Robert Sloan | c9abfe4 | 2018-11-26 12:19:07 -0800 | [diff] [blame] | 1 | # This file is generated from a similarly-named Perl script in the BoringSSL |
| 2 | # source tree. Do not edit by hand. |
| 3 | |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 4 | #if defined(__i386__) |
Robert Sloan | 6e8c959 | 2018-12-03 11:20:49 -0800 | [diff] [blame] | 5 | #if defined(BORINGSSL_PREFIX) |
| 6 | #include <boringssl_prefix_symbols_asm.h> |
| 7 | #endif |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 8 | .text |
| 9 | .globl _ChaCha20_ctr32 |
| 10 | .private_extern _ChaCha20_ctr32 |
| 11 | .align 4 |
| 12 | _ChaCha20_ctr32: |
| 13 | L_ChaCha20_ctr32_begin: |
| 14 | pushl %ebp |
| 15 | pushl %ebx |
| 16 | pushl %esi |
| 17 | pushl %edi |
| 18 | xorl %eax,%eax |
| 19 | cmpl 28(%esp),%eax |
| 20 | je L000no_data |
| 21 | call Lpic_point |
| 22 | Lpic_point: |
| 23 | popl %eax |
| 24 | movl L_OPENSSL_ia32cap_P$non_lazy_ptr-Lpic_point(%eax),%ebp |
| 25 | testl $16777216,(%ebp) |
| 26 | jz L001x86 |
| 27 | testl $512,4(%ebp) |
| 28 | jz L001x86 |
| 29 | jmp Lssse3_shortcut |
| 30 | L001x86: |
| 31 | movl 32(%esp),%esi |
| 32 | movl 36(%esp),%edi |
| 33 | subl $132,%esp |
| 34 | movl (%esi),%eax |
| 35 | movl 4(%esi),%ebx |
| 36 | movl 8(%esi),%ecx |
| 37 | movl 12(%esi),%edx |
| 38 | movl %eax,80(%esp) |
| 39 | movl %ebx,84(%esp) |
| 40 | movl %ecx,88(%esp) |
| 41 | movl %edx,92(%esp) |
| 42 | movl 16(%esi),%eax |
| 43 | movl 20(%esi),%ebx |
| 44 | movl 24(%esi),%ecx |
| 45 | movl 28(%esi),%edx |
| 46 | movl %eax,96(%esp) |
| 47 | movl %ebx,100(%esp) |
| 48 | movl %ecx,104(%esp) |
| 49 | movl %edx,108(%esp) |
| 50 | movl (%edi),%eax |
| 51 | movl 4(%edi),%ebx |
| 52 | movl 8(%edi),%ecx |
| 53 | movl 12(%edi),%edx |
| 54 | subl $1,%eax |
| 55 | movl %eax,112(%esp) |
| 56 | movl %ebx,116(%esp) |
| 57 | movl %ecx,120(%esp) |
| 58 | movl %edx,124(%esp) |
| 59 | jmp L002entry |
| 60 | .align 4,0x90 |
| 61 | L003outer_loop: |
| 62 | movl %ebx,156(%esp) |
| 63 | movl %eax,152(%esp) |
| 64 | movl %ecx,160(%esp) |
| 65 | L002entry: |
| 66 | movl $1634760805,%eax |
| 67 | movl $857760878,4(%esp) |
| 68 | movl $2036477234,8(%esp) |
| 69 | movl $1797285236,12(%esp) |
| 70 | movl 84(%esp),%ebx |
| 71 | movl 88(%esp),%ebp |
| 72 | movl 104(%esp),%ecx |
| 73 | movl 108(%esp),%esi |
| 74 | movl 116(%esp),%edx |
| 75 | movl 120(%esp),%edi |
| 76 | movl %ebx,20(%esp) |
| 77 | movl %ebp,24(%esp) |
| 78 | movl %ecx,40(%esp) |
| 79 | movl %esi,44(%esp) |
| 80 | movl %edx,52(%esp) |
| 81 | movl %edi,56(%esp) |
| 82 | movl 92(%esp),%ebx |
| 83 | movl 124(%esp),%edi |
| 84 | movl 112(%esp),%edx |
| 85 | movl 80(%esp),%ebp |
| 86 | movl 96(%esp),%ecx |
| 87 | movl 100(%esp),%esi |
| 88 | addl $1,%edx |
| 89 | movl %ebx,28(%esp) |
| 90 | movl %edi,60(%esp) |
| 91 | movl %edx,112(%esp) |
| 92 | movl $10,%ebx |
| 93 | jmp L004loop |
| 94 | .align 4,0x90 |
| 95 | L004loop: |
| 96 | addl %ebp,%eax |
| 97 | movl %ebx,128(%esp) |
| 98 | movl %ebp,%ebx |
| 99 | xorl %eax,%edx |
| 100 | roll $16,%edx |
| 101 | addl %edx,%ecx |
| 102 | xorl %ecx,%ebx |
| 103 | movl 52(%esp),%edi |
| 104 | roll $12,%ebx |
| 105 | movl 20(%esp),%ebp |
| 106 | addl %ebx,%eax |
| 107 | xorl %eax,%edx |
| 108 | movl %eax,(%esp) |
| 109 | roll $8,%edx |
| 110 | movl 4(%esp),%eax |
| 111 | addl %edx,%ecx |
| 112 | movl %edx,48(%esp) |
| 113 | xorl %ecx,%ebx |
| 114 | addl %ebp,%eax |
| 115 | roll $7,%ebx |
| 116 | xorl %eax,%edi |
| 117 | movl %ecx,32(%esp) |
| 118 | roll $16,%edi |
| 119 | movl %ebx,16(%esp) |
| 120 | addl %edi,%esi |
| 121 | movl 40(%esp),%ecx |
| 122 | xorl %esi,%ebp |
| 123 | movl 56(%esp),%edx |
| 124 | roll $12,%ebp |
| 125 | movl 24(%esp),%ebx |
| 126 | addl %ebp,%eax |
| 127 | xorl %eax,%edi |
| 128 | movl %eax,4(%esp) |
| 129 | roll $8,%edi |
| 130 | movl 8(%esp),%eax |
| 131 | addl %edi,%esi |
| 132 | movl %edi,52(%esp) |
