OLD | NEW |
(Empty) | |
| 1 default rel |
| 2 %define XMMWORD |
| 3 %define YMMWORD |
| 4 %define ZMMWORD |
| 5 section .data data align=8 |
| 6 |
| 7 |
| 8 ALIGN 16 |
| 9 one: |
| 10 DQ 1,0 |
| 11 two: |
| 12 DQ 2,0 |
| 13 three: |
| 14 DQ 3,0 |
| 15 four: |
| 16 DQ 4,0 |
| 17 five: |
| 18 DQ 5,0 |
| 19 six: |
| 20 DQ 6,0 |
| 21 seven: |
| 22 DQ 7,0 |
| 23 eight: |
| 24 DQ 8,0 |
| 25 |
| 26 OR_MASK: |
| 27 DD 0x00000000,0x00000000,0x00000000,0x80000000 |
| 28 poly: |
| 29 DQ 0x1,0xc200000000000000 |
| 30 mask: |
| 31 DD 0x0c0f0e0d,0x0c0f0e0d,0x0c0f0e0d,0x0c0f0e0d |
| 32 con1: |
| 33 DD 1,1,1,1 |
| 34 con2: |
| 35 DD 0x1b,0x1b,0x1b,0x1b |
| 36 con3: |
| 37 DB -1,-1,-1,-1,-1,-1,-1,-1,4,5,6,7,4,5,6,7 |
| 38 and_mask: |
| 39 DD 0,0xffffffff,0xffffffff,0xffffffff |
| 40 section .text code align=64 |
| 41 |
| 42 |
| 43 ALIGN 16 |
| 44 GFMUL: |
| 45 |
| 46 vpclmulqdq xmm2,xmm0,xmm1,0x00 |
| 47 vpclmulqdq xmm5,xmm0,xmm1,0x11 |
| 48 vpclmulqdq xmm3,xmm0,xmm1,0x10 |
| 49 vpclmulqdq xmm4,xmm0,xmm1,0x01 |
| 50 vpxor xmm3,xmm3,xmm4 |
| 51 vpslldq xmm4,xmm3,8 |
| 52 vpsrldq xmm3,xmm3,8 |
| 53 vpxor xmm2,xmm2,xmm4 |
| 54 vpxor xmm5,xmm5,xmm3 |
| 55 |
| 56 vpclmulqdq xmm3,xmm2,XMMWORD[poly],0x10 |
| 57 vpshufd xmm4,xmm2,78 |
| 58 vpxor xmm2,xmm3,xmm4 |
| 59 |
| 60 vpclmulqdq xmm3,xmm2,XMMWORD[poly],0x10 |
| 61 vpshufd xmm4,xmm2,78 |
| 62 vpxor xmm2,xmm3,xmm4 |
| 63 |
| 64 vpxor xmm0,xmm2,xmm5 |
| 65 DB 0F3h,0C3h ;repret |
| 66 |
| 67 |
| 68 global aesgcmsiv_htable_init |
| 69 |
| 70 ALIGN 16 |
| 71 aesgcmsiv_htable_init: |
| 72 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 73 mov QWORD[16+rsp],rsi |
| 74 mov rax,rsp |
| 75 $L$SEH_begin_aesgcmsiv_htable_init: |
| 76 mov rdi,rcx |
| 77 mov rsi,rdx |
| 78 |
| 79 |
| 80 |
| 81 vmovdqa xmm0,XMMWORD[rsi] |
| 82 vmovdqa xmm1,xmm0 |
| 83 vmovdqa XMMWORD[rdi],xmm0 |
| 84 call GFMUL |
| 85 vmovdqa XMMWORD[16+rdi],xmm0 |
| 86 call GFMUL |
| 87 vmovdqa XMMWORD[32+rdi],xmm0 |
| 88 call GFMUL |
| 89 vmovdqa XMMWORD[48+rdi],xmm0 |
| 90 call GFMUL |
| 91 vmovdqa XMMWORD[64+rdi],xmm0 |
| 92 call GFMUL |
| 93 vmovdqa XMMWORD[80+rdi],xmm0 |
| 94 call GFMUL |
| 95 vmovdqa XMMWORD[96+rdi],xmm0 |
| 96 call GFMUL |
| 97 vmovdqa XMMWORD[112+rdi],xmm0 |
| 98 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 99 mov rsi,QWORD[16+rsp] |
| 100 DB 0F3h,0C3h ;repret |
| 101 |
| 102 $L$SEH_end_aesgcmsiv_htable_init: |
| 103 global aesgcmsiv_htable6_init |
| 104 |
| 105 ALIGN 16 |
| 106 aesgcmsiv_htable6_init: |
| 107 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 108 mov QWORD[16+rsp],rsi |
| 109 mov rax,rsp |
| 110 $L$SEH_begin_aesgcmsiv_htable6_init: |
| 111 mov rdi,rcx |
| 112 mov rsi,rdx |
| 113 |
| 114 |
| 115 |
| 116 vmovdqa xmm0,XMMWORD[rsi] |
| 117 vmovdqa xmm1,xmm0 |
| 118 vmovdqa XMMWORD[rdi],xmm0 |
| 119 call GFMUL |
| 120 vmovdqa XMMWORD[16+rdi],xmm0 |
| 121 call GFMUL |
| 122 vmovdqa XMMWORD[32+rdi],xmm0 |
| 123 call GFMUL |
| 124 vmovdqa XMMWORD[48+rdi],xmm0 |
| 125 call GFMUL |
| 126 vmovdqa XMMWORD[64+rdi],xmm0 |
| 127 call GFMUL |
| 128 vmovdqa XMMWORD[80+rdi],xmm0 |
| 129 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 130 mov rsi,QWORD[16+rsp] |
| 131 DB 0F3h,0C3h ;repret |
| 132 |
| 133 $L$SEH_end_aesgcmsiv_htable6_init: |
| 134 global aesgcmsiv_htable_polyval |
| 135 |
| 136 ALIGN 16 |
| 137 aesgcmsiv_htable_polyval: |
| 138 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 139 mov QWORD[16+rsp],rsi |
| 140 mov rax,rsp |
| 141 $L$SEH_begin_aesgcmsiv_htable_polyval: |
| 142 mov rdi,rcx |
| 143 mov rsi,rdx |
| 144 mov rdx,r8 |
| 145 mov rcx,r9 |
| 146 |
| 147 |
| 148 |
| 149 test rdx,rdx |
| 150 jnz NEAR $L$htable_polyval_start |
| 151 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 152 mov rsi,QWORD[16+rsp] |
| 153 DB 0F3h,0C3h ;repret |
| 154 |
| 155 $L$htable_polyval_start: |
| 156 vzeroall |
| 157 |
| 158 |
| 159 |
| 160 mov r11,rdx |
| 161 and r11,127 |
| 162 |
| 163 jz NEAR $L$htable_polyval_no_prefix |
| 164 |
| 165 vpxor xmm9,xmm9,xmm9 |
| 166 vmovdqa xmm1,XMMWORD[rcx] |
| 167 sub rdx,r11 |
| 168 |
| 169 sub r11,16 |
| 170 |
| 171 |
| 172 vmovdqu xmm0,XMMWORD[rsi] |
| 173 vpxor xmm0,xmm0,xmm1 |
| 174 |
| 175 vpclmulqdq xmm5,xmm0,XMMWORD[r11*1+rdi],0x01 |
| 176 vpclmulqdq xmm3,xmm0,XMMWORD[r11*1+rdi],0x00 |
| 177 vpclmulqdq xmm4,xmm0,XMMWORD[r11*1+rdi],0x11 |
| 178 vpclmulqdq xmm6,xmm0,XMMWORD[r11*1+rdi],0x10 |
| 179 vpxor xmm5,xmm5,xmm6 |
| 180 |
| 181 lea rsi,[16+rsi] |
| 182 test r11,r11 |
| 183 jnz NEAR $L$htable_polyval_prefix_loop |
| 184 jmp NEAR $L$htable_polyval_prefix_complete |
| 185 |
| 186 |
| 187 ALIGN 64 |
| 188 $L$htable_polyval_prefix_loop: |
| 189 sub r11,16 |
| 190 |
| 191 vmovdqu xmm0,XMMWORD[rsi] |
| 192 |
| 193 vpclmulqdq xmm6,xmm0,XMMWORD[r11*1+rdi],0x00 |
| 194 vpxor xmm3,xmm3,xmm6 |
| 195 vpclmulqdq xmm6,xmm0,XMMWORD[r11*1+rdi],0x11 |
| 196 vpxor xmm4,xmm4,xmm6 |
| 197 vpclmulqdq xmm6,xmm0,XMMWORD[r11*1+rdi],0x01 |
| 198 vpxor xmm5,xmm5,xmm6 |
| 199 vpclmulqdq xmm6,xmm0,XMMWORD[r11*1+rdi],0x10 |
| 200 vpxor xmm5,xmm5,xmm6 |
| 201 |
| 202 test r11,r11 |
| 203 |
| 204 lea rsi,[16+rsi] |
| 205 |
| 206 jnz NEAR $L$htable_polyval_prefix_loop |
| 207 |
| 208 $L$htable_polyval_prefix_complete: |
| 209 vpsrldq xmm6,xmm5,8 |
| 210 vpslldq xmm5,xmm5,8 |
| 211 |
| 212 vpxor xmm9,xmm4,xmm6 |
| 213 vpxor xmm1,xmm3,xmm5 |
| 214 |
| 215 jmp NEAR $L$htable_polyval_main_loop |
| 216 |
| 217 $L$htable_polyval_no_prefix: |
| 218 |
| 219 |
| 220 |
| 221 |
| 222 vpxor xmm1,xmm1,xmm1 |
| 223 vmovdqa xmm9,XMMWORD[rcx] |
| 224 |
| 225 ALIGN 64 |
| 226 $L$htable_polyval_main_loop: |
| 227 sub rdx,0x80 |
| 228 jb NEAR $L$htable_polyval_out |
| 229 |
| 230 vmovdqu xmm0,XMMWORD[112+rsi] |
| 231 |
| 232 vpclmulqdq xmm5,xmm0,XMMWORD[rdi],0x01 |
| 233 vpclmulqdq xmm3,xmm0,XMMWORD[rdi],0x00 |
| 234 vpclmulqdq xmm4,xmm0,XMMWORD[rdi],0x11 |
| 235 vpclmulqdq xmm6,xmm0,XMMWORD[rdi],0x10 |
| 236 vpxor xmm5,xmm5,xmm6 |
| 237 |
| 238 |
| 239 vmovdqu xmm0,XMMWORD[96+rsi] |
| 240 vpclmulqdq xmm6,xmm0,XMMWORD[16+rdi],0x01 |
| 241 vpxor xmm5,xmm5,xmm6 |
| 242 vpclmulqdq xmm6,xmm0,XMMWORD[16+rdi],0x00 |
| 243 vpxor xmm3,xmm3,xmm6 |
| 244 vpclmulqdq xmm6,xmm0,XMMWORD[16+rdi],0x11 |
| 245 vpxor xmm4,xmm4,xmm6 |
| 246 vpclmulqdq xmm6,xmm0,XMMWORD[16+rdi],0x10 |
| 247 vpxor xmm5,xmm5,xmm6 |
| 248 |
| 249 |
| 250 |
| 251 vmovdqu xmm0,XMMWORD[80+rsi] |
| 252 |
| 253 vpclmulqdq xmm7,xmm1,XMMWORD[poly],0x10 |
| 254 vpalignr xmm1,xmm1,xmm1,8 |
| 255 |
| 256 vpclmulqdq xmm6,xmm0,XMMWORD[32+rdi],0x01 |
| 257 vpxor xmm5,xmm5,xmm6 |
| 258 vpclmulqdq xmm6,xmm0,XMMWORD[32+rdi],0x00 |
| 259 vpxor xmm3,xmm3,xmm6 |
| 260 vpclmulqdq xmm6,xmm0,XMMWORD[32+rdi],0x11 |
| 261 vpxor xmm4,xmm4,xmm6 |
| 262 vpclmulqdq xmm6,xmm0,XMMWORD[32+rdi],0x10 |
| 263 vpxor xmm5,xmm5,xmm6 |
| 264 |
| 265 |
| 266 vpxor xmm1,xmm1,xmm7 |
| 267 |
| 268 vmovdqu xmm0,XMMWORD[64+rsi] |
| 269 |
| 270 vpclmulqdq xmm6,xmm0,XMMWORD[48+rdi],0x01 |
| 271 vpxor xmm5,xmm5,xmm6 |
| 272 vpclmulqdq xmm6,xmm0,XMMWORD[48+rdi],0x00 |
| 273 vpxor xmm3,xmm3,xmm6 |
| 274 vpclmulqdq xmm6,xmm0,XMMWORD[48+rdi],0x11 |
| 275 vpxor xmm4,xmm4,xmm6 |
| 276 vpclmulqdq xmm6,xmm0,XMMWORD[48+rdi],0x10 |
| 277 vpxor xmm5,xmm5,xmm6 |
| 278 |
| 279 |
| 280 vmovdqu xmm0,XMMWORD[48+rsi] |
| 281 |
| 282 vpclmulqdq xmm7,xmm1,XMMWORD[poly],0x10 |
| 283 vpalignr xmm1,xmm1,xmm1,8 |
| 284 |
| 285 vpclmulqdq xmm6,xmm0,XMMWORD[64+rdi],0x01 |
| 286 vpxor xmm5,xmm5,xmm6 |
| 287 vpclmulqdq xmm6,xmm0,XMMWORD[64+rdi],0x00 |
| 288 vpxor xmm3,xmm3,xmm6 |
| 289 vpclmulqdq xmm6,xmm0,XMMWORD[64+rdi],0x11 |
| 290 vpxor xmm4,xmm4,xmm6 |
| 291 vpclmulqdq xmm6,xmm0,XMMWORD[64+rdi],0x10 |
| 292 vpxor xmm5,xmm5,xmm6 |
| 293 |
| 294 |
| 295 vpxor xmm1,xmm1,xmm7 |
| 296 |
| 297 vmovdqu xmm0,XMMWORD[32+rsi] |
| 298 |
| 299 vpclmulqdq xmm6,xmm0,XMMWORD[80+rdi],0x01 |
| 300 vpxor xmm5,xmm5,xmm6 |
| 301 vpclmulqdq xmm6,xmm0,XMMWORD[80+rdi],0x00 |
| 302 vpxor xmm3,xmm3,xmm6 |
| 303 vpclmulqdq xmm6,xmm0,XMMWORD[80+rdi],0x11 |
| 304 vpxor xmm4,xmm4,xmm6 |
| 305 vpclmulqdq xmm6,xmm0,XMMWORD[80+rdi],0x10 |
| 306 vpxor xmm5,xmm5,xmm6 |
| 307 |
| 308 |
| 309 vpxor xmm1,xmm1,xmm9 |
| 310 |
| 311 vmovdqu xmm0,XMMWORD[16+rsi] |
| 312 |
| 313 vpclmulqdq xmm6,xmm0,XMMWORD[96+rdi],0x01 |
| 314 vpxor xmm5,xmm5,xmm6 |
| 315 vpclmulqdq xmm6,xmm0,XMMWORD[96+rdi],0x00 |
| 316 vpxor xmm3,xmm3,xmm6 |
| 317 vpclmulqdq xmm6,xmm0,XMMWORD[96+rdi],0x11 |
| 318 vpxor xmm4,xmm4,xmm6 |
| 319 vpclmulqdq xmm6,xmm0,XMMWORD[96+rdi],0x10 |
| 320 vpxor xmm5,xmm5,xmm6 |
| 321 |
| 322 |
| 323 vmovdqu xmm0,XMMWORD[rsi] |
| 324 vpxor xmm0,xmm0,xmm1 |
| 325 |
| 326 vpclmulqdq xmm6,xmm0,XMMWORD[112+rdi],0x01 |
| 327 vpxor xmm5,xmm5,xmm6 |
| 328 vpclmulqdq xmm6,xmm0,XMMWORD[112+rdi],0x00 |
| 329 vpxor xmm3,xmm3,xmm6 |
| 330 vpclmulqdq xmm6,xmm0,XMMWORD[112+rdi],0x11 |
| 331 vpxor xmm4,xmm4,xmm6 |
| 332 vpclmulqdq xmm6,xmm0,XMMWORD[112+rdi],0x10 |
| 333 vpxor xmm5,xmm5,xmm6 |
| 334 |
| 335 |
| 336 vpsrldq xmm6,xmm5,8 |
| 337 vpslldq xmm5,xmm5,8 |
| 338 |
| 339 vpxor xmm9,xmm4,xmm6 |
| 340 vpxor xmm1,xmm3,xmm5 |
| 341 |
| 342 lea rsi,[128+rsi] |
| 343 jmp NEAR $L$htable_polyval_main_loop |
| 344 |
| 345 |
| 346 |
| 347 $L$htable_polyval_out: |
| 348 vpclmulqdq xmm6,xmm1,XMMWORD[poly],0x10 |
| 349 vpalignr xmm1,xmm1,xmm1,8 |
| 350 vpxor xmm1,xmm1,xmm6 |
| 351 |
| 352 vpclmulqdq xmm6,xmm1,XMMWORD[poly],0x10 |
| 353 vpalignr xmm1,xmm1,xmm1,8 |
| 354 vpxor xmm1,xmm1,xmm6 |
| 355 vpxor xmm1,xmm1,xmm9 |
| 356 |
| 357 vmovdqu XMMWORD[rcx],xmm1 |
| 358 vzeroupper |
| 359 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 360 mov rsi,QWORD[16+rsp] |
| 361 DB 0F3h,0C3h ;repret |
| 362 |
| 363 $L$SEH_end_aesgcmsiv_htable_polyval: |
| 364 global aesgcmsiv_polyval_horner |
| 365 |
| 366 ALIGN 16 |
| 367 aesgcmsiv_polyval_horner: |
| 368 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 369 mov QWORD[16+rsp],rsi |
| 370 mov rax,rsp |
| 371 $L$SEH_begin_aesgcmsiv_polyval_horner: |
| 372 mov rdi,rcx |
| 373 mov rsi,rdx |
| 374 mov rdx,r8 |
| 375 mov rcx,r9 |
| 376 |
| 377 |
| 378 |
| 379 test rcx,rcx |
| 380 jnz NEAR $L$polyval_horner_start |
| 381 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 382 mov rsi,QWORD[16+rsp] |
| 383 DB 0F3h,0C3h ;repret |
| 384 |
| 385 $L$polyval_horner_start: |
| 386 |
| 387 |
| 388 |
| 389 xor r10,r10 |
| 390 shl rcx,4 |
| 391 |
| 392 vmovdqa xmm1,XMMWORD[rsi] |
| 393 vmovdqa xmm0,XMMWORD[rdi] |
| 394 |
| 395 $L$polyval_horner_loop: |
| 396 vpxor xmm0,xmm0,XMMWORD[r10*1+rdx] |
| 397 call GFMUL |
| 398 |
| 399 add r10,16 |
| 400 cmp rcx,r10 |
| 401 jne NEAR $L$polyval_horner_loop |
| 402 |
| 403 |
| 404 vmovdqa XMMWORD[rdi],xmm0 |
| 405 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 406 mov rsi,QWORD[16+rsp] |
| 407 DB 0F3h,0C3h ;repret |
| 408 |
| 409 $L$SEH_end_aesgcmsiv_polyval_horner: |
| 410 global aes128gcmsiv_aes_ks |
| 411 |
| 412 ALIGN 16 |
| 413 aes128gcmsiv_aes_ks: |
| 414 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 415 mov QWORD[16+rsp],rsi |
| 416 mov rax,rsp |
| 417 $L$SEH_begin_aes128gcmsiv_aes_ks: |
| 418 mov rdi,rcx |
| 419 mov rsi,rdx |
| 420 |
| 421 |
| 422 |
| 423 vmovdqa xmm1,XMMWORD[rdi] |
| 424 vmovdqa XMMWORD[rsi],xmm1 |
| 425 |
| 426 vmovdqa xmm0,XMMWORD[con1] |
| 427 vmovdqa xmm15,XMMWORD[mask] |
| 428 |
| 429 mov rax,8 |
| 430 |
| 431 $L$ks128_loop: |
| 432 add rsi,16 |
| 433 sub rax,1 |
| 434 vpshufb xmm2,xmm1,xmm15 |
| 435 vaesenclast xmm2,xmm2,xmm0 |
| 436 vpslld xmm0,xmm0,1 |
| 437 vpslldq xmm3,xmm1,4 |
| 438 vpxor xmm1,xmm1,xmm3 |
| 439 vpslldq xmm3,xmm3,4 |
| 440 vpxor xmm1,xmm1,xmm3 |
| 441 vpslldq xmm3,xmm3,4 |
| 442 vpxor xmm1,xmm1,xmm3 |
| 443 vpxor xmm1,xmm1,xmm2 |
| 444 vmovdqa XMMWORD[rsi],xmm1 |
| 445 jne NEAR $L$ks128_loop |
| 446 |
| 447 vmovdqa xmm0,XMMWORD[con2] |
| 448 vpshufb xmm2,xmm1,xmm15 |
| 449 vaesenclast xmm2,xmm2,xmm0 |
| 450 vpslld xmm0,xmm0,1 |
| 451 vpslldq xmm3,xmm1,4 |
| 452 vpxor xmm1,xmm1,xmm3 |
| 453 vpslldq xmm3,xmm3,4 |
| 454 vpxor xmm1,xmm1,xmm3 |
| 455 vpslldq xmm3,xmm3,4 |
| 456 vpxor xmm1,xmm1,xmm3 |
| 457 vpxor xmm1,xmm1,xmm2 |
