edi) |
| 801 movups %xmm4,32(%edi) |
| 802 paddd %xmm0,%xmm1 |
| 803 paddd 64(%esp),%xmm0 |
| 804 movdqa (%esp),%xmm2 |
| 805 movups 48(%esi),%xmm3 |
| 806 movups 64(%esi),%xmm4 |
| 807 xorps %xmm3,%xmm5 |
| 808 movups 80(%esi),%xmm3 |
| 809 leal 96(%esi),%esi |
| 810 movdqa %xmm1,48(%esp) |
| 811 .byte 102,15,56,0,202 |
| 812 xorps %xmm4,%xmm6 |
| 813 movups %xmm5,48(%edi) |
| 814 xorps %xmm3,%xmm7 |
| 815 movdqa %xmm0,64(%esp) |
| 816 .byte 102,15,56,0,194 |
| 817 movups %xmm6,64(%edi) |
| 818 pshufd $192,%xmm1,%xmm2 |
| 819 movups %xmm7,80(%edi) |
| 820 leal 96(%edi),%edi |
| 821 movl %ebx,%ecx |
| 822 pshufd $128,%xmm1,%xmm3 |
| 823 subl $6,%eax |
| 824 jnc L035ctr32_loop6 |
| 825 addl $6,%eax |
| 826 jz L036ctr32_ret |
| 827 movl %ebp,%edx |
| 828 leal 1(,%ecx,2),%ecx |
| 829 movdqa 32(%esp),%xmm7 |
| 830 L034ctr32_tail: |
| 831 por %xmm7,%xmm2 |
| 832 cmpl $2,%eax |
| 833 jb L037ctr32_one |
| 834 pshufd $64,%xmm1,%xmm4 |
| 835 por %xmm7,%xmm3 |
| 836 je L038ctr32_two |
| 837 pshufd $192,%xmm0,%xmm5 |
| 838 por %xmm7,%xmm4 |
| 839 cmpl $4,%eax |
| 840 jb L039ctr32_three |
| 841 pshufd $128,%xmm0,%xmm6 |
| 842 por %xmm7,%xmm5 |
| 843 je L040ctr32_four |
| 844 por %xmm7,%xmm6 |
| 845 call __aesni_encrypt6 |
| 846 movups (%esi),%xmm1 |
| 847 movups 16(%esi),%xmm0 |
| 848 xorps %xmm1,%xmm2 |
| 849 movups 32(%esi),%xmm1 |
| 850 xorps %xmm0,%xmm3 |
| 851 movups 48(%esi),%xmm0 |
| 852 xorps %xmm1,%xmm4 |
| 853 movups 64(%esi),%xmm1 |
| 854 xorps %xmm0,%xmm5 |
| 855 movups %xmm2,(%edi) |
| 856 xorps %xmm1,%xmm6 |
| 857 movups %xmm3,16(%edi) |
| 858 movups %xmm4,32(%edi) |
| 859 movups %xmm5,48(%edi) |
| 860 movups %xmm6,64(%edi) |
| 861 jmp L036ctr32_ret |
| 862 .align 4,0x90 |
| 863 L033ctr32_one_shortcut: |
| 864 movups (%ebx),%xmm2 |
| 865 movl 240(%edx),%ecx |
| 866 L037ctr32_one: |
| 867 movups (%edx),%xmm0 |
| 868 movups 16(%edx),%xmm1 |
| 869 leal 32(%edx),%edx |
| 870 xorps %xmm0,%xmm2 |
| 871 L041enc1_loop_7: |
| 872 .byte 102,15,56,220,209 |
| 873 decl %ecx |
| 874 movups (%edx),%xmm1 |
| 875 leal 16(%edx),%edx |
| 876 jnz L041enc1_loop_7 |
| 877 .byte 102,15,56,221,209 |
| 878 movups (%esi),%xmm6 |
| 879 xorps %xmm2,%xmm6 |
| 880 movups %xmm6,(%edi) |
| 881 jmp L036ctr32_ret |
| 882 .align 4,0x90 |
| 883 L038ctr32_two: |
| 884 call __aesni_encrypt3 |
| 885 movups (%esi),%xmm5 |
| 886 movups 16(%esi),%xmm6 |
| 887 xorps %xmm5,%xmm2 |
| 888 xorps %xmm6,%xmm3 |
| 889 movups %xmm2,(%edi) |
| 890 movups %xmm3,16(%edi) |
| 891 jmp L036ctr32_ret |
| 892 .align 4,0x90 |
| 893 L039ctr32_three: |
| 894 call __aesni_encrypt3 |
| 895 movups (%esi),%xmm5 |
| 896 movups 16(%esi),%xmm6 |
| 897 xorps %xmm5,%xmm2 |
| 898 movups 32(%esi),%xmm7 |
| 899 xorps %xmm6,%xmm3 |
| 900 movups %xmm2,(%edi) |
| 901 xorps %xmm7,%xmm4 |
| 902 movups %xmm3,16(%edi) |
| 903 movups %xmm4,32(%edi) |
| 904 jmp L036ctr32_ret |
| 905 .align 4,0x90 |
| 906 L040ctr32_four: |
| 907 call __aesni_encrypt4 |
| 908 movups (%esi),%xmm6 |
| 909 movups 16(%esi),%xmm7 |
| 910 movups 32(%esi),%xmm1 |
| 911 xorps %xmm6,%xmm2 |
| 912 movups 48(%esi),%xmm0 |
| 913 xorps %xmm7,%xmm3 |
| 914 movups %xmm2,(%edi) |
| 915 xorps %xmm1,%xmm4 |
| 916 movups %xmm3,16(%edi) |
| 917 xorps %xmm0,%xmm5 |
| 918 movups %xmm4,32(%edi) |
| 919 movups %xmm5,48(%edi) |
| 920 L036ctr32_ret: |
| 921 movl 80(%esp),%esp |
| 922 popl %edi |
| 923 popl %esi |
| 924 popl %ebx |
| 925 popl %ebp |
| 926 ret |
| 927 .globl _aesni_xts_encrypt |
| 928 .align 4 |
| 929 _aesni_xts_encrypt: |
| 930 L_aesni_xts_encrypt_begin: |
| 931 pushl %ebp |
| 932 pushl %ebx |
| 933 pushl %esi |
| 934 pushl %edi |
| 935 movl 36(%esp),%edx |
| 936 movl 40(%esp),%esi |
| 937 movl 240(%edx),%ecx |
| 938 movups (%esi),%xmm2 |
| 939 movups (%edx),%xmm0 |
| 940 movups 16(%edx),%xmm1 |
| 941 leal 32(%edx),%edx |
| 942 xorps %xmm0,%xmm2 |
| 943 L042enc1_loop_8: |
| 944 .byte 102,15,56,220,209 |
| 945 decl %ecx |
| 946 movups (%edx),%xmm1 |
| 947 leal 16(%edx),%edx |
| 948 jnz L042enc1_loop_8 |
| 949 .byte 102,15,56,221,209 |
| 950 movl 20(%esp),%esi |
| 951 movl 24(%esp),%edi |
| 952 movl 28(%esp),%eax |
| 953 movl 32(%esp),%edx |
| 954 movl %esp,%ebp |
| 955 subl $120,%esp |
| 956 movl 240(%edx),%ecx |
| 957 andl $-16,%esp |
| 958 movl $135,96(%esp) |
| 959 movl $0,100(%esp) |
| 960 movl $1,104(%esp) |
| 961 movl $0,108(%esp) |
| 962 movl %eax,112(%esp) |
| 963 movl %ebp,116(%esp) |
| 964 movdqa %xmm2,%xmm1 |
| 965 pxor %xmm0,%xmm0 |
| 966 movdqa 96(%esp),%xmm3 |
| 967 pcmpgtd %xmm1,%xmm0 |
| 968 andl $-16,%eax |
| 969 movl %edx,%ebp |
| 970 movl %ecx,%ebx |
| 971 subl $96,%eax |