| 133 | xorl %esi,%ebp |
| 134 | addl %ebx,%eax |
| 135 | roll $7,%ebp |
| 136 | xorl %eax,%edx |
| 137 | movl %esi,36(%esp) |
| 138 | roll $16,%edx |
| 139 | movl %ebp,20(%esp) |
| 140 | addl %edx,%ecx |
| 141 | movl 44(%esp),%esi |
| 142 | xorl %ecx,%ebx |
| 143 | movl 60(%esp),%edi |
| 144 | roll $12,%ebx |
| 145 | movl 28(%esp),%ebp |
| 146 | addl %ebx,%eax |
| 147 | xorl %eax,%edx |
| 148 | movl %eax,8(%esp) |
| 149 | roll $8,%edx |
| 150 | movl 12(%esp),%eax |
| 151 | addl %edx,%ecx |
| 152 | movl %edx,56(%esp) |
| 153 | xorl %ecx,%ebx |
| 154 | addl %ebp,%eax |
| 155 | roll $7,%ebx |
| 156 | xorl %eax,%edi |
| 157 | roll $16,%edi |
| 158 | movl %ebx,24(%esp) |
| 159 | addl %edi,%esi |
| 160 | xorl %esi,%ebp |
| 161 | roll $12,%ebp |
| 162 | movl 20(%esp),%ebx |
| 163 | addl %ebp,%eax |
| 164 | xorl %eax,%edi |
| 165 | movl %eax,12(%esp) |
| 166 | roll $8,%edi |
| 167 | movl (%esp),%eax |
| 168 | addl %edi,%esi |
| 169 | movl %edi,%edx |
| 170 | xorl %esi,%ebp |
| 171 | addl %ebx,%eax |
| 172 | roll $7,%ebp |
| 173 | xorl %eax,%edx |
| 174 | roll $16,%edx |
| 175 | movl %ebp,28(%esp) |
| 176 | addl %edx,%ecx |
| 177 | xorl %ecx,%ebx |
| 178 | movl 48(%esp),%edi |
| 179 | roll $12,%ebx |
| 180 | movl 24(%esp),%ebp |
| 181 | addl %ebx,%eax |
| 182 | xorl %eax,%edx |
| 183 | movl %eax,(%esp) |
| 184 | roll $8,%edx |
| 185 | movl 4(%esp),%eax |
| 186 | addl %edx,%ecx |
| 187 | movl %edx,60(%esp) |
| 188 | xorl %ecx,%ebx |
| 189 | addl %ebp,%eax |
| 190 | roll $7,%ebx |
| 191 | xorl %eax,%edi |
| 192 | movl %ecx,40(%esp) |
| 193 | roll $16,%edi |
| 194 | movl %ebx,20(%esp) |
| 195 | addl %edi,%esi |
| 196 | movl 32(%esp),%ecx |
| 197 | xorl %esi,%ebp |
| 198 | movl 52(%esp),%edx |
| 199 | roll $12,%ebp |
| 200 | movl 28(%esp),%ebx |
| 201 | addl %ebp,%eax |
| 202 | xorl %eax,%edi |
| 203 | movl %eax,4(%esp) |
| 204 | roll $8,%edi |
| 205 | movl 8(%esp),%eax |
| 206 | addl %edi,%esi |
| 207 | movl %edi,48(%esp) |
| 208 | xorl %esi,%ebp |
| 209 | addl %ebx,%eax |
| 210 | roll $7,%ebp |
| 211 | xorl %eax,%edx |
| 212 | movl %esi,44(%esp) |
| 213 | roll $16,%edx |
| 214 | movl %ebp,24(%esp) |
| 215 | addl %edx,%ecx |
| 216 | movl 36(%esp),%esi |
| 217 | xorl %ecx,%ebx |
| 218 | movl 56(%esp),%edi |
| 219 | roll $12,%ebx |
| 220 | movl 16(%esp),%ebp |
| 221 | addl %ebx,%eax |
| 222 | xorl %eax,%edx |
| 223 | movl %eax,8(%esp) |
| 224 | roll $8,%edx |
| 225 | movl 12(%esp),%eax |
| 226 | addl %edx,%ecx |
| 227 | movl %edx,52(%esp) |
| 228 | xorl %ecx,%ebx |
| 229 | addl %ebp,%eax |
| 230 | roll $7,%ebx |
| 231 | xorl %eax,%edi |
| 232 | roll $16,%edi |
| 233 | movl %ebx,28(%esp) |
| 234 | addl %edi,%esi |
| 235 | xorl %esi,%ebp |
| 236 | movl 48(%esp),%edx |
| 237 | roll $12,%ebp |
| 238 | movl 128(%esp),%ebx |
| 239 | addl %ebp,%eax |
| 240 | xorl %eax,%edi |
| 241 | movl %eax,12(%esp) |
| 242 | roll $8,%edi |
| 243 | movl (%esp),%eax |
| 244 | addl %edi,%esi |
| 245 | movl %edi,56(%esp) |
| 246 | xorl %esi,%ebp |
| 247 | roll $7,%ebp |
| 248 | decl %ebx |
| 249 | jnz L004loop |
| 250 | movl 160(%esp),%ebx |
| 251 | addl $1634760805,%eax |
| 252 | addl 80(%esp),%ebp |
| 253 | addl 96(%esp),%ecx |
| 254 | addl 100(%esp),%esi |
| 255 | cmpl $64,%ebx |
| 256 | jb L005tail |
| 257 | movl 156(%esp),%ebx |
| 258 | addl 112(%esp),%edx |
| 259 | addl 120(%esp),%edi |
| 260 | xorl (%ebx),%eax |
| 261 | xorl 16(%ebx),%ebp |
| 262 | movl %eax,(%esp) |
| 263 | movl 152(%esp),%eax |
| 264 | xorl 32(%ebx),%ecx |
| 265 | xorl 36(%ebx),%esi |
| 266 | xorl 48(%ebx),%edx |
| 267 | xorl 56(%ebx),%edi |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 268 | movl %ebp,16(%eax) |
| 269 | movl %ecx,32(%eax) |
| 270 | movl %esi,36(%eax) |
| 271 | movl %edx,48(%eax) |
| 272 | movl %edi,56(%eax) |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 273 | movl 4(%esp),%ebp |
| 274 | movl 8(%esp),%ecx |
