third_party/boringssl/win-x86/crypto/aes/aesni-x86.asm

   1 %ifidn __OUTPUT_FORMAT__,obj
   2 section code    use32 class=code align=64
   3 %elifidn __OUTPUT_FORMAT__,win32
   4 %ifdef __YASM_VERSION_ID__
   5 %if __YASM_VERSION_ID__ < 01010000h
   6 %error yasm version 1.1.0 or later needed.
   7 %endif
   8 ; Yasm automatically includes .00 and complains about redefining it.
   9 ; https://www.tortall.net/projects/yasm/manual/html/objfmt-win32-safeseh.html
  10 %else
  11 $@feat.00 equ 1
  12 %endif
  13 section .text   code align=64
  14 %else
  15 section .text   code
  16 %endif
  17 global  _aesni_encrypt
  18 align   16
  19 _aesni_encrypt:
  20 L$_aesni_encrypt_begin:
  21         mov     eax,DWORD [4+esp]
  22         mov     edx,DWORD [12+esp]
  23         movups  xmm2,[eax]
  24         mov     ecx,DWORD [240+edx]
  25         mov     eax,DWORD [8+esp]
  26         movups  xmm0,[edx]
  27         movups  xmm1,[16+edx]
  28         lea     edx,[32+edx]
  29         xorps   xmm2,xmm0
  30 L$000enc1_loop_1:
  31 db      102,15,56,220,209
  32         dec     ecx
  33         movups  xmm1,[edx]
  34         lea     edx,[16+edx]
  35         jnz     NEAR L$000enc1_loop_1
  36 db      102,15,56,221,209
  37         movups  [eax],xmm2
  38         ret
  39 global  _aesni_decrypt
  40 align   16
  41 _aesni_decrypt:
  42 L$_aesni_decrypt_begin:
  43         mov     eax,DWORD [4+esp]
  44         mov     edx,DWORD [12+esp]
  45         movups  xmm2,[eax]
  46         mov     ecx,DWORD [240+edx]
  47         mov     eax,DWORD [8+esp]
  48         movups  xmm0,[edx]
  49         movups  xmm1,[16+edx]
  50         lea     edx,[32+edx]
  51         xorps   xmm2,xmm0
  52 L$001dec1_loop_2:
  53 db      102,15,56,222,209
  54         dec     ecx
  55         movups  xmm1,[edx]
  56         lea     edx,[16+edx]
  57         jnz     NEAR L$001dec1_loop_2
  58 db      102,15,56,223,209
  59         movups  [eax],xmm2
  60         ret
  61 align   16
  62 __aesni_encrypt2:
  63         movups  xmm0,[edx]
  64         shl     ecx,4
  65         movups  xmm1,[16+edx]
  66         xorps   xmm2,xmm0
  67         pxor    xmm3,xmm0
  68         movups  xmm0,[32+edx]
  69         lea     edx,[32+ecx*1+edx]
  70         neg     ecx
  71         add     ecx,16
  72 L$002enc2_loop:
  73 db      102,15,56,220,209
  74 db      102,15,56,220,217
  75         movups  xmm1,[ecx*1+edx]
  76         add     ecx,32
  77 db      102,15,56,220,208
  78 db      102,15,56,220,216
  79         movups  xmm0,[ecx*1+edx-16]
  80         jnz     NEAR L$002enc2_loop
  81 db      102,15,56,220,209
  82 db      102,15,56,220,217
  83 db      102,15,56,221,208
  84 db      102,15,56,221,216
  85         ret
  86 align   16
  87 __aesni_decrypt2:
  88         movups  xmm0,[edx]
  89         shl     ecx,4
  90         movups  xmm1,[16+edx]
  91         xorps   xmm2,xmm0
  92         pxor    xmm3,xmm0
  93         movups  xmm0,[32+edx]
  94         lea     edx,[32+ecx*1+edx]
  95         neg     ecx
  96         add     ecx,16
  97 L$003dec2_loop:
  98 db      102,15,56,222,209
  99 db      102,15,56,222,217
 100         movups  xmm1,[ecx*1+edx]
 101         add     ecx,32
 102 db      102,15,56,222,208
 103 db      102,15,56,222,216
 104         movups  xmm0,[ecx*1+edx-16]
 105         jnz     NEAR L$003dec2_loop
 106 db      102,15,56,222,209
 107 db      102,15,56,222,217
 108 db      102,15,56,223,208
 109 db      102,15,56,223,216
 110         ret
 111 align   16
 112 __aesni_encrypt3:
 113         movups  xmm0,[edx]
 114         shl     ecx,4
 115         movups  xmm1,[16+edx]
 116         xorps   xmm2,xmm0
 117         pxor    xmm3,xmm0
 118         pxor    xmm4,xmm0
 119         movups  xmm0,[32+edx]
 120         lea     edx,[32+ecx*1+edx]
 121         neg     ecx
 122         add     ecx,16
 123 L$004enc3_loop:
 124 db      102,15,56,220,209
 125 db      102,15,56,220,217
 126 db      102,15,56,220,225
 127         movups  xmm1,[ecx*1+edx]
 128         add     ecx,32
 129 db      102,15,56,220,208
 130 db      102,15,56,220,216
 131 db      102,15,56,220,224
 132         movups  xmm0,[ecx*1+edx-16]
 133         jnz     NEAR L$004enc3_loop
 134 db      102,15,56,220,209
 135 db      102,15,56,220,217
 136 db      102,15,56,220,225
 137 db      102,15,56,221,208
 138 db      102,15,56,221,216
 139 db      102,15,56,221,224
 140         ret
 141 align   16
 142 __aesni_decrypt3:
 143         movups  xmm0,[edx]
 144         shl     ecx,4
 145         movups  xmm1,[16+edx]
 146         xorps   xmm2,xmm0
 147         pxor    xmm3,xmm0
 148         pxor    xmm4,xmm0
 149         movups  xmm0,[32+edx]
 150         lea     edx,[32+ecx*1+edx]
 151         neg     ecx
 152         add     ecx,16
 153 L$005dec3_loop:
 154 db      102,15,56,222,209
 155 db      102,15,56,222,217
 156 db      102,15,56,222,225
 157         movups  xmm1,[ecx*1+edx]
 158         add     ecx,32
 159 db      102,15,56,222,208
 160 db      102,15,56,222,216
 161 db      102,15,56,222,224
 162         movups  xmm0,[ecx*1+edx-16]
 163         jnz     NEAR L$005dec3_loop
 164 db      102,15,56,222,209
 165 db      102,15,56,222,217
 166 db      102,15,56,222,225
 167 db      102,15,56,223,208
 168 db      102,15,56,223,216
 169 db      102,15,56,223,224
 170         ret
 171 align   16
 172 __aesni_encrypt4:
 173         movups  xmm0,[edx]
 174         movups  xmm1,[16+edx]
 175         shl     ecx,4
 176         xorps   xmm2,xmm0
 177         pxor    xmm3,xmm0
 178         pxor    xmm4,xmm0
 179         pxor    xmm5,xmm0
 180         movups  xmm0,[32+edx]
 181         lea     edx,[32+ecx*1+edx]
 182         neg     ecx
 183 db      15,31,64,0
 184         add     ecx,16
 185 L$006enc4_loop:
 186 db      102,15,56,220,209
 187 db      102,15,56,220,217
 188 db      102,15,56,220,225
 189 db      102,15,56,220,233
 190         movups  xmm1,[ecx*1+edx]
 191         add     ecx,32
 192 db      102,15,56,220,208
 193 db      102,15,56,220,216
 194 db      102,15,56,220,224
 195 db      102,15,56,220,232
 196         movups  xmm0,[ecx*1+edx-16]
 197         jnz     NEAR L$006enc4_loop
 198 db      102,15,56,220,209
 199 db      102,15,56,220,217
 200 db      102,15,56,220,225
 201 db      102,15,56,220,233
 202 db      102,15,56,221,208
 203 db      102,15,56,221,216
 204 db      102,15,56,221,224
 205 db      102,15,56,221,232
 206         ret
 207 align   16
 208 __aesni_decrypt4:
 209         movups  xmm0,[edx]
 210         movups  xmm1,[16+edx]
 211         shl     ecx,4
 212         xorps   xmm2,xmm0
 213         pxor    xmm3,xmm0
 214         pxor    xmm4,xmm0
 215         pxor    xmm5,xmm0
 216         movups  xmm0,[32+edx]
 217         lea     edx,[32+ecx*1+edx]
 218         neg     ecx
 219 db      15,31,64,0
 220         add     ecx,16
 221 L$007dec4_loop:
 222 db      102,15,56,222,209
 223 db      102,15,56,222,217
 224 db      102,15,56,222,225
 225 db      102,15,56,222,233
 226         movups  xmm1,[ecx*1+edx]
 227         add     ecx,32
 228 db      102,15,56,222,208
 229 db      102,15,56,222,216
 230 db      102,15,56,222,224
 231 db      102,15,56,222,232
 232         movups  xmm0,[ecx*1+edx-16]
 233         jnz     NEAR L$007dec4_loop
 234 db      102,15,56,222,209
 235 db      102,15,56,222,217
 236 db      102,15,56,222,225
 237 db      102,15,56,222,233
 238 db      102,15,56,223,208
 239 db      102,15,56,223,216
 240 db      102,15,56,223,224
 241 db      102,15,56,223,232
 242         ret
 243 align   16
 244 __aesni_encrypt6:
 245         movups  xmm0,[edx]
 246         shl     ecx,4
 247         movups  xmm1,[16+edx]
 248         xorps   xmm2,xmm0
 249         pxor    xmm3,xmm0
 250         pxor    xmm4,xmm0
 251 db      102,15,56,220,209
 252         pxor    xmm5,xmm0
 253         pxor    xmm6,xmm0
 254 db      102,15,56,220,217
 255         lea     edx,[32+ecx*1+edx]
 256         neg     ecx
 257 db      102,15,56,220,225
 258         pxor    xmm7,xmm0
 259         add     ecx,16
 260 db      102,15,56,220,233
 261 db      102,15,56,220,241
 262 db      102,15,56,220,249
 263         movups  xmm0,[ecx*1+edx-16]
 264         jmp     NEAR L$_aesni_encrypt6_enter
 265 align   16
 266 L$008enc6_loop:
 267 db      102,15,56,220,209
 268 db      102,15,56,220,217
 269 db      102,15,56,220,225
 270 db      102,15,56,220,233
 271 db      102,15,56,220,241
 272 db      102,15,56,220,249
 273 L$_aesni_encrypt6_enter:
 274         movups  xmm1,[ecx*1+edx]
 275         add     ecx,32
 276 db      102,15,56,220,208
 277 db      102,15,56,220,216
 278 db      102,15,56,220,224
 279 db      102,15,56,220,232
 280 db      102,15,56,220,240
 281 db      102,15,56,220,248
 282         movups  xmm0,[ecx*1+edx-16]
