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 .section .rodata 39 .align 64 40 .type K256,@object 41 K256: 42 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5 43 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5 44 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5 45 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5 46 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3 47 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3 48 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174 49 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174 50 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc 51 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc 52 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da 53 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da 54 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7 55 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7 56 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967 57 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967 58 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13 59 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13 60 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85 61 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85 62 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3 63 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3 64 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070 65 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070 66 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5 67 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5 68 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3 69 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3 70 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208 71 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208 72 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2 73 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2 74 75 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f 76 .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f 77 .long 0,0,0,0, 0,0,0,0, -1,-1,-1,-1 78 .long 0,0,0,0, 0,0,0,0 79 .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 80 .align 64 81 .previous 82 .type aesni_cbc_sha256_enc_xop,@function 83 .align 64 84 aesni_cbc_sha256_enc_xop: 85 .cfi_startproc 86 .Lxop_shortcut: 87 movq 8(%rsp),%r10 88 movq %rsp,%rax 89 .cfi_def_cfa_register %rax 90 pushq %rbx 91 .cfi_offset %rbx,-16 92 pushq %rbp 93 .cfi_offset %rbp,-24 94 pushq %r12 95 .cfi_offset %r12,-32 96 pushq %r13 97 .cfi_offset %r13,-40 98 pushq %r14 99 .cfi_offset %r14,-48 100 pushq %r15 101 .cfi_offset %r15,-56 102 subq $128,%rsp 103 andq $-64,%rsp 104 105 shlq $6,%rdx 106 subq %rdi,%rsi 107 subq %rdi,%r10 108 addq %rdi,%rdx 109 110 111 movq %rsi,64+8(%rsp) 112 movq %rdx,64+16(%rsp) 113 114 movq %r8,64+32(%rsp) 115 movq %r9,64+40(%rsp) 116 movq %r10,64+48(%rsp) 117 movq %rax,120(%rsp) 118 .cfi_escape 0x0f,0x06,0x77,0xf8,0x00,0x06,0x23,0x08 119 .Lprologue_xop: 120 vzeroall 121 122 movq %rdi,%r12 123 leaq 128(%rcx),%rdi 124 leaq K256+544(%rip),%r13 125 movl 240-128(%rdi),%r14d 126 movq %r9,%r15 127 movq %r10,%rsi 128 vmovdqu (%r8),%xmm8 129 subq $9,%r14 130 131 movl 0(%r15),%eax 132 movl 4(%r15),%ebx 133 movl 8(%r15),%ecx 134 movl 12(%r15),%edx 135 movl 16(%r15),%r8d 136 movl 20(%r15),%r9d 137 movl 24(%r15),%r10d 138 movl 28(%r15),%r11d 139 140 vmovdqa 0(%r13,%r14,8),%xmm14 141 vmovdqa 16(%r13,%r14,8),%xmm13 142 vmovdqa 32(%r13,%r14,8),%xmm12 143 vmovdqu 0-128(%rdi),%xmm10 144 jmp .Lloop_xop 145 .align 16 146 .Lloop_xop: 147 vmovdqa K256+512(%rip),%xmm7 148 vmovdqu 0(%rsi,%r12,1),%xmm0 149 vmovdqu 16(%rsi,%r12,1),%xmm1 150 vmovdqu 32(%rsi,%r12,1),%xmm2 151 vmovdqu 48(%rsi,%r12,1),%xmm3 152 vpshufb %xmm7,%xmm0,%xmm0 153 leaq K256(%rip),%rbp 154 vpshufb %xmm7,%xmm1,%xmm1 155 vpshufb %xmm7,%xmm2,%xmm2 156 vpaddd 0(%rbp),%xmm0,%xmm4 157 vpshufb %xmm7,%xmm3,%xmm3 158 vpaddd 32(%rbp),%xmm1,%xmm5 159 vpaddd 64(%rbp),%xmm2,%xmm6 160 vpaddd 96(%rbp),%xmm3,%xmm7 161 vmovdqa %xmm4,0(%rsp) 162 movl %eax,%r14d 163 vmovdqa %xmm5,16(%rsp) 164 movl %ebx,%esi 165 vmovdqa %xmm6,32(%rsp) 166 xorl %ecx,%esi 167 vmovdqa %xmm7,48(%rsp) 168 movl %r8d,%r13d 169 jmp .Lxop_00_47 170 171 .align 16 172 .Lxop_00_47: 173 subq $-32*4,%rbp 174 vmovdqu (%r12),%xmm9 175 movq %r12,64+0(%rsp) 176 vpalignr $4,%xmm0,%xmm1,%xmm4 177 rorl $14,%r13d 178 movl %r14d,%eax 179 vpalignr $4,%xmm2,%xmm3,%xmm7 180 movl %r9d,%r12d 181 xorl %r8d,%r13d 182 .byte 143,232,120,194,236,14 183 rorl $9,%r14d 184 xorl %r10d,%r12d 185 vpsrld $3,%xmm4,%xmm4 186 rorl $5,%r13d 187 xorl %eax,%r14d 188 vpaddd %xmm7,%xmm0,%xmm0 189 andl %r8d,%r12d 190 vpxor %xmm10,%xmm9,%xmm9 191 vmovdqu 16-128(%rdi),%xmm10 192 xorl %r8d,%r13d 193 addl 0(%rsp),%r11d 194 movl %eax,%r15d 195 .byte 143,232,120,194,245,11 196 rorl $11,%r14d 197 xorl %r10d,%r12d 198 vpxor %xmm5,%xmm4,%xmm4 199 xorl %ebx,%r15d 200 rorl $6,%r13d 201 addl %r12d,%r11d 202 andl %r15d,%esi 203 .byte 143,232,120,194,251,13 204 xorl %eax,%r14d 205 addl %r13d,%r11d 206 vpxor %xmm6,%xmm4,%xmm4 207 xorl %ebx,%esi 208 addl %r11d,%edx 209 vpsrld $10,%xmm3,%xmm6 210 rorl $2,%r14d 211 addl %esi,%r11d 212 vpaddd %xmm4,%xmm0,%xmm0 213 movl %edx,%r13d 214 addl %r11d,%r14d 215 .byte 143,232,120,194,239,2 216 rorl $14,%r13d 217 movl %r14d,%r11d 218 vpxor %xmm6,%xmm7,%xmm7 219 movl %r8d,%r12d 220 xorl %edx,%r13d 221 rorl $9,%r14d 222 xorl %r9d,%r12d 223 vpxor %xmm5,%xmm7,%xmm7 224 rorl $5,%r13d 225 xorl %r11d,%r14d 226 andl %edx,%r12d 227 vpxor %xmm8,%xmm9,%xmm9 228 xorl %edx,%r13d 229 vpsrldq $8,%xmm7,%xmm7 230 addl 4(%rsp),%r10d 231 movl %r11d,%esi 232 rorl $11,%r14d 233 xorl %r9d,%r12d 234 vpaddd %xmm7,%xmm0,%xmm0 235 xorl %eax,%esi 236 rorl $6,%r13d 237 addl %r12d,%r10d 238 andl %esi,%r15d 239 .byte 143,232,120,194,248,13 240 xorl %r11d,%r14d 241 addl %r13d,%r10d 242 vpsrld $10,%xmm0,%xmm6 243 xorl %eax,%r15d 244 addl %r10d,%ecx 245 .byte 143,232,120,194,239,2 246 rorl $2,%r14d 247 addl %r15d,%r10d 248 vpxor %xmm6,%xmm7,%xmm7 249 movl %ecx,%r13d 250 addl %r10d,%r14d 251 rorl $14,%r13d 252 movl %r14d,%r10d 253 vpxor %xmm5,%xmm7,%xmm7 254 movl %edx,%r12d 255 xorl %ecx,%r13d 256 rorl $9,%r14d 257 xorl %r8d,%r12d 258 vpslldq $8,%xmm7,%xmm7 259 rorl $5,%r13d 260 xorl %r10d,%r14d 261 andl %ecx,%r12d 262 vaesenc %xmm10,%xmm9,%xmm9 263 vmovdqu 32-128(%rdi),%xmm10 264 xorl %ecx,%r13d 265 vpaddd %xmm7,%xmm0,%xmm0 266 addl 8(%rsp),%r9d 267 movl %r10d,%r15d 268 rorl $11,%r14d 269 xorl %r8d,%r12d 270 vpaddd 0(%rbp),%xmm0,%xmm6 271 xorl %r11d,%r15d 272 rorl $6,%r13d 273 addl %r12d,%r9d 274 andl %r15d,%esi 275 xorl %r10d,%r14d 276 addl %r13d,%r9d 277 xorl %r11d,%esi 278 addl %r9d,%ebx 279 rorl $2,%r14d 280 addl %esi,%r9d 281 movl %ebx,%r13d 282 addl %r9d,%r14d 283 rorl $14,%r13d 284 movl %r14d,%r9d 285 movl %ecx,%r12d 286 xorl %ebx,%r13d 287 rorl $9,%r14d 288 xorl %edx,%r12d 289 rorl $5,%r13d 290 xorl %r9d,%r14d 291 andl %ebx,%r12d 292 vaesenc %xmm10,%xmm9,%xmm9 293 vmovdqu 48-128(%rdi),%xmm10 294 xorl %ebx,%r13d 295 addl 12(%rsp),%r8d 296 movl %r9d,%esi 297 rorl $11,%r14d 298 xorl %edx,%r12d 299 xorl %r10d,%esi 300 rorl $6,%r13d 301 addl %r12d,%r8d 302 andl %esi,%r15d 303 xorl %r9d,%r14d 304 addl %r13d,%r8d 305 xorl %r10d,%r15d 306 addl %r8d,%eax 307 rorl $2,%r14d 308 addl %r15d,%r8d 309 movl %eax,%r13d 310 addl %r8d,%r14d 311 vmovdqa %xmm6,0(%rsp) 312 vpalignr $4,%xmm1,%xmm2,%xmm4 313 rorl $14,%r13d 314 movl %r14d,%r8d 315 vpalignr $4,%xmm3,%xmm0,%xmm7 316 movl %ebx,%r12d 317 xorl %eax,%r13d 318 .byte 143,232,120,194,236,14 319 rorl $9,%r14d 320 xorl %ecx,%r12d 321 vpsrld $3,%xmm4,%xmm4 322 rorl $5,%r13d 323 xorl %r8d,%r14d 324 vpaddd %xmm7,%xmm1,%xmm1 325 andl %eax,%r12d 326 vaesenc %xmm10,%xmm9,%xmm9 327 vmovdqu 64-128(%rdi),%xmm10 328 xorl %eax,%r13d 329 addl 16(%rsp),%edx 330 movl %r8d,%r15d 331 .byte 143,232,120,194,245,11 332 rorl $11,%r14d 333 xorl %ecx,%r12d 334 vpxor %xmm5,%xmm4,%xmm4 335 xorl %r9d,%r15d 336 rorl $6,%r13d 337 addl %r12d,%edx 338 andl %r15d,%esi 339 .byte 143,232,120,194,248,13 340 xorl %r8d,%r14d 341 addl %r13d,%edx 342 vpxor %xmm6,%xmm4,%xmm4 343 xorl %r9d,%esi 344 addl %edx,%r11d 345 vpsrld $10,%xmm0,%xmm6 346 rorl $2,%r14d 347 addl %esi,%edx 348 vpaddd %xmm4,%xmm1,%xmm1 349 movl %r11d,%r13d 350 addl %edx,%r14d 351 .byte 143,232,120,194,239,2 352 rorl $14,%r13d 353 movl %r14d,%edx 354 vpxor %xmm6,%xmm7,%xmm7 355 movl %eax,%r12d 356 xorl %r11d,%r13d 357 rorl $9,%r14d 358 xorl %ebx,%r12d 359 vpxor %xmm5,%xmm7,%xmm7 360 rorl $5,%r13d 361 xorl %edx,%r14d 362 andl %r11d,%r12d 363 vaesenc %xmm10,%xmm9,%xmm9 364 vmovdqu 80-128(%rdi),%xmm10 365 xorl %r11d,%r13d 366 vpsrldq $8,%xmm7,%xmm7 367 addl 20(%rsp),%ecx 368 movl %edx,%esi 369 rorl $11,%r14d 370 xorl %ebx,%r12d 371 vpaddd %xmm7,%xmm1,%xmm1 372 xorl %r8d,%esi 373 rorl $6,%r13d 374 addl %r12d,%ecx 375 andl %esi,%r15d 376 .byte 143,232,120,194,249,13 377 xorl %edx,%r14d 378 addl %r13d,%ecx 379 vpsrld $10,%xmm1,%xmm6 380 xorl %r8d,%r15d 381 addl %ecx,%r10d 382 .byte 143,232,120,194,239,2 383 rorl $2,%r14d 384 addl %r15d,%ecx 385 vpxor %xmm6,%xmm7,%xmm7 386 movl %r10d,%r13d 387 addl %ecx,%r14d 388 rorl $14,%r13d 389 movl %r14d,%ecx 390 vpxor %xmm5,%xmm7,%xmm7 391 movl %r11d,%r12d 392 xorl %r10d,%r13d 393 rorl $9,%r14d 394 xorl %eax,%r12d 395 vpslldq $8,%xmm7,%xmm7 396 rorl $5,%r13d 397 xorl %ecx,%r14d 398 andl %r10d,%r12d 399 vaesenc %xmm10,%xmm9,%xmm9 400 vmovdqu 96-128(%rdi),%xmm10 401 xorl %r10d,%r13d 402 vpaddd %xmm7,%xmm1,%xmm1 403 addl 24(%rsp),%ebx 404 movl %ecx,%r15d 405 rorl $11,%r14d 406 xorl %eax,%r12d 407 vpaddd 32(%rbp),%xmm1,%xmm6 408 xorl %edx,%r15d 409 rorl $6,%r13d 410 addl %r12d,%ebx 411 andl %r15d,%esi 412 xorl %ecx,%r14d 413 addl %r13d,%ebx 414 xorl %edx,%esi 415 addl %ebx,%r9d 416 rorl $2,%r14d 417 addl %esi,%ebx 418 movl %r9d,%r13d 419 addl %ebx,%r14d 420 rorl $14,%r13d 421 movl %r14d,%ebx 422 movl %r10d,%r12d 423 xorl %r9d,%r13d 424 rorl $9,%r14d 425 xorl %r11d,%r12d 426 rorl $5,%r13d 427 xorl %ebx,%r14d 428 andl %r9d,%r12d 429 vaesenc %xmm10,%xmm9,%xmm9 430 vmovdqu 112-128(%rdi),%xmm10 431 xorl %r9d,%r13d 432 addl 28(%rsp),%eax 433 movl %ebx,%esi 434 rorl $11,%r14d 435 xorl %r11d,%r12d 436 xorl %ecx,%esi 437 rorl $6,%r13d 438 addl %r12d,%eax 439 andl %esi,%r15d 440 xorl %ebx,%r14d 441 addl %r13d,%eax 442 xorl %ecx,%r15d 443 addl %eax,%r8d 444 rorl $2,%r14d 445 addl %r15d,%eax 446 movl %r8d,%r13d 447 addl %eax,%r14d 448 vmovdqa %xmm6,16(%rsp) 449 vpalignr $4,%xmm2,%xmm3,%xmm4 450 rorl $14,%r13d 451 movl %r14d,%eax 452 vpalignr $4,%xmm0,%xmm1,%xmm7 453 movl %r9d,%r12d 454 xorl %r8d,%r13d 455 .byte 143,232,120,194,236,14 456 rorl $9,%r14d 457 xorl %r10d,%r12d 458 vpsrld $3,%xmm4,%xmm4 459 rorl $5,%r13d 460 xorl %eax,%r14d 461 vpaddd %xmm7,%xmm2,%xmm2 462 andl %r8d,%r12d 463 vaesenc %xmm10,%xmm9,%xmm9 464 vmovdqu 128-128(%rdi),%xmm10 465 xorl %r8d,%r13d 466 addl 32(%rsp),%r11d 467 movl %eax,%r15d 468 .byte 143,232,120,194,245,11 469 rorl $11,%r14d 470 xorl %r10d,%r12d 471 vpxor %xmm5,%xmm4,%xmm4 472 xorl %ebx,%r15d 473 rorl $6,%r13d 474 addl %r12d,%r11d 475 andl %r15d,%esi 476 .byte 143,232,120,194,249,13 477 xorl %eax,%r14d 478 addl %r13d,%r11d 479 vpxor %xmm6,%xmm4,%xmm4 480 xorl %ebx,%esi 481 addl %r11d,%edx 482 vpsrld $10,%xmm1,%xmm6 483 rorl $2,%r14d 484 addl %esi,%r11d 485 vpaddd %xmm4,%xmm2,%xmm2 486 movl %edx,%r13d 487 addl %r11d,%r14d 488 .byte 143,232,120,194,239,2 489 rorl $14,%r13d 490 movl %r14d,%r11d 491 vpxor %xmm6,%xmm7,%xmm7 492 movl %r8d,%r12d 493 xorl %edx,%r13d 494 rorl $9,%r14d 495 xorl %r9d,%r12d 496 vpxor %xmm5,%xmm7,%xmm7 497 rorl $5,%r13d 498 xorl %r11d,%r14d 499 andl %edx,%r12d 500 vaesenc %xmm10,%xmm9,%xmm9 501 vmovdqu 144-128(%rdi),%xmm10 502 xorl %edx,%r13d 503 vpsrldq $8,%xmm7,%xmm7 504 addl 36(%rsp),%r10d 505 movl %r11d,%esi 506 rorl $11,%r14d 507 xorl %r9d,%r12d 508 vpaddd %xmm7,%xmm2,%xmm2 509 xorl %eax,%esi 510 rorl $6,%r13d 511 addl %r12d,%r10d 512 andl %esi,%r15d 513 .byte 143,232,120,194,250,13 514 xorl %r11d,%r14d 515 addl %r13d,%r10d 516 vpsrld $10,%xmm2,%xmm6 517 xorl %eax,%r15d 518 addl %r10d,%ecx 519 .byte 143,232,120,194,239,2 520 rorl $2,%r14d 521 addl %r15d,%r10d 522 vpxor %xmm6,%xmm7,%xmm7 523 movl %ecx,%r13d 524 addl %r10d,%r14d 525 rorl $14,%r13d 526 movl %r14d,%r10d 527 vpxor %xmm5,%xmm7,%xmm7 528 movl %edx,%r12d 529 xorl %ecx,%r13d 530 rorl $9,%r14d 531 xorl %r8d,%r12d 532 vpslldq $8,%xmm7,%xmm7 533 rorl $5,%r13d 534 xorl %r10d,%r14d 535 andl %ecx,%r12d 536 vaesenc %xmm10,%xmm9,%xmm9 537 vmovdqu 160-128(%rdi),%xmm10 538 xorl %ecx,%r13d 539 vpaddd %xmm7,%xmm2,%xmm2 540 addl 40(%rsp),%r9d 541 movl %r10d,%r15d 542 rorl $11,%r14d 543 xorl %r8d,%r12d 544 vpaddd 64(%rbp),%xmm2,%xmm6 545 xorl %r11d,%r15d 546 rorl $6,%r13d 547 addl %r12d,%r9d 548 andl %r15d,%esi 549 xorl %r10d,%r14d 550 addl %r13d,%r9d 551 xorl %r11d,%esi 552 addl %r9d,%ebx 553 rorl $2,%r14d 554 addl %esi,%r9d 555 movl %ebx,%r13d 556 addl %r9d,%r14d 557 rorl $14,%r13d 558 movl %r14d,%r9d 559 movl %ecx,%r12d 560 xorl %ebx,%r13d 561 rorl $9,%r14d 562 xorl %edx,%r12d 563 rorl $5,%r13d 564 xorl %r9d,%r14d 565 andl %ebx,%r12d 566 vaesenclast %xmm10,%xmm9,%xmm11 567 vaesenc %xmm10,%xmm9,%xmm9 568 vmovdqu 176-128(%rdi),%xmm10 569 xorl %ebx,%r13d 570 addl 44(%rsp),%r8d 571 movl %r9d,%esi 572 rorl $11,%r14d 573 xorl %edx,%r12d 574 xorl %r10d,%esi 575 rorl $6,%r13d 576 addl %r12d,%r8d 577 andl %esi,%r15d 578 xorl %r9d,%r14d 579 addl %r13d,%r8d 580 xorl %r10d,%r15d 581 addl %r8d,%eax 582 rorl $2,%r14d 583 addl %r15d,%r8d 584 movl %eax,%r13d 