| 275 | movl 12(%esp),%esi |
| 276 | movl 20(%esp),%edx |
| 277 | movl 24(%esp),%edi |
| 278 | addl $857760878,%ebp |
| 279 | addl $2036477234,%ecx |
| 280 | addl $1797285236,%esi |
| 281 | addl 84(%esp),%edx |
| 282 | addl 88(%esp),%edi |
| 283 | xorl 4(%ebx),%ebp |
| 284 | xorl 8(%ebx),%ecx |
| 285 | xorl 12(%ebx),%esi |
| 286 | xorl 20(%ebx),%edx |
| 287 | xorl 24(%ebx),%edi |
| 288 | movl %ebp,4(%eax) |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 289 | movl %ecx,8(%eax) |
| 290 | movl %esi,12(%eax) |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 291 | movl %edx,20(%eax) |
| 292 | movl %edi,24(%eax) |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 293 | movl 28(%esp),%ebp |
| 294 | movl 40(%esp),%ecx |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 295 | movl 44(%esp),%esi |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 296 | movl 52(%esp),%edx |
| 297 | movl 60(%esp),%edi |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 298 | addl 92(%esp),%ebp |
| 299 | addl 104(%esp),%ecx |
| 300 | addl 108(%esp),%esi |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 301 | addl 116(%esp),%edx |
| 302 | addl 124(%esp),%edi |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 303 | xorl 28(%ebx),%ebp |
| 304 | xorl 40(%ebx),%ecx |
| 305 | xorl 44(%ebx),%esi |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 306 | xorl 52(%ebx),%edx |
| 307 | xorl 60(%ebx),%edi |
| 308 | leal 64(%ebx),%ebx |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 309 | movl %ebp,28(%eax) |
| 310 | movl (%esp),%ebp |
| 311 | movl %ecx,40(%eax) |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 312 | movl 160(%esp),%ecx |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 313 | movl %esi,44(%eax) |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 314 | movl %edx,52(%eax) |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 315 | movl %edi,60(%eax) |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 316 | movl %ebp,(%eax) |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 317 | leal 64(%eax),%eax |
| 318 | subl $64,%ecx |
| 319 | jnz L003outer_loop |
| 320 | jmp L006done |
| 321 | L005tail: |
| 322 | addl 112(%esp),%edx |
| 323 | addl 120(%esp),%edi |
| 324 | movl %eax,(%esp) |
| 325 | movl %ebp,16(%esp) |
| 326 | movl %ecx,32(%esp) |
| 327 | movl %esi,36(%esp) |
| 328 | movl %edx,48(%esp) |
| 329 | movl %edi,56(%esp) |
| 330 | movl 4(%esp),%ebp |
| 331 | movl 8(%esp),%ecx |
| 332 | movl 12(%esp),%esi |
| 333 | movl 20(%esp),%edx |
| 334 | movl 24(%esp),%edi |
| 335 | addl $857760878,%ebp |
| 336 | addl $2036477234,%ecx |
| 337 | addl $1797285236,%esi |
| 338 | addl 84(%esp),%edx |
| 339 | addl 88(%esp),%edi |
| 340 | movl %ebp,4(%esp) |
| 341 | movl %ecx,8(%esp) |
| 342 | movl %esi,12(%esp) |
| 343 | movl %edx,20(%esp) |
| 344 | movl %edi,24(%esp) |
| 345 | movl 28(%esp),%ebp |
| 346 | movl 40(%esp),%ecx |
| 347 | movl 44(%esp),%esi |
| 348 | movl 52(%esp),%edx |
| 349 | movl 60(%esp),%edi |
| 350 | addl 92(%esp),%ebp |
| 351 | addl 104(%esp),%ecx |
| 352 | addl 108(%esp),%esi |
| 353 | addl 116(%esp),%edx |
| 354 | addl 124(%esp),%edi |
| 355 | movl %ebp,28(%esp) |
| 356 | movl 156(%esp),%ebp |
| 357 | movl %ecx,40(%esp) |
| 358 | movl 152(%esp),%ecx |
| 359 | movl %esi,44(%esp) |
| 360 | xorl %esi,%esi |
| 361 | movl %edx,52(%esp) |
| 362 | movl %edi,60(%esp) |
| 363 | xorl %eax,%eax |
| 364 | xorl %edx,%edx |
| 365 | L007tail_loop: |
| 366 | movb (%esi,%ebp,1),%al |
| 367 | movb (%esp,%esi,1),%dl |
| 368 | leal 1(%esi),%esi |
| 369 | xorb %dl,%al |
| 370 | movb %al,-1(%ecx,%esi,1) |
| 371 | decl %ebx |
| 372 | jnz L007tail_loop |
| 373 | L006done: |
| 374 | addl $132,%esp |
| 375 | L000no_data: |
| 376 | popl %edi |
| 377 | popl %esi |
| 378 | popl %ebx |
| 379 | popl %ebp |
| 380 | ret |
| 381 | .globl _ChaCha20_ssse3 |
| 382 | .private_extern _ChaCha20_ssse3 |
