blob: 345893a29f91fcc100cb1f29b57aae95d78f41d3 [file] [log] [blame]
Paul Bakker5121ce52009-01-03 21:22:43 +00001/**
2 * \file bn_mul.h
Paul Bakkere0ccd0a2009-01-04 16:27:10 +00003 *
Paul Bakker37ca75d2011-01-06 12:28:03 +00004 * \brief Multi-precision integer library
5 *
Paul Bakker84f12b72010-07-18 10:13:04 +00006 * Copyright (C) 2006-2010, Brainspark B.V.
Paul Bakkerb96f1542010-07-18 20:36:00 +00007 *
8 * This file is part of PolarSSL (http://www.polarssl.org)
Paul Bakker84f12b72010-07-18 10:13:04 +00009 * Lead Maintainer: Paul Bakker <polarssl_maintainer at polarssl.org>
Paul Bakkerb96f1542010-07-18 20:36:00 +000010 *
Paul Bakker77b385e2009-07-28 17:23:11 +000011 * All rights reserved.
Paul Bakkere0ccd0a2009-01-04 16:27:10 +000012 *
Paul Bakkere0ccd0a2009-01-04 16:27:10 +000013 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
17 *
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
22 *
23 * You should have received a copy of the GNU General Public License along
24 * with this program; if not, write to the Free Software Foundation, Inc.,
25 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
Paul Bakker5121ce52009-01-03 21:22:43 +000026 */
27/*
28 * Multiply source vector [s] with b, add result
29 * to destination vector [d] and set carry c.
30 *
31 * Currently supports:
32 *
33 * . IA-32 (386+) . AMD64 / EM64T
34 * . IA-32 (SSE2) . Motorola 68000
35 * . PowerPC, 32-bit . MicroBlaze
36 * . PowerPC, 64-bit . TriCore
37 * . SPARC v8 . ARM v3+
38 * . Alpha . MIPS32
39 * . C, longlong . C, generic
40 */
Paul Bakker40e46942009-01-03 21:51:57 +000041#ifndef POLARSSL_BN_MUL_H
42#define POLARSSL_BN_MUL_H
Paul Bakker5121ce52009-01-03 21:22:43 +000043
Paul Bakker66219872012-01-22 20:38:13 +000044#include "bignum.h"
Paul Bakker5121ce52009-01-03 21:22:43 +000045
Paul Bakker40e46942009-01-03 21:51:57 +000046#if defined(POLARSSL_HAVE_ASM)
Paul Bakker5121ce52009-01-03 21:22:43 +000047
48#if defined(__GNUC__)
49#if defined(__i386__)
50
Paul Bakkerc89cf7c2009-07-19 21:37:39 +000051#define MULADDC_INIT \
52 asm( " \
53 movl %%ebx, %0; \
54 movl %5, %%esi; \
55 movl %6, %%edi; \
56 movl %7, %%ecx; \
57 movl %8, %%ebx; \
58 "
Paul Bakker5121ce52009-01-03 21:22:43 +000059
Paul Bakker40fe2b02009-07-08 19:49:01 +000060#define MULADDC_CORE \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +000061 " \
Paul Bakker40fe2b02009-07-08 19:49:01 +000062 lodsl; \
63 mull %%ebx; \
64 addl %%ecx, %%eax; \
65 adcl $0, %%edx; \
66 addl (%%edi), %%eax; \
67 adcl $0, %%edx; \
68 movl %%edx, %%ecx; \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +000069 stosl; \
70 "
Paul Bakker5121ce52009-01-03 21:22:43 +000071
Paul Bakker40e46942009-01-03 21:51:57 +000072#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +000073
Paul Bakkerc89cf7c2009-07-19 21:37:39 +000074#define MULADDC_HUIT \
75 " \
76 movd %%ecx, %%mm1; \
77 movd %%ebx, %%mm0; \
78 movd (%%edi), %%mm3; \
79 paddq %%mm3, %%mm1; \
80 movd (%%esi), %%mm2; \
81 pmuludq %%mm0, %%mm2; \
82 movd 4(%%esi), %%mm4; \
83 pmuludq %%mm0, %%mm4; \
84 movd 8(%%esi), %%mm6; \
85 pmuludq %%mm0, %%mm6; \
86 movd 12(%%esi), %%mm7; \
87 pmuludq %%mm0, %%mm7; \
88 paddq %%mm2, %%mm1; \
89 movd 4(%%edi), %%mm3; \
90 paddq %%mm4, %%mm3; \
91 movd 8(%%edi), %%mm5; \
92 paddq %%mm6, %%mm5; \
93 movd 12(%%edi), %%mm4; \
94 paddq %%mm4, %%mm7; \
95 movd %%mm1, (%%edi); \
96 movd 16(%%esi), %%mm2; \
97 pmuludq %%mm0, %%mm2; \
98 psrlq $32, %%mm1; \
99 movd 20(%%esi), %%mm4; \
100 pmuludq %%mm0, %%mm4; \
101 paddq %%mm3, %%mm1; \
102 movd 24(%%esi), %%mm6; \
103 pmuludq %%mm0, %%mm6; \
104 movd %%mm1, 4(%%edi); \
105 psrlq $32, %%mm1; \
106 movd 28(%%esi), %%mm3; \
107 pmuludq %%mm0, %%mm3; \
108 paddq %%mm5, %%mm1; \