| 458 vmovdqa XMMWORD[16+rsi],xmm1 |
| 459 |
| 460 vpshufb xmm2,xmm1,xmm15 |
| 461 vaesenclast xmm2,xmm2,xmm0 |
| 462 vpslldq xmm3,xmm1,4 |
| 463 vpxor xmm1,xmm1,xmm3 |
| 464 vpslldq xmm3,xmm3,4 |
| 465 vpxor xmm1,xmm1,xmm3 |
| 466 vpslldq xmm3,xmm3,4 |
| 467 vpxor xmm1,xmm1,xmm3 |
| 468 vpxor xmm1,xmm1,xmm2 |
| 469 vmovdqa XMMWORD[32+rsi],xmm1 |
| 470 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 471 mov rsi,QWORD[16+rsp] |
| 472 DB 0F3h,0C3h ;repret |
| 473 |
| 474 $L$SEH_end_aes128gcmsiv_aes_ks: |
| 475 global aes256gcmsiv_aes_ks |
| 476 |
| 477 ALIGN 16 |
| 478 aes256gcmsiv_aes_ks: |
| 479 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 480 mov QWORD[16+rsp],rsi |
| 481 mov rax,rsp |
| 482 $L$SEH_begin_aes256gcmsiv_aes_ks: |
| 483 mov rdi,rcx |
| 484 mov rsi,rdx |
| 485 |
| 486 |
| 487 |
| 488 vmovdqa xmm1,XMMWORD[rdi] |
| 489 vmovdqa xmm3,XMMWORD[16+rdi] |
| 490 vmovdqa XMMWORD[rsi],xmm1 |
| 491 vmovdqa XMMWORD[16+rsi],xmm3 |
| 492 vmovdqa xmm0,XMMWORD[con1] |
| 493 vmovdqa xmm15,XMMWORD[mask] |
| 494 vpxor xmm14,xmm14,xmm14 |
| 495 mov rax,6 |
| 496 |
| 497 $L$ks256_loop: |
| 498 add rsi,32 |
| 499 sub rax,1 |
| 500 vpshufb xmm2,xmm3,xmm15 |
| 501 vaesenclast xmm2,xmm2,xmm0 |
| 502 vpslld xmm0,xmm0,1 |
| 503 vpsllq xmm4,xmm1,32 |
| 504 vpxor xmm1,xmm1,xmm4 |
| 505 vpshufb xmm4,xmm1,XMMWORD[con3] |
| 506 vpxor xmm1,xmm1,xmm4 |
| 507 vpxor xmm1,xmm1,xmm2 |
| 508 vmovdqa XMMWORD[rsi],xmm1 |
| 509 vpshufd xmm2,xmm1,0xff |
| 510 vaesenclast xmm2,xmm2,xmm14 |
| 511 vpsllq xmm4,xmm3,32 |
| 512 vpxor xmm3,xmm3,xmm4 |
| 513 vpshufb xmm4,xmm3,XMMWORD[con3] |
| 514 vpxor xmm3,xmm3,xmm4 |
| 515 vpxor xmm3,xmm3,xmm2 |
| 516 vmovdqa XMMWORD[16+rsi],xmm3 |
| 517 jne NEAR $L$ks256_loop |
| 518 |
| 519 vpshufb xmm2,xmm3,xmm15 |
| 520 vaesenclast xmm2,xmm2,xmm0 |
| 521 vpsllq xmm4,xmm1,32 |
| 522 vpxor xmm1,xmm1,xmm4 |
| 523 vpshufb xmm4,xmm1,XMMWORD[con3] |
| 524 vpxor xmm1,xmm1,xmm4 |
| 525 vpxor xmm1,xmm1,xmm2 |
| 526 vmovdqa XMMWORD[32+rsi],xmm1 |
| 527 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 528 mov rsi,QWORD[16+rsp] |
| 529 DB 0F3h,0C3h ;repret |
| 530 |
| 531 global aes128gcmsiv_aes_ks_enc_x1 |
| 532 |
| 533 ALIGN 16 |
| 534 aes128gcmsiv_aes_ks_enc_x1: |
| 535 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 536 mov QWORD[16+rsp],rsi |
| 537 mov rax,rsp |
| 538 $L$SEH_begin_aes128gcmsiv_aes_ks_enc_x1: |
| 539 mov rdi,rcx |
| 540 mov rsi,rdx |
| 541 mov rdx,r8 |
| 542 mov rcx,r9 |
| 543 |
| 544 |
| 545 |
| 546 vmovdqa xmm1,XMMWORD[rcx] |
| 547 vmovdqa xmm4,XMMWORD[rdi] |
| 548 |
| 549 vmovdqa XMMWORD[rdx],xmm1 |
| 550 vpxor xmm4,xmm4,xmm1 |
| 551 |
| 552 vmovdqa xmm0,XMMWORD[con1] |
| 553 vmovdqa xmm15,XMMWORD[mask] |
| 554 |
| 555 vpshufb xmm2,xmm1,xmm15 |
| 556 vaesenclast xmm2,xmm2,xmm0 |
| 557 vpslld xmm0,xmm0,1 |
| 558 vpsllq xmm3,xmm1,32 |
| 559 vpxor xmm1,xmm1,xmm3 |
| 560 vpshufb xmm3,xmm1,XMMWORD[con3] |
| 561 vpxor xmm1,xmm1,xmm3 |
| 562 vpxor xmm1,xmm1,xmm2 |
| 563 |
| 564 vaesenc xmm4,xmm4,xmm1 |
| 565 vmovdqa XMMWORD[16+rdx],xmm1 |
| 566 |
| 567 vpshufb xmm2,xmm1,xmm15 |
| 568 vaesenclast xmm2,xmm2,xmm0 |
| 569 vpslld xmm0,xmm0,1 |
| 570 vpsllq xmm3,xmm1,32 |
| 571 vpxor xmm1,xmm1,xmm3 |
| 572 vpshufb xmm3,xmm1,XMMWORD[con3] |
| 573 vpxor xmm1,xmm1,xmm3 |
| 574 vpxor xmm1,xmm1,xmm2 |
| 575 |
| 576 vaesenc xmm4,xmm4,xmm1 |
| 577 vmovdqa XMMWORD[32+rdx],xmm1 |
| 578 |
| 579 vpshufb xmm2,xmm1,xmm15 |
| 580 vaesenclast xmm2,xmm2,xmm0 |
| 581 vpslld xmm0,xmm0,1 |
| 582 vpsllq xmm3,xmm1,32 |
| 583 vpxor xmm1,xmm1,xmm3 |
| 584 vpshufb xmm3,xmm1,XMMWORD[con3] |
| 585 vpxor xmm1,xmm1,xmm3 |
| 586 vpxor xmm1,xmm1,xmm2 |
| 587 |
| 588 vaesenc xmm4,xmm4,xmm1 |
| 589 vmovdqa XMMWORD[48+rdx],xmm1 |
| 590 |
| 591 vpshufb xmm2,xmm1,xmm15 |
| 592 vaesenclast xmm2,xmm2,xmm0 |
| 593 vpslld xmm0,xmm0,1 |
| 594 vpsllq xmm3,xmm1,32 |
| 595 vpxor xmm1,xmm1,xmm3 |
| 596 vpshufb xmm3,xmm1,XMMWORD[con3] |
| 597 vpxor xmm1,xmm1,xmm3 |
| 598 vpxor xmm1,xmm1,xmm2 |
| 599 |
| 600 vaesenc xmm4,xmm4,xmm1 |
| 601 vmovdqa XMMWORD[64+rdx],xmm1 |
| 602 |
| 603 vpshufb xmm2,xmm1,xmm15 |
| 604 vaesenclast xmm2,xmm2,xmm0 |
| 605 vpslld xmm0,xmm0,1 |
| 606 vpsllq xmm3,xmm1,32 |
| 607 vpxor xmm1,xmm1,xmm3 |
| 608 vpshufb xmm3,xmm1,XMMWORD[con3] |
| 609 vpxor xmm1,xmm1,xmm3 |
| 610 vpxor xmm1,xmm1,xmm2 |
| 611 |
| 612 vaesenc xmm4,xmm4,xmm1 |
| 613 vmovdqa XMMWORD[80+rdx],xmm1 |
| 614 |
| 615 vpshufb xmm2,xmm1,xmm15 |
| 616 vaesenclast xmm2,xmm2,xmm0 |
| 617 vpslld xmm0,xmm0,1 |
| 618 vpsllq xmm3,xmm1,32 |
| 619 vpxor xmm1,xmm1,xmm3 |
| 620 vpshufb xmm3,xmm1,XMMWORD[con3] |
| 621 vpxor xmm1,xmm1,xmm3 |
| 622 vpxor xmm1,xmm1,xmm2 |
| 623 |
| 624 vaesenc xmm4,xmm4,xmm1 |
| 625 vmovdqa XMMWORD[96+rdx],xmm1 |
| 626 |
| 627 vpshufb xmm2,xmm1,xmm15 |
| 628 vaesenclast xmm2,xmm2,xmm0 |
| 629 vpslld xmm0,xmm0,1 |
| 630 vpsllq xmm3,xmm1,32 |
| 631 vpxor xmm1,xmm1,xmm3 |
| 632 vpshufb xmm3,xmm1,XMMWORD[con3] |
| 633 vpxor xmm1,xmm1,xmm3 |
| 634 vpxor xmm1,xmm1,xmm2 |
| 635 |
| 636 vaesenc xmm4,xmm4,xmm1 |
| 637 vmovdqa XMMWORD[112+rdx],xmm1 |
| 638 |
| 639 vpshufb xmm2,xmm1,xmm15 |
| 640 vaesenclast xmm2,xmm2,xmm0 |
| 641 vpslld xmm0,xmm0,1 |
| 642 vpsllq xmm3,xmm1,32 |
| 643 vpxor xmm1,xmm1,xmm3 |
| 644 vpshufb xmm3,xmm1,XMMWORD[con3] |
| 645 vpxor xmm1,xmm1,xmm3 |
| 646 vpxor xmm1,xmm1,xmm2 |
| 647 |
| 648 vaesenc xmm4,xmm4,xmm1 |
| 649 vmovdqa XMMWORD[128+rdx],xmm1 |
| 650 |
| 651 |
| 652 vmovdqa xmm0,XMMWORD[con2] |
| 653 |
| 654 vpshufb xmm2,xmm1,xmm15 |
| 655 vaesenclast xmm2,xmm2,xmm0 |
| 656 vpslld xmm0,xmm0,1 |
| 657 vpsllq xmm3,xmm1,32 |
| 658 vpxor xmm1,xmm1,xmm3 |
| 659 vpshufb xmm3,xmm1,XMMWORD[con3] |
| 660 vpxor xmm1,xmm1,xmm3 |
| 661 vpxor xmm1,xmm1,xmm2 |
| 662 |
| 663 vaesenc xmm4,xmm4,xmm1 |
| 664 vmovdqa XMMWORD[144+rdx],xmm1 |
| 665 |
| 666 vpshufb xmm2,xmm1,xmm15 |
| 667 vaesenclast xmm2,xmm2,xmm0 |
| 668 vpsllq xmm3,xmm1,32 |
| 669 vpxor xmm1,xmm1,xmm3 |
| 670 vpshufb xmm3,xmm1,XMMWORD[con3] |
| 671 vpxor xmm1,xmm1,xmm3 |
| 672 vpxor xmm1,xmm1,xmm2 |
| 673 |
| 674 vaesenclast xmm4,xmm4,xmm1 |
| 675 vmovdqa XMMWORD[160+rdx],xmm1 |
| 676 |
| 677 |
| 678 vmovdqa XMMWORD[rsi],xmm4 |
| 679 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 680 mov rsi,QWORD[16+rsp] |
| 681 DB 0F3h,0C3h ;repret |
| 682 |
| 683 $L$SEH_end_aes128gcmsiv_aes_ks_enc_x1: |
| 684 global aes128gcmsiv_kdf |
| 685 |
| 686 ALIGN 16 |
| 687 aes128gcmsiv_kdf: |
| 688 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 689 mov QWORD[16+rsp],rsi |
| 690 mov rax,rsp |
| 691 $L$SEH_begin_aes128gcmsiv_kdf: |
| 692 mov rdi,rcx |
| 693 mov rsi,rdx |
| 694 mov rdx,r8 |
| 695 |
| 696 |
| 697 |
| 698 |
| 699 |
| 700 |
| 701 |
| 702 vmovdqa xmm1,XMMWORD[rdx] |
| 703 vmovdqa xmm9,XMMWORD[rdi] |
| 704 vmovdqa xmm12,XMMWORD[and_mask] |
| 705 vmovdqa xmm13,XMMWORD[one] |
| 706 vpshufd xmm9,xmm9,0x90 |
| 707 vpand xmm9,xmm9,xmm12 |
| 708 vpaddd xmm10,xmm9,xmm13 |
| 709 vpaddd xmm11,xmm10,xmm13 |
| 710 vpaddd xmm12,xmm11,xmm13 |
| 711 |
| 712 vpxor xmm9,xmm9,xmm1 |
| 713 vpxor xmm10,xmm10,xmm1 |
| 714 vpxor xmm11,xmm11,xmm1 |
| 715 vpxor xmm12,xmm12,xmm1 |
| 716 |
| 717 vmovdqa xmm1,XMMWORD[16+rdx] |
| 718 vaesenc xmm9,xmm9,xmm1 |
| 719 vaesenc xmm10,xmm10,xmm1 |
| 720 vaesenc xmm11,xmm11,xmm1 |
| 721 vaesenc xmm12,xmm12,xmm1 |
| 722 |
| 723 vmovdqa xmm2,XMMWORD[32+rdx] |
| 724 vaesenc xmm9,xmm9,xmm2 |
| 725 vaesenc xmm10,xmm10,xmm2 |
| 726 vaesenc xmm11,xmm11,xmm2 |
| 727 vaesenc xmm12,xmm12,xmm2 |
| 728 |
| 729 vmovdqa xmm1,XMMWORD[48+rdx] |
| 730 vaesenc xmm9,xmm9,xmm1 |
| 731 vaesenc xmm10,xmm10,xmm1 |
| 732 vaesenc xmm11,xmm11,xmm1 |
| 733 vaesenc xmm12,xmm12,xmm1 |
| 734 |
| 735 vmovdqa xmm2,XMMWORD[64+rdx] |
| 736 vaesenc xmm9,xmm9,xmm2 |
| 737 vaesenc xmm10,xmm10,xmm2 |
| 738 vaesenc xmm11,xmm11,xmm2 |
| 739 vaesenc xmm12,xmm12,xmm2 |
| 740 |
| 741 vmovdqa xmm1,XMMWORD[80+rdx] |
| 742 vaesenc xmm9,xmm9,xmm1 |
| 743 vaesenc xmm10,xmm10,xmm1 |
| 744 vaesenc xmm11,xmm11,xmm1 |
| 745 vaesenc xmm12,xmm12,xmm1 |
| 746 |
| 747 vmovdqa xmm2,XMMWORD[96+rdx] |
| 748 vaesenc xmm9,xmm9,xmm2 |
| 749 vaesenc xmm10,xmm10,xmm2 |
| 750 vaesenc xmm11,xmm11,xmm2 |
| 751 vaesenc xmm12,xmm12,xmm2 |
| 752 |
| 753 vmovdqa xmm1,XMMWORD[112+rdx] |
| 754 vaesenc xmm9,xmm9,xmm1 |
| 755 vaesenc xmm10,xmm10,xmm1 |
| 756 vaesenc xmm11,xmm11,xmm1 |
| 757 vaesenc xmm12,xmm12,xmm1 |
| 758 |
| 759 vmovdqa xmm2,XMMWORD[128+rdx] |
| 760 vaesenc xmm9,xmm9,xmm2 |
| 761 vaesenc xmm10,xmm10,xmm2 |
| 762 vaesenc xmm11,xmm11,xmm2 |
| 763 vaesenc xmm12,xmm12,xmm2 |
| 764 |
| 765 vmovdqa xmm1,XMMWORD[144+rdx] |
| 766 vaesenc xmm9,xmm9,xmm1 |
| 767 vaesenc xmm10,xmm10,xmm1 |
| 768 vaesenc xmm11,xmm11,xmm1 |
| 769 vaesenc xmm12,xmm12,xmm1 |
| 770 |
| 771 vmovdqa xmm2,XMMWORD[160+rdx] |
| 772 vaesenclast xmm9,xmm9,xmm2 |
| 773 vaesenclast xmm10,xmm10,xmm2 |
| 774 vaesenclast xmm11,xmm11,xmm2 |
| 775 vaesenclast xmm12,xmm12,xmm2 |
| 776 |
| 777 |
| 778 vmovdqa XMMWORD[rsi],xmm9 |
| 779 vmovdqa XMMWORD[16+rsi],xmm10 |
| 780 vmovdqa XMMWORD[32+rsi],xmm11 |
| 781 vmovdqa XMMWORD[48+rsi],xmm12 |
| 782 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 783 mov rsi,QWORD[16+rsp] |
| 784 DB 0F3h,0C3h ;repret |
| 785 |
| 786 $L$SEH_end_aes128gcmsiv_kdf: |
| 787 global aes128gcmsiv_enc_msg_x4 |
| 788 |
| 789 ALIGN 16 |
| 790 aes128gcmsiv_enc_msg_x4: |
| 791 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 792 mov QWORD[16+rsp],rsi |
| 793 mov rax,rsp |
| 794 $L$SEH_begin_aes128gcmsiv_enc_msg_x4: |
| 795 mov rdi,rcx |
| 796 mov rsi,rdx |
| 797 mov rdx,r8 |
| 798 mov rcx,r9 |
| 799 mov r8,QWORD[40+rsp] |
| 800 |
| 801 |
| 802 |
| 803 test r8,r8 |
| 804 jnz NEAR $L$128_enc_msg_x4_start |
| 805 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 806 mov rsi,QWORD[16+rsp] |
| 807 DB 0F3h,0C3h ;repret |
| 808 |
| 809 $L$128_enc_msg_x4_start: |
| 810 push r12 |
| 811 |
| 812 push r13 |
| 813 |
| 814 |
| 815 shr r8,4 |
| 816 mov r10,r8 |
| 817 shl r10,62 |
| 818 shr r10,62 |
| 819 |
| 820 |
| 821 vmovdqa xmm15,XMMWORD[rdx] |
| 822 vpor xmm15,xmm15,XMMWORD[OR_MASK] |
| 823 |
| 824 vmovdqu xmm4,XMMWORD[four] |
| 825 vmovdqa xmm0,xmm15 |
| 826 vpaddd xmm1,xmm15,XMMWORD[one] |
| 827 vpaddd xmm2,xmm15,XMMWORD[two] |
| 828 vpaddd xmm3,xmm15,XMMWORD[three] |
| 829 |
| 830 shr r8,2 |
| 831 je NEAR $L$128_enc_msg_x4_check_remainder |
| 832 |
| 833 sub rsi,64 |
| 834 sub rdi,64 |
| 835 |
| 836 $L$128_enc_msg_x4_loop1: |
| 837 add rsi,64 |
| 838 add rdi,64 |
| 839 |
| 840 vmovdqa xmm5,xmm0 |
| 841 vmovdqa xmm6,xmm1 |
| 842 vmovdqa xmm7,xmm2 |
| 843 vmovdqa xmm8,xmm3 |
| 844 |
| 845 vpxor xmm5,xmm5,XMMWORD[rcx] |
| 846 vpxor xmm6,xmm6,XMMWORD[rcx] |
| 847 vpxor xmm7,xmm7,XMMWORD[rcx] |
| 848 vpxor xmm8,xmm8,XMMWORD[rcx] |
| 849 |
| 850 vmovdqu xmm12,XMMWORD[16+rcx] |
| 851 vaesenc xmm5,xmm5,xmm12 |
| 852 vaesenc xmm6,xmm6,xmm12 |
| 853 vaesenc xmm7,xmm7,xmm12 |
| 854 vaesenc xmm8,xmm8,xmm12 |
| 855 |
| 856 vpaddd xmm0,xmm0,xmm4 |
| 857 vmovdqu xmm12,XMMWORD[32+rcx] |
| 858 vaesenc xmm5,xmm5,xmm12 |
| 859 vaesenc xmm6,xmm6,xmm12 |
| 860 vaesenc xmm7,xmm7,xmm12 |
| 861 vaesenc xmm8,xmm8,xmm12 |
| 862 |
| 863 vpaddd xmm1,xmm1,xmm4 |
| 864 vmovdqu xmm12,XMMWORD[48+rcx] |
| 865 vaesenc xmm5,xmm5,xmm12 |
| 866 vaesenc xmm6,xmm6,xmm12 |
| 867 vaesenc xmm7,xmm7,xmm12 |
| 868 vaesenc xmm8,xmm8,xmm12 |
| 869 |
| 870 vpaddd xmm2,xmm2,xmm4 |
| 871 vmovdqu xmm12,XMMWORD[64+rcx] |
| 872 vaesenc xmm5,xmm5,xmm12 |
| 873 vaesenc xmm6,xmm6,xmm12 |
| 874 vaesenc xmm7,xmm7,xmm12 |
| 875 vaesenc xmm8,xmm8,xmm12 |
| 876 |
| 877 vpaddd xmm3,xmm3,xmm4 |
| 878 |
| 879 vmovdqu xmm12,XMMWORD[80+rcx] |
| 880 vaesenc xmm5,xmm5,xmm12 |
| 881 vaesenc xmm6,xmm6,xmm12 |
| 882 vaesenc xmm7,xmm7,xmm12 |
| 883 vaesenc xmm8,xmm8,xmm12 |
| 884 |
| 885 vmovdqu xmm12,XMMWORD[96+rcx] |
| 886 vaesenc xmm5,xmm5,xmm12 |
| 887 vaesenc xmm6,xmm6,xmm12 |
| 888 vaesenc xmm7,xmm7,xmm12 |
| 889 vaesenc xmm8,xmm8,xmm12 |
| 890 |
| 891 vmovdqu xmm12,XMMWORD[112+rcx] |
| 892 vaesenc xmm5,xmm5,xmm12 |
| 893 vaesenc xmm6,xmm6,xmm12 |
| 894 vaesenc xmm7,xmm7,xmm12 |
