OLD | NEW |
(Empty) | |
| 1 #if defined(__x86_64__) |
| 2 .text |
| 3 |
| 4 |
| 5 |
| 6 .p2align 6 |
| 7 L$zero: |
| 8 .long 0,0,0,0 |
| 9 L$one: |
| 10 .long 1,0,0,0 |
| 11 L$inc: |
| 12 .long 0,1,2,3 |
| 13 L$four: |
| 14 .long 4,4,4,4 |
| 15 L$incy: |
| 16 .long 0,2,4,6,1,3,5,7 |
| 17 L$eight: |
| 18 .long 8,8,8,8,8,8,8,8 |
| 19 L$rot16: |
| 20 .byte 0x2,0x3,0x0,0x1, 0x6,0x7,0x4,0x5, 0xa,0xb,0x8,0x9, 0xe,0xf,0xc,0xd |
| 21 L$rot24: |
| 22 .byte 0x3,0x0,0x1,0x2, 0x7,0x4,0x5,0x6, 0xb,0x8,0x9,0xa, 0xf,0xc,0xd,0xe |
| 23 L$sigma: |
| 24 .byte 101,120,112,97,110,100,32,51,50,45,98,121,116,101,32,107,0 |
| 25 .byte 67,104,97,67,104,97,50,48,32,102,111,114,32,120,56,54,95,54,52,44,32,67,
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 |
| 26 .globl _ChaCha20_ctr32 |
| 27 .private_extern _ChaCha20_ctr32 |
| 28 |
| 29 .p2align 6 |
| 30 _ChaCha20_ctr32: |
| 31 cmpq $0,%rdx |
| 32 je L$no_data |
| 33 movq _OPENSSL_ia32cap_P+4(%rip),%r10 |
| 34 testl $512,%r10d |
| 35 jnz L$ChaCha20_ssse3 |
| 36 |
| 37 pushq %rbx |
| 38 pushq %rbp |
| 39 pushq %r12 |
| 40 pushq %r13 |
| 41 pushq %r14 |
| 42 pushq %r15 |
| 43 subq $64+24,%rsp |
| 44 |
| 45 |
| 46 movdqu (%rcx),%xmm1 |
| 47 movdqu 16(%rcx),%xmm2 |
| 48 movdqu (%r8),%xmm3 |
| 49 movdqa L$one(%rip),%xmm4 |
| 50 |
| 51 |
| 52 movdqa %xmm1,16(%rsp) |
| 53 movdqa %xmm2,32(%rsp) |
| 54 movdqa %xmm3,48(%rsp) |
| 55 movq %rdx,%rbp |
| 56 jmp L$oop_outer |
| 57 |
| 58 .p2align 5 |
| 59 L$oop_outer: |
| 60 movl $0x61707865,%eax |
| 61 movl $0x3320646e,%ebx |
| 62 movl $0x79622d32,%ecx |
| 63 movl $0x6b206574,%edx |
| 64 movl 16(%rsp),%r8d |
| 65 movl 20(%rsp),%r9d |
| 66 movl 24(%rsp),%r10d |
| 67 movl 28(%rsp),%r11d |
| 68 movd %xmm3,%r12d |
| 69 movl 52(%rsp),%r13d |
| 70 movl 56(%rsp),%r14d |
| 71 movl 60(%rsp),%r15d |
| 72 |
| 73 movq %rbp,64+0(%rsp) |
| 74 movl $10,%ebp |
| 75 movq %rsi,64+8(%rsp) |
| 76 .byte 102,72,15,126,214 |
| 77 movq %rdi,64+16(%rsp) |
| 78 movq %rsi,%rdi |
| 79 shrq $32,%rdi |
| 80 jmp L$oop |
| 81 |
| 82 .p2align 5 |
| 83 L$oop: |
| 84 addl %r8d,%eax |
| 85 xorl %eax,%r12d |
| 86 roll $16,%r12d |
| 87 addl %r9d,%ebx |
| 88 xorl %ebx,%r13d |
| 89 roll $16,%r13d |
| 90 addl %r12d,%esi |
| 91 xorl %esi,%r8d |
| 92 roll $12,%r8d |
| 93 addl %r13d,%edi |
| 94 xorl %edi,%r9d |
| 95 roll $12,%r9d |
| 96 addl %r8d,%eax |
| 97 xorl %eax,%r12d |
| 98 roll $8,%r12d |
| 99 addl %r9d,%ebx |
| 100 xorl %ebx,%r13d |
| 101 roll $8,%r13d |
| 102 addl %r12d,%esi |
| 103 xorl %esi,%r8d |
| 104 roll $7,%r8d |
| 105 addl %r13d,%edi |
| 106 xorl %edi,%r9d |
| 107 roll $7,%r9d |
| 108 movl %esi,32(%rsp) |
| 109 movl %edi,36(%rsp) |
| 110 movl 40(%rsp),%esi |
| 111 movl 44(%rsp),%edi |
| 112 addl %r10d,%ecx |
| 113 xorl %ecx,%r14d |
| 114 roll $16,%r14d |
| 115 addl %r11d,%edx |
| 116 xorl %edx,%r15d |
| 117 roll $16,%r15d |
| 118 addl %r14d,%esi |
| 119 xorl %esi,%r10d |
| 120 roll $12,%r10d |
| 121 addl %r15d,%edi |
| 122 xorl %edi,%r11d |
| 123 roll $12,%r11d |
| 124 addl %r10d,%ecx |
| 125 xorl %ecx,%r14d |
| 126 roll $8,%r14d |
| 127 addl %r11d,%edx |
| 128 xorl %edx,%r15d |
| 129 roll $8,%r15d |
| 130 addl %r14d,%esi |
| 131 xorl %esi,%r10d |
| 132 roll $7,%r10d |
| 133 addl %r15d,%edi |
| 134 xorl %edi,%r11d |
| 135 roll $7,%r11d |
| 136 addl %r9d,%eax |
| 137 xorl %eax,%r15d |
| 138 roll $16,%r15d |
| 139 addl %r10d,%ebx |
| 140 xorl %ebx,%r12d |
| 141 roll $16,%r12d |
| 142 addl %r15d,%esi |
| 143 xorl %esi,%r9d |
| 144 roll $12,%r9d |
| 145 addl %r12d,%edi |
| 146 xorl %edi,%r10d |
| 147 roll $12,%r10d |
| 148 addl %r9d,%eax |
| 149 xorl %eax,%r15d |
| 150 roll $8,%r15d |
| 151 addl %r10d,%ebx |
| 152 xorl %ebx,%r12d |
| 153 roll $8,%r12d |
| 154 addl %r15d,%esi |
| 155 xorl %esi,%r9d |
| 156 roll $7,%r9d |
| 157 addl %r12d,%edi |
| 158 xorl %edi,%r10d |
| 159 roll $7,%r10d |
| 160 movl %esi,40(%rsp) |
| 161 movl %edi,44(%rsp) |
| 162 movl 32(%rsp),%esi |
| 163 movl 36(%rsp),%edi |
| 164 addl %r11d,%ecx |
| 165 xorl %ecx,%r13d |
| 166 roll $16,%r13d |
| 167 addl %r8d,%edx |
| 168 xorl %edx,%r14d |
| 169 roll $16,%r14d |
| 170 addl %r13d,%esi |
| 171 xorl %esi,%r11d |
| 172 roll $12,%r11d |
| 173 addl %r14d,%edi |
| 174 xorl %edi,%r8d |
| 175 roll $12,%r8d |
| 176 addl %r11d,%ecx |
| 177 xorl %ecx,%r13d |
| 178 roll $8,%r13d |
| 179 addl %r8d,%edx |
| 180 xorl %edx,%r14d |
| 181 roll $8,%r14d |
| 182 addl %r13d,%esi |
| 183 xorl %esi,%r11d |
| 184 roll $7,%r11d |
| 185 addl %r14d,%edi |
| 186 xorl %edi,%r8d |
| 187 roll $7,%r8d |
| 188 decl %ebp |
| 189 jnz L$oop |
| 190 movl %edi,36(%rsp) |
| 191 movl %esi,32(%rsp) |
| 192 movq 64(%rsp),%rbp |
| 193 movdqa %xmm2,%xmm1 |
| 194 movq 64+8(%rsp),%rsi |
| 195 paddd %xmm4,%xmm3 |
| 196 movq 64+16(%rsp),%rdi |
| 197 |
| 198 addl $0x61707865,%eax |
| 199 addl $0x3320646e,%ebx |
| 200 addl $0x79622d32,%ecx |
| 201 addl $0x6b206574,%edx |
| 202 addl 16(%rsp),%r8d |
| 203 addl 20(%rsp),%r9d |
| 204 addl 24(%rsp),%r10d |
| 205 addl 28(%rsp),%r11d |
| 206 addl 48(%rsp),%r12d |
| 207 addl 52(%rsp),%r13d |
| 208 addl 56(%rsp),%r14d |
| 209 addl 60(%rsp),%r15d |
| 210 paddd 32(%rsp),%xmm1 |
| 211 |
| 212 cmpq $64,%rbp |
| 213 jb L$tail |
| 214 |
| 215 xorl 0(%rsi),%eax |
| 216 xorl 4(%rsi),%ebx |
| 217 xorl 8(%rsi),%ecx |
| 218 xorl 12(%rsi),%edx |
| 219 xorl 16(%rsi),%r8d |
| 220 xorl 20(%rsi),%r9d |
| 221 xorl 24(%rsi),%r10d |
| 222 xorl 28(%rsi),%r11d |
| 223 movdqu 32(%rsi),%xmm0 |
| 224 xorl 48(%rsi),%r12d |
| 225 xorl 52(%rsi),%r13d |
| 226 xorl 56(%rsi),%r14d |
| 227 xorl 60(%rsi),%r15d |
