1 /** 2 * \file bn_mul.h 3 * 4 * \brief Multi-precision integer library 5 */ 6 /* 7 * Copyright The Mbed TLS Contributors 8 * SPDX-License-Identifier: Apache-2.0 9 * 10 * Licensed under the Apache License, Version 2.0 (the "License"); you may 11 * not use this file except in compliance with the License. 12 * You may obtain a copy of the License at 13 * 14 * http://www.apache.org/licenses/LICENSE-2.0 15 * 16 * Unless required by applicable law or agreed to in writing, software 17 * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT 18 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 19 * See the License for the specific language governing permissions and 20 * limitations under the License. 21 */ 22 /* 23 * Multiply source vector [s] with b, add result 24 * to destination vector [d] and set carry c. 25 * 26 * Currently supports: 27 * 28 * . IA-32 (386+) . AMD64 / EM64T 29 * . IA-32 (SSE2) . Motorola 68000 30 * . PowerPC, 32-bit . MicroBlaze 31 * . PowerPC, 64-bit . TriCore 32 * . SPARC v8 . ARM v3+ 33 * . Alpha . MIPS32 34 * . C, longlong . C, generic 35 */ 36 #ifndef MBEDTLS_BN_MUL_H 37 #define MBEDTLS_BN_MUL_H 38 39 #include "mbedtls/build_info.h" 40 41 #include "mbedtls/bignum.h" 42 43 44 /* 45 * Conversion macros for embedded constants: 46 * build lists of mbedtls_mpi_uint's from lists of unsigned char's grouped by 8, 4 or 2 47 */ 48 #if defined(MBEDTLS_HAVE_INT32) 49 50 #define MBEDTLS_BYTES_TO_T_UINT_4( a, b, c, d ) \ 51 ( (mbedtls_mpi_uint) (a) << 0 ) | \ 52 ( (mbedtls_mpi_uint) (b) << 8 ) | \ 53 ( (mbedtls_mpi_uint) (c) << 16 ) | \ 54 ( (mbedtls_mpi_uint) (d) << 24 ) 55 56 #define MBEDTLS_BYTES_TO_T_UINT_2( a, b ) \ 57 MBEDTLS_BYTES_TO_T_UINT_4( a, b, 0, 0 ) 58 59 #define MBEDTLS_BYTES_TO_T_UINT_8( a, b, c, d, e, f, g, h ) \ 60 MBEDTLS_BYTES_TO_T_UINT_4( a, b, c, d ), \ 61 MBEDTLS_BYTES_TO_T_UINT_4( e, f, g, h ) 62 63 #else /* 64-bits */ 64 65 #define MBEDTLS_BYTES_TO_T_UINT_8( a, b, c, d, e, f, g, h ) \ 66 ( (mbedtls_mpi_uint) (a) << 0 ) | \ 67 ( (mbedtls_mpi_uint) (b) << 8 ) | \ 68 ( (mbedtls_mpi_uint) (c) << 16 ) | \ 69 ( (mbedtls_mpi_uint) (d) << 24 ) | \ 70 ( (mbedtls_mpi_uint) (e) << 32 ) | \ 71 ( (mbedtls_mpi_uint) (f) << 40 ) | \ 72 ( (mbedtls_mpi_uint) (g) << 48 ) | \ 73 ( (mbedtls_mpi_uint) (h) << 56 ) 74 75 #define MBEDTLS_BYTES_TO_T_UINT_4( a, b, c, d ) \ 76 MBEDTLS_BYTES_TO_T_UINT_8( a, b, c, d, 0, 0, 0, 0 ) 77 78 #define MBEDTLS_BYTES_TO_T_UINT_2( a, b ) \ 79 MBEDTLS_BYTES_TO_T_UINT_8( a, b, 0, 0, 0, 0, 0, 0 ) 80 81 #endif /* bits in mbedtls_mpi_uint */ 82 83 #if defined(MBEDTLS_HAVE_ASM) 84 85 #ifndef asm 86 #define asm __asm 87 #endif 88 89 /* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */ 90 #if defined(__GNUC__) && \ 91 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 ) 92 93 /* 94 * Disable use of the i386 assembly code below if option -O0, to disable all 95 * compiler optimisations, is passed, detected with __OPTIMIZE__ 96 * This is done as the number of registers used in the assembly code doesn't 97 * work with the -O0 option. 