| 383 | .align 4 |
| 384 | _ChaCha20_ssse3: |
| 385 | L_ChaCha20_ssse3_begin: |
| 386 | pushl %ebp |
| 387 | pushl %ebx |
| 388 | pushl %esi |
| 389 | pushl %edi |
| 390 | Lssse3_shortcut: |
| 391 | movl 20(%esp),%edi |
| 392 | movl 24(%esp),%esi |
| 393 | movl 28(%esp),%ecx |
| 394 | movl 32(%esp),%edx |
| 395 | movl 36(%esp),%ebx |
| 396 | movl %esp,%ebp |
| 397 | subl $524,%esp |
| 398 | andl $-64,%esp |
| 399 | movl %ebp,512(%esp) |
| 400 | leal Lssse3_data-Lpic_point(%eax),%eax |
| 401 | movdqu (%ebx),%xmm3 |
| 402 | cmpl $256,%ecx |
| 403 | jb L0081x |
| 404 | movl %edx,516(%esp) |
| 405 | movl %ebx,520(%esp) |
| 406 | subl $256,%ecx |
| 407 | leal 384(%esp),%ebp |
| 408 | movdqu (%edx),%xmm7 |
| 409 | pshufd $0,%xmm3,%xmm0 |
| 410 | pshufd $85,%xmm3,%xmm1 |
| 411 | pshufd $170,%xmm3,%xmm2 |
| 412 | pshufd $255,%xmm3,%xmm3 |
| 413 | paddd 48(%eax),%xmm0 |
| 414 | pshufd $0,%xmm7,%xmm4 |
| 415 | pshufd $85,%xmm7,%xmm5 |
| 416 | psubd 64(%eax),%xmm0 |
| 417 | pshufd $170,%xmm7,%xmm6 |
| 418 | pshufd $255,%xmm7,%xmm7 |
| 419 | movdqa %xmm0,64(%ebp) |
| 420 | movdqa %xmm1,80(%ebp) |
| 421 | movdqa %xmm2,96(%ebp) |
| 422 | movdqa %xmm3,112(%ebp) |
| 423 | movdqu 16(%edx),%xmm3 |
| 424 | movdqa %xmm4,-64(%ebp) |
| 425 | movdqa %xmm5,-48(%ebp) |
| 426 | movdqa %xmm6,-32(%ebp) |
| 427 | movdqa %xmm7,-16(%ebp) |
| 428 | movdqa 32(%eax),%xmm7 |
| 429 | leal 128(%esp),%ebx |
| 430 | pshufd $0,%xmm3,%xmm0 |
| 431 | pshufd $85,%xmm3,%xmm1 |
| 432 | pshufd $170,%xmm3,%xmm2 |
| 433 | pshufd $255,%xmm3,%xmm3 |
| 434 | pshufd $0,%xmm7,%xmm4 |
| 435 | pshufd $85,%xmm7,%xmm5 |
| 436 | pshufd $170,%xmm7,%xmm6 |
| 437 | pshufd $255,%xmm7,%xmm7 |
| 438 | movdqa %xmm0,(%ebp) |
| 439 | movdqa %xmm1,16(%ebp) |
| 440 | movdqa %xmm2,32(%ebp) |
| 441 | movdqa %xmm3,48(%ebp) |
| 442 | movdqa %xmm4,-128(%ebp) |
| 443 | movdqa %xmm5,-112(%ebp) |
| 444 | movdqa %xmm6,-96(%ebp) |
| 445 | movdqa %xmm7,-80(%ebp) |
| 446 | leal 128(%esi),%esi |
| 447 | leal 128(%edi),%edi |
| 448 | jmp L009outer_loop |
| 449 | .align 4,0x90 |
| 450 | L009outer_loop: |
| 451 | movdqa -112(%ebp),%xmm1 |
| 452 | movdqa -96(%ebp),%xmm2 |
| 453 | movdqa -80(%ebp),%xmm3 |
| 454 | movdqa -48(%ebp),%xmm5 |
| 455 | movdqa -32(%ebp),%xmm6 |
| 456 | movdqa -16(%ebp),%xmm7 |
| 457 | movdqa %xmm1,-112(%ebx) |
| 458 | movdqa %xmm2,-96(%ebx) |
| 459 | movdqa %xmm3,-80(%ebx) |
| 460 | movdqa %xmm5,-48(%ebx) |
| 461 | movdqa %xmm6,-32(%ebx) |
| 462 | movdqa %xmm7,-16(%ebx) |
| 463 | movdqa 32(%ebp),%xmm2 |
| 464 | movdqa 48(%ebp),%xmm3 |
| 465 | movdqa 64(%ebp),%xmm4 |
| 466 | movdqa 80(%ebp),%xmm5 |
| 467 | movdqa 96(%ebp),%xmm6 |
| 468 | movdqa 112(%ebp),%xmm7 |
| 469 | paddd 64(%eax),%xmm4 |
| 470 | movdqa %xmm2,32(%ebx) |
| 471 | movdqa %xmm3,48(%ebx) |
| 472 | movdqa %xmm4,64(%ebx) |
| 473 | movdqa %xmm5,80(%ebx) |
| 474 | movdqa %xmm6,96(%ebx) |
| 475 | movdqa %xmm7,112(%ebx) |
| 476 | movdqa %xmm4,64(%ebp) |
| 477 | movdqa -128(%ebp),%xmm0 |
| 478 | movdqa %xmm4,%xmm6 |
| 479 | movdqa -64(%ebp),%xmm3 |
| 480 | movdqa (%ebp),%xmm4 |
| 481 | movdqa 16(%ebp),%xmm5 |
| 482 | movl $10,%edx |
| 483 | nop |
| 484 | .align 4,0x90 |
| 485 | L010loop: |
| 486 | paddd %xmm3,%xmm0 |
| 487 | movdqa %xmm3,%xmm2 |
| 488 | pxor %xmm0,%xmm6 |
| 489 | pshufb (%eax),%xmm6 |
| 490 | paddd %xmm6,%xmm4 |
| 491 | pxor %xmm4,%xmm2 |
| 492 | movdqa -48(%ebx),%xmm3 |
| 493 | movdqa %xmm2,%xmm1 |
| 494 | pslld $12,%xmm2 |
| 495 | psrld $20,%xmm1 |
| 496 | por %xmm1,%xmm2 |
| 497 | movdqa -112(%ebx),%xmm1 |
| 498 | paddd %xmm2,%xmm0 |
| 499 | movdqa 80(%ebx),%xmm7 |
| 500 | pxor %xmm0,%xmm6 |
| 501 | movdqa %xmm0,-128(%ebx) |
| 502 | pshufb 16(%eax),%xmm6 |
| 503 | paddd %xmm6,%xmm4 |
| 504 | movdqa %xmm6,64(%ebx) |
| 505 | pxor %xmm4,%xmm2 |
| 506 | paddd %xmm3,%xmm1 |
| 507 | movdqa %xmm2,%xmm0 |
| 508 | pslld $7,%xmm2 |
| 509 | psrld $25,%xmm0 |
| 510 | pxor %xmm1,%xmm7 |
