third_party/boringssl/win-x86_64/crypto/aes/bsaes-x86_64.asm

   1 default rel
   2 %define XMMWORD
   3 %define YMMWORD
   4 %define ZMMWORD
   5 section .text code align=64
   6
   7
   8 EXTERN  asm_AES_encrypt
   9 EXTERN  asm_AES_decrypt
  10
  11
  12 ALIGN   64
  13 _bsaes_encrypt8:
  14         lea     r11,[$L$BS0]
  15
  16         movdqa  xmm8,XMMWORD[rax]
  17         lea     rax,[16+rax]
  18         movdqa  xmm7,XMMWORD[80+r11]
  19         pxor    xmm15,xmm8
  20         pxor    xmm0,xmm8
  21         pxor    xmm1,xmm8
  22         pxor    xmm2,xmm8
  23 DB      102,68,15,56,0,255
  24 DB      102,15,56,0,199
  25         pxor    xmm3,xmm8
  26         pxor    xmm4,xmm8
  27 DB      102,15,56,0,207
  28 DB      102,15,56,0,215
  29         pxor    xmm5,xmm8
  30         pxor    xmm6,xmm8
  31 DB      102,15,56,0,223
  32 DB      102,15,56,0,231
  33 DB      102,15,56,0,239
  34 DB      102,15,56,0,247
  35 _bsaes_encrypt8_bitslice:
  36         movdqa  xmm7,XMMWORD[r11]
  37         movdqa  xmm8,XMMWORD[16+r11]
  38         movdqa  xmm9,xmm5
  39         psrlq   xmm5,1
  40         movdqa  xmm10,xmm3
  41         psrlq   xmm3,1
  42         pxor    xmm5,xmm6
  43         pxor    xmm3,xmm4
  44         pand    xmm5,xmm7
  45         pand    xmm3,xmm7
  46         pxor    xmm6,xmm5
  47         psllq   xmm5,1
  48         pxor    xmm4,xmm3
  49         psllq   xmm3,1
  50         pxor    xmm5,xmm9
  51         pxor    xmm3,xmm10
  52         movdqa  xmm9,xmm1
  53         psrlq   xmm1,1
  54         movdqa  xmm10,xmm15
  55         psrlq   xmm15,1
  56         pxor    xmm1,xmm2
  57         pxor    xmm15,xmm0
  58         pand    xmm1,xmm7
  59         pand    xmm15,xmm7
  60         pxor    xmm2,xmm1
  61         psllq   xmm1,1
  62         pxor    xmm0,xmm15
  63         psllq   xmm15,1
  64         pxor    xmm1,xmm9
  65         pxor    xmm15,xmm10
  66         movdqa  xmm7,XMMWORD[32+r11]
  67         movdqa  xmm9,xmm4
  68         psrlq   xmm4,2
  69         movdqa  xmm10,xmm3
  70         psrlq   xmm3,2
  71         pxor    xmm4,xmm6
  72         pxor    xmm3,xmm5
  73         pand    xmm4,xmm8
  74         pand    xmm3,xmm8
  75         pxor    xmm6,xmm4
  76         psllq   xmm4,2
  77         pxor    xmm5,xmm3
  78         psllq   xmm3,2
  79         pxor    xmm4,xmm9
  80         pxor    xmm3,xmm10
  81         movdqa  xmm9,xmm0
  82         psrlq   xmm0,2
  83         movdqa  xmm10,xmm15
  84         psrlq   xmm15,2
  85         pxor    xmm0,xmm2
  86         pxor    xmm15,xmm1
  87         pand    xmm0,xmm8
  88         pand    xmm15,xmm8
  89         pxor    xmm2,xmm0
  90         psllq   xmm0,2
  91         pxor    xmm1,xmm15
  92         psllq   xmm15,2
  93         pxor    xmm0,xmm9
  94         pxor    xmm15,xmm10
  95         movdqa  xmm9,xmm2
  96         psrlq   xmm2,4
  97         movdqa  xmm10,xmm1
  98         psrlq   xmm1,4
  99         pxor    xmm2,xmm6
 100         pxor    xmm1,xmm5
 101         pand    xmm2,xmm7
 102         pand    xmm1,xmm7
 103         pxor    xmm6,xmm2
 104         psllq   xmm2,4
 105         pxor    xmm5,xmm1
 106         psllq   xmm1,4
 107         pxor    xmm2,xmm9
 108         pxor    xmm1,xmm10
 109         movdqa  xmm9,xmm0
 110         psrlq   xmm0,4
 111         movdqa  xmm10,xmm15
 112         psrlq   xmm15,4
 113         pxor    xmm0,xmm4
 114         pxor    xmm15,xmm3
 115         pand    xmm0,xmm7
 116         pand    xmm15,xmm7
 117         pxor    xmm4,xmm0
 118         psllq   xmm0,4
 119         pxor    xmm3,xmm15
 120         psllq   xmm15,4
 121         pxor    xmm0,xmm9
 122         pxor    xmm15,xmm10
 123         dec     r10d
 124         jmp     NEAR $L$enc_sbox
 125 ALIGN   16
 126 $L$enc_loop:
 127         pxor    xmm15,XMMWORD[rax]
 128         pxor    xmm0,XMMWORD[16+rax]
 129         pxor    xmm1,XMMWORD[32+rax]
 130         pxor    xmm2,XMMWORD[48+rax]
 131 DB      102,68,15,56,0,255
 132 DB      102,15,56,0,199
 133         pxor    xmm3,XMMWORD[64+rax]
 134         pxor    xmm4,XMMWORD[80+rax]
 135 DB      102,15,56,0,207
 136 DB      102,15,56,0,215
 137         pxor    xmm5,XMMWORD[96+rax]
 138         pxor    xmm6,XMMWORD[112+rax]
 139 DB      102,15,56,0,223
 140 DB      102,15,56,0,231
 141 DB      102,15,56,0,239
 142 DB      102,15,56,0,247
 143         lea     rax,[128+rax]
 144 $L$enc_sbox:
 145         pxor    xmm4,xmm5
 146         pxor    xmm1,xmm0
 147         pxor    xmm2,xmm15
 148         pxor    xmm5,xmm1
 149         pxor    xmm4,xmm15
 150
 151         pxor    xmm5,xmm2
 152         pxor    xmm2,xmm6
 153         pxor    xmm6,xmm4
 154         pxor    xmm2,xmm3
 155         pxor    xmm3,xmm4
 156         pxor    xmm2,xmm0
 157
 158         pxor    xmm1,xmm6
 159         pxor    xmm0,xmm4
 160         movdqa  xmm10,xmm6
 161         movdqa  xmm9,xmm0
 162         movdqa  xmm8,xmm4
 163         movdqa  xmm12,xmm1
 164         movdqa  xmm11,xmm5
 165
 166         pxor    xmm10,xmm3
 167         pxor    xmm9,xmm1
 168         pxor    xmm8,xmm2
 169         movdqa  xmm13,xmm10
 170         pxor    xmm12,xmm3
 171         movdqa  xmm7,xmm9
 172         pxor    xmm11,xmm15
 173         movdqa  xmm14,xmm10
 174
 175         por     xmm9,xmm8
 176         por     xmm10,xmm11
 177         pxor    xmm14,xmm7
 178         pand    xmm13,xmm11
 179         pxor    xmm11,xmm8
 180         pand    xmm7,xmm8
 181         pand    xmm14,xmm11
 182         movdqa  xmm11,xmm2
 183         pxor    xmm11,xmm15
 184         pand    xmm12,xmm11
 185         pxor    xmm10,xmm12
 186         pxor    xmm9,xmm12
 187         movdqa  xmm12,xmm6
 188         movdqa  xmm11,xmm4
 189         pxor    xmm12,xmm0
 190         pxor    xmm11,xmm5
 191         movdqa  xmm8,xmm12
 192         pand    xmm12,xmm11
 193         por     xmm8,xmm11
 194         pxor    xmm7,xmm12
 195         pxor    xmm10,xmm14
 196         pxor    xmm9,xmm13
 197         pxor    xmm8,xmm14
 198         movdqa  xmm11,xmm1
 199         pxor    xmm7,xmm13
 200         movdqa  xmm12,xmm3
 201         pxor    xmm8,xmm13
 202         movdqa  xmm13,xmm0
 203         pand    xmm11,xmm2
 204         movdqa  xmm14,xmm6
 205         pand    xmm12,xmm15
 206         pand    xmm13,xmm4
 207         por     xmm14,xmm5
 208         pxor    xmm10,xmm11
 209         pxor    xmm9,xmm12
 210         pxor    xmm8,xmm13
 211         pxor    xmm7,xmm14
 212
 213
 214
 215
 216
 217         movdqa  xmm11,xmm10
 218         pand    xmm10,xmm8
 219         pxor    xmm11,xmm9
 220
 221         movdqa  xmm13,xmm7
 222         movdqa  xmm14,xmm11
 223         pxor    xmm13,xmm10
 224         pand    xmm14,xmm13
 225
 226         movdqa  xmm12,xmm8
 227         pxor    xmm14,xmm9
 228         pxor    xmm12,xmm7
 229
 230         pxor    xmm10,xmm9
 231
 232         pand    xmm12,xmm10
 233
 234         movdqa  xmm9,xmm13
 235         pxor    xmm12,xmm7
 236
 237         pxor    xmm9,xmm12
 238         pxor    xmm8,xmm12
 239
 240         pand    xmm9,xmm7
 241
 242         pxor    xmm13,xmm9
 243         pxor    xmm8,xmm9
 244
 245         pand    xmm13,xmm14
 246
 247         pxor    xmm13,xmm11
 248         movdqa  xmm11,xmm5
 249         movdqa  xmm7,xmm4
 250         movdqa  xmm9,xmm14
 251         pxor    xmm9,xmm13
 252         pand    xmm9,xmm5
 253         pxor    xmm5,xmm4
 254         pand    xmm4,xmm14
 255         pand    xmm5,xmm13
 256         pxor    xmm5,xmm4
 257         pxor    xmm4,xmm9
 258         pxor    xmm11,xmm15
 259         pxor    xmm7,xmm2
 260         pxor    xmm14,xmm12
 261         pxor    xmm13,xmm8
 262         movdqa  xmm10,xmm14
 263         movdqa  xmm9,xmm12
 264         pxor    xmm10,xmm13
 265         pxor    xmm9,xmm8
 266         pand    xmm10,xmm11
 267         pand    xmm9,xmm15
 268         pxor    xmm11,xmm7
 269         pxor    xmm15,xmm2
 270         pand    xmm7,xmm14
 271         pand    xmm2,xmm12
 272         pand    xmm11,xmm13
 273         pand    xmm15,xmm8
 274         pxor    xmm7,xmm11
 275         pxor    xmm15,xmm2
 276         pxor    xmm11,xmm10
 277         pxor    xmm2,xmm9
 278         pxor    xmm5,xmm11
 279         pxor    xmm15,xmm11
 280         pxor    xmm4,xmm7
 281         pxor    xmm2,xmm7
 282
 283         movdqa  xmm11,xmm6
 284         movdqa  xmm7,xmm0
 285         pxor    xmm11,xmm3
 286         pxor    xmm7,xmm1
 287         movdqa  xmm10,xmm14
 288         movdqa  xmm9,xmm12
 289         pxor    xmm10,xmm13
 290         pxor    xmm9,xmm8
 291         pand    xmm10,xmm11
 292         pand    xmm9,xmm3
 293         pxor    xmm11,xmm7
 294         pxor    xmm3,xmm1
 295         pand    xmm7,xmm14
 296         pand    xmm1,xmm12
 297         pand    xmm11,xmm13
 298         pand    xmm3,xmm8
 299         pxor    xmm7,xmm11
 300         pxor    xmm3,xmm1
 301         pxor    xmm11,xmm10
 302         pxor    xmm1,xmm9
 303         pxor    xmm14,xmm12
 304         pxor    xmm13,xmm8
 305         movdqa  xmm10,xmm14
 306         pxor    xmm10,xmm13
 307         pand    xmm10,xmm6
 308         pxor    xmm6,xmm0
 309         pand    xmm0,xmm14
 310         pand    xmm6,xmm13
 311         pxor    xmm6,xmm0
 312         pxor    xmm0,xmm10
 313         pxor    xmm6,xmm11
 314         pxor    xmm3,xmm11
 315         pxor    xmm0,xmm7
 316         pxor    xmm1,xmm7
 317         pxor    xmm6,xmm15
 318         pxor    xmm0,xmm5
 319         pxor    xmm3,xmm6
 320         pxor    xmm5,xmm15
 321         pxor    xmm15,xmm0
 322
 323         pxor    xmm0,xmm4
 324         pxor    xmm4,xmm1
 325         pxor    xmm1,xmm2
 326         pxor    xmm2,xmm4
 327         pxor    xmm3,xmm4
 328
 329         pxor    xmm5,xmm2
 330         dec     r10d
 331         jl      NEAR $L$enc_done
 332         pshufd  xmm7,xmm15,0x93
 333         pshufd  xmm8,xmm0,0x93
 334         pxor    xmm15,xmm7
 335         pshufd  xmm9,xmm3,0x93
 336         pxor    xmm0,xmm8
 337         pshufd  xmm10,xmm5,0x93
 338         pxor    xmm3,xmm9
 339         pshufd  xmm11,xmm2,0x93
 340         pxor    xmm5,xmm10
 341         pshufd  xmm12,xmm6,0x93
 342         pxor    xmm2,xmm11
 343         pshufd  xmm13,xmm1,0x93
 344         pxor    xmm6,xmm12
 345         pshufd  xmm14,xmm4,0x93
 346         pxor    xmm1,xmm13
 347         pxor    xmm4,xmm14
 348
 349         pxor    xmm8,xmm15
 350         pxor    xmm7,xmm4
 351         pxor    xmm8,xmm4
 352         pshufd  xmm15,xmm15,0x4E
 353         pxor    xmm9,xmm0
 354         pshufd  xmm0,xmm0,0x4E
 355         pxor    xmm12,xmm2
 356         pxor    xmm15,xmm7
 357         pxor    xmm13,xmm6
 358         pxor    xmm0,xmm8
 359         pxor    xmm11,xmm5
 360         pshufd  xmm7,xmm2,0x4E
 361         pxor    xmm14,xmm1
 362         pshufd  xmm8,xmm6,0x4E
 363         pxor    xmm10,xmm3
 364         pshufd  xmm2,xmm5,0x4E
 365         pxor    xmm10,xmm4
 366         pshufd  xmm6,xmm4,0x4E
 367         pxor    xmm11,xmm4
 368         pshufd  xmm5,xmm1,0x4E