| 228 leaq 64(%rsi),%rsi |
| 229 pxor %xmm1,%xmm0 |
| 230 |
| 231 movdqa %xmm2,32(%rsp) |
| 232 movd %xmm3,48(%rsp) |
| 233 |
| 234 movl %eax,0(%rdi) |
| 235 movl %ebx,4(%rdi) |
| 236 movl %ecx,8(%rdi) |
| 237 movl %edx,12(%rdi) |
| 238 movl %r8d,16(%rdi) |
| 239 movl %r9d,20(%rdi) |
| 240 movl %r10d,24(%rdi) |
| 241 movl %r11d,28(%rdi) |
| 242 movdqu %xmm0,32(%rdi) |
| 243 movl %r12d,48(%rdi) |
| 244 movl %r13d,52(%rdi) |
| 245 movl %r14d,56(%rdi) |
| 246 movl %r15d,60(%rdi) |
| 247 leaq 64(%rdi),%rdi |
| 248 |
| 249 subq $64,%rbp |
| 250 jnz L$oop_outer |
| 251 |
| 252 jmp L$done |
| 253 |
| 254 .p2align 4 |
| 255 L$tail: |
| 256 movl %eax,0(%rsp) |
| 257 movl %ebx,4(%rsp) |
| 258 xorq %rbx,%rbx |
| 259 movl %ecx,8(%rsp) |
| 260 movl %edx,12(%rsp) |
| 261 movl %r8d,16(%rsp) |
| 262 movl %r9d,20(%rsp) |
| 263 movl %r10d,24(%rsp) |
| 264 movl %r11d,28(%rsp) |
| 265 movdqa %xmm1,32(%rsp) |
| 266 movl %r12d,48(%rsp) |
| 267 movl %r13d,52(%rsp) |
| 268 movl %r14d,56(%rsp) |
| 269 movl %r15d,60(%rsp) |
| 270 |
| 271 L$oop_tail: |
| 272 movzbl (%rsi,%rbx,1),%eax |
| 273 movzbl (%rsp,%rbx,1),%edx |
| 274 leaq 1(%rbx),%rbx |
| 275 xorl %edx,%eax |
| 276 movb %al,-1(%rdi,%rbx,1) |
| 277 decq %rbp |
| 278 jnz L$oop_tail |
| 279 |
| 280 L$done: |
| 281 addq $64+24,%rsp |
| 282 popq %r15 |
| 283 popq %r14 |
| 284 popq %r13 |
| 285 popq %r12 |
| 286 popq %rbp |
| 287 popq %rbx |
| 288 L$no_data: |
| 289 .byte 0xf3,0xc3 |
| 290 |
| 291 |
| 292 .p2align 5 |
| 293 ChaCha20_ssse3: |
| 294 L$ChaCha20_ssse3: |
| 295 cmpq $128,%rdx |
| 296 ja L$ChaCha20_4x |
| 297 |
| 298 L$do_sse3_after_all: |
| 299 pushq %rbx |
| 300 pushq %rbp |
| 301 pushq %r12 |
| 302 pushq %r13 |
| 303 pushq %r14 |
| 304 pushq %r15 |
| 305 |
| 306 subq $64+24,%rsp |
| 307 movdqa L$sigma(%rip),%xmm0 |
| 308 movdqu (%rcx),%xmm1 |
| 309 movdqu 16(%rcx),%xmm2 |
| 310 movdqu (%r8),%xmm3 |
| 311 movdqa L$rot16(%rip),%xmm6 |
| 312 movdqa L$rot24(%rip),%xmm7 |
| 313 |
| 314 movdqa %xmm0,0(%rsp) |
| 315 movdqa %xmm1,16(%rsp) |
| 316 movdqa %xmm2,32(%rsp) |
| 317 movdqa %xmm3,48(%rsp) |
| 318 movl $10,%ebp |
| 319 jmp L$oop_ssse3 |
| 320 |
| 321 .p2align 5 |
| 322 L$oop_outer_ssse3: |
| 323 movdqa L$one(%rip),%xmm3 |
| 324 movdqa 0(%rsp),%xmm0 |
| 325 movdqa 16(%rsp),%xmm1 |
| 326 movdqa 32(%rsp),%xmm2 |
| 327 paddd 48(%rsp),%xmm3 |
| 328 movl $10,%ebp |
| 329 movdqa %xmm3,48(%rsp) |
| 330 jmp L$oop_ssse3 |
| 331 |
| 332 .p2align 5 |
| 333 L$oop_ssse3: |
| 334 paddd %xmm1,%xmm0 |
| 335 pxor %xmm0,%xmm3 |
| 336 .byte 102,15,56,0,222 |
| 337 paddd %xmm3,%xmm2 |
| 338 pxor %xmm2,%xmm1 |
| 339 movdqa %xmm1,%xmm4 |
| 340 psrld $20,%xmm1 |
| 341 pslld $12,%xmm4 |
| 342 por %xmm4,%xmm1 |
| 343 paddd %xmm1,%xmm0 |
| 344 pxor %xmm0,%xmm3 |
| 345 .byte 102,15,56,0,223 |
| 346 paddd %xmm3,%xmm2 |
| 347 pxor %xmm2,%xmm1 |
| 348 movdqa %xmm1,%xmm4 |
| 349 psrld $25,%xmm1 |
| 350 pslld $7,%xmm4 |
| 351 por %xmm4,%xmm1 |
| 352 pshufd $78,%xmm2,%xmm2 |
| 353 pshufd $57,%xmm1,%xmm1 |
| 354 pshufd $147,%xmm3,%xmm3 |
| 355 nop |
| 356 paddd %xmm1,%xmm0 |
| 357 pxor %xmm0,%xmm3 |
| 358 .byte 102,15,56,0,222 |
| 359 paddd %xmm3,%xmm2 |
| 360 pxor %xmm2,%xmm1 |
| 361 movdqa %xmm1,%xmm4 |
| 362 psrld $20,%xmm1 |
| 363 pslld $12,%xmm4 |
| 364 por %xmm4,%xmm1 |
| 365 paddd %xmm1,%xmm0 |
| 366 pxor %xmm0,%xmm3 |
| 367 .byte 102,15,56,0,223 |
| 368 paddd %xmm3,%xmm2 |
| 369 pxor %xmm2,%xmm1 |
| 370 movdqa %xmm1,%xmm4 |
| 371 psrld $25,%xmm1 |
| 372 pslld $7,%xmm4 |
| 373 por %xmm4,%xmm1 |
| 374 pshufd $78,%xmm2,%xmm2 |
| 375 pshufd $147,%xmm1,%xmm1 |
| 376 pshufd $57,%xmm3,%xmm3 |
| 377 decl %ebp |
| 378 jnz L$oop_ssse3 |
| 379 paddd 0(%rsp),%xmm0 |
| 380 paddd 16(%rsp),%xmm1 |
| 381 paddd 32(%rsp),%xmm2 |
| 382 paddd 48(%rsp),%xmm3 |
| 383 |
| 384 cmpq $64,%rdx |
| 385 jb L$tail_ssse3 |
| 386 |
| 387 movdqu 0(%rsi),%xmm4 |
| 388 movdqu 16(%rsi),%xmm5 |
| 389 pxor %xmm4,%xmm0 |
| 390 movdqu 32(%rsi),%xmm4 |
| 391 pxor %xmm5,%xmm1 |
| 392 movdqu 48(%rsi),%xmm5 |
| 393 leaq 64(%rsi),%rsi |
| 394 pxor %xmm4,%xmm2 |
| 395 pxor %xmm5,%xmm3 |
| 396 |
| 397 movdqu %xmm0,0(%rdi) |
| 398 movdqu %xmm1,16(%rdi) |
| 399 movdqu %xmm2,32(%rdi) |
| 400 movdqu %xmm3,48(%rdi) |
| 401 leaq 64(%rdi),%rdi |
| 402 |
| 403 subq $64,%rdx |
| 404 jnz L$oop_outer_ssse3 |
| 405 |
| 406 jmp L$done_ssse3 |
| 407 |
| 408 .p2align 4 |
| 409 L$tail_ssse3: |
| 410 movdqa %xmm0,0(%rsp) |
| 411 movdqa %xmm1,16(%rsp) |
| 412 movdqa %xmm2,32(%rsp) |
| 413 movdqa %xmm3,48(%rsp) |
| 414 xorq %rbx,%rbx |
| 415 |
| 416 L$oop_tail_ssse3: |
| 417 movzbl (%rsi,%rbx,1),%eax |
| 418 movzbl (%rsp,%rbx,1),%ecx |
| 419 leaq 1(%rbx),%rbx |
| 420 xorl %ecx,%eax |
| 421 movb %al,-1(%rdi,%rbx,1) |
| 422 decq %rdx |
| 423 jnz L$oop_tail_ssse3 |
| 424 |
| 425 L$done_ssse3: |
| 426 addq $64+24,%rsp |
| 427 popq %r15 |
| 428 popq %r14 |
| 429 popq %r13 |
| 430 popq %r12 |
| 431 popq %rbp |
| 432 popq %rbx |
| 433 .byte 0xf3,0xc3 |
| 434 |
| 435 |
| 436 .p2align 5 |
| 437 ChaCha20_4x: |
| 438 L$ChaCha20_4x: |
| 439 movq %r10,%r11 |
| 440 shrq $32,%r10 |
| 441 testq $32,%r10 |
| 442 jnz L$ChaCha20_8x |
| 443 cmpq $192,%rdx |
| 444 ja L$proceed4x |
| 445 |
| 446 andq $71303168,%r11 |
| 447 cmpq $4194304,%r11 |
| 448 je L$do_sse3_after_all |
| 449 |
| 450 L$proceed4x: |
| 451 leaq -120(%rsp),%r11 |
| 452 subq $0x148+0,%rsp |
| 453 movdqa L$sigma(%rip),%xmm11 |
| 454 movdqu (%rcx),%xmm15 |
| 455 movdqu 16(%rcx),%xmm7 |
| 456 movdqu (%r8),%xmm3 |
| 457 leaq 256(%rsp),%rcx |
| 458 leaq L$rot16(%rip),%r10 |
| 459 leaq L$rot24(%rip),%r11 |
