takashi kadono
/
Nucleo_446
Important changes to repositories hosted on mbed.com
Mbed hosted mercurial repositories are deprecated and are due to be permanently deleted in July 2026.
To keep a copy of this software download the repository Zip archive or clone locally using Mercurial.
It is also possible to export all your personal repositories from the account settings page.
Embed:
(wiki syntax)
Show/hide line numbers
bn_mul.h
Go to the documentation of this file.
00001 /** 00002 * \file bn_mul.h 00003 * 00004 * \brief Multi-precision integer library 00005 */ 00006 /* 00007 * Copyright (C) 2006-2015, ARM Limited, All Rights Reserved 00008 * SPDX-License-Identifier: Apache-2.0 00009 * 00010 * Licensed under the Apache License, Version 2.0 (the "License"); you may 00011 * not use this file except in compliance with the License. 00012 * You may obtain a copy of the License at 00013 * 00014 * http://www.apache.org/licenses/LICENSE-2.0 00015 * 00016 * Unless required by applicable law or agreed to in writing, software 00017 * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT 00018 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 00019 * See the License for the specific language governing permissions and 00020 * limitations under the License. 00021 * 00022 * This file is part of mbed TLS (https://tls.mbed.org) 00023 */ 00024 /* 00025 * Multiply source vector [s] with b, add result 00026 * to destination vector [d] and set carry c. 00027 * 00028 * Currently supports: 00029 * 00030 * . IA-32 (386+) . AMD64 / EM64T 00031 * . IA-32 (SSE2) . Motorola 68000 00032 * . PowerPC, 32-bit . MicroBlaze 00033 * . PowerPC, 64-bit . TriCore 00034 * . SPARC v8 . ARM v3+ 00035 * . Alpha . MIPS32 00036 * . C, longlong . C, generic 00037 */ 00038 #ifndef MBEDTLS_BN_MUL_H 00039 #define MBEDTLS_BN_MUL_H 00040 00041 #include "bignum.h" 00042 00043 #if defined(MBEDTLS_HAVE_ASM) 00044 00045 #ifndef asm 00046 #define asm __asm 00047 #endif 00048 00049 /* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */ 00050 #if defined(__GNUC__) && \ 00051 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 ) 00052 00053 /* 00054 * Disable use of the i386 assembly code below if option -O0, to disable all 00055 * compiler optimisations, is passed, detected with __OPTIMIZE__ 00056 * This is done as the number of registers used in the assembly code doesn't 00057 * work with the -O0 option. 00058 */ 00059 #if defined(__i386__) && defined(__OPTIMIZE__) 00060 00061 #define MULADDC_INIT \ 00062 asm( \ 00063 "movl %%ebx, %0 \n\t" \ 00064 "movl %5, %%esi \n\t" \ 00065 "movl %6, %%edi \n\t" \ 00066 "movl %7, %%ecx \n\t" \ 00067 "movl %8, %%ebx \n\t" 00068 00069 #define MULADDC_CORE \ 00070 "lodsl \n\t" \ 00071 "mull %%ebx \n\t" \ 00072 "addl %%ecx, %%eax \n\t" \ 00073 "adcl $0, %%edx \n\t" \ 00074 "addl (%%edi), %%eax \n\t" \ 00075 "adcl $0, %%edx \n\t" \ 00076 "movl %%edx, %%ecx \n\t" \ 00077 "stosl \n\t" 00078 00079 #if defined(MBEDTLS_HAVE_SSE2) 00080 00081 #define MULADDC_HUIT \ 00082 "movd %%ecx, %%mm1 \n\t" \ 00083 "movd %%ebx, %%mm0 \n\t" \ 00084 "movd (%%edi), %%mm3 \n\t" \ 00085 "paddq %%mm3, %%mm1 \n\t" \ 00086 "movd (%%esi), %%mm2 \n\t" \ 00087 "pmuludq %%mm0, %%mm2 \n\t" \ 00088 "movd 4(%%esi), %%mm4 \n\t" \ 00089 "pmuludq %%mm0, %%mm4 \n\t" \ 00090 "movd 8(%%esi), %%mm6 \n\t" \ 00091 "pmuludq %%mm0, %%mm6 \n\t" \ 00092 "movd 12(%%esi), %%mm7 \n\t" \ 00093 "pmuludq %%mm0, %%mm7 \n\t" \ 00094 "paddq %%mm2, %%mm1 \n\t" \ 00095 "movd 4(%%edi), %%mm3 \n\t" \ 