1 # This file is generated from a similarly-named Perl script in the BoringSSL 2 # source tree. Do not edit by hand. 3 4 #if defined(__i386__) 5 #if defined(BORINGSSL_PREFIX) 6 #include <boringssl_prefix_symbols_asm.h> 7 #endif 8 .text 9 .globl bn_mul_mont 10 .hidden bn_mul_mont 11 .type bn_mul_mont,@function 12 .align 16 13 bn_mul_mont: 14 .L_bn_mul_mont_begin: 15 pushl %ebp 16 pushl %ebx 17 pushl %esi 18 pushl %edi 19 xorl %eax,%eax 20 movl 40(%esp),%edi 21 cmpl $4,%edi 22 jl .L000just_leave 23 leal 20(%esp),%esi 24 leal 24(%esp),%edx 25 addl $2,%edi 26 negl %edi 27 leal -32(%esp,%edi,4),%ebp 28 negl %edi 29 movl %ebp,%eax 30 subl %edx,%eax 31 andl $2047,%eax 32 subl %eax,%ebp 33 xorl %ebp,%edx 34 andl $2048,%edx 35 xorl $2048,%edx 36 subl %edx,%ebp 37 andl $-64,%ebp 38 movl %esp,%eax 39 subl %ebp,%eax 40 andl $-4096,%eax 41 movl %esp,%edx 42 leal (%ebp,%eax,1),%esp 43 movl (%esp),%eax 44 cmpl %ebp,%esp 45 ja .L001page_walk 46 jmp .L002page_walk_done 47 .align 16 48 .L001page_walk: 49 leal -4096(%esp),%esp 50 movl (%esp),%eax 51 cmpl %ebp,%esp 52 ja .L001page_walk 53 .L002page_walk_done: 54 movl (%esi),%eax 55 movl 4(%esi),%ebx 56 movl 8(%esi),%ecx 57 movl 12(%esi),%ebp 58 movl 16(%esi),%esi 59 movl (%esi),%esi 60 movl %eax,4(%esp) 61 movl %ebx,8(%esp) 62 movl %ecx,12(%esp) 63 movl %ebp,16(%esp) 64 movl %esi,20(%esp) 65 leal -3(%edi),%ebx 66 movl %edx,24(%esp) 67 call .L003PIC_me_up 68 .L003PIC_me_up: 69 popl %eax 70 leal OPENSSL_ia32cap_P-.L003PIC_me_up(%eax),%eax 71 btl $26,(%eax) 72 jnc .L004non_sse2 73 movl $-1,%eax 74 movd %eax,%mm7 75 movl 8(%esp),%esi 76 movl 12(%esp),%edi 77 movl 16(%esp),%ebp 78 xorl %edx,%edx 79 xorl %ecx,%ecx 80 movd (%edi),%mm4 81 movd (%esi),%mm5 82 movd (%ebp),%mm3 83 pmuludq %mm4,%mm5 84 movq %mm5,%mm2 85 movq %mm5,%mm0 86 pand %mm7,%mm0 87 pmuludq 20(%esp),%mm5 88 pmuludq %mm5,%mm3 89 paddq %mm0,%mm3 90 movd 4(%ebp),%mm1 91 movd 4(%esi),%mm0 92 psrlq $32,%mm2 93 psrlq $32,%mm3 94 incl %ecx 95 .align 16 96 .L0051st: 97 pmuludq %mm4,%mm0 98 pmuludq %mm5,%mm1 99 paddq %mm0,%mm2 100 paddq %mm1,%mm3 101 movq %mm2,%mm0 102 pand %mm7,%mm0 103 movd 4(%ebp,%ecx,4),%mm1 104 paddq %mm0,%mm3 105 movd 4(%esi,%ecx,4),%mm0 106 psrlq $32,%mm2 107 movd %mm3,28(%esp,%ecx,4) 108 psrlq $32,%mm3 109 leal 1(%ecx),%ecx 110 cmpl %ebx,%ecx 111 jl .L0051st 112 pmuludq %mm4,%mm0 113 pmuludq %mm5,%mm1 114 paddq %mm0,%mm2 115 paddq %mm1,%mm3 116 movq %mm2,%mm0 117 pand %mm7,%mm0 118 paddq %mm0,%mm3 119 movd %mm3,28(%esp,%ecx,4) 120 psrlq $32,%mm2 121 psrlq $32,%mm3 122 paddq %mm2,%mm3 123 movq %mm3,32(%esp,%ebx,4) 124 incl %edx 125 .L006outer: 126 xorl %ecx,%ecx 127 movd (%edi,%edx,4),%mm4 128 movd (%esi),%mm5 129 movd 32(%esp),%mm6 130 movd (%ebp),%mm3 131 pmuludq %mm4,%mm5 132 paddq %mm6,%mm5 133 movq %mm5,%mm0 134 movq %mm5,%mm2 135 pand %mm7,%mm0 136 pmuludq 20(%esp),%mm5 137 pmuludq %mm5,%mm3 138 paddq %mm0,%mm3 139 movd 36(%esp),%mm6 140 movd 4(%ebp),%mm1 141 movd 4(%esi),%mm0 142 psrlq $32,%mm2 143 psrlq $32,%mm3 144 paddq %mm6,%mm2 145 incl %ecx 146 decl %ebx 147 .L007inner: 148 pmuludq %mm4,%mm0 149 pmuludq %mm5,%mm1 150 paddq %mm0,%mm2 151 paddq %mm1,%mm3 152 movq %mm2,%mm0 153 movd 36(%esp,%ecx,4),%mm6 154 pand %mm7,%mm0 155 movd 4(%ebp,%ecx,4),%mm1 156 paddq %mm0,%mm3 157 movd 4(%esi,%ecx,4),%mm0 158 psrlq $32,%mm2 159 movd %mm3,28(%esp,%ecx,4) 160 psrlq $32,%mm3 161 paddq %mm6,%mm2 162 decl %ebx 163 leal 1(%ecx),%ecx 164 jnz .L007inner 165 movl %ecx,%ebx 166 pmuludq %mm4,%mm0 167 pmuludq %mm5,%mm1 168 paddq %mm0,%mm2 169 paddq %mm1,%mm3 170 movq %mm2,%mm0 171 pand %mm7,%mm0 172 paddq %mm0,%mm3 173 movd %mm3,28(%esp,%ecx,4) 174 psrlq $32,%mm2 175 psrlq $32,%mm3 176 movd 36(%esp,%ebx,4),%mm6 177 paddq %mm2,%mm3 178 paddq %mm6,%mm3 179 movq %mm3,32(%esp,%ebx,4) 180 leal 1(%edx),%edx 181 cmpl %ebx,%edx 182 jle .L006outer 183 emms 184 jmp .L008common_tail 185 .align 16 186 .L004non_sse2: 187 movl 8(%esp),%esi 188 leal 1(%ebx),%ebp 189 movl 12(%esp),%edi 190 xorl %ecx,%ecx 191 movl %esi,%edx 192 andl $1,%ebp 193 subl %edi,%edx 194 leal 4(%edi,%ebx,4),%eax 195 orl %edx,%ebp 196 movl (%edi),%edi 197 jz .L009bn_sqr_mont 198 movl %eax,28(%esp) 199 movl (%esi),%eax 200 xorl %edx,%edx 201 .align 16 202 .L010mull: 203 movl %edx,%ebp 204 mull %edi 205 addl %eax,%ebp 206 leal 1(%ecx),%ecx 207 adcl $0,%edx 208 movl (%esi,%ecx,4),%eax 209 cmpl %ebx,%ecx 210 movl %ebp,28(%esp,%ecx,4) 211 jl .L010mull 212 movl %edx,%ebp 213 mull %edi 214 movl 20(%esp),%edi 215 addl %ebp,%eax 216 movl 16(%esp),%esi 217 adcl $0,%edx 218 imull 32(%esp),%edi 219 movl %eax,32(%esp,%ebx,4) 220 xorl %ecx,%ecx 221 movl %edx,36(%esp,%ebx,4) 222 movl %ecx,40(%esp,%ebx,4) 223 movl (%esi),%eax 224 mull %edi 225 addl 32(%esp),%eax 226 movl 4(%esi),%eax 227 adcl $0,%edx 228 incl %ecx 229 jmp .L0112ndmadd 230 .align 16 231 .L0121stmadd: 232 movl %edx,%ebp 233 mull %edi 234 addl 32(%esp,%ecx,4),%ebp 235 leal 1(%ecx),%ecx 236 adcl $0,%edx 237 addl %eax,%ebp 238 movl (%esi,%ecx,4),%eax 239 adcl $0,%edx 240 cmpl %ebx,%ecx 241 movl %ebp,28(%esp,%ecx,4) 242 jl .L0121stmadd 243 movl %edx,%ebp 244 mull %edi 245 addl 32(%esp,%ebx,4),%eax 246 movl 20(%esp),%edi 247 adcl $0,%edx 248 movl 16(%esp),%esi 