 369         pxor    xmm7,xmm11
 370         pshufd  xmm1,xmm3,0x4E
 371         pxor    xmm8,xmm12
 372         pxor    xmm2,xmm10
 373         pxor    xmm6,xmm14
 374         pxor    xmm5,xmm13
 375         movdqa  xmm3,xmm7
 376         pxor    xmm1,xmm9
 377         movdqa  xmm4,xmm8
 378         movdqa  xmm7,XMMWORD[48+r11]
 379         jnz     NEAR $L$enc_loop
 380         movdqa  xmm7,XMMWORD[64+r11]
 381         jmp     NEAR $L$enc_loop
 382 ALIGN   16
 383 $L$enc_done:
 384         movdqa  xmm7,XMMWORD[r11]
 385         movdqa  xmm8,XMMWORD[16+r11]
 386         movdqa  xmm9,xmm1
 387         psrlq   xmm1,1
 388         movdqa  xmm10,xmm2
 389         psrlq   xmm2,1
 390         pxor    xmm1,xmm4
 391         pxor    xmm2,xmm6
 392         pand    xmm1,xmm7
 393         pand    xmm2,xmm7
 394         pxor    xmm4,xmm1
 395         psllq   xmm1,1
 396         pxor    xmm6,xmm2
 397         psllq   xmm2,1
 398         pxor    xmm1,xmm9
 399         pxor    xmm2,xmm10
 400         movdqa  xmm9,xmm3
 401         psrlq   xmm3,1
 402         movdqa  xmm10,xmm15
 403         psrlq   xmm15,1
 404         pxor    xmm3,xmm5
 405         pxor    xmm15,xmm0
 406         pand    xmm3,xmm7
 407         pand    xmm15,xmm7
 408         pxor    xmm5,xmm3
 409         psllq   xmm3,1
 410         pxor    xmm0,xmm15
 411         psllq   xmm15,1
 412         pxor    xmm3,xmm9
 413         pxor    xmm15,xmm10
 414         movdqa  xmm7,XMMWORD[32+r11]
 415         movdqa  xmm9,xmm6
 416         psrlq   xmm6,2
 417         movdqa  xmm10,xmm2
 418         psrlq   xmm2,2
 419         pxor    xmm6,xmm4
 420         pxor    xmm2,xmm1
 421         pand    xmm6,xmm8
 422         pand    xmm2,xmm8
 423         pxor    xmm4,xmm6
 424         psllq   xmm6,2
 425         pxor    xmm1,xmm2
 426         psllq   xmm2,2
 427         pxor    xmm6,xmm9
 428         pxor    xmm2,xmm10
 429         movdqa  xmm9,xmm0
 430         psrlq   xmm0,2
 431         movdqa  xmm10,xmm15
 432         psrlq   xmm15,2
 433         pxor    xmm0,xmm5
 434         pxor    xmm15,xmm3
 435         pand    xmm0,xmm8
 436         pand    xmm15,xmm8
 437         pxor    xmm5,xmm0
 438         psllq   xmm0,2
 439         pxor    xmm3,xmm15
 440         psllq   xmm15,2
 441         pxor    xmm0,xmm9
 442         pxor    xmm15,xmm10
 443         movdqa  xmm9,xmm5
 444         psrlq   xmm5,4
 445         movdqa  xmm10,xmm3
 446         psrlq   xmm3,4
 447         pxor    xmm5,xmm4
 448         pxor    xmm3,xmm1
 449         pand    xmm5,xmm7
 450         pand    xmm3,xmm7
 451         pxor    xmm4,xmm5
 452         psllq   xmm5,4
 453         pxor    xmm1,xmm3
 454         psllq   xmm3,4
 455         pxor    xmm5,xmm9
 456         pxor    xmm3,xmm10
 457         movdqa  xmm9,xmm0
 458         psrlq   xmm0,4
 459         movdqa  xmm10,xmm15
 460         psrlq   xmm15,4
 461         pxor    xmm0,xmm6
 462         pxor    xmm15,xmm2
 463         pand    xmm0,xmm7
 464         pand    xmm15,xmm7
 465         pxor    xmm6,xmm0
 466         psllq   xmm0,4
 467         pxor    xmm2,xmm15
 468         psllq   xmm15,4
 469         pxor    xmm0,xmm9
 470         pxor    xmm15,xmm10
 471         movdqa  xmm7,XMMWORD[rax]
 472         pxor    xmm3,xmm7
 473         pxor    xmm5,xmm7
 474         pxor    xmm2,xmm7
 475         pxor    xmm6,xmm7
 476         pxor    xmm1,xmm7
 477         pxor    xmm4,xmm7
 478         pxor    xmm15,xmm7
 479         pxor    xmm0,xmm7
 480         DB      0F3h,0C3h               ;repret
 481
 482
 483
 484 ALIGN   64
 485 _bsaes_decrypt8:
 486         lea     r11,[$L$BS0]
 487
 488         movdqa  xmm8,XMMWORD[rax]
 489         lea     rax,[16+rax]
 490         movdqa  xmm7,XMMWORD[((-48))+r11]
 491         pxor    xmm15,xmm8
 492         pxor    xmm0,xmm8
 493         pxor    xmm1,xmm8
 494         pxor    xmm2,xmm8
 495 DB      102,68,15,56,0,255
 496 DB      102,15,56,0,199
 497         pxor    xmm3,xmm8
 498         pxor    xmm4,xmm8
 499 DB      102,15,56,0,207
 500 DB      102,15,56,0,215
 501         pxor    xmm5,xmm8
 502         pxor    xmm6,xmm8
 503 DB      102,15,56,0,223
 504 DB      102,15,56,0,231
 505 DB      102,15,56,0,239
 506 DB      102,15,56,0,247
 507         movdqa  xmm7,XMMWORD[r11]
 508         movdqa  xmm8,XMMWORD[16+r11]
 509         movdqa  xmm9,xmm5
 510         psrlq   xmm5,1
 511         movdqa  xmm10,xmm3
 512         psrlq   xmm3,1
 513         pxor    xmm5,xmm6
 514         pxor    xmm3,xmm4
 515         pand    xmm5,xmm7
 516         pand    xmm3,xmm7
 517         pxor    xmm6,xmm5
 518         psllq   xmm5,1
 519         pxor    xmm4,xmm3
 520         psllq   xmm3,1
 521         pxor    xmm5,xmm9
 522         pxor    xmm3,xmm10
 523         movdqa  xmm9,xmm1
 524         psrlq   xmm1,1
 525         movdqa  xmm10,xmm15
 526         psrlq   xmm15,1
 527         pxor    xmm1,xmm2
 528         pxor    xmm15,xmm0
 529         pand    xmm1,xmm7
 530         pand    xmm15,xmm7
 531         pxor    xmm2,xmm1
 532         psllq   xmm1,1
 533         pxor    xmm0,xmm15
 534         psllq   xmm15,1
 535         pxor    xmm1,xmm9
 536         pxor    xmm15,xmm10
 537         movdqa  xmm7,XMMWORD[32+r11]
 538         movdqa  xmm9,xmm4
 539         psrlq   xmm4,2
 540         movdqa  xmm10,xmm3
 541         psrlq   xmm3,2
 542         pxor    xmm4,xmm6
 543         pxor    xmm3,xmm5
 544         pand    xmm4,xmm8
 545         pand    xmm3,xmm8
 546         pxor    xmm6,xmm4
 547         psllq   xmm4,2
 548         pxor    xmm5,xmm3
 549         psllq   xmm3,2
 550         pxor    xmm4,xmm9
 551         pxor    xmm3,xmm10
 552         movdqa  xmm9,xmm0
 553         psrlq   xmm0,2
 554         movdqa  xmm10,xmm15
 555         psrlq   xmm15,2
 556         pxor    xmm0,xmm2
 557         pxor    xmm15,xmm1
 558         pand    xmm0,xmm8
 559         pand    xmm15,xmm8
 560         pxor    xmm2,xmm0
 561         psllq   xmm0,2
 562         pxor    xmm1,xmm15
 563         psllq   xmm15,2
 564         pxor    xmm0,xmm9
 565         pxor    xmm15,xmm10
 566         movdqa  xmm9,xmm2
 567         psrlq   xmm2,4
 568         movdqa  xmm10,xmm1
 569         psrlq   xmm1,4
 570         pxor    xmm2,xmm6
 571         pxor    xmm1,xmm5
 572         pand    xmm2,xmm7
 573         pand    xmm1,xmm7
 574         pxor    xmm6,xmm2
 575         psllq   xmm2,4
 576         pxor    xmm5,xmm1
 577         psllq   xmm1,4
 578         pxor    xmm2,xmm9
 579         pxor    xmm1,xmm10
 580         movdqa  xmm9,xmm0
 581         psrlq   xmm0,4
 582         movdqa  xmm10,xmm15
 583         psrlq   xmm15,4
 584         pxor    xmm0,xmm4
 585         pxor    xmm15,xmm3
 586         pand    xmm0,xmm7
 587         pand    xmm15,xmm7
 588         pxor    xmm4,xmm0
 589         psllq   xmm0,4
 590         pxor    xmm3,xmm15
 591         psllq   xmm15,4
 592         pxor    xmm0,xmm9
 593         pxor    xmm15,xmm10
 594         dec     r10d
 595         jmp     NEAR $L$dec_sbox
 596 ALIGN   16
 597 $L$dec_loop:
 598         pxor    xmm15,XMMWORD[rax]
 599         pxor    xmm0,XMMWORD[16+rax]
 600         pxor    xmm1,XMMWORD[32+rax]
 601         pxor    xmm2,XMMWORD[48+rax]
 602 DB      102,68,15,56,0,255
 603 DB      102,15,56,0,199
 604         pxor    xmm3,XMMWORD[64+rax]
 605         pxor    xmm4,XMMWORD[80+rax]
 606 DB      102,15,56,0,207
 607 DB      102,15,56,0,215
 608         pxor    xmm5,XMMWORD[96+rax]
 609         pxor    xmm6,XMMWORD[112+rax]
 610 DB      102,15,56,0,223
 611 DB      102,15,56,0,231
 612 DB      102,15,56,0,239
 613 DB      102,15,56,0,247
 614         lea     rax,[128+rax]
 615 $L$dec_sbox:
 616         pxor    xmm2,xmm3
 617
 618         pxor    xmm3,xmm6
 619         pxor    xmm1,xmm6
 620         pxor    xmm5,xmm3
 621         pxor    xmm6,xmm5
 622         pxor    xmm0,xmm6
 623
 624         pxor    xmm15,xmm0
 625         pxor    xmm1,xmm4
 626         pxor    xmm2,xmm15
 627         pxor    xmm4,xmm15
 628         pxor    xmm0,xmm2
 629         movdqa  xmm10,xmm2
 630         movdqa  xmm9,xmm6
 631         movdqa  xmm8,xmm0
 632         movdqa  xmm12,xmm3
 633         movdqa  xmm11,xmm4
 634
 635         pxor    xmm10,xmm15
 636         pxor    xmm9,xmm3
 637         pxor    xmm8,xmm5
 638         movdqa  xmm13,xmm10
 639         pxor    xmm12,xmm15
 640         movdqa  xmm7,xmm9
 641         pxor    xmm11,xmm1
 642         movdqa  xmm14,xmm10
 643
 644         por     xmm9,xmm8
 645         por     xmm10,xmm11
 646         pxor    xmm14,xmm7
 647         pand    xmm13,xmm11
 648         pxor    xmm11,xmm8
 649         pand    xmm7,xmm8
 650         pand    xmm14,xmm11
 651         movdqa  xmm11,xmm5
 652         pxor    xmm11,xmm1
 653         pand    xmm12,xmm11
 654         pxor    xmm10,xmm12
 655         pxor    xmm9,xmm12
 656         movdqa  xmm12,xmm2
 657         movdqa  xmm11,xmm0
 658         pxor    xmm12,xmm6
 659         pxor    xmm11,xmm4
 660         movdqa  xmm8,xmm12
 661         pand    xmm12,xmm11
 662         por     xmm8,xmm11
 663         pxor    xmm7,xmm12
 664         pxor    xmm10,xmm14
 665         pxor    xmm9,xmm13
 666         pxor    xmm8,xmm14
 667         movdqa  xmm11,xmm3
 668         pxor    xmm7,xmm13
 669         movdqa  xmm12,xmm15
 670         pxor    xmm8,xmm13
 671         movdqa  xmm13,xmm6
 672         pand    xmm11,xmm5
 673         movdqa  xmm14,xmm2
 674         pand    xmm12,xmm1
 675         pand    xmm13,xmm0
 676         por     xmm14,xmm4
 677         pxor    xmm10,xmm11
 678         pxor    xmm9,xmm12
 679         pxor    xmm8,xmm13
 680         pxor    xmm7,xmm14
 681
 682
 683
 684
 685
 686         movdqa  xmm11,xmm10
 687         pand    xmm10,xmm8
 688         pxor    xmm11,xmm9
 689
 690         movdqa  xmm13,xmm7
 691         movdqa  xmm14,xmm11
 692         pxor    xmm13,xmm10
 693         pand    xmm14,xmm13
 694
 695         movdqa  xmm12,xmm8
 696         pxor    xmm14,xmm9
 697         pxor    xmm12,xmm7
 698
 699         pxor    xmm10,xmm9
 700
 701         pand    xmm12,xmm10
 702
 703         movdqa  xmm9,xmm13
 704         pxor    xmm12,xmm7
 705
 706         pxor    xmm9,xmm12
 707         pxor    xmm8,xmm12
 708
 709         pand    xmm9,xmm7
 710
 711         pxor    xmm13,xmm9
 712         pxor    xmm8,xmm9
 713
 714         pand    xmm13,xmm14
 715
 716         pxor    xmm13,xmm11
 717         movdqa  xmm11,xmm4
 718         movdqa  xmm7,xmm0
 719         movdqa  xmm9,xmm14
 720         pxor    xmm9,xmm13
 721         pand    xmm9,xmm4
 722         pxor    xmm4,xmm0
 723         pand    xmm0,xmm14
 724         pand    xmm4,xmm13
 725         pxor    xmm4,xmm0
 726         pxor    xmm0,xmm9
 727         pxor    xmm11,xmm1
 728         pxor    xmm7,xmm5
 729         pxor    xmm14,xmm12
 730         pxor    xmm13,xmm8
 731         movdqa  xmm10,xmm14
 732         movdqa  xmm9,xmm12
 733         pxor    xmm10,xmm13
 734         pxor    xmm9,xmm8
 735         pand    xmm10,xmm11