| 460 |
| 461 pshufd $0x00,%xmm11,%xmm8 |
| 462 pshufd $0x55,%xmm11,%xmm9 |
| 463 movdqa %xmm8,64(%rsp) |
| 464 pshufd $0xaa,%xmm11,%xmm10 |
| 465 movdqa %xmm9,80(%rsp) |
| 466 pshufd $0xff,%xmm11,%xmm11 |
| 467 movdqa %xmm10,96(%rsp) |
| 468 movdqa %xmm11,112(%rsp) |
| 469 |
| 470 pshufd $0x00,%xmm15,%xmm12 |
| 471 pshufd $0x55,%xmm15,%xmm13 |
| 472 movdqa %xmm12,128-256(%rcx) |
| 473 pshufd $0xaa,%xmm15,%xmm14 |
| 474 movdqa %xmm13,144-256(%rcx) |
| 475 pshufd $0xff,%xmm15,%xmm15 |
| 476 movdqa %xmm14,160-256(%rcx) |
| 477 movdqa %xmm15,176-256(%rcx) |
| 478 |
| 479 pshufd $0x00,%xmm7,%xmm4 |
| 480 pshufd $0x55,%xmm7,%xmm5 |
| 481 movdqa %xmm4,192-256(%rcx) |
| 482 pshufd $0xaa,%xmm7,%xmm6 |
| 483 movdqa %xmm5,208-256(%rcx) |
| 484 pshufd $0xff,%xmm7,%xmm7 |
| 485 movdqa %xmm6,224-256(%rcx) |
| 486 movdqa %xmm7,240-256(%rcx) |
| 487 |
| 488 pshufd $0x00,%xmm3,%xmm0 |
| 489 pshufd $0x55,%xmm3,%xmm1 |
| 490 paddd L$inc(%rip),%xmm0 |
| 491 pshufd $0xaa,%xmm3,%xmm2 |
| 492 movdqa %xmm1,272-256(%rcx) |
| 493 pshufd $0xff,%xmm3,%xmm3 |
| 494 movdqa %xmm2,288-256(%rcx) |
| 495 movdqa %xmm3,304-256(%rcx) |
| 496 |
| 497 jmp L$oop_enter4x |
| 498 |
| 499 .p2align 5 |
| 500 L$oop_outer4x: |
| 501 movdqa 64(%rsp),%xmm8 |
| 502 movdqa 80(%rsp),%xmm9 |
| 503 movdqa 96(%rsp),%xmm10 |
| 504 movdqa 112(%rsp),%xmm11 |
| 505 movdqa 128-256(%rcx),%xmm12 |
| 506 movdqa 144-256(%rcx),%xmm13 |
| 507 movdqa 160-256(%rcx),%xmm14 |
| 508 movdqa 176-256(%rcx),%xmm15 |
| 509 movdqa 192-256(%rcx),%xmm4 |
| 510 movdqa 208-256(%rcx),%xmm5 |
| 511 movdqa 224-256(%rcx),%xmm6 |
| 512 movdqa 240-256(%rcx),%xmm7 |
| 513 movdqa 256-256(%rcx),%xmm0 |
| 514 movdqa 272-256(%rcx),%xmm1 |
| 515 movdqa 288-256(%rcx),%xmm2 |
| 516 movdqa 304-256(%rcx),%xmm3 |
| 517 paddd L$four(%rip),%xmm0 |
| 518 |
| 519 L$oop_enter4x: |
| 520 movdqa %xmm6,32(%rsp) |
| 521 movdqa %xmm7,48(%rsp) |
| 522 movdqa (%r10),%xmm7 |
| 523 movl $10,%eax |
| 524 movdqa %xmm0,256-256(%rcx) |
| 525 jmp L$oop4x |
| 526 |
| 527 .p2align 5 |
| 528 L$oop4x: |
| 529 paddd %xmm12,%xmm8 |
| 530 paddd %xmm13,%xmm9 |
| 531 pxor %xmm8,%xmm0 |
| 532 pxor %xmm9,%xmm1 |
| 533 .byte 102,15,56,0,199 |
| 534 .byte 102,15,56,0,207 |
| 535 paddd %xmm0,%xmm4 |
| 536 paddd %xmm1,%xmm5 |
| 537 pxor %xmm4,%xmm12 |
| 538 pxor %xmm5,%xmm13 |
| 539 movdqa %xmm12,%xmm6 |
| 540 pslld $12,%xmm12 |
| 541 psrld $20,%xmm6 |
| 542 movdqa %xmm13,%xmm7 |
| 543 pslld $12,%xmm13 |
| 544 por %xmm6,%xmm12 |
| 545 psrld $20,%xmm7 |
| 546 movdqa (%r11),%xmm6 |
| 547 por %xmm7,%xmm13 |
| 548 paddd %xmm12,%xmm8 |
| 549 paddd %xmm13,%xmm9 |
| 550 pxor %xmm8,%xmm0 |
| 551 pxor %xmm9,%xmm1 |
| 552 .byte 102,15,56,0,198 |
| 553 .byte 102,15,56,0,206 |
| 554 paddd %xmm0,%xmm4 |
| 555 paddd %xmm1,%xmm5 |
| 556 pxor %xmm4,%xmm12 |
| 557 pxor %xmm5,%xmm13 |
| 558 movdqa %xmm12,%xmm7 |
| 559 pslld $7,%xmm12 |
| 560 psrld $25,%xmm7 |
| 561 movdqa %xmm13,%xmm6 |
| 562 pslld $7,%xmm13 |
| 563 por %xmm7,%xmm12 |
| 564 psrld $25,%xmm6 |
| 565 movdqa (%r10),%xmm7 |
| 566 por %xmm6,%xmm13 |
| 567 movdqa %xmm4,0(%rsp) |
| 568 movdqa %xmm5,16(%rsp) |
| 569 movdqa 32(%rsp),%xmm4 |
| 570 movdqa 48(%rsp),%xmm5 |
| 571 paddd %xmm14,%xmm10 |
| 572 paddd %xmm15,%xmm11 |
| 573 pxor %xmm10,%xmm2 |
| 574 pxor %xmm11,%xmm3 |
| 575 .byte 102,15,56,0,215 |
| 576 .byte 102,15,56,0,223 |
| 577 paddd %xmm2,%xmm4 |
| 578 paddd %xmm3,%xmm5 |
| 579 pxor %xmm4,%xmm14 |
| 580 pxor %xmm5,%xmm15 |
| 581 movdqa %xmm14,%xmm6 |
| 582 pslld $12,%xmm14 |
| 583 psrld $20,%xmm6 |
| 584 movdqa %xmm15,%xmm7 |
| 585 pslld $12,%xmm15 |
| 586 por %xmm6,%xmm14 |
| 587 psrld $20,%xmm7 |
| 588 movdqa (%r11),%xmm6 |
| 589 por %xmm7,%xmm15 |
| 590 paddd %xmm14,%xmm10 |
| 591 paddd %xmm15,%xmm11 |
| 592 pxor %xmm10,%xmm2 |
| 593 pxor %xmm11,%xmm3 |
| 594 .byte 102,15,56,0,214 |
| 595 .byte 102,15,56,0,222 |
| 596 paddd %xmm2,%xmm4 |
| 597 paddd %xmm3,%xmm5 |
| 598 pxor %xmm4,%xmm14 |
| 599 pxor %xmm5,%xmm15 |
| 600 movdqa %xmm14,%xmm7 |
| 601 pslld $7,%xmm14 |
| 602 psrld $25,%xmm7 |
| 603 movdqa %xmm15,%xmm6 |
| 604 pslld $7,%xmm15 |
| 605 por %xmm7,%xmm14 |
| 606 psrld $25,%xmm6 |
| 607 movdqa (%r10),%xmm7 |
| 608 por %xmm6,%xmm15 |
| 609 paddd %xmm13,%xmm8 |
| 610 paddd %xmm14,%xmm9 |
| 611 pxor %xmm8,%xmm3 |
| 612 pxor %xmm9,%xmm0 |
| 613 .byte 102,15,56,0,223 |
| 614 .byte 102,15,56,0,199 |
| 615 paddd %xmm3,%xmm4 |
| 616 paddd %xmm0,%xmm5 |
| 617 pxor %xmm4,%xmm13 |
| 618 pxor %xmm5,%xmm14 |
| 619 movdqa %xmm13,%xmm6 |
| 620 pslld $12,%xmm13 |
| 621 psrld $20,%xmm6 |
| 622 movdqa %xmm14,%xmm7 |
| 623 pslld $12,%xmm14 |
| 624 por %xmm6,%xmm13 |
| 625 psrld $20,%xmm7 |
| 626 movdqa (%r11),%xmm6 |
| 627 por %xmm7,%xmm14 |
| 628 paddd %xmm13,%xmm8 |
| 629 paddd %xmm14,%xmm9 |
| 630 pxor %xmm8,%xmm3 |
| 631 pxor %xmm9,%xmm0 |
| 632 .byte 102,15,56,0,222 |
| 633 .byte 102,15,56,0,198 |
| 634 paddd %xmm3,%xmm4 |
| 635 paddd %xmm0,%xmm5 |
| 636 pxor %xmm4,%xmm13 |
| 637 pxor %xmm5,%xmm14 |
| 638 movdqa %xmm13,%xmm7 |
| 639 pslld $7,%xmm13 |
| 640 psrld $25,%xmm7 |
| 641 movdqa %xmm14,%xmm6 |
| 642 pslld $7,%xmm14 |
| 643 por %xmm7,%xmm13 |
| 644 psrld $25,%xmm6 |
| 645 movdqa (%r10),%xmm7 |
| 646 por %xmm6,%xmm14 |
| 647 movdqa %xmm4,32(%rsp) |
| 648 movdqa %xmm5,48(%rsp) |
| 649 movdqa 0(%rsp),%xmm4 |
| 650 movdqa 16(%rsp),%xmm5 |
| 651 paddd %xmm15,%xmm10 |
| 652 paddd %xmm12,%xmm11 |
| 653 pxor %xmm10,%xmm1 |
| 654 pxor %xmm11,%xmm2 |
| 655 .byte 102,15,56,0,207 |
| 656 .byte 102,15,56,0,215 |
| 657 paddd %xmm1,%xmm4 |
| 658 paddd %xmm2,%xmm5 |