| 895 vaesenc xmm8,xmm8,xmm12 |
| 896 |
| 897 vmovdqu xmm12,XMMWORD[128+rcx] |
| 898 vaesenc xmm5,xmm5,xmm12 |
| 899 vaesenc xmm6,xmm6,xmm12 |
| 900 vaesenc xmm7,xmm7,xmm12 |
| 901 vaesenc xmm8,xmm8,xmm12 |
| 902 |
| 903 vmovdqu xmm12,XMMWORD[144+rcx] |
| 904 vaesenc xmm5,xmm5,xmm12 |
| 905 vaesenc xmm6,xmm6,xmm12 |
| 906 vaesenc xmm7,xmm7,xmm12 |
| 907 vaesenc xmm8,xmm8,xmm12 |
| 908 |
| 909 vmovdqu xmm12,XMMWORD[160+rcx] |
| 910 vaesenclast xmm5,xmm5,xmm12 |
| 911 vaesenclast xmm6,xmm6,xmm12 |
| 912 vaesenclast xmm7,xmm7,xmm12 |
| 913 vaesenclast xmm8,xmm8,xmm12 |
| 914 |
| 915 |
| 916 |
| 917 vpxor xmm5,xmm5,XMMWORD[rdi] |
| 918 vpxor xmm6,xmm6,XMMWORD[16+rdi] |
| 919 vpxor xmm7,xmm7,XMMWORD[32+rdi] |
| 920 vpxor xmm8,xmm8,XMMWORD[48+rdi] |
| 921 |
| 922 sub r8,1 |
| 923 |
| 924 vmovdqu XMMWORD[rsi],xmm5 |
| 925 vmovdqu XMMWORD[16+rsi],xmm6 |
| 926 vmovdqu XMMWORD[32+rsi],xmm7 |
| 927 vmovdqu XMMWORD[48+rsi],xmm8 |
| 928 |
| 929 jne NEAR $L$128_enc_msg_x4_loop1 |
| 930 |
| 931 add rsi,64 |
| 932 add rdi,64 |
| 933 |
| 934 $L$128_enc_msg_x4_check_remainder: |
| 935 cmp r10,0 |
| 936 je NEAR $L$128_enc_msg_x4_out |
| 937 |
| 938 $L$128_enc_msg_x4_loop2: |
| 939 |
| 940 |
| 941 vmovdqa xmm5,xmm0 |
| 942 vpaddd xmm0,xmm0,XMMWORD[one] |
| 943 |
| 944 vpxor xmm5,xmm5,XMMWORD[rcx] |
| 945 vaesenc xmm5,xmm5,XMMWORD[16+rcx] |
| 946 vaesenc xmm5,xmm5,XMMWORD[32+rcx] |
| 947 vaesenc xmm5,xmm5,XMMWORD[48+rcx] |
| 948 vaesenc xmm5,xmm5,XMMWORD[64+rcx] |
| 949 vaesenc xmm5,xmm5,XMMWORD[80+rcx] |
| 950 vaesenc xmm5,xmm5,XMMWORD[96+rcx] |
| 951 vaesenc xmm5,xmm5,XMMWORD[112+rcx] |
| 952 vaesenc xmm5,xmm5,XMMWORD[128+rcx] |
| 953 vaesenc xmm5,xmm5,XMMWORD[144+rcx] |
| 954 vaesenclast xmm5,xmm5,XMMWORD[160+rcx] |
| 955 |
| 956 |
| 957 vpxor xmm5,xmm5,XMMWORD[rdi] |
| 958 vmovdqu XMMWORD[rsi],xmm5 |
| 959 |
| 960 add rdi,16 |
| 961 add rsi,16 |
| 962 |
| 963 sub r10,1 |
| 964 jne NEAR $L$128_enc_msg_x4_loop2 |
| 965 |
| 966 $L$128_enc_msg_x4_out: |
| 967 pop r13 |
| 968 |
| 969 pop r12 |
| 970 |
| 971 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 972 mov rsi,QWORD[16+rsp] |
| 973 DB 0F3h,0C3h ;repret |
| 974 |
| 975 $L$SEH_end_aes128gcmsiv_enc_msg_x4: |
| 976 global aes128gcmsiv_enc_msg_x8 |
| 977 |
| 978 ALIGN 16 |
| 979 aes128gcmsiv_enc_msg_x8: |
| 980 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 981 mov QWORD[16+rsp],rsi |
| 982 mov rax,rsp |
| 983 $L$SEH_begin_aes128gcmsiv_enc_msg_x8: |
| 984 mov rdi,rcx |
| 985 mov rsi,rdx |
| 986 mov rdx,r8 |
| 987 mov rcx,r9 |
| 988 mov r8,QWORD[40+rsp] |
| 989 |
| 990 |
| 991 |
| 992 test r8,r8 |
| 993 jnz NEAR $L$128_enc_msg_x8_start |
| 994 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 995 mov rsi,QWORD[16+rsp] |
| 996 DB 0F3h,0C3h ;repret |
| 997 |
| 998 $L$128_enc_msg_x8_start: |
| 999 push r12 |
| 1000 |
| 1001 push r13 |
| 1002 |
| 1003 push rbp |
| 1004 |
| 1005 mov rbp,rsp |
| 1006 |
| 1007 |
| 1008 |
| 1009 sub rsp,128 |
| 1010 and rsp,-64 |
| 1011 |
| 1012 shr r8,4 |
| 1013 mov r10,r8 |
| 1014 shl r10,61 |
| 1015 shr r10,61 |
| 1016 |
| 1017 |
| 1018 vmovdqu xmm1,XMMWORD[rdx] |
| 1019 vpor xmm1,xmm1,XMMWORD[OR_MASK] |
| 1020 |
| 1021 |
| 1022 vpaddd xmm0,xmm1,XMMWORD[seven] |
| 1023 vmovdqu XMMWORD[rsp],xmm0 |
| 1024 vpaddd xmm9,xmm1,XMMWORD[one] |
| 1025 vpaddd xmm10,xmm1,XMMWORD[two] |
| 1026 vpaddd xmm11,xmm1,XMMWORD[three] |
| 1027 vpaddd xmm12,xmm1,XMMWORD[four] |
| 1028 vpaddd xmm13,xmm1,XMMWORD[five] |
| 1029 vpaddd xmm14,xmm1,XMMWORD[six] |
| 1030 vmovdqa xmm0,xmm1 |
| 1031 |
| 1032 shr r8,3 |
| 1033 je NEAR $L$128_enc_msg_x8_check_remainder |
| 1034 |
| 1035 sub rsi,128 |
| 1036 sub rdi,128 |
| 1037 |
| 1038 $L$128_enc_msg_x8_loop1: |
| 1039 add rsi,128 |
| 1040 add rdi,128 |
| 1041 |
| 1042 vmovdqa xmm1,xmm0 |
| 1043 vmovdqa xmm2,xmm9 |
| 1044 vmovdqa xmm3,xmm10 |
| 1045 vmovdqa xmm4,xmm11 |
| 1046 vmovdqa xmm5,xmm12 |
| 1047 vmovdqa xmm6,xmm13 |
| 1048 vmovdqa xmm7,xmm14 |
| 1049 |
| 1050 vmovdqu xmm8,XMMWORD[rsp] |
| 1051 |
| 1052 vpxor xmm1,xmm1,XMMWORD[rcx] |
| 1053 vpxor xmm2,xmm2,XMMWORD[rcx] |
| 1054 vpxor xmm3,xmm3,XMMWORD[rcx] |
| 1055 vpxor xmm4,xmm4,XMMWORD[rcx] |
| 1056 vpxor xmm5,xmm5,XMMWORD[rcx] |
| 1057 vpxor xmm6,xmm6,XMMWORD[rcx] |
| 1058 vpxor xmm7,xmm7,XMMWORD[rcx] |
| 1059 vpxor xmm8,xmm8,XMMWORD[rcx] |
| 1060 |
| 1061 vmovdqu xmm15,XMMWORD[16+rcx] |
| 1062 vaesenc xmm1,xmm1,xmm15 |
| 1063 vaesenc xmm2,xmm2,xmm15 |
| 1064 vaesenc xmm3,xmm3,xmm15 |
| 1065 vaesenc xmm4,xmm4,xmm15 |
| 1066 vaesenc xmm5,xmm5,xmm15 |
| 1067 vaesenc xmm6,xmm6,xmm15 |
| 1068 vaesenc xmm7,xmm7,xmm15 |
| 1069 vaesenc xmm8,xmm8,xmm15 |
| 1070 |
| 1071 vmovdqu xmm14,XMMWORD[rsp] |
| 1072 vpaddd xmm14,xmm14,XMMWORD[eight] |
| 1073 vmovdqu XMMWORD[rsp],xmm14 |
| 1074 vmovdqu xmm15,XMMWORD[32+rcx] |
| 1075 vaesenc xmm1,xmm1,xmm15 |
| 1076 vaesenc xmm2,xmm2,xmm15 |
| 1077 vaesenc xmm3,xmm3,xmm15 |
| 1078 vaesenc xmm4,xmm4,xmm15 |
| 1079 vaesenc xmm5,xmm5,xmm15 |
| 1080 vaesenc xmm6,xmm6,xmm15 |
| 1081 vaesenc xmm7,xmm7,xmm15 |
| 1082 vaesenc xmm8,xmm8,xmm15 |
| 1083 |
| 1084 vpsubd xmm14,xmm14,XMMWORD[one] |
| 1085 vmovdqu xmm15,XMMWORD[48+rcx] |
| 1086 vaesenc xmm1,xmm1,xmm15 |
| 1087 vaesenc xmm2,xmm2,xmm15 |
| 1088 vaesenc xmm3,xmm3,xmm15 |
| 1089 vaesenc xmm4,xmm4,xmm15 |
| 1090 vaesenc xmm5,xmm5,xmm15 |
| 1091 vaesenc xmm6,xmm6,xmm15 |
| 1092 vaesenc xmm7,xmm7,xmm15 |
| 1093 vaesenc xmm8,xmm8,xmm15 |
| 1094 |
| 1095 vpaddd xmm0,xmm0,XMMWORD[eight] |
| 1096 vmovdqu xmm15,XMMWORD[64+rcx] |
| 1097 vaesenc xmm1,xmm1,xmm15 |
| 1098 vaesenc xmm2,xmm2,xmm15 |
| 1099 vaesenc xmm3,xmm3,xmm15 |
| 1100 vaesenc xmm4,xmm4,xmm15 |
| 1101 vaesenc xmm5,xmm5,xmm15 |
| 1102 vaesenc xmm6,xmm6,xmm15 |
| 1103 vaesenc xmm7,xmm7,xmm15 |
| 1104 vaesenc xmm8,xmm8,xmm15 |
| 1105 |
| 1106 vpaddd xmm9,xmm9,XMMWORD[eight] |
| 1107 vmovdqu xmm15,XMMWORD[80+rcx] |
| 1108 vaesenc xmm1,xmm1,xmm15 |
| 1109 vaesenc xmm2,xmm2,xmm15 |
| 1110 vaesenc xmm3,xmm3,xmm15 |
| 1111 vaesenc xmm4,xmm4,xmm15 |
| 1112 vaesenc xmm5,xmm5,xmm15 |
| 1113 vaesenc xmm6,xmm6,xmm15 |
| 1114 vaesenc xmm7,xmm7,xmm15 |
| 1115 vaesenc xmm8,xmm8,xmm15 |
| 1116 |
| 1117 vpaddd xmm10,xmm10,XMMWORD[eight] |
| 1118 vmovdqu xmm15,XMMWORD[96+rcx] |
| 1119 vaesenc xmm1,xmm1,xmm15 |
| 1120 vaesenc xmm2,xmm2,xmm15 |
| 1121 vaesenc xmm3,xmm3,xmm15 |
| 1122 vaesenc xmm4,xmm4,xmm15 |
| 1123 vaesenc xmm5,xmm5,xmm15 |
| 1124 vaesenc xmm6,xmm6,xmm15 |
| 1125 vaesenc xmm7,xmm7,xmm15 |
| 1126 vaesenc xmm8,xmm8,xmm15 |
| 1127 |
| 1128 vpaddd xmm11,xmm11,XMMWORD[eight] |
| 1129 vmovdqu xmm15,XMMWORD[112+rcx] |
| 1130 vaesenc xmm1,xmm1,xmm15 |
| 1131 vaesenc xmm2,xmm2,xmm15 |
| 1132 vaesenc xmm3,xmm3,xmm15 |
| 1133 vaesenc xmm4,xmm4,xmm15 |
| 1134 vaesenc xmm5,xmm5,xmm15 |
| 1135 vaesenc xmm6,xmm6,xmm15 |
| 1136 vaesenc xmm7,xmm7,xmm15 |
| 1137 vaesenc xmm8,xmm8,xmm15 |
| 1138 |
| 1139 vpaddd xmm12,xmm12,XMMWORD[eight] |
| 1140 vmovdqu xmm15,XMMWORD[128+rcx] |
| 1141 vaesenc xmm1,xmm1,xmm15 |
| 1142 vaesenc xmm2,xmm2,xmm15 |
| 1143 vaesenc xmm3,xmm3,xmm15 |
| 1144 vaesenc xmm4,xmm4,xmm15 |
| 1145 vaesenc xmm5,xmm5,xmm15 |
| 1146 vaesenc xmm6,xmm6,xmm15 |
| 1147 vaesenc xmm7,xmm7,xmm15 |
| 1148 vaesenc xmm8,xmm8,xmm15 |
| 1149 |
| 1150 vpaddd xmm13,xmm13,XMMWORD[eight] |
| 1151 vmovdqu xmm15,XMMWORD[144+rcx] |
| 1152 vaesenc xmm1,xmm1,xmm15 |
| 1153 vaesenc xmm2,xmm2,xmm15 |
| 1154 vaesenc xmm3,xmm3,xmm15 |
| 1155 vaesenc xmm4,xmm4,xmm15 |
| 1156 vaesenc xmm5,xmm5,xmm15 |
| 1157 vaesenc xmm6,xmm6,xmm15 |
| 1158 vaesenc xmm7,xmm7,xmm15 |
| 1159 vaesenc xmm8,xmm8,xmm15 |
| 1160 |
| 1161 vmovdqu xmm15,XMMWORD[160+rcx] |
| 1162 vaesenclast xmm1,xmm1,xmm15 |
| 1163 vaesenclast xmm2,xmm2,xmm15 |
| 1164 vaesenclast xmm3,xmm3,xmm15 |
| 1165 vaesenclast xmm4,xmm4,xmm15 |
| 1166 vaesenclast xmm5,xmm5,xmm15 |
| 1167 vaesenclast xmm6,xmm6,xmm15 |
| 1168 vaesenclast xmm7,xmm7,xmm15 |
| 1169 vaesenclast xmm8,xmm8,xmm15 |
| 1170 |
| 1171 |
| 1172 |
| 1173 vpxor xmm1,xmm1,XMMWORD[rdi] |
| 1174 vpxor xmm2,xmm2,XMMWORD[16+rdi] |
| 1175 vpxor xmm3,xmm3,XMMWORD[32+rdi] |
| 1176 vpxor xmm4,xmm4,XMMWORD[48+rdi] |
| 1177 vpxor xmm5,xmm5,XMMWORD[64+rdi] |
| 1178 vpxor xmm6,xmm6,XMMWORD[80+rdi] |
| 1179 vpxor xmm7,xmm7,XMMWORD[96+rdi] |
| 1180 vpxor xmm8,xmm8,XMMWORD[112+rdi] |
| 1181 |
| 1182 dec r8 |
| 1183 |
| 1184 vmovdqu XMMWORD[rsi],xmm1 |
| 1185 vmovdqu XMMWORD[16+rsi],xmm2 |
| 1186 vmovdqu XMMWORD[32+rsi],xmm3 |
| 1187 vmovdqu XMMWORD[48+rsi],xmm4 |
| 1188 vmovdqu XMMWORD[64+rsi],xmm5 |
| 1189 vmovdqu XMMWORD[80+rsi],xmm6 |
| 1190 vmovdqu XMMWORD[96+rsi],xmm7 |
| 1191 vmovdqu XMMWORD[112+rsi],xmm8 |
| 1192 |
| 1193 jne NEAR $L$128_enc_msg_x8_loop1 |
| 1194 |
| 1195 add rsi,128 |
| 1196 add rdi,128 |
| 1197 |
| 1198 $L$128_enc_msg_x8_check_remainder: |
| 1199 cmp r10,0 |
| 1200 je NEAR $L$128_enc_msg_x8_out |
| 1201 |
| 1202 $L$128_enc_msg_x8_loop2: |
| 1203 |
| 1204 |
| 1205 vmovdqa xmm1,xmm0 |
| 1206 vpaddd xmm0,xmm0,XMMWORD[one] |
| 1207 |
| 1208 vpxor xmm1,xmm1,XMMWORD[rcx] |
| 1209 vaesenc xmm1,xmm1,XMMWORD[16+rcx] |
| 1210 vaesenc xmm1,xmm1,XMMWORD[32+rcx] |
| 1211 vaesenc xmm1,xmm1,XMMWORD[48+rcx] |
| 1212 vaesenc xmm1,xmm1,XMMWORD[64+rcx] |
| 1213 vaesenc xmm1,xmm1,XMMWORD[80+rcx] |
| 1214 vaesenc xmm1,xmm1,XMMWORD[96+rcx] |
| 1215 vaesenc xmm1,xmm1,XMMWORD[112+rcx] |
| 1216 vaesenc xmm1,xmm1,XMMWORD[128+rcx] |
| 1217 vaesenc xmm1,xmm1,XMMWORD[144+rcx] |
| 1218 vaesenclast xmm1,xmm1,XMMWORD[160+rcx] |
| 1219 |
| 1220 |
| 1221 vpxor xmm1,xmm1,XMMWORD[rdi] |
| 1222 |
| 1223 vmovdqu XMMWORD[rsi],xmm1 |
| 1224 |
| 1225 add rdi,16 |
| 1226 add rsi,16 |
| 1227 |
| 1228 dec r10 |
| 1229 jne NEAR $L$128_enc_msg_x8_loop2 |
| 1230 |
| 1231 $L$128_enc_msg_x8_out: |
| 1232 mov rsp,rbp |
| 1233 |
| 1234 pop rbp |
| 1235 |
| 1236 pop r13 |
| 1237 |
| 1238 pop r12 |
| 1239 |
| 1240 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 1241 mov rsi,QWORD[16+rsp] |
| 1242 DB 0F3h,0C3h ;repret |
| 1243 |
| 1244 $L$SEH_end_aes128gcmsiv_enc_msg_x8: |
| 1245 global aes128gcmsiv_dec |
| 1246 |
| 1247 ALIGN 16 |
| 1248 aes128gcmsiv_dec: |
| 1249 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 1250 mov QWORD[16+rsp],rsi |
| 1251 mov rax,rsp |
| 1252 $L$SEH_begin_aes128gcmsiv_dec: |
| 1253 mov rdi,rcx |
| 1254 mov rsi,rdx |
| 1255 mov rdx,r8 |
| 1256 mov rcx,r9 |
| 1257 mov r8,QWORD[40+rsp] |
| 1258 mov r9,QWORD[48+rsp] |
| 1259 |
| 1260 |
| 1261 |
| 1262 test r9,~15 |
| 1263 jnz NEAR $L$128_dec_start |
| 1264 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 1265 mov rsi,QWORD[16+rsp] |
| 1266 DB 0F3h,0C3h ;repret |
| 1267 |
| 1268 $L$128_dec_start: |
| 1269 vzeroupper |
| 1270 vmovdqa xmm0,XMMWORD[rdx] |
| 1271 mov rax,rdx |
| 1272 |
| 1273 lea rax,[32+rax] |
| 1274 lea rcx,[32+rcx] |
| 1275 |
| 1276 |
| 1277 vmovdqu xmm15,XMMWORD[r9*1+rdi] |
| 1278 vpor xmm15,xmm15,XMMWORD[OR_MASK] |
| 1279 and r9,~15 |
| 1280 |
| 1281 |
| 1282 cmp r9,96 |
| 1283 jb NEAR $L$128_dec_loop2 |
| 1284 |
| 1285 |
| 1286 sub r9,96 |
| 1287 vmovdqa xmm7,xmm15 |
| 1288 vpaddd xmm8,xmm7,XMMWORD[one] |
| 1289 vpaddd xmm9,xmm7,XMMWORD[two] |
| 1290 vpaddd xmm10,xmm9,XMMWORD[one] |
| 1291 vpaddd xmm11,xmm9,XMMWORD[two] |
| 1292 vpaddd xmm12,xmm11,XMMWORD[one] |
| 1293 vpaddd xmm15,xmm11,XMMWORD[two] |
| 1294 |
| 1295 vpxor xmm7,xmm7,XMMWORD[r8] |
| 1296 vpxor xmm8,xmm8,XMMWORD[r8] |
| 1297 vpxor xmm9,xmm9,XMMWORD[r8] |
| 1298 vpxor xmm10,xmm10,XMMWORD[r8] |
| 1299 vpxor xmm11,xmm11,XMMWORD[r8] |
| 1300 vpxor xmm12,xmm12,XMMWORD[r8] |
| 1301 |
| 1302 vmovdqu xmm4,XMMWORD[16+r8] |
| 1303 vaesenc xmm7,xmm7,xmm4 |
| 1304 vaesenc xmm8,xmm8,xmm4 |
| 1305 vaesenc xmm9,xmm9,xmm4 |
| 1306 vaesenc xmm10,xmm10,xmm4 |
| 1307 vaesenc xmm11,xmm11,xmm4 |
| 1308 vaesenc xmm12,xmm12,xmm4 |
| 1309 |
| 1310 vmovdqu xmm4,XMMWORD[32+r8] |
| 1311 vaesenc xmm7,xmm7,xmm4 |
| 1312 vaesenc xmm8,xmm8,xmm4 |
| 1313 vaesenc xmm9,xmm9,xmm4 |
| 1314 vaesenc xmm10,xmm10,xmm4 |
| 1315 vaesenc xmm11,xmm11,xmm4 |
| 1316 vaesenc xmm12,xmm12,xmm4 |
| 1317 |
| 1318 vmovdqu xmm4,XMMWORD[48+r8] |
| 1319 vaesenc xmm7,xmm7,xmm4 |