109 movd 16(%%edi), %%mm5; \
110 paddq %%mm5, %%mm2; \
111 movd %%mm1, 8(%%edi); \
112 psrlq $32, %%mm1; \
113 paddq %%mm7, %%mm1; \
114 movd 20(%%edi), %%mm5; \
115 paddq %%mm5, %%mm4; \
116 movd %%mm1, 12(%%edi); \
117 psrlq $32, %%mm1; \
118 paddq %%mm2, %%mm1; \
119 movd 24(%%edi), %%mm5; \
120 paddq %%mm5, %%mm6; \
121 movd %%mm1, 16(%%edi); \
122 psrlq $32, %%mm1; \
123 paddq %%mm4, %%mm1; \
124 movd 28(%%edi), %%mm5; \
125 paddq %%mm5, %%mm3; \
126 movd %%mm1, 20(%%edi); \
127 psrlq $32, %%mm1; \
128 paddq %%mm6, %%mm1; \
129 movd %%mm1, 24(%%edi); \
130 psrlq $32, %%mm1; \
131 paddq %%mm3, %%mm1; \
132 movd %%mm1, 28(%%edi); \
133 addl $32, %%edi; \
134 addl $32, %%esi; \
135 psrlq $32, %%mm1; \
136 movd %%mm1, %%ecx; \
137 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000138
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000139#define MULADDC_STOP \
140 " \
141 emms; \
142 movl %4, %%ebx; \
143 movl %%ecx, %1; \
144 movl %%edi, %2; \
145 movl %%esi, %3; \
146 " \
147 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
148 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
149 : "eax", "ecx", "edx", "esi", "edi" \
150 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000151
152#else
153
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000154#define MULADDC_STOP \
155 " \
156 movl %4, %%ebx; \
157 movl %%ecx, %1; \
158 movl %%edi, %2; \
159 movl %%esi, %3; \
160 " \
161 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
162 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
163 : "eax", "ecx", "edx", "esi", "edi" \
164 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000165#endif /* SSE2 */
166#endif /* i386 */
167
168#if defined(__amd64__) || defined (__x86_64__)
169
Manuel Pégourié-Gonnarddef018d2014-01-07 17:50:46 +0100170#define MULADDC_INIT \
171 asm( \
172 " \
173 movq %3, %%rsi; \
174 movq %4, %%rdi; \
175 movq %5, %%rcx; \
176 movq %6, %%rbx; \
177 xorq %%r8, %%r8; \
178 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000179
Manuel Pégourié-Gonnarddef018d2014-01-07 17:50:46 +0100180#define MULADDC_CORE \
181 " \
182 movq (%%rsi), %%rax; \
183 mulq %%rbx; \
184 addq $8, %%rsi; \
185 addq %%rcx, %%rax; \
186 movq %%r8, %%rcx; \
187 adcq $0, %%rdx; \
188 nop; \
189 addq %%rax, (%%rdi); \
190 adcq %%rdx, %%rcx; \
191 addq $8, %%rdi; \
192 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000193
Manuel Pégourié-Gonnarddef018d2014-01-07 17:50:46 +0100194#define MULADDC_STOP \
195 " \
196 movq %%rcx, %0; \
197 movq %%rdi, %1; \
198 movq %%rsi, %2; \
199 " \
200 : "=m" (c), "=m" (d), "=m" (s) \
201 : "m" (s), "m" (d), "m" (c), "m" (b) \
202 : "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" \
203 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000204
205#endif /* AMD64 */
206
207#if defined(__mc68020__) || defined(__mcpu32__)
208
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100209#define MULADDC_INIT \
210 asm( \
211 " \
212 movl %3, %%a2; \
213 movl %4, %%a3; \
214 movl %5, %%d3; \
215 movl %6, %%d2; \
216 moveq #0, %%d0; \
217 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000218
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100219#define MULADDC_CORE \
220 " \
221 movel %%a2@+, %%d1; \
222 mulul %%d2, %%d4:%%d1; \
223 addl %%d3, %%d1; \
224 addxl %%d0, %%d4; \
225 moveq #0, %%d3; \
226 addl %%d1, %%a3@+; \
227 addxl %%d4, %%d3; \
228 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000229
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100230#define MULADDC_STOP \
231 " \
232 movl %%d3, %0; \
233 movl %%a3, %1; \
234 movl %%a2, %2; \
235 " \
236 : "=m" (c), "=m" (d), "=m" (s) \
237 : "m" (s), "m" (d), "m" (c), "m" (b) \
