1 #include <machine/asm.h> 2 .text 3 4 5 .globl aesni_cbc_sha256_enc 6 .type aesni_cbc_sha256_enc,@function 7 .align 16 8 aesni_cbc_sha256_enc: 9 .cfi_startproc 10 leaq OPENSSL_ia32cap_P(%rip),%r11 11 movl $1,%eax 12 cmpq $0,%rdi 13 je .Lprobe 14 movl 0(%r11),%eax 15 movq 4(%r11),%r10 16 btq $61,%r10 17 jc aesni_cbc_sha256_enc_shaext 18 movq %r10,%r11 19 shrq $32,%r11 20 21 testl $2048,%r10d 22 jnz aesni_cbc_sha256_enc_xop 23 andl $296,%r11d 24 cmpl $296,%r11d 25 je aesni_cbc_sha256_enc_avx2 26 andl $268435456,%r10d 27 jnz aesni_cbc_sha256_enc_avx 28 ud2 29 xorl %eax,%eax 30 cmpq $0,%rdi 31 je .Lprobe 32 ud2 33 .Lprobe: 34 .byte 0xf3,0xc3 35 .cfi_endproc 36 .size aesni_cbc_sha256_enc,.-aesni_cbc_sha256_enc 37 38 .align 64 39 .type K256,@object 40 K256: 41 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5 42 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5 43 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5 44 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5 45 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3 46 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3 47 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174 48 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174 49 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc 50 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc 51 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da 52 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da 53 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7 54 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7 55 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967 56 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967 57 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13 58 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13 59 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85 60 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85 61 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3 62 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3 63 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070 64 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070 65 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5 66 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5 67 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3 68 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3 69 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208 70 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208 71 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2 72 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2 73 74 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f 75 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f 76 .long 0,0,0,0, 0,0,0,0, -1,-1,-1,-1 77 .long 0,0,0,0, 0,0,0,0 78 .byte 65,69,83,78,73,45,67,66,67,43,83,72,65,50,53,54,32,115,116,105,116,99,104,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0 79 .align 64 80 .type aesni_cbc_sha256_enc_xop,@function 81 .align 64 82 aesni_cbc_sha256_enc_xop: 83 .cfi_startproc 84 .Lxop_shortcut: 85 movq 8(%rsp),%r10 86 movq %rsp,%rax 87 .cfi_def_cfa_register %rax 88 pushq %rbx 89 .cfi_offset %rbx,-16 90 pushq %rbp 91 .cfi_offset %rbp,-24 92 pushq %r12 93 .cfi_offset %r12,-32 94 pushq %r13 95 .cfi_offset %r13,-40 96 pushq %r14 97 .cfi_offset %r14,-48 98 pushq %r15 99 .cfi_offset %r15,-56 100 subq $128,%rsp 101 andq $-64,%rsp 102 103 shlq $6,%rdx 104 subq %rdi,%rsi 105 subq %rdi,%r10 106 addq %rdi,%rdx 107 108 109 movq %rsi,64+8(%rsp) 110 movq %rdx,64+16(%rsp) 111 112 movq %r8,64+32(%rsp) 113 movq %r9,64+40(%rsp) 114 movq %r10,64+48(%rsp) 115 movq %rax,120(%rsp) 116 .cfi_escape 0x0f,0x06,0x77,0xf8,0x00,0x06,0x23,0x08 117 .Lprologue_xop: 118 vzeroall 119 120 movq %rdi,%r12 121 leaq 128(%rcx),%rdi 122 leaq K256+544(%rip),%r13 123 movl 240-128(%rdi),%r14d 124 movq %r9,%r15 125 movq %r10,%rsi 126 vmovdqu (%r8),%xmm8 127 subq $9,%r14 128 129 movl 0(%r15),%eax 130 movl 4(%r15),%ebx 131 movl 8(%r15),%ecx 132 movl 12(%r15),%edx 133 movl 16(%r15),%r8d 134 movl 20(%r15),%r9d 135 movl 24(%r15),%r10d 136 movl 28(%r15),%r11d 137 138 vmovdqa 0(%r13,%r14,8),%xmm14 139 vmovdqa 16(%r13,%r14,8),%xmm13 140 vmovdqa 32(%r13,%r14,8),%xmm12 141 vmovdqu 0-128(%rdi),%xmm10 142 jmp .Lloop_xop 143 .align 16 144 .Lloop_xop: 145 vmovdqa K256+512(%rip),%xmm7 146 vmovdqu 0(%rsi,%r12,1),%xmm0 147 vmovdqu 16(%rsi,%r12,1),%xmm1 148 vmovdqu 32(%rsi,%r12,1),%xmm2 149 vmovdqu 48(%rsi,%r12,1),%xmm3 150 vpshufb %xmm7,%xmm0,%xmm0 151 leaq K256(%rip),%rbp 152 vpshufb %xmm7,%xmm1,%xmm1 153 vpshufb %xmm7,%xmm2,%xmm2 154 vpaddd 0(%rbp),%xmm0,%xmm4 155 vpshufb %xmm7,%xmm3,%xmm3 156 vpaddd 32(%rbp),%xmm1,%xmm5 157 vpaddd 64(%rbp),%xmm2,%xmm6 158 vpaddd 96(%rbp),%xmm3,%xmm7 159 vmovdqa %xmm4,0(%rsp) 160 movl %eax,%r14d 161 vmovdqa %xmm5,16(%rsp) 162 movl %ebx,%esi 163 vmovdqa %xmm6,32(%rsp) 164 xorl %ecx,%esi 165 vmovdqa %xmm7,48(%rsp) 166 movl %r8d,%r13d 167 jmp .Lxop_00_47 168 169 .align 16 170 .Lxop_00_47: 171 subq $-32*4,%rbp 172 vmovdqu (%r12),%xmm9 173 movq %r12,64+0(%rsp) 174 vpalignr $4,%xmm0,%xmm1,%xmm4 175 rorl $14,%r13d 176 movl %r14d,%eax 177 vpalignr $4,%xmm2,%xmm3,%xmm7 178 movl %r9d,%r12d 179 xorl %r8d,%r13d 180 .byte 143,232,120,194,236,14 181 rorl $9,%r14d 182 xorl %r10d,%r12d 183 vpsrld $3,%xmm4,%xmm4 184 rorl $5,%r13d 185 xorl %eax,%r14d 186 vpaddd %xmm7,%xmm0,%xmm0 187 andl %r8d,%r12d 188 vpxor %xmm10,%xmm9,%xmm9 189 vmovdqu 16-128(%rdi),%xmm10 190 xorl %r8d,%r13d 191 addl 0(%rsp),%r11d 192 movl %eax,%r15d 193 .byte 143,232,120,194,245,11 194 rorl $11,%r14d 195 xorl %r10d,%r12d 196 vpxor %xmm5,%xmm4,%xmm4 197 xorl %ebx,%r15d 198 rorl $6,%r13d 199 addl %r12d,%r11d 200 andl %r15d,%esi 201 .byte 143,232,120,194,251,13 202 xorl %eax,%r14d 203 addl %r13d,%r11d 204 vpxor %xmm6,%xmm4,%xmm4 205 xorl %ebx,%esi 206 addl %r11d,%edx 207 vpsrld $10,%xmm3,%xmm6 208 rorl $2,%r14d 209 addl %esi,%r11d 210 vpaddd %xmm4,%xmm0,%xmm0 211 movl %edx,%r13d 212 addl %r11d,%r14d 213 .byte 143,232,120,194,239,2 214 rorl $14,%r13d 215 movl %r14d,%r11d 216 vpxor %xmm6,%xmm7,%xmm7 217 movl %r8d,%r12d 218 xorl %edx,%r13d 219 rorl $9,%r14d 220 xorl %r9d,%r12d 221 vpxor %xmm5,%xmm7,%xmm7 222 rorl $5,%r13d 223 xorl %r11d,%r14d 224 andl %edx,%r12d 225 vpxor %xmm8,%xmm9,%xmm9 226 xorl %edx,%r13d 227 vpsrldq $8,%xmm7,%xmm7 228 addl 4(%rsp),%r10d 229 movl %r11d,%esi 230 rorl $11,%r14d 231 xorl %r9d,%r12d 232 vpaddd %xmm7,%xmm0,%xmm0 233 xorl %eax,%esi 234 rorl $6,%r13d 235 addl %r12d,%r10d 236 andl %esi,%r15d 237 .byte 143,232,120,194,248,13 238 xorl %r11d,%r14d 239 addl %r13d,%r10d 240 vpsrld $10,%xmm0,%xmm6 241 xorl %eax,%r15d 242 addl %r10d,%ecx 243 .byte 143,232,120,194,239,2 244 rorl $2,%r14d 245 addl %r15d,%r10d 246 vpxor %xmm6,%xmm7,%xmm7 247 movl %ecx,%r13d 248 addl %r10d,%r14d 249 rorl $14,%r13d 250 movl %r14d,%r10d 251 vpxor %xmm5,%xmm7,%xmm7 252 movl %edx,%r12d 253 xorl %ecx,%r13d 254 rorl $9,%r14d 255 xorl %r8d,%r12d 256 vpslldq $8,%xmm7,%xmm7 257 rorl $5,%r13d 258 xorl %r10d,%r14d 259 andl %ecx,%r12d 260 vaesenc %xmm10,%xmm9,%xmm9 261 vmovdqu 32-128(%rdi),%xmm10 262 xorl %ecx,%r13d 263 vpaddd %xmm7,%xmm0,%xmm0 264 addl 8(%rsp),%r9d 265 movl %r10d,%r15d 266 rorl $11,%r14d 267 xorl %r8d,%r12d 268 vpaddd 0(%rbp),%xmm0,%xmm6 269 xorl %r11d,%r15d 270 rorl $6,%r13d 271 addl %r12d,%r9d 272 andl %r15d,%esi 273 xorl %r10d,%r14d 274 addl %r13d,%r9d 275 xorl %r11d,%esi 276 addl %r9d,%ebx 277 rorl $2,%r14d 278 addl %esi,%r9d 279 movl %ebx,%r13d 280 addl %r9d,%r14d 281 rorl $14,%r13d 282 movl %r14d,%r9d 283 movl %ecx,%r12d 284 xorl %ebx,%r13d 285 rorl $9,%r14d 286 xorl %edx,%r12d 287 rorl $5,%r13d 288 xorl %r9d,%r14d 289 andl %ebx,%r12d 290 vaesenc %xmm10,%xmm9,%xmm9 291 vmovdqu 48-128(%rdi),%xmm10 292 xorl %ebx,%r13d 293 addl 12(%rsp),%r8d 294 movl %r9d,%esi 295 rorl $11,%r14d 296 xorl %edx,%r12d 297 xorl %r10d,%esi 298 rorl $6,%r13d 299 addl %r12d,%r8d 300 andl %esi,%r15d 301 xorl %r9d,%r14d 302 addl %r13d,%r8d 303 xorl %r10d,%r15d 304 addl %r8d,%eax 305 rorl $2,%r14d 306 addl %r15d,%r8d 307 movl %eax,%r13d 308 addl %r8d,%r14d 309 vmovdqa %xmm6,0(%rsp) 310 vpalignr $4,%xmm1,%xmm2,%xmm4 311 rorl $14,%r13d 312 movl %r14d,%r8d 313 vpalignr $4,%xmm3,%xmm0,%xmm7 314 movl %ebx,%r12d 315 xorl %eax,%r13d 316 .byte 143,232,120,194,236,14 317 rorl $9,%r14d 318 xorl %ecx,%r12d 319 vpsrld $3,%xmm4,%xmm4 320 rorl $5,%r13d 321 xorl %r8d,%r14d 322 vpaddd %xmm7,%xmm1,%xmm1 323 andl %eax,%r12d 324 vaesenc %xmm10,%xmm9,%xmm9 325 vmovdqu 64-128(%rdi),%xmm10 326 xorl %eax,%r13d 327 addl 16(%rsp),%edx 328 movl %r8d,%r15d 329 .byte 143,232,120,194,245,11 330 rorl $11,%r14d 331 xorl %ecx,%r12d 332 vpxor %xmm5,%xmm4,%xmm4 333 xorl %r9d,%r15d 334 rorl $6,%r13d 335 addl %r12d,%edx 336 andl %r15d,%esi 337 .byte 143,232,120,194,248,13 338 xorl %r8d,%r14d 339 addl %r13d,%edx 340 vpxor %xmm6,%xmm4,%xmm4 341 xorl %r9d,%esi 342 addl %edx,%r11d 343 vpsrld $10,%xmm0,%xmm6 344 rorl $2,%r14d 345 addl %esi,%edx 346 vpaddd %xmm4,%xmm1,%xmm1 347 movl %r11d,%r13d 348 addl %edx,%r14d 349 .byte 143,232,120,194,239,2 350 rorl $14,%r13d 351 movl %r14d,%edx 352 vpxor %xmm6,%xmm7,%xmm7 353 movl %eax,%r12d 354 xorl %r11d,%r13d 355 rorl $9,%r14d 356 xorl %ebx,%r12d 357 vpxor %xmm5,%xmm7,%xmm7 358 rorl $5,%r13d 359 xorl %edx,%r14d 360 andl %r11d,%r12d 361 vaesenc %xmm10,%xmm9,%xmm9 362 vmovdqu 80-128(%rdi),%xmm10 363 xorl %r11d,%r13d 364 vpsrldq $8,%xmm7,%xmm7 365 addl 20(%rsp),%ecx 366 movl %edx,%esi 367 rorl $11,%r14d 368 xorl %ebx,%r12d 369 vpaddd %xmm7,%xmm1,%xmm1 370 xorl %r8d,%esi 371 rorl $6,%r13d 372 addl %r12d,%ecx 373 andl %esi,%r15d 374 .byte 143,232,120,194,249,13 375 xorl %edx,%r14d 376 addl %r13d,%ecx 377 vpsrld $10,%xmm1,%xmm6 378 xorl %r8d,%r15d 379 addl %ecx,%r10d 380 .byte 143,232,120,194,239,2 381 rorl $2,%r14d 382 addl %r15d,%ecx 383 vpxor %xmm6,%xmm7,%xmm7 384 movl %r10d,%r13d 385 addl %ecx,%r14d 386 rorl $14,%r13d 387 movl %r14d,%ecx 388 vpxor %xmm5,%xmm7,%xmm7 389 movl %r11d,%r12d 390 xorl %r10d,%r13d 391 rorl $9,%r14d 392 xorl %eax,%r12d 393 vpslldq $8,%xmm7,%xmm7 394 rorl $5,%r13d 395 xorl %ecx,%r14d 396 andl %r10d,%r12d 397 vaesenc %xmm10,%xmm9,%xmm9 398 vmovdqu 96-128(%rdi),%xmm10 399 xorl %r10d,%r13d 400 vpaddd %xmm7,%xmm1,%xmm1 401 addl 24(%rsp),%ebx 402 movl %ecx,%r15d 403 rorl $11,%r14d 404 xorl %eax,%r12d 405 vpaddd 32(%rbp),%xmm1,%xmm6 406 xorl %edx,%r15d 407 rorl $6,%r13d 408 addl %r12d,%ebx 409 andl %r15d,%esi 410 xorl %ecx,%r14d 411 addl %r13d,%ebx 412 xorl %edx,%esi 413 addl %ebx,%r9d 414 rorl $2,%r14d 415 addl %esi,%ebx 416 movl %r9d,%r13d 417 addl %ebx,%r14d 418 rorl $14,%r13d 419 movl %r14d,%ebx 420 movl %r10d,%r12d 421 xorl %r9d,%r13d 422 rorl $9,%r14d 423 xorl %r11d,%r12d 424 rorl $5,%r13d 425 xorl %ebx,%r14d 426 andl %r9d,%r12d 427 vaesenc %xmm10,%xmm9,%xmm9 428 vmovdqu 112-128(%rdi),%xmm10 429 xorl %r9d,%r13d 430 addl 28(%rsp),%eax 431 movl %ebx,%esi 432 rorl $11,%r14d 433 xorl %r11d,%r12d 434 xorl %ecx,%esi 435 rorl $6,%r13d 436 addl %r12d,%eax 437 andl %esi,%r15d 438 xorl %ebx,%r14d 439 addl %r13d,%eax 440 xorl %ecx,%r15d 441 addl %eax,%r8d 442 rorl $2,%r14d 443 addl %r15d,%eax 444 movl %r8d,%r13d 445 addl %eax,%r14d 446 vmovdqa %xmm6,16(%rsp) 447 vpalignr $4,%xmm2,%xmm3,%xmm4 448 rorl $14,%r13d 449 movl %r14d,%eax 450 vpalignr $4,%xmm0,%xmm1,%xmm7 451 movl %r9d,%r12d 452 xorl %r8d,%r13d 453 .byte 143,232,120,194,236,14 454 rorl $9,%r14d 455 xorl %r10d,%r12d 456 vpsrld $3,%xmm4,%xmm4 457 rorl $5,%r13d 458 xorl %eax,%r14d 459 vpaddd %xmm7,%xmm2,%xmm2 460 andl %r8d,%r12d 461 vaesenc %xmm10,%xmm9,%xmm9 462 vmovdqu 128-128(%rdi),%xmm10 463 xorl %r8d,%r13d 464 addl 32(%rsp),%r11d 465 movl %eax,%r15d 466 .byte 143,232,120,194,245,11 467 rorl $11,%r14d 468 xorl %r10d,%r12d 469 vpxor %xmm5,%xmm4,%xmm4 470 xorl %ebx,%r15d 471 rorl $6,%r13d 472 addl %r12d,%r11d 473 andl %r15d,%esi 474 .byte 143,232,120,194,249,13 475 xorl %eax,%r14d 476 addl %r13d,%r11d 477 vpxor %xmm6,%xmm4,%xmm4 478 xorl %ebx,%esi 479 addl %r11d,%edx 480 vpsrld $10,%xmm1,%xmm6 481 rorl $2,%r14d 482 addl %esi,%r11d 483 vpaddd %xmm4,%xmm2,%xmm2 484 movl %edx,%r13d 485 addl %r11d,%r14d 486 .byte 143,232,120,194,239,2 487 rorl $14,%r13d 488 movl %r14d,%r11d 489 vpxor %xmm6,%xmm7,%xmm7 490 movl %r8d,%r12d 491 xorl %edx,%r13d 492 rorl $9,%r14d 493 xorl %r9d,%r12d 494 vpxor %xmm5,%xmm7,%xmm7 495 rorl $5,%r13d 496 xorl %r11d,%r14d 497 andl %edx,%r12d 498 vaesenc %xmm10,%xmm9,%xmm9 499 vmovdqu 144-128(%rdi),%xmm10 500 xorl %edx,%r13d 501 vpsrldq $8,%xmm7,%xmm7 502 addl 36(%rsp),%r10d 503 movl %r11d,%esi 504 rorl $11,%r14d 505 xorl %r9d,%r12d 506 vpaddd %xmm7,%xmm2,%xmm2 507 xorl %eax,%esi 508 rorl $6,%r13d 509 addl %r12d,%r10d 510 andl %esi,%r15d 511 .byte 143,232,120,194,250,13 512 xorl %r11d,%r14d 513 addl %r13d,%r10d 514 vpsrld $10,%xmm2,%xmm6 515 xorl %eax,%r15d 516 addl %r10d,%ecx 517 .byte 143,232,120,194,239,2 518 rorl $2,%r14d 519 addl %r15d,%r10d 520 vpxor %xmm6,%xmm7,%xmm7 521 movl %ecx,%r13d 522 addl %r10d,%r14d 523 rorl $14,%r13d 524 movl %r14d,%r10d 525 vpxor %xmm5,%xmm7,%xmm7 526 movl %edx,%r12d 527 xorl %ecx,%r13d 528 rorl $9,%r14d 529 xorl %r8d,%r12d 530 vpslldq $8,%xmm7,%xmm7 531 rorl $5,%r13d 532 xorl %r10d,%r14d 533 andl %ecx,%r12d 534 vaesenc %xmm10,%xmm9,%xmm9 535 vmovdqu 160-128(%rdi),%xmm10 536 xorl %ecx,%r13d 537 vpaddd %xmm7,%xmm2,%xmm2 538 addl 40(%rsp),%r9d 539 movl %r10d,%r15d 540 rorl $11,%r14d 541 xorl %r8d,%r12d 542 vpaddd 64(%rbp),%xmm2,%xmm6 543 xorl %r11d,%r15d 544 rorl $6,%r13d 545 addl %r12d,%r9d 546 andl %r15d,%esi 547 xorl %r10d,%r14d 548 addl %r13d,%r9d 549 xorl %r11d,%esi 550 addl %r9d,%ebx 