 283         jnz     NEAR L$008enc6_loop
 284 db      102,15,56,220,209
 285 db      102,15,56,220,217
 286 db      102,15,56,220,225
 287 db      102,15,56,220,233
 288 db      102,15,56,220,241
 289 db      102,15,56,220,249
 290 db      102,15,56,221,208
 291 db      102,15,56,221,216
 292 db      102,15,56,221,224
 293 db      102,15,56,221,232
 294 db      102,15,56,221,240
 295 db      102,15,56,221,248
 296         ret
 297 align   16
 298 __aesni_decrypt6:
 299         movups  xmm0,[edx]
 300         shl     ecx,4
 301         movups  xmm1,[16+edx]
 302         xorps   xmm2,xmm0
 303         pxor    xmm3,xmm0
 304         pxor    xmm4,xmm0
 305 db      102,15,56,222,209
 306         pxor    xmm5,xmm0
 307         pxor    xmm6,xmm0
 308 db      102,15,56,222,217
 309         lea     edx,[32+ecx*1+edx]
 310         neg     ecx
 311 db      102,15,56,222,225
 312         pxor    xmm7,xmm0
 313         add     ecx,16
 314 db      102,15,56,222,233
 315 db      102,15,56,222,241
 316 db      102,15,56,222,249
 317         movups  xmm0,[ecx*1+edx-16]
 318         jmp     NEAR L$_aesni_decrypt6_enter
 319 align   16
 320 L$009dec6_loop:
 321 db      102,15,56,222,209
 322 db      102,15,56,222,217
 323 db      102,15,56,222,225
 324 db      102,15,56,222,233
 325 db      102,15,56,222,241
 326 db      102,15,56,222,249
 327 L$_aesni_decrypt6_enter:
 328         movups  xmm1,[ecx*1+edx]
 329         add     ecx,32
 330 db      102,15,56,222,208
 331 db      102,15,56,222,216
 332 db      102,15,56,222,224
 333 db      102,15,56,222,232
 334 db      102,15,56,222,240
 335 db      102,15,56,222,248
 336         movups  xmm0,[ecx*1+edx-16]
 337         jnz     NEAR L$009dec6_loop
 338 db      102,15,56,222,209
 339 db      102,15,56,222,217
 340 db      102,15,56,222,225
 341 db      102,15,56,222,233
 342 db      102,15,56,222,241
 343 db      102,15,56,222,249
 344 db      102,15,56,223,208
 345 db      102,15,56,223,216
 346 db      102,15,56,223,224
 347 db      102,15,56,223,232
 348 db      102,15,56,223,240
 349 db      102,15,56,223,248
 350         ret
 351 global  _aesni_ecb_encrypt
 352 align   16
 353 _aesni_ecb_encrypt:
 354 L$_aesni_ecb_encrypt_begin:
 355         push    ebp
 356         push    ebx
 357         push    esi
 358         push    edi
 359         mov     esi,DWORD [20+esp]
 360         mov     edi,DWORD [24+esp]
 361         mov     eax,DWORD [28+esp]
 362         mov     edx,DWORD [32+esp]
 363         mov     ebx,DWORD [36+esp]
 364         and     eax,-16
 365         jz      NEAR L$010ecb_ret
 366         mov     ecx,DWORD [240+edx]
 367         test    ebx,ebx
 368         jz      NEAR L$011ecb_decrypt
 369         mov     ebp,edx
 370         mov     ebx,ecx
 371         cmp     eax,96
 372         jb      NEAR L$012ecb_enc_tail
 373         movdqu  xmm2,[esi]
 374         movdqu  xmm3,[16+esi]
 375         movdqu  xmm4,[32+esi]
 376         movdqu  xmm5,[48+esi]
 377         movdqu  xmm6,[64+esi]
 378         movdqu  xmm7,[80+esi]
 379         lea     esi,[96+esi]
 380         sub     eax,96
 381         jmp     NEAR L$013ecb_enc_loop6_enter
 382 align   16
 383 L$014ecb_enc_loop6:
 384         movups  [edi],xmm2
 385         movdqu  xmm2,[esi]
 386         movups  [16+edi],xmm3
 387         movdqu  xmm3,[16+esi]
 388         movups  [32+edi],xmm4
 389         movdqu  xmm4,[32+esi]
 390         movups  [48+edi],xmm5
 391         movdqu  xmm5,[48+esi]
 392         movups  [64+edi],xmm6
 393         movdqu  xmm6,[64+esi]
 394         movups  [80+edi],xmm7
 395         lea     edi,[96+edi]
 396         movdqu  xmm7,[80+esi]
 397         lea     esi,[96+esi]
 398 L$013ecb_enc_loop6_enter:
 399         call    __aesni_encrypt6
 400         mov     edx,ebp
 401         mov     ecx,ebx
 402         sub     eax,96
 403         jnc     NEAR L$014ecb_enc_loop6
 404         movups  [edi],xmm2
 405         movups  [16+edi],xmm3
 406         movups  [32+edi],xmm4
 407         movups  [48+edi],xmm5
 408         movups  [64+edi],xmm6
 409         movups  [80+edi],xmm7
 410         lea     edi,[96+edi]
 411         add     eax,96
 412         jz      NEAR L$010ecb_ret
 413 L$012ecb_enc_tail:
 414         movups  xmm2,[esi]
 415         cmp     eax,32
 416         jb      NEAR L$015ecb_enc_one
 417         movups  xmm3,[16+esi]
 418         je      NEAR L$016ecb_enc_two
 419         movups  xmm4,[32+esi]
 420         cmp     eax,64
 421         jb      NEAR L$017ecb_enc_three
 422         movups  xmm5,[48+esi]
 423         je      NEAR L$018ecb_enc_four
 424         movups  xmm6,[64+esi]
 425         xorps   xmm7,xmm7
 426         call    __aesni_encrypt6
 427         movups  [edi],xmm2
 428         movups  [16+edi],xmm3
 429         movups  [32+edi],xmm4
 430         movups  [48+edi],xmm5
 431         movups  [64+edi],xmm6
 432         jmp     NEAR L$010ecb_ret
 433 align   16
 434 L$015ecb_enc_one:
 435         movups  xmm0,[edx]
 436         movups  xmm1,[16+edx]
 437         lea     edx,[32+edx]
 438         xorps   xmm2,xmm0
 439 L$019enc1_loop_3:
 440 db      102,15,56,220,209
 441         dec     ecx
 442         movups  xmm1,[edx]
 443         lea     edx,[16+edx]
 444         jnz     NEAR L$019enc1_loop_3
 445 db      102,15,56,221,209
 446         movups  [edi],xmm2
 447         jmp     NEAR L$010ecb_ret
 448 align   16
 449 L$016ecb_enc_two:
 450         call    __aesni_encrypt2
 451         movups  [edi],xmm2
 452         movups  [16+edi],xmm3
 453         jmp     NEAR L$010ecb_ret
 454 align   16
 455 L$017ecb_enc_three:
 456         call    __aesni_encrypt3
 457         movups  [edi],xmm2
 458         movups  [16+edi],xmm3
 459         movups  [32+edi],xmm4
 460         jmp     NEAR L$010ecb_ret
 461 align   16
 462 L$018ecb_enc_four:
 463         call    __aesni_encrypt4
 464         movups  [edi],xmm2
 465         movups  [16+edi],xmm3
 466         movups  [32+edi],xmm4
 467         movups  [48+edi],xmm5
 468         jmp     NEAR L$010ecb_ret
 469 align   16
 470 L$011ecb_decrypt:
 471         mov     ebp,edx
 472         mov     ebx,ecx
 473         cmp     eax,96
 474         jb      NEAR L$020ecb_dec_tail
 475         movdqu  xmm2,[esi]
 476         movdqu  xmm3,[16+esi]
 477         movdqu  xmm4,[32+esi]
 478         movdqu  xmm5,[48+esi]
 479         movdqu  xmm6,[64+esi]
 480         movdqu  xmm7,[80+esi]
 481         lea     esi,[96+esi]
 482         sub     eax,96
 483         jmp     NEAR L$021ecb_dec_loop6_enter
 484 align   16
 485 L$022ecb_dec_loop6:
 486         movups  [edi],xmm2
 487         movdqu  xmm2,[esi]
 488         movups  [16+edi],xmm3
 489         movdqu  xmm3,[16+esi]
 490         movups  [32+edi],xmm4
 491         movdqu  xmm4,[32+esi]
 492         movups  [48+edi],xmm5
 493         movdqu  xmm5,[48+esi]
 494         movups  [64+edi],xmm6
 495         movdqu  xmm6,[64+esi]
 496         movups  [80+edi],xmm7
 497         lea     edi,[96+edi]
 498         movdqu  xmm7,[80+esi]
 499         lea     esi,[96+esi]
 500 L$021ecb_dec_loop6_enter:
 501         call    __aesni_decrypt6
 502         mov     edx,ebp
 503         mov     ecx,ebx
 504         sub     eax,96
 505         jnc     NEAR L$022ecb_dec_loop6
 506         movups  [edi],xmm2
 507         movups  [16+edi],xmm3
 508         movups  [32+edi],xmm4
 509         movups  [48+edi],xmm5
 510         movups  [64+edi],xmm6
 511         movups  [80+edi],xmm7
 512         lea     edi,[96+edi]
 513         add     eax,96
 514         jz      NEAR L$010ecb_ret
 515 L$020ecb_dec_tail:
 516         movups  xmm2,[esi]
 517         cmp     eax,32
 518         jb      NEAR L$023ecb_dec_one
 519         movups  xmm3,[16+esi]
 520         je      NEAR L$024ecb_dec_two
 521         movups  xmm4,[32+esi]
 522         cmp     eax,64
 523         jb      NEAR L$025ecb_dec_three
 524         movups  xmm5,[48+esi]
 525         je      NEAR L$026ecb_dec_four
 526         movups  xmm6,[64+esi]
 527         xorps   xmm7,xmm7
 528         call    __aesni_decrypt6
 529         movups  [edi],xmm2
 530         movups  [16+edi],xmm3
 531         movups  [32+edi],xmm4
 532         movups  [48+edi],xmm5
 533         movups  [64+edi],xmm6
 534         jmp     NEAR L$010ecb_ret
 535 align   16
 536 L$023ecb_dec_one:
 537         movups  xmm0,[edx]
 538         movups  xmm1,[16+edx]
 539         lea     edx,[32+edx]
 540         xorps   xmm2,xmm0
 541 L$027dec1_loop_4:
 542 db      102,15,56,222,209
 543         dec     ecx
 544         movups  xmm1,[edx]
 545         lea     edx,[16+edx]
 546         jnz     NEAR L$027dec1_loop_4
 547 db      102,15,56,223,209
 548         movups  [edi],xmm2
 549         jmp     NEAR L$010ecb_ret
 550 align   16
 551 L$024ecb_dec_two:
 552         call    __aesni_decrypt2