238 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
239 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000240
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100241#define MULADDC_HUIT \
242 " \
243 movel %%a2@+, %%d1; \
244 mulul %%d2, %%d4:%%d1; \
245 addxl %%d3, %%d1; \
246 addxl %%d0, %%d4; \
247 addl %%d1, %%a3@+; \
248 movel %%a2@+, %%d1; \
249 mulul %%d2, %%d3:%%d1; \
250 addxl %%d4, %%d1; \
251 addxl %%d0, %%d3; \
252 addl %%d1, %%a3@+; \
253 movel %%a2@+, %%d1; \
254 mulul %%d2, %%d4:%%d1; \
255 addxl %%d3, %%d1; \
256 addxl %%d0, %%d4; \
257 addl %%d1, %%a3@+; \
258 movel %%a2@+, %%d1; \
259 mulul %%d2, %%d3:%%d1; \
260 addxl %%d4, %%d1; \
261 addxl %%d0, %%d3; \
262 addl %%d1, %%a3@+; \
263 movel %%a2@+, %%d1; \
264 mulul %%d2, %%d4:%%d1; \
265 addxl %%d3, %%d1; \
266 addxl %%d0, %%d4; \
267 addl %%d1, %%a3@+; \
268 movel %%a2@+, %%d1; \
269 mulul %%d2, %%d3:%%d1; \
270 addxl %%d4, %%d1; \
271 addxl %%d0, %%d3; \
272 addl %%d1, %%a3@+; \
273 movel %%a2@+, %%d1; \
274 mulul %%d2, %%d4:%%d1; \
275 addxl %%d3, %%d1; \
276 addxl %%d0, %%d4; \
277 addl %%d1, %%a3@+; \
278 movel %%a2@+, %%d1; \
279 mulul %%d2, %%d3:%%d1; \
280 addxl %%d4, %%d1; \
281 addxl %%d0, %%d3; \
282 addl %%d1, %%a3@+; \
283 addxl %%d0, %%d3; \
284 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000285
286#endif /* MC68000 */
287
Paul Bakker5121ce52009-01-03 21:22:43 +0000288#if defined(__powerpc64__) || defined(__ppc64__)
289
290#if defined(__MACH__) && defined(__APPLE__)
291
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100292#define MULADDC_INIT \
293 asm( \
294 " \
295 ld r3, %3; \
296 ld r4, %4; \
297 ld r5, %5; \
298 ld r6, %6; \
299 addi r3, r3, -8; \
300 addi r4, r4, -8; \
301 addic r5, r5, 0; \
302 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000303
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100304#define MULADDC_CORE \
305 " \
306 ldu r7, 8(r3); \
307 mulld r8, r7, r6; \
308 mulhdu r9, r7, r6; \
309 adde r8, r8, r5; \
310 ld r7, 8(r4); \
311 addze r5, r9; \
312 addc r8, r8, r7; \
313 stdu r8, 8(r4); \
314 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000315
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100316#define MULADDC_STOP \
317 " \
318 addze r5, r5; \
319 addi r4, r4, 8; \
320 addi r3, r3, 8; \
321 std r5, %0; \
322 std r4, %1; \
323 std r3, %2; \
324 " \
325 : "=m" (c), "=m" (d), "=m" (s) \
326 : "m" (s), "m" (d), "m" (c), "m" (b) \
327 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
328 );
329
Paul Bakker5121ce52009-01-03 21:22:43 +0000330
Paul Bakker9af723c2014-05-01 13:03:14 +0200331#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000332
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100333#define MULADDC_INIT \
334 asm( \
335 " \
336 ld %%r3, %3; \
337 ld %%r4, %4; \
338 ld %%r5, %5; \
339 ld %%r6, %6; \
340 addi %%r3, %%r3, -8; \
341 addi %%r4, %%r4, -8; \
342 addic %%r5, %%r5, 0; \
343 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000344
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100345#define MULADDC_CORE \
346 " \
347 ldu %%r7, 8(%%r3); \
348 mulld %%r8, %%r7, %%r6; \
349 mulhdu %%r9, %%r7, %%r6; \
350 adde %%r8, %%r8, %%r5; \
351 ld %%r7, 8(%%r4); \
352 addze %%r5, %%r9; \
353 addc %%r8, %%r8, %%r7; \
354 stdu %%r8, 8(%%r4); \
355 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000356
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100357#define MULADDC_STOP \
358 " \
359 addze %%r5, %%r5; \
360 addi %%r4, %%r4, 8; \
361 addi %%r3, %%r3, 8; \
362 std %%r5, %0; \
363 std %%r4, %1; \
364 std %%r3, %2; \
365 " \
366 : "=m" (c), "=m" (d), "=m" (s) \
367 : "m" (s), "m" (d), "m" (c), "m" (b) \