| 972 jc L043xts_enc_short |
| 973 shrl $1,%ecx |
| 974 movl %ecx,%ebx |
| 975 jmp L044xts_enc_loop6 |
| 976 .align 4,0x90 |
| 977 L044xts_enc_loop6: |
| 978 pshufd $19,%xmm0,%xmm2 |
| 979 pxor %xmm0,%xmm0 |
| 980 movdqa %xmm1,(%esp) |
| 981 paddq %xmm1,%xmm1 |
| 982 pand %xmm3,%xmm2 |
| 983 pcmpgtd %xmm1,%xmm0 |
| 984 pxor %xmm2,%xmm1 |
| 985 pshufd $19,%xmm0,%xmm2 |
| 986 pxor %xmm0,%xmm0 |
| 987 movdqa %xmm1,16(%esp) |
| 988 paddq %xmm1,%xmm1 |
| 989 pand %xmm3,%xmm2 |
| 990 pcmpgtd %xmm1,%xmm0 |
| 991 pxor %xmm2,%xmm1 |
| 992 pshufd $19,%xmm0,%xmm2 |
| 993 pxor %xmm0,%xmm0 |
| 994 movdqa %xmm1,32(%esp) |
| 995 paddq %xmm1,%xmm1 |
| 996 pand %xmm3,%xmm2 |
| 997 pcmpgtd %xmm1,%xmm0 |
| 998 pxor %xmm2,%xmm1 |
| 999 pshufd $19,%xmm0,%xmm2 |
| 1000 pxor %xmm0,%xmm0 |
| 1001 movdqa %xmm1,48(%esp) |
| 1002 paddq %xmm1,%xmm1 |
| 1003 pand %xmm3,%xmm2 |
| 1004 pcmpgtd %xmm1,%xmm0 |
| 1005 pxor %xmm2,%xmm1 |
| 1006 pshufd $19,%xmm0,%xmm7 |
| 1007 movdqa %xmm1,64(%esp) |
| 1008 paddq %xmm1,%xmm1 |
| 1009 movups (%ebp),%xmm0 |
| 1010 pand %xmm3,%xmm7 |
| 1011 movups (%esi),%xmm2 |
| 1012 pxor %xmm1,%xmm7 |
| 1013 movdqu 16(%esi),%xmm3 |
| 1014 xorps %xmm0,%xmm2 |
| 1015 movdqu 32(%esi),%xmm4 |
| 1016 pxor %xmm0,%xmm3 |
| 1017 movdqu 48(%esi),%xmm5 |
| 1018 pxor %xmm0,%xmm4 |
| 1019 movdqu 64(%esi),%xmm6 |
| 1020 pxor %xmm0,%xmm5 |
| 1021 movdqu 80(%esi),%xmm1 |
| 1022 pxor %xmm0,%xmm6 |
| 1023 leal 96(%esi),%esi |
| 1024 pxor (%esp),%xmm2 |
| 1025 movdqa %xmm7,80(%esp) |
| 1026 pxor %xmm1,%xmm7 |
| 1027 movups 16(%ebp),%xmm1 |
| 1028 leal 32(%ebp),%edx |
| 1029 pxor 16(%esp),%xmm3 |
| 1030 .byte 102,15,56,220,209 |
| 1031 pxor 32(%esp),%xmm4 |
| 1032 .byte 102,15,56,220,217 |
| 1033 pxor 48(%esp),%xmm5 |
| 1034 decl %ecx |
| 1035 .byte 102,15,56,220,225 |
| 1036 pxor 64(%esp),%xmm6 |
| 1037 .byte 102,15,56,220,233 |
| 1038 pxor %xmm0,%xmm7 |
| 1039 .byte 102,15,56,220,241 |
| 1040 movups (%edx),%xmm0 |
| 1041 .byte 102,15,56,220,249 |
| 1042 call L_aesni_encrypt6_enter |
| 1043 movdqa 80(%esp),%xmm1 |
| 1044 pxor %xmm0,%xmm0 |
| 1045 xorps (%esp),%xmm2 |
| 1046 pcmpgtd %xmm1,%xmm0 |
| 1047 xorps 16(%esp),%xmm3 |
| 1048 movups %xmm2,(%edi) |
| 1049 xorps 32(%esp),%xmm4 |
| 1050 movups %xmm3,16(%edi) |
| 1051 xorps 48(%esp),%xmm5 |
| 1052 movups %xmm4,32(%edi) |
| 1053 xorps 64(%esp),%xmm6 |
| 1054 movups %xmm5,48(%edi) |
| 1055 xorps %xmm1,%xmm7 |
| 1056 movups %xmm6,64(%edi) |
| 1057 pshufd $19,%xmm0,%xmm2 |
| 1058 movups %xmm7,80(%edi) |
| 1059 leal 96(%edi),%edi |
| 1060 movdqa 96(%esp),%xmm3 |
| 1061 pxor %xmm0,%xmm0 |
| 1062 paddq %xmm1,%xmm1 |
| 1063 pand %xmm3,%xmm2 |
| 1064 pcmpgtd %xmm1,%xmm0 |
| 1065 movl %ebx,%ecx |
| 1066 pxor %xmm2,%xmm1 |
| 1067 subl $96,%eax |
| 1068 jnc L044xts_enc_loop6 |
| 1069 leal 1(,%ecx,2),%ecx |
| 1070 movl %ebp,%edx |
| 1071 movl %ecx,%ebx |
| 1072 L043xts_enc_short: |
| 1073 addl $96,%eax |
| 1074 jz L045xts_enc_done6x |
| 1075 movdqa %xmm1,%xmm5 |
| 1076 cmpl $32,%eax |
| 1077 jb L046xts_enc_one |
| 1078 pshufd $19,%xmm0,%xmm2 |
| 1079 pxor %xmm0,%xmm0 |
| 1080 paddq %xmm1,%xmm1 |
| 1081 pand %xmm3,%xmm2 |
| 1082 pcmpgtd %xmm1,%xmm0 |
| 1083 pxor %xmm2,%xmm1 |
| 1084 je L047xts_enc_two |
| 1085 pshufd $19,%xmm0,%xmm2 |
| 1086 pxor %xmm0,%xmm0 |
| 1087 movdqa %xmm1,%xmm6 |
| 1088 paddq %xmm1,%xmm1 |
| 1089 pand %xmm3,%xmm2 |
| 1090 pcmpgtd %xmm1,%xmm0 |
| 1091 pxor %xmm2,%xmm1 |
| 1092 cmpl $64,%eax |
| 1093 jb L048xts_enc_three |
| 1094 pshufd $19,%xmm0,%xmm2 |
| 1095 pxor %xmm0,%xmm0 |
| 1096 movdqa %xmm1,%xmm7 |
| 1097 paddq %xmm1,%xmm1 |
| 1098 pand %xmm3,%xmm2 |
| 1099 pcmpgtd %xmm1,%xmm0 |
| 1100 pxor %xmm2,%xmm1 |
| 1101 movdqa %xmm5,(%esp) |
| 1102 movdqa %xmm6,16(%esp) |
| 1103 je L049xts_enc_four |
| 1104 movdqa %xmm7,32(%esp) |
| 1105 pshufd $19,%xmm0,%xmm7 |
| 1106 movdqa %xmm1,48(%esp) |
| 1107 paddq %xmm1,%xmm1 |
| 1108 pand %xmm3,%xmm7 |
| 1109 pxor %xmm1,%xmm7 |
| 1110 movdqu (%esi),%xmm2 |
| 1111 movdqu 16(%esi),%xmm3 |
| 1112 movdqu 32(%esi),%xmm4 |
| 1113 pxor (%esp),%xmm2 |
| 1114 movdqu 48(%esi),%xmm5 |
| 1115 pxor 16(%esp),%xmm3 |
| 1116 movdqu 64(%esi),%xmm6 |
| 1117 pxor 32(%esp),%xmm4 |
| 1118 leal 80(%esi),%esi |
| 1119 pxor 48(%esp),%xmm5 |
| 1120 movdqa %xmm7,64(%esp) |
| 1121 pxor %xmm7,%xmm6 |
| 1122 call __aesni_encrypt6 |
| 1123 movaps 64(%esp),%xmm1 |
| 1124 xorps (%esp),%xmm2 |
| 1125 xorps 16(%esp),%xmm3 |
| 1126 xorps 32(%esp),%xmm4 |
| 1127 movups %xmm2,(%edi) |
| 1128 xorps 48(%esp),%xmm5 |
| 1129 movups %xmm3,16(%edi) |
| 1130 xorps %xmm1,%xmm6 |
| 1131 movups %xmm4,32(%edi) |
| 1132 movups %xmm5,48(%edi) |