 553         movups  [edi],xmm2
 554         movups  [16+edi],xmm3
 555         jmp     NEAR L$010ecb_ret
 556 align   16
 557 L$025ecb_dec_three:
 558         call    __aesni_decrypt3
 559         movups  [edi],xmm2
 560         movups  [16+edi],xmm3
 561         movups  [32+edi],xmm4
 562         jmp     NEAR L$010ecb_ret
 563 align   16
 564 L$026ecb_dec_four:
 565         call    __aesni_decrypt4
 566         movups  [edi],xmm2
 567         movups  [16+edi],xmm3
 568         movups  [32+edi],xmm4
 569         movups  [48+edi],xmm5
 570 L$010ecb_ret:
 571         pop     edi
 572         pop     esi
 573         pop     ebx
 574         pop     ebp
 575         ret
 576 global  _aesni_ccm64_encrypt_blocks
 577 align   16
 578 _aesni_ccm64_encrypt_blocks:
 579 L$_aesni_ccm64_encrypt_blocks_begin:
 580         push    ebp
 581         push    ebx
 582         push    esi
 583         push    edi
 584         mov     esi,DWORD [20+esp]
 585         mov     edi,DWORD [24+esp]
 586         mov     eax,DWORD [28+esp]
 587         mov     edx,DWORD [32+esp]
 588         mov     ebx,DWORD [36+esp]
 589         mov     ecx,DWORD [40+esp]
 590         mov     ebp,esp
 591         sub     esp,60
 592         and     esp,-16
 593         mov     DWORD [48+esp],ebp
 594         movdqu  xmm7,[ebx]
 595         movdqu  xmm3,[ecx]
 596         mov     ecx,DWORD [240+edx]
 597         mov     DWORD [esp],202182159
 598         mov     DWORD [4+esp],134810123
 599         mov     DWORD [8+esp],67438087
 600         mov     DWORD [12+esp],66051
 601         mov     ebx,1
 602         xor     ebp,ebp
 603         mov     DWORD [16+esp],ebx
 604         mov     DWORD [20+esp],ebp
 605         mov     DWORD [24+esp],ebp
 606         mov     DWORD [28+esp],ebp
 607         shl     ecx,4
 608         mov     ebx,16
 609         lea     ebp,[edx]
 610         movdqa  xmm5,[esp]
 611         movdqa  xmm2,xmm7
 612         lea     edx,[32+ecx*1+edx]
 613         sub     ebx,ecx
 614 db      102,15,56,0,253
 615 L$028ccm64_enc_outer:
 616         movups  xmm0,[ebp]
 617         mov     ecx,ebx
 618         movups  xmm6,[esi]
 619         xorps   xmm2,xmm0
 620         movups  xmm1,[16+ebp]
 621         xorps   xmm0,xmm6
 622         xorps   xmm3,xmm0
 623         movups  xmm0,[32+ebp]
 624 L$029ccm64_enc2_loop:
 625 db      102,15,56,220,209
 626 db      102,15,56,220,217
 627         movups  xmm1,[ecx*1+edx]
 628         add     ecx,32
 629 db      102,15,56,220,208
 630 db      102,15,56,220,216
 631         movups  xmm0,[ecx*1+edx-16]
 632         jnz     NEAR L$029ccm64_enc2_loop
 633 db      102,15,56,220,209
 634 db      102,15,56,220,217
 635         paddq   xmm7,[16+esp]
 636         dec     eax
 637 db      102,15,56,221,208
 638 db      102,15,56,221,216
 639         lea     esi,[16+esi]
 640         xorps   xmm6,xmm2
 641         movdqa  xmm2,xmm7
 642         movups  [edi],xmm6
 643 db      102,15,56,0,213
 644         lea     edi,[16+edi]
 645         jnz     NEAR L$028ccm64_enc_outer
 646         mov     esp,DWORD [48+esp]
 647         mov     edi,DWORD [40+esp]
 648         movups  [edi],xmm3
 649         pop     edi
 650         pop     esi
 651         pop     ebx
 652         pop     ebp
 653         ret
 654 global  _aesni_ccm64_decrypt_blocks
 655 align   16
 656 _aesni_ccm64_decrypt_blocks:
 657 L$_aesni_ccm64_decrypt_blocks_begin:
 658         push    ebp
 659         push    ebx
 660         push    esi
 661         push    edi
 662         mov     esi,DWORD [20+esp]
 663         mov     edi,DWORD [24+esp]
 664         mov     eax,DWORD [28+esp]
 665         mov     edx,DWORD [32+esp]
 666         mov     ebx,DWORD [36+esp]
 667         mov     ecx,DWORD [40+esp]
 668         mov     ebp,esp
 669         sub     esp,60
 670         and     esp,-16
 671         mov     DWORD [48+esp],ebp
 672         movdqu  xmm7,[ebx]
 673         movdqu  xmm3,[ecx]
 674         mov     ecx,DWORD [240+edx]
 675         mov     DWORD [esp],202182159
 676         mov     DWORD [4+esp],134810123
 677         mov     DWORD [8+esp],67438087
 678         mov     DWORD [12+esp],66051
 679         mov     ebx,1
 680         xor     ebp,ebp
 681         mov     DWORD [16+esp],ebx
 682         mov     DWORD [20+esp],ebp
 683         mov     DWORD [24+esp],ebp
 684         mov     DWORD [28+esp],ebp
 685         movdqa  xmm5,[esp]
 686         movdqa  xmm2,xmm7
 687         mov     ebp,edx
 688         mov     ebx,ecx
 689 db      102,15,56,0,253
 690         movups  xmm0,[edx]
 691         movups  xmm1,[16+edx]
 692         lea     edx,[32+edx]
 693         xorps   xmm2,xmm0
 694 L$030enc1_loop_5:
 695 db      102,15,56,220,209
 696         dec     ecx
 697         movups  xmm1,[edx]
 698         lea     edx,[16+edx]
 699         jnz     NEAR L$030enc1_loop_5
 700 db      102,15,56,221,209
 701         shl     ebx,4
 702         mov     ecx,16
 703         movups  xmm6,[esi]
 704         paddq   xmm7,[16+esp]
 705         lea     esi,[16+esi]
 706         sub     ecx,ebx
 707         lea     edx,[32+ebx*1+ebp]
 708         mov     ebx,ecx
 709         jmp     NEAR L$031ccm64_dec_outer
 710 align   16
 711 L$031ccm64_dec_outer:
 712         xorps   xmm6,xmm2
 713         movdqa  xmm2,xmm7
 714         movups  [edi],xmm6
 715         lea     edi,[16+edi]
 716 db      102,15,56,0,213
 717         sub     eax,1
 718         jz      NEAR L$032ccm64_dec_break
 719         movups  xmm0,[ebp]
 720         mov     ecx,ebx
 721         movups  xmm1,[16+ebp]
 722         xorps   xmm6,xmm0
 723         xorps   xmm2,xmm0
 724         xorps   xmm3,xmm6
 725         movups  xmm0,[32+ebp]
 726 L$033ccm64_dec2_loop:
 727 db      102,15,56,220,209
 728 db      102,15,56,220,217
 729         movups  xmm1,[ecx*1+edx]
 730         add     ecx,32
 731 db      102,15,56,220,208
 732 db      102,15,56,220,216
 733         movups  xmm0,[ecx*1+edx-16]
 734         jnz     NEAR L$033ccm64_dec2_loop
 735         movups  xmm6,[esi]
 736         paddq   xmm7,[16+esp]
 737 db      102,15,56,220,209
 738 db      102,15,56,220,217
 739 db      102,15,56,221,208
 740 db      102,15,56,221,216
 741         lea     esi,[16+esi]
 742         jmp     NEAR L$031ccm64_dec_outer
 743 align   16
 744 L$032ccm64_dec_break:
 745         mov     ecx,DWORD [240+ebp]
 746         mov     edx,ebp
 747         movups  xmm0,[edx]
 748         movups  xmm1,[16+edx]
 749         xorps   xmm6,xmm0
 750         lea     edx,[32+edx]
 751         xorps   xmm3,xmm6
 752 L$034enc1_loop_6:
 753 db      102,15,56,220,217
 754         dec     ecx
 755         movups  xmm1,[edx]
 756         lea     edx,[16+edx]
 757         jnz     NEAR L$034enc1_loop_6
 758 db      102,15,56,221,217
 759         mov     esp,DWORD [48+esp]
 760         mov     edi,DWORD [40+esp]
 761         movups  [edi],xmm3
 762         pop     edi
 763         pop     esi
 764         pop     ebx
 765         pop     ebp
 766         ret
 767 global  _aesni_ctr32_encrypt_blocks
 768 align   16
 769 _aesni_ctr32_encrypt_blocks:
 770 L$_aesni_ctr32_encrypt_blocks_begin:
 771         push    ebp
 772         push    ebx
 773         push    esi
 774         push    edi
 775         mov     esi,DWORD [20+esp]
 776         mov     edi,DWORD [24+esp]
 777         mov     eax,DWORD [28+esp]
 778         mov     edx,DWORD [32+esp]
 779         mov     ebx,DWORD [36+esp]
 780         mov     ebp,esp
 781         sub     esp,88
 782         and     esp,-16
 783         mov     DWORD [80+esp],ebp
 784         cmp     eax,1
 785         je      NEAR L$035ctr32_one_shortcut
 786         movdqu  xmm7,[ebx]
 787         mov     DWORD [esp],202182159
 788         mov     DWORD [4+esp],134810123
 789         mov     DWORD [8+esp],67438087
 790         mov     DWORD [12+esp],66051
 791         mov     ecx,6
 792         xor     ebp,ebp
 793         mov     DWORD [16+esp],ecx
 794         mov     DWORD [20+esp],ecx
 795         mov     DWORD [24+esp],ecx
 796         mov     DWORD [28+esp],ebp
 797 db      102,15,58,22,251,3
 798 db      102,15,58,34,253,3
 799         mov     ecx,DWORD [240+edx]
 800         bswap   ebx
 801         pxor    xmm0,xmm0
 802         pxor    xmm1,xmm1
 803         movdqa  xmm2,[esp]
 804 db      102,15,58,34,195,0
 805         lea     ebp,[3+ebx]
 806 db      102,15,58,34,205,0
 807         inc     ebx
 808 db      102,15,58,34,195,1
 809         inc     ebp
 810 db      102,15,58,34,205,1
 811         inc     ebx
 812 db      102,15,58,34,195,2
 813         inc     ebp
 814 db      102,15,58,34,205,2
 815         movdqa  [48+esp],xmm0
 816 db      102,15,56,0,194
 817         movdqu  xmm6,[edx]
 818         movdqa  [64+esp],xmm1
 819 db      102,15,56,0,202
