bn_mul.h 35 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916
  1. /**
  2. * \file bn_mul.h
  3. *
  4. * \brief Multi-precision integer library
  5. */
  6. /*
  7. * Copyright (C) 2006-2015, ARM Limited, All Rights Reserved
  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. * This file is part of mbed TLS (https://tls.mbed.org)
  23. */
  24. /*
  25. * Multiply source vector [s] with b, add result
  26. * to destination vector [d] and set carry c.
  27. *
  28. * Currently supports:
  29. *
  30. * . IA-32 (386+) . AMD64 / EM64T
  31. * . IA-32 (SSE2) . Motorola 68000
  32. * . PowerPC, 32-bit . MicroBlaze
  33. * . PowerPC, 64-bit . TriCore
  34. * . SPARC v8 . ARM v3+
  35. * . Alpha . MIPS32
  36. * . C, longlong . C, generic
  37. */
  38. #ifndef MBEDTLS_BN_MUL_H
  39. #define MBEDTLS_BN_MUL_H
  40. #if !defined(MBEDTLS_CONFIG_FILE)
  41. #include "config.h"
  42. #else
  43. #include MBEDTLS_CONFIG_FILE
  44. #endif
  45. #include "bignum.h"
  46. #if defined(MBEDTLS_HAVE_ASM)
  47. #ifndef asm
  48. #define asm __asm
  49. #endif
  50. /* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */
  51. #if defined(__GNUC__) && \
  52. ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
  53. /*
  54. * Disable use of the i386 assembly code below if option -O0, to disable all
  55. * compiler optimisations, is passed, detected with __OPTIMIZE__
  56. * This is done as the number of registers used in the assembly code doesn't
  57. * work with the -O0 option.
  58. */
  59. #if defined(__i386__) && defined(__OPTIMIZE__)
  60. #define MULADDC_INIT \
  61. asm( \
  62. "movl %%ebx, %0 \n\t" \
  63. "movl %5, %%esi \n\t" \
  64. "movl %6, %%edi \n\t" \
  65. "movl %7, %%ecx \n\t" \
  66. "movl %8, %%ebx \n\t"
  67. #define MULADDC_CORE \
  68. "lodsl \n\t" \
  69. "mull %%ebx \n\t" \
  70. "addl %%ecx, %%eax \n\t" \
  71. "adcl $0, %%edx \n\t" \
  72. "addl (%%edi), %%eax \n\t" \
  73. "adcl $0, %%edx \n\t" \
  74. "movl %%edx, %%ecx \n\t" \
  75. "stosl \n\t"
  76. #if defined(MBEDTLS_HAVE_SSE2)
  77. #define MULADDC_HUIT \
  78. "movd %%ecx, %%mm1 \n\t" \
  79. "movd %%ebx, %%mm0 \n\t" \
  80. "movd (%%edi), %%mm3 \n\t" \
  81. "paddq %%mm3, %%mm1 \n\t" \
  82. "movd (%%esi), %%mm2 \n\t" \
  83. "pmuludq %%mm0, %%mm2 \n\t" \
  84. "movd 4(%%esi), %%mm4 \n\t" \
  85. "pmuludq %%mm0, %%mm4 \n\t" \
  86. "movd 8(%%esi), %%mm6 \n\t" \
  87. "pmuludq %%mm0, %%mm6 \n\t" \
  88. "movd 12(%%esi), %%mm7 \n\t" \
  89. "pmuludq %%mm0, %%mm7 \n\t" \
  90. "paddq %%mm2, %%mm1 \n\t" \
  91. "movd 4(%%edi), %%mm3 \n\t" \
  92. "paddq %%mm4, %%mm3 \n\t" \
  93. "movd 8(%%edi), %%mm5 \n\t" \
  94. "paddq %%mm6, %%mm5 \n\t" \
  95. "movd 12(%%edi), %%mm4 \n\t" \