| 1133 movups %xmm6,64(%edi) |
| 1134 leal 80(%edi),%edi |
| 1135 jmp L050xts_enc_done |
| 1136 .align 4,0x90 |
| 1137 L046xts_enc_one: |
| 1138 movups (%esi),%xmm2 |
| 1139 leal 16(%esi),%esi |
| 1140 xorps %xmm5,%xmm2 |
| 1141 movups (%edx),%xmm0 |
| 1142 movups 16(%edx),%xmm1 |
| 1143 leal 32(%edx),%edx |
| 1144 xorps %xmm0,%xmm2 |
| 1145 L051enc1_loop_9: |
| 1146 .byte 102,15,56,220,209 |
| 1147 decl %ecx |
| 1148 movups (%edx),%xmm1 |
| 1149 leal 16(%edx),%edx |
| 1150 jnz L051enc1_loop_9 |
| 1151 .byte 102,15,56,221,209 |
| 1152 xorps %xmm5,%xmm2 |
| 1153 movups %xmm2,(%edi) |
| 1154 leal 16(%edi),%edi |
| 1155 movdqa %xmm5,%xmm1 |
| 1156 jmp L050xts_enc_done |
| 1157 .align 4,0x90 |
| 1158 L047xts_enc_two: |
| 1159 movaps %xmm1,%xmm6 |
| 1160 movups (%esi),%xmm2 |
| 1161 movups 16(%esi),%xmm3 |
| 1162 leal 32(%esi),%esi |
| 1163 xorps %xmm5,%xmm2 |
| 1164 xorps %xmm6,%xmm3 |
| 1165 xorps %xmm4,%xmm4 |
| 1166 call __aesni_encrypt3 |
| 1167 xorps %xmm5,%xmm2 |
| 1168 xorps %xmm6,%xmm3 |
| 1169 movups %xmm2,(%edi) |
| 1170 movups %xmm3,16(%edi) |
| 1171 leal 32(%edi),%edi |
| 1172 movdqa %xmm6,%xmm1 |
| 1173 jmp L050xts_enc_done |
| 1174 .align 4,0x90 |
| 1175 L048xts_enc_three: |
| 1176 movaps %xmm1,%xmm7 |
| 1177 movups (%esi),%xmm2 |
| 1178 movups 16(%esi),%xmm3 |
| 1179 movups 32(%esi),%xmm4 |
| 1180 leal 48(%esi),%esi |
| 1181 xorps %xmm5,%xmm2 |
| 1182 xorps %xmm6,%xmm3 |
| 1183 xorps %xmm7,%xmm4 |
| 1184 call __aesni_encrypt3 |
| 1185 xorps %xmm5,%xmm2 |
| 1186 xorps %xmm6,%xmm3 |
| 1187 xorps %xmm7,%xmm4 |
| 1188 movups %xmm2,(%edi) |
| 1189 movups %xmm3,16(%edi) |
| 1190 movups %xmm4,32(%edi) |
| 1191 leal 48(%edi),%edi |
| 1192 movdqa %xmm7,%xmm1 |
| 1193 jmp L050xts_enc_done |
| 1194 .align 4,0x90 |
| 1195 L049xts_enc_four: |
| 1196 movaps %xmm1,%xmm6 |
| 1197 movups (%esi),%xmm2 |
| 1198 movups 16(%esi),%xmm3 |
| 1199 movups 32(%esi),%xmm4 |
| 1200 xorps (%esp),%xmm2 |
| 1201 movups 48(%esi),%xmm5 |
| 1202 leal 64(%esi),%esi |
| 1203 xorps 16(%esp),%xmm3 |
| 1204 xorps %xmm7,%xmm4 |
| 1205 xorps %xmm6,%xmm5 |
| 1206 call __aesni_encrypt4 |
| 1207 xorps (%esp),%xmm2 |
| 1208 xorps 16(%esp),%xmm3 |
| 1209 xorps %xmm7,%xmm4 |
| 1210 movups %xmm2,(%edi) |
| 1211 xorps %xmm6,%xmm5 |
| 1212 movups %xmm3,16(%edi) |
| 1213 movups %xmm4,32(%edi) |
| 1214 movups %xmm5,48(%edi) |
| 1215 leal 64(%edi),%edi |
| 1216 movdqa %xmm6,%xmm1 |
| 1217 jmp L050xts_enc_done |
| 1218 .align 4,0x90 |
| 1219 L045xts_enc_done6x: |
| 1220 movl 112(%esp),%eax |
| 1221 andl $15,%eax |
| 1222 jz L052xts_enc_ret |
| 1223 movdqa %xmm1,%xmm5 |
| 1224 movl %eax,112(%esp) |
| 1225 jmp L053xts_enc_steal |
| 1226 .align 4,0x90 |
| 1227 L050xts_enc_done: |
| 1228 movl 112(%esp),%eax |
| 1229 pxor %xmm0,%xmm0 |
| 1230 andl $15,%eax |
| 1231 jz L052xts_enc_ret |
| 1232 pcmpgtd %xmm1,%xmm0 |
| 1233 movl %eax,112(%esp) |
| 1234 pshufd $19,%xmm0,%xmm5 |
| 1235 paddq %xmm1,%xmm1 |
| 1236 pand 96(%esp),%xmm5 |
| 1237 pxor %xmm1,%xmm5 |
| 1238 L053xts_enc_steal: |
| 1239 movzbl (%esi),%ecx |
| 1240 movzbl -16(%edi),%edx |
| 1241 leal 1(%esi),%esi |
| 1242 movb %cl,-16(%edi) |
| 1243 movb %dl,(%edi) |
| 1244 leal 1(%edi),%edi |
| 1245 subl $1,%eax |
| 1246 jnz L053xts_enc_steal |
| 1247 subl 112(%esp),%edi |
| 1248 movl %ebp,%edx |
| 1249 movl %ebx,%ecx |
| 1250 movups -16(%edi),%xmm2 |
| 1251 xorps %xmm5,%xmm2 |
| 1252 movups (%edx),%xmm0 |
| 1253 movups 16(%edx),%xmm1 |
| 1254 leal 32(%edx),%edx |
| 1255 xorps %xmm0,%xmm2 |
| 1256 L054enc1_loop_10: |
| 1257 .byte 102,15,56,220,209 |
| 1258 decl %ecx |
| 1259 movups (%edx),%xmm1 |
| 1260 leal 16(%edx),%edx |
| 1261 jnz L054enc1_loop_10 |
| 1262 .byte 102,15,56,221,209 |
| 1263 xorps %xmm5,%xmm2 |
| 1264 movups %xmm2,-16(%edi) |
| 1265 L052xts_enc_ret: |
| 1266 movl 116(%esp),%esp |
| 1267 popl %edi |
| 1268 popl %esi |
| 1269 popl %ebx |
| 1270 popl %ebp |
| 1271 ret |
| 1272 .globl _aesni_xts_decrypt |
| 1273 .align 4 |
| 1274 _aesni_xts_decrypt: |
| 1275 L_aesni_xts_decrypt_begin: |
| 1276 pushl %ebp |
| 1277 pushl %ebx |
| 1278 pushl %esi |
| 1279 pushl %edi |
| 1280 movl 36(%esp),%edx |
| 1281 movl 40(%esp),%esi |
| 1282 movl 240(%edx),%ecx |
| 1283 movups (%esi),%xmm2 |
| 1284 movups (%edx),%xmm0 |
| 1285 movups 16(%edx),%xmm1 |
| 1286 leal 32(%edx),%edx |
| 1287 xorps %xmm0,%xmm2 |
| 1288 L055enc1_loop_11: |
| 1289 .byte 102,15,56,220,209 |
| 1290 decl %ecx |
| 1291 movups (%edx),%xmm1 |
| 1292 leal 16(%edx),%edx |
| 1293 jnz L055enc1_loop_11 |
| 1294 .byte 102,15,56,221,209 |
| 1295 movl 20(%esp),%esi |
| 1296 movl 24(%esp),%edi |