 820         pshufd  xmm2,xmm0,192
 821         pshufd  xmm3,xmm0,128
 822         cmp     eax,6
 823         jb      NEAR L$036ctr32_tail
 824         pxor    xmm7,xmm6
 825         shl     ecx,4
 826         mov     ebx,16
 827         movdqa  [32+esp],xmm7
 828         mov     ebp,edx
 829         sub     ebx,ecx
 830         lea     edx,[32+ecx*1+edx]
 831         sub     eax,6
 832         jmp     NEAR L$037ctr32_loop6
 833 align   16
 834 L$037ctr32_loop6:
 835         pshufd  xmm4,xmm0,64
 836         movdqa  xmm0,[32+esp]
 837         pshufd  xmm5,xmm1,192
 838         pxor    xmm2,xmm0
 839         pshufd  xmm6,xmm1,128
 840         pxor    xmm3,xmm0
 841         pshufd  xmm7,xmm1,64
 842         movups  xmm1,[16+ebp]
 843         pxor    xmm4,xmm0
 844         pxor    xmm5,xmm0
 845 db      102,15,56,220,209
 846         pxor    xmm6,xmm0
 847         pxor    xmm7,xmm0
 848 db      102,15,56,220,217
 849         movups  xmm0,[32+ebp]
 850         mov     ecx,ebx
 851 db      102,15,56,220,225
 852 db      102,15,56,220,233
 853 db      102,15,56,220,241
 854 db      102,15,56,220,249
 855         call    L$_aesni_encrypt6_enter
 856         movups  xmm1,[esi]
 857         movups  xmm0,[16+esi]
 858         xorps   xmm2,xmm1
 859         movups  xmm1,[32+esi]
 860         xorps   xmm3,xmm0
 861         movups  [edi],xmm2
 862         movdqa  xmm0,[16+esp]
 863         xorps   xmm4,xmm1
 864         movdqa  xmm1,[64+esp]
 865         movups  [16+edi],xmm3
 866         movups  [32+edi],xmm4
 867         paddd   xmm1,xmm0
 868         paddd   xmm0,[48+esp]
 869         movdqa  xmm2,[esp]
 870         movups  xmm3,[48+esi]
 871         movups  xmm4,[64+esi]
 872         xorps   xmm5,xmm3
 873         movups  xmm3,[80+esi]
 874         lea     esi,[96+esi]
 875         movdqa  [48+esp],xmm0
 876 db      102,15,56,0,194
 877         xorps   xmm6,xmm4
 878         movups  [48+edi],xmm5
 879         xorps   xmm7,xmm3
 880         movdqa  [64+esp],xmm1
 881 db      102,15,56,0,202
 882         movups  [64+edi],xmm6
 883         pshufd  xmm2,xmm0,192
 884         movups  [80+edi],xmm7
 885         lea     edi,[96+edi]
 886         pshufd  xmm3,xmm0,128
 887         sub     eax,6
 888         jnc     NEAR L$037ctr32_loop6
 889         add     eax,6
 890         jz      NEAR L$038ctr32_ret
 891         movdqu  xmm7,[ebp]
 892         mov     edx,ebp
 893         pxor    xmm7,[32+esp]
 894         mov     ecx,DWORD [240+ebp]
 895 L$036ctr32_tail:
 896         por     xmm2,xmm7
 897         cmp     eax,2
 898         jb      NEAR L$039ctr32_one
 899         pshufd  xmm4,xmm0,64
 900         por     xmm3,xmm7
 901         je      NEAR L$040ctr32_two
 902         pshufd  xmm5,xmm1,192
 903         por     xmm4,xmm7
 904         cmp     eax,4
 905         jb      NEAR L$041ctr32_three
 906         pshufd  xmm6,xmm1,128
 907         por     xmm5,xmm7
 908         je      NEAR L$042ctr32_four
 909         por     xmm6,xmm7
 910         call    __aesni_encrypt6
 911         movups  xmm1,[esi]
 912         movups  xmm0,[16+esi]
 913         xorps   xmm2,xmm1
 914         movups  xmm1,[32+esi]
 915         xorps   xmm3,xmm0
 916         movups  xmm0,[48+esi]
 917         xorps   xmm4,xmm1
 918         movups  xmm1,[64+esi]
 919         xorps   xmm5,xmm0
 920         movups  [edi],xmm2
 921         xorps   xmm6,xmm1
 922         movups  [16+edi],xmm3
 923         movups  [32+edi],xmm4
 924         movups  [48+edi],xmm5
 925         movups  [64+edi],xmm6
 926         jmp     NEAR L$038ctr32_ret
 927 align   16
 928 L$035ctr32_one_shortcut:
 929         movups  xmm2,[ebx]
 930         mov     ecx,DWORD [240+edx]
 931 L$039ctr32_one:
 932         movups  xmm0,[edx]
 933         movups  xmm1,[16+edx]
 934         lea     edx,[32+edx]
 935         xorps   xmm2,xmm0
 936 L$043enc1_loop_7:
 937 db      102,15,56,220,209
 938         dec     ecx
 939         movups  xmm1,[edx]
 940         lea     edx,[16+edx]
 941         jnz     NEAR L$043enc1_loop_7
 942 db      102,15,56,221,209
 943         movups  xmm6,[esi]
 944         xorps   xmm6,xmm2
 945         movups  [edi],xmm6
 946         jmp     NEAR L$038ctr32_ret
 947 align   16
 948 L$040ctr32_two:
 949         call    __aesni_encrypt2
 950         movups  xmm5,[esi]
 951         movups  xmm6,[16+esi]
 952         xorps   xmm2,xmm5
 953         xorps   xmm3,xmm6
 954         movups  [edi],xmm2
 955         movups  [16+edi],xmm3
 956         jmp     NEAR L$038ctr32_ret
 957 align   16
 958 L$041ctr32_three:
 959         call    __aesni_encrypt3
 960         movups  xmm5,[esi]
 961         movups  xmm6,[16+esi]
 962         xorps   xmm2,xmm5
 963         movups  xmm7,[32+esi]
 964         xorps   xmm3,xmm6
 965         movups  [edi],xmm2
 966         xorps   xmm4,xmm7
 967         movups  [16+edi],xmm3
 968         movups  [32+edi],xmm4
 969         jmp     NEAR L$038ctr32_ret
 970 align   16
 971 L$042ctr32_four:
 972         call    __aesni_encrypt4
 973         movups  xmm6,[esi]
 974         movups  xmm7,[16+esi]
 975         movups  xmm1,[32+esi]
 976         xorps   xmm2,xmm6
 977         movups  xmm0,[48+esi]
 978         xorps   xmm3,xmm7
 979         movups  [edi],xmm2
 980         xorps   xmm4,xmm1
 981         movups  [16+edi],xmm3
 982         xorps   xmm5,xmm0
 983         movups  [32+edi],xmm4
 984         movups  [48+edi],xmm5
 985 L$038ctr32_ret:
 986         mov     esp,DWORD [80+esp]
 987         pop     edi
 988         pop     esi
 989         pop     ebx
 990         pop     ebp
 991         ret
 992 global  _aesni_xts_encrypt
 993 align   16
 994 _aesni_xts_encrypt:
 995 L$_aesni_xts_encrypt_begin:
 996         push    ebp
 997         push    ebx
 998         push    esi
 999         push    edi
1000         mov     edx,DWORD [36+esp]
1001         mov     esi,DWORD [40+esp]
1002         mov     ecx,DWORD [240+edx]
1003         movups  xmm2,[esi]
1004         movups  xmm0,[edx]
1005         movups  xmm1,[16+edx]
1006         lea     edx,[32+edx]
1007         xorps   xmm2,xmm0
1008 L$044enc1_loop_8:
1009 db      102,15,56,220,209
1010         dec     ecx
1011         movups  xmm1,[edx]
1012         lea     edx,[16+edx]
1013         jnz     NEAR L$044enc1_loop_8
1014 db      102,15,56,221,209
1015         mov     esi,DWORD [20+esp]
1016         mov     edi,DWORD [24+esp]
1017         mov     eax,DWORD [28+esp]
1018         mov     edx,DWORD [32+esp]
1019         mov     ebp,esp
1020         sub     esp,120
1021         mov     ecx,DWORD [240+edx]
1022         and     esp,-16
1023         mov     DWORD [96+esp],135
1024         mov     DWORD [100+esp],0
1025         mov     DWORD [104+esp],1
1026         mov     DWORD [108+esp],0
1027         mov     DWORD [112+esp],eax
1028         mov     DWORD [116+esp],ebp
1029         movdqa  xmm1,xmm2
1030         pxor    xmm0,xmm0
1031         movdqa  xmm3,[96+esp]
1032         pcmpgtd xmm0,xmm1
1033         and     eax,-16
1034         mov     ebp,edx
1035         mov     ebx,ecx
1036         sub     eax,96
1037         jc      NEAR L$045xts_enc_short
1038         shl     ecx,4
1039         mov     ebx,16
1040         sub     ebx,ecx
1041         lea     edx,[32+ecx*1+edx]
1042         jmp     NEAR L$046xts_enc_loop6
1043 align   16
1044 L$046xts_enc_loop6:
1045         pshufd  xmm2,xmm0,19
1046         pxor    xmm0,xmm0
1047         movdqa  [esp],xmm1
1048         paddq   xmm1,xmm1
1049         pand    xmm2,xmm3
1050         pcmpgtd xmm0,xmm1
1051         pxor    xmm1,xmm2
1052         pshufd  xmm2,xmm0,19
1053         pxor    xmm0,xmm0
1054         movdqa  [16+esp],xmm1
1055         paddq   xmm1,xmm1
1056         pand    xmm2,xmm3
1057         pcmpgtd xmm0,xmm1
1058         pxor    xmm1,xmm2
1059         pshufd  xmm2,xmm0,19
1060         pxor    xmm0,xmm0
1061         movdqa  [32+esp],xmm1
1062         paddq   xmm1,xmm1
1063         pand    xmm2,xmm3
1064         pcmpgtd xmm0,xmm1
1065         pxor    xmm1,xmm2
1066         pshufd  xmm2,xmm0,19
1067         pxor    xmm0,xmm0
1068         movdqa  [48+esp],xmm1
1069         paddq   xmm1,xmm1
1070         pand    xmm2,xmm3
1071         pcmpgtd xmm0,xmm1
1072         pxor    xmm1,xmm2
1073         pshufd  xmm7,xmm0,19
1074         movdqa  [64+esp],xmm1
1075         paddq   xmm1,xmm1
1076         movups  xmm0,[ebp]
1077         pand    xmm7,xmm3
1078         movups  xmm2,[esi]
1079         pxor    xmm7,xmm1
1080         mov     ecx,ebx
1081         movdqu  xmm3,[16+esi]
1082         xorps   xmm2,xmm0
1083         movdqu  xmm4,[32+esi]
1084         pxor    xmm3,xmm0
1085         movdqu  xmm5,[48+esi]
1086         pxor    xmm4,xmm0
1087         movdqu  xmm6,[64+esi]
1088         pxor    xmm5,xmm0
1089         movdqu  xmm1,[80+esi]