00096 "paddq %%mm4, %%mm3 \n\t" \ 00097 "movd 8(%%edi), %%mm5 \n\t" \ 00098 "paddq %%mm6, %%mm5 \n\t" \ 00099 "movd 12(%%edi), %%mm4 \n\t" \ 00100 "paddq %%mm4, %%mm7 \n\t" \ 00101 "movd %%mm1, (%%edi) \n\t" \ 00102 "movd 16(%%esi), %%mm2 \n\t" \ 00103 "pmuludq %%mm0, %%mm2 \n\t" \ 00104 "psrlq $32, %%mm1 \n\t" \ 00105 "movd 20(%%esi), %%mm4 \n\t" \ 00106 "pmuludq %%mm0, %%mm4 \n\t" \ 00107 "paddq %%mm3, %%mm1 \n\t" \ 00108 "movd 24(%%esi), %%mm6 \n\t" \ 00109 "pmuludq %%mm0, %%mm6 \n\t" \ 00110 "movd %%mm1, 4(%%edi) \n\t" \ 00111 "psrlq $32, %%mm1 \n\t" \ 00112 "movd 28(%%esi), %%mm3 \n\t" \ 00113 "pmuludq %%mm0, %%mm3 \n\t" \ 00114 "paddq %%mm5, %%mm1 \n\t" \ 00115 "movd 16(%%edi), %%mm5 \n\t" \ 00116 "paddq %%mm5, %%mm2 \n\t" \ 00117 "movd %%mm1, 8(%%edi) \n\t" \ 00118 "psrlq $32, %%mm1 \n\t" \ 00119 "paddq %%mm7, %%mm1 \n\t" \ 00120 "movd 20(%%edi), %%mm5 \n\t" \ 00121 "paddq %%mm5, %%mm4 \n\t" \ 00122 "movd %%mm1, 12(%%edi) \n\t" \ 00123 "psrlq $32, %%mm1 \n\t" \ 00124 "paddq %%mm2, %%mm1 \n\t" \ 00125 "movd 24(%%edi), %%mm5 \n\t" \ 00126 "paddq %%mm5, %%mm6 \n\t" \ 00127 "movd %%mm1, 16(%%edi) \n\t" \ 00128 "psrlq $32, %%mm1 \n\t" \ 00129 "paddq %%mm4, %%mm1 \n\t" \ 00130 "movd 28(%%edi), %%mm5 \n\t" \ 00131 "paddq %%mm5, %%mm3 \n\t" \ 00132 "movd %%mm1, 20(%%edi) \n\t" \ 00133 "psrlq $32, %%mm1 \n\t" \ 00134 "paddq %%mm6, %%mm1 \n\t" \ 00135 "movd %%mm1, 24(%%edi) \n\t" \ 00136 "psrlq $32, %%mm1 \n\t" \ 00137 "paddq %%mm3, %%mm1 \n\t" \ 00138 "movd %%mm1, 28(%%edi) \n\t" \ 00139 "addl $32, %%edi \n\t" \ 00140 "addl $32, %%esi \n\t" \ 00141 "psrlq $32, %%mm1 \n\t" \ 00142 "movd %%mm1, %%ecx \n\t" 00143 00144 #define MULADDC_STOP \ 00145 "emms \n\t" \ 00146 "movl %4, %%ebx \n\t" \ 00147 "movl %%ecx, %1 \n\t" \ 00148 "movl %%edi, %2 \n\t" \ 00149 "movl %%esi, %3 \n\t" \ 00150 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ 00151 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ 00152 : "eax", "ebx", "ecx", "edx", "esi", "edi" \ 00153 ); 00154 00155 #else 00156 00157 #define MULADDC_STOP \ 00158 "movl %4, %%ebx \n\t" \ 00159 "movl %%ecx, %1 \n\t" \ 00160 "movl %%edi, %2 \n\t" \ 00161 "movl %%esi, %3 \n\t" \ 00162 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ 00163 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ 00164 : "eax", "ebx", "ecx", "edx", "esi", "edi" \ 00165 ); 00166 #endif /* SSE2 */ 00167 #endif /* i386 */ 00168 00169 #if defined(__amd64__) || defined (__x86_64__) 00170 00171 #define MULADDC_INIT \ 00172 asm( \ 00173 "xorq %%r8, %%r8 \n\t" 00174 00175 #define MULADDC_CORE \ 00176 "movq (%%rsi), %%rax \n\t" \ 00177 "mulq %%rbx \n\t" \ 00178 "addq $8, %%rsi \n\t" \ 00179 "addq %%rcx, %%rax \n\t" \ 00180 "movq %%r8, %%rcx \n\t" \ 00181 "adcq $0, %%rdx \n\t" \ 00182 "nop \n\t" \ 00183 "addq %%rax, (%%rdi) \n\t" \ 00184 "adcq %%rdx, %%rcx \n\t" \ 00185 "addq $8, %%rdi \n\t" 00186 00187 #define MULADDC_STOP \ 00188 : "+c" (c), "+D" (d), "+S" (s) \ 00189 : "b" (b) \ 00190 : "rax", "rdx", "r8" \ 00191 ); 00192 00193 #endif /* AMD64 */ 00194 00195 #if defined(__mc68020__) || defined(__mcpu32__) 00196 00197 #define MULADDC_INIT \ 00198 asm( \ 00199 "movl %3, %%a2 \n\t" \ 00200 "movl %4, %%a3 \n\t" \ 00201 "movl %5, %%d3 \n\t" \ 00202 "movl %6, %%d2 \n\t" \ 00203 "moveq #0, %%d0 \n\t" 00204 00205 #define MULADDC_CORE \ 00206 "movel %%a2@+, %%d1 \n\t" \ 00207 "mulul %%d2, %%d4:%%d1 \n\t" \ 00208 "addl %%d3, %%d1 \n\t" \ 00209 "addxl %%d0, %%d4 \n\t" \ 00210 "moveq #0, %%d3 \n\t" \ 00211 "addl %%d1, %%a3@+ \n\t" \ 00212 "addxl %%d4, %%d3 \n\t" 00213 00214 #define MULADDC_STOP \ 00215 "movl %%d3, %0 \n\t" \ 00216 "movl %%a3, %1 \n\t" \ 00217 "movl %%a2, %2 \n\t" \ 00218 : "=m" (c), "=m" (d), "=m" (s) \ 00219 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00220 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \ 00221 ); 00222 00223 #define MULADDC_HUIT \ 00224 "movel %%a2@+, %%d1 \n\t" \ 00225 "mulul %%d2, %%d4:%%d1 \n\t" \ 00226 "addxl %%d3, %%d1 \n\t" \ 00227 "addxl %%d0, %%d4 \n\t" \ 00228 "addl %%d1, %%a3@+ \n\t" \ 00229 "movel %%a2@+, %%d1 \n\t" \ 00230 "mulul %%d2, %%d3:%%d1 \n\t" \ 00231 "addxl %%d4, %%d1 \n\t" \ 00232 "addxl %%d0, %%d3 \n\t" \ 00233 "addl %%d1, %%a3@+ \n\t" \ 00234 "movel %%a2@+, %%d1 \n\t" \ 00235 "mulul %%d2, %%d4:%%d1 \n\t" \ 00236 "addxl %%d3, %%d1 \n\t" \ 00237 "addxl %%d0, %%d4 \n\t" \ 00238 "addl %%d1, %%a3@+ \n\t" \ 00239 "movel %%a2@+, %%d1 \n\t" \ 00240 "mulul %%d2, %%d3:%%d1 \n\t" \ 00241 "addxl %%d4, %%d1 \n\t" \ 00242 "addxl %%d0, %%d3 \n\t" \ 00243 "addl %%d1, %%a3@+ \n\t" \ 00244 "movel %%a2@+, %%d1 \n\t" \ 00245 "mulul %%d2, %%d4:%%d1 \n\t" \ 00246 "addxl %%d3, %%d1 \n\t" \ 00247 "addxl %%d0, %%d4 \n\t" \ 00248 "addl %%d1, %%a3@+ \n\t" \ 00249 "movel %%a2@+, %%d1 \n\t" \ 00250 "mulul %%d2, %%d3:%%d1 \n\t" \ 00251 "addxl %%d4, %%d1 \n\t" \ 00252 "addxl %%d0, %%d3 \n\t" \ 00253 "addl %%d1, %%a3@+ \n\t" \ 00254 "movel %%a2@+, %%d1 \n\t" \ 00255 "mulul %%d2, %%d4:%%d1 \n\t" \ 00256 "addxl %%d3, %%d1 \n\t" \ 00257 "addxl %%d0, %%d4 \n\t" \ 00258 "addl %%d1, %%a3@+ \n\t" \ 00259 "movel %%a2@+, %%d1 \n\t" \ 00260 "mulul %%d2, %%d3:%%d1 \n\t" \ 00261 "addxl %%d4, %%d1 \n\t" \ 00262 "addxl %%d0, %%d3 \n\t" \ 00263 "addl %%d1, %%a3@+ \n\t" \ 00264 "addxl %%d0, %%d3 \n\t" 00265 00266 #endif /* MC68000 */ 00267 00268 #if defined(__powerpc64__) || defined(__ppc64__) 00269 00270 #if defined(__MACH__) && defined(__APPLE__) 00271 00272 #define MULADDC_INIT \ 00273 asm( \ 00274 "ld r3, %3 \n\t" \ 00275 "ld r4, %4 \n\t" \ 00276 "ld r5, %5 \n\t" \ 00277 "ld r6, %6 \n\t" \ 00278 "addi r3, r3, -8 \n\t" \ 00279 "addi r4, r4, -8 \n\t" \ 00280 "addic r5, r5, 0 \n\t" 00281 00282 #define MULADDC_CORE \ 00283 "ldu r7, 8(r3) \n\t" \ 00284 "mulld r8, r7, r6 \n\t" \ 00285 "mulhdu r9, r7, r6 \n\t" \ 00286 "adde r8, r8, r5 \n\t" \ 00287 "ld r7, 8(r4) \n\t" \ 00288 "addze r5, r9 \n\t" \ 00289 "addc r8, r8, r7 \n\t" \ 00290 "stdu r8, 8(r4) \n\t" 00291 00292 #define MULADDC_STOP \ 00293 "addze r5, r5 \n\t" \ 00294 "addi r4, r4, 8 \n\t" \ 00295 "addi r3, r3, 8 \n\t" \ 00296 "std r5, %0 \n\t" \ 00297 "std r4, %1 \n\t" \ 00298 "std r3, %2 \n\t" \ 00299 : "=m" (c), "=m" (d), "=m" (s) \ 00300 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00301 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 00302 ); 00303 00304 00305 #else /* __MACH__ && __APPLE__ */ 00306 00307 #define MULADDC_INIT \ 00308 asm( \ 00309 "ld %%r3, %3 \n\t" \ 00310 "ld %%r4, %4 \n\t" \ 00311 "ld %%r5, %5 \n\t" \ 00312 "ld %%r6, %6 \n\t" \ 00313 "addi %%r3, %%r3, -8 \n\t" \ 00314 "addi %%r4, %%r4, -8 \n\t" \ 00315 "addic %%r5, %%r5, 0 \n\t" 00316 00317 #define MULADDC_CORE \ 00318 "ldu %%r7, 8(%%r3) \n\t" \ 00319 "mulld %%r8, %%r7, %%r6 \n\t" \ 00320 "mulhdu %%r9, %%r7, %%r6 \n\t" \ 00321 "adde %%r8, %%r8, %%r5 \n\t" \ 00322 "ld %%r7, 8(%%r4) \n\t" \ 00323 "addze %%r5, %%r9 \n\t" \ 00324 "addc %%r8, %%r8, %%r7 \n\t" \ 00325 "stdu %%r8, 8(%%r4) \n\t" 00326 00327 #define MULADDC_STOP \ 00328 "addze %%r5, %%r5 \n\t" \ 00329 "addi %%r4, %%r4, 