| 1 | default rel
|
|---|
| 2 | %define XMMWORD
|
|---|
| 3 | %define YMMWORD
|
|---|
| 4 | %define ZMMWORD
|
|---|
| 5 | section .text code align=64
|
|---|
| 6 |
|
|---|
| 7 |
|
|---|
| 8 | EXTERN OPENSSL_ia32cap_P
|
|---|
| 9 |
|
|---|
| 10 | global rsaz_512_sqr
|
|---|
| 11 |
|
|---|
| 12 | ALIGN 32
|
|---|
| 13 | rsaz_512_sqr:
|
|---|
| 14 | mov QWORD[8+rsp],rdi ;WIN64 prologue
|
|---|
| 15 | mov QWORD[16+rsp],rsi
|
|---|
| 16 | mov rax,rsp
|
|---|
| 17 | $L$SEH_begin_rsaz_512_sqr:
|
|---|
| 18 | mov rdi,rcx
|
|---|
| 19 | mov rsi,rdx
|
|---|
| 20 | mov rdx,r8
|
|---|
| 21 | mov rcx,r9
|
|---|
| 22 | mov r8,QWORD[40+rsp]
|
|---|
| 23 |
|
|---|
| 24 |
|
|---|
| 25 |
|
|---|
| 26 | push rbx
|
|---|
| 27 |
|
|---|
| 28 | push rbp
|
|---|
| 29 |
|
|---|
| 30 | push r12
|
|---|
| 31 |
|
|---|
| 32 | push r13
|
|---|
| 33 |
|
|---|
| 34 | push r14
|
|---|
| 35 |
|
|---|
| 36 | push r15
|
|---|
| 37 |
|
|---|
| 38 |
|
|---|
| 39 | sub rsp,128+24
|
|---|
| 40 |
|
|---|
| 41 | $L$sqr_body:
|
|---|
| 42 | DB 102,72,15,110,202
|
|---|
| 43 | mov rdx,QWORD[rsi]
|
|---|
| 44 | mov rax,QWORD[8+rsi]
|
|---|
| 45 | mov QWORD[128+rsp],rcx
|
|---|
| 46 | mov r11d,0x80100
|
|---|
| 47 | and r11d,DWORD[((OPENSSL_ia32cap_P+8))]
|
|---|
| 48 | cmp r11d,0x80100
|
|---|
| 49 | je NEAR $L$oop_sqrx
|
|---|
| 50 | jmp NEAR $L$oop_sqr
|
|---|
| 51 |
|
|---|
| 52 | ALIGN 32
|
|---|
| 53 | $L$oop_sqr:
|
|---|
| 54 | mov DWORD[((128+8))+rsp],r8d
|
|---|
| 55 |
|
|---|
| 56 | mov rbx,rdx
|
|---|
| 57 | mov rbp,rax
|
|---|
| 58 | mul rdx
|
|---|
| 59 | mov r8,rax
|
|---|
| 60 | mov rax,QWORD[16+rsi]
|
|---|
| 61 | mov r9,rdx
|
|---|
| 62 |
|
|---|
| 63 | mul rbx
|
|---|
| 64 | add r9,rax
|
|---|
| 65 | mov rax,QWORD[24+rsi]
|
|---|
| 66 | mov r10,rdx
|
|---|
| 67 | adc r10,0
|
|---|
| 68 |
|
|---|
| 69 | mul rbx
|
|---|
| 70 | add r10,rax
|
|---|
| 71 | mov rax,QWORD[32+rsi]
|
|---|
| 72 | mov r11,rdx
|
|---|
| 73 | adc r11,0
|
|---|
| 74 |
|
|---|
| 75 | mul rbx
|
|---|
| 76 | add r11,rax
|
|---|
| 77 | mov rax,QWORD[40+rsi]
|
|---|
| 78 | mov r12,rdx
|
|---|
| 79 | adc r12,0
|
|---|
| 80 |
|
|---|
| 81 | mul rbx
|
|---|
| 82 | add r12,rax
|
|---|
| 83 | mov rax,QWORD[48+rsi]
|
|---|
| 84 | mov r13,rdx
|
|---|
| 85 | adc r13,0
|
|---|
| 86 |
|
|---|
| 87 | mul rbx
|
|---|
| 88 | add r13,rax
|
|---|
| 89 | mov rax,QWORD[56+rsi]
|
|---|
| 90 | mov r14,rdx
|
|---|
| 91 | adc r14,0
|
|---|
| 92 |
|
|---|
| 93 | mul rbx
|
|---|
| 94 | add r14,rax
|
|---|
| 95 | mov rax,rbx
|
|---|
| 96 | adc rdx,0
|
|---|
| 97 |
|
|---|
| 98 | xor rcx,rcx
|
|---|
| 99 | add r8,r8
|
|---|
| 100 | mov r15,rdx
|
|---|
| 101 | adc rcx,0
|
|---|
| 102 |
|
|---|
| 103 | mul rax
|
|---|
| 104 | add rdx,r8
|
|---|
| 105 | adc rcx,0
|
|---|
| 106 |
|
|---|
| 107 | mov QWORD[rsp],rax
|
|---|
| 108 | mov QWORD[8+rsp],rdx
|
|---|
| 109 |
|
|---|
| 110 |
|
|---|
| 111 | mov rax,QWORD[16+rsi]
|
|---|
| 112 | mul rbp
|
|---|
| 113 | add r10,rax
|
|---|
| 114 | mov rax,QWORD[24+rsi]
|
|---|
| 115 | mov rbx,rdx
|
|---|
| 116 | adc rbx,0
|
|---|
| 117 |
|
|---|
| 118 | mul rbp
|
|---|
| 119 | add r11,rax
|
|---|
| 120 | mov rax,QWORD[32+rsi]
|
|---|
| 121 | adc rdx,0
|
|---|
| 122 | add r11,rbx
|
|---|
| 123 | mov rbx,rdx
|
|---|
| 124 | adc rbx,0
|
|---|
| 125 |
|
|---|
| 126 | mul rbp
|
|---|
| 127 | add r12,rax
|
|---|
| 128 | mov rax,QWORD[40+rsi]
|
|---|
| 129 | adc rdx,0
|
|---|
| 130 | add r12,rbx
|
|---|
| 131 | mov rbx,rdx
|
|---|
| 132 | adc rbx,0
|
|---|
| 133 |
|
|---|
| 134 | mul rbp
|
|---|
| 135 | add r13,rax
|
|---|
| 136 | mov rax,QWORD[48+rsi]
|
|---|
| 137 | adc rdx,0
|
|---|
| 138 | add r13,rbx
|
|---|
| 139 | mov rbx,rdx
|
|---|
| 140 | adc rbx,0
|
|---|
| 141 |
|
|---|
| 142 | mul rbp
|
|---|
| 143 | add r14,rax
|
|---|
| 144 | mov rax,QWORD[56+rsi]
|
|---|
| 145 | adc rdx,0
|
|---|
| 146 | add r14,rbx
|
|---|
| 147 | mov rbx,rdx
|
|---|
| 148 | adc rbx,0
|
|---|
| 149 |
|
|---|
| 150 | mul rbp
|
|---|
| 151 | add r15,rax
|
|---|
| 152 | mov rax,rbp
|
|---|
| 153 | adc rdx,0
|
|---|
| 154 | add r15,rbx
|
|---|
| 155 | adc rdx,0
|
|---|
| 156 |
|
|---|
| 157 | xor rbx,rbx
|
|---|
| 158 | add r9,r9
|
|---|
| 159 | mov r8,rdx
|
|---|
| 160 | adc r10,r10
|
|---|
| 161 | adc rbx,0
|
|---|
| 162 |
|
|---|
| 163 | mul rax
|
|---|
| 164 |
|
|---|
| 165 | add rax,rcx
|
|---|
| 166 | mov rbp,QWORD[16+rsi]
|
|---|
| 167 | add r9,rax
|
|---|
| 168 | mov rax,QWORD[24+rsi]
|
|---|
| 169 | adc r10,rdx
|
|---|
| 170 | adc rbx,0
|
|---|
| 171 |
|
|---|
| 172 | mov QWORD[16+rsp],r9
|
|---|
| 173 | mov QWORD[24+rsp],r10
|
|---|
| 174 |
|
|---|
| 175 |
|
|---|
| 176 | mul rbp
|
|---|
| 177 | add r12,rax
|
|---|
| 178 | mov rax,QWORD[32+rsi]
|
|---|
| 179 | mov rcx,rdx
|
|---|
| 180 | adc rcx,0
|
|---|
| 181 |
|
|---|
| 182 | mul rbp
|
|---|
| 183 | add r13,rax
|
|---|
| 184 | mov rax,QWORD[40+rsi]
|
|---|
| 185 | adc rdx,0
|
|---|
| 186 | add r13,rcx
|
|---|
| 187 | mov rcx,rdx
|
|---|
| 188 | adc rcx,0
|
|---|
| 189 |
|
|---|
| 190 | mul rbp
|
|---|
| 191 | add r14,rax
|
|---|
| 192 | mov rax,QWORD[48+rsi]
|
|---|
| 193 | adc rdx,0
|
|---|
| 194 | add r14,rcx
|
|---|
| 195 | mov rcx,rdx
|
|---|
| 196 | adc rcx,0
|
|---|
| 197 |
|
|---|
| 198 | mul rbp
|
|---|
| 199 | add r15,rax
|
|---|
| 200 | mov rax,QWORD[56+rsi]
|
|---|
| 201 | adc rdx,0
|
|---|
| 202 | add r15,rcx
|
|---|
| 203 | mov rcx,rdx
|
|---|
| 204 | adc rcx,0
|
|---|
| 205 |
|
|---|
| 206 | mul rbp
|
|---|
| 207 | add r8,rax
|
|---|
| 208 | mov rax,rbp
|
|---|
| 209 | adc rdx,0
|
|---|
| 210 | add r8,rcx
|
|---|
| 211 | adc rdx,0
|
|---|
| 212 |
|
|---|
| 213 | xor rcx,rcx
|
|---|
| 214 | add r11,r11
|
|---|
| 215 | mov r9,rdx
|
|---|
| 216 | adc r12,r12
|
|---|
| 217 | adc rcx,0
|
|---|
| 218 |
|
|---|
| 219 | mul rax
|
|---|
| 220 |
|
|---|
| 221 | add rax,rbx
|
|---|
| 222 | mov r10,QWORD[24+rsi]
|
|---|
| 223 | add r11,rax
|
|---|
| 224 | mov rax,QWORD[32+rsi]
|
|---|
| 225 | adc r12,rdx
|
|---|
| 226 | adc rcx,0
|
|---|
| 227 |
|
|---|
| 228 | mov QWORD[32+rsp],r11
|
|---|
| 229 | mov QWORD[40+rsp],r12
|
|---|
| 230 |
|
|---|
| 231 |
|
|---|
| 232 | mov r11,rax
|
|---|
| 233 | mul r10
|
|---|
| 234 | add r14,rax
|
|---|
| 235 | mov rax,QWORD[40+rsi]
|
|---|
| 236 | mov rbx,rdx
|
|---|
| 237 | adc rbx,0
|
|---|
| 238 |
|
|---|
| 239 | mov r12,rax
|
|---|
| 240 | mul r10
|
|---|
| 241 | add r15,rax
|
|---|
| 242 | mov rax,QWORD[48+rsi]
|
|---|
| 243 | adc rdx,0
|
|---|
| 244 | add r15,rbx
|
|---|
| 245 | mov rbx,rdx
|
|---|
| 246 | adc rbx,0
|
|---|
| 247 |
|
|---|
| 248 | mov rbp,rax
|
|---|
| 249 | mul r10
|
|---|
| 250 | add r8,rax
|
|---|
| 251 | mov rax,QWORD[56+rsi]
|
|---|
| 252 | adc rdx,0
|
|---|
| 253 | add r8,rbx
|
|---|
| 254 | mov rbx,rdx
|
|---|
| 255 | adc rbx,0
|
|---|
| 256 |
|
|---|
| 257 | mul r10
|
|---|
| 258 | add r9,rax
|
|---|
| 259 | mov rax,r10
|
|---|
| 260 | adc rdx,0
|
|---|
| 261 | add r9,rbx
|
|---|
| 262 | adc rdx,0
|
|---|
| 263 |
|
|---|
| 264 | xor rbx,rbx
|
|---|
| 265 | add r13,r13
|
|---|
| 266 | mov r10,rdx
|
|---|
| 267 | adc r14,r14
|
|---|
| 268 | adc rbx,0
|
|---|
| 269 |
|
|---|
| 270 | mul rax
|
|---|
| 271 |
|
|---|
| 272 | add rax,rcx
|
|---|
| 273 | add r13,rax
|
|---|
| 274 | mov rax,r12
|
|---|
| 275 | adc r14,rdx
|
|---|
| 276 | adc rbx,0
|
|---|
| 277 |
|
|---|
| 278 | mov QWORD[48+rsp],r13
|
|---|
| 279 | mov QWORD[56+rsp],r14
|
|---|
| 280 |
|
|---|
| 281 |
|
|---|
| 282 | mul r11
|
|---|
| 283 | add r8,rax
|
|---|
| 284 | mov rax,rbp
|
|---|
| 285 | mov rcx,rdx
|
|---|
| 286 | adc rcx,0
|
|---|
| 287 |
|
|---|
| 288 | mul r11
|
|---|
| 289 | add r9,rax
|
|---|
| 290 | mov rax,QWORD[56+rsi]
|
|---|
| 291 | adc rdx,0
|
|---|
| 292 | add r9,rcx
|
|---|
| 293 | mov rcx,rdx
|
|---|
| 294 | adc rcx,0
|
|---|
| 295 |
|
|---|
| 296 | mov r14,rax
|
|---|
| 297 | mul r11
|
|---|
| 298 | add r10,rax
|
|---|
| 299 | mov rax,r11
|
|---|
| 300 | adc rdx,0
|
|---|
| 301 | add r10,rcx
|
|---|
| 302 | adc rdx,0
|
|---|
| 303 |
|
|---|
| 304 | xor rcx,rcx
|
|---|
| 305 | add r15,r15
|
|---|
| 306 | mov r11,rdx
|
|---|
| 307 | adc r8,r8
|
|---|
| 308 | adc rcx,0
|
|---|
| 309 |
|
|---|
| 310 | mul rax
|
|---|
| 311 |
|
|---|
| 312 | add rax,rbx
|
|---|
| 313 | add r15,rax
|
|---|
| 314 | mov rax,rbp
|
|---|
| 315 | adc r8,rdx
|
|---|
| 316 | adc rcx,0
|
|---|
| 317 |
|
|---|
| 318 | mov QWORD[64+rsp],r15
|
|---|
| 319 | mov QWORD[72+rsp],r8
|
|---|
| 320 |
|
|---|
| 321 |
|
|---|
| 322 | mul r12
|
|---|
| 323 | add r10,rax
|
|---|
| 324 | mov rax,r14
|
|---|
| 325 | mov rbx,rdx
|
|---|
| 326 | adc rbx,0
|
|---|
| 327 |
|
|---|
| 328 | mul r12
|
|---|
| 329 | add r11,rax
|
|---|
| 330 | mov rax,r12
|
|---|
| 331 | adc rdx,0
|
|---|
| 332 | add r11,rbx
|
|---|
| 333 | adc rdx,0
|
|---|
| 334 |
|
|---|
| 335 | xor rbx,rbx
|
|---|