  96. "paddq %%mm4, %%mm7 \n\t" \
  97. "movd %%mm1, (%%edi) \n\t" \
  98. "movd 16(%%esi), %%mm2 \n\t" \
  99. "pmuludq %%mm0, %%mm2 \n\t" \
  100. "psrlq $32, %%mm1 \n\t" \
  101. "movd 20(%%esi), %%mm4 \n\t" \
  102. "pmuludq %%mm0, %%mm4 \n\t" \
  103. "paddq %%mm3, %%mm1 \n\t" \
  104. "movd 24(%%esi), %%mm6 \n\t" \
  105. "pmuludq %%mm0, %%mm6 \n\t" \
  106. "movd %%mm1, 4(%%edi) \n\t" \
  107. "psrlq $32, %%mm1 \n\t" \
  108. "movd 28(%%esi), %%mm3 \n\t" \
  109. "pmuludq %%mm0, %%mm3 \n\t" \
  110. "paddq %%mm5, %%mm1 \n\t" \
  111. "movd 16(%%edi), %%mm5 \n\t" \
  112. "paddq %%mm5, %%mm2 \n\t" \
  113. "movd %%mm1, 8(%%edi) \n\t" \
  114. "psrlq $32, %%mm1 \n\t" \
  115. "paddq %%mm7, %%mm1 \n\t" \
  116. "movd 20(%%edi), %%mm5 \n\t" \
  117. "paddq %%mm5, %%mm4 \n\t" \
  118. "movd %%mm1, 12(%%edi) \n\t" \
  119. "psrlq $32, %%mm1 \n\t" \
  120. "paddq %%mm2, %%mm1 \n\t" \
  121. "movd 24(%%edi), %%mm5 \n\t" \
  122. "paddq %%mm5, %%mm6 \n\t" \
  123. "movd %%mm1, 16(%%edi) \n\t" \
  124. "psrlq $32, %%mm1 \n\t" \
  125. "paddq %%mm4, %%mm1 \n\t" \
  126. "movd 28(%%edi), %%mm5 \n\t" \
  127. "paddq %%mm5, %%mm3 \n\t" \
  128. "movd %%mm1, 20(%%edi) \n\t" \
  129. "psrlq $32, %%mm1 \n\t" \
  130. "paddq %%mm6, %%mm1 \n\t" \
  131. "movd %%mm1, 24(%%edi) \n\t" \
  132. "psrlq $32, %%mm1 \n\t" \
  133. "paddq %%mm3, %%mm1 \n\t" \
  134. "movd %%mm1, 28(%%edi) \n\t" \
  135. "addl $32, %%edi \n\t" \
  136. "addl $32, %%esi \n\t" \
  137. "psrlq $32, %%mm1 \n\t" \
  138. "movd %%mm1, %%ecx \n\t"
  139. #define MULADDC_STOP \
  140. "emms \n\t" \
  141. "movl %4, %%ebx \n\t" \
  142. "movl %%ecx, %1 \n\t" \
  143. "movl %%edi, %2 \n\t" \
  144. "movl %%esi, %3 \n\t" \
  145. : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
  146. : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
  147. : "eax", "ebx", "ecx", "edx", "esi", "edi" \
  148. );
  149. #else
  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", "ebx", "ecx", "edx", "esi", "edi" \
  158. );
  159. #endif /* SSE2 */
  160. #endif /* i386 */
  161. #if defined(__amd64__) || defined (__x86_64__)
  162. #define MULADDC_INIT \
  163. asm( \
  164. "xorq %%r8, %%r8\n"
  165. #define MULADDC_CORE \
  166. "movq (%%rsi), %%rax\n" \
  167. "mulq %%rbx\n" \
  168. "addq $8, %%rsi\n" \
  169. "addq %%rcx, %%rax\n" \
  170. "movq %%r8, %%rcx\n" \
  171. "adcq $0, %%rdx\n" \
  172. "nop \n" \
  173. "addq %%rax, (%%rdi)\n" \
  174. "adcq %%rdx, %%rcx\n" \
  175. "addq $8, %%rdi\n"
  176. #define MULADDC_STOP \
  177. : "+c" (c), "+D" (d), "+S" (s) \
  178. : "b" (b) \
  179. : "rax", "rdx", "r8" \
  180. );
  181. #endif /* AMD64 */
  182. #if defined(__mc68020__) || defined(__mcpu32__)
  183. #define MULADDC_INIT \