551 rorl $2,%r14d 552 addl %esi,%r9d 553 movl %ebx,%r13d 554 addl %r9d,%r14d 555 rorl $14,%r13d 556 movl %r14d,%r9d 557 movl %ecx,%r12d 558 xorl %ebx,%r13d 559 rorl $9,%r14d 560 xorl %edx,%r12d 561 rorl $5,%r13d 562 xorl %r9d,%r14d 563 andl %ebx,%r12d 564 vaesenclast %xmm10,%xmm9,%xmm11 565 vaesenc %xmm10,%xmm9,%xmm9 566 vmovdqu 176-128(%rdi),%xmm10 567 xorl %ebx,%r13d 568 addl 44(%rsp),%r8d 569 movl %r9d,%esi 570 rorl $11,%r14d 571 xorl %edx,%r12d 572 xorl %r10d,%esi 573 rorl $6,%r13d 574 addl %r12d,%r8d 575 andl %esi,%r15d 576 xorl %r9d,%r14d 577 addl %r13d,%r8d 578 xorl %r10d,%r15d 579 addl %r8d,%eax 580 rorl $2,%r14d 581 addl %r15d,%r8d 582 movl %eax,%r13d 583 addl %r8d,%r14d 584 vmovdqa %xmm6,32(%rsp) 585 vpalignr $4,%xmm3,%xmm0,%xmm4 586 rorl $14,%r13d 587 movl %r14d,%r8d 588 vpalignr $4,%xmm1,%xmm2,%xmm7 589 movl %ebx,%r12d 590 xorl %eax,%r13d 591 .byte 143,232,120,194,236,14 592 rorl $9,%r14d 593 xorl %ecx,%r12d 594 vpsrld $3,%xmm4,%xmm4 595 rorl $5,%r13d 596 xorl %r8d,%r14d 597 vpaddd %xmm7,%xmm3,%xmm3 598 andl %eax,%r12d 599 vpand %xmm12,%xmm11,%xmm8 600 vaesenc %xmm10,%xmm9,%xmm9 601 vmovdqu 192-128(%rdi),%xmm10 602 xorl %eax,%r13d 603 addl 48(%rsp),%edx 604 movl %r8d,%r15d 605 .byte 143,232,120,194,245,11 606 rorl $11,%r14d 607 xorl %ecx,%r12d 608 vpxor %xmm5,%xmm4,%xmm4 609 xorl %r9d,%r15d 610 rorl $6,%r13d 611 addl %r12d,%edx 612 andl %r15d,%esi 613 .byte 143,232,120,194,250,13 614 xorl %r8d,%r14d 615 addl %r13d,%edx 616 vpxor %xmm6,%xmm4,%xmm4 617 xorl %r9d,%esi 618 addl %edx,%r11d 619 vpsrld $10,%xmm2,%xmm6 620 rorl $2,%r14d 621 addl %esi,%edx 622 vpaddd %xmm4,%xmm3,%xmm3 623 movl %r11d,%r13d 624 addl %edx,%r14d 625 .byte 143,232,120,194,239,2 626 rorl $14,%r13d 627 movl %r14d,%edx 628 vpxor %xmm6,%xmm7,%xmm7 629 movl %eax,%r12d 630 xorl %r11d,%r13d 631 rorl $9,%r14d 632 xorl %ebx,%r12d 633 vpxor %xmm5,%xmm7,%xmm7 634 rorl $5,%r13d 635 xorl %edx,%r14d 636 andl %r11d,%r12d 637 vaesenclast %xmm10,%xmm9,%xmm11 638 vaesenc %xmm10,%xmm9,%xmm9 639 vmovdqu 208-128(%rdi),%xmm10 640 xorl %r11d,%r13d 641 vpsrldq $8,%xmm7,%xmm7 642 addl 52(%rsp),%ecx 643 movl %edx,%esi 644 rorl $11,%r14d 645 xorl %ebx,%r12d 646 vpaddd %xmm7,%xmm3,%xmm3 647 xorl %r8d,%esi 648 rorl $6,%r13d 649 addl %r12d,%ecx 650 andl %esi,%r15d 651 .byte 143,232,120,194,251,13 652 xorl %edx,%r14d 653 addl %r13d,%ecx 654 vpsrld $10,%xmm3,%xmm6 655 xorl %r8d,%r15d 656 addl %ecx,%r10d 657 .byte 143,232,120,194,239,2 658 rorl $2,%r14d 659 addl %r15d,%ecx 660 vpxor %xmm6,%xmm7,%xmm7 661 movl %r10d,%r13d 662 addl %ecx,%r14d 663 rorl $14,%r13d 664 movl %r14d,%ecx 665 vpxor %xmm5,%xmm7,%xmm7 666 movl %r11d,%r12d 667 xorl %r10d,%r13d 668 rorl $9,%r14d 669 xorl %eax,%r12d 670 vpslldq $8,%xmm7,%xmm7 671 rorl $5,%r13d 672 xorl %ecx,%r14d 673 andl %r10d,%r12d 674 vpand %xmm13,%xmm11,%xmm11 675 vaesenc %xmm10,%xmm9,%xmm9 676 vmovdqu 224-128(%rdi),%xmm10 677 xorl %r10d,%r13d 678 vpaddd %xmm7,%xmm3,%xmm3 679 addl 56(%rsp),%ebx 680 movl %ecx,%r15d 681 rorl $11,%r14d 682 xorl %eax,%r12d 683 vpaddd 96(%rbp),%xmm3,%xmm6 684 xorl %edx,%r15d 685 rorl $6,%r13d 686 addl %r12d,%ebx 687 andl %r15d,%esi 688 xorl %ecx,%r14d 689 addl %r13d,%ebx 690 xorl %edx,%esi 691 addl %ebx,%r9d 692 rorl $2,%r14d 693 addl %esi,%ebx 694 movl %r9d,%r13d 695 addl %ebx,%r14d 696 rorl $14,%r13d 697 movl %r14d,%ebx 698 movl %r10d,%r12d 699 xorl %r9d,%r13d 700 rorl $9,%r14d 701 xorl %r11d,%r12d 702 rorl $5,%r13d 703 xorl %ebx,%r14d 704 andl %r9d,%r12d 705 vpor %xmm11,%xmm8,%xmm8 706 vaesenclast %xmm10,%xmm9,%xmm11 707 vmovdqu 0-128(%rdi),%xmm10 708 xorl %r9d,%r13d 709 addl 60(%rsp),%eax 710 movl %ebx,%esi 711 rorl $11,%r14d 712 xorl %r11d,%r12d 713 xorl %ecx,%esi 714 rorl $6,%r13d 715 addl %r12d,%eax 716 andl %esi,%r15d 717 xorl %ebx,%r14d 718 addl %r13d,%eax 719 xorl %ecx,%r15d 720 addl %eax,%r8d 721 rorl $2,%r14d 722 addl %r15d,%eax 723 movl %r8d,%r13d 724 addl %eax,%r14d 725 vmovdqa %xmm6,48(%rsp) 726 movq 64+0(%rsp),%r12 727 vpand %xmm14,%xmm11,%xmm11 728 movq 64+8(%rsp),%r15 729 vpor %xmm11,%xmm8,%xmm8 730 vmovdqu %xmm8,(%r15,%r12,1) 731 leaq 16(%r12),%r12 732 cmpb $0,131(%rbp) 733 jne .Lxop_00_47 734 vmovdqu (%r12),%xmm9 735 movq %r12,64+0(%rsp) 736 rorl $14,%r13d 737 movl %r14d,%eax 738 movl %r9d,%r12d 739 xorl %r8d,%r13d 740 rorl $9,%r14d 741 xorl %r10d,%r12d 742 rorl $5,%r13d 743 xorl %eax,%r14d 744 andl %r8d,%r12d 745 vpxor %xmm10,%xmm9,%xmm9 746 vmovdqu 16-128(%rdi),%xmm10 747 xorl %r8d,%r13d 748 addl 0(%rsp),%r11d 749 movl %eax,%r15d 750 rorl $11,%r14d 751 xorl %r10d,%r12d 752 xorl %ebx,%r15d 753 rorl $6,%r13d 754 addl %r12d,%r11d 755 andl %r15d,%esi 756 xorl %eax,%r14d 757 addl %r13d,%r11d 758 xorl %ebx,%esi 759 addl %r11d,%edx 760 rorl $2,%r14d 761 addl %esi,%r11d 762 movl %edx,%r13d 763 addl %r11d,%r14d 764 rorl $14,%r13d 765 movl %r14d,%r11d 766 movl %r8d,%r12d 767 xorl %edx,%r13d 768 rorl $9,%r14d 769 xorl %r9d,%r12d 770 rorl $5,%r13d 771 xorl %r11d,%r14d 772 andl %edx,%r12d 773 vpxor %xmm8,%xmm9,%xmm9 774 xorl %edx,%r13d 775 addl 4(%rsp),%r10d 776 movl %r11d,%esi 777 rorl $11,%r14d 778 xorl %r9d,%r12d 779 xorl %eax,%esi 780 rorl $6,%r13d 781 addl %r12d,%r10d 782 andl %esi,%r15d 783 xorl %r11d,%r14d 784 addl %r13d,%r10d 785 xorl %eax,%r15d 786 addl %r10d,%ecx 787 rorl $2,%r14d 788 addl %r15d,%r10d 789 movl %ecx,%r13d 790 addl %r10d,%r14d 791 rorl $14,%r13d 792 movl %r14d,%r10d 793 movl %edx,%r12d 794 xorl %ecx,%r13d 795 rorl $9,%r14d 796 xorl %r8d,%r12d 797 rorl $5,%r13d 798 xorl %r10d,%r14d 799 andl %ecx,%r12d 800 vaesenc %xmm10,%xmm9,%xmm9 801 vmovdqu 32-128(%rdi),%xmm10 802 xorl %ecx,%r13d 803 addl 8(%rsp),%r9d 804 movl %r10d,%r15d 805 rorl $11,%r14d 806 xorl %r8d,%r12d 807 xorl %r11d,%r15d 808 rorl $6,%r13d 809 addl %r12d,%r9d 810 andl %r15d,%esi 811 xorl %r10d,%r14d 812 addl %r13d,%r9d 813 xorl %r11d,%esi 814 addl %r9d,%ebx 815 rorl $2,%r14d 816 addl %esi,%r9d 817 movl %ebx,%r13d 818 addl %r9d,%r14d 819 rorl $14,%r13d 820 movl %r14d,%r9d 821 movl %ecx,%r12d 822 xorl %ebx,%r13d 823 rorl $9,%r14d 824 xorl %edx,%r12d 825 rorl $5,%r13d 826 xorl %r9d,%r14d 827 andl %ebx,%r12d 828 vaesenc %xmm10,%xmm9,%xmm9 829 vmovdqu 48-128(%rdi),%xmm10 830 xorl %ebx,%r13d 831 addl 12(%rsp),%r8d 832 movl %r9d,%esi 833 rorl $11,%r14d 834 xorl %edx,%r12d 835 xorl %r10d,%esi 836 rorl $6,%r13d 837 addl %r12d,%r8d 838 andl %esi,%r15d 839 xorl %r9d,%r14d 840 addl %r13d,%r8d 841 xorl %r10d,%r15d 842 addl %r8d,%eax 843 rorl $2,%r14d 844 addl %r15d,%r8d 845 movl %eax,%r13d 846 addl %r8d,%r14d 847 rorl $14,%r13d 848 movl %r14d,%r8d 849 movl %ebx,%r12d 850 xorl %eax,%r13d 851 rorl $9,%r14d 852 xorl %ecx,%r12d 853 rorl $5,%r13d 854 xorl %r8d,%r14d 855 andl %eax,%r12d 856 vaesenc %xmm10,%xmm9,%xmm9 857 vmovdqu 64-128(%rdi),%xmm10 858 xorl %eax,%r13d 859 addl 16(%rsp),%edx 860 movl %r8d,%r15d 861 rorl $11,%r14d 862 xorl %ecx,%r12d 863 xorl %r9d,%r15d 864 rorl $6,%r13d 865 addl %r12d,%edx 866 andl %r15d,%esi 867 xorl %r8d,%r14d 868 addl %r13d,%edx 869 xorl %r9d,%esi 870 addl %edx,%r11d 871 rorl $2,%r14d 872 addl %esi,%edx 873 movl %r11d,%r13d 874 addl %edx,%r14d 875 rorl $14,%r13d 876 movl %r14d,%edx 877 movl %eax,%r12d 878 xorl %r11d,%r13d 879 rorl $9,%r14d 880 xorl %ebx,%r12d 881 rorl $5,%r13d 882 xorl %edx,%r14d 883 andl %r11d,%r12d 884 vaesenc %xmm10,%xmm9,%xmm9 885 vmovdqu 80-128(%rdi),%xmm10 886 xorl %r11d,%r13d 887 addl 20(%rsp),%ecx 888 movl %edx,%esi 889 rorl $11,%r14d 890 xorl %ebx,%r12d 891 xorl %r8d,%esi 892 rorl $6,%r13d 893 addl %r12d,%ecx 894 andl %esi,%r15d 895 xorl %edx,%r14d 896 addl %r13d,%ecx 897 xorl %r8d,%r15d 898 addl %ecx,%r10d 899 rorl $2,%r14d 900 addl %r15d,%ecx 901 movl %r10d,%r13d 902 addl %ecx,%r14d 903 rorl $14,%r13d 904 movl %r14d,%ecx 905 movl %r11d,%r12d 906 xorl %r10d,%r13d 907 rorl $9,%r14d 908 xorl %eax,%r12d 909 rorl $5,%r13d 910 xorl %ecx,%r14d 911 andl %r10d,%r12d 912 vaesenc %xmm10,%xmm9,%xmm9 913 vmovdqu 96-128(%rdi),%xmm10 914 xorl %r10d,%r13d 915 addl 24(%rsp),%ebx 916 movl %ecx,%r15d 917 rorl $11,%r14d 918 xorl %eax,%r12d 919 xorl %edx,%r15d 920 rorl $6,%r13d 921 addl %r12d,%ebx 922 andl %r15d,%esi 923 xorl %ecx,%r14d 924 addl %r13d,%ebx 925 xorl %edx,%esi 926 addl %ebx,%r9d 927 rorl $2,%r14d 928 addl %esi,%ebx 929 movl %r9d,%r13d 930 addl %ebx,%r14d 931 rorl $14,%r13d 932 movl %r14d,%ebx 933 movl %r10d,%r12d 934 xorl %r9d,%r13d 935 rorl $9,%r14d 936 xorl %r11d,%r12d 937 rorl $5,%r13d 938 xorl %ebx,%r14d 939 andl %r9d,%r12d 940 vaesenc %xmm10,%xmm9,%xmm9 941 vmovdqu 112-128(%rdi),%xmm10 942 xorl %r9d,%r13d 943 addl 28(%rsp),%eax 944 movl %ebx,%esi 945 rorl $11,%r14d 946 xorl %r11d,%r12d 947 xorl %ecx,%esi 948 rorl $6,%r13d 949 addl %r12d,%eax 950 andl %esi,%r15d 951 xorl %ebx,%r14d 952 addl %r13d,%eax 953 xorl %ecx,%r15d 954 addl %eax,%r8d 955 rorl $2,%r14d 956 addl %r15d,%eax 957 movl %r8d,%r13d 958 addl %eax,%r14d 959 rorl $14,%r13d 960 movl %r14d,%eax 961 movl %r9d,%r12d 962 xorl %r8d,%r13d 963 rorl $9,%r14d 964 xorl %r10d,%r12d 965 rorl $5,%r13d 966 xorl %eax,%r14d 967 andl %r8d,%r12d 968 vaesenc %xmm10,%xmm9,%xmm9 969 vmovdqu 128-128(%rdi),%xmm10 970 xorl %r8d,%r13d 971 addl 32(%rsp),%r11d 972 movl %eax,%r15d 973 rorl $11,%r14d 974 xorl %r10d,%r12d 975 xorl %ebx,%r15d 976 rorl $6,%r13d 977 addl %r12d,%r11d 978 andl %r15d,%esi 979 xorl %eax,%r14d 980 addl %r13d,%r11d 981 xorl %ebx,%esi 982 addl %r11d,%edx 983 rorl $2,%r14d 984 addl %esi,%r11d 985 movl %edx,%r13d 986 addl %r11d,%r14d 987 rorl $14,%r13d 988 movl %r14d,%r11d 989 movl %r8d,%r12d 990 xorl %edx,%r13d 991 rorl $9,%r14d 992 xorl %r9d,%r12d 993 rorl $5,%r13d 994 xorl %r11d,%r14d 995 andl %edx,%r12d 996 vaesenc %xmm10,%xmm9,%xmm9 997 vmovdqu 144-128(%rdi),%xmm10 998 xorl %edx,%r13d 999 addl 36(%rsp),%r10d 1000 movl %r11d,%esi 1001 rorl $11,%r14d 1002 xorl %r9d,%r12d 1003 xorl %eax,%esi 1004 rorl $6,%r13d 1005 addl %r12d,%r10d 1006 andl %esi,%r15d 1007 xorl %r11d,%r14d 1008 addl %r13d,%r10d 1009 xorl %eax,%r15d 1010 addl %r10d,%ecx 1011 rorl $2,%r14d 1012 addl %r15d,%r10d 1013 movl %ecx,%r13d 1014 addl %r10d,%r14d 1015 rorl $14,%r13d 1016 movl %r14d,%r10d 1017 movl %edx,%r12d 1018 xorl %ecx,%r13d 1019 rorl $9,%r14d 1020 xorl %r8d,%r12d 1021 rorl $5,%r13d 1022 xorl %r10d,%r14d 1023 andl %ecx,%r12d 1024 vaesenc %xmm10,%xmm9,%xmm9 1025 vmovdqu 160-128(%rdi),%xmm10 1026 xorl %ecx,%r13d 1027 addl 40(%rsp),%r9d 1028 movl %r10d,%r15d 1029 rorl $11,%r14d 1030 xorl %r8d,%r12d 1031 xorl %r11d,%r15d 1032 rorl $6,%r13d 1033 addl %r12d,%r9d 1034 andl %r15d,%esi 1035 xorl %r10d,%r14d 1036 addl %r13d,%r9d 1037 xorl %r11d,%esi 1038 addl %r9d,%ebx 1039 rorl $2,%r14d 1040 addl %esi,%r9d 1041 movl %ebx,%r13d 1042 addl %r9d,%r14d 1043 rorl $14,%r13d 1044 movl %r14d,%r9d 1045 movl %ecx,%r12d 1046 xorl %ebx,%r13d 1047 rorl $9,%r14d 1048 xorl %edx,%r12d 1049 rorl $5,%r13d 1050 xorl %r9d,%r14d 1051 andl %ebx,%r12d 1052 vaesenclast %xmm10,%xmm9,%xmm11 1053 vaesenc %xmm10,%xmm9,%xmm9 1054 vmovdqu 176-128(%rdi),%xmm10 1055 xorl %ebx,%r13d 1056 addl 44(%rsp),%r8d 1057 movl %r9d,%esi 1058 rorl $11,%r14d 1059 xorl %edx,%r12d 1060 xorl %r10d,%esi 1061 rorl $6,%r13d 1062 addl %r12d,%r8d 1063 andl %esi,%r15d 1064 xorl %r9d,%r14d 1065 addl %r13d,%r8d 1066 xorl %r10d,%r15d 1067 addl %r8d,%eax 1068 rorl $2,%r14d 1069 addl %r15d,%r8d 1070 movl %eax,%r13d 1071 addl %r8d,%r14d 1072 rorl $14,%r13d 1073 movl %r14d,%r8d 1074 movl %ebx,%r12d 1075 xorl %eax,%r13d 1076 rorl $9,%r14d 1077 xorl %ecx,%r12d 1078 rorl $5,%r13d 1079 xorl %r8d,%r14d 1080 andl %eax,%r12d 1081 vpand %xmm12,%xmm11,%xmm8 1082 vaesenc %xmm10,%xmm9,%xmm9 1083 vmovdqu 192-128(%rdi),%xmm10 1084 xorl %eax,%r13d 1085 addl 48(%rsp),%edx 1086 movl %r8d,%r15d 1087 rorl $11,%r14d 1088 xorl %ecx,%r12d 1089 xorl %r9d,%r15d 1090 rorl $6,%r13d 1091 addl %r12d,%edx 1092 andl %r15d,%esi 1093 xorl %r8d,%r14d 1094 addl %r13d,%edx 1095 xorl %r9d,%esi 1096 addl %edx,%r11d 1097 rorl $2,%r14d 1098 addl %esi,%edx 1099 movl %r11d,%r13d 1100 addl %edx,%r14d 1101 rorl $14,%r13d 1102 movl %r14d,%edx 1103 movl %eax,%r12d 1104 xorl %r11d,%r13d 1105 rorl $9,%r14d 1106 xorl %ebx,%r12d 1107 rorl $5,%r13d 1108 xorl %edx,%r14d 1109 andl %r11d,%r12d 1110 vaesenclast %xmm10,%xmm9,%xmm11 1111 vaesenc %xmm10,%xmm9,%xmm9 1112 vmovdqu 208-128(%rdi),%xmm10 1113 xorl %r11d,%r13d 1114 addl 52(%rsp),%ecx 1115 movl %edx,%esi 1116 rorl $11,%r14d 1117 xorl %ebx,%r12d 1118 xorl %r8d,%esi 1119 rorl $6,%r13d 1120 addl %r12d,%ecx 1121 andl %esi,%r15d 1122 xorl %edx,%r14d 1123 addl %r13d,%ecx 1124 xorl %r8d,%r15d 1125 addl %ecx,%r10d 1126 rorl $2,%r14d 1127 addl %r15d,%ecx 1128 movl %r10d,%r13d 1129 addl %ecx,%r14d 1130 rorl $14,%r13d 1131 movl %r14d,%ecx 1132 movl %r11d,%r12d 1133 xorl %r10d,%r13d 1134 rorl $9,%r14d 1135 xorl %eax,%r12d 1136 rorl $5,%r13d 1137 xorl %ecx,%r14d 1138 andl %r10d,%r12d 1139 vpand %xmm13,%xmm11,%xmm11 1140 vaesenc %xmm10,%xmm9,%xmm9 1141 vmovdqu 224-128(%rdi),%xmm10 1142 xorl %r10d,%r13d 1143 addl 56(%rsp),%ebx 1144 movl %ecx,%r15d 1145 rorl $11,%r14d 1146 xorl %eax,%r12d 1147 xorl %edx,%r15d 1148 rorl $6,%r13d 1149 addl %r12d,%ebx 1150 andl %r15d,%esi 1151 xorl %ecx,%r14d 1152 addl %r13d,%ebx 1153 xorl %edx,%esi 1154 addl %ebx,%r9d 1155 rorl $2,%r14d 1156 addl %esi,%ebx 1157 movl %r9d,%r13d 1158 addl %ebx,%r14d 1159 rorl $14,%r13d 1160 movl %r14d,%ebx 1161 movl %r10d,%r12d 1162 xorl %r9d,%r13d 1163 rorl $9,%r14d 1164 xorl %r11d,%r12d 1165 rorl $5,%r13d 1166 xorl %ebx,%r14d 1167 andl %r9d,%r12d 1168 vpor %xmm11,%xmm8,%xmm8 1169 vaesenclast %xmm10,%xmm9,%xmm11 1170 vmovdqu 0-128(%rdi),%xmm10 1171 xorl %r9d,%r13d 1172 addl 60(%rsp),%eax 1173 movl %ebx,%esi 1174 rorl $11,%r14d 1175 xorl %r11d,%r12d 1176 xorl %ecx,%esi 1177 rorl $6,%r13d 1178 addl %r12d,%eax 1179 andl %esi,%r15d 1180 xorl %ebx,%r14d 1181 addl %r13d,%eax 1182 xorl %ecx,%r15d 1183 addl %eax,%r8d 1184 rorl $2,%r14d 1185 addl %r15d,%eax 1186 movl %r8d,%r13d 1187 addl %eax,%r14d 1188 movq 64+0(%rsp),%r12 1189 movq 64+8(%rsp),%r13 1190 movq 64+40(%rsp),%r15 1191 movq 64+48(%rsp),%rsi 1192 1193 vpand %xmm14,%xmm11,%xmm11 1194 movl %r14d,%eax 1195 vpor %xmm11,%xmm8,%xmm8 1196 vmovdqu %xmm8,(%r12,%r13,1) 1197 leaq 16(%r12),%r12 1198 1199 addl 0(%r15),%eax 1200 addl 4(%r15),%ebx 1201 addl 8(%r15),%ecx 1202 addl 12(%r15),%edx 1203 addl 16(%r15),%r8d 1204 addl 20(%r15),%r9d 1205 addl 24(%r15),%r10d 1206 addl 28(%r15),%r11d 1207 1208 cmpq 64+16(%rsp),%r12 1209 1210 movl %eax,0(%r15) 1211 movl %ebx,4(%r15) 1212 movl %ecx,8(%r15) 1213 movl %edx,12(%r15) 1214 movl %r8d,16(%r15) 1215 movl %r9d,20(%r15) 1216 movl %r10d,24(%r15) 1217 movl %r11d,28(%r15) 1218 1219 jb .Lloop_xop 1220 1221 movq 64+32(%rsp),%r8 1222 movq 120(%rsp),%rsi 1223 .cfi_def_cfa %rsi,8 1224 vmovdqu %xmm8,(%r8) 1225 vzeroall 1226 movq -48(%rsi),%r15 1227 .cfi_restore %r15 1228 movq -40(%rsi),%r14 1229 .cfi_restore %r14 1230 movq -32(%rsi),%r13 1231 .cfi_restore %r13 1232 movq -24(%rsi),%r12 1233 .cfi_restore %r12 1234 movq -16(%rsi),%rbp 1235 .cfi_restore %rbp 1236 movq -8(%rsi),%rbx 1237 .cfi_restore %rbx 1238 leaq (%rsi),%rsp 1239 .cfi_def_cfa_register %rsp 1240 .Lepilogue_xop: 1241 .byte 0xf3,0xc3 1242 .cfi_endproc 1243 .size aesni_cbc_sha256_enc_xop,.