1090         pxor    xmm6,xmm0
1091         lea     esi,[96+esi]
1092         pxor    xmm2,[esp]
1093         movdqa  [80+esp],xmm7
1094         pxor    xmm7,xmm1
1095         movups  xmm1,[16+ebp]
1096         pxor    xmm3,[16+esp]
1097         pxor    xmm4,[32+esp]
1098 db      102,15,56,220,209
1099         pxor    xmm5,[48+esp]
1100         pxor    xmm6,[64+esp]
1101 db      102,15,56,220,217
1102         pxor    xmm7,xmm0
1103         movups  xmm0,[32+ebp]
1104 db      102,15,56,220,225
1105 db      102,15,56,220,233
1106 db      102,15,56,220,241
1107 db      102,15,56,220,249
1108         call    L$_aesni_encrypt6_enter
1109         movdqa  xmm1,[80+esp]
1110         pxor    xmm0,xmm0
1111         xorps   xmm2,[esp]
1112         pcmpgtd xmm0,xmm1
1113         xorps   xmm3,[16+esp]
1114         movups  [edi],xmm2
1115         xorps   xmm4,[32+esp]
1116         movups  [16+edi],xmm3
1117         xorps   xmm5,[48+esp]
1118         movups  [32+edi],xmm4
1119         xorps   xmm6,[64+esp]
1120         movups  [48+edi],xmm5
1121         xorps   xmm7,xmm1
1122         movups  [64+edi],xmm6
1123         pshufd  xmm2,xmm0,19
1124         movups  [80+edi],xmm7
1125         lea     edi,[96+edi]
1126         movdqa  xmm3,[96+esp]
1127         pxor    xmm0,xmm0
1128         paddq   xmm1,xmm1
1129         pand    xmm2,xmm3
1130         pcmpgtd xmm0,xmm1
1131         pxor    xmm1,xmm2
1132         sub     eax,96
1133         jnc     NEAR L$046xts_enc_loop6
1134         mov     ecx,DWORD [240+ebp]
1135         mov     edx,ebp
1136         mov     ebx,ecx
1137 L$045xts_enc_short:
1138         add     eax,96
1139         jz      NEAR L$047xts_enc_done6x
1140         movdqa  xmm5,xmm1
1141         cmp     eax,32
1142         jb      NEAR L$048xts_enc_one
1143         pshufd  xmm2,xmm0,19
1144         pxor    xmm0,xmm0
1145         paddq   xmm1,xmm1
1146         pand    xmm2,xmm3
1147         pcmpgtd xmm0,xmm1
1148         pxor    xmm1,xmm2
1149         je      NEAR L$049xts_enc_two
1150         pshufd  xmm2,xmm0,19
1151         pxor    xmm0,xmm0
1152         movdqa  xmm6,xmm1
1153         paddq   xmm1,xmm1
1154         pand    xmm2,xmm3
1155         pcmpgtd xmm0,xmm1
1156         pxor    xmm1,xmm2
1157         cmp     eax,64
1158         jb      NEAR L$050xts_enc_three
1159         pshufd  xmm2,xmm0,19
1160         pxor    xmm0,xmm0
1161         movdqa  xmm7,xmm1
1162         paddq   xmm1,xmm1
1163         pand    xmm2,xmm3
1164         pcmpgtd xmm0,xmm1
1165         pxor    xmm1,xmm2
1166         movdqa  [esp],xmm5
1167         movdqa  [16+esp],xmm6
1168         je      NEAR L$051xts_enc_four
1169         movdqa  [32+esp],xmm7
1170         pshufd  xmm7,xmm0,19
1171         movdqa  [48+esp],xmm1
1172         paddq   xmm1,xmm1
1173         pand    xmm7,xmm3
1174         pxor    xmm7,xmm1
1175         movdqu  xmm2,[esi]
1176         movdqu  xmm3,[16+esi]
1177         movdqu  xmm4,[32+esi]
1178         pxor    xmm2,[esp]
1179         movdqu  xmm5,[48+esi]
1180         pxor    xmm3,[16+esp]
1181         movdqu  xmm6,[64+esi]
1182         pxor    xmm4,[32+esp]
1183         lea     esi,[80+esi]
1184         pxor    xmm5,[48+esp]
1185         movdqa  [64+esp],xmm7
1186         pxor    xmm6,xmm7
1187         call    __aesni_encrypt6
1188         movaps  xmm1,[64+esp]
1189         xorps   xmm2,[esp]
1190         xorps   xmm3,[16+esp]
1191         xorps   xmm4,[32+esp]
1192         movups  [edi],xmm2
1193         xorps   xmm5,[48+esp]
1194         movups  [16+edi],xmm3
1195         xorps   xmm6,xmm1
1196         movups  [32+edi],xmm4
1197         movups  [48+edi],xmm5
1198         movups  [64+edi],xmm6
1199         lea     edi,[80+edi]
1200         jmp     NEAR L$052xts_enc_done
1201 align   16
1202 L$048xts_enc_one:
1203         movups  xmm2,[esi]
1204         lea     esi,[16+esi]
1205         xorps   xmm2,xmm5
1206         movups  xmm0,[edx]
1207         movups  xmm1,[16+edx]
1208         lea     edx,[32+edx]
1209         xorps   xmm2,xmm0
1210 L$053enc1_loop_9:
1211 db      102,15,56,220,209
1212         dec     ecx
1213         movups  xmm1,[edx]
1214         lea     edx,[16+edx]
1215         jnz     NEAR L$053enc1_loop_9
1216 db      102,15,56,221,209
1217         xorps   xmm2,xmm5
1218         movups  [edi],xmm2
1219         lea     edi,[16+edi]
1220         movdqa  xmm1,xmm5
1221         jmp     NEAR L$052xts_enc_done
1222 align   16
1223 L$049xts_enc_two:
1224         movaps  xmm6,xmm1
1225         movups  xmm2,[esi]
1226         movups  xmm3,[16+esi]
1227         lea     esi,[32+esi]
1228         xorps   xmm2,xmm5
1229         xorps   xmm3,xmm6
1230         call    __aesni_encrypt2
1231         xorps   xmm2,xmm5
1232         xorps   xmm3,xmm6
1233         movups  [edi],xmm2
1234         movups  [16+edi],xmm3
1235         lea     edi,[32+edi]
1236         movdqa  xmm1,xmm6
1237         jmp     NEAR L$052xts_enc_done
1238 align   16
1239 L$050xts_enc_three:
1240         movaps  xmm7,xmm1
1241         movups  xmm2,[esi]
1242         movups  xmm3,[16+esi]
1243         movups  xmm4,[32+esi]
1244         lea     esi,[48+esi]
1245         xorps   xmm2,xmm5
1246         xorps   xmm3,xmm6
1247         xorps   xmm4,xmm7
1248         call    __aesni_encrypt3
1249         xorps   xmm2,xmm5
1250         xorps   xmm3,xmm6
1251         xorps   xmm4,xmm7
1252         movups  [edi],xmm2
1253         movups  [16+edi],xmm3
1254         movups  [32+edi],xmm4
1255         lea     edi,[48+edi]
1256         movdqa  xmm1,xmm7
1257         jmp     NEAR L$052xts_enc_done
1258 align   16
1259 L$051xts_enc_four:
1260         movaps  xmm6,xmm1
1261         movups  xmm2,[esi]
1262         movups  xmm3,[16+esi]
1263         movups  xmm4,[32+esi]
1264         xorps   xmm2,[esp]
1265         movups  xmm5,[48+esi]
1266         lea     esi,[64+esi]
1267         xorps   xmm3,[16+esp]
1268         xorps   xmm4,xmm7
1269         xorps   xmm5,xmm6
1270         call    __aesni_encrypt4
1271         xorps   xmm2,[esp]
1272         xorps   xmm3,[16+esp]
1273         xorps   xmm4,xmm7
1274         movups  [edi],xmm2
1275         xorps   xmm5,xmm6
1276         movups  [16+edi],xmm3
1277         movups  [32+edi],xmm4
1278         movups  [48+edi],xmm5
1279         lea     edi,[64+edi]
1280         movdqa  xmm1,xmm6
1281         jmp     NEAR L$052xts_enc_done
1282 align   16
1283 L$047xts_enc_done6x:
1284         mov     eax,DWORD [112+esp]
1285         and     eax,15
1286         jz      NEAR L$054xts_enc_ret
1287         movdqa  xmm5,xmm1
1288         mov     DWORD [112+esp],eax
1289         jmp     NEAR L$055xts_enc_steal
1290 align   16
1291 L$052xts_enc_done:
1292         mov     eax,DWORD [112+esp]
1293         pxor    xmm0,xmm0
1294         and     eax,15
1295         jz      NEAR L$054xts_enc_ret
1296         pcmpgtd xmm0,xmm1
1297         mov     DWORD [112+esp],eax
1298         pshufd  xmm5,xmm0,19
1299         paddq   xmm1,xmm1
1300         pand    xmm5,[96+esp]
1301         pxor    xmm5,xmm1
1302 L$055xts_enc_steal:
1303         movzx   ecx,BYTE [esi]
1304         movzx   edx,BYTE [edi-16]
1305         lea     esi,[1+esi]
1306         mov     BYTE [edi-16],cl
1307         mov     BYTE [edi],dl
1308         lea     edi,[1+edi]
1309         sub     eax,1
1310         jnz     NEAR L$055xts_enc_steal
1311         sub     edi,DWORD [112+esp]
1312         mov     edx,ebp
1313         mov     ecx,ebx
1314         movups  xmm2,[edi-16]
1315         xorps   xmm2,xmm5
1316         movups  xmm0,[edx]
1317         movups  xmm1,[16+edx]
1318         lea     edx,[32+edx]
1319         xorps   xmm2,xmm0
1320 L$056enc1_loop_10:
1321 db      102,15,56,220,209
1322         dec     ecx
1323         movups  xmm1,[edx]
1324         lea     edx,[16+edx]
1325         jnz     NEAR L$056enc1_loop_10
1326 db      102,15,56,221,209
1327         xorps   xmm2,xmm5
1328         movups  [edi-16],xmm2
1329 L$054xts_enc_ret:
1330         mov     esp,DWORD [116+esp]
1331         pop     edi
1332         pop     esi
1333         pop     ebx
1334         pop     ebp
1335         ret
1336 global  _aesni_xts_decrypt
1337 align   16
1338 _aesni_xts_decrypt:
1339 L$_aesni_xts_decrypt_begin:
1340         push    ebp
1341         push    ebx
1342         push    esi
1343         push    edi
1344         mov     edx,DWORD [36+esp]
1345         mov     esi,DWORD [40+esp]
1346         mov     ecx,DWORD [240+edx]
1347         movups  xmm2,[esi]
1348         movups  xmm0,[edx]
1349         movups  xmm1,[16+edx]
1350         lea     edx,[32+edx]
1351         xorps   xmm2,xmm0
1352 L$057enc1_loop_11:
1353 db      102,15,56,220,209
1354         dec     ecx
1355         movups  xmm1,[edx]
1356         lea     edx,[16+edx]
1357         jnz     NEAR L$057enc1_loop_11