| 336 | add r9,r9
|
|---|
| 337 | mov r12,rdx
|
|---|
| 338 | adc r10,r10
|
|---|
| 339 | adc rbx,0
|
|---|
| 340 |
|
|---|
| 341 | mul rax
|
|---|
| 342 |
|
|---|
| 343 | add rax,rcx
|
|---|
| 344 | add r9,rax
|
|---|
| 345 | mov rax,r14
|
|---|
| 346 | adc r10,rdx
|
|---|
| 347 | adc rbx,0
|
|---|
| 348 |
|
|---|
| 349 | mov QWORD[80+rsp],r9
|
|---|
| 350 | mov QWORD[88+rsp],r10
|
|---|
| 351 |
|
|---|
| 352 |
|
|---|
| 353 | mul rbp
|
|---|
| 354 | add r12,rax
|
|---|
| 355 | mov rax,rbp
|
|---|
| 356 | adc rdx,0
|
|---|
| 357 |
|
|---|
| 358 | xor rcx,rcx
|
|---|
| 359 | add r11,r11
|
|---|
| 360 | mov r13,rdx
|
|---|
| 361 | adc r12,r12
|
|---|
| 362 | adc rcx,0
|
|---|
| 363 |
|
|---|
| 364 | mul rax
|
|---|
| 365 |
|
|---|
| 366 | add rax,rbx
|
|---|
| 367 | add r11,rax
|
|---|
| 368 | mov rax,r14
|
|---|
| 369 | adc r12,rdx
|
|---|
| 370 | adc rcx,0
|
|---|
| 371 |
|
|---|
| 372 | mov QWORD[96+rsp],r11
|
|---|
| 373 | mov QWORD[104+rsp],r12
|
|---|
| 374 |
|
|---|
| 375 |
|
|---|
| 376 | xor rbx,rbx
|
|---|
| 377 | add r13,r13
|
|---|
| 378 | adc rbx,0
|
|---|
| 379 |
|
|---|
| 380 | mul rax
|
|---|
| 381 |
|
|---|
| 382 | add rax,rcx
|
|---|
| 383 | add rax,r13
|
|---|
| 384 | adc rdx,rbx
|
|---|
| 385 |
|
|---|
| 386 | mov r8,QWORD[rsp]
|
|---|
| 387 | mov r9,QWORD[8+rsp]
|
|---|
| 388 | mov r10,QWORD[16+rsp]
|
|---|
| 389 | mov r11,QWORD[24+rsp]
|
|---|
| 390 | mov r12,QWORD[32+rsp]
|
|---|
| 391 | mov r13,QWORD[40+rsp]
|
|---|
| 392 | mov r14,QWORD[48+rsp]
|
|---|
| 393 | mov r15,QWORD[56+rsp]
|
|---|
| 394 | DB 102,72,15,126,205
|
|---|
| 395 |
|
|---|
| 396 | mov QWORD[112+rsp],rax
|
|---|
| 397 | mov QWORD[120+rsp],rdx
|
|---|
| 398 |
|
|---|
| 399 | call __rsaz_512_reduce
|
|---|
| 400 |
|
|---|
| 401 | add r8,QWORD[64+rsp]
|
|---|
| 402 | adc r9,QWORD[72+rsp]
|
|---|
| 403 | adc r10,QWORD[80+rsp]
|
|---|
| 404 | adc r11,QWORD[88+rsp]
|
|---|
| 405 | adc r12,QWORD[96+rsp]
|
|---|
| 406 | adc r13,QWORD[104+rsp]
|
|---|
| 407 | adc r14,QWORD[112+rsp]
|
|---|
| 408 | adc r15,QWORD[120+rsp]
|
|---|
| 409 | sbb rcx,rcx
|
|---|
| 410 |
|
|---|
| 411 | call __rsaz_512_subtract
|
|---|
| 412 |
|
|---|
| 413 | mov rdx,r8
|
|---|
| 414 | mov rax,r9
|
|---|
| 415 | mov r8d,DWORD[((128+8))+rsp]
|
|---|
| 416 | mov rsi,rdi
|
|---|
| 417 |
|
|---|
| 418 | dec r8d
|
|---|
| 419 | jnz NEAR $L$oop_sqr
|
|---|
| 420 | jmp NEAR $L$sqr_tail
|
|---|
| 421 |
|
|---|
| 422 | ALIGN 32
|
|---|
| 423 | $L$oop_sqrx:
|
|---|
| 424 | mov DWORD[((128+8))+rsp],r8d
|
|---|
| 425 | DB 102,72,15,110,199
|
|---|
| 426 |
|
|---|
| 427 | mulx r9,r8,rax
|
|---|
| 428 | mov rbx,rax
|
|---|
| 429 |
|
|---|
| 430 | mulx r10,rcx,QWORD[16+rsi]
|
|---|
| 431 | xor rbp,rbp
|
|---|
| 432 |
|
|---|
| 433 | mulx r11,rax,QWORD[24+rsi]
|
|---|
| 434 | adcx r9,rcx
|
|---|
| 435 |
|
|---|
| 436 | DB 0xc4,0x62,0xf3,0xf6,0xa6,0x20,0x00,0x00,0x00
|
|---|
| 437 | adcx r10,rax
|
|---|
| 438 |
|
|---|
| 439 | DB 0xc4,0x62,0xfb,0xf6,0xae,0x28,0x00,0x00,0x00
|
|---|
| 440 | adcx r11,rcx
|
|---|
| 441 |
|
|---|
| 442 | mulx r14,rcx,QWORD[48+rsi]
|
|---|
| 443 | adcx r12,rax
|
|---|
| 444 | adcx r13,rcx
|
|---|
| 445 |
|
|---|
| 446 | mulx r15,rax,QWORD[56+rsi]
|
|---|
| 447 | adcx r14,rax
|
|---|
| 448 | adcx r15,rbp
|
|---|
| 449 |
|
|---|
| 450 | mulx rdi,rax,rdx
|
|---|
| 451 | mov rdx,rbx
|
|---|
| 452 | xor rcx,rcx
|
|---|
| 453 | adox r8,r8
|
|---|
| 454 | adcx r8,rdi
|
|---|
| 455 | adox rcx,rbp
|
|---|
| 456 | adcx rcx,rbp
|
|---|
| 457 |
|
|---|
| 458 | mov QWORD[rsp],rax
|
|---|
| 459 | mov QWORD[8+rsp],r8
|
|---|
| 460 |
|
|---|
| 461 |
|
|---|
| 462 | DB 0xc4,0xe2,0xfb,0xf6,0x9e,0x10,0x00,0x00,0x00
|
|---|
| 463 | adox r10,rax
|
|---|
| 464 | adcx r11,rbx
|
|---|
| 465 |
|
|---|
| 466 | mulx r8,rdi,QWORD[24+rsi]
|
|---|
| 467 | adox r11,rdi
|
|---|
| 468 | DB 0x66
|
|---|
| 469 | adcx r12,r8
|
|---|
| 470 |
|
|---|
| 471 | mulx rbx,rax,QWORD[32+rsi]
|
|---|
| 472 | adox r12,rax
|
|---|
| 473 | adcx r13,rbx
|
|---|
| 474 |
|
|---|
| 475 | mulx r8,rdi,QWORD[40+rsi]
|
|---|
| 476 | adox r13,rdi
|
|---|
| 477 | adcx r14,r8
|
|---|
| 478 |
|
|---|
| 479 | DB 0xc4,0xe2,0xfb,0xf6,0x9e,0x30,0x00,0x00,0x00
|
|---|
| 480 | adox r14,rax
|
|---|
| 481 | adcx r15,rbx
|
|---|
| 482 |
|
|---|
| 483 | DB 0xc4,0x62,0xc3,0xf6,0x86,0x38,0x00,0x00,0x00
|
|---|
| 484 | adox r15,rdi
|
|---|
| 485 | adcx r8,rbp
|
|---|
| 486 | mulx rdi,rax,rdx
|
|---|
| 487 | adox r8,rbp
|
|---|
| 488 | DB 0x48,0x8b,0x96,0x10,0x00,0x00,0x00
|
|---|
| 489 |
|
|---|
| 490 | xor rbx,rbx
|
|---|
| 491 | adox r9,r9
|
|---|
| 492 |
|
|---|
| 493 | adcx rax,rcx
|
|---|
| 494 | adox r10,r10
|
|---|
| 495 | adcx r9,rax
|
|---|
| 496 | adox rbx,rbp
|
|---|
| 497 | adcx r10,rdi
|
|---|
| 498 | adcx rbx,rbp
|
|---|
| 499 |
|
|---|
| 500 | mov QWORD[16+rsp],r9
|
|---|
| 501 | DB 0x4c,0x89,0x94,0x24,0x18,0x00,0x00,0x00
|
|---|
| 502 |
|
|---|
| 503 |
|
|---|
| 504 | mulx r9,rdi,QWORD[24+rsi]
|
|---|
| 505 | adox r12,rdi
|
|---|
| 506 | adcx r13,r9
|
|---|
| 507 |
|
|---|
| 508 | mulx rcx,rax,QWORD[32+rsi]
|
|---|
| 509 | adox r13,rax
|
|---|
| 510 | adcx r14,rcx
|
|---|
| 511 |
|
|---|
| 512 | DB 0xc4,0x62,0xc3,0xf6,0x8e,0x28,0x00,0x00,0x00
|
|---|
| 513 | adox r14,rdi
|
|---|
| 514 | adcx r15,r9
|
|---|
| 515 |
|
|---|
| 516 | DB 0xc4,0xe2,0xfb,0xf6,0x8e,0x30,0x00,0x00,0x00
|
|---|
| 517 | adox r15,rax
|
|---|
| 518 | adcx r8,rcx
|
|---|
| 519 |
|
|---|
| 520 | mulx r9,rdi,QWORD[56+rsi]
|
|---|
| 521 | adox r8,rdi
|
|---|
| 522 | adcx r9,rbp
|
|---|
| 523 | mulx rdi,rax,rdx
|
|---|
| 524 | adox r9,rbp
|
|---|
| 525 | mov rdx,QWORD[24+rsi]
|
|---|
| 526 |
|
|---|
| 527 | xor rcx,rcx
|
|---|
| 528 | adox r11,r11
|
|---|
| 529 |
|
|---|
| 530 | adcx rax,rbx
|
|---|
| 531 | adox r12,r12
|
|---|
| 532 | adcx r11,rax
|
|---|
| 533 | adox rcx,rbp
|
|---|
| 534 | adcx r12,rdi
|
|---|
| 535 | adcx rcx,rbp
|
|---|
| 536 |
|
|---|
| 537 | mov QWORD[32+rsp],r11
|
|---|
| 538 | mov QWORD[40+rsp],r12
|
|---|
| 539 |
|
|---|
| 540 |
|
|---|
| 541 | mulx rbx,rax,QWORD[32+rsi]
|
|---|
| 542 | adox r14,rax
|
|---|
| 543 | adcx r15,rbx
|
|---|
| 544 |
|
|---|
| 545 | mulx r10,rdi,QWORD[40+rsi]
|
|---|
| 546 | adox r15,rdi
|
|---|
| 547 | adcx r8,r10
|
|---|
| 548 |
|
|---|
| 549 | mulx rbx,rax,QWORD[48+rsi]
|
|---|
| 550 | adox r8,rax
|
|---|
| 551 | adcx r9,rbx
|
|---|
| 552 |
|
|---|
| 553 | mulx r10,rdi,QWORD[56+rsi]
|
|---|
| 554 | adox r9,rdi
|
|---|
| 555 | adcx r10,rbp
|
|---|
| 556 | mulx rdi,rax,rdx
|
|---|
| 557 | adox r10,rbp
|
|---|
| 558 | mov rdx,QWORD[32+rsi]
|
|---|
| 559 |
|
|---|
| 560 | xor rbx,rbx
|
|---|
| 561 | adox r13,r13
|
|---|
| 562 |
|
|---|
| 563 | adcx rax,rcx
|
|---|
| 564 | adox r14,r14
|
|---|
| 565 | adcx r13,rax
|
|---|
| 566 | adox rbx,rbp
|
|---|
| 567 | adcx r14,rdi
|
|---|
| 568 | adcx rbx,rbp
|
|---|
| 569 |
|
|---|
| 570 | mov QWORD[48+rsp],r13
|
|---|
| 571 | mov QWORD[56+rsp],r14
|
|---|
| 572 |
|
|---|
| 573 |
|
|---|
| 574 | mulx r11,rdi,QWORD[40+rsi]
|
|---|
| 575 | adox r8,rdi
|
|---|
| 576 | adcx r9,r11
|
|---|
| 577 |
|
|---|
| 578 | mulx rcx,rax,QWORD[48+rsi]
|
|---|
| 579 | adox r9,rax
|
|---|
| 580 | adcx r10,rcx
|
|---|
| 581 |
|
|---|
| 582 | mulx r11,rdi,QWORD[56+rsi]
|
|---|
| 583 | adox r10,rdi
|
|---|
| 584 | adcx r11,rbp
|
|---|
| 585 | mulx rdi,rax,rdx
|
|---|
| 586 | mov rdx,QWORD[40+rsi]
|
|---|
| 587 | adox r11,rbp
|
|---|
| 588 |
|
|---|
| 589 | xor rcx,rcx
|
|---|
| 590 | adox r15,r15
|
|---|
| 591 |
|
|---|
| 592 | adcx rax,rbx
|
|---|
| 593 | adox r8,r8
|
|---|
| 594 | adcx r15,rax
|
|---|
| 595 | adox rcx,rbp
|
|---|
| 596 | adcx r8,rdi
|
|---|
| 597 | adcx rcx,rbp
|
|---|
| 598 |
|
|---|
| 599 | mov QWORD[64+rsp],r15
|
|---|
| 600 | mov QWORD[72+rsp],r8
|
|---|
| 601 |
|
|---|
| 602 |
|
|---|
| 603 | DB 0xc4,0xe2,0xfb,0xf6,0x9e,0x30,0x00,0x00,0x00
|
|---|
| 604 | adox r10,rax
|
|---|
| 605 | adcx r11,rbx
|
|---|
| 606 |
|
|---|
| 607 | DB 0xc4,0x62,0xc3,0xf6,0xa6,0x38,0x00,0x00,0x00
|
|---|
| 608 | adox r11,rdi
|
|---|
| 609 | adcx r12,rbp
|
|---|
| 610 | mulx rdi,rax,rdx
|
|---|
| 611 | adox r12,rbp
|
|---|
| 612 | mov rdx,QWORD[48+rsi]
|
|---|
| 613 |
|
|---|
| 614 | xor rbx,rbx
|
|---|
| 615 | adox r9,r9
|
|---|
| 616 |
|
|---|
| 617 | adcx rax,rcx
|
|---|
| 618 | adox r10,r10
|
|---|
| 619 | adcx r9,rax
|
|---|
| 620 | adcx r10,rdi
|
|---|
| 621 | adox rbx,rbp
|
|---|
| 622 | adcx rbx,rbp
|
|---|
| 623 |
|
|---|
| 624 | mov QWORD[80+rsp],r9
|
|---|
| 625 | mov QWORD[88+rsp],r10
|
|---|
| 626 |
|
|---|
| 627 |
|
|---|
| 628 | DB 0xc4,0x62,0xfb,0xf6,0xae,0x38,0x00,0x00,0x00
|
|---|
| 629 | adox r12,rax
|
|---|
| 630 | adox r13,rbp
|
|---|
| 631 |
|
|---|
| 632 | mulx rdi,rax,rdx
|
|---|
| 633 | xor rcx,rcx
|
|---|
| 634 | mov rdx,QWORD[56+rsi]
|
|---|
| 635 | adox r11,r11
|
|---|
| 636 |
|
|---|