  184. asm( \
  185. "movl %3, %%a2 \n\t" \
  186. "movl %4, %%a3 \n\t" \
  187. "movl %5, %%d3 \n\t" \
  188. "movl %6, %%d2 \n\t" \
  189. "moveq #0, %%d0 \n\t"
  190. #define MULADDC_CORE \
  191. "movel %%a2@+, %%d1 \n\t" \
  192. "mulul %%d2, %%d4:%%d1 \n\t" \
  193. "addl %%d3, %%d1 \n\t" \
  194. "addxl %%d0, %%d4 \n\t" \
  195. "moveq #0, %%d3 \n\t" \
  196. "addl %%d1, %%a3@+ \n\t" \
  197. "addxl %%d4, %%d3 \n\t"
  198. #define MULADDC_STOP \
  199. "movl %%d3, %0 \n\t" \
  200. "movl %%a3, %1 \n\t" \
  201. "movl %%a2, %2 \n\t" \
  202. : "=m" (c), "=m" (d), "=m" (s) \
  203. : "m" (s), "m" (d), "m" (c), "m" (b) \
  204. : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
  205. );
  206. #define MULADDC_HUIT \
  207. "movel %%a2@+, %%d1 \n\t" \
  208. "mulul %%d2, %%d4:%%d1 \n\t" \
  209. "addxl %%d3, %%d1 \n\t" \
  210. "addxl %%d0, %%d4 \n\t" \
  211. "addl %%d1, %%a3@+ \n\t" \
  212. "movel %%a2@+, %%d1 \n\t" \
  213. "mulul %%d2, %%d3:%%d1 \n\t" \
  214. "addxl %%d4, %%d1 \n\t" \
  215. "addxl %%d0, %%d3 \n\t" \
  216. "addl %%d1, %%a3@+ \n\t" \
  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. "addxl %%d0, %%d3 \n\t"
  248. #endif /* MC68000 */
  249. #if defined(__powerpc64__) || defined(__ppc64__)
  250. #if defined(__MACH__) && defined(__APPLE__)
  251. #define MULADDC_INIT \
  252. asm( \
  253. "ld r3, %3 \n\t" \
  254. "ld r4, %4 \n\t" \
  255. "ld r5, %5 \n\t" \
  256. "ld r6, %6 \n\t" \
  257. "addi r3, r3, -8 \n\t" \
  258. "addi r4, r4, -8 \n\t" \
  259. "addic r5, r5, 0 \n\t"
  260. #define MULADDC_CORE \
  261. "ldu r7, 8(r3) \n\t" \
  262. "mulld r8, r7, r6 \n\t" \
  263. "mulhdu r9, r7, r6 \n\t" \
  264. "adde r8, r8, r5 \n\t" \
  265. "ld r7, 8(r4) \n\t" \
  266. "addze r5, r9 \n\t" \
  267. "addc r8, r8, r7 \n\t" \
  268. "stdu r8, 8(r4) \n\t"
  269. #define MULADDC_STOP \
  270. "addze r5, r5 \n\t" \
  271. "addi r4, r4, 8 \n\t" \
  272. "addi r3, r3, 8 \n\t" \
  273. "std r5, %0 \n\t" \
  274. "std r4, %1 \n\t" \
  275. "std r3, %2 \n\t" \
  276. : "=m" (c), "=m" (d), "=m" (s) \
  277. : "m" (s), "m" (d), "m" (c), "m" (b) \
  278. : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
  279. );
  280. #else /* __MACH__ && __APPLE__ */
  281. #define MULADDC_INIT \
  282. asm( \
  283. "ld %%r3, %3 \n\t" \
  284. "ld %%r4, %4 \n\t" \
  285. "ld %%r5, %5 \n\t" \
  286. "ld %%r6, %6 \n\t" \
  287. "addi %%r3, %%r3, -8 \n\t" \
  288. "addi %%r4, %%r4, -8 \n\t" \
  289. "addic %%r5, %%r5, 0 \n\t"
  290. #define MULADDC_CORE \
  291. "ldu %%r7, 8(%%r3) \n\t" \
  292. "mulld %%r8, %%r7, %%r6 \n\t" \
  293. "mulhdu %%r9, %%r7, %%r6 \n\t" \
  294. "adde %%r8, %%r8, %%r5 \n\t" \
  295. "ld %%r7, 8(%%r4) \n\t" \