| 1297 movl 28(%esp),%eax |
| 1298 movl 32(%esp),%edx |
| 1299 movl %esp,%ebp |
| 1300 subl $120,%esp |
| 1301 andl $-16,%esp |
| 1302 xorl %ebx,%ebx |
| 1303 testl $15,%eax |
| 1304 setnz %bl |
| 1305 shll $4,%ebx |
| 1306 subl %ebx,%eax |
| 1307 movl $135,96(%esp) |
| 1308 movl $0,100(%esp) |
| 1309 movl $1,104(%esp) |
| 1310 movl $0,108(%esp) |
| 1311 movl %eax,112(%esp) |
| 1312 movl %ebp,116(%esp) |
| 1313 movl 240(%edx),%ecx |
| 1314 movl %edx,%ebp |
| 1315 movl %ecx,%ebx |
| 1316 movdqa %xmm2,%xmm1 |
| 1317 pxor %xmm0,%xmm0 |
| 1318 movdqa 96(%esp),%xmm3 |
| 1319 pcmpgtd %xmm1,%xmm0 |
| 1320 andl $-16,%eax |
| 1321 subl $96,%eax |
| 1322 jc L056xts_dec_short |
| 1323 shrl $1,%ecx |
| 1324 movl %ecx,%ebx |
| 1325 jmp L057xts_dec_loop6 |
| 1326 .align 4,0x90 |
| 1327 L057xts_dec_loop6: |
| 1328 pshufd $19,%xmm0,%xmm2 |
| 1329 pxor %xmm0,%xmm0 |
| 1330 movdqa %xmm1,(%esp) |
| 1331 paddq %xmm1,%xmm1 |
| 1332 pand %xmm3,%xmm2 |
| 1333 pcmpgtd %xmm1,%xmm0 |
| 1334 pxor %xmm2,%xmm1 |
| 1335 pshufd $19,%xmm0,%xmm2 |
| 1336 pxor %xmm0,%xmm0 |
| 1337 movdqa %xmm1,16(%esp) |
| 1338 paddq %xmm1,%xmm1 |
| 1339 pand %xmm3,%xmm2 |
| 1340 pcmpgtd %xmm1,%xmm0 |
| 1341 pxor %xmm2,%xmm1 |
| 1342 pshufd $19,%xmm0,%xmm2 |
| 1343 pxor %xmm0,%xmm0 |
| 1344 movdqa %xmm1,32(%esp) |
| 1345 paddq %xmm1,%xmm1 |
| 1346 pand %xmm3,%xmm2 |
| 1347 pcmpgtd %xmm1,%xmm0 |
| 1348 pxor %xmm2,%xmm1 |
| 1349 pshufd $19,%xmm0,%xmm2 |
| 1350 pxor %xmm0,%xmm0 |
| 1351 movdqa %xmm1,48(%esp) |
| 1352 paddq %xmm1,%xmm1 |
| 1353 pand %xmm3,%xmm2 |
| 1354 pcmpgtd %xmm1,%xmm0 |
| 1355 pxor %xmm2,%xmm1 |
| 1356 pshufd $19,%xmm0,%xmm7 |
| 1357 movdqa %xmm1,64(%esp) |
| 1358 paddq %xmm1,%xmm1 |
| 1359 movups (%ebp),%xmm0 |
| 1360 pand %xmm3,%xmm7 |
| 1361 movups (%esi),%xmm2 |
| 1362 pxor %xmm1,%xmm7 |
| 1363 movdqu 16(%esi),%xmm3 |
| 1364 xorps %xmm0,%xmm2 |
| 1365 movdqu 32(%esi),%xmm4 |
| 1366 pxor %xmm0,%xmm3 |
| 1367 movdqu 48(%esi),%xmm5 |
| 1368 pxor %xmm0,%xmm4 |
| 1369 movdqu 64(%esi),%xmm6 |
| 1370 pxor %xmm0,%xmm5 |
| 1371 movdqu 80(%esi),%xmm1 |
| 1372 pxor %xmm0,%xmm6 |
| 1373 leal 96(%esi),%esi |
| 1374 pxor (%esp),%xmm2 |
| 1375 movdqa %xmm7,80(%esp) |
| 1376 pxor %xmm1,%xmm7 |
| 1377 movups 16(%ebp),%xmm1 |
| 1378 leal 32(%ebp),%edx |
| 1379 pxor 16(%esp),%xmm3 |
| 1380 .byte 102,15,56,222,209 |
| 1381 pxor 32(%esp),%xmm4 |
| 1382 .byte 102,15,56,222,217 |
| 1383 pxor 48(%esp),%xmm5 |
| 1384 decl %ecx |
| 1385 .byte 102,15,56,222,225 |
| 1386 pxor 64(%esp),%xmm6 |
| 1387 .byte 102,15,56,222,233 |
| 1388 pxor %xmm0,%xmm7 |
| 1389 .byte 102,15,56,222,241 |
| 1390 movups (%edx),%xmm0 |
| 1391 .byte 102,15,56,222,249 |
| 1392 call L_aesni_decrypt6_enter |
| 1393 movdqa 80(%esp),%xmm1 |
| 1394 pxor %xmm0,%xmm0 |
| 1395 xorps (%esp),%xmm2 |
| 1396 pcmpgtd %xmm1,%xmm0 |
| 1397 xorps 16(%esp),%xmm3 |
| 1398 movups %xmm2,(%edi) |
| 1399 xorps 32(%esp),%xmm4 |
| 1400 movups %xmm3,16(%edi) |
| 1401 xorps 48(%esp),%xmm5 |
| 1402 movups %xmm4,32(%edi) |
| 1403 xorps 64(%esp),%xmm6 |
| 1404 movups %xmm5,48(%edi) |
| 1405 xorps %xmm1,%xmm7 |
| 1406 movups %xmm6,64(%edi) |
| 1407 pshufd $19,%xmm0,%xmm2 |
| 1408 movups %xmm7,80(%edi) |
| 1409 leal 96(%edi),%edi |
| 1410 movdqa 96(%esp),%xmm3 |
| 1411 pxor %xmm0,%xmm0 |
| 1412 paddq %xmm1,%xmm1 |
| 1413 pand %xmm3,%xmm2 |
| 1414 pcmpgtd %xmm1,%xmm0 |
| 1415 movl %ebx,%ecx |
| 1416 pxor %xmm2,%xmm1 |
| 1417 subl $96,%eax |
| 1418 jnc L057xts_dec_loop6 |
| 1419 leal 1(,%ecx,2),%ecx |
| 1420 movl %ebp,%edx |
| 1421 movl %ecx,%ebx |
| 1422 L056xts_dec_short: |
| 1423 addl $96,%eax |
| 1424 jz L058xts_dec_done6x |
| 1425 movdqa %xmm1,%xmm5 |
| 1426 cmpl $32,%eax |
| 1427 jb L059xts_dec_one |
| 1428 pshufd $19,%xmm0,%xmm2 |
| 1429 pxor %xmm0,%xmm0 |
| 1430 paddq %xmm1,%xmm1 |
| 1431 pand %xmm3,%xmm2 |
| 1432 pcmpgtd %xmm1,%xmm0 |
| 1433 pxor %xmm2,%xmm1 |
| 1434 je L060xts_dec_two |
| 1435 pshufd $19,%xmm0,%xmm2 |
| 1436 pxor %xmm0,%xmm0 |
| 1437 movdqa %xmm1,%xmm6 |
| 1438 paddq %xmm1,%xmm1 |
| 1439 pand %xmm3,%xmm2 |
| 1440 pcmpgtd %xmm1,%xmm0 |
| 1441 pxor %xmm2,%xmm1 |
| 1442 cmpl $64,%eax |
| 1443 jb L061xts_dec_three |
| 1444 pshufd $19,%xmm0,%xmm2 |
| 1445 pxor %xmm0,%xmm0 |
| 1446 movdqa %xmm1,%xmm7 |
| 1447 paddq %xmm1,%xmm1 |
| 1448 pand %xmm3,%xmm2 |
| 1449 pcmpgtd %xmm1,%xmm0 |
| 1450 pxor %xmm2,%xmm1 |
| 1451 movdqa %xmm5,(%esp) |
| 1452 movdqa %xmm6,16(%esp) |
| 1453 je L062xts_dec_four |
| 1454 movdqa %xmm7,32(%esp) |
| 1455 pshufd $19,%xmm0,%xmm7 |
| 1456 movdqa %xmm1,48(%esp) |
| 1457 paddq %xmm1,%xmm1 |
| 1458 pand %xmm3,%xmm7 |
| 1459 pxor %xmm1,%xmm7 |
| 1460 movdqu (%esi),%xmm2 |