| 1320 vaesenc xmm8,xmm8,xmm4 |
| 1321 vaesenc xmm9,xmm9,xmm4 |
| 1322 vaesenc xmm10,xmm10,xmm4 |
| 1323 vaesenc xmm11,xmm11,xmm4 |
| 1324 vaesenc xmm12,xmm12,xmm4 |
| 1325 |
| 1326 vmovdqu xmm4,XMMWORD[64+r8] |
| 1327 vaesenc xmm7,xmm7,xmm4 |
| 1328 vaesenc xmm8,xmm8,xmm4 |
| 1329 vaesenc xmm9,xmm9,xmm4 |
| 1330 vaesenc xmm10,xmm10,xmm4 |
| 1331 vaesenc xmm11,xmm11,xmm4 |
| 1332 vaesenc xmm12,xmm12,xmm4 |
| 1333 |
| 1334 vmovdqu xmm4,XMMWORD[80+r8] |
| 1335 vaesenc xmm7,xmm7,xmm4 |
| 1336 vaesenc xmm8,xmm8,xmm4 |
| 1337 vaesenc xmm9,xmm9,xmm4 |
| 1338 vaesenc xmm10,xmm10,xmm4 |
| 1339 vaesenc xmm11,xmm11,xmm4 |
| 1340 vaesenc xmm12,xmm12,xmm4 |
| 1341 |
| 1342 vmovdqu xmm4,XMMWORD[96+r8] |
| 1343 vaesenc xmm7,xmm7,xmm4 |
| 1344 vaesenc xmm8,xmm8,xmm4 |
| 1345 vaesenc xmm9,xmm9,xmm4 |
| 1346 vaesenc xmm10,xmm10,xmm4 |
| 1347 vaesenc xmm11,xmm11,xmm4 |
| 1348 vaesenc xmm12,xmm12,xmm4 |
| 1349 |
| 1350 vmovdqu xmm4,XMMWORD[112+r8] |
| 1351 vaesenc xmm7,xmm7,xmm4 |
| 1352 vaesenc xmm8,xmm8,xmm4 |
| 1353 vaesenc xmm9,xmm9,xmm4 |
| 1354 vaesenc xmm10,xmm10,xmm4 |
| 1355 vaesenc xmm11,xmm11,xmm4 |
| 1356 vaesenc xmm12,xmm12,xmm4 |
| 1357 |
| 1358 vmovdqu xmm4,XMMWORD[128+r8] |
| 1359 vaesenc xmm7,xmm7,xmm4 |
| 1360 vaesenc xmm8,xmm8,xmm4 |
| 1361 vaesenc xmm9,xmm9,xmm4 |
| 1362 vaesenc xmm10,xmm10,xmm4 |
| 1363 vaesenc xmm11,xmm11,xmm4 |
| 1364 vaesenc xmm12,xmm12,xmm4 |
| 1365 |
| 1366 vmovdqu xmm4,XMMWORD[144+r8] |
| 1367 vaesenc xmm7,xmm7,xmm4 |
| 1368 vaesenc xmm8,xmm8,xmm4 |
| 1369 vaesenc xmm9,xmm9,xmm4 |
| 1370 vaesenc xmm10,xmm10,xmm4 |
| 1371 vaesenc xmm11,xmm11,xmm4 |
| 1372 vaesenc xmm12,xmm12,xmm4 |
| 1373 |
| 1374 vmovdqu xmm4,XMMWORD[160+r8] |
| 1375 vaesenclast xmm7,xmm7,xmm4 |
| 1376 vaesenclast xmm8,xmm8,xmm4 |
| 1377 vaesenclast xmm9,xmm9,xmm4 |
| 1378 vaesenclast xmm10,xmm10,xmm4 |
| 1379 vaesenclast xmm11,xmm11,xmm4 |
| 1380 vaesenclast xmm12,xmm12,xmm4 |
| 1381 |
| 1382 |
| 1383 vpxor xmm7,xmm7,XMMWORD[rdi] |
| 1384 vpxor xmm8,xmm8,XMMWORD[16+rdi] |
| 1385 vpxor xmm9,xmm9,XMMWORD[32+rdi] |
| 1386 vpxor xmm10,xmm10,XMMWORD[48+rdi] |
| 1387 vpxor xmm11,xmm11,XMMWORD[64+rdi] |
| 1388 vpxor xmm12,xmm12,XMMWORD[80+rdi] |
| 1389 |
| 1390 vmovdqu XMMWORD[rsi],xmm7 |
| 1391 vmovdqu XMMWORD[16+rsi],xmm8 |
| 1392 vmovdqu XMMWORD[32+rsi],xmm9 |
| 1393 vmovdqu XMMWORD[48+rsi],xmm10 |
| 1394 vmovdqu XMMWORD[64+rsi],xmm11 |
| 1395 vmovdqu XMMWORD[80+rsi],xmm12 |
| 1396 |
| 1397 add rdi,96 |
| 1398 add rsi,96 |
| 1399 jmp NEAR $L$128_dec_loop1 |
| 1400 |
| 1401 |
| 1402 ALIGN 64 |
| 1403 $L$128_dec_loop1: |
| 1404 cmp r9,96 |
| 1405 jb NEAR $L$128_dec_finish_96 |
| 1406 sub r9,96 |
| 1407 |
| 1408 vmovdqa xmm6,xmm12 |
| 1409 vmovdqa XMMWORD[(16-32)+rax],xmm11 |
| 1410 vmovdqa XMMWORD[(32-32)+rax],xmm10 |
| 1411 vmovdqa XMMWORD[(48-32)+rax],xmm9 |
| 1412 vmovdqa XMMWORD[(64-32)+rax],xmm8 |
| 1413 vmovdqa XMMWORD[(80-32)+rax],xmm7 |
| 1414 |
| 1415 vmovdqa xmm7,xmm15 |
| 1416 vpaddd xmm8,xmm7,XMMWORD[one] |
| 1417 vpaddd xmm9,xmm7,XMMWORD[two] |
| 1418 vpaddd xmm10,xmm9,XMMWORD[one] |
| 1419 vpaddd xmm11,xmm9,XMMWORD[two] |
| 1420 vpaddd xmm12,xmm11,XMMWORD[one] |
| 1421 vpaddd xmm15,xmm11,XMMWORD[two] |
| 1422 |
| 1423 vmovdqa xmm4,XMMWORD[r8] |
| 1424 vpxor xmm7,xmm7,xmm4 |
| 1425 vpxor xmm8,xmm8,xmm4 |
| 1426 vpxor xmm9,xmm9,xmm4 |
| 1427 vpxor xmm10,xmm10,xmm4 |
| 1428 vpxor xmm11,xmm11,xmm4 |
| 1429 vpxor xmm12,xmm12,xmm4 |
| 1430 |
| 1431 vmovdqu xmm4,XMMWORD[((0-32))+rcx] |
| 1432 vpclmulqdq xmm2,xmm6,xmm4,0x11 |
| 1433 vpclmulqdq xmm3,xmm6,xmm4,0x00 |
| 1434 vpclmulqdq xmm1,xmm6,xmm4,0x01 |
| 1435 vpclmulqdq xmm4,xmm6,xmm4,0x10 |
| 1436 vpxor xmm1,xmm1,xmm4 |
| 1437 |
| 1438 vmovdqu xmm4,XMMWORD[16+r8] |
| 1439 vaesenc xmm7,xmm7,xmm4 |
| 1440 vaesenc xmm8,xmm8,xmm4 |
| 1441 vaesenc xmm9,xmm9,xmm4 |
| 1442 vaesenc xmm10,xmm10,xmm4 |
| 1443 vaesenc xmm11,xmm11,xmm4 |
| 1444 vaesenc xmm12,xmm12,xmm4 |
| 1445 |
| 1446 vmovdqu xmm6,XMMWORD[((-16))+rax] |
| 1447 vmovdqu xmm13,XMMWORD[((-16))+rcx] |
| 1448 |
| 1449 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 1450 vpxor xmm1,xmm1,xmm4 |
| 1451 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 1452 vpxor xmm2,xmm2,xmm4 |
| 1453 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 1454 vpxor xmm3,xmm3,xmm4 |
| 1455 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 1456 vpxor xmm1,xmm1,xmm4 |
| 1457 |
| 1458 |
| 1459 vmovdqu xmm4,XMMWORD[32+r8] |
| 1460 vaesenc xmm7,xmm7,xmm4 |
| 1461 vaesenc xmm8,xmm8,xmm4 |
| 1462 vaesenc xmm9,xmm9,xmm4 |
| 1463 vaesenc xmm10,xmm10,xmm4 |
| 1464 vaesenc xmm11,xmm11,xmm4 |
| 1465 vaesenc xmm12,xmm12,xmm4 |
| 1466 |
| 1467 vmovdqu xmm6,XMMWORD[rax] |
| 1468 vmovdqu xmm13,XMMWORD[rcx] |
| 1469 |
| 1470 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 1471 vpxor xmm1,xmm1,xmm4 |
| 1472 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 1473 vpxor xmm2,xmm2,xmm4 |
| 1474 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 1475 vpxor xmm3,xmm3,xmm4 |
| 1476 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 1477 vpxor xmm1,xmm1,xmm4 |
| 1478 |
| 1479 |
| 1480 vmovdqu xmm4,XMMWORD[48+r8] |
| 1481 vaesenc xmm7,xmm7,xmm4 |
| 1482 vaesenc xmm8,xmm8,xmm4 |
| 1483 vaesenc xmm9,xmm9,xmm4 |
| 1484 vaesenc xmm10,xmm10,xmm4 |
| 1485 vaesenc xmm11,xmm11,xmm4 |
| 1486 vaesenc xmm12,xmm12,xmm4 |
| 1487 |
| 1488 vmovdqu xmm6,XMMWORD[16+rax] |
| 1489 vmovdqu xmm13,XMMWORD[16+rcx] |
| 1490 |
| 1491 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 1492 vpxor xmm1,xmm1,xmm4 |
| 1493 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 1494 vpxor xmm2,xmm2,xmm4 |
| 1495 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 1496 vpxor xmm3,xmm3,xmm4 |
| 1497 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 1498 vpxor xmm1,xmm1,xmm4 |
| 1499 |
| 1500 |
| 1501 vmovdqu xmm4,XMMWORD[64+r8] |
| 1502 vaesenc xmm7,xmm7,xmm4 |
| 1503 vaesenc xmm8,xmm8,xmm4 |
| 1504 vaesenc xmm9,xmm9,xmm4 |
| 1505 vaesenc xmm10,xmm10,xmm4 |
| 1506 vaesenc xmm11,xmm11,xmm4 |
| 1507 vaesenc xmm12,xmm12,xmm4 |
| 1508 |
| 1509 vmovdqu xmm6,XMMWORD[32+rax] |
| 1510 vmovdqu xmm13,XMMWORD[32+rcx] |
| 1511 |
| 1512 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 1513 vpxor xmm1,xmm1,xmm4 |
| 1514 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 1515 vpxor xmm2,xmm2,xmm4 |
| 1516 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 1517 vpxor xmm3,xmm3,xmm4 |
| 1518 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 1519 vpxor xmm1,xmm1,xmm4 |
| 1520 |
| 1521 |
| 1522 vmovdqu xmm4,XMMWORD[80+r8] |
| 1523 vaesenc xmm7,xmm7,xmm4 |
| 1524 vaesenc xmm8,xmm8,xmm4 |
| 1525 vaesenc xmm9,xmm9,xmm4 |
| 1526 vaesenc xmm10,xmm10,xmm4 |
| 1527 vaesenc xmm11,xmm11,xmm4 |
| 1528 vaesenc xmm12,xmm12,xmm4 |
| 1529 |
| 1530 vmovdqu xmm4,XMMWORD[96+r8] |
| 1531 vaesenc xmm7,xmm7,xmm4 |
| 1532 vaesenc xmm8,xmm8,xmm4 |
| 1533 vaesenc xmm9,xmm9,xmm4 |
| 1534 vaesenc xmm10,xmm10,xmm4 |
| 1535 vaesenc xmm11,xmm11,xmm4 |
| 1536 vaesenc xmm12,xmm12,xmm4 |
| 1537 |
| 1538 vmovdqu xmm4,XMMWORD[112+r8] |
| 1539 vaesenc xmm7,xmm7,xmm4 |
| 1540 vaesenc xmm8,xmm8,xmm4 |
| 1541 vaesenc xmm9,xmm9,xmm4 |
| 1542 vaesenc xmm10,xmm10,xmm4 |
| 1543 vaesenc xmm11,xmm11,xmm4 |
| 1544 vaesenc xmm12,xmm12,xmm4 |
| 1545 |
| 1546 |
| 1547 vmovdqa xmm6,XMMWORD[((80-32))+rax] |
| 1548 vpxor xmm6,xmm6,xmm0 |
| 1549 vmovdqu xmm5,XMMWORD[((80-32))+rcx] |
| 1550 |
| 1551 vpclmulqdq xmm4,xmm6,xmm5,0x01 |
| 1552 vpxor xmm1,xmm1,xmm4 |
| 1553 vpclmulqdq xmm4,xmm6,xmm5,0x11 |
| 1554 vpxor xmm2,xmm2,xmm4 |
| 1555 vpclmulqdq xmm4,xmm6,xmm5,0x00 |
| 1556 vpxor xmm3,xmm3,xmm4 |
| 1557 vpclmulqdq xmm4,xmm6,xmm5,0x10 |
| 1558 vpxor xmm1,xmm1,xmm4 |
| 1559 |
| 1560 vmovdqu xmm4,XMMWORD[128+r8] |
| 1561 vaesenc xmm7,xmm7,xmm4 |
| 1562 vaesenc xmm8,xmm8,xmm4 |
| 1563 vaesenc xmm9,xmm9,xmm4 |
| 1564 vaesenc xmm10,xmm10,xmm4 |
| 1565 vaesenc xmm11,xmm11,xmm4 |
| 1566 vaesenc xmm12,xmm12,xmm4 |
| 1567 |
| 1568 |
| 1569 vpsrldq xmm4,xmm1,8 |
| 1570 vpxor xmm5,xmm2,xmm4 |
| 1571 vpslldq xmm4,xmm1,8 |
| 1572 vpxor xmm0,xmm3,xmm4 |
| 1573 |
| 1574 vmovdqa xmm3,XMMWORD[poly] |
| 1575 |
| 1576 vmovdqu xmm4,XMMWORD[144+r8] |
| 1577 vaesenc xmm7,xmm7,xmm4 |
| 1578 vaesenc xmm8,xmm8,xmm4 |
| 1579 vaesenc xmm9,xmm9,xmm4 |
| 1580 vaesenc xmm10,xmm10,xmm4 |
| 1581 vaesenc xmm11,xmm11,xmm4 |
| 1582 vaesenc xmm12,xmm12,xmm4 |
| 1583 |
| 1584 vmovdqu xmm6,XMMWORD[160+r8] |
| 1585 vpalignr xmm2,xmm0,xmm0,8 |
| 1586 vpclmulqdq xmm0,xmm0,xmm3,0x10 |
| 1587 vpxor xmm0,xmm2,xmm0 |
| 1588 |
| 1589 vpxor xmm4,xmm6,XMMWORD[rdi] |
| 1590 vaesenclast xmm7,xmm7,xmm4 |
| 1591 vpxor xmm4,xmm6,XMMWORD[16+rdi] |
| 1592 vaesenclast xmm8,xmm8,xmm4 |
| 1593 vpxor xmm4,xmm6,XMMWORD[32+rdi] |
| 1594 vaesenclast xmm9,xmm9,xmm4 |
| 1595 vpxor xmm4,xmm6,XMMWORD[48+rdi] |
| 1596 vaesenclast xmm10,xmm10,xmm4 |
| 1597 vpxor xmm4,xmm6,XMMWORD[64+rdi] |
| 1598 vaesenclast xmm11,xmm11,xmm4 |
| 1599 vpxor xmm4,xmm6,XMMWORD[80+rdi] |
| 1600 vaesenclast xmm12,xmm12,xmm4 |
| 1601 |
| 1602 vpalignr xmm2,xmm0,xmm0,8 |
| 1603 vpclmulqdq xmm0,xmm0,xmm3,0x10 |
| 1604 vpxor xmm0,xmm2,xmm0 |
| 1605 |
| 1606 vmovdqu XMMWORD[rsi],xmm7 |
| 1607 vmovdqu XMMWORD[16+rsi],xmm8 |
| 1608 vmovdqu XMMWORD[32+rsi],xmm9 |
| 1609 vmovdqu XMMWORD[48+rsi],xmm10 |
| 1610 vmovdqu XMMWORD[64+rsi],xmm11 |
| 1611 vmovdqu XMMWORD[80+rsi],xmm12 |
| 1612 |
| 1613 vpxor xmm0,xmm0,xmm5 |
| 1614 |
| 1615 lea rdi,[96+rdi] |
| 1616 lea rsi,[96+rsi] |
| 1617 jmp NEAR $L$128_dec_loop1 |
| 1618 |
| 1619 $L$128_dec_finish_96: |
| 1620 vmovdqa xmm6,xmm12 |
| 1621 vmovdqa XMMWORD[(16-32)+rax],xmm11 |
| 1622 vmovdqa XMMWORD[(32-32)+rax],xmm10 |
| 1623 vmovdqa XMMWORD[(48-32)+rax],xmm9 |
| 1624 vmovdqa XMMWORD[(64-32)+rax],xmm8 |
| 1625 vmovdqa XMMWORD[(80-32)+rax],xmm7 |
| 1626 |
| 1627 vmovdqu xmm4,XMMWORD[((0-32))+rcx] |
| 1628 vpclmulqdq xmm1,xmm6,xmm4,0x10 |
| 1629 vpclmulqdq xmm2,xmm6,xmm4,0x11 |
| 1630 vpclmulqdq xmm3,xmm6,xmm4,0x00 |
| 1631 vpclmulqdq xmm4,xmm6,xmm4,0x01 |
| 1632 vpxor xmm1,xmm1,xmm4 |
| 1633 |
| 1634 vmovdqu xmm6,XMMWORD[((-16))+rax] |
| 1635 vmovdqu xmm13,XMMWORD[((-16))+rcx] |
| 1636 |
| 1637 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 1638 vpxor xmm1,xmm1,xmm4 |
| 1639 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 1640 vpxor xmm2,xmm2,xmm4 |
| 1641 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 1642 vpxor xmm3,xmm3,xmm4 |
| 1643 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 1644 vpxor xmm1,xmm1,xmm4 |
| 1645 |
| 1646 vmovdqu xmm6,XMMWORD[rax] |
| 1647 vmovdqu xmm13,XMMWORD[rcx] |
| 1648 |
| 1649 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 1650 vpxor xmm1,xmm1,xmm4 |
| 1651 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 1652 vpxor xmm2,xmm2,xmm4 |
| 1653 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 1654 vpxor xmm3,xmm3,xmm4 |
| 1655 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 1656 vpxor xmm1,xmm1,xmm4 |
| 1657 |
| 1658 vmovdqu xmm6,XMMWORD[16+rax] |
| 1659 vmovdqu xmm13,XMMWORD[16+rcx] |
| 1660 |
| 1661 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 1662 vpxor xmm1,xmm1,xmm4 |
| 1663 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 1664 vpxor xmm2,xmm2,xmm4 |
| 1665 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 1666 vpxor xmm3,xmm3,xmm4 |
| 1667 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 1668 vpxor xmm1,xmm1,xmm4 |
| 1669 |
| 1670 vmovdqu xmm6,XMMWORD[32+rax] |
| 1671 vmovdqu xmm13,XMMWORD[32+rcx] |
| 1672 |
| 1673 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 1674 vpxor xmm1,xmm1,xmm4 |
| 1675 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 1676 vpxor xmm2,xmm2,xmm4 |
| 1677 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 1678 vpxor xmm3,xmm3,xmm4 |
| 1679 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 1680 vpxor xmm1,xmm1,xmm4 |
| 1681 |
| 1682 |
| 1683 vmovdqu xmm6,XMMWORD[((80-32))+rax] |
| 1684 vpxor xmm6,xmm6,xmm0 |
| 1685 vmovdqu xmm5,XMMWORD[((80-32))+rcx] |
| 1686 vpclmulqdq xmm4,xmm6,xmm5,0x11 |
| 1687 vpxor xmm2,xmm2,xmm4 |
| 1688 vpclmulqdq xmm4,xmm6,xmm5,0x00 |