1358 db      102,15,56,221,209
1359         mov     esi,DWORD [20+esp]
1360         mov     edi,DWORD [24+esp]
1361         mov     eax,DWORD [28+esp]
1362         mov     edx,DWORD [32+esp]
1363         mov     ebp,esp
1364         sub     esp,120
1365         and     esp,-16
1366         xor     ebx,ebx
1367         test    eax,15
1368         setnz   bl
1369         shl     ebx,4
1370         sub     eax,ebx
1371         mov     DWORD [96+esp],135
1372         mov     DWORD [100+esp],0
1373         mov     DWORD [104+esp],1
1374         mov     DWORD [108+esp],0
1375         mov     DWORD [112+esp],eax
1376         mov     DWORD [116+esp],ebp
1377         mov     ecx,DWORD [240+edx]
1378         mov     ebp,edx
1379         mov     ebx,ecx
1380         movdqa  xmm1,xmm2
1381         pxor    xmm0,xmm0
1382         movdqa  xmm3,[96+esp]
1383         pcmpgtd xmm0,xmm1
1384         and     eax,-16
1385         sub     eax,96
1386         jc      NEAR L$058xts_dec_short
1387         shl     ecx,4
1388         mov     ebx,16
1389         sub     ebx,ecx
1390         lea     edx,[32+ecx*1+edx]
1391         jmp     NEAR L$059xts_dec_loop6
1392 align   16
1393 L$059xts_dec_loop6:
1394         pshufd  xmm2,xmm0,19
1395         pxor    xmm0,xmm0
1396         movdqa  [esp],xmm1
1397         paddq   xmm1,xmm1
1398         pand    xmm2,xmm3
1399         pcmpgtd xmm0,xmm1
1400         pxor    xmm1,xmm2
1401         pshufd  xmm2,xmm0,19
1402         pxor    xmm0,xmm0
1403         movdqa  [16+esp],xmm1
1404         paddq   xmm1,xmm1
1405         pand    xmm2,xmm3
1406         pcmpgtd xmm0,xmm1
1407         pxor    xmm1,xmm2
1408         pshufd  xmm2,xmm0,19
1409         pxor    xmm0,xmm0
1410         movdqa  [32+esp],xmm1
1411         paddq   xmm1,xmm1
1412         pand    xmm2,xmm3
1413         pcmpgtd xmm0,xmm1
1414         pxor    xmm1,xmm2
1415         pshufd  xmm2,xmm0,19
1416         pxor    xmm0,xmm0
1417         movdqa  [48+esp],xmm1
1418         paddq   xmm1,xmm1
1419         pand    xmm2,xmm3
1420         pcmpgtd xmm0,xmm1
1421         pxor    xmm1,xmm2
1422         pshufd  xmm7,xmm0,19
1423         movdqa  [64+esp],xmm1
1424         paddq   xmm1,xmm1
1425         movups  xmm0,[ebp]
1426         pand    xmm7,xmm3
1427         movups  xmm2,[esi]
1428         pxor    xmm7,xmm1
1429         mov     ecx,ebx
1430         movdqu  xmm3,[16+esi]
1431         xorps   xmm2,xmm0
1432         movdqu  xmm4,[32+esi]
1433         pxor    xmm3,xmm0
1434         movdqu  xmm5,[48+esi]
1435         pxor    xmm4,xmm0
1436         movdqu  xmm6,[64+esi]
1437         pxor    xmm5,xmm0
1438         movdqu  xmm1,[80+esi]
1439         pxor    xmm6,xmm0
1440         lea     esi,[96+esi]
1441         pxor    xmm2,[esp]
1442         movdqa  [80+esp],xmm7
1443         pxor    xmm7,xmm1
1444         movups  xmm1,[16+ebp]
1445         pxor    xmm3,[16+esp]
1446         pxor    xmm4,[32+esp]
1447 db      102,15,56,222,209
1448         pxor    xmm5,[48+esp]
1449         pxor    xmm6,[64+esp]
1450 db      102,15,56,222,217
1451         pxor    xmm7,xmm0
1452         movups  xmm0,[32+ebp]
1453 db      102,15,56,222,225
1454 db      102,15,56,222,233
1455 db      102,15,56,222,241
1456 db      102,15,56,222,249
1457         call    L$_aesni_decrypt6_enter
1458         movdqa  xmm1,[80+esp]
1459         pxor    xmm0,xmm0
1460         xorps   xmm2,[esp]
1461         pcmpgtd xmm0,xmm1
1462         xorps   xmm3,[16+esp]
1463         movups  [edi],xmm2
1464         xorps   xmm4,[32+esp]
1465         movups  [16+edi],xmm3
1466         xorps   xmm5,[48+esp]
1467         movups  [32+edi],xmm4
1468         xorps   xmm6,[64+esp]
1469         movups  [48+edi],xmm5
1470         xorps   xmm7,xmm1
1471         movups  [64+edi],xmm6
1472         pshufd  xmm2,xmm0,19
1473         movups  [80+edi],xmm7
1474         lea     edi,[96+edi]
1475         movdqa  xmm3,[96+esp]
1476         pxor    xmm0,xmm0
1477         paddq   xmm1,xmm1
1478         pand    xmm2,xmm3
1479         pcmpgtd xmm0,xmm1
1480         pxor    xmm1,xmm2
1481         sub     eax,96
1482         jnc     NEAR L$059xts_dec_loop6
1483         mov     ecx,DWORD [240+ebp]
1484         mov     edx,ebp
1485         mov     ebx,ecx
1486 L$058xts_dec_short:
1487         add     eax,96
1488         jz      NEAR L$060xts_dec_done6x
1489         movdqa  xmm5,xmm1
1490         cmp     eax,32
1491         jb      NEAR L$061xts_dec_one
1492         pshufd  xmm2,xmm0,19
1493         pxor    xmm0,xmm0
1494         paddq   xmm1,xmm1
1495         pand    xmm2,xmm3
1496         pcmpgtd xmm0,xmm1
1497         pxor    xmm1,xmm2
1498         je      NEAR L$062xts_dec_two
1499         pshufd  xmm2,xmm0,19
1500         pxor    xmm0,xmm0
1501         movdqa  xmm6,xmm1
1502         paddq   xmm1,xmm1
1503         pand    xmm2,xmm3
1504         pcmpgtd xmm0,xmm1
1505         pxor    xmm1,xmm2
1506         cmp     eax,64
1507         jb      NEAR L$063xts_dec_three
1508         pshufd  xmm2,xmm0,19
1509         pxor    xmm0,xmm0
1510         movdqa  xmm7,xmm1
1511         paddq   xmm1,xmm1
1512         pand    xmm2,xmm3
1513         pcmpgtd xmm0,xmm1
1514         pxor    xmm1,xmm2
1515         movdqa  [esp],xmm5
1516         movdqa  [16+esp],xmm6
1517         je      NEAR L$064xts_dec_four
1518         movdqa  [32+esp],xmm7
1519         pshufd  xmm7,xmm0,19
1520         movdqa  [48+esp],xmm1
1521         paddq   xmm1,xmm1
1522         pand    xmm7,xmm3
1523         pxor    xmm7,xmm1
1524         movdqu  xmm2,[esi]
1525         movdqu  xmm3,[16+esi]
1526         movdqu  xmm4,[32+esi]
1527         pxor    xmm2,[esp]
1528         movdqu  xmm5,[48+esi]
1529         pxor    xmm3,[16+esp]
1530         movdqu  xmm6,[64+esi]
1531         pxor    xmm4,[32+esp]
1532         lea     esi,[80+esi]
1533         pxor    xmm5,[48+esp]
1534         movdqa  [64+esp],xmm7
1535         pxor    xmm6,xmm7
1536         call    __aesni_decrypt6
1537         movaps  xmm1,[64+esp]
1538         xorps   xmm2,[esp]
1539         xorps   xmm3,[16+esp]
1540         xorps   xmm4,[32+esp]
1541         movups  [edi],xmm2
1542         xorps   xmm5,[48+esp]
1543         movups  [16+edi],xmm3
1544         xorps   xmm6,xmm1
1545         movups  [32+edi],xmm4
1546         movups  [48+edi],xmm5
1547         movups  [64+edi],xmm6
1548         lea     edi,[80+edi]
1549         jmp     NEAR L$065xts_dec_done
1550 align   16
1551 L$061xts_dec_one:
1552         movups  xmm2,[esi]
1553         lea     esi,[16+esi]
1554         xorps   xmm2,xmm5
1555         movups  xmm0,[edx]
1556         movups  xmm1,[16+edx]
1557         lea     edx,[32+edx]
1558         xorps   xmm2,xmm0
1559 L$066dec1_loop_12:
1560 db      102,15,56,222,209
1561         dec     ecx
1562         movups  xmm1,[edx]
1563         lea     edx,[16+edx]
1564         jnz     NEAR L$066dec1_loop_12
1565 db      102,15,56,223,209
1566         xorps   xmm2,xmm5
1567         movups  [edi],xmm2
1568         lea     edi,[16+edi]
1569         movdqa  xmm1,xmm5
1570         jmp     NEAR L$065xts_dec_done
1571 align   16
1572 L$062xts_dec_two:
1573         movaps  xmm6,xmm1
1574         movups  xmm2,[esi]
1575         movups  xmm3,[16+esi]
1576         lea     esi,[32+esi]
1577         xorps   xmm2,xmm5
1578         xorps   xmm3,xmm6
1579         call    __aesni_decrypt2
1580         xorps   xmm2,xmm5
1581         xorps   xmm3,xmm6
1582         movups  [edi],xmm2
1583         movups  [16+edi],xmm3
1584         lea     edi,[32+edi]
1585         movdqa  xmm1,xmm6
1586         jmp     NEAR L$065xts_dec_done
1587 align   16
1588 L$063xts_dec_three:
1589         movaps  xmm7,xmm1
1590         movups  xmm2,[esi]
1591         movups  xmm3,[16+esi]
1592         movups  xmm4,[32+esi]
1593         lea     esi,[48+esi]
1594         xorps   xmm2,xmm5
1595         xorps   xmm3,xmm6
1596         xorps   xmm4,xmm7
1597         call    __aesni_decrypt3
1598         xorps   xmm2,xmm5
1599         xorps   xmm3,xmm6
1600         xorps   xmm4,xmm7
1601         movups  [edi],xmm2
1602         movups  [16+edi],xmm3
1603         movups  [32+edi],xmm4
1604         lea     edi,[48+edi]
1605         movdqa  xmm1,xmm7
1606         jmp     NEAR L$065xts_dec_done
1607 align   16
1608 L$064xts_dec_four:
1609         movaps  xmm6,xmm1
1610         movups  xmm2,[esi]
1611         movups  xmm3,[16+esi]
1612         movups  xmm4,[32+esi]
1613         xorps   xmm2,[esp]
1614         movups  xmm5,[48+esi]
1615         lea     esi,[64+esi]
1616         xorps   xmm3,[16+esp]
1617         xorps   xmm4,xmm7
1618         xorps   xmm5,xmm6
1619         call    __aesni_decrypt4
1620         xorps   xmm2,[esp]
1621         xorps   xmm3,[16+esp]
1622         xorps   xmm4,xmm7
1623         movups  [edi],xmm2