98 */ 99 #if defined(__i386__) && defined(__OPTIMIZE__) 100 101 #define MULADDC_INIT \ 102 asm( \ 103 "movl %%ebx, %0 \n\t" \ 104 "movl %5, %%esi \n\t" \ 105 "movl %6, %%edi \n\t" \ 106 "movl %7, %%ecx \n\t" \ 107 "movl %8, %%ebx \n\t" 108 109 #define MULADDC_CORE \ 110 "lodsl \n\t" \ 111 "mull %%ebx \n\t" \ 112 "addl %%ecx, %%eax \n\t" \ 113 "adcl $0, %%edx \n\t" \ 114 "addl (%%edi), %%eax \n\t" \ 115 "adcl $0, %%edx \n\t" \ 116 "movl %%edx, %%ecx \n\t" \ 117 "stosl \n\t" 118 119 #if defined(MBEDTLS_HAVE_SSE2) 120 121 #define MULADDC_HUIT \ 122 "movd %%ecx, %%mm1 \n\t" \ 123 "movd %%ebx, %%mm0 \n\t" \ 124 "movd (%%edi), %%mm3 \n\t" \ 125 "paddq %%mm3, %%mm1 \n\t" \ 126 "movd (%%esi), %%mm2 \n\t" \ 127 "pmuludq %%mm0, %%mm2 \n\t" \ 128 "movd 4(%%esi), %%mm4 \n\t" \ 129 "pmuludq %%mm0, %%mm4 \n\t" \ 130 "movd 8(%%esi), %%mm6 \n\t" \ 131 "pmuludq %%mm0, %%mm6 \n\t" \ 132 "movd 12(%%esi), %%mm7 \n\t" \ 133 "pmuludq %%mm0, %%mm7 \n\t" \ 134 "paddq %%mm2, %%mm1 \n\t" \ 135 "movd 4(%%edi), %%mm3 \n\t" \ 136 "paddq %%mm4, %%mm3 \n\t" \ 137 "movd 8(%%edi), %%mm5 \n\t" \ 138 "paddq %%mm6, %%mm5 \n\t" \ 139 "movd 12(%%edi), %%mm4 \n\t" \ 140 "paddq %%mm4, %%mm7 \n\t" \ 141 "movd %%mm1, (%%edi) \n\t" \ 142 "movd 16(%%esi), %%mm2 \n\t" \ 143 "pmuludq %%mm0, %%mm2 \n\t" \ 144 "psrlq $32, %%mm1 \n\t" \ 145 "movd 20(%%esi), %%mm4 \n\t" \ 146 "pmuludq %%mm0, %%mm4 \n\t" \ 147 "paddq %%mm3, %%mm1 \n\t" \ 148 "movd 24(%%esi), %%mm6 \n\t" \ 149 "pmuludq %%mm0, %%mm6 \n\t" \ 150 "movd %%mm1, 4(%%edi) \n\t" \ 151 "psrlq $32, %%mm1 \n\t" \ 152 "movd 28(%%esi), %%mm3 \n\t" \ 153 "pmuludq %%mm0, %%mm3 \n\t" \ 154 "paddq %%mm5, %%mm1 \n\t" \ 155 "movd 16(%%edi), %%mm5 \n\t" \ 156 "paddq %%mm5, %%mm2 \n\t" \ 157 "movd %%mm1, 8(%%edi) \n\t" \ 158 "psrlq $32, %%mm1 \n\t" \ 159 "paddq %%mm7, %%mm1 \n\t" \ 160 "movd 20(%%edi), %%mm5 \n\t" \ 161 "paddq %%mm5, %%mm4 \n\t" \ 162 "movd %%mm1, 12(%%edi) \n\t" \ 163 "psrlq $32, %%mm1 \n\t" \ 164 "paddq %%mm2, %%mm1 \n\t" \ 165 "movd 24(%%edi), %%mm5 \n\t" \ 166 "paddq %%mm5, %%mm6 \n\t" \ 167 "movd %%mm1, 16(%%edi) \n\t" \ 168 "psrlq $32, %%mm1 \n\t" \ 169 "paddq %%mm4, %%mm1 \n\t" \ 170 "movd 28(%%edi), %%mm5 \n\t" \ 171 "paddq %%mm5, %%mm3 \n\t" \ 172 "movd %%mm1, 20(%%edi) \n\t" \ 173 "psrlq $32, %%mm1 \n\t" \ 174 "paddq %%mm6, %%mm1 \n\t" \ 175 "movd %%mm1, 24(%%edi) \n\t" \ 176 "psrlq $32, %%mm1 \n\t" \ 177 "paddq %%mm3, %%mm1 \n\t" \ 178 "movd %%mm1, 28(%%edi) \n\t" \ 179 "addl $32, %%edi \n\t" \ 180 "addl $32, %%esi \n\t" \ 181 "psrlq $32, %%mm1 \n\t" \ 182 "movd %%mm1, %%ecx \n\t" 183 184 #define MULADDC_STOP \ 185 "emms \n\t" \ 186 "movl %4, %%ebx \n\t" \ 187 "movl %%ecx, %1 \n\t" \ 188 "movl %%edi, %2 \n\t" \ 189 "movl %%esi, %3 \n\t" \ 190 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ 191 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ 192 : "eax", "ebx", "ecx", "edx", "esi", "edi" \ 193 ); 194 195 #else 196 197 #define MULADDC_STOP \ 198 "movl %4, %%ebx \n\t" \ 199 "movl %%ecx, %1 \n\t" \ 200 "movl %%edi, %2 \n\t" \ 201 "movl %%esi, %3 \n\t" \ 202 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ 203 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ 204 : "eax", "ebx", "ecx", "edx", "esi", "edi" \ 205 ); 206 #endif /* SSE2 */ 207 #endif /* i386 */ 208 209 #if defined(__amd64__) || defined (__x86_64__) 210 211 #define MULADDC_INIT \ 212 asm( \ 213 "xorq %%r8, %%r8\n" 214 215 #define MULADDC_CORE \ 216 "movq (%%rsi), %%rax\n" \ 217 "mulq %%rbx\n" \ 218 "addq $8, %%rsi\n" \ 219 "addq %%rcx, %%rax\n" \ 220 "movq %%r8, %%rcx\n" \ 221 "adcq $0, %%rdx\n" \ 222 "nop \n" \ 223 "addq %%rax, (%%rdi)\n" \ 224 "adcq %%rdx, %%rcx\n" \ 225 "addq $8, %%rdi\n" 226 227 #define MULADDC_STOP \ 228 : "+c" (c), "+D" (d), "+S" (s), "+m" (*(uint64_t (*)[16]) d) \ 229 : "b" (b), "m" (*(const uint64_t (*)[16]) s) \ 230 : "rax", "rdx", "r8" \ 231 ); 232 233 #endif /* AMD64 */ 234 235 #ifndef HW_MBEDTLS_ATLAS_U_BOOT 236 #if defined(__aarch64__) 237 238 #define MULADDC_INIT \ 239 asm( 240 241 #define MULADDC_CORE \ 242 "ldr x4, [%2], #8 \n\t" \ 243 "ldr x5, [%1] \n\t" \ 244 "mul x6, x4, %4 \n\t" \ 245 "umulh x7, x4, %4 \n\t" \ 246 "adds x5, x5, x6 \n\t" \ 247 "adc x7, x7, xzr \n\t" \ 248 "adds x5, x5, %0 \n\t" \ 249 "adc %0, x7, xzr \n\t" \ 250 "str x5, [%1], #8 \n\t" 251 252 #define MULADDC_STOP \ 253 : "+r" (c), "+r" (d), "+r" (s), "+m" (*(uint64_t (*)[16]) d) \ 254 : "r" (b), "m" (*(const uint64_t (*)[16]) s) \ 255 : "x4", "x5", "x6", "x7", "cc" \ 256 ); 257 258 #endif /* Aarch64 */ 259 #endif 260 261 #if defined(__mc68020__) || defined(__mcpu32__) 262 263 #define MULADDC_INIT \ 264 asm( \ 265 "movl %3, %%a2 \n\t" \ 266 "movl %4, %%a3 \n\t" \ 267 "movl %5, %%d3 \n\t" \ 268 "movl %6, %%d2 \n\t" \ 269 "moveq #0, %%d0 \n\t" 270 271 #define MULADDC_CORE \ 272 "movel %%a2@+, %%d1 \n\t" \ 273 "mulul %%d2, %%d4:%%d1 \n\t" \ 274 "addl %%d3, %%d1 \n\t" \ 275 "addxl %%d0, %%d4 \n\t" \ 276 "moveq #0, %%d3 \n\t" \ 277 "addl %%d1, %%a3@+ \n\t" \ 278 "addxl %%d4, %%d3 \n\t" 279 280 #define MULADDC_STOP \ 281 "movl %%d3, %0 \n\t" \ 282 "movl %%a3, %1 \n\t" \ 283 "movl %%a2, %2 \n\t" \ 284 : "=m" (c), "=m" (d), "=m" (s) \ 285 : "m" (s), "m" (d), "m" (c), "m" (b) \ 286 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \ 287 ); 288 289 #define MULADDC_HUIT \ 290 "movel %%a2@+, %%d1 \n\t" \ 291 "mulul %%d2, %%d4:%%d1 \n\t" \ 292 "addxl %%d3, %%d1 \n\t" \ 293 "addxl %%d0, %%d4 \n\t" \ 294 "addl %%d1, %%a3@+ \n\t" \ 295 "movel %%a2@+, %%d1 \n\t" \ 296 "mulul %%d2, %%d3:%%d1 \n\t" \ 297 "addxl %%d4, %%d1 \n\t" \ 298 "addxl %%d0, %%d3 \n\t" \ 299 "addl %%d1, %%a3@+ \n\t" \ 300 "movel %%a2@+, %%d1 \n\t" \ 301 "mulul %%d2, %%d4:%%d1 \n\t" \ 302 "addxl %%d3, %%d1 \n\t" \ 303 "addxl %%d0, %%d4 \n\t" \ 304 "addl %%d1, %%a3@+ \n\t" \ 305 "movel %%a2@+, %%d1 \n\t" \ 306 "mulul %%d2, %%d3:%%d1 \n\t" \ 307 "addxl %%d4, %%d1 \n\t" \ 308 "addxl %%d0, %%d3 \n\t" \ 309 "addl %%d1, %%a3@+ \n\t" \ 310 "movel %%a2@+, %%d1 \n\t" \ 311 "mulul %%d2, %%d4:%%d1 \n\t" \ 312 "addxl %%d3, %%d1 \n\t" \ 313 "addxl %%d0, %%d4 \n\t" \ 314 "addl %%d1, %%a3@+ \n\t" \ 315 "movel %%a2@+, %%d1 \n\t" \ 316 "mulul %%d2, %%d3:%%d1 \n\t" \ 317 "addxl %%d4, %%d1 \n\t" \ 318 "addxl %%d0, %%d3 \n\t" \ 319 "addl %%d1, %%a3@+ \n\t" \ 320 "movel %%a2@+, %%d1 \n\t" \ 321 "mulul %%d2, %%d4:%%d1 \n\t" \ 322 "addxl %%d3, %%d1 \n\t" \ 323 "addxl %%d0, %%d4 \n\t" \ 324 "addl %%d1, %%a3@+ \n\t" \ 325 "movel %%a2@+, %%d1 \n\t" \ 326 "mulul %%d2, %%d3:%%d1 \n\t" \ 327 "addxl %%d4, %%d1 \n\t" \ 328 "addxl %%d0, %%d3 \n\t" \ 329 "addl %%d1, %%a3@+ \n\t" \ 330 "addxl %%d0, %%d3 \n\t" 331 332 #endif /* MC68000 */ 333 334 #if defined(__powerpc64__) || defined(__ppc64__) 335 336 #if defined(__MACH__) && defined(__APPLE__) 337 338 #define MULADDC_INIT \ 339 asm( \ 340 "ld r3, %3 \n\t" \ 341 "ld r4, %4 \n\t" \ 342 "ld r5, %5 \n\t" \ 343 "ld r6, %6 \n\t" \ 344 "addi r3, r3, -8 \n\t" \ 345 "addi r4, r4, -8 \n\t" \ 346 "addic r5, r5, 0 \n\t" 347 348 #define MULADDC_CORE \ 349 "ldu r7, 8(r3) \n\t" \ 350 "mulld r8, r7, r6 \n\t" \ 351 "mulhdu r9, r7, r6 \n\t" \ 352 "adde r8, r8, r5 \n\t" \ 353 "ld r7, 8(r4) \n\t" \ 354 "addze r5, r9 \n\t" \ 355 "addc r8, r8, r7 \n\t" \ 356 "stdu r8, 8(r4) \n\t" 357 358 #define MULADDC_STOP \ 359 "addze r5, r5 \n\t" \ 360 "addi r4, r4, 8 \n\t" \ 361 "addi r3, r3, 8 \n\t" \ 362 "std r5, %0 \n\t" \ 363 "std r4, %1 \n\t" \ 