-aesni_cbc_sha256_enc_xop 1244 .type aesni_cbc_sha256_enc_avx,@function 1245 .align 64 1246 aesni_cbc_sha256_enc_avx: 1247 .cfi_startproc 1248 .Lavx_shortcut: 1249 movq 8(%rsp),%r10 1250 movq %rsp,%rax 1251 .cfi_def_cfa_register %rax 1252 pushq %rbx 1253 .cfi_offset %rbx,-16 1254 pushq %rbp 1255 .cfi_offset %rbp,-24 1256 pushq %r12 1257 .cfi_offset %r12,-32 1258 pushq %r13 1259 .cfi_offset %r13,-40 1260 pushq %r14 1261 .cfi_offset %r14,-48 1262 pushq %r15 1263 .cfi_offset %r15,-56 1264 subq $128,%rsp 1265 andq $-64,%rsp 1266 1267 shlq $6,%rdx 1268 subq %rdi,%rsi 1269 subq %rdi,%r10 1270 addq %rdi,%rdx 1271 1272 1273 movq %rsi,64+8(%rsp) 1274 movq %rdx,64+16(%rsp) 1275 1276 movq %r8,64+32(%rsp) 1277 movq %r9,64+40(%rsp) 1278 movq %r10,64+48(%rsp) 1279 movq %rax,120(%rsp) 1280 .cfi_escape 0x0f,0x06,0x77,0xf8,0x00,0x06,0x23,0x08 1281 .Lprologue_avx: 1282 vzeroall 1283 1284 movq %rdi,%r12 1285 leaq 128(%rcx),%rdi 1286 leaq K256+544(%rip),%r13 1287 movl 240-128(%rdi),%r14d 1288 movq %r9,%r15 1289 movq %r10,%rsi 1290 vmovdqu (%r8),%xmm8 1291 subq $9,%r14 1292 1293 movl 0(%r15),%eax 1294 movl 4(%r15),%ebx 1295 movl 8(%r15),%ecx 1296 movl 12(%r15),%edx 1297 movl 16(%r15),%r8d 1298 movl 20(%r15),%r9d 1299 movl 24(%r15),%r10d 1300 movl 28(%r15),%r11d 1301 1302 vmovdqa 0(%r13,%r14,8),%xmm14 1303 vmovdqa 16(%r13,%r14,8),%xmm13 1304 vmovdqa 32(%r13,%r14,8),%xmm12 1305 vmovdqu 0-128(%rdi),%xmm10 1306 jmp .Lloop_avx 1307 .align 16 1308 .Lloop_avx: 1309 vmovdqa K256+512(%rip),%xmm7 1310 vmovdqu 0(%rsi,%r12,1),%xmm0 1311 vmovdqu 16(%rsi,%r12,1),%xmm1 1312 vmovdqu 32(%rsi,%r12,1),%xmm2 1313 vmovdqu 48(%rsi,%r12,1),%xmm3 1314 vpshufb %xmm7,%xmm0,%xmm0 1315 leaq K256(%rip),%rbp 1316 vpshufb %xmm7,%xmm1,%xmm1 1317 vpshufb %xmm7,%xmm2,%xmm2 1318 vpaddd 0(%rbp),%xmm0,%xmm4 1319 vpshufb %xmm7,%xmm3,%xmm3 1320 vpaddd 32(%rbp),%xmm1,%xmm5 1321 vpaddd 64(%rbp),%xmm2,%xmm6 1322 vpaddd 96(%rbp),%xmm3,%xmm7 1323 vmovdqa %xmm4,0(%rsp) 1324 movl %eax,%r14d 1325 vmovdqa %xmm5,16(%rsp) 1326 movl %ebx,%esi 1327 vmovdqa %xmm6,32(%rsp) 1328 xorl %ecx,%esi 1329 vmovdqa %xmm7,48(%rsp) 1330 movl %r8d,%r13d 1331 jmp .Lavx_00_47 1332 1333 .align 16 1334 .Lavx_00_47: 1335 subq $-32*4,%rbp 1336 vmovdqu (%r12),%xmm9 1337 movq %r12,64+0(%rsp) 1338 vpalignr $4,%xmm0,%xmm1,%xmm4 1339 shrdl $14,%r13d,%r13d 1340 movl %r14d,%eax 1341 movl %r9d,%r12d 1342 vpalignr $4,%xmm2,%xmm3,%xmm7 1343 xorl %r8d,%r13d 1344 shrdl $9,%r14d,%r14d 1345 xorl %r10d,%r12d 1346 vpsrld $7,%xmm4,%xmm6 1347 shrdl $5,%r13d,%r13d 1348 xorl %eax,%r14d 1349 andl %r8d,%r12d 1350 vpaddd %xmm7,%xmm0,%xmm0 1351 vpxor %xmm10,%xmm9,%xmm9 1352 vmovdqu 16-128(%rdi),%xmm10 1353 xorl %r8d,%r13d 1354 addl 0(%rsp),%r11d 1355 movl %eax,%r15d 1356 vpsrld $3,%xmm4,%xmm7 1357 shrdl $11,%r14d,%r14d 1358 xorl %r10d,%r12d 1359 xorl %ebx,%r15d 1360 vpslld $14,%xmm4,%xmm5 1361 shrdl $6,%r13d,%r13d 1362 addl %r12d,%r11d 1363 andl %r15d,%esi 1364 vpxor %xmm6,%xmm7,%xmm4 1365 xorl %eax,%r14d 1366 addl %r13d,%r11d 1367 xorl %ebx,%esi 1368 vpshufd $250,%xmm3,%xmm7 1369 addl %r11d,%edx 1370 shrdl $2,%r14d,%r14d 1371 addl %esi,%r11d 1372 vpsrld $11,%xmm6,%xmm6 1373 movl %edx,%r13d 1374 addl %r11d,%r14d 1375 shrdl $14,%r13d,%r13d 1376 vpxor %xmm5,%xmm4,%xmm4 1377 movl %r14d,%r11d 1378 movl %r8d,%r12d 1379 xorl %edx,%r13d 1380 vpslld $11,%xmm5,%xmm5 1381 shrdl $9,%r14d,%r14d 1382 xorl %r9d,%r12d 1383 shrdl $5,%r13d,%r13d 1384 vpxor %xmm6,%xmm4,%xmm4 1385 xorl %r11d,%r14d 1386 andl %edx,%r12d 1387 vpxor %xmm8,%xmm9,%xmm9 1388 xorl %edx,%r13d 1389 vpsrld $10,%xmm7,%xmm6 1390 addl 4(%rsp),%r10d 1391 movl %r11d,%esi 1392 shrdl $11,%r14d,%r14d 1393 vpxor %xmm5,%xmm4,%xmm4 1394 xorl %r9d,%r12d 1395 xorl %eax,%esi 1396 shrdl $6,%r13d,%r13d 1397 vpsrlq $17,%xmm7,%xmm7 1398 addl %r12d,%r10d 1399 andl %esi,%r15d 1400 xorl %r11d,%r14d 1401 vpaddd %xmm4,%xmm0,%xmm0 1402 addl %r13d,%r10d 1403 xorl %eax,%r15d 1404 addl %r10d,%ecx 1405 vpxor %xmm7,%xmm6,%xmm6 1406 shrdl $2,%r14d,%r14d 1407 addl %r15d,%r10d 1408 movl %ecx,%r13d 1409 vpsrlq $2,%xmm7,%xmm7 1410 addl %r10d,%r14d 1411 shrdl $14,%r13d,%r13d 1412 movl %r14d,%r10d 1413 vpxor %xmm7,%xmm6,%xmm6 1414 movl %edx,%r12d 1415 xorl %ecx,%r13d 1416 shrdl $9,%r14d,%r14d 1417 vpshufd $132,%xmm6,%xmm6 1418 xorl %r8d,%r12d 1419 shrdl $5,%r13d,%r13d 1420 xorl %r10d,%r14d 1421 vpsrldq $8,%xmm6,%xmm6 1422 andl %ecx,%r12d 1423 vaesenc %xmm10,%xmm9,%xmm9 1424 vmovdqu 32-128(%rdi),%xmm10 1425 xorl %ecx,%r13d 1426 addl 8(%rsp),%r9d 1427 vpaddd %xmm6,%xmm0,%xmm0 1428 movl %r10d,%r15d 1429 shrdl $11,%r14d,%r14d 1430 xorl %r8d,%r12d 1431 vpshufd $80,%xmm0,%xmm7 1432 xorl %r11d,%r15d 1433 shrdl $6,%r13d,%r13d 1434 addl %r12d,%r9d 1435 vpsrld $10,%xmm7,%xmm6 1436 andl %r15d,%esi 1437 xorl %r10d,%r14d 1438 addl %r13d,%r9d 1439 vpsrlq $17,%xmm7,%xmm7 1440 xorl %r11d,%esi 1441 addl %r9d,%ebx 1442 shrdl $2,%r14d,%r14d 1443 vpxor %xmm7,%xmm6,%xmm6 1444 addl %esi,%r9d 1445 movl %ebx,%r13d 1446 addl %r9d,%r14d 1447 vpsrlq $2,%xmm7,%xmm7 1448 shrdl $14,%r13d,%r13d 1449 movl %r14d,%r9d 1450 movl %ecx,%r12d 1451 vpxor %xmm7,%xmm6,%xmm6 1452 xorl %ebx,%r13d 1453 shrdl $9,%r14d,%r14d 1454 xorl %edx,%r12d 1455 vpshufd $232,%xmm6,%xmm6 1456 shrdl $5,%r13d,%r13d 1457 xorl %r9d,%r14d 1458 andl %ebx,%r12d 1459 vpslldq $8,%xmm6,%xmm6 1460 vaesenc %xmm10,%xmm9,%xmm9 1461 vmovdqu 48-128(%rdi),%xmm10 1462 xorl %ebx,%r13d 1463 addl 12(%rsp),%r8d 1464 movl %r9d,%esi 1465 vpaddd %xmm6,%xmm0,%xmm0 1466 shrdl $11,%r14d,%r14d 1467 xorl %edx,%r12d 1468 xorl %r10d,%esi 1469 vpaddd 0(%rbp),%xmm0,%xmm6 1470 shrdl $6,%r13d,%r13d 1471 addl %r12d,%r8d 1472 andl %esi,%r15d 1473 xorl %r9d,%r14d 1474 addl %r13d,%r8d 1475 xorl %r10d,%r15d 1476 addl %r8d,%eax 1477 shrdl $2,%r14d,%r14d 1478 addl %r15d,%r8d 1479 movl %eax,%r13d 1480 addl %r8d,%r14d 1481 vmovdqa %xmm6,0(%rsp) 1482 vpalignr $4,%xmm1,%xmm2,%xmm4 1483 shrdl $14,%r13d,%r13d 1484 movl %r14d,%r8d 1485 movl %ebx,%r12d 1486 vpalignr $4,%xmm3,%xmm0,%xmm7 1487 xorl %eax,%r13d 1488 shrdl $9,%r14d,%r14d 1489 xorl %ecx,%r12d 1490 vpsrld $7,%xmm4,%xmm6 1491 shrdl $5,%r13d,%r13d 1492 xorl %r8d,%r14d 1493 andl %eax,%r12d 1494 vpaddd %xmm7,%xmm1,%xmm1 1495 vaesenc %xmm10,%xmm9,%xmm9 1496 vmovdqu 64-128(%rdi),%xmm10 1497 xorl %eax,%r13d 1498 addl 16(%rsp),%edx 1499 movl %r8d,%r15d 1500 vpsrld $3,%xmm4,%xmm7 1501 shrdl $11,%r14d,%r14d 1502 xorl %ecx,%r12d 1503 xorl %r9d,%r15d 1504 vpslld $14,%xmm4,%xmm5 1505 shrdl $6,%r13d,%r13d 1506 addl %r12d,%edx 1507 andl %r15d,%esi 1508 vpxor %xmm6,%xmm7,%xmm4 1509 xorl %r8d,%r14d 1510 addl %r13d,%edx 1511 xorl %r9d,%esi 1512 vpshufd $250,%xmm0,%xmm7 1513 addl %edx,%r11d 1514 shrdl $2,%r14d,%r14d 1515 addl %esi,%edx 1516 vpsrld $11,%xmm6,%xmm6 1517 movl %r11d,%r13d 1518 addl %edx,%r14d 1519 shrdl $14,%r13d,%r13d 1520 vpxor %xmm5,%xmm4,%xmm4 1521 movl %r14d,%edx 1522 movl %eax,%r12d 1523 xorl %r11d,%r13d 1524 vpslld $11,%xmm5,%xmm5 1525 shrdl $9,%r14d,%r14d 1526 xorl %ebx,%r12d 1527 shrdl $5,%r13d,%r13d 1528 vpxor %xmm6,%xmm4,%xmm4 1529 xorl %edx,%r14d 1530 andl %r11d,%r12d 1531 vaesenc %xmm10,%xmm9,%xmm9 1532 vmovdqu 80-128(%rdi),%xmm10 1533 xorl %r11d,%r13d 1534 vpsrld $10,%xmm7,%xmm6 1535 addl 20(%rsp),%ecx 1536 movl %edx,%esi 1537 shrdl $11,%r14d,%r14d 1538 vpxor %xmm5,%xmm4,%xmm4 1539 xorl %ebx,%r12d 1540 xorl %r8d,%esi 1541 shrdl $6,%r13d,%r13d 1542 vpsrlq $17,%xmm7,%xmm7 1543 addl %r12d,%ecx 1544 andl %esi,%r15d 1545 xorl %edx,%r14d 1546 vpaddd %xmm4,%xmm1,%xmm1 1547 addl %r13d,%ecx 1548 xorl %r8d,%r15d 1549 addl %ecx,%r10d 1550 vpxor %xmm7,%xmm6,%xmm6 1551 shrdl $2,%r14d,%r14d 1552 addl %r15d,%ecx 1553 movl %r10d,%r13d 1554 vpsrlq $2,%xmm7,%xmm7 1555 addl %ecx,%r14d 1556 shrdl $14,%r13d,%r13d 1557 movl %r14d,%ecx 1558 vpxor %xmm7,%xmm6,%xmm6 1559 movl %r11d,%r12d 1560 xorl %r10d,%r13d 1561 shrdl $9,%r14d,%r14d 1562 vpshufd $132,%xmm6,%xmm6 1563 xorl %eax,%r12d 1564 shrdl $5,%r13d,%r13d 1565 xorl %ecx,%r14d 1566 vpsrldq $8,%xmm6,%xmm6 1567 andl %r10d,%r12d 1568 vaesenc %xmm10,%xmm9,%xmm9 1569 vmovdqu 96-128(%rdi),%xmm10 1570 xorl %r10d,%r13d 1571 addl 24(%rsp),%ebx 1572 vpaddd %xmm6,%xmm1,%xmm1 1573 movl %ecx,%r15d 1574 shrdl $11,%r14d,%r14d 1575 xorl %eax,%r12d 1576 vpshufd $80,%xmm1,%xmm7 1577 xorl %edx,%r15d 1578 shrdl $6,%r13d,%r13d 1579 addl %r12d,%ebx 1580 vpsrld $10,%xmm7,%xmm6 1581 andl %r15d,%esi 1582 xorl %ecx,%r14d 1583 addl %r13d,%ebx 1584 vpsrlq $17,%xmm7,%xmm7 1585 xorl %edx,%esi 1586 addl %ebx,%r9d 1587 shrdl $2,%r14d,%r14d 1588 vpxor %xmm7,%xmm6,%xmm6 1589 addl %esi,%ebx 1590 movl %r9d,%r13d 1591 addl %ebx,%r14d 1592 vpsrlq $2,%xmm7,%xmm7 1593 shrdl $14,%r13d,%r13d 1594 movl %r14d,%ebx 1595 movl %r10d,%r12d 1596 vpxor %xmm7,%xmm6,%xmm6 1597 xorl %r9d,%r13d 1598 shrdl $9,%r14d,%r14d 1599 xorl %r11d,%r12d 1600 vpshufd $232,%xmm6,%xmm6 1601 shrdl $5,%r13d,%r13d 1602 xorl %ebx,%r14d 1603 andl %r9d,%r12d 1604 vpslldq $8,%xmm6,%xmm6 1605 vaesenc %xmm10,%xmm9,%xmm9 1606 vmovdqu 112-128(%rdi),%xmm10 1607 xorl %r9d,%r13d 1608 addl 28(%rsp),%eax 1609 movl %ebx,%esi 1610 vpaddd %xmm6,%xmm1,%xmm1 1611 shrdl $11,%r14d,%r14d 1612 xorl %r11d,%r12d 1613 xorl %ecx,%esi 1614 vpaddd 32(%rbp),%xmm1,%xmm6 1615 shrdl $6,%r13d,%r13d 1616 addl %r12d,%eax 1617 andl %esi,%r15d 1618 xorl %ebx,%r14d 1619 addl %r13d,%eax 1620 xorl %ecx,%r15d 1621 addl %eax,%r8d 1622 shrdl $2,%r14d,%r14d 1623 addl %r15d,%eax 1624 movl %r8d,%r13d 1625 addl %eax,%r14d 1626 vmovdqa %xmm6,16(%rsp) 1627 vpalignr $4,%xmm2,%xmm3,%xmm4 1628 shrdl $14,%r13d,%r13d 1629 movl %r14d,%eax 1630 movl %r9d,%r12d 1631 vpalignr $4,%xmm0,%xmm1,%xmm7 1632 xorl %r8d,%r13d 1633 shrdl $9,%r14d,%r14d 1634 xorl %r10d,%r12d 1635 vpsrld $7,%xmm4,%xmm6 1636 shrdl $5,%r13d,%r13d 1637 xorl %eax,%r14d 1638 andl %r8d,%r12d 1639 vpaddd %xmm7,%xmm2,%xmm2 1640 vaesenc %xmm10,%xmm9,%xmm9 1641 vmovdqu 128-128(%rdi),%xmm10 1642 xorl %r8d,%r13d 1643 addl 32(%rsp),%r11d 1644 movl %eax,%r15d 1645 vpsrld $3,%xmm4,%xmm7 1646 shrdl $11,%r14d,%r14d 1647 xorl %r10d,%r12d 1648 xorl %ebx,%r15d 1649 vpslld $14,%xmm4,%xmm5 1650 shrdl $6,%r13d,%r13d 1651 addl %r12d,%r11d 1652 andl %r15d,%esi 1653 vpxor %xmm6,%xmm7,%xmm4 1654 xorl %eax,%r14d 1655 addl %r13d,%r11d 1656 xorl %ebx,%esi 1657 vpshufd $250,%xmm1,%xmm7 1658 addl %r11d,%edx 1659 shrdl $2,%r14d,%r14d 1660 addl %esi,%r11d 1661 vpsrld $11,%xmm6,%xmm6 1662 movl %edx,%r13d 1663 addl %r11d,%r14d 1664 shrdl $14,%r13d,%r13d 1665 vpxor %xmm5,%xmm4,%xmm4 1666 movl %r14d,%r11d 1667 movl %r8d,%r12d 1668 xorl %edx,%r13d 1669 vpslld $11,%xmm5,%xmm5 1670 shrdl $9,%r14d,%r14d 1671 xorl %r9d,%r12d 1672 shrdl $5,%r13d,%r13d 1673 vpxor %xmm6,%xmm4,%xmm4 1674 xorl %r11d,%r14d 1675 andl %edx,%r12d 1676 vaesenc %xmm10,%xmm9,%xmm9 1677 vmovdqu 144-128(%rdi),%xmm10 1678 xorl %edx,%r13d 1679 vpsrld $10,%xmm7,%xmm6 1680 addl 36(%rsp),%r10d 1681 movl %r11d,%esi 1682 shrdl $11,%r14d,%r14d 1683 vpxor %xmm5,%xmm4,%xmm4 1684 xorl %r9d,%r12d 1685 xorl %eax,%esi 1686 shrdl $6,%r13d,%r13d 1687 vpsrlq $17,%xmm7,%xmm7 1688 addl %r12d,%r10d 1689 andl %esi,%r15d 1690 xorl %r11d,%r14d 1691 vpaddd %xmm4,%xmm2,%xmm2 1692 addl %r13d,%r10d 1693 xorl %eax,%r15d 1694 addl %r10d,%ecx 1695 vpxor %xmm7,%xmm6,%xmm6 1696 shrdl $2,%r14d,%r14d 1697 addl %r15d,%r10d 1698 movl %ecx,%r13d 1699 vpsrlq $2,%xmm7,%xmm7 1700 addl %r10d,%r14d 1701 shrdl $14,%r13d,%r13d 1702 movl %r14d,%r10d 1703 vpxor %xmm7,%xmm6,%xmm6 1704 movl %edx,%r12d 1705 xorl %ecx,%r13d 1706 shrdl $9,%r14d,%r14d 1707 vpshufd $132,%xmm6,%xmm6 1708 xorl %r8d,%r12d 1709 shrdl $5,%r13d,%r13d 1710 xorl %r10d,%r14d 1711 vpsrldq $8,%xmm6,%xmm6 1712 andl %ecx,%r12d 1713 vaesenc %xmm10,%xmm9,%xmm9 1714 vmovdqu 160-128(%rdi),%xmm10 1715 xorl %ecx,%r13d 1716 addl 40(%rsp),%r9d 1717 vpaddd %xmm6,%xmm2,%xmm2 1718 movl %r10d,%r15d 1719 shrdl $11,%r14d,%r14d 1720 xorl %r8d,%r12d 1721 vpshufd $80,%xmm2,%xmm7 1722 xorl %r11d,%r15d 1723 shrdl $6,%r13d,%r13d 1724 addl %r12d,%r9d 1725 vpsrld $10,%xmm7,%xmm6 1726 andl %r15d,%esi 1727 xorl %r10d,%r14d 1728 addl %r13d,%r9d 1729 vpsrlq $17,%xmm7,%xmm7 1730 xorl %r11d,%esi 1731 addl %r9d,%ebx 1732 shrdl $2,%r14d,%r14d 1733 vpxor %xmm7,%xmm6,%xmm6 1734 addl %esi,%r9d 1735 movl %ebx,%r13d 1736 addl %r9d,%r14d 1737 vpsrlq $2,%xmm7,%xmm7 1738 shrdl $14,%r13d,%r13d 