8 \n\t" \ 00330 "addi %%r3, %%r3, 8 \n\t" \ 00331 "std %%r5, %0 \n\t" \ 00332 "std %%r4, %1 \n\t" \ 00333 "std %%r3, %2 \n\t" \ 00334 : "=m" (c), "=m" (d), "=m" (s) \ 00335 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00336 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 00337 ); 00338 00339 #endif /* __MACH__ && __APPLE__ */ 00340 00341 #elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */ 00342 00343 #if defined(__MACH__) && defined(__APPLE__) 00344 00345 #define MULADDC_INIT \ 00346 asm( \ 00347 "lwz r3, %3 \n\t" \ 00348 "lwz r4, %4 \n\t" \ 00349 "lwz r5, %5 \n\t" \ 00350 "lwz r6, %6 \n\t" \ 00351 "addi r3, r3, -4 \n\t" \ 00352 "addi r4, r4, -4 \n\t" \ 00353 "addic r5, r5, 0 \n\t" 00354 00355 #define MULADDC_CORE \ 00356 "lwzu r7, 4(r3) \n\t" \ 00357 "mullw r8, r7, r6 \n\t" \ 00358 "mulhwu r9, r7, r6 \n\t" \ 00359 "adde r8, r8, r5 \n\t" \ 00360 "lwz r7, 4(r4) \n\t" \ 00361 "addze r5, r9 \n\t" \ 00362 "addc r8, r8, r7 \n\t" \ 00363 "stwu r8, 4(r4) \n\t" 00364 00365 #define MULADDC_STOP \ 00366 "addze r5, r5 \n\t" \ 00367 "addi r4, r4, 4 \n\t" \ 00368 "addi r3, r3, 4 \n\t" \ 00369 "stw r5, %0 \n\t" \ 00370 "stw r4, %1 \n\t" \ 00371 "stw r3, %2 \n\t" \ 00372 : "=m" (c), "=m" (d), "=m" (s) \ 00373 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00374 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 00375 ); 00376 00377 #else /* __MACH__ && __APPLE__ */ 00378 00379 #define MULADDC_INIT \ 00380 asm( \ 00381 "lwz %%r3, %3 \n\t" \ 00382 "lwz %%r4, %4 \n\t" \ 00383 "lwz %%r5, %5 \n\t" \ 00384 "lwz %%r6, %6 \n\t" \ 00385 "addi %%r3, %%r3, -4 \n\t" \ 00386 "addi %%r4, %%r4, -4 \n\t" \ 00387 "addic %%r5, %%r5, 0 \n\t" 00388 00389 #define MULADDC_CORE \ 00390 "lwzu %%r7, 4(%%r3) \n\t" \ 00391 "mullw %%r8, %%r7, %%r6 \n\t" \ 00392 "mulhwu %%r9, %%r7, %%r6 \n\t" \ 00393 "adde %%r8, %%r8, %%r5 \n\t" \ 00394 "lwz %%r7, 4(%%r4) \n\t" \ 00395 "addze %%r5, %%r9 \n\t" \ 00396 "addc %%r8, %%r8, %%r7 \n\t" \ 00397 "stwu %%r8, 4(%%r4) \n\t" 00398 00399 #define MULADDC_STOP \ 00400 "addze %%r5, %%r5 \n\t" \ 00401 "addi %%r4, %%r4, 4 \n\t" \ 00402 "addi %%r3, %%r3, 4 \n\t" \ 00403 "stw %%r5, %0 \n\t" \ 00404 "stw %%r4, %1 \n\t" \ 00405 "stw %%r3, %2 \n\t" \ 00406 : "=m" (c), "=m" (d), "=m" (s) \ 00407 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00408 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ 00409 ); 00410 00411 #endif /* __MACH__ && __APPLE__ */ 00412 00413 #endif /* PPC32 */ 00414 00415 /* 00416 * The Sparc(64) assembly is reported to be broken. 00417 * Disable it for now, until we're able to fix it. 00418 */ 00419 #if 0 && defined(__sparc__) 00420 #if defined(__sparc64__) 00421 00422 #define MULADDC_INIT \ 00423 asm( \ 00424 "ldx %3, %%o0 \n\t" \ 00425 "ldx %4, %%o1 \n\t" \ 00426 "ld %5, %%o2 \n\t" \ 00427 "ld %6, %%o3 \n\t" 00428 00429 #define MULADDC_CORE \ 00430 "ld [%%o0], %%o4 \n\t" \ 00431 "inc 4, %%o0 \n\t" \ 00432 "ld [%%o1], %%o5 \n\t" \ 00433 "umul %%o3, %%o4, %%o4 \n\t" \ 00434 "addcc %%o4, %%o2, %%o4 \n\t" \ 00435 "rd %%y, %%g1 \n\t" \ 00436 "addx %%g1, 0, %%g1 \n\t" \ 00437 "addcc %%o4, %%o5, %%o4 \n\t" \ 00438 "st %%o4, [%%o1] \n\t" \ 00439 "addx %%g1, 0, %%o2 \n\t" \ 00440 "inc 4, %%o1 \n\t" 00441 00442 #define MULADDC_STOP \ 00443 "st %%o2, %0 \n\t" \ 00444 "stx %%o1, %1 \n\t" \ 00445 "stx %%o0, %2 \n\t" \ 00446 : "=m" (c), "=m" (d), "=m" (s) \ 00447 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00448 : "g1", "o0", "o1", "o2", "o3", "o4", \ 00449 "o5" \ 00450 ); 00451 00452 #else /* __sparc64__ */ 00453 00454 #define MULADDC_INIT \ 00455 asm( \ 00456 "ld %3, %%o0 \n\t" \ 00457 "ld %4, %%o1 \n\t" \ 00458 "ld %5, %%o2 \n\t" \ 00459 "ld %6, %%o3 \n\t" 00460 00461 #define MULADDC_CORE \ 00462 "ld [%%o0], %%o4 \n\t" \ 00463 "inc 4, %%o0 \n\t" \ 00464 "ld [%%o1], %%o5 \n\t" \ 00465 "umul %%o3, %%o4, %%o4 \n\t" \ 00466 "addcc %%o4, %%o2, %%o4 \n\t" \ 00467 "rd %%y, %%g1 \n\t" \ 00468 "addx %%g1, 0, %%g1 \n\t" \ 00469 "addcc %%o4, %%o5, %%o4 \n\t" \ 00470 "st %%o4, [%%o1] \n\t" \ 00471 "addx %%g1, 0, %%o2 \n\t" \ 00472 "inc 4, %%o1 \n\t" 00473 00474 #define MULADDC_STOP \ 00475 "st %%o2, %0 \n\t" \ 00476 "st %%o1, %1 \n\t" \ 00477 "st %%o0, %2 \n\t" \ 00478 : "=m" (c), "=m" (d), "=m" (s) \ 00479 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00480 : "g1", "o0", "o1", "o2", "o3", "o4", \ 00481 "o5" \ 00482 ); 00483 00484 #endif /* __sparc64__ */ 00485 #endif /* __sparc__ */ 00486 00487 #if defined(__microblaze__) || defined(microblaze) 00488 00489 #define MULADDC_INIT \ 00490 asm( \ 00491 "lwi r3, %3 \n\t" \ 00492 "lwi r4, %4 \n\t" \ 00493 "lwi r5, %5 \n\t" \ 00494 "lwi r6, %6 \n\t" \ 00495 "andi r7, r6, 0xffff \n\t" \ 00496 "bsrli r6, r6, 16 \n\t" 00497 00498 #define MULADDC_CORE \ 00499 "lhui r8, r3, 0 \n\t" \ 00500 "addi r3, r3, 2 \n\t" \ 00501 "lhui r9, r3, 0 \n\t" \ 00502 "addi r3, r3, 2 \n\t" \ 00503 "mul r10, r9, r6 \n\t" \ 00504 "mul r11, r8, r7 \n\t" \ 00505 "mul r12, r9, r7 \n\t" \ 00506 "mul r13, r8, r6 \n\t" \ 00507 "bsrli r8, r10, 16 \n\t" \ 00508 "bsrli r9, r11, 16 \n\t" \ 00509 "add r13, r13, r8 \n\t" \ 00510 "add r13, r13, r9 \n\t" \ 00511 "bslli r10, r10, 16 \n\t" \ 00512 "bslli r11, r11, 16 \n\t" \ 00513 "add r12, r12, r10 \n\t" \ 00514 "addc r13, r13, r0 \n\t" \ 00515 "add r12, r12, r11 \n\t" \ 00516 "addc r13, r13, r0 \n\t" \ 00517 "lwi r10, r4, 0 \n\t" \ 00518 "add r12, r12, r10 \n\t" \ 00519 "addc r13, r13, r0 \n\t" \ 00520 "add r12, r12, r5 \n\t" \ 00521 "addc r5, r13, r0 \n\t" \ 00522 "swi r12, r4, 0 \n\t" \ 00523 "addi r4, r4, 4 \n\t" 00524 00525 #define MULADDC_STOP \ 00526 "swi r5, %0 \n\t" \ 00527 "swi r4, %1 \n\t" \ 00528 "swi r3, %2 \n\t" \ 00529 : "=m" (c), "=m" (d), "=m" (s) \ 00530 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00531 : "r3", "r4", "r5", "r6", "r7", "r8", \ 00532 "r9", "r10", "r11", "r12", "r13" \ 00533 ); 00534 00535 #endif /* MicroBlaze */ 00536 00537 #if defined(__tricore__) 00538 00539 #define MULADDC_INIT \ 00540 asm( \ 00541 "ld.a %%a2, %3 \n\t" \ 00542 "ld.a %%a3, %4 \n\t" \ 00543 "ld.w %%d4, %5 \n\t" \ 00544 "ld.w %%d1, %6 \n\t" \ 00545 "xor %%d5, %%d5 \n\t" 00546 00547 #define MULADDC_CORE \ 00548 "ld.w %%d0, [%%a2+] \n\t" \ 00549 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \ 00550 "ld.w %%d0, [%%a3] \n\t" \ 00551 "addx %%d2, %%d2, %%d0 \n\t" \ 00552 "addc %%d3, %%d3, 0 \n\t" \ 00553 "mov %%d4, %%d3 \n\t" \ 00554 "st.w [%%a3+], %%d2 \n\t" 00555 00556 #define MULADDC_STOP \ 00557 "st.w %0, %%d4 \n\t" \ 00558 "st.a %1, %%a3 \n\t" \ 00559 "st.a %2, %%a2 \n\t" \ 00560 : "=m" (c), "=m" (d), "=m" (s) \ 00561 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00562 : "d0", "d1", "e2", "d4", "a2", "a3" \ 00563 ); 00564 00565 #endif /* TriCore */ 00566 00567 /* 00568 * gcc -O0 by default uses r7 for the frame pointer, so it complains about our 00569 * use of r7 below, unless -fomit-frame-pointer is passed. Unfortunately, 00570 * passing that option is not easy when building with yotta. 