  296. "addze %%r5, %%r9 \n\t" \
  297. "addc %%r8, %%r8, %%r7 \n\t" \
  298. "stdu %%r8, 8(%%r4) \n\t"
  299. #define MULADDC_STOP \
  300. "addze %%r5, %%r5 \n\t" \
  301. "addi %%r4, %%r4, 8 \n\t" \
  302. "addi %%r3, %%r3, 8 \n\t" \
  303. "std %%r5, %0 \n\t" \
  304. "std %%r4, %1 \n\t" \
  305. "std %%r3, %2 \n\t" \
  306. : "=m" (c), "=m" (d), "=m" (s) \
  307. : "m" (s), "m" (d), "m" (c), "m" (b) \
  308. : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
  309. );
  310. #endif /* __MACH__ && __APPLE__ */
  311. #elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
  312. #if defined(__MACH__) && defined(__APPLE__)
  313. #define MULADDC_INIT \
  314. asm( \
  315. "lwz r3, %3 \n\t" \
  316. "lwz r4, %4 \n\t" \
  317. "lwz r5, %5 \n\t" \
  318. "lwz r6, %6 \n\t" \
  319. "addi r3, r3, -4 \n\t" \
  320. "addi r4, r4, -4 \n\t" \
  321. "addic r5, r5, 0 \n\t"
  322. #define MULADDC_CORE \
  323. "lwzu r7, 4(r3) \n\t" \
  324. "mullw r8, r7, r6 \n\t" \
  325. "mulhwu r9, r7, r6 \n\t" \
  326. "adde r8, r8, r5 \n\t" \
  327. "lwz r7, 4(r4) \n\t" \
  328. "addze r5, r9 \n\t" \
  329. "addc r8, r8, r7 \n\t" \
  330. "stwu r8, 4(r4) \n\t"
  331. #define MULADDC_STOP \
  332. "addze r5, r5 \n\t" \
  333. "addi r4, r4, 4 \n\t" \
  334. "addi r3, r3, 4 \n\t" \
  335. "stw r5, %0 \n\t" \
  336. "stw r4, %1 \n\t" \
  337. "stw r3, %2 \n\t" \
  338. : "=m" (c), "=m" (d), "=m" (s) \
  339. : "m" (s), "m" (d), "m" (c), "m" (b) \
  340. : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
  341. );
  342. #else /* __MACH__ && __APPLE__ */
  343. #define MULADDC_INIT \
  344. asm( \
  345. "lwz %%r3, %3 \n\t" \
  346. "lwz %%r4, %4 \n\t" \
  347. "lwz %%r5, %5 \n\t" \
  348. "lwz %%r6, %6 \n\t" \
  349. "addi %%r3, %%r3, -4 \n\t" \
  350. "addi %%r4, %%r4, -4 \n\t" \
  351. "addic %%r5, %%r5, 0 \n\t"
  352. #define MULADDC_CORE \
  353. "lwzu %%r7, 4(%%r3) \n\t" \
  354. "mullw %%r8, %%r7, %%r6 \n\t" \
  355. "mulhwu %%r9, %%r7, %%r6 \n\t" \
  356. "adde %%r8, %%r8, %%r5 \n\t" \
  357. "lwz %%r7, 4(%%r4) \n\t" \
  358. "addze %%r5, %%r9 \n\t" \
  359. "addc %%r8, %%r8, %%r7 \n\t" \
  360. "stwu %%r8, 4(%%r4) \n\t"
  361. #define MULADDC_STOP \
  362. "addze %%r5, %%r5 \n\t" \
  363. "addi %%r4, %%r4, 4 \n\t" \
  364. "addi %%r3, %%r3, 4 \n\t" \
  365. "stw %%r5, %0 \n\t" \
  366. "stw %%r4, %1 \n\t" \
  367. "stw %%r3, %2 \n\t" \
  368. : "=m" (c), "=m" (d), "=m" (s) \
  369. : "m" (s), "m" (d), "m" (c), "m" (b) \
  370. : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
  371. );
  372. #endif /* __MACH__ && __APPLE__ */
  373. #endif /* PPC32 */
  374. /*
  375. * The Sparc(64) assembly is reported to be broken.
  376. * Disable it for now, until we're able to fix it.