| 1689 vpxor xmm3,xmm3,xmm4 |
| 1690 vpclmulqdq xmm4,xmm6,xmm5,0x10 |
| 1691 vpxor xmm1,xmm1,xmm4 |
| 1692 vpclmulqdq xmm4,xmm6,xmm5,0x01 |
| 1693 vpxor xmm1,xmm1,xmm4 |
| 1694 |
| 1695 vpsrldq xmm4,xmm1,8 |
| 1696 vpxor xmm5,xmm2,xmm4 |
| 1697 vpslldq xmm4,xmm1,8 |
| 1698 vpxor xmm0,xmm3,xmm4 |
| 1699 |
| 1700 vmovdqa xmm3,XMMWORD[poly] |
| 1701 |
| 1702 vpalignr xmm2,xmm0,xmm0,8 |
| 1703 vpclmulqdq xmm0,xmm0,xmm3,0x10 |
| 1704 vpxor xmm0,xmm2,xmm0 |
| 1705 |
| 1706 vpalignr xmm2,xmm0,xmm0,8 |
| 1707 vpclmulqdq xmm0,xmm0,xmm3,0x10 |
| 1708 vpxor xmm0,xmm2,xmm0 |
| 1709 |
| 1710 vpxor xmm0,xmm0,xmm5 |
| 1711 |
| 1712 $L$128_dec_loop2: |
| 1713 |
| 1714 |
| 1715 |
| 1716 cmp r9,16 |
| 1717 jb NEAR $L$128_dec_out |
| 1718 sub r9,16 |
| 1719 |
| 1720 vmovdqa xmm2,xmm15 |
| 1721 vpaddd xmm15,xmm15,XMMWORD[one] |
| 1722 |
| 1723 vpxor xmm2,xmm2,XMMWORD[r8] |
| 1724 vaesenc xmm2,xmm2,XMMWORD[16+r8] |
| 1725 vaesenc xmm2,xmm2,XMMWORD[32+r8] |
| 1726 vaesenc xmm2,xmm2,XMMWORD[48+r8] |
| 1727 vaesenc xmm2,xmm2,XMMWORD[64+r8] |
| 1728 vaesenc xmm2,xmm2,XMMWORD[80+r8] |
| 1729 vaesenc xmm2,xmm2,XMMWORD[96+r8] |
| 1730 vaesenc xmm2,xmm2,XMMWORD[112+r8] |
| 1731 vaesenc xmm2,xmm2,XMMWORD[128+r8] |
| 1732 vaesenc xmm2,xmm2,XMMWORD[144+r8] |
| 1733 vaesenclast xmm2,xmm2,XMMWORD[160+r8] |
| 1734 vpxor xmm2,xmm2,XMMWORD[rdi] |
| 1735 vmovdqu XMMWORD[rsi],xmm2 |
| 1736 add rdi,16 |
| 1737 add rsi,16 |
| 1738 |
| 1739 vpxor xmm0,xmm0,xmm2 |
| 1740 vmovdqa xmm1,XMMWORD[((-32))+rcx] |
| 1741 call GFMUL |
| 1742 |
| 1743 jmp NEAR $L$128_dec_loop2 |
| 1744 |
| 1745 $L$128_dec_out: |
| 1746 vmovdqu XMMWORD[rdx],xmm0 |
| 1747 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 1748 mov rsi,QWORD[16+rsp] |
| 1749 DB 0F3h,0C3h ;repret |
| 1750 |
| 1751 $L$SEH_end_aes128gcmsiv_dec: |
| 1752 global aes128gcmsiv_ecb_enc_block |
| 1753 |
| 1754 ALIGN 16 |
| 1755 aes128gcmsiv_ecb_enc_block: |
| 1756 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 1757 mov QWORD[16+rsp],rsi |
| 1758 mov rax,rsp |
| 1759 $L$SEH_begin_aes128gcmsiv_ecb_enc_block: |
| 1760 mov rdi,rcx |
| 1761 mov rsi,rdx |
| 1762 mov rdx,r8 |
| 1763 |
| 1764 |
| 1765 |
| 1766 vmovdqa xmm1,XMMWORD[rdi] |
| 1767 |
| 1768 vpxor xmm1,xmm1,XMMWORD[rdx] |
| 1769 vaesenc xmm1,xmm1,XMMWORD[16+rdx] |
| 1770 vaesenc xmm1,xmm1,XMMWORD[32+rdx] |
| 1771 vaesenc xmm1,xmm1,XMMWORD[48+rdx] |
| 1772 vaesenc xmm1,xmm1,XMMWORD[64+rdx] |
| 1773 vaesenc xmm1,xmm1,XMMWORD[80+rdx] |
| 1774 vaesenc xmm1,xmm1,XMMWORD[96+rdx] |
| 1775 vaesenc xmm1,xmm1,XMMWORD[112+rdx] |
| 1776 vaesenc xmm1,xmm1,XMMWORD[128+rdx] |
| 1777 vaesenc xmm1,xmm1,XMMWORD[144+rdx] |
| 1778 vaesenclast xmm1,xmm1,XMMWORD[160+rdx] |
| 1779 |
| 1780 vmovdqa XMMWORD[rsi],xmm1 |
| 1781 |
| 1782 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 1783 mov rsi,QWORD[16+rsp] |
| 1784 DB 0F3h,0C3h ;repret |
| 1785 |
| 1786 $L$SEH_end_aes128gcmsiv_ecb_enc_block: |
| 1787 global aes256gcmsiv_aes_ks_enc_x1 |
| 1788 |
| 1789 ALIGN 16 |
| 1790 aes256gcmsiv_aes_ks_enc_x1: |
| 1791 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 1792 mov QWORD[16+rsp],rsi |
| 1793 mov rax,rsp |
| 1794 $L$SEH_begin_aes256gcmsiv_aes_ks_enc_x1: |
| 1795 mov rdi,rcx |
| 1796 mov rsi,rdx |
| 1797 mov rdx,r8 |
| 1798 mov rcx,r9 |
| 1799 |
| 1800 |
| 1801 |
| 1802 vmovdqa xmm0,XMMWORD[con1] |
| 1803 vmovdqa xmm15,XMMWORD[mask] |
| 1804 vmovdqa xmm8,XMMWORD[rdi] |
| 1805 vmovdqa xmm1,XMMWORD[rcx] |
| 1806 vmovdqa xmm3,XMMWORD[16+rcx] |
| 1807 vpxor xmm8,xmm8,xmm1 |
| 1808 vaesenc xmm8,xmm8,xmm3 |
| 1809 vmovdqu XMMWORD[rdx],xmm1 |
| 1810 vmovdqu XMMWORD[16+rdx],xmm3 |
| 1811 vpxor xmm14,xmm14,xmm14 |
| 1812 |
| 1813 vpshufb xmm2,xmm3,xmm15 |
| 1814 vaesenclast xmm2,xmm2,xmm0 |
| 1815 vpslld xmm0,xmm0,1 |
| 1816 vpslldq xmm4,xmm1,4 |
| 1817 vpxor xmm1,xmm1,xmm4 |
| 1818 vpslldq xmm4,xmm4,4 |
| 1819 vpxor xmm1,xmm1,xmm4 |
| 1820 vpslldq xmm4,xmm4,4 |
| 1821 vpxor xmm1,xmm1,xmm4 |
| 1822 vpxor xmm1,xmm1,xmm2 |
| 1823 vaesenc xmm8,xmm8,xmm1 |
| 1824 vmovdqu XMMWORD[32+rdx],xmm1 |
| 1825 |
| 1826 vpshufd xmm2,xmm1,0xff |
| 1827 vaesenclast xmm2,xmm2,xmm14 |
| 1828 vpslldq xmm4,xmm3,4 |
| 1829 vpxor xmm3,xmm3,xmm4 |
| 1830 vpslldq xmm4,xmm4,4 |
| 1831 vpxor xmm3,xmm3,xmm4 |
| 1832 vpslldq xmm4,xmm4,4 |
| 1833 vpxor xmm3,xmm3,xmm4 |
| 1834 vpxor xmm3,xmm3,xmm2 |
| 1835 vaesenc xmm8,xmm8,xmm3 |
| 1836 vmovdqu XMMWORD[48+rdx],xmm3 |
| 1837 |
| 1838 vpshufb xmm2,xmm3,xmm15 |
| 1839 vaesenclast xmm2,xmm2,xmm0 |
| 1840 vpslld xmm0,xmm0,1 |
| 1841 vpslldq xmm4,xmm1,4 |
| 1842 vpxor xmm1,xmm1,xmm4 |
| 1843 vpslldq xmm4,xmm4,4 |
| 1844 vpxor xmm1,xmm1,xmm4 |
| 1845 vpslldq xmm4,xmm4,4 |
| 1846 vpxor xmm1,xmm1,xmm4 |
| 1847 vpxor xmm1,xmm1,xmm2 |
| 1848 vaesenc xmm8,xmm8,xmm1 |
| 1849 vmovdqu XMMWORD[64+rdx],xmm1 |
| 1850 |
| 1851 vpshufd xmm2,xmm1,0xff |
| 1852 vaesenclast xmm2,xmm2,xmm14 |
| 1853 vpslldq xmm4,xmm3,4 |
| 1854 vpxor xmm3,xmm3,xmm4 |
| 1855 vpslldq xmm4,xmm4,4 |
| 1856 vpxor xmm3,xmm3,xmm4 |
| 1857 vpslldq xmm4,xmm4,4 |
| 1858 vpxor xmm3,xmm3,xmm4 |
| 1859 vpxor xmm3,xmm3,xmm2 |
| 1860 vaesenc xmm8,xmm8,xmm3 |
| 1861 vmovdqu XMMWORD[80+rdx],xmm3 |
| 1862 |
| 1863 vpshufb xmm2,xmm3,xmm15 |
| 1864 vaesenclast xmm2,xmm2,xmm0 |
| 1865 vpslld xmm0,xmm0,1 |
| 1866 vpslldq xmm4,xmm1,4 |
| 1867 vpxor xmm1,xmm1,xmm4 |
| 1868 vpslldq xmm4,xmm4,4 |
| 1869 vpxor xmm1,xmm1,xmm4 |
| 1870 vpslldq xmm4,xmm4,4 |
| 1871 vpxor xmm1,xmm1,xmm4 |
| 1872 vpxor xmm1,xmm1,xmm2 |
| 1873 vaesenc xmm8,xmm8,xmm1 |
| 1874 vmovdqu XMMWORD[96+rdx],xmm1 |
| 1875 |
| 1876 vpshufd xmm2,xmm1,0xff |
| 1877 vaesenclast xmm2,xmm2,xmm14 |
| 1878 vpslldq xmm4,xmm3,4 |
| 1879 vpxor xmm3,xmm3,xmm4 |
| 1880 vpslldq xmm4,xmm4,4 |
| 1881 vpxor xmm3,xmm3,xmm4 |
| 1882 vpslldq xmm4,xmm4,4 |
| 1883 vpxor xmm3,xmm3,xmm4 |
| 1884 vpxor xmm3,xmm3,xmm2 |
| 1885 vaesenc xmm8,xmm8,xmm3 |
| 1886 vmovdqu XMMWORD[112+rdx],xmm3 |
| 1887 |
| 1888 vpshufb xmm2,xmm3,xmm15 |
| 1889 vaesenclast xmm2,xmm2,xmm0 |
| 1890 vpslld xmm0,xmm0,1 |
| 1891 vpslldq xmm4,xmm1,4 |
| 1892 vpxor xmm1,xmm1,xmm4 |
| 1893 vpslldq xmm4,xmm4,4 |
| 1894 vpxor xmm1,xmm1,xmm4 |
| 1895 vpslldq xmm4,xmm4,4 |
| 1896 vpxor xmm1,xmm1,xmm4 |
| 1897 vpxor xmm1,xmm1,xmm2 |
| 1898 vaesenc xmm8,xmm8,xmm1 |
| 1899 vmovdqu XMMWORD[128+rdx],xmm1 |
| 1900 |
| 1901 vpshufd xmm2,xmm1,0xff |
| 1902 vaesenclast xmm2,xmm2,xmm14 |
| 1903 vpslldq xmm4,xmm3,4 |
| 1904 vpxor xmm3,xmm3,xmm4 |
| 1905 vpslldq xmm4,xmm4,4 |
| 1906 vpxor xmm3,xmm3,xmm4 |
| 1907 vpslldq xmm4,xmm4,4 |
| 1908 vpxor xmm3,xmm3,xmm4 |
| 1909 vpxor xmm3,xmm3,xmm2 |
| 1910 vaesenc xmm8,xmm8,xmm3 |
| 1911 vmovdqu XMMWORD[144+rdx],xmm3 |
| 1912 |
| 1913 vpshufb xmm2,xmm3,xmm15 |
| 1914 vaesenclast xmm2,xmm2,xmm0 |
| 1915 vpslld xmm0,xmm0,1 |
| 1916 vpslldq xmm4,xmm1,4 |
| 1917 vpxor xmm1,xmm1,xmm4 |
| 1918 vpslldq xmm4,xmm4,4 |
| 1919 vpxor xmm1,xmm1,xmm4 |
| 1920 vpslldq xmm4,xmm4,4 |
| 1921 vpxor xmm1,xmm1,xmm4 |
| 1922 vpxor xmm1,xmm1,xmm2 |
| 1923 vaesenc xmm8,xmm8,xmm1 |
| 1924 vmovdqu XMMWORD[160+rdx],xmm1 |
| 1925 |
| 1926 vpshufd xmm2,xmm1,0xff |
| 1927 vaesenclast xmm2,xmm2,xmm14 |
| 1928 vpslldq xmm4,xmm3,4 |
| 1929 vpxor xmm3,xmm3,xmm4 |
| 1930 vpslldq xmm4,xmm4,4 |
| 1931 vpxor xmm3,xmm3,xmm4 |
| 1932 vpslldq xmm4,xmm4,4 |
| 1933 vpxor xmm3,xmm3,xmm4 |
| 1934 vpxor xmm3,xmm3,xmm2 |
| 1935 vaesenc xmm8,xmm8,xmm3 |
| 1936 vmovdqu XMMWORD[176+rdx],xmm3 |
| 1937 |
| 1938 vpshufb xmm2,xmm3,xmm15 |
| 1939 vaesenclast xmm2,xmm2,xmm0 |
| 1940 vpslld xmm0,xmm0,1 |
| 1941 vpslldq xmm4,xmm1,4 |
| 1942 vpxor xmm1,xmm1,xmm4 |
| 1943 vpslldq xmm4,xmm4,4 |
| 1944 vpxor xmm1,xmm1,xmm4 |
| 1945 vpslldq xmm4,xmm4,4 |
| 1946 vpxor xmm1,xmm1,xmm4 |
| 1947 vpxor xmm1,xmm1,xmm2 |
| 1948 vaesenc xmm8,xmm8,xmm1 |
| 1949 vmovdqu XMMWORD[192+rdx],xmm1 |
| 1950 |
| 1951 vpshufd xmm2,xmm1,0xff |
| 1952 vaesenclast xmm2,xmm2,xmm14 |
| 1953 vpslldq xmm4,xmm3,4 |
| 1954 vpxor xmm3,xmm3,xmm4 |
| 1955 vpslldq xmm4,xmm4,4 |
| 1956 vpxor xmm3,xmm3,xmm4 |
| 1957 vpslldq xmm4,xmm4,4 |
| 1958 vpxor xmm3,xmm3,xmm4 |
| 1959 vpxor xmm3,xmm3,xmm2 |
| 1960 vaesenc xmm8,xmm8,xmm3 |
| 1961 vmovdqu XMMWORD[208+rdx],xmm3 |
| 1962 |
| 1963 vpshufb xmm2,xmm3,xmm15 |
| 1964 vaesenclast xmm2,xmm2,xmm0 |
| 1965 vpslldq xmm4,xmm1,4 |
| 1966 vpxor xmm1,xmm1,xmm4 |
| 1967 vpslldq xmm4,xmm4,4 |
| 1968 vpxor xmm1,xmm1,xmm4 |
| 1969 vpslldq xmm4,xmm4,4 |
| 1970 vpxor xmm1,xmm1,xmm4 |
| 1971 vpxor xmm1,xmm1,xmm2 |
| 1972 vaesenclast xmm8,xmm8,xmm1 |
| 1973 vmovdqu XMMWORD[224+rdx],xmm1 |
| 1974 |
| 1975 vmovdqa XMMWORD[rsi],xmm8 |
| 1976 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 1977 mov rsi,QWORD[16+rsp] |
| 1978 DB 0F3h,0C3h ;repret |
| 1979 |
| 1980 $L$SEH_end_aes256gcmsiv_aes_ks_enc_x1: |
| 1981 global aes256gcmsiv_ecb_enc_block |
| 1982 |
| 1983 ALIGN 16 |
| 1984 aes256gcmsiv_ecb_enc_block: |
| 1985 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 1986 mov QWORD[16+rsp],rsi |
| 1987 mov rax,rsp |
| 1988 $L$SEH_begin_aes256gcmsiv_ecb_enc_block: |
| 1989 mov rdi,rcx |
| 1990 mov rsi,rdx |
| 1991 mov rdx,r8 |
| 1992 |
| 1993 |
| 1994 |
| 1995 vmovdqa xmm1,XMMWORD[rdi] |
| 1996 vpxor xmm1,xmm1,XMMWORD[rdx] |
| 1997 vaesenc xmm1,xmm1,XMMWORD[16+rdx] |
| 1998 vaesenc xmm1,xmm1,XMMWORD[32+rdx] |
| 1999 vaesenc xmm1,xmm1,XMMWORD[48+rdx] |
| 2000 vaesenc xmm1,xmm1,XMMWORD[64+rdx] |
| 2001 vaesenc xmm1,xmm1,XMMWORD[80+rdx] |
| 2002 vaesenc xmm1,xmm1,XMMWORD[96+rdx] |
| 2003 vaesenc xmm1,xmm1,XMMWORD[112+rdx] |
| 2004 vaesenc xmm1,xmm1,XMMWORD[128+rdx] |
| 2005 vaesenc xmm1,xmm1,XMMWORD[144+rdx] |
| 2006 vaesenc xmm1,xmm1,XMMWORD[160+rdx] |
| 2007 vaesenc xmm1,xmm1,XMMWORD[176+rdx] |
| 2008 vaesenc xmm1,xmm1,XMMWORD[192+rdx] |
| 2009 vaesenc xmm1,xmm1,XMMWORD[208+rdx] |
| 2010 vaesenclast xmm1,xmm1,XMMWORD[224+rdx] |
| 2011 vmovdqa XMMWORD[rsi],xmm1 |
| 2012 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 2013 mov rsi,QWORD[16+rsp] |
| 2014 DB 0F3h,0C3h ;repret |
| 2015 |
| 2016 $L$SEH_end_aes256gcmsiv_ecb_enc_block: |
| 2017 global aes256gcmsiv_enc_msg_x4 |
| 2018 |
| 2019 ALIGN 16 |
| 2020 aes256gcmsiv_enc_msg_x4: |
| 2021 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 2022 mov QWORD[16+rsp],rsi |
| 2023 mov rax,rsp |
| 2024 $L$SEH_begin_aes256gcmsiv_enc_msg_x4: |
| 2025 mov rdi,rcx |
| 2026 mov rsi,rdx |
| 2027 mov rdx,r8 |
| 2028 mov rcx,r9 |
| 2029 mov r8,QWORD[40+rsp] |
| 2030 |
| 2031 |
| 2032 |
| 2033 test r8,r8 |
| 2034 jnz NEAR $L$256_enc_msg_x4_start |
| 2035 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 2036 mov rsi,QWORD[16+rsp] |
| 2037 DB 0F3h,0C3h ;repret |
| 2038 |
| 2039 $L$256_enc_msg_x4_start: |
| 2040 mov r10,r8 |
| 2041 shr r8,4 |
| 2042 shl r10,60 |
| 2043 jz NEAR $L$256_enc_msg_x4_start2 |
| 2044 add r8,1 |
| 2045 |
| 2046 $L$256_enc_msg_x4_start2: |
| 2047 mov r10,r8 |
| 2048 shl r10,62 |
| 2049 shr r10,62 |
| 2050 |
| 2051 |
| 2052 vmovdqa xmm15,XMMWORD[rdx] |
| 2053 vpor xmm15,xmm15,XMMWORD[OR_MASK] |
| 2054 |
| 2055 vmovdqa xmm4,XMMWORD[four] |
| 2056 vmovdqa xmm0,xmm15 |
| 2057 vpaddd xmm1,xmm15,XMMWORD[one] |
| 2058 vpaddd xmm2,xmm15,XMMWORD[two] |
| 2059 vpaddd xmm3,xmm15,XMMWORD[three] |
| 2060 |
| 2061 shr r8,2 |
| 2062 je NEAR $L$256_enc_msg_x4_check_remainder |
| 2063 |
| 2064 sub rsi,64 |
| 2065 sub rdi,64 |
| 2066 |
| 2067 $L$256_enc_msg_x4_loop1: |
| 2068 add rsi,64 |
| 2069 add rdi,64 |
| 2070 |
| 2071 vmovdqa xmm5,xmm0 |
| 2072 vmovdqa xmm6,xmm1 |
| 2073 vmovdqa xmm7,xmm2 |
| 2074 vmovdqa xmm8,xmm3 |
| 2075 |
| 2076 vpxor xmm5,xmm5,XMMWORD[rcx] |
| 2077 vpxor xmm6,xmm6,XMMWORD[rcx] |
| 2078 vpxor xmm7,xmm7,XMMWORD[rcx] |
| 2079 vpxor xmm8,xmm8,XMMWORD[rcx] |
| 2080 |
| 2081 vmovdqu xmm12,XMMWORD[16+rcx] |
| 2082 vaesenc xmm5,xmm5,xmm12 |
| 2083 vaesenc xmm6,xmm6,xmm12 |
| 2084 vaesenc xmm7,xmm7,xmm12 |
| 2085 vaesenc xmm8,xmm8,xmm12 |
| 2086 |
| 2087 vpaddd xmm0,xmm0,xmm4 |
| 2088 vmovdqu xmm12,XMMWORD[32+rcx] |
| 2089 vaesenc xmm5,xmm5,xmm12 |
| 2090 vaesenc xmm6,xmm6,xmm12 |
| 2091 vaesenc xmm7,xmm7,xmm12 |
| 2092 vaesenc xmm8,xmm8,xmm12 |
| 2093 |
| 2094 vpaddd xmm1,xmm1,xmm4 |
