blob: e59b5a03f89d9281b5e57744e022d899c833c9e1 [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
288#if defined(__powerpc__) || defined(__ppc__)
289#if defined(__powerpc64__) || defined(__ppc64__)
290
291#if defined(__MACH__) && defined(__APPLE__)
292
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100293#define MULADDC_INIT \
294 asm( \
295 " \
296 ld r3, %3; \
297 ld r4, %4; \
298 ld r5, %5; \
299 ld r6, %6; \
300 addi r3, r3, -8; \
301 addi r4, r4, -8; \
302 addic r5, r5, 0; \
303 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000304
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100305#define MULADDC_CORE \
306 " \
307 ldu r7, 8(r3); \
308 mulld r8, r7, r6; \
309 mulhdu r9, r7, r6; \
310 adde r8, r8, r5; \
311 ld r7, 8(r4); \
312 addze r5, r9; \
313 addc r8, r8, r7; \
314 stdu r8, 8(r4); \
315 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000316
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100317#define MULADDC_STOP \
318 " \
319 addze r5, r5; \
320 addi r4, r4, 8; \
321 addi r3, r3, 8; \
322 std r5, %0; \
323 std r4, %1; \
324 std r3, %2; \
325 " \
326 : "=m" (c), "=m" (d), "=m" (s) \
327 : "m" (s), "m" (d), "m" (c), "m" (b) \
328 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
329 );
330
Paul Bakker5121ce52009-01-03 21:22:43 +0000331
332#else
333
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100334#define MULADDC_INIT \
335 asm( \
336 " \
337 ld %%r3, %3; \
338 ld %%r4, %4; \
339 ld %%r5, %5; \
340 ld %%r6, %6; \
341 addi %%r3, %%r3, -8; \
342 addi %%r4, %%r4, -8; \
343 addic %%r5, %%r5, 0; \
344 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000345
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100346#define MULADDC_CORE \
347 " \
348 ldu %%r7, 8(%%r3); \
349 mulld %%r8, %%r7, %%r6; \
350 mulhdu %%r9, %%r7, %%r6; \
351 adde %%r8, %%r8, %%r5; \
352 ld %%r7, 8(%%r4); \
353 addze %%r5, %%r9; \
354 addc %%r8, %%r8, %%r7; \
355 stdu %%r8, 8(%%r4); \
356 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000357
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100358#define MULADDC_STOP \
359 " \
360 addze %%r5, %%r5; \
361 addi %%r4, %%r4, 8; \
362 addi %%r3, %%r3, 8; \
363 std %%r5, %0; \
364 std %%r4, %1; \
365 std %%r3, %2; \
366 " \
367 : "=m" (c), "=m" (d), "=m" (s) \
368 : "m" (s), "m" (d), "m" (c), "m" (b) \
369 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
370 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000371
372#endif
373
374#else /* PPC32 */
375
376#if defined(__MACH__) && defined(__APPLE__)
377
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100378#define MULADDC_INIT \
379 asm( \
380 " \
381 lwz r3, %3; \
382 lwz r4, %4; \
383 lwz r5, %5; \
384 lwz r6, %6; \
385 addi r3, r3, -4; \
386 addi r4, r4, -4; \
387 addic r5, r5, 0; \
388 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000389
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100390#define MULADDC_CORE \
391 " \
392 lwzu r7, 4(r3); \
393 mullw r8, r7, r6; \
394 mulhwu r9, r7, r6; \
395 adde r8, r8, r5; \
396 lwz r7, 4(r4); \
397 addze r5, r9; \
398 addc r8, r8, r7; \
399 stwu r8, 4(r4); \
400 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000401
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100402#define MULADDC_STOP \
403 " \
404 addze r5, r5; \
405 addi r4, r4, 4; \
406 addi r3, r3, 4; \
407 stw r5, %0; \
408 stw r4, %1; \
409 stw r3, %2; \
410 " \
411 : "=m" (c), "=m" (d), "=m" (s) \
412 : "m" (s), "m" (d), "m" (c), "m" (b) \