 736         pand    xmm9,xmm1
 737         pxor    xmm11,xmm7
 738         pxor    xmm1,xmm5
 739         pand    xmm7,xmm14
 740         pand    xmm5,xmm12
 741         pand    xmm11,xmm13
 742         pand    xmm1,xmm8
 743         pxor    xmm7,xmm11
 744         pxor    xmm1,xmm5
 745         pxor    xmm11,xmm10
 746         pxor    xmm5,xmm9
 747         pxor    xmm4,xmm11
 748         pxor    xmm1,xmm11
 749         pxor    xmm0,xmm7
 750         pxor    xmm5,xmm7
 751
 752         movdqa  xmm11,xmm2
 753         movdqa  xmm7,xmm6
 754         pxor    xmm11,xmm15
 755         pxor    xmm7,xmm3
 756         movdqa  xmm10,xmm14
 757         movdqa  xmm9,xmm12
 758         pxor    xmm10,xmm13
 759         pxor    xmm9,xmm8
 760         pand    xmm10,xmm11
 761         pand    xmm9,xmm15
 762         pxor    xmm11,xmm7
 763         pxor    xmm15,xmm3
 764         pand    xmm7,xmm14
 765         pand    xmm3,xmm12
 766         pand    xmm11,xmm13
 767         pand    xmm15,xmm8
 768         pxor    xmm7,xmm11
 769         pxor    xmm15,xmm3
 770         pxor    xmm11,xmm10
 771         pxor    xmm3,xmm9
 772         pxor    xmm14,xmm12
 773         pxor    xmm13,xmm8
 774         movdqa  xmm10,xmm14
 775         pxor    xmm10,xmm13
 776         pand    xmm10,xmm2
 777         pxor    xmm2,xmm6
 778         pand    xmm6,xmm14
 779         pand    xmm2,xmm13
 780         pxor    xmm2,xmm6
 781         pxor    xmm6,xmm10
 782         pxor    xmm2,xmm11
 783         pxor    xmm15,xmm11
 784         pxor    xmm6,xmm7
 785         pxor    xmm3,xmm7
 786         pxor    xmm0,xmm6
 787         pxor    xmm5,xmm4
 788
 789         pxor    xmm3,xmm0
 790         pxor    xmm1,xmm6
 791         pxor    xmm4,xmm6
 792         pxor    xmm3,xmm1
 793         pxor    xmm6,xmm15
 794         pxor    xmm3,xmm4
 795         pxor    xmm2,xmm5
 796         pxor    xmm5,xmm0
 797         pxor    xmm2,xmm3
 798
 799         pxor    xmm3,xmm15
 800         pxor    xmm6,xmm2
 801         dec     r10d
 802         jl      NEAR $L$dec_done
 803
 804         pshufd  xmm7,xmm15,0x4E
 805         pshufd  xmm13,xmm2,0x4E
 806         pxor    xmm7,xmm15
 807         pshufd  xmm14,xmm4,0x4E
 808         pxor    xmm13,xmm2
 809         pshufd  xmm8,xmm0,0x4E
 810         pxor    xmm14,xmm4
 811         pshufd  xmm9,xmm5,0x4E
 812         pxor    xmm8,xmm0
 813         pshufd  xmm10,xmm3,0x4E
 814         pxor    xmm9,xmm5
 815         pxor    xmm15,xmm13
 816         pxor    xmm0,xmm13
 817         pshufd  xmm11,xmm1,0x4E
 818         pxor    xmm10,xmm3
 819         pxor    xmm5,xmm7
 820         pxor    xmm3,xmm8
 821         pshufd  xmm12,xmm6,0x4E
 822         pxor    xmm11,xmm1
 823         pxor    xmm0,xmm14
 824         pxor    xmm1,xmm9
 825         pxor    xmm12,xmm6
 826
 827         pxor    xmm5,xmm14
 828         pxor    xmm3,xmm13
 829         pxor    xmm1,xmm13
 830         pxor    xmm6,xmm10
 831         pxor    xmm2,xmm11
 832         pxor    xmm1,xmm14
 833         pxor    xmm6,xmm14
 834         pxor    xmm4,xmm12
 835         pshufd  xmm7,xmm15,0x93
 836         pshufd  xmm8,xmm0,0x93
 837         pxor    xmm15,xmm7
 838         pshufd  xmm9,xmm5,0x93
 839         pxor    xmm0,xmm8
 840         pshufd  xmm10,xmm3,0x93
 841         pxor    xmm5,xmm9
 842         pshufd  xmm11,xmm1,0x93
 843         pxor    xmm3,xmm10
 844         pshufd  xmm12,xmm6,0x93
 845         pxor    xmm1,xmm11
 846         pshufd  xmm13,xmm2,0x93
 847         pxor    xmm6,xmm12
 848         pshufd  xmm14,xmm4,0x93
 849         pxor    xmm2,xmm13
 850         pxor    xmm4,xmm14
 851
 852         pxor    xmm8,xmm15
 853         pxor    xmm7,xmm4
 854         pxor    xmm8,xmm4
 855         pshufd  xmm15,xmm15,0x4E
 856         pxor    xmm9,xmm0
 857         pshufd  xmm0,xmm0,0x4E
 858         pxor    xmm12,xmm1
 859         pxor    xmm15,xmm7
 860         pxor    xmm13,xmm6
 861         pxor    xmm0,xmm8
 862         pxor    xmm11,xmm3
 863         pshufd  xmm7,xmm1,0x4E
 864         pxor    xmm14,xmm2
 865         pshufd  xmm8,xmm6,0x4E
 866         pxor    xmm10,xmm5
 867         pshufd  xmm1,xmm3,0x4E
 868         pxor    xmm10,xmm4
 869         pshufd  xmm6,xmm4,0x4E
 870         pxor    xmm11,xmm4
 871         pshufd  xmm3,xmm2,0x4E
 872         pxor    xmm7,xmm11
 873         pshufd  xmm2,xmm5,0x4E
 874         pxor    xmm8,xmm12
 875         pxor    xmm10,xmm1
 876         pxor    xmm6,xmm14
 877         pxor    xmm13,xmm3
 878         movdqa  xmm3,xmm7
 879         pxor    xmm2,xmm9
 880         movdqa  xmm5,xmm13
 881         movdqa  xmm4,xmm8
 882         movdqa  xmm1,xmm2
 883         movdqa  xmm2,xmm10
 884         movdqa  xmm7,XMMWORD[((-16))+r11]
 885         jnz     NEAR $L$dec_loop
 886         movdqa  xmm7,XMMWORD[((-32))+r11]
 887         jmp     NEAR $L$dec_loop
 888 ALIGN   16
 889 $L$dec_done:
 890         movdqa  xmm7,XMMWORD[r11]
 891         movdqa  xmm8,XMMWORD[16+r11]
 892         movdqa  xmm9,xmm2
 893         psrlq   xmm2,1
 894         movdqa  xmm10,xmm1
 895         psrlq   xmm1,1
 896         pxor    xmm2,xmm4
 897         pxor    xmm1,xmm6
 898         pand    xmm2,xmm7
 899         pand    xmm1,xmm7
 900         pxor    xmm4,xmm2
 901         psllq   xmm2,1
 902         pxor    xmm6,xmm1
 903         psllq   xmm1,1
 904         pxor    xmm2,xmm9
 905         pxor    xmm1,xmm10
 906         movdqa  xmm9,xmm5
 907         psrlq   xmm5,1
 908         movdqa  xmm10,xmm15
 909         psrlq   xmm15,1
 910         pxor    xmm5,xmm3
 911         pxor    xmm15,xmm0
 912         pand    xmm5,xmm7
 913         pand    xmm15,xmm7
 914         pxor    xmm3,xmm5
 915         psllq   xmm5,1
 916         pxor    xmm0,xmm15
 917         psllq   xmm15,1
 918         pxor    xmm5,xmm9
 919         pxor    xmm15,xmm10
 920         movdqa  xmm7,XMMWORD[32+r11]
 921         movdqa  xmm9,xmm6
 922         psrlq   xmm6,2
 923         movdqa  xmm10,xmm1
 924         psrlq   xmm1,2
 925         pxor    xmm6,xmm4
 926         pxor    xmm1,xmm2
 927         pand    xmm6,xmm8
 928         pand    xmm1,xmm8
 929         pxor    xmm4,xmm6
 930         psllq   xmm6,2
 931         pxor    xmm2,xmm1
 932         psllq   xmm1,2
 933         pxor    xmm6,xmm9
 934         pxor    xmm1,xmm10
 935         movdqa  xmm9,xmm0
 936         psrlq   xmm0,2
 937         movdqa  xmm10,xmm15
 938         psrlq   xmm15,2
 939         pxor    xmm0,xmm3
 940         pxor    xmm15,xmm5
 941         pand    xmm0,xmm8
 942         pand    xmm15,xmm8
 943         pxor    xmm3,xmm0
 944         psllq   xmm0,2
 945         pxor    xmm5,xmm15
 946         psllq   xmm15,2
 947         pxor    xmm0,xmm9
 948         pxor    xmm15,xmm10
 949         movdqa  xmm9,xmm3
 950         psrlq   xmm3,4
 951         movdqa  xmm10,xmm5
 952         psrlq   xmm5,4
 953         pxor    xmm3,xmm4
 954         pxor    xmm5,xmm2
 955         pand    xmm3,xmm7
 956         pand    xmm5,xmm7
 957         pxor    xmm4,xmm3
 958         psllq   xmm3,4
 959         pxor    xmm2,xmm5
 960         psllq   xmm5,4
 961         pxor    xmm3,xmm9
 962         pxor    xmm5,xmm10
 963         movdqa  xmm9,xmm0
 964         psrlq   xmm0,4
 965         movdqa  xmm10,xmm15
 966         psrlq   xmm15,4
 967         pxor    xmm0,xmm6
 968         pxor    xmm15,xmm1
 969         pand    xmm0,xmm7
 970         pand    xmm15,xmm7
 971         pxor    xmm6,xmm0
 972         psllq   xmm0,4
 973         pxor    xmm1,xmm15
 974         psllq   xmm15,4
 975         pxor    xmm0,xmm9
 976         pxor    xmm15,xmm10
 977         movdqa  xmm7,XMMWORD[rax]
 978         pxor    xmm5,xmm7
 979         pxor    xmm3,xmm7
 980         pxor    xmm1,xmm7
 981         pxor    xmm6,xmm7
 982         pxor    xmm2,xmm7
 983         pxor    xmm4,xmm7
 984         pxor    xmm15,xmm7
 985         pxor    xmm0,xmm7
 986         DB      0F3h,0C3h               ;repret
 987
 988
 989 ALIGN   16
 990 _bsaes_key_convert:
 991         lea     r11,[$L$masks]
 992         movdqu  xmm7,XMMWORD[rcx]
 993         lea     rcx,[16+rcx]
 994         movdqa  xmm0,XMMWORD[r11]
 995         movdqa  xmm1,XMMWORD[16+r11]
 996         movdqa  xmm2,XMMWORD[32+r11]
 997         movdqa  xmm3,XMMWORD[48+r11]
 998         movdqa  xmm4,XMMWORD[64+r11]
 999         pcmpeqd xmm5,xmm5
1000
1001         movdqu  xmm6,XMMWORD[rcx]
1002         movdqa  XMMWORD[rax],xmm7
1003         lea     rax,[16+rax]
1004         dec     r10d
1005         jmp     NEAR $L$key_loop
1006 ALIGN   16
1007 $L$key_loop:
1008 DB      102,15,56,0,244
1009
1010         movdqa  xmm8,xmm0
1011         movdqa  xmm9,xmm1
1012
1013         pand    xmm8,xmm6
1014         pand    xmm9,xmm6
1015         movdqa  xmm10,xmm2
1016         pcmpeqb xmm8,xmm0
1017         psllq   xmm0,4
1018         movdqa  xmm11,xmm3
1019         pcmpeqb xmm9,xmm1
1020         psllq   xmm1,4
1021
1022         pand    xmm10,xmm6
1023         pand    xmm11,xmm6
1024         movdqa  xmm12,xmm0
1025         pcmpeqb xmm10,xmm2
1026         psllq   xmm2,4
1027         movdqa  xmm13,xmm1
1028         pcmpeqb xmm11,xmm3
1029         psllq   xmm3,4
1030
1031         movdqa  xmm14,xmm2
1032         movdqa  xmm15,xmm3
1033         pxor    xmm8,xmm5
1034         pxor    xmm9,xmm5
1035
1036         pand    xmm12,xmm6
1037         pand    xmm13,xmm6
1038         movdqa  XMMWORD[rax],xmm8
1039         pcmpeqb xmm12,xmm0
1040         psrlq   xmm0,4
1041         movdqa  XMMWORD[16+rax],xmm9
1042         pcmpeqb xmm13,xmm1
1043         psrlq   xmm1,4
1044         lea     rcx,[16+rcx]
1045
1046         pand    xmm14,xmm6
1047         pand    xmm15,xmm6
1048         movdqa  XMMWORD[32+rax],xmm10
1049         pcmpeqb xmm14,xmm2
1050         psrlq   xmm2,4
1051         movdqa  XMMWORD[48+rax],xmm11
1052         pcmpeqb xmm15,xmm3
1053         psrlq   xmm3,4
1054         movdqu  xmm6,XMMWORD[rcx]
1055
1056         pxor    xmm13,xmm5
1057         pxor    xmm14,xmm5
1058         movdqa  XMMWORD[64+rax],xmm12
1059         movdqa  XMMWORD[80+rax],xmm13
1060         movdqa  XMMWORD[96+rax],xmm14
1061         movdqa  XMMWORD[112+rax],xmm15
1062         lea     rax,[128+rax]
1063         dec     r10d
1064         jnz     NEAR $L$key_loop
1065
1066         movdqa  xmm7,XMMWORD[80+r11]
1067
1068         DB      0F3h,0C3h               ;repret
1069
1070 EXTERN  asm_AES_cbc_encrypt
1071 global  bsaes_cbc_encrypt
1072
1073 ALIGN   16
1074 bsaes_cbc_encrypt:
1075         mov     r11d,DWORD[48+rsp]
1076         cmp     r11d,0
1077         jne     NEAR asm_AES_cbc_encrypt
1078         cmp     r8,128
1079         jb      NEAR asm_AES_cbc_encrypt
1080
1081         mov     rax,rsp
1082 $L$cbc_dec_prologue:
1083         push    rbp
1084         push    rbx
1085         push    r12
1086         push    r13
1087         push    r14
1088         push    r15
1089         lea     rsp,[((-72))+rsp]
1090         mov     r10,QWORD[160+rsp]
1091         lea     rsp,[((-160))+rsp]