1739 movl %r14d,%r9d 1740 movl %ecx,%r12d 1741 vpxor %xmm7,%xmm6,%xmm6 1742 xorl %ebx,%r13d 1743 shrdl $9,%r14d,%r14d 1744 xorl %edx,%r12d 1745 vpshufd $232,%xmm6,%xmm6 1746 shrdl $5,%r13d,%r13d 1747 xorl %r9d,%r14d 1748 andl %ebx,%r12d 1749 vpslldq $8,%xmm6,%xmm6 1750 vaesenclast %xmm10,%xmm9,%xmm11 1751 vaesenc %xmm10,%xmm9,%xmm9 1752 vmovdqu 176-128(%rdi),%xmm10 1753 xorl %ebx,%r13d 1754 addl 44(%rsp),%r8d 1755 movl %r9d,%esi 1756 vpaddd %xmm6,%xmm2,%xmm2 1757 shrdl $11,%r14d,%r14d 1758 xorl %edx,%r12d 1759 xorl %r10d,%esi 1760 vpaddd 64(%rbp),%xmm2,%xmm6 1761 shrdl $6,%r13d,%r13d 1762 addl %r12d,%r8d 1763 andl %esi,%r15d 1764 xorl %r9d,%r14d 1765 addl %r13d,%r8d 1766 xorl %r10d,%r15d 1767 addl %r8d,%eax 1768 shrdl $2,%r14d,%r14d 1769 addl %r15d,%r8d 1770 movl %eax,%r13d 1771 addl %r8d,%r14d 1772 vmovdqa %xmm6,32(%rsp) 1773 vpalignr $4,%xmm3,%xmm0,%xmm4 1774 shrdl $14,%r13d,%r13d 1775 movl %r14d,%r8d 1776 movl %ebx,%r12d 1777 vpalignr $4,%xmm1,%xmm2,%xmm7 1778 xorl %eax,%r13d 1779 shrdl $9,%r14d,%r14d 1780 xorl %ecx,%r12d 1781 vpsrld $7,%xmm4,%xmm6 1782 shrdl $5,%r13d,%r13d 1783 xorl %r8d,%r14d 1784 andl %eax,%r12d 1785 vpaddd %xmm7,%xmm3,%xmm3 1786 vpand %xmm12,%xmm11,%xmm8 1787 vaesenc %xmm10,%xmm9,%xmm9 1788 vmovdqu 192-128(%rdi),%xmm10 1789 xorl %eax,%r13d 1790 addl 48(%rsp),%edx 1791 movl %r8d,%r15d 1792 vpsrld $3,%xmm4,%xmm7 1793 shrdl $11,%r14d,%r14d 1794 xorl %ecx,%r12d 1795 xorl %r9d,%r15d 1796 vpslld $14,%xmm4,%xmm5 1797 shrdl $6,%r13d,%r13d 1798 addl %r12d,%edx 1799 andl %r15d,%esi 1800 vpxor %xmm6,%xmm7,%xmm4 1801 xorl %r8d,%r14d 1802 addl %r13d,%edx 1803 xorl %r9d,%esi 1804 vpshufd $250,%xmm2,%xmm7 1805 addl %edx,%r11d 1806 shrdl $2,%r14d,%r14d 1807 addl %esi,%edx 1808 vpsrld $11,%xmm6,%xmm6 1809 movl %r11d,%r13d 1810 addl %edx,%r14d 1811 shrdl $14,%r13d,%r13d 1812 vpxor %xmm5,%xmm4,%xmm4 1813 movl %r14d,%edx 1814 movl %eax,%r12d 1815 xorl %r11d,%r13d 1816 vpslld $11,%xmm5,%xmm5 1817 shrdl $9,%r14d,%r14d 1818 xorl %ebx,%r12d 1819 shrdl $5,%r13d,%r13d 1820 vpxor %xmm6,%xmm4,%xmm4 1821 xorl %edx,%r14d 1822 andl %r11d,%r12d 1823 vaesenclast %xmm10,%xmm9,%xmm11 1824 vaesenc %xmm10,%xmm9,%xmm9 1825 vmovdqu 208-128(%rdi),%xmm10 1826 xorl %r11d,%r13d 1827 vpsrld $10,%xmm7,%xmm6 1828 addl 52(%rsp),%ecx 1829 movl %edx,%esi 1830 shrdl $11,%r14d,%r14d 1831 vpxor %xmm5,%xmm4,%xmm4 1832 xorl %ebx,%r12d 1833 xorl %r8d,%esi 1834 shrdl $6,%r13d,%r13d 1835 vpsrlq $17,%xmm7,%xmm7 1836 addl %r12d,%ecx 1837 andl %esi,%r15d 1838 xorl %edx,%r14d 1839 vpaddd %xmm4,%xmm3,%xmm3 1840 addl %r13d,%ecx 1841 xorl %r8d,%r15d 1842 addl %ecx,%r10d 1843 vpxor %xmm7,%xmm6,%xmm6 1844 shrdl $2,%r14d,%r14d 1845 addl %r15d,%ecx 1846 movl %r10d,%r13d 1847 vpsrlq $2,%xmm7,%xmm7 1848 addl %ecx,%r14d 1849 shrdl $14,%r13d,%r13d 1850 movl %r14d,%ecx 1851 vpxor %xmm7,%xmm6,%xmm6 1852 movl %r11d,%r12d 1853 xorl %r10d,%r13d 1854 shrdl $9,%r14d,%r14d 1855 vpshufd $132,%xmm6,%xmm6 1856 xorl %eax,%r12d 1857 shrdl $5,%r13d,%r13d 1858 xorl %ecx,%r14d 1859 vpsrldq $8,%xmm6,%xmm6 1860 andl %r10d,%r12d 1861 vpand %xmm13,%xmm11,%xmm11 1862 vaesenc %xmm10,%xmm9,%xmm9 1863 vmovdqu 224-128(%rdi),%xmm10 1864 xorl %r10d,%r13d 1865 addl 56(%rsp),%ebx 1866 vpaddd %xmm6,%xmm3,%xmm3 1867 movl %ecx,%r15d 1868 shrdl $11,%r14d,%r14d 1869 xorl %eax,%r12d 1870 vpshufd $80,%xmm3,%xmm7 1871 xorl %edx,%r15d 1872 shrdl $6,%r13d,%r13d 1873 addl %r12d,%ebx 1874 vpsrld $10,%xmm7,%xmm6 1875 andl %r15d,%esi 1876 xorl %ecx,%r14d 1877 addl %r13d,%ebx 1878 vpsrlq $17,%xmm7,%xmm7 1879 xorl %edx,%esi 1880 addl %ebx,%r9d 1881 shrdl $2,%r14d,%r14d 1882 vpxor %xmm7,%xmm6,%xmm6 1883 addl %esi,%ebx 1884 movl %r9d,%r13d 1885 addl %ebx,%r14d 1886 vpsrlq $2,%xmm7,%xmm7 1887 shrdl $14,%r13d,%r13d 1888 movl %r14d,%ebx 1889 movl %r10d,%r12d 1890 vpxor %xmm7,%xmm6,%xmm6 1891 xorl %r9d,%r13d 1892 shrdl $9,%r14d,%r14d 1893 xorl %r11d,%r12d 1894 vpshufd $232,%xmm6,%xmm6 1895 shrdl $5,%r13d,%r13d 1896 xorl %ebx,%r14d 1897 andl %r9d,%r12d 1898 vpslldq $8,%xmm6,%xmm6 1899 vpor %xmm11,%xmm8,%xmm8 1900 vaesenclast %xmm10,%xmm9,%xmm11 1901 vmovdqu 0-128(%rdi),%xmm10 1902 xorl %r9d,%r13d 1903 addl 60(%rsp),%eax 1904 movl %ebx,%esi 1905 vpaddd %xmm6,%xmm3,%xmm3 1906 shrdl $11,%r14d,%r14d 1907 xorl %r11d,%r12d 1908 xorl %ecx,%esi 1909 vpaddd 96(%rbp),%xmm3,%xmm6 1910 shrdl $6,%r13d,%r13d 1911 addl %r12d,%eax 1912 andl %esi,%r15d 1913 xorl %ebx,%r14d 1914 addl %r13d,%eax 1915 xorl %ecx,%r15d 1916 addl %eax,%r8d 1917 shrdl $2,%r14d,%r14d 1918 addl %r15d,%eax 1919 movl %r8d,%r13d 1920 addl %eax,%r14d 1921 vmovdqa %xmm6,48(%rsp) 1922 movq 64+0(%rsp),%r12 1923 vpand %xmm14,%xmm11,%xmm11 1924 movq 64+8(%rsp),%r15 1925 vpor %xmm11,%xmm8,%xmm8 1926 vmovdqu %xmm8,(%r15,%r12,1) 1927 leaq 16(%r12),%r12 1928 cmpb $0,131(%rbp) 1929 jne .Lavx_00_47 1930 vmovdqu (%r12),%xmm9 1931 movq %r12,64+0(%rsp) 1932 shrdl $14,%r13d,%r13d 1933 movl %r14d,%eax 1934 movl %r9d,%r12d 1935 xorl %r8d,%r13d 1936 shrdl $9,%r14d,%r14d 1937 xorl %r10d,%r12d 1938 shrdl $5,%r13d,%r13d 1939 xorl %eax,%r14d 1940 andl %r8d,%r12d 1941 vpxor %xmm10,%xmm9,%xmm9 1942 vmovdqu 16-128(%rdi),%xmm10 1943 xorl %r8d,%r13d 1944 addl 0(%rsp),%r11d 1945 movl %eax,%r15d 1946 shrdl $11,%r14d,%r14d 1947 xorl %r10d,%r12d 1948 xorl %ebx,%r15d 1949 shrdl $6,%r13d,%r13d 1950 addl %r12d,%r11d 1951 andl %r15d,%esi 1952 xorl %eax,%r14d 1953 addl %r13d,%r11d 1954 xorl %ebx,%esi 1955 addl %r11d,%edx 1956 shrdl $2,%r14d,%r14d 1957 addl %esi,%r11d 1958 movl %edx,%r13d 1959 addl %r11d,%r14d 1960 shrdl $14,%r13d,%r13d 1961 movl %r14d,%r11d 1962 movl %r8d,%r12d 1963 xorl %edx,%r13d 1964 shrdl $9,%r14d,%r14d 1965 xorl %r9d,%r12d 1966 shrdl $5,%r13d,%r13d 1967 xorl %r11d,%r14d 1968 andl %edx,%r12d 1969 vpxor %xmm8,%xmm9,%xmm9 1970 xorl %edx,%r13d 1971 addl 4(%rsp),%r10d 1972 movl %r11d,%esi 1973 shrdl $11,%r14d,%r14d 1974 xorl %r9d,%r12d 1975 xorl %eax,%esi 1976 shrdl $6,%r13d,%r13d 1977 addl %r12d,%r10d 1978 andl %esi,%r15d 1979 xorl %r11d,%r14d 1980 addl %r13d,%r10d 1981 xorl %eax,%r15d 1982 addl %r10d,%ecx 1983 shrdl $2,%r14d,%r14d 1984 addl %r15d,%r10d 1985 movl %ecx,%r13d 1986 addl %r10d,%r14d 1987 shrdl $14,%r13d,%r13d 1988 movl %r14d,%r10d 1989 movl %edx,%r12d 1990 xorl %ecx,%r13d 1991 shrdl $9,%r14d,%r14d 1992 xorl %r8d,%r12d 1993 shrdl $5,%r13d,%r13d 1994 xorl %r10d,%r14d 1995 andl %ecx,%r12d 1996 vaesenc %xmm10,%xmm9,%xmm9 1997 vmovdqu 32-128(%rdi),%xmm10 1998 xorl %ecx,%r13d 1999 addl 8(%rsp),%r9d 2000 movl %r10d,%r15d 2001 shrdl $11,%r14d,%r14d 2002 xorl %r8d,%r12d 2003 xorl %r11d,%r15d 2004 shrdl $6,%r13d,%r13d 2005 addl %r12d,%r9d 2006 andl %r15d,%esi 2007 xorl %r10d,%r14d 2008 addl %r13d,%r9d 2009 xorl %r11d,%esi 2010 addl %r9d,%ebx 2011 shrdl $2,%r14d,%r14d 2012 addl %esi,%r9d 2013 movl %ebx,%r13d 2014 addl %r9d,%r14d 2015 shrdl $14,%r13d,%r13d 2016 movl %r14d,%r9d 2017 movl %ecx,%r12d 2018 xorl %ebx,%r13d 2019 shrdl $9,%r14d,%r14d 2020 xorl %edx,%r12d 2021 shrdl $5,%r13d,%r13d 2022 xorl %r9d,%r14d 2023 andl %ebx,%r12d 2024 vaesenc %xmm10,%xmm9,%xmm9 2025 vmovdqu 48-128(%rdi),%xmm10 2026 xorl %ebx,%r13d 2027 addl 12(%rsp),%r8d 2028 movl %r9d,%esi 2029 shrdl $11,%r14d,%r14d 2030 xorl %edx,%r12d 2031 xorl %r10d,%esi 2032 shrdl $6,%r13d,%r13d 2033 addl %r12d,%r8d 2034 andl %esi,%r15d 2035 xorl %r9d,%r14d 2036 addl %r13d,%r8d 2037 xorl %r10d,%r15d 2038 addl %r8d,%eax 2039 shrdl $2,%r14d,%r14d 2040 addl %r15d,%r8d 2041 movl %eax,%r13d 2042 addl %r8d,%r14d 2043 shrdl $14,%r13d,%r13d 2044 movl %r14d,%r8d 2045 movl %ebx,%r12d 2046 xorl %eax,%r13d 2047 shrdl $9,%r14d,%r14d 2048 xorl %ecx,%r12d 2049 shrdl $5,%r13d,%r13d 2050 xorl %r8d,%r14d 2051 andl %eax,%r12d 2052 vaesenc %xmm10,%xmm9,%xmm9 2053 vmovdqu 64-128(%rdi),%xmm10 2054 xorl %eax,%r13d 2055 addl 16(%rsp),%edx 2056 movl %r8d,%r15d 2057 shrdl $11,%r14d,%r14d 2058 xorl %ecx,%r12d 2059 xorl %r9d,%r15d 2060 shrdl $6,%r13d,%r13d 2061 addl %r12d,%edx 2062 andl %r15d,%esi 2063 xorl %r8d,%r14d 2064 addl %r13d,%edx 2065 xorl %r9d,%esi 2066 addl %edx,%r11d 2067 shrdl $2,%r14d,%r14d 2068 addl %esi,%edx 2069 movl %r11d,%r13d 2070 addl %edx,%r14d 2071 shrdl $14,%r13d,%r13d 2072 movl %r14d,%edx 2073 movl %eax,%r12d 2074 xorl %r11d,%r13d 2075 shrdl $9,%r14d,%r14d 2076 xorl %ebx,%r12d 2077 shrdl $5,%r13d,%r13d 2078 xorl %edx,%r14d 2079 andl %r11d,%r12d 2080 vaesenc %xmm10,%xmm9,%xmm9 2081 vmovdqu 80-128(%rdi),%xmm10 2082 xorl %r11d,%r13d 2083 addl 20(%rsp),%ecx 2084 movl %edx,%esi 2085 shrdl $11,%r14d,%r14d 2086 xorl %ebx,%r12d 2087 xorl %r8d,%esi 2088 shrdl $6,%r13d,%r13d 2089 addl %r12d,%ecx 2090 andl %esi,%r15d 2091 xorl %edx,%r14d 2092 addl %r13d,%ecx 2093 xorl %r8d,%r15d 2094 addl %ecx,%r10d 2095 shrdl $2,%r14d,%r14d 2096 addl %r15d,%ecx 2097 movl %r10d,%r13d 2098 addl %ecx,%r14d 2099 shrdl $14,%r13d,%r13d 2100 movl %r14d,%ecx 2101 movl %r11d,%r12d 2102 xorl %r10d,%r13d 2103 shrdl $9,%r14d,%r14d 2104 xorl %eax,%r12d 2105 shrdl $5,%r13d,%r13d 2106 xorl %ecx,%r14d 2107 andl %r10d,%r12d 2108 vaesenc %xmm10,%xmm9,%xmm9 2109 vmovdqu 96-128(%rdi),%xmm10 2110 xorl %r10d,%r13d 2111 addl 24(%rsp),%ebx 2112 movl %ecx,%r15d 2113 shrdl $11,%r14d,%r14d 2114 xorl %eax,%r12d 2115 xorl %edx,%r15d 2116 shrdl $6,%r13d,%r13d 2117 addl %r12d,%ebx 2118 andl %r15d,%esi 2119 xorl %ecx,%r14d 2120 addl %r13d,%ebx 2121 xorl %edx,%esi 2122 addl %ebx,%r9d 2123 shrdl $2,%r14d,%r14d 2124 addl %esi,%ebx 2125 movl %r9d,%r13d 2126 addl %ebx,%r14d 2127 shrdl $14,%r13d,%r13d 2128 movl %r14d,%ebx 2129 movl %r10d,%r12d 2130 xorl %r9d,%r13d 2131 shrdl $9,%r14d,%r14d 2132 xorl %r11d,%r12d 2133 shrdl $5,%r13d,%r13d 2134 xorl %ebx,%r14d 2135 andl %r9d,%r12d 2136 vaesenc %xmm10,%xmm9,%xmm9 2137 vmovdqu 112-128(%rdi),%xmm10 2138 xorl %r9d,%r13d 2139 addl 28(%rsp),%eax 2140 movl %ebx,%esi 2141 shrdl $11,%r14d,%r14d 2142 xorl %r11d,%r12d 2143 xorl %ecx,%esi 2144 shrdl $6,%r13d,%r13d 2145 addl %r12d,%eax 2146 andl %esi,%r15d 2147 xorl %ebx,%r14d 2148 addl %r13d,%eax 2149 xorl %ecx,%r15d 2150 addl %eax,%r8d 2151 shrdl $2,%r14d,%r14d 2152 addl %r15d,%eax 2153 movl %r8d,%r13d 2154 addl %eax,%r14d 2155 shrdl $14,%r13d,%r13d 2156 movl %r14d,%eax 2157 movl %r9d,%r12d 2158 xorl %r8d,%r13d 2159 shrdl $9,%r14d,%r14d 2160 xorl %r10d,%r12d 2161 shrdl $5,%r13d,%r13d 2162 xorl %eax,%r14d 2163 andl %r8d,%r12d 2164 vaesenc %xmm10,%xmm9,%xmm9 2165 vmovdqu 128-128(%rdi),%xmm10 2166 xorl %r8d,%r13d 2167 addl 32(%rsp),%r11d 2168 movl %eax,%r15d 2169 shrdl $11,%r14d,%r14d 2170 xorl %r10d,%r12d 2171 xorl %ebx,%r15d 2172 shrdl $6,%r13d,%r13d 2173 addl %r12d,%r11d 2174 andl %r15d,%esi 2175 xorl %eax,%r14d 2176 addl %r13d,%r11d 2177 xorl %ebx,%esi 2178 addl %r11d,%edx 2179 shrdl $2,%r14d,%r14d 2180 addl %esi,%r11d 2181 movl %edx,%r13d 2182 addl %r11d,%r14d 2183 shrdl $14,%r13d,%r13d 2184 movl %r14d,%r11d 2185 movl %r8d,%r12d 2186 xorl %edx,%r13d 2187 shrdl $9,%r14d,%r14d 2188 xorl %r9d,%r12d 2189 shrdl $5,%r13d,%r13d 2190 xorl %r11d,%r14d 2191 andl %edx,%r12d 2192 vaesenc %xmm10,%xmm9,%xmm9 2193 vmovdqu 144-128(%rdi),%xmm10 2194 xorl %edx,%r13d 2195 addl 36(%rsp),%r10d 2196 movl %r11d,%esi 2197 shrdl $11,%r14d,%r14d 2198 xorl %r9d,%r12d 2199 xorl %eax,%esi 2200 shrdl $6,%r13d,%r13d 2201 addl %r12d,%r10d 2202 andl %esi,%r15d 2203 xorl %r11d,%r14d 2204 addl %r13d,%r10d 2205 xorl %eax,%r15d 2206 addl %r10d,%ecx 2207 shrdl $2,%r14d,%r14d 2208 addl %r15d,%r10d 2209 movl %ecx,%r13d 2210 addl %r10d,%r14d 2211 shrdl $14,%r13d,%r13d 2212 movl %r14d,%r10d 2213 movl %edx,%r12d 2214 xorl %ecx,%r13d 2215 shrdl $9,%r14d,%r14d 2216 xorl %r8d,%r12d 2217 shrdl $5,%r13d,%r13d 2218 xorl %r10d,%r14d 2219 andl %ecx,%r12d 2220 vaesenc %xmm10,%xmm9,%xmm9 2221 vmovdqu 160-128(%rdi),%xmm10 2222 xorl %ecx,%r13d 2223 addl 40(%rsp),%r9d 2224 movl %r10d,%r15d 2225 shrdl $11,%r14d,%r14d 2226 xorl %r8d,%r12d 2227 xorl %r11d,%r15d 2228 shrdl $6,%r13d,%r13d 2229 addl %r12d,%r9d 2230 andl %r15d,%esi 2231 xorl %r10d,%r14d 2232 addl %r13d,%r9d 2233 xorl %r11d,%esi 2234 addl %r9d,%ebx 2235 shrdl $2,%r14d,%r14d 2236 addl %esi,%r9d 2237 movl %ebx,%r13d 2238 addl %r9d,%r14d 2239 shrdl $14,%r13d,%r13d 2240 movl %r14d,%r9d 2241 movl %ecx,%r12d 2242 xorl %ebx,%r13d 2243 shrdl $9,%r14d,%r14d 2244 xorl %edx,%r12d 2245 shrdl $5,%r13d,%r13d 2246 xorl %r9d,%r14d 2247 andl %ebx,%r12d 2248 vaesenclast %xmm10,%xmm9,%xmm11 2249 vaesenc %xmm10,%xmm9,%xmm9 2250 vmovdqu 176-128(%rdi),%xmm10 2251 xorl %ebx,%r13d 2252 addl 44(%rsp),%r8d 2253 movl %r9d,%esi 2254 shrdl $11,%r14d,%r14d 2255 xorl %edx,%r12d 2256 xorl %r10d,%esi 2257 shrdl $6,%r13d,%r13d 2258 addl %r12d,%r8d 2259 andl %esi,%r15d 2260 xorl %r9d,%r14d 2261 addl %r13d,%r8d 2262 xorl %r10d,%r15d 2263 addl %r8d,%eax 2264 shrdl $2,%r14d,%r14d 2265 addl %r15d,%r8d 2266 movl %eax,%r13d 2267 addl %r8d,%r14d 2268 shrdl $14,%r13d,%r13d 2269 movl %r14d,%r8d 2270 movl %ebx,%r12d 2271 xorl %eax,%r13d 2272 shrdl $9,%r14d,%r14d 2273 xorl %ecx,%r12d 2274 shrdl $5,%r13d,%r13d 2275 xorl %r8d,%r14d 2276 andl %eax,%r12d 2277 vpand %xmm12,%xmm11,%xmm8 2278 vaesenc %xmm10,%xmm9,%xmm9 2279 vmovdqu 192-128(%rdi),%xmm10 2280 xorl %eax,%r13d 2281 addl 48(%rsp),%edx 2282 movl %r8d,%r15d 2283 shrdl $11,%r14d,%r14d 2284 xorl %ecx,%r12d 2285 xorl %r9d,%r15d 2286 shrdl $6,%r13d,%r13d 2287 addl %r12d,%edx 2288 andl %r15d,%esi 2289 xorl %r8d,%r14d 2290 addl %r13d,%edx 2291 xorl %r9d,%esi 2292 addl %edx,%r11d 2293 shrdl $2,%r14d,%r14d 2294 addl %esi,%edx 2295 movl %r11d,%r13d 2296 addl %edx,%r14d 2297 shrdl $14,%r13d,%r13d 2298 movl %r14d,%edx 2299 movl %eax,%r12d 2300 xorl %r11d,%r13d 2301 shrdl $9,%r14d,%r14d 2302 xorl %ebx,%r12d 2303 shrdl $5,%r13d,%r13d 2304 xorl %edx,%r14d 2305 andl %r11d,%r12d 2306 vaesenclast %xmm10,%xmm9,%xmm11 2307 vaesenc %xmm10,%xmm9,%xmm9 