| 659 pxor %xmm4,%xmm15 |
| 660 pxor %xmm5,%xmm12 |
| 661 movdqa %xmm15,%xmm6 |
| 662 pslld $12,%xmm15 |
| 663 psrld $20,%xmm6 |
| 664 movdqa %xmm12,%xmm7 |
| 665 pslld $12,%xmm12 |
| 666 por %xmm6,%xmm15 |
| 667 psrld $20,%xmm7 |
| 668 movdqa (%r11),%xmm6 |
| 669 por %xmm7,%xmm12 |
| 670 paddd %xmm15,%xmm10 |
| 671 paddd %xmm12,%xmm11 |
| 672 pxor %xmm10,%xmm1 |
| 673 pxor %xmm11,%xmm2 |
| 674 .byte 102,15,56,0,206 |
| 675 .byte 102,15,56,0,214 |
| 676 paddd %xmm1,%xmm4 |
| 677 paddd %xmm2,%xmm5 |
| 678 pxor %xmm4,%xmm15 |
| 679 pxor %xmm5,%xmm12 |
| 680 movdqa %xmm15,%xmm7 |
| 681 pslld $7,%xmm15 |
| 682 psrld $25,%xmm7 |
| 683 movdqa %xmm12,%xmm6 |
| 684 pslld $7,%xmm12 |
| 685 por %xmm7,%xmm15 |
| 686 psrld $25,%xmm6 |
| 687 movdqa (%r10),%xmm7 |
| 688 por %xmm6,%xmm12 |
| 689 decl %eax |
| 690 jnz L$oop4x |
| 691 |
| 692 paddd 64(%rsp),%xmm8 |
| 693 paddd 80(%rsp),%xmm9 |
| 694 paddd 96(%rsp),%xmm10 |
| 695 paddd 112(%rsp),%xmm11 |
| 696 |
| 697 movdqa %xmm8,%xmm6 |
| 698 punpckldq %xmm9,%xmm8 |
| 699 movdqa %xmm10,%xmm7 |
| 700 punpckldq %xmm11,%xmm10 |
| 701 punpckhdq %xmm9,%xmm6 |
| 702 punpckhdq %xmm11,%xmm7 |
| 703 movdqa %xmm8,%xmm9 |
| 704 punpcklqdq %xmm10,%xmm8 |
| 705 movdqa %xmm6,%xmm11 |
| 706 punpcklqdq %xmm7,%xmm6 |
| 707 punpckhqdq %xmm10,%xmm9 |
| 708 punpckhqdq %xmm7,%xmm11 |
| 709 paddd 128-256(%rcx),%xmm12 |
| 710 paddd 144-256(%rcx),%xmm13 |
| 711 paddd 160-256(%rcx),%xmm14 |
| 712 paddd 176-256(%rcx),%xmm15 |
| 713 |
| 714 movdqa %xmm8,0(%rsp) |
| 715 movdqa %xmm9,16(%rsp) |
| 716 movdqa 32(%rsp),%xmm8 |
| 717 movdqa 48(%rsp),%xmm9 |
| 718 |
| 719 movdqa %xmm12,%xmm10 |
| 720 punpckldq %xmm13,%xmm12 |
| 721 movdqa %xmm14,%xmm7 |
| 722 punpckldq %xmm15,%xmm14 |
| 723 punpckhdq %xmm13,%xmm10 |
| 724 punpckhdq %xmm15,%xmm7 |
| 725 movdqa %xmm12,%xmm13 |
| 726 punpcklqdq %xmm14,%xmm12 |
| 727 movdqa %xmm10,%xmm15 |
| 728 punpcklqdq %xmm7,%xmm10 |
| 729 punpckhqdq %xmm14,%xmm13 |
| 730 punpckhqdq %xmm7,%xmm15 |
| 731 paddd 192-256(%rcx),%xmm4 |
| 732 paddd 208-256(%rcx),%xmm5 |
| 733 paddd 224-256(%rcx),%xmm8 |
| 734 paddd 240-256(%rcx),%xmm9 |
| 735 |
| 736 movdqa %xmm6,32(%rsp) |
| 737 movdqa %xmm11,48(%rsp) |
| 738 |
| 739 movdqa %xmm4,%xmm14 |
| 740 punpckldq %xmm5,%xmm4 |
| 741 movdqa %xmm8,%xmm7 |
| 742 punpckldq %xmm9,%xmm8 |
| 743 punpckhdq %xmm5,%xmm14 |
| 744 punpckhdq %xmm9,%xmm7 |
| 745 movdqa %xmm4,%xmm5 |
| 746 punpcklqdq %xmm8,%xmm4 |
| 747 movdqa %xmm14,%xmm9 |
| 748 punpcklqdq %xmm7,%xmm14 |
| 749 punpckhqdq %xmm8,%xmm5 |
| 750 punpckhqdq %xmm7,%xmm9 |
| 751 paddd 256-256(%rcx),%xmm0 |
| 752 paddd 272-256(%rcx),%xmm1 |
| 753 paddd 288-256(%rcx),%xmm2 |
| 754 paddd 304-256(%rcx),%xmm3 |
| 755 |
| 756 movdqa %xmm0,%xmm8 |
| 757 punpckldq %xmm1,%xmm0 |
| 758 movdqa %xmm2,%xmm7 |
| 759 punpckldq %xmm3,%xmm2 |
| 760 punpckhdq %xmm1,%xmm8 |
| 761 punpckhdq %xmm3,%xmm7 |
| 762 movdqa %xmm0,%xmm1 |
| 763 punpcklqdq %xmm2,%xmm0 |
| 764 movdqa %xmm8,%xmm3 |
| 765 punpcklqdq %xmm7,%xmm8 |
| 766 punpckhqdq %xmm2,%xmm1 |
| 767 punpckhqdq %xmm7,%xmm3 |
| 768 cmpq $256,%rdx |
| 769 jb L$tail4x |
| 770 |
| 771 movdqu 0(%rsi),%xmm6 |
| 772 movdqu 16(%rsi),%xmm11 |
| 773 movdqu 32(%rsi),%xmm2 |
| 774 movdqu 48(%rsi),%xmm7 |
| 775 pxor 0(%rsp),%xmm6 |
| 776 pxor %xmm12,%xmm11 |
| 777 pxor %xmm4,%xmm2 |
| 778 pxor %xmm0,%xmm7 |
| 779 |
| 780 movdqu %xmm6,0(%rdi) |
| 781 movdqu 64(%rsi),%xmm6 |
| 782 movdqu %xmm11,16(%rdi) |
| 783 movdqu 80(%rsi),%xmm11 |
| 784 movdqu %xmm2,32(%rdi) |
| 785 movdqu 96(%rsi),%xmm2 |
| 786 movdqu %xmm7,48(%rdi) |
| 787 movdqu 112(%rsi),%xmm7 |
| 788 leaq 128(%rsi),%rsi |
| 789 pxor 16(%rsp),%xmm6 |
| 790 pxor %xmm13,%xmm11 |
| 791 pxor %xmm5,%xmm2 |
| 792 pxor %xmm1,%xmm7 |
| 793 |
| 794 movdqu %xmm6,64(%rdi) |
| 795 movdqu 0(%rsi),%xmm6 |
| 796 movdqu %xmm11,80(%rdi) |
| 797 movdqu 16(%rsi),%xmm11 |
| 798 movdqu %xmm2,96(%rdi) |
| 799 movdqu 32(%rsi),%xmm2 |
| 800 movdqu %xmm7,112(%rdi) |
| 801 leaq 128(%rdi),%rdi |
| 802 movdqu 48(%rsi),%xmm7 |
| 803 pxor 32(%rsp),%xmm6 |
| 804 pxor %xmm10,%xmm11 |
| 805 pxor %xmm14,%xmm2 |
| 806 pxor %xmm8,%xmm7 |
| 807 |
| 808 movdqu %xmm6,0(%rdi) |
| 809 movdqu 64(%rsi),%xmm6 |
| 810 movdqu %xmm11,16(%rdi) |
| 811 movdqu 80(%rsi),%xmm11 |
| 812 movdqu %xmm2,32(%rdi) |
| 813 movdqu 96(%rsi),%xmm2 |
| 814 movdqu %xmm7,48(%rdi) |
| 815 movdqu 112(%rsi),%xmm7 |
| 816 leaq 128(%rsi),%rsi |
| 817 pxor 48(%rsp),%xmm6 |
| 818 pxor %xmm15,%xmm11 |
| 819 pxor %xmm9,%xmm2 |
| 820 pxor %xmm3,%xmm7 |
| 821 movdqu %xmm6,64(%rdi) |
| 822 movdqu %xmm11,80(%rdi) |
| 823 movdqu %xmm2,96(%rdi) |
| 824 movdqu %xmm7,112(%rdi) |
| 825 leaq 128(%rdi),%rdi |
| 826 |
| 827 subq $256,%rdx |
| 828 jnz L$oop_outer4x |
| 829 |
| 830 jmp L$done4x |
| 831 |
| 832 L$tail4x: |
| 833 cmpq $192,%rdx |
| 834 jae L$192_or_more4x |
| 835 cmpq $128,%rdx |
| 836 jae L$128_or_more4x |
| 837 cmpq $64,%rdx |
| 838 jae L$64_or_more4x |
| 839 |
| 840 |
| 841 xorq %r10,%r10 |
| 842 |
| 843 movdqa %xmm12,16(%rsp) |
| 844 movdqa %xmm4,32(%rsp) |
| 845 movdqa %xmm0,48(%rsp) |
| 846 jmp L$oop_tail4x |
| 847 |
| 848 .p2align 5 |
| 849 L$64_or_more4x: |
| 850 movdqu 0(%rsi),%xmm6 |
| 851 movdqu 16(%rsi),%xmm11 |
| 852 movdqu 32(%rsi),%xmm2 |
| 853 movdqu 48(%rsi),%xmm7 |
| 854 pxor 0(%rsp),%xmm6 |
| 855 pxor %xmm12,%xmm11 |
| 856 pxor %xmm4,%xmm2 |
| 857 pxor %xmm0,%xmm7 |
| 858 movdqu %xmm6,0(%rdi) |
| 859 movdqu %xmm11,16(%rdi) |
| 860 movdqu %xmm2,32(%rdi) |
| 861 movdqu %xmm7,48(%rdi) |
| 862 je L$done4x |
| 863 |
| 864 movdqa 16(%rsp),%xmm6 |
| 865 leaq 64(%rsi),%rsi |