1092         movaps  XMMWORD[64+rsp],xmm6
1093         movaps  XMMWORD[80+rsp],xmm7
1094         movaps  XMMWORD[96+rsp],xmm8
1095         movaps  XMMWORD[112+rsp],xmm9
1096         movaps  XMMWORD[128+rsp],xmm10
1097         movaps  XMMWORD[144+rsp],xmm11
1098         movaps  XMMWORD[160+rsp],xmm12
1099         movaps  XMMWORD[176+rsp],xmm13
1100         movaps  XMMWORD[192+rsp],xmm14
1101         movaps  XMMWORD[208+rsp],xmm15
1102 $L$cbc_dec_body:
1103         mov     rbp,rsp
1104         mov     eax,DWORD[240+r9]
1105         mov     r12,rcx
1106         mov     r13,rdx
1107         mov     r14,r8
1108         mov     r15,r9
1109         mov     rbx,r10
1110         shr     r14,4
1111
1112         mov     edx,eax
1113         shl     rax,7
1114         sub     rax,96
1115         sub     rsp,rax
1116
1117         mov     rax,rsp
1118         mov     rcx,r15
1119         mov     r10d,edx
1120         call    _bsaes_key_convert
1121         pxor    xmm7,XMMWORD[rsp]
1122         movdqa  XMMWORD[rax],xmm6
1123         movdqa  XMMWORD[rsp],xmm7
1124
1125         movdqu  xmm14,XMMWORD[rbx]
1126         sub     r14,8
1127 $L$cbc_dec_loop:
1128         movdqu  xmm15,XMMWORD[r12]
1129         movdqu  xmm0,XMMWORD[16+r12]
1130         movdqu  xmm1,XMMWORD[32+r12]
1131         movdqu  xmm2,XMMWORD[48+r12]
1132         movdqu  xmm3,XMMWORD[64+r12]
1133         movdqu  xmm4,XMMWORD[80+r12]
1134         mov     rax,rsp
1135         movdqu  xmm5,XMMWORD[96+r12]
1136         mov     r10d,edx
1137         movdqu  xmm6,XMMWORD[112+r12]
1138         movdqa  XMMWORD[32+rbp],xmm14
1139
1140         call    _bsaes_decrypt8
1141
1142         pxor    xmm15,XMMWORD[32+rbp]
1143         movdqu  xmm7,XMMWORD[r12]
1144         movdqu  xmm8,XMMWORD[16+r12]
1145         pxor    xmm0,xmm7
1146         movdqu  xmm9,XMMWORD[32+r12]
1147         pxor    xmm5,xmm8
1148         movdqu  xmm10,XMMWORD[48+r12]
1149         pxor    xmm3,xmm9
1150         movdqu  xmm11,XMMWORD[64+r12]
1151         pxor    xmm1,xmm10
1152         movdqu  xmm12,XMMWORD[80+r12]
1153         pxor    xmm6,xmm11
1154         movdqu  xmm13,XMMWORD[96+r12]
1155         pxor    xmm2,xmm12
1156         movdqu  xmm14,XMMWORD[112+r12]
1157         pxor    xmm4,xmm13
1158         movdqu  XMMWORD[r13],xmm15
1159         lea     r12,[128+r12]
1160         movdqu  XMMWORD[16+r13],xmm0
1161         movdqu  XMMWORD[32+r13],xmm5
1162         movdqu  XMMWORD[48+r13],xmm3
1163         movdqu  XMMWORD[64+r13],xmm1
1164         movdqu  XMMWORD[80+r13],xmm6
1165         movdqu  XMMWORD[96+r13],xmm2
1166         movdqu  XMMWORD[112+r13],xmm4
1167         lea     r13,[128+r13]
1168         sub     r14,8
1169         jnc     NEAR $L$cbc_dec_loop
1170
1171         add     r14,8
1172         jz      NEAR $L$cbc_dec_done
1173
1174         movdqu  xmm15,XMMWORD[r12]
1175         mov     rax,rsp
1176         mov     r10d,edx
1177         cmp     r14,2
1178         jb      NEAR $L$cbc_dec_one
1179         movdqu  xmm0,XMMWORD[16+r12]
1180         je      NEAR $L$cbc_dec_two
1181         movdqu  xmm1,XMMWORD[32+r12]
1182         cmp     r14,4
1183         jb      NEAR $L$cbc_dec_three
1184         movdqu  xmm2,XMMWORD[48+r12]
1185         je      NEAR $L$cbc_dec_four
1186         movdqu  xmm3,XMMWORD[64+r12]
1187         cmp     r14,6
1188         jb      NEAR $L$cbc_dec_five
1189         movdqu  xmm4,XMMWORD[80+r12]
1190         je      NEAR $L$cbc_dec_six
1191         movdqu  xmm5,XMMWORD[96+r12]
1192         movdqa  XMMWORD[32+rbp],xmm14
1193         call    _bsaes_decrypt8
1194         pxor    xmm15,XMMWORD[32+rbp]
1195         movdqu  xmm7,XMMWORD[r12]
1196         movdqu  xmm8,XMMWORD[16+r12]
1197         pxor    xmm0,xmm7
1198         movdqu  xmm9,XMMWORD[32+r12]
1199         pxor    xmm5,xmm8
1200         movdqu  xmm10,XMMWORD[48+r12]
1201         pxor    xmm3,xmm9
1202         movdqu  xmm11,XMMWORD[64+r12]
1203         pxor    xmm1,xmm10
1204         movdqu  xmm12,XMMWORD[80+r12]
1205         pxor    xmm6,xmm11
1206         movdqu  xmm14,XMMWORD[96+r12]
1207         pxor    xmm2,xmm12
1208         movdqu  XMMWORD[r13],xmm15
1209         movdqu  XMMWORD[16+r13],xmm0
1210         movdqu  XMMWORD[32+r13],xmm5
1211         movdqu  XMMWORD[48+r13],xmm3
1212         movdqu  XMMWORD[64+r13],xmm1
1213         movdqu  XMMWORD[80+r13],xmm6
1214         movdqu  XMMWORD[96+r13],xmm2
1215         jmp     NEAR $L$cbc_dec_done
1216 ALIGN   16
1217 $L$cbc_dec_six:
1218         movdqa  XMMWORD[32+rbp],xmm14
1219         call    _bsaes_decrypt8
1220         pxor    xmm15,XMMWORD[32+rbp]
1221         movdqu  xmm7,XMMWORD[r12]
1222         movdqu  xmm8,XMMWORD[16+r12]
1223         pxor    xmm0,xmm7
1224         movdqu  xmm9,XMMWORD[32+r12]
1225         pxor    xmm5,xmm8
1226         movdqu  xmm10,XMMWORD[48+r12]
1227         pxor    xmm3,xmm9
1228         movdqu  xmm11,XMMWORD[64+r12]
1229         pxor    xmm1,xmm10
1230         movdqu  xmm14,XMMWORD[80+r12]
1231         pxor    xmm6,xmm11
1232         movdqu  XMMWORD[r13],xmm15
1233         movdqu  XMMWORD[16+r13],xmm0
1234         movdqu  XMMWORD[32+r13],xmm5
1235         movdqu  XMMWORD[48+r13],xmm3
1236         movdqu  XMMWORD[64+r13],xmm1
1237         movdqu  XMMWORD[80+r13],xmm6
1238         jmp     NEAR $L$cbc_dec_done
1239 ALIGN   16
1240 $L$cbc_dec_five:
1241         movdqa  XMMWORD[32+rbp],xmm14
1242         call    _bsaes_decrypt8
1243         pxor    xmm15,XMMWORD[32+rbp]
1244         movdqu  xmm7,XMMWORD[r12]
1245         movdqu  xmm8,XMMWORD[16+r12]
1246         pxor    xmm0,xmm7
1247         movdqu  xmm9,XMMWORD[32+r12]
1248         pxor    xmm5,xmm8
1249         movdqu  xmm10,XMMWORD[48+r12]
1250         pxor    xmm3,xmm9
1251         movdqu  xmm14,XMMWORD[64+r12]
1252         pxor    xmm1,xmm10
1253         movdqu  XMMWORD[r13],xmm15
1254         movdqu  XMMWORD[16+r13],xmm0
1255         movdqu  XMMWORD[32+r13],xmm5
1256         movdqu  XMMWORD[48+r13],xmm3
1257         movdqu  XMMWORD[64+r13],xmm1
1258         jmp     NEAR $L$cbc_dec_done
1259 ALIGN   16
1260 $L$cbc_dec_four:
1261         movdqa  XMMWORD[32+rbp],xmm14
1262         call    _bsaes_decrypt8
1263         pxor    xmm15,XMMWORD[32+rbp]
1264         movdqu  xmm7,XMMWORD[r12]
1265         movdqu  xmm8,XMMWORD[16+r12]
1266         pxor    xmm0,xmm7
1267         movdqu  xmm9,XMMWORD[32+r12]
1268         pxor    xmm5,xmm8
1269         movdqu  xmm14,XMMWORD[48+r12]
1270         pxor    xmm3,xmm9
1271         movdqu  XMMWORD[r13],xmm15
1272         movdqu  XMMWORD[16+r13],xmm0
1273         movdqu  XMMWORD[32+r13],xmm5
1274         movdqu  XMMWORD[48+r13],xmm3
1275         jmp     NEAR $L$cbc_dec_done
1276 ALIGN   16
1277 $L$cbc_dec_three:
1278         movdqa  XMMWORD[32+rbp],xmm14
1279         call    _bsaes_decrypt8
1280         pxor    xmm15,XMMWORD[32+rbp]
1281         movdqu  xmm7,XMMWORD[r12]
1282         movdqu  xmm8,XMMWORD[16+r12]
1283         pxor    xmm0,xmm7
1284         movdqu  xmm14,XMMWORD[32+r12]
1285         pxor    xmm5,xmm8
1286         movdqu  XMMWORD[r13],xmm15
1287         movdqu  XMMWORD[16+r13],xmm0
1288         movdqu  XMMWORD[32+r13],xmm5
1289         jmp     NEAR $L$cbc_dec_done
1290 ALIGN   16
1291 $L$cbc_dec_two:
1292         movdqa  XMMWORD[32+rbp],xmm14
1293         call    _bsaes_decrypt8
1294         pxor    xmm15,XMMWORD[32+rbp]
1295         movdqu  xmm7,XMMWORD[r12]
1296         movdqu  xmm14,XMMWORD[16+r12]
1297         pxor    xmm0,xmm7
1298         movdqu  XMMWORD[r13],xmm15
1299         movdqu  XMMWORD[16+r13],xmm0
1300         jmp     NEAR $L$cbc_dec_done
1301 ALIGN   16
1302 $L$cbc_dec_one:
1303         lea     rcx,[r12]
1304         lea     rdx,[32+rbp]
1305         lea     r8,[r15]
1306         call    asm_AES_decrypt
1307         pxor    xmm14,XMMWORD[32+rbp]
1308         movdqu  XMMWORD[r13],xmm14
1309         movdqa  xmm14,xmm15
1310
1311 $L$cbc_dec_done:
1312         movdqu  XMMWORD[rbx],xmm14
1313         lea     rax,[rsp]
1314         pxor    xmm0,xmm0
1315 $L$cbc_dec_bzero:
1316         movdqa  XMMWORD[rax],xmm0
1317         movdqa  XMMWORD[16+rax],xmm0
1318         lea     rax,[32+rax]
1319         cmp     rbp,rax
1320         ja      NEAR $L$cbc_dec_bzero
1321
1322         lea     rsp,[rbp]
1323         movaps  xmm6,XMMWORD[64+rbp]
1324         movaps  xmm7,XMMWORD[80+rbp]
1325         movaps  xmm8,XMMWORD[96+rbp]
1326         movaps  xmm9,XMMWORD[112+rbp]
1327         movaps  xmm10,XMMWORD[128+rbp]
1328         movaps  xmm11,XMMWORD[144+rbp]
1329         movaps  xmm12,XMMWORD[160+rbp]
1330         movaps  xmm13,XMMWORD[176+rbp]
1331         movaps  xmm14,XMMWORD[192+rbp]
1332         movaps  xmm15,XMMWORD[208+rbp]
1333         lea     rsp,[160+rbp]
1334         mov     r15,QWORD[72+rsp]
1335         mov     r14,QWORD[80+rsp]
1336         mov     r13,QWORD[88+rsp]
1337         mov     r12,QWORD[96+rsp]
1338         mov     rbx,QWORD[104+rsp]
1339         mov     rax,QWORD[112+rsp]
1340         lea     rsp,[120+rsp]
1341         mov     rbp,rax
1342 $L$cbc_dec_epilogue:
1343         DB      0F3h,0C3h               ;repret
1344
1345
1346 global  bsaes_ctr32_encrypt_blocks
1347
1348 ALIGN   16
1349 bsaes_ctr32_encrypt_blocks:
1350         mov     rax,rsp
1351 $L$ctr_enc_prologue:
1352         push    rbp
1353         push    rbx
1354         push    r12
1355         push    r13
1356         push    r14
1357         push    r15
1358         lea     rsp,[((-72))+rsp]
1359         mov     r10,QWORD[160+rsp]
1360         lea     rsp,[((-160))+rsp]
1361         movaps  XMMWORD[64+rsp],xmm6
1362         movaps  XMMWORD[80+rsp],xmm7
1363         movaps  XMMWORD[96+rsp],xmm8
1364         movaps  XMMWORD[112+rsp],xmm9
1365         movaps  XMMWORD[128+rsp],xmm10
1366         movaps  XMMWORD[144+rsp],xmm11
1367         movaps  XMMWORD[160+rsp],xmm12
1368         movaps  XMMWORD[176+rsp],xmm13
1369         movaps  XMMWORD[192+rsp],xmm14
1370         movaps  XMMWORD[208+rsp],xmm15
1371 $L$ctr_enc_body:
1372         mov     rbp,rsp
1373         movdqu  xmm0,XMMWORD[r10]
1374         mov     eax,DWORD[240+r9]
1375         mov     r12,rcx
1376         mov     r13,rdx
1377         mov     r14,r8
1378         mov     r15,r9
1379         movdqa  XMMWORD[32+rbp],xmm0
1380         cmp     r8,8
1381         jb      NEAR $L$ctr_enc_short
1382
1383         mov     ebx,eax
1384         shl     rax,7
1385         sub     rax,96
1386         sub     rsp,rax
1387
1388         mov     rax,rsp
1389         mov     rcx,r15
1390         mov     r10d,ebx
1391         call    _bsaes_key_convert
1392         pxor    xmm7,xmm6
1393         movdqa  XMMWORD[rax],xmm7
1394
1395         movdqa  xmm8,XMMWORD[rsp]
1396         lea     r11,[$L$ADD1]
1397         movdqa  xmm15,XMMWORD[32+rbp]
1398         movdqa  xmm7,XMMWORD[((-32))+r11]
1399 DB      102,68,15,56,0,199
1400 DB      102,68,15,56,0,255
1401         movdqa  XMMWORD[rsp],xmm8
1402         jmp     NEAR $L$ctr_enc_loop
1403 ALIGN   16
1404 $L$ctr_enc_loop:
1405         movdqa  XMMWORD[32+rbp],xmm15
1406         movdqa  xmm0,xmm15
1407         movdqa  xmm1,xmm15
1408         paddd   xmm0,XMMWORD[r11]
1409         movdqa  xmm2,xmm15
1410         paddd   xmm1,XMMWORD[16+r11]
1411         movdqa  xmm3,xmm15
1412         paddd   xmm2,XMMWORD[32+r11]
1413         movdqa  xmm4,xmm15
1414         paddd   xmm3,XMMWORD[48+r11]
1415         movdqa  xmm5,xmm15
1416         paddd   xmm4,XMMWORD[64+r11]
1417         movdqa  xmm6,xmm15
1418         paddd   xmm5,XMMWORD[80+r11]
1419         paddd   xmm6,XMMWORD[96+r11]
1420
1421
1422
1423         movdqa  xmm8,XMMWORD[rsp]
1424         lea     rax,[16+rsp]
1425         movdqa  xmm7,XMMWORD[((-16))+r11]
1426         pxor    xmm15,xmm8
1427         pxor    xmm0,xmm8
1428         pxor    xmm1,xmm8
1429         pxor    xmm2,xmm8
1430 DB      102,68,15,56,0,255
1431 DB      102,15,56,0,199
1432         pxor    xmm3,xmm8
1433         pxor    xmm4,xmm8
1434 DB      102,15,56,0,207
1435 DB      102,15,56,0,215
1436         pxor    xmm5,xmm8
1437         pxor    xmm6,xmm8
1438 DB      102,15,56,0,223
1439 DB      102,15,56,0,231
1440 DB      102,15,56,0,239
1441 DB      102,15,56,0,247
1442         lea     r11,[$L$BS0]
1443         mov     r10d,ebx
1444
1445         call    _bsaes_encrypt8_bitslice
1446
1447         sub     r14,8
1448         jc      NEAR $L$ctr_enc_loop_done
1449
1450         movdqu  xmm7,XMMWORD[r12]
1451         movdqu  xmm8,XMMWORD[16+r12]
1452         movdqu  xmm9,XMMWORD[32+r12]
1453         movdqu  xmm10,XMMWORD[48+r12]
1454         movdqu  xmm11,XMMWORD[64+r12]
1455         movdqu  xmm12,XMMWORD[80+r12]
1456         movdqu  xmm13,XMMWORD[96+r12]
1457         movdqu  xmm14,XMMWORD[112+r12]
1458         lea     r12,[128+r12]
1459         pxor    xmm7,xmm15
1460         movdqa  xmm15,XMMWORD[32+rbp]
1461         pxor    xmm0,xmm8
1462         movdqu  XMMWORD[r13],xmm7
1463         pxor    xmm3,xmm9
1464         movdqu  XMMWORD[16+r13],xmm0
1465         pxor    xmm5,xmm10
1466         movdqu  XMMWORD[32+r13],xmm3
1467         pxor    xmm2,xmm11
1468         movdqu  XMMWORD[48+r13],xmm5
1469         pxor    xmm6,xmm12
1470         movdqu  XMMWORD[64+r13],xmm2
1471         pxor    xmm1,xmm13
1472         movdqu  XMMWORD[80+r13],xmm6
1473         pxor    xmm4,xmm14
1474         movdqu  XMMWORD[96+r13],xmm1
1475         lea     r11,[$L$ADD1]
1476         movdqu  XMMWORD[112+r13],xmm4
1477         lea     r13,[128+r13]
1478         paddd   xmm15,XMMWORD[112+r11]
1479         jnz     NEAR $L$ctr_enc_loop
1480
1481         jmp     NEAR $L$ctr_enc_done
1482 ALIGN   16
1483 $L$ctr_enc_loop_done:
1484         add     r14,8
1485         movdqu  xmm7,XMMWORD[r12]
1486         pxor    xmm15,xmm7
1487         movdqu  XMMWORD[r13],xmm15
1488         cmp     r14,2
1489         jb      NEAR $L$ctr_enc_done
1490         movdqu  xmm8,XMMWORD[16+r12]
1491         pxor    xmm0,xmm8
1492         movdqu  XMMWORD[16+r13],xmm0
1493         je      NEAR $L$ctr_enc_done
1494         movdqu  xmm9,XMMWORD[32+r12]
1495         pxor    xmm3,xmm9
1496         movdqu  XMMWORD[32+r13],xmm3
1497         cmp     r14,4
1498         jb      NEAR $L$ctr_enc_done
1499         movdqu  xmm10,XMMWORD[48+r12]
1500         pxor    xmm5,xmm10
1501         movdqu  XMMWORD[48+r13],xmm5
1502         je      NEAR $L$ctr_enc_done
1503         movdqu  xmm11,XMMWORD[64+r12]
1504         pxor    xmm2,xmm11
1505         movdqu  XMMWORD[64+r13],xmm2
1506         cmp     r14,6
1507         jb      NEAR $L$ctr_enc_done
1508         movdqu  xmm12,XMMWORD[80+r12]
1509         pxor    xmm6,xmm12
1510         movdqu  XMMWORD[80+r13],xmm6
1511         je      NEAR $L$ctr_enc_done
1512         movdqu  xmm13,XMMWORD[96+r12]
1513         pxor    xmm1,xmm13
1514         movdqu  XMMWORD[96+r13],xmm1
1515         jmp     NEAR $L$ctr_enc_done
1516
1517 ALIGN   16
1518 $L$ctr_enc_short:
1519         lea     rcx,[32+rbp]
1520         lea     rdx,[48+rbp]
1521         lea     r8,[r15]
1522         call    asm_AES_encrypt
1523         movdqu  xmm0,XMMWORD[r12]
1524         lea     r12,[16+r12]
1525         mov     eax,DWORD[44+rbp]
1526         bswap   eax
1527         pxor    xmm0,XMMWORD[48+rbp]
1528         inc     eax
1529         movdqu  XMMWORD[r13],xmm0
1530         bswap   eax
1531         lea     r13,[16+r13]
1532         mov     DWORD[44+rsp],eax
1533         dec     r14
1534         jnz     NEAR $L$ctr_enc_short
1535
1536 $L$ctr_enc_done:
1537         lea     rax,[rsp]
1538         pxor    xmm0,xmm0
1539 $L$ctr_enc_bzero:
1540         movdqa  XMMWORD[rax],xmm0
1541         movdqa  XMMWORD[16+rax],xmm0
1542         lea     rax,[32+rax]
1543         cmp     rbp,rax
1544         ja      NEAR $L$ctr_enc_bzero
1545
1546         lea     rsp,[rbp]
1547         movaps  xmm6,XMMWORD[64+rbp]
1548         movaps  xmm7,XMMWORD[80+rbp]
1549         movaps  xmm8,XMMWORD[96+rbp]
1550         movaps  xmm9,XMMWORD[112+rbp]
1551         movaps  xmm10,XMMWORD[128+rbp]
1552         movaps  xmm11,XMMWORD[144+rbp]
1553         movaps  xmm12,XMMWORD[160+rbp]
1554         movaps  xmm13,XMMWORD[176+rbp]
1555         movaps  xmm14,XMMWORD[192+rbp]
1556         movaps  xmm15,XMMWORD[208+rbp]
1557         lea     rsp,[160+rbp]
1558         mov     r15,QWORD[72+rsp]
1559         mov     r14,QWORD[80+rsp]
1560         mov     r13,QWORD[88+rsp]
1561         mov     r12,QWORD[96+rsp]
1562         mov     rbx,QWORD[104+rsp]
1563         mov     rax,QWORD[112+rsp]
1564         lea     rsp,[120+rsp]
1565         mov     rbp,rax
1566 $L$ctr_enc_epilogue:
1567         DB      0F3h,0C3h               ;repret
1568
1569 global  bsaes_xts_encrypt
1570
1571 ALIGN   16
1572 bsaes_xts_encrypt:
1573         mov     rax,rsp
1574 $L$xts_enc_prologue:
1575         push    rbp
1576         push    rbx
1577         push    r12
1578         push    r13
1579         push    r14
1580         push    r15
1581         lea     rsp,[((-72))+rsp]
1582         mov     r10,QWORD[160+rsp]
1583         mov     r11,QWORD[168+rsp]
1584         lea     rsp,[((-160))+rsp]
1585         movaps  XMMWORD[64+rsp],xmm6
1586         movaps  XMMWORD[80+rsp],xmm7
1587         movaps  XMMWORD[96+rsp],xmm8
1588         movaps  XMMWORD[112+rsp],xmm9
1589         movaps  XMMWORD[128+rsp],xmm10
1590         movaps  XMMWORD[144+rsp],xmm11
1591         movaps  XMMWORD[160+rsp],xmm12
1592         movaps  XMMWORD[176+rsp],xmm13
1593         movaps  XMMWORD[192+rsp],xmm14
1594         movaps  XMMWORD[208+rsp],xmm15
1595 $L$xts_enc_body:
1596         mov     rbp,rsp
1597         mov     r12,rcx
1598         mov     r13,rdx
1599         mov     r14,r8
1600         mov     r15,r9
1601
1602         lea     rcx,[r11]
1603         lea     rdx,[32+rbp]
1604         lea     r8,[r10]
1605         call    asm_AES_encrypt
1606
1607         mov     eax,DWORD[240+r15]
1608         mov     rbx,r14
1609
1610         mov     edx,eax
1611         shl     rax,7
1612         sub     rax,96
1613         sub     rsp,rax
1614
1615         mov     rax,rsp
1616         mov     rcx,r15
1617         mov     r10d,edx
1618         call    _bsaes_key_convert
1619         pxor    xmm7,xmm6
1620         movdqa  XMMWORD[rax],xmm7
1621
1622         and     r14,-16
1623         sub     rsp,0x80
1624         movdqa  xmm6,XMMWORD[32+rbp]
1625
1626         pxor    xmm14,xmm14
1627         movdqa  xmm12,XMMWORD[$L$xts_magic]
1628         pcmpgtd xmm14,xmm6
1629
1630         sub     r14,0x80
1631         jc      NEAR $L$xts_enc_short
1632         jmp     NEAR $L$xts_enc_loop
1633
1634 ALIGN   16
1635 $L$xts_enc_loop:
1636         pshufd  xmm13,xmm14,0x13
1637         pxor    xmm14,xmm14
1638         movdqa  xmm15,xmm6
1639         movdqa  XMMWORD[rsp],xmm6
1640         paddq   xmm6,xmm6
1641         pand    xmm13,xmm12
1642         pcmpgtd xmm14,xmm6
1643         pxor    xmm6,xmm13
1644         pshufd  xmm13,xmm14,0x13
1645         pxor    xmm14,xmm14
1646         movdqa  xmm0,xmm6
1647         movdqa  XMMWORD[16+rsp],xmm6
1648         paddq   xmm6,xmm6
1649         pand    xmm13,xmm12
1650         pcmpgtd xmm14,xmm6
1651         pxor    xmm6,xmm13
1652         movdqu  xmm7,XMMWORD[r12]
1653         pshufd  xmm13,xmm14,0x13
1654         pxor    xmm14,xmm14
1655         movdqa  xmm1,xmm6
1656         movdqa  XMMWORD[32+rsp],xmm6
1657         paddq   xmm6,xmm6
1658         pand    xmm13,xmm12
1659         pcmpgtd xmm14,xmm6
1660         pxor    xmm6,xmm13
1661         movdqu  xmm8,XMMWORD[16+r12]
1662         pxor    xmm15,xmm7
1663         pshufd  xmm13,xmm14,0x13
1664         pxor    xmm14,xmm14
1665         movdqa  xmm2,xmm6
1666         movdqa  XMMWORD[48+rsp],xmm6
1667         paddq   xmm6,xmm6
1668         pand    xmm13,xmm12
1669         pcmpgtd xmm14,xmm6
1670         pxor    xmm6,xmm13
1671         movdqu  xmm9,XMMWORD[32+r12]
1672         pxor    xmm0,xmm8
1673         pshufd  xmm13,xmm14,0x13
1674         pxor    xmm14,xmm14
1675         movdqa  xmm3,xmm6
1676         movdqa  XMMWORD[64+rsp],xmm6
1677         paddq   xmm6,xmm6
1678         pand    xmm13,xmm12
1679         pcmpgtd xmm14,xmm6
1680         pxor    xmm6,xmm13
1681         movdqu  xmm10,XMMWORD[48+r12]
1682         pxor    xmm1,xmm9
1683         pshufd  xmm13,xmm14,0x13
1684         pxor    xmm14,xmm14
1685         movdqa  xmm4,xmm6
1686         movdqa  XMMWORD[80+rsp],xmm6
1687         paddq   xmm6,xmm6
1688         pand    xmm13,xmm12
1689         pcmpgtd xmm14,xmm6
1690         pxor    xmm6,xmm13
1691         movdqu  xmm11,XMMWORD[64+r12]
1692         pxor    xmm2,xmm10
1693         pshufd  xmm13,xmm14,0x13
1694         pxor    xmm14,xmm14
1695         movdqa  xmm5,xmm6
1696         movdqa  XMMWORD[96+rsp],xmm6
1697         paddq   xmm6,xmm6
1698         pand    xmm13,xmm12
1699         pcmpgtd xmm14,xmm6
1700         pxor    xmm6,xmm13
1701         movdqu  xmm12,XMMWORD[80+r12]
1702         pxor    xmm3,xmm11
1703         movdqu  xmm13,XMMWORD[96+r12]
1704         pxor    xmm4,xmm12
1705         movdqu  xmm14,XMMWORD[112+r12]
1706         lea     r12,[128+r12]
1707         movdqa  XMMWORD[112+rsp],xmm6
1708         pxor    xmm5,xmm13
1709         lea     rax,[128+rsp]
1710         pxor    xmm6,xmm14
1711         mov     r10d,edx
1712
1713         call    _bsaes_encrypt8
1714
1715         pxor    xmm15,XMMWORD[rsp]
1716         pxor    xmm0,XMMWORD[16+rsp]
1717         movdqu  XMMWORD[r13],xmm15
1718         pxor    xmm3,XMMWORD[32+rsp]
1719         movdqu  XMMWORD[16+r13],xmm0
1720         pxor    xmm5,XMMWORD[48+rsp]
1721         movdqu  XMMWORD[32+r13],xmm3
1722         pxor    xmm2,XMMWORD[64+rsp]
1723         movdqu  XMMWORD[48+r13],xmm5
1724         pxor    xmm6,XMMWORD[80+rsp]
1725         movdqu  XMMWORD[64+r13],xmm2
1726         pxor    xmm1,XMMWORD[96+rsp]
1727         movdqu  XMMWORD[80+r13],xmm6
1728         pxor    xmm4,XMMWORD[112+rsp]
1729         movdqu  XMMWORD[96+r13],xmm1
1730         movdqu  XMMWORD[112+r13],xmm4
1731         lea     r13,[128+r13]