  377. */
  378. #if 0 && defined(__sparc__)
  379. #if defined(__sparc64__)
  380. #define MULADDC_INIT \
  381. asm( \
  382. "ldx %3, %%o0 \n\t" \
  383. "ldx %4, %%o1 \n\t" \
  384. "ld %5, %%o2 \n\t" \
  385. "ld %6, %%o3 \n\t"
  386. #define MULADDC_CORE \
  387. "ld [%%o0], %%o4 \n\t" \
  388. "inc 4, %%o0 \n\t" \
  389. "ld [%%o1], %%o5 \n\t" \
  390. "umul %%o3, %%o4, %%o4 \n\t" \
  391. "addcc %%o4, %%o2, %%o4 \n\t" \
  392. "rd %%y, %%g1 \n\t" \
  393. "addx %%g1, 0, %%g1 \n\t" \
  394. "addcc %%o4, %%o5, %%o4 \n\t" \
  395. "st %%o4, [%%o1] \n\t" \
  396. "addx %%g1, 0, %%o2 \n\t" \
  397. "inc 4, %%o1 \n\t"
  398. #define MULADDC_STOP \
  399. "st %%o2, %0 \n\t" \
  400. "stx %%o1, %1 \n\t" \
  401. "stx %%o0, %2 \n\t" \
  402. : "=m" (c), "=m" (d), "=m" (s) \
  403. : "m" (s), "m" (d), "m" (c), "m" (b) \
  404. : "g1", "o0", "o1", "o2", "o3", "o4", \
  405. "o5" \
  406. );
  407. #else /* __sparc64__ */
  408. #define MULADDC_INIT \
  409. asm( \
  410. "ld %3, %%o0 \n\t" \
  411. "ld %4, %%o1 \n\t" \
  412. "ld %5, %%o2 \n\t" \
  413. "ld %6, %%o3 \n\t"
  414. #define MULADDC_CORE \
  415. "ld [%%o0], %%o4 \n\t" \
  416. "inc 4, %%o0 \n\t" \
  417. "ld [%%o1], %%o5 \n\t" \
  418. "umul %%o3, %%o4, %%o4 \n\t" \
  419. "addcc %%o4, %%o2, %%o4 \n\t" \
  420. "rd %%y, %%g1 \n\t" \
  421. "addx %%g1, 0, %%g1 \n\t" \
  422. "addcc %%o4, %%o5, %%o4 \n\t" \
  423. "st %%o4, [%%o1] \n\t" \
  424. "addx %%g1, 0, %%o2 \n\t" \
  425. "inc 4, %%o1 \n\t"
  426. #define MULADDC_STOP \
  427. "st %%o2, %0 \n\t" \
  428. "st %%o1, %1 \n\t" \
  429. "st %%o0, %2 \n\t" \
  430. : "=m" (c), "=m" (d), "=m" (s) \
  431. : "m" (s), "m" (d), "m" (c), "m" (b) \
  432. : "g1", "o0", "o1", "o2", "o3", "o4", \
  433. "o5" \
  434. );
  435. #endif /* __sparc64__ */
  436. #endif /* __sparc__ */
  437. #if defined(__microblaze__) || defined(microblaze)
  438. #define MULADDC_INIT \
  439. asm( \
  440. "lwi r3, %3 \n\t" \
  441. "lwi r4, %4 \n\t" \
  442. "lwi r5, %5 \n\t" \
  443. "lwi r6, %6 \n\t" \
  444. "andi r7, r6, 0xffff \n\t" \
  445. "bsrli r6, r6, 16 \n\t"
  446. #define MULADDC_CORE \
  447. "lhui r8, r3, 0 \n\t" \
  448. "addi r3, r3, 2 \n\t" \
  449. "lhui r9, r3, 0 \n\t" \
  450. "addi r3, r3, 2 \n\t" \
  451. "mul r10, r9, r6 \n\t" \
  452. "mul r11, r8, r7 \n\t" \
  453. "mul r12, r9, r7 \n\t" \
  454. "mul r13, r8, r6 \n\t" \
  455. "bsrli r8, r10, 16 \n\t" \
  456. "bsrli r9, r11, 16 \n\t" \
  457. "add r13, r13, r8 \n\t" \
  458. "add r13, r13, r9 \n\t" \
  459. "bslli r10, r10, 16 \n\t" \
  460. "bslli r11, r11, 16 \n\t" \
  461. "add r12, r12, r10 \n\t" \
  462. "addc r13, r13, r0 \n\t" \
  463. "add r12, r12, r11 \n\t" \
  464. "addc r13, r13, r0 \n\t" \
  465. "lwi r10, r4, 0 \n\t" \
  466. "add r12, r12, r10 \n\t" \
  467. "addc r13, r13, r0 \n\t" \
  468. "add r12, r12, r5 \n\t" \
  469. "addc r5, r13, r0 \n\t" \
  470. "swi r12, r4, 0 \n\t" \
  471. "addi r4, r4, 4 \n\t"
  472. #define MULADDC_STOP \
  473. "swi r5, %0 \n\t" \
  474. "swi r4, %1 \n\t" \
  475. "swi r3, %2 \n\t" \
  476. : "=m" (c), "=m" (d), "=m" (s) \
  477. : "m" (s), "m" (d), "m" (c), "m" (b) \
  478. : "r3", "r4", "r5", "r6", "r7", "r8", \
  479. "r9", "r10", "r11", "r12", "r13" \
  480. );
  481. #endif /* MicroBlaze */
  482. #if defined(__tricore__)
  483. #define MULADDC_INIT \
  484. asm( \
  485. "ld.a %%a2, %3 \n\t" \
  486. "ld.a %%a3, %4 \n\t" \
  487. "ld.w %%d4, %5 \n\t" \
  488. "ld.w %%d1, %6 \n\t" \
  489. "xor %%d5, %%d5 \n\t"
  490. #define MULADDC_CORE \
  491. "ld.w %%d0, [%%a2+] \n\t" \
  492. "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
  493. "ld.w %%d0, [%%a3] \n\t" \
  494. "addx %%d2, %%d2, %%d0 \n\t" \
  495. "addc %%d3, %%d3, 0 \n\t" \
  496. "mov %%d4, %%d3 \n\t" \
  497. "st.w [%%a3+], %%d2 \n\t"
  498. #define MULADDC_STOP \
  499. "st.w %0, %%d4 \n\t" \
  500. "st.a %1, %%a3 \n\t" \
  501. "st.a %2, %%a2 \n\t" \
  502. : "=m" (c), "=m" (d), "=m" (s) \
  503. : "m" (s), "m" (d), "m" (c), "m" (b) \
  504. : "d0", "d1", "e2", "d4", "a2", "a3" \
  505. );
  506. #endif /* TriCore */
  507. /*
  508. * Note, gcc -O0 by default uses r7 for the frame pointer, so it complains about
  509. * our use of r7 below, unless -fomit-frame-pointer is passed.