| 1461 movdqu 16(%esi),%xmm3 |
| 1462 movdqu 32(%esi),%xmm4 |
| 1463 pxor (%esp),%xmm2 |
| 1464 movdqu 48(%esi),%xmm5 |
| 1465 pxor 16(%esp),%xmm3 |
| 1466 movdqu 64(%esi),%xmm6 |
| 1467 pxor 32(%esp),%xmm4 |
| 1468 leal 80(%esi),%esi |
| 1469 pxor 48(%esp),%xmm5 |
| 1470 movdqa %xmm7,64(%esp) |
| 1471 pxor %xmm7,%xmm6 |
| 1472 call __aesni_decrypt6 |
| 1473 movaps 64(%esp),%xmm1 |
| 1474 xorps (%esp),%xmm2 |
| 1475 xorps 16(%esp),%xmm3 |
| 1476 xorps 32(%esp),%xmm4 |
| 1477 movups %xmm2,(%edi) |
| 1478 xorps 48(%esp),%xmm5 |
| 1479 movups %xmm3,16(%edi) |
| 1480 xorps %xmm1,%xmm6 |
| 1481 movups %xmm4,32(%edi) |
| 1482 movups %xmm5,48(%edi) |
| 1483 movups %xmm6,64(%edi) |
| 1484 leal 80(%edi),%edi |
| 1485 jmp L063xts_dec_done |
| 1486 .align 4,0x90 |
| 1487 L059xts_dec_one: |
| 1488 movups (%esi),%xmm2 |
| 1489 leal 16(%esi),%esi |
| 1490 xorps %xmm5,%xmm2 |
| 1491 movups (%edx),%xmm0 |
| 1492 movups 16(%edx),%xmm1 |
| 1493 leal 32(%edx),%edx |
| 1494 xorps %xmm0,%xmm2 |
| 1495 L064dec1_loop_12: |
| 1496 .byte 102,15,56,222,209 |
| 1497 decl %ecx |
| 1498 movups (%edx),%xmm1 |
| 1499 leal 16(%edx),%edx |
| 1500 jnz L064dec1_loop_12 |
| 1501 .byte 102,15,56,223,209 |
| 1502 xorps %xmm5,%xmm2 |
| 1503 movups %xmm2,(%edi) |
| 1504 leal 16(%edi),%edi |
| 1505 movdqa %xmm5,%xmm1 |
| 1506 jmp L063xts_dec_done |
| 1507 .align 4,0x90 |
| 1508 L060xts_dec_two: |
| 1509 movaps %xmm1,%xmm6 |
| 1510 movups (%esi),%xmm2 |
| 1511 movups 16(%esi),%xmm3 |
| 1512 leal 32(%esi),%esi |
| 1513 xorps %xmm5,%xmm2 |
| 1514 xorps %xmm6,%xmm3 |
| 1515 call __aesni_decrypt3 |
| 1516 xorps %xmm5,%xmm2 |
| 1517 xorps %xmm6,%xmm3 |
| 1518 movups %xmm2,(%edi) |
| 1519 movups %xmm3,16(%edi) |
| 1520 leal 32(%edi),%edi |
| 1521 movdqa %xmm6,%xmm1 |
| 1522 jmp L063xts_dec_done |
| 1523 .align 4,0x90 |
| 1524 L061xts_dec_three: |
| 1525 movaps %xmm1,%xmm7 |
| 1526 movups (%esi),%xmm2 |
| 1527 movups 16(%esi),%xmm3 |
| 1528 movups 32(%esi),%xmm4 |
| 1529 leal 48(%esi),%esi |
| 1530 xorps %xmm5,%xmm2 |
| 1531 xorps %xmm6,%xmm3 |
| 1532 xorps %xmm7,%xmm4 |
| 1533 call __aesni_decrypt3 |
| 1534 xorps %xmm5,%xmm2 |
| 1535 xorps %xmm6,%xmm3 |
| 1536 xorps %xmm7,%xmm4 |
| 1537 movups %xmm2,(%edi) |
| 1538 movups %xmm3,16(%edi) |
| 1539 movups %xmm4,32(%edi) |
| 1540 leal 48(%edi),%edi |
| 1541 movdqa %xmm7,%xmm1 |
| 1542 jmp L063xts_dec_done |
| 1543 .align 4,0x90 |
| 1544 L062xts_dec_four: |
| 1545 movaps %xmm1,%xmm6 |
| 1546 movups (%esi),%xmm2 |
| 1547 movups 16(%esi),%xmm3 |
| 1548 movups 32(%esi),%xmm4 |
| 1549 xorps (%esp),%xmm2 |
| 1550 movups 48(%esi),%xmm5 |
| 1551 leal 64(%esi),%esi |
| 1552 xorps 16(%esp),%xmm3 |
| 1553 xorps %xmm7,%xmm4 |
| 1554 xorps %xmm6,%xmm5 |
| 1555 call __aesni_decrypt4 |
| 1556 xorps (%esp),%xmm2 |
| 1557 xorps 16(%esp),%xmm3 |
| 1558 xorps %xmm7,%xmm4 |
| 1559 movups %xmm2,(%edi) |
| 1560 xorps %xmm6,%xmm5 |
| 1561 movups %xmm3,16(%edi) |
| 1562 movups %xmm4,32(%edi) |
| 1563 movups %xmm5,48(%edi) |
| 1564 leal 64(%edi),%edi |
| 1565 movdqa %xmm6,%xmm1 |
| 1566 jmp L063xts_dec_done |
| 1567 .align 4,0x90 |
| 1568 L058xts_dec_done6x: |
| 1569 movl 112(%esp),%eax |
| 1570 andl $15,%eax |
| 1571 jz L065xts_dec_ret |
| 1572 movl %eax,112(%esp) |
| 1573 jmp L066xts_dec_only_one_more |
| 1574 .align 4,0x90 |
| 1575 L063xts_dec_done: |
| 1576 movl 112(%esp),%eax |
| 1577 pxor %xmm0,%xmm0 |
| 1578 andl $15,%eax |
| 1579 jz L065xts_dec_ret |
| 1580 pcmpgtd %xmm1,%xmm0 |
| 1581 movl %eax,112(%esp) |
| 1582 pshufd $19,%xmm0,%xmm2 |
| 1583 pxor %xmm0,%xmm0 |
| 1584 movdqa 96(%esp),%xmm3 |
| 1585 paddq %xmm1,%xmm1 |
| 1586 pand %xmm3,%xmm2 |
| 1587 pcmpgtd %xmm1,%xmm0 |
| 1588 pxor %xmm2,%xmm1 |
| 1589 L066xts_dec_only_one_more: |
| 1590 pshufd $19,%xmm0,%xmm5 |
| 1591 movdqa %xmm1,%xmm6 |
| 1592 paddq %xmm1,%xmm1 |
| 1593 pand %xmm3,%xmm5 |
| 1594 pxor %xmm1,%xmm5 |
| 1595 movl %ebp,%edx |
| 1596 movl %ebx,%ecx |
| 1597 movups (%esi),%xmm2 |
| 1598 xorps %xmm5,%xmm2 |
| 1599 movups (%edx),%xmm0 |
| 1600 movups 16(%edx),%xmm1 |
| 1601 leal 32(%edx),%edx |
| 1602 xorps %xmm0,%xmm2 |
| 1603 L067dec1_loop_13: |
| 1604 .byte 102,15,56,222,209 |
| 1605 decl %ecx |
| 1606 movups (%edx),%xmm1 |
| 1607 leal 16(%edx),%edx |
| 1608 jnz L067dec1_loop_13 |
| 1609 .byte 102,15,56,223,209 |
| 1610 xorps %xmm5,%xmm2 |
| 1611 movups %xmm2,(%edi) |
| 1612 L068xts_dec_steal: |
| 1613 movzbl 16(%esi),%ecx |
| 1614 movzbl (%edi),%edx |
| 1615 leal 1(%esi),%esi |
| 1616 movb %cl,(%edi) |
| 1617 movb %dl,16(%edi) |
| 1618 leal 1(%edi),%edi |
| 1619 subl $1,%eax |
| 1620 jnz L068xts_dec_steal |