585 addl %r8d,%r14d 586 vmovdqa %xmm6,32(%rsp) 587 vpalignr $4,%xmm3,%xmm0,%xmm4 588 rorl $14,%r13d 589 movl %r14d,%r8d 590 vpalignr $4,%xmm1,%xmm2,%xmm7 591 movl %ebx,%r12d 592 xorl %eax,%r13d 593 .byte 143,232,120,194,236,14 594 rorl $9,%r14d 595 xorl %ecx,%r12d 596 vpsrld $3,%xmm4,%xmm4 597 rorl $5,%r13d 598 xorl %r8d,%r14d 599 vpaddd %xmm7,%xmm3,%xmm3 600 andl %eax,%r12d 601 vpand %xmm12,%xmm11,%xmm8 602 vaesenc %xmm10,%xmm9,%xmm9 603 vmovdqu 192-128(%rdi),%xmm10 604 xorl %eax,%r13d 605 addl 48(%rsp),%edx 606 movl %r8d,%r15d 607 .byte 143,232,120,194,245,11 608 rorl $11,%r14d 609 xorl %ecx,%r12d 610 vpxor %xmm5,%xmm4,%xmm4 611 xorl %r9d,%r15d 612 rorl $6,%r13d 613 addl %r12d,%edx 614 andl %r15d,%esi 615 .byte 143,232,120,194,250,13 616 xorl %r8d,%r14d 617 addl %r13d,%edx 618 vpxor %xmm6,%xmm4,%xmm4 619 xorl %r9d,%esi 620 addl %edx,%r11d 621 vpsrld $10,%xmm2,%xmm6 622 rorl $2,%r14d 623 addl %esi,%edx 624 vpaddd %xmm4,%xmm3,%xmm3 625 movl %r11d,%r13d 626 addl %edx,%r14d 627 .byte 143,232,120,194,239,2 628 rorl $14,%r13d 629 movl %r14d,%edx 630 vpxor %xmm6,%xmm7,%xmm7 631 movl %eax,%r12d 632 xorl %r11d,%r13d 633 rorl $9,%r14d 634 xorl %ebx,%r12d 635 vpxor %xmm5,%xmm7,%xmm7 636 rorl $5,%r13d 637 xorl %edx,%r14d 638 andl %r11d,%r12d 639 vaesenclast %xmm10,%xmm9,%xmm11 640 vaesenc %xmm10,%xmm9,%xmm9 641 vmovdqu 208-128(%rdi),%xmm10 642 xorl %r11d,%r13d 643 vpsrldq $8,%xmm7,%xmm7 644 addl 52(%rsp),%ecx 645 movl %edx,%esi 646 rorl $11,%r14d 647 xorl %ebx,%r12d 648 vpaddd %xmm7,%xmm3,%xmm3 649 xorl %r8d,%esi 650 rorl $6,%r13d 651 addl %r12d,%ecx 652 andl %esi,%r15d 653 .byte 143,232,120,194,251,13 654 xorl %edx,%r14d 655 addl %r13d,%ecx 656 vpsrld $10,%xmm3,%xmm6 657 xorl %r8d,%r15d 658 addl %ecx,%r10d 659 .byte 143,232,120,194,239,2 660 rorl $2,%r14d 661 addl %r15d,%ecx 662 vpxor %xmm6,%xmm7,%xmm7 663 movl %r10d,%r13d 664 addl %ecx,%r14d 665 rorl $14,%r13d 666 movl %r14d,%ecx 667 vpxor %xmm5,%xmm7,%xmm7 668 movl %r11d,%r12d 669 xorl %r10d,%r13d 670 rorl $9,%r14d 671 xorl %eax,%r12d 672 vpslldq $8,%xmm7,%xmm7 673 rorl $5,%r13d 674 xorl %ecx,%r14d 675 andl %r10d,%r12d 676 vpand %xmm13,%xmm11,%xmm11 677 vaesenc %xmm10,%xmm9,%xmm9 678 vmovdqu 224-128(%rdi),%xmm10 679 xorl %r10d,%r13d 680 vpaddd %xmm7,%xmm3,%xmm3 681 addl 56(%rsp),%ebx 682 movl %ecx,%r15d 683 rorl $11,%r14d 684 xorl %eax,%r12d 685 vpaddd 96(%rbp),%xmm3,%xmm6 686 xorl %edx,%r15d 687 rorl $6,%r13d 688 addl %r12d,%ebx 689 andl %r15d,%esi 690 xorl %ecx,%r14d 691 addl %r13d,%ebx 692 xorl %edx,%esi 693 addl %ebx,%r9d 694 rorl $2,%r14d 695 addl %esi,%ebx 696 movl %r9d,%r13d 697 addl %ebx,%r14d 698 rorl $14,%r13d 699 movl %r14d,%ebx 700 movl %r10d,%r12d 701 xorl %r9d,%r13d 702 rorl $9,%r14d 703 xorl %r11d,%r12d 704 rorl $5,%r13d 705 xorl %ebx,%r14d 706 andl %r9d,%r12d 707 vpor %xmm11,%xmm8,%xmm8 708 vaesenclast %xmm10,%xmm9,%xmm11 709 vmovdqu 0-128(%rdi),%xmm10 710 xorl %r9d,%r13d 711 addl 60(%rsp),%eax 712 movl %ebx,%esi 713 rorl $11,%r14d 714 xorl %r11d,%r12d 715 xorl %ecx,%esi 716 rorl $6,%r13d 717 addl %r12d,%eax 718 andl %esi,%r15d 719 xorl %ebx,%r14d 720 addl %r13d,%eax 721 xorl %ecx,%r15d 722 addl %eax,%r8d 723 rorl $2,%r14d 724 addl %r15d,%eax 725 movl %r8d,%r13d 726 addl %eax,%r14d 727 vmovdqa %xmm6,48(%rsp) 728 movq 64+0(%rsp),%r12 729 vpand %xmm14,%xmm11,%xmm11 730 movq 64+8(%rsp),%r15 731 vpor %xmm11,%xmm8,%xmm8 732 vmovdqu %xmm8,(%r15,%r12,1) 733 leaq 16(%r12),%r12 734 cmpb $0,131(%rbp) 735 jne .Lxop_00_47 736 vmovdqu (%r12),%xmm9 737 movq %r12,64+0(%rsp) 738 rorl $14,%r13d 739 movl %r14d,%eax 740 movl %r9d,%r12d 741 xorl %r8d,%r13d 742 rorl $9,%r14d 743 xorl %r10d,%r12d 744 rorl $5,%r13d 745 xorl %eax,%r14d 746 andl %r8d,%r12d 747 vpxor %xmm10,%xmm9,%xmm9 748 vmovdqu 16-128(%rdi),%xmm10 749 xorl %r8d,%r13d 750 addl 0(%rsp),%r11d 751 movl %eax,%r15d 752 rorl $11,%r14d 753 xorl %r10d,%r12d 754 xorl %ebx,%r15d 755 rorl $6,%r13d 756 addl %r12d,%r11d 757 andl %r15d,%esi 758 xorl %eax,%r14d 759 addl %r13d,%r11d 760 xorl %ebx,%esi 761 addl %r11d,%edx 762 rorl $2,%r14d 763 addl %esi,%r11d 764 movl %edx,%r13d 765 addl %r11d,%r14d 766 rorl $14,%r13d 767 movl %r14d,%r11d 768 movl %r8d,%r12d 769 xorl %edx,%r13d 770 rorl $9,%r14d 771 xorl %r9d,%r12d 772 rorl $5,%r13d 773 xorl %r11d,%r14d 774 andl %edx,%r12d 775 vpxor %xmm8,%xmm9,%xmm9 776 xorl %edx,%r13d 777 addl 4(%rsp),%r10d 778 movl %r11d,%esi 779 rorl $11,%r14d 780 xorl %r9d,%r12d 781 xorl %eax,%esi 782 rorl $6,%r13d 783 addl %r12d,%r10d 784 andl %esi,%r15d 785 xorl %r11d,%r14d 786 addl %r13d,%r10d 787 xorl %eax,%r15d 788 addl %r10d,%ecx 789 rorl $2,%r14d 790 addl %r15d,%r10d 791 movl %ecx,%r13d 792 addl %r10d,%r14d 793 rorl $14,%r13d 794 movl %r14d,%r10d 795 movl %edx,%r12d 796 xorl %ecx,%r13d 797 rorl $9,%r14d 798 xorl %r8d,%r12d 799 rorl $5,%r13d 800 xorl %r10d,%r14d 801 andl %ecx,%r12d 802 vaesenc %xmm10,%xmm9,%xmm9 803 vmovdqu 32-128(%rdi),%xmm10 804 xorl %ecx,%r13d 805 addl 8(%rsp),%r9d 806 movl %r10d,%r15d 807 rorl $11,%r14d 808 xorl %r8d,%r12d 809 xorl %r11d,%r15d 810 rorl $6,%r13d 811 addl %r12d,%r9d 812 andl %r15d,%esi 813 xorl %r10d,%r14d 814 addl %r13d,%r9d 815 xorl %r11d,%esi 816 addl %r9d,%ebx 817 rorl $2,%r14d 818 addl %esi,%r9d 819 movl %ebx,%r13d 820 addl %r9d,%r14d 821 rorl $14,%r13d 822 movl %r14d,%r9d 823 movl %ecx,%r12d 824 xorl %ebx,%r13d 825 rorl $9,%r14d 826 xorl %edx,%r12d 827 rorl $5,%r13d 828 xorl %r9d,%r14d 829 andl %ebx,%r12d 830 vaesenc %xmm10,%xmm9,%xmm9 831 vmovdqu 48-128(%rdi),%xmm10 832 xorl %ebx,%r13d 833 addl 12(%rsp),%r8d 834 movl %r9d,%esi 835 rorl $11,%r14d 836 xorl %edx,%r12d 837 xorl %r10d,%esi 838 rorl $6,%r13d 839 addl %r12d,%r8d 840 andl %esi,%r15d 841 xorl %r9d,%r14d 842 addl %r13d,%r8d 843 xorl %r10d,%r15d 844 addl %r8d,%eax 845 rorl $2,%r14d 846 addl %r15d,%r8d 847 movl %eax,%r13d 848 addl %r8d,%r14d 849 rorl $14,%r13d 850 movl %r14d,%r8d 851 movl %ebx,%r12d 852 xorl %eax,%r13d 853 rorl $9,%r14d 854 xorl %ecx,%r12d 855 rorl $5,%r13d 856 xorl %r8d,%r14d 857 andl %eax,%r12d 858 vaesenc %xmm10,%xmm9,%xmm9 859 vmovdqu 64-128(%rdi),%xmm10 860 xorl %eax,%r13d 861 addl 16(%rsp),%edx 862 movl %r8d,%r15d 863 rorl $11,%r14d 864 xorl %ecx,%r12d 865 xorl %r9d,%r15d 866 rorl $6,%r13d 867 addl %r12d,%edx 868 andl %r15d,%esi 869 xorl %r8d,%r14d 870 addl %r13d,%edx 871 xorl %r9d,%esi 872 addl %edx,%r11d 873 rorl $2,%r14d 874 addl %esi,%edx 875 movl %r11d,%r13d 876 addl %edx,%r14d 877 rorl $14,%r13d 878 movl %r14d,%edx 879 movl %eax,%r12d 880 xorl %r11d,%r13d 881 rorl $9,%r14d 882 xorl %ebx,%r12d 883 rorl $5,%r13d 884 xorl %edx,%r14d 885 andl %r11d,%r12d 886 vaesenc %xmm10,%xmm9,%xmm9 887 vmovdqu 80-128(%rdi),%xmm10 888 xorl %r11d,%r13d 889 addl 20(%rsp),%ecx 890 movl %edx,%esi 891 rorl $11,%r14d 892 xorl %ebx,%r12d 893 xorl %r8d,%esi 894 rorl $6,%r13d 895 addl %r12d,%ecx 896 andl %esi,%r15d 897 xorl %edx,%r14d 898 addl %r13d,%ecx 899 xorl %r8d,%r15d 900 addl %ecx,%r10d 901 rorl $2,%r14d 902 addl %r15d,%ecx 903 movl %r10d,%r13d 904 addl %ecx,%r14d 905 rorl $14,%r13d 906 movl %r14d,%ecx 907 movl %r11d,%r12d 908 xorl %r10d,%r13d 909 rorl $9,%r14d 910 xorl %eax,%r12d 911 rorl $5,%r13d 912 xorl %ecx,%r14d 913 andl %r10d,%r12d 914 vaesenc %xmm10,%xmm9,%xmm9 915 vmovdqu 96-128(%rdi),%xmm10 916 xorl %r10d,%r13d 917 addl 24(%rsp),%ebx 918 movl %ecx,%r15d 919 rorl $11,%r14d 920 xorl %eax,%r12d 921 xorl %edx,%r15d 922 rorl $6,%r13d 923 addl %r12d,%ebx 924 andl %r15d,%esi 925 xorl %ecx,%r14d 926 addl %r13d,%ebx 927 xorl %edx,%esi 928 addl %ebx,%r9d 929 rorl $2,%r14d 930 addl %esi,%ebx 931 movl %r9d,%r13d 932 addl %ebx,%r14d 933 rorl $14,%r13d 934 movl %r14d,%ebx 935 movl %r10d,%r12d 936 xorl %r9d,%r13d 937 rorl $9,%r14d 938 xorl %r11d,%r12d 939 rorl $5,%r13d 940 xorl %ebx,%r14d 941 andl %r9d,%r12d 942 vaesenc %xmm10,%xmm9,%xmm9 943 vmovdqu 112-128(%rdi),%xmm10 944 xorl %r9d,%r13d 945 addl 28(%rsp),%eax 946 movl %ebx,%esi 947 rorl $11,%r14d 948 xorl %r11d,%r12d 949 xorl %ecx,%esi 950 rorl $6,%r13d 951 addl %r12d,%eax 952 andl %esi,%r15d 953 xorl %ebx,%r14d 954 addl %r13d,%eax 955 xorl %ecx,%r15d 956 addl %eax,%r8d 957 rorl $2,%r14d 958 addl %r15d,%eax 959 movl %r8d,%r13d 960 addl %eax,%r14d 961 rorl $14,%r13d 962 movl %r14d,%eax 963 movl %r9d,%r12d 964 xorl %r8d,%r13d 965 rorl $9,%r14d 966 xorl %r10d,%r12d 967 rorl $5,%r13d 968 xorl %eax,%r14d 969 andl %r8d,%r12d 970 vaesenc %xmm10,%xmm9,%xmm9 971 vmovdqu 128-128(%rdi),%xmm10 972 xorl %r8d,%r13d 973 addl 32(%rsp),%r11d 974 movl %eax,%r15d 975 rorl $11,%r14d 976 xorl %r10d,%r12d 977 xorl %ebx,%r15d 978 rorl $6,%r13d 979 addl %r12d,%r11d 980 andl %r15d,%esi 981 xorl %eax,%r14d 982 addl %r13d,%r11d 983 xorl %ebx,%esi 984 addl %r11d,%edx 985 rorl $2,%r14d 986 addl %esi,%r11d 987 movl %edx,%r13d 988 addl %r11d,%r14d 989 rorl $14,%r13d 990 movl %r14d,%r11d 991 movl %r8d,%r12d 992 xorl %edx,%r13d 993 rorl $9,%r14d 994 xorl %r9d,%r12d 995 rorl $5,%r13d 996 xorl %r11d,%r14d 997 andl %edx,%r12d 998 vaesenc %xmm10,%xmm9,%xmm9 999 vmovdqu 144-128(%rdi),%xmm10 1000 xorl %edx,%r13d 1001 addl 36(%rsp),%r10d 1002 movl %r11d,%esi 1003 rorl $11,%r14d 1004 xorl %r9d,%r12d 1005 xorl %eax,%esi 1006 rorl $6,%r13d 1007 addl %r12d,%r10d 1008 andl %esi,%r15d 1009 xorl %r11d,%r14d 1010 addl %r13d,%r10d 1011 xorl %eax,%r15d 1012 addl %r10d,%ecx 1013 rorl $2,%r14d 1014 addl %r15d,%r10d 1015 movl %ecx,%r13d 1016 addl %r10d,%r14d 1017 rorl $14,%r13d 1018 movl %r14d,%r10d 1019 movl %edx,%r12d 1020 xorl %ecx,%r13d 1021 rorl $9,%r14d 1022 xorl %r8d,%r12d 1023 rorl $5,%r13d 1024 xorl %r10d,%r14d 1025 andl %ecx,%r12d 1026 vaesenc %xmm10,%xmm9,%xmm9 1027 vmovdqu 160-128(%rdi),%xmm10 1028 xorl %ecx,%r13d 1029 addl 40(%rsp),%r9d 1030 movl %r10d,%r15d 1031 rorl $11,%r14d 1032 xorl %r8d,%r12d 1033 xorl %r11d,%r15d 1034 rorl $6,%r13d 1035 addl %r12d,%r9d 1036 andl %r15d,%esi 1037 xorl %r10d,%r14d 1038 addl %r13d,%r9d 1039 xorl %r11d,%esi 1040 addl %r9d,%ebx 1041 rorl $2,%r14d 1042 addl %esi,%r9d 1043 movl %ebx,%r13d 1044 addl %r9d,%r14d 1045 rorl $14,%r13d 1046 movl %r14d,%r9d 1047 movl %ecx,%r12d 1048 xorl %ebx,%r13d 1049 rorl $9,%r14d 1050 xorl %edx,%r12d 1051 rorl $5,%r13d 1052 xorl %r9d,%r14d 1053 andl %ebx,%r12d 1054 vaesenclast %xmm10,%xmm9,%xmm11 1055 vaesenc %xmm10,%xmm9,%xmm9 1056 vmovdqu 176-128(%rdi),%xmm10 1057 xorl %ebx,%r13d 1058 addl 44(%rsp),%r8d 1059 movl %r9d,%esi 1060 rorl $11,%r14d 1061 xorl %edx,%r12d 1062 xorl %r10d,%esi 1063 rorl $6,%r13d 1064 addl %r12d,%r8d 1065 andl %esi,%r15d 1066 xorl %r9d,%r14d 1067 addl %r13d,%r8d 1068 xorl %r10d,%r15d 1069 addl %r8d,%eax 1070 rorl $2,%r14d 1071 addl %r15d,%r8d 1072 movl %eax,%r13d 1073 addl %r8d,%r14d 1074 rorl $14,%r13d 1075 movl %r14d,%r8d 1076 movl %ebx,%r12d 1077 xorl %eax,%r13d 1078 rorl $9,%r14d 1079 xorl %ecx,%r12d 1080 rorl $5,%r13d 1081 xorl %r8d,%r14d 1082 andl %eax,%r12d 1083 vpand %xmm12,%xmm11,%xmm8 1084 vaesenc %xmm10,%xmm9,%xmm9 1085 vmovdqu 192-128(%rdi),%xmm10 1086 xorl %eax,%r13d 1087 addl 48(%rsp),%edx 1088 movl %r8d,%r15d 1089 rorl $11,%r14d 1090 xorl %ecx,%r12d 1091 xorl %r9d,%r15d 1092 rorl $6,%r13d 1093 addl %r12d,%edx 1094 andl %r15d,%esi 1095 xorl %r8d,%r14d 1096 addl %r13d,%edx 1097 xorl %r9d,%esi 1098 addl %edx,%r11d 1099 rorl $2,%r14d 1100 addl %esi,%edx 1101 movl %r11d,%r13d 1102 addl %edx,%r14d 1103 rorl $14,%r13d 1104 movl %r14d,%edx 1105 movl %eax,%r12d 1106 xorl %r11d,%r13d 1107 rorl $9,%r14d 1108 xorl %ebx,%r12d 1109 rorl $5,%r13d 1110 xorl %edx,%r14d 1111 andl %r11d,%r12d 1112 vaesenclast %xmm10,%xmm9,%xmm11 1113 vaesenc %xmm10,%xmm9,%xmm9 1114 vmovdqu 208-128(%rdi),%xmm10 1115 xorl %r11d,%r13d 1116 addl 52(%rsp),%ecx 1117 movl %edx,%esi 1118 rorl $11,%r14d 1119 xorl %ebx,%r12d 1120 xorl %r8d,%esi 1121 rorl $6,%r13d 1122 addl %r12d,%ecx 1123 andl %esi,%r15d 1124 xorl %edx,%r14d 1125 addl %r13d,%ecx 1126 xorl %r8d,%r15d 1127 addl %ecx,%r10d 1128 rorl $2,%r14d 1129 addl %r15d,%ecx 1130 movl %r10d,%r13d 1131 addl %ecx,%r14d 1132 rorl $14,%r13d 1133 movl %r14d,%ecx 1134 movl %r11d,%r12d 1135 xorl %r10d,%r13d 1136 rorl $9,%r14d 1137 xorl %eax,%r12d 1138 rorl $5,%r13d 1139 xorl %ecx,%r14d 1140 andl %r10d,%r12d 1141 vpand %xmm13,%xmm11,%xmm11 1142 vaesenc %xmm10,%xmm9,%xmm9 1143 vmovdqu 224-128(%rdi),%xmm10 1144 xorl %r10d,%r13d 1145 addl 56(%rsp),%ebx 1146 movl %ecx,%r15d 1147 rorl $11,%r14d 1148 xorl %eax,%r12d 1149 xorl %edx,%r15d 1150 rorl $6,%r13d 1151 addl %r12d,%ebx 1152 andl %r15d,%esi 1153 xorl %ecx,%r14d 1154 addl %r13d,%ebx 1155 xorl %edx,%esi 1156 addl %ebx,%r9d 1157 rorl $2,%r14d 1158 addl %esi,%ebx 1159 movl %r9d,%r13d 1160 addl %ebx,%r14d 1161 rorl $14,%r13d 1162 movl %r14d,%ebx 1163 movl %r10d,%r12d 1164 xorl %r9d,%r13d 1165 rorl $9,%r14d 1166 xorl %r11d,%r12d 1167 rorl $5,%r13d 1168 xorl %ebx,%r14d 1169 andl %r9d,%r12d 1170 vpor %xmm11,%xmm8,%xmm8 1171 vaesenclast %xmm10,%xmm9,%xmm11 1172 vmovdqu 0-128(%rdi),%xmm10 1173 xorl %r9d,%r13d 1174 addl 60(%rsp),%eax 1175 movl %ebx,%esi 1176 rorl $11,%r14d 1177 xorl %r11d,%r12d 1178 xorl %ecx,%esi 1179 rorl $6,%r13d 1180 addl %r12d,%eax 1181 andl %esi,%r15d 1182 xorl %ebx,%r14d 1183 addl %r13d,%eax 1184 xorl %ecx,%r15d 1185 addl %eax,%r8d 1186 rorl $2,%r14d 1187 addl %r15d,%eax 1188 movl %r8d,%r13d 1189 addl %eax,%r14d 1190 movq 64+0(%rsp),%r12 1191 movq 64+8(%rsp),%r13 1192 movq 64+40(%rsp),%r15 1193 movq 64+48(%rsp),%rsi 1194 1195 vpand %xmm14,%xmm11,%xmm11 1196 movl %r14d,%eax 1197 vpor %xmm11,%xmm8,%xmm8 1198 vmovdqu %xmm8,(%r12,%r13,1) 1199 leaq 16(%r12),%r12 1200 1201 addl 0(%r15),%eax 1202 addl 4(%r15),%ebx 1203 addl 8(%r15),%ecx 1204 addl 12(%r15),%edx 1205 addl 16(%r15),%r8d 1206 addl 20(%r15),%r9d 1207 addl 24(%r15),%r10d 1208 addl 28(%r15),%r11d 1209 1210 cmpq 64+16(%rsp),%r12 1211 1212 movl %eax,0(%r15) 1213 movl %ebx,4(%r15) 1214 movl %ecx,8(%r15) 1215 movl %edx,12(%r15) 1216 movl %r8d,16(%r15) 1217 movl %r9d,20(%r15) 1218 movl %r10d,24(%r15) 1219 movl %r11d,28(%r15) 1220 1221 jb .Lloop_xop 1222 1223 movq 64+32(%rsp),%r8 1224 movq 120(%rsp),%rsi 1225 .cfi_def_cfa %rsi,8 1226 vmovdqu %xmm8,(%r8) 1227 vzeroall 1228 movq -48(%rsi),%r15 1229 .cfi_restore %r15 1230 movq -40(%rsi),%r14 1231 .cfi_restore %r14 1232 movq -32(%rsi),%r13 1233 .cfi_restore %r13 1234 movq -24(%rsi),%r12 1235 .cfi_restore %r12 1236 movq -16(%rsi),%rbp 1237 .cfi_restore %rbp 1238 movq -8(%rsi),%rbx 1239 .cfi_restore %rbx 1240 leaq (%rsi),%rsp 1241 .cfi_def_cfa_register %rsp 1242 .Lepilogue_xop: 1243 .byte 0xf3,0xc3 1244 .cfi_endproc 1245 .size aesni_cbc_sha256_enc_xop,.