1624         xorps   xmm5,xmm6
1625         movups  [16+edi],xmm3
1626         movups  [32+edi],xmm4
1627         movups  [48+edi],xmm5
1628         lea     edi,[64+edi]
1629         movdqa  xmm1,xmm6
1630         jmp     NEAR L$065xts_dec_done
1631 align   16
1632 L$060xts_dec_done6x:
1633         mov     eax,DWORD [112+esp]
1634         and     eax,15
1635         jz      NEAR L$067xts_dec_ret
1636         mov     DWORD [112+esp],eax
1637         jmp     NEAR L$068xts_dec_only_one_more
1638 align   16
1639 L$065xts_dec_done:
1640         mov     eax,DWORD [112+esp]
1641         pxor    xmm0,xmm0
1642         and     eax,15
1643         jz      NEAR L$067xts_dec_ret
1644         pcmpgtd xmm0,xmm1
1645         mov     DWORD [112+esp],eax
1646         pshufd  xmm2,xmm0,19
1647         pxor    xmm0,xmm0
1648         movdqa  xmm3,[96+esp]
1649         paddq   xmm1,xmm1
1650         pand    xmm2,xmm3
1651         pcmpgtd xmm0,xmm1
1652         pxor    xmm1,xmm2
1653 L$068xts_dec_only_one_more:
1654         pshufd  xmm5,xmm0,19
1655         movdqa  xmm6,xmm1
1656         paddq   xmm1,xmm1
1657         pand    xmm5,xmm3
1658         pxor    xmm5,xmm1
1659         mov     edx,ebp
1660         mov     ecx,ebx
1661         movups  xmm2,[esi]
1662         xorps   xmm2,xmm5
1663         movups  xmm0,[edx]
1664         movups  xmm1,[16+edx]
1665         lea     edx,[32+edx]
1666         xorps   xmm2,xmm0
1667 L$069dec1_loop_13:
1668 db      102,15,56,222,209
1669         dec     ecx
1670         movups  xmm1,[edx]
1671         lea     edx,[16+edx]
1672         jnz     NEAR L$069dec1_loop_13
1673 db      102,15,56,223,209
1674         xorps   xmm2,xmm5
1675         movups  [edi],xmm2
1676 L$070xts_dec_steal:
1677         movzx   ecx,BYTE [16+esi]
1678         movzx   edx,BYTE [edi]
1679         lea     esi,[1+esi]
1680         mov     BYTE [edi],cl
1681         mov     BYTE [16+edi],dl
1682         lea     edi,[1+edi]
1683         sub     eax,1
1684         jnz     NEAR L$070xts_dec_steal
1685         sub     edi,DWORD [112+esp]
1686         mov     edx,ebp
1687         mov     ecx,ebx
1688         movups  xmm2,[edi]
1689         xorps   xmm2,xmm6
1690         movups  xmm0,[edx]
1691         movups  xmm1,[16+edx]
1692         lea     edx,[32+edx]
1693         xorps   xmm2,xmm0
1694 L$071dec1_loop_14:
1695 db      102,15,56,222,209
1696         dec     ecx
1697         movups  xmm1,[edx]
1698         lea     edx,[16+edx]
1699         jnz     NEAR L$071dec1_loop_14
1700 db      102,15,56,223,209
1701         xorps   xmm2,xmm6
1702         movups  [edi],xmm2
1703 L$067xts_dec_ret:
1704         mov     esp,DWORD [116+esp]
1705         pop     edi
1706         pop     esi
1707         pop     ebx
1708         pop     ebp
1709         ret
1710 global  _aesni_cbc_encrypt
1711 align   16
1712 _aesni_cbc_encrypt:
1713 L$_aesni_cbc_encrypt_begin:
1714         push    ebp
1715         push    ebx
1716         push    esi
1717         push    edi
1718         mov     esi,DWORD [20+esp]
1719         mov     ebx,esp
1720         mov     edi,DWORD [24+esp]
1721         sub     ebx,24
1722         mov     eax,DWORD [28+esp]
1723         and     ebx,-16
1724         mov     edx,DWORD [32+esp]
1725         mov     ebp,DWORD [36+esp]
1726         test    eax,eax
1727         jz      NEAR L$072cbc_abort
1728         cmp     DWORD [40+esp],0
1729         xchg    ebx,esp
1730         movups  xmm7,[ebp]
1731         mov     ecx,DWORD [240+edx]
1732         mov     ebp,edx
1733         mov     DWORD [16+esp],ebx
1734         mov     ebx,ecx
1735         je      NEAR L$073cbc_decrypt
1736         movaps  xmm2,xmm7
1737         cmp     eax,16
1738         jb      NEAR L$074cbc_enc_tail
1739         sub     eax,16
1740         jmp     NEAR L$075cbc_enc_loop
1741 align   16
1742 L$075cbc_enc_loop:
1743         movups  xmm7,[esi]
1744         lea     esi,[16+esi]
1745         movups  xmm0,[edx]
1746         movups  xmm1,[16+edx]
1747         xorps   xmm7,xmm0
1748         lea     edx,[32+edx]
1749         xorps   xmm2,xmm7
1750 L$076enc1_loop_15:
1751 db      102,15,56,220,209
1752         dec     ecx
1753         movups  xmm1,[edx]
1754         lea     edx,[16+edx]
1755         jnz     NEAR L$076enc1_loop_15
1756 db      102,15,56,221,209
1757         mov     ecx,ebx
1758         mov     edx,ebp
1759         movups  [edi],xmm2
1760         lea     edi,[16+edi]
1761         sub     eax,16
1762         jnc     NEAR L$075cbc_enc_loop
1763         add     eax,16
1764         jnz     NEAR L$074cbc_enc_tail
1765         movaps  xmm7,xmm2
1766         jmp     NEAR L$077cbc_ret
1767 L$074cbc_enc_tail:
1768         mov     ecx,eax
1769 dd      2767451785
1770         mov     ecx,16
1771         sub     ecx,eax
1772         xor     eax,eax
1773 dd      2868115081
1774         lea     edi,[edi-16]
1775         mov     ecx,ebx
1776         mov     esi,edi
1777         mov     edx,ebp
1778         jmp     NEAR L$075cbc_enc_loop
1779 align   16
1780 L$073cbc_decrypt:
1781         cmp     eax,80
1782         jbe     NEAR L$078cbc_dec_tail
1783         movaps  [esp],xmm7
1784         sub     eax,80
1785         jmp     NEAR L$079cbc_dec_loop6_enter
1786 align   16
1787 L$080cbc_dec_loop6:
1788         movaps  [esp],xmm0
1789         movups  [edi],xmm7
1790         lea     edi,[16+edi]
1791 L$079cbc_dec_loop6_enter:
1792         movdqu  xmm2,[esi]
1793         movdqu  xmm3,[16+esi]
1794         movdqu  xmm4,[32+esi]
1795         movdqu  xmm5,[48+esi]
1796         movdqu  xmm6,[64+esi]
1797         movdqu  xmm7,[80+esi]
1798         call    __aesni_decrypt6
1799         movups  xmm1,[esi]
1800         movups  xmm0,[16+esi]
1801         xorps   xmm2,[esp]
1802         xorps   xmm3,xmm1
1803         movups  xmm1,[32+esi]
1804         xorps   xmm4,xmm0
1805         movups  xmm0,[48+esi]
1806         xorps   xmm5,xmm1
1807         movups  xmm1,[64+esi]
1808         xorps   xmm6,xmm0
1809         movups  xmm0,[80+esi]
1810         xorps   xmm7,xmm1
1811         movups  [edi],xmm2
1812         movups  [16+edi],xmm3
1813         lea     esi,[96+esi]
1814         movups  [32+edi],xmm4
1815         mov     ecx,ebx
1816         movups  [48+edi],xmm5
1817         mov     edx,ebp
1818         movups  [64+edi],xmm6
1819         lea     edi,[80+edi]
1820         sub     eax,96
1821         ja      NEAR L$080cbc_dec_loop6
1822         movaps  xmm2,xmm7
1823         movaps  xmm7,xmm0
1824         add     eax,80
1825         jle     NEAR L$081cbc_dec_tail_collected
1826         movups  [edi],xmm2
1827         lea     edi,[16+edi]
1828 L$078cbc_dec_tail:
1829         movups  xmm2,[esi]
1830         movaps  xmm6,xmm2
1831         cmp     eax,16
1832         jbe     NEAR L$082cbc_dec_one
1833         movups  xmm3,[16+esi]
1834         movaps  xmm5,xmm3
1835         cmp     eax,32
1836         jbe     NEAR L$083cbc_dec_two
1837         movups  xmm4,[32+esi]
1838         cmp     eax,48
1839         jbe     NEAR L$084cbc_dec_three
1840         movups  xmm5,[48+esi]
1841         cmp     eax,64
1842         jbe     NEAR L$085cbc_dec_four
1843         movups  xmm6,[64+esi]
1844         movaps  [esp],xmm7
1845         movups  xmm2,[esi]
1846         xorps   xmm7,xmm7
1847         call    __aesni_decrypt6
1848         movups  xmm1,[esi]
1849         movups  xmm0,[16+esi]
1850         xorps   xmm2,[esp]
1851         xorps   xmm3,xmm1
1852         movups  xmm1,[32+esi]
1853         xorps   xmm4,xmm0
1854         movups  xmm0,[48+esi]
1855         xorps   xmm5,xmm1
1856         movups  xmm7,[64+esi]
1857         xorps   xmm6,xmm0
1858         movups  [edi],xmm2
1859         movups  [16+edi],xmm3
1860         movups  [32+edi],xmm4
1861         movups  [48+edi],xmm5
1862         lea     edi,[64+edi]
1863         movaps  xmm2,xmm6
1864         sub     eax,80
1865         jmp     NEAR L$081cbc_dec_tail_collected
1866 align   16
1867 L$082cbc_dec_one:
1868         movups  xmm0,[edx]
1869         movups  xmm1,[16+edx]
1870         lea     edx,[32+edx]
1871         xorps   xmm2,xmm0
1872 L$086dec1_loop_16:
1873 db      102,15,56,222,209
1874         dec     ecx
1875         movups  xmm1,[edx]
1876         lea     edx,[16+edx]
1877         jnz     NEAR L$086dec1_loop_16
1878 db      102,15,56,223,209
1879         xorps   xmm2,xmm7
1880         movaps  xmm7,xmm6
1881         sub     eax,16
1882         jmp     NEAR L$081cbc_dec_tail_collected
1883 align   16
1884 L$083cbc_dec_two:
1885         call    __aesni_decrypt2
1886         xorps   xmm2,xmm7
1887         xorps   xmm3,xmm6
1888         movups  [edi],xmm2
1889         movaps  xmm2,xmm3
1890         lea     edi,[16+edi]
1891         movaps  xmm7,xmm5
1892         sub     eax,32
1893         jmp     NEAR L$081cbc_dec_tail_collected