| 866 xorq %r10,%r10 |
| 867 movdqa %xmm6,0(%rsp) |
| 868 movdqa %xmm13,16(%rsp) |
| 869 leaq 64(%rdi),%rdi |
| 870 movdqa %xmm5,32(%rsp) |
| 871 subq $64,%rdx |
| 872 movdqa %xmm1,48(%rsp) |
| 873 jmp L$oop_tail4x |
| 874 |
| 875 .p2align 5 |
| 876 L$128_or_more4x: |
| 877 movdqu 0(%rsi),%xmm6 |
| 878 movdqu 16(%rsi),%xmm11 |
| 879 movdqu 32(%rsi),%xmm2 |
| 880 movdqu 48(%rsi),%xmm7 |
| 881 pxor 0(%rsp),%xmm6 |
| 882 pxor %xmm12,%xmm11 |
| 883 pxor %xmm4,%xmm2 |
| 884 pxor %xmm0,%xmm7 |
| 885 |
| 886 movdqu %xmm6,0(%rdi) |
| 887 movdqu 64(%rsi),%xmm6 |
| 888 movdqu %xmm11,16(%rdi) |
| 889 movdqu 80(%rsi),%xmm11 |
| 890 movdqu %xmm2,32(%rdi) |
| 891 movdqu 96(%rsi),%xmm2 |
| 892 movdqu %xmm7,48(%rdi) |
| 893 movdqu 112(%rsi),%xmm7 |
| 894 pxor 16(%rsp),%xmm6 |
| 895 pxor %xmm13,%xmm11 |
| 896 pxor %xmm5,%xmm2 |
| 897 pxor %xmm1,%xmm7 |
| 898 movdqu %xmm6,64(%rdi) |
| 899 movdqu %xmm11,80(%rdi) |
| 900 movdqu %xmm2,96(%rdi) |
| 901 movdqu %xmm7,112(%rdi) |
| 902 je L$done4x |
| 903 |
| 904 movdqa 32(%rsp),%xmm6 |
| 905 leaq 128(%rsi),%rsi |
| 906 xorq %r10,%r10 |
| 907 movdqa %xmm6,0(%rsp) |
| 908 movdqa %xmm10,16(%rsp) |
| 909 leaq 128(%rdi),%rdi |
| 910 movdqa %xmm14,32(%rsp) |
| 911 subq $128,%rdx |
| 912 movdqa %xmm8,48(%rsp) |
| 913 jmp L$oop_tail4x |
| 914 |
| 915 .p2align 5 |
| 916 L$192_or_more4x: |
| 917 movdqu 0(%rsi),%xmm6 |
| 918 movdqu 16(%rsi),%xmm11 |
| 919 movdqu 32(%rsi),%xmm2 |
| 920 movdqu 48(%rsi),%xmm7 |
| 921 pxor 0(%rsp),%xmm6 |
| 922 pxor %xmm12,%xmm11 |
| 923 pxor %xmm4,%xmm2 |
| 924 pxor %xmm0,%xmm7 |
| 925 |
| 926 movdqu %xmm6,0(%rdi) |
| 927 movdqu 64(%rsi),%xmm6 |
| 928 movdqu %xmm11,16(%rdi) |
| 929 movdqu 80(%rsi),%xmm11 |
| 930 movdqu %xmm2,32(%rdi) |
| 931 movdqu 96(%rsi),%xmm2 |
| 932 movdqu %xmm7,48(%rdi) |
| 933 movdqu 112(%rsi),%xmm7 |
| 934 leaq 128(%rsi),%rsi |
| 935 pxor 16(%rsp),%xmm6 |
| 936 pxor %xmm13,%xmm11 |
| 937 pxor %xmm5,%xmm2 |
| 938 pxor %xmm1,%xmm7 |
| 939 |
| 940 movdqu %xmm6,64(%rdi) |
| 941 movdqu 0(%rsi),%xmm6 |
| 942 movdqu %xmm11,80(%rdi) |
| 943 movdqu 16(%rsi),%xmm11 |
| 944 movdqu %xmm2,96(%rdi) |
| 945 movdqu 32(%rsi),%xmm2 |
| 946 movdqu %xmm7,112(%rdi) |
| 947 leaq 128(%rdi),%rdi |
| 948 movdqu 48(%rsi),%xmm7 |
| 949 pxor 32(%rsp),%xmm6 |
| 950 pxor %xmm10,%xmm11 |
| 951 pxor %xmm14,%xmm2 |
| 952 pxor %xmm8,%xmm7 |
| 953 movdqu %xmm6,0(%rdi) |
| 954 movdqu %xmm11,16(%rdi) |
| 955 movdqu %xmm2,32(%rdi) |
| 956 movdqu %xmm7,48(%rdi) |
| 957 je L$done4x |
| 958 |
| 959 movdqa 48(%rsp),%xmm6 |
| 960 leaq 64(%rsi),%rsi |
| 961 xorq %r10,%r10 |
| 962 movdqa %xmm6,0(%rsp) |
| 963 movdqa %xmm15,16(%rsp) |
| 964 leaq 64(%rdi),%rdi |
| 965 movdqa %xmm9,32(%rsp) |
| 966 subq $192,%rdx |
| 967 movdqa %xmm3,48(%rsp) |
| 968 |
| 969 L$oop_tail4x: |
| 970 movzbl (%rsi,%r10,1),%eax |
| 971 movzbl (%rsp,%r10,1),%ecx |
| 972 leaq 1(%r10),%r10 |
| 973 xorl %ecx,%eax |
| 974 movb %al,-1(%rdi,%r10,1) |
| 975 decq %rdx |
| 976 jnz L$oop_tail4x |
| 977 |
| 978 L$done4x: |
| 979 addq $0x148+0,%rsp |
| 980 .byte 0xf3,0xc3 |
| 981 |
| 982 |
| 983 .p2align 5 |
| 984 ChaCha20_8x: |
| 985 L$ChaCha20_8x: |
| 986 movq %rsp,%r10 |
| 987 subq $0x280+8,%rsp |
| 988 andq $-32,%rsp |
| 989 vzeroupper |
| 990 movq %r10,640(%rsp) |
| 991 |
| 992 |
| 993 |
| 994 |
| 995 |
| 996 |
| 997 |
| 998 |
| 999 |
| 1000 |
| 1001 vbroadcasti128 L$sigma(%rip),%ymm11 |
| 1002 vbroadcasti128 (%rcx),%ymm3 |
| 1003 vbroadcasti128 16(%rcx),%ymm15 |
| 1004 vbroadcasti128 (%r8),%ymm7 |
| 1005 leaq 256(%rsp),%rcx |
| 1006 leaq 512(%rsp),%rax |
| 1007 leaq L$rot16(%rip),%r10 |
| 1008 leaq L$rot24(%rip),%r11 |
| 1009 |
| 1010 vpshufd $0x00,%ymm11,%ymm8 |
| 1011 vpshufd $0x55,%ymm11,%ymm9 |
| 1012 vmovdqa %ymm8,128-256(%rcx) |
| 1013 vpshufd $0xaa,%ymm11,%ymm10 |
| 1014 vmovdqa %ymm9,160-256(%rcx) |
| 1015 vpshufd $0xff,%ymm11,%ymm11 |
| 1016 vmovdqa %ymm10,192-256(%rcx) |
| 1017 vmovdqa %ymm11,224-256(%rcx) |
| 1018 |
| 1019 vpshufd $0x00,%ymm3,%ymm0 |
| 1020 vpshufd $0x55,%ymm3,%ymm1 |
| 1021 vmovdqa %ymm0,256-256(%rcx) |
| 1022 vpshufd $0xaa,%ymm3,%ymm2 |
| 1023 vmovdqa %ymm1,288-256(%rcx) |
| 1024 vpshufd $0xff,%ymm3,%ymm3 |
| 1025 vmovdqa %ymm2,320-256(%rcx) |
| 1026 vmovdqa %ymm3,352-256(%rcx) |
| 1027 |
| 1028 vpshufd $0x00,%ymm15,%ymm12 |
| 1029 vpshufd $0x55,%ymm15,%ymm13 |
| 1030 vmovdqa %ymm12,384-512(%rax) |
| 1031 vpshufd $0xaa,%ymm15,%ymm14 |
| 1032 vmovdqa %ymm13,416-512(%rax) |
| 1033 vpshufd $0xff,%ymm15,%ymm15 |
| 1034 vmovdqa %ymm14,448-512(%rax) |
| 1035 vmovdqa %ymm15,480-512(%rax) |
| 1036 |
| 1037 vpshufd $0x00,%ymm7,%ymm4 |
| 1038 vpshufd $0x55,%ymm7,%ymm5 |
| 1039 vpaddd L$incy(%rip),%ymm4,%ymm4 |
| 1040 vpshufd $0xaa,%ymm7,%ymm6 |
| 1041 vmovdqa %ymm5,544-512(%rax) |
| 1042 vpshufd $0xff,%ymm7,%ymm7 |
| 1043 vmovdqa %ymm6,576-512(%rax) |
| 1044 vmovdqa %ymm7,608-512(%rax) |
| 1045 |
| 1046 jmp L$oop_enter8x |
| 1047 |
| 1048 .p2align 5 |
| 1049 L$oop_outer8x: |
| 1050 vmovdqa 128-256(%rcx),%ymm8 |
| 1051 vmovdqa 160-256(%rcx),%ymm9 |
| 1052 vmovdqa 192-256(%rcx),%ymm10 |
| 1053 vmovdqa 224-256(%rcx),%ymm11 |
| 1054 vmovdqa 256-256(%rcx),%ymm0 |
| 1055 vmovdqa 288-256(%rcx),%ymm1 |
| 1056 vmovdqa 320-256(%rcx),%ymm2 |
| 1057 vmovdqa 352-256(%rcx),%ymm3 |
| 1058 vmovdqa 384-512(%rax),%ymm12 |
| 1059 vmovdqa 416-512(%rax),%ymm13 |
| 1060 vmovdqa 448-512(%rax),%ymm14 |
| 1061 vmovdqa 480-512(%rax),%ymm15 |
| 1062 vmovdqa 512-512(%rax),%ymm4 |
| 1063 vmovdqa 544-512(%rax),%ymm5 |
| 1064 vmovdqa 576-512(%rax),%ymm6 |
| 1065 vmovdqa 608-512(%rax),%ymm7 |