  510. *
  511. * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
  512. * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
  513. * clang and armcc5 under the same conditions).
  514. *
  515. * So, only use the optimized assembly below for optimized build, which avoids
  516. * the build error and is pretty reasonable anyway.
  517. */
  518. #if defined(__GNUC__) && !defined(__OPTIMIZE__)
  519. #define MULADDC_CANNOT_USE_R7
  520. #endif
  521. #if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
  522. #if defined(__thumb__) && !defined(__thumb2__)
  523. #define MULADDC_INIT \
  524. asm( \
  525. "ldr r0, %3 \n\t" \
  526. "ldr r1, %4 \n\t" \
  527. "ldr r2, %5 \n\t" \
  528. "ldr r3, %6 \n\t" \
  529. "lsr r7, r3, #16 \n\t" \
  530. "mov r9, r7 \n\t" \
  531. "lsl r7, r3, #16 \n\t" \
  532. "lsr r7, r7, #16 \n\t" \
  533. "mov r8, r7 \n\t"
  534. #define MULADDC_CORE \
  535. "ldmia r0!, {r6} \n\t" \
  536. "lsr r7, r6, #16 \n\t" \
  537. "lsl r6, r6, #16 \n\t" \
  538. "lsr r6, r6, #16 \n\t" \
  539. "mov r4, r8 \n\t" \
  540. "mul r4, r6 \n\t" \
  541. "mov r3, r9 \n\t" \
  542. "mul r6, r3 \n\t" \
  543. "mov r5, r9 \n\t" \
  544. "mul r5, r7 \n\t" \
  545. "mov r3, r8 \n\t" \
  546. "mul r7, r3 \n\t" \
  547. "lsr r3, r6, #16 \n\t" \
  548. "add r5, r5, r3 \n\t" \
  549. "lsr r3, r7, #16 \n\t" \
  550. "add r5, r5, r3 \n\t" \
  551. "add r4, r4, r2 \n\t" \
  552. "mov r2, #0 \n\t" \
  553. "adc r5, r2 \n\t" \
  554. "lsl r3, r6, #16 \n\t" \
  555. "add r4, r4, r3 \n\t" \
  556. "adc r5, r2 \n\t" \
  557. "lsl r3, r7, #16 \n\t" \
  558. "add r4, r4, r3 \n\t" \
  559. "adc r5, r2 \n\t" \
  560. "ldr r3, [r1] \n\t" \
  561. "add r4, r4, r3 \n\t" \
  562. "adc r2, r5 \n\t" \
  563. "stmia r1!, {r4} \n\t"
  564. #define MULADDC_STOP \
  565. "str r2, %0 \n\t" \
  566. "str r1, %1 \n\t" \
  567. "str r0, %2 \n\t" \
  568. : "=m" (c), "=m" (d), "=m" (s) \
  569. : "m" (s), "m" (d), "m" (c), "m" (b) \
  570. : "r0", "r1", "r2", "r3", "r4", "r5", \
  571. "r6", "r7", "r8", "r9", "cc" \
  572. );
  573. #elif (__ARM_ARCH >= 6) && \
  574. defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)
  575. #define MULADDC_INIT \
  576. asm(
  577. #define MULADDC_CORE \
  578. "ldr r0, [%0], #4 \n\t" \
  579. "ldr r1, [%1] \n\t" \
  580. "umaal r1, %2, %3, r0 \n\t" \
  581. "str r1, [%1], #4 \n\t"
  582. #define MULADDC_STOP \
  583. : "=r" (s), "=r" (d), "=r" (c) \
  584. : "r" (b), "0" (s), "1" (d), "2" (c) \
  585. : "r0", "r1", "memory" \
  586. );
  587. #else
  588. #define MULADDC_INIT \
  589. asm( \
  590. "ldr r0, %3 \n\t" \
  591. "ldr r1, %4 \n\t" \
  592. "ldr r2, %5 \n\t" \
  593. "ldr r3, %6 \n\t"
  594. #define MULADDC_CORE \
  595. "ldr r4, [r0], #4 \n\t" \
  596. "mov r5, #0 \n\t" \
  597. "ldr r6, [r1] \n\t" \
  598. "umlal r2, r5, r3, r4 \n\t" \
  599. "adds r7, r6, r2 \n\t" \
  600. "adc r2, r5, #0 \n\t" \
  601. "str r7, [r1], #4 \n\t"
  602. #define MULADDC_STOP \
  603. "str r2, %0 \n\t" \
  604. "str r1, %1 \n\t" \
  605. "str r0, %2 \n\t" \
  606. : "=m" (c), "=m" (d), "=m" (s) \
  607. : "m" (s), "m" (d), "m" (c), "m" (b) \
  608. : "r0", "r1", "r2", "r3", "r4", "r5", \
  609. "r6", "r7", "cc" \
  610. );
  611. #endif /* Thumb */
  612. #endif /* ARMv3 */
  613. #if defined(__alpha__)
  614. #define MULADDC_INIT \
  615. asm( \
  616. "ldq $1, %3 \n\t" \
  617. "ldq $2, %4 \n\t" \
  618. "ldq $3, %5 \n\t" \
  619. "ldq $4, %6 \n\t"
  620. #define MULADDC_CORE \
  621. "ldq $6, 0($1) \n\t" \
  622. "addq $1, 8, $1 \n\t" \
  623. "mulq $6, $4, $7 \n\t" \
  624. "umulh $6, $4, $6 \n\t" \
  625. "addq $7, $3, $7 \n\t" \
  626. "cmpult $7, $3, $3 \n\t" \
