38 #ifndef MBEDTLS_BN_MUL_H 39 #define MBEDTLS_BN_MUL_H 43 #if defined(MBEDTLS_HAVE_ASM) 50 #if defined(__GNUC__) && \ 51 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 ) 54 #define MULADDC_INIT \ 56 "movl %%ebx, %0 \n\t" \ 57 "movl %5, %%esi \n\t" \ 58 "movl %6, %%edi \n\t" \ 59 "movl %7, %%ecx \n\t" \ 62 #define MULADDC_CORE \ 65 "addl %%ecx, %%eax \n\t" \ 66 "adcl $0, %%edx \n\t" \ 67 "addl (%%edi), %%eax \n\t" \ 68 "adcl $0, %%edx \n\t" \ 69 "movl %%edx, %%ecx \n\t" \ 72 #if defined(MBEDTLS_HAVE_SSE2) 74 #define MULADDC_HUIT \ 75 "movd %%ecx, %%mm1 \n\t" \ 76 "movd %%ebx, %%mm0 \n\t" \ 77 "movd (%%edi), %%mm3 \n\t" \ 78 "paddq %%mm3, %%mm1 \n\t" \ 79 "movd (%%esi), %%mm2 \n\t" \ 80 "pmuludq %%mm0, %%mm2 \n\t" \ 81 "movd 4(%%esi), %%mm4 \n\t" \ 82 "pmuludq %%mm0, %%mm4 \n\t" \ 83 "movd 8(%%esi), %%mm6 \n\t" \ 84 "pmuludq %%mm0, %%mm6 \n\t" \ 85 "movd 12(%%esi), %%mm7 \n\t" \ 86 "pmuludq %%mm0, %%mm7 \n\t" \ 87 "paddq %%mm2, %%mm1 \n\t" \ 88 "movd 4(%%edi), %%mm3 \n\t" \ 89 "paddq %%mm4, %%mm3 \n\t" \ 90 "movd 8(%%edi), %%mm5 \n\t" \ 91 "paddq %%mm6, %%mm5 \n\t" \ 92 "movd 12(%%edi), %%mm4 \n\t" \ 93 "paddq %%mm4, %%mm7 \n\t" \ 94 "movd %%mm1, (%%edi) \n\t" \ 95 "movd 16(%%esi), %%mm2 \n\t" \ 96 "pmuludq %%mm0, %%mm2 \n\t" \ 97 "psrlq $32, %%mm1 \n\t" \ 98 "movd 20(%%esi), %%mm4 \n\t" \ 99 "pmuludq %%mm0, %%mm4 \n\t" \ 100 "paddq %%mm3, %%mm1 \n\t" \ 101 "movd 24(%%esi), %%mm6 \n\t" \ 102 "pmuludq %%mm0, %%mm6 \n\t" \ 103 "movd %%mm1, 4(%%edi) \n\t" \ 104 "psrlq $32, %%mm1 \n\t" \ 105 "movd 28(%%esi), %%mm3 \n\t" \ 106 "pmuludq %%mm0, %%mm3 \n\t" \ 107 "paddq %%mm5, %%mm1 \n\t" \ 108 "movd 16(%%edi), %%mm5 \n\t" \ 109 "paddq %%mm5, %%mm2 \n\t" \ 110 "movd %%mm1, 8(%%edi) \n\t" \ 111 "psrlq $32, %%mm1 \n\t" \ 112 "paddq %%mm7, %%mm1 \n\t" \ 113 "movd 20(%%edi), %%mm5 \n\t" \ 114 "paddq %%mm5, %%mm4 \n\t" \ 115 "movd %%mm1, 12(%%edi) \n\t" \ 116 "psrlq $32, %%mm1 \n\t" \ 117 "paddq %%mm2, %%mm1 \n\t" \ 118 "movd 24(%%edi), %%mm5 \n\t" \ 119 "paddq %%mm5, %%mm6 \n\t" \ 120 "movd %%mm1, 16(%%edi) \n\t" \ 121 "psrlq $32, %%mm1 \n\t" \ 122 "paddq %%mm4, %%mm1 \n\t" \ 123 "movd 28(%%edi), %%mm5 \n\t" \ 124 "paddq %%mm5, %%mm3 \n\t" \ 125 "movd %%mm1, 20(%%edi) \n\t" \ 126 "psrlq $32, %%mm1 \n\t" \ 127 "paddq %%mm6, %%mm1 \n\t" \ 128 "movd %%mm1, 24(%%edi) \n\t" \ 129 "psrlq $32, %%mm1 \n\t" \ 130 "paddq %%mm3, %%mm1 \n\t" \ 131 "movd %%mm1, 28(%%edi) \n\t" \ 132 "addl $32, %%edi \n\t" \ 133 "addl $32, %%esi \n\t" \ 134 "psrlq $32, %%mm1 \n\t" \ 135 "movd %%mm1, %%ecx \n\t" 137 #define MULADDC_STOP \ 139 "movl %4, %%ebx \n\t" \ 140 "movl %%ecx, %1 \n\t" \ 141 "movl %%edi, %2 \n\t" \ 142 "movl %%esi, %3 \n\t" \ 143 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ 144 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ 145 : "eax", "ecx", "edx", "esi", "edi" \ 150 #define MULADDC_STOP \ 151 "movl %4, %%ebx \n\t" \ 152 "movl %%ecx, %1 \n\t" \ 153 "movl %%edi, %2 \n\t" \ 154 "movl %%esi, %3 \n\t" \ 155 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ 156 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ 157 : "eax", "ecx", "edx", "esi", "edi" \ 162 #if defined(__amd64__) || defined (__x86_64__) 164 #define MULADDC_INIT \ 166 "xorq %%r8, %%r8 \n\t" 168 #define MULADDC_CORE \ 169 "movq (%%rsi), %%rax \n\t" \ 171 "addq $8, %%rsi \n\t" \ 172 "addq %%rcx, %%rax \n\t" \ 173 "movq %%r8, %%rcx \n\t" \ 174 "adcq $0, %%rdx \n\t" \ 176 "addq %%rax, (%%rdi) \n\t" \ 177 "adcq %%rdx, %%rcx \n\t" \ 178 "addq $8, %%rdi \n\t" 180 #define MULADDC_STOP \ 181 : "+c" (c), "+D" (d), "+S" (s) \ 183 : "rax", "rdx", "r8" \ 188 #if defined(__mc68020__) || defined(__mcpu32__) 190 #define MULADDC_INIT \ 192 "movl %3, %%a2 \n\t" \ 193 "movl %4, %%a3 \n\t" \ 194 "movl %5, %%d3 \n\t" \ 195 "movl %6, %%d2 \n\t" \ 196 "moveq #0, %%d0 \n\t" 198 #define MULADDC_CORE \ 199 "movel %%a2@+, %%d1 \n\t" \ 200 "mulul %%d2, %%d4:%%d1 \n\t" \ 201 "addl %%d3, %%d1 \n\t" \ 202 "addxl %%d0, %%d4 \n\t" \ 203 "moveq #0, %%d3 \n\t" \ 204 "addl %%d1, %%a3@+ \n\t" \ 205 "addxl %%d4, %%d3 \n\t" 207 #define MULADDC_STOP \ 208 "movl %%d3, %0 \n\t" \ 209 "movl %%a3, %1 \n\t" \ 210 "movl %%a2, %2 \n\t" \ 211 : "=m" (c), "=m" (d), "=m" (s) \ 212 : "m" (s), "m" (d), "m" (c), "m" (b) \ 213 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \ 216 #define MULADDC_HUIT \ 217 "movel %%a2@+, %%d1 \n\t" \ 218 "mulul %%d2, %%d4:%%d1 \n\t" \ 219 "addxl %%d3, %%d1 \n\t" \ 220 "addxl %%d0, %%d4 \n\t" \ 221 "addl %%d1, %%a3@+ \n\t" \ 222 "movel %%a2@+, %%d1 \n\t" \ 223 "mulul %%d2, %%d3:%%d1 \n\t" \ 224 "addxl %%d4, %%d1 \n\t" \ 225 "addxl %%d0, %%d3 \n\t" \ 226 "addl %%d1, %%a3@+ \n\t" \ 227 "movel %%a2@+, %%d1 \n\t" \ 228 "mulul %%d2, %%d4:%%d1 \n\t" \ 229 "addxl %%d3, %%d1 \n\t" \ 230 "addxl %%d0, %%d4 \n\t" \ 231 "addl %%d1, %%a3@+ \n\t" \ 232 "movel %%a2@+, %%d1 \n\t" \ 233 "mulul %%d2, %%d3:%%d1 \n\t" \ 234 "addxl %%d4, %%d1 \n\t" \ 235 "addxl %%d0, %%d3 \n\t" \ 236 "addl %%d1, %%a3@+ \n\t" \ 237 "movel %%a2@+, %%d1 \n\t" \ 238 "mulul %%d2, %%d4:%%d1 \n\t" \ 239 "addxl %%d3, %%d1 \n\t" \ 240 "addxl %%d0, %%d4 \n\t" \ 241 "addl %%d1, %%a3@+ \n\t" \ 242 "movel %%a2@+, %%d1 \n\t" \ 243 "mulul %%d2, %%d3:%%d1 \n\t" \ 244 "addxl %%d4, %%d1 \n\t" \ 245 "addxl %%d0, %%d3 \n\t" \ 246 "addl %%d1, %%a3@+ \n\t" \ 247 "movel %%a2@+, %%d1 \n\t" \ 248 "mulul %%d2, %%d4:%%d1 \n\t" \ 249 "addxl %%d3, %%d1 \n\t" \ 250 "addxl %%d0, %%d4 \n\t" \ 251 "addl %%d1, %%a3@+ \n\t" \ 252 "movel %%a2@+, %%d1 \n\t" \ 253 "mulul %%d2, %%d3:%%d1 \n\t" \ 254 "addxl %%d4, %%d1 \n\t" \ 255 "addxl %%d0, %%d3 \n\t" \ 256 "addl %%d1, %%a3@+ \n\t" \ 257 "addxl %%d0, %%d3 \n\t" 261 #if defined(__powerpc64__) || defined(__ppc64__) 263 #if defined(__MACH__) && defined(__APPLE__) 265 #define MULADDC_INIT \ 271 "addi r3, r3, -8 \n\t" \ 272 "addi r4, r4, -8 \n\t" \ 273 "addic r5, r5, 0 \n\t" 275 #define MULADDC_CORE \ 276 "ldu r7, 8(r3) \n\t" \ 277 "mulld r8, r7, r6 \n\t" \ 278 "mulhdu r9, r7, r6 \n\t" \ 279 "adde r8, r8, r5 \n\t" \ 280 "ld r7, 8(r4) \n\t" \ 281 "addze r5, r9 \n\t" \ 282 "addc r8, r8, r7 \n\t" \ 283 "stdu r8, 8(r4) \n\t" 285 #define MULADDC_STOP \ 286 "addze r5, r5 \n\t" \ 287 "addi r4, r4, 8 \n\t" \ 288 "addi r3, r3, 8 \n\t" \ 292 : "=m" (c), "=m" (d), "=m" (s) \ 293 : "m" (s), "m" (d), "m" (c), "m" (b) \ 294 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 300 #define MULADDC_INIT \ 306 "addi %%r3, %%r3, -8 \n\t" \ 307 "addi %%r4, %%r4, -8 \n\t" \ 308 "addic %%r5, %%r5, 0 \n\t" 310 #define MULADDC_CORE \ 311 "ldu %%r7, 8(%%r3) \n\t" \ 312 "mulld %%r8, %%r7, %%r6 \n\t" \ 313 "mulhdu %%r9, %%r7, %%r6 \n\t" \ 314 "adde %%r8, %%r8, %%r5 \n\t" \ 315 "ld %%r7, 8(%%r4) \n\t" \ 316 "addze %%r5, %%r9 \n\t" \ 317 "addc %%r8, %%r8, %%r7 \n\t" \ 318 "stdu %%r8, 8(%%r4) \n\t" 320 #define MULADDC_STOP \ 321 "addze %%r5, %%r5 \n\t" \ 322 "addi %%r4, %%r4, 8 \n\t" \ 323 "addi %%r3, %%r3, 8 \n\t" \ 324 "std %%r5, %0 \n\t" \ 325 "std %%r4, %1 \n\t" \ 326 "std %%r3, %2 \n\t" \ 327 : "=m" (c), "=m" (d), "=m" (s) \ 328 : "m" (s), "m" (d), "m" (c), "m" (b) \ 329 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 334 #elif defined(__powerpc__) || defined(__ppc__) 336 #if defined(__MACH__) && defined(__APPLE__) 338 #define MULADDC_INIT \ 344 "addi r3, r3, -4 \n\t" \ 345 "addi r4, r4, -4 \n\t" \ 346 "addic r5, r5, 0 \n\t" 348 #define MULADDC_CORE \ 349 "lwzu r7, 4(r3) \n\t" \ 350 "mullw r8, r7, r6 \n\t" \ 351 "mulhwu r9, r7, r6 \n\t" \ 352 "adde r8, r8, r5 \n\t" \ 353 "lwz r7, 4(r4) \n\t" \ 354 "addze r5, r9 \n\t" \ 355 "addc r8, r8, r7 \n\t" \ 356 "stwu r8, 4(r4) \n\t" 358 #define MULADDC_STOP \ 359 "addze r5, r5 \n\t" \ 360 "addi r4, r4, 4 \n\t" \ 361 "addi r3, r3, 4 \n\t" \ 365 : "=m" (c), "=m" (d), "=m" (s) \ 366 : "m" (s), "m" (d), "m" (c), "m" (b) \ 367 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 372 #define MULADDC_INIT \ 374 "lwz %%r3, %3 \n\t" \ 375 "lwz %%r4, %4 \n\t" \ 376 "lwz %%r5, %5 \n\t" \ 377 "lwz %%r6, %6 \n\t" \ 378 "addi %%r3, %%r3, -4 \n\t" \ 379 "addi %%r4, %%r4, -4 \n\t" \ 380 "addic %%r5, %%r5, 0 \n\t" 382 #define MULADDC_CORE \ 383 "lwzu %%r7, 4(%%r3) \n\t" \ 384 "mullw %%r8, %%r7, %%r6 \n\t" \ 385 "mulhwu %%r9, %%r7, %%r6 \n\t" \ 386 "adde %%r8, %%r8, %%r5 \n\t" \ 387 "lwz %%r7, 4(%%r4) \n\t" \ 388 "addze %%r5, %%r9 \n\t" \ 389 "addc %%r8, %%r8, %%r7 \n\t" \ 390 "stwu %%r8, 4(%%r4) \n\t" 392 #define MULADDC_STOP \ 393 "addze %%r5, %%r5 \n\t" \ 394 "addi %%r4, %%r4, 4 \n\t" \ 395 "addi %%r3, %%r3, 4 \n\t" \ 396 "stw %%r5, %0 \n\t" \ 397 "stw %%r4, %1 \n\t" \ 398 "stw %%r3, %2 \n\t" \ 399 : "=m" (c), "=m" (d), "=m" (s) \ 400 : "m" (s), "m" (d), "m" (c), "m" (b) \ 401 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 412 #if 0 && defined(__sparc__) 413 #if defined(__sparc64__) 415 #define MULADDC_INIT \ 417 "ldx %3, %%o0 \n\t" \ 418 "ldx %4, %%o1 \n\t" \ 422 #define MULADDC_CORE \ 423 "ld [%%o0], %%o4 \n\t" \ 425 "ld [%%o1], %%o5 \n\t" \ 426 "umul %%o3, %%o4, %%o4 \n\t" \ 427 "addcc %%o4, %%o2, %%o4 \n\t" \ 428 "rd %%y, %%g1 \n\t" \ 429 "addx %%g1, 0, %%g1 \n\t" \ 430 "addcc %%o4, %%o5, %%o4 \n\t" \ 431 "st %%o4, [%%o1] \n\t" \ 432 "addx %%g1, 0, %%o2 \n\t" \ 435 #define MULADDC_STOP \ 437 "stx %%o1, %1 \n\t" \ 438 "stx %%o0, %2 \n\t" \ 439 : "=m" (c), "=m" (d), "=m" (s) \ 440 : "m" (s), "m" (d), "m" (c), "m" (b) \ 441 : "g1", "o0", "o1", "o2", "o3", "o4", \ 447 #define MULADDC_INIT \ 454 #define MULADDC_CORE \ 455 "ld [%%o0], %%o4 \n\t" \ 457 "ld [%%o1], %%o5 \n\t" \ 458 "umul %%o3, %%o4, %%o4 \n\t" \ 459 "addcc %%o4, %%o2, %%o4 \n\t" \ 460 "rd %%y, %%g1 \n\t" \ 461 "addx %%g1, 0, %%g1 \n\t" \ 462 "addcc %%o4, %%o5, %%o4 \n\t" \ 463 "st %%o4, [%%o1] \n\t" \ 464 "addx %%g1, 0, %%o2 \n\t" \ 467 #define MULADDC_STOP \ 471 : "=m" (c), "=m" (d), "=m" (s) \ 472 : "m" (s), "m" (d), "m" (c), "m" (b) \ 473 : "g1", "o0", "o1", "o2", "o3", "o4", \ 480 #if defined(__microblaze__) || defined(microblaze) 482 #define MULADDC_INIT \ 488 "andi r7, r6, 0xffff \n\t" \ 489 "bsrli r6, r6, 16 \n\t" 491 #define MULADDC_CORE \ 492 "lhui r8, r3, 0 \n\t" \ 493 "addi r3, r3, 2 \n\t" \ 494 "lhui r9, r3, 0 \n\t" \ 495 "addi r3, r3, 2 \n\t" \ 496 "mul r10, r9, r6 \n\t" \ 497 "mul r11, r8, r7 \n\t" \ 498 "mul r12, r9, r7 \n\t" \ 499 "mul r13, r8, r6 \n\t" \ 500 "bsrli r8, r10, 16 \n\t" \ 501 "bsrli r9, r11, 16 \n\t" \ 502 "add r13, r13, r8 \n\t" \ 503 "add r13, r13, r9 \n\t" \ 504 "bslli r10, r10, 16 \n\t" \ 505 "bslli r11, r11, 16 \n\t" \ 506 "add r12, r12, r10 \n\t" \ 507 "addc r13, r13, r0 \n\t" \ 508 "add r12, r12, r11 \n\t" \ 509 "addc r13, r13, r0 \n\t" \ 510 "lwi r10, r4, 0 \n\t" \ 511 "add r12, r12, r10 \n\t" \ 512 "addc r13, r13, r0 \n\t" \ 513 "add r12, r12, r5 \n\t" \ 514 "addc r5, r13, r0 \n\t" \ 515 "swi r12, r4, 0 \n\t" \ 516 "addi r4, r4, 4 \n\t" 518 #define MULADDC_STOP \ 522 : "=m" (c), "=m" (d), "=m" (s) \ 523 : "m" (s), "m" (d), "m" (c), "m" (b) \ 524 : "r3", "r4" "r5", "r6", "r7", "r8", \ 525 "r9", "r10", "r11", "r12", "r13" \ 530 #if defined(__tricore__) 532 #define MULADDC_INIT \ 534 "ld.a %%a2, %3 \n\t" \ 535 "ld.a %%a3, %4 \n\t" \ 536 "ld.w %%d4, %5 \n\t" \ 537 "ld.w %%d1, %6 \n\t" \ 538 "xor %%d5, %%d5 \n\t" 540 #define MULADDC_CORE \ 541 "ld.w %%d0, [%%a2+] \n\t" \ 542 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \ 543 "ld.w %%d0, [%%a3] \n\t" \ 544 "addx %%d2, %%d2, %%d0 \n\t" \ 545 "addc %%d3, %%d3, 0 \n\t" \ 546 "mov %%d4, %%d3 \n\t" \ 547 "st.w [%%a3+], %%d2 \n\t" 549 #define MULADDC_STOP \ 550 "st.w %0, %%d4 \n\t" \ 551 "st.a %1, %%a3 \n\t" \ 552 "st.a %2, %%a2 \n\t" \ 553 : "=m" (c), "=m" (d), "=m" (s) \ 554 : "m" (s), "m" (d), "m" (c), "m" (b) \ 555 : "d0", "d1", "e2", "d4", "a2", "a3" \ 572 #if defined(__GNUC__) && !defined(__OPTIMIZE__) 573 #define MULADDC_CANNOT_USE_R7 576 #if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7) 578 #if defined(__thumb__) && !defined(__thumb2__) 580 #define MULADDC_INIT \ 586 "lsr r7, r3, #16 \n\t" \ 588 "lsl r7, r3, #16 \n\t" \ 589 "lsr r7, r7, #16 \n\t" \ 592 #define MULADDC_CORE \ 593 "ldmia r0!, {r6} \n\t" \ 594 "lsr r7, r6, #16 \n\t" \ 595 "lsl r6, r6, #16 \n\t" \ 596 "lsr r6, r6, #16 \n\t" \ 605 "lsr r3, r6, #16 \n\t" \ 606 "add r5, r5, r3 \n\t" \ 607 "lsr r3, r7, #16 \n\t" \ 608 "add r5, r5, r3 \n\t" \ 609 "add r4, r4, r2 \n\t" \ 612 "lsl r3, r6, #16 \n\t" \ 613 "add r4, r4, r3 \n\t" \ 615 "lsl r3, r7, #16 \n\t" \ 616 "add r4, r4, r3 \n\t" \ 618 "ldr r3, [r1] \n\t" \ 619 "add r4, r4, r3 \n\t" \ 621 "stmia r1!, {r4} \n\t" 623 #define MULADDC_STOP \ 627 : "=m" (c), "=m" (d), "=m" (s) \ 628 : "m" (s), "m" (d), "m" (c), "m" (b) \ 629 : "r0", "r1", "r2", "r3", "r4", "r5", \ 630 "r6", "r7", "r8", "r9", "cc" \ 635 #define MULADDC_INIT \ 642 #define MULADDC_CORE \ 643 "ldr r4, [r0], #4 \n\t" \ 645 "ldr r6, [r1] \n\t" \ 646 "umlal r2, r5, r3, r4 \n\t" \ 647 "adds r7, r6, r2 \n\t" \ 648 "adc r2, r5, #0 \n\t" \ 649 "str r7, [r1], #4 \n\t" 651 #define MULADDC_STOP \ 655 : "=m" (c), "=m" (d), "=m" (s) \ 656 : "m" (s), "m" (d), "m" (c), "m" (b) \ 657 : "r0", "r1", "r2", "r3", "r4", "r5", \ 665 #if defined(__alpha__) 667 #define MULADDC_INIT \ 674 #define MULADDC_CORE \ 675 "ldq $6, 0($1) \n\t" \ 676 "addq $1, 8, $1 \n\t" \ 677 "mulq $6, $4, $7 \n\t" \ 678 "umulh $6, $4, $6 \n\t" \ 679 "addq $7, $3, $7 \n\t" \ 680 "cmpult $7, $3, $3 \n\t" \ 681 "ldq $5, 0($2) \n\t" \ 682 "addq $7, $5, $7 \n\t" \ 683 "cmpult $7, $5, $5 \n\t" \ 684 "stq $7, 0($2) \n\t" \ 685 "addq $2, 8, $2 \n\t" \ 686 "addq $6, $3, $3 \n\t" \ 687 "addq $5, $3, $3 \n\t" 689 #define MULADDC_STOP \ 693 : "=m" (c), "=m" (d), "=m" (s) \ 694 : "m" (s), "m" (d), "m" (c), "m" (b) \ 695 