-aesni_cbc_sha256_enc_xop 1246 .type aesni_cbc_sha256_enc_avx,@function 1247 .align 64 1248 aesni_cbc_sha256_enc_avx: 1249 .cfi_startproc 1250 .Lavx_shortcut: 1251 movq 8(%rsp),%r10 1252 movq %rsp,%rax 1253 .cfi_def_cfa_register %rax 1254 pushq %rbx 1255 .cfi_offset %rbx,-16 1256 pushq %rbp 1257 .cfi_offset %rbp,-24 1258 pushq %r12 1259 .cfi_offset %r12,-32 1260 pushq %r13 1261 .cfi_offset %r13,-40 1262 pushq %r14 1263 .cfi_offset %r14,-48 1264 pushq %r15 1265 .cfi_offset %r15,-56 1266 subq $128,%rsp 1267 andq $-64,%rsp 1268 1269 shlq $6,%rdx 1270 subq %rdi,%rsi 1271 subq %rdi,%r10 1272 addq %rdi,%rdx 1273 1274 1275 movq %rsi,64+8(%rsp) 1276 movq %rdx,64+16(%rsp) 1277 1278 movq %r8,64+32(%rsp) 1279 movq %r9,64+40(%rsp) 1280 movq %r10,64+48(%rsp) 1281 movq %rax,120(%rsp) 1282 .cfi_escape 0x0f,0x06,0x77,0xf8,0x00,0x06,0x23,0x08 1283 .Lprologue_avx: 1284 vzeroall 1285 1286 movq %rdi,%r12 1287 leaq 128(%rcx),%rdi 1288 leaq K256+544(%rip),%r13 1289 movl 240-128(%rdi),%r14d 1290 movq %r9,%r15 1291 movq %r10,%rsi 1292 vmovdqu (%r8),%xmm8 1293 subq $9,%r14 1294 1295 movl 0(%r15),%eax 1296 movl 4(%r15),%ebx 1297 movl 8(%r15),%ecx 1298 movl 12(%r15),%edx 1299 movl 16(%r15),%r8d 1300 movl 20(%r15),%r9d 1301 movl 24(%r15),%r10d 1302 movl 28(%r15),%r11d 1303 1304 vmovdqa 0(%r13,%r14,8),%xmm14 1305 vmovdqa 16(%r13,%r14,8),%xmm13 1306 vmovdqa 32(%r13,%r14,8),%xmm12 1307 vmovdqu 0-128(%rdi),%xmm10 1308 jmp .Lloop_avx 1309 .align 16 1310 .Lloop_avx: 1311 vmovdqa K256+512(%rip),%xmm7 1312 vmovdqu 0(%rsi,%r12,1),%xmm0 1313 vmovdqu 16(%rsi,%r12,1),%xmm1 1314 vmovdqu 32(%rsi,%r12,1),%xmm2 1315 vmovdqu 48(%rsi,%r12,1),%xmm3 1316 vpshufb %xmm7,%xmm0,%xmm0 1317 leaq K256(%rip),%rbp 1318 vpshufb %xmm7,%xmm1,%xmm1 1319 vpshufb %xmm7,%xmm2,%xmm2 1320 vpaddd 0(%rbp),%xmm0,%xmm4 1321 vpshufb %xmm7,%xmm3,%xmm3 1322 vpaddd 32(%rbp),%xmm1,%xmm5 1323 vpaddd 64(%rbp),%xmm2,%xmm6 1324 vpaddd 96(%rbp),%xmm3,%xmm7 1325 vmovdqa %xmm4,0(%rsp) 1326 movl %eax,%r14d 1327 vmovdqa %xmm5,16(%rsp) 1328 movl %ebx,%esi 1329 vmovdqa %xmm6,32(%rsp) 1330 xorl %ecx,%esi 1331 vmovdqa %xmm7,48(%rsp) 1332 movl %r8d,%r13d 1333 jmp .Lavx_00_47 1334 1335 .align 16 1336 .Lavx_00_47: 1337 subq $-32*4,%rbp 1338 vmovdqu (%r12),%xmm9 1339 movq %r12,64+0(%rsp) 1340 vpalignr $4,%xmm0,%xmm1,%xmm4 1341 shrdl $14,%r13d,%r13d 1342 movl %r14d,%eax 1343 movl %r9d,%r12d 1344 vpalignr $4,%xmm2,%xmm3,%xmm7 1345 xorl %r8d,%r13d 1346 shrdl $9,%r14d,%r14d 1347 xorl %r10d,%r12d 1348 vpsrld $7,%xmm4,%xmm6 1349 shrdl $5,%r13d,%r13d 1350 xorl %eax,%r14d 1351 andl %r8d,%r12d 1352 vpaddd %xmm7,%xmm0,%xmm0 1353 vpxor %xmm10,%xmm9,%xmm9 1354 vmovdqu 16-128(%rdi),%xmm10 1355 xorl %r8d,%r13d 1356 addl 0(%rsp),%r11d 1357 movl %eax,%r15d 1358 vpsrld $3,%xmm4,%xmm7 1359 shrdl $11,%r14d,%r14d 1360 xorl %r10d,%r12d 1361 xorl %ebx,%r15d 1362 vpslld $14,%xmm4,%xmm5 1363 shrdl $6,%r13d,%r13d 1364 addl %r12d,%r11d 1365 andl %r15d,%esi 1366 vpxor %xmm6,%xmm7,%xmm4 1367 xorl %eax,%r14d 1368 addl %r13d,%r11d 1369 xorl %ebx,%esi 1370 vpshufd $250,%xmm3,%xmm7 1371 addl %r11d,%edx 1372 shrdl $2,%r14d,%r14d 1373 addl %esi,%r11d 1374 vpsrld $11,%xmm6,%xmm6 1375 movl %edx,%r13d 1376 addl %r11d,%r14d 1377 shrdl $14,%r13d,%r13d 1378 vpxor %xmm5,%xmm4,%xmm4 1379 movl %r14d,%r11d 1380 movl %r8d,%r12d 1381 xorl %edx,%r13d 1382 vpslld $11,%xmm5,%xmm5 1383 shrdl $9,%r14d,%r14d 1384 xorl %r9d,%r12d 1385 shrdl $5,%r13d,%r13d 1386 vpxor %xmm6,%xmm4,%xmm4 1387 xorl %r11d,%r14d 1388 andl %edx,%r12d 1389 vpxor %xmm8,%xmm9,%xmm9 1390 xorl %edx,%r13d 1391 vpsrld $10,%xmm7,%xmm6 1392 addl 4(%rsp),%r10d 1393 movl %r11d,%esi 1394 shrdl $11,%r14d,%r14d 1395 vpxor %xmm5,%xmm4,%xmm4 1396 xorl %r9d,%r12d 1397 xorl %eax,%esi 1398 shrdl $6,%r13d,%r13d 1399 vpsrlq $17,%xmm7,%xmm7 1400 addl %r12d,%r10d 1401 andl %esi,%r15d 1402 xorl %r11d,%r14d 1403 vpaddd %xmm4,%xmm0,%xmm0 1404 addl %r13d,%r10d 1405 xorl %eax,%r15d 1406 addl %r10d,%ecx 1407 vpxor %xmm7,%xmm6,%xmm6 1408 shrdl $2,%r14d,%r14d 1409 addl %r15d,%r10d 1410 movl %ecx,%r13d 1411 vpsrlq $2,%xmm7,%xmm7 1412 addl %r10d,%r14d 1413 shrdl $14,%r13d,%r13d 1414 movl %r14d,%r10d 1415 vpxor %xmm7,%xmm6,%xmm6 1416 movl %edx,%r12d 1417 xorl %ecx,%r13d 1418 shrdl $9,%r14d,%r14d 1419 vpshufd $132,%xmm6,%xmm6 1420 xorl %r8d,%r12d 1421 shrdl $5,%r13d,%r13d 1422 xorl %r10d,%r14d 1423 vpsrldq $8,%xmm6,%xmm6 1424 andl %ecx,%r12d 1425 vaesenc %xmm10,%xmm9,%xmm9 1426 vmovdqu 32-128(%rdi),%xmm10 1427 xorl %ecx,%r13d 1428 addl 8(%rsp),%r9d 1429 vpaddd %xmm6,%xmm0,%xmm0 1430 movl %r10d,%r15d 1431 shrdl $11,%r14d,%r14d 1432 xorl %r8d,%r12d 1433 vpshufd $80,%xmm0,%xmm7 1434 xorl %r11d,%r15d 1435 shrdl $6,%r13d,%r13d 1436 addl %r12d,%r9d 1437 vpsrld $10,%xmm7,%xmm6 1438 andl %r15d,%esi 1439 xorl %r10d,%r14d 1440 addl %r13d,%r9d 1441 vpsrlq $17,%xmm7,%xmm7 1442 xorl %r11d,%esi 1443 addl %r9d,%ebx 1444 shrdl $2,%r14d,%r14d 1445 vpxor %xmm7,%xmm6,%xmm6 1446 addl %esi,%r9d 1447 movl %ebx,%r13d 1448 addl %r9d,%r14d 1449 vpsrlq $2,%xmm7,%xmm7 1450 shrdl $14,%r13d,%r13d 1451 movl %r14d,%r9d 1452 movl %ecx,%r12d 1453 vpxor %xmm7,%xmm6,%xmm6 1454 xorl %ebx,%r13d 1455 shrdl $9,%r14d,%r14d 1456 xorl %edx,%r12d 1457 vpshufd $232,%xmm6,%xmm6 1458 shrdl $5,%r13d,%r13d 1459 xorl %r9d,%r14d 1460 andl %ebx,%r12d 1461 vpslldq $8,%xmm6,%xmm6 1462 vaesenc %xmm10,%xmm9,%xmm9 1463 vmovdqu 48-128(%rdi),%xmm10 1464 xorl %ebx,%r13d 1465 addl 12(%rsp),%r8d 1466 movl %r9d,%esi 1467 vpaddd %xmm6,%xmm0,%xmm0 1468 shrdl $11,%r14d,%r14d 1469 xorl %edx,%r12d 1470 xorl %r10d,%esi 1471 vpaddd 0(%rbp),%xmm0,%xmm6 1472 shrdl $6,%r13d,%r13d 1473 addl %r12d,%r8d 1474 andl %esi,%r15d 1475 xorl %r9d,%r14d 1476 addl %r13d,%r8d 1477 xorl %r10d,%r15d 1478 addl %r8d,%eax 1479 shrdl $2,%r14d,%r14d 1480 addl %r15d,%r8d 1481 movl %eax,%r13d 1482 addl %r8d,%r14d 1483 vmovdqa %xmm6,0(%rsp) 1484 vpalignr $4,%xmm1,%xmm2,%xmm4 1485 shrdl $14,%r13d,%r13d 1486 movl %r14d,%r8d 1487 movl %ebx,%r12d 1488 vpalignr $4,%xmm3,%xmm0,%xmm7 1489 xorl %eax,%r13d 1490 shrdl $9,%r14d,%r14d 1491 xorl %ecx,%r12d 1492 vpsrld $7,%xmm4,%xmm6 1493 shrdl $5,%r13d,%r13d 1494 xorl %r8d,%r14d 1495 andl %eax,%r12d 1496 vpaddd %xmm7,%xmm1,%xmm1 1497 vaesenc %xmm10,%xmm9,%xmm9 1498 vmovdqu 64-128(%rdi),%xmm10 1499 xorl %eax,%r13d 1500 addl 16(%rsp),%edx 1501 movl %r8d,%r15d 1502 vpsrld $3,%xmm4,%xmm7 1503 shrdl $11,%r14d,%r14d 1504 xorl %ecx,%r12d 1505 xorl %r9d,%r15d 1506 vpslld $14,%xmm4,%xmm5 1507 shrdl $6,%r13d,%r13d 1508 addl %r12d,%edx 1509 andl %r15d,%esi 1510 vpxor %xmm6,%xmm7,%xmm4 1511 xorl %r8d,%r14d 1512 addl %r13d,%edx 1513 xorl %r9d,%esi 1514 vpshufd $250,%xmm0,%xmm7 1515 addl %edx,%r11d 1516 shrdl $2,%r14d,%r14d 1517 addl %esi,%edx 1518 vpsrld $11,%xmm6,%xmm6 1519 movl %r11d,%r13d 1520 addl %edx,%r14d 1521 shrdl $14,%r13d,%r13d 1522 vpxor %xmm5,%xmm4,%xmm4 1523 movl %r14d,%edx 1524 movl %eax,%r12d 1525 xorl %r11d,%r13d 1526 vpslld $11,%xmm5,%xmm5 1527 shrdl $9,%r14d,%r14d 1528 xorl %ebx,%r12d 1529 shrdl $5,%r13d,%r13d 1530 vpxor %xmm6,%xmm4,%xmm4 1531 xorl %edx,%r14d 1532 andl %r11d,%r12d 1533 vaesenc %xmm10,%xmm9,%xmm9 1534 vmovdqu 80-128(%rdi),%xmm10 1535 xorl %r11d,%r13d 1536 vpsrld $10,%xmm7,%xmm6 1537 addl 20(%rsp),%ecx 1538 movl %edx,%esi 1539 shrdl $11,%r14d,%r14d 1540 vpxor %xmm5,%xmm4,%xmm4 1541 xorl %ebx,%r12d 1542 xorl %r8d,%esi 1543 shrdl $6,%r13d,%r13d 1544 vpsrlq $17,%xmm7,%xmm7 1545 addl %r12d,%ecx 1546 andl %esi,%r15d 1547 xorl %edx,%r14d 1548 vpaddd %xmm4,%xmm1,%xmm1 1549 addl %r13d,%ecx 1550 xorl %r8d,%r15d 1551 addl %ecx,%r10d 1552 vpxor %xmm7,%xmm6,%xmm6 1553 shrdl $2,%r14d,%r14d 1554 addl %r15d,%ecx 1555 movl %r10d,%r13d 1556 vpsrlq $2,%xmm7,%xmm7 1557 addl %ecx,%r14d 1558 shrdl $14,%r13d,%r13d 1559 movl %r14d,%ecx 1560 vpxor %xmm7,%xmm6,%xmm6 1561 movl %r11d,%r12d 1562 xorl %r10d,%r13d 1563 shrdl $9,%r14d,%r14d 1564 vpshufd $132,%xmm6,%xmm6 1565 xorl %eax,%r12d 1566 shrdl $5,%r13d,%r13d 1567 xorl %ecx,%r14d 1568 vpsrldq $8,%xmm6,%xmm6 1569 andl %r10d,%r12d 1570 vaesenc %xmm10,%xmm9,%xmm9 1571 vmovdqu 96-128(%rdi),%xmm10 1572 xorl %r10d,%r13d 1573 addl 24(%rsp),%ebx 1574 vpaddd %xmm6,%xmm1,%xmm1 1575 movl %ecx,%r15d 1576 shrdl $11,%r14d,%r14d 1577 xorl %eax,%r12d 1578 vpshufd $80,%xmm1,%xmm7 1579 xorl %edx,%r15d 1580 shrdl $6,%r13d,%r13d 1581 addl %r12d,%ebx 1582 vpsrld $10,%xmm7,%xmm6 1583 andl %r15d,%esi 1584 xorl %ecx,%r14d 1585 addl %r13d,%ebx 1586 vpsrlq $17,%xmm7,%xmm7 1587 xorl %edx,%esi 1588 addl %ebx,%r9d 1589 shrdl $2,%r14d,%r14d 1590 vpxor %xmm7,%xmm6,%xmm6 1591 addl %esi,%ebx 1592 movl %r9d,%r13d 1593 addl %ebx,%r14d 1594 vpsrlq $2,%xmm7,%xmm7 1595 shrdl $14,%r13d,%r13d 1596 movl %r14d,%ebx 1597 movl %r10d,%r12d 1598 vpxor %xmm7,%xmm6,%xmm6 1599 xorl %r9d,%r13d 1600 shrdl $9,%r14d,%r14d 1601 xorl %r11d,%r12d 1602 vpshufd $232,%xmm6,%xmm6 1603 shrdl $5,%r13d,%r13d 1604 xorl %ebx,%r14d 1605 andl %r9d,%r12d 1606 vpslldq $8,%xmm6,%xmm6 1607 vaesenc %xmm10,%xmm9,%xmm9 1608 vmovdqu 112-128(%rdi),%xmm10 1609 xorl %r9d,%r13d 1610 addl 28(%rsp),%eax 1611 movl %ebx,%esi 1612 vpaddd %xmm6,%xmm1,%xmm1 1613 shrdl $11,%r14d,%r14d 1614 xorl %r11d,%r12d 1615 xorl %ecx,%esi 1616 vpaddd 32(%rbp),%xmm1,%xmm6 1617 shrdl $6,%r13d,%r13d 1618 addl %r12d,%eax 1619 andl %esi,%r15d 1620 xorl %ebx,%r14d 1621 addl %r13d,%eax 1622 xorl %ecx,%r15d 1623 addl %eax,%r8d 1624 shrdl $2,%r14d,%r14d 1625 addl %r15d,%eax 1626 movl %r8d,%r13d 1627 addl %eax,%r14d 1628 vmovdqa %xmm6,16(%rsp) 1629 vpalignr $4,%xmm2,%xmm3,%xmm4 1630 shrdl $14,%r13d,%r13d 1631 movl %r14d,%eax 1632 movl %r9d,%r12d 1633 vpalignr $4,%xmm0,%xmm1,%xmm7 1634 xorl %r8d,%r13d 1635 shrdl $9,%r14d,%r14d 1636 xorl %r10d,%r12d 1637 vpsrld $7,%xmm4,%xmm6 1638 shrdl $5,%r13d,%r13d 1639 xorl %eax,%r14d 1640 andl %r8d,%r12d 1641 vpaddd %xmm7,%xmm2,%xmm2 1642 vaesenc %xmm10,%xmm9,%xmm9 1643 vmovdqu 128-128(%rdi),%xmm10 1644 xorl %r8d,%r13d 1645 addl 32(%rsp),%r11d 1646 movl %eax,%r15d 1647 vpsrld $3,%xmm4,%xmm7 1648 shrdl $11,%r14d,%r14d 1649 xorl %r10d,%r12d 1650 xorl %ebx,%r15d 1651 vpslld $14,%xmm4,%xmm5 1652 shrdl $6,%r13d,%r13d 1653 addl %r12d,%r11d 1654 andl %r15d,%esi 1655 vpxor %xmm6,%xmm7,%xmm4 1656 xorl %eax,%r14d 1657 addl %r13d,%r11d 1658 xorl %ebx,%esi 1659 vpshufd $250,%xmm1,%xmm7 1660 addl %r11d,%edx 1661 shrdl $2,%r14d,%r14d 1662 addl %esi,%r11d 1663 vpsrld $11,%xmm6,%xmm6 1664 movl %edx,%r13d 1665 addl %r11d,%r14d 1666 shrdl $14,%r13d,%r13d 1667 vpxor %xmm5,%xmm4,%xmm4 1668 movl %r14d,%r11d 1669 movl %r8d,%r12d 1670 xorl %edx,%r13d 1671 vpslld $11,%xmm5,%xmm5 1672 shrdl $9,%r14d,%r14d 1673 xorl %r9d,%r12d 1674 shrdl $5,%r13d,%r13d 1675 vpxor %xmm6,%xmm4,%xmm4 1676 xorl %r11d,%r14d 1677 andl %edx,%r12d 1678 vaesenc %xmm10,%xmm9,%xmm9 1679 vmovdqu 144-128(%rdi),%xmm10 1680 xorl %edx,%r13d 1681 vpsrld $10,%xmm7,%xmm6 1682 addl 36(%rsp),%r10d 1683 movl %r11d,%esi 1684 shrdl $11,%r14d,%r14d 1685 vpxor %xmm5,%xmm4,%xmm4 1686 xorl %r9d,%r12d 1687 xorl %eax,%esi 1688 shrdl $6,%r13d,%r13d 1689 vpsrlq $17,%xmm7,%xmm7 1690 addl %r12d,%r10d 1691 andl %esi,%r15d 1692 xorl %r11d,%r14d 1693 vpaddd %xmm4,%xmm2,%xmm2 1694 addl %r13d,%r10d 1695 xorl %eax,%r15d 1696 addl %r10d,%ecx 1697 vpxor %xmm7,%xmm6,%xmm6 1698 shrdl $2,%r14d,%r14d 1699 addl %r15d,%r10d 1700 movl %ecx,%r13d 1701 vpsrlq $2,%xmm7,%xmm7 1702 addl %r10d,%r14d 1703 shrdl $14,%r13d,%r13d 1704 movl %r14d,%r10d 1705 vpxor %xmm7,%xmm6,%xmm6 1706 movl %edx,%r12d 1707 xorl %ecx,%r13d 1708 shrdl $9,%r14d,%r14d 1709 vpshufd $132,%xmm6,%xmm6 1710 xorl %r8d,%r12d 1711 shrdl $5,%r13d,%r13d 1712 xorl %r10d,%r14d 1713 vpsrldq $8,%xmm6,%xmm6 1714 andl %ecx,%r12d 1715 vaesenc %xmm10,%xmm9,%xmm9 1716 vmovdqu 160-128(%rdi),%xmm10 1717 xorl %ecx,%r13d 1718 addl 40(%rsp),%r9d 1719 vpaddd %xmm6,%xmm2,%xmm2 1720 movl %r10d,%r15d 1721 shrdl $11,%r14d,%r14d 1722 xorl %r8d,%r12d 1723 vpshufd $80,%xmm2,%xmm7 1724 xorl %r11d,%r15d 1725 shrdl $6,%r13d,%r13d 1726 addl %r12d,%r9d 1727 vpsrld $10,%xmm7,%xmm6 1728 andl %r15d,%esi 1729 xorl %r10d,%r14d 1730 addl %r13d,%r9d 1731 vpsrlq $17,%xmm7,%xmm7 1732 xorl %r11d,%esi 1733 addl %r9d,%ebx 1734 shrdl $2,%r14d,%r14d 1735 vpxor %xmm7,%xmm6,%xmm6 1736 addl %esi,%r9d 1737 movl %ebx,%r13d 1738 addl %r9d,%r14d 1739 vpsrlq $2,%xmm7,%xmm7 1740 shrdl $14,%r13d,%r13d 1741 movl %r14d,%r9d 1742 movl %ecx,%r12d 1743 vpxor %xmm7,%xmm6,%xmm6 1744 xorl %ebx,%r13d 1745 shrdl $9,%r14d,%r14d 1746 xorl %edx,%r12d 1747 vpshufd $232,%xmm6,%xmm6 1748 shrdl $5,%r13d,%r13d 1749 xorl %r9d,%r14d 1750 andl %ebx,%r12d 1751 vpslldq $8,%xmm6,%xmm6 1752 vaesenclast %xmm10,%xmm9,%xmm11 1753 vaesenc %xmm10,%xmm9,%xmm9 1754 vmovdqu 176-128(%rdi),%xmm10 1755 xorl %ebx,%r13d 1756 addl 44(%rsp),%r8d 1757 movl %r9d,%esi 1758 vpaddd %xmm6,%xmm2,%xmm2 1759 shrdl $11,%r14d,%r14d 1760 xorl %edx,%r12d 1761 xorl %r10d,%esi 1762 vpaddd 64(%rbp),%xmm2,%xmm6 1763 shrdl $6,%r13d,%r13d 1764 addl %r12d,%r8d 1765 andl %esi,%r15d 1766 xorl %r9d,%r14d 1767 addl %r13d,%r8d 1768 xorl %r10d,%r15d 1769 addl %r8d,%eax 1770 shrdl $2,%r14d,%r14d 1771 addl %r15d,%r8d 1772 movl %eax,%r13d 1773 addl %r8d,%r14d 1774 vmovdqa %xmm6,32(%rsp) 1775 vpalignr $4,%xmm3,%xmm0,%xmm4 1776 shrdl $14,%r13d,%r13d 1777 movl %r14d,%r8d 1778 movl %ebx,%r12d 1779 vpalignr $4,%xmm1,%xmm2,%xmm7 1780 xorl %eax,%r13d 1781 shrdl $9,%r14d,%r14d 1782 xorl %ecx,%r12d 1783 vpsrld $7,%xmm4,%xmm6 1784 shrdl $5,%r13d,%r13d 1785 xorl %r8d,%r14d 1786 andl %eax,%r12d 1787 vpaddd %xmm7,%xmm3,%xmm3 1788 vpand %xmm12,%xmm11,%xmm8 1789 vaesenc %xmm10,%xmm9,%xmm9 1790 vmovdqu 192-128(%rdi),%xmm10 1791 xorl %eax,%r13d 1792 addl 48(%rsp),%edx 1793 movl %r8d,%r15d 1794 vpsrld $3,%xmm4,%xmm7 1795 shrdl $11,%r14d,%r14d 1796 xorl %ecx,%r12d 1797 xorl %r9d,%r15d 1798 vpslld $14,%xmm4,%xmm5 1799 shrdl $6,%r13d,%r13d 1800 addl %r12d,%edx 1801 andl %r15d,%esi 1802 vpxor %xmm6,%xmm7,%xmm4 1803 xorl %r8d,%r14d 1804 addl %r13d,%edx 1805 xorl %r9d,%esi 1806 vpshufd $250,%xmm2,%xmm7 1807 addl %edx,%r11d 1808 shrdl $2,%r14d,%r14d 1809 addl %esi,%edx 1810 vpsrld $11,%xmm6,%xmm6 1811 movl %r11d,%r13d 1812 addl %edx,%r14d 1813 shrdl $14,%r13d,%r13d 1814 vpxor %xmm5,%xmm4,%xmm4 1815 movl %r14d,%edx 1816 movl %eax,%r12d 1817 xorl %r11d,%r13d 1818 vpslld $11,%xmm5,%xmm5 1819 shrdl $9,%r14d,%r14d 1820 xorl %ebx,%r12d 1821 shrdl $5,%r13d,%r13d 1822 vpxor %xmm6,%xmm4,%xmm4 1823 xorl %edx,%r14d 1824 andl %r11d,%r12d 1825 vaesenclast %xmm10,%xmm9,%xmm11 1826 vaesenc %xmm10,%xmm9,%xmm9 1827 vmovdqu 208-128(%rdi),%xmm10 1828 xorl %r11d,%r13d 1829 vpsrld $10,%xmm7,%xmm6 1830 addl 52(%rsp),%ecx 1831 movl %edx,%esi 1832 shrdl $11,%r14d,%r14d 1833 vpxor %xmm5,%xmm4,%xmm4 1834 xorl %ebx,%r12d 1835 xorl %r8d,%esi 1836 shrdl $6,%r13d,%r13d 1837 vpsrlq $17,%xmm7,%xmm7 1838 addl %r12d,%ecx 1839 andl %esi,%r15d 1840 xorl %edx,%r14d 1841 vpaddd %xmm4,%xmm3,%xmm3 1842 addl %r13d,%ecx 1843 xorl %r8d,%r15d 1844 addl %ecx,%r10d 1845 vpxor %xmm7,%xmm6,%xmm6 1846 shrdl $2,%r14d,%r14d 1847 addl %r15d,%ecx 1848 movl %r10d,%r13d 1849 vpsrlq $2,%xmm7,%xmm7 1850 addl %ecx,%r14d 1851 shrdl $14,%r13d,%r13d 1852 movl %r14d,%ecx 1853 vpxor %xmm7,%xmm6,%xmm6 1854 movl %r11d,%r12d 1855 xorl %r10d,%r13d 1856 shrdl $9,%r14d,%r14d 1857 vpshufd $132,%xmm6,%xmm6 1858 xorl %eax,%r12d 1859 shrdl $5,%r13d,%r13d 1860 xorl %ecx,%r14d 1861 vpsrldq $8,%xmm6,%xmm6 1862 andl %r10d,%r12d 1863 vpand %xmm13,%xmm11,%xmm11 1864 vaesenc %xmm10,%xmm9,%xmm9 1865 vmovdqu 224-128(%rdi),%xmm10 1866 xorl %r10d,%r13d 1867 addl 56(%rsp),%ebx 1868 vpaddd %xmm6,%xmm3,%xmm3 1869 movl %ecx,%r15d 1870 shrdl $11,%r14d,%r14d 1871 xorl %eax,%r12d 1872 vpshufd $80,%xmm3,%xmm7 1873 xorl %edx,%r15d 1874 shrdl $6,%r13d,%r13d 1875 addl %r12d,%ebx 1876 vpsrld $10,%xmm7,%xmm6 1877 andl %r15d,%esi 1878 xorl %ecx,%r14d 1879 addl %r13d,%ebx 1880 vpsrlq $17,%xmm7,%xmm7 1881 xorl %edx,%esi 1882 addl %ebx,%r9d 1883 shrdl $2,%r14d,%r14d 1884 vpxor %xmm7,%xmm6,%xmm6 1885 addl %esi,%ebx 1886 movl %r9d,%r13d 1887 addl %ebx,%r14d 1888 vpsrlq $2,%xmm7,%xmm7 1889 shrdl $14,%r13d,%r13d 1890 movl %r14d,%ebx 1891 movl %r10d,%r12d 1892 vpxor %xmm7,%xmm6,%xmm6 1893 xorl %r9d,%r13d 1894 shrdl $9,%r14d,%r14d 1895 xorl %r11d,%r12d 1896 vpshufd $232,%xmm6,%xmm6 1897 shrdl $5,%r13d,%r13d 1898 xorl %ebx,%r14d 1899 andl %r9d,%r12d 1900 vpslldq $8,%xmm6,%xmm6 1901 vpor %xmm11,%xmm8,%xmm8 1902 vaesenclast %xmm10,%xmm9,%xmm11 1903 vmovdqu 0-128(%rdi),%xmm10 1904 xorl %r9d,%r13d 1905 addl 60(%rsp),%eax 1906 movl %ebx,%esi 1907 vpaddd %xmm6,%xmm3,%xmm3 1908 shrdl $11,%r14d,%r14d 1909 xorl %r11d,%r12d 1910 xorl %ecx,%esi 1911 vpaddd 96(%rbp),%xmm3,%xmm6 1912 shrdl $6,%r13d,%r13d 1913 addl %r12d,%eax 1914 andl %esi,%r15d 1915 xorl %ebx,%r14d 1916 addl %r13d,%eax 1917 xorl %ecx,%r15d 1918 addl %eax,%r8d 1919 shrdl $2,%r14d,%r14d 1920 addl %r15d,%eax 1921 movl %r8d,%r13d 1922 addl %eax,%r14d 1923 vmovdqa %xmm6,48(%rsp) 1924 movq 64+0(%rsp),%r12 1925 vpand %xmm14,%xmm11,%xmm11 1926 movq 64+8(%rsp),%r15 1927 vpor %xmm11,%xmm8,%xmm8 1928 vmovdqu %xmm8,(%r15,%r12,1) 1929 leaq 16(%r12),%r12 1930 cmpb $0,131(%rbp) 1931 jne .Lavx_00_47 1932 vmovdqu (%r12),%xmm9 1933 movq %r12,64+0(%rsp) 1934 shrdl $14,%r13d,%r13d 1935 movl %r14d,%eax 1936 movl %r9d,%r12d 1937 xorl %r8d,%r13d 1938 shrdl $9,%r14d,%r14d 1939 xorl %r10d,%r12d 1940 shrdl $5,%r13d,%r13d 1941 xorl %eax,%r14d 1942 andl %r8d,%r12d 1943 vpxor %xmm10,%xmm9,%xmm9 1944 vmovdqu 16-128(%rdi),%xmm10 1945 xorl %r8d,%r13d 1946 addl 0(%rsp),%r11d 1947 movl %eax,%r15d 1948 shrdl $11,%r14d,%r14d 1949 xorl %r10d,%r12d 1950 xorl %ebx,%r15d 1951 shrdl $6,%r13d,%r13d 1952 addl %r12d,%r11d 1953 andl %r15d,%esi 1954 xorl %eax,%r14d 1955 addl %r13d,%r11d 1956 xorl %ebx,%esi 1957 addl %r11d,%edx 1958 shrdl $2,%r14d,%r14d 1959 addl %esi,%r11d 1960 movl %edx,%r13d 1961 addl %r11d,%r14d 1962 shrdl $14,%r13d,%r13d 1963 movl %r14d,%r11d 1964 movl %r8d,%r12d 1965 xorl %edx,%r13d 1966 shrdl $9,%r14d,%r14d 1967 xorl %r9d,%r12d 1968 shrdl $5,%r13d,%r13d 1969 xorl %r11d,%r14d 1970 andl %edx,%r12d 1971 vpxor %xmm8,%xmm9,%xmm9 1972 xorl %edx,%r13d 1973 addl 4(%rsp),%r10d 1974 movl %r11d,%esi 1975 shrdl $11,%r14d,%r14d 1976 xorl %r9d,%r12d 1977 xorl %eax,%esi 1978 shrdl $6,%r13d,%r13d 1979 addl %r12d,%r10d 1980 andl %esi,%r15d 1981 xorl %r11d,%r14d 1982 addl %r13d,%r10d 1983 xorl %eax,%r15d 1984 addl %r10d,%ecx 1985 shrdl $2,%r14d,%r14d 1986 addl %r15d,%r10d 1987 movl %ecx,%r13d 1988 addl %r10d,%r14d 1989 shrdl $14,%r13d,%r13d 1990 movl %r14d,%r10d 1991 movl %edx,%r12d 1992 xorl %ecx,%r13d 1993 shrdl $9,%r14d,%r14d 1994 xorl %r8d,%r12d 1995 shrdl $5,%r13d,%r13d 1996 xorl %r10d,%r14d 1997 andl %ecx,%r12d 1998 vaesenc %xmm10,%xmm9,%xmm9 1999 vmovdqu 32-128(%rdi),%xmm10 2000 xorl %ecx,%r13d 2001 addl 8(%rsp),%r9d 2002 movl %r10d,%r15d 2003 shrdl $11,%r14d,%r14d 2004 xorl %r8d,%r12d 2005 xorl %r11d,%r15d 2006 shrdl $6,%r13d,%r13d 2007 addl %r12d,%r9d 2008 andl %r15d,%esi 2009 xorl %r10d,%r14d 2010 addl %r13d,%r9d 2011 xorl %r11d,%esi 2012 addl %r9d,%ebx 2013 shrdl $2,%r14d,%r14d 2014 addl %esi,%r9d 2015 movl %ebx,%r13d 2016 addl %r9d,%r14d 2017 shrdl $14,%r13d,%r13d 2018 movl %r14d,%r9d 2019 movl %ecx,%r12d 2020 xorl %ebx,%r13d 2021 shrdl $9,%r14d,%r14d 2022 xorl %edx,%r12d 2023 shrdl $5,%r13d,%r13d 2024 xorl %r9d,%r14d 2025 andl %ebx,%r12d 2026 vaesenc %xmm10,%xmm9,%xmm9 2027 vmovdqu 48-128(%rdi),%xmm10 2028 xorl %ebx,%r13d 2029 addl 12(%rsp),%r8d 2030 movl %r9d,%esi 2031 shrdl $11,%r14d,%r14d 2032 xorl %edx,%r12d 2033 xorl %r10d,%esi 2034 shrdl $6,%r13d,%r13d 2035 addl %r12d,%r8d 2036 andl %esi,%r15d 2037 xorl %r9d,%r14d 2038 addl %r13d,%r8d 2039 xorl %r10d,%r15d 2040 addl %r8d,%eax 2041 shrdl $2,%r14d,%r14d 2042 addl %r15d,%r8d 2043 movl %eax,%r13d 2044 addl %r8d,%r14d 2045 shrdl $14,%r13d,%r13d 2046 movl %r14d,%r8d 2047 movl %ebx,%r12d 2048 xorl %eax,%r13d 2049 shrdl $9,%r14d,%r14d 2050 xorl %ecx,%r12d 2051 shrdl $5,%r13d,%r13d 2052 xorl %r8d,%r14d 2053 andl %eax,%r12d 2054 vaesenc %xmm10,%xmm9,%xmm9 2055 vmovdqu 64-128(%rdi),%xmm10 2056 xorl %eax,%r13d 2057 addl 16(%rsp),%edx 2058 movl %r8d,%r15d 2059 shrdl $11,%r14d,%r14d 2060 xorl %ecx,%r12d 2061 xorl %r9d,%r15d 2062 shrdl $6,%r13d,%r13d 2063 addl %r12d,%edx 2064 andl %r15d,%esi 2065 xorl %r8d,%r14d 2066 addl %r13d,%edx 2067 xorl %r9d,%esi 2068 addl %edx,%r11d 2069 shrdl $2,%r14d,%r14d 2070 addl %esi,%edx 2071 movl %r11d,%r13d 2072 addl %edx,%r14d 2073 shrdl $14,%r13d,%r13d 2074 movl %r14d,%edx 2075 movl %eax,%r12d 2076 xorl %r11d,%r13d 2077 shrdl $9,%r14d,%r14d 2078 xorl %ebx,%r12d 2079 shrdl $5,%r13d,%r13d 2080 xorl %edx,%r14d 2081 andl %r11d,%r12d 2082 vaesenc %xmm10,%xmm9,%xmm9 2083 vmovdqu 80-128(%rdi),%xmm10 2084 xorl %r11d,%r13d 2085 addl 20(%rsp),%ecx 2086 movl %edx,%esi 2087 shrdl $11,%r14d,%r14d 2088 xorl %ebx,%r12d 2089 xorl %r8d,%esi 2090 shrdl $6,%r13d,%r13d 2091 addl %r12d,%ecx 2092 andl %esi,%r15d 2093 xorl %edx,%r14d 2094 addl %r13d,%ecx 2095 xorl %r8d,%r15d 2096 addl %ecx,%r10d 2097 shrdl $2,%r14d,%r14d 2098 addl %r15d,%ecx 2099 movl %r10d,%r13d 2100 addl %ecx,%r14d 2101 shrdl $14,%r13d,%r13d 2102 movl %r14d,%ecx 2103 movl %r11d,%r12d 2104 xorl %r10d,%r13d 2105 shrdl $9,%r14d,%r14d 2106 xorl %eax,%r12d 2107 shrdl $5,%r13d,%r13d 2108 xorl %ecx,%r14d 2109 andl %r10d,%r12d 2110 vaesenc %xmm10,%xmm9,%xmm9 2111 vmovdqu 96-128(%rdi),%xmm10 2112 xorl %r10d,%r13d 2113 addl 24(%rsp),%ebx 2114 movl %ecx,%r15d 2115 shrdl $11,%r14d,%r14d 2116 xorl %eax,%r12d 2117 xorl %edx,%r15d 2118 shrdl $6,%r13d,%r13d 2119 addl %r12d,%ebx 2120 andl %r15d,%esi 2121 xorl %ecx,%r14d 2122 addl %r13d,%ebx 2123 xorl %edx,%esi 2124 addl %ebx,%r9d 2125 shrdl $2,%r14d,%r14d 2126 addl %esi,%ebx 2127 movl %r9d,%r13d 2128 addl %ebx,%r14d 2129 shrdl $14,%r13d,%r13d 2130 movl %r14d,%ebx 2131 movl %r10d,%r12d 2132 xorl %r9d,%r13d 2133 shrdl $9,%r14d,%r14d 2134 xorl %r11d,%r12d 2135 shrdl $5,%r13d,%r13d 2136 xorl %ebx,%r14d 2137 andl %r9d,%r12d 2138 vaesenc %xmm10,%xmm9,%xmm9 2139 vmovdqu 112-128(%rdi),%xmm10 2140 xorl %r9d,%r13d 2141 addl 28(%rsp),%eax 2142 movl %ebx,%esi 2143 shrdl $11,%r14d,%r14d 2144 xorl %r11d,%r12d 2145 xorl %ecx,%esi 2146 shrdl $6,%r13d,%r13d 2147 addl %r12d,%eax 2148 andl %esi,%r15d 2149 xorl %ebx,%r14d 2150 addl %r13d,%eax 2151 xorl %ecx,%r15d 2152 addl %eax,%r8d 2153 shrdl $2,%r14d,%r14d 2154 addl %r15d,%eax 2155 movl %r8d,%r13d 2156 addl %eax,%r14d 2157 shrdl $14,%r13d,%r13d 2158 movl %r14d,%eax 2159 movl %r9d,%r12d 2160 xorl %r8d,%r13d 2161 shrdl $9,%r14d,%r14d 2162 xorl %r10d,%r12d 2163 shrdl $5,%r13d,%r13d 2164 xorl %eax,%r14d 2165 andl %r8d,%r12d 2166 vaesenc %xmm10,%xmm9,%xmm9 2167 vmovdqu 128-128(%rdi),%xmm10 2168 xorl %r8d,%r13d 2169 addl 32(%rsp),%r11d 2170 movl %eax,%r15d 2171 shrdl $11,%r14d,%r14d 2172 xorl %r10d,%r12d 2173 xorl %ebx,%r15d 2174 shrdl $6,%r13d,%r13d 2175 addl %r12d,%r11d 2176 andl %r15d,%esi 2177 xorl %eax,%r14d 2178 addl %r13d,%r11d 2179 xorl %ebx,%esi 2180 addl %r11d,%edx 2181 shrdl $2,%r14d,%r14d 2182 addl %esi,%r11d 2183 movl %edx,%r13d 2184 addl %r11d,%r14d 2185 shrdl $14,%r13d,%r13d 2186 movl %r14d,%r11d 2187 movl %r8d,%r12d 2188 xorl %edx,%r13d 2189 shrdl $9,%r14d,%r14d 2190 xorl %r9d,%r12d 2191 shrdl $5,%r13d,%r13d 2192 xorl %r11d,%r14d 2193 andl %edx,%r12d 2194 vaesenc %xmm10,%xmm9,%xmm9 2195 vmovdqu 144-128(%rdi),%xmm10 2196 xorl %edx,%r13d 2197 addl 36(%rsp),%r10d 2198 movl %r11d,%esi 2199 shrdl $11,%r14d,%r14d 2200 xorl %r9d,%r12d 2201 xorl %eax,%esi 2202 shrdl $6,%r13d,%r13d 2203 addl %r12d,%r10d 2204 andl %esi,%r15d 2205 xorl %r11d,%r14d 2206 addl %r13d,%r10d 2207 xorl %eax,%r15d 2208 addl %r10d,%ecx 2209 shrdl $2,%r14d,%r14d 2210 addl %r15d,%r10d 2211 movl %ecx,%r13d 2212 addl %r10d,%r14d 2213 shrdl $14,%r13d,%r13d 2214 movl %r14d,%r10d 2215 movl %edx,%r12d 2216 xorl %ecx,%r13d 2217 shrdl $9,%r14d,%r14d 2218 xorl %r8d,%r12d 2219 shrdl $5,%r13d,%r13d 2220 xorl %r10d,%r14d 2221 andl %ecx,%r12d 2222 vaesenc %xmm10,%xmm9,%xmm9 2223 vmovdqu 160-128(%rdi),%xmm10 2224 xorl %ecx,%r13d 2225 addl 40(%rsp),%r9d 2226 movl %r10d,%r15d 2227 shrdl $11,%r14d,%r14d 2228 xorl %r8d,%r12d 2229 xorl %r11d,%r15d 2230 shrdl $6,%r13d,%r13d 2231 addl %r12d,%r9d 2232 andl %r15d,%esi 2233 xorl %r10d,%r14d 2234 addl %r13d,%r9d 2235 xorl %r11d,%esi 2236 addl %r9d,%ebx 2237 shrdl $2,%r14d,%r14d 2238 addl %esi,%r9d 2239 movl %ebx,%r13d 2240 addl %r9d,%r14d 2241 shrdl $14,%r13d,%r13d 2242 movl %r14d,%r9d 2243 movl %ecx,%r12d 2244 xorl %ebx,%r13d 2245 shrdl $9,%r14d,%r14d 2246 xorl %edx,%r12d 2247 shrdl $5,%r13d,%r13d 2248 xorl %r9d,%r14d 2249 andl %ebx,%r12d 2250 vaesenclast %xmm10,%xmm9,%xmm11 2251 vaesenc %xmm10,%xmm9,%xmm9 2252 vmovdqu 176-128(%rdi),%xmm10 2253 xorl %ebx,%r13d 2254 addl 44(%rsp),%r8d 2255 movl %r9d,%esi 2256 shrdl $11,%r14d,%r14d 2257 xorl %edx,%r12d 2258 xorl %r10d,%esi 2259 shrdl $6,%r13d,%r13d 2260 addl %r12d,%r8d 2261 andl %esi,%r15d 2262 xorl %r9d,%r14d 2263 addl %r13d,%r8d 2264 xorl %r10d,%r15d 2265 addl %r8d,%eax 2266 shrdl $2,%r14d,%r14d 2267 addl %r15d,%r8d 2268 movl %eax,%r13d 2269 addl %r8d,%r14d 2270 shrdl $14,%r13d,%r13d 2271 movl %r14d,%r8d 2272 movl %ebx,%r12d 2273 xorl %eax,%r13d 2274 shrdl $9,%r14d,%r14d 2275 xorl %ecx,%r12d 2276 shrdl $5,%r13d,%r13d 2277 xorl %r8d,%r14d 2278 andl %eax,%r12d 2279 vpand %xmm12,%xmm11,%xmm8 2280 vaesenc %xmm10,%xmm9,%xmm9 2281 vmovdqu 192-128(%rdi),%xmm10 2282 xorl %eax,%r13d 2283 addl 48(%rsp),%edx 2284 movl %r8d,%r15d 2285 shrdl $11,%r14d,%r14d 2286 xorl %ecx,%r12d 2287 xorl %r9d,%r15d 2288 shrdl $6,%r13d,%r13d 2289 addl %r12d,%edx 2290 andl %r15d,%esi 2291 xorl %r8d,%r14d 2292 addl %r13d,%edx 2293 xorl %r9d,%esi 2294 addl %edx,%r11d 2295 shrdl $2,%r14d,%r14d 2296 addl %esi,%edx 2297 movl %r11d,%r13d 2298 addl %edx,%r14d 2299 shrdl $14,%r13d,%r13d 2300 movl %r14d,%edx 2301 movl %eax,%r12d 2302 xorl %r11d,%r13d 2303 shrdl $9,%r14d,%r14d 2304 xorl %ebx,%r12d 2305 shrdl $5,%r13d,%r13d 2306 xorl %edx,%r14d 2307 andl %r11d,%r12d 2308 vaesenclast %xmm10,%xmm9,%xmm11 2309 vaesenc %xmm10,%xmm9,%xmm9 2310 vmovdqu 208-128(%rdi),%xmm10 2311 xorl %r11d,%r13d 2312 addl 52(%rsp),%ecx 2313 movl %edx,%esi 2314 shrdl $11,%r14d,%r14d 2315 xorl %ebx,%r12d 2316 xorl %r8d,%esi 2317 shrdl $6,%r13d,%r13d 2318 addl %r12d,%ecx 2319 andl %esi,%r15d 2320 xorl %edx,%r14d 2321 addl %r13d,%ecx 2322 xorl %r8d,%r15d 2323 addl %ecx,%r10d 2324 shrdl $2,%r14d,%r14d 2325 addl %r15d,%ecx 2326 movl %r10d,%r13d 2327 addl %ecx,%r14d 2328 shrdl $14,%r13d,%r13d 2329 movl %r14d,%ecx 2330 movl %r11d,%r12d 2331 xorl %r10d,%r13d 2332 shrdl $9,%r14d,%r14d 2333 xorl %eax,%r12d 2334 shrdl $5,%r13d,%r13d 2335 xorl %ecx,%r14d 2336 andl %r10d,%r12d 2337 vpand %xmm13,%xmm11,%xmm11 2338 vaesenc %xmm10,%xmm9,%xmm9 2339 vmovdqu 224-128(%rdi),%xmm10 2340 xorl %r10d,%r13d 2341 addl 56(%rsp),%ebx 2342 movl %ecx,%r15d 2343 shrdl $11,%r14d,%r14d 2344 xorl %eax,%r12d 2345 xorl %edx,%r15d 2346 shrdl $6,%r13d,%r13d 2347 addl %r12d,%ebx 2348 andl %r15d,%esi 2349 xorl %ecx,%r14d 2350 addl %r13d,%ebx 2351 xorl %edx,%esi 2352 addl %ebx,%r9d 2353 shrdl $2,%r14d,%r14d 2354 addl %esi,%ebx 2355 movl %r9d,%r13d 2356 addl %ebx,%r14d 2357 shrdl $14,%r13d,%r13d 2358 movl %r14d,%ebx 2359 movl %r10d,%r12d 2360 xorl %r9d,%r13d 2361 shrdl $9,%r14d,%r14d 2362 xorl %r11d,%r12d 2363 shrdl $5,%r13d,%r13d 2364 xorl %ebx,%r14d 2365 andl %r9d,%r12d 2366 vpor %xmm11,%xmm8,%xmm8 2367 vaesenclast %xmm10,%xmm9,%xmm11 2368 vmovdqu 0-128(%rdi),%xmm10 2369 xorl %r9d,%r13d 2370 addl 60(%rsp),%eax 2371 movl %ebx,%esi 2372 shrdl $11,%r14d,%r14d 2373 xorl %r11d,%r12d 2374 xorl %ecx,%esi 2375 shrdl $6,%r13d,%r13d 2376 addl %r12d,%eax 2377 andl %esi,%r15d 2378 xorl %ebx,%r14d 2379 addl %r13d,%eax 2380 xorl %ecx,%r15d 2381 addl %eax,%r8d 2382 shrdl $2,%r14d,%r14d 2383 addl %r15d,%eax 2384 movl %r8d,%r13d 2385 addl %eax,%r14d 2386 movq 64+0(%rsp),%r12 2387 movq 64+8(%rsp),%r13 2388 movq 64+40(%rsp),%r15 2389 movq 64+48(%rsp),%rsi 2390 2391 vpand %xmm14,%xmm11,%xmm11 2392 movl %r14d,%eax 2393 vpor %xmm11,%xmm8,%xmm8 2394 vmovdqu %xmm8,(%r12,%r13,1) 2395 leaq 16(%r12),%r12 2396 2397 addl 0(%r15),%eax 2398 addl 4(%r15),%ebx 2399 addl 8(%r15),%ecx 2400 addl 12(%r15),%edx 2401 addl 16(%r15),%r8d 2402 addl 20(%r15),%r9d 2403 addl 24(%r15),%r10d 2404 addl 28(%r15),%r11d 2405 2406 cmpq 64+16(%rsp),%r12 2407 2408 movl %eax,0(%r15) 2409 movl %ebx,4(%r15) 2410 movl %ecx,8(%r15) 2411 movl %edx,12(%r15) 2412 movl %r8d,16(%r15) 2413 movl %r9d,20(%r15) 2414 movl %r10d,24(%r15) 2415 movl %r11d,28(%r15) 2416 jb .Lloop_avx 2417 2418 movq 64+32(%rsp),%r8 2419 movq 120(%rsp),%rsi 2420 .cfi_def_cfa %rsi,8 2421 vmovdqu %xmm8,(%r8) 2422 vzeroall 2423 movq -48(%rsi),%r15 2424 .cfi_restore %r15 2425 movq -40(%rsi),%r14 2426 .cfi_restore %r14 2427 movq -32(%rsi),%r13 2428 .cfi_restore %r13 2429 movq -24(%rsi),%r12 2430 .cfi_restore %r12 2431 movq -16(%rsi),%rbp 2432 .cfi_restore %rbp 2433 movq -8(%rsi),%rbx 2434 .cfi_restore %rbx 2435 leaq (%rsi),%rsp 2436 .cfi_def_cfa_register %rsp 2437 .Lepilogue_avx: 2438 .byte 0xf3,0xc3 2439 .cfi_endproc 2440 .size aesni_cbc_sha256_enc_avx,.-aesni_cbc_sha256_enc_avx 2441 .type aesni_cbc_sha256_enc_avx2,@function 2442 .align 64 2443 aesni_cbc_sha256_enc_avx2: 2444 .cfi_startproc 2445 .Lavx2_shortcut: 2446 movq 8(%rsp),%r10 2447 movq %rsp,%rax 2448 .cfi_def_cfa_register %rax 2449 pushq %rbx 2450 .cfi_offset %rbx,-16 2451 pushq %rbp 2452 .cfi_offset %rbp,-24 2453 pushq %r12 2454 .cfi_offset %r12,-32 2455 pushq %r13 2456 .cfi_offset %r13,-40 2457 pushq %r14 2458 .cfi_offset %r14,-48 2459 pushq %r15 2460 .cfi_offset %r15,-56 2461 subq $576,%rsp 2462 andq $-1024,%rsp 2463 addq $448,%rsp 2464 2465 shlq $6,%rdx 2466 subq %rdi,%rsi 2467 subq %rdi,%r10 2468 addq %rdi,%rdx 2469 2470 2471 2472 movq %rdx,64+16(%rsp) 2473 2474 movq %r8,64+32(%rsp) 2475 movq %r9,64+40(%rsp) 2476 movq %r10,64+48(%rsp) 2477 movq %rax,120(%rsp) 2478 .cfi_escape 0x0f,0x06,0x77,0xf8,0x00,0x06,0x23,0x08 2479 .Lprologue_avx2: 2480 vzeroall 2481 2482 movq %rdi,%r13 2483 vpinsrq $1,%rsi,%xmm15,%xmm15 2484 leaq 128(%rcx),%rdi 2485 leaq K256+544(%rip),%r12 2486 movl 240-128(%rdi),%r14d 2487 movq %r9,%r15 2488 movq %r10,%rsi 2489 vmovdqu (%r8),%xmm8 2490 leaq -9(%r14),%r14 2491 2492 vmovdqa 0(%r12,%r14,8),%xmm14 2493 vmovdqa 16(%r12,%r14,8),%xmm13 2494 vmovdqa 32(%r12,%r14,8),%xmm12 2495 2496 subq $-64,%r13 2497 movl 0(%r15),%eax 2498 leaq (%rsi,%r13,1),%r12 2499 movl 4(%r15),%ebx 2500 cmpq %rdx,%r13 2501 movl 8(%r15),%ecx 2502 cmoveq %rsp,%r12 2503 movl 12(%r15),%edx 2504 movl 16(%r15),%r8d 2505 movl 20(%r15),%r9d 2506 movl 24(%r15),%r10d 2507 movl 28(%r15),%r11d 2508 vmovdqu 0-128(%rdi),%xmm10 2509 jmp .Loop_avx2 2510 .align 16 2511 .Loop_avx2: 2512 vmovdqa K256+512(%rip),%ymm7 2513 vmovdqu -64+0(%rsi,%r13,1),%xmm0 2514 vmovdqu -64+16(%rsi,%r13,1),%xmm1 2515 vmovdqu -64+32(%rsi,%r13,1),%xmm2 2516 vmovdqu -64+48(%rsi,%r13,1),%xmm3 2517 2518 vinserti128 $1,(%r12),%ymm0,%ymm0 2519 vinserti128 $1,16(%r12),%ymm1,%ymm1 2520 vpshufb %ymm7,%ymm0,%ymm0 2521 vinserti128 $1,32(%r12),%ymm2,%ymm2 2522 vpshufb %ymm7,%ymm1,%ymm1 2523 vinserti128 $1,48(%r12),%ymm3,%ymm3 2524 2525 leaq K256(%rip),%rbp 2526 vpshufb %ymm7,%ymm2,%ymm2 2527 leaq -64(%r13),%r13 2528 vpaddd 0(%rbp),%ymm0,%ymm4 2529 vpshufb %ymm7,%ymm3,%ymm3 2530 vpaddd 32(%rbp),%ymm1,%ymm5 2531 vpaddd 64(%rbp),%ymm2,%ymm6 2532 vpaddd 96(%rbp),%ymm3,%ymm7 2533 vmovdqa %ymm4,0(%rsp) 2534 xorl %r14d,%r14d 2535 vmovdqa %ymm5,32(%rsp) 2536 2537 movq 120(%rsp),%rsi 2538 .cfi_def_cfa %rsi,8 2539 leaq -64(%rsp),%rsp 2540 2541 2542 2543 movq %rsi,-8(%rsp) 2544 .cfi_escape 0x0f,0x05,0x77,0x78,0x06,0x23,0x08 2545 movl %ebx,%esi 2546 vmovdqa %ymm6,0(%rsp) 2547 xorl %ecx,%esi 2548 vmovdqa %ymm7,32(%rsp) 2549 movl %r9d,%r12d 2550 subq $-32*4,%rbp 2551 jmp .Lavx2_00_47 2552 2553 .align 16 2554 .Lavx2_00_47: 2555 vmovdqu (%r13),%xmm9 2556 vpinsrq $0,%r13,%xmm15,%xmm15 2557 leaq -64(%rsp),%rsp 2558 .cfi_escape 0x0f,0x05,0x77,0x38,0x06,0x23,0x08 2559 2560 pushq 64-8(%rsp) 2561 .cfi_escape 0x0f,0x05,0x77,0x00,0x06,0x23,0x08 2562 leaq 8(%rsp),%rsp 2563 .cfi_escape 0x0f,0x05,0x77,0x78,0x06,0x23,0x08 2564 vpalignr $4,%ymm0,%ymm1,%ymm4 2565 addl 0+128(%rsp),%r11d 2566 andl %r8d,%r12d 2567 rorxl $25,%r8d,%r13d 2568 vpalignr $4,%ymm2,%ymm3,%ymm7 2569 rorxl $11,%r8d,%r15d 2570 leal (%rax,%r14,1),%eax 2571 leal (%r11,%r12,1),%r11d 2572 vpsrld $7,%ymm4,%ymm6 2573 andnl %r10d,%r8d,%r12d 2574 xorl %r15d,%r13d 2575 rorxl $6,%r8d,%r14d 2576 vpaddd %ymm7,%ymm0,%ymm0 2577 leal (%r11,%r12,1),%r11d 2578 xorl %r14d,%r13d 2579 movl %eax,%r15d 2580 vpsrld $3,%ymm4,%ymm7 2581 rorxl $22,%eax,%r12d 2582 leal (%r11,%r13,1),%r11d 2583 xorl %ebx,%r15d 2584 vpslld $14,%ymm4,%ymm5 2585 rorxl $13,%eax,%r14d 2586 rorxl $2,%eax,%r13d 2587 leal (%rdx,%r11,1),%edx 2588 vpxor %ymm6,%ymm7,%ymm4 2589 andl %r15d,%esi 2590 vpxor %xmm10,%xmm9,%xmm9 2591 vmovdqu 16-128(%rdi),%xmm10 2592 xorl %r12d,%r14d 2593 xorl %ebx,%esi 2594 vpshufd $250,%ymm3,%ymm7 2595 xorl %r13d,%r14d 2596 leal (%r11,%rsi,1),%r11d 2597 movl %r8d,%r12d 2598 vpsrld $11,%ymm6,%ymm6 2599 addl 4+128(%rsp),%r10d 2600 andl %edx,%r12d 2601 rorxl $25,%edx,%r13d 2602 vpxor %ymm5,%ymm4,%ymm4 2603 rorxl $11,%edx,%esi 2604 leal (%r11,%r14,1),%r11d 2605 leal (%r10,%r12,1),%r10d 2606 vpslld $11,%ymm5,%ymm5 2607 andnl %r9d,%edx,%r12d 2608 xorl %esi,%r13d 2609 rorxl $6,%edx,%r14d 2610 vpxor %ymm6,%ymm4,%ymm4 2611 leal (%r10,%r12,1),%r10d 2612 xorl %r14d,%r13d 2613 movl %r11d,%esi 2614 vpsrld $10,%ymm7,%ymm6 2615 rorxl $22,%r11d,%r12d 2616 leal (%r10,%r13,1),%r10d 2617 xorl %eax,%esi 2618 vpxor %ymm5,%ymm4,%ymm4 2619 rorxl $13,%r11d,%r14d 2620 rorxl $2,%r11d,%r13d 2621 leal (%rcx,%r10,1),%ecx 2622 vpsrlq $17,%ymm7,%ymm7 2623 andl %esi,%r15d 2624 vpxor %xmm8,%xmm9,%xmm9 2625 xorl %r12d,%r14d 2626 xorl %eax,%r15d 2627 vpaddd %ymm4,%ymm0,%ymm0 2628 xorl %r13d,%r14d 2629 leal (%r10,%r15,1),%r10d 2630 movl %edx,%r12d 2631 vpxor %ymm7,%ymm6,%ymm6 2632 addl 8+128(%rsp),%r9d 2633 andl %ecx,%r12d 2634 rorxl $25,%ecx,%r13d 2635 vpsrlq $2,%ymm7,%ymm7 2636 rorxl $11,%ecx,%r15d 2637 leal (%r10,%r14,1),%r10d 2638 leal (%r9,%r12,1),%r9d 2639 vpxor %ymm7,%ymm6,%ymm6 2640 andnl %r8d,%ecx,%r12d 2641 xorl %r15d,%r13d 2642 rorxl $6,%ecx,%r14d 2643 vpshufd $132,%ymm6,%ymm6 2644 leal (%r9,%r12,1),%r9d 2645 xorl %r14d,%r13d 2646 movl %r10d,%r15d 2647 vpsrldq $8,%ymm6,%ymm6 2648 rorxl $22,%r10d,%r12d 2649 leal (%r9,%r13,1),%r9d 2650 xorl %r11d,%r15d 2651 vpaddd %ymm6,%ymm0,%ymm0 2652 rorxl $13,%r10d,%r14d 2653 rorxl $2,%r10d,%r13d 2654 leal (%rbx,%r9,1),%ebx 2655 vpshufd $80,%ymm0,%ymm7 2656 andl %r15d,%esi 2657 vaesenc %xmm10,%xmm9,%xmm9 2658 vmovdqu 32-128(%rdi),%xmm10 2659 xorl %r12d,%r14d 2660 xorl %r11d,%esi 2661 vpsrld $10,%ymm7,%ymm6 2662 xorl %r13d,%r14d 2663 leal (%r9,%rsi,1),%r9d 2664 movl %ecx,%r12d 2665 vpsrlq $17,%ymm7,%ymm7 2666 addl 12+128(%rsp),%r8d 2667 andl %ebx,%r12d 2668 rorxl $25,%ebx,%r13d 2669 vpxor %ymm7,%ymm6,%ymm6 2670 rorxl $11,%ebx,%esi 2671 leal (%r9,%r14,1),%r9d 2672 leal (%r8,%r12,1),%r8d 2673 vpsrlq $2,%ymm7,%ymm7 2674 andnl %edx,%ebx,%r12d 2675 xorl %esi,%r13d 2676 rorxl $6,%ebx,%r14d 2677 vpxor %ymm7,%ymm6,%ymm6 2678 leal (%r8,%r12,1),%r8d 2679 xorl %r14d,%r13d 2680 movl %r9d,%esi 2681 vpshufd $232,%ymm6,%ymm6 2682 rorxl $22,%r9d,%r12d 2683 leal (%r8,%r13,1),%r8d 2684 xorl %r10d,%esi 2685 vpslldq $8,%ymm6,%ymm6 2686 rorxl $13,%r9d,%r14d 2687 rorxl $2,%r9d,%r13d 2688 leal (%rax,%r8,1),%eax 2689 vpaddd %ymm6,%ymm0,%ymm0 2690 andl %esi,%r15d 2691 vaesenc %xmm10,%xmm9,%xmm9 2692 vmovdqu 48-128(%rdi),%xmm10 2693 xorl %r12d,%r14d 2694 xorl %r10d,%r15d 2695 vpaddd 0(%rbp),%ymm0,%ymm6 2696 xorl %r13d,%r14d 2697 leal (%r8,%r15,1),%r8d 2698 movl %ebx,%r12d 2699 vmovdqa %ymm6,0(%rsp) 2700 vpalignr $4,%ymm1,%ymm2,%ymm4 2701 addl 32+128(%rsp),%edx 2702 andl %eax,%r12d 2703 rorxl $25,%eax,%r13d 2704 vpalignr $4,%ymm3,%ymm0,%ymm7 2705 rorxl $11,%eax,%r15d 2706 leal (%r8,%r14,1),%r8d 2707 leal (%rdx,%r12,1),%edx 2708 vpsrld $7,%ymm4,%ymm6 2709 andnl %ecx,%eax,%r12d 2710 xorl %r15d,%r13d 2711 rorxl $6,%eax,%r14d 2712 vpaddd %ymm7,%ymm1,%ymm1 2713 leal (%rdx,%r12,1),%edx 2714 xorl %r14d,%r13d 2715 movl %r8d,%r15d 2716 vpsrld $3,%ymm4,%ymm7 2717 rorxl $22,%r8d,%r12d 2718 leal (%rdx,%r13,1),%edx 2719 xorl %r9d,%r15d 2720 vpslld $14,%ymm4,%ymm5 2721 rorxl $13,%r8d,%r14d 2722 rorxl $2,%r8d,%r13d 2723 leal (%r11,%rdx,1),%r11d 2724 vpxor %ymm6,%ymm7,%ymm4 2725 andl %r15d,%esi 2726 vaesenc %xmm10,%xmm9,%xmm9 2727 vmovdqu 64-128(%rdi),%xmm10 2728 xorl %r12d,%r14d 2729 xorl %r9d,%esi 2730 vpshufd $250,%ymm0,%ymm7 2731 xorl %r13d,%r14d 2732 leal (%rdx,%rsi,1),%edx 2733 movl %eax,%r12d 2734 vpsrld $11,%ymm6,%ymm6 2735 addl 36+128(%rsp),%ecx 2736 andl %r11d,%r12d 2737 rorxl $25,%r11d,%r13d 2738 vpxor %ymm5,%ymm4,%ymm4 2739 rorxl $11,%r11d,%esi 2740 leal (%rdx,%r14,1),%edx 2741 leal (%rcx,%r12,1),%ecx 2742 vpslld $11,%ymm5,%ymm5 2743 andnl %ebx,%r11d,%r12d 2744 xorl %esi,%r13d 2745 rorxl $6,%r11d,%r14d 2746 vpxor %ymm6,%ymm4,%ymm4 2747 leal (%rcx,%r12,1),%ecx 2748 xorl %r14d,%r13d 2749 movl %edx,%esi 2750 vpsrld $10,%ymm7,%ymm6 2751 rorxl $22,%edx,%r12d 2752 leal (%rcx,%r13,1),%ecx 2753 xorl %r8d,%esi 2754 vpxor %ymm5,%ymm4,%ymm4 2755 rorxl $13,%edx,%r14d 2756 rorxl $2,%edx,%r13d 2757 leal (%r10,%rcx,1),%r10d 2758 vpsrlq $17,%ymm7,%ymm7 2759 andl %esi,%r15d 2760 vaesenc %xmm10,%xmm9,%xmm9 2761 vmovdqu 80-128(%rdi),%xmm10 2762 xorl %r12d,%r14d 2763 xorl %r8d,%r15d 2764 vpaddd %ymm4,%ymm1,%ymm1 2765 xorl %r13d,%r14d 2766 leal (%rcx,%r15,1),%ecx 2767 movl %r11d,%r12d 2768 vpxor %ymm7,%ymm6,%ymm6 2769 addl 40+128(%rsp),%ebx 2770 andl %r10d,%r12d 2771 rorxl $25,%r10d,%r13d 2772 vpsrlq $2,%ymm7,%ymm7 2773 rorxl $11,%r10d,%r15d 2774 leal (%rcx,%r14,1),%ecx 2775 leal (%rbx,%r12,1),%ebx 2776 vpxor %ymm7,%ymm6,%ymm6 2777 andnl %eax,%r10d,%r12d 2778 xorl %r15d,%r13d 2779 rorxl $6,%r10d,%r14d 2780 vpshufd $132,%ymm6,%ymm6 2781 leal (%rbx,%r12,1),%ebx 2782 xorl %r14d,%r13d 2783 movl %ecx,%r15d 2784 vpsrldq $8,%ymm6,%ymm6 2785 rorxl $22,%ecx,%r12d 2786 leal (%rbx,%r13,1),%ebx 2787 xorl %edx,%r15d 2788 vpaddd %ymm6,%ymm1,%ymm1 2789 rorxl $13,%ecx,%r14d 2790 rorxl $2,%ecx,%r13d 2791 leal (%r9,%rbx,1),%r9d 2792 vpshufd $80,%ymm1,%ymm7 2793 andl %r15d,%esi 2794 vaesenc %xmm10,%xmm9,%xmm9 2795 vmovdqu 96-128(%rdi),%xmm10 2796 xorl %r12d,%r14d 2797 xorl %edx,%esi 2798 vpsrld $10,%ymm7,%ymm6 2799 xorl %r13d,%r14d 2800 leal (%rbx,%rsi,1),%ebx 2801 movl %r10d,%r12d 2802 vpsrlq $17,%ymm7,%ymm7 2803 addl 44+128(%rsp),%eax 2804 andl %r9d,%r12d 2805 rorxl $25,%r9d,%r13d 2806 vpxor %ymm7,%ymm6,%ymm6 2807 rorxl $11,%r9d,%esi 2808 leal (%rbx,%r14,1),%ebx 2809 leal (%rax,%r12,1),%eax 2810 vpsrlq $2,%ymm7,%ymm7 2811 andnl %r11d,%r9d,%r12d 2812 xorl %esi,%r13d 2813 rorxl $6,%r9d,%r14d 2814 vpxor %ymm7,%ymm6,%ymm6 2815 leal (%rax,%r12,1),%eax 2816 xorl %r14d,%r13d 2817 movl %ebx,%esi 2818 vpshufd $232,%ymm6,%ymm6 2819 rorxl $22,%ebx,%r12d 2820 leal (%rax,%r13,1),%eax 2821 xorl %ecx,%esi 2822 vpslldq $8,%ymm6,%ymm6 2823 rorxl $13,%ebx,%r14d 2824 rorxl $2,%ebx,%r13d 2825 leal (%r8,%rax,1),%r8d 2826 vpaddd %ymm6,%ymm1,%ymm1 2827 andl %esi,%r15d 2828 vaesenc %xmm10,%xmm9,%xmm9 2829 vmovdqu 112-128(%rdi),%xmm10 2830 xorl %r12d,%r14d 2831 xorl %ecx,%r15d 2832 vpaddd 32(%rbp),%ymm1,%ymm6 2833 xorl %r13d,%r14d 2834 leal (%rax,%r15,1),%eax 2835 movl %r9d,%r12d 2836 vmovdqa %ymm6,32(%rsp) 2837 leaq -64(%rsp),%rsp 2838 .cfi_escape 0x0f,0x05,0x77,0x38,0x06,0x23,0x08 2839 2840 pushq 64-8(%rsp) 2841 .cfi_escape 0x0f,0x05,0x77,0x00,0x06,0x23,0x08 2842 leaq 8(%rsp),%rsp 2843 .cfi_escape 0x0f,0x05,0x77,0x78,0x06,0x23,0x08 2844 vpalignr $4,%ymm2,%ymm3,%ymm4 2845 addl 0+128(%rsp),%r11d 2846 andl %r8d,%r12d 2847 rorxl $25,%r8d,%r13d 2848 vpalignr $4,%ymm0,%ymm1,%ymm7 2849 rorxl $11,%r8d,%r15d 2850 leal (%rax,%r14,1),%eax 2851 leal (%r11,%r12,1),%r11d 2852 vpsrld $7,%ymm4,%ymm6 2853 andnl %r10d,%r8d,%r12d 2854 xorl %r15d,%r13d 2855 rorxl $6,%r8d,%r14d 2856 vpaddd %ymm7,%ymm2,%ymm2 2857 leal (%r11,%r12,1),%r11d 2858 xorl %r14d,%r13d 2859 movl %eax,%r15d 2860 vpsrld $3,%ymm4,%ymm7 2861 rorxl $22,%eax,%r12d 2862 leal (%r11,%r13,1),%r11d 2863 xorl %ebx,%r15d 2864 vpslld $14,%ymm4,%ymm5 2865 rorxl $13,%eax,%r14d 