  627. "ldq $5, 0($2) \n\t" \
  628. "addq $7, $5, $7 \n\t" \
  629. "cmpult $7, $5, $5 \n\t" \
  630. "stq $7, 0($2) \n\t" \
  631. "addq $2, 8, $2 \n\t" \
  632. "addq $6, $3, $3 \n\t" \
  633. "addq $5, $3, $3 \n\t"
  634. #define MULADDC_STOP \
  635. "stq $3, %0 \n\t" \
  636. "stq $2, %1 \n\t" \
  637. "stq $1, %2 \n\t" \
  638. : "=m" (c), "=m" (d), "=m" (s) \
  639. : "m" (s), "m" (d), "m" (c), "m" (b) \
  640. : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
  641. );
  642. #endif /* Alpha */
  643. #if defined(__mips__) && !defined(__mips64)
  644. #define MULADDC_INIT \
  645. asm( \
  646. "lw $10, %3 \n\t" \
  647. "lw $11, %4 \n\t" \
  648. "lw $12, %5 \n\t" \
  649. "lw $13, %6 \n\t"
  650. #define MULADDC_CORE \
  651. "lw $14, 0($10) \n\t" \
  652. "multu $13, $14 \n\t" \
  653. "addi $10, $10, 4 \n\t" \
  654. "mflo $14 \n\t" \
  655. "mfhi $9 \n\t" \
  656. "addu $14, $12, $14 \n\t" \
  657. "lw $15, 0($11) \n\t" \
  658. "sltu $12, $14, $12 \n\t" \
  659. "addu $15, $14, $15 \n\t" \
  660. "sltu $14, $15, $14 \n\t" \
  661. "addu $12, $12, $9 \n\t" \
  662. "sw $15, 0($11) \n\t" \
  663. "addu $12, $12, $14 \n\t" \
  664. "addi $11, $11, 4 \n\t"
  665. #define MULADDC_STOP \
  666. "sw $12, %0 \n\t" \
  667. "sw $11, %1 \n\t" \
  668. "sw $10, %2 \n\t" \
  669. : "=m" (c), "=m" (d), "=m" (s) \
  670. : "m" (s), "m" (d), "m" (c), "m" (b) \
  671. : "$9", "$10", "$11", "$12", "$13", "$14", "$15", "lo", "hi" \
  672. );
  673. #endif /* MIPS */
  674. #endif /* GNUC */
  675. #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
  676. #define MULADDC_INIT \
  677. __asm mov esi, s \
  678. __asm mov edi, d \
  679. __asm mov ecx, c \
  680. __asm mov ebx, b
  681. #define MULADDC_CORE \
  682. __asm lodsd \
  683. __asm mul ebx \
  684. __asm add eax, ecx \
  685. __asm adc edx, 0 \
  686. __asm add eax, [edi] \
  687. __asm adc edx, 0 \
  688. __asm mov ecx, edx \
  689. __asm stosd
  690. #if defined(MBEDTLS_HAVE_SSE2)
  691. #define EMIT __asm _emit
  692. #define MULADDC_HUIT \
  693. EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
  694. EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
  695. EMIT 0x0F EMIT 0x6E EMIT 0x1F \
  696. EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
  697. EMIT 0x0F EMIT 0x6E EMIT 0x16 \
  698. EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
  699. EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
  700. EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
  701. EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
  702. EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
  703. EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
  704. EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
  705. EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
  706. EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
  707. EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
  708. EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
  709. EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
  710. EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
  711. EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
  712. EMIT 0x0F EMIT 0x7E EMIT 0x0F \
  713. EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
  714. EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
  715. EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
  716. EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
  717. EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
  718. EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
  719. EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