1732
1733         movdqa  xmm6,XMMWORD[112+rsp]
1734         pxor    xmm14,xmm14
1735         movdqa  xmm12,XMMWORD[$L$xts_magic]
1736         pcmpgtd xmm14,xmm6
1737         pshufd  xmm13,xmm14,0x13
1738         pxor    xmm14,xmm14
1739         paddq   xmm6,xmm6
1740         pand    xmm13,xmm12
1741         pcmpgtd xmm14,xmm6
1742         pxor    xmm6,xmm13
1743
1744         sub     r14,0x80
1745         jnc     NEAR $L$xts_enc_loop
1746
1747 $L$xts_enc_short:
1748         add     r14,0x80
1749         jz      NEAR $L$xts_enc_done
1750         pshufd  xmm13,xmm14,0x13
1751         pxor    xmm14,xmm14
1752         movdqa  xmm15,xmm6
1753         movdqa  XMMWORD[rsp],xmm6
1754         paddq   xmm6,xmm6
1755         pand    xmm13,xmm12
1756         pcmpgtd xmm14,xmm6
1757         pxor    xmm6,xmm13
1758         pshufd  xmm13,xmm14,0x13
1759         pxor    xmm14,xmm14
1760         movdqa  xmm0,xmm6
1761         movdqa  XMMWORD[16+rsp],xmm6
1762         paddq   xmm6,xmm6
1763         pand    xmm13,xmm12
1764         pcmpgtd xmm14,xmm6
1765         pxor    xmm6,xmm13
1766         movdqu  xmm7,XMMWORD[r12]
1767         cmp     r14,16
1768         je      NEAR $L$xts_enc_1
1769         pshufd  xmm13,xmm14,0x13
1770         pxor    xmm14,xmm14
1771         movdqa  xmm1,xmm6
1772         movdqa  XMMWORD[32+rsp],xmm6
1773         paddq   xmm6,xmm6
1774         pand    xmm13,xmm12
1775         pcmpgtd xmm14,xmm6
1776         pxor    xmm6,xmm13
1777         movdqu  xmm8,XMMWORD[16+r12]
1778         cmp     r14,32
1779         je      NEAR $L$xts_enc_2
1780         pxor    xmm15,xmm7
1781         pshufd  xmm13,xmm14,0x13
1782         pxor    xmm14,xmm14
1783         movdqa  xmm2,xmm6
1784         movdqa  XMMWORD[48+rsp],xmm6
1785         paddq   xmm6,xmm6
1786         pand    xmm13,xmm12
1787         pcmpgtd xmm14,xmm6
1788         pxor    xmm6,xmm13
1789         movdqu  xmm9,XMMWORD[32+r12]
1790         cmp     r14,48
1791         je      NEAR $L$xts_enc_3
1792         pxor    xmm0,xmm8
1793         pshufd  xmm13,xmm14,0x13
1794         pxor    xmm14,xmm14
1795         movdqa  xmm3,xmm6
1796         movdqa  XMMWORD[64+rsp],xmm6
1797         paddq   xmm6,xmm6
1798         pand    xmm13,xmm12
1799         pcmpgtd xmm14,xmm6
1800         pxor    xmm6,xmm13
1801         movdqu  xmm10,XMMWORD[48+r12]
1802         cmp     r14,64
1803         je      NEAR $L$xts_enc_4
1804         pxor    xmm1,xmm9
1805         pshufd  xmm13,xmm14,0x13
1806         pxor    xmm14,xmm14
1807         movdqa  xmm4,xmm6
1808         movdqa  XMMWORD[80+rsp],xmm6
1809         paddq   xmm6,xmm6
1810         pand    xmm13,xmm12
1811         pcmpgtd xmm14,xmm6
1812         pxor    xmm6,xmm13
1813         movdqu  xmm11,XMMWORD[64+r12]
1814         cmp     r14,80
1815         je      NEAR $L$xts_enc_5
1816         pxor    xmm2,xmm10
1817         pshufd  xmm13,xmm14,0x13
1818         pxor    xmm14,xmm14
1819         movdqa  xmm5,xmm6
1820         movdqa  XMMWORD[96+rsp],xmm6
1821         paddq   xmm6,xmm6
1822         pand    xmm13,xmm12
1823         pcmpgtd xmm14,xmm6
1824         pxor    xmm6,xmm13
1825         movdqu  xmm12,XMMWORD[80+r12]
1826         cmp     r14,96
1827         je      NEAR $L$xts_enc_6
1828         pxor    xmm3,xmm11
1829         movdqu  xmm13,XMMWORD[96+r12]
1830         pxor    xmm4,xmm12
1831         movdqa  XMMWORD[112+rsp],xmm6
1832         lea     r12,[112+r12]
1833         pxor    xmm5,xmm13
1834         lea     rax,[128+rsp]
1835         mov     r10d,edx
1836
1837         call    _bsaes_encrypt8
1838
1839         pxor    xmm15,XMMWORD[rsp]
1840         pxor    xmm0,XMMWORD[16+rsp]
1841         movdqu  XMMWORD[r13],xmm15
1842         pxor    xmm3,XMMWORD[32+rsp]
1843         movdqu  XMMWORD[16+r13],xmm0
1844         pxor    xmm5,XMMWORD[48+rsp]
1845         movdqu  XMMWORD[32+r13],xmm3
1846         pxor    xmm2,XMMWORD[64+rsp]
1847         movdqu  XMMWORD[48+r13],xmm5
1848         pxor    xmm6,XMMWORD[80+rsp]
1849         movdqu  XMMWORD[64+r13],xmm2
1850         pxor    xmm1,XMMWORD[96+rsp]
1851         movdqu  XMMWORD[80+r13],xmm6
1852         movdqu  XMMWORD[96+r13],xmm1
1853         lea     r13,[112+r13]
1854
1855         movdqa  xmm6,XMMWORD[112+rsp]
1856         jmp     NEAR $L$xts_enc_done
1857 ALIGN   16
1858 $L$xts_enc_6:
1859         pxor    xmm3,xmm11
1860         lea     r12,[96+r12]
1861         pxor    xmm4,xmm12
1862         lea     rax,[128+rsp]
1863         mov     r10d,edx
1864
1865         call    _bsaes_encrypt8
1866
1867         pxor    xmm15,XMMWORD[rsp]
1868         pxor    xmm0,XMMWORD[16+rsp]
1869         movdqu  XMMWORD[r13],xmm15
1870         pxor    xmm3,XMMWORD[32+rsp]
1871         movdqu  XMMWORD[16+r13],xmm0
1872         pxor    xmm5,XMMWORD[48+rsp]
1873         movdqu  XMMWORD[32+r13],xmm3
1874         pxor    xmm2,XMMWORD[64+rsp]
1875         movdqu  XMMWORD[48+r13],xmm5
1876         pxor    xmm6,XMMWORD[80+rsp]
1877         movdqu  XMMWORD[64+r13],xmm2
1878         movdqu  XMMWORD[80+r13],xmm6
1879         lea     r13,[96+r13]
1880
1881         movdqa  xmm6,XMMWORD[96+rsp]
1882         jmp     NEAR $L$xts_enc_done
1883 ALIGN   16
1884 $L$xts_enc_5:
1885         pxor    xmm2,xmm10
1886         lea     r12,[80+r12]
1887         pxor    xmm3,xmm11
1888         lea     rax,[128+rsp]
1889         mov     r10d,edx
1890
1891         call    _bsaes_encrypt8
1892
1893         pxor    xmm15,XMMWORD[rsp]
1894         pxor    xmm0,XMMWORD[16+rsp]
1895         movdqu  XMMWORD[r13],xmm15
1896         pxor    xmm3,XMMWORD[32+rsp]
1897         movdqu  XMMWORD[16+r13],xmm0
1898         pxor    xmm5,XMMWORD[48+rsp]
1899         movdqu  XMMWORD[32+r13],xmm3
1900         pxor    xmm2,XMMWORD[64+rsp]
1901         movdqu  XMMWORD[48+r13],xmm5
1902         movdqu  XMMWORD[64+r13],xmm2
1903         lea     r13,[80+r13]
1904
1905         movdqa  xmm6,XMMWORD[80+rsp]
1906         jmp     NEAR $L$xts_enc_done
1907 ALIGN   16
1908 $L$xts_enc_4:
1909         pxor    xmm1,xmm9
1910         lea     r12,[64+r12]
1911         pxor    xmm2,xmm10
1912         lea     rax,[128+rsp]
1913         mov     r10d,edx
1914
1915         call    _bsaes_encrypt8
1916
1917         pxor    xmm15,XMMWORD[rsp]
1918         pxor    xmm0,XMMWORD[16+rsp]
1919         movdqu  XMMWORD[r13],xmm15
1920         pxor    xmm3,XMMWORD[32+rsp]
1921         movdqu  XMMWORD[16+r13],xmm0
1922         pxor    xmm5,XMMWORD[48+rsp]
1923         movdqu  XMMWORD[32+r13],xmm3
1924         movdqu  XMMWORD[48+r13],xmm5
1925         lea     r13,[64+r13]
1926
1927         movdqa  xmm6,XMMWORD[64+rsp]
1928         jmp     NEAR $L$xts_enc_done
1929 ALIGN   16
1930 $L$xts_enc_3:
1931         pxor    xmm0,xmm8
1932         lea     r12,[48+r12]
1933         pxor    xmm1,xmm9
1934         lea     rax,[128+rsp]
1935         mov     r10d,edx
1936
1937         call    _bsaes_encrypt8
1938
1939         pxor    xmm15,XMMWORD[rsp]
1940         pxor    xmm0,XMMWORD[16+rsp]
1941         movdqu  XMMWORD[r13],xmm15
1942         pxor    xmm3,XMMWORD[32+rsp]
1943         movdqu  XMMWORD[16+r13],xmm0
1944         movdqu  XMMWORD[32+r13],xmm3
1945         lea     r13,[48+r13]
1946
1947         movdqa  xmm6,XMMWORD[48+rsp]
1948         jmp     NEAR $L$xts_enc_done
1949 ALIGN   16
1950 $L$xts_enc_2:
1951         pxor    xmm15,xmm7
1952         lea     r12,[32+r12]
1953         pxor    xmm0,xmm8
1954         lea     rax,[128+rsp]
1955         mov     r10d,edx
1956
1957         call    _bsaes_encrypt8
1958
1959         pxor    xmm15,XMMWORD[rsp]
1960         pxor    xmm0,XMMWORD[16+rsp]
1961         movdqu  XMMWORD[r13],xmm15
1962         movdqu  XMMWORD[16+r13],xmm0
1963         lea     r13,[32+r13]
1964
1965         movdqa  xmm6,XMMWORD[32+rsp]
1966         jmp     NEAR $L$xts_enc_done
1967 ALIGN   16
1968 $L$xts_enc_1:
1969         pxor    xmm7,xmm15
1970         lea     r12,[16+r12]
1971         movdqa  XMMWORD[32+rbp],xmm7
1972         lea     rcx,[32+rbp]
1973         lea     rdx,[32+rbp]
1974         lea     r8,[r15]
1975         call    asm_AES_encrypt
1976         pxor    xmm15,XMMWORD[32+rbp]
1977
1978
1979
1980
1981
1982         movdqu  XMMWORD[r13],xmm15
1983         lea     r13,[16+r13]
1984
1985         movdqa  xmm6,XMMWORD[16+rsp]
1986
1987 $L$xts_enc_done:
1988         and     ebx,15
1989         jz      NEAR $L$xts_enc_ret
1990         mov     rdx,r13
1991
1992 $L$xts_enc_steal:
1993         movzx   eax,BYTE[r12]
1994         movzx   ecx,BYTE[((-16))+rdx]
1995         lea     r12,[1+r12]
1996         mov     BYTE[((-16))+rdx],al
1997         mov     BYTE[rdx],cl
1998         lea     rdx,[1+rdx]
1999         sub     ebx,1
2000         jnz     NEAR $L$xts_enc_steal
2001
2002         movdqu  xmm15,XMMWORD[((-16))+r13]
2003         lea     rcx,[32+rbp]
2004         pxor    xmm15,xmm6
2005         lea     rdx,[32+rbp]
2006         movdqa  XMMWORD[32+rbp],xmm15
2007         lea     r8,[r15]
2008         call    asm_AES_encrypt
2009         pxor    xmm6,XMMWORD[32+rbp]
2010         movdqu  XMMWORD[(-16)+r13],xmm6
2011
2012 $L$xts_enc_ret:
2013         lea     rax,[rsp]
2014         pxor    xmm0,xmm0
2015 $L$xts_enc_bzero:
2016         movdqa  XMMWORD[rax],xmm0
2017         movdqa  XMMWORD[16+rax],xmm0
2018         lea     rax,[32+rax]
2019         cmp     rbp,rax
2020         ja      NEAR $L$xts_enc_bzero
2021
2022         lea     rsp,[rbp]
2023         movaps  xmm6,XMMWORD[64+rbp]
2024         movaps  xmm7,XMMWORD[80+rbp]
2025         movaps  xmm8,XMMWORD[96+rbp]
2026         movaps  xmm9,XMMWORD[112+rbp]
2027         movaps  xmm10,XMMWORD[128+rbp]
2028         movaps  xmm11,XMMWORD[144+rbp]
2029         movaps  xmm12,XMMWORD[160+rbp]
2030         movaps  xmm13,XMMWORD[176+rbp]
2031         movaps  xmm14,XMMWORD[192+rbp]
2032         movaps  xmm15,XMMWORD[208+rbp]
2033         lea     rsp,[160+rbp]
2034         mov     r15,QWORD[72+rsp]
2035         mov     r14,QWORD[80+rsp]
2036         mov     r13,QWORD[88+rsp]
2037         mov     r12,QWORD[96+rsp]
2038         mov     rbx,QWORD[104+rsp]
2039         mov     rax,QWORD[112+rsp]
2040         lea     rsp,[120+rsp]
2041         mov     rbp,rax
2042 $L$xts_enc_epilogue:
2043         DB      0F3h,0C3h               ;repret
2044
2045
2046 global  bsaes_xts_decrypt
2047
2048 ALIGN   16
2049 bsaes_xts_decrypt:
2050         mov     rax,rsp
2051 $L$xts_dec_prologue:
2052         push    rbp
2053         push    rbx
2054         push    r12
2055         push    r13
2056         push    r14
2057         push    r15
2058         lea     rsp,[((-72))+rsp]
2059         mov     r10,QWORD[160+rsp]
2060         mov     r11,QWORD[168+rsp]
2061         lea     rsp,[((-160))+rsp]
2062         movaps  XMMWORD[64+rsp],xmm6
2063         movaps  XMMWORD[80+rsp],xmm7
2064         movaps  XMMWORD[96+rsp],xmm8
2065         movaps  XMMWORD[112+rsp],xmm9
2066         movaps  XMMWORD[128+rsp],xmm10
2067         movaps  XMMWORD[144+rsp],xmm11
2068         movaps  XMMWORD[160+rsp],xmm12