368 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
369 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000370
Paul Bakker9af723c2014-05-01 13:03:14 +0200371#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000372
Barry K. Nathan35e7cb92014-05-05 23:26:13 -0700373#elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000374
375#if defined(__MACH__) && defined(__APPLE__)
376
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100377#define MULADDC_INIT \
378 asm( \
379 " \
380 lwz r3, %3; \
381 lwz r4, %4; \
382 lwz r5, %5; \
383 lwz r6, %6; \
384 addi r3, r3, -4; \
385 addi r4, r4, -4; \
386 addic r5, r5, 0; \
387 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000388
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100389#define MULADDC_CORE \
390 " \
391 lwzu r7, 4(r3); \
392 mullw r8, r7, r6; \
393 mulhwu r9, r7, r6; \
394 adde r8, r8, r5; \
395 lwz r7, 4(r4); \
396 addze r5, r9; \
397 addc r8, r8, r7; \
398 stwu r8, 4(r4); \
399 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000400
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100401#define MULADDC_STOP \
402 " \
403 addze r5, r5; \
404 addi r4, r4, 4; \
405 addi r3, r3, 4; \
406 stw r5, %0; \
407 stw r4, %1; \
408 stw r3, %2; \
409 " \
410 : "=m" (c), "=m" (d), "=m" (s) \
411 : "m" (s), "m" (d), "m" (c), "m" (b) \
412 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
413 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000414
Paul Bakker9af723c2014-05-01 13:03:14 +0200415#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000416
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100417#define MULADDC_INIT \
418 asm( \
419 " \
420 lwz %%r3, %3; \
421 lwz %%r4, %4; \
422 lwz %%r5, %5; \
423 lwz %%r6, %6; \
424 addi %%r3, %%r3, -4; \
425 addi %%r4, %%r4, -4; \
426 addic %%r5, %%r5, 0; \
427 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000428
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100429#define MULADDC_CORE \
430 " \
431 lwzu %%r7, 4(%%r3); \
432 mullw %%r8, %%r7, %%r6; \
433 mulhwu %%r9, %%r7, %%r6; \
434 adde %%r8, %%r8, %%r5; \
435 lwz %%r7, 4(%%r4); \
436 addze %%r5, %%r9; \
437 addc %%r8, %%r8, %%r7; \
438 stwu %%r8, 4(%%r4); \
439 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000440
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100441#define MULADDC_STOP \
442 " \
443 addze %%r5, %%r5; \
444 addi %%r4, %%r4, 4; \
445 addi %%r3, %%r3, 4; \
446 stw %%r5, %0; \
447 stw %%r4, %1; \
448 stw %%r3, %2; \
449 " \
450 : "=m" (c), "=m" (d), "=m" (s) \
451 : "m" (s), "m" (d), "m" (c), "m" (b) \
452 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
453 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000454
Paul Bakker9af723c2014-05-01 13:03:14 +0200455#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000456
457#endif /* PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000458
Paul Bakker4f024b72012-10-30 07:29:57 +0000459#if defined(__sparc__) && defined(__sparc64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000460
461#define MULADDC_INIT \
Paul Bakker4f024b72012-10-30 07:29:57 +0000462 asm( \
463 " \
464 ldx %3, %%o0; \
465 ldx %4, %%o1; \
466 ld %5, %%o2; \
467 ld %6, %%o3; \
468 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000469
470#define MULADDC_CORE \
Paul Bakker4f024b72012-10-30 07:29:57 +0000471 " \
472 ld [%%o0], %%o4; \
473 inc 4, %%o0; \
474 ld [%%o1], %%o5; \
475 umul %%o3, %%o4, %%o4; \
476 addcc %%o4, %%o2, %%o4; \
477 rd %%y, %%g1; \
478 addx %%g1, 0, %%g1; \
479 addcc %%o4, %%o5, %%o4; \
480 st %%o4, [%%o1]; \
481 addx %%g1, 0, %%o2; \
482 inc 4, %%o1; \
483 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000484
485#define MULADDC_STOP \
Paul Bakker4f024b72012-10-30 07:29:57 +0000486 " \