| 1621 subl 112(%esp),%edi |
| 1622 movl %ebp,%edx |
| 1623 movl %ebx,%ecx |
| 1624 movups (%edi),%xmm2 |
| 1625 xorps %xmm6,%xmm2 |
| 1626 movups (%edx),%xmm0 |
| 1627 movups 16(%edx),%xmm1 |
| 1628 leal 32(%edx),%edx |
| 1629 xorps %xmm0,%xmm2 |
| 1630 L069dec1_loop_14: |
| 1631 .byte 102,15,56,222,209 |
| 1632 decl %ecx |
| 1633 movups (%edx),%xmm1 |
| 1634 leal 16(%edx),%edx |
| 1635 jnz L069dec1_loop_14 |
| 1636 .byte 102,15,56,223,209 |
| 1637 xorps %xmm6,%xmm2 |
| 1638 movups %xmm2,(%edi) |
| 1639 L065xts_dec_ret: |
| 1640 movl 116(%esp),%esp |
| 1641 popl %edi |
| 1642 popl %esi |
| 1643 popl %ebx |
| 1644 popl %ebp |
| 1645 ret |
| 1646 .globl _aesni_cbc_encrypt |
| 1647 .align 4 |
| 1648 _aesni_cbc_encrypt: |
| 1649 L_aesni_cbc_encrypt_begin: |
| 1650 pushl %ebp |
| 1651 pushl %ebx |
| 1652 pushl %esi |
| 1653 pushl %edi |
| 1654 movl 20(%esp),%esi |
| 1655 movl %esp,%ebx |
| 1656 movl 24(%esp),%edi |
| 1657 subl $24,%ebx |
| 1658 movl 28(%esp),%eax |
| 1659 andl $-16,%ebx |
| 1660 movl 32(%esp),%edx |
| 1661 movl 36(%esp),%ebp |
| 1662 testl %eax,%eax |
| 1663 jz L070cbc_abort |
| 1664 cmpl $0,40(%esp) |
| 1665 xchgl %esp,%ebx |
| 1666 movups (%ebp),%xmm7 |
| 1667 movl 240(%edx),%ecx |
| 1668 movl %edx,%ebp |
| 1669 movl %ebx,16(%esp) |
| 1670 movl %ecx,%ebx |
| 1671 je L071cbc_decrypt |
| 1672 movaps %xmm7,%xmm2 |
| 1673 cmpl $16,%eax |
| 1674 jb L072cbc_enc_tail |
| 1675 subl $16,%eax |
| 1676 jmp L073cbc_enc_loop |
| 1677 .align 4,0x90 |
| 1678 L073cbc_enc_loop: |
| 1679 movups (%esi),%xmm7 |
| 1680 leal 16(%esi),%esi |
| 1681 movups (%edx),%xmm0 |
| 1682 movups 16(%edx),%xmm1 |
| 1683 xorps %xmm0,%xmm7 |
| 1684 leal 32(%edx),%edx |
| 1685 xorps %xmm7,%xmm2 |
| 1686 L074enc1_loop_15: |
| 1687 .byte 102,15,56,220,209 |
| 1688 decl %ecx |
| 1689 movups (%edx),%xmm1 |
| 1690 leal 16(%edx),%edx |
| 1691 jnz L074enc1_loop_15 |
| 1692 .byte 102,15,56,221,209 |
| 1693 movl %ebx,%ecx |
| 1694 movl %ebp,%edx |
| 1695 movups %xmm2,(%edi) |
| 1696 leal 16(%edi),%edi |
| 1697 subl $16,%eax |
| 1698 jnc L073cbc_enc_loop |
| 1699 addl $16,%eax |
| 1700 jnz L072cbc_enc_tail |
| 1701 movaps %xmm2,%xmm7 |
| 1702 jmp L075cbc_ret |
| 1703 L072cbc_enc_tail: |
| 1704 movl %eax,%ecx |
| 1705 .long 2767451785 |
| 1706 movl $16,%ecx |
| 1707 subl %eax,%ecx |
| 1708 xorl %eax,%eax |
| 1709 .long 2868115081 |
| 1710 leal -16(%edi),%edi |
| 1711 movl %ebx,%ecx |
| 1712 movl %edi,%esi |
| 1713 movl %ebp,%edx |
| 1714 jmp L073cbc_enc_loop |
| 1715 .align 4,0x90 |
| 1716 L071cbc_decrypt: |
| 1717 cmpl $80,%eax |
| 1718 jbe L076cbc_dec_tail |
| 1719 movaps %xmm7,(%esp) |
| 1720 subl $80,%eax |
| 1721 jmp L077cbc_dec_loop6_enter |
| 1722 .align 4,0x90 |
| 1723 L078cbc_dec_loop6: |
| 1724 movaps %xmm0,(%esp) |
| 1725 movups %xmm7,(%edi) |
| 1726 leal 16(%edi),%edi |
| 1727 L077cbc_dec_loop6_enter: |
| 1728 movdqu (%esi),%xmm2 |
| 1729 movdqu 16(%esi),%xmm3 |
| 1730 movdqu 32(%esi),%xmm4 |
| 1731 movdqu 48(%esi),%xmm5 |
| 1732 movdqu 64(%esi),%xmm6 |
| 1733 movdqu 80(%esi),%xmm7 |
| 1734 call __aesni_decrypt6 |
| 1735 movups (%esi),%xmm1 |
| 1736 movups 16(%esi),%xmm0 |
| 1737 xorps (%esp),%xmm2 |
| 1738 xorps %xmm1,%xmm3 |
| 1739 movups 32(%esi),%xmm1 |
| 1740 xorps %xmm0,%xmm4 |
| 1741 movups 48(%esi),%xmm0 |
| 1742 xorps %xmm1,%xmm5 |
| 1743 movups 64(%esi),%xmm1 |
| 1744 xorps %xmm0,%xmm6 |
| 1745 movups 80(%esi),%xmm0 |
| 1746 xorps %xmm1,%xmm7 |
| 1747 movups %xmm2,(%edi) |
| 1748 movups %xmm3,16(%edi) |
| 1749 leal 96(%esi),%esi |
| 1750 movups %xmm4,32(%edi) |
| 1751 movl %ebx,%ecx |
| 1752 movups %xmm5,48(%edi) |
| 1753 movl %ebp,%edx |
| 1754 movups %xmm6,64(%edi) |
| 1755 leal 80(%edi),%edi |
| 1756 subl $96,%eax |
| 1757 ja L078cbc_dec_loop6 |
| 1758 movaps %xmm7,%xmm2 |
| 1759 movaps %xmm0,%xmm7 |
| 1760 addl $80,%eax |
| 1761 jle L079cbc_dec_tail_collected |
| 1762 movups %xmm2,(%edi) |
| 1763 leal 16(%edi),%edi |
| 1764 L076cbc_dec_tail: |
| 1765 movups (%esi),%xmm2 |
| 1766 movaps %xmm2,%xmm6 |
| 1767 cmpl $16,%eax |
| 1768 jbe L080cbc_dec_one |
| 1769 movups 16(%esi),%xmm3 |
| 1770 movaps %xmm3,%xmm5 |
| 1771 cmpl $32,%eax |
| 1772 jbe L081cbc_dec_two |
| 1773 movups 32(%esi),%xmm4 |
| 1774 cmpl $48,%eax |
| 1775 jbe L082cbc_dec_three |
| 1776 movups 48(%esi),%xmm5 |
| 1777 cmpl $64,%eax |
| 1778 jbe L083cbc_dec_four |
| 1779 movups 64(%esi),%xmm6 |
| 1780 movaps %xmm7,(%esp) |
| 1781 movups (%esi),%xmm2 |
| 1782 xorps %xmm7,%xmm7 |
| 1783 call __aesni_decrypt6 |
| 1784 movups (%esi),%xmm1 |
| 1785 movups 16(%esi),%xmm0 |
| 1786 xorps (%esp),%xmm2 |
| 1787 xorps %xmm1,%xmm3 |