2069         movaps  XMMWORD[176+rsp],xmm13
2070         movaps  XMMWORD[192+rsp],xmm14
2071         movaps  XMMWORD[208+rsp],xmm15
2072 $L$xts_dec_body:
2073         mov     rbp,rsp
2074         mov     r12,rcx
2075         mov     r13,rdx
2076         mov     r14,r8
2077         mov     r15,r9
2078
2079         lea     rcx,[r11]
2080         lea     rdx,[32+rbp]
2081         lea     r8,[r10]
2082         call    asm_AES_encrypt
2083
2084         mov     eax,DWORD[240+r15]
2085         mov     rbx,r14
2086
2087         mov     edx,eax
2088         shl     rax,7
2089         sub     rax,96
2090         sub     rsp,rax
2091
2092         mov     rax,rsp
2093         mov     rcx,r15
2094         mov     r10d,edx
2095         call    _bsaes_key_convert
2096         pxor    xmm7,XMMWORD[rsp]
2097         movdqa  XMMWORD[rax],xmm6
2098         movdqa  XMMWORD[rsp],xmm7
2099
2100         xor     eax,eax
2101         and     r14,-16
2102         test    ebx,15
2103         setnz   al
2104         shl     rax,4
2105         sub     r14,rax
2106
2107         sub     rsp,0x80
2108         movdqa  xmm6,XMMWORD[32+rbp]
2109
2110         pxor    xmm14,xmm14
2111         movdqa  xmm12,XMMWORD[$L$xts_magic]
2112         pcmpgtd xmm14,xmm6
2113
2114         sub     r14,0x80
2115         jc      NEAR $L$xts_dec_short
2116         jmp     NEAR $L$xts_dec_loop
2117
2118 ALIGN   16
2119 $L$xts_dec_loop:
2120         pshufd  xmm13,xmm14,0x13
2121         pxor    xmm14,xmm14
2122         movdqa  xmm15,xmm6
2123         movdqa  XMMWORD[rsp],xmm6
2124         paddq   xmm6,xmm6
2125         pand    xmm13,xmm12
2126         pcmpgtd xmm14,xmm6
2127         pxor    xmm6,xmm13
2128         pshufd  xmm13,xmm14,0x13
2129         pxor    xmm14,xmm14
2130         movdqa  xmm0,xmm6
2131         movdqa  XMMWORD[16+rsp],xmm6
2132         paddq   xmm6,xmm6
2133         pand    xmm13,xmm12
2134         pcmpgtd xmm14,xmm6
2135         pxor    xmm6,xmm13
2136         movdqu  xmm7,XMMWORD[r12]
2137         pshufd  xmm13,xmm14,0x13
2138         pxor    xmm14,xmm14
2139         movdqa  xmm1,xmm6
2140         movdqa  XMMWORD[32+rsp],xmm6
2141         paddq   xmm6,xmm6
2142         pand    xmm13,xmm12
2143         pcmpgtd xmm14,xmm6
2144         pxor    xmm6,xmm13
2145         movdqu  xmm8,XMMWORD[16+r12]
2146         pxor    xmm15,xmm7
2147         pshufd  xmm13,xmm14,0x13
2148         pxor    xmm14,xmm14
2149         movdqa  xmm2,xmm6
2150         movdqa  XMMWORD[48+rsp],xmm6
2151         paddq   xmm6,xmm6
2152         pand    xmm13,xmm12
2153         pcmpgtd xmm14,xmm6
2154         pxor    xmm6,xmm13
2155         movdqu  xmm9,XMMWORD[32+r12]
2156         pxor    xmm0,xmm8
2157         pshufd  xmm13,xmm14,0x13
2158         pxor    xmm14,xmm14
2159         movdqa  xmm3,xmm6
2160         movdqa  XMMWORD[64+rsp],xmm6
2161         paddq   xmm6,xmm6
2162         pand    xmm13,xmm12
2163         pcmpgtd xmm14,xmm6
2164         pxor    xmm6,xmm13
2165         movdqu  xmm10,XMMWORD[48+r12]
2166         pxor    xmm1,xmm9
2167         pshufd  xmm13,xmm14,0x13
2168         pxor    xmm14,xmm14
2169         movdqa  xmm4,xmm6
2170         movdqa  XMMWORD[80+rsp],xmm6
2171         paddq   xmm6,xmm6
2172         pand    xmm13,xmm12
2173         pcmpgtd xmm14,xmm6
2174         pxor    xmm6,xmm13
2175         movdqu  xmm11,XMMWORD[64+r12]
2176         pxor    xmm2,xmm10
2177         pshufd  xmm13,xmm14,0x13
2178         pxor    xmm14,xmm14
2179         movdqa  xmm5,xmm6
2180         movdqa  XMMWORD[96+rsp],xmm6
2181         paddq   xmm6,xmm6
2182         pand    xmm13,xmm12
2183         pcmpgtd xmm14,xmm6
2184         pxor    xmm6,xmm13
2185         movdqu  xmm12,XMMWORD[80+r12]
2186         pxor    xmm3,xmm11
2187         movdqu  xmm13,XMMWORD[96+r12]
2188         pxor    xmm4,xmm12
2189         movdqu  xmm14,XMMWORD[112+r12]
2190         lea     r12,[128+r12]
2191         movdqa  XMMWORD[112+rsp],xmm6
2192         pxor    xmm5,xmm13
2193         lea     rax,[128+rsp]
2194         pxor    xmm6,xmm14
2195         mov     r10d,edx
2196
2197         call    _bsaes_decrypt8
2198
2199         pxor    xmm15,XMMWORD[rsp]
2200         pxor    xmm0,XMMWORD[16+rsp]
2201         movdqu  XMMWORD[r13],xmm15
2202         pxor    xmm5,XMMWORD[32+rsp]
2203         movdqu  XMMWORD[16+r13],xmm0
2204         pxor    xmm3,XMMWORD[48+rsp]
2205         movdqu  XMMWORD[32+r13],xmm5
2206         pxor    xmm1,XMMWORD[64+rsp]
2207         movdqu  XMMWORD[48+r13],xmm3
2208         pxor    xmm6,XMMWORD[80+rsp]
2209         movdqu  XMMWORD[64+r13],xmm1
2210         pxor    xmm2,XMMWORD[96+rsp]
2211         movdqu  XMMWORD[80+r13],xmm6
2212         pxor    xmm4,XMMWORD[112+rsp]
2213         movdqu  XMMWORD[96+r13],xmm2
2214         movdqu  XMMWORD[112+r13],xmm4
2215         lea     r13,[128+r13]
2216
2217         movdqa  xmm6,XMMWORD[112+rsp]
2218         pxor    xmm14,xmm14
2219         movdqa  xmm12,XMMWORD[$L$xts_magic]
2220         pcmpgtd xmm14,xmm6
2221         pshufd  xmm13,xmm14,0x13
2222         pxor    xmm14,xmm14
2223         paddq   xmm6,xmm6
2224         pand    xmm13,xmm12
2225         pcmpgtd xmm14,xmm6
2226         pxor    xmm6,xmm13
2227
2228         sub     r14,0x80
2229         jnc     NEAR $L$xts_dec_loop
2230
2231 $L$xts_dec_short:
2232         add     r14,0x80
2233         jz      NEAR $L$xts_dec_done
2234         pshufd  xmm13,xmm14,0x13
2235         pxor    xmm14,xmm14
2236         movdqa  xmm15,xmm6
2237         movdqa  XMMWORD[rsp],xmm6
2238         paddq   xmm6,xmm6
2239         pand    xmm13,xmm12
2240         pcmpgtd xmm14,xmm6
2241         pxor    xmm6,xmm13
2242         pshufd  xmm13,xmm14,0x13
2243         pxor    xmm14,xmm14
2244         movdqa  xmm0,xmm6
2245         movdqa  XMMWORD[16+rsp],xmm6
2246         paddq   xmm6,xmm6
2247         pand    xmm13,xmm12
2248         pcmpgtd xmm14,xmm6
2249         pxor    xmm6,xmm13
2250         movdqu  xmm7,XMMWORD[r12]
2251         cmp     r14,16
2252         je      NEAR $L$xts_dec_1
2253         pshufd  xmm13,xmm14,0x13
2254         pxor    xmm14,xmm14
2255         movdqa  xmm1,xmm6
2256         movdqa  XMMWORD[32+rsp],xmm6
2257         paddq   xmm6,xmm6
2258         pand    xmm13,xmm12
2259         pcmpgtd xmm14,xmm6
2260         pxor    xmm6,xmm13
2261         movdqu  xmm8,XMMWORD[16+r12]
2262         cmp     r14,32
2263         je      NEAR $L$xts_dec_2
2264         pxor    xmm15,xmm7
2265         pshufd  xmm13,xmm14,0x13
2266         pxor    xmm14,xmm14
2267         movdqa  xmm2,xmm6
2268         movdqa  XMMWORD[48+rsp],xmm6
2269         paddq   xmm6,xmm6
2270         pand    xmm13,xmm12
2271         pcmpgtd xmm14,xmm6
2272         pxor    xmm6,xmm13
2273         movdqu  xmm9,XMMWORD[32+r12]
2274         cmp     r14,48
2275         je      NEAR $L$xts_dec_3
2276         pxor    xmm0,xmm8
2277         pshufd  xmm13,xmm14,0x13
2278         pxor    xmm14,xmm14
2279         movdqa  xmm3,xmm6
2280         movdqa  XMMWORD[64+rsp],xmm6
2281         paddq   xmm6,xmm6
2282         pand    xmm13,xmm12
2283         pcmpgtd xmm14,xmm6
2284         pxor    xmm6,xmm13
2285         movdqu  xmm10,XMMWORD[48+r12]
2286         cmp     r14,64
2287         je      NEAR $L$xts_dec_4
2288         pxor    xmm1,xmm9
2289         pshufd  xmm13,xmm14,0x13
2290         pxor    xmm14,xmm14
2291         movdqa  xmm4,xmm6
2292         movdqa  XMMWORD[80+rsp],xmm6
2293         paddq   xmm6,xmm6
2294         pand    xmm13,xmm12
2295         pcmpgtd xmm14,xmm6
2296         pxor    xmm6,xmm13
2297         movdqu  xmm11,XMMWORD[64+r12]
2298         cmp     r14,80
2299         je      NEAR $L$xts_dec_5
2300         pxor    xmm2,xmm10
2301         pshufd  xmm13,xmm14,0x13
2302         pxor    xmm14,xmm14
2303         movdqa  xmm5,xmm6
2304         movdqa  XMMWORD[96+rsp],xmm6
2305         paddq   xmm6,xmm6
2306         pand    xmm13,xmm12
2307         pcmpgtd xmm14,xmm6
2308         pxor    xmm6,xmm13
2309         movdqu  xmm12,XMMWORD[80+r12]
2310         cmp     r14,96
2311         je      NEAR $L$xts_dec_6
2312         pxor    xmm3,xmm11
2313         movdqu  xmm13,XMMWORD[96+r12]
2314         pxor    xmm4,xmm12
2315         movdqa  XMMWORD[112+rsp],xmm6
2316         lea     r12,[112+r12]
2317         pxor    xmm5,xmm13
2318         lea     rax,[128+rsp]
2319         mov     r10d,edx
2320
2321         call    _bsaes_decrypt8
2322
2323         pxor    xmm15,XMMWORD[rsp]
2324         pxor    xmm0,XMMWORD[16+rsp]
2325         movdqu  XMMWORD[r13],xmm15
2326         pxor    xmm5,XMMWORD[32+rsp]
2327         movdqu  XMMWORD[16+r13],xmm0
2328         pxor    xmm3,XMMWORD[48+rsp]
2329         movdqu  XMMWORD[32+r13],xmm5
2330         pxor    xmm1,XMMWORD[64+rsp]
2331         movdqu  XMMWORD[48+r13],xmm3
2332         pxor    xmm6,XMMWORD[80+rsp]
2333         movdqu  XMMWORD[64+r13],xmm1
2334         pxor    xmm2,XMMWORD[96+rsp]
2335         movdqu  XMMWORD[80+r13],xmm6
2336         movdqu  XMMWORD[96+r13],xmm2
2337         lea     r13,[112+r13]
2338
2339         movdqa  xmm6,XMMWORD[112+rsp]
2340         jmp     NEAR $L$xts_dec_done
2341 ALIGN   16
2342 $L$xts_dec_6:
2343         pxor    xmm3,xmm11
2344         lea     r12,[96+r12]
2345         pxor    xmm4,xmm12
2346         lea     rax,[128+rsp]
2347         mov     r10d,edx
2348
2349         call    _bsaes_decrypt8
2350
2351         pxor    xmm15,XMMWORD[rsp]
2352         pxor    xmm0,XMMWORD[16+rsp]
2353         movdqu  XMMWORD[r13],xmm15
2354         pxor    xmm5,XMMWORD[32+rsp]
2355         movdqu  XMMWORD[16+r13],xmm0
2356         pxor    xmm3,XMMWORD[48+rsp]
2357         movdqu  XMMWORD[32+r13],xmm5
2358         pxor    xmm1,XMMWORD[64+rsp]
2359         movdqu  XMMWORD[48+r13],xmm3
2360         pxor    xmm6,XMMWORD[80+rsp]
2361         movdqu  XMMWORD[64+r13],xmm1
2362         movdqu  XMMWORD[80+r13],xmm6
2363         lea     r13,[96+r13]
2364
2365         movdqa  xmm6,XMMWORD[96+rsp]
2366         jmp     NEAR $L$xts_dec_done
2367 ALIGN   16
2368 $L$xts_dec_5:
2369         pxor    xmm2,xmm10
2370         lea     r12,[80+r12]
2371         pxor    xmm3,xmm11
2372         lea     rax,[128+rsp]
2373         mov     r10d,edx
2374
2375         call    _bsaes_decrypt8
2376
2377         pxor    xmm15,XMMWORD[rsp]
2378         pxor    xmm0,XMMWORD[16+rsp]
2379         movdqu  XMMWORD[r13],xmm15
2380         pxor    xmm5,XMMWORD[32+rsp]
2381         movdqu  XMMWORD[16+r13],xmm0
2382         pxor    xmm3,XMMWORD[48+rsp]
2383         movdqu  XMMWORD[32+r13],xmm5
2384         pxor    xmm1,XMMWORD[64+rsp]
2385         movdqu  XMMWORD[48+r13],xmm3
2386         movdqu  XMMWORD[64+r13],xmm1
2387         lea     r13,[80+r13]
2388
2389         movdqa  xmm6,XMMWORD[80+rsp]
2390         jmp     NEAR $L$xts_dec_done
2391 ALIGN   16
2392 $L$xts_dec_4:
2393         pxor    xmm1,xmm9
2394         lea     r12,[64+r12]
2395         pxor    xmm2,xmm10
2396         lea     rax,[128+rsp]
2397         mov     r10d,edx
2398
2399         call    _bsaes_decrypt8
2400