| 1066 vpaddd L$eight(%rip),%ymm4,%ymm4 |
| 1067 |
| 1068 L$oop_enter8x: |
| 1069 vmovdqa %ymm14,64(%rsp) |
| 1070 vmovdqa %ymm15,96(%rsp) |
| 1071 vbroadcasti128 (%r10),%ymm15 |
| 1072 vmovdqa %ymm4,512-512(%rax) |
| 1073 movl $10,%eax |
| 1074 jmp L$oop8x |
| 1075 |
| 1076 .p2align 5 |
| 1077 L$oop8x: |
| 1078 vpaddd %ymm0,%ymm8,%ymm8 |
| 1079 vpxor %ymm4,%ymm8,%ymm4 |
| 1080 vpshufb %ymm15,%ymm4,%ymm4 |
| 1081 vpaddd %ymm1,%ymm9,%ymm9 |
| 1082 vpxor %ymm5,%ymm9,%ymm5 |
| 1083 vpshufb %ymm15,%ymm5,%ymm5 |
| 1084 vpaddd %ymm4,%ymm12,%ymm12 |
| 1085 vpxor %ymm0,%ymm12,%ymm0 |
| 1086 vpslld $12,%ymm0,%ymm14 |
| 1087 vpsrld $20,%ymm0,%ymm0 |
| 1088 vpor %ymm0,%ymm14,%ymm0 |
| 1089 vbroadcasti128 (%r11),%ymm14 |
| 1090 vpaddd %ymm5,%ymm13,%ymm13 |
| 1091 vpxor %ymm1,%ymm13,%ymm1 |
| 1092 vpslld $12,%ymm1,%ymm15 |
| 1093 vpsrld $20,%ymm1,%ymm1 |
| 1094 vpor %ymm1,%ymm15,%ymm1 |
| 1095 vpaddd %ymm0,%ymm8,%ymm8 |
| 1096 vpxor %ymm4,%ymm8,%ymm4 |
| 1097 vpshufb %ymm14,%ymm4,%ymm4 |
| 1098 vpaddd %ymm1,%ymm9,%ymm9 |
| 1099 vpxor %ymm5,%ymm9,%ymm5 |
| 1100 vpshufb %ymm14,%ymm5,%ymm5 |
| 1101 vpaddd %ymm4,%ymm12,%ymm12 |
| 1102 vpxor %ymm0,%ymm12,%ymm0 |
| 1103 vpslld $7,%ymm0,%ymm15 |
| 1104 vpsrld $25,%ymm0,%ymm0 |
| 1105 vpor %ymm0,%ymm15,%ymm0 |
| 1106 vbroadcasti128 (%r10),%ymm15 |
| 1107 vpaddd %ymm5,%ymm13,%ymm13 |
| 1108 vpxor %ymm1,%ymm13,%ymm1 |
| 1109 vpslld $7,%ymm1,%ymm14 |
| 1110 vpsrld $25,%ymm1,%ymm1 |
| 1111 vpor %ymm1,%ymm14,%ymm1 |
| 1112 vmovdqa %ymm12,0(%rsp) |
| 1113 vmovdqa %ymm13,32(%rsp) |
| 1114 vmovdqa 64(%rsp),%ymm12 |
| 1115 vmovdqa 96(%rsp),%ymm13 |
| 1116 vpaddd %ymm2,%ymm10,%ymm10 |
| 1117 vpxor %ymm6,%ymm10,%ymm6 |
| 1118 vpshufb %ymm15,%ymm6,%ymm6 |
| 1119 vpaddd %ymm3,%ymm11,%ymm11 |
| 1120 vpxor %ymm7,%ymm11,%ymm7 |
| 1121 vpshufb %ymm15,%ymm7,%ymm7 |
| 1122 vpaddd %ymm6,%ymm12,%ymm12 |
| 1123 vpxor %ymm2,%ymm12,%ymm2 |
| 1124 vpslld $12,%ymm2,%ymm14 |
| 1125 vpsrld $20,%ymm2,%ymm2 |
| 1126 vpor %ymm2,%ymm14,%ymm2 |
| 1127 vbroadcasti128 (%r11),%ymm14 |
| 1128 vpaddd %ymm7,%ymm13,%ymm13 |
| 1129 vpxor %ymm3,%ymm13,%ymm3 |
| 1130 vpslld $12,%ymm3,%ymm15 |
| 1131 vpsrld $20,%ymm3,%ymm3 |
| 1132 vpor %ymm3,%ymm15,%ymm3 |
| 1133 vpaddd %ymm2,%ymm10,%ymm10 |
| 1134 vpxor %ymm6,%ymm10,%ymm6 |
| 1135 vpshufb %ymm14,%ymm6,%ymm6 |
| 1136 vpaddd %ymm3,%ymm11,%ymm11 |
| 1137 vpxor %ymm7,%ymm11,%ymm7 |
| 1138 vpshufb %ymm14,%ymm7,%ymm7 |
| 1139 vpaddd %ymm6,%ymm12,%ymm12 |
| 1140 vpxor %ymm2,%ymm12,%ymm2 |
| 1141 vpslld $7,%ymm2,%ymm15 |
| 1142 vpsrld $25,%ymm2,%ymm2 |
| 1143 vpor %ymm2,%ymm15,%ymm2 |
| 1144 vbroadcasti128 (%r10),%ymm15 |
| 1145 vpaddd %ymm7,%ymm13,%ymm13 |
| 1146 vpxor %ymm3,%ymm13,%ymm3 |
| 1147 vpslld $7,%ymm3,%ymm14 |
| 1148 vpsrld $25,%ymm3,%ymm3 |
| 1149 vpor %ymm3,%ymm14,%ymm3 |
| 1150 vpaddd %ymm1,%ymm8,%ymm8 |
| 1151 vpxor %ymm7,%ymm8,%ymm7 |
| 1152 vpshufb %ymm15,%ymm7,%ymm7 |
| 1153 vpaddd %ymm2,%ymm9,%ymm9 |
| 1154 vpxor %ymm4,%ymm9,%ymm4 |
| 1155 vpshufb %ymm15,%ymm4,%ymm4 |
| 1156 vpaddd %ymm7,%ymm12,%ymm12 |
| 1157 vpxor %ymm1,%ymm12,%ymm1 |
| 1158 vpslld $12,%ymm1,%ymm14 |
| 1159 vpsrld $20,%ymm1,%ymm1 |
| 1160 vpor %ymm1,%ymm14,%ymm1 |
| 1161 vbroadcasti128 (%r11),%ymm14 |
| 1162 vpaddd %ymm4,%ymm13,%ymm13 |
| 1163 vpxor %ymm2,%ymm13,%ymm2 |
| 1164 vpslld $12,%ymm2,%ymm15 |
| 1165 vpsrld $20,%ymm2,%ymm2 |
| 1166 vpor %ymm2,%ymm15,%ymm2 |
| 1167 vpaddd %ymm1,%ymm8,%ymm8 |
| 1168 vpxor %ymm7,%ymm8,%ymm7 |
| 1169 vpshufb %ymm14,%ymm7,%ymm7 |
| 1170 vpaddd %ymm2,%ymm9,%ymm9 |
| 1171 vpxor %ymm4,%ymm9,%ymm4 |
| 1172 vpshufb %ymm14,%ymm4,%ymm4 |
| 1173 vpaddd %ymm7,%ymm12,%ymm12 |
| 1174 vpxor %ymm1,%ymm12,%ymm1 |
| 1175 vpslld $7,%ymm1,%ymm15 |
| 1176 vpsrld $25,%ymm1,%ymm1 |
| 1177 vpor %ymm1,%ymm15,%ymm1 |
| 1178 vbroadcasti128 (%r10),%ymm15 |
| 1179 vpaddd %ymm4,%ymm13,%ymm13 |
| 1180 vpxor %ymm2,%ymm13,%ymm2 |
| 1181 vpslld $7,%ymm2,%ymm14 |
| 1182 vpsrld $25,%ymm2,%ymm2 |
| 1183 vpor %ymm2,%ymm14,%ymm2 |
| 1184 vmovdqa %ymm12,64(%rsp) |
| 1185 vmovdqa %ymm13,96(%rsp) |
| 1186 vmovdqa 0(%rsp),%ymm12 |
| 1187 vmovdqa 32(%rsp),%ymm13 |
| 1188 vpaddd %ymm3,%ymm10,%ymm10 |
| 1189 vpxor %ymm5,%ymm10,%ymm5 |
| 1190 vpshufb %ymm15,%ymm5,%ymm5 |
| 1191 vpaddd %ymm0,%ymm11,%ymm11 |
| 1192 vpxor %ymm6,%ymm11,%ymm6 |
| 1193 vpshufb %ymm15,%ymm6,%ymm6 |
| 1194 vpaddd %ymm5,%ymm12,%ymm12 |
| 1195 vpxor %ymm3,%ymm12,%ymm3 |
| 1196 vpslld $12,%ymm3,%ymm14 |
| 1197 vpsrld $20,%ymm3,%ymm3 |
| 1198 vpor %ymm3,%ymm14,%ymm3 |
| 1199 vbroadcasti128 (%r11),%ymm14 |
| 1200 vpaddd %ymm6,%ymm13,%ymm13 |
| 1201 vpxor %ymm0,%ymm13,%ymm0 |
| 1202 vpslld $12,%ymm0,%ymm15 |
| 1203 vpsrld $20,%ymm0,%ymm0 |
| 1204 vpor %ymm0,%ymm15,%ymm0 |
| 1205 vpaddd %ymm3,%ymm10,%ymm10 |
| 1206 vpxor %ymm5,%ymm10,%ymm5 |
| 1207 vpshufb %ymm14,%ymm5,%ymm5 |
| 1208 vpaddd %ymm0,%ymm11,%ymm11 |
| 1209 vpxor %ymm6,%ymm11,%ymm6 |
| 1210 vpshufb %ymm14,%ymm6,%ymm6 |
| 1211 vpaddd %ymm5,%ymm12,%ymm12 |
| 1212 vpxor %ymm3,%ymm12,%ymm3 |
| 1213 vpslld $7,%ymm3,%ymm15 |
| 1214 vpsrld $25,%ymm3,%ymm3 |
| 1215 vpor %ymm3,%ymm15,%ymm3 |
| 1216 vbroadcasti128 (%r10),%ymm15 |
| 1217 vpaddd %ymm6,%ymm13,%ymm13 |
| 1218 vpxor %ymm0,%ymm13,%ymm0 |
| 1219 vpslld $7,%ymm0,%ymm14 |
| 1220 vpsrld $25,%ymm0,%ymm0 |
| 1221 vpor %ymm0,%ymm14,%ymm0 |
| 1222 decl %eax |
| 1223 jnz L$oop8x |
| 1224 |
| 1225 leaq 512(%rsp),%rax |
| 1226 vpaddd 128-256(%rcx),%ymm8,%ymm8 |
| 1227 vpaddd 160-256(%rcx),%ymm9,%ymm9 |
| 1228 vpaddd 192-256(%rcx),%ymm10,%ymm10 |
| 1229 vpaddd 224-256(%rcx),%ymm11,%ymm11 |
| 1230 |