2308 vmovdqu 208-128(%rdi),%xmm10 2309 xorl %r11d,%r13d 2310 addl 52(%rsp),%ecx 2311 movl %edx,%esi 2312 shrdl $11,%r14d,%r14d 2313 xorl %ebx,%r12d 2314 xorl %r8d,%esi 2315 shrdl $6,%r13d,%r13d 2316 addl %r12d,%ecx 2317 andl %esi,%r15d 2318 xorl %edx,%r14d 2319 addl %r13d,%ecx 2320 xorl %r8d,%r15d 2321 addl %ecx,%r10d 2322 shrdl $2,%r14d,%r14d 2323 addl %r15d,%ecx 2324 movl %r10d,%r13d 2325 addl %ecx,%r14d 2326 shrdl $14,%r13d,%r13d 2327 movl %r14d,%ecx 2328 movl %r11d,%r12d 2329 xorl %r10d,%r13d 2330 shrdl $9,%r14d,%r14d 2331 xorl %eax,%r12d 2332 shrdl $5,%r13d,%r13d 2333 xorl %ecx,%r14d 2334 andl %r10d,%r12d 2335 vpand %xmm13,%xmm11,%xmm11 2336 vaesenc %xmm10,%xmm9,%xmm9 2337 vmovdqu 224-128(%rdi),%xmm10 2338 xorl %r10d,%r13d 2339 addl 56(%rsp),%ebx 2340 movl %ecx,%r15d 2341 shrdl $11,%r14d,%r14d 2342 xorl %eax,%r12d 2343 xorl %edx,%r15d 2344 shrdl $6,%r13d,%r13d 2345 addl %r12d,%ebx 2346 andl %r15d,%esi 2347 xorl %ecx,%r14d 2348 addl %r13d,%ebx 2349 xorl %edx,%esi 2350 addl %ebx,%r9d 2351 shrdl $2,%r14d,%r14d 2352 addl %esi,%ebx 2353 movl %r9d,%r13d 2354 addl %ebx,%r14d 2355 shrdl $14,%r13d,%r13d 2356 movl %r14d,%ebx 2357 movl %r10d,%r12d 2358 xorl %r9d,%r13d 2359 shrdl $9,%r14d,%r14d 2360 xorl %r11d,%r12d 2361 shrdl $5,%r13d,%r13d 2362 xorl %ebx,%r14d 2363 andl %r9d,%r12d 2364 vpor %xmm11,%xmm8,%xmm8 2365 vaesenclast %xmm10,%xmm9,%xmm11 2366 vmovdqu 0-128(%rdi),%xmm10 2367 xorl %r9d,%r13d 2368 addl 60(%rsp),%eax 2369 movl %ebx,%esi 2370 shrdl $11,%r14d,%r14d 2371 xorl %r11d,%r12d 2372 xorl %ecx,%esi 2373 shrdl $6,%r13d,%r13d 2374 addl %r12d,%eax 2375 andl %esi,%r15d 2376 xorl %ebx,%r14d 2377 addl %r13d,%eax 2378 xorl %ecx,%r15d 2379 addl %eax,%r8d 2380 shrdl $2,%r14d,%r14d 2381 addl %r15d,%eax 2382 movl %r8d,%r13d 2383 addl %eax,%r14d 2384 movq 64+0(%rsp),%r12 2385 movq 64+8(%rsp),%r13 2386 movq 64+40(%rsp),%r15 2387 movq 64+48(%rsp),%rsi 2388 2389 vpand %xmm14,%xmm11,%xmm11 2390 movl %r14d,%eax 2391 vpor %xmm11,%xmm8,%xmm8 2392 vmovdqu %xmm8,(%r12,%r13,1) 2393 leaq 16(%r12),%r12 2394 2395 addl 0(%r15),%eax 2396 addl 4(%r15),%ebx 2397 addl 8(%r15),%ecx 2398 addl 12(%r15),%edx 2399 addl 16(%r15),%r8d 2400 addl 20(%r15),%r9d 2401 addl 24(%r15),%r10d 2402 addl 28(%r15),%r11d 2403 2404 cmpq 64+16(%rsp),%r12 2405 2406 movl %eax,0(%r15) 2407 movl %ebx,4(%r15) 2408 movl %ecx,8(%r15) 2409 movl %edx,12(%r15) 2410 movl %r8d,16(%r15) 2411 movl %r9d,20(%r15) 2412 movl %r10d,24(%r15) 2413 movl %r11d,28(%r15) 2414 jb .Lloop_avx 2415 2416 movq 64+32(%rsp),%r8 2417 movq 120(%rsp),%rsi 2418 .cfi_def_cfa %rsi,8 2419 vmovdqu %xmm8,(%r8) 2420 vzeroall 2421 movq -48(%rsi),%r15 2422 .cfi_restore %r15 2423 movq -40(%rsi),%r14 2424 .cfi_restore %r14 2425 movq -32(%rsi),%r13 2426 .cfi_restore %r13 2427 movq -24(%rsi),%r12 2428 .cfi_restore %r12 2429 movq -16(%rsi),%rbp 2430 .cfi_restore %rbp 2431 movq -8(%rsi),%rbx 2432 .cfi_restore %rbx 2433 leaq (%rsi),%rsp 2434 .cfi_def_cfa_register %rsp 2435 .Lepilogue_avx: 2436 .byte 0xf3,0xc3 2437 .cfi_endproc 2438 .size aesni_cbc_sha256_enc_avx,.-aesni_cbc_sha256_enc_avx 2439 .type aesni_cbc_sha256_enc_avx2,@function 2440 .align 64 2441 aesni_cbc_sha256_enc_avx2: 2442 .cfi_startproc 2443 .Lavx2_shortcut: 2444 movq 8(%rsp),%r10 2445 movq %rsp,%rax 2446 .cfi_def_cfa_register %rax 2447 pushq %rbx 2448 .cfi_offset %rbx,-16 2449 pushq %rbp 2450 .cfi_offset %rbp,-24 2451 pushq %r12 2452 .cfi_offset %r12,-32 2453 pushq %r13 2454 .cfi_offset %r13,-40 2455 pushq %r14 2456 .cfi_offset %r14,-48 2457 pushq %r15 2458 .cfi_offset %r15,-56 2459 subq $576,%rsp 2460 andq $-1024,%rsp 2461 addq $448,%rsp 2462 2463 shlq $6,%rdx 2464 subq %rdi,%rsi 2465 subq %rdi,%r10 2466 addq %rdi,%rdx 2467 2468 2469 2470 movq %rdx,64+16(%rsp) 2471 2472 movq %r8,64+32(%rsp) 2473 movq %r9,64+40(%rsp) 2474 movq %r10,64+48(%rsp) 2475 movq %rax,120(%rsp) 2476 .cfi_escape 0x0f,0x06,0x77,0xf8,0x00,0x06,0x23,0x08 2477 .Lprologue_avx2: 2478 vzeroall 2479 2480 movq %rdi,%r13 2481 vpinsrq $1,%rsi,%xmm15,%xmm15 2482 leaq 128(%rcx),%rdi 2483 leaq K256+544(%rip),%r12 2484 movl 240-128(%rdi),%r14d 2485 movq %r9,%r15 2486 movq %r10,%rsi 2487 vmovdqu (%r8),%xmm8 2488 leaq -9(%r14),%r14 2489 2490 vmovdqa 0(%r12,%r14,8),%xmm14 2491 vmovdqa 16(%r12,%r14,8),%xmm13 2492 vmovdqa 32(%r12,%r14,8),%xmm12 2493 2494 subq $-64,%r13 2495 movl 0(%r15),%eax 2496 leaq (%rsi,%r13,1),%r12 2497 movl 4(%r15),%ebx 2498 cmpq %rdx,%r13 2499 movl 8(%r15),%ecx 2500 cmoveq %rsp,%r12 2501 movl 12(%r15),%edx 2502 movl 16(%r15),%r8d 2503 movl 20(%r15),%r9d 2504 movl 24(%r15),%r10d 2505 movl 28(%r15),%r11d 2506 vmovdqu 0-128(%rdi),%xmm10 2507 jmp .Loop_avx2 2508 .align 16 2509 .Loop_avx2: 2510 vmovdqa K256+512(%rip),%ymm7 2511 vmovdqu -64+0(%rsi,%r13,1),%xmm0 2512 vmovdqu -64+16(%rsi,%r13,1),%xmm1 2513 vmovdqu -64+32(%rsi,%r13,1),%xmm2 2514 vmovdqu -64+48(%rsi,%r13,1),%xmm3 2515 2516 vinserti128 $1,(%r12),%ymm0,%ymm0 2517 vinserti128 $1,16(%r12),%ymm1,%ymm1 2518 vpshufb %ymm7,%ymm0,%ymm0 2519 vinserti128 $1,32(%r12),%ymm2,%ymm2 2520 vpshufb %ymm7,%ymm1,%ymm1 2521 vinserti128 $1,48(%r12),%ymm3,%ymm3 2522 2523 leaq K256(%rip),%rbp 2524 vpshufb %ymm7,%ymm2,%ymm2 2525 leaq -64(%r13),%r13 2526 vpaddd 0(%rbp),%ymm0,%ymm4 2527 vpshufb %ymm7,%ymm3,%ymm3 2528 vpaddd 32(%rbp),%ymm1,%ymm5 2529 vpaddd 64(%rbp),%ymm2,%ymm6 2530 vpaddd 96(%rbp),%ymm3,%ymm7 2531 vmovdqa %ymm4,0(%rsp) 2532 xorl %r14d,%r14d 2533 vmovdqa %ymm5,32(%rsp) 2534 2535 movq 120(%rsp),%rsi 2536 .cfi_def_cfa %rsi,8 2537 leaq -64(%rsp),%rsp 2538 2539 2540 2541 movq %rsi,-8(%rsp) 2542 .cfi_escape 0x0f,0x05,0x77,0x78,0x06,0x23,0x08 2543 movl %ebx,%esi 2544 vmovdqa %ymm6,0(%rsp) 2545 xorl %ecx,%esi 2546 vmovdqa %ymm7,32(%rsp) 2547 movl %r9d,%r12d 2548 subq $-32*4,%rbp 2549 jmp .Lavx2_00_47 2550 2551 .align 16 2552 .Lavx2_00_47: 2553 vmovdqu (%r13),%xmm9 2554 vpinsrq $0,%r13,%xmm15,%xmm15 2555 leaq -64(%rsp),%rsp 2556 .cfi_escape 0x0f,0x05,0x77,0x38,0x06,0x23,0x08 2557 2558 pushq 64-8(%rsp) 2559 .cfi_escape 0x0f,0x05,0x77,0x00,0x06,0x23,0x08 2560 leaq 8(%rsp),%rsp 2561 .cfi_escape 0x0f,0x05,0x77,0x78,0x06,0x23,0x08 2562 vpalignr $4,%ymm0,%ymm1,%ymm4 2563 addl 0+128(%rsp),%r11d 2564 andl %r8d,%r12d 2565 rorxl $25,%r8d,%r13d 2566 vpalignr $4,%ymm2,%ymm3,%ymm7 2567 rorxl $11,%r8d,%r15d 2568 leal (%rax,%r14,1),%eax 2569 leal (%r11,%r12,1),%r11d 2570 vpsrld $7,%ymm4,%ymm6 2571 andnl %r10d,%r8d,%r12d 2572 xorl %r15d,%r13d 2573 rorxl $6,%r8d,%r14d 2574 vpaddd %ymm7,%ymm0,%ymm0 2575 leal (%r11,%r12,1),%r11d 2576 xorl %r14d,%r13d 2577 movl %eax,%r15d 2578 vpsrld $3,%ymm4,%ymm7 2579 rorxl $22,%eax,%r12d 2580 leal (%r11,%r13,1),%r11d 2581 xorl %ebx,%r15d 2582 vpslld $14,%ymm4,%ymm5 2583 rorxl $13,%eax,%r14d 2584 rorxl $2,%eax,%r13d 2585 leal (%rdx,%r11,1),%edx 2586 vpxor %ymm6,%ymm7,%ymm4 2587 andl %r15d,%esi 2588 vpxor %xmm10,%xmm9,%xmm9 2589 vmovdqu 16-128(%rdi),%xmm10 2590 xorl %r12d,%r14d 2591 xorl %ebx,%esi 2592 vpshufd $250,%ymm3,%ymm7 2593 xorl %r13d,%r14d 2594 leal (%r11,%rsi,1),%r11d 2595 movl %r8d,%r12d 2596 vpsrld $11,%ymm6,%ymm6 2597 addl 4+128(%rsp),%r10d 2598 andl %edx,%r12d 2599 rorxl $25,%edx,%r13d 2600 vpxor %ymm5,%ymm4,%ymm4 2601 rorxl $11,%edx,%esi 2602 leal (%r11,%r14,1),%r11d 2603 leal (%r10,%r12,1),%r10d 2604 vpslld $11,%ymm5,%ymm5 2605 andnl %r9d,%edx,%r12d 2606 xorl %esi,%r13d 2607 rorxl $6,%edx,%r14d 2608 vpxor %ymm6,%ymm4,%ymm4 2609 leal (%r10,%r12,1),%r10d 2610 xorl %r14d,%r13d 2611 movl %r11d,%esi 2612 vpsrld $10,%ymm7,%ymm6 2613 rorxl $22,%r11d,%r12d 2614 leal (%r10,%r13,1),%r10d 2615 xorl %eax,%esi 2616 vpxor %ymm5,%ymm4,%ymm4 2617 rorxl $13,%r11d,%r14d 2618 rorxl $2,%r11d,%r13d 2619 leal (%rcx,%r10,1),%ecx 2620 vpsrlq $17,%ymm7,%ymm7 2621 andl %esi,%r15d 2622 vpxor %xmm8,%xmm9,%xmm9 2623 xorl %r12d,%r14d 2624 xorl %eax,%r15d 2625 vpaddd %ymm4,%ymm0,%ymm0 2626 xorl %r13d,%r14d 2627 leal (%r10,%r15,1),%r10d 2628 movl %edx,%r12d 2629 vpxor %ymm7,%ymm6,%ymm6 2630 addl 8+128(%rsp),%r9d 2631 andl %ecx,%r12d 2632 rorxl $25,%ecx,%r13d 2633 vpsrlq $2,%ymm7,%ymm7 2634 rorxl $11,%ecx,%r15d 2635 leal (%r10,%r14,1),%r10d 2636 leal (%r9,%r12,1),%r9d 2637 vpxor %ymm7,%ymm6,%ymm6 2638 andnl %r8d,%ecx,%r12d 2639 xorl %r15d,%r13d 2640 rorxl $6,%ecx,%r14d 2641 vpshufd $132,%ymm6,%ymm6 2642 leal (%r9,%r12,1),%r9d 2643 xorl %r14d,%r13d 2644 movl %r10d,%r15d 2645 vpsrldq $8,%ymm6,%ymm6 2646 rorxl $22,%r10d,%r12d 2647 leal (%r9,%r13,1),%r9d 2648 xorl %r11d,%r15d 2649 vpaddd %ymm6,%ymm0,%ymm0 2650 rorxl $13,%r10d,%r14d 2651 rorxl $2,%r10d,%r13d 2652 leal (%rbx,%r9,1),%ebx 2653 vpshufd $80,%ymm0,%ymm7 2654 andl %r15d,%esi 2655 vaesenc %xmm10,%xmm9,%xmm9 2656 vmovdqu 32-128(%rdi),%xmm10 2657 xorl %r12d,%r14d 2658 xorl %r11d,%esi 2659 vpsrld $10,%ymm7,%ymm6 2660 xorl %r13d,%r14d 2661 leal (%r9,%rsi,1),%r9d 2662 movl %ecx,%r12d 2663 vpsrlq $17,%ymm7,%ymm7 2664 addl 12+128(%rsp),%r8d 2665 andl %ebx,%r12d 2666 rorxl $25,%ebx,%r13d 2667 vpxor %ymm7,%ymm6,%ymm6 2668 rorxl $11,%ebx,%esi 2669 leal (%r9,%r14,1),%r9d 2670 leal (%r8,%r12,1),%r8d 2671 vpsrlq $2,%ymm7,%ymm7 2672 andnl %edx,%ebx,%r12d 2673 xorl %esi,%r13d 2674 rorxl $6,%ebx,%r14d 2675 vpxor %ymm7,%ymm6,%ymm6 2676 leal (%r8,%r12,1),%r8d 2677 xorl %r14d,%r13d 2678 movl %r9d,%esi 2679 vpshufd $232,%ymm6,%ymm6 2680 rorxl $22,%r9d,%r12d 2681 leal (%r8,%r13,1),%r8d 2682 xorl %r10d,%esi 2683 vpslldq $8,%ymm6,%ymm6 2684 rorxl $13,%r9d,%r14d 2685 rorxl $2,%r9d,%r13d 2686 leal (%rax,%r8,1),%eax 2687 vpaddd %ymm6,%ymm0,%ymm0 2688 andl %esi,%r15d 2689 vaesenc %xmm10,%xmm9,%xmm9 2690 vmovdqu 48-128(%rdi),%xmm10 2691 xorl %r12d,%r14d 2692 xorl %r10d,%r15d 2693 vpaddd 0(%rbp),%ymm0,%ymm6 2694 xorl %r13d,%r14d 2695 leal (%r8,%r15,1),%r8d 2696 movl %ebx,%r12d 2697 vmovdqa %ymm6,0(%rsp) 2698 vpalignr $4,%ymm1,%ymm2,%ymm4 2699 addl 32+128(%rsp),%edx 2700 andl %eax,%r12d 2701 rorxl $25,%eax,%r13d 2702 vpalignr $4,%ymm3,%ymm0,%ymm7 2703 rorxl $11,%eax,%r15d 2704 leal (%r8,%r14,1),%r8d 2705 leal (%rdx,%r12,1),%edx 2706 vpsrld $7,%ymm4,%ymm6 2707 andnl %ecx,%eax,%r12d 2708 xorl %r15d,%r13d 2709 rorxl $6,%eax,%r14d 2710 vpaddd %ymm7,%ymm1,%ymm1 2711 leal (%rdx,%r12,1),%edx 2712 xorl %r14d,%r13d 2713 movl %r8d,%r15d 2714 vpsrld $3,%ymm4,%ymm7 2715 rorxl $22,%r8d,%r12d 2716 leal (%rdx,%r13,1),%edx 2717 xorl %r9d,%r15d 2718 vpslld $14,%ymm4,%ymm5 2719 rorxl $13,%r8d,%r14d 2720 rorxl $2,%r8d,%r13d 2721 leal (%r11,%rdx,1),%r11d 2722 vpxor %ymm6,%ymm7,%ymm4 2723 andl %r15d,%esi 2724 vaesenc %xmm10,%xmm9,%xmm9 2725 vmovdqu 64-128(%rdi),%xmm10 2726 xorl %r12d,%r14d 2727 xorl %r9d,%esi 2728 vpshufd $250,%ymm0,%ymm7 2729 xorl %r13d,%r14d 2730 leal (%rdx,%rsi,1),%edx 2731 movl %eax,%r12d 2732 vpsrld $11,%ymm6,%ymm6 2733 addl 36+128(%rsp),%ecx 2734 andl %r11d,%r12d 2735 rorxl $25,%r11d,%r13d 2736 vpxor %ymm5,%ymm4,%ymm4 2737 rorxl $11,%r11d,%esi 2738 leal (%rdx,%r14,1),%edx 2739 leal (%rcx,%r12,1),%ecx 2740 vpslld $11,%ymm5,%ymm5 2741 andnl %ebx,%r11d,%r12d 2742 xorl %esi,%r13d 2743 rorxl $6,%r11d,%r14d 2744 vpxor %ymm6,%ymm4,%ymm4 2745 leal (%rcx,%r12,1),%ecx 2746 xorl %r14d,%r13d 2747 movl %edx,%esi 2748 vpsrld $10,%ymm7,%ymm6 2749 rorxl $22,%edx,%r12d 2750 leal (%rcx,%r13,1),%ecx 2751 xorl %r8d,%esi 2752 vpxor %ymm5,%ymm4,%ymm4 2753 rorxl $13,%edx,%r14d 2754 rorxl $2,%edx,%r13d 2755 leal (%r10,%rcx,1),%r10d 2756 vpsrlq $17,%ymm7,%ymm7 2757 andl %esi,%r15d 2758 vaesenc %xmm10,%xmm9,%xmm9 2759 vmovdqu 80-128(%rdi),%xmm10 2760 xorl %r12d,%r14d 2761 xorl %r8d,%r15d 2762 vpaddd %ymm4,%ymm1,%ymm1 2763 xorl %r13d,%r14d 2764 leal (%rcx,%r15,1),%ecx 2765 movl %r11d,%r12d 2766 vpxor %ymm7,%ymm6,%ymm6 2767 addl 40+128(%rsp),%ebx 2768 andl %r10d,%r12d 2769 rorxl $25,%r10d,%r13d 2770 vpsrlq $2,%ymm7,%ymm7 2771 rorxl $11,%r10d,%r15d 2772 leal (%rcx,%r14,1),%ecx 2773 leal (%rbx,%r12,1),%ebx 2774 vpxor %ymm7,%ymm6,%ymm6 2775 andnl %eax,%r10d,%r12d 2776 xorl %r15d,%r13d 2777 rorxl $6,%r10d,%r14d 2778 vpshufd $132,%ymm6,%ymm6 2779 leal (%rbx,%r12,1),%ebx 2780 xorl %r14d,%r13d 2781 movl %ecx,%r15d 2782 vpsrldq $8,%ymm6,%ymm6 2783 rorxl $22,%ecx,%r12d 2784 leal (%rbx,%r13,1),%ebx 2785 xorl %edx,%r15d 2786 vpaddd %ymm6,%ymm1,%ymm1 2787 rorxl $13,%ecx,%r14d 2788 rorxl $2,%ecx,%r13d 2789 leal (%r9,%rbx,1),%r9d 2790 vpshufd $80,%ymm1,%ymm7 2791 andl %r15d,%esi 2792 vaesenc %xmm10,%xmm9,%xmm9 2793 vmovdqu 96-128(%rdi),%xmm10 2794 xorl %r12d,%r14d 2795 xorl %edx,%esi 2796 vpsrld $10,%ymm7,%ymm6 2797 xorl %r13d,%r14d 2798 leal (%rbx,%rsi,1),%ebx 2799 movl %r10d,%r12d 2800 vpsrlq $17,%ymm7,%ymm7 2801 addl 44+128(%rsp),%eax 2802 andl %r9d,%r12d 2803 rorxl $25,%r9d,%r13d 2804 vpxor %ymm7,%ymm6,%ymm6 2805 rorxl $11,%r9d,%esi 2806 leal (%rbx,%r14,1),%ebx 2807 leal (%rax,%r12,1),%eax 2808 vpsrlq $2,%ymm7,%ymm7 2809 andnl %r11d,%r9d,%r12d 2810 xorl %esi,%r13d 2811 rorxl $6,%r9d,%r14d 2812 vpxor %ymm7,%ymm6,%ymm6 2813 leal (%rax,%r12,1),%eax 2814 xorl %r14d,%r13d 2815 movl %ebx,%esi 2816 vpshufd $232,%ymm6,%ymm6 2817 rorxl $22,%ebx,%r12d 2818 leal (%rax,%r13,1),%eax 2819 xorl %ecx,%esi 2820 vpslldq $8,%ymm6,%ymm6 2821 rorxl $13,%ebx,%r14d 2822 rorxl $2,%ebx,%r13d 2823 leal (%r8,%rax,1),%r8d 2824 vpaddd %ymm6,%ymm1,%ymm1 2825 andl %esi,%r15d 2826 vaesenc %xmm10,%xmm9,%xmm9 2827 vmovdqu 112-128(%rdi),%xmm10 2828 xorl %r12d,%r14d 2829 xorl %ecx,%r15d 2830 vpaddd 32(%rbp),%ymm1,%ymm6 2831 xorl %r13d,%r14d 2832 leal (%rax,%r15,1),%eax 2833 movl %r9d,%r12d 2834 vmovdqa %ymm6,32(%rsp) 2835 leaq -64(%rsp),%rsp 2836 .cfi_escape 0x0f,0x05,0x77,0x38,0x06,0x23,0x08 2837 2838 pushq 64-8(%rsp) 2839 .cfi_escape 0x0f,0x05,0x77,0x00,0x06,0x23,0x08 2840 leaq 8(%rsp),%rsp 2841 .cfi_escape 0x0f,0x05,0x77,0x78,0x06,0x23,0x08 2842 vpalignr $4,%ymm2,%ymm3,%ymm4 2843 addl 0+128(%rsp),%r11d 2844 andl %r8d,%r12d 2845 