2401         pxor    xmm15,XMMWORD[rsp]
2402         pxor    xmm0,XMMWORD[16+rsp]
2403         movdqu  XMMWORD[r13],xmm15
2404         pxor    xmm5,XMMWORD[32+rsp]
2405         movdqu  XMMWORD[16+r13],xmm0
2406         pxor    xmm3,XMMWORD[48+rsp]
2407         movdqu  XMMWORD[32+r13],xmm5
2408         movdqu  XMMWORD[48+r13],xmm3
2409         lea     r13,[64+r13]
2410
2411         movdqa  xmm6,XMMWORD[64+rsp]
2412         jmp     NEAR $L$xts_dec_done
2413 ALIGN   16
2414 $L$xts_dec_3:
2415         pxor    xmm0,xmm8
2416         lea     r12,[48+r12]
2417         pxor    xmm1,xmm9
2418         lea     rax,[128+rsp]
2419         mov     r10d,edx
2420
2421         call    _bsaes_decrypt8
2422
2423         pxor    xmm15,XMMWORD[rsp]
2424         pxor    xmm0,XMMWORD[16+rsp]
2425         movdqu  XMMWORD[r13],xmm15
2426         pxor    xmm5,XMMWORD[32+rsp]
2427         movdqu  XMMWORD[16+r13],xmm0
2428         movdqu  XMMWORD[32+r13],xmm5
2429         lea     r13,[48+r13]
2430
2431         movdqa  xmm6,XMMWORD[48+rsp]
2432         jmp     NEAR $L$xts_dec_done
2433 ALIGN   16
2434 $L$xts_dec_2:
2435         pxor    xmm15,xmm7
2436         lea     r12,[32+r12]
2437         pxor    xmm0,xmm8
2438         lea     rax,[128+rsp]
2439         mov     r10d,edx
2440
2441         call    _bsaes_decrypt8
2442
2443         pxor    xmm15,XMMWORD[rsp]
2444         pxor    xmm0,XMMWORD[16+rsp]
2445         movdqu  XMMWORD[r13],xmm15
2446         movdqu  XMMWORD[16+r13],xmm0
2447         lea     r13,[32+r13]
2448
2449         movdqa  xmm6,XMMWORD[32+rsp]
2450         jmp     NEAR $L$xts_dec_done
2451 ALIGN   16
2452 $L$xts_dec_1:
2453         pxor    xmm7,xmm15
2454         lea     r12,[16+r12]
2455         movdqa  XMMWORD[32+rbp],xmm7
2456         lea     rcx,[32+rbp]
2457         lea     rdx,[32+rbp]
2458         lea     r8,[r15]
2459         call    asm_AES_decrypt
2460         pxor    xmm15,XMMWORD[32+rbp]
2461
2462
2463
2464
2465
2466         movdqu  XMMWORD[r13],xmm15
2467         lea     r13,[16+r13]
2468
2469         movdqa  xmm6,XMMWORD[16+rsp]
2470
2471 $L$xts_dec_done:
2472         and     ebx,15
2473         jz      NEAR $L$xts_dec_ret
2474
2475         pxor    xmm14,xmm14
2476         movdqa  xmm12,XMMWORD[$L$xts_magic]
2477         pcmpgtd xmm14,xmm6
2478         pshufd  xmm13,xmm14,0x13
2479         movdqa  xmm5,xmm6
2480         paddq   xmm6,xmm6
2481         pand    xmm13,xmm12
2482         movdqu  xmm15,XMMWORD[r12]
2483         pxor    xmm6,xmm13
2484
2485         lea     rcx,[32+rbp]
2486         pxor    xmm15,xmm6
2487         lea     rdx,[32+rbp]
2488         movdqa  XMMWORD[32+rbp],xmm15
2489         lea     r8,[r15]
2490         call    asm_AES_decrypt
2491         pxor    xmm6,XMMWORD[32+rbp]
2492         mov     rdx,r13
2493         movdqu  XMMWORD[r13],xmm6
2494
2495 $L$xts_dec_steal:
2496         movzx   eax,BYTE[16+r12]
2497         movzx   ecx,BYTE[rdx]
2498         lea     r12,[1+r12]
2499         mov     BYTE[rdx],al
2500         mov     BYTE[16+rdx],cl
2501         lea     rdx,[1+rdx]
2502         sub     ebx,1
2503         jnz     NEAR $L$xts_dec_steal
2504
2505         movdqu  xmm15,XMMWORD[r13]
2506         lea     rcx,[32+rbp]
2507         pxor    xmm15,xmm5
2508         lea     rdx,[32+rbp]
2509         movdqa  XMMWORD[32+rbp],xmm15
2510         lea     r8,[r15]
2511         call    asm_AES_decrypt
2512         pxor    xmm5,XMMWORD[32+rbp]
2513         movdqu  XMMWORD[r13],xmm5
2514
2515 $L$xts_dec_ret:
2516         lea     rax,[rsp]
2517         pxor    xmm0,xmm0
2518 $L$xts_dec_bzero:
2519         movdqa  XMMWORD[rax],xmm0
2520         movdqa  XMMWORD[16+rax],xmm0
2521         lea     rax,[32+rax]
2522         cmp     rbp,rax
2523         ja      NEAR $L$xts_dec_bzero
2524
2525         lea     rsp,[rbp]
2526         movaps  xmm6,XMMWORD[64+rbp]
2527         movaps  xmm7,XMMWORD[80+rbp]
2528         movaps  xmm8,XMMWORD[96+rbp]
2529         movaps  xmm9,XMMWORD[112+rbp]
2530         movaps  xmm10,XMMWORD[128+rbp]
2531         movaps  xmm11,XMMWORD[144+rbp]
2532         movaps  xmm12,XMMWORD[160+rbp]
2533         movaps  xmm13,XMMWORD[176+rbp]
2534         movaps  xmm14,XMMWORD[192+rbp]
2535         movaps  xmm15,XMMWORD[208+rbp]
2536         lea     rsp,[160+rbp]
2537         mov     r15,QWORD[72+rsp]
2538         mov     r14,QWORD[80+rsp]
2539         mov     r13,QWORD[88+rsp]
2540         mov     r12,QWORD[96+rsp]
2541         mov     rbx,QWORD[104+rsp]
2542         mov     rax,QWORD[112+rsp]
2543         lea     rsp,[120+rsp]
2544         mov     rbp,rax
2545 $L$xts_dec_epilogue:
2546         DB      0F3h,0C3h               ;repret
2547
2548
2549 ALIGN   64
2550 _bsaes_const:
2551 $L$M0ISR:
2552         DQ      0x0a0e0206070b0f03,0x0004080c0d010509
2553 $L$ISRM0:
2554         DQ      0x01040b0e0205080f,0x0306090c00070a0d
2555 $L$ISR:
2556         DQ      0x0504070602010003,0x0f0e0d0c080b0a09
2557 $L$BS0:
2558         DQ      0x5555555555555555,0x5555555555555555
2559 $L$BS1:
2560         DQ      0x3333333333333333,0x3333333333333333
2561 $L$BS2:
2562         DQ      0x0f0f0f0f0f0f0f0f,0x0f0f0f0f0f0f0f0f
2563 $L$SR:
2564         DQ      0x0504070600030201,0x0f0e0d0c0a09080b
2565 $L$SRM0:
2566         DQ      0x0304090e00050a0f,0x01060b0c0207080d
2567 $L$M0SR:
2568         DQ      0x0a0e02060f03070b,0x0004080c05090d01
2569 $L$SWPUP:
2570         DQ      0x0706050403020100,0x0c0d0e0f0b0a0908
2571 $L$SWPUPM0SR:
2572         DQ      0x0a0d02060c03070b,0x0004080f05090e01
2573 $L$ADD1:
2574         DQ      0x0000000000000000,0x0000000100000000
2575 $L$ADD2:
2576         DQ      0x0000000000000000,0x0000000200000000
2577 $L$ADD3:
2578         DQ      0x0000000000000000,0x0000000300000000
2579 $L$ADD4:
2580         DQ      0x0000000000000000,0x0000000400000000
2581 $L$ADD5:
2582         DQ      0x0000000000000000,0x0000000500000000
2583 $L$ADD6:
2584         DQ      0x0000000000000000,0x0000000600000000
2585 $L$ADD7:
2586         DQ      0x0000000000000000,0x0000000700000000
2587 $L$ADD8:
2588         DQ      0x0000000000000000,0x0000000800000000
2589 $L$xts_magic:
2590         DD      0x87,0,1,0
2591 $L$masks:
2592         DQ      0x0101010101010101,0x0101010101010101
2593         DQ      0x0202020202020202,0x0202020202020202
2594         DQ      0x0404040404040404,0x0404040404040404
2595         DQ      0x0808080808080808,0x0808080808080808
2596 $L$M0:
2597         DQ      0x02060a0e03070b0f,0x0004080c0105090d
2598 $L$63:
2599         DQ      0x6363636363636363,0x6363636363636363
2600 DB      66,105,116,45,115,108,105,99,101,100,32,65,69,83,32,102
2601 DB      111,114,32,120,56,54,95,54,52,47,83,83,83,69,51,44
2602 DB      32,69,109,105,108,105,97,32,75,195,164,115,112,101,114,44
2603 DB      32,80,101,116,101,114,32,83,99,104,119,97,98,101,44,32
2604 DB      65,110,100,121,32,80,111,108,121,97,107,111,118,0
2605 ALIGN   64
2606
2607 EXTERN  __imp_RtlVirtualUnwind
2608
2609 ALIGN   16
2610 se_handler:
2611         push    rsi
2612         push    rdi
2613         push    rbx
2614         push    rbp
2615         push    r12
2616         push    r13
2617         push    r14
2618         push    r15
2619         pushfq
2620         sub     rsp,64
2621
2622         mov     rax,QWORD[120+r8]
2623         mov     rbx,QWORD[248+r8]
2624
2625         mov     rsi,QWORD[8+r9]
2626         mov     r11,QWORD[56+r9]
2627
2628         mov     r10d,DWORD[r11]
2629         lea     r10,[r10*1+rsi]
2630         cmp     rbx,r10
2631         jb      NEAR $L$in_prologue
2632
2633         mov     rax,QWORD[152+r8]
2634
2635         mov     r10d,DWORD[4+r11]
2636         lea     r10,[r10*1+rsi]
2637         cmp     rbx,r10
2638         jae     NEAR $L$in_prologue
2639
2640         mov     rax,QWORD[160+r8]
2641
2642         lea     rsi,[64+rax]
2643         lea     rdi,[512+r8]
2644         mov     ecx,20
2645         DD      0xa548f3fc
2646         lea     rax,[160+rax]
2647
2648         mov     rbp,QWORD[112+rax]
2649         mov     rbx,QWORD[104+rax]
2650         mov     r12,QWORD[96+rax]
2651         mov     r13,QWORD[88+rax]
2652         mov     r14,QWORD[80+rax]
2653         mov     r15,QWORD[72+rax]
2654         lea     rax,[120+rax]
2655         mov     QWORD[144+r8],rbx
2656         mov     QWORD[160+r8],rbp
2657         mov     QWORD[216+r8],r12
2658         mov     QWORD[224+r8],r13
2659         mov     QWORD[232+r8],r14
2660         mov     QWORD[240+r8],r15
2661
2662 $L$in_prologue:
2663         mov     QWORD[152+r8],rax
2664
2665         mov     rdi,QWORD[40+r9]
2666         mov     rsi,r8
2667         mov     ecx,154
2668         DD      0xa548f3fc
2669
2670         mov     rsi,r9
2671         xor     rcx,rcx
2672         mov     rdx,QWORD[8+rsi]
2673         mov     r8,QWORD[rsi]
2674         mov     r9,QWORD[16+rsi]
2675         mov     r10,QWORD[40+rsi]
2676         lea     r11,[56+rsi]
2677         lea     r12,[24+rsi]
2678         mov     QWORD[32+rsp],r10
2679         mov     QWORD[40+rsp],r11
2680         mov     QWORD[48+rsp],r12
2681         mov     QWORD[56+rsp],rcx
2682         call    QWORD[__imp_RtlVirtualUnwind]
2683
2684         mov     eax,1
2685         add     rsp,64
2686         popfq
2687         pop     r15
2688         pop     r14
2689         pop     r13
2690         pop     r12
2691         pop     rbp
2692         pop     rbx
2693         pop     rdi
2694         pop     rsi
2695         DB      0F3h,0C3h               ;repret
2696
2697
2698 section .pdata rdata align=4
2699 ALIGN   4
2700         DD      $L$cbc_dec_prologue wrt ..imagebase
2701         DD      $L$cbc_dec_epilogue wrt ..imagebase
2702         DD      $L$cbc_dec_info wrt ..imagebase
2703
2704         DD      $L$ctr_enc_prologue wrt ..imagebase
2705         DD      $L$ctr_enc_epilogue wrt ..imagebase
2706         DD      $L$ctr_enc_info wrt ..imagebase
2707
2708         DD      $L$xts_enc_prologue wrt ..imagebase
2709         DD      $L$xts_enc_epilogue wrt ..imagebase
2710         DD      $L$xts_enc_info wrt ..imagebase
2711
2712         DD      $L$xts_dec_prologue wrt ..imagebase
2713         DD      $L$xts_dec_epilogue wrt ..imagebase
2714         DD      $L$xts_dec_info wrt ..imagebase
2715
2716 section .xdata rdata align=8
2717 ALIGN   8
2718 $L$cbc_dec_info:
2719 DB      9,0,0,0
2720         DD      se_handler wrt ..imagebase
2721         DD      $L$cbc_dec_body wrt ..imagebase,$L$cbc_dec_epilogue wrt ..imagebase
2722 $L$ctr_enc_info:
2723 DB      9,0,0,0
2724         DD      se_handler wrt ..imagebase
2725         DD      $L$ctr_enc_body wrt ..imagebase,$L$ctr_enc_epilogue wrt ..imagebase
2726 $L$xts_enc_info:
2727 DB      9,0,0,0
2728         DD      se_handler wrt ..imagebase
2729         DD      $L$xts_enc_body wrt ..imagebase,$L$xts_enc_epilogue wrt ..imagebase
2730 $L$xts_dec_info:
2731 DB      9,0,0,0
2732         DD      se_handler wrt ..imagebase
2733         DD      $L$xts_dec_body wrt ..imagebase,$L$xts_dec_epilogue wrt ..imagebase