00571 * 00572 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with 00573 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by 00574 * clang and armcc5 under the same conditions). 00575 * 00576 * So, only use the optimized assembly below for optimized build, which avoids 00577 * the build error and is pretty reasonable anyway. 00578 */ 00579 #if defined(__GNUC__) && !defined(__OPTIMIZE__) 00580 #define MULADDC_CANNOT_USE_R7 00581 #endif 00582 00583 #if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7) 00584 00585 #if defined(__thumb__) && !defined(__thumb2__) 00586 00587 #define MULADDC_INIT \ 00588 asm( \ 00589 "ldr r0, %3 \n\t" \ 00590 "ldr r1, %4 \n\t" \ 00591 "ldr r2, %5 \n\t" \ 00592 "ldr r3, %6 \n\t" \ 00593 "lsr r7, r3, #16 \n\t" \ 00594 "mov r9, r7 \n\t" \ 00595 "lsl r7, r3, #16 \n\t" \ 00596 "lsr r7, r7, #16 \n\t" \ 00597 "mov r8, r7 \n\t" 00598 00599 #define MULADDC_CORE \ 00600 "ldmia r0!, {r6} \n\t" \ 00601 "lsr r7, r6, #16 \n\t" \ 00602 "lsl r6, r6, #16 \n\t" \ 00603 "lsr r6, r6, #16 \n\t" \ 00604 "mov r4, r8 \n\t" \ 00605 "mul r4, r6 \n\t" \ 00606 "mov r3, r9 \n\t" \ 00607 "mul r6, r3 \n\t" \ 00608 "mov r5, r9 \n\t" \ 00609 "mul r5, r7 \n\t" \ 00610 "mov r3, r8 \n\t" \ 00611 "mul r7, r3 \n\t" \ 00612 "lsr r3, r6, #16 \n\t" \ 00613 "add r5, r5, r3 \n\t" \ 00614 "lsr r3, r7, #16 \n\t" \ 00615 "add r5, r5, r3 \n\t" \ 00616 "add r4, r4, r2 \n\t" \ 00617 "mov r2, #0 \n\t" \ 00618 "adc r5, r2 \n\t" \ 00619 "lsl r3, r6, #16 \n\t" \ 00620 "add r4, r4, r3 \n\t" \ 00621 "adc r5, r2 \n\t" \ 00622 "lsl r3, r7, #16 \n\t" \ 00623 "add r4, r4, r3 \n\t" \ 00624 "adc r5, r2 \n\t" \ 00625 "ldr r3, [r1] \n\t" \ 00626 "add r4, r4, r3 \n\t" \ 00627 "adc r2, r5 \n\t" \ 00628 "stmia r1!, {r4} \n\t" 00629 00630 #define MULADDC_STOP \ 00631 "str r2, %0 \n\t" \ 00632 "str r1, %1 \n\t" \ 00633 "str r0, %2 \n\t" \ 00634 : "=m" (c), "=m" (d), "=m" (s) \ 00635 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00636 : "r0", "r1", "r2", "r3", "r4", "r5", \ 00637 "r6", "r7", "r8", "r9", "cc" \ 00638 ); 00639 00640 #else 00641 00642 #define MULADDC_INIT \ 00643 asm( \ 00644 "ldr r0, %3 \n\t" \ 00645 "ldr r1, %4 \n\t" \ 00646 "ldr r2, %5 \n\t" \ 00647 "ldr r3, %6 \n\t" 00648 00649 #define MULADDC_CORE \ 00650 "ldr r4, [r0], #4 \n\t" \ 00651 "mov r5, #0 \n\t" \ 00652 "ldr r6, [r1] \n\t" \ 00653 "umlal r2, r5, r3, r4 \n\t" \ 00654 "adds r7, r6, r2 \n\t" \ 00655 "adc r2, r5, #0 \n\t" \ 00656 "str r7, [r1], #4 \n\t" 00657 00658 #define MULADDC_STOP \ 00659 "str r2, %0 \n\t" \ 00660 "str r1, %1 \n\t" \ 00661 "str r0, %2 \n\t" \ 00662 : "=m" (c), "=m" (d), "=m" (s) \ 00663 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00664 : "r0", "r1", "r2", "r3", "r4", "r5", \ 00665 "r6", "r7", "cc" \ 00666 ); 00667 00668 #endif /* Thumb */ 00669 00670 #endif /* ARMv3 */ 00671 00672 #if defined(__alpha__) 00673 00674 #define MULADDC_INIT \ 00675 asm( \ 00676 "ldq $1, %3 \n\t" \ 00677 "ldq $2, %4 \n\t" \ 00678 "ldq $3, %5 \n\t" \ 00679 "ldq $4, %6 \n\t" 00680 00681 #define MULADDC_CORE \ 00682 "ldq $6, 0($1) \n\t" \ 00683 "addq $1, 8, $1 \n\t" \ 00684 "mulq $6, $4, $7 \n\t" \ 00685 "umulh $6, $4, $6 \n\t" \ 00686 "addq $7, $3, $7 \n\t" \ 00687 "cmpult $7, $3, $3 \n\t" \ 00688 "ldq $5, 0($2) \n\t" \ 00689 "addq $7, $5, $7 \n\t" \ 00690 "cmpult $7, $5, $5 \n\t" \ 00691 "stq $7, 0($2) \n\t" \ 00692 "addq $2, 8, $2 \n\t" \ 00693 "addq $6, $3, $3 \n\t" \ 00694 "addq $5, $3, $3 \n\t" 00695 00696 #define MULADDC_STOP \ 00697 "stq $3, %0 \n\t" \ 00698 "stq $2, %1 \n\t" \ 00699 "stq $1, %2 \n\t" \ 00700 : "=m" (c), "=m" (d), "=m" (s) \ 00701 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00702 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \ 00703 ); 00704 #endif /* Alpha */ 00705 00706 #if defined(__mips__) && !defined(__mips64) 00707 00708 #define MULADDC_INIT \ 00709 asm( \ 00710 "lw $10, %3 \n\t" \ 00711 "lw $11, %4 \n\t" \ 00712 "lw $12, %5 \n\t" \ 00713 "lw $13, %6 \n\t" 00714 00715 #define MULADDC_CORE \ 00716 "lw $14, 0($10) \n\t" \ 00717 "multu $13, $14 \n\t" \ 00718 "addi $10, $10, 4 \n\t" \ 00719 "mflo $14 \n\t" \ 00720 "mfhi $9 \n\t" \ 00721 "addu $14, $12, $14 \n\t" \ 00722 "lw $15, 0($11) \n\t" \ 00723 "sltu $12, $14, $12 \n\t" \ 00724 "addu $15, $14, $15 \n\t" \ 00725 "sltu $14, $15, $14 \n\t" \ 00726 "addu $12, $12, $9 \n\t" \ 00727 "sw $15, 0($11) \n\t" \ 00728 "addu $12, $12, $14 \n\t" \ 00729 "addi $11, $11, 4 \n\t" 00730 00731 #define MULADDC_STOP \ 00732 "sw $12, %0 \n\t" \ 00733 "sw $11, %1 \n\t" \ 00734 "sw $10, %2 \n\t" \ 00735 : "=m" (c), "=m" (d), "=m" (s) \ 00736 : "m" (s), "m" (d), "m" (c), "m" (b) \ 00737 : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \ 00738 ); 00739 00740 #endif /* MIPS */ 00741 #endif /* GNUC */ 00742 00743 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__) 00744 00745 #define MULADDC_INIT \ 00746 __asm mov esi, s \ 00747 __asm mov edi, d \ 00748 __asm mov ecx, c \ 00749 __asm mov ebx, b 00750 00751 #define MULADDC_CORE \ 00752 __asm lodsd \ 00753 __asm mul ebx \ 00754 __asm add eax, ecx \ 00755 __asm adc edx, 0 \ 00756 __asm add eax, [edi] \ 00757 __asm adc edx, 0 \ 00758 __asm mov ecx, edx \ 00759 __asm stosd 00760 00761 #if defined(MBEDTLS_HAVE_SSE2) 00762 00763 #define EMIT __asm _emit 00764 00765 #define MULADDC_HUIT \ 00766 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \ 00767 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \ 00768 EMIT 0x0F EMIT 0x6E EMIT 0x1F \ 00769 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 00770 EMIT 0x0F EMIT 0x6E EMIT 0x16 \ 00771 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ 00772 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \ 00773 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ 00774 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \ 00775 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ 00776 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \ 00777 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \ 00778 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ 00779 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \ 00780 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \ 00781 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \ 00782 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \ 00783 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \ 00784 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \ 00785 EMIT 0x0F EMIT 0x7E EMIT 0x0F \ 00786 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \ 00787 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ 00788 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00789 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \ 00790 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ 00791 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 00792 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \ 00793 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ 00794 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \ 00795 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00796 