| 637 | adcx rax,rbx
|
|---|
| 638 | adox r12,r12
|
|---|
| 639 | adcx r11,rax
|
|---|
| 640 | adox rcx,rbp
|
|---|
| 641 | adcx r12,rdi
|
|---|
| 642 | adcx rcx,rbp
|
|---|
| 643 |
|
|---|
| 644 | DB 0x4c,0x89,0x9c,0x24,0x60,0x00,0x00,0x00
|
|---|
| 645 | DB 0x4c,0x89,0xa4,0x24,0x68,0x00,0x00,0x00
|
|---|
| 646 |
|
|---|
| 647 |
|
|---|
| 648 | mulx rdx,rax,rdx
|
|---|
| 649 | xor rbx,rbx
|
|---|
| 650 | adox r13,r13
|
|---|
| 651 |
|
|---|
| 652 | adcx rax,rcx
|
|---|
| 653 | adox rbx,rbp
|
|---|
| 654 | adcx rax,r13
|
|---|
| 655 | adcx rbx,rdx
|
|---|
| 656 |
|
|---|
| 657 | DB 102,72,15,126,199
|
|---|
| 658 | DB 102,72,15,126,205
|
|---|
| 659 |
|
|---|
| 660 | mov rdx,QWORD[128+rsp]
|
|---|
| 661 | mov r8,QWORD[rsp]
|
|---|
| 662 | mov r9,QWORD[8+rsp]
|
|---|
| 663 | mov r10,QWORD[16+rsp]
|
|---|
| 664 | mov r11,QWORD[24+rsp]
|
|---|
| 665 | mov r12,QWORD[32+rsp]
|
|---|
| 666 | mov r13,QWORD[40+rsp]
|
|---|
| 667 | mov r14,QWORD[48+rsp]
|
|---|
| 668 | mov r15,QWORD[56+rsp]
|
|---|
| 669 |
|
|---|
| 670 | mov QWORD[112+rsp],rax
|
|---|
| 671 | mov QWORD[120+rsp],rbx
|
|---|
| 672 |
|
|---|
| 673 | call __rsaz_512_reducex
|
|---|
| 674 |
|
|---|
| 675 | add r8,QWORD[64+rsp]
|
|---|
| 676 | adc r9,QWORD[72+rsp]
|
|---|
| 677 | adc r10,QWORD[80+rsp]
|
|---|
| 678 | adc r11,QWORD[88+rsp]
|
|---|
| 679 | adc r12,QWORD[96+rsp]
|
|---|
| 680 | adc r13,QWORD[104+rsp]
|
|---|
| 681 | adc r14,QWORD[112+rsp]
|
|---|
| 682 | adc r15,QWORD[120+rsp]
|
|---|
| 683 | sbb rcx,rcx
|
|---|
| 684 |
|
|---|
| 685 | call __rsaz_512_subtract
|
|---|
| 686 |
|
|---|
| 687 | mov rdx,r8
|
|---|
| 688 | mov rax,r9
|
|---|
| 689 | mov r8d,DWORD[((128+8))+rsp]
|
|---|
| 690 | mov rsi,rdi
|
|---|
| 691 |
|
|---|
| 692 | dec r8d
|
|---|
| 693 | jnz NEAR $L$oop_sqrx
|
|---|
| 694 |
|
|---|
| 695 | $L$sqr_tail:
|
|---|
| 696 |
|
|---|
| 697 | lea rax,[((128+24+48))+rsp]
|
|---|
| 698 |
|
|---|
| 699 | mov r15,QWORD[((-48))+rax]
|
|---|
| 700 |
|
|---|
| 701 | mov r14,QWORD[((-40))+rax]
|
|---|
| 702 |
|
|---|
| 703 | mov r13,QWORD[((-32))+rax]
|
|---|
| 704 |
|
|---|
| 705 | mov r12,QWORD[((-24))+rax]
|
|---|
| 706 |
|
|---|
| 707 | mov rbp,QWORD[((-16))+rax]
|
|---|
| 708 |
|
|---|
| 709 | mov rbx,QWORD[((-8))+rax]
|
|---|
| 710 |
|
|---|
| 711 | lea rsp,[rax]
|
|---|
| 712 |
|
|---|
| 713 | $L$sqr_epilogue:
|
|---|
| 714 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
|---|
| 715 | mov rsi,QWORD[16+rsp]
|
|---|
| 716 | DB 0F3h,0C3h ;repret
|
|---|
| 717 |
|
|---|
| 718 | $L$SEH_end_rsaz_512_sqr:
|
|---|
| 719 | global rsaz_512_mul
|
|---|
| 720 |
|
|---|
| 721 | ALIGN 32
|
|---|
| 722 | rsaz_512_mul:
|
|---|
| 723 | mov QWORD[8+rsp],rdi ;WIN64 prologue
|
|---|
| 724 | mov QWORD[16+rsp],rsi
|
|---|
| 725 | mov rax,rsp
|
|---|
| 726 | $L$SEH_begin_rsaz_512_mul:
|
|---|
| 727 | mov rdi,rcx
|
|---|
| 728 | mov rsi,rdx
|
|---|
| 729 | mov rdx,r8
|
|---|
| 730 | mov rcx,r9
|
|---|
| 731 | mov r8,QWORD[40+rsp]
|
|---|
| 732 |
|
|---|
| 733 |
|
|---|
| 734 |
|
|---|
| 735 | push rbx
|
|---|
| 736 |
|
|---|
| 737 | push rbp
|
|---|
| 738 |
|
|---|
| 739 | push r12
|
|---|
| 740 |
|
|---|
| 741 | push r13
|
|---|
| 742 |
|
|---|
| 743 | push r14
|
|---|
| 744 |
|
|---|
| 745 | push r15
|
|---|
| 746 |
|
|---|
| 747 |
|
|---|
| 748 | sub rsp,128+24
|
|---|
| 749 |
|
|---|
| 750 | $L$mul_body:
|
|---|
| 751 | DB 102,72,15,110,199
|
|---|
| 752 | DB 102,72,15,110,201
|
|---|
| 753 | mov QWORD[128+rsp],r8
|
|---|
| 754 | mov r11d,0x80100
|
|---|
| 755 | and r11d,DWORD[((OPENSSL_ia32cap_P+8))]
|
|---|
| 756 | cmp r11d,0x80100
|
|---|
| 757 | je NEAR $L$mulx
|
|---|
| 758 | mov rbx,QWORD[rdx]
|
|---|
| 759 | mov rbp,rdx
|
|---|
| 760 | call __rsaz_512_mul
|
|---|
| 761 |
|
|---|
| 762 | DB 102,72,15,126,199
|
|---|
| 763 | DB 102,72,15,126,205
|
|---|
| 764 |
|
|---|
| 765 | mov r8,QWORD[rsp]
|
|---|
| 766 | mov r9,QWORD[8+rsp]
|
|---|
| 767 | mov r10,QWORD[16+rsp]
|
|---|
| 768 | mov r11,QWORD[24+rsp]
|
|---|
| 769 | mov r12,QWORD[32+rsp]
|
|---|
| 770 | mov r13,QWORD[40+rsp]
|
|---|
| 771 | mov r14,QWORD[48+rsp]
|
|---|
| 772 | mov r15,QWORD[56+rsp]
|
|---|
| 773 |
|
|---|
| 774 | call __rsaz_512_reduce
|
|---|
| 775 | jmp NEAR $L$mul_tail
|
|---|
| 776 |
|
|---|
| 777 | ALIGN 32
|
|---|
| 778 | $L$mulx:
|
|---|
| 779 | mov rbp,rdx
|
|---|
| 780 | mov rdx,QWORD[rdx]
|
|---|
| 781 | call __rsaz_512_mulx
|
|---|
| 782 |
|
|---|
| 783 | DB 102,72,15,126,199
|
|---|
| 784 | DB 102,72,15,126,205
|
|---|
| 785 |
|
|---|
| 786 | mov rdx,QWORD[128+rsp]
|
|---|
| 787 | mov r8,QWORD[rsp]
|
|---|
| 788 | mov r9,QWORD[8+rsp]
|
|---|
| 789 | mov r10,QWORD[16+rsp]
|
|---|
| 790 | mov r11,QWORD[24+rsp]
|
|---|
| 791 | mov r12,QWORD[32+rsp]
|
|---|
| 792 | mov r13,QWORD[40+rsp]
|
|---|
| 793 | mov r14,QWORD[48+rsp]
|
|---|
| 794 | mov r15,QWORD[56+rsp]
|
|---|
| 795 |
|
|---|
| 796 | call __rsaz_512_reducex
|
|---|
| 797 | $L$mul_tail:
|
|---|
| 798 | add r8,QWORD[64+rsp]
|
|---|
| 799 | adc r9,QWORD[72+rsp]
|
|---|
| 800 | adc r10,QWORD[80+rsp]
|
|---|
| 801 | adc r11,QWORD[88+rsp]
|
|---|
| 802 | adc r12,QWORD[96+rsp]
|
|---|
| 803 | adc r13,QWORD[104+rsp]
|
|---|
| 804 | adc r14,QWORD[112+rsp]
|
|---|
| 805 | adc r15,QWORD[120+rsp]
|
|---|
| 806 | sbb rcx,rcx
|
|---|
| 807 |
|
|---|
| 808 | call __rsaz_512_subtract
|
|---|
| 809 |
|
|---|
| 810 | lea rax,[((128+24+48))+rsp]
|
|---|
| 811 |
|
|---|
| 812 | mov r15,QWORD[((-48))+rax]
|
|---|
| 813 |
|
|---|
| 814 | mov r14,QWORD[((-40))+rax]
|
|---|
| 815 |
|
|---|
| 816 | mov r13,QWORD[((-32))+rax]
|
|---|
| 817 |
|
|---|
| 818 | mov r12,QWORD[((-24))+rax]
|
|---|
| 819 |
|
|---|
| 820 | mov rbp,QWORD[((-16))+rax]
|
|---|
| 821 |
|
|---|
| 822 | mov rbx,QWORD[((-8))+rax]
|
|---|
| 823 |
|
|---|
| 824 | lea rsp,[rax]
|
|---|
| 825 |
|
|---|
| 826 | $L$mul_epilogue:
|
|---|
| 827 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
|---|
| 828 | mov rsi,QWORD[16+rsp]
|
|---|
| 829 | DB 0F3h,0C3h ;repret
|
|---|
| 830 |
|
|---|
| 831 | $L$SEH_end_rsaz_512_mul:
|
|---|
| 832 | global rsaz_512_mul_gather4
|
|---|
| 833 |
|
|---|
| 834 | ALIGN 32
|
|---|
| 835 | rsaz_512_mul_gather4:
|
|---|
| 836 | mov QWORD[8+rsp],rdi ;WIN64 prologue
|
|---|
| 837 | mov QWORD[16+rsp],rsi
|
|---|
| 838 | mov rax,rsp
|
|---|
| 839 | $L$SEH_begin_rsaz_512_mul_gather4:
|
|---|
| 840 | mov rdi,rcx
|
|---|
| 841 | mov rsi,rdx
|
|---|
| 842 | mov rdx,r8
|
|---|
| 843 | mov rcx,r9
|
|---|
| 844 | mov r8,QWORD[40+rsp]
|
|---|
| 845 | mov r9,QWORD[48+rsp]
|
|---|
| 846 |
|
|---|
| 847 |
|
|---|
| 848 |
|
|---|
| 849 | push rbx
|
|---|
| 850 |
|
|---|
| 851 | push rbp
|
|---|
| 852 |
|
|---|
| 853 | push r12
|
|---|
| 854 |
|
|---|
| 855 | push r13
|
|---|
| 856 |
|
|---|
| 857 | push r14
|
|---|
| 858 |
|
|---|
| 859 | push r15
|
|---|
| 860 |
|
|---|
| 861 |
|
|---|
| 862 | sub rsp,328
|
|---|
| 863 |
|
|---|
| 864 | movaps XMMWORD[160+rsp],xmm6
|
|---|
| 865 | movaps XMMWORD[176+rsp],xmm7
|
|---|
| 866 | movaps XMMWORD[192+rsp],xmm8
|
|---|
| 867 | movaps XMMWORD[208+rsp],xmm9
|
|---|
| 868 | movaps XMMWORD[224+rsp],xmm10
|
|---|
| 869 | movaps XMMWORD[240+rsp],xmm11
|
|---|
| 870 | movaps XMMWORD[256+rsp],xmm12
|
|---|
| 871 | movaps XMMWORD[272+rsp],xmm13
|
|---|
| 872 | movaps XMMWORD[288+rsp],xmm14
|
|---|
| 873 | movaps XMMWORD[304+rsp],xmm15
|
|---|
| 874 | $L$mul_gather4_body:
|
|---|
| 875 | movd xmm8,r9d
|
|---|
| 876 | movdqa xmm1,XMMWORD[(($L$inc+16))]
|
|---|
| 877 | movdqa xmm0,XMMWORD[$L$inc]
|
|---|
| 878 |
|
|---|
| 879 | pshufd xmm8,xmm8,0
|
|---|
| 880 | movdqa xmm7,xmm1
|
|---|
| 881 | movdqa xmm2,xmm1
|
|---|
| 882 | paddd xmm1,xmm0
|
|---|
| 883 | pcmpeqd xmm0,xmm8
|
|---|
| 884 | movdqa xmm3,xmm7
|
|---|
| 885 | paddd xmm2,xmm1
|
|---|
| 886 | pcmpeqd xmm1,xmm8
|
|---|
| 887 | movdqa xmm4,xmm7
|
|---|
| 888 | paddd xmm3,xmm2
|
|---|
| 889 | pcmpeqd xmm2,xmm8
|
|---|
| 890 | movdqa xmm5,xmm7
|
|---|
| 891 | paddd xmm4,xmm3
|
|---|
| 892 | pcmpeqd xmm3,xmm8
|
|---|
| 893 | movdqa xmm6,xmm7
|
|---|
| 894 | paddd xmm5,xmm4
|
|---|
| 895 | pcmpeqd xmm4,xmm8
|
|---|
| 896 | paddd xmm6,xmm5
|
|---|
| 897 | pcmpeqd xmm5,xmm8
|
|---|
| 898 | paddd xmm7,xmm6
|
|---|
| 899 | pcmpeqd xmm6,xmm8
|
|---|
| 900 | pcmpeqd xmm7,xmm8
|
|---|
| 901 |
|
|---|
| 902 | movdqa xmm8,XMMWORD[rdx]
|
|---|
| 903 | movdqa xmm9,XMMWORD[16+rdx]
|
|---|
| 904 | movdqa xmm10,XMMWORD[32+rdx]
|
|---|
| 905 | movdqa xmm11,XMMWORD[48+rdx]
|
|---|
| 906 | pand xmm8,xmm0
|
|---|
| 907 | movdqa xmm12,XMMWORD[64+rdx]
|
|---|
| 908 | pand xmm9,xmm1
|
|---|
| 909 | movdqa xmm13,XMMWORD[80+rdx]
|
|---|
| 910 | pand xmm10,xmm2
|
|---|
| 911 | movdqa xmm14,XMMWORD[96+rdx]
|
|---|
| 912 | pand xmm11,xmm3
|
|---|
| 913 | movdqa xmm15,XMMWORD[112+rdx]
|
|---|
| 914 | lea rbp,[128+rdx]
|
|---|
| 915 | pand xmm12,xmm4
|
|---|