| 1788 movups 32(%esi),%xmm1 |
| 1789 xorps %xmm0,%xmm4 |
| 1790 movups 48(%esi),%xmm0 |
| 1791 xorps %xmm1,%xmm5 |
| 1792 movups 64(%esi),%xmm7 |
| 1793 xorps %xmm0,%xmm6 |
| 1794 movups %xmm2,(%edi) |
| 1795 movups %xmm3,16(%edi) |
| 1796 movups %xmm4,32(%edi) |
| 1797 movups %xmm5,48(%edi) |
| 1798 leal 64(%edi),%edi |
| 1799 movaps %xmm6,%xmm2 |
| 1800 subl $80,%eax |
| 1801 jmp L079cbc_dec_tail_collected |
| 1802 .align 4,0x90 |
| 1803 L080cbc_dec_one: |
| 1804 movups (%edx),%xmm0 |
| 1805 movups 16(%edx),%xmm1 |
| 1806 leal 32(%edx),%edx |
| 1807 xorps %xmm0,%xmm2 |
| 1808 L084dec1_loop_16: |
| 1809 .byte 102,15,56,222,209 |
| 1810 decl %ecx |
| 1811 movups (%edx),%xmm1 |
| 1812 leal 16(%edx),%edx |
| 1813 jnz L084dec1_loop_16 |
| 1814 .byte 102,15,56,223,209 |
| 1815 xorps %xmm7,%xmm2 |
| 1816 movaps %xmm6,%xmm7 |
| 1817 subl $16,%eax |
| 1818 jmp L079cbc_dec_tail_collected |
| 1819 .align 4,0x90 |
| 1820 L081cbc_dec_two: |
| 1821 xorps %xmm4,%xmm4 |
| 1822 call __aesni_decrypt3 |
| 1823 xorps %xmm7,%xmm2 |
| 1824 xorps %xmm6,%xmm3 |
| 1825 movups %xmm2,(%edi) |
| 1826 movaps %xmm3,%xmm2 |
| 1827 leal 16(%edi),%edi |
| 1828 movaps %xmm5,%xmm7 |
| 1829 subl $32,%eax |
| 1830 jmp L079cbc_dec_tail_collected |
| 1831 .align 4,0x90 |
| 1832 L082cbc_dec_three: |
| 1833 call __aesni_decrypt3 |
| 1834 xorps %xmm7,%xmm2 |
| 1835 xorps %xmm6,%xmm3 |
| 1836 xorps %xmm5,%xmm4 |
| 1837 movups %xmm2,(%edi) |
| 1838 movaps %xmm4,%xmm2 |
| 1839 movups %xmm3,16(%edi) |
| 1840 leal 32(%edi),%edi |
| 1841 movups 32(%esi),%xmm7 |
| 1842 subl $48,%eax |
| 1843 jmp L079cbc_dec_tail_collected |
| 1844 .align 4,0x90 |
| 1845 L083cbc_dec_four: |
| 1846 call __aesni_decrypt4 |
| 1847 movups 16(%esi),%xmm1 |
| 1848 movups 32(%esi),%xmm0 |
| 1849 xorps %xmm7,%xmm2 |
| 1850 movups 48(%esi),%xmm7 |
| 1851 xorps %xmm6,%xmm3 |
| 1852 movups %xmm2,(%edi) |
| 1853 xorps %xmm1,%xmm4 |
| 1854 movups %xmm3,16(%edi) |
| 1855 xorps %xmm0,%xmm5 |
| 1856 movups %xmm4,32(%edi) |
| 1857 leal 48(%edi),%edi |
| 1858 movaps %xmm5,%xmm2 |
| 1859 subl $64,%eax |
| 1860 L079cbc_dec_tail_collected: |
| 1861 andl $15,%eax |
| 1862 jnz L085cbc_dec_tail_partial |
| 1863 movups %xmm2,(%edi) |
| 1864 jmp L075cbc_ret |
| 1865 .align 4,0x90 |
| 1866 L085cbc_dec_tail_partial: |
| 1867 movaps %xmm2,(%esp) |
| 1868 movl $16,%ecx |
| 1869 movl %esp,%esi |
| 1870 subl %eax,%ecx |
| 1871 .long 2767451785 |
| 1872 L075cbc_ret: |
| 1873 movl 16(%esp),%esp |
| 1874 movl 36(%esp),%ebp |
| 1875 movups %xmm7,(%ebp) |
| 1876 L070cbc_abort: |
| 1877 popl %edi |
| 1878 popl %esi |
| 1879 popl %ebx |
| 1880 popl %ebp |
| 1881 ret |
| 1882 .align 4 |
| 1883 __aesni_set_encrypt_key: |
| 1884 testl %eax,%eax |
| 1885 jz L086bad_pointer |
| 1886 testl %edx,%edx |
| 1887 jz L086bad_pointer |
| 1888 movups (%eax),%xmm0 |
| 1889 xorps %xmm4,%xmm4 |
| 1890 leal 16(%edx),%edx |
| 1891 cmpl $256,%ecx |
| 1892 je L08714rounds |
| 1893 cmpl $192,%ecx |
| 1894 je L08812rounds |
| 1895 cmpl $128,%ecx |
| 1896 jne L089bad_keybits |
| 1897 .align 4,0x90 |
| 1898 L09010rounds: |
| 1899 movl $9,%ecx |
| 1900 movups %xmm0,-16(%edx) |
| 1901 .byte 102,15,58,223,200,1 |
| 1902 call L091key_128_cold |
| 1903 .byte 102,15,58,223,200,2 |
| 1904 call L092key_128 |
| 1905 .byte 102,15,58,223,200,4 |
| 1906 call L092key_128 |
| 1907 .byte 102,15,58,223,200,8 |
| 1908 call L092key_128 |
| 1909 .byte 102,15,58,223,200,16 |
| 1910 call L092key_128 |
| 1911 .byte 102,15,58,223,200,32 |
| 1912 call L092key_128 |
| 1913 .byte 102,15,58,223,200,64 |
| 1914 call L092key_128 |
| 1915 .byte 102,15,58,223,200,128 |
| 1916 call L092key_128 |
| 1917 .byte 102,15,58,223,200,27 |
| 1918 call L092key_128 |
| 1919 .byte 102,15,58,223,200,54 |
| 1920 call L092key_128 |
| 1921 movups %xmm0,(%edx) |
| 1922 movl %ecx,80(%edx) |
| 1923 xorl %eax,%eax |
| 1924 ret |
| 1925 .align 4,0x90 |
| 1926 L092key_128: |
| 1927 movups %xmm0,(%edx) |
| 1928 leal 16(%edx),%edx |
| 1929 L091key_128_cold: |
| 1930 shufps $16,%xmm0,%xmm4 |
| 1931 xorps %xmm4,%xmm0 |
| 1932 shufps $140,%xmm0,%xmm4 |
| 1933 xorps %xmm4,%xmm0 |
| 1934 shufps $255,%xmm1,%xmm1 |
| 1935 xorps %xmm1,%xmm0 |
| 1936 ret |
| 1937 .align 4,0x90 |
| 1938 L08812rounds: |
| 1939 movq 16(%eax),%xmm2 |
| 1940 movl $11,%ecx |
| 1941 movups %xmm0,-16(%edx) |
| 1942 .byte 102,15,58,223,202,1 |
| 1943 call L093key_192a_cold |
| 1944 .byte 102,15,58,223,202,2 |
| 1945 call L094key_192b |
| 1946 .byte 102,15,58,223,202,4 |
| 1947 call L095key_192a |
| 1948 .byte 102,15,58,223,202,8 |
| 1949 call L094key_192b |
| 1950 .byte 102,15,58,223,202,16 |