487 st %%o2, %0; \
488 stx %%o1, %1; \
489 stx %%o0, %2; \
490 " \
491 : "=m" (c), "=m" (d), "=m" (s) \
492 : "m" (s), "m" (d), "m" (c), "m" (b) \
493 : "g1", "o0", "o1", "o2", "o3", "o4", \
494 "o5" \
495 );
496#endif /* SPARCv9 */
497
498#if defined(__sparc__) && !defined(__sparc64__)
499
500#define MULADDC_INIT \
501 asm( \
502 " \
503 ld %3, %%o0; \
504 ld %4, %%o1; \
505 ld %5, %%o2; \
506 ld %6, %%o3; \
507 "
508
509#define MULADDC_CORE \
510 " \
511 ld [%%o0], %%o4; \
512 inc 4, %%o0; \
513 ld [%%o1], %%o5; \
514 umul %%o3, %%o4, %%o4; \
515 addcc %%o4, %%o2, %%o4; \
516 rd %%y, %%g1; \
517 addx %%g1, 0, %%g1; \
518 addcc %%o4, %%o5, %%o4; \
519 st %%o4, [%%o1]; \
520 addx %%g1, 0, %%o2; \
521 inc 4, %%o1; \
522 "
523
524#define MULADDC_STOP \
525 " \
526 st %%o2, %0; \
527 st %%o1, %1; \
528 st %%o0, %2; \
529 " \
530 : "=m" (c), "=m" (d), "=m" (s) \
531 : "m" (s), "m" (d), "m" (c), "m" (b) \
532 : "g1", "o0", "o1", "o2", "o3", "o4", \
533 "o5" \
534 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000535
536#endif /* SPARCv8 */
537
538#if defined(__microblaze__) || defined(microblaze)
539
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100540#define MULADDC_INIT \
541 asm( \
542 " \
543 lwi r3, %3; \
544 lwi r4, %4; \
545 lwi r5, %5; \
546 lwi r6, %6; \
547 andi r7, r6, 0xffff; \
548 bsrli r6, r6, 16; \
549 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000550
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100551#define MULADDC_CORE \
552 " \
553 lhui r8, r3, 0; \
554 addi r3, r3, 2; \
555 lhui r9, r3, 0; \
556 addi r3, r3, 2; \
557 mul r10, r9, r6; \
558 mul r11, r8, r7; \
559 mul r12, r9, r7; \
560 mul r13, r8, r6; \
561 bsrli r8, r10, 16; \
562 bsrli r9, r11, 16; \
563 add r13, r13, r8; \
564 add r13, r13, r9; \
565 bslli r10, r10, 16; \
566 bslli r11, r11, 16; \
567 add r12, r12, r10; \
568 addc r13, r13, r0; \
569 add r12, r12, r11; \
570 addc r13, r13, r0; \
571 lwi r10, r4, 0; \
572 add r12, r12, r10; \
573 addc r13, r13, r0; \
574 add r12, r12, r5; \
575 addc r5, r13, r0; \
576 swi r12, r4, 0; \
577 addi r4, r4, 4; \
578 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000579
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100580#define MULADDC_STOP \
581 " \
582 swi r5, %0; \
583 swi r4, %1; \
584 swi r3, %2; \
585 " \
586 : "=m" (c), "=m" (d), "=m" (s) \
587 : "m" (s), "m" (d), "m" (c), "m" (b) \
588 : "r3", "r4" "r5", "r6", "r7", "r8", \
589 "r9", "r10", "r11", "r12", "r13" \
590 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000591
592#endif /* MicroBlaze */
593
594#if defined(__tricore__)
595
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100596#define MULADDC_INIT \
597 asm( \
598 " \
599 ld.a %%a2, %3; \
600 ld.a %%a3, %4; \
601 ld.w %%d4, %5; \
602 ld.w %%d1, %6; \
603 xor %%d5, %%d5; \
604 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000605
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100606#define MULADDC_CORE \
607 " \
608 ld.w %%d0, [%%a2+]; \
609 madd.u %%e2, %%e4, %%d0, %%d1; \
610 ld.w %%d0, [%%a3]; \
611 addx %%d2, %%d2, %%d0; \
612 addc %%d3, %%d3, 0; \
613 mov %%d4, %%d3; \
614 st.w [%%a3+], %%d2; \
615 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000616
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100617#define MULADDC_STOP \
618 " \
619 st.w %0, %%d4; \
620 st.a %1, %%a3; \
621 st.a %2, %%a2; \
622 " \
623 : "=m" (c), "=m" (d), "=m" (s) \
624 : "m" (s), "m" (d), "m" (c), "m" (b) \
625 : "d0", "d1", "e2", "d4", "a2", "a3" \
626 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000627
628#endif /* TriCore */
629
630#if defined(__arm__)
631
Paul Bakkerfc4f46f2013-06-24 19:23:56 +0200632#if defined(__thumb__) && !defined(__thumb2__)