2866 rorxl $2,%eax,%r13d 2867 leal (%rdx,%r11,1),%edx 2868 vpxor %ymm6,%ymm7,%ymm4 2869 andl %r15d,%esi 2870 vaesenc %xmm10,%xmm9,%xmm9 2871 vmovdqu 128-128(%rdi),%xmm10 2872 xorl %r12d,%r14d 2873 xorl %ebx,%esi 2874 vpshufd $250,%ymm1,%ymm7 2875 xorl %r13d,%r14d 2876 leal (%r11,%rsi,1),%r11d 2877 movl %r8d,%r12d 2878 vpsrld $11,%ymm6,%ymm6 2879 addl 4+128(%rsp),%r10d 2880 andl %edx,%r12d 2881 rorxl $25,%edx,%r13d 2882 vpxor %ymm5,%ymm4,%ymm4 2883 rorxl $11,%edx,%esi 2884 leal (%r11,%r14,1),%r11d 2885 leal (%r10,%r12,1),%r10d 2886 vpslld $11,%ymm5,%ymm5 2887 andnl %r9d,%edx,%r12d 2888 xorl %esi,%r13d 2889 rorxl $6,%edx,%r14d 2890 vpxor %ymm6,%ymm4,%ymm4 2891 leal (%r10,%r12,1),%r10d 2892 xorl %r14d,%r13d 2893 movl %r11d,%esi 2894 vpsrld $10,%ymm7,%ymm6 2895 rorxl $22,%r11d,%r12d 2896 leal (%r10,%r13,1),%r10d 2897 xorl %eax,%esi 2898 vpxor %ymm5,%ymm4,%ymm4 2899 rorxl $13,%r11d,%r14d 2900 rorxl $2,%r11d,%r13d 2901 leal (%rcx,%r10,1),%ecx 2902 vpsrlq $17,%ymm7,%ymm7 2903 andl %esi,%r15d 2904 vaesenc %xmm10,%xmm9,%xmm9 2905 vmovdqu 144-128(%rdi),%xmm10 2906 xorl %r12d,%r14d 2907 xorl %eax,%r15d 2908 vpaddd %ymm4,%ymm2,%ymm2 2909 xorl %r13d,%r14d 2910 leal (%r10,%r15,1),%r10d 2911 movl %edx,%r12d 2912 vpxor %ymm7,%ymm6,%ymm6 2913 addl 8+128(%rsp),%r9d 2914 andl %ecx,%r12d 2915 rorxl $25,%ecx,%r13d 2916 vpsrlq $2,%ymm7,%ymm7 2917 rorxl $11,%ecx,%r15d 2918 leal (%r10,%r14,1),%r10d 2919 leal (%r9,%r12,1),%r9d 2920 vpxor %ymm7,%ymm6,%ymm6 2921 andnl %r8d,%ecx,%r12d 2922 xorl %r15d,%r13d 2923 rorxl $6,%ecx,%r14d 2924 vpshufd $132,%ymm6,%ymm6 2925 leal (%r9,%r12,1),%r9d 2926 xorl %r14d,%r13d 2927 movl %r10d,%r15d 2928 vpsrldq $8,%ymm6,%ymm6 2929 rorxl $22,%r10d,%r12d 2930 leal (%r9,%r13,1),%r9d 2931 xorl %r11d,%r15d 2932 vpaddd %ymm6,%ymm2,%ymm2 2933 rorxl $13,%r10d,%r14d 2934 rorxl $2,%r10d,%r13d 2935 leal (%rbx,%r9,1),%ebx 2936 vpshufd $80,%ymm2,%ymm7 2937 andl %r15d,%esi 2938 vaesenc %xmm10,%xmm9,%xmm9 2939 vmovdqu 160-128(%rdi),%xmm10 2940 xorl %r12d,%r14d 2941 xorl %r11d,%esi 2942 vpsrld $10,%ymm7,%ymm6 2943 xorl %r13d,%r14d 2944 leal (%r9,%rsi,1),%r9d 2945 movl %ecx,%r12d 2946 vpsrlq $17,%ymm7,%ymm7 2947 addl 12+128(%rsp),%r8d 2948 andl %ebx,%r12d 2949 rorxl $25,%ebx,%r13d 2950 vpxor %ymm7,%ymm6,%ymm6 2951 rorxl $11,%ebx,%esi 2952 leal (%r9,%r14,1),%r9d 2953 leal (%r8,%r12,1),%r8d 2954 vpsrlq $2,%ymm7,%ymm7 2955 andnl %edx,%ebx,%r12d 2956 xorl %esi,%r13d 2957 rorxl $6,%ebx,%r14d 2958 vpxor %ymm7,%ymm6,%ymm6 2959 leal (%r8,%r12,1),%r8d 2960 xorl %r14d,%r13d 2961 movl %r9d,%esi 2962 vpshufd $232,%ymm6,%ymm6 2963 rorxl $22,%r9d,%r12d 2964 leal (%r8,%r13,1),%r8d 2965 xorl %r10d,%esi 2966 vpslldq $8,%ymm6,%ymm6 2967 rorxl $13,%r9d,%r14d 2968 rorxl $2,%r9d,%r13d 2969 leal (%rax,%r8,1),%eax 2970 vpaddd %ymm6,%ymm2,%ymm2 2971 andl %esi,%r15d 2972 vaesenclast %xmm10,%xmm9,%xmm11 2973 vaesenc %xmm10,%xmm9,%xmm9 2974 vmovdqu 176-128(%rdi),%xmm10 2975 xorl %r12d,%r14d 2976 xorl %r10d,%r15d 2977 vpaddd 64(%rbp),%ymm2,%ymm6 2978 xorl %r13d,%r14d 2979 leal (%r8,%r15,1),%r8d 2980 movl %ebx,%r12d 2981 vmovdqa %ymm6,0(%rsp) 2982 vpalignr $4,%ymm3,%ymm0,%ymm4 2983 addl 32+128(%rsp),%edx 2984 andl %eax,%r12d 2985 rorxl $25,%eax,%r13d 2986 vpalignr $4,%ymm1,%ymm2,%ymm7 2987 rorxl $11,%eax,%r15d 2988 leal (%r8,%r14,1),%r8d 2989 leal (%rdx,%r12,1),%edx 2990 vpsrld $7,%ymm4,%ymm6 2991 andnl %ecx,%eax,%r12d 2992 xorl %r15d,%r13d 2993 rorxl $6,%eax,%r14d 2994 vpaddd %ymm7,%ymm3,%ymm3 2995 leal (%rdx,%r12,1),%edx 2996 xorl %r14d,%r13d 2997 movl %r8d,%r15d 2998 vpsrld $3,%ymm4,%ymm7 2999 rorxl $22,%r8d,%r12d 3000 leal (%rdx,%r13,1),%edx 3001 xorl %r9d,%r15d 3002 vpslld $14,%ymm4,%ymm5 3003 rorxl $13,%r8d,%r14d 3004 rorxl $2,%r8d,%r13d 3005 leal (%r11,%rdx,1),%r11d 3006 vpxor %ymm6,%ymm7,%ymm4 3007 andl %r15d,%esi 3008 vpand %xmm12,%xmm11,%xmm8 3009 vaesenc %xmm10,%xmm9,%xmm9 3010 vmovdqu 192-128(%rdi),%xmm10 3011 xorl %r12d,%r14d 3012 xorl %r9d,%esi 3013 vpshufd $250,%ymm2,%ymm7 3014 xorl %r13d,%r14d 3015 leal (%rdx,%rsi,1),%edx 3016 movl %eax,%r12d 3017 vpsrld $11,%ymm6,%ymm6 3018 addl 36+128(%rsp),%ecx 3019 andl %r11d,%r12d 3020 rorxl $25,%r11d,%r13d 3021 vpxor %ymm5,%ymm4,%ymm4 3022 rorxl $11,%r11d,%esi 3023 leal (%rdx,%r14,1),%edx 3024 leal (%rcx,%r12,1),%ecx 3025 vpslld $11,%ymm5,%ymm5 3026 andnl %ebx,%r11d,%r12d 3027 xorl %esi,%r13d 3028 rorxl $6,%r11d,%r14d 3029 vpxor %ymm6,%ymm4,%ymm4 3030 leal (%rcx,%r12,1),%ecx 3031 xorl %r14d,%r13d 3032 movl %edx,%esi 3033 vpsrld $10,%ymm7,%ymm6 3034 rorxl $22,%edx,%r12d 3035 leal (%rcx,%r13,1),%ecx 3036 xorl %r8d,%esi 3037 vpxor %ymm5,%ymm4,%ymm4 3038 rorxl $13,%edx,%r14d 3039 rorxl $2,%edx,%r13d 3040 leal (%r10,%rcx,1),%r10d 3041 vpsrlq $17,%ymm7,%ymm7 3042 andl %esi,%r15d 3043 vaesenclast %xmm10,%xmm9,%xmm11 3044 vaesenc %xmm10,%xmm9,%xmm9 3045 vmovdqu 208-128(%rdi),%xmm10 3046 xorl %r12d,%r14d 3047 xorl %r8d,%r15d 3048 vpaddd %ymm4,%ymm3,%ymm3 3049 xorl %r13d,%r14d 3050 leal (%rcx,%r15,1),%ecx 3051 movl %r11d,%r12d 3052 vpxor %ymm7,%ymm6,%ymm6 3053 addl 40+128(%rsp),%ebx 3054 andl %r10d,%r12d 3055 rorxl $25,%r10d,%r13d 3056 vpsrlq $2,%ymm7,%ymm7 3057 rorxl $11,%r10d,%r15d 3058 leal (%rcx,%r14,1),%ecx 3059 leal (%rbx,%r12,1),%ebx 3060 vpxor %ymm7,%ymm6,%ymm6 3061 andnl %eax,%r10d,%r12d 3062 xorl %r15d,%r13d 3063 rorxl $6,%r10d,%r14d 3064 vpshufd $132,%ymm6,%ymm6 3065 leal (%rbx,%r12,1),%ebx 3066 xorl %r14d,%r13d 3067 movl %ecx,%r15d 3068 vpsrldq $8,%ymm6,%ymm6 3069 rorxl $22,%ecx,%r12d 3070 leal (%rbx,%r13,1),%ebx 3071 xorl %edx,%r15d 3072 vpaddd %ymm6,%ymm3,%ymm3 3073 rorxl $13,%ecx,%r14d 3074 rorxl $2,%ecx,%r13d 3075 leal (%r9,%rbx,1),%r9d 3076 vpshufd $80,%ymm3,%ymm7 3077 andl %r15d,%esi 3078 vpand %xmm13,%xmm11,%xmm11 3079 vaesenc %xmm10,%xmm9,%xmm9 3080 vmovdqu 224-128(%rdi),%xmm10 3081 xorl %r12d,%r14d 3082 xorl %edx,%esi 3083 vpsrld $10,%ymm7,%ymm6 3084 xorl %r13d,%r14d 3085 leal (%rbx,%rsi,1),%ebx 3086 movl %r10d,%r12d 3087 vpsrlq $17,%ymm7,%ymm7 3088 addl 44+128(%rsp),%eax 3089 andl %r9d,%r12d 3090 rorxl $25,%r9d,%r13d 3091 vpxor %ymm7,%ymm6,%ymm6 3092 rorxl $11,%r9d,%esi 3093 leal (%rbx,%r14,1),%ebx 3094 leal (%rax,%r12,1),%eax 3095 vpsrlq $2,%ymm7,%ymm7 3096 andnl %r11d,%r9d,%r12d 3097 xorl %esi,%r13d 3098 rorxl $6,%r9d,%r14d 3099 vpxor %ymm7,%ymm6,%ymm6 3100 leal (%rax,%r12,1),%eax 3101 xorl %r14d,%r13d 3102 movl %ebx,%esi 3103 vpshufd $232,%ymm6,%ymm6 3104 rorxl $22,%ebx,%r12d 3105 leal (%rax,%r13,1),%eax 3106 xorl %ecx,%esi 3107 vpslldq $8,%ymm6,%ymm6 3108 rorxl $13,%ebx,%r14d 3109 rorxl $2,%ebx,%r13d 3110 leal (%r8,%rax,1),%r8d 3111 vpaddd %ymm6,%ymm3,%ymm3 3112 andl %esi,%r15d 3113 vpor %xmm11,%xmm8,%xmm8 3114 vaesenclast %xmm10,%xmm9,%xmm11 3115 vmovdqu 0-128(%rdi),%xmm10 3116 xorl %r12d,%r14d 3117 xorl %ecx,%r15d 3118 vpaddd 96(%rbp),%ymm3,%ymm6 3119 xorl %r13d,%r14d 3120 leal (%rax,%r15,1),%eax 3121 movl %r9d,%r12d 3122 vmovdqa %ymm6,32(%rsp) 3123 vmovq %xmm15,%r13 3124 vpextrq $1,%xmm15,%r15 3125 vpand %xmm14,%xmm11,%xmm11 3126 vpor %xmm11,%xmm8,%xmm8 3127 vmovdqu %xmm8,(%r15,%r13,1) 3128 leaq 16(%r13),%r13 3129 leaq 128(%rbp),%rbp 3130 cmpb $0,3(%rbp) 3131 jne .Lavx2_00_47 3132 vmovdqu (%r13),%xmm9 3133 vpinsrq $0,%r13,%xmm15,%xmm15 3134 addl 0+64(%rsp),%r11d 3135 andl %r8d,%r12d 3136 rorxl $25,%r8d,%r13d 3137 rorxl $11,%r8d,%r15d 3138 leal (%rax,%r14,1),%eax 3139 leal (%r11,%r12,1),%r11d 3140 andnl %r10d,%r8d,%r12d 3141 xorl %r15d,%r13d 3142 rorxl $6,%r8d,%r14d 3143 leal (%r11,%r12,1),%r11d 3144 xorl %r14d,%r13d 3145 movl %eax,%r15d 3146 rorxl $22,%eax,%r12d 3147 leal (%r11,%r13,1),%r11d 3148 xorl %ebx,%r15d 3149 rorxl $13,%eax,%r14d 3150 rorxl $2,%eax,%r13d 3151 leal (%rdx,%r11,1),%edx 3152 andl %r15d,%esi 3153 vpxor %xmm10,%xmm9,%xmm9 3154 vmovdqu 16-128(%rdi),%xmm10 3155 xorl %r12d,%r14d 3156 xorl %ebx,%esi 3157 xorl %r13d,%r14d 3158 leal (%r11,%rsi,1),%r11d 3159 movl %r8d,%r12d 3160 addl 4+64(%rsp),%r10d 3161 andl %edx,%r12d 3162 rorxl $25,%edx,%r13d 3163 rorxl $11,%edx,%esi 3164 leal (%r11,%r14,1),%r11d 3165 leal (%r10,%r12,1),%r10d 3166 andnl %r9d,%edx,%r12d 3167 xorl %esi,%r13d 3168 rorxl $6,%edx,%r14d 3169 leal (%r10,%r12,1),%r10d 3170 xorl %r14d,%r13d 3171 movl %r11d,%esi 3172 rorxl $22,%r11d,%r12d 3173 leal (%r10,%r13,1),%r10d 3174 xorl %eax,%esi 3175 rorxl $13,%r11d,%r14d 3176 rorxl $2,%r11d,%r13d 3177 leal (%rcx,%r10,1),%ecx 3178 andl %esi,%r15d 3179 vpxor %xmm8,%xmm9,%xmm9 3180 xorl %r12d,%r14d 3181 xorl %eax,%r15d 3182 xorl %r13d,%r14d 3183 leal (%r10,%r15,1),%r10d 3184 movl %edx,%r12d 3185 addl 8+64(%rsp),%r9d 3186 andl %ecx,%r12d 3187 rorxl $25,%ecx,%r13d 3188 rorxl $11,%ecx,%r15d 3189 leal (%r10,%r14,1),%r10d 3190 leal (%r9,%r12,1),%r9d 3191 andnl %r8d,%ecx,%r12d 3192 xorl %r15d,%r13d 3193 rorxl $6,%ecx,%r14d 3194 leal (%r9,%r12,1),%r9d 3195 xorl %r14d,%r13d 3196 movl %r10d,%r15d 3197 rorxl $22,%r10d,%r12d 3198 leal (%r9,%r13,1),%r9d 3199 xorl %r11d,%r15d 3200 rorxl $13,%r10d,%r14d 3201 rorxl $2,%r10d,%r13d 3202 leal (%rbx,%r9,1),%ebx 3203 andl %r15d,%esi 3204 vaesenc %xmm10,%xmm9,%xmm9 3205 vmovdqu 32-128(%rdi),%xmm10 3206 xorl %r12d,%r14d 3207 xorl %r11d,%esi 3208 xorl %r13d,%r14d 3209 leal (%r9,%rsi,1),%r9d 3210 movl %ecx,%r12d 3211 addl 12+64(%rsp),%r8d 3212 andl %ebx,%r12d 3213 rorxl $25,%ebx,%r13d 3214 rorxl $11,%ebx,%esi 3215 leal (%r9,%r14,1),%r9d 3216 leal (%r8,%r12,1),%r8d 3217 andnl %edx,%ebx,%r12d 3218 xorl %esi,%r13d 3219 rorxl $6,%ebx,%r14d 3220 leal (%r8,%r12,1),%r8d 3221 xorl %r14d,%r13d 3222 movl %r9d,%esi 3223 rorxl $22,%r9d,%r12d 3224 leal (%r8,%r13,1),%r8d 3225 xorl %r10d,%esi 3226 rorxl $13,%r9d,%r14d 3227 rorxl $2,%r9d,%r13d 3228 leal (%rax,%r8,1),%eax 3229 andl %esi,%r15d 3230 vaesenc %xmm10,%xmm9,%xmm9 3231 vmovdqu 48-128(%rdi),%xmm10 3232 xorl %r12d,%r14d 3233 xorl %r10d,%r15d 3234 xorl %r13d,%r14d 3235 leal (%r8,%r15,1),%r8d 3236 movl %ebx,%r12d 3237 addl 32+64(%rsp),%edx 3238 andl %eax,%r12d 3239 rorxl $25,%eax,%r13d 3240 rorxl $11,%eax,%r15d 3241 leal (%r8,%r14,1),%r8d 3242 leal (%rdx,%r12,1),%edx 3243 andnl %ecx,%eax,%r12d 3244 xorl %r15d,%r13d 3245 rorxl $6,%eax,%r14d 3246 leal (%rdx,%r12,1),%edx 3247 xorl %r14d,%r13d 3248 movl %r8d,%r15d 3249 rorxl $22,%r8d,%r12d 3250 leal (%rdx,%r13,1),%edx 3251 xorl %r9d,%r15d 3252 rorxl $13,%r8d,%r14d 3253 rorxl $2,%r8d,%r13d 3254 leal (%r11,%rdx,1),%r11d 3255 andl %r15d,%esi 3256 vaesenc %xmm10,%xmm9,%xmm9 3257 vmovdqu 64-128(%rdi),%xmm10 3258 xorl %r12d,%r14d 3259 xorl %r9d,%esi 3260 xorl %r13d,%r14d 3261 leal (%rdx,%rsi,1),%edx 3262 movl %eax,%r12d 3263 addl 36+64(%rsp),%ecx 3264 andl %r11d,%r12d 3265 rorxl $25,%r11d,%r13d 3266 rorxl $11,%r11d,%esi 3267 leal (%rdx,%r14,1),%edx 3268 leal (%rcx,%r12,1),%ecx 3269 andnl %ebx,%r11d,%r12d 3270 xorl %esi,%r13d 3271 rorxl $6,%r11d,%r14d 3272 leal (%rcx,%r12,1),%ecx 3273 xorl %r14d,%r13d 3274 movl %edx,%esi 3275 rorxl $22,%edx,%r12d 3276 leal (%rcx,%r13,1),%ecx 3277 xorl %r8d,%esi 3278 rorxl $13,%edx,%r14d 3279 rorxl $2,%edx,%r13d 3280 leal (%r10,%rcx,1),%r10d 3281 andl %esi,%r15d 3282 vaesenc %xmm10,%xmm9,%xmm9 3283 vmovdqu 80-128(%rdi),%xmm10 3284 xorl %r12d,%r14d 3285 xorl %r8d,%r15d 3286 xorl %r13d,%r14d 3287 leal (%rcx,%r15,1),%ecx 3288 movl %r11d,%r12d 3289 addl 40+64(%rsp),%ebx 3290 andl %r10d,%r12d 3291 rorxl $25,%r10d,%r13d 3292 rorxl $11,%r10d,%r15d 3293 leal (%rcx,%r14,1),%ecx 3294 leal (%rbx,%r12,1),%ebx 3295 andnl %eax,%r10d,%r12d 3296 xorl %r15d,%r13d 3297 rorxl $6,%r10d,%r14d 3298 leal (%rbx,%r12,1),%ebx 3299 xorl %r14d,%r13d 3300 movl %ecx,%r15d 3301 rorxl $22,%ecx,%r12d 3302 leal (%rbx,%r13,1),%ebx 3303 xorl %edx,%r15d 3304 rorxl $13,%ecx,%r14d 3305 rorxl $2,%ecx,%r13d 3306 leal (%r9,%rbx,1),%r9d 3307 andl %r15d,%esi 3308 vaesenc %xmm10,%xmm9,%xmm9 3309 vmovdqu 96-128(%rdi),%xmm10 3310 xorl %r12d,%r14d 3311 xorl %edx,%esi 3312 xorl %r13d,%r14d 3313 leal (%rbx,%rsi,1),%ebx 3314 movl %r10d,%r12d 3315 addl 44+64(%rsp),%eax 3316 andl %r9d,%r12d 3317 rorxl $25,%r9d,%r13d 3318 rorxl $11,%r9d,%esi 3319 leal (%rbx,%r14,1),%ebx 3320 leal (%rax,%r12,1),%eax 3321 andnl %r11d,%r9d,%r12d 3322 xorl %esi,%r13d 3323 rorxl $6,%r9d,%r14d 3324 leal (%rax,%r12,1),%eax 3325 xorl %r14d,%r13d 3326 movl %ebx,%esi 3327 rorxl $22,%ebx,%r12d 3328 leal (%rax,%r13,1),%eax 3329 xorl %ecx,%esi 3330 rorxl $13,%ebx,%r14d 3331 rorxl $2,%ebx,%r13d 3332 leal (%r8,%rax,1),%r8d 3333 andl %esi,%r15d 3334 vaesenc %xmm10,%xmm9,%xmm9 3335 vmovdqu 112-128(%rdi),%xmm10 3336 xorl %r12d,%r14d 3337 xorl %ecx,%r15d 3338 xorl %r13d,%r14d 3339 leal (%rax,%r15,1),%eax 3340 movl %r9d,%r12d 3341 addl 0(%rsp),%r11d 3342 andl %r8d,%r12d 3343 rorxl $25,%r8d,%r13d 3344 rorxl $11,%r8d,%r15d 3345 leal (%rax,%r14,1),%eax 3346 leal (%r11,%r12,1),%r11d 3347 andnl %r10d,%r8d,%r12d 3348 xorl %r15d,%r13d 3349 rorxl $6,%r8d,%r14d 3350 leal (%r11,%r12,1),%r11d 3351 xorl %r14d,%r13d 3352 movl %eax,%r15d 3353 rorxl $22,%eax,%r12d 3354 leal (%r11,%r13,1),%r11d 3355 xorl %ebx,%r15d 3356 rorxl $13,%eax,%r14d 3357 rorxl $2,%eax,%r13d 3358 leal (%rdx,%r11,1),%edx 3359 andl %r15d,%esi 3360 vaesenc %xmm10,%xmm9,%xmm9 3361 vmovdqu 128-128(%rdi),%xmm10 3362 xorl %r12d,%r14d 3363 xorl %ebx,%esi 3364 xorl %r13d,%r14d 3365 leal (%r11,%rsi,1),%r11d 3366 movl %r8d,%r12d 3367 addl 4(%rsp),%r10d 3368 andl %edx,%r12d 3369 rorxl $25,%edx,%r13d 3370 rorxl $11,%edx,%esi 3371 leal (%r11,%r14,1),%r11d 3372 leal (%r10,%r12,1),%r10d 3373 andnl %r9d,%edx,%r12d 3374 xorl %esi,%r13d 3375 rorxl $6,%edx,%r14d 3376 leal (%r10,%r12,1),%r10d 3377 xorl %r14d,%r13d 3378 movl %r11d,%esi 3379 rorxl $22,%r11d,%r12d 3380 leal (%r10,%r13,1),%r10d 3381 xorl %eax,%esi 3382 rorxl $13,%r11d,%r14d 3383 rorxl $2,%r11d,%r13d 3384 leal (%rcx,%r10,1),%ecx 