1894 align   16
1895 L$084cbc_dec_three:
1896         call    __aesni_decrypt3
1897         xorps   xmm2,xmm7
1898         xorps   xmm3,xmm6
1899         xorps   xmm4,xmm5
1900         movups  [edi],xmm2
1901         movaps  xmm2,xmm4
1902         movups  [16+edi],xmm3
1903         lea     edi,[32+edi]
1904         movups  xmm7,[32+esi]
1905         sub     eax,48
1906         jmp     NEAR L$081cbc_dec_tail_collected
1907 align   16
1908 L$085cbc_dec_four:
1909         call    __aesni_decrypt4
1910         movups  xmm1,[16+esi]
1911         movups  xmm0,[32+esi]
1912         xorps   xmm2,xmm7
1913         movups  xmm7,[48+esi]
1914         xorps   xmm3,xmm6
1915         movups  [edi],xmm2
1916         xorps   xmm4,xmm1
1917         movups  [16+edi],xmm3
1918         xorps   xmm5,xmm0
1919         movups  [32+edi],xmm4
1920         lea     edi,[48+edi]
1921         movaps  xmm2,xmm5
1922         sub     eax,64
1923 L$081cbc_dec_tail_collected:
1924         and     eax,15
1925         jnz     NEAR L$087cbc_dec_tail_partial
1926         movups  [edi],xmm2
1927         jmp     NEAR L$077cbc_ret
1928 align   16
1929 L$087cbc_dec_tail_partial:
1930         movaps  [esp],xmm2
1931         mov     ecx,16
1932         mov     esi,esp
1933         sub     ecx,eax
1934 dd      2767451785
1935 L$077cbc_ret:
1936         mov     esp,DWORD [16+esp]
1937         mov     ebp,DWORD [36+esp]
1938         movups  [ebp],xmm7
1939 L$072cbc_abort:
1940         pop     edi
1941         pop     esi
1942         pop     ebx
1943         pop     ebp
1944         ret
1945 align   16
1946 __aesni_set_encrypt_key:
1947         test    eax,eax
1948         jz      NEAR L$088bad_pointer
1949         test    edx,edx
1950         jz      NEAR L$088bad_pointer
1951         movups  xmm0,[eax]
1952         xorps   xmm4,xmm4
1953         lea     edx,[16+edx]
1954         cmp     ecx,256
1955         je      NEAR L$08914rounds
1956         cmp     ecx,192
1957         je      NEAR L$09012rounds
1958         cmp     ecx,128
1959         jne     NEAR L$091bad_keybits
1960 align   16
1961 L$09210rounds:
1962         mov     ecx,9
1963         movups  [edx-16],xmm0
1964 db      102,15,58,223,200,1
1965         call    L$093key_128_cold
1966 db      102,15,58,223,200,2
1967         call    L$094key_128
1968 db      102,15,58,223,200,4
1969         call    L$094key_128
1970 db      102,15,58,223,200,8
1971         call    L$094key_128
1972 db      102,15,58,223,200,16
1973         call    L$094key_128
1974 db      102,15,58,223,200,32
1975         call    L$094key_128
1976 db      102,15,58,223,200,64
1977         call    L$094key_128
1978 db      102,15,58,223,200,128
1979         call    L$094key_128
1980 db      102,15,58,223,200,27
1981         call    L$094key_128
1982 db      102,15,58,223,200,54
1983         call    L$094key_128
1984         movups  [edx],xmm0
1985         mov     DWORD [80+edx],ecx
1986         xor     eax,eax
1987         ret
1988 align   16
1989 L$094key_128:
1990         movups  [edx],xmm0
1991         lea     edx,[16+edx]
1992 L$093key_128_cold:
1993         shufps  xmm4,xmm0,16
1994         xorps   xmm0,xmm4
1995         shufps  xmm4,xmm0,140
1996         xorps   xmm0,xmm4
1997         shufps  xmm1,xmm1,255
1998         xorps   xmm0,xmm1
1999         ret
2000 align   16
2001 L$09012rounds:
2002         movq    xmm2,[16+eax]
2003         mov     ecx,11
2004         movups  [edx-16],xmm0
2005 db      102,15,58,223,202,1
2006         call    L$095key_192a_cold
2007 db      102,15,58,223,202,2
2008         call    L$096key_192b
2009 db      102,15,58,223,202,4
2010         call    L$097key_192a
2011 db      102,15,58,223,202,8
2012         call    L$096key_192b
2013 db      102,15,58,223,202,16
2014         call    L$097key_192a
2015 db      102,15,58,223,202,32
2016         call    L$096key_192b
2017 db      102,15,58,223,202,64
2018         call    L$097key_192a
2019 db      102,15,58,223,202,128
2020         call    L$096key_192b
2021         movups  [edx],xmm0
2022         mov     DWORD [48+edx],ecx
2023         xor     eax,eax
2024         ret
2025 align   16
2026 L$097key_192a:
2027         movups  [edx],xmm0
2028         lea     edx,[16+edx]
2029 align   16
2030 L$095key_192a_cold:
2031         movaps  xmm5,xmm2
2032 L$098key_192b_warm:
2033         shufps  xmm4,xmm0,16
2034         movdqa  xmm3,xmm2
2035         xorps   xmm0,xmm4
2036         shufps  xmm4,xmm0,140
2037         pslldq  xmm3,4
2038         xorps   xmm0,xmm4
2039         pshufd  xmm1,xmm1,85
2040         pxor    xmm2,xmm3
2041         pxor    xmm0,xmm1
2042         pshufd  xmm3,xmm0,255
2043         pxor    xmm2,xmm3
2044         ret
2045 align   16
2046 L$096key_192b:
2047         movaps  xmm3,xmm0
2048         shufps  xmm5,xmm0,68
2049         movups  [edx],xmm5
2050         shufps  xmm3,xmm2,78
2051         movups  [16+edx],xmm3
2052         lea     edx,[32+edx]
2053         jmp     NEAR L$098key_192b_warm
2054 align   16
2055 L$08914rounds:
2056         movups  xmm2,[16+eax]
2057         mov     ecx,13
2058         lea     edx,[16+edx]
2059         movups  [edx-32],xmm0
2060         movups  [edx-16],xmm2
2061 db      102,15,58,223,202,1
2062         call    L$099key_256a_cold
2063 db      102,15,58,223,200,1
2064         call    L$100key_256b
2065 db      102,15,58,223,202,2
2066         call    L$101key_256a
2067 db      102,15,58,223,200,2
2068         call    L$100key_256b
2069 db      102,15,58,223,202,4
2070         call    L$101key_256a
2071 db      102,15,58,223,200,4
2072         call    L$100key_256b
2073 db      102,15,58,223,202,8
2074         call    L$101key_256a
2075 db      102,15,58,223,200,8
2076         call    L$100key_256b
2077 db      102,15,58,223,202,16
2078         call    L$101key_256a
2079 db      102,15,58,223,200,16
2080         call    L$100key_256b
2081 db      102,15,58,223,202,32
2082         call    L$101key_256a
2083 db      102,15,58,223,200,32
2084         call    L$100key_256b
2085 db      102,15,58,223,202,64
2086         call    L$101key_256a
2087         movups  [edx],xmm0
2088         mov     DWORD [16+edx],ecx
2089         xor     eax,eax
2090         ret
2091 align   16
2092 L$101key_256a:
2093         movups  [edx],xmm2
2094         lea     edx,[16+edx]
2095 L$099key_256a_cold:
2096         shufps  xmm4,xmm0,16
2097         xorps   xmm0,xmm4
2098         shufps  xmm4,xmm0,140
2099         xorps   xmm0,xmm4
2100         shufps  xmm1,xmm1,255
2101         xorps   xmm0,xmm1
2102         ret
2103 align   16
2104 L$100key_256b:
2105         movups  [edx],xmm0
2106         lea     edx,[16+edx]
2107         shufps  xmm4,xmm2,16
2108         xorps   xmm2,xmm4
2109         shufps  xmm4,xmm2,140
2110         xorps   xmm2,xmm4
2111         shufps  xmm1,xmm1,170
2112         xorps   xmm2,xmm1
2113         ret
2114 align   4
2115 L$088bad_pointer:
2116         mov     eax,-1
2117         ret
2118 align   4
2119 L$091bad_keybits:
2120         mov     eax,-2
2121         ret
2122 global  _aesni_set_encrypt_key
2123 align   16
2124 _aesni_set_encrypt_key:
2125 L$_aesni_set_encrypt_key_begin:
2126         mov     eax,DWORD [4+esp]
2127         mov     ecx,DWORD [8+esp]
2128         mov     edx,DWORD [12+esp]
2129         call    __aesni_set_encrypt_key
2130         ret
2131 global  _aesni_set_decrypt_key
2132 align   16
2133 _aesni_set_decrypt_key:
2134 L$_aesni_set_decrypt_key_begin:
2135         mov     eax,DWORD [4+esp]
2136         mov     ecx,DWORD [8+esp]
2137         mov     edx,DWORD [12+esp]
2138         call    __aesni_set_encrypt_key
2139         mov     edx,DWORD [12+esp]
2140         shl     ecx,4
2141         test    eax,eax
2142         jnz     NEAR L$102dec_key_ret
2143         lea     eax,[16+ecx*1+edx]
2144         movups  xmm0,[edx]
2145         movups  xmm1,[eax]
2146         movups  [eax],xmm0
2147         movups  [edx],xmm1
2148         lea     edx,[16+edx]
2149         lea     eax,[eax-16]
2150 L$103dec_key_inverse:
2151         movups  xmm0,[edx]
2152         movups  xmm1,[eax]
2153 db      102,15,56,219,192
2154 db      102,15,56,219,201
2155         lea     edx,[16+edx]
2156         lea     eax,[eax-16]
2157         movups  [16+eax],xmm0
2158         movups  [edx-16],xmm1
2159         cmp     eax,edx
2160         ja      NEAR L$103dec_key_inverse
2161         movups  xmm0,[edx]
2162 db      102,15,56,219,192
2163         movups  [edx],xmm0
2164         xor     eax,eax
2165 L$102dec_key_ret:
2166         ret
2167 db      65,69,83,32,102,111,114,32,73,110,116,101,108,32,65,69
2168 db      83,45,78,73,44,32,67,82,89,80,84,79,71,65,77,83
2169 db      32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
2170 db      115,108,46,111,114,103,62,0