  720. EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
  721. EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
  722. EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
  723. EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
  724. EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
  725. EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
  726. EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
  727. EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
  728. EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
  729. EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
  730. EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
  731. EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
  732. EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
  733. EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
  734. EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
  735. EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
  736. EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
  737. EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
  738. EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
  739. EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
  740. EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
  741. EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
  742. EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
  743. EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
  744. EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
  745. EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
  746. EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
  747. EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
  748. EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
  749. EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
  750. EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
  751. EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
  752. EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
  753. EMIT 0x0F EMIT 0x7E EMIT 0xC9
  754. #define MULADDC_STOP \
  755. EMIT 0x0F EMIT 0x77 \
  756. __asm mov c, ecx \
  757. __asm mov d, edi \
  758. __asm mov s, esi \
  759. #else
  760. #define MULADDC_STOP \
  761. __asm mov c, ecx \
  762. __asm mov d, edi \
  763. __asm mov s, esi \
  764. #endif /* SSE2 */
  765. #endif /* MSVC */
  766. #endif /* MBEDTLS_HAVE_ASM */
  767. #if !defined(MULADDC_CORE)
  768. #if defined(MBEDTLS_HAVE_UDBL)
  769. #define MULADDC_INIT \
  770. { \
  771. mbedtls_t_udbl r; \
  772. mbedtls_mpi_uint r0, r1;
  773. #define MULADDC_CORE \
  774. r = *(s++) * (mbedtls_t_udbl) b; \
  775. r0 = (mbedtls_mpi_uint) r; \
  776. r1 = (mbedtls_mpi_uint)( r >> biL ); \
  777. r0 += c; r1 += (r0 < c); \
  778. r0 += *d; r1 += (r0 < *d); \
  779. c = r1; *(d++) = r0;
  780. #define MULADDC_STOP \
  781. }
  782. #else
  783. #define MULADDC_INIT \
  784. { \
  785. mbedtls_mpi_uint s0, s1, b0, b1; \
  786. mbedtls_mpi_uint r0, r1, rx, ry; \
  787. b0 = ( b << biH ) >> biH; \
  788. b1 = ( b >> biH );
  789. #define MULADDC_CORE \
  790. s0 = ( *s << biH ) >> biH; \
  791. s1 = ( *s >> biH ); s++; \
  792. rx = s0 * b1; r0 = s0 * b0; \
  793. ry = s1 * b0; r1 = s1 * b1; \
  794. r1 += ( rx >> biH ); \
  795. r1 += ( ry >> biH ); \
  796. rx <<= biH; ry <<= biH; \
  797. r0 += rx; r1 += (r0 < rx); \
  798. r0 += ry; r1 += (r0 < ry); \
  799. r0 += c; r1 += (r0 < c); \
  800. r0 += *d; r1 += (r0 < *d); \
  801. c = r1; *(d++) = r0;
  802. #define MULADDC_STOP \
  803. }
  804. #endif /* C (generic) */
  805. #endif /* C (longlong) */
  806. #endif /* bn_mul.h */