| 1231 vpunpckldq %ymm9,%ymm8,%ymm14 |
| 1232 vpunpckldq %ymm11,%ymm10,%ymm15 |
| 1233 vpunpckhdq %ymm9,%ymm8,%ymm8 |
| 1234 vpunpckhdq %ymm11,%ymm10,%ymm10 |
| 1235 vpunpcklqdq %ymm15,%ymm14,%ymm9 |
| 1236 vpunpckhqdq %ymm15,%ymm14,%ymm14 |
| 1237 vpunpcklqdq %ymm10,%ymm8,%ymm11 |
| 1238 vpunpckhqdq %ymm10,%ymm8,%ymm8 |
| 1239 vpaddd 256-256(%rcx),%ymm0,%ymm0 |
| 1240 vpaddd 288-256(%rcx),%ymm1,%ymm1 |
| 1241 vpaddd 320-256(%rcx),%ymm2,%ymm2 |
| 1242 vpaddd 352-256(%rcx),%ymm3,%ymm3 |
| 1243 |
| 1244 vpunpckldq %ymm1,%ymm0,%ymm10 |
| 1245 vpunpckldq %ymm3,%ymm2,%ymm15 |
| 1246 vpunpckhdq %ymm1,%ymm0,%ymm0 |
| 1247 vpunpckhdq %ymm3,%ymm2,%ymm2 |
| 1248 vpunpcklqdq %ymm15,%ymm10,%ymm1 |
| 1249 vpunpckhqdq %ymm15,%ymm10,%ymm10 |
| 1250 vpunpcklqdq %ymm2,%ymm0,%ymm3 |
| 1251 vpunpckhqdq %ymm2,%ymm0,%ymm0 |
| 1252 vperm2i128 $0x20,%ymm1,%ymm9,%ymm15 |
| 1253 vperm2i128 $0x31,%ymm1,%ymm9,%ymm1 |
| 1254 vperm2i128 $0x20,%ymm10,%ymm14,%ymm9 |
| 1255 vperm2i128 $0x31,%ymm10,%ymm14,%ymm10 |
| 1256 vperm2i128 $0x20,%ymm3,%ymm11,%ymm14 |
| 1257 vperm2i128 $0x31,%ymm3,%ymm11,%ymm3 |
| 1258 vperm2i128 $0x20,%ymm0,%ymm8,%ymm11 |
| 1259 vperm2i128 $0x31,%ymm0,%ymm8,%ymm0 |
| 1260 vmovdqa %ymm15,0(%rsp) |
| 1261 vmovdqa %ymm9,32(%rsp) |
| 1262 vmovdqa 64(%rsp),%ymm15 |
| 1263 vmovdqa 96(%rsp),%ymm9 |
| 1264 |
| 1265 vpaddd 384-512(%rax),%ymm12,%ymm12 |
| 1266 vpaddd 416-512(%rax),%ymm13,%ymm13 |
| 1267 vpaddd 448-512(%rax),%ymm15,%ymm15 |
| 1268 vpaddd 480-512(%rax),%ymm9,%ymm9 |
| 1269 |
| 1270 vpunpckldq %ymm13,%ymm12,%ymm2 |
| 1271 vpunpckldq %ymm9,%ymm15,%ymm8 |
| 1272 vpunpckhdq %ymm13,%ymm12,%ymm12 |
| 1273 vpunpckhdq %ymm9,%ymm15,%ymm15 |
| 1274 vpunpcklqdq %ymm8,%ymm2,%ymm13 |
| 1275 vpunpckhqdq %ymm8,%ymm2,%ymm2 |
| 1276 vpunpcklqdq %ymm15,%ymm12,%ymm9 |
| 1277 vpunpckhqdq %ymm15,%ymm12,%ymm12 |
| 1278 vpaddd 512-512(%rax),%ymm4,%ymm4 |
| 1279 vpaddd 544-512(%rax),%ymm5,%ymm5 |
| 1280 vpaddd 576-512(%rax),%ymm6,%ymm6 |
| 1281 vpaddd 608-512(%rax),%ymm7,%ymm7 |
| 1282 |
| 1283 vpunpckldq %ymm5,%ymm4,%ymm15 |
| 1284 vpunpckldq %ymm7,%ymm6,%ymm8 |
| 1285 vpunpckhdq %ymm5,%ymm4,%ymm4 |
| 1286 vpunpckhdq %ymm7,%ymm6,%ymm6 |
| 1287 vpunpcklqdq %ymm8,%ymm15,%ymm5 |
| 1288 vpunpckhqdq %ymm8,%ymm15,%ymm15 |
| 1289 vpunpcklqdq %ymm6,%ymm4,%ymm7 |
| 1290 vpunpckhqdq %ymm6,%ymm4,%ymm4 |
| 1291 vperm2i128 $0x20,%ymm5,%ymm13,%ymm8 |
| 1292 vperm2i128 $0x31,%ymm5,%ymm13,%ymm5 |
| 1293 vperm2i128 $0x20,%ymm15,%ymm2,%ymm13 |
| 1294 vperm2i128 $0x31,%ymm15,%ymm2,%ymm15 |
| 1295 vperm2i128 $0x20,%ymm7,%ymm9,%ymm2 |
| 1296 vperm2i128 $0x31,%ymm7,%ymm9,%ymm7 |
| 1297 vperm2i128 $0x20,%ymm4,%ymm12,%ymm9 |
| 1298 vperm2i128 $0x31,%ymm4,%ymm12,%ymm4 |
| 1299 vmovdqa 0(%rsp),%ymm6 |
| 1300 vmovdqa 32(%rsp),%ymm12 |
| 1301 |
| 1302 cmpq $512,%rdx |
| 1303 jb L$tail8x |
| 1304 |
| 1305 vpxor 0(%rsi),%ymm6,%ymm6 |
| 1306 vpxor 32(%rsi),%ymm8,%ymm8 |
| 1307 vpxor 64(%rsi),%ymm1,%ymm1 |
| 1308 vpxor 96(%rsi),%ymm5,%ymm5 |
| 1309 leaq 128(%rsi),%rsi |
| 1310 vmovdqu %ymm6,0(%rdi) |
| 1311 vmovdqu %ymm8,32(%rdi) |
| 1312 vmovdqu %ymm1,64(%rdi) |
| 1313 vmovdqu %ymm5,96(%rdi) |
| 1314 leaq 128(%rdi),%rdi |
| 1315 |
| 1316 vpxor 0(%rsi),%ymm12,%ymm12 |
| 1317 vpxor 32(%rsi),%ymm13,%ymm13 |
| 1318 vpxor 64(%rsi),%ymm10,%ymm10 |
| 1319 vpxor 96(%rsi),%ymm15,%ymm15 |
| 1320 leaq 128(%rsi),%rsi |
| 1321 vmovdqu %ymm12,0(%rdi) |
| 1322 vmovdqu %ymm13,32(%rdi) |
| 1323 vmovdqu %ymm10,64(%rdi) |
| 1324 vmovdqu %ymm15,96(%rdi) |
| 1325 leaq 128(%rdi),%rdi |
| 1326 |
| 1327 vpxor 0(%rsi),%ymm14,%ymm14 |
| 1328 vpxor 32(%rsi),%ymm2,%ymm2 |
| 1329 vpxor 64(%rsi),%ymm3,%ymm3 |
| 1330 vpxor 96(%rsi),%ymm7,%ymm7 |
| 1331 leaq 128(%rsi),%rsi |
| 1332 vmovdqu %ymm14,0(%rdi) |
| 1333 vmovdqu %ymm2,32(%rdi) |
| 1334 vmovdqu %ymm3,64(%rdi) |
| 1335 vmovdqu %ymm7,96(%rdi) |
| 1336 leaq 128(%rdi),%rdi |
| 1337 |
| 1338 vpxor 0(%rsi),%ymm11,%ymm11 |
| 1339 vpxor 32(%rsi),%ymm9,%ymm9 |
| 1340 vpxor 64(%rsi),%ymm0,%ymm0 |
| 1341 vpxor 96(%rsi),%ymm4,%ymm4 |
| 1342 leaq 128(%rsi),%rsi |
| 1343 vmovdqu %ymm11,0(%rdi) |
| 1344 vmovdqu %ymm9,32(%rdi) |
| 1345 vmovdqu %ymm0,64(%rdi) |
| 1346 vmovdqu %ymm4,96(%rdi) |
| 1347 leaq 128(%rdi),%rdi |
| 1348 |
| 1349 subq $512,%rdx |
| 1350 jnz L$oop_outer8x |
| 1351 |
| 1352 jmp L$done8x |
| 1353 |
| 1354 L$tail8x: |
| 1355 cmpq $448,%rdx |
| 1356 jae L$448_or_more8x |
| 1357 cmpq $384,%rdx |
| 1358 jae L$384_or_more8x |
| 1359 cmpq $320,%rdx |
| 1360 jae L$320_or_more8x |
| 1361 cmpq $256,%rdx |
| 1362 jae L$256_or_more8x |
| 1363 cmpq $192,%rdx |
| 1364 jae L$192_or_more8x |
| 1365 cmpq $128,%rdx |
| 1366 jae L$128_or_more8x |
| 1367 cmpq $64,%rdx |
| 1368 jae L$64_or_more8x |
| 1369 |
| 1370 xorq %r10,%r10 |
| 1371 vmovdqa %ymm6,0(%rsp) |
| 1372 vmovdqa %ymm8,32(%rsp) |
| 1373 jmp L$oop_tail8x |
| 1374 |
| 1375 .p2align 5 |
| 1376 L$64_or_more8x: |
| 1377 vpxor 0(%rsi),%ymm6,%ymm6 |
| 1378 vpxor 32(%rsi),%ymm8,%ymm8 |
| 1379 vmovdqu %ymm6,0(%rdi) |
| 1380 vmovdqu %ymm8,32(%rdi) |
| 1381 je L$done8x |
| 1382 |
| 1383 leaq 64(%rsi),%rsi |
| 1384 xorq %r10,%r10 |
| 1385 vmovdqa %ymm1,0(%rsp) |
| 1386 leaq 64(%rdi),%rdi |
| 1387 subq $64,%rdx |
| 1388 vmovdqa %ymm5,32(%rsp) |
| 1389 jmp L$oop_tail8x |
| 1390 |
| 1391 .p2align 5 |
| 1392 L$128_or_more8x: |
| 1393 vpxor 0(%rsi),%ymm6,%ymm6 |
| 1394 vpxor 32(%rsi),%ymm8,%ymm8 |
| 1395 vpxor 64(%rsi),%ymm1,%ymm1 |
| 1396 vpxor 96(%rsi),%ymm5,%ymm5 |
| 1397 vmovdqu %ymm6,0(%rdi) |
| 1398 vmovdqu %ymm8,32(%rdi) |
| 1399 vmovdqu %ymm1,64(%rdi) |
| 1400 vmovdqu %ymm5,96(%rdi) |