3385 andl %esi,%r15d 3386 vaesenc %xmm10,%xmm9,%xmm9 3387 vmovdqu 144-128(%rdi),%xmm10 3388 xorl %r12d,%r14d 3389 xorl %eax,%r15d 3390 xorl %r13d,%r14d 3391 leal (%r10,%r15,1),%r10d 3392 movl %edx,%r12d 3393 addl 8(%rsp),%r9d 3394 andl %ecx,%r12d 3395 rorxl $25,%ecx,%r13d 3396 rorxl $11,%ecx,%r15d 3397 leal (%r10,%r14,1),%r10d 3398 leal (%r9,%r12,1),%r9d 3399 andnl %r8d,%ecx,%r12d 3400 xorl %r15d,%r13d 3401 rorxl $6,%ecx,%r14d 3402 leal (%r9,%r12,1),%r9d 3403 xorl %r14d,%r13d 3404 movl %r10d,%r15d 3405 rorxl $22,%r10d,%r12d 3406 leal (%r9,%r13,1),%r9d 3407 xorl %r11d,%r15d 3408 rorxl $13,%r10d,%r14d 3409 rorxl $2,%r10d,%r13d 3410 leal (%rbx,%r9,1),%ebx 3411 andl %r15d,%esi 3412 vaesenc %xmm10,%xmm9,%xmm9 3413 vmovdqu 160-128(%rdi),%xmm10 3414 xorl %r12d,%r14d 3415 xorl %r11d,%esi 3416 xorl %r13d,%r14d 3417 leal (%r9,%rsi,1),%r9d 3418 movl %ecx,%r12d 3419 addl 12(%rsp),%r8d 3420 andl %ebx,%r12d 3421 rorxl $25,%ebx,%r13d 3422 rorxl $11,%ebx,%esi 3423 leal (%r9,%r14,1),%r9d 3424 leal (%r8,%r12,1),%r8d 3425 andnl %edx,%ebx,%r12d 3426 xorl %esi,%r13d 3427 rorxl $6,%ebx,%r14d 3428 leal (%r8,%r12,1),%r8d 3429 xorl %r14d,%r13d 3430 movl %r9d,%esi 3431 rorxl $22,%r9d,%r12d 3432 leal (%r8,%r13,1),%r8d 3433 xorl %r10d,%esi 3434 rorxl $13,%r9d,%r14d 3435 rorxl $2,%r9d,%r13d 3436 leal (%rax,%r8,1),%eax 3437 andl %esi,%r15d 3438 vaesenclast %xmm10,%xmm9,%xmm11 3439 vaesenc %xmm10,%xmm9,%xmm9 3440 vmovdqu 176-128(%rdi),%xmm10 3441 xorl %r12d,%r14d 3442 xorl %r10d,%r15d 3443 xorl %r13d,%r14d 3444 leal (%r8,%r15,1),%r8d 3445 movl %ebx,%r12d 3446 addl 32(%rsp),%edx 3447 andl %eax,%r12d 3448 rorxl $25,%eax,%r13d 3449 rorxl $11,%eax,%r15d 3450 leal (%r8,%r14,1),%r8d 3451 leal (%rdx,%r12,1),%edx 3452 andnl %ecx,%eax,%r12d 3453 xorl %r15d,%r13d 3454 rorxl $6,%eax,%r14d 3455 leal (%rdx,%r12,1),%edx 3456 xorl %r14d,%r13d 3457 movl %r8d,%r15d 3458 rorxl $22,%r8d,%r12d 3459 leal (%rdx,%r13,1),%edx 3460 xorl %r9d,%r15d 3461 rorxl $13,%r8d,%r14d 3462 rorxl $2,%r8d,%r13d 3463 leal (%r11,%rdx,1),%r11d 3464 andl %r15d,%esi 3465 vpand %xmm12,%xmm11,%xmm8 3466 vaesenc %xmm10,%xmm9,%xmm9 3467 vmovdqu 192-128(%rdi),%xmm10 3468 xorl %r12d,%r14d 3469 xorl %r9d,%esi 3470 xorl %r13d,%r14d 3471 leal (%rdx,%rsi,1),%edx 3472 movl %eax,%r12d 3473 addl 36(%rsp),%ecx 3474 andl %r11d,%r12d 3475 rorxl $25,%r11d,%r13d 3476 rorxl $11,%r11d,%esi 3477 leal (%rdx,%r14,1),%edx 3478 leal (%rcx,%r12,1),%ecx 3479 andnl %ebx,%r11d,%r12d 3480 xorl %esi,%r13d 3481 rorxl $6,%r11d,%r14d 3482 leal (%rcx,%r12,1),%ecx 3483 xorl %r14d,%r13d 3484 movl %edx,%esi 3485 rorxl $22,%edx,%r12d 3486 leal (%rcx,%r13,1),%ecx 3487 xorl %r8d,%esi 3488 rorxl $13,%edx,%r14d 3489 rorxl $2,%edx,%r13d 3490 leal (%r10,%rcx,1),%r10d 3491 andl %esi,%r15d 3492 vaesenclast %xmm10,%xmm9,%xmm11 3493 vaesenc %xmm10,%xmm9,%xmm9 3494 vmovdqu 208-128(%rdi),%xmm10 3495 xorl %r12d,%r14d 3496 xorl %r8d,%r15d 3497 xorl %r13d,%r14d 3498 leal (%rcx,%r15,1),%ecx 3499 movl %r11d,%r12d 3500 addl 40(%rsp),%ebx 3501 andl %r10d,%r12d 3502 rorxl $25,%r10d,%r13d 3503 rorxl $11,%r10d,%r15d 3504 leal (%rcx,%r14,1),%ecx 3505 leal (%rbx,%r12,1),%ebx 3506 andnl %eax,%r10d,%r12d 3507 xorl %r15d,%r13d 3508 rorxl $6,%r10d,%r14d 3509 leal (%rbx,%r12,1),%ebx 3510 xorl %r14d,%r13d 3511 movl %ecx,%r15d 3512 rorxl $22,%ecx,%r12d 3513 leal (%rbx,%r13,1),%ebx 3514 xorl %edx,%r15d 3515 rorxl $13,%ecx,%r14d 3516 rorxl $2,%ecx,%r13d 3517 leal (%r9,%rbx,1),%r9d 3518 andl %r15d,%esi 3519 vpand %xmm13,%xmm11,%xmm11 3520 vaesenc %xmm10,%xmm9,%xmm9 3521 vmovdqu 224-128(%rdi),%xmm10 3522 xorl %r12d,%r14d 3523 xorl %edx,%esi 3524 xorl %r13d,%r14d 3525 leal (%rbx,%rsi,1),%ebx 3526 movl %r10d,%r12d 3527 addl 44(%rsp),%eax 3528 andl %r9d,%r12d 3529 rorxl $25,%r9d,%r13d 3530 rorxl $11,%r9d,%esi 3531 leal (%rbx,%r14,1),%ebx 3532 leal (%rax,%r12,1),%eax 3533 andnl %r11d,%r9d,%r12d 3534 xorl %esi,%r13d 3535 rorxl $6,%r9d,%r14d 3536 leal (%rax,%r12,1),%eax 3537 xorl %r14d,%r13d 3538 movl %ebx,%esi 3539 rorxl $22,%ebx,%r12d 3540 leal (%rax,%r13,1),%eax 3541 xorl %ecx,%esi 3542 rorxl $13,%ebx,%r14d 3543 rorxl $2,%ebx,%r13d 3544 leal (%r8,%rax,1),%r8d 3545 andl %esi,%r15d 3546 vpor %xmm11,%xmm8,%xmm8 3547 vaesenclast %xmm10,%xmm9,%xmm11 3548 vmovdqu 0-128(%rdi),%xmm10 3549 xorl %r12d,%r14d 3550 xorl %ecx,%r15d 3551 xorl %r13d,%r14d 3552 leal (%rax,%r15,1),%eax 3553 movl %r9d,%r12d 3554 vpextrq $1,%xmm15,%r12 3555 vmovq %xmm15,%r13 3556 movq 552(%rsp),%r15 3557 addl %r14d,%eax 3558 leaq 448(%rsp),%rbp 3559 3560 vpand %xmm14,%xmm11,%xmm11 3561 vpor %xmm11,%xmm8,%xmm8 3562 vmovdqu %xmm8,(%r12,%r13,1) 3563 leaq 16(%r13),%r13 3564 3565 addl 0(%r15),%eax 3566 addl 4(%r15),%ebx 3567 addl 8(%r15),%ecx 3568 addl 12(%r15),%edx 3569 addl 16(%r15),%r8d 3570 addl 20(%r15),%r9d 3571 addl 24(%r15),%r10d 3572 addl 28(%r15),%r11d 3573 3574 movl %eax,0(%r15) 3575 movl %ebx,4(%r15) 3576 movl %ecx,8(%r15) 3577 movl %edx,12(%r15) 3578 movl %r8d,16(%r15) 3579 movl %r9d,20(%r15) 3580 movl %r10d,24(%r15) 3581 movl %r11d,28(%r15) 3582 3583 cmpq 80(%rbp),%r13 3584 je .Ldone_avx2 3585 3586 xorl %r14d,%r14d 3587 movl %ebx,%esi 3588 movl %r9d,%r12d 3589 xorl %ecx,%esi 3590 jmp .Lower_avx2 3591 .align 16 3592 .Lower_avx2: 3593 vmovdqu (%r13),%xmm9 3594 vpinsrq $0,%r13,%xmm15,%xmm15 3595 addl 0+16(%rbp),%r11d 3596 andl %r8d,%r12d 3597 rorxl $25,%r8d,%r13d 3598 rorxl $11,%r8d,%r15d 3599 leal (%rax,%r14,1),%eax 3600 leal (%r11,%r12,1),%r11d 3601 andnl %r10d,%r8d,%r12d 3602 xorl %r15d,%r13d 3603 rorxl $6,%r8d,%r14d 3604 leal (%r11,%r12,1),%r11d 3605 xorl %r14d,%r13d 3606 movl %eax,%r15d 3607 rorxl $22,%eax,%r12d 3608 leal (%r11,%r13,1),%r11d 3609 xorl %ebx,%r15d 3610 rorxl $13,%eax,%r14d 3611 rorxl $2,%eax,%r13d 3612 leal (%rdx,%r11,1),%edx 3613 andl %r15d,%esi 3614 vpxor %xmm10,%xmm9,%xmm9 3615 vmovdqu 16-128(%rdi),%xmm10 3616 xorl %r12d,%r14d 3617 xorl %ebx,%esi 3618 xorl %r13d,%r14d 3619 leal (%r11,%rsi,1),%r11d 3620 movl %r8d,%r12d 3621 addl 4+16(%rbp),%r10d 3622 andl %edx,%r12d 3623 rorxl $25,%edx,%r13d 3624 rorxl $11,%edx,%esi 3625 leal (%r11,%r14,1),%r11d 3626 leal (%r10,%r12,1),%r10d 3627 andnl %r9d,%edx,%r12d 3628 xorl %esi,%r13d 3629 rorxl $6,%edx,%r14d 3630 leal (%r10,%r12,1),%r10d 3631 xorl %r14d,%r13d 3632 movl %r11d,%esi 3633 rorxl $22,%r11d,%r12d 3634 leal (%r10,%r13,1),%r10d 3635 xorl %eax,%esi 3636 rorxl $13,%r11d,%r14d 3637 rorxl $2,%r11d,%r13d 3638 leal (%rcx,%r10,1),%ecx 3639 andl %esi,%r15d 3640 vpxor %xmm8,%xmm9,%xmm9 3641 xorl %r12d,%r14d 3642 xorl %eax,%r15d 3643 xorl %r13d,%r14d 3644 leal (%r10,%r15,1),%r10d 3645 movl %edx,%r12d 3646 addl 8+16(%rbp),%r9d 3647 andl %ecx,%r12d 3648 rorxl $25,%ecx,%r13d 3649 rorxl $11,%ecx,%r15d 3650 leal (%r10,%r14,1),%r10d 3651 leal (%r9,%r12,1),%r9d 3652 andnl %r8d,%ecx,%r12d 3653 xorl %r15d,%r13d 3654 rorxl $6,%ecx,%r14d 3655 leal (%r9,%r12,1),%r9d 3656 xorl %r14d,%r13d 3657 movl %r10d,%r15d 3658 rorxl $22,%r10d,%r12d 3659 leal (%r9,%r13,1),%r9d 3660 xorl %r11d,%r15d 3661 rorxl $13,%r10d,%r14d 3662 rorxl $2,%r10d,%r13d 3663 leal (%rbx,%r9,1),%ebx 3664 andl %r15d,%esi 3665 vaesenc %xmm10,%xmm9,%xmm9 3666 vmovdqu 32-128(%rdi),%xmm10 3667 xorl %r12d,%r14d 3668 xorl %r11d,%esi 3669 xorl %r13d,%r14d 3670 leal (%r9,%rsi,1),%r9d 3671 movl %ecx,%r12d 3672 addl 12+16(%rbp),%r8d 3673 andl %ebx,%r12d 3674 rorxl $25,%ebx,%r13d 3675 rorxl $11,%ebx,%esi 3676 leal (%r9,%r14,1),%r9d 3677 leal (%r8,%r12,1),%r8d 3678 andnl %edx,%ebx,%r12d 3679 xorl %esi,%r13d 3680 rorxl $6,%ebx,%r14d 3681 leal (%r8,%r12,1),%r8d 3682 xorl %r14d,%r13d 3683 movl %r9d,%esi 3684 rorxl $22,%r9d,%r12d 3685 leal (%r8,%r13,1),%r8d 3686 xorl %r10d,%esi 3687 rorxl $13,%r9d,%r14d 3688 rorxl $2,%r9d,%r13d 3689 leal (%rax,%r8,1),%eax 3690 andl %esi,%r15d 3691 vaesenc %xmm10,%xmm9,%xmm9 3692 vmovdqu 48-128(%rdi),%xmm10 3693 xorl %r12d,%r14d 3694 xorl %r10d,%r15d 3695 xorl %r13d,%r14d 3696 leal (%r8,%r15,1),%r8d 3697 movl %ebx,%r12d 3698 addl 32+16(%rbp),%edx 3699 andl %eax,%r12d 3700 rorxl $25,%eax,%r13d 3701 rorxl $11,%eax,%r15d 3702 leal (%r8,%r14,1),%r8d 3703 leal (%rdx,%r12,1),%edx 3704 andnl %ecx,%eax,%r12d 3705 xorl %r15d,%r13d 3706 rorxl $6,%eax,%r14d 3707 leal (%rdx,%r12,1),%edx 3708 xorl %r14d,%r13d 3709 movl %r8d,%r15d 3710 rorxl $22,%r8d,%r12d 3711 leal (%rdx,%r13,1),%edx 3712 xorl %r9d,%r15d 3713 rorxl $13,%r8d,%r14d 3714 rorxl $2,%r8d,%r13d 3715 leal (%r11,%rdx,1),%r11d 3716 andl %r15d,%esi 3717 vaesenc %xmm10,%xmm9,%xmm9 3718 vmovdqu 64-128(%rdi),%xmm10 3719 xorl %r12d,%r14d 3720 xorl %r9d,%esi 3721 xorl %r13d,%r14d 3722 leal (%rdx,%rsi,1),%edx 3723 movl %eax,%r12d 3724 addl 36+16(%rbp),%ecx 3725 andl %r11d,%r12d 3726 rorxl $25,%r11d,%r13d 3727 rorxl $11,%r11d,%esi 3728 leal (%rdx,%r14,1),%edx 3729 leal (%rcx,%r12,1),%ecx 3730 andnl %ebx,%r11d,%r12d 3731 xorl %esi,%r13d 3732 rorxl $6,%r11d,%r14d 3733 leal (%rcx,%r12,1),%ecx 3734 xorl %r14d,%r13d 3735 movl %edx,%esi 3736 rorxl $22,%edx,%r12d 3737 leal (%rcx,%r13,1),%ecx 3738 xorl %r8d,%esi 3739 rorxl $13,%edx,%r14d 3740 rorxl $2,%edx,%r13d 3741 leal (%r10,%rcx,1),%r10d 3742 andl %esi,%r15d 3743 vaesenc %xmm10,%xmm9,%xmm9 3744 vmovdqu 80-128(%rdi),%xmm10 3745 xorl %r12d,%r14d 3746 xorl %r8d,%r15d 3747 xorl %r13d,%r14d 3748 leal (%rcx,%r15,1),%ecx 3749 movl %r11d,%r12d 3750 addl 40+16(%rbp),%ebx 3751 andl %r10d,%r12d 3752 rorxl $25,%r10d,%r13d 3753 rorxl $11,%r10d,%r15d 3754 leal (%rcx,%r14,1),%ecx 3755 leal (%rbx,%r12,1),%ebx 3756 andnl %eax,%r10d,%r12d 3757 xorl %r15d,%r13d 3758 rorxl $6,%r10d,%r14d 3759 leal (%rbx,%r12,1),%ebx 3760 xorl %r14d,%r13d 3761 movl %ecx,%r15d 3762 rorxl $22,%ecx,%r12d 3763 leal (%rbx,%r13,1),%ebx 3764 xorl %edx,%r15d 3765 rorxl $13,%ecx,%r14d 3766 rorxl $2,%ecx,%r13d 3767 leal (%r9,%rbx,1),%r9d 3768 andl %r15d,%esi 3769 vaesenc %xmm10,%xmm9,%xmm9 3770 vmovdqu 96-128(%rdi),%xmm10 3771 xorl %r12d,%r14d 3772 xorl %edx,%esi 3773 xorl %r13d,%r14d 3774 leal (%rbx,%rsi,1),%ebx 3775 movl %r10d,%r12d 3776 addl 44+16(%rbp),%eax 3777 andl %r9d,%r12d 3778 rorxl $25,%r9d,%r13d 3779 rorxl $11,%r9d,%esi 3780 leal (%rbx,%r14,1),%ebx 3781 leal (%rax,%r12,1),%eax 3782 andnl %r11d,%r9d,%r12d 3783 xorl %esi,%r13d 3784 rorxl $6,%r9d,%r14d 3785 leal (%rax,%r12,1),%eax 3786 xorl %r14d,%r13d 3787 movl %ebx,%esi 3788 rorxl $22,%ebx,%r12d 3789 leal (%rax,%r13,1),%eax 3790 xorl %ecx,%esi 3791 rorxl $13,%ebx,%r14d 3792 rorxl $2,%ebx,%r13d 3793 leal (%r8,%rax,1),%r8d 3794 andl %esi,%r15d 3795 vaesenc %xmm10,%xmm9,%xmm9 3796 vmovdqu 112-128(%rdi),%xmm10 3797 xorl %r12d,%r14d 3798 xorl %ecx,%r15d 3799 xorl %r13d,%r14d 3800 leal (%rax,%r15,1),%eax 3801 movl %r9d,%r12d 3802 leaq -64(%rbp),%rbp 3803 addl 0+16(%rbp),%r11d 3804 andl %r8d,%r12d 3805 rorxl $25,%r8d,%r13d 3806 rorxl $11,%r8d,%r15d 3807 leal (%rax,%r14,1),%eax 3808 leal (%r11,%r12,1),%r11d 3809 andnl %r10d,%r8d,%r12d 3810 xorl %r15d,%r13d 3811 rorxl $6,%r8d,%r14d 3812 leal (%r11,%r12,1),%r11d 3813 xorl %r14d,%r13d 3814 movl %eax,%r15d 3815 rorxl $22,%eax,%r12d 3816 leal (%r11,%r13,1),%r11d 3817 xorl %ebx,%r15d 3818 rorxl $13,%eax,%r14d 3819 rorxl $2,%eax,%r13d 3820 leal (%rdx,%r11,1),%edx 3821 andl %r15d,%esi 3822 vaesenc %xmm10,%xmm9,%xmm9 3823 vmovdqu 128-128(%rdi),%xmm10 3824 xorl %r12d,%r14d 3825 xorl %ebx,%esi 3826 xorl %r13d,%r14d 3827 leal (%r11,%rsi,1),%r11d 3828 movl %r8d,%r12d 3829 addl 4+16(%rbp),%r10d 3830 andl %edx,%r12d 3831 rorxl $25,%edx,%r13d 3832 rorxl $11,%edx,%esi 3833 leal (%r11,%r14,1),%r11d 3834 leal (%r10,%r12,1),%r10d 3835 andnl %r9d,%edx,%r12d 3836 xorl %esi,%r13d 3837 rorxl $6,%edx,%r14d 3838 leal (%r10,%r12,1),%r10d 3839 xorl %r14d,%r13d 3840 movl %r11d,%esi 3841 rorxl $22,%r11d,%r12d 3842 leal (%r10,%r13,1),%r10d 3843 xorl %eax,%esi 3844 rorxl $13,%r11d,%r14d 3845 rorxl $2,%r11d,%r13d 3846 leal (%rcx,%r10,1),%ecx 3847 andl %esi,%r15d 3848 vaesenc %xmm10,%xmm9,%xmm9 3849 vmovdqu 144-128(%rdi),%xmm10 3850 xorl %r12d,%r14d 3851 xorl %eax,%r15d 3852 xorl %r13d,%r14d 3853 leal (%r10,%r15,1),%r10d 3854 movl %edx,%r12d 3855 addl 8+16(%rbp),%r9d 3856 andl %ecx,%r12d 3857 rorxl $25,%ecx,%r13d 3858 rorxl $11,%ecx,%r15d 3859 leal (%r10,%r14,1),%r10d 3860 leal (%r9,%r12,1),%r9d 3861 andnl %r8d,%ecx,%r12d 3862 xorl %r15d,%r13d 3863 rorxl $6,%ecx,%r14d 3864 leal (%r9,%r12,1),%r9d 3865 xorl %r14d,%r13d 3866 movl %r10d,%r15d 3867 rorxl $22,%r10d,%r12d 3868 leal (%r9,%r13,1),%r9d 3869 xorl %r11d,%r15d 3870 rorxl $13,%r10d,%r14d 3871 rorxl $2,%r10d,%r13d 3872 leal (%rbx,%r9,1),%ebx 3873 andl %r15d,%esi 3874 vaesenc %xmm10,%xmm9,%xmm9 3875 vmovdqu 160-128(%rdi),%xmm10 3876 xorl %r12d,%r14d 3877 xorl %r11d,%esi 3878 xorl %r13d,%r14d 3879 leal (%r9,%rsi,1),%r9d 3880 movl %ecx,%r12d 3881 addl 12+16(%rbp),%r8d 3882 andl %ebx,%r12d 3883 rorxl $25,%ebx,%r13d 3884 rorxl $11,%ebx,%esi 3885 leal (%r9,%r14,1),%r9d 3886 leal (%r8,%r12,1),%r8d 3887 andnl %edx,%ebx,%r12d 3888 xorl %esi,%r13d 3889 rorxl $6,%ebx,%r14d 3890 leal (%r8,%r12,1),%r8d 3891 xorl %r14d,%r13d 3892 movl %r9d,%esi 3893 rorxl $22,%r9d,%r12d 3894 leal (%r8,%r13,1),%r8d 3895 xorl %r10d,%esi 3896 rorxl $13,%r9d,%r14d 3897 rorxl $2,%r9d,%r13d 3898 leal (%rax,%r8,1),%eax 3899 andl %esi,%r15d 3900 vaesenclast %xmm10,%xmm9,%xmm11 3901 vaesenc %xmm10,%xmm9,%xmm9 3902 vmovdqu 176-128(%rdi),%xmm10 3903 xorl %r12d,%r14d 3904 xorl %r10d,%r15d 3905 xorl %r13d,%r14d 3906 leal (%r8,%r15,1),%r8d 3907 movl %ebx,%r12d 3908 addl 32+16(%rbp),%edx 3909 andl %eax,%r12d 3910 rorxl $25,%eax,%r13d 3911 rorxl $11,%eax,%r15d 3912 leal (%r8,%r14,1),%r8d 3913 leal (%rdx,%r12,1),%edx 3914 andnl %ecx,%eax,%r12d 3915 xorl %r15d,%r13d 3916 rorxl $6,%eax,%r14d 3917 leal (%rdx,%r12,1),%edx 3918 xorl %r14d,%r13d 3919 movl %r8d,%r15d 