| 916 | pand xmm13,xmm5
|
|---|
| 917 | pand xmm14,xmm6
|
|---|
| 918 | pand xmm15,xmm7
|
|---|
| 919 | por xmm8,xmm10
|
|---|
| 920 | por xmm9,xmm11
|
|---|
| 921 | por xmm8,xmm12
|
|---|
| 922 | por xmm9,xmm13
|
|---|
| 923 | por xmm8,xmm14
|
|---|
| 924 | por xmm9,xmm15
|
|---|
| 925 |
|
|---|
| 926 | por xmm8,xmm9
|
|---|
| 927 | pshufd xmm9,xmm8,0x4e
|
|---|
| 928 | por xmm8,xmm9
|
|---|
| 929 | mov r11d,0x80100
|
|---|
| 930 | and r11d,DWORD[((OPENSSL_ia32cap_P+8))]
|
|---|
| 931 | cmp r11d,0x80100
|
|---|
| 932 | je NEAR $L$mulx_gather
|
|---|
| 933 | DB 102,76,15,126,195
|
|---|
| 934 |
|
|---|
| 935 | mov QWORD[128+rsp],r8
|
|---|
| 936 | mov QWORD[((128+8))+rsp],rdi
|
|---|
| 937 | mov QWORD[((128+16))+rsp],rcx
|
|---|
| 938 |
|
|---|
| 939 | mov rax,QWORD[rsi]
|
|---|
| 940 | mov rcx,QWORD[8+rsi]
|
|---|
| 941 | mul rbx
|
|---|
| 942 | mov QWORD[rsp],rax
|
|---|
| 943 | mov rax,rcx
|
|---|
| 944 | mov r8,rdx
|
|---|
| 945 |
|
|---|
| 946 | mul rbx
|
|---|
| 947 | add r8,rax
|
|---|
| 948 | mov rax,QWORD[16+rsi]
|
|---|
| 949 | mov r9,rdx
|
|---|
| 950 | adc r9,0
|
|---|
| 951 |
|
|---|
| 952 | mul rbx
|
|---|
| 953 | add r9,rax
|
|---|
| 954 | mov rax,QWORD[24+rsi]
|
|---|
| 955 | mov r10,rdx
|
|---|
| 956 | adc r10,0
|
|---|
| 957 |
|
|---|
| 958 | mul rbx
|
|---|
| 959 | add r10,rax
|
|---|
| 960 | mov rax,QWORD[32+rsi]
|
|---|
| 961 | mov r11,rdx
|
|---|
| 962 | adc r11,0
|
|---|
| 963 |
|
|---|
| 964 | mul rbx
|
|---|
| 965 | add r11,rax
|
|---|
| 966 | mov rax,QWORD[40+rsi]
|
|---|
| 967 | mov r12,rdx
|
|---|
| 968 | adc r12,0
|
|---|
| 969 |
|
|---|
| 970 | mul rbx
|
|---|
| 971 | add r12,rax
|
|---|
| 972 | mov rax,QWORD[48+rsi]
|
|---|
| 973 | mov r13,rdx
|
|---|
| 974 | adc r13,0
|
|---|
| 975 |
|
|---|
| 976 | mul rbx
|
|---|
| 977 | add r13,rax
|
|---|
| 978 | mov rax,QWORD[56+rsi]
|
|---|
| 979 | mov r14,rdx
|
|---|
| 980 | adc r14,0
|
|---|
| 981 |
|
|---|
| 982 | mul rbx
|
|---|
| 983 | add r14,rax
|
|---|
| 984 | mov rax,QWORD[rsi]
|
|---|
| 985 | mov r15,rdx
|
|---|
| 986 | adc r15,0
|
|---|
| 987 |
|
|---|
| 988 | lea rdi,[8+rsp]
|
|---|
| 989 | mov ecx,7
|
|---|
| 990 | jmp NEAR $L$oop_mul_gather
|
|---|
| 991 |
|
|---|
| 992 | ALIGN 32
|
|---|
| 993 | $L$oop_mul_gather:
|
|---|
| 994 | movdqa xmm8,XMMWORD[rbp]
|
|---|
| 995 | movdqa xmm9,XMMWORD[16+rbp]
|
|---|
| 996 | movdqa xmm10,XMMWORD[32+rbp]
|
|---|
| 997 | movdqa xmm11,XMMWORD[48+rbp]
|
|---|
| 998 | pand xmm8,xmm0
|
|---|
| 999 | movdqa xmm12,XMMWORD[64+rbp]
|
|---|
| 1000 | pand xmm9,xmm1
|
|---|
| 1001 | movdqa xmm13,XMMWORD[80+rbp]
|
|---|
| 1002 | pand xmm10,xmm2
|
|---|
| 1003 | movdqa xmm14,XMMWORD[96+rbp]
|
|---|
| 1004 | pand xmm11,xmm3
|
|---|
| 1005 | movdqa xmm15,XMMWORD[112+rbp]
|
|---|
| 1006 | lea rbp,[128+rbp]
|
|---|
| 1007 | pand xmm12,xmm4
|
|---|
| 1008 | pand xmm13,xmm5
|
|---|
| 1009 | pand xmm14,xmm6
|
|---|
| 1010 | pand xmm15,xmm7
|
|---|
| 1011 | por xmm8,xmm10
|
|---|
| 1012 | por xmm9,xmm11
|
|---|
| 1013 | por xmm8,xmm12
|
|---|
| 1014 | por xmm9,xmm13
|
|---|
| 1015 | por xmm8,xmm14
|
|---|
| 1016 | por xmm9,xmm15
|
|---|
| 1017 |
|
|---|
| 1018 | por xmm8,xmm9
|
|---|
| 1019 | pshufd xmm9,xmm8,0x4e
|
|---|
| 1020 | por xmm8,xmm9
|
|---|
| 1021 | DB 102,76,15,126,195
|
|---|
| 1022 |
|
|---|
| 1023 | mul rbx
|
|---|
| 1024 | add r8,rax
|
|---|
| 1025 | mov rax,QWORD[8+rsi]
|
|---|
| 1026 | mov QWORD[rdi],r8
|
|---|
| 1027 | mov r8,rdx
|
|---|
| 1028 | adc r8,0
|
|---|
| 1029 |
|
|---|
| 1030 | mul rbx
|
|---|
| 1031 | add r9,rax
|
|---|
| 1032 | mov rax,QWORD[16+rsi]
|
|---|
| 1033 | adc rdx,0
|
|---|
| 1034 | add r8,r9
|
|---|
| 1035 | mov r9,rdx
|
|---|
| 1036 | adc r9,0
|
|---|
| 1037 |
|
|---|
| 1038 | mul rbx
|
|---|
| 1039 | add r10,rax
|
|---|
| 1040 | mov rax,QWORD[24+rsi]
|
|---|
| 1041 | adc rdx,0
|
|---|
| 1042 | add r9,r10
|
|---|
| 1043 | mov r10,rdx
|
|---|
| 1044 | adc r10,0
|
|---|
| 1045 |
|
|---|
| 1046 | mul rbx
|
|---|
| 1047 | add r11,rax
|
|---|
| 1048 | mov rax,QWORD[32+rsi]
|
|---|
| 1049 | adc rdx,0
|
|---|
| 1050 | add r10,r11
|
|---|
| 1051 | mov r11,rdx
|
|---|
| 1052 | adc r11,0
|
|---|
| 1053 |
|
|---|
| 1054 | mul rbx
|
|---|
| 1055 | add r12,rax
|
|---|
| 1056 | mov rax,QWORD[40+rsi]
|
|---|
| 1057 | adc rdx,0
|
|---|
| 1058 | add r11,r12
|
|---|
| 1059 | mov r12,rdx
|
|---|
| 1060 | adc r12,0
|
|---|
| 1061 |
|
|---|
| 1062 | mul rbx
|
|---|
| 1063 | add r13,rax
|
|---|
| 1064 | mov rax,QWORD[48+rsi]
|
|---|
| 1065 | adc rdx,0
|
|---|
| 1066 | add r12,r13
|
|---|
| 1067 | mov r13,rdx
|
|---|
| 1068 | adc r13,0
|
|---|
| 1069 |
|
|---|
| 1070 | mul rbx
|
|---|
| 1071 | add r14,rax
|
|---|
| 1072 | mov rax,QWORD[56+rsi]
|
|---|
| 1073 | adc rdx,0
|
|---|
| 1074 | add r13,r14
|
|---|
| 1075 | mov r14,rdx
|
|---|
| 1076 | adc r14,0
|
|---|
| 1077 |
|
|---|
| 1078 | mul rbx
|
|---|
| 1079 | add r15,rax
|
|---|
| 1080 | mov rax,QWORD[rsi]
|
|---|
| 1081 | adc rdx,0
|
|---|
| 1082 | add r14,r15
|
|---|
| 1083 | mov r15,rdx
|
|---|
| 1084 | adc r15,0
|
|---|
| 1085 |
|
|---|
| 1086 | lea rdi,[8+rdi]
|
|---|
| 1087 |
|
|---|
| 1088 | dec ecx
|
|---|
| 1089 | jnz NEAR $L$oop_mul_gather
|
|---|
| 1090 |
|
|---|
| 1091 | mov QWORD[rdi],r8
|
|---|
| 1092 | mov QWORD[8+rdi],r9
|
|---|
| 1093 | mov QWORD[16+rdi],r10
|
|---|
| 1094 | mov QWORD[24+rdi],r11
|
|---|
| 1095 | mov QWORD[32+rdi],r12
|
|---|
| 1096 | mov QWORD[40+rdi],r13
|
|---|
| 1097 | mov QWORD[48+rdi],r14
|
|---|
| 1098 | mov QWORD[56+rdi],r15
|
|---|
| 1099 |
|
|---|
| 1100 | mov rdi,QWORD[((128+8))+rsp]
|
|---|
| 1101 | mov rbp,QWORD[((128+16))+rsp]
|
|---|
| 1102 |
|
|---|
| 1103 | mov r8,QWORD[rsp]
|
|---|
| 1104 | mov r9,QWORD[8+rsp]
|
|---|
| 1105 | mov r10,QWORD[16+rsp]
|
|---|
| 1106 | mov r11,QWORD[24+rsp]
|
|---|
| 1107 | mov r12,QWORD[32+rsp]
|
|---|
| 1108 | mov r13,QWORD[40+rsp]
|
|---|
| 1109 | mov r14,QWORD[48+rsp]
|
|---|
| 1110 | mov r15,QWORD[56+rsp]
|
|---|
| 1111 |
|
|---|
| 1112 | call __rsaz_512_reduce
|
|---|
| 1113 | jmp NEAR $L$mul_gather_tail
|
|---|
| 1114 |
|
|---|
| 1115 | ALIGN 32
|
|---|
| 1116 | $L$mulx_gather:
|
|---|
| 1117 | DB 102,76,15,126,194
|
|---|
| 1118 |
|
|---|
| 1119 | mov QWORD[128+rsp],r8
|
|---|
| 1120 | mov QWORD[((128+8))+rsp],rdi
|
|---|
| 1121 | mov QWORD[((128+16))+rsp],rcx
|
|---|
| 1122 |
|
|---|
| 1123 | mulx r8,rbx,QWORD[rsi]
|
|---|
| 1124 | mov QWORD[rsp],rbx
|
|---|
| 1125 | xor edi,edi
|
|---|
| 1126 |
|
|---|
| 1127 | mulx r9,rax,QWORD[8+rsi]
|
|---|
| 1128 |
|
|---|
| 1129 | mulx r10,rbx,QWORD[16+rsi]
|
|---|
| 1130 | adcx r8,rax
|
|---|
| 1131 |
|
|---|
| 1132 | mulx r11,rax,QWORD[24+rsi]
|
|---|
| 1133 | adcx r9,rbx
|
|---|
| 1134 |
|
|---|
| 1135 | mulx r12,rbx,QWORD[32+rsi]
|
|---|
| 1136 | adcx r10,rax
|
|---|
| 1137 |
|
|---|
| 1138 | mulx r13,rax,QWORD[40+rsi]
|
|---|
| 1139 | adcx r11,rbx
|
|---|
| 1140 |
|
|---|
| 1141 | mulx r14,rbx,QWORD[48+rsi]
|
|---|
| 1142 | adcx r12,rax
|
|---|
| 1143 |
|
|---|
| 1144 | mulx r15,rax,QWORD[56+rsi]
|
|---|
| 1145 | adcx r13,rbx
|
|---|
| 1146 | adcx r14,rax
|
|---|
| 1147 | DB 0x67
|
|---|
| 1148 | mov rbx,r8
|
|---|
| 1149 | adcx r15,rdi
|
|---|
| 1150 |
|
|---|
| 1151 | mov rcx,-7
|
|---|
| 1152 | jmp NEAR $L$oop_mulx_gather
|
|---|
| 1153 |
|
|---|
| 1154 | ALIGN 32
|
|---|
| 1155 | $L$oop_mulx_gather:
|
|---|
| 1156 | movdqa xmm8,XMMWORD[rbp]
|
|---|
| 1157 | movdqa xmm9,XMMWORD[16+rbp]
|
|---|
| 1158 | movdqa xmm10,XMMWORD[32+rbp]
|
|---|
| 1159 | movdqa xmm11,XMMWORD[48+rbp]
|
|---|
| 1160 | pand xmm8,xmm0
|
|---|
| 1161 | movdqa xmm12,XMMWORD[64+rbp]
|
|---|
| 1162 | pand xmm9,xmm1
|
|---|
| 1163 | movdqa xmm13,XMMWORD[80+rbp]
|
|---|
| 1164 | pand xmm10,xmm2
|
|---|
| 1165 | movdqa xmm14,XMMWORD[96+rbp]
|
|---|
| 1166 | pand xmm11,xmm3
|
|---|
| 1167 | movdqa xmm15,XMMWORD[112+rbp]
|
|---|
| 1168 | lea rbp,[128+rbp]
|
|---|
| 1169 | pand xmm12,xmm4
|
|---|
| 1170 | pand xmm13,xmm5
|
|---|
| 1171 | pand xmm14,xmm6
|
|---|
| 1172 | pand xmm15,xmm7
|
|---|
| 1173 | por xmm8,xmm10
|
|---|
| 1174 | por xmm9,xmm11
|
|---|
| 1175 | por xmm8,xmm12
|
|---|
| 1176 | por xmm9,xmm13
|
|---|
| 1177 | por xmm8,xmm14
|
|---|
| 1178 | por xmm9,xmm15
|
|---|
| 1179 |
|
|---|
| 1180 | por xmm8,xmm9
|
|---|
| 1181 | pshufd xmm9,xmm8,0x4e
|
|---|
| 1182 | por xmm8,xmm9
|
|---|
| 1183 | DB 102,76,15,126,194
|
|---|
| 1184 |
|
|---|
| 1185 | DB 0xc4,0x62,0xfb,0xf6,0x86,0x00,0x00,0x00,0x00
|
|---|
| 1186 | adcx rbx,rax
|
|---|
| 1187 | adox r8,r9
|
|---|
| 1188 |
|
|---|
| 1189 | mulx r9,rax,QWORD[8+rsi]
|
|---|
| 1190 | adcx r8,rax
|
|---|
| 1191 | adox r9,r10
|
|---|
| 1192 |
|
|---|
| 1193 | mulx r10,rax,QWORD[16+rsi]
|
|---|
| 1194 | adcx r9,rax
|
|---|
| 1195 | adox r10,r11
|
|---|
| 1196 |
|
|---|
| 1197 | DB 0xc4,0x62,0xfb,0xf6,0x9e,0x18,0x00,0x00,0x00
|
|---|
| 1198 | adcx r10,rax