| 511 | por %xmm0,%xmm2 |
| 512 | movdqa %xmm4,(%ebx) |
| 513 | pshufb (%eax),%xmm7 |
| 514 | movdqa %xmm2,-64(%ebx) |
| 515 | paddd %xmm7,%xmm5 |
| 516 | movdqa 32(%ebx),%xmm4 |
| 517 | pxor %xmm5,%xmm3 |
| 518 | movdqa -32(%ebx),%xmm2 |
| 519 | movdqa %xmm3,%xmm0 |
| 520 | pslld $12,%xmm3 |
| 521 | psrld $20,%xmm0 |
| 522 | por %xmm0,%xmm3 |
| 523 | movdqa -96(%ebx),%xmm0 |
| 524 | paddd %xmm3,%xmm1 |
| 525 | movdqa 96(%ebx),%xmm6 |
| 526 | pxor %xmm1,%xmm7 |
| 527 | movdqa %xmm1,-112(%ebx) |
| 528 | pshufb 16(%eax),%xmm7 |
| 529 | paddd %xmm7,%xmm5 |
| 530 | movdqa %xmm7,80(%ebx) |
| 531 | pxor %xmm5,%xmm3 |
| 532 | paddd %xmm2,%xmm0 |
| 533 | movdqa %xmm3,%xmm1 |
| 534 | pslld $7,%xmm3 |
| 535 | psrld $25,%xmm1 |
| 536 | pxor %xmm0,%xmm6 |
| 537 | por %xmm1,%xmm3 |
| 538 | movdqa %xmm5,16(%ebx) |
| 539 | pshufb (%eax),%xmm6 |
| 540 | movdqa %xmm3,-48(%ebx) |
| 541 | paddd %xmm6,%xmm4 |
| 542 | movdqa 48(%ebx),%xmm5 |
| 543 | pxor %xmm4,%xmm2 |
| 544 | movdqa -16(%ebx),%xmm3 |
| 545 | movdqa %xmm2,%xmm1 |
| 546 | pslld $12,%xmm2 |
| 547 | psrld $20,%xmm1 |
| 548 | por %xmm1,%xmm2 |
| 549 | movdqa -80(%ebx),%xmm1 |
| 550 | paddd %xmm2,%xmm0 |
| 551 | movdqa 112(%ebx),%xmm7 |
| 552 | pxor %xmm0,%xmm6 |
| 553 | movdqa %xmm0,-96(%ebx) |
| 554 | pshufb 16(%eax),%xmm6 |
| 555 | paddd %xmm6,%xmm4 |
| 556 | movdqa %xmm6,96(%ebx) |
| 557 | pxor %xmm4,%xmm2 |
| 558 | paddd %xmm3,%xmm1 |
| 559 | movdqa %xmm2,%xmm0 |
| 560 | pslld $7,%xmm2 |
| 561 | psrld $25,%xmm0 |
| 562 | pxor %xmm1,%xmm7 |
| 563 | por %xmm0,%xmm2 |
| 564 | pshufb (%eax),%xmm7 |
| 565 | movdqa %xmm2,-32(%ebx) |
| 566 | paddd %xmm7,%xmm5 |
| 567 | pxor %xmm5,%xmm3 |
| 568 | movdqa -48(%ebx),%xmm2 |
| 569 | movdqa %xmm3,%xmm0 |
| 570 | pslld $12,%xmm3 |
| 571 | psrld $20,%xmm0 |
| 572 | por %xmm0,%xmm3 |
| 573 | movdqa -128(%ebx),%xmm0 |
| 574 | paddd %xmm3,%xmm1 |
| 575 | pxor %xmm1,%xmm7 |
| 576 | movdqa %xmm1,-80(%ebx) |
| 577 | pshufb 16(%eax),%xmm7 |
| 578 | paddd %xmm7,%xmm5 |
| 579 | movdqa %xmm7,%xmm6 |
| 580 | pxor %xmm5,%xmm3 |
| 581 | paddd %xmm2,%xmm0 |
| 582 | movdqa %xmm3,%xmm1 |
| 583 | pslld $7,%xmm3 |
| 584 | psrld $25,%xmm1 |
| 585 | pxor %xmm0,%xmm6 |
| 586 | por %xmm1,%xmm3 |
| 587 | pshufb (%eax),%xmm6 |
| 588 | movdqa %xmm3,-16(%ebx) |
| 589 | paddd %xmm6,%xmm4 |
| 590 | pxor %xmm4,%xmm2 |
| 591 | movdqa -32(%ebx),%xmm3 |
| 592 | movdqa %xmm2,%xmm1 |
| 593 | pslld $12,%xmm2 |
| 594 | psrld $20,%xmm1 |
| 595 | por %xmm1,%xmm2 |
| 596 | movdqa -112(%ebx),%xmm1 |
| 597 | paddd %xmm2,%xmm0 |
| 598 | movdqa 64(%ebx),%xmm7 |
| 599 | pxor %xmm0,%xmm6 |
| 600 | movdqa %xmm0,-128(%ebx) |
| 601 | pshufb 16(%eax),%xmm6 |
| 602 | paddd %xmm6,%xmm4 |
| 603 | movdqa %xmm6,112(%ebx) |
| 604 | pxor %xmm4,%xmm2 |
| 605 | paddd %xmm3,%xmm1 |
| 606 | movdqa %xmm2,%xmm0 |
| 607 | pslld $7,%xmm2 |
| 608 | psrld $25,%xmm0 |
| 609 | pxor %xmm1,%xmm7 |
| 610 | por %xmm0,%xmm2 |
| 611 | movdqa %xmm4,32(%ebx) |
| 612 | pshufb (%eax),%xmm7 |
| 613 | movdqa %xmm2,-48(%ebx) |
| 614 | paddd %xmm7,%xmm5 |
| 615 | movdqa (%ebx),%xmm4 |
| 616 | pxor %xmm5,%xmm3 |
| 617 | movdqa -16(%ebx),%xmm2 |
| 618 | movdqa %xmm3,%xmm0 |
| 619 | pslld $12,%xmm3 |
| 620 | psrld $20,%xmm0 |
| 621 | por %xmm0,%xmm3 |
| 622 | movdqa -96(%ebx),%xmm0 |
| 623 | paddd %xmm3,%xmm1 |
| 624 | movdqa 80(%ebx),%xmm6 |
| 625 | pxor %xmm1,%xmm7 |
| 626 | movdqa %xmm1,-112(%ebx) |
| 627 | pshufb 16(%eax),%xmm7 |
| 628 | paddd %xmm7,%xmm5 |
| 629 | movdqa %xmm7,64(%ebx) |
| 630 | pxor %xmm5,%xmm3 |
| 631 | paddd %xmm2,%xmm0 |
| 632 | movdqa %xmm3,%xmm1 |
| 633 | pslld $7,%xmm3 |
| 634 | psrld $25,%xmm1 |
| 635 | pxor %xmm0,%xmm6 |
| 636 | por %xmm1,%xmm3 |
| 637 | movdqa %xmm5,48(%ebx) |
| 638 | pshufb (%eax),%xmm6 |
| 639 | movdqa %xmm3,-32(%ebx) |
| 640 | paddd %xmm6,%xmm4 |
| 641 | movdqa 16(%ebx),%xmm5 |
| 642 | pxor %xmm4,%xmm2 |
| 643 | movdqa -64(%ebx),%xmm3 |