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000633
634#define MULADDC_INIT \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100635 asm( \
636 " \
637 ldr r0, %3; \
638 ldr r1, %4; \
639 ldr r2, %5; \
640 ldr r3, %6; \
641 lsr r7, r3, #16; \
642 mov r9, r7; \
643 lsl r7, r3, #16; \
644 lsr r7, r7, #16; \
645 mov r8, r7; \
646 "
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000647
648#define MULADDC_CORE \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100649 " \
650 ldmia r0!, {r6}; \
651 lsr r7, r6, #16; \
652 lsl r6, r6, #16; \
653 lsr r6, r6, #16; \
654 mov r4, r8; \
655 mul r4, r6; \
656 mov r3, r9; \
657 mul r6, r3; \
658 mov r5, r9; \
659 mul r5, r7; \
660 mov r3, r8; \
661 mul r7, r3; \
662 lsr r3, r6, #16; \
663 add r5, r5, r3; \
664 lsr r3, r7, #16; \
665 add r5, r5, r3; \
666 add r4, r4, r2; \
667 mov r2, #0; \
668 adc r5, r2; \
669 lsl r3, r6, #16; \
670 add r4, r4, r3; \
671 adc r5, r2; \
672 lsl r3, r7, #16; \
673 add r4, r4, r3; \
674 adc r5, r2; \
675 ldr r3, [r1]; \
676 add r4, r4, r3; \
677 adc r2, r5; \
678 stmia r1!, {r4}; \
679 "
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000680
681#define MULADDC_STOP \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100682 " \
683 str r2, %0; \
684 str r1, %1; \
685 str r0, %2; \
686 " \
687 : "=m" (c), "=m" (d), "=m" (s) \
688 : "m" (s), "m" (d), "m" (c), "m" (b) \
689 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200690 "r6", "r7", "r8", "r9", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100691 );
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000692
693#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000694
Paul Bakker5121ce52009-01-03 21:22:43 +0000695#define MULADDC_INIT \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100696 asm( \
697 " \
698 ldr r0, %3; \
699 ldr r1, %4; \
700 ldr r2, %5; \
701 ldr r3, %6; \
702 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000703
704#define MULADDC_CORE \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100705 " \
706 ldr r4, [r0], #4; \
707 mov r5, #0; \
708 ldr r6, [r1]; \
709 umlal r2, r5, r3, r4; \
710 adds r7, r6, r2; \
711 adc r2, r5, #0; \
712 str r7, [r1], #4; \
713 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000714
715#define MULADDC_STOP \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100716 " \
717 str r2, %0; \
718 str r1, %1; \
719 str r0, %2; \
720 " \
721 : "=m" (c), "=m" (d), "=m" (s) \
722 : "m" (s), "m" (d), "m" (c), "m" (b) \
723 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200724 "r6", "r7", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100725 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000726
Paul Bakkera2713a32011-11-18 12:47:23 +0000727#endif /* Thumb */
728
Paul Bakker5121ce52009-01-03 21:22:43 +0000729#endif /* ARMv3 */
730
731#if defined(__alpha__)
732
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100733#define MULADDC_INIT \
734 asm( \
735 " \
736 ldq $1, %3; \
737 ldq $2, %4; \
738 ldq $3, %5; \
739 ldq $4, %6; \
740 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000741
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100742#define MULADDC_CORE \
743 " \
744 ldq $6, 0($1); \
745 addq $1, 8, $1; \
746 mulq $6, $4, $7; \
747 umulh $6, $4, $6; \
748 addq $7, $3, $7; \
749 cmpult $7, $3, $3; \
750 ldq $5, 0($2); \
751 addq $7, $5, $7; \
752 cmpult $7, $5, $5; \
753 stq $7, 0($2); \
754 addq $2, 8, $2; \
755 addq $6, $3, $3; \
756 addq $5, $3, $3; \
757 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000758
759#define MULADDC_STOP \
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100760 " \
761 stq $3, %0; \
762 stq $2, %1; \