|
|---|
| 1199 | adox r11,r12
|
|---|
| 1200 |
|
|---|
| 1201 | mulx r12,rax,QWORD[32+rsi]
|
|---|
| 1202 | adcx r11,rax
|
|---|
| 1203 | adox r12,r13
|
|---|
| 1204 |
|
|---|
| 1205 | mulx r13,rax,QWORD[40+rsi]
|
|---|
| 1206 | adcx r12,rax
|
|---|
| 1207 | adox r13,r14
|
|---|
| 1208 |
|
|---|
| 1209 | DB 0xc4,0x62,0xfb,0xf6,0xb6,0x30,0x00,0x00,0x00
|
|---|
| 1210 | adcx r13,rax
|
|---|
| 1211 | DB 0x67
|
|---|
| 1212 | adox r14,r15
|
|---|
| 1213 |
|
|---|
| 1214 | mulx r15,rax,QWORD[56+rsi]
|
|---|
| 1215 | mov QWORD[64+rcx*8+rsp],rbx
|
|---|
| 1216 | adcx r14,rax
|
|---|
| 1217 | adox r15,rdi
|
|---|
| 1218 | mov rbx,r8
|
|---|
| 1219 | adcx r15,rdi
|
|---|
| 1220 |
|
|---|
| 1221 | inc rcx
|
|---|
| 1222 | jnz NEAR $L$oop_mulx_gather
|
|---|
| 1223 |
|
|---|
| 1224 | mov QWORD[64+rsp],r8
|
|---|
| 1225 | mov QWORD[((64+8))+rsp],r9
|
|---|
| 1226 | mov QWORD[((64+16))+rsp],r10
|
|---|
| 1227 | mov QWORD[((64+24))+rsp],r11
|
|---|
| 1228 | mov QWORD[((64+32))+rsp],r12
|
|---|
| 1229 | mov QWORD[((64+40))+rsp],r13
|
|---|
| 1230 | mov QWORD[((64+48))+rsp],r14
|
|---|
| 1231 | mov QWORD[((64+56))+rsp],r15
|
|---|
| 1232 |
|
|---|
| 1233 | mov rdx,QWORD[128+rsp]
|
|---|
| 1234 | mov rdi,QWORD[((128+8))+rsp]
|
|---|
| 1235 | mov rbp,QWORD[((128+16))+rsp]
|
|---|
| 1236 |
|
|---|
| 1237 | mov r8,QWORD[rsp]
|
|---|
| 1238 | mov r9,QWORD[8+rsp]
|
|---|
| 1239 | mov r10,QWORD[16+rsp]
|
|---|
| 1240 | mov r11,QWORD[24+rsp]
|
|---|
| 1241 | mov r12,QWORD[32+rsp]
|
|---|
| 1242 | mov r13,QWORD[40+rsp]
|
|---|
| 1243 | mov r14,QWORD[48+rsp]
|
|---|
| 1244 | mov r15,QWORD[56+rsp]
|
|---|
| 1245 |
|
|---|
| 1246 | call __rsaz_512_reducex
|
|---|
| 1247 |
|
|---|
| 1248 | $L$mul_gather_tail:
|
|---|
| 1249 | add r8,QWORD[64+rsp]
|
|---|
| 1250 | adc r9,QWORD[72+rsp]
|
|---|
| 1251 | adc r10,QWORD[80+rsp]
|
|---|
| 1252 | adc r11,QWORD[88+rsp]
|
|---|
| 1253 | adc r12,QWORD[96+rsp]
|
|---|
| 1254 | adc r13,QWORD[104+rsp]
|
|---|
| 1255 | adc r14,QWORD[112+rsp]
|
|---|
| 1256 | adc r15,QWORD[120+rsp]
|
|---|
| 1257 | sbb rcx,rcx
|
|---|
| 1258 |
|
|---|
| 1259 | call __rsaz_512_subtract
|
|---|
| 1260 |
|
|---|
| 1261 | lea rax,[((128+24+48))+rsp]
|
|---|
| 1262 | movaps xmm6,XMMWORD[((160-200))+rax]
|
|---|
| 1263 | movaps xmm7,XMMWORD[((176-200))+rax]
|
|---|
| 1264 | movaps xmm8,XMMWORD[((192-200))+rax]
|
|---|
| 1265 | movaps xmm9,XMMWORD[((208-200))+rax]
|
|---|
| 1266 | movaps xmm10,XMMWORD[((224-200))+rax]
|
|---|
| 1267 | movaps xmm11,XMMWORD[((240-200))+rax]
|
|---|
| 1268 | movaps xmm12,XMMWORD[((256-200))+rax]
|
|---|
| 1269 | movaps xmm13,XMMWORD[((272-200))+rax]
|
|---|
| 1270 | movaps xmm14,XMMWORD[((288-200))+rax]
|
|---|
| 1271 | movaps xmm15,XMMWORD[((304-200))+rax]
|
|---|
| 1272 | lea rax,[176+rax]
|
|---|
| 1273 |
|
|---|
| 1274 | mov r15,QWORD[((-48))+rax]
|
|---|
| 1275 |
|
|---|
| 1276 | mov r14,QWORD[((-40))+rax]
|
|---|
| 1277 |
|
|---|
| 1278 | mov r13,QWORD[((-32))+rax]
|
|---|
| 1279 |
|
|---|
| 1280 | mov r12,QWORD[((-24))+rax]
|
|---|
| 1281 |
|
|---|
| 1282 | mov rbp,QWORD[((-16))+rax]
|
|---|
| 1283 |
|
|---|
| 1284 | mov rbx,QWORD[((-8))+rax]
|
|---|
| 1285 |
|
|---|
| 1286 | lea rsp,[rax]
|
|---|
| 1287 |
|
|---|
| 1288 | $L$mul_gather4_epilogue:
|
|---|
| 1289 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
|---|
| 1290 | mov rsi,QWORD[16+rsp]
|
|---|
| 1291 | DB 0F3h,0C3h ;repret
|
|---|
| 1292 |
|
|---|
| 1293 | $L$SEH_end_rsaz_512_mul_gather4:
|
|---|
| 1294 | global rsaz_512_mul_scatter4
|
|---|
| 1295 |
|
|---|
| 1296 | ALIGN 32
|
|---|
| 1297 | rsaz_512_mul_scatter4:
|
|---|
| 1298 | mov QWORD[8+rsp],rdi ;WIN64 prologue
|
|---|
| 1299 | mov QWORD[16+rsp],rsi
|
|---|
| 1300 | mov rax,rsp
|
|---|
| 1301 | $L$SEH_begin_rsaz_512_mul_scatter4:
|
|---|
| 1302 | mov rdi,rcx
|
|---|
| 1303 | mov rsi,rdx
|
|---|
| 1304 | mov rdx,r8
|
|---|
| 1305 | mov rcx,r9
|
|---|
| 1306 | mov r8,QWORD[40+rsp]
|
|---|
| 1307 | mov r9,QWORD[48+rsp]
|
|---|
| 1308 |
|
|---|
| 1309 |
|
|---|
| 1310 |
|
|---|
| 1311 | push rbx
|
|---|
| 1312 |
|
|---|
| 1313 | push rbp
|
|---|
| 1314 |
|
|---|
| 1315 | push r12
|
|---|
| 1316 |
|
|---|
| 1317 | push r13
|
|---|
| 1318 |
|
|---|
| 1319 | push r14
|
|---|
| 1320 |
|
|---|
| 1321 | push r15
|
|---|
| 1322 |
|
|---|
| 1323 |
|
|---|
| 1324 | mov r9d,r9d
|
|---|
| 1325 | sub rsp,128+24
|
|---|
| 1326 |
|
|---|
| 1327 | $L$mul_scatter4_body:
|
|---|
| 1328 | lea r8,[r9*8+r8]
|
|---|
| 1329 | DB 102,72,15,110,199
|
|---|
| 1330 | DB 102,72,15,110,202
|
|---|
| 1331 | DB 102,73,15,110,208
|
|---|
| 1332 | mov QWORD[128+rsp],rcx
|
|---|
| 1333 |
|
|---|
| 1334 | mov rbp,rdi
|
|---|
| 1335 | mov r11d,0x80100
|
|---|
| 1336 | and r11d,DWORD[((OPENSSL_ia32cap_P+8))]
|
|---|
| 1337 | cmp r11d,0x80100
|
|---|
| 1338 | je NEAR $L$mulx_scatter
|
|---|
| 1339 | mov rbx,QWORD[rdi]
|
|---|
| 1340 | call __rsaz_512_mul
|
|---|
| 1341 |
|
|---|
| 1342 | DB 102,72,15,126,199
|
|---|
| 1343 | DB 102,72,15,126,205
|
|---|
| 1344 |
|
|---|
| 1345 | mov r8,QWORD[rsp]
|
|---|
| 1346 | mov r9,QWORD[8+rsp]
|
|---|
| 1347 | mov r10,QWORD[16+rsp]
|
|---|
| 1348 | mov r11,QWORD[24+rsp]
|
|---|
| 1349 | mov r12,QWORD[32+rsp]
|
|---|
| 1350 | mov r13,QWORD[40+rsp]
|
|---|
| 1351 | mov r14,QWORD[48+rsp]
|
|---|
| 1352 | mov r15,QWORD[56+rsp]
|
|---|
| 1353 |
|
|---|
| 1354 | call __rsaz_512_reduce
|
|---|
| 1355 | jmp NEAR $L$mul_scatter_tail
|
|---|
| 1356 |
|
|---|
| 1357 | ALIGN 32
|
|---|
| 1358 | $L$mulx_scatter:
|
|---|
| 1359 | mov rdx,QWORD[rdi]
|
|---|
| 1360 | call __rsaz_512_mulx
|
|---|
| 1361 |
|
|---|
| 1362 | DB 102,72,15,126,199
|
|---|
| 1363 | DB 102,72,15,126,205
|
|---|
| 1364 |
|
|---|
| 1365 | mov rdx,QWORD[128+rsp]
|
|---|
| 1366 | mov r8,QWORD[rsp]
|
|---|
| 1367 | mov r9,QWORD[8+rsp]
|
|---|
| 1368 | mov r10,QWORD[16+rsp]
|
|---|
| 1369 | mov r11,QWORD[24+rsp]
|
|---|
| 1370 | mov r12,QWORD[32+rsp]
|
|---|
| 1371 | mov r13,QWORD[40+rsp]
|
|---|
| 1372 | mov r14,QWORD[48+rsp]
|
|---|
| 1373 | mov r15,QWORD[56+rsp]
|
|---|
| 1374 |
|
|---|
| 1375 | call __rsaz_512_reducex
|
|---|
| 1376 |
|
|---|
| 1377 | $L$mul_scatter_tail:
|
|---|
| 1378 | add r8,QWORD[64+rsp]
|
|---|
| 1379 | adc r9,QWORD[72+rsp]
|
|---|
| 1380 | adc r10,QWORD[80+rsp]
|
|---|
| 1381 | adc r11,QWORD[88+rsp]
|
|---|
| 1382 | adc r12,QWORD[96+rsp]
|
|---|
| 1383 | adc r13,QWORD[104+rsp]
|
|---|
| 1384 | adc r14,QWORD[112+rsp]
|
|---|
| 1385 | adc r15,QWORD[120+rsp]
|
|---|
| 1386 | DB 102,72,15,126,214
|
|---|
| 1387 | sbb rcx,rcx
|
|---|
| 1388 |
|
|---|
| 1389 | call __rsaz_512_subtract
|
|---|
| 1390 |
|
|---|
| 1391 | mov QWORD[rsi],r8
|
|---|
| 1392 | mov QWORD[128+rsi],r9
|
|---|
| 1393 | mov QWORD[256+rsi],r10
|
|---|
| 1394 | mov QWORD[384+rsi],r11
|
|---|
| 1395 | mov QWORD[512+rsi],r12
|
|---|
| 1396 | mov QWORD[640+rsi],r13
|
|---|
| 1397 | mov QWORD[768+rsi],r14
|
|---|
| 1398 | mov QWORD[896+rsi],r15
|
|---|
| 1399 |
|
|---|
| 1400 | lea rax,[((128+24+48))+rsp]
|
|---|
| 1401 |
|
|---|
| 1402 | mov r15,QWORD[((-48))+rax]
|
|---|
| 1403 |
|
|---|
| 1404 | mov r14,QWORD[((-40))+rax]
|
|---|
| 1405 |
|
|---|
| 1406 | mov r13,QWORD[((-32))+rax]
|
|---|
| 1407 |
|
|---|
| 1408 | mov r12,QWORD[((-24))+rax]
|
|---|
| 1409 |
|
|---|
| 1410 | mov rbp,QWORD[((-16))+rax]
|
|---|
| 1411 |
|
|---|
| 1412 | mov rbx,QWORD[((-8))+rax]
|
|---|
| 1413 |
|
|---|
| 1414 | lea rsp,[rax]
|
|---|
| 1415 |
|
|---|
| 1416 | $L$mul_scatter4_epilogue:
|
|---|
| 1417 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
|---|
| 1418 | mov rsi,QWORD[16+rsp]
|
|---|
| 1419 | DB 0F3h,0C3h ;repret
|
|---|
| 1420 |
|
|---|
| 1421 | $L$SEH_end_rsaz_512_mul_scatter4:
|
|---|
| 1422 | global rsaz_512_mul_by_one
|
|---|
| 1423 |
|
|---|
| 1424 | ALIGN 32
|
|---|
| 1425 | rsaz_512_mul_by_one:
|
|---|
| 1426 | mov QWORD[8+rsp],rdi ;WIN64 prologue
|
|---|
| 1427 | mov QWORD[16+rsp],rsi
|
|---|
| 1428 | mov rax,rsp
|
|---|
| 1429 | $L$SEH_begin_rsaz_512_mul_by_one:
|
|---|
| 1430 | mov rdi,rcx
|
|---|
| 1431 | mov rsi,rdx
|
|---|
| 1432 | mov rdx,r8
|
|---|
| 1433 | mov rcx,r9
|
|---|
| 1434 |
|
|---|
| 1435 |
|
|---|
| 1436 |
|
|---|
| 1437 | push rbx
|
|---|
| 1438 |
|
|---|
| 1439 | push rbp
|
|---|
| 1440 |
|
|---|
| 1441 | push r12
|
|---|
| 1442 |
|
|---|
| 1443 | push r13
|
|---|
| 1444 |
|
|---|
| 1445 | push r14
|
|---|
| 1446 |
|
|---|
| 1447 | push r15
|
|---|
| 1448 |
|
|---|
| 1449 |
|
|---|
| 1450 | sub rsp,128+24
|
|---|
| 1451 |
|
|---|
| 1452 | $L$mul_by_one_body:
|
|---|
| 1453 | mov eax,DWORD[((OPENSSL_ia32cap_P+8))]
|
|---|
| 1454 | mov rbp,rdx
|
|---|
| 1455 | mov QWORD[128+rsp],rcx
|
|---|
| 1456 |
|
|---|
| 1457 | mov r8,QWORD[rsi]
|
|---|
| 1458 | pxor xmm0,xmm0
|
|---|
| 1459 | mov r9,QWORD[8+rsi]
|
|---|
| 1460 | mov r10,QWORD[16+rsi]
|
|---|