| 2095 vmovdqu xmm12,XMMWORD[48+rcx] |
| 2096 vaesenc xmm5,xmm5,xmm12 |
| 2097 vaesenc xmm6,xmm6,xmm12 |
| 2098 vaesenc xmm7,xmm7,xmm12 |
| 2099 vaesenc xmm8,xmm8,xmm12 |
| 2100 |
| 2101 vpaddd xmm2,xmm2,xmm4 |
| 2102 vmovdqu xmm12,XMMWORD[64+rcx] |
| 2103 vaesenc xmm5,xmm5,xmm12 |
| 2104 vaesenc xmm6,xmm6,xmm12 |
| 2105 vaesenc xmm7,xmm7,xmm12 |
| 2106 vaesenc xmm8,xmm8,xmm12 |
| 2107 |
| 2108 vpaddd xmm3,xmm3,xmm4 |
| 2109 |
| 2110 vmovdqu xmm12,XMMWORD[80+rcx] |
| 2111 vaesenc xmm5,xmm5,xmm12 |
| 2112 vaesenc xmm6,xmm6,xmm12 |
| 2113 vaesenc xmm7,xmm7,xmm12 |
| 2114 vaesenc xmm8,xmm8,xmm12 |
| 2115 |
| 2116 vmovdqu xmm12,XMMWORD[96+rcx] |
| 2117 vaesenc xmm5,xmm5,xmm12 |
| 2118 vaesenc xmm6,xmm6,xmm12 |
| 2119 vaesenc xmm7,xmm7,xmm12 |
| 2120 vaesenc xmm8,xmm8,xmm12 |
| 2121 |
| 2122 vmovdqu xmm12,XMMWORD[112+rcx] |
| 2123 vaesenc xmm5,xmm5,xmm12 |
| 2124 vaesenc xmm6,xmm6,xmm12 |
| 2125 vaesenc xmm7,xmm7,xmm12 |
| 2126 vaesenc xmm8,xmm8,xmm12 |
| 2127 |
| 2128 vmovdqu xmm12,XMMWORD[128+rcx] |
| 2129 vaesenc xmm5,xmm5,xmm12 |
| 2130 vaesenc xmm6,xmm6,xmm12 |
| 2131 vaesenc xmm7,xmm7,xmm12 |
| 2132 vaesenc xmm8,xmm8,xmm12 |
| 2133 |
| 2134 vmovdqu xmm12,XMMWORD[144+rcx] |
| 2135 vaesenc xmm5,xmm5,xmm12 |
| 2136 vaesenc xmm6,xmm6,xmm12 |
| 2137 vaesenc xmm7,xmm7,xmm12 |
| 2138 vaesenc xmm8,xmm8,xmm12 |
| 2139 |
| 2140 vmovdqu xmm12,XMMWORD[160+rcx] |
| 2141 vaesenc xmm5,xmm5,xmm12 |
| 2142 vaesenc xmm6,xmm6,xmm12 |
| 2143 vaesenc xmm7,xmm7,xmm12 |
| 2144 vaesenc xmm8,xmm8,xmm12 |
| 2145 |
| 2146 vmovdqu xmm12,XMMWORD[176+rcx] |
| 2147 vaesenc xmm5,xmm5,xmm12 |
| 2148 vaesenc xmm6,xmm6,xmm12 |
| 2149 vaesenc xmm7,xmm7,xmm12 |
| 2150 vaesenc xmm8,xmm8,xmm12 |
| 2151 |
| 2152 vmovdqu xmm12,XMMWORD[192+rcx] |
| 2153 vaesenc xmm5,xmm5,xmm12 |
| 2154 vaesenc xmm6,xmm6,xmm12 |
| 2155 vaesenc xmm7,xmm7,xmm12 |
| 2156 vaesenc xmm8,xmm8,xmm12 |
| 2157 |
| 2158 vmovdqu xmm12,XMMWORD[208+rcx] |
| 2159 vaesenc xmm5,xmm5,xmm12 |
| 2160 vaesenc xmm6,xmm6,xmm12 |
| 2161 vaesenc xmm7,xmm7,xmm12 |
| 2162 vaesenc xmm8,xmm8,xmm12 |
| 2163 |
| 2164 vmovdqu xmm12,XMMWORD[224+rcx] |
| 2165 vaesenclast xmm5,xmm5,xmm12 |
| 2166 vaesenclast xmm6,xmm6,xmm12 |
| 2167 vaesenclast xmm7,xmm7,xmm12 |
| 2168 vaesenclast xmm8,xmm8,xmm12 |
| 2169 |
| 2170 |
| 2171 |
| 2172 vpxor xmm5,xmm5,XMMWORD[rdi] |
| 2173 vpxor xmm6,xmm6,XMMWORD[16+rdi] |
| 2174 vpxor xmm7,xmm7,XMMWORD[32+rdi] |
| 2175 vpxor xmm8,xmm8,XMMWORD[48+rdi] |
| 2176 |
| 2177 sub r8,1 |
| 2178 |
| 2179 vmovdqu XMMWORD[rsi],xmm5 |
| 2180 vmovdqu XMMWORD[16+rsi],xmm6 |
| 2181 vmovdqu XMMWORD[32+rsi],xmm7 |
| 2182 vmovdqu XMMWORD[48+rsi],xmm8 |
| 2183 |
| 2184 jne NEAR $L$256_enc_msg_x4_loop1 |
| 2185 |
| 2186 add rsi,64 |
| 2187 add rdi,64 |
| 2188 |
| 2189 $L$256_enc_msg_x4_check_remainder: |
| 2190 cmp r10,0 |
| 2191 je NEAR $L$256_enc_msg_x4_out |
| 2192 |
| 2193 $L$256_enc_msg_x4_loop2: |
| 2194 |
| 2195 |
| 2196 |
| 2197 vmovdqa xmm5,xmm0 |
| 2198 vpaddd xmm0,xmm0,XMMWORD[one] |
| 2199 vpxor xmm5,xmm5,XMMWORD[rcx] |
| 2200 vaesenc xmm5,xmm5,XMMWORD[16+rcx] |
| 2201 vaesenc xmm5,xmm5,XMMWORD[32+rcx] |
| 2202 vaesenc xmm5,xmm5,XMMWORD[48+rcx] |
| 2203 vaesenc xmm5,xmm5,XMMWORD[64+rcx] |
| 2204 vaesenc xmm5,xmm5,XMMWORD[80+rcx] |
| 2205 vaesenc xmm5,xmm5,XMMWORD[96+rcx] |
| 2206 vaesenc xmm5,xmm5,XMMWORD[112+rcx] |
| 2207 vaesenc xmm5,xmm5,XMMWORD[128+rcx] |
| 2208 vaesenc xmm5,xmm5,XMMWORD[144+rcx] |
| 2209 vaesenc xmm5,xmm5,XMMWORD[160+rcx] |
| 2210 vaesenc xmm5,xmm5,XMMWORD[176+rcx] |
| 2211 vaesenc xmm5,xmm5,XMMWORD[192+rcx] |
| 2212 vaesenc xmm5,xmm5,XMMWORD[208+rcx] |
| 2213 vaesenclast xmm5,xmm5,XMMWORD[224+rcx] |
| 2214 |
| 2215 |
| 2216 vpxor xmm5,xmm5,XMMWORD[rdi] |
| 2217 |
| 2218 vmovdqu XMMWORD[rsi],xmm5 |
| 2219 |
| 2220 add rdi,16 |
| 2221 add rsi,16 |
| 2222 |
| 2223 sub r10,1 |
| 2224 jne NEAR $L$256_enc_msg_x4_loop2 |
| 2225 |
| 2226 $L$256_enc_msg_x4_out: |
| 2227 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 2228 mov rsi,QWORD[16+rsp] |
| 2229 DB 0F3h,0C3h ;repret |
| 2230 |
| 2231 $L$SEH_end_aes256gcmsiv_enc_msg_x4: |
| 2232 global aes256gcmsiv_enc_msg_x8 |
| 2233 |
| 2234 ALIGN 16 |
| 2235 aes256gcmsiv_enc_msg_x8: |
| 2236 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 2237 mov QWORD[16+rsp],rsi |
| 2238 mov rax,rsp |
| 2239 $L$SEH_begin_aes256gcmsiv_enc_msg_x8: |
| 2240 mov rdi,rcx |
| 2241 mov rsi,rdx |
| 2242 mov rdx,r8 |
| 2243 mov rcx,r9 |
| 2244 mov r8,QWORD[40+rsp] |
| 2245 |
| 2246 |
| 2247 |
| 2248 test r8,r8 |
| 2249 jnz NEAR $L$256_enc_msg_x8_start |
| 2250 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 2251 mov rsi,QWORD[16+rsp] |
| 2252 DB 0F3h,0C3h ;repret |
| 2253 |
| 2254 $L$256_enc_msg_x8_start: |
| 2255 |
| 2256 mov r11,rsp |
| 2257 sub r11,16 |
| 2258 and r11,-64 |
| 2259 |
| 2260 mov r10,r8 |
| 2261 shr r8,4 |
| 2262 shl r10,60 |
| 2263 jz NEAR $L$256_enc_msg_x8_start2 |
| 2264 add r8,1 |
| 2265 |
| 2266 $L$256_enc_msg_x8_start2: |
| 2267 mov r10,r8 |
| 2268 shl r10,61 |
| 2269 shr r10,61 |
| 2270 |
| 2271 |
| 2272 vmovdqa xmm1,XMMWORD[rdx] |
| 2273 vpor xmm1,xmm1,XMMWORD[OR_MASK] |
| 2274 |
| 2275 |
| 2276 vpaddd xmm0,xmm1,XMMWORD[seven] |
| 2277 vmovdqa XMMWORD[r11],xmm0 |
| 2278 vpaddd xmm9,xmm1,XMMWORD[one] |
| 2279 vpaddd xmm10,xmm1,XMMWORD[two] |
| 2280 vpaddd xmm11,xmm1,XMMWORD[three] |
| 2281 vpaddd xmm12,xmm1,XMMWORD[four] |
| 2282 vpaddd xmm13,xmm1,XMMWORD[five] |
| 2283 vpaddd xmm14,xmm1,XMMWORD[six] |
| 2284 vmovdqa xmm0,xmm1 |
| 2285 |
| 2286 shr r8,3 |
| 2287 jz NEAR $L$256_enc_msg_x8_check_remainder |
| 2288 |
| 2289 sub rsi,128 |
| 2290 sub rdi,128 |
| 2291 |
| 2292 $L$256_enc_msg_x8_loop1: |
| 2293 add rsi,128 |
| 2294 add rdi,128 |
| 2295 |
| 2296 vmovdqa xmm1,xmm0 |
| 2297 vmovdqa xmm2,xmm9 |
| 2298 vmovdqa xmm3,xmm10 |
| 2299 vmovdqa xmm4,xmm11 |
| 2300 vmovdqa xmm5,xmm12 |
| 2301 vmovdqa xmm6,xmm13 |
| 2302 vmovdqa xmm7,xmm14 |
| 2303 |
| 2304 vmovdqa xmm8,XMMWORD[r11] |
| 2305 |
| 2306 vpxor xmm1,xmm1,XMMWORD[rcx] |
| 2307 vpxor xmm2,xmm2,XMMWORD[rcx] |
| 2308 vpxor xmm3,xmm3,XMMWORD[rcx] |
| 2309 vpxor xmm4,xmm4,XMMWORD[rcx] |
| 2310 vpxor xmm5,xmm5,XMMWORD[rcx] |
| 2311 vpxor xmm6,xmm6,XMMWORD[rcx] |
| 2312 vpxor xmm7,xmm7,XMMWORD[rcx] |
| 2313 vpxor xmm8,xmm8,XMMWORD[rcx] |
| 2314 |
| 2315 vmovdqu xmm15,XMMWORD[16+rcx] |
| 2316 vaesenc xmm1,xmm1,xmm15 |
| 2317 vaesenc xmm2,xmm2,xmm15 |
| 2318 vaesenc xmm3,xmm3,xmm15 |
| 2319 vaesenc xmm4,xmm4,xmm15 |
| 2320 vaesenc xmm5,xmm5,xmm15 |
| 2321 vaesenc xmm6,xmm6,xmm15 |
| 2322 vaesenc xmm7,xmm7,xmm15 |
| 2323 vaesenc xmm8,xmm8,xmm15 |
| 2324 |
| 2325 vmovdqa xmm14,XMMWORD[r11] |
| 2326 vpaddd xmm14,xmm14,XMMWORD[eight] |
| 2327 vmovdqa XMMWORD[r11],xmm14 |
| 2328 vmovdqu xmm15,XMMWORD[32+rcx] |
| 2329 vaesenc xmm1,xmm1,xmm15 |
| 2330 vaesenc xmm2,xmm2,xmm15 |
| 2331 vaesenc xmm3,xmm3,xmm15 |
| 2332 vaesenc xmm4,xmm4,xmm15 |
| 2333 vaesenc xmm5,xmm5,xmm15 |
| 2334 vaesenc xmm6,xmm6,xmm15 |
| 2335 vaesenc xmm7,xmm7,xmm15 |
| 2336 vaesenc xmm8,xmm8,xmm15 |
| 2337 |
| 2338 vpsubd xmm14,xmm14,XMMWORD[one] |
| 2339 vmovdqu xmm15,XMMWORD[48+rcx] |
| 2340 vaesenc xmm1,xmm1,xmm15 |
| 2341 vaesenc xmm2,xmm2,xmm15 |
| 2342 vaesenc xmm3,xmm3,xmm15 |
| 2343 vaesenc xmm4,xmm4,xmm15 |
| 2344 vaesenc xmm5,xmm5,xmm15 |
| 2345 vaesenc xmm6,xmm6,xmm15 |
| 2346 vaesenc xmm7,xmm7,xmm15 |
| 2347 vaesenc xmm8,xmm8,xmm15 |
| 2348 |
| 2349 vpaddd xmm0,xmm0,XMMWORD[eight] |
| 2350 vmovdqu xmm15,XMMWORD[64+rcx] |
| 2351 vaesenc xmm1,xmm1,xmm15 |
| 2352 vaesenc xmm2,xmm2,xmm15 |
| 2353 vaesenc xmm3,xmm3,xmm15 |
| 2354 vaesenc xmm4,xmm4,xmm15 |
| 2355 vaesenc xmm5,xmm5,xmm15 |
| 2356 vaesenc xmm6,xmm6,xmm15 |
| 2357 vaesenc xmm7,xmm7,xmm15 |
| 2358 vaesenc xmm8,xmm8,xmm15 |
| 2359 |
| 2360 vpaddd xmm9,xmm9,XMMWORD[eight] |
| 2361 vmovdqu xmm15,XMMWORD[80+rcx] |
| 2362 vaesenc xmm1,xmm1,xmm15 |
| 2363 vaesenc xmm2,xmm2,xmm15 |
| 2364 vaesenc xmm3,xmm3,xmm15 |
| 2365 vaesenc xmm4,xmm4,xmm15 |
| 2366 vaesenc xmm5,xmm5,xmm15 |
| 2367 vaesenc xmm6,xmm6,xmm15 |
| 2368 vaesenc xmm7,xmm7,xmm15 |
| 2369 vaesenc xmm8,xmm8,xmm15 |
| 2370 |
| 2371 vpaddd xmm10,xmm10,XMMWORD[eight] |
| 2372 vmovdqu xmm15,XMMWORD[96+rcx] |
| 2373 vaesenc xmm1,xmm1,xmm15 |
| 2374 vaesenc xmm2,xmm2,xmm15 |
| 2375 vaesenc xmm3,xmm3,xmm15 |
| 2376 vaesenc xmm4,xmm4,xmm15 |
| 2377 vaesenc xmm5,xmm5,xmm15 |
| 2378 vaesenc xmm6,xmm6,xmm15 |
| 2379 vaesenc xmm7,xmm7,xmm15 |
| 2380 vaesenc xmm8,xmm8,xmm15 |
| 2381 |
| 2382 vpaddd xmm11,xmm11,XMMWORD[eight] |
| 2383 vmovdqu xmm15,XMMWORD[112+rcx] |
| 2384 vaesenc xmm1,xmm1,xmm15 |
| 2385 vaesenc xmm2,xmm2,xmm15 |
| 2386 vaesenc xmm3,xmm3,xmm15 |
| 2387 vaesenc xmm4,xmm4,xmm15 |
| 2388 vaesenc xmm5,xmm5,xmm15 |
| 2389 vaesenc xmm6,xmm6,xmm15 |
| 2390 vaesenc xmm7,xmm7,xmm15 |
| 2391 vaesenc xmm8,xmm8,xmm15 |
| 2392 |
| 2393 vpaddd xmm12,xmm12,XMMWORD[eight] |
| 2394 vmovdqu xmm15,XMMWORD[128+rcx] |
| 2395 vaesenc xmm1,xmm1,xmm15 |
| 2396 vaesenc xmm2,xmm2,xmm15 |
| 2397 vaesenc xmm3,xmm3,xmm15 |
| 2398 vaesenc xmm4,xmm4,xmm15 |
| 2399 vaesenc xmm5,xmm5,xmm15 |
| 2400 vaesenc xmm6,xmm6,xmm15 |
| 2401 vaesenc xmm7,xmm7,xmm15 |
| 2402 vaesenc xmm8,xmm8,xmm15 |
| 2403 |
| 2404 vpaddd xmm13,xmm13,XMMWORD[eight] |
| 2405 vmovdqu xmm15,XMMWORD[144+rcx] |
| 2406 vaesenc xmm1,xmm1,xmm15 |
| 2407 vaesenc xmm2,xmm2,xmm15 |
| 2408 vaesenc xmm3,xmm3,xmm15 |
| 2409 vaesenc xmm4,xmm4,xmm15 |
| 2410 vaesenc xmm5,xmm5,xmm15 |
| 2411 vaesenc xmm6,xmm6,xmm15 |
| 2412 vaesenc xmm7,xmm7,xmm15 |
| 2413 vaesenc xmm8,xmm8,xmm15 |
| 2414 |
| 2415 vmovdqu xmm15,XMMWORD[160+rcx] |
| 2416 vaesenc xmm1,xmm1,xmm15 |
| 2417 vaesenc xmm2,xmm2,xmm15 |
| 2418 vaesenc xmm3,xmm3,xmm15 |
| 2419 vaesenc xmm4,xmm4,xmm15 |
| 2420 vaesenc xmm5,xmm5,xmm15 |
| 2421 vaesenc xmm6,xmm6,xmm15 |
| 2422 vaesenc xmm7,xmm7,xmm15 |
| 2423 vaesenc xmm8,xmm8,xmm15 |
| 2424 |
| 2425 vmovdqu xmm15,XMMWORD[176+rcx] |
| 2426 vaesenc xmm1,xmm1,xmm15 |
| 2427 vaesenc xmm2,xmm2,xmm15 |
| 2428 vaesenc xmm3,xmm3,xmm15 |
| 2429 vaesenc xmm4,xmm4,xmm15 |
| 2430 vaesenc xmm5,xmm5,xmm15 |
| 2431 vaesenc xmm6,xmm6,xmm15 |
| 2432 vaesenc xmm7,xmm7,xmm15 |
| 2433 vaesenc xmm8,xmm8,xmm15 |
| 2434 |
| 2435 vmovdqu xmm15,XMMWORD[192+rcx] |
| 2436 vaesenc xmm1,xmm1,xmm15 |
| 2437 vaesenc xmm2,xmm2,xmm15 |
| 2438 vaesenc xmm3,xmm3,xmm15 |
| 2439 vaesenc xmm4,xmm4,xmm15 |
| 2440 vaesenc xmm5,xmm5,xmm15 |
| 2441 vaesenc xmm6,xmm6,xmm15 |
| 2442 vaesenc xmm7,xmm7,xmm15 |
| 2443 vaesenc xmm8,xmm8,xmm15 |
| 2444 |
| 2445 vmovdqu xmm15,XMMWORD[208+rcx] |
| 2446 vaesenc xmm1,xmm1,xmm15 |
| 2447 vaesenc xmm2,xmm2,xmm15 |
| 2448 vaesenc xmm3,xmm3,xmm15 |
| 2449 vaesenc xmm4,xmm4,xmm15 |
| 2450 vaesenc xmm5,xmm5,xmm15 |
| 2451 vaesenc xmm6,xmm6,xmm15 |
| 2452 vaesenc xmm7,xmm7,xmm15 |
| 2453 vaesenc xmm8,xmm8,xmm15 |
| 2454 |
| 2455 vmovdqu xmm15,XMMWORD[224+rcx] |
| 2456 vaesenclast xmm1,xmm1,xmm15 |
| 2457 vaesenclast xmm2,xmm2,xmm15 |
| 2458 vaesenclast xmm3,xmm3,xmm15 |
| 2459 vaesenclast xmm4,xmm4,xmm15 |
| 2460 vaesenclast xmm5,xmm5,xmm15 |
| 2461 vaesenclast xmm6,xmm6,xmm15 |
| 2462 vaesenclast xmm7,xmm7,xmm15 |
| 2463 vaesenclast xmm8,xmm8,xmm15 |
| 2464 |
| 2465 |
| 2466 |
| 2467 vpxor xmm1,xmm1,XMMWORD[rdi] |
| 2468 vpxor xmm2,xmm2,XMMWORD[16+rdi] |
| 2469 vpxor xmm3,xmm3,XMMWORD[32+rdi] |
| 2470 vpxor xmm4,xmm4,XMMWORD[48+rdi] |
| 2471 vpxor xmm5,xmm5,XMMWORD[64+rdi] |
| 2472 vpxor xmm6,xmm6,XMMWORD[80+rdi] |
| 2473 vpxor xmm7,xmm7,XMMWORD[96+rdi] |
| 2474 vpxor xmm8,xmm8,XMMWORD[112+rdi] |
| 2475 |
| 2476 sub r8,1 |
| 2477 |
| 2478 vmovdqu XMMWORD[rsi],xmm1 |
| 2479 vmovdqu XMMWORD[16+rsi],xmm2 |
| 2480 vmovdqu XMMWORD[32+rsi],xmm3 |
| 2481 vmovdqu XMMWORD[48+rsi],xmm4 |
| 2482 vmovdqu XMMWORD[64+rsi],xmm5 |
| 2483 vmovdqu XMMWORD[80+rsi],xmm6 |
| 2484 vmovdqu XMMWORD[96+rsi],xmm7 |
| 2485 vmovdqu XMMWORD[112+rsi],xmm8 |
| 2486 |
| 2487 jne NEAR $L$256_enc_msg_x8_loop1 |
| 2488 |
| 2489 add rsi,128 |
| 2490 add rdi,128 |
| 2491 |
| 2492 $L$256_enc_msg_x8_check_remainder: |
| 2493 cmp r10,0 |
| 2494 je NEAR $L$256_enc_msg_x8_out |
| 2495 |