763 stq $1, %2; \
764 " \
765 : "=m" (c), "=m" (d), "=m" (s) \
766 : "m" (s), "m" (d), "m" (c), "m" (b) \
767 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
768 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000769#endif /* Alpha */
770
Alex Wilson73491422014-03-06 00:04:09 +1000771#if defined(__mips__) && !defined(__mips64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000772
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100773#define MULADDC_INIT \
774 asm( \
775 " \
776 lw $10, %3; \
777 lw $11, %4; \
778 lw $12, %5; \
779 lw $13, %6; \
780 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000781
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100782#define MULADDC_CORE \
783 " \
784 lw $14, 0($10); \
785 multu $13, $14; \
786 addi $10, $10, 4; \
787 mflo $14; \
788 mfhi $9; \
789 addu $14, $12, $14; \
790 lw $15, 0($11); \
791 sltu $12, $14, $12; \
792 addu $15, $14, $15; \
793 sltu $14, $15, $14; \
794 addu $12, $12, $9; \
795 sw $15, 0($11); \
796 addu $12, $12, $14; \
797 addi $11, $11, 4; \
798 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000799
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100800#define MULADDC_STOP \
801 " \
802 sw $12, %0; \
803 sw $11, %1; \
804 sw $10, %2; \
805 " \
806 : "=m" (c), "=m" (d), "=m" (s) \
807 : "m" (s), "m" (d), "m" (c), "m" (b) \
808 : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
809 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000810
811#endif /* MIPS */
812#endif /* GNUC */
813
814#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
815
816#define MULADDC_INIT \
817 __asm mov esi, s \
818 __asm mov edi, d \
819 __asm mov ecx, c \
820 __asm mov ebx, b
821
822#define MULADDC_CORE \
823 __asm lodsd \
824 __asm mul ebx \
825 __asm add eax, ecx \
826 __asm adc edx, 0 \
827 __asm add eax, [edi] \
828 __asm adc edx, 0 \
829 __asm mov ecx, edx \
830 __asm stosd
831
Paul Bakker40e46942009-01-03 21:51:57 +0000832#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000833
834#define EMIT __asm _emit
835
836#define MULADDC_HUIT \
837 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
838 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
839 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
840 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
841 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
842 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
843 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
844 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
845 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
846 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
847 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
848 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
849 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
850 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
851 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
852 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
853 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
854 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
855 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
856 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
857 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
858 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
859 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
860 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
861 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
862 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
863 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
864 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