| 1461 | mov r11,QWORD[24+rsi]
|
|---|
| 1462 | mov r12,QWORD[32+rsi]
|
|---|
| 1463 | mov r13,QWORD[40+rsi]
|
|---|
| 1464 | mov r14,QWORD[48+rsi]
|
|---|
| 1465 | mov r15,QWORD[56+rsi]
|
|---|
| 1466 |
|
|---|
| 1467 | movdqa XMMWORD[rsp],xmm0
|
|---|
| 1468 | movdqa XMMWORD[16+rsp],xmm0
|
|---|
| 1469 | movdqa XMMWORD[32+rsp],xmm0
|
|---|
| 1470 | movdqa XMMWORD[48+rsp],xmm0
|
|---|
| 1471 | movdqa XMMWORD[64+rsp],xmm0
|
|---|
| 1472 | movdqa XMMWORD[80+rsp],xmm0
|
|---|
| 1473 | movdqa XMMWORD[96+rsp],xmm0
|
|---|
| 1474 | and eax,0x80100
|
|---|
| 1475 | cmp eax,0x80100
|
|---|
| 1476 | je NEAR $L$by_one_callx
|
|---|
| 1477 | call __rsaz_512_reduce
|
|---|
| 1478 | jmp NEAR $L$by_one_tail
|
|---|
| 1479 | ALIGN 32
|
|---|
| 1480 | $L$by_one_callx:
|
|---|
| 1481 | mov rdx,QWORD[128+rsp]
|
|---|
| 1482 | call __rsaz_512_reducex
|
|---|
| 1483 | $L$by_one_tail:
|
|---|
| 1484 | mov QWORD[rdi],r8
|
|---|
| 1485 | mov QWORD[8+rdi],r9
|
|---|
| 1486 | mov QWORD[16+rdi],r10
|
|---|
| 1487 | mov QWORD[24+rdi],r11
|
|---|
| 1488 | mov QWORD[32+rdi],r12
|
|---|
| 1489 | mov QWORD[40+rdi],r13
|
|---|
| 1490 | mov QWORD[48+rdi],r14
|
|---|
| 1491 | mov QWORD[56+rdi],r15
|
|---|
| 1492 |
|
|---|
| 1493 | lea rax,[((128+24+48))+rsp]
|
|---|
| 1494 |
|
|---|
| 1495 | mov r15,QWORD[((-48))+rax]
|
|---|
| 1496 |
|
|---|
| 1497 | mov r14,QWORD[((-40))+rax]
|
|---|
| 1498 |
|
|---|
| 1499 | mov r13,QWORD[((-32))+rax]
|
|---|
| 1500 |
|
|---|
| 1501 | mov r12,QWORD[((-24))+rax]
|
|---|
| 1502 |
|
|---|
| 1503 | mov rbp,QWORD[((-16))+rax]
|
|---|
| 1504 |
|
|---|
| 1505 | mov rbx,QWORD[((-8))+rax]
|
|---|
| 1506 |
|
|---|
| 1507 | lea rsp,[rax]
|
|---|
| 1508 |
|
|---|
| 1509 | $L$mul_by_one_epilogue:
|
|---|
| 1510 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
|---|
| 1511 | mov rsi,QWORD[16+rsp]
|
|---|
| 1512 | DB 0F3h,0C3h ;repret
|
|---|
| 1513 |
|
|---|
| 1514 | $L$SEH_end_rsaz_512_mul_by_one:
|
|---|
| 1515 |
|
|---|
| 1516 | ALIGN 32
|
|---|
| 1517 | __rsaz_512_reduce:
|
|---|
| 1518 |
|
|---|
| 1519 | mov rbx,r8
|
|---|
| 1520 | imul rbx,QWORD[((128+8))+rsp]
|
|---|
| 1521 | mov rax,QWORD[rbp]
|
|---|
| 1522 | mov ecx,8
|
|---|
| 1523 | jmp NEAR $L$reduction_loop
|
|---|
| 1524 |
|
|---|
| 1525 | ALIGN 32
|
|---|
| 1526 | $L$reduction_loop:
|
|---|
| 1527 | mul rbx
|
|---|
| 1528 | mov rax,QWORD[8+rbp]
|
|---|
| 1529 | neg r8
|
|---|
| 1530 | mov r8,rdx
|
|---|
| 1531 | adc r8,0
|
|---|
| 1532 |
|
|---|
| 1533 | mul rbx
|
|---|
| 1534 | add r9,rax
|
|---|
| 1535 | mov rax,QWORD[16+rbp]
|
|---|
| 1536 | adc rdx,0
|
|---|
| 1537 | add r8,r9
|
|---|
| 1538 | mov r9,rdx
|
|---|
| 1539 | adc r9,0
|
|---|
| 1540 |
|
|---|
| 1541 | mul rbx
|
|---|
| 1542 | add r10,rax
|
|---|
| 1543 | mov rax,QWORD[24+rbp]
|
|---|
| 1544 | adc rdx,0
|
|---|
| 1545 | add r9,r10
|
|---|
| 1546 | mov r10,rdx
|
|---|
| 1547 | adc r10,0
|
|---|
| 1548 |
|
|---|
| 1549 | mul rbx
|
|---|
| 1550 | add r11,rax
|
|---|
| 1551 | mov rax,QWORD[32+rbp]
|
|---|
| 1552 | adc rdx,0
|
|---|
| 1553 | add r10,r11
|
|---|
| 1554 | mov rsi,QWORD[((128+8))+rsp]
|
|---|
| 1555 |
|
|---|
| 1556 |
|
|---|
| 1557 | adc rdx,0
|
|---|
| 1558 | mov r11,rdx
|
|---|
| 1559 |
|
|---|
| 1560 | mul rbx
|
|---|
| 1561 | add r12,rax
|
|---|
| 1562 | mov rax,QWORD[40+rbp]
|
|---|
| 1563 | adc rdx,0
|
|---|
| 1564 | imul rsi,r8
|
|---|
| 1565 | add r11,r12
|
|---|
| 1566 | mov r12,rdx
|
|---|
| 1567 | adc r12,0
|
|---|
| 1568 |
|
|---|
| 1569 | mul rbx
|
|---|
| 1570 | add r13,rax
|
|---|
| 1571 | mov rax,QWORD[48+rbp]
|
|---|
| 1572 | adc rdx,0
|
|---|
| 1573 | add r12,r13
|
|---|
| 1574 | mov r13,rdx
|
|---|
| 1575 | adc r13,0
|
|---|
| 1576 |
|
|---|
| 1577 | mul rbx
|
|---|
| 1578 | add r14,rax
|
|---|
| 1579 | mov rax,QWORD[56+rbp]
|
|---|
| 1580 | adc rdx,0
|
|---|
| 1581 | add r13,r14
|
|---|
| 1582 | mov r14,rdx
|
|---|
| 1583 | adc r14,0
|
|---|
| 1584 |
|
|---|
| 1585 | mul rbx
|
|---|
| 1586 | mov rbx,rsi
|
|---|
| 1587 | add r15,rax
|
|---|
| 1588 | mov rax,QWORD[rbp]
|
|---|
| 1589 | adc rdx,0
|
|---|
| 1590 | add r14,r15
|
|---|
| 1591 | mov r15,rdx
|
|---|
| 1592 | adc r15,0
|
|---|
| 1593 |
|
|---|
| 1594 | dec ecx
|
|---|
| 1595 | jne NEAR $L$reduction_loop
|
|---|
| 1596 |
|
|---|
| 1597 | DB 0F3h,0C3h ;repret
|
|---|
| 1598 |
|
|---|
| 1599 |
|
|---|
| 1600 |
|
|---|
| 1601 | ALIGN 32
|
|---|
| 1602 | __rsaz_512_reducex:
|
|---|
| 1603 |
|
|---|
| 1604 |
|
|---|
| 1605 | imul rdx,r8
|
|---|
| 1606 | xor rsi,rsi
|
|---|
| 1607 | mov ecx,8
|
|---|
| 1608 | jmp NEAR $L$reduction_loopx
|
|---|
| 1609 |
|
|---|
| 1610 | ALIGN 32
|
|---|
| 1611 | $L$reduction_loopx:
|
|---|
| 1612 | mov rbx,r8
|
|---|
| 1613 | mulx r8,rax,QWORD[rbp]
|
|---|
| 1614 | adcx rax,rbx
|
|---|
| 1615 | adox r8,r9
|
|---|
| 1616 |
|
|---|
| 1617 | mulx r9,rax,QWORD[8+rbp]
|
|---|
| 1618 | adcx r8,rax
|
|---|
| 1619 | adox r9,r10
|
|---|
| 1620 |
|
|---|
| 1621 | mulx r10,rbx,QWORD[16+rbp]
|
|---|
| 1622 | adcx r9,rbx
|
|---|
| 1623 | adox r10,r11
|
|---|
| 1624 |
|
|---|
| 1625 | mulx r11,rbx,QWORD[24+rbp]
|
|---|
| 1626 | adcx r10,rbx
|
|---|
| 1627 | adox r11,r12
|
|---|
| 1628 |
|
|---|
| 1629 | DB 0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00
|
|---|
| 1630 | mov rax,rdx
|
|---|
| 1631 | mov rdx,r8
|
|---|
| 1632 | adcx r11,rbx
|
|---|
| 1633 | adox r12,r13
|
|---|
| 1634 |
|
|---|
| 1635 | mulx rdx,rbx,QWORD[((128+8))+rsp]
|
|---|
| 1636 | mov rdx,rax
|
|---|
| 1637 |
|
|---|
| 1638 | mulx r13,rax,QWORD[40+rbp]
|
|---|
| 1639 | adcx r12,rax
|
|---|
| 1640 | adox r13,r14
|
|---|
| 1641 |
|
|---|
| 1642 | DB 0xc4,0x62,0xfb,0xf6,0xb5,0x30,0x00,0x00,0x00
|
|---|
| 1643 | adcx r13,rax
|
|---|
| 1644 | adox r14,r15
|
|---|
| 1645 |
|
|---|
| 1646 | mulx r15,rax,QWORD[56+rbp]
|
|---|
| 1647 | mov rdx,rbx
|
|---|
| 1648 | adcx r14,rax
|
|---|
| 1649 | adox r15,rsi
|
|---|
| 1650 | adcx r15,rsi
|
|---|
| 1651 |
|
|---|
| 1652 | dec ecx
|
|---|
| 1653 | jne NEAR $L$reduction_loopx
|
|---|
| 1654 |
|
|---|
| 1655 | DB 0F3h,0C3h ;repret
|
|---|
| 1656 |
|
|---|
| 1657 |
|
|---|
| 1658 |
|
|---|
| 1659 | ALIGN 32
|
|---|
| 1660 | __rsaz_512_subtract:
|
|---|
| 1661 |
|
|---|
| 1662 | mov QWORD[rdi],r8
|
|---|
| 1663 | mov QWORD[8+rdi],r9
|
|---|
| 1664 | mov QWORD[16+rdi],r10
|
|---|
| 1665 | mov QWORD[24+rdi],r11
|
|---|
| 1666 | mov QWORD[32+rdi],r12
|
|---|
| 1667 | mov QWORD[40+rdi],r13
|
|---|
| 1668 | mov QWORD[48+rdi],r14
|
|---|
| 1669 | mov QWORD[56+rdi],r15
|
|---|
| 1670 |
|
|---|
| 1671 | mov r8,QWORD[rbp]
|
|---|
| 1672 | mov r9,QWORD[8+rbp]
|
|---|
| 1673 | neg r8
|
|---|
| 1674 | not r9
|
|---|
| 1675 | and r8,rcx
|
|---|
| 1676 | mov r10,QWORD[16+rbp]
|
|---|
| 1677 | and r9,rcx
|
|---|
| 1678 | not r10
|
|---|
| 1679 | mov r11,QWORD[24+rbp]
|
|---|
| 1680 | and r10,rcx
|
|---|
| 1681 | not r11
|
|---|
| 1682 | mov r12,QWORD[32+rbp]
|
|---|
| 1683 | and r11,rcx
|
|---|
| 1684 | not r12
|
|---|
| 1685 | mov r13,QWORD[40+rbp]
|
|---|
| 1686 | and r12,rcx
|
|---|
| 1687 | not r13
|
|---|
| 1688 | mov r14,QWORD[48+rbp]
|
|---|
| 1689 | and r13,rcx
|
|---|
| 1690 | not r14
|
|---|
| 1691 | mov r15,QWORD[56+rbp]
|
|---|
| 1692 | and r14,rcx
|
|---|
| 1693 | not r15
|
|---|
| 1694 | and r15,rcx
|
|---|
| 1695 |
|
|---|
| 1696 | add r8,QWORD[rdi]
|
|---|
| 1697 | adc r9,QWORD[8+rdi]
|
|---|
| 1698 | adc r10,QWORD[16+rdi]
|
|---|
| 1699 | adc r11,QWORD[24+rdi]
|
|---|
| 1700 | adc r12,QWORD[32+rdi]
|
|---|
| 1701 | adc r13,QWORD[40+rdi]
|
|---|
| 1702 | adc r14,QWORD[48+rdi]
|
|---|
| 1703 | adc r15,QWORD[56+rdi]
|
|---|
| 1704 |
|
|---|
| 1705 | mov QWORD[rdi],r8
|
|---|
| 1706 | mov QWORD[8+rdi],r9
|
|---|
| 1707 | mov QWORD[16+rdi],r10
|
|---|
| 1708 | mov QWORD[24+rdi],r11
|
|---|
| 1709 | mov QWORD[32+rdi],r12
|
|---|
| 1710 | mov QWORD[40+rdi],r13
|
|---|
| 1711 | mov QWORD[48+rdi],r14
|
|---|
| 1712 | mov QWORD[56+rdi],r15
|
|---|
| 1713 |
|
|---|
| 1714 | DB 0F3h,0C3h ;repret
|
|---|
| 1715 |
|
|---|
| 1716 |
|
|---|
| 1717 |
|
|---|
| 1718 | ALIGN 32
|
|---|
| 1719 | __rsaz_512_mul:
|
|---|
| 1720 |
|
|---|
| 1721 | lea rdi,[8+rsp]
|
|---|
| 1722 |
|
|---|
| 1723 | mov rax,QWORD[rsi]
|
|---|
| 1724 | mul rbx
|
|---|
| 1725 | mov QWORD[rdi],rax
|
|---|
| 1726 | mov rax,QWORD[8+rsi]
|
|---|
| 1727 | mov r8,rdx
|
|---|
| 1728 |
|
|---|
| 1729 | mul rbx
|
|---|
| 1730 | add r8,rax
|
|---|
| 1731 | mov rax,QWORD[16+rsi]
|
|---|
| 1732 | mov r9,rdx
|
|---|
| 1733 | adc r9,0
|
|---|
| 1734 |
|
|---|
| 1735 | mul rbx
|
|---|
| 1736 | add r9,rax
|
|---|
| 1737 | mov rax,QWORD[24+rsi]
|
|---|
| 1738 | mov r10,rdx
|
|---|
| 1739 | adc r10,0
|
|---|
| 1740 |
|
|---|
| 1741 | mul rbx
|
|---|
| 1742 | add r10,rax
|
|---|
| 1743 | mov rax,QWORD[32+rsi]
|
|---|
| 1744 | mov r11,rdx
|
|---|
| 1745 | adc r11,0
|
|---|
| 1746 |
|
|---|
| 1747 | mul rbx