| 1401 je L$done8x |
| 1402 |
| 1403 leaq 128(%rsi),%rsi |
| 1404 xorq %r10,%r10 |
| 1405 vmovdqa %ymm12,0(%rsp) |
| 1406 leaq 128(%rdi),%rdi |
| 1407 subq $128,%rdx |
| 1408 vmovdqa %ymm13,32(%rsp) |
| 1409 jmp L$oop_tail8x |
| 1410 |
| 1411 .p2align 5 |
| 1412 L$192_or_more8x: |
| 1413 vpxor 0(%rsi),%ymm6,%ymm6 |
| 1414 vpxor 32(%rsi),%ymm8,%ymm8 |
| 1415 vpxor 64(%rsi),%ymm1,%ymm1 |
| 1416 vpxor 96(%rsi),%ymm5,%ymm5 |
| 1417 vpxor 128(%rsi),%ymm12,%ymm12 |
| 1418 vpxor 160(%rsi),%ymm13,%ymm13 |
| 1419 vmovdqu %ymm6,0(%rdi) |
| 1420 vmovdqu %ymm8,32(%rdi) |
| 1421 vmovdqu %ymm1,64(%rdi) |
| 1422 vmovdqu %ymm5,96(%rdi) |
| 1423 vmovdqu %ymm12,128(%rdi) |
| 1424 vmovdqu %ymm13,160(%rdi) |
| 1425 je L$done8x |
| 1426 |
| 1427 leaq 192(%rsi),%rsi |
| 1428 xorq %r10,%r10 |
| 1429 vmovdqa %ymm10,0(%rsp) |
| 1430 leaq 192(%rdi),%rdi |
| 1431 subq $192,%rdx |
| 1432 vmovdqa %ymm15,32(%rsp) |
| 1433 jmp L$oop_tail8x |
| 1434 |
| 1435 .p2align 5 |
| 1436 L$256_or_more8x: |
| 1437 vpxor 0(%rsi),%ymm6,%ymm6 |
| 1438 vpxor 32(%rsi),%ymm8,%ymm8 |
| 1439 vpxor 64(%rsi),%ymm1,%ymm1 |
| 1440 vpxor 96(%rsi),%ymm5,%ymm5 |
| 1441 vpxor 128(%rsi),%ymm12,%ymm12 |
| 1442 vpxor 160(%rsi),%ymm13,%ymm13 |
| 1443 vpxor 192(%rsi),%ymm10,%ymm10 |
| 1444 vpxor 224(%rsi),%ymm15,%ymm15 |
| 1445 vmovdqu %ymm6,0(%rdi) |
| 1446 vmovdqu %ymm8,32(%rdi) |
| 1447 vmovdqu %ymm1,64(%rdi) |
| 1448 vmovdqu %ymm5,96(%rdi) |
| 1449 vmovdqu %ymm12,128(%rdi) |
| 1450 vmovdqu %ymm13,160(%rdi) |
| 1451 vmovdqu %ymm10,192(%rdi) |
| 1452 vmovdqu %ymm15,224(%rdi) |
| 1453 je L$done8x |
| 1454 |
| 1455 leaq 256(%rsi),%rsi |
| 1456 xorq %r10,%r10 |
| 1457 vmovdqa %ymm14,0(%rsp) |
| 1458 leaq 256(%rdi),%rdi |
| 1459 subq $256,%rdx |
| 1460 vmovdqa %ymm2,32(%rsp) |
| 1461 jmp L$oop_tail8x |
| 1462 |
| 1463 .p2align 5 |
| 1464 L$320_or_more8x: |
| 1465 vpxor 0(%rsi),%ymm6,%ymm6 |
| 1466 vpxor 32(%rsi),%ymm8,%ymm8 |
| 1467 vpxor 64(%rsi),%ymm1,%ymm1 |
| 1468 vpxor 96(%rsi),%ymm5,%ymm5 |
| 1469 vpxor 128(%rsi),%ymm12,%ymm12 |
| 1470 vpxor 160(%rsi),%ymm13,%ymm13 |
| 1471 vpxor 192(%rsi),%ymm10,%ymm10 |
| 1472 vpxor 224(%rsi),%ymm15,%ymm15 |
| 1473 vpxor 256(%rsi),%ymm14,%ymm14 |
| 1474 vpxor 288(%rsi),%ymm2,%ymm2 |
| 1475 vmovdqu %ymm6,0(%rdi) |
| 1476 vmovdqu %ymm8,32(%rdi) |
| 1477 vmovdqu %ymm1,64(%rdi) |
| 1478 vmovdqu %ymm5,96(%rdi) |
| 1479 vmovdqu %ymm12,128(%rdi) |
| 1480 vmovdqu %ymm13,160(%rdi) |
| 1481 vmovdqu %ymm10,192(%rdi) |
| 1482 vmovdqu %ymm15,224(%rdi) |
| 1483 vmovdqu %ymm14,256(%rdi) |
| 1484 vmovdqu %ymm2,288(%rdi) |
| 1485 je L$done8x |
| 1486 |
| 1487 leaq 320(%rsi),%rsi |
| 1488 xorq %r10,%r10 |
| 1489 vmovdqa %ymm3,0(%rsp) |
| 1490 leaq 320(%rdi),%rdi |
| 1491 subq $320,%rdx |
| 1492 vmovdqa %ymm7,32(%rsp) |
| 1493 jmp L$oop_tail8x |
| 1494 |
| 1495 .p2align 5 |
| 1496 L$384_or_more8x: |
| 1497 vpxor 0(%rsi),%ymm6,%ymm6 |
| 1498 vpxor 32(%rsi),%ymm8,%ymm8 |
| 1499 vpxor 64(%rsi),%ymm1,%ymm1 |
| 1500 vpxor 96(%rsi),%ymm5,%ymm5 |
| 1501 vpxor 128(%rsi),%ymm12,%ymm12 |
| 1502 vpxor 160(%rsi),%ymm13,%ymm13 |
| 1503 vpxor 192(%rsi),%ymm10,%ymm10 |
| 1504 vpxor 224(%rsi),%ymm15,%ymm15 |
| 1505 vpxor 256(%rsi),%ymm14,%ymm14 |
| 1506 vpxor 288(%rsi),%ymm2,%ymm2 |
| 1507 vpxor 320(%rsi),%ymm3,%ymm3 |
| 1508 vpxor 352(%rsi),%ymm7,%ymm7 |
| 1509 vmovdqu %ymm6,0(%rdi) |
| 1510 vmovdqu %ymm8,32(%rdi) |
| 1511 vmovdqu %ymm1,64(%rdi) |
| 1512 vmovdqu %ymm5,96(%rdi) |
| 1513 vmovdqu %ymm12,128(%rdi) |
| 1514 vmovdqu %ymm13,160(%rdi) |
| 1515 vmovdqu %ymm10,192(%rdi) |
| 1516 vmovdqu %ymm15,224(%rdi) |
| 1517 vmovdqu %ymm14,256(%rdi) |
| 1518 vmovdqu %ymm2,288(%rdi) |
| 1519 vmovdqu %ymm3,320(%rdi) |
| 1520 vmovdqu %ymm7,352(%rdi) |
| 1521 je L$done8x |
| 1522 |
| 1523 leaq 384(%rsi),%rsi |
| 1524 xorq %r10,%r10 |
| 1525 vmovdqa %ymm11,0(%rsp) |
| 1526 leaq 384(%rdi),%rdi |
| 1527 subq $384,%rdx |
| 1528 vmovdqa %ymm9,32(%rsp) |
| 1529 jmp L$oop_tail8x |
| 1530 |
| 1531 .p2align 5 |
| 1532 L$448_or_more8x: |
| 1533 vpxor 0(%rsi),%ymm6,%ymm6 |
| 1534 vpxor 32(%rsi),%ymm8,%ymm8 |
| 1535 vpxor 64(%rsi),%ymm1,%ymm1 |
| 1536 vpxor 96(%rsi),%ymm5,%ymm5 |
| 1537 vpxor 128(%rsi),%ymm12,%ymm12 |
| 1538 vpxor 160(%rsi),%ymm13,%ymm13 |
| 1539 vpxor 192(%rsi),%ymm10,%ymm10 |
| 1540 vpxor 224(%rsi),%ymm15,%ymm15 |
| 1541 vpxor 256(%rsi),%ymm14,%ymm14 |
| 1542 vpxor 288(%rsi),%ymm2,%ymm2 |
| 1543 vpxor 320(%rsi),%ymm3,%ymm3 |
| 1544 vpxor 352(%rsi),%ymm7,%ymm7 |
| 1545 vpxor 384(%rsi),%ymm11,%ymm11 |
| 1546 vpxor 416(%rsi),%ymm9,%ymm9 |
| 1547 vmovdqu %ymm6,0(%rdi) |
| 1548 vmovdqu %ymm8,32(%rdi) |
| 1549 vmovdqu %ymm1,64(%rdi) |
| 1550 vmovdqu %ymm5,96(%rdi) |
| 1551 vmovdqu %ymm12,128(%rdi) |
| 1552 vmovdqu %ymm13,160(%rdi) |
| 1553 vmovdqu %ymm10,192(%rdi) |
| 1554 vmovdqu %ymm15,224(%rdi) |
| 1555 vmovdqu %ymm14,256(%rdi) |
| 1556 vmovdqu %ymm2,288(%rdi) |
| 1557 vmovdqu %ymm3,320(%rdi) |
| 1558 vmovdqu %ymm7,352(%rdi) |
| 1559 vmovdqu %ymm11,384(%rdi) |
| 1560 vmovdqu %ymm9,416(%rdi) |
| 1561 je L$done8x |
| 1562 |
| 1563 leaq 448(%rsi),%rsi |
| 1564 xorq %r10,%r10 |
| 1565 vmovdqa %ymm0,0(%rsp) |
| 1566 leaq 448(%rdi),%rdi |
| 1567 subq $448,%rdx |
| 1568 vmovdqa %ymm4,32(%rsp) |
| 1569 |
| 1570 L$oop_tail8x: |
| 1571 movzbl (%rsi,%r10,1),%eax |
| 1572 movzbl (%rsp,%r10,1),%ecx |
| 1573 leaq 1(%r10),%r10 |
| 1574 xorl %ecx,%eax |
| 1575 movb %al,-1(%rdi,%r10,1) |
| 1576 decq %rdx |
| 1577 jnz L$oop_tail8x |
| 1578 |
| 1579 L$done8x: |
| 1580 vzeroall |
| 1581 movq 640(%rsp),%rsp |
| 1582 .byte 0xf3,0xc3 |
| 1583 |
| 1584 #endif |
OLD | NEW |