249 addl %eax,%ebp 250 adcl $0,%edx 251 imull 32(%esp),%edi 252 xorl %ecx,%ecx 253 addl 36(%esp,%ebx,4),%edx 254 movl %ebp,32(%esp,%ebx,4) 255 adcl $0,%ecx 256 movl (%esi),%eax 257 movl %edx,36(%esp,%ebx,4) 258 movl %ecx,40(%esp,%ebx,4) 259 mull %edi 260 addl 32(%esp),%eax 261 movl 4(%esi),%eax 262 adcl $0,%edx 263 movl $1,%ecx 264 .align 16 265 .L0112ndmadd: 266 movl %edx,%ebp 267 mull %edi 268 addl 32(%esp,%ecx,4),%ebp 269 leal 1(%ecx),%ecx 270 adcl $0,%edx 271 addl %eax,%ebp 272 movl (%esi,%ecx,4),%eax 273 adcl $0,%edx 274 cmpl %ebx,%ecx 275 movl %ebp,24(%esp,%ecx,4) 276 jl .L0112ndmadd 277 movl %edx,%ebp 278 mull %edi 279 addl 32(%esp,%ebx,4),%ebp 280 adcl $0,%edx 281 addl %eax,%ebp 282 adcl $0,%edx 283 movl %ebp,28(%esp,%ebx,4) 284 xorl %eax,%eax 285 movl 12(%esp),%ecx 286 addl 36(%esp,%ebx,4),%edx 287 adcl 40(%esp,%ebx,4),%eax 288 leal 4(%ecx),%ecx 289 movl %edx,32(%esp,%ebx,4) 290 cmpl 28(%esp),%ecx 291 movl %eax,36(%esp,%ebx,4) 292 je .L008common_tail 293 movl (%ecx),%edi 294 movl 8(%esp),%esi 295 movl %ecx,12(%esp) 296 xorl %ecx,%ecx 297 xorl %edx,%edx 298 movl (%esi),%eax 299 jmp .L0121stmadd 300 .align 16 301 .L009bn_sqr_mont: 302 movl %ebx,(%esp) 303 movl %ecx,12(%esp) 304 movl %edi,%eax 305 mull %edi 306 movl %eax,32(%esp) 307 movl %edx,%ebx 308 shrl $1,%edx 309 andl $1,%ebx 310 incl %ecx 311 .align 16 312 .L013sqr: 313 movl (%esi,%ecx,4),%eax 314 movl %edx,%ebp 315 mull %edi 316 addl %ebp,%eax 317 leal 1(%ecx),%ecx 318 adcl $0,%edx 319 leal (%ebx,%eax,2),%ebp 320 shrl $31,%eax 321 cmpl (%esp),%ecx 322 movl %eax,%ebx 323 movl %ebp,28(%esp,%ecx,4) 324 jl .L013sqr 325 movl (%esi,%ecx,4),%eax 326 movl %edx,%ebp 327 mull %edi 328 addl %ebp,%eax 329 movl 20(%esp),%edi 330 adcl $0,%edx 331 movl 16(%esp),%esi 332 leal (%ebx,%eax,2),%ebp 333 imull 32(%esp),%edi 334 shrl $31,%eax 335 movl %ebp,32(%esp,%ecx,4) 336 leal (%eax,%edx,2),%ebp 337 movl (%esi),%eax 338 shrl $31,%edx 339 movl %ebp,36(%esp,%ecx,4) 340 movl %edx,40(%esp,%ecx,4) 341 mull %edi 342 addl 32(%esp),%eax 343 movl %ecx,%ebx 344 adcl $0,%edx 345 movl 4(%esi),%eax 346 movl $1,%ecx 347 .align 16 348 .L0143rdmadd: 349 movl %edx,%ebp 350 mull %edi 351 addl 32(%esp,%ecx,4),%ebp 352 adcl $0,%edx 353 addl %eax,%ebp 354 movl 4(%esi,%ecx,4),%eax 355 adcl $0,%edx 356 movl %ebp,28(%esp,%ecx,4) 357 movl %edx,%ebp 358 mull %edi 359 addl 36(%esp,%ecx,4),%ebp 360 leal 2(%ecx),%ecx 361 adcl $0,%edx 362 addl %eax,%ebp 363 movl (%esi,%ecx,4),%eax 364 adcl $0,%edx 365 cmpl %ebx,%ecx 366 movl %ebp,24(%esp,%ecx,4) 367 jl .L0143rdmadd 368 movl %edx,%ebp 369 mull %edi 370 addl 