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \ 00797 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \ 00798 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \ 00799 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \ 00800 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \ 00801 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \ 00802 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00803 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \ 00804 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \ 00805 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \ 00806 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \ 00807 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00808 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ 00809 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \ 00810 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \ 00811 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \ 00812 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00813 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \ 00814 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \ 00815 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \ 00816 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \ 00817 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00818 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \ 00819 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \ 00820 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00821 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 00822 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \ 00823 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \ 00824 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \ 00825 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00826 EMIT 0x0F EMIT 0x7E EMIT 0xC9 00827 00828 #define MULADDC_STOP \ 00829 EMIT 0x0F EMIT 0x77 \ 00830 __asm mov c, ecx \ 00831 __asm mov d, edi \ 00832 __asm mov s, esi \ 00833 00834 #else 00835 00836 #define MULADDC_STOP \ 00837 __asm mov c, ecx \ 00838 __asm mov d, edi \ 00839 __asm mov s, esi \ 00840 00841 #endif /* SSE2 */ 00842 #endif /* MSVC */ 00843 00844 #endif /* MBEDTLS_HAVE_ASM */ 00845 00846 #if !defined(MULADDC_CORE) 00847 #if defined(MBEDTLS_HAVE_UDBL) 00848 00849 #define MULADDC_INIT \ 00850 { \ 00851 mbedtls_t_udbl r; \ 00852 mbedtls_mpi_uint r0, r1; 00853 00854 #define MULADDC_CORE \ 00855 r = *(s++) * (mbedtls_t_udbl) b; \ 00856 r0 = (mbedtls_mpi_uint) r; \ 00857 r1 = (mbedtls_mpi_uint)( r >> biL ); \ 00858 r0 += c; r1 += (r0 < c); \ 00859 r0 += *d; r1 += (r0 < *d); \ 00860 c = r1; *(d++) = r0; 00861 00862 #define MULADDC_STOP \ 00863 } 00864 00865 #else 00866 #define MULADDC_INIT \ 00867 { \ 00868 mbedtls_mpi_uint s0, s1, b0, b1; \ 00869 mbedtls_mpi_uint r0, r1, rx, ry; \ 00870 b0 = ( b << biH ) >> biH; \ 00871 b1 = ( b >> biH ); 00872 00873 #define MULADDC_CORE \ 00874 s0 = ( *s << biH ) >> biH; \ 00875 s1 = ( *s >> biH ); s++; \ 00876 rx = s0 * b1; r0 = s0 * b0; \ 00877 ry = s1 * b0; r1 = s1 * b1; \ 00878 r1 += ( rx >> biH ); \ 00879 r1 += ( ry >> biH ); \ 00880 rx <<= biH; ry <<= biH; \ 00881 r0 += rx; r1 += (r0 < rx); \ 00882 r0 += ry; r1 += (r0 < ry); \ 00883 r0 += c; r1 += (r0 < c); \ 00884 r0 += *d; r1 += (r0 < *d); \ 00885 c = r1; *(d++) = r0; 00886 00887 #define MULADDC_STOP \ 00888 } 00889 00890 #endif /* C (generic) */ 00891 #endif /* C (longlong) */ 00892 00893 #endif /* bn_mul.h */
Generated on Tue Aug 9 2022 00:37:03 by
1.7.2