| 644 | movdqa %xmm2,%xmm1 |
| 645 | pslld $12,%xmm2 |
| 646 | psrld $20,%xmm1 |
| 647 | por %xmm1,%xmm2 |
| 648 | movdqa -80(%ebx),%xmm1 |
| 649 | paddd %xmm2,%xmm0 |
| 650 | movdqa 96(%ebx),%xmm7 |
| 651 | pxor %xmm0,%xmm6 |
| 652 | movdqa %xmm0,-96(%ebx) |
| 653 | pshufb 16(%eax),%xmm6 |
| 654 | paddd %xmm6,%xmm4 |
| 655 | movdqa %xmm6,80(%ebx) |
| 656 | pxor %xmm4,%xmm2 |
| 657 | paddd %xmm3,%xmm1 |
| 658 | movdqa %xmm2,%xmm0 |
| 659 | pslld $7,%xmm2 |
| 660 | psrld $25,%xmm0 |
| 661 | pxor %xmm1,%xmm7 |
| 662 | por %xmm0,%xmm2 |
| 663 | pshufb (%eax),%xmm7 |
| 664 | movdqa %xmm2,-16(%ebx) |
| 665 | paddd %xmm7,%xmm5 |
| 666 | pxor %xmm5,%xmm3 |
| 667 | movdqa %xmm3,%xmm0 |
| 668 | pslld $12,%xmm3 |
| 669 | psrld $20,%xmm0 |
| 670 | por %xmm0,%xmm3 |
| 671 | movdqa -128(%ebx),%xmm0 |
| 672 | paddd %xmm3,%xmm1 |
| 673 | movdqa 64(%ebx),%xmm6 |
| 674 | pxor %xmm1,%xmm7 |
| 675 | movdqa %xmm1,-80(%ebx) |
| 676 | pshufb 16(%eax),%xmm7 |
| 677 | paddd %xmm7,%xmm5 |
| 678 | movdqa %xmm7,96(%ebx) |
| 679 | pxor %xmm5,%xmm3 |
| 680 | movdqa %xmm3,%xmm1 |
| 681 | pslld $7,%xmm3 |
| 682 | psrld $25,%xmm1 |
| 683 | por %xmm1,%xmm3 |
| 684 | decl %edx |
| 685 | jnz L010loop |
| 686 | movdqa %xmm3,-64(%ebx) |
| 687 | movdqa %xmm4,(%ebx) |
| 688 | movdqa %xmm5,16(%ebx) |
| 689 | movdqa %xmm6,64(%ebx) |
| 690 | movdqa %xmm7,96(%ebx) |
| 691 | movdqa -112(%ebx),%xmm1 |
| 692 | movdqa -96(%ebx),%xmm2 |
| 693 | movdqa -80(%ebx),%xmm3 |
| 694 | paddd -128(%ebp),%xmm0 |
| 695 | paddd -112(%ebp),%xmm1 |
| 696 | paddd -96(%ebp),%xmm2 |
| 697 | paddd -80(%ebp),%xmm3 |
| 698 | movdqa %xmm0,%xmm6 |
| 699 | punpckldq %xmm1,%xmm0 |
| 700 | movdqa %xmm2,%xmm7 |
| 701 | punpckldq %xmm3,%xmm2 |
| 702 | punpckhdq %xmm1,%xmm6 |
| 703 | punpckhdq %xmm3,%xmm7 |
| 704 | movdqa %xmm0,%xmm1 |
| 705 | punpcklqdq %xmm2,%xmm0 |
| 706 | movdqa %xmm6,%xmm3 |
| 707 | punpcklqdq %xmm7,%xmm6 |
| 708 | punpckhqdq %xmm2,%xmm1 |
| 709 | punpckhqdq %xmm7,%xmm3 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 710 | movdqu -128(%esi),%xmm4 |
| 711 | movdqu -64(%esi),%xmm5 |
| 712 | movdqu (%esi),%xmm2 |
| 713 | movdqu 64(%esi),%xmm7 |
| 714 | leal 16(%esi),%esi |
| 715 | pxor %xmm0,%xmm4 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 716 | movdqa -64(%ebx),%xmm0 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 717 | pxor %xmm1,%xmm5 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 718 | movdqa -48(%ebx),%xmm1 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 719 | pxor %xmm2,%xmm6 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 720 | movdqa -32(%ebx),%xmm2 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 721 | pxor %xmm3,%xmm7 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 722 | movdqa -16(%ebx),%xmm3 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 723 | movdqu %xmm4,-128(%edi) |
| 724 | movdqu %xmm5,-64(%edi) |
| 725 | movdqu %xmm6,(%edi) |
| 726 | movdqu %xmm7,64(%edi) |
| 727 | leal 16(%edi),%edi |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 728 | paddd -64(%ebp),%xmm0 |
| 729 | paddd -48(%ebp),%xmm1 |
| 730 | paddd -32(%ebp),%xmm2 |
| 731 | paddd -16(%ebp),%xmm3 |
| 732 | movdqa %xmm0,%xmm6 |
| 733 | punpckldq %xmm1,%xmm0 |
| 734 | movdqa %xmm2,%xmm7 |
| 735 | punpckldq %xmm3,%xmm2 |
| 736 | punpckhdq %xmm1,%xmm6 |
| 737 | punpckhdq %xmm3,%xmm7 |
| 738 | movdqa %xmm0,%xmm1 |
| 739 | punpcklqdq %xmm2,%xmm0 |
| 740 | movdqa %xmm6,%xmm3 |
| 741 | punpcklqdq %xmm7,%xmm6 |
| 742 | punpckhqdq %xmm2,%xmm1 |
| 743 | punpckhqdq %xmm7,%xmm3 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 744 | movdqu -128(%esi),%xmm4 |
| 745 | movdqu -64(%esi),%xmm5 |
| 746 | movdqu (%esi),%xmm2 |
| 747 | movdqu 64(%esi),%xmm7 |
| 748 | leal 16(%esi),%esi |