364 "std r3, %2 \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" \ 368 ); 369 370 371 #else /* __MACH__ && __APPLE__ */ 372 373 #define MULADDC_INIT \ 374 asm( \ 375 "ld %%r3, %3 \n\t" \ 376 "ld %%r4, %4 \n\t" \ 377 "ld %%r5, %5 \n\t" \ 378 "ld %%r6, %6 \n\t" \ 379 "addi %%r3, %%r3, -8 \n\t" \ 380 "addi %%r4, %%r4, -8 \n\t" \ 381 "addic %%r5, %%r5, 0 \n\t" 382 383 #define MULADDC_CORE \ 384 "ldu %%r7, 8(%%r3) \n\t" \ 385 "mulld %%r8, %%r7, %%r6 \n\t" \ 386 "mulhdu %%r9, %%r7, %%r6 \n\t" \ 387 "adde %%r8, %%r8, %%r5 \n\t" \ 388 "ld %%r7, 8(%%r4) \n\t" \ 389 "addze %%r5, %%r9 \n\t" \ 390 "addc %%r8, %%r8, %%r7 \n\t" \ 391 "stdu %%r8, 8(%%r4) \n\t" 392 393 #define MULADDC_STOP \ 394 "addze %%r5, %%r5 \n\t" \ 395 "addi %%r4, %%r4, 8 \n\t" \ 396 "addi %%r3, %%r3, 8 \n\t" \ 397 "std %%r5, %0 \n\t" \ 398 "std %%r4, %1 \n\t" \ 399 "std %%r3, %2 \n\t" \ 400 : "=m" (c), "=m" (d), "=m" (s) \ 401 : "m" (s), "m" (d), "m" (c), "m" (b) \ 402 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 403 ); 404 405 #endif /* __MACH__ && __APPLE__ */ 406 407 #elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */ 408 409 #if defined(__MACH__) && defined(__APPLE__) 410 411 #define MULADDC_INIT \ 412 asm( \ 413 "lwz r3, %3 \n\t" \ 414 "lwz r4, %4 \n\t" \ 415 "lwz r5, %5 \n\t" \ 416 "lwz r6, %6 \n\t" \ 417 "addi r3, r3, -4 \n\t" \ 418 "addi r4, r4, -4 \n\t" \ 419 "addic r5, r5, 0 \n\t" 420 421 #define MULADDC_CORE \ 422 "lwzu r7, 4(r3) \n\t" \ 423 "mullw r8, r7, r6 \n\t" \ 424 "mulhwu r9, r7, r6 \n\t" \ 425 "adde r8, r8, r5 \n\t" \ 426 "lwz r7, 4(r4) \n\t" \ 427 "addze r5, r9 \n\t" \ 428 "addc r8, r8, r7 \n\t" \ 429 "stwu r8, 4(r4) \n\t" 430 431 #define MULADDC_STOP \ 432 "addze r5, r5 \n\t" \ 433 "addi r4, r4, 4 \n\t" \ 434 "addi r3, r3, 4 \n\t" \ 435 "stw r5, %0 \n\t" \ 436 "stw r4, %1 \n\t" \ 437 "stw r3, %2 \n\t" \ 438 : "=m" (c), "=m" (d), "=m" (s) \ 439 : "m" (s), "m" (d), "m" (c), "m" (b) \ 440 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 441 ); 442 443 #else /* __MACH__ && __APPLE__ */ 444 445 #define MULADDC_INIT \ 446 asm( \ 447 "lwz %%r3, %3 \n\t" \ 448 "lwz %%r4, %4 \n\t" \ 449 "lwz %%r5, %5 \n\t" \ 450 "lwz %%r6, %6 \n\t" \ 451 "addi %%r3, %%r3, -4 \n\t" \ 452 "addi %%r4, %%r4, -4 \n\t" \ 453 "addic %%r5, %%r5, 0 \n\t" 454 455 #define MULADDC_CORE \ 456 "lwzu %%r7, 4(%%r3) \n\t" \ 457 "mullw %%r8, %%r7, %%r6 \n\t" \ 458 "mulhwu %%r9, %%r7, %%r6 \n\t" \ 459 "adde %%r8, %%r8, %%r5 \n\t" \ 460 "lwz %%r7, 4(%%r4) \n\t" \ 461 "addze %%r5, %%r9 \n\t" \ 462 "addc %%r8, %%r8, %%r7 \n\t" \ 463 "stwu %%r8, 4(%%r4) \n\t" 464 465 #define MULADDC_STOP \ 466 "addze %%r5, %%r5 \n\t" \ 467 "addi %%r4, %%r4, 4 \n\t" \ 468 "addi %%r3, %%r3, 4 \n\t" \ 469 "stw %%r5, %0 \n\t" \ 470 "stw %%r4, %1 \n\t" \ 471 "stw %%r3, %2 \n\t" \ 472 : "=m" (c), "=m" (d), "=m" (s) \ 473 : "m" (s), "m" (d), "m" (c), "m" (b) \ 474 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 475 ); 476 477 #endif /* __MACH__ && __APPLE__ */ 478 479 #endif /* PPC32 */ 480 481 /* 482 * The Sparc(64) assembly is reported to be broken. 