rorxl $25,%r8d,%r13d 2846 vpalignr $4,%ymm0,%ymm1,%ymm7 2847 rorxl $11,%r8d,%r15d 2848 leal (%rax,%r14,1),%eax 2849 leal (%r11,%r12,1),%r11d 2850 vpsrld $7,%ymm4,%ymm6 2851 andnl %r10d,%r8d,%r12d 2852 xorl %r15d,%r13d 2853 rorxl $6,%r8d,%r14d 2854 vpaddd %ymm7,%ymm2,%ymm2 2855 leal (%r11,%r12,1),%r11d 2856 xorl %r14d,%r13d 2857 movl %eax,%r15d 2858 vpsrld $3,%ymm4,%ymm7 2859 rorxl $22,%eax,%r12d 2860 leal (%r11,%r13,1),%r11d 2861 xorl %ebx,%r15d 2862 vpslld $14,%ymm4,%ymm5 2863 rorxl $13,%eax,%r14d 2864 rorxl $2,%eax,%r13d 2865 leal (%rdx,%r11,1),%edx 2866 vpxor %ymm6,%ymm7,%ymm4 2867 andl %r15d,%esi 2868 vaesenc %xmm10,%xmm9,%xmm9 2869 vmovdqu 128-128(%rdi),%xmm10 2870 xorl %r12d,%r14d 2871 xorl %ebx,%esi 2872 vpshufd $250,%ymm1,%ymm7 2873 xorl %r13d,%r14d 2874 leal (%r11,%rsi,1),%r11d 2875 movl %r8d,%r12d 2876 vpsrld $11,%ymm6,%ymm6 2877 addl 4+128(%rsp),%r10d 2878 andl %edx,%r12d 2879 rorxl $25,%edx,%r13d 2880 vpxor %ymm5,%ymm4,%ymm4 2881 rorxl $11,%edx,%esi 2882 leal (%r11,%r14,1),%r11d 2883 leal (%r10,%r12,1),%r10d 2884 vpslld $11,%ymm5,%ymm5 2885 andnl %r9d,%edx,%r12d 2886 xorl %esi,%r13d 2887 rorxl $6,%edx,%r14d 2888 vpxor %ymm6,%ymm4,%ymm4 2889 leal (%r10,%r12,1),%r10d 2890 xorl %r14d,%r13d 2891 movl %r11d,%esi 2892 vpsrld $10,%ymm7,%ymm6 2893 rorxl $22,%r11d,%r12d 2894 leal (%r10,%r13,1),%r10d 2895 xorl %eax,%esi 2896 vpxor %ymm5,%ymm4,%ymm4 2897 rorxl $13,%r11d,%r14d 2898 rorxl $2,%r11d,%r13d 2899 leal (%rcx,%r10,1),%ecx 2900 vpsrlq $17,%ymm7,%ymm7 2901 andl %esi,%r15d 2902 vaesenc %xmm10,%xmm9,%xmm9 2903 vmovdqu 144-128(%rdi),%xmm10 2904 xorl %r12d,%r14d 2905 xorl %eax,%r15d 2906 vpaddd %ymm4,%ymm2,%ymm2 2907 xorl %r13d,%r14d 2908 leal (%r10,%r15,1),%r10d 2909 movl %edx,%r12d 2910 vpxor %ymm7,%ymm6,%ymm6 2911 addl 8+128(%rsp),%r9d 2912 andl %ecx,%r12d 2913 rorxl $25,%ecx,%r13d 2914 vpsrlq $2,%ymm7,%ymm7 2915 rorxl $11,%ecx,%r15d 2916 leal (%r10,%r14,1),%r10d 2917 leal (%r9,%r12,1),%r9d 2918 vpxor %ymm7,%ymm6,%ymm6 2919 andnl %r8d,%ecx,%r12d 2920 xorl %r15d,%r13d 2921 rorxl $6,%ecx,%r14d 2922 vpshufd $132,%ymm6,%ymm6 2923 leal (%r9,%r12,1),%r9d 2924 xorl %r14d,%r13d 2925 movl %r10d,%r15d 2926 vpsrldq $8,%ymm6,%ymm6 2927 rorxl $22,%r10d,%r12d 2928 leal (%r9,%r13,1),%r9d 2929 xorl %r11d,%r15d 2930 vpaddd %ymm6,%ymm2,%ymm2 2931 rorxl $13,%r10d,%r14d 2932 rorxl $2,%r10d,%r13d 2933 leal (%rbx,%r9,1),%ebx 2934 vpshufd $80,%ymm2,%ymm7 2935 andl %r15d,%esi 2936 vaesenc %xmm10,%xmm9,%xmm9 2937 vmovdqu 160-128(%rdi),%xmm10 2938 xorl %r12d,%r14d 2939 xorl %r11d,%esi 2940 vpsrld $10,%ymm7,%ymm6 2941 xorl %r13d,%r14d 2942 leal (%r9,%rsi,1),%r9d 2943 movl %ecx,%r12d 2944 vpsrlq $17,%ymm7,%ymm7 2945 addl 12+128(%rsp),%r8d 2946 andl %ebx,%r12d 2947 rorxl $25,%ebx,%r13d 2948 vpxor %ymm7,%ymm6,%ymm6 2949 rorxl $11,%ebx,%esi 2950 leal (%r9,%r14,1),%r9d 2951 leal (%r8,%r12,1),%r8d 2952 vpsrlq $2,%ymm7,%ymm7 2953 andnl %edx,%ebx,%r12d 2954 xorl %esi,%r13d 2955 rorxl $6,%ebx,%r14d 2956 vpxor %ymm7,%ymm6,%ymm6 2957 leal (%r8,%r12,1),%r8d 2958 xorl %r14d,%r13d 2959 movl %r9d,%esi 2960 vpshufd $232,%ymm6,%ymm6 2961 rorxl $22,%r9d,%r12d 2962 leal (%r8,%r13,1),%r8d 2963 xorl %r10d,%esi 2964 vpslldq $8,%ymm6,%ymm6 2965 rorxl $13,%r9d,%r14d 2966 rorxl $2,%r9d,%r13d 2967 leal (%rax,%r8,1),%eax 2968 vpaddd %ymm6,%ymm2,%ymm2 2969 andl %esi,%r15d 2970 vaesenclast %xmm10,%xmm9,%xmm11 2971 vaesenc %xmm10,%xmm9,%xmm9 2972 vmovdqu 176-128(%rdi),%xmm10 2973 xorl %r12d,%r14d 2974 xorl %r10d,%r15d 2975 vpaddd 64(%rbp),%ymm2,%ymm6 2976 xorl %r13d,%r14d 2977 leal (%r8,%r15,1),%r8d 2978 movl %ebx,%r12d 2979 vmovdqa %ymm6,0(%rsp) 2980 vpalignr $4,%ymm3,%ymm0,%ymm4 2981 addl 32+128(%rsp),%edx 2982 andl %eax,%r12d 2983 rorxl $25,%eax,%r13d 2984 vpalignr $4,%ymm1,%ymm2,%ymm7 2985 rorxl $11,%eax,%r15d 2986 leal (%r8,%r14,1),%r8d 2987 leal (%rdx,%r12,1),%edx 2988 vpsrld $7,%ymm4,%ymm6 2989 andnl %ecx,%eax,%r12d 2990 xorl %r15d,%r13d 2991 rorxl $6,%eax,%r14d 2992 vpaddd %ymm7,%ymm3,%ymm3 2993 leal (%rdx,%r12,1),%edx 2994 xorl %r14d,%r13d 2995 movl %r8d,%r15d 2996 vpsrld $3,%ymm4,%ymm7 2997 rorxl $22,%r8d,%r12d 2998 leal (%rdx,%r13,1),%edx 2999 xorl %r9d,%r15d 3000 vpslld $14,%ymm4,%ymm5 3001 rorxl $13,%r8d,%r14d 3002 rorxl $2,%r8d,%r13d 3003 leal (%r11,%rdx,1),%r11d 3004 vpxor %ymm6,%ymm7,%ymm4 3005 andl %r15d,%esi 3006 vpand %xmm12,%xmm11,%xmm8 3007 vaesenc %xmm10,%xmm9,%xmm9 3008 vmovdqu 192-128(%rdi),%xmm10 3009 xorl %r12d,%r14d 3010 xorl %r9d,%esi 3011 vpshufd $250,%ymm2,%ymm7 3012 xorl %r13d,%r14d 3013 leal (%rdx,%rsi,1),%edx 3014 movl %eax,%r12d 3015 vpsrld $11,%ymm6,%ymm6 3016 addl 36+128(%rsp),%ecx 3017 andl %r11d,%r12d 3018 rorxl $25,%r11d,%r13d 3019 vpxor %ymm5,%ymm4,%ymm4 3020 rorxl $11,%r11d,%esi 3021 leal (%rdx,%r14,1),%edx 3022 leal (%rcx,%r12,1),%ecx 3023 vpslld $11,%ymm5,%ymm5 3024 andnl %ebx,%r11d,%r12d 3025 xorl %esi,%r13d 3026 rorxl $6,%r11d,%r14d 3027 vpxor %ymm6,%ymm4,%ymm4 3028 leal (%rcx,%r12,1),%ecx 3029 xorl %r14d,%r13d 3030 movl %edx,%esi 3031 vpsrld $10,%ymm7,%ymm6 3032 rorxl $22,%edx,%r12d 3033 leal (%rcx,%r13,1),%ecx 3034 xorl %r8d,%esi 3035 vpxor %ymm5,%ymm4,%ymm4 3036 rorxl $13,%edx,%r14d 3037 rorxl $2,%edx,%r13d 3038 leal (%r10,%rcx,1),%r10d 3039 vpsrlq $17,%ymm7,%ymm7 3040 andl %esi,%r15d 3041 vaesenclast %xmm10,%xmm9,%xmm11 3042 vaesenc %xmm10,%xmm9,%xmm9 3043 vmovdqu 208-128(%rdi),%xmm10 3044 xorl %r12d,%r14d 3045 xorl %r8d,%r15d 3046 vpaddd %ymm4,%ymm3,%ymm3 3047 xorl %r13d,%r14d 3048 leal (%rcx,%r15,1),%ecx 3049 movl %r11d,%r12d 3050 vpxor %ymm7,%ymm6,%ymm6 3051 addl 40+128(%rsp),%ebx 3052 andl %r10d,%r12d 3053 rorxl $25,%r10d,%r13d 3054 vpsrlq $2,%ymm7,%ymm7 3055 rorxl $11,%r10d,%r15d 3056 leal (%rcx,%r14,1),%ecx 3057 leal (%rbx,%r12,1),%ebx 3058 vpxor %ymm7,%ymm6,%ymm6 3059 andnl %eax,%r10d,%r12d 3060 xorl %r15d,%r13d 3061 rorxl $6,%r10d,%r14d 3062 vpshufd $132,%ymm6,%ymm6 3063 leal (%rbx,%r12,1),%ebx 3064 xorl %r14d,%r13d 3065 movl %ecx,%r15d 3066 vpsrldq $8,%ymm6,%ymm6 3067 rorxl $22,%ecx,%r12d 3068 leal (%rbx,%r13,1),%ebx 3069 xorl %edx,%r15d 3070 vpaddd %ymm6,%ymm3,%ymm3 3071 rorxl $13,%ecx,%r14d 3072 rorxl $2,%ecx,%r13d 3073 leal (%r9,%rbx,1),%r9d 3074 vpshufd $80,%ymm3,%ymm7 3075 andl %r15d,%esi 3076 vpand %xmm13,%xmm11,%xmm11 3077 vaesenc %xmm10,%xmm9,%xmm9 3078 vmovdqu 224-128(%rdi),%xmm10 3079 xorl %r12d,%r14d 3080 xorl %edx,%esi 3081 vpsrld $10,%ymm7,%ymm6 3082 xorl %r13d,%r14d 3083 leal (%rbx,%rsi,1),%ebx 3084 movl %r10d,%r12d 3085 vpsrlq $17,%ymm7,%ymm7 3086 addl 44+128(%rsp),%eax 3087 andl %r9d,%r12d 3088 rorxl $25,%r9d,%r13d 3089 vpxor %ymm7,%ymm6,%ymm6 3090 rorxl $11,%r9d,%esi 3091 leal (%rbx,%r14,1),%ebx 3092 leal (%rax,%r12,1),%eax 3093 vpsrlq $2,%ymm7,%ymm7 3094 andnl %r11d,%r9d,%r12d 3095 xorl %esi,%r13d 3096 rorxl $6,%r9d,%r14d 3097 vpxor %ymm7,%ymm6,%ymm6 3098 leal (%rax,%r12,1),%eax 3099 xorl %r14d,%r13d 3100 movl %ebx,%esi 3101 vpshufd $232,%ymm6,%ymm6 3102 rorxl $22,%ebx,%r12d 3103 leal (%rax,%r13,1),%eax 3104 xorl %ecx,%esi 3105 vpslldq $8,%ymm6,%ymm6 3106 rorxl $13,%ebx,%r14d 3107 rorxl $2,%ebx,%r13d 3108 leal (%r8,%rax,1),%r8d 3109 vpaddd %ymm6,%ymm3,%ymm3 3110 andl %esi,%r15d 3111 vpor %xmm11,%xmm8,%xmm8 3112 vaesenclast %xmm10,%xmm9,%xmm11 3113 vmovdqu 0-128(%rdi),%xmm10 3114 xorl %r12d,%r14d 3115 xorl %ecx,%r15d 3116 vpaddd 96(%rbp),%ymm3,%ymm6 3117 xorl %r13d,%r14d 3118 leal (%rax,%r15,1),%eax 3119 movl %r9d,%r12d 3120 vmovdqa %ymm6,32(%rsp) 3121 vmovq %xmm15,%r13 3122 vpextrq $1,%xmm15,%r15 3123 vpand %xmm14,%xmm11,%xmm11 3124 vpor %xmm11,%xmm8,%xmm8 3125 vmovdqu %xmm8,(%r15,%r13,1) 3126 leaq 16(%r13),%r13 3127 leaq 128(%rbp),%rbp 3128 cmpb $0,3(%rbp) 3129 jne .Lavx2_00_47 3130 vmovdqu (%r13),%xmm9 3131 vpinsrq $0,%r13,%xmm15,%xmm15 3132 addl 0+64(%rsp),%r11d 3133 andl %r8d,%r12d 3134 rorxl $25,%r8d,%r13d 3135 rorxl $11,%r8d,%r15d 3136 leal (%rax,%r14,1),%eax 3137 leal (%r11,%r12,1),%r11d 3138 andnl %r10d,%r8d,%r12d 3139 xorl %r15d,%r13d 3140 rorxl $6,%r8d,%r14d 3141 leal (%r11,%r12,1),%r11d 3142 xorl %r14d,%r13d 3143 movl %eax,%r15d 3144 rorxl $22,%eax,%r12d 3145 leal (%r11,%r13,1),%r11d 3146 xorl %ebx,%r15d 3147 rorxl $13,%eax,%r14d 3148 rorxl $2,%eax,%r13d 3149 leal (%rdx,%r11,1),%edx 3150 andl %r15d,%esi 3151 vpxor %xmm10,%xmm9,%xmm9 3152 vmovdqu 16-128(%rdi),%xmm10 3153 xorl %r12d,%r14d 3154 xorl %ebx,%esi 3155 xorl %r13d,%r14d 3156 leal (%r11,%rsi,1),%r11d 3157 movl %r8d,%r12d 3158 addl 4+64(%rsp),%r10d 3159 andl %edx,%r12d 3160 rorxl $25,%edx,%r13d 3161 rorxl $11,%edx,%esi 3162 leal (%r11,%r14,1),%r11d 3163 leal (%r10,%r12,1),%r10d 3164 andnl %r9d,%edx,%r12d 3165 xorl %esi,%r13d 3166 rorxl $6,%edx,%r14d 3167 leal (%r10,%r12,1),%r10d 3168 xorl %r14d,%r13d 3169 movl %r11d,%esi 3170 rorxl $22,%r11d,%r12d 3171 leal (%r10,%r13,1),%r10d 3172 xorl %eax,%esi 3173 rorxl $13,%r11d,%r14d 3174 rorxl $2,%r11d,%r13d 3175 leal (%rcx,%r10,1),%ecx 3176 andl %esi,%r15d 3177 vpxor %xmm8,%xmm9,%xmm9 3178 xorl %r12d,%r14d 3179 xorl %eax,%r15d 3180 xorl %r13d,%r14d 3181 leal (%r10,%r15,1),%r10d 3182 movl %edx,%r12d 3183 addl 8+64(%rsp),%r9d 3184 andl %ecx,%r12d 3185 rorxl $25,%ecx,%r13d 3186 rorxl $11,%ecx,%r15d 3187 leal (%r10,%r14,1),%r10d 3188 leal (%r9,%r12,1),%r9d 3189 andnl %r8d,%ecx,%r12d 3190 xorl %r15d,%r13d 3191 rorxl $6,%ecx,%r14d 3192 leal (%r9,%r12,1),%r9d 3193 xorl %r14d,%r13d 3194 movl %r10d,%r15d 3195 rorxl $22,%r10d,%r12d 3196 leal (%r9,%r13,1),%r9d 3197 xorl %r11d,%r15d 3198 rorxl $13,%r10d,%r14d 3199 rorxl $2,%r10d,%r13d 3200 leal (%rbx,%r9,1),%ebx 3201 andl %r15d,%esi 3202 vaesenc %xmm10,%xmm9,%xmm9 3203 vmovdqu 32-128(%rdi),%xmm10 3204 xorl %r12d,%r14d 3205 xorl %r11d,%esi 3206 xorl %r13d,%r14d 3207 leal (%r9,%rsi,1),%r9d 3208 movl %ecx,%r12d 3209 addl 12+64(%rsp),%r8d 3210 andl %ebx,%r12d 3211 rorxl $25,%ebx,%r13d 3212 rorxl $11,%ebx,%esi 3213 leal (%r9,%r14,1),%r9d 3214 leal (%r8,%r12,1),%r8d 3215 andnl %edx,%ebx,%r12d 3216 xorl %esi,%r13d 3217 rorxl $6,%ebx,%r14d 3218 leal (%r8,%r12,1),%r8d 3219 xorl %r14d,%r13d 3220 movl %r9d,%esi 3221 rorxl $22,%r9d,%r12d 3222 leal (%r8,%r13,1),%r8d 3223 xorl %r10d,%esi 3224 rorxl $13,%r9d,%r14d 3225 rorxl $2,%r9d,%r13d 3226 leal (%rax,%r8,1),%eax 3227 andl %esi,%r15d 3228 vaesenc %xmm10,%xmm9,%xmm9 3229 vmovdqu 48-128(%rdi),%xmm10 3230 xorl %r12d,%r14d 3231 xorl %r10d,%r15d 3232 xorl %r13d,%r14d 3233 leal (%r8,%r15,1),%r8d 3234 movl %ebx,%r12d 3235 addl 32+64(%rsp),%edx 3236 andl %eax,%r12d 3237 rorxl $25,%eax,%r13d 3238 rorxl $11,%eax,%r15d 3239 leal (%r8,%r14,1),%r8d 3240 leal (%rdx,%r12,1),%edx 3241 andnl %ecx,%eax,%r12d 3242 xorl %r15d,%r13d 3243 rorxl $6,%eax,%r14d 3244 leal (%rdx,%r12,1),%edx 3245 xorl %r14d,%r13d 3246 movl %r8d,%r15d 3247 rorxl $22,%r8d,%r12d 3248 leal (%rdx,%r13,1),%edx 3249 xorl %r9d,%r15d 3250 rorxl $13,%r8d,%r14d 3251 rorxl $2,%r8d,%r13d 3252 leal (%r11,%rdx,1),%r11d 3253 andl %r15d,%esi 3254 vaesenc %xmm10,%xmm9,%xmm9 3255 vmovdqu 64-128(%rdi),%xmm10 3256 xorl %r12d,%r14d 3257 xorl %r9d,%esi 3258 xorl %r13d,%r14d 3259 leal (%rdx,%rsi,1),%edx 3260 movl %eax,%r12d 3261 addl 36+64(%rsp),%ecx 3262 andl %r11d,%r12d 3263 rorxl $25,%r11d,%r13d 3264 rorxl $11,%r11d,%esi 3265 leal (%rdx,%r14,1),%edx 3266 leal (%rcx,%r12,1),%ecx 3267 andnl %ebx,%r11d,%r12d 3268 xorl %esi,%r13d 3269 rorxl $6,%r11d,%r14d 3270 leal (%rcx,%r12,1),%ecx 3271 xorl %r14d,%r13d 3272 movl %edx,%esi 3273 rorxl $22,%edx,%r12d 3274 leal (%rcx,%r13,1),%ecx 3275 xorl %r8d,%esi 3276 rorxl $13,%edx,%r14d 3277 rorxl $2,%edx,%r13d 3278 leal (%r10,%rcx,1),%r10d 3279 andl %esi,%r15d 3280 vaesenc %xmm10,%xmm9,%xmm9 3281 vmovdqu 80-128(%rdi),%xmm10 3282 xorl %r12d,%r14d 3283 xorl %r8d,%r15d 3284 xorl %r13d,%r14d 3285 leal (%rcx,%r15,1),%ecx 3286 movl %r11d,%r12d 3287 addl 40+64(%rsp),%ebx 3288 andl %r10d,%r12d 3289 rorxl $25,%r10d,%r13d 3290 rorxl $11,%r10d,%r15d 3291 leal (%rcx,%r14,1),%ecx 3292 leal (%rbx,%r12,1),%ebx 3293 andnl %eax,%r10d,%r12d 3294 xorl %r15d,%r13d 3295 rorxl $6,%r10d,%r14d 3296 leal (%rbx,%r12,1),%ebx 3297 xorl %r14d,%r13d 3298 movl %ecx,%r15d 3299 rorxl $22,%ecx,%r12d 3300 leal (%rbx,%r13,1),%ebx 3301 xorl %edx,%r15d 3302 rorxl $13,%ecx,%r14d 3303 rorxl $2,%ecx,%r13d 3304 leal (%r9,%rbx,1),%r9d 3305 andl %r15d,%esi 3306 vaesenc %xmm10,%xmm9,%xmm9 3307 vmovdqu 96-128(%rdi),%xmm10 3308 xorl %r12d,%r14d 3309 xorl %edx,%esi 3310 xorl %r13d,%r14d 3311 leal (%rbx,%rsi,1),%ebx 3312 movl %r10d,%r12d 3313 addl 44+64(%rsp),%eax 3314 andl %r9d,%r12d 3315 rorxl $25,%r9d,%r13d 3316 rorxl $11,%r9d,%esi 3317 leal (%rbx,%r14,1),%ebx 3318 leal (%rax,%r12,1),%eax 3319 andnl %r11d,%r9d,%r12d 3320 xorl %esi,%r13d 3321 rorxl $6,%r9d,%r14d 3322 leal (%rax,%r12,1),%eax 3323 xorl %r14d,%r13d 3324 movl %ebx,%esi 3325 rorxl $22,%ebx,%r12d 3326 leal (%rax,%r13,1),%eax 3327 xorl %ecx,%esi 3328 rorxl $13,%ebx,%r14d 3329 rorxl $2,%ebx,%r13d 3330 leal (%r8,%rax,1),%r8d 3331 andl %esi,%r15d 3332 vaesenc %xmm10,%xmm9,%xmm9 3333 vmovdqu 112-128(%rdi),%xmm10 3334 xorl %r12d,%r14d 3335 xorl %ecx,%r15d 3336 xorl %r13d,%r14d 3337 leal (%rax,%r15,1),%eax 3338 movl %r9d,%r12d 3339 addl 0(%rsp),%r11d 3340 andl %r8d,%r12d 3341 rorxl $25,%r8d,%r13d 3342 rorxl $11,%r8d,%r15d 3343 leal (%rax,%r14,1),%eax 3344 leal (%r11,%r12,1),%r11d 3345 andnl %r10d,%r8d,%r12d 3346 xorl %r15d,%r13d 3347 rorxl $6,%r8d,%r14d 3348 leal (%r11,%r12,1),%r11d 3349 xorl %r14d,%r13d 3350 movl %eax,%r15d 3351 rorxl $22,%eax,%r12d 3352 leal (%r11,%r13,1),%r11d 3353 xorl %ebx,%r15d 3354 rorxl $13,%eax,%r14d 3355 rorxl $2,%eax,%r13d 3356 leal (%rdx,%r11,1),%edx 3357 andl %r15d,%esi 3358 vaesenc %xmm10,%xmm9,%xmm9 3359 vmovdqu 128-128(%rdi),%xmm10 3360 xorl %r12d,%r14d 3361 xorl %ebx,%esi 3362 xorl %r13d,%r14d 3363 leal (%r11,%rsi,1),%r11d 3364 movl %r8d,%r12d 3365 addl 4(%rsp),%r10d 3366 andl %edx,%r12d 3367 rorxl $25,%edx,%r13d 3368 rorxl $11,%edx,%esi 3369 leal (%r11,%r14,1),%r11d 3370 leal (%r10,%r12,1),%r10d 3371 andnl %r9d,%edx,%r12d 3372 xorl %esi,%r13d 3373 rorxl $6,%edx,%r14d 3374 leal (%r10,%r12,1),%r10d 3375 xorl %r14d,%r13d 3376 movl %r11d,%esi 3377 rorxl $22,%r11d,%r12d 3378 leal (%r10,%r13,1),%r10d 3379 xorl %eax,%esi 3380 rorxl $13,%r11d,%r14d 3381 rorxl $2,%r11d,%r13d 3382 leal (%rcx,%r10,1),%ecx 3383 andl %esi,%r15d 3384 vaesenc %xmm10,%xmm9,%xmm9 3385 vmovdqu 144-128(%rdi),%xmm10 3386 xorl %r12d,%r14d 3387 xorl %eax,%r15d 3388 xorl %r13d,%r14d 3389 leal (%r10,%r15,1),%r10d 3390 movl %edx,%r12d 3391 addl 8(%rsp),%r9d 3392 andl %ecx,%r12d 3393 rorxl $25,%ecx,%r13d 3394 rorxl $11,%ecx,%r15d 3395 leal (%r10,%r14,1),%r10d 3396 leal (%r9,%r12,1),%r9d 3397 andnl %r8d,%ecx,%r12d 3398 xorl %r15d,%r13d 3399 rorxl $6,%ecx,%r14d 3400 leal (%r9,%r12,1),%r9d 3401 xorl %r14d,%r13d 3402 movl %r10d,%r15d 3403 rorxl $22,%r10d,%r12d 3404 leal (%r9,%r13,1),%r9d 3405 xorl %r11d,%r15d 3406 rorxl $13,%r10d,%r14d 3407 rorxl $2,%r10d,%r13d 3408 leal (%rbx,%r9,1),%ebx 3409 andl %r15d,%esi 3410 vaesenc %xmm10,%xmm9,%xmm9 3411 vmovdqu 160-128(%rdi),%xmm10 3412 xorl %r12d,%r14d 3413 xorl %r11d,%esi 3414 xorl %r13d,%r14d 3415 leal (%r9,%rsi,1),%r9d 3416 movl %ecx,%r12d 3417 addl 12(%rsp),%r8d 3418 andl %ebx,%r12d 3419 rorxl $25,%ebx,%r13d 3420 rorxl $11,%ebx,%esi 3421 leal (%r9,%r14,1),%r9d 3422 leal (%r8,%r12,1),%r8d 3423 andnl %edx,%ebx,%r12d 3424 xorl %esi,%r13d 3425 rorxl $6,%ebx,%r14d 3426 leal (%r8,%r12,1),%r8d 3427 xorl %r14d,%r13d 3428 movl %r9d,%esi 3429 rorxl $22,%r9d,%r12d 3430 leal (%r8,%r13,1),%r8d 3431 xorl %r10d,%esi 3432 rorxl $13,%r9d,%r14d 3433 rorxl $2,%r9d,%r13d 3434 leal (%rax,%r8,1),%eax 3435 andl %esi,%r15d 3436 vaesenclast %xmm10,%xmm9,%xmm11 3437 vaesenc %xmm10,%xmm9,%xmm9 3438 vmovdqu 176-128(%rdi),%xmm10 3439 xorl %r12d,%r14d 3440 xorl %r10d,%r15d 3441 xorl %r13d,%r14d 3442 leal (%r8,%r15,1),%r8d 3443 movl %ebx,%r12d 3444 addl 32(%rsp),%edx 3445 andl %eax,%r12d 3446 rorxl $25,%eax,%r13d 3447 rorxl $11,%eax,%r15d 3448 leal (%r8,%r14,1),%r8d 3449 leal (%rdx,%r12,1),%edx 3450 andnl %ecx,%eax,%r12d 3451 xorl %r15d,%r13d 3452 rorxl $6,%eax,%r14d 3453 leal (%rdx,%r12,1),%edx 3454 xorl %r14d,%r13d 3455 movl %r8d,%r15d 3456 rorxl $22,%r8d,%r12d 3457 leal (%rdx,%r13,1),%edx 3458 xorl %r9d,%r15d 3459 rorxl $13,%r8d,%r14d 3460 rorxl $2,%r8d,%r13d 3461 leal (%r11,%rdx,1),%r11d 3462 andl %r15d,%esi 3463 vpand %xmm12,%xmm11,%xmm8 3464 vaesenc %xmm10,%xmm9,%xmm9 3465 vmovdqu 192-128(%rdi),%xmm10 3466 xorl %r12d,%r14d 3467 xorl %r9d,%esi 3468 xorl %r13d,%r14d 3469 leal (%rdx,%rsi,1),%edx 3470 movl %eax,%r12d 3471 addl 36(%rsp),%ecx 3472 andl %r11d,%r12d 3473 rorxl $25,%r11d,%r13d 3474 rorxl $11,%r11d,%esi 3475 leal (%rdx,%r14,1),%edx 3476 leal (%rcx,%r12,1),%ecx 3477 andnl %ebx,%r11d,%r12d 3478 xorl %esi,%r13d 3479 rorxl $6,%r11d,%r14d 3480 leal (%rcx,%r12,1),%ecx 3481 xorl %r14d,%r13d 3482 movl %edx,%esi 3483 rorxl $22,%edx,%r12d 3484 leal (%rcx,%r13,1),%ecx 3485 xorl %r8d,%esi 3486 rorxl $13,%edx,%r14d 3487 rorxl $2,%edx,%r13d 3488 leal (%r10,%rcx,1),%r10d 3489 andl %esi,%r15d 3490 vaesenclast %xmm10,%xmm9,%xmm11 3491 vaesenc %xmm10,%xmm9,%xmm9 3492 vmovdqu 208-128(%rdi),%xmm10 3493 xorl %r12d,%r14d 3494 xorl %r8d,%r15d 3495 xorl %r13d,%r14d 3496 leal (%rcx,%r15,1),%ecx 3497 movl %r11d,%r12d 3498 addl 40(%rsp),%ebx 3499 andl %r10d,%r12d 3500 rorxl $25,%r10d,%r13d 3501 rorxl $11,%r10d,%r15d 3502 leal (%rcx,%r14,1),%ecx 3503 leal (%rbx,%r12,1),%ebx 3504 andnl %eax,%r10d,%r12d 3505 xorl %r15d,%r13d 3506 rorxl $6,%r10d,%r14d 3507 leal (%rbx,%r12,1),%ebx 3508 xorl %r14d,%r13d 3509 movl %ecx,%r15d 3510 rorxl $22,%ecx,%r12d 3511 leal (%rbx,%r13,1),%ebx 3512 xorl %edx,%r15d 3513 rorxl $13,%ecx,%r14d 3514 rorxl $2,%ecx,%r13d 3515 leal (%r9,%rbx,1),%r9d 3516 andl %r15d,%esi 3517 vpand %xmm13,%xmm11,%xmm11 3518 vaesenc %xmm10,%xmm9,%xmm9 3519 vmovdqu 224-128(%rdi),%xmm10 3520 xorl %r12d,%r14d 3521 xorl %edx,%esi 3522 xorl %r13d,%r14d 3523 leal (%rbx,%rsi,1),%ebx 3524 movl %r10d,%r12d 3525 addl 44(%rsp),%eax 3526 andl %r9d,%r12d 3527 rorxl $25,%r9d,%r13d 3528 rorxl $11,%r9d,%esi 3529 leal (%rbx,%r14,1),%ebx 3530 leal (%rax,%r12,1),%eax 3531 andnl %r11d,%r9d,%r12d 3532 xorl %esi,%r13d 3533 rorxl $6,%r9d,%r14d 3534 leal (%rax,%r12,1),%eax 3535 xorl %r14d,%r13d 3536 movl %ebx,%esi 3537 rorxl $22,%ebx,%r12d 3538 leal (%rax,%r13,1),%eax 3539 xorl %ecx,%esi 3540 rorxl $13,%ebx,%r14d 3541 rorxl $2,%ebx,%r13d 3542 leal (%r8,%rax,1),%r8d 3543 andl %esi,%r15d 3544 vpor %xmm11,%xmm8,%xmm8 3545 vaesenclast %xmm10,%xmm9,%xmm11 3546 vmovdqu 0-128(%rdi),%xmm10 3547 xorl %r12d,%r14d 3548 xorl %ecx,%r15d 3549 xorl %r13d,%r14d 3550 leal (%rax,%r15,1),%eax 3551 movl %r9d,%r12d 3552 vpextrq $1,%xmm15,%r12 3553 vmovq %xmm15,%r13 3554 movq 552(%rsp),%r15 3555 addl %r14d,%eax 3556 leaq 448(%rsp),%rbp 3557 3558 vpand %xmm14,%xmm11,%xmm11 3559 vpor %xmm11,%xmm8,%xmm8 3560 vmovdqu %xmm8,(%r12,%r13,1) 3561 leaq 16(%r13),%r13 3562 3563 addl 0(%r15),%eax 3564 addl 4(%r15),%ebx 3565 addl 8(%r15),%ecx 3566 addl 12(%r15),%edx 3567 addl 16(%r15),%r8d 3568 addl 20(%r15),%r9d 3569 addl 24(%r15),%r10d 3570 addl 28(%r15),%r11d 3571 3572 movl %eax,0(%r15) 3573 movl %ebx,4(%r15) 3574 movl %ecx,8(%r15) 3575 movl %edx,12(%r15) 3576 movl %r8d,16(%r15) 3577 movl %r9d,20(%r15) 3578 movl %r10d,24(%r15) 3579 movl %r11d,28(%r15) 3580 3581 cmpq 80(%rbp),%r13 3582 je .Ldone_avx2 3583 3584 xorl %r14d,%r14d 3585 movl %ebx,%esi 3586 movl %r9d,%r12d 3587 xorl %ecx,%esi 3588 jmp .Lower_avx2 3589 .align 16 3590 .Lower_avx2: 3591 vmovdqu (%r13),%xmm9 3592 vpinsrq $0,%r13,%xmm15,%xmm15 3593 addl 0+16(%rbp),%r11d 3594 andl %r8d,%r12d 3595 rorxl $25,%r8d,%r13d 3596 rorxl $11,%r8d,%r15d 3597 leal (%rax,%r14,1),%eax 3598 leal (%r11,%r12,1),%r11d 3599 andnl %r10d,%r8d,%r12d 3600 xorl %r15d,%r13d 3601 rorxl $6,%r8d,%r14d 3602 leal (%r11,%r12,1),%r11d 3603 xorl %r14d,%r13d 3604 movl %eax,%r15d 3605 rorxl $22,%eax,%r12d 3606 leal (%r11,%r13,1),%r11d 3607 xorl %ebx,%r15d 3608 rorxl $13,%eax,%r14d 3609 rorxl $2,%eax,%r13d 3610 leal (%rdx,%r11,1),%edx 3611 andl %r15d,%esi 3612 vpxor %xmm10,%xmm9,%xmm9 3613 vmovdqu 16-128(%rdi),%xmm10 3614 xorl %r12d,%r14d 3615 xorl %ebx,%esi 3616 xorl %r13d,%r14d 3617 leal (%r11,%rsi,1),%r11d 3618 movl %r8d,%r12d 3619 addl 4+16(%rbp),%r10d 3620 andl %edx,%r12d 3621 rorxl $25,%edx,%r13d 3622 rorxl $11,%edx,%esi 3623 leal (%r11,%r14,1),%r11d 3624 leal (%r10,%r12,1),%r10d 3625 andnl %r9d,%edx,%r12d 3626 xorl %esi,%r13d 3627 rorxl $6,%edx,%r14d 3628 leal (%r10,%r12,1),%r10d 3629 xorl %r14d,%r13d 3630 movl %r11d,%esi 3631 rorxl $22,%r11d,%r12d 3632 leal (%r10,%r13,1),%r10d 3633 xorl %eax,%esi 3634 rorxl $13,%r11d,%r14d 3635 rorxl $2,%r11d,%r13d 3636 leal (%rcx,%r10,1),%ecx 3637 andl %esi,%r15d 3638 vpxor %xmm8,%xmm9,%xmm9 3639 xorl %r12d,%r14d 3640 xorl %eax,%r15d 3641 xorl %r13d,%r14d 3642 leal (%r10,%r15,1),%r10d 3643 movl %edx,%r12d 3644 addl 8+16(%rbp),%r9d 3645 andl %ecx,%r12d 3646 rorxl $25,%ecx,%r13d 3647 rorxl $11,%ecx,%r15d 3648 leal (%r10,%r14,1),%r10d 3649 leal (%r9,%r12,1),%r9d 3650 andnl %r8d,%ecx,%r12d 3651 xorl %r15d,%r13d 3652 rorxl $6,%ecx,%r14d 3653 leal (%r9,%r12,1),%r9d 3654 xorl %r14d,%r13d 3655 movl %r10d,%r15d 3656 rorxl $22,%r10d,%r12d 3657 leal (%r9,%r13,1),%r9d 3658 xorl %r11d,%r15d 3659 rorxl $13,%r10d,%r14d 3660 rorxl $2,%r10d,%r13d 3661 leal (%rbx,%r9,1),%ebx 3662 andl %r15d,%esi 3663 vaesenc %xmm10,%xmm9,%xmm9 3664 vmovdqu 32-128(%rdi),%xmm10 3665 xorl %r12d,%r14d 3666 xorl %r11d,%esi 3667 xorl %r13d,%r14d 3668 leal (%r9,%rsi,1),%r9d 3669 movl %ecx,%r12d 3670 addl 12+16(%rbp),%r8d 3671 andl %ebx,%r12d 3672 rorxl $25,%ebx,%r13d 3673 rorxl $11,%ebx,%esi 3674 leal (%r9,%r14,1),%r9d 3675 leal (%r8,%r12,1),%r8d 3676 andnl %edx,%ebx,%r12d 3677 xorl %esi,%r13d 3678 rorxl $6,%ebx,%r14d 3679 leal (%r8,%r12,1),%r8d 3680 xorl %r14d,%r13d 3681 movl %r9d,%esi 3682 rorxl $22,%r9d,%r12d 3683 leal (%r8,%r13,1),%r8d 3684 xorl %r10d,%esi 3685 rorxl $13,%r9d,%r14d 3686 rorxl $2,%r9d,%r13d 3687 leal (%rax,%r8,1),%eax 3688 andl %esi,%r15d 3689 vaesenc %xmm10,%xmm9,%xmm9 3690 vmovdqu 48-128(%rdi),%xmm10 3691 xorl %r12d,%r14d 3692 xorl %r10d,%r15d 3693 xorl %r13d,%r14d 3694 leal (%r8,%r15,1),%r8d 3695 movl %ebx,%r12d 3696 addl 32+16(%rbp),%edx 3697 andl %eax,%r12d 3698 rorxl $25,%eax,%r13d 3699 rorxl $11,%eax,%r15d 3700 leal (%r8,%r14,1),%r8d 3701 leal (%rdx,%r12,1),%edx 3702 andnl %ecx,%eax,%r12d 3703 xorl %r15d,%r13d 3704 rorxl $6,%eax,%r14d 3705 leal (%rdx,%r12,1),%edx 3706 xorl %r14d,%r13d 3707 movl %r8d,%r15d 3708 rorxl $22,%r8d,%r12d 3709 leal (%rdx,%r13,1),%edx 3710 xorl %r9d,%r15d 3711 rorxl $13,%r8d,%r14d 3712 rorxl $2,%r8d,%r13d 3713 leal (%r11,%rdx,1),%r11d 3714 andl %r15d,%esi 3715 vaesenc %xmm10,%xmm9,%xmm9 3716 vmovdqu 64-128(%rdi),%xmm10 3717 xorl %r12d,%r14d 3718 xorl %r9d,%esi 3719 xorl %r13d,%r14d 3720 leal (%rdx,%rsi,1),%edx 3721 movl %eax,%r12d 3722 addl 36+16(%rbp),%ecx 3723 andl %r11d,%r12d 3724 rorxl $25,%r11d,%r13d 3725 rorxl $11,%r11d,%esi 3726 leal (%rdx,%r14,1),%edx 3727 leal (%rcx,%r12,1),%ecx 3728 andnl %ebx,%r11d,%r12d 3729 xorl %esi,%r13d 3730 rorxl $6,%r11d,%r14d 3731 leal (%rcx,%r12,1),%ecx 3732 xorl %r14d,%r13d 3733 movl %edx,%esi 3734 rorxl $22,%edx,%r12d 3735 leal (%rcx,%r13,1),%ecx 3736 xorl %r8d,%esi 3737 rorxl $13,%edx,%r14d 3738 rorxl $2,%edx,%r13d 3739 leal (%r10,%rcx,1),%r10d 3740 andl %esi,%r15d 3741 vaesenc %xmm10,%xmm9,%xmm9 3742 vmovdqu 80-128(%rdi),%xmm10 3743 xorl %r12d,%r14d 3744 xorl %r8d,%r15d 3745 xorl %r13d,%r14d 3746 leal (%rcx,%r15,1),%ecx 3747 movl %r11d,%r12d 3748 addl 40+16(%rbp),%ebx 3749 andl %r10d,%r12d 3750 rorxl $25,%r10d,%r13d 3751 rorxl $11,%r10d,%r15d 3752 leal (%rcx,%r14,1),%ecx 3753 leal (%rbx,%r12,1),%ebx 3754 andnl %eax,%r10d,%r12d 3755 xorl %r15d,%r13d 3756 rorxl $6,%r10d,%r14d 3757 leal (%rbx,%r12,1),%ebx 3758 xorl %r14d,%r13d 3759 movl %ecx,%r15d 3760 rorxl $22,%ecx,%r12d 3761 leal (%rbx,%r13,1),%ebx 3762 xorl %edx,%r15d 3763 rorxl $13,%ecx,%r14d 3764 rorxl $2,%ecx,%r13d 3765 leal (%r9,%rbx,1),%r9d 3766 andl %r15d,%esi 3767 vaesenc %xmm10,%xmm9,%xmm9 3768 vmovdqu 96-128(%rdi),%xmm10 3769 xorl %r12d,%r14d 3770 xorl %edx,%esi 3771 xorl %r13d,%r14d 3772 leal (%rbx,%rsi,1),%ebx 3773 movl %r10d,%r12d 3774 addl 44+16(%rbp),%eax 3775 andl %r9d,%r12d 3776 rorxl $25,%r9d,%r13d 3777 rorxl $11,%r9d,%esi 3778 leal (%rbx,%r14,1),%ebx 3779 leal (%rax,%r12,1),%eax 3780 andnl %r11d,%r9d,%r12d 3781 xorl %esi,%r13d 3782 rorxl $6,%r9d,%r14d 3783 leal (%rax,%r12,1),%eax 3784 xorl %r14d,%r13d 3785 movl %ebx,%esi 3786 rorxl $22,%ebx,%r12d 3787 leal (%rax,%r13,1),%eax 3788 xorl %ecx,%esi 3789 rorxl $13,%ebx,%r14d 3790 rorxl $2,%ebx,%r13d 3791 leal (%r8,%rax,1),%r8d 3792 andl %esi,%r15d 3793 vaesenc %xmm10,%xmm9,%xmm9 3794 vmovdqu 112-128(%rdi),%xmm10 3795 xorl %r12d,%r14d 3796 xorl %ecx,%r15d 3797 xorl %r13d,%r14d 3798 leal (%rax,%r15,1),%eax 3799 movl %r9d,%r12d 3800 leaq -64(%rbp),%rbp 3801 addl 0+16(%rbp),%r11d 3802 andl %r8d,%r12d 3803 rorxl $25,%r8d,%r13d 3804 rorxl $11,%r8d,%r15d 3805 leal (%rax,%r14,1),%eax 3806 leal (%r11,%r12,1),%r11d 3807 andnl %r10d,%r8d,%r12d 3808 xorl %r15d,%r13d 3809 rorxl $6,%r8d,%r14d 3810 leal (%r11,%r12,1),%r11d 3811 xorl %r14d,%r13d 3812 movl %eax,%r15d 3813 rorxl $22,%eax,%r12d 3814 leal (%r11,%r13,1),%r11d 3815 xorl %ebx,%r15d 3816 rorxl $13,%eax,%r14d 3817 rorxl $2,%eax,%r13d 3818 leal (%rdx,%r11,1),%edx 3819 andl %r15d,%esi 3820 vaesenc %xmm10,%xmm9,%xmm9 3821 vmovdqu 128-128(%rdi),%xmm10 3822 xorl %r12d,%r14d 3823 xorl %ebx,%esi 3824 xorl %r13d,%r14d 3825 leal (%r11,%rsi,1),%r11d 3826 movl %r8d,%r12d 3827 addl 4+16(%rbp),%r10d 3828 andl %edx,%r12d 3829 rorxl $25,%edx,%r13d 3830 rorxl $11,%edx,%esi 3831 leal (%r11,%r14,1),%r11d 3832 leal (%r10,%r12,1),%r10d 3833 andnl %r9d,%edx,%r12d 3834 xorl %esi,%r13d 3835 rorxl $6,%edx,%r14d 3836 leal (%r10,%r12,1),%r10d 3837 xorl %r14d,%r13d 3838 movl %r11d,%esi 3839 rorxl $22,%r11d,%r12d 3840 leal (%r10,%r13,1),%r10d 3841 xorl %eax,%esi 3842 rorxl $13,%r11d,%r14d 3843 rorxl $2,%r11d,%r13d 3844 leal (%rcx,%r10,1),%ecx 3845 andl %esi,%r15d 3846 vaesenc %xmm10,%xmm9,%xmm9 3847 vmovdqu 144-128(%rdi),%xmm10 3848 xorl %r12d,%r14d 3849 xorl %eax,%r15d 3850 xorl %r13d,%r14d 3851 leal (%r10,%r15,1),%r10d 3852 movl %edx,%r12d 3853 addl 8+16(%rbp),%r9d 3854 andl %ecx,%r12d 3855 rorxl $25,%ecx,%r13d 3856 rorxl $11,%ecx,%r15d 3857 leal (%r10,%r14,1),%r10d 3858 leal (%r9,%r12,1),%r9d 3859 andnl %r8d,%ecx,%r12d 3860 xorl %r15d,%r13d 3861 rorxl $6,%ecx,%r14d 3862 leal (%r9,%r12,1),%r9d 3863 xorl %r14d,%r13d 3864 movl %r10d,%r15d 3865 rorxl $22,%r10d,%r12d 3866 leal (%r9,%r13,1),%r9d 3867 xorl %r11d,%r15d 3868 rorxl $13,%r10d,%r14d 3869 rorxl $2,%r10d,%r13d 3870 leal (%rbx,%r9,1),%ebx 3871 andl %r15d,%esi 3872 vaesenc %xmm10,%xmm9,%xmm9 3873 vmovdqu 160-128(%rdi),%xmm10 3874 xorl %r12d,%r14d 3875 xorl %r11d,%esi 3876 xorl %r13d,%r14d 3877 leal (%r9,%rsi,1),%r9d 3878 movl %ecx,%r12d 3879 addl 12+16(%rbp),%r8d 3880 andl %ebx,%r12d 3881 rorxl $25,%ebx,%r13d 3882 rorxl $11,%ebx,%esi 3883 leal (%r9,%r14,1),%r9d 3884 leal (%r8,%r12,1),%r8d 3885 andnl %edx,%ebx,%r12d 3886 xorl %esi,%r13d 3887 rorxl $6,%ebx,%r14d 3888 leal (%r8,%r12,1),%r8d 3889 xorl %r14d,%r13d 3890 movl %r9d,%esi 3891 rorxl $22,%r9d,%r12d 3892 leal (%r8,%r13,1),%r8d 3893 xorl %r10d,%esi 3894 rorxl $13,%r9d,%r14d 3895 rorxl $2,%r9d,%r13d 3896 leal (%rax,%r8,1),%eax 3897 andl %esi,%r15d 3898 vaesenclast %xmm10,%xmm9,%xmm11 3899 vaesenc %xmm10,%xmm9,%xmm9 3900 vmovdqu 176-128(%rdi),%xmm10 