| 749 | pxor %xmm0,%xmm4 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 750 | movdqa (%ebx),%xmm0 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 751 | pxor %xmm1,%xmm5 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 752 | movdqa 16(%ebx),%xmm1 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 753 | pxor %xmm2,%xmm6 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 754 | movdqa 32(%ebx),%xmm2 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 755 | pxor %xmm3,%xmm7 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 756 | movdqa 48(%ebx),%xmm3 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 757 | movdqu %xmm4,-128(%edi) |
| 758 | movdqu %xmm5,-64(%edi) |
| 759 | movdqu %xmm6,(%edi) |
| 760 | movdqu %xmm7,64(%edi) |
| 761 | leal 16(%edi),%edi |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 762 | paddd (%ebp),%xmm0 |
| 763 | paddd 16(%ebp),%xmm1 |
| 764 | paddd 32(%ebp),%xmm2 |
| 765 | paddd 48(%ebp),%xmm3 |
| 766 | movdqa %xmm0,%xmm6 |
| 767 | punpckldq %xmm1,%xmm0 |
| 768 | movdqa %xmm2,%xmm7 |
| 769 | punpckldq %xmm3,%xmm2 |
| 770 | punpckhdq %xmm1,%xmm6 |
| 771 | punpckhdq %xmm3,%xmm7 |
| 772 | movdqa %xmm0,%xmm1 |
| 773 | punpcklqdq %xmm2,%xmm0 |
| 774 | movdqa %xmm6,%xmm3 |
| 775 | punpcklqdq %xmm7,%xmm6 |
| 776 | punpckhqdq %xmm2,%xmm1 |
| 777 | punpckhqdq %xmm7,%xmm3 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 778 | movdqu -128(%esi),%xmm4 |
| 779 | movdqu -64(%esi),%xmm5 |
| 780 | movdqu (%esi),%xmm2 |
| 781 | movdqu 64(%esi),%xmm7 |
| 782 | leal 16(%esi),%esi |
| 783 | pxor %xmm0,%xmm4 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 784 | movdqa 64(%ebx),%xmm0 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 785 | pxor %xmm1,%xmm5 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 786 | movdqa 80(%ebx),%xmm1 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 787 | pxor %xmm2,%xmm6 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 788 | movdqa 96(%ebx),%xmm2 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 789 | pxor %xmm3,%xmm7 |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 790 | movdqa 112(%ebx),%xmm3 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 791 | movdqu %xmm4,-128(%edi) |
| 792 | movdqu %xmm5,-64(%edi) |
| 793 | movdqu %xmm6,(%edi) |
| 794 | movdqu %xmm7,64(%edi) |
| 795 | leal 16(%edi),%edi |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 796 | paddd 64(%ebp),%xmm0 |
| 797 | paddd 80(%ebp),%xmm1 |
| 798 | paddd 96(%ebp),%xmm2 |
| 799 | paddd 112(%ebp),%xmm3 |
| 800 | movdqa %xmm0,%xmm6 |
| 801 | punpckldq %xmm1,%xmm0 |
| 802 | movdqa %xmm2,%xmm7 |
| 803 | punpckldq %xmm3,%xmm2 |
| 804 | punpckhdq %xmm1,%xmm6 |
| 805 | punpckhdq %xmm3,%xmm7 |
| 806 | movdqa %xmm0,%xmm1 |
| 807 | punpcklqdq %xmm2,%xmm0 |
| 808 | movdqa %xmm6,%xmm3 |
| 809 | punpcklqdq %xmm7,%xmm6 |
| 810 | punpckhqdq %xmm2,%xmm1 |
| 811 | punpckhqdq %xmm7,%xmm3 |
David Benjamin | 6e899c7 | 2016-06-09 18:02:18 -0400 | [diff] [blame] | 812 | movdqu -128(%esi),%xmm4 |
| 813 | movdqu -64(%esi),%xmm5 |
| 814 | movdqu (%esi),%xmm2 |
| 815 | movdqu 64(%esi),%xmm7 |
| 816 | leal 208(%esi),%esi |
| 817 | pxor %xmm0,%xmm4 |
| 818 | pxor %xmm1,%xmm5 |
| 819 | pxor %xmm2,%xmm6 |
| 820 | pxor %xmm3,%xmm7 |
| 821 | movdqu %xmm4,-128(%edi) |
| 822 | movdqu %xmm5,-64(%edi) |
| 823 | movdqu %xmm6,(%edi) |
| 824 | movdqu %xmm7,64(%edi) |
| 825 | leal 208(%edi),%edi |
David Benjamin | 4969cc9 | 2016-04-22 15:02:23 -0400 | [diff] [blame] | 826 | subl $256,%ecx |
| 827 | jnc L009outer_loop |
| 828 | addl $256,%ecx |
| 829 | jz L011done |
| 830 | movl 520(%esp),%ebx |
| 831 | leal -128(%esi),%esi |
| 832 | movl 516(%esp),%edx |
| 833 | leal -128(%edi),%edi |
| 834 | movd 64(%ebp),%xmm2 |
| 835 | movdqu (%ebx),%xmm3 |
| 836 | paddd 96(%eax),%xmm2 |
| 837 | pand 112(%eax),%xmm3 |
| 838 | por %xmm2,%xmm3 |
| 839 | L0081x: |
| 840 | movdqa 32(%eax),%xmm0 |
| 841 | movdqu (%edx),%xmm1 |
| 842 | movdqu 16(%edx),%xmm2 |
| 843 | movdqa (%eax),%xmm6 |