|
|---|
| 1748 | add r11,rax
|
|---|
| 1749 | mov rax,QWORD[40+rsi]
|
|---|
| 1750 | mov r12,rdx
|
|---|
| 1751 | adc r12,0
|
|---|
| 1752 |
|
|---|
| 1753 | mul rbx
|
|---|
| 1754 | add r12,rax
|
|---|
| 1755 | mov rax,QWORD[48+rsi]
|
|---|
| 1756 | mov r13,rdx
|
|---|
| 1757 | adc r13,0
|
|---|
| 1758 |
|
|---|
| 1759 | mul rbx
|
|---|
| 1760 | add r13,rax
|
|---|
| 1761 | mov rax,QWORD[56+rsi]
|
|---|
| 1762 | mov r14,rdx
|
|---|
| 1763 | adc r14,0
|
|---|
| 1764 |
|
|---|
| 1765 | mul rbx
|
|---|
| 1766 | add r14,rax
|
|---|
| 1767 | mov rax,QWORD[rsi]
|
|---|
| 1768 | mov r15,rdx
|
|---|
| 1769 | adc r15,0
|
|---|
| 1770 |
|
|---|
| 1771 | lea rbp,[8+rbp]
|
|---|
| 1772 | lea rdi,[8+rdi]
|
|---|
| 1773 |
|
|---|
| 1774 | mov ecx,7
|
|---|
| 1775 | jmp NEAR $L$oop_mul
|
|---|
| 1776 |
|
|---|
| 1777 | ALIGN 32
|
|---|
| 1778 | $L$oop_mul:
|
|---|
| 1779 | mov rbx,QWORD[rbp]
|
|---|
| 1780 | mul rbx
|
|---|
| 1781 | add r8,rax
|
|---|
| 1782 | mov rax,QWORD[8+rsi]
|
|---|
| 1783 | mov QWORD[rdi],r8
|
|---|
| 1784 | mov r8,rdx
|
|---|
| 1785 | adc r8,0
|
|---|
| 1786 |
|
|---|
| 1787 | mul rbx
|
|---|
| 1788 | add r9,rax
|
|---|
| 1789 | mov rax,QWORD[16+rsi]
|
|---|
| 1790 | adc rdx,0
|
|---|
| 1791 | add r8,r9
|
|---|
| 1792 | mov r9,rdx
|
|---|
| 1793 | adc r9,0
|
|---|
| 1794 |
|
|---|
| 1795 | mul rbx
|
|---|
| 1796 | add r10,rax
|
|---|
| 1797 | mov rax,QWORD[24+rsi]
|
|---|
| 1798 | adc rdx,0
|
|---|
| 1799 | add r9,r10
|
|---|
| 1800 | mov r10,rdx
|
|---|
| 1801 | adc r10,0
|
|---|
| 1802 |
|
|---|
| 1803 | mul rbx
|
|---|
| 1804 | add r11,rax
|
|---|
| 1805 | mov rax,QWORD[32+rsi]
|
|---|
| 1806 | adc rdx,0
|
|---|
| 1807 | add r10,r11
|
|---|
| 1808 | mov r11,rdx
|
|---|
| 1809 | adc r11,0
|
|---|
| 1810 |
|
|---|
| 1811 | mul rbx
|
|---|
| 1812 | add r12,rax
|
|---|
| 1813 | mov rax,QWORD[40+rsi]
|
|---|
| 1814 | adc rdx,0
|
|---|
| 1815 | add r11,r12
|
|---|
| 1816 | mov r12,rdx
|
|---|
| 1817 | adc r12,0
|
|---|
| 1818 |
|
|---|
| 1819 | mul rbx
|
|---|
| 1820 | add r13,rax
|
|---|
| 1821 | mov rax,QWORD[48+rsi]
|
|---|
| 1822 | adc rdx,0
|
|---|
| 1823 | add r12,r13
|
|---|
| 1824 | mov r13,rdx
|
|---|
| 1825 | adc r13,0
|
|---|
| 1826 |
|
|---|
| 1827 | mul rbx
|
|---|
| 1828 | add r14,rax
|
|---|
| 1829 | mov rax,QWORD[56+rsi]
|
|---|
| 1830 | adc rdx,0
|
|---|
| 1831 | add r13,r14
|
|---|
| 1832 | mov r14,rdx
|
|---|
| 1833 | lea rbp,[8+rbp]
|
|---|
| 1834 | adc r14,0
|
|---|
| 1835 |
|
|---|
| 1836 | mul rbx
|
|---|
| 1837 | add r15,rax
|
|---|
| 1838 | mov rax,QWORD[rsi]
|
|---|
| 1839 | adc rdx,0
|
|---|
| 1840 | add r14,r15
|
|---|
| 1841 | mov r15,rdx
|
|---|
| 1842 | adc r15,0
|
|---|
| 1843 |
|
|---|
| 1844 | lea rdi,[8+rdi]
|
|---|
| 1845 |
|
|---|
| 1846 | dec ecx
|
|---|
| 1847 | jnz NEAR $L$oop_mul
|
|---|
| 1848 |
|
|---|
| 1849 | mov QWORD[rdi],r8
|
|---|
| 1850 | mov QWORD[8+rdi],r9
|
|---|
| 1851 | mov QWORD[16+rdi],r10
|
|---|
| 1852 | mov QWORD[24+rdi],r11
|
|---|
| 1853 | mov QWORD[32+rdi],r12
|
|---|
| 1854 | mov QWORD[40+rdi],r13
|
|---|
| 1855 | mov QWORD[48+rdi],r14
|
|---|
| 1856 | mov QWORD[56+rdi],r15
|
|---|
| 1857 |
|
|---|
| 1858 | DB 0F3h,0C3h ;repret
|
|---|
| 1859 |
|
|---|
| 1860 |
|
|---|
| 1861 |
|
|---|
| 1862 | ALIGN 32
|
|---|
| 1863 | __rsaz_512_mulx:
|
|---|
| 1864 |
|
|---|
| 1865 | mulx r8,rbx,QWORD[rsi]
|
|---|
| 1866 | mov rcx,-6
|
|---|
| 1867 |
|
|---|
| 1868 | mulx r9,rax,QWORD[8+rsi]
|
|---|
| 1869 | mov QWORD[8+rsp],rbx
|
|---|
| 1870 |
|
|---|
| 1871 | mulx r10,rbx,QWORD[16+rsi]
|
|---|
| 1872 | adc r8,rax
|
|---|
| 1873 |
|
|---|
| 1874 | mulx r11,rax,QWORD[24+rsi]
|
|---|
| 1875 | adc r9,rbx
|
|---|
| 1876 |
|
|---|
| 1877 | mulx r12,rbx,QWORD[32+rsi]
|
|---|
| 1878 | adc r10,rax
|
|---|
| 1879 |
|
|---|
| 1880 | mulx r13,rax,QWORD[40+rsi]
|
|---|
| 1881 | adc r11,rbx
|
|---|
| 1882 |
|
|---|
| 1883 | mulx r14,rbx,QWORD[48+rsi]
|
|---|
| 1884 | adc r12,rax
|
|---|
| 1885 |
|
|---|
| 1886 | mulx r15,rax,QWORD[56+rsi]
|
|---|
| 1887 | mov rdx,QWORD[8+rbp]
|
|---|
| 1888 | adc r13,rbx
|
|---|
| 1889 | adc r14,rax
|
|---|
| 1890 | adc r15,0
|
|---|
| 1891 |
|
|---|
| 1892 | xor rdi,rdi
|
|---|
| 1893 | jmp NEAR $L$oop_mulx
|
|---|
| 1894 |
|
|---|
| 1895 | ALIGN 32
|
|---|
| 1896 | $L$oop_mulx:
|
|---|
| 1897 | mov rbx,r8
|
|---|
| 1898 | mulx r8,rax,QWORD[rsi]
|
|---|
| 1899 | adcx rbx,rax
|
|---|
| 1900 | adox r8,r9
|
|---|
| 1901 |
|
|---|
| 1902 | mulx r9,rax,QWORD[8+rsi]
|
|---|
| 1903 | adcx r8,rax
|
|---|
| 1904 | adox r9,r10
|
|---|
| 1905 |
|
|---|
| 1906 | mulx r10,rax,QWORD[16+rsi]
|
|---|
| 1907 | adcx r9,rax
|
|---|
| 1908 | adox r10,r11
|
|---|
| 1909 |
|
|---|
| 1910 | mulx r11,rax,QWORD[24+rsi]
|
|---|
| 1911 | adcx r10,rax
|
|---|
| 1912 | adox r11,r12
|
|---|
| 1913 |
|
|---|
| 1914 | DB 0x3e,0xc4,0x62,0xfb,0xf6,0xa6,0x20,0x00,0x00,0x00
|
|---|
| 1915 | adcx r11,rax
|
|---|
| 1916 | adox r12,r13
|
|---|
| 1917 |
|
|---|
| 1918 | mulx r13,rax,QWORD[40+rsi]
|
|---|
| 1919 | adcx r12,rax
|
|---|
| 1920 | adox r13,r14
|
|---|
| 1921 |
|
|---|
| 1922 | mulx r14,rax,QWORD[48+rsi]
|
|---|
| 1923 | adcx r13,rax
|
|---|
| 1924 | adox r14,r15
|
|---|
| 1925 |
|
|---|
| 1926 | mulx r15,rax,QWORD[56+rsi]
|
|---|
| 1927 | mov rdx,QWORD[64+rcx*8+rbp]
|
|---|
| 1928 | mov QWORD[((8+64-8))+rcx*8+rsp],rbx
|
|---|
| 1929 | adcx r14,rax
|
|---|
| 1930 | adox r15,rdi
|
|---|
| 1931 | adcx r15,rdi
|
|---|
| 1932 |
|
|---|
| 1933 | inc rcx
|
|---|
| 1934 | jnz NEAR $L$oop_mulx
|
|---|
| 1935 |
|
|---|
| 1936 | mov rbx,r8
|
|---|
| 1937 | mulx r8,rax,QWORD[rsi]
|
|---|
| 1938 | adcx rbx,rax
|
|---|
| 1939 | adox r8,r9
|
|---|
| 1940 |
|
|---|
| 1941 | DB 0xc4,0x62,0xfb,0xf6,0x8e,0x08,0x00,0x00,0x00
|
|---|
| 1942 | adcx r8,rax
|
|---|
| 1943 | adox r9,r10
|
|---|
| 1944 |
|
|---|
| 1945 | DB 0xc4,0x62,0xfb,0xf6,0x96,0x10,0x00,0x00,0x00
|
|---|
| 1946 | adcx r9,rax
|
|---|
| 1947 | adox r10,r11
|
|---|
| 1948 |
|
|---|
| 1949 | mulx r11,rax,QWORD[24+rsi]
|
|---|
| 1950 | adcx r10,rax
|
|---|
| 1951 | adox r11,r12
|
|---|
| 1952 |
|
|---|
| 1953 | mulx r12,rax,QWORD[32+rsi]
|
|---|
| 1954 | adcx r11,rax
|
|---|
| 1955 | adox r12,r13
|
|---|
| 1956 |
|
|---|
| 1957 | mulx r13,rax,QWORD[40+rsi]
|
|---|
| 1958 | adcx r12,rax
|
|---|
| 1959 | adox r13,r14
|
|---|
| 1960 |
|
|---|
| 1961 | DB 0xc4,0x62,0xfb,0xf6,0xb6,0x30,0x00,0x00,0x00
|
|---|
| 1962 | adcx r13,rax
|
|---|
| 1963 | adox r14,r15
|
|---|
| 1964 |
|
|---|
| 1965 | DB 0xc4,0x62,0xfb,0xf6,0xbe,0x38,0x00,0x00,0x00
|
|---|
| 1966 | adcx r14,rax
|
|---|
| 1967 | adox r15,rdi
|
|---|
| 1968 | adcx r15,rdi
|
|---|
| 1969 |
|
|---|
| 1970 | mov QWORD[((8+64-8))+rsp],rbx
|
|---|
| 1971 | mov QWORD[((8+64))+rsp],r8
|
|---|
| 1972 | mov QWORD[((8+64+8))+rsp],r9
|
|---|
| 1973 | mov QWORD[((8+64+16))+rsp],r10
|
|---|
| 1974 | mov QWORD[((8+64+24))+rsp],r11
|
|---|
| 1975 | mov QWORD[((8+64+32))+rsp],r12
|
|---|
| 1976 | mov QWORD[((8+64+40))+rsp],r13
|
|---|
| 1977 | mov QWORD[((8+64+48))+rsp],r14
|
|---|
| 1978 | mov QWORD[((8+64+56))+rsp],r15
|
|---|
| 1979 |
|
|---|
| 1980 | DB 0F3h,0C3h ;repret
|
|---|
| 1981 |
|
|---|
| 1982 |
|
|---|
| 1983 | global rsaz_512_scatter4
|
|---|
| 1984 |
|
|---|
| 1985 | ALIGN 16
|
|---|
| 1986 | rsaz_512_scatter4:
|
|---|
| 1987 |
|
|---|
| 1988 | lea rcx,[r8*8+rcx]
|
|---|
| 1989 | mov r9d,8
|
|---|
| 1990 | jmp NEAR $L$oop_scatter
|
|---|
| 1991 | ALIGN 16
|
|---|
| 1992 | $L$oop_scatter:
|
|---|
| 1993 | mov rax,QWORD[rdx]
|
|---|
| 1994 | lea rdx,[8+rdx]
|
|---|
| 1995 | mov QWORD[rcx],rax
|
|---|
| 1996 | lea rcx,[128+rcx]
|
|---|
| 1997 | dec r9d
|
|---|
| 1998 | jnz NEAR $L$oop_scatter
|
|---|
| 1999 | DB 0F3h,0C3h ;repret
|
|---|
| 2000 |
|
|---|
| 2001 |
|
|---|
| 2002 |
|
|---|
| 2003 | global rsaz_512_gather4
|
|---|
| 2004 |
|
|---|
| 2005 | ALIGN 16
|
|---|
| 2006 | rsaz_512_gather4:
|
|---|
| 2007 |
|
|---|
| 2008 | $L$SEH_begin_rsaz_512_gather4:
|
|---|
| 2009 | DB 0x48,0x81,0xec,0xa8,0x00,0x00,0x00
|
|---|
| 2010 | DB 0x0f,0x29,0x34,0x24
|
|---|
| 2011 | DB 0x0f,0x29,0x7c,0x24,0x10
|
|---|
| 2012 | DB 0x44,0x0f,0x29,0x44,0x24,0x20
|
|---|
| 2013 | DB 0x44,0x0f,0x29,0x4c,0x24,0x30
|
|---|
| 2014 | DB 0x44,0x0f,0x29,0x54,0x24,0x40
|
|---|
| 2015 | DB 0x44,0x0f,0x29,0x5c,0x24,0x50
|
|---|
| 2016 | DB 0x44,0x0f,0x29,0x64,0x24,0x60
|
|---|
| 2017 | DB 0x44,0x0f,0x29,0x6c,0x24,0x70
|
|---|
| 2018 | DB 0x44,0x0f,0x29,0xb4,0x24,0x80,0,0,0
|
|---|
| 2019 | DB 0x44,0x0f,0x29,0xbc,0x24,0x90,0,0,0
|
|---|
| 2020 | movd xmm8,r8d
|
|---|
| 2021 | movdqa xmm1,XMMWORD[(($L$inc+16))]
|
|---|
| 2022 | movdqa xmm0,XMMWORD[$L$inc]
|
|---|
| 2023 |
|
|---|
| 2024 | pshufd xmm8,xmm8,0
|
|---|
| 2025 | movdqa xmm7,xmm1
|
|---|
| 2026 | movdqa xmm2,xmm1
|
|---|
| 2027 | paddd xmm1,xmm0
|
|---|
| 2028 | pcmpeqd xmm0,xmm8
|
|---|