3901 xorl %r12d,%r14d 3902 xorl %r10d,%r15d 3903 xorl %r13d,%r14d 3904 leal (%r8,%r15,1),%r8d 3905 movl %ebx,%r12d 3906 addl 32+16(%rbp),%edx 3907 andl %eax,%r12d 3908 rorxl $25,%eax,%r13d 3909 rorxl $11,%eax,%r15d 3910 leal (%r8,%r14,1),%r8d 3911 leal (%rdx,%r12,1),%edx 3912 andnl %ecx,%eax,%r12d 3913 xorl %r15d,%r13d 3914 rorxl $6,%eax,%r14d 3915 leal (%rdx,%r12,1),%edx 3916 xorl %r14d,%r13d 3917 movl %r8d,%r15d 3918 rorxl $22,%r8d,%r12d 3919 leal (%rdx,%r13,1),%edx 3920 xorl %r9d,%r15d 3921 rorxl $13,%r8d,%r14d 3922 rorxl $2,%r8d,%r13d 3923 leal (%r11,%rdx,1),%r11d 3924 andl %r15d,%esi 3925 vpand %xmm12,%xmm11,%xmm8 3926 vaesenc %xmm10,%xmm9,%xmm9 3927 vmovdqu 192-128(%rdi),%xmm10 3928 xorl %r12d,%r14d 3929 xorl %r9d,%esi 3930 xorl %r13d,%r14d 3931 leal (%rdx,%rsi,1),%edx 3932 movl %eax,%r12d 3933 addl 36+16(%rbp),%ecx 3934 andl %r11d,%r12d 3935 rorxl $25,%r11d,%r13d 3936 rorxl $11,%r11d,%esi 3937 leal (%rdx,%r14,1),%edx 3938 leal (%rcx,%r12,1),%ecx 3939 andnl %ebx,%r11d,%r12d 3940 xorl %esi,%r13d 3941 rorxl $6,%r11d,%r14d 3942 leal (%rcx,%r12,1),%ecx 3943 xorl %r14d,%r13d 3944 movl %edx,%esi 3945 rorxl $22,%edx,%r12d 3946 leal (%rcx,%r13,1),%ecx 3947 xorl %r8d,%esi 3948 rorxl $13,%edx,%r14d 3949 rorxl $2,%edx,%r13d 3950 leal (%r10,%rcx,1),%r10d 3951 andl %esi,%r15d 3952 vaesenclast %xmm10,%xmm9,%xmm11 3953 vaesenc %xmm10,%xmm9,%xmm9 3954 vmovdqu 208-128(%rdi),%xmm10 3955 xorl %r12d,%r14d 3956 xorl %r8d,%r15d 3957 xorl %r13d,%r14d 3958 leal (%rcx,%r15,1),%ecx 3959 movl %r11d,%r12d 3960 addl 40+16(%rbp),%ebx 3961 andl %r10d,%r12d 3962 rorxl $25,%r10d,%r13d 3963 rorxl $11,%r10d,%r15d 3964 leal (%rcx,%r14,1),%ecx 3965 leal (%rbx,%r12,1),%ebx 3966 andnl %eax,%r10d,%r12d 3967 xorl %r15d,%r13d 3968 rorxl $6,%r10d,%r14d 3969 leal (%rbx,%r12,1),%ebx 3970 xorl %r14d,%r13d 3971 movl %ecx,%r15d 3972 rorxl $22,%ecx,%r12d 3973 leal (%rbx,%r13,1),%ebx 3974 xorl %edx,%r15d 3975 rorxl $13,%ecx,%r14d 3976 rorxl $2,%ecx,%r13d 3977 leal (%r9,%rbx,1),%r9d 3978 andl %r15d,%esi 3979 vpand %xmm13,%xmm11,%xmm11 3980 vaesenc %xmm10,%xmm9,%xmm9 3981 vmovdqu 224-128(%rdi),%xmm10 3982 xorl %r12d,%r14d 3983 xorl %edx,%esi 3984 xorl %r13d,%r14d 3985 leal (%rbx,%rsi,1),%ebx 3986 movl %r10d,%r12d 3987 addl 44+16(%rbp),%eax 3988 andl %r9d,%r12d 3989 rorxl $25,%r9d,%r13d 3990 rorxl $11,%r9d,%esi 3991 leal (%rbx,%r14,1),%ebx 3992 leal (%rax,%r12,1),%eax 3993 andnl %r11d,%r9d,%r12d 3994 xorl %esi,%r13d 3995 rorxl $6,%r9d,%r14d 3996 leal (%rax,%r12,1),%eax 3997 xorl %r14d,%r13d 3998 movl %ebx,%esi 3999 rorxl $22,%ebx,%r12d 4000 leal (%rax,%r13,1),%eax 4001 xorl %ecx,%esi 4002 rorxl $13,%ebx,%r14d 4003 rorxl $2,%ebx,%r13d 4004 leal (%r8,%rax,1),%r8d 4005 andl %esi,%r15d 4006 vpor %xmm11,%xmm8,%xmm8 4007 vaesenclast %xmm10,%xmm9,%xmm11 4008 vmovdqu 0-128(%rdi),%xmm10 4009 xorl %r12d,%r14d 4010 xorl %ecx,%r15d 4011 xorl %r13d,%r14d 4012 leal (%rax,%r15,1),%eax 4013 movl %r9d,%r12d 4014 vmovq %xmm15,%r13 4015 vpextrq $1,%xmm15,%r15 4016 vpand %xmm14,%xmm11,%xmm11 4017 vpor %xmm11,%xmm8,%xmm8 4018 leaq -64(%rbp),%rbp 4019 vmovdqu %xmm8,(%r15,%r13,1) 4020 leaq 16(%r13),%r13 4021 cmpq %rsp,%rbp 4022 jae .Lower_avx2 4023 4024 movq 552(%rsp),%r15 4025 leaq 64(%r13),%r13 4026 movq 560(%rsp),%rsi 4027 addl %r14d,%eax 4028 leaq 448(%rsp),%rsp 4029 4030 addl 0(%r15),%eax 4031 addl 4(%r15),%ebx 4032 addl 8(%r15),%ecx 4033 addl 12(%r15),%edx 4034 addl 16(%r15),%r8d 4035 addl 20(%r15),%r9d 4036 addl 24(%r15),%r10d 4037 leaq (%rsi,%r13,1),%r12 4038 addl 28(%r15),%r11d 4039 4040 cmpq 64+16(%rsp),%r13 4041 4042 movl %eax,0(%r15) 4043 cmoveq %rsp,%r12 4044 movl %ebx,4(%r15) 4045 movl %ecx,8(%r15) 4046 movl %edx,12(%r15) 4047 movl %r8d,16(%r15) 4048 movl %r9d,20(%r15) 4049 movl %r10d,24(%r15) 4050 movl %r11d,28(%r15) 4051 4052 jbe .Loop_avx2 4053 leaq (%rsp),%rbp 4054 4055 4056 .cfi_escape 0x0f,0x06,0x76,0xf8,0x00,0x06,0x23,0x08 4057 4058 .Ldone_avx2: 4059 movq 64+32(%rbp),%r8 4060 movq 64+56(%rbp),%rsi 4061 .cfi_def_cfa %rsi,8 4062 vmovdqu %xmm8,(%r8) 4063 vzeroall 4064 movq -48(%rsi),%r15 4065 .cfi_restore %r15 4066 movq -40(%rsi),%r14 4067 .cfi_restore %r14 4068 movq -32(%rsi),%r13 4069 .cfi_restore %r13 4070 movq -24(%rsi),%r12 4071 .cfi_restore %r12 4072 movq -16(%rsi),%rbp 4073 .cfi_restore %rbp 4074 movq -8(%rsi),%rbx 4075 .cfi_restore %rbx 4076 leaq (%rsi),%rsp 4077 .cfi_def_cfa_register %rsp 4078 .Lepilogue_avx2: 4079 .byte 0xf3,0xc3 4080 .cfi_endproc 4081 .size aesni_cbc_sha256_enc_avx2,.-aesni_cbc_sha256_enc_avx2 4082 .type aesni_cbc_sha256_enc_shaext,@function 4083 .align 32 4084 aesni_cbc_sha256_enc_shaext: 4085 .cfi_startproc 4086 movq 8(%rsp),%r10 4087 leaq K256+128(%rip),%rax 4088 movdqu (%r9),%xmm1 4089 movdqu 16(%r9),%xmm2 4090 movdqa 512-128(%rax),%xmm3 4091 4092 movl 240(%rcx),%r11d 4093 subq %rdi,%rsi 4094 movups (%rcx),%xmm15 4095 movups (%r8),%xmm6 4096 movups 16(%rcx),%xmm4 4097 leaq 112(%rcx),%rcx 4098 4099 pshufd $0x1b,%xmm1,%xmm0 4100 pshufd $0xb1,%xmm1,%xmm1 4101 pshufd $0x1b,%xmm2,%xmm2 4102 movdqa %xmm3,%xmm7 4103 .byte 102,15,58,15,202,8 4104 punpcklqdq %xmm0,%xmm2 4105 4106 jmp .Loop_shaext 4107 4108 .align 16 4109 .Loop_shaext: 4110 movdqu (%r10),%xmm10 4111 movdqu 16(%r10),%xmm11 4112 movdqu 32(%r10),%xmm12 4113 .byte 102,68,15,56,0,211 4114 movdqu 48(%r10),%xmm13 4115 4116 movdqa 0-128(%rax),%xmm0 4117 paddd %xmm10,%xmm0 4118 .byte 102,68,15,56,0,219 4119 movdqa %xmm2,%xmm9 4120 movdqa %xmm1,%xmm8 4121 movups 0(%rdi),%xmm14 4122 xorps %xmm15,%xmm14 4123 xorps %xmm14,%xmm6 4124 movups -80(%rcx),%xmm5 4125 aesenc %xmm4,%xmm6 4126 .byte 15,56,203,209 4127 pshufd $0x0e,%xmm0,%xmm0 4128 movups -64(%rcx),%xmm4 4129 aesenc %xmm5,%xmm6 4130 .byte 15,56,203,202 4131 4132 movdqa 32-128(%rax),%xmm0 4133 paddd %xmm11,%xmm0 4134 .byte 102,68,15,56,0,227 4135 leaq 64(%r10),%r10 4136 movups -48(%rcx),%xmm5 4137 aesenc %xmm4,%xmm6 4138 .byte 15,56,203,209 4139 pshufd $0x0e,%xmm0,%xmm0 4140 movups -32(%rcx),%xmm4 4141 aesenc %xmm5,%xmm6 4142 .byte 15,56,203,202 4143 4144 movdqa 64-128(%rax),%xmm0 4145 paddd %xmm12,%xmm0 4146 .byte 102,68,15,56,0,235 4147 .byte 69,15,56,204,211 4148 movups -16(%rcx),%xmm5 4149 aesenc %xmm4,%xmm6 4150 .byte 15,56,203,209 4151 pshufd $0x0e,%xmm0,%xmm0 4152 movdqa %xmm13,%xmm3 4153 .byte 102,65,15,58,15,220,4 4154 paddd %xmm3,%xmm10 4155 movups 0(%rcx),%xmm4 4156 aesenc %xmm5,%xmm6 4157 .byte 15,56,203,202 4158 4159 movdqa 96-128(%rax),%xmm0 4160 paddd %xmm13,%xmm0 4161 .byte 69,15,56,205,213 4162 .byte 69,15,56,204,220 4163 movups 16(%rcx),%xmm5 4164 aesenc %xmm4,%xmm6 4165 .byte 15,56,203,209 4166 pshufd $0x0e,%xmm0,%xmm0 4167 movups 32(%rcx),%xmm4 4168 aesenc %xmm5,%xmm6 4169 movdqa %xmm10,%xmm3 4170 .byte 102,65,15,58,15,221,4 4171 paddd %xmm3,%xmm11 4172 .byte 15,56,203,202 4173 movdqa 128-128(%rax),%xmm0 4174 paddd %xmm10,%xmm0 4175 .byte 69,15,56,205,218 4176 .byte 69,15,56,204,229 4177 movups 48(%rcx),%xmm5 4178 aesenc %xmm4,%xmm6 4179 .byte 15,56,203,209 4180 pshufd $0x0e,%xmm0,%xmm0 4181 movdqa %xmm11,%xmm3 4182 .byte 102,65,15,58,15,218,4 4183 paddd %xmm3,%xmm12 4184 cmpl $11,%r11d 4185 jb .Laesenclast1 4186 movups 64(%rcx),%xmm4 4187 aesenc %xmm5,%xmm6 4188 movups 80(%rcx),%xmm5 4189 aesenc %xmm4,%xmm6 4190 je .Laesenclast1 4191 movups 96(%rcx),%xmm4 4192 aesenc %xmm5,%xmm6 4193 movups 112(%rcx),%xmm5 4194 aesenc %xmm4,%xmm6 4195 .Laesenclast1: 4196 aesenclast %xmm5,%xmm6 4197 movups 16-112(%rcx),%xmm4 4198 nop 4199 .byte 15,56,203,202 4200 movups 16(%rdi),%xmm14 4201 xorps %xmm15,%xmm14 4202 movups %xmm6,0(%rsi,%rdi,1) 4203 xorps %xmm14,%xmm6 4204 movups -80(%rcx),%xmm5 4205 aesenc %xmm4,%xmm6 4206 movdqa 160-128(%rax),%xmm0 4207 paddd %xmm11,%xmm0 4208 .byte 69,15,56,205,227 4209 .byte 69,15,56,204,234 4210 movups -64(%rcx),%xmm4 4211 aesenc %xmm5,%xmm6 4212 .byte 15,56,203,209 4213 pshufd $0x0e,%xmm0,%xmm0 4214 movdqa %xmm12,%xmm3 4215 .byte 102,65,15,58,15,219,4 4216 paddd %xmm3,%xmm13 4217 movups -48(%rcx),%xmm5 4218 aesenc %xmm4,%xmm6 4219 .byte 15,56,203,202 4220 movdqa 192-128(%rax),%xmm0 4221 paddd %xmm12,%xmm0 4222 .byte 69,15,56,205,236 4223 .byte 69,15,56,204,211 4224 movups -32(%rcx),%xmm4 4225 aesenc %xmm5,%xmm6 4226 .byte 15,56,203,209 4227 pshufd $0x0e,%xmm0,%xmm0 4228 movdqa %xmm13,%xmm3 4229 .byte 102,65,15,58,15,220,4 4230 paddd %xmm3,%xmm10 4231 movups -16(%rcx),%xmm5 4232 aesenc %xmm4,%xmm6 4233 .byte 15,56,203,202 4234 movdqa 224-128(%rax),%xmm0 4235 paddd %xmm13,%xmm0 4236 .byte 69,15,56,205,213 4237 .byte 69,15,56,204,220 4238 movups 0(%rcx),%xmm4 4239 aesenc %xmm5,%xmm6 4240 .byte 15,56,203,209 4241 pshufd $0x0e,%xmm0,%xmm0 4242 movdqa %xmm10,%xmm3 4243 .byte 102,65,15,58,15,221,4 4244 paddd %xmm3,%xmm11 4245 movups 16(%rcx),%xmm5 4246 aesenc %xmm4,%xmm6 4247 .byte 15,56,203,202 4248 movdqa 256-128(%rax),%xmm0 4249 paddd %xmm10,%xmm0 4250 .byte 69,15,56,205,218 4251 .byte 69,15,56,204,229 4252 movups 32(%rcx),%xmm4 4253 aesenc %xmm5,%xmm6 4254 .byte 15,56,203,209 4255 pshufd $0x0e,%xmm0,%xmm0 4256 movdqa %xmm11,%xmm3 4257 .byte 102,65,15,58,15,218,4 4258 paddd %xmm3,%xmm12 4259 movups 48(%rcx),%xmm5 4260 aesenc %xmm4,%xmm6 4261 cmpl $11,%r11d 4262 jb .Laesenclast2 4263 movups 64(%rcx),%xmm4 4264 aesenc %xmm5,%xmm6 4265 movups 80(%rcx),%xmm5 4266 aesenc %xmm4,%xmm6 4267 je .Laesenclast2 4268 movups 96(%rcx),%xmm4 4269 aesenc %xmm5,%xmm6 4270 movups 112(%rcx),%xmm5 4271 aesenc %xmm4,%xmm6 4272 .Laesenclast2: 4273 aesenclast %xmm5,%xmm6 4274 movups 16-112(%rcx),%xmm4 4275 nop 4276 .byte 15,56,203,202 4277 movups 32(%rdi),%xmm14 4278 xorps %xmm15,%xmm14 4279 movups %xmm6,16(%rsi,%rdi,1) 4280 xorps %xmm14,%xmm6 4281 movups -80(%rcx),%xmm5 4282 aesenc %xmm4,%xmm6 4283 movdqa 288-128(%rax),%xmm0 4284 paddd %xmm11,%xmm0 4285 .byte 69,15,56,205,227 4286 .byte 69,15,56,204,234 4287 movups -64(%rcx),%xmm4 4288 aesenc %xmm5,%xmm6 4289 .byte 15,56,203,209 4290 pshufd $0x0e,%xmm0,%xmm0 4291 movdqa %xmm12,%xmm3 4292 .byte 102,65,15,58,15,219,4 4293 paddd %xmm3,%xmm13 4294 movups -48(%rcx),%xmm5 4295 aesenc %xmm4,%xmm6 4296 .byte 15,56,203,202 4297 movdqa 320-128(%rax),%xmm0 4298 paddd %xmm12,%xmm0 4299 .byte 69,15,56,205,236 4300 .byte 69,15,56,204,211 4301 movups -32(%rcx),%xmm4 4302 aesenc %xmm5,%xmm6 4303 .byte 15,56,203,209 4304 pshufd $0x0e,%xmm0,%xmm0 4305 movdqa %xmm13,%xmm3 4306 .byte 102,65,15,58,15,220,4 4307 paddd %xmm3,%xmm10 4308 movups -16(%rcx),%xmm5 4309 aesenc %xmm4,%xmm6 4310 .byte 15,56,203,202 4311 movdqa 352-128(%rax),%xmm0 4312 paddd %xmm13,%xmm0 4313 .byte 69,15,56,205,213 4314 .byte 69,15,56,204,220 4315 movups 0(%rcx),%xmm4 4316 aesenc %xmm5,%xmm6 4317 .byte 15,56,203,209 4318 pshufd $0x0e,%xmm0,%xmm0 4319 movdqa %xmm10,%xmm3 4320 .byte 102,65,15,58,15,221,4 4321 paddd %xmm3,%xmm11 4322 movups 16(%rcx),%xmm5 4323 aesenc %xmm4,%xmm6 4324 .byte 15,56,203,202 4325 movdqa 384-128(%rax),%xmm0 4326 paddd %xmm10,%xmm0 4327 .byte 69,15,56,205,218 4328 .byte 69,15,56,204,229 4329 movups 32(%rcx),%xmm4 4330 aesenc %xmm5,%xmm6 4331 .byte 15,56,203,209 4332 pshufd $0x0e,%xmm0,%xmm0 4333 movdqa %xmm11,%xmm3 4334 .byte 102,65,15,58,15,218,4 4335 paddd %xmm3,%xmm12 4336 movups 48(%rcx),%xmm5 4337 aesenc %xmm4,%xmm6 4338 .byte 15,56,203,202 4339 movdqa 416-128(%rax),%xmm0 4340 paddd %xmm11,%xmm0 4341 .byte 69,15,56,205,227 4342 .byte 69,15,56,204,234 4343 cmpl $11,%r11d 4344 jb .Laesenclast3 4345 movups 64(%rcx),%xmm4 4346 aesenc %xmm5,%xmm6 4347 movups 80(%rcx),%xmm5 4348 aesenc %xmm4,%xmm6 4349 je .Laesenclast3 4350 movups 96(%rcx),%xmm4 4351 aesenc %xmm5,%xmm6 4352 movups 112(%rcx),%xmm5 4353 aesenc %xmm4,%xmm6 4354 .Laesenclast3: 4355 aesenclast %xmm5,%xmm6 4356 movups 16-112(%rcx),%xmm4 4357 nop 4358 .byte 15,56,203,209 4359 pshufd $0x0e,%xmm0,%xmm0 4360 movdqa %xmm12,%xmm3 4361 .byte 102,65,15,58,15,219,4 4362 paddd %xmm3,%xmm13 4363 movups 48(%rdi),%xmm14 4364 xorps %xmm15,%xmm14 4365 movups %xmm6,32(%rsi,%rdi,1) 4366 xorps %xmm14,%xmm6 4367 movups -80(%rcx),%xmm5 4368 aesenc %xmm4,%xmm6 4369 movups -64(%rcx),%xmm4 4370 aesenc %xmm5,%xmm6 4371 .byte 15,56,203,202 4372 4373 movdqa 448-128(%rax),%xmm0 4374 paddd %xmm12,%xmm0 4375 .byte 69,15,56,205,236 4376 movdqa %xmm7,%xmm3 4377 movups -48(%rcx),%xmm5 4378 aesenc %xmm4,%xmm6 4379 .byte 15,56,203,209 4380 pshufd $0x0e,%xmm0,%xmm0 4381 movups -32(%rcx),%xmm4 4382 aesenc %xmm5,%xmm6 4383 .byte 15,56,203,202 4384 4385 movdqa 480-128(%rax),%xmm0 4386 paddd %xmm13,%xmm0 4387 movups -16(%rcx),%xmm5 4388 aesenc %xmm4,%xmm6 4389 movups 0(%rcx),%xmm4 4390 aesenc %xmm5,%xmm6 4391 .byte 15,56,203,209 4392 pshufd $0x0e,%xmm0,%xmm0 4393 movups 16(%rcx),%xmm5 4394 aesenc %xmm4,%xmm6 4395 .byte 15,56,203,202 4396 4397 movups 32(%rcx),%xmm4 4398 aesenc %xmm5,%xmm6 4399 movups 48(%rcx),%xmm5 4400 aesenc %xmm4,%xmm6 4401 cmpl $11,%r11d 4402 jb .Laesenclast4 4403 movups 64(%rcx),%xmm4 4404 aesenc %xmm5,%xmm6 4405 movups 80(%rcx),%xmm5 4406 aesenc %xmm4,%xmm6 4407 je .Laesenclast4 4408 movups 96(%rcx),%xmm4 4409 aesenc %xmm5,%xmm6 4410 movups 112(%rcx),%xmm5 4411 aesenc %xmm4,%xmm6 4412 .Laesenclast4: 4413 aesenclast %xmm5,%xmm6 4414 movups 16-112(%rcx),%xmm4 4415 nop 4416 4417 paddd %xmm9,%xmm2 4418 paddd %xmm8,%xmm1 4419 4420 decq %rdx 4421 movups %xmm6,48(%rsi,%rdi,1) 4422 leaq 64(%rdi),%rdi 4423 jnz .Loop_shaext 4424 4425 pshufd $0xb1,%xmm2,%xmm2 4426 pshufd $0x1b,%xmm1,%xmm3 4427 pshufd $0xb1,%xmm1,%xmm1 4428 punpckhqdq %xmm2,%xmm1 4429 .byte 102,15,58,15,211,8 4430 4431 movups %xmm6,(%r8) 4432 movdqu %xmm1,(%r9) 4433 movdqu %xmm2,16(%r9) 4434 .byte 0xf3,0xc3 4435 .cfi_endproc 4436 .size aesni_cbc_sha256_enc_shaext,.-aesni_cbc_sha256_enc_shaext 4437
Indexes created Sat Feb 28 05:31:39 UTC 2026