483 * Disable it for now, until we're able to fix it. 484 */ 485 #if 0 && defined(__sparc__) 486 #if defined(__sparc64__) 487 488 #define MULADDC_INIT \ 489 asm( \ 490 "ldx %3, %%o0 \n\t" \ 491 "ldx %4, %%o1 \n\t" \ 492 "ld %5, %%o2 \n\t" \ 493 "ld %6, %%o3 \n\t" 494 495 #define MULADDC_CORE \ 496 "ld [%%o0], %%o4 \n\t" \ 497 "inc 4, %%o0 \n\t" \ 498 "ld [%%o1], %%o5 \n\t" \ 499 "umul %%o3, %%o4, %%o4 \n\t" \ 500 "addcc %%o4, %%o2, %%o4 \n\t" \ 501 "rd %%y, %%g1 \n\t" \ 502 "addx %%g1, 0, %%g1 \n\t" \ 503 "addcc %%o4, %%o5, %%o4 \n\t" \ 504 "st %%o4, [%%o1] \n\t" \ 505 "addx %%g1, 0, %%o2 \n\t" \ 506 "inc 4, %%o1 \n\t" 507 508 #define MULADDC_STOP \ 509 "st %%o2, %0 \n\t" \ 510 "stx %%o1, %1 \n\t" \ 511 "stx %%o0, %2 \n\t" \ 512 : "=m" (c), "=m" (d), "=m" (s) \ 513 : "m" (s), "m" (d), "m" (c), "m" (b) \ 514 : "g1", "o0", "o1", "o2", "o3", "o4", \ 515 "o5" \ 516 ); 517 518 #else /* __sparc64__ */ 519 520 #define MULADDC_INIT \ 521 asm( \ 522 "ld %3, %%o0 \n\t" \ 523 "ld %4, %%o1 \n\t" \ 524 "ld %5, %%o2 \n\t" \ 525 "ld %6, %%o3 \n\t" 526 527 #define MULADDC_CORE \ 528 "ld [%%o0], %%o4 \n\t" \ 529 "inc 4, %%o0 \n\t" \ 530 "ld [%%o1], %%o5 \n\t" \ 531 "umul %%o3, %%o4, %%o4 \n\t" \ 532 "addcc %%o4, %%o2, %%o4 \n\t" \ 533 "rd %%y, %%g1 \n\t" \ 534 "addx %%g1, 0, %%g1 \n\t" \ 535 "addcc %%o4, %%o5, %%o4 \n\t" \ 536 "st %%o4, [%%o1] \n\t" \ 537 "addx %%g1, 0, %%o2 \n\t" \ 538 "inc 4, %%o1 \n\t" 539 540 #define MULADDC_STOP \ 541 "st %%o2, %0 \n\t" \ 542 "st %%o1, %1 \n\t" \ 543 "st %%o0, %2 \n\t" \ 544 : "=m" (c), "=m" (d), "=m" (s) \ 545 : "m" (s), "m" (d), "m" (c), "m" (b) \ 546 : "g1", "o0", "o1", "o2", "o3", "o4", \ 547 "o5" \ 548 ); 549 550 #endif /* __sparc64__ */ 551 #endif /* __sparc__ */ 552 553 #if defined(__microblaze__) || defined(microblaze) 554 555 #define MULADDC_INIT \ 556 asm( \ 557 "lwi r3, %3 \n\t" \ 558 "lwi r4, %4 \n\t" \ 559 "lwi r5, %5 \n\t" \ 560 "lwi r6, %6 \n\t" \ 561 "andi r7, r6, 0xffff \n\t" \ 562 "bsrli r6, r6, 16 \n\t" 563 564 #define MULADDC_CORE \ 565 "lhui r8, r3, 0 \n\t" \ 566 "addi r3, r3, 2 \n\t" \ 567 "lhui r9, r3, 0 \n\t" \ 568 "addi r3, r3, 2 \n\t" \ 569 "mul r10, r9, r6 \n\t" \ 570 "mul r11, r8, r7 \n\t" \ 571 "mul r12, r9, r7 \n\t" \ 572 "mul r13, r8, r6 \n\t" \ 573 "bsrli r8, r10, 16 \n\t" \ 574 "bsrli r9, r11, 16 \n\t" \ 575 "add r13, r13, r8 \n\t" \ 576 "add r13, r13, r9 \n\t" \ 577 "bslli r10, r10, 16 \n\t" \ 578 "bslli r11, r11, 16 \n\t" \ 579 "add r12, r12, r10 \n\t" \ 580 "addc r13, r13, r0 \n\t" \ 581 "add r12, r12, r11 \n\t" \ 582 "addc r13, r13, r0 \n\t" \ 583 "lwi r10, r4, 0 \n\t" \ 584 "add r12, r12, r10 \n\t" \ 585 "addc r13, r13, r0 \n\t" \ 586 "add r12, r12, r5 \n\t" \ 587 "addc r5, r13, r0 \n\t" \ 588 "swi r12, r4, 0 \n\t" \ 589 "addi r4, r4, 4 \n\t" 590 591 #define MULADDC_STOP \ 592 "swi r5, %0 \n\t" \ 593 "swi r4, %1 \n\t" \ 594 "swi r3, %2 \n\t" \ 595 : "=m" (c), "=m" (d), "=m" (s) \ 596 : "m" (s), "m" (d), "m" (c), "m" (b) \ 597 : "r3", "r4", "r5", "r6", "r7", "r8", \ 598 "r9", "r10", "r11", "r12", "r13" \ 599 ); 600 601 #endif /* MicroBlaze */ 602 603 #if defined(__tricore__) 604 605 #define MULADDC_INIT \ 606 asm( \ 607 "ld.a %%a2, %3 \n\t" \ 608 "ld.a %%a3, %4 \n\t" \ 609 "ld.w %%d4, %5 \n\t" \ 610 "ld.w %%d1, %6 \n\t" \ 611 "xor %%d5, %%d5 \n\t" 612 613 #define MULADDC_CORE \ 614 "ld.w %%d0, [%%a2+] \n\t" \ 615 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \ 616 "ld.w %%d0, [%%a3] \n\t" \ 617 "addx %%d2, %%d2, %%d0 \n\t" \ 618 "addc %%d3, %%d3, 0 \n\t" \ 619 "mov %%d4, %%d3 \n\t" \ 620 "st.w [%%a3+], %%d2 \n\t" 621 622 #define MULADDC_STOP \ 623 "st.w %0, %%d4 \n\t" \ 624 "st.a %1, %%a3 \n\t" \ 625 "st.a %2, %%a2 \n\t" \ 626 : "=m" (c), "=m" (d), "=m" (s) \ 627 : "m" (s), "m" (d), "m" (c), "m" (b) \ 628 : "d0", "d1", "e2", "d4", "a2", "a3" \ 629 ); 630 631 #endif /* TriCore */ 632 633 /* 634 * Note, gcc -O0 by default uses r7 for the frame pointer, so it complains about 635 * our use of r7 below, unless -fomit-frame-pointer is passed. 