| 1951 call L095key_192a |
| 1952 .byte 102,15,58,223,202,32 |
| 1953 call L094key_192b |
| 1954 .byte 102,15,58,223,202,64 |
| 1955 call L095key_192a |
| 1956 .byte 102,15,58,223,202,128 |
| 1957 call L094key_192b |
| 1958 movups %xmm0,(%edx) |
| 1959 movl %ecx,48(%edx) |
| 1960 xorl %eax,%eax |
| 1961 ret |
| 1962 .align 4,0x90 |
| 1963 L095key_192a: |
| 1964 movups %xmm0,(%edx) |
| 1965 leal 16(%edx),%edx |
| 1966 .align 4,0x90 |
| 1967 L093key_192a_cold: |
| 1968 movaps %xmm2,%xmm5 |
| 1969 L096key_192b_warm: |
| 1970 shufps $16,%xmm0,%xmm4 |
| 1971 movdqa %xmm2,%xmm3 |
| 1972 xorps %xmm4,%xmm0 |
| 1973 shufps $140,%xmm0,%xmm4 |
| 1974 pslldq $4,%xmm3 |
| 1975 xorps %xmm4,%xmm0 |
| 1976 pshufd $85,%xmm1,%xmm1 |
| 1977 pxor %xmm3,%xmm2 |
| 1978 pxor %xmm1,%xmm0 |
| 1979 pshufd $255,%xmm0,%xmm3 |
| 1980 pxor %xmm3,%xmm2 |
| 1981 ret |
| 1982 .align 4,0x90 |
| 1983 L094key_192b: |
| 1984 movaps %xmm0,%xmm3 |
| 1985 shufps $68,%xmm0,%xmm5 |
| 1986 movups %xmm5,(%edx) |
| 1987 shufps $78,%xmm2,%xmm3 |
| 1988 movups %xmm3,16(%edx) |
| 1989 leal 32(%edx),%edx |
| 1990 jmp L096key_192b_warm |
| 1991 .align 4,0x90 |
| 1992 L08714rounds: |
| 1993 movups 16(%eax),%xmm2 |
| 1994 movl $13,%ecx |
| 1995 leal 16(%edx),%edx |
| 1996 movups %xmm0,-32(%edx) |
| 1997 movups %xmm2,-16(%edx) |
| 1998 .byte 102,15,58,223,202,1 |
| 1999 call L097key_256a_cold |
| 2000 .byte 102,15,58,223,200,1 |
| 2001 call L098key_256b |
| 2002 .byte 102,15,58,223,202,2 |
| 2003 call L099key_256a |
| 2004 .byte 102,15,58,223,200,2 |
| 2005 call L098key_256b |
| 2006 .byte 102,15,58,223,202,4 |
| 2007 call L099key_256a |
| 2008 .byte 102,15,58,223,200,4 |
| 2009 call L098key_256b |
| 2010 .byte 102,15,58,223,202,8 |
| 2011 call L099key_256a |
| 2012 .byte 102,15,58,223,200,8 |
| 2013 call L098key_256b |
| 2014 .byte 102,15,58,223,202,16 |
| 2015 call L099key_256a |
| 2016 .byte 102,15,58,223,200,16 |
| 2017 call L098key_256b |
| 2018 .byte 102,15,58,223,202,32 |
| 2019 call L099key_256a |
| 2020 .byte 102,15,58,223,200,32 |
| 2021 call L098key_256b |
| 2022 .byte 102,15,58,223,202,64 |
| 2023 call L099key_256a |
| 2024 movups %xmm0,(%edx) |
| 2025 movl %ecx,16(%edx) |
| 2026 xorl %eax,%eax |
| 2027 ret |
| 2028 .align 4,0x90 |
| 2029 L099key_256a: |
| 2030 movups %xmm2,(%edx) |
| 2031 leal 16(%edx),%edx |
| 2032 L097key_256a_cold: |
| 2033 shufps $16,%xmm0,%xmm4 |
| 2034 xorps %xmm4,%xmm0 |
| 2035 shufps $140,%xmm0,%xmm4 |
| 2036 xorps %xmm4,%xmm0 |
| 2037 shufps $255,%xmm1,%xmm1 |
| 2038 xorps %xmm1,%xmm0 |
| 2039 ret |
| 2040 .align 4,0x90 |
| 2041 L098key_256b: |
| 2042 movups %xmm0,(%edx) |
| 2043 leal 16(%edx),%edx |
| 2044 shufps $16,%xmm2,%xmm4 |
| 2045 xorps %xmm4,%xmm2 |
| 2046 shufps $140,%xmm2,%xmm4 |
| 2047 xorps %xmm4,%xmm2 |
| 2048 shufps $170,%xmm1,%xmm1 |
| 2049 xorps %xmm1,%xmm2 |
| 2050 ret |
| 2051 .align 2,0x90 |
| 2052 L086bad_pointer: |
| 2053 movl $-1,%eax |
| 2054 ret |
| 2055 .align 2,0x90 |
| 2056 L089bad_keybits: |
| 2057 movl $-2,%eax |
| 2058 ret |
| 2059 .globl _aesni_set_encrypt_key |
| 2060 .align 4 |
| 2061 _aesni_set_encrypt_key: |
| 2062 L_aesni_set_encrypt_key_begin: |
| 2063 movl 4(%esp),%eax |
| 2064 movl 8(%esp),%ecx |
| 2065 movl 12(%esp),%edx |
| 2066 call __aesni_set_encrypt_key |
| 2067 ret |
| 2068 .globl _aesni_set_decrypt_key |
| 2069 .align 4 |
| 2070 _aesni_set_decrypt_key: |
| 2071 L_aesni_set_decrypt_key_begin: |
| 2072 movl 4(%esp),%eax |
| 2073 movl 8(%esp),%ecx |
| 2074 movl 12(%esp),%edx |
| 2075 call __aesni_set_encrypt_key |
| 2076 movl 12(%esp),%edx |
| 2077 shll $4,%ecx |
| 2078 testl %eax,%eax |
| 2079 jnz L100dec_key_ret |
| 2080 leal 16(%edx,%ecx,1),%eax |
| 2081 movups (%edx),%xmm0 |
| 2082 movups (%eax),%xmm1 |
| 2083 movups %xmm0,(%eax) |
| 2084 movups %xmm1,(%edx) |
| 2085 leal 16(%edx),%edx |
| 2086 leal -16(%eax),%eax |
| 2087 L101dec_key_inverse: |
| 2088 movups (%edx),%xmm0 |
| 2089 movups (%eax),%xmm1 |
| 2090 .byte 102,15,56,219,192 |
| 2091 .byte 102,15,56,219,201 |
| 2092 leal 16(%edx),%edx |
| 2093 leal -16(%eax),%eax |
| 2094 movups %xmm0,16(%eax) |
| 2095 movups %xmm1,-16(%edx) |
| 2096 cmpl %edx,%eax |
| 2097 ja L101dec_key_inverse |
| 2098 movups (%edx),%xmm0 |
| 2099 .byte 102,15,56,219,192 |
| 2100 movups %xmm0,(%edx) |
| 2101 xorl %eax,%eax |
| 2102 L100dec_key_ret: |
| 2103 ret |
| 2104 .byte 65,69,83,32,102,111,114,32,73,110,116,101,108,32,65,69 |
| 2105 .byte 83,45,78,73,44,32,67,82,89,80,84,79,71,65,77,83 |
| 2106 .byte 32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115 |
| 2107 .byte 115,108,46,111,114,103,62,0 |
OLD | NEW |