32(%esp,%ebx,4),%ebp 371 adcl $0,%edx 372 addl %eax,%ebp 373 adcl $0,%edx 374 movl %ebp,28(%esp,%ebx,4) 375 movl 12(%esp),%ecx 376 xorl %eax,%eax 377 movl 8(%esp),%esi 378 addl 36(%esp,%ebx,4),%edx 379 adcl 40(%esp,%ebx,4),%eax 380 movl %edx,32(%esp,%ebx,4) 381 cmpl %ebx,%ecx 382 movl %eax,36(%esp,%ebx,4) 383 je .L008common_tail 384 movl 4(%esi,%ecx,4),%edi 385 leal 1(%ecx),%ecx 386 movl %edi,%eax 387 movl %ecx,12(%esp) 388 mull %edi 389 addl 32(%esp,%ecx,4),%eax 390 adcl $0,%edx 391 movl %eax,32(%esp,%ecx,4) 392 xorl %ebp,%ebp 393 cmpl %ebx,%ecx 394 leal 1(%ecx),%ecx 395 je .L015sqrlast 396 movl %edx,%ebx 397 shrl $1,%edx 398 andl $1,%ebx 399 .align 16 400 .L016sqradd: 401 movl (%esi,%ecx,4),%eax 402 movl %edx,%ebp 403 mull %edi 404 addl %ebp,%eax 405 leal (%eax,%eax,1),%ebp 406 adcl $0,%edx 407 shrl $31,%eax 408 addl 32(%esp,%ecx,4),%ebp 409 leal 1(%ecx),%ecx 410 adcl $0,%eax 411 addl %ebx,%ebp 412 adcl $0,%eax 413 cmpl (%esp),%ecx 414 movl %ebp,28(%esp,%ecx,4) 415 movl %eax,%ebx 416 jle .L016sqradd 417 movl %edx,%ebp 418 addl %edx,%edx 419 shrl $31,%ebp 420 addl %ebx,%edx 421 adcl $0,%ebp 422 .L015sqrlast: 423 movl 20(%esp),%edi 424 movl 16(%esp),%esi 425 imull 32(%esp),%edi 426 addl 32(%esp,%ecx,4),%edx 427 movl (%esi),%eax 428 adcl $0,%ebp 429 movl %edx,32(%esp,%ecx,4) 430 movl %ebp,36(%esp,%ecx,4) 431 mull %edi 432 addl 32(%esp),%eax 433 leal -1(%ecx),%ebx 434 adcl $0,%edx 435 movl $1,%ecx 436 movl 4(%esi),%eax 437 jmp .L0143rdmadd 438 .align 16 439 .L008common_tail: 440 movl 16(%esp),%ebp 441 movl 4(%esp),%edi 442 leal 32(%esp),%esi 443 movl (%esi),%eax 444 movl %ebx,%ecx 445 xorl %edx,%edx 446 .align 16 447 .L017sub: 448 sbbl (%ebp,%edx,4),%eax 449 movl %eax,(%edi,%edx,4) 450 decl %ecx 451 movl 4(%esi,%edx,4),%eax 452 leal 1(%edx),%edx 453 jge .L017sub 454 sbbl $0,%eax 455 movl $-1,%edx 456 xorl %eax,%edx 457 jmp .L018copy 458 .align 16 459 .L018copy: 460 movl 32(%esp,%ebx,4),%esi 461 movl (%edi,%ebx,4),%ebp 462 movl %ecx,32(%esp,%ebx,4) 463 andl %eax,%esi 464 andl %edx,%ebp 465 orl %esi,%ebp 466 movl %ebp,(%edi,%ebx,4) 467 decl %ebx 468 jge .L018copy 469 movl 24(%esp),%esp 470 movl $1,%eax 471 .L000just_leave: 472 popl %edi 473 popl %esi 474 popl %ebx 475 popl %ebp 476 ret 477 .size bn_mul_mont,.-.L_bn_mul_mont_begin 478 .byte 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105 479 .byte 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56 480 .byte 54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121 481 .byte 32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46 482 .byte 111,114,103,62,0 483 #endif 484