| 844 | movdqa 16(%eax),%xmm7 |
| 845 | movl %ebp,48(%esp) |
| 846 | movdqa %xmm0,(%esp) |
| 847 | movdqa %xmm1,16(%esp) |
| 848 | movdqa %xmm2,32(%esp) |
| 849 | movdqa %xmm3,48(%esp) |
| 850 | movl $10,%edx |
| 851 | jmp L012loop1x |
| 852 | .align 4,0x90 |
| 853 | L013outer1x: |
| 854 | movdqa 80(%eax),%xmm3 |
| 855 | movdqa (%esp),%xmm0 |
| 856 | movdqa 16(%esp),%xmm1 |
| 857 | movdqa 32(%esp),%xmm2 |
| 858 | paddd 48(%esp),%xmm3 |
| 859 | movl $10,%edx |
| 860 | movdqa %xmm3,48(%esp) |
| 861 | jmp L012loop1x |
| 862 | .align 4,0x90 |
| 863 | L012loop1x: |
| 864 | paddd %xmm1,%xmm0 |
| 865 | pxor %xmm0,%xmm3 |
| 866 | .byte 102,15,56,0,222 |
| 867 | paddd %xmm3,%xmm2 |
| 868 | pxor %xmm2,%xmm1 |
| 869 | movdqa %xmm1,%xmm4 |
| 870 | psrld $20,%xmm1 |
| 871 | pslld $12,%xmm4 |
| 872 | por %xmm4,%xmm1 |
| 873 | paddd %xmm1,%xmm0 |
| 874 | pxor %xmm0,%xmm3 |
| 875 | .byte 102,15,56,0,223 |
| 876 | paddd %xmm3,%xmm2 |
| 877 | pxor %xmm2,%xmm1 |
| 878 | movdqa %xmm1,%xmm4 |
| 879 | psrld $25,%xmm1 |
| 880 | pslld $7,%xmm4 |
| 881 | por %xmm4,%xmm1 |
| 882 | pshufd $78,%xmm2,%xmm2 |
| 883 | pshufd $57,%xmm1,%xmm1 |
| 884 | pshufd $147,%xmm3,%xmm3 |
| 885 | nop |
| 886 | paddd %xmm1,%xmm0 |
| 887 | pxor %xmm0,%xmm3 |
| 888 | .byte 102,15,56,0,222 |
| 889 | paddd %xmm3,%xmm2 |
| 890 | pxor %xmm2,%xmm1 |
| 891 | movdqa %xmm1,%xmm4 |
| 892 | psrld $20,%xmm1 |
| 893 | pslld $12,%xmm4 |
| 894 | por %xmm4,%xmm1 |
| 895 | paddd %xmm1,%xmm0 |
| 896 | pxor %xmm0,%xmm3 |
| 897 | .byte 102,15,56,0,223 |
| 898 | paddd %xmm3,%xmm2 |
| 899 | pxor %xmm2,%xmm1 |
| 900 | movdqa %xmm1,%xmm4 |
| 901 | psrld $25,%xmm1 |
| 902 | pslld $7,%xmm4 |
| 903 | por %xmm4,%xmm1 |
| 904 | pshufd $78,%xmm2,%xmm2 |
| 905 | pshufd $147,%xmm1,%xmm1 |
| 906 | pshufd $57,%xmm3,%xmm3 |
| 907 | decl %edx |
| 908 | jnz L012loop1x |
| 909 | paddd (%esp),%xmm0 |
| 910 | paddd 16(%esp),%xmm1 |
| 911 | paddd 32(%esp),%xmm2 |
| 912 | paddd 48(%esp),%xmm3 |
| 913 | cmpl $64,%ecx |
| 914 | jb L014tail |
| 915 | movdqu (%esi),%xmm4 |
| 916 | movdqu 16(%esi),%xmm5 |
| 917 | pxor %xmm4,%xmm0 |
| 918 | movdqu 32(%esi),%xmm4 |
| 919 | pxor %xmm5,%xmm1 |
| 920 | movdqu 48(%esi),%xmm5 |
| 921 | pxor %xmm4,%xmm2 |
| 922 | pxor %xmm5,%xmm3 |
| 923 | leal 64(%esi),%esi |
| 924 | movdqu %xmm0,(%edi) |
| 925 | movdqu %xmm1,16(%edi) |
| 926 | movdqu %xmm2,32(%edi) |
| 927 | movdqu %xmm3,48(%edi) |
| 928 | leal 64(%edi),%edi |
| 929 | subl $64,%ecx |
| 930 | jnz L013outer1x |
| 931 | jmp L011done |
| 932 | L014tail: |
| 933 | movdqa %xmm0,(%esp) |
| 934 | movdqa %xmm1,16(%esp) |
| 935 | movdqa %xmm2,32(%esp) |
| 936 | movdqa %xmm3,48(%esp) |
| 937 | xorl %eax,%eax |
| 938 | xorl %edx,%edx |
| 939 | xorl %ebp,%ebp |
| 940 | L015tail_loop: |
| 941 | movb (%esp,%ebp,1),%al |
| 942 | movb (%esi,%ebp,1),%dl |
| 943 | leal 1(%ebp),%ebp |
| 944 | xorb %dl,%al |
| 945 | movb %al,-1(%edi,%ebp,1) |
| 946 | decl %ecx |
| 947 | jnz L015tail_loop |
| 948 | L011done: |
| 949 | movl 512(%esp),%esp |
| 950 | popl %edi |
| 951 | popl %esi |
| 952 | popl %ebx |
| 953 | popl %ebp |
| 954 | ret |
| 955 | .align 6,0x90 |
| 956 | Lssse3_data: |
| 957 | .byte 2,3,0,1,6,7,4,5,10,11,8,9,14,15,12,13 |
| 958 | .byte 3,0,1,2,7,4,5,6,11,8,9,10,15,12,13,14 |
| 959 | .long 1634760805,857760878,2036477234,1797285236 |
| 960 | .long 0,1,2,3 |
| 961 | .long 4,4,4,4 |
| 962 | .long 1,0,0,0 |
| 963 | .long 4,0,0,0 |
| 964 | .long 0,-1,-1,-1 |
| 965 | .align 6,0x90 |
| 966 | .byte 67,104,97,67,104,97,50,48,32,102,111,114,32,120,56,54 |
| 967 | .byte 44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32 |
| 968 | .byte 60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111 |
| 969 | .byte 114,103,62,0 |
| 970 | .section __IMPORT,__pointers,non_lazy_symbol_pointers |
| 971 | L_OPENSSL_ia32cap_P$non_lazy_ptr: |
| 972 | .indirect_symbol _OPENSSL_ia32cap_P |
| 973 | .long 0 |
| 974 | #endif |