3920 rorxl $22,%r8d,%r12d 3921 leal (%rdx,%r13,1),%edx 3922 xorl %r9d,%r15d 3923 rorxl $13,%r8d,%r14d 3924 rorxl $2,%r8d,%r13d 3925 leal (%r11,%rdx,1),%r11d 3926 andl %r15d,%esi 3927 vpand %xmm12,%xmm11,%xmm8 3928 vaesenc %xmm10,%xmm9,%xmm9 3929 vmovdqu 192-128(%rdi),%xmm10 3930 xorl %r12d,%r14d 3931 xorl %r9d,%esi 3932 xorl %r13d,%r14d 3933 leal (%rdx,%rsi,1),%edx 3934 movl %eax,%r12d 3935 addl 36+16(%rbp),%ecx 3936 andl %r11d,%r12d 3937 rorxl $25,%r11d,%r13d 3938 rorxl $11,%r11d,%esi 3939 leal (%rdx,%r14,1),%edx 3940 leal (%rcx,%r12,1),%ecx 3941 andnl %ebx,%r11d,%r12d 3942 xorl %esi,%r13d 3943 rorxl $6,%r11d,%r14d 3944 leal (%rcx,%r12,1),%ecx 3945 xorl %r14d,%r13d 3946 movl %edx,%esi 3947 rorxl $22,%edx,%r12d 3948 leal (%rcx,%r13,1),%ecx 3949 xorl %r8d,%esi 3950 rorxl $13,%edx,%r14d 3951 rorxl $2,%edx,%r13d 3952 leal (%r10,%rcx,1),%r10d 3953 andl %esi,%r15d 3954 vaesenclast %xmm10,%xmm9,%xmm11 3955 vaesenc %xmm10,%xmm9,%xmm9 3956 vmovdqu 208-128(%rdi),%xmm10 3957 xorl %r12d,%r14d 3958 xorl %r8d,%r15d 3959 xorl %r13d,%r14d 3960 leal (%rcx,%r15,1),%ecx 3961 movl %r11d,%r12d 3962 addl 40+16(%rbp),%ebx 3963 andl %r10d,%r12d 3964 rorxl $25,%r10d,%r13d 3965 rorxl $11,%r10d,%r15d 3966 leal (%rcx,%r14,1),%ecx 3967 leal (%rbx,%r12,1),%ebx 3968 andnl %eax,%r10d,%r12d 3969 xorl %r15d,%r13d 3970 rorxl $6,%r10d,%r14d 3971 leal (%rbx,%r12,1),%ebx 3972 xorl %r14d,%r13d 3973 movl %ecx,%r15d 3974 rorxl $22,%ecx,%r12d 3975 leal (%rbx,%r13,1),%ebx 3976 xorl %edx,%r15d 3977 rorxl $13,%ecx,%r14d 3978 rorxl $2,%ecx,%r13d 3979 leal (%r9,%rbx,1),%r9d 3980 andl %r15d,%esi 3981 vpand %xmm13,%xmm11,%xmm11 3982 vaesenc %xmm10,%xmm9,%xmm9 3983 vmovdqu 224-128(%rdi),%xmm10 3984 xorl %r12d,%r14d 3985 xorl %edx,%esi 3986 xorl %r13d,%r14d 3987 leal (%rbx,%rsi,1),%ebx 3988 movl %r10d,%r12d 3989 addl 44+16(%rbp),%eax 3990 andl %r9d,%r12d 3991 rorxl $25,%r9d,%r13d 3992 rorxl $11,%r9d,%esi 3993 leal (%rbx,%r14,1),%ebx 3994 leal (%rax,%r12,1),%eax 3995 andnl %r11d,%r9d,%r12d 3996 xorl %esi,%r13d 3997 rorxl $6,%r9d,%r14d 3998 leal (%rax,%r12,1),%eax 3999 xorl %r14d,%r13d 4000 movl %ebx,%esi 4001 rorxl $22,%ebx,%r12d 4002 leal (%rax,%r13,1),%eax 4003 xorl %ecx,%esi 4004 rorxl $13,%ebx,%r14d 4005 rorxl $2,%ebx,%r13d 4006 leal (%r8,%rax,1),%r8d 4007 andl %esi,%r15d 4008 vpor %xmm11,%xmm8,%xmm8 4009 vaesenclast %xmm10,%xmm9,%xmm11 4010 vmovdqu 0-128(%rdi),%xmm10 4011 xorl %r12d,%r14d 4012 xorl %ecx,%r15d 4013 xorl %r13d,%r14d 4014 leal (%rax,%r15,1),%eax 4015 movl %r9d,%r12d 4016 vmovq %xmm15,%r13 4017 vpextrq $1,%xmm15,%r15 4018 vpand %xmm14,%xmm11,%xmm11 4019 vpor %xmm11,%xmm8,%xmm8 4020 leaq -64(%rbp),%rbp 4021 vmovdqu %xmm8,(%r15,%r13,1) 4022 leaq 16(%r13),%r13 4023 cmpq %rsp,%rbp 4024 jae .Lower_avx2 4025 4026 movq 552(%rsp),%r15 4027 leaq 64(%r13),%r13 4028 movq 560(%rsp),%rsi 4029 addl %r14d,%eax 4030 leaq 448(%rsp),%rsp 4031 4032 addl 0(%r15),%eax 4033 addl 4(%r15),%ebx 4034 addl 8(%r15),%ecx 4035 addl 12(%r15),%edx 4036 addl 16(%r15),%r8d 4037 addl 20(%r15),%r9d 4038 addl 24(%r15),%r10d 4039 leaq (%rsi,%r13,1),%r12 4040 addl 28(%r15),%r11d 4041 4042 cmpq 64+16(%rsp),%r13 4043 4044 movl %eax,0(%r15) 4045 cmoveq %rsp,%r12 4046 movl %ebx,4(%r15) 4047 movl %ecx,8(%r15) 4048 movl %edx,12(%r15) 4049 movl %r8d,16(%r15) 4050 movl %r9d,20(%r15) 4051 movl %r10d,24(%r15) 4052 movl %r11d,28(%r15) 4053 4054 jbe .Loop_avx2 4055 leaq (%rsp),%rbp 4056 4057 4058 .cfi_escape 0x0f,0x06,0x76,0xf8,0x00,0x06,0x23,0x08 4059 4060 .Ldone_avx2: 4061 movq 64+32(%rbp),%r8 4062 movq 64+56(%rbp),%rsi 4063 .cfi_def_cfa %rsi,8 4064 vmovdqu %xmm8,(%r8) 4065 vzeroall 4066 movq -48(%rsi),%r15 4067 .cfi_restore %r15 4068 movq -40(%rsi),%r14 4069 .cfi_restore %r14 4070 movq -32(%rsi),%r13 4071 .cfi_restore %r13 4072 movq -24(%rsi),%r12 4073 .cfi_restore %r12 4074 movq -16(%rsi),%rbp 4075 .cfi_restore %rbp 4076 movq -8(%rsi),%rbx 4077 .cfi_restore %rbx 4078 leaq (%rsi),%rsp 4079 .cfi_def_cfa_register %rsp 4080 .Lepilogue_avx2: 4081 .byte 0xf3,0xc3 4082 .cfi_endproc 4083 .size aesni_cbc_sha256_enc_avx2,.-aesni_cbc_sha256_enc_avx2 4084 .type aesni_cbc_sha256_enc_shaext,@function 4085 .align 32 4086 aesni_cbc_sha256_enc_shaext: 4087 .cfi_startproc 4088 movq 8(%rsp),%r10 4089 leaq K256+128(%rip),%rax 4090 movdqu (%r9),%xmm1 4091 movdqu 16(%r9),%xmm2 4092 movdqa 512-128(%rax),%xmm3 4093 4094 movl 240(%rcx),%r11d 4095 subq %rdi,%rsi 4096 movups (%rcx),%xmm15 4097 movups (%r8),%xmm6 4098 movups 16(%rcx),%xmm4 4099 leaq 112(%rcx),%rcx 4100 4101 pshufd $0x1b,%xmm1,%xmm0 4102 pshufd $0xb1,%xmm1,%xmm1 4103 pshufd $0x1b,%xmm2,%xmm2 4104 movdqa %xmm3,%xmm7 4105 .byte 102,15,58,15,202,8 4106 punpcklqdq %xmm0,%xmm2 4107 4108 jmp .Loop_shaext 4109 4110 .align 16 4111 .Loop_shaext: 4112 movdqu (%r10),%xmm10 4113 movdqu 16(%r10),%xmm11 4114 movdqu 32(%r10),%xmm12 4115 .byte 102,68,15,56,0,211 4116 movdqu 48(%r10),%xmm13 4117 4118 movdqa 0-128(%rax),%xmm0 4119 paddd %xmm10,%xmm0 4120 .byte 102,68,15,56,0,219 4121 movdqa %xmm2,%xmm9 4122 movdqa %xmm1,%xmm8 4123 movups 0(%rdi),%xmm14 4124 xorps %xmm15,%xmm14 4125 xorps %xmm14,%xmm6 4126 movups -80(%rcx),%xmm5 4127 aesenc %xmm4,%xmm6 4128 .byte 15,56,203,209 4129 pshufd $0x0e,%xmm0,%xmm0 4130 movups -64(%rcx),%xmm4 4131 aesenc %xmm5,%xmm6 4132 .byte 15,56,203,202 4133 4134 movdqa 32-128(%rax),%xmm0 4135 paddd %xmm11,%xmm0 4136 .byte 102,68,15,56,0,227 4137 leaq 64(%r10),%r10 4138 movups -48(%rcx),%xmm5 4139 aesenc %xmm4,%xmm6 4140 .byte 15,56,203,209 4141 pshufd $0x0e,%xmm0,%xmm0 4142 movups -32(%rcx),%xmm4 4143 aesenc %xmm5,%xmm6 4144 .byte 15,56,203,202 4145 4146 movdqa 64-128(%rax),%xmm0 4147 paddd %xmm12,%xmm0 4148 .byte 102,68,15,56,0,235 4149 .byte 69,15,56,204,211 4150 movups -16(%rcx),%xmm5 4151 aesenc %xmm4,%xmm6 4152 .byte 15,56,203,209 4153 pshufd $0x0e,%xmm0,%xmm0 4154 movdqa %xmm13,%xmm3 4155 .byte 102,65,15,58,15,220,4 4156 paddd %xmm3,%xmm10 4157 movups 0(%rcx),%xmm4 4158 aesenc %xmm5,%xmm6 4159 .byte 15,56,203,202 4160 4161 movdqa 96-128(%rax),%xmm0 4162 paddd %xmm13,%xmm0 4163 .byte 69,15,56,205,213 4164 .byte 69,15,56,204,220 4165 movups 16(%rcx),%xmm5 4166 aesenc %xmm4,%xmm6 4167 .byte 15,56,203,209 4168 pshufd $0x0e,%xmm0,%xmm0 4169 movups 32(%rcx),%xmm4 4170 aesenc %xmm5,%xmm6 4171 movdqa %xmm10,%xmm3 4172 .byte 102,65,15,58,15,221,4 4173 paddd %xmm3,%xmm11 4174 .byte 15,56,203,202 4175 movdqa 128-128(%rax),%xmm0 4176 paddd %xmm10,%xmm0 4177 .byte 69,15,56,205,218 4178 .byte 69,15,56,204,229 4179 movups 48(%rcx),%xmm5 4180 aesenc %xmm4,%xmm6 4181 .byte 15,56,203,209 4182 pshufd $0x0e,%xmm0,%xmm0 4183 movdqa %xmm11,%xmm3 4184 .byte 102,65,15,58,15,218,4 4185 paddd %xmm3,%xmm12 4186 cmpl $11,%r11d 4187 jb .Laesenclast1 4188 movups 64(%rcx),%xmm4 4189 aesenc %xmm5,%xmm6 4190 movups 80(%rcx),%xmm5 4191 aesenc %xmm4,%xmm6 4192 je .Laesenclast1 4193 movups 96(%rcx),%xmm4 4194 aesenc %xmm5,%xmm6 4195 movups 112(%rcx),%xmm5 4196 aesenc %xmm4,%xmm6 4197 .Laesenclast1: 4198 aesenclast %xmm5,%xmm6 4199 movups 16-112(%rcx),%xmm4 4200 nop 4201 .byte 15,56,203,202 4202 movups 16(%rdi),%xmm14 4203 xorps %xmm15,%xmm14 4204 movups %xmm6,0(%rsi,%rdi,1) 4205 xorps %xmm14,%xmm6 4206 movups -80(%rcx),%xmm5 4207 aesenc %xmm4,%xmm6 4208 movdqa 160-128(%rax),%xmm0 4209 paddd %xmm11,%xmm0 4210 .byte 69,15,56,205,227 4211 .byte 69,15,56,204,234 4212 movups -64(%rcx),%xmm4 4213 aesenc %xmm5,%xmm6 4214 .byte 15,56,203,209 4215 pshufd $0x0e,%xmm0,%xmm0 4216 movdqa %xmm12,%xmm3 4217 .byte 102,65,15,58,15,219,4 4218 paddd %xmm3,%xmm13 4219 movups -48(%rcx),%xmm5 4220 aesenc %xmm4,%xmm6 4221 .byte 15,56,203,202 4222 movdqa 192-128(%rax),%xmm0 4223 paddd %xmm12,%xmm0 4224 .byte 69,15,56,205,236 4225 .byte 69,15,56,204,211 4226 movups -32(%rcx),%xmm4 4227 aesenc %xmm5,%xmm6 4228 .byte 15,56,203,209 4229 pshufd $0x0e,%xmm0,%xmm0 4230 movdqa %xmm13,%xmm3 4231 .byte 102,65,15,58,15,220,4 4232 paddd %xmm3,%xmm10 4233 movups -16(%rcx),%xmm5 4234 aesenc %xmm4,%xmm6 4235 .byte 15,56,203,202 4236 movdqa 224-128(%rax),%xmm0 4237 paddd %xmm13,%xmm0 4238 .byte 69,15,56,205,213 4239 .byte 69,15,56,204,220 4240 movups 0(%rcx),%xmm4 4241 aesenc %xmm5,%xmm6 4242 .byte 15,56,203,209 4243 pshufd $0x0e,%xmm0,%xmm0 4244 movdqa %xmm10,%xmm3 4245 .byte 102,65,15,58,15,221,4 4246 paddd %xmm3,%xmm11 4247 movups 16(%rcx),%xmm5 4248 aesenc %xmm4,%xmm6 4249 .byte 15,56,203,202 4250 movdqa 256-128(%rax),%xmm0 4251 paddd %xmm10,%xmm0 4252 .byte 69,15,56,205,218 4253 .byte 69,15,56,204,229 4254 movups 32(%rcx),%xmm4 4255 aesenc %xmm5,%xmm6 4256 .byte 15,56,203,209 4257 pshufd $0x0e,%xmm0,%xmm0 4258 movdqa %xmm11,%xmm3 4259 .byte 102,65,15,58,15,218,4 4260 paddd %xmm3,%xmm12 4261 movups 48(%rcx),%xmm5 4262 aesenc %xmm4,%xmm6 4263 cmpl $11,%r11d 4264 jb .Laesenclast2 4265 movups 64(%rcx),%xmm4 4266 aesenc %xmm5,%xmm6 4267 movups 80(%rcx),%xmm5 4268 aesenc %xmm4,%xmm6 4269 je .Laesenclast2 4270 movups 96(%rcx),%xmm4 4271 aesenc %xmm5,%xmm6 4272 movups 112(%rcx),%xmm5 4273 aesenc %xmm4,%xmm6 4274 .Laesenclast2: 4275 aesenclast %xmm5,%xmm6 4276 movups 16-112(%rcx),%xmm4 4277 nop 4278 .byte 15,56,203,202 4279 movups 32(%rdi),%xmm14 4280 xorps %xmm15,%xmm14 4281 movups %xmm6,16(%rsi,%rdi,1) 4282 xorps %xmm14,%xmm6 4283 movups -80(%rcx),%xmm5 4284 aesenc %xmm4,%xmm6 4285 movdqa 288-128(%rax),%xmm0 4286 paddd %xmm11,%xmm0 4287 .byte 69,15,56,205,227 4288 .byte 69,15,56,204,234 4289 movups -64(%rcx),%xmm4 4290 aesenc %xmm5,%xmm6 4291 .byte 15,56,203,209 4292 pshufd $0x0e,%xmm0,%xmm0 4293 movdqa %xmm12,%xmm3 4294 .byte 102,65,15,58,15,219,4 4295 paddd %xmm3,%xmm13 4296 movups -48(%rcx),%xmm5 4297 aesenc %xmm4,%xmm6 4298 .byte 15,56,203,202 4299 movdqa 320-128(%rax),%xmm0 4300 paddd %xmm12,%xmm0 4301 .byte 69,15,56,205,236 4302 .byte 69,15,56,204,211 4303 movups -32(%rcx),%xmm4 4304 aesenc %xmm5,%xmm6 4305 .byte 15,56,203,209 4306 pshufd $0x0e,%xmm0,%xmm0 4307 movdqa %xmm13,%xmm3 4308 .byte 102,65,15,58,15,220,4 4309 paddd %xmm3,%xmm10 4310 movups -16(%rcx),%xmm5 4311 aesenc %xmm4,%xmm6 4312 .byte 15,56,203,202 4313 movdqa 352-128(%rax),%xmm0 4314 paddd %xmm13,%xmm0 4315 .byte 69,15,56,205,213 4316 .byte 69,15,56,204,220 4317 movups 0(%rcx),%xmm4 4318 aesenc %xmm5,%xmm6 4319 .byte 15,56,203,209 4320 pshufd $0x0e,%xmm0,%xmm0 4321 movdqa %xmm10,%xmm3 4322 .byte 102,65,15,58,15,221,4 4323 paddd %xmm3,%xmm11 4324 movups 16(%rcx),%xmm5 4325 aesenc %xmm4,%xmm6 4326 .byte 15,56,203,202 4327 movdqa 384-128(%rax),%xmm0 4328 paddd %xmm10,%xmm0 4329 .byte 69,15,56,205,218 4330 .byte 69,15,56,204,229 4331 movups 32(%rcx),%xmm4 4332 aesenc %xmm5,%xmm6 4333 .byte 15,56,203,209 4334 pshufd $0x0e,%xmm0,%xmm0 4335 movdqa %xmm11,%xmm3 4336 .byte 102,65,15,58,15,218,4 4337 paddd %xmm3,%xmm12 4338 movups 48(%rcx),%xmm5 4339 aesenc %xmm4,%xmm6 4340 .byte 15,56,203,202 4341 movdqa 416-128(%rax),%xmm0 4342 paddd %xmm11,%xmm0 4343 .byte 69,15,56,205,227 4344 .byte 69,15,56,204,234 4345 cmpl $11,%r11d 4346 jb .Laesenclast3 4347 movups 64(%rcx),%xmm4 4348 aesenc %xmm5,%xmm6 4349 movups 80(%rcx),%xmm5 4350 aesenc %xmm4,%xmm6 4351 je .Laesenclast3 4352 movups 96(%rcx),%xmm4 4353 aesenc %xmm5,%xmm6 4354 movups 112(%rcx),%xmm5 4355 aesenc %xmm4,%xmm6 4356 .Laesenclast3: 4357 aesenclast %xmm5,%xmm6 4358 movups 16-112(%rcx),%xmm4 4359 nop 4360 .byte 15,56,203,209 4361 pshufd $0x0e,%xmm0,%xmm0 4362 movdqa %xmm12,%xmm3 4363 .byte 102,65,15,58,15,219,4 4364 paddd %xmm3,%xmm13 4365 movups 48(%rdi),%xmm14 4366 xorps %xmm15,%xmm14 4367 movups %xmm6,32(%rsi,%rdi,1) 4368 xorps %xmm14,%xmm6 4369 movups -80(%rcx),%xmm5 4370 aesenc %xmm4,%xmm6 4371 movups -64(%rcx),%xmm4 4372 aesenc %xmm5,%xmm6 4373 .byte 15,56,203,202 4374 4375 movdqa 448-128(%rax),%xmm0 4376 paddd %xmm12,%xmm0 4377 .byte 69,15,56,205,236 4378 movdqa %xmm7,%xmm3 4379 movups -48(%rcx),%xmm5 4380 aesenc %xmm4,%xmm6 4381 .byte 15,56,203,209 4382 pshufd $0x0e,%xmm0,%xmm0 4383 movups -32(%rcx),%xmm4 4384 aesenc %xmm5,%xmm6 4385 .byte 15,56,203,202 4386 4387 movdqa 480-128(%rax),%xmm0 4388 paddd %xmm13,%xmm0 4389 movups -16(%rcx),%xmm5 4390 aesenc %xmm4,%xmm6 4391 movups 0(%rcx),%xmm4 4392 aesenc %xmm5,%xmm6 4393 .byte 15,56,203,209 4394 pshufd $0x0e,%xmm0,%xmm0 4395 movups 16(%rcx),%xmm5 4396 aesenc %xmm4,%xmm6 4397 .byte 15,56,203,202 4398 4399 movups 32(%rcx),%xmm4 4400 aesenc %xmm5,%xmm6 4401 movups 48(%rcx),%xmm5 4402 aesenc %xmm4,%xmm6 4403 cmpl $11,%r11d 4404 jb .Laesenclast4 4405 movups 64(%rcx),%xmm4 4406 aesenc %xmm5,%xmm6 4407 movups 80(%rcx),%xmm5 4408 aesenc %xmm4,%xmm6 4409 je .Laesenclast4 4410 movups 96(%rcx),%xmm4 4411 aesenc %xmm5,%xmm6 4412 movups 112(%rcx),%xmm5 4413 aesenc %xmm4,%xmm6 4414 .Laesenclast4: 4415 aesenclast %xmm5,%xmm6 4416 movups 16-112(%rcx),%xmm4 4417 nop 4418 4419 paddd %xmm9,%xmm2 4420 paddd %xmm8,%xmm1 4421 4422 decq %rdx 4423 movups %xmm6,48(%rsi,%rdi,1) 4424 leaq 64(%rdi),%rdi 4425 jnz .Loop_shaext 4426 4427 pshufd $0xb1,%xmm2,%xmm2 4428 pshufd $0x1b,%xmm1,%xmm3 4429 pshufd $0xb1,%xmm1,%xmm1 4430 punpckhqdq %xmm2,%xmm1 4431 .byte 102,15,58,15,211,8 4432 4433 movups %xmm6,(%r8) 4434 movdqu %xmm1,(%r9) 4435 movdqu %xmm2,16(%r9) 4436 .byte 0xf3,0xc3 4437 .cfi_endproc 4438 .size aesni_cbc_sha256_enc_shaext,.-aesni_cbc_sha256_enc_shaext 4439 .section ".note.gnu.property", "a" 4440 .p2align 3 4441 .long 1f - 0f 4442 .long 4f - 1f 4443 .long 5 4444 0: 4445 # "GNU" encoded with .byte, since .asciz isn't supported 4446 # on Solaris. 4447 .byte 0x47 4448 .byte 0x4e 4449 .byte 0x55 4450 .byte 0 4451 1: 4452 .p2align 3 4453 .long 0xc0000002 4454 .long 3f - 2f 4455 2: 4456 .long 3 4457 3: 4458 .p2align 3 4459 4: 4460
Indexes created Sat Feb 28 05:31:39 UTC 2026