865 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
866 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
867 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
868 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
869 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
870 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
871 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
872 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
873 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
874 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
875 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
876 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
877 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
878 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
879 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
880 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
881 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
882 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
883 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
884 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
885 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
886 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
887 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
888 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
889 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
890 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
891 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
892 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
893 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
894 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
895 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
896 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
897 EMIT 0x0F EMIT 0x7E EMIT 0xC9
898
899#define MULADDC_STOP \
900 EMIT 0x0F EMIT 0x77 \
901 __asm mov c, ecx \
902 __asm mov d, edi \
903 __asm mov s, esi \
904
905#else
906
907#define MULADDC_STOP \
908 __asm mov c, ecx \
909 __asm mov d, edi \
910 __asm mov s, esi \
911
912#endif /* SSE2 */
913#endif /* MSVC */
914
Paul Bakker40e46942009-01-03 21:51:57 +0000915#endif /* POLARSSL_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000916
917#if !defined(MULADDC_CORE)
Paul Bakker0fd018e2012-10-23 12:44:47 +0000918#if defined(POLARSSL_HAVE_UDBL)
Paul Bakker5121ce52009-01-03 21:22:43 +0000919
920#define MULADDC_INIT \
921{ \
Paul Bakkera755ca12011-04-24 09:11:17 +0000922 t_udbl r; \
923 t_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000924
925#define MULADDC_CORE \
Paul Bakker5c2364c2012-10-01 14:41:15 +0000926 r = *(s++) * (t_udbl) b; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000927 r0 = r; \
928 r1 = r >> biL; \
929 r0 += c; r1 += (r0 < c); \
930 r0 += *d; r1 += (r0 < *d); \
931 c = r1; *(d++) = r0;
932
933#define MULADDC_STOP \
934}
935
936#else
937#define MULADDC_INIT \
938{ \
Paul Bakkera755ca12011-04-24 09:11:17 +0000939 t_uint s0, s1, b0, b1; \
940 t_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000941 b0 = ( b << biH ) >> biH; \
942 b1 = ( b >> biH );
943
944#define MULADDC_CORE \
945 s0 = ( *s << biH ) >> biH; \
946 s1 = ( *s >> biH ); s++; \
947 rx = s0 * b1; r0 = s0 * b0; \
948 ry = s1 * b0; r1 = s1 * b1; \
949 r1 += ( rx >> biH ); \
950 r1 += ( ry >> biH ); \
951 rx <<= biH; ry <<= biH; \
952 r0 += rx; r1 += (r0 < rx); \
953 r0 += ry; r1 += (r0 < ry); \
954 r0 += c; r1 += (r0 < c); \
955 r0 += *d; r1 += (r0 < *d); \
956 c = r1; *(d++) = r0;
957
958#define MULADDC_STOP \
959}
960
961#endif /* C (generic) */
962#endif /* C (longlong) */
963
964#endif /* bn_mul.h */