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \ 699 #if defined(__mips__) && !defined(__mips64) 701 #define MULADDC_INIT \ 708 #define MULADDC_CORE \ 709 "lw $14, 0($10) \n\t" \ 710 "multu $13, $14 \n\t" \ 711 "addi $10, $10, 4 \n\t" \ 714 "addu $14, $12, $14 \n\t" \ 715 "lw $15, 0($11) \n\t" \ 716 "sltu $12, $14, $12 \n\t" \ 717 "addu $15, $14, $15 \n\t" \ 718 "sltu $14, $15, $14 \n\t" \ 719 "addu $12, $12, $9 \n\t" \ 720 "sw $15, 0($11) \n\t" \ 721 "addu $12, $12, $14 \n\t" \ 722 "addi $11, $11, 4 \n\t" 724 #define MULADDC_STOP \ 728 : "=m" (c), "=m" (d), "=m" (s) \ 729 : "m" (s), "m" (d), "m" (c), "m" (b) \ 730 : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \ 736 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__) 738 #define MULADDC_INIT \ 744 #define MULADDC_CORE \ 749 __asm add eax, [edi] \ 754 #if defined(MBEDTLS_HAVE_SSE2) 756 #define EMIT __asm _emit 758 #define MULADDC_HUIT \ 759 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \ 760 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \ 761 EMIT 0x0F EMIT 0x6E EMIT 0x1F \ 762 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 763 EMIT 0x0F EMIT 0x6E EMIT 0x16 \ 764 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ 765 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \ 766 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ 767 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \ 768 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ 769 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \ 770 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \ 771 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ 772 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \ 773 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \ 774 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \ 775 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \ 776 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \ 777 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \ 778 EMIT 0x0F EMIT 0x7E EMIT 0x0F \ 779 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \ 780 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ 781 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 782 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \ 783 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ 784 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 785 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \ 786 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ 787 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \ 788 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 789 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \ 790 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \ 791 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \ 792 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \ 793 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \ 794 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \ 795 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 796 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \ 797 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \ 798 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \ 799 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \ 800 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 801 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ 802 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \ 803 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \ 804 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \ 805 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 806 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \ 807 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \ 808 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \ 809 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \ 810 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 811 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \ 812 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \ 813 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 814 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 815 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \ 816 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \ 817 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \ 818 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 819 EMIT 0x0F EMIT 0x7E EMIT 0xC9 821 #define MULADDC_STOP \ 822 EMIT 0x0F EMIT 0x77 \ 829 #define MULADDC_STOP \ 839 #if !defined(MULADDC_CORE) 840 #if defined(MBEDTLS_HAVE_UDBL) 842 #define MULADDC_INIT \ 845 mbedtls_mpi_uint r0, r1; 847 #define MULADDC_CORE \ 848 r = *(s++) * (mbedtls_t_udbl) b; \ 849 r0 = (mbedtls_mpi_uint) r; \ 850 r1 = (mbedtls_mpi_uint)( r >> biL ); \ 851 r0 += c; r1 += (r0 < c); \ 852 r0 += *d; r1 += (r0 < *d); \ 855 #define MULADDC_STOP \ 859 #define MULADDC_INIT \ 861 mbedtls_mpi_uint s0, s1, b0, b1; \ 862 mbedtls_mpi_uint r0, r1, rx, ry; \ 863 b0 = ( b << biH ) >> biH; \ 866 #define MULADDC_CORE \ 867 s0 = ( *s << biH ) >> biH; \ 868 s1 = ( *s >> biH ); s++; \ 869 rx = s0 * b1; r0 = s0 * b0; \ 870 ry = s1 * b0; r1 = s1 * b1; \ 871 r1 += ( rx >> biH ); \ 872 r1 += ( ry >> biH ); \ 873 rx <<= biH; ry <<= biH; \ 874 r0 += rx; r1 += (r0 < rx); \ 875 r0 += ry; r1 += (r0 < ry); \ 876 r0 += c; r1 += (r0 < c); \ 877 r0 += *d; r1 += (r0 < *d); \ 880 #define MULADDC_STOP \ Multi-precision integer library.