| 2496 $L$256_enc_msg_x8_loop2: |
| 2497 |
| 2498 |
| 2499 vmovdqa xmm1,xmm0 |
| 2500 vpaddd xmm0,xmm0,XMMWORD[one] |
| 2501 |
| 2502 vpxor xmm1,xmm1,XMMWORD[rcx] |
| 2503 vaesenc xmm1,xmm1,XMMWORD[16+rcx] |
| 2504 vaesenc xmm1,xmm1,XMMWORD[32+rcx] |
| 2505 vaesenc xmm1,xmm1,XMMWORD[48+rcx] |
| 2506 vaesenc xmm1,xmm1,XMMWORD[64+rcx] |
| 2507 vaesenc xmm1,xmm1,XMMWORD[80+rcx] |
| 2508 vaesenc xmm1,xmm1,XMMWORD[96+rcx] |
| 2509 vaesenc xmm1,xmm1,XMMWORD[112+rcx] |
| 2510 vaesenc xmm1,xmm1,XMMWORD[128+rcx] |
| 2511 vaesenc xmm1,xmm1,XMMWORD[144+rcx] |
| 2512 vaesenc xmm1,xmm1,XMMWORD[160+rcx] |
| 2513 vaesenc xmm1,xmm1,XMMWORD[176+rcx] |
| 2514 vaesenc xmm1,xmm1,XMMWORD[192+rcx] |
| 2515 vaesenc xmm1,xmm1,XMMWORD[208+rcx] |
| 2516 vaesenclast xmm1,xmm1,XMMWORD[224+rcx] |
| 2517 |
| 2518 |
| 2519 vpxor xmm1,xmm1,XMMWORD[rdi] |
| 2520 |
| 2521 vmovdqu XMMWORD[rsi],xmm1 |
| 2522 |
| 2523 add rdi,16 |
| 2524 add rsi,16 |
| 2525 sub r10,1 |
| 2526 jnz NEAR $L$256_enc_msg_x8_loop2 |
| 2527 |
| 2528 $L$256_enc_msg_x8_out: |
| 2529 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 2530 mov rsi,QWORD[16+rsp] |
| 2531 DB 0F3h,0C3h ;repret |
| 2532 |
| 2533 |
| 2534 $L$SEH_end_aes256gcmsiv_enc_msg_x8: |
| 2535 global aes256gcmsiv_dec |
| 2536 |
| 2537 ALIGN 16 |
| 2538 aes256gcmsiv_dec: |
| 2539 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 2540 mov QWORD[16+rsp],rsi |
| 2541 mov rax,rsp |
| 2542 $L$SEH_begin_aes256gcmsiv_dec: |
| 2543 mov rdi,rcx |
| 2544 mov rsi,rdx |
| 2545 mov rdx,r8 |
| 2546 mov rcx,r9 |
| 2547 mov r8,QWORD[40+rsp] |
| 2548 mov r9,QWORD[48+rsp] |
| 2549 |
| 2550 |
| 2551 |
| 2552 test r9,~15 |
| 2553 jnz NEAR $L$256_dec_start |
| 2554 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 2555 mov rsi,QWORD[16+rsp] |
| 2556 DB 0F3h,0C3h ;repret |
| 2557 |
| 2558 $L$256_dec_start: |
| 2559 vzeroupper |
| 2560 vmovdqa xmm0,XMMWORD[rdx] |
| 2561 mov rax,rdx |
| 2562 |
| 2563 lea rax,[32+rax] |
| 2564 lea rcx,[32+rcx] |
| 2565 |
| 2566 |
| 2567 vmovdqu xmm15,XMMWORD[r9*1+rdi] |
| 2568 vpor xmm15,xmm15,XMMWORD[OR_MASK] |
| 2569 and r9,~15 |
| 2570 |
| 2571 |
| 2572 cmp r9,96 |
| 2573 jb NEAR $L$256_dec_loop2 |
| 2574 |
| 2575 |
| 2576 sub r9,96 |
| 2577 vmovdqa xmm7,xmm15 |
| 2578 vpaddd xmm8,xmm7,XMMWORD[one] |
| 2579 vpaddd xmm9,xmm7,XMMWORD[two] |
| 2580 vpaddd xmm10,xmm9,XMMWORD[one] |
| 2581 vpaddd xmm11,xmm9,XMMWORD[two] |
| 2582 vpaddd xmm12,xmm11,XMMWORD[one] |
| 2583 vpaddd xmm15,xmm11,XMMWORD[two] |
| 2584 |
| 2585 vpxor xmm7,xmm7,XMMWORD[r8] |
| 2586 vpxor xmm8,xmm8,XMMWORD[r8] |
| 2587 vpxor xmm9,xmm9,XMMWORD[r8] |
| 2588 vpxor xmm10,xmm10,XMMWORD[r8] |
| 2589 vpxor xmm11,xmm11,XMMWORD[r8] |
| 2590 vpxor xmm12,xmm12,XMMWORD[r8] |
| 2591 |
| 2592 vmovdqu xmm4,XMMWORD[16+r8] |
| 2593 vaesenc xmm7,xmm7,xmm4 |
| 2594 vaesenc xmm8,xmm8,xmm4 |
| 2595 vaesenc xmm9,xmm9,xmm4 |
| 2596 vaesenc xmm10,xmm10,xmm4 |
| 2597 vaesenc xmm11,xmm11,xmm4 |
| 2598 vaesenc xmm12,xmm12,xmm4 |
| 2599 |
| 2600 vmovdqu xmm4,XMMWORD[32+r8] |
| 2601 vaesenc xmm7,xmm7,xmm4 |
| 2602 vaesenc xmm8,xmm8,xmm4 |
| 2603 vaesenc xmm9,xmm9,xmm4 |
| 2604 vaesenc xmm10,xmm10,xmm4 |
| 2605 vaesenc xmm11,xmm11,xmm4 |
| 2606 vaesenc xmm12,xmm12,xmm4 |
| 2607 |
| 2608 vmovdqu xmm4,XMMWORD[48+r8] |
| 2609 vaesenc xmm7,xmm7,xmm4 |
| 2610 vaesenc xmm8,xmm8,xmm4 |
| 2611 vaesenc xmm9,xmm9,xmm4 |
| 2612 vaesenc xmm10,xmm10,xmm4 |
| 2613 vaesenc xmm11,xmm11,xmm4 |
| 2614 vaesenc xmm12,xmm12,xmm4 |
| 2615 |
| 2616 vmovdqu xmm4,XMMWORD[64+r8] |
| 2617 vaesenc xmm7,xmm7,xmm4 |
| 2618 vaesenc xmm8,xmm8,xmm4 |
| 2619 vaesenc xmm9,xmm9,xmm4 |
| 2620 vaesenc xmm10,xmm10,xmm4 |
| 2621 vaesenc xmm11,xmm11,xmm4 |
| 2622 vaesenc xmm12,xmm12,xmm4 |
| 2623 |
| 2624 vmovdqu xmm4,XMMWORD[80+r8] |
| 2625 vaesenc xmm7,xmm7,xmm4 |
| 2626 vaesenc xmm8,xmm8,xmm4 |
| 2627 vaesenc xmm9,xmm9,xmm4 |
| 2628 vaesenc xmm10,xmm10,xmm4 |
| 2629 vaesenc xmm11,xmm11,xmm4 |
| 2630 vaesenc xmm12,xmm12,xmm4 |
| 2631 |
| 2632 vmovdqu xmm4,XMMWORD[96+r8] |
| 2633 vaesenc xmm7,xmm7,xmm4 |
| 2634 vaesenc xmm8,xmm8,xmm4 |
| 2635 vaesenc xmm9,xmm9,xmm4 |
| 2636 vaesenc xmm10,xmm10,xmm4 |
| 2637 vaesenc xmm11,xmm11,xmm4 |
| 2638 vaesenc xmm12,xmm12,xmm4 |
| 2639 |
| 2640 vmovdqu xmm4,XMMWORD[112+r8] |
| 2641 vaesenc xmm7,xmm7,xmm4 |
| 2642 vaesenc xmm8,xmm8,xmm4 |
| 2643 vaesenc xmm9,xmm9,xmm4 |
| 2644 vaesenc xmm10,xmm10,xmm4 |
| 2645 vaesenc xmm11,xmm11,xmm4 |
| 2646 vaesenc xmm12,xmm12,xmm4 |
| 2647 |
| 2648 vmovdqu xmm4,XMMWORD[128+r8] |
| 2649 vaesenc xmm7,xmm7,xmm4 |
| 2650 vaesenc xmm8,xmm8,xmm4 |
| 2651 vaesenc xmm9,xmm9,xmm4 |
| 2652 vaesenc xmm10,xmm10,xmm4 |
| 2653 vaesenc xmm11,xmm11,xmm4 |
| 2654 vaesenc xmm12,xmm12,xmm4 |
| 2655 |
| 2656 vmovdqu xmm4,XMMWORD[144+r8] |
| 2657 vaesenc xmm7,xmm7,xmm4 |
| 2658 vaesenc xmm8,xmm8,xmm4 |
| 2659 vaesenc xmm9,xmm9,xmm4 |
| 2660 vaesenc xmm10,xmm10,xmm4 |
| 2661 vaesenc xmm11,xmm11,xmm4 |
| 2662 vaesenc xmm12,xmm12,xmm4 |
| 2663 |
| 2664 vmovdqu xmm4,XMMWORD[160+r8] |
| 2665 vaesenc xmm7,xmm7,xmm4 |
| 2666 vaesenc xmm8,xmm8,xmm4 |
| 2667 vaesenc xmm9,xmm9,xmm4 |
| 2668 vaesenc xmm10,xmm10,xmm4 |
| 2669 vaesenc xmm11,xmm11,xmm4 |
| 2670 vaesenc xmm12,xmm12,xmm4 |
| 2671 |
| 2672 vmovdqu xmm4,XMMWORD[176+r8] |
| 2673 vaesenc xmm7,xmm7,xmm4 |
| 2674 vaesenc xmm8,xmm8,xmm4 |
| 2675 vaesenc xmm9,xmm9,xmm4 |
| 2676 vaesenc xmm10,xmm10,xmm4 |
| 2677 vaesenc xmm11,xmm11,xmm4 |
| 2678 vaesenc xmm12,xmm12,xmm4 |
| 2679 |
| 2680 vmovdqu xmm4,XMMWORD[192+r8] |
| 2681 vaesenc xmm7,xmm7,xmm4 |
| 2682 vaesenc xmm8,xmm8,xmm4 |
| 2683 vaesenc xmm9,xmm9,xmm4 |
| 2684 vaesenc xmm10,xmm10,xmm4 |
| 2685 vaesenc xmm11,xmm11,xmm4 |
| 2686 vaesenc xmm12,xmm12,xmm4 |
| 2687 |
| 2688 vmovdqu xmm4,XMMWORD[208+r8] |
| 2689 vaesenc xmm7,xmm7,xmm4 |
| 2690 vaesenc xmm8,xmm8,xmm4 |
| 2691 vaesenc xmm9,xmm9,xmm4 |
| 2692 vaesenc xmm10,xmm10,xmm4 |
| 2693 vaesenc xmm11,xmm11,xmm4 |
| 2694 vaesenc xmm12,xmm12,xmm4 |
| 2695 |
| 2696 vmovdqu xmm4,XMMWORD[224+r8] |
| 2697 vaesenclast xmm7,xmm7,xmm4 |
| 2698 vaesenclast xmm8,xmm8,xmm4 |
| 2699 vaesenclast xmm9,xmm9,xmm4 |
| 2700 vaesenclast xmm10,xmm10,xmm4 |
| 2701 vaesenclast xmm11,xmm11,xmm4 |
| 2702 vaesenclast xmm12,xmm12,xmm4 |
| 2703 |
| 2704 |
| 2705 vpxor xmm7,xmm7,XMMWORD[rdi] |
| 2706 vpxor xmm8,xmm8,XMMWORD[16+rdi] |
| 2707 vpxor xmm9,xmm9,XMMWORD[32+rdi] |
| 2708 vpxor xmm10,xmm10,XMMWORD[48+rdi] |
| 2709 vpxor xmm11,xmm11,XMMWORD[64+rdi] |
| 2710 vpxor xmm12,xmm12,XMMWORD[80+rdi] |
| 2711 |
| 2712 vmovdqu XMMWORD[rsi],xmm7 |
| 2713 vmovdqu XMMWORD[16+rsi],xmm8 |
| 2714 vmovdqu XMMWORD[32+rsi],xmm9 |
| 2715 vmovdqu XMMWORD[48+rsi],xmm10 |
| 2716 vmovdqu XMMWORD[64+rsi],xmm11 |
| 2717 vmovdqu XMMWORD[80+rsi],xmm12 |
| 2718 |
| 2719 add rdi,96 |
| 2720 add rsi,96 |
| 2721 jmp NEAR $L$256_dec_loop1 |
| 2722 |
| 2723 |
| 2724 ALIGN 64 |
| 2725 $L$256_dec_loop1: |
| 2726 cmp r9,96 |
| 2727 jb NEAR $L$256_dec_finish_96 |
| 2728 sub r9,96 |
| 2729 |
| 2730 vmovdqa xmm6,xmm12 |
| 2731 vmovdqa XMMWORD[(16-32)+rax],xmm11 |
| 2732 vmovdqa XMMWORD[(32-32)+rax],xmm10 |
| 2733 vmovdqa XMMWORD[(48-32)+rax],xmm9 |
| 2734 vmovdqa XMMWORD[(64-32)+rax],xmm8 |
| 2735 vmovdqa XMMWORD[(80-32)+rax],xmm7 |
| 2736 |
| 2737 vmovdqa xmm7,xmm15 |
| 2738 vpaddd xmm8,xmm7,XMMWORD[one] |
| 2739 vpaddd xmm9,xmm7,XMMWORD[two] |
| 2740 vpaddd xmm10,xmm9,XMMWORD[one] |
| 2741 vpaddd xmm11,xmm9,XMMWORD[two] |
| 2742 vpaddd xmm12,xmm11,XMMWORD[one] |
| 2743 vpaddd xmm15,xmm11,XMMWORD[two] |
| 2744 |
| 2745 vmovdqa xmm4,XMMWORD[r8] |
| 2746 vpxor xmm7,xmm7,xmm4 |
| 2747 vpxor xmm8,xmm8,xmm4 |
| 2748 vpxor xmm9,xmm9,xmm4 |
| 2749 vpxor xmm10,xmm10,xmm4 |
| 2750 vpxor xmm11,xmm11,xmm4 |
| 2751 vpxor xmm12,xmm12,xmm4 |
| 2752 |
| 2753 vmovdqu xmm4,XMMWORD[((0-32))+rcx] |
| 2754 vpclmulqdq xmm2,xmm6,xmm4,0x11 |
| 2755 vpclmulqdq xmm3,xmm6,xmm4,0x00 |
| 2756 vpclmulqdq xmm1,xmm6,xmm4,0x01 |
| 2757 vpclmulqdq xmm4,xmm6,xmm4,0x10 |
| 2758 vpxor xmm1,xmm1,xmm4 |
| 2759 |
| 2760 vmovdqu xmm4,XMMWORD[16+r8] |
| 2761 vaesenc xmm7,xmm7,xmm4 |
| 2762 vaesenc xmm8,xmm8,xmm4 |
| 2763 vaesenc xmm9,xmm9,xmm4 |
| 2764 vaesenc xmm10,xmm10,xmm4 |
| 2765 vaesenc xmm11,xmm11,xmm4 |
| 2766 vaesenc xmm12,xmm12,xmm4 |
| 2767 |
| 2768 vmovdqu xmm6,XMMWORD[((-16))+rax] |
| 2769 vmovdqu xmm13,XMMWORD[((-16))+rcx] |
| 2770 |
| 2771 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 2772 vpxor xmm1,xmm1,xmm4 |
| 2773 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 2774 vpxor xmm2,xmm2,xmm4 |
| 2775 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 2776 vpxor xmm3,xmm3,xmm4 |
| 2777 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 2778 vpxor xmm1,xmm1,xmm4 |
| 2779 |
| 2780 |
| 2781 vmovdqu xmm4,XMMWORD[32+r8] |
| 2782 vaesenc xmm7,xmm7,xmm4 |
| 2783 vaesenc xmm8,xmm8,xmm4 |
| 2784 vaesenc xmm9,xmm9,xmm4 |
| 2785 vaesenc xmm10,xmm10,xmm4 |
| 2786 vaesenc xmm11,xmm11,xmm4 |
| 2787 vaesenc xmm12,xmm12,xmm4 |
| 2788 |
| 2789 vmovdqu xmm6,XMMWORD[rax] |
| 2790 vmovdqu xmm13,XMMWORD[rcx] |
| 2791 |
| 2792 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 2793 vpxor xmm1,xmm1,xmm4 |
| 2794 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 2795 vpxor xmm2,xmm2,xmm4 |
| 2796 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 2797 vpxor xmm3,xmm3,xmm4 |
| 2798 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 2799 vpxor xmm1,xmm1,xmm4 |
| 2800 |
| 2801 |
| 2802 vmovdqu xmm4,XMMWORD[48+r8] |
| 2803 vaesenc xmm7,xmm7,xmm4 |
| 2804 vaesenc xmm8,xmm8,xmm4 |
| 2805 vaesenc xmm9,xmm9,xmm4 |
| 2806 vaesenc xmm10,xmm10,xmm4 |
| 2807 vaesenc xmm11,xmm11,xmm4 |
| 2808 vaesenc xmm12,xmm12,xmm4 |
| 2809 |
| 2810 vmovdqu xmm6,XMMWORD[16+rax] |
| 2811 vmovdqu xmm13,XMMWORD[16+rcx] |
| 2812 |
| 2813 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 2814 vpxor xmm1,xmm1,xmm4 |
| 2815 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 2816 vpxor xmm2,xmm2,xmm4 |
| 2817 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 2818 vpxor xmm3,xmm3,xmm4 |
| 2819 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 2820 vpxor xmm1,xmm1,xmm4 |
| 2821 |
| 2822 |
| 2823 vmovdqu xmm4,XMMWORD[64+r8] |
| 2824 vaesenc xmm7,xmm7,xmm4 |
| 2825 vaesenc xmm8,xmm8,xmm4 |
| 2826 vaesenc xmm9,xmm9,xmm4 |
| 2827 vaesenc xmm10,xmm10,xmm4 |
| 2828 vaesenc xmm11,xmm11,xmm4 |
| 2829 vaesenc xmm12,xmm12,xmm4 |
| 2830 |
| 2831 vmovdqu xmm6,XMMWORD[32+rax] |
| 2832 vmovdqu xmm13,XMMWORD[32+rcx] |
| 2833 |
| 2834 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 2835 vpxor xmm1,xmm1,xmm4 |
| 2836 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 2837 vpxor xmm2,xmm2,xmm4 |
| 2838 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 2839 vpxor xmm3,xmm3,xmm4 |
| 2840 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 2841 vpxor xmm1,xmm1,xmm4 |
| 2842 |
| 2843 |
| 2844 vmovdqu xmm4,XMMWORD[80+r8] |
| 2845 vaesenc xmm7,xmm7,xmm4 |
| 2846 vaesenc xmm8,xmm8,xmm4 |
| 2847 vaesenc xmm9,xmm9,xmm4 |
| 2848 vaesenc xmm10,xmm10,xmm4 |
| 2849 vaesenc xmm11,xmm11,xmm4 |
| 2850 vaesenc xmm12,xmm12,xmm4 |
| 2851 |
| 2852 vmovdqu xmm4,XMMWORD[96+r8] |
| 2853 vaesenc xmm7,xmm7,xmm4 |
| 2854 vaesenc xmm8,xmm8,xmm4 |
| 2855 vaesenc xmm9,xmm9,xmm4 |
| 2856 vaesenc xmm10,xmm10,xmm4 |
| 2857 vaesenc xmm11,xmm11,xmm4 |
| 2858 vaesenc xmm12,xmm12,xmm4 |
| 2859 |
| 2860 vmovdqu xmm4,XMMWORD[112+r8] |
| 2861 vaesenc xmm7,xmm7,xmm4 |
| 2862 vaesenc xmm8,xmm8,xmm4 |
| 2863 vaesenc xmm9,xmm9,xmm4 |
| 2864 vaesenc xmm10,xmm10,xmm4 |
| 2865 vaesenc xmm11,xmm11,xmm4 |
| 2866 vaesenc xmm12,xmm12,xmm4 |
| 2867 |
| 2868 |
| 2869 vmovdqa xmm6,XMMWORD[((80-32))+rax] |
| 2870 vpxor xmm6,xmm6,xmm0 |
| 2871 vmovdqu xmm5,XMMWORD[((80-32))+rcx] |
| 2872 |
| 2873 vpclmulqdq xmm4,xmm6,xmm5,0x01 |
| 2874 vpxor xmm1,xmm1,xmm4 |
| 2875 vpclmulqdq xmm4,xmm6,xmm5,0x11 |
| 2876 vpxor xmm2,xmm2,xmm4 |
| 2877 vpclmulqdq xmm4,xmm6,xmm5,0x00 |
| 2878 vpxor xmm3,xmm3,xmm4 |
| 2879 vpclmulqdq xmm4,xmm6,xmm5,0x10 |
| 2880 vpxor xmm1,xmm1,xmm4 |
| 2881 |
| 2882 vmovdqu xmm4,XMMWORD[128+r8] |
| 2883 vaesenc xmm7,xmm7,xmm4 |
| 2884 vaesenc xmm8,xmm8,xmm4 |