636 * 637 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with 638 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by 639 * clang and armcc5 under the same conditions). 640 * 641 * So, only use the optimized assembly below for optimized build, which avoids 642 * the build error and is pretty reasonable anyway. 643 */ 644 #if defined(__GNUC__) && !defined(__OPTIMIZE__) 645 #define MULADDC_CANNOT_USE_R7 646 #endif 647 648 #if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7) 649 650 #if defined(__thumb__) && !defined(__thumb2__) 651 652 #define MULADDC_INIT \ 653 asm( \ 654 "ldr r0, %3 \n\t" \ 655 "ldr r1, %4 \n\t" \ 656 "ldr r2, %5 \n\t" \ 657 "ldr r3, %6 \n\t" \ 658 "lsr r7, r3, #16 \n\t" \ 659 "mov r9, r7 \n\t" \ 660 "lsl r7, r3, #16 \n\t" \ 661 "lsr r7, r7, #16 \n\t" \ 662 "mov r8, r7 \n\t" 663 664 #define MULADDC_CORE \ 665 "ldmia r0!, {r6} \n\t" \ 666 "lsr r7, r6, #16 \n\t" \ 667 "lsl r6, r6, #16 \n\t" \ 668 "lsr r6, r6, #16 \n\t" \ 669 "mov r4, r8 \n\t" \ 670 "mul r4, r6 \n\t" \ 671 "mov r3, r9 \n\t" \ 672 "mul r6, r3 \n\t" \ 673 "mov r5, r9 \n\t" \ 674 "mul r5, r7 \n\t" \ 675 "mov r3, r8 \n\t" \ 676 "mul r7, r3 \n\t" \ 677 "lsr r3, r6, #16 \n\t" \ 678 "add r5, r5, r3 \n\t" \ 679 "lsr r3, r7, #16 \n\t" \ 680 "add r5, r5, r3 \n\t" \ 681 "add r4, r4, r2 \n\t" \ 682 "mov r2, #0 \n\t" \ 683 "adc r5, r2 \n\t" \ 684 "lsl r3, r6, #16 \n\t" \ 685 "add r4, r4, r3 \n\t" \ 686 "adc r5, r2 \n\t" \ 687 "lsl r3, r7, #16 \n\t" \ 688 "add r4, r4, r3 \n\t" \ 689 "adc r5, r2 \n\t" \ 690 "ldr r3, [r1] \n\t" \ 691 "add r4, r4, r3 \n\t" \ 692 "adc r2, r5 \n\t" \ 693 "stmia r1!, {r4} \n\t" 694 695 #define MULADDC_STOP \ 696 "str r2, %0 \n\t" \ 697 "str r1, %1 \n\t" \ 698 "str r0, %2 \n\t" \ 699 : "=m" (c), "=m" (d), "=m" (s) \ 700 : "m" (s), "m" (d), "m" (c), "m" (b) \ 701 : "r0", "r1", "r2", "r3", "r4", "r5", \ 702 "r6", "r7", "r8", "r9", "cc" \ 703 ); 704 705 #elif (__ARM_ARCH >= 6) && \ 706 defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1) 707 708 #define MULADDC_INIT \ 709 asm( 710 711 #define MULADDC_CORE \ 712 "ldr r0, [%0], #4 \n\t" \ 713 "ldr r1, [%1] \n\t" \ 714 "umaal r1, %2, %3, r0 \n\t" \ 715 "str r1, [%1], #4 \n\t" 716 717 #define MULADDC_STOP \ 718 : "=r" (s), "=r" (d), "=r" (c) \ 719 : "r" (b), "0" (s), "1" (d), "2" (c) \ 720 : "r0", "r1", "memory" \ 721 ); 722 723 #else 724 725 #define MULADDC_INIT \ 726 asm( \ 727 "ldr r0, %3 \n\t" \ 728 "ldr r1, %4 \n\t" \ 729 "ldr r2, %5 \n\t" \ 730 "ldr r3, %6 \n\t" 731 732 #define MULADDC_CORE \ 733 "ldr r4, [r0], #4 \n\t" \ 734 "mov r5, #0 \n\t" \ 735 "ldr r6, [r1] \n\t" \ 736 "umlal r2, r5, r3, r4 \n\t" \ 737 "adds r7, r6, r2 \n\t" \ 738 "adc r2, r5, #0 \n\t" \ 739 "str r7, [r1], #4 \n\t" 740 741 #define MULADDC_STOP \ 742 "str r2, %0 \n\t" \ 743 "str r1, %1 \n\t" \ 744 "str r0, %2 \n\t" \ 745 : "=m" (c), "=m" (d), "=m" (s) \ 746 : "m" (s), "m" (d), "m" (c), "m" (b) \ 747 : "r0", "r1", "r2", "r3", "r4", "r5", \ 748 "r6", "r7", "cc" \ 749 ); 750 751 #endif /* Thumb */ 752 753 #endif /* ARMv3 */ 754 755 #if defined(__alpha__) 756 757 #define MULADDC_INIT \ 758 asm( \ 759 "ldq $1, %3 \n\t" \ 760 "ldq $2, %4 \n\t" \ 761 "ldq $3, %5 \n\t" \ 762 "ldq $4, %6 \n\t" 763 764 #define MULADDC_CORE \ 765 "ldq $6, 0($1) \n\t" \ 766 "addq $1, 8, $1 \n\t" \ 767 "mulq $6, $4, $7 \n\t" \ 768 "umulh $6, $4, $6 \n\t" \ 769 "addq $7, $3, $7 \n\t" \ 770 "cmpult $7, $3, $3 \n\t" \ 771 "ldq $5, 0($2) \n\t" \ 772 "addq $7, $5, $7 \n\t" \ 773 "cmpult $7, $5, $5 \n\t" \ 774 "stq $7, 0($2) \n\t" \ 775 "addq $2, 8, $2 \n\t" \ 776 "addq $6, $3, $3 \n\t" \ 777 "addq $5, $3, $3 \n\t" 778 779 #define MULADDC_STOP \ 780 "stq $3, %0 \n\t" \ 781 "stq $2, %1 \n\t" \ 782 "stq $1, %2 \n\t" \ 783 : "=m" (c), "=m" (d), "=m" (s) \ 784 : "m" (s), "m" (d), "m" (c), "m" (b) \ 785 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \ 786 ); 787 #endif /* Alpha */ 788 789 #if defined(__mips__) && !defined(__mips64) 790 791 #define MULADDC_INIT \ 792 asm( \ 793 "lw $10, %3 \n\t" \ 794 "lw $11, %4 \n\t" \ 795 "lw $12, %5 \n\t" \ 796 "lw $13, %6 \n\t" 797 798 #define MULADDC_CORE \ 799 "lw $14, 0($10) \n\t" \ 800 "multu $13, $14 \n\t" \ 801 "addi $10, $10, 4 \n\t" \ 802 "mflo $14 \n\t" \ 803 "mfhi $9 \n\t" \ 804 "addu $14, $12, $14 \n\t" \ 805 "lw $15, 0($11) \n\t" \ 806 "sltu $12, $14, $12 \n\t" \ 807 "addu $15, $14, $15 \n\t" \ 808 "sltu $14, $15, $14 \n\t" \ 809 "addu $12, $12, $9 \n\t" \ 810 "sw $15, 0($11) \n\t" \ 811 "addu $12, $12, $14 \n\t" \ 812 "addi $11, $11, 4 \n\t" 813 814 #define MULADDC_STOP \ 815 "sw $12, %0 \n\t" \ 816 "sw $11, %1 \n\t" \ 817 "sw $10, %2 \n\t" \ 818 : "=m" (c), "=m" (d), "=m" (s) \ 819 : "m" (s), "m" (d), "m" (c), "m" (b) \ 820 : "$9", "$10", "$11", "$12", "$13", "$14", "$15", "lo", "hi" \ 821 ); 822 823 #endif /* MIPS */ 824 #endif /* GNUC */ 825 826 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__) 827 828 #define MULADDC_INIT \ 829 __asm mov esi, s \ 830 __asm mov edi, d \ 831 __asm mov ecx, c \ 832 __asm mov ebx, b 833 834 #define MULADDC_CORE \ 835 __asm lodsd \ 836 __asm mul ebx \ 837 __asm add eax, ecx \ 838 __asm adc edx, 0 \ 839 __asm add eax, [edi] \ 840 __asm adc edx, 0 \ 841 __asm mov ecx, edx \ 842 __asm stosd 843 844 #if defined(MBEDTLS_HAVE_SSE2) 845 846 #define EMIT __asm _emit 847 848 #define MULADDC_HUIT \ 849 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \ 850 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \ 851 EMIT 0x0F EMIT 0x6E EMIT 0x1F \ 852 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 853 EMIT 0x0F EMIT 0x6E EMIT 0x16 \ 854 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ 855 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \ 856 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ 857 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \ 858 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ 859 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \ 860 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \ 861 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ 862 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \ 863 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \ 864 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \ 865 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \ 866 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \ 867 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \ 868 EMIT 0x0F EMIT 0x7E EMIT 0x0F \ 869 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \ 870 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ 871 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 872 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \ 873 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ 874 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 875 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \ 876 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ 877 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \ 878 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 879 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \ 880 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \ 881 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \ 882 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \ 883 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \ 884 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \ 885 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 886 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \ 887 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \ 888 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \ 889 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \ 890 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 891 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ 892 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \ 893 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \ 894 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \ 895 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 896 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \ 897 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \ 898 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \ 899 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \ 900 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 901 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \ 902 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \ 903 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 904 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 905 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \ 906 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \ 907 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \ 908 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 909 EMIT 0x0F EMIT 0x7E EMIT 0xC9 910 911 #define MULADDC_STOP \ 912 EMIT 0x0F EMIT 0x77 \ 913 __asm mov c, ecx \ 914 __asm mov d, edi \ 915 __asm mov s, esi \ 916 917 #else 918 919 #define MULADDC_STOP \ 920 __asm mov c, ecx \ 921 __asm mov d, edi \ 922 __asm mov s, esi \ 923 924 #endif /* SSE2 */ 925 #endif /* MSVC */ 926 927 #endif /* MBEDTLS_HAVE_ASM */ 928 929 #if !defined(MULADDC_CORE) 930 #if defined(MBEDTLS_HAVE_UDBL) 931 932 #define MULADDC_INIT \ 933 { \ 934 mbedtls_t_udbl r; \ 935 mbedtls_mpi_uint r0, r1; 936 937 #define MULADDC_CORE \ 938 r = *(s++) * (mbedtls_t_udbl) b; \ 939 r0 = (mbedtls_mpi_uint) r; \ 940 r1 = (mbedtls_mpi_uint)( r >> biL ); \ 941 r0 += c; r1 += (r0 < c); \ 942 r0 += *d; r1 += (r0 < *d); \ 943 c = r1; *(d++) = r0; 944 945 #define MULADDC_STOP \ 946 } 947 948 #else 949 #define MULADDC_INIT \ 950 { \ 951 mbedtls_mpi_uint s0, s1, b0, b1; \ 952 mbedtls_mpi_uint r0, r1, rx, ry; \ 953 b0 = ( b << biH ) >> biH; \ 954 b1 = ( b >> biH ); 955 956 #define MULADDC_CORE \ 957 s0 = ( *s << biH ) >> biH; \ 958 s1 = ( *s >> biH ); s++; \ 959 rx = s0 * b1; r0 = s0 * b0; \ 960 ry = s1 * b0; r1 = s1 * b1; \ 961 r1 += ( rx >> biH ); \ 962 r1 += ( ry >> biH ); \ 963 rx <<= biH; ry <<= biH; \ 964 r0 += rx; r1 += (r0 < rx); \ 965 r0 += ry; r1 += (r0 < ry); \ 966 r0 += c; r1 += (r0 < c); \ 967 r0 += *d; r1 += (r0 < *d); \ 968 c = r1; *(d++) = r0; 969 970 #define MULADDC_STOP \ 971 } 972 973 #endif /* C (generic) */ 974 #endif /* C (longlong) */ 975 976 #endif /* bn_mul.h */ 977