| 2029 | movdqa xmm3,xmm7
|
|---|
| 2030 | paddd xmm2,xmm1
|
|---|
| 2031 | pcmpeqd xmm1,xmm8
|
|---|
| 2032 | movdqa xmm4,xmm7
|
|---|
| 2033 | paddd xmm3,xmm2
|
|---|
| 2034 | pcmpeqd xmm2,xmm8
|
|---|
| 2035 | movdqa xmm5,xmm7
|
|---|
| 2036 | paddd xmm4,xmm3
|
|---|
| 2037 | pcmpeqd xmm3,xmm8
|
|---|
| 2038 | movdqa xmm6,xmm7
|
|---|
| 2039 | paddd xmm5,xmm4
|
|---|
| 2040 | pcmpeqd xmm4,xmm8
|
|---|
| 2041 | paddd xmm6,xmm5
|
|---|
| 2042 | pcmpeqd xmm5,xmm8
|
|---|
| 2043 | paddd xmm7,xmm6
|
|---|
| 2044 | pcmpeqd xmm6,xmm8
|
|---|
| 2045 | pcmpeqd xmm7,xmm8
|
|---|
| 2046 | mov r9d,8
|
|---|
| 2047 | jmp NEAR $L$oop_gather
|
|---|
| 2048 | ALIGN 16
|
|---|
| 2049 | $L$oop_gather:
|
|---|
| 2050 | movdqa xmm8,XMMWORD[rdx]
|
|---|
| 2051 | movdqa xmm9,XMMWORD[16+rdx]
|
|---|
| 2052 | movdqa xmm10,XMMWORD[32+rdx]
|
|---|
| 2053 | movdqa xmm11,XMMWORD[48+rdx]
|
|---|
| 2054 | pand xmm8,xmm0
|
|---|
| 2055 | movdqa xmm12,XMMWORD[64+rdx]
|
|---|
| 2056 | pand xmm9,xmm1
|
|---|
| 2057 | movdqa xmm13,XMMWORD[80+rdx]
|
|---|
| 2058 | pand xmm10,xmm2
|
|---|
| 2059 | movdqa xmm14,XMMWORD[96+rdx]
|
|---|
| 2060 | pand xmm11,xmm3
|
|---|
| 2061 | movdqa xmm15,XMMWORD[112+rdx]
|
|---|
| 2062 | lea rdx,[128+rdx]
|
|---|
| 2063 | pand xmm12,xmm4
|
|---|
| 2064 | pand xmm13,xmm5
|
|---|
| 2065 | pand xmm14,xmm6
|
|---|
| 2066 | pand xmm15,xmm7
|
|---|
| 2067 | por xmm8,xmm10
|
|---|
| 2068 | por xmm9,xmm11
|
|---|
| 2069 | por xmm8,xmm12
|
|---|
| 2070 | por xmm9,xmm13
|
|---|
| 2071 | por xmm8,xmm14
|
|---|
| 2072 | por xmm9,xmm15
|
|---|
| 2073 |
|
|---|
| 2074 | por xmm8,xmm9
|
|---|
| 2075 | pshufd xmm9,xmm8,0x4e
|
|---|
| 2076 | por xmm8,xmm9
|
|---|
| 2077 | movq QWORD[rcx],xmm8
|
|---|
| 2078 | lea rcx,[8+rcx]
|
|---|
| 2079 | dec r9d
|
|---|
| 2080 | jnz NEAR $L$oop_gather
|
|---|
| 2081 | movaps xmm6,XMMWORD[rsp]
|
|---|
| 2082 | movaps xmm7,XMMWORD[16+rsp]
|
|---|
| 2083 | movaps xmm8,XMMWORD[32+rsp]
|
|---|
| 2084 | movaps xmm9,XMMWORD[48+rsp]
|
|---|
| 2085 | movaps xmm10,XMMWORD[64+rsp]
|
|---|
| 2086 | movaps xmm11,XMMWORD[80+rsp]
|
|---|
| 2087 | movaps xmm12,XMMWORD[96+rsp]
|
|---|
| 2088 | movaps xmm13,XMMWORD[112+rsp]
|
|---|
| 2089 | movaps xmm14,XMMWORD[128+rsp]
|
|---|
| 2090 | movaps xmm15,XMMWORD[144+rsp]
|
|---|
| 2091 | add rsp,0xa8
|
|---|
| 2092 | DB 0F3h,0C3h ;repret
|
|---|
| 2093 | $L$SEH_end_rsaz_512_gather4:
|
|---|
| 2094 |
|
|---|
| 2095 |
|
|---|
| 2096 |
|
|---|
| 2097 | ALIGN 64
|
|---|
| 2098 | $L$inc:
|
|---|
| 2099 | DD 0,0,1,1
|
|---|
| 2100 | DD 2,2,2,2
|
|---|
| 2101 | EXTERN __imp_RtlVirtualUnwind
|
|---|
| 2102 |
|
|---|
| 2103 | ALIGN 16
|
|---|
| 2104 | se_handler:
|
|---|
| 2105 | push rsi
|
|---|
| 2106 | push rdi
|
|---|
| 2107 | push rbx
|
|---|
| 2108 | push rbp
|
|---|
| 2109 | push r12
|
|---|
| 2110 | push r13
|
|---|
| 2111 | push r14
|
|---|
| 2112 | push r15
|
|---|
| 2113 | pushfq
|
|---|
| 2114 | sub rsp,64
|
|---|
| 2115 |
|
|---|
| 2116 | mov rax,QWORD[120+r8]
|
|---|
| 2117 | mov rbx,QWORD[248+r8]
|
|---|
| 2118 |
|
|---|
| 2119 | mov rsi,QWORD[8+r9]
|
|---|
| 2120 | mov r11,QWORD[56+r9]
|
|---|
| 2121 |
|
|---|
| 2122 | mov r10d,DWORD[r11]
|
|---|
| 2123 | lea r10,[r10*1+rsi]
|
|---|
| 2124 | cmp rbx,r10
|
|---|
| 2125 | jb NEAR $L$common_seh_tail
|
|---|
| 2126 |
|
|---|
| 2127 | mov rax,QWORD[152+r8]
|
|---|
| 2128 |
|
|---|
| 2129 | mov r10d,DWORD[4+r11]
|
|---|
| 2130 | lea r10,[r10*1+rsi]
|
|---|
| 2131 | cmp rbx,r10
|
|---|
| 2132 | jae NEAR $L$common_seh_tail
|
|---|
| 2133 |
|
|---|
| 2134 | lea rax,[((128+24+48))+rax]
|
|---|
| 2135 |
|
|---|
| 2136 | lea rbx,[$L$mul_gather4_epilogue]
|
|---|
| 2137 | cmp rbx,r10
|
|---|
| 2138 | jne NEAR $L$se_not_in_mul_gather4
|
|---|
| 2139 |
|
|---|
| 2140 | lea rax,[176+rax]
|
|---|
| 2141 |
|
|---|
| 2142 | lea rsi,[((-48-168))+rax]
|
|---|
| 2143 | lea rdi,[512+r8]
|
|---|
| 2144 | mov ecx,20
|
|---|
| 2145 | DD 0xa548f3fc
|
|---|
| 2146 |
|
|---|
| 2147 | $L$se_not_in_mul_gather4:
|
|---|
| 2148 | mov rbx,QWORD[((-8))+rax]
|
|---|
| 2149 | mov rbp,QWORD[((-16))+rax]
|
|---|
| 2150 | mov r12,QWORD[((-24))+rax]
|
|---|
| 2151 | mov r13,QWORD[((-32))+rax]
|
|---|
| 2152 | mov r14,QWORD[((-40))+rax]
|
|---|
| 2153 | mov r15,QWORD[((-48))+rax]
|
|---|
| 2154 | mov QWORD[144+r8],rbx
|
|---|
| 2155 | mov QWORD[160+r8],rbp
|
|---|
| 2156 | mov QWORD[216+r8],r12
|
|---|
| 2157 | mov QWORD[224+r8],r13
|
|---|
| 2158 | mov QWORD[232+r8],r14
|
|---|
| 2159 | mov QWORD[240+r8],r15
|
|---|
| 2160 |
|
|---|
| 2161 | $L$common_seh_tail:
|
|---|
| 2162 | mov rdi,QWORD[8+rax]
|
|---|
| 2163 | mov rsi,QWORD[16+rax]
|
|---|
| 2164 | mov QWORD[152+r8],rax
|
|---|
| 2165 | mov QWORD[168+r8],rsi
|
|---|
| 2166 | mov QWORD[176+r8],rdi
|
|---|
| 2167 |
|
|---|
| 2168 | mov rdi,QWORD[40+r9]
|
|---|
| 2169 | mov rsi,r8
|
|---|
| 2170 | mov ecx,154
|
|---|
| 2171 | DD 0xa548f3fc
|
|---|
| 2172 |
|
|---|
| 2173 | mov rsi,r9
|
|---|
| 2174 | xor rcx,rcx
|
|---|
| 2175 | mov rdx,QWORD[8+rsi]
|
|---|
| 2176 | mov r8,QWORD[rsi]
|
|---|
| 2177 | mov r9,QWORD[16+rsi]
|
|---|
| 2178 | mov r10,QWORD[40+rsi]
|
|---|
| 2179 | lea r11,[56+rsi]
|
|---|
| 2180 | lea r12,[24+rsi]
|
|---|
| 2181 | mov QWORD[32+rsp],r10
|
|---|
| 2182 | mov QWORD[40+rsp],r11
|
|---|
| 2183 | mov QWORD[48+rsp],r12
|
|---|
| 2184 | mov QWORD[56+rsp],rcx
|
|---|
| 2185 | call QWORD[__imp_RtlVirtualUnwind]
|
|---|
| 2186 |
|
|---|
| 2187 | mov eax,1
|
|---|
| 2188 | add rsp,64
|
|---|
| 2189 | popfq
|
|---|
| 2190 | pop r15
|
|---|
| 2191 | pop r14
|
|---|
| 2192 | pop r13
|
|---|
| 2193 | pop r12
|
|---|
| 2194 | pop rbp
|
|---|
| 2195 | pop rbx
|
|---|
| 2196 | pop rdi
|
|---|
| 2197 | pop rsi
|
|---|
| 2198 | DB 0F3h,0C3h ;repret
|
|---|
| 2199 |
|
|---|
| 2200 |
|
|---|
| 2201 | section .pdata rdata align=4
|
|---|
| 2202 | ALIGN 4
|
|---|
| 2203 | DD $L$SEH_begin_rsaz_512_sqr wrt ..imagebase
|
|---|
| 2204 | DD $L$SEH_end_rsaz_512_sqr wrt ..imagebase
|
|---|
| 2205 | DD $L$SEH_info_rsaz_512_sqr wrt ..imagebase
|
|---|
| 2206 |
|
|---|
| 2207 | DD $L$SEH_begin_rsaz_512_mul wrt ..imagebase
|
|---|
| 2208 | DD $L$SEH_end_rsaz_512_mul wrt ..imagebase
|
|---|
| 2209 | DD $L$SEH_info_rsaz_512_mul wrt ..imagebase
|
|---|
| 2210 |
|
|---|
| 2211 | DD $L$SEH_begin_rsaz_512_mul_gather4 wrt ..imagebase
|
|---|
| 2212 | DD $L$SEH_end_rsaz_512_mul_gather4 wrt ..imagebase
|
|---|
| 2213 | DD $L$SEH_info_rsaz_512_mul_gather4 wrt ..imagebase
|
|---|
| 2214 |
|
|---|
| 2215 | DD $L$SEH_begin_rsaz_512_mul_scatter4 wrt ..imagebase
|
|---|
| 2216 | DD $L$SEH_end_rsaz_512_mul_scatter4 wrt ..imagebase
|
|---|
| 2217 | DD $L$SEH_info_rsaz_512_mul_scatter4 wrt ..imagebase
|
|---|
| 2218 |
|
|---|
| 2219 | DD $L$SEH_begin_rsaz_512_mul_by_one wrt ..imagebase
|
|---|
| 2220 | DD $L$SEH_end_rsaz_512_mul_by_one wrt ..imagebase
|
|---|
| 2221 | DD $L$SEH_info_rsaz_512_mul_by_one wrt ..imagebase
|
|---|
| 2222 |
|
|---|
| 2223 | DD $L$SEH_begin_rsaz_512_gather4 wrt ..imagebase
|
|---|
| 2224 | DD $L$SEH_end_rsaz_512_gather4 wrt ..imagebase
|
|---|
| 2225 | DD $L$SEH_info_rsaz_512_gather4 wrt ..imagebase
|
|---|
| 2226 |
|
|---|
| 2227 | section .xdata rdata align=8
|
|---|
| 2228 | ALIGN 8
|
|---|
| 2229 | $L$SEH_info_rsaz_512_sqr:
|
|---|
| 2230 | DB 9,0,0,0
|
|---|
| 2231 | DD se_handler wrt ..imagebase
|
|---|
| 2232 | DD $L$sqr_body wrt ..imagebase,$L$sqr_epilogue wrt ..imagebase
|
|---|
| 2233 | $L$SEH_info_rsaz_512_mul:
|
|---|
| 2234 | DB 9,0,0,0
|
|---|
| 2235 | DD se_handler wrt ..imagebase
|
|---|
| 2236 | DD $L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase
|
|---|
| 2237 | $L$SEH_info_rsaz_512_mul_gather4:
|
|---|
| 2238 | DB 9,0,0,0
|
|---|
| 2239 | DD se_handler wrt ..imagebase
|
|---|
| 2240 | DD $L$mul_gather4_body wrt ..imagebase,$L$mul_gather4_epilogue wrt ..imagebase
|
|---|
| 2241 | $L$SEH_info_rsaz_512_mul_scatter4:
|
|---|
| 2242 | DB 9,0,0,0
|
|---|
| 2243 | DD se_handler wrt ..imagebase
|
|---|
| 2244 | DD $L$mul_scatter4_body wrt ..imagebase,$L$mul_scatter4_epilogue wrt ..imagebase
|
|---|
| 2245 | $L$SEH_info_rsaz_512_mul_by_one:
|
|---|
| 2246 | DB 9,0,0,0
|
|---|
| 2247 | DD se_handler wrt ..imagebase
|
|---|
| 2248 | DD $L$mul_by_one_body wrt ..imagebase,$L$mul_by_one_epilogue wrt ..imagebase
|
|---|
| 2249 | $L$SEH_info_rsaz_512_gather4:
|
|---|
| 2250 | DB 0x01,0x46,0x16,0x00
|
|---|
| 2251 | DB 0x46,0xf8,0x09,0x00
|
|---|
| 2252 | DB 0x3d,0xe8,0x08,0x00
|
|---|
| 2253 | DB 0x34,0xd8,0x07,0x00
|
|---|
| 2254 | DB 0x2e,0xc8,0x06,0x00
|
|---|
| 2255 | DB 0x28,0xb8,0x05,0x00
|
|---|
| 2256 | DB 0x22,0xa8,0x04,0x00
|
|---|
| 2257 | DB 0x1c,0x98,0x03,0x00
|
|---|
| 2258 | DB 0x16,0x88,0x02,0x00
|
|---|
| 2259 | DB 0x10,0x78,0x01,0x00
|
|---|
| 2260 | DB 0x0b,0x68,0x00,0x00
|
|---|
| 2261 | DB 0x07,0x01,0x15,0x00
|
|---|