| 2885 vaesenc xmm9,xmm9,xmm4 |
| 2886 vaesenc xmm10,xmm10,xmm4 |
| 2887 vaesenc xmm11,xmm11,xmm4 |
| 2888 vaesenc xmm12,xmm12,xmm4 |
| 2889 |
| 2890 |
| 2891 vpsrldq xmm4,xmm1,8 |
| 2892 vpxor xmm5,xmm2,xmm4 |
| 2893 vpslldq xmm4,xmm1,8 |
| 2894 vpxor xmm0,xmm3,xmm4 |
| 2895 |
| 2896 vmovdqa xmm3,XMMWORD[poly] |
| 2897 |
| 2898 vmovdqu xmm4,XMMWORD[144+r8] |
| 2899 vaesenc xmm7,xmm7,xmm4 |
| 2900 vaesenc xmm8,xmm8,xmm4 |
| 2901 vaesenc xmm9,xmm9,xmm4 |
| 2902 vaesenc xmm10,xmm10,xmm4 |
| 2903 vaesenc xmm11,xmm11,xmm4 |
| 2904 vaesenc xmm12,xmm12,xmm4 |
| 2905 |
| 2906 vmovdqu xmm4,XMMWORD[160+r8] |
| 2907 vaesenc xmm7,xmm7,xmm4 |
| 2908 vaesenc xmm8,xmm8,xmm4 |
| 2909 vaesenc xmm9,xmm9,xmm4 |
| 2910 vaesenc xmm10,xmm10,xmm4 |
| 2911 vaesenc xmm11,xmm11,xmm4 |
| 2912 vaesenc xmm12,xmm12,xmm4 |
| 2913 |
| 2914 vmovdqu xmm4,XMMWORD[176+r8] |
| 2915 vaesenc xmm7,xmm7,xmm4 |
| 2916 vaesenc xmm8,xmm8,xmm4 |
| 2917 vaesenc xmm9,xmm9,xmm4 |
| 2918 vaesenc xmm10,xmm10,xmm4 |
| 2919 vaesenc xmm11,xmm11,xmm4 |
| 2920 vaesenc xmm12,xmm12,xmm4 |
| 2921 |
| 2922 vmovdqu xmm4,XMMWORD[192+r8] |
| 2923 vaesenc xmm7,xmm7,xmm4 |
| 2924 vaesenc xmm8,xmm8,xmm4 |
| 2925 vaesenc xmm9,xmm9,xmm4 |
| 2926 vaesenc xmm10,xmm10,xmm4 |
| 2927 vaesenc xmm11,xmm11,xmm4 |
| 2928 vaesenc xmm12,xmm12,xmm4 |
| 2929 |
| 2930 vmovdqu xmm4,XMMWORD[208+r8] |
| 2931 vaesenc xmm7,xmm7,xmm4 |
| 2932 vaesenc xmm8,xmm8,xmm4 |
| 2933 vaesenc xmm9,xmm9,xmm4 |
| 2934 vaesenc xmm10,xmm10,xmm4 |
| 2935 vaesenc xmm11,xmm11,xmm4 |
| 2936 vaesenc xmm12,xmm12,xmm4 |
| 2937 |
| 2938 vmovdqu xmm6,XMMWORD[224+r8] |
| 2939 vpalignr xmm2,xmm0,xmm0,8 |
| 2940 vpclmulqdq xmm0,xmm0,xmm3,0x10 |
| 2941 vpxor xmm0,xmm2,xmm0 |
| 2942 |
| 2943 vpxor xmm4,xmm6,XMMWORD[rdi] |
| 2944 vaesenclast xmm7,xmm7,xmm4 |
| 2945 vpxor xmm4,xmm6,XMMWORD[16+rdi] |
| 2946 vaesenclast xmm8,xmm8,xmm4 |
| 2947 vpxor xmm4,xmm6,XMMWORD[32+rdi] |
| 2948 vaesenclast xmm9,xmm9,xmm4 |
| 2949 vpxor xmm4,xmm6,XMMWORD[48+rdi] |
| 2950 vaesenclast xmm10,xmm10,xmm4 |
| 2951 vpxor xmm4,xmm6,XMMWORD[64+rdi] |
| 2952 vaesenclast xmm11,xmm11,xmm4 |
| 2953 vpxor xmm4,xmm6,XMMWORD[80+rdi] |
| 2954 vaesenclast xmm12,xmm12,xmm4 |
| 2955 |
| 2956 vpalignr xmm2,xmm0,xmm0,8 |
| 2957 vpclmulqdq xmm0,xmm0,xmm3,0x10 |
| 2958 vpxor xmm0,xmm2,xmm0 |
| 2959 |
| 2960 vmovdqu XMMWORD[rsi],xmm7 |
| 2961 vmovdqu XMMWORD[16+rsi],xmm8 |
| 2962 vmovdqu XMMWORD[32+rsi],xmm9 |
| 2963 vmovdqu XMMWORD[48+rsi],xmm10 |
| 2964 vmovdqu XMMWORD[64+rsi],xmm11 |
| 2965 vmovdqu XMMWORD[80+rsi],xmm12 |
| 2966 |
| 2967 vpxor xmm0,xmm0,xmm5 |
| 2968 |
| 2969 lea rdi,[96+rdi] |
| 2970 lea rsi,[96+rsi] |
| 2971 jmp NEAR $L$256_dec_loop1 |
| 2972 |
| 2973 $L$256_dec_finish_96: |
| 2974 vmovdqa xmm6,xmm12 |
| 2975 vmovdqa XMMWORD[(16-32)+rax],xmm11 |
| 2976 vmovdqa XMMWORD[(32-32)+rax],xmm10 |
| 2977 vmovdqa XMMWORD[(48-32)+rax],xmm9 |
| 2978 vmovdqa XMMWORD[(64-32)+rax],xmm8 |
| 2979 vmovdqa XMMWORD[(80-32)+rax],xmm7 |
| 2980 |
| 2981 vmovdqu xmm4,XMMWORD[((0-32))+rcx] |
| 2982 vpclmulqdq xmm1,xmm6,xmm4,0x10 |
| 2983 vpclmulqdq xmm2,xmm6,xmm4,0x11 |
| 2984 vpclmulqdq xmm3,xmm6,xmm4,0x00 |
| 2985 vpclmulqdq xmm4,xmm6,xmm4,0x01 |
| 2986 vpxor xmm1,xmm1,xmm4 |
| 2987 |
| 2988 vmovdqu xmm6,XMMWORD[((-16))+rax] |
| 2989 vmovdqu xmm13,XMMWORD[((-16))+rcx] |
| 2990 |
| 2991 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 2992 vpxor xmm1,xmm1,xmm4 |
| 2993 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 2994 vpxor xmm2,xmm2,xmm4 |
| 2995 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 2996 vpxor xmm3,xmm3,xmm4 |
| 2997 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 2998 vpxor xmm1,xmm1,xmm4 |
| 2999 |
| 3000 vmovdqu xmm6,XMMWORD[rax] |
| 3001 vmovdqu xmm13,XMMWORD[rcx] |
| 3002 |
| 3003 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 3004 vpxor xmm1,xmm1,xmm4 |
| 3005 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 3006 vpxor xmm2,xmm2,xmm4 |
| 3007 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 3008 vpxor xmm3,xmm3,xmm4 |
| 3009 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 3010 vpxor xmm1,xmm1,xmm4 |
| 3011 |
| 3012 vmovdqu xmm6,XMMWORD[16+rax] |
| 3013 vmovdqu xmm13,XMMWORD[16+rcx] |
| 3014 |
| 3015 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 3016 vpxor xmm1,xmm1,xmm4 |
| 3017 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 3018 vpxor xmm2,xmm2,xmm4 |
| 3019 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 3020 vpxor xmm3,xmm3,xmm4 |
| 3021 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 3022 vpxor xmm1,xmm1,xmm4 |
| 3023 |
| 3024 vmovdqu xmm6,XMMWORD[32+rax] |
| 3025 vmovdqu xmm13,XMMWORD[32+rcx] |
| 3026 |
| 3027 vpclmulqdq xmm4,xmm6,xmm13,0x10 |
| 3028 vpxor xmm1,xmm1,xmm4 |
| 3029 vpclmulqdq xmm4,xmm6,xmm13,0x11 |
| 3030 vpxor xmm2,xmm2,xmm4 |
| 3031 vpclmulqdq xmm4,xmm6,xmm13,0x00 |
| 3032 vpxor xmm3,xmm3,xmm4 |
| 3033 vpclmulqdq xmm4,xmm6,xmm13,0x01 |
| 3034 vpxor xmm1,xmm1,xmm4 |
| 3035 |
| 3036 |
| 3037 vmovdqu xmm6,XMMWORD[((80-32))+rax] |
| 3038 vpxor xmm6,xmm6,xmm0 |
| 3039 vmovdqu xmm5,XMMWORD[((80-32))+rcx] |
| 3040 vpclmulqdq xmm4,xmm6,xmm5,0x11 |
| 3041 vpxor xmm2,xmm2,xmm4 |
| 3042 vpclmulqdq xmm4,xmm6,xmm5,0x00 |
| 3043 vpxor xmm3,xmm3,xmm4 |
| 3044 vpclmulqdq xmm4,xmm6,xmm5,0x10 |
| 3045 vpxor xmm1,xmm1,xmm4 |
| 3046 vpclmulqdq xmm4,xmm6,xmm5,0x01 |
| 3047 vpxor xmm1,xmm1,xmm4 |
| 3048 |
| 3049 vpsrldq xmm4,xmm1,8 |
| 3050 vpxor xmm5,xmm2,xmm4 |
| 3051 vpslldq xmm4,xmm1,8 |
| 3052 vpxor xmm0,xmm3,xmm4 |
| 3053 |
| 3054 vmovdqa xmm3,XMMWORD[poly] |
| 3055 |
| 3056 vpalignr xmm2,xmm0,xmm0,8 |
| 3057 vpclmulqdq xmm0,xmm0,xmm3,0x10 |
| 3058 vpxor xmm0,xmm2,xmm0 |
| 3059 |
| 3060 vpalignr xmm2,xmm0,xmm0,8 |
| 3061 vpclmulqdq xmm0,xmm0,xmm3,0x10 |
| 3062 vpxor xmm0,xmm2,xmm0 |
| 3063 |
| 3064 vpxor xmm0,xmm0,xmm5 |
| 3065 |
| 3066 $L$256_dec_loop2: |
| 3067 |
| 3068 |
| 3069 |
| 3070 cmp r9,16 |
| 3071 jb NEAR $L$256_dec_out |
| 3072 sub r9,16 |
| 3073 |
| 3074 vmovdqa xmm2,xmm15 |
| 3075 vpaddd xmm15,xmm15,XMMWORD[one] |
| 3076 |
| 3077 vpxor xmm2,xmm2,XMMWORD[r8] |
| 3078 vaesenc xmm2,xmm2,XMMWORD[16+r8] |
| 3079 vaesenc xmm2,xmm2,XMMWORD[32+r8] |
| 3080 vaesenc xmm2,xmm2,XMMWORD[48+r8] |
| 3081 vaesenc xmm2,xmm2,XMMWORD[64+r8] |
| 3082 vaesenc xmm2,xmm2,XMMWORD[80+r8] |
| 3083 vaesenc xmm2,xmm2,XMMWORD[96+r8] |
| 3084 vaesenc xmm2,xmm2,XMMWORD[112+r8] |
| 3085 vaesenc xmm2,xmm2,XMMWORD[128+r8] |
| 3086 vaesenc xmm2,xmm2,XMMWORD[144+r8] |
| 3087 vaesenc xmm2,xmm2,XMMWORD[160+r8] |
| 3088 vaesenc xmm2,xmm2,XMMWORD[176+r8] |
| 3089 vaesenc xmm2,xmm2,XMMWORD[192+r8] |
| 3090 vaesenc xmm2,xmm2,XMMWORD[208+r8] |
| 3091 vaesenclast xmm2,xmm2,XMMWORD[224+r8] |
| 3092 vpxor xmm2,xmm2,XMMWORD[rdi] |
| 3093 vmovdqu XMMWORD[rsi],xmm2 |
| 3094 add rdi,16 |
| 3095 add rsi,16 |
| 3096 |
| 3097 vpxor xmm0,xmm0,xmm2 |
| 3098 vmovdqa xmm1,XMMWORD[((-32))+rcx] |
| 3099 call GFMUL |
| 3100 |
| 3101 jmp NEAR $L$256_dec_loop2 |
| 3102 |
| 3103 $L$256_dec_out: |
| 3104 vmovdqu XMMWORD[rdx],xmm0 |
| 3105 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 3106 mov rsi,QWORD[16+rsp] |
| 3107 DB 0F3h,0C3h ;repret |
| 3108 |
| 3109 $L$SEH_end_aes256gcmsiv_dec: |
| 3110 global aes256gcmsiv_kdf |
| 3111 |
| 3112 ALIGN 16 |
| 3113 aes256gcmsiv_kdf: |
| 3114 mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 3115 mov QWORD[16+rsp],rsi |
| 3116 mov rax,rsp |
| 3117 $L$SEH_begin_aes256gcmsiv_kdf: |
| 3118 mov rdi,rcx |
| 3119 mov rsi,rdx |
| 3120 mov rdx,r8 |
| 3121 |
| 3122 |
| 3123 |
| 3124 |
| 3125 |
| 3126 |
| 3127 |
| 3128 vmovdqa xmm1,XMMWORD[rdx] |
| 3129 vmovdqa xmm4,XMMWORD[rdi] |
| 3130 vmovdqa xmm11,XMMWORD[and_mask] |
| 3131 vmovdqa xmm8,XMMWORD[one] |
| 3132 vpshufd xmm4,xmm4,0x90 |
| 3133 vpand xmm4,xmm4,xmm11 |
| 3134 vpaddd xmm6,xmm4,xmm8 |
| 3135 vpaddd xmm7,xmm6,xmm8 |
| 3136 vpaddd xmm11,xmm7,xmm8 |
| 3137 vpaddd xmm12,xmm11,xmm8 |
| 3138 vpaddd xmm13,xmm12,xmm8 |
| 3139 |
| 3140 vpxor xmm4,xmm4,xmm1 |
| 3141 vpxor xmm6,xmm6,xmm1 |
| 3142 vpxor xmm7,xmm7,xmm1 |
| 3143 vpxor xmm11,xmm11,xmm1 |
| 3144 vpxor xmm12,xmm12,xmm1 |
| 3145 vpxor xmm13,xmm13,xmm1 |
| 3146 |
| 3147 vmovdqa xmm1,XMMWORD[16+rdx] |
| 3148 vaesenc xmm4,xmm4,xmm1 |
| 3149 vaesenc xmm6,xmm6,xmm1 |
| 3150 vaesenc xmm7,xmm7,xmm1 |
| 3151 vaesenc xmm11,xmm11,xmm1 |
| 3152 vaesenc xmm12,xmm12,xmm1 |
| 3153 vaesenc xmm13,xmm13,xmm1 |
| 3154 |
| 3155 vmovdqa xmm2,XMMWORD[32+rdx] |
| 3156 vaesenc xmm4,xmm4,xmm2 |
| 3157 vaesenc xmm6,xmm6,xmm2 |
| 3158 vaesenc xmm7,xmm7,xmm2 |
| 3159 vaesenc xmm11,xmm11,xmm2 |
| 3160 vaesenc xmm12,xmm12,xmm2 |
| 3161 vaesenc xmm13,xmm13,xmm2 |
| 3162 |
| 3163 vmovdqa xmm1,XMMWORD[48+rdx] |
| 3164 vaesenc xmm4,xmm4,xmm1 |
| 3165 vaesenc xmm6,xmm6,xmm1 |
| 3166 vaesenc xmm7,xmm7,xmm1 |
| 3167 vaesenc xmm11,xmm11,xmm1 |
| 3168 vaesenc xmm12,xmm12,xmm1 |
| 3169 vaesenc xmm13,xmm13,xmm1 |
| 3170 |
| 3171 vmovdqa xmm2,XMMWORD[64+rdx] |
| 3172 vaesenc xmm4,xmm4,xmm2 |
| 3173 vaesenc xmm6,xmm6,xmm2 |
| 3174 vaesenc xmm7,xmm7,xmm2 |
| 3175 vaesenc xmm11,xmm11,xmm2 |
| 3176 vaesenc xmm12,xmm12,xmm2 |
| 3177 vaesenc xmm13,xmm13,xmm2 |
| 3178 |
| 3179 vmovdqa xmm1,XMMWORD[80+rdx] |
| 3180 vaesenc xmm4,xmm4,xmm1 |
| 3181 vaesenc xmm6,xmm6,xmm1 |
| 3182 vaesenc xmm7,xmm7,xmm1 |
| 3183 vaesenc xmm11,xmm11,xmm1 |
| 3184 vaesenc xmm12,xmm12,xmm1 |
| 3185 vaesenc xmm13,xmm13,xmm1 |
| 3186 |
| 3187 vmovdqa xmm2,XMMWORD[96+rdx] |
| 3188 vaesenc xmm4,xmm4,xmm2 |
| 3189 vaesenc xmm6,xmm6,xmm2 |
| 3190 vaesenc xmm7,xmm7,xmm2 |
| 3191 vaesenc xmm11,xmm11,xmm2 |
| 3192 vaesenc xmm12,xmm12,xmm2 |
| 3193 vaesenc xmm13,xmm13,xmm2 |
| 3194 |
| 3195 vmovdqa xmm1,XMMWORD[112+rdx] |
| 3196 vaesenc xmm4,xmm4,xmm1 |
| 3197 vaesenc xmm6,xmm6,xmm1 |
| 3198 vaesenc xmm7,xmm7,xmm1 |
| 3199 vaesenc xmm11,xmm11,xmm1 |
| 3200 vaesenc xmm12,xmm12,xmm1 |
| 3201 vaesenc xmm13,xmm13,xmm1 |
| 3202 |
| 3203 vmovdqa xmm2,XMMWORD[128+rdx] |
| 3204 vaesenc xmm4,xmm4,xmm2 |
| 3205 vaesenc xmm6,xmm6,xmm2 |
| 3206 vaesenc xmm7,xmm7,xmm2 |
| 3207 vaesenc xmm11,xmm11,xmm2 |
| 3208 vaesenc xmm12,xmm12,xmm2 |
| 3209 vaesenc xmm13,xmm13,xmm2 |
| 3210 |
| 3211 vmovdqa xmm1,XMMWORD[144+rdx] |
| 3212 vaesenc xmm4,xmm4,xmm1 |
| 3213 vaesenc xmm6,xmm6,xmm1 |
| 3214 vaesenc xmm7,xmm7,xmm1 |
| 3215 vaesenc xmm11,xmm11,xmm1 |
| 3216 vaesenc xmm12,xmm12,xmm1 |
| 3217 vaesenc xmm13,xmm13,xmm1 |
| 3218 |
| 3219 vmovdqa xmm2,XMMWORD[160+rdx] |
| 3220 vaesenc xmm4,xmm4,xmm2 |
| 3221 vaesenc xmm6,xmm6,xmm2 |
| 3222 vaesenc xmm7,xmm7,xmm2 |
| 3223 vaesenc xmm11,xmm11,xmm2 |
| 3224 vaesenc xmm12,xmm12,xmm2 |
| 3225 vaesenc xmm13,xmm13,xmm2 |
| 3226 |
| 3227 vmovdqa xmm1,XMMWORD[176+rdx] |
| 3228 vaesenc xmm4,xmm4,xmm1 |
| 3229 vaesenc xmm6,xmm6,xmm1 |
| 3230 vaesenc xmm7,xmm7,xmm1 |
| 3231 vaesenc xmm11,xmm11,xmm1 |
| 3232 vaesenc xmm12,xmm12,xmm1 |
| 3233 vaesenc xmm13,xmm13,xmm1 |
| 3234 |
| 3235 vmovdqa xmm2,XMMWORD[192+rdx] |
| 3236 vaesenc xmm4,xmm4,xmm2 |
| 3237 vaesenc xmm6,xmm6,xmm2 |
| 3238 vaesenc xmm7,xmm7,xmm2 |
| 3239 vaesenc xmm11,xmm11,xmm2 |
| 3240 vaesenc xmm12,xmm12,xmm2 |
| 3241 vaesenc xmm13,xmm13,xmm2 |
| 3242 |
| 3243 vmovdqa xmm1,XMMWORD[208+rdx] |
| 3244 vaesenc xmm4,xmm4,xmm1 |
| 3245 vaesenc xmm6,xmm6,xmm1 |
| 3246 vaesenc xmm7,xmm7,xmm1 |
| 3247 vaesenc xmm11,xmm11,xmm1 |
| 3248 vaesenc xmm12,xmm12,xmm1 |
| 3249 vaesenc xmm13,xmm13,xmm1 |
| 3250 |
| 3251 vmovdqa xmm2,XMMWORD[224+rdx] |
| 3252 vaesenclast xmm4,xmm4,xmm2 |
| 3253 vaesenclast xmm6,xmm6,xmm2 |
| 3254 vaesenclast xmm7,xmm7,xmm2 |
| 3255 vaesenclast xmm11,xmm11,xmm2 |
| 3256 vaesenclast xmm12,xmm12,xmm2 |
| 3257 vaesenclast xmm13,xmm13,xmm2 |
| 3258 |
| 3259 |
| 3260 vmovdqa XMMWORD[rsi],xmm4 |
| 3261 vmovdqa XMMWORD[16+rsi],xmm6 |
| 3262 vmovdqa XMMWORD[32+rsi],xmm7 |
| 3263 vmovdqa XMMWORD[48+rsi],xmm11 |
| 3264 vmovdqa XMMWORD[64+rsi],xmm12 |
| 3265 vmovdqa XMMWORD[80+rsi],xmm13 |
| 3266 mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 3267 mov rsi,QWORD[16+rsp] |
| 3268 DB 0F3h,0C3h ;repret |
| 3269 |
| 3270 $L$SEH_end_aes256gcmsiv_kdf: |
OLD | NEW |