413 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
414 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000415
416#else
417
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100418#define MULADDC_INIT \
419 asm( \
420 " \
421 lwz %%r3, %3; \
422 lwz %%r4, %4; \
423 lwz %%r5, %5; \
424 lwz %%r6, %6; \
425 addi %%r3, %%r3, -4; \
426 addi %%r4, %%r4, -4; \
427 addic %%r5, %%r5, 0; \
428 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000429
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100430#define MULADDC_CORE \
431 " \
432 lwzu %%r7, 4(%%r3); \
433 mullw %%r8, %%r7, %%r6; \
434 mulhwu %%r9, %%r7, %%r6; \
435 adde %%r8, %%r8, %%r5; \
436 lwz %%r7, 4(%%r4); \
437 addze %%r5, %%r9; \
438 addc %%r8, %%r8, %%r7; \
439 stwu %%r8, 4(%%r4); \
440 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000441
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100442#define MULADDC_STOP \
443 " \
444 addze %%r5, %%r5; \
445 addi %%r4, %%r4, 4; \
446 addi %%r3, %%r3, 4; \
447 stw %%r5, %0; \
448 stw %%r4, %1; \
449 stw %%r3, %2; \
450 " \
451 : "=m" (c), "=m" (d), "=m" (s) \
452 : "m" (s), "m" (d), "m" (c), "m" (b) \
453 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
454 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000455
456#endif
457
458#endif /* PPC32 */
459#endif /* PPC64 */
460
Paul Bakker4f024b72012-10-30 07:29:57 +0000461#if defined(__sparc__) && defined(__sparc64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000462
463#define MULADDC_INIT \
Paul Bakker4f024b72012-10-30 07:29:57 +0000464 asm( \
465 " \
466 ldx %3, %%o0; \
467 ldx %4, %%o1; \
468 ld %5, %%o2; \
469 ld %6, %%o3; \
470 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000471
472#define MULADDC_CORE \
Paul Bakker4f024b72012-10-30 07:29:57 +0000473 " \
474 ld [%%o0], %%o4; \
475 inc 4, %%o0; \
476 ld [%%o1], %%o5; \
477 umul %%o3, %%o4, %%o4; \
478 addcc %%o4, %%o2, %%o4; \
479 rd %%y, %%g1; \
480 addx %%g1, 0, %%g1; \
481 addcc %%o4, %%o5, %%o4; \
482 st %%o4, [%%o1]; \
483 addx %%g1, 0, %%o2; \
484 inc 4, %%o1; \
485 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000486
487#define MULADDC_STOP \
Paul Bakker4f024b72012-10-30 07:29:57 +0000488 " \
489 st %%o2, %0; \
490 stx %%o1, %1; \
491 stx %%o0, %2; \
492 " \
493 : "=m" (c), "=m" (d), "=m" (s) \
494 : "m" (s), "m" (d), "m" (c), "m" (b) \
495 : "g1", "o0", "o1", "o2", "o3", "o4", \
496 "o5" \
497 );
498#endif /* SPARCv9 */
499
500#if defined(__sparc__) && !defined(__sparc64__)
501
502#define MULADDC_INIT \
503 asm( \
504 " \
505 ld %3, %%o0; \
506 ld %4, %%o1; \
507 ld %5, %%o2; \
508 ld %6, %%o3; \
509 "
510
511#define MULADDC_CORE \
512 " \
513 ld [%%o0], %%o4; \
514 inc 4, %%o0; \
515 ld [%%o1], %%o5; \
516 umul %%o3, %%o4, %%o4; \
517 addcc %%o4, %%o2, %%o4; \
518 rd %%y, %%g1; \
519 addx %%g1, 0, %%g1; \
520 addcc %%o4, %%o5, %%o4; \
521 st %%o4, [%%o1]; \
522 addx %%g1, 0, %%o2; \
523 inc 4, %%o1; \
524 "
525
526#define MULADDC_STOP \
527 " \
528 st %%o2, %0; \
529 st %%o1, %1; \
530 st %%o0, %2; \
531 " \
532 : "=m" (c), "=m" (d), "=m" (s) \
533 : "m" (s), "m" (d), "m" (c), "m" (b) \
534 : "g1", "o0", "o1", "o2", "o3", "o4", \
535 "o5" \
536 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000537
538#endif /* SPARCv8 */
539
540#if defined(__microblaze__) || defined(microblaze)
541
542#define MULADDC_INIT \
543 asm( "lwi r3, %0 " :: "m" (s)); \
544 asm( "lwi r4, %0 " :: "m" (d)); \
545 asm( "lwi r5, %0 " :: "m" (c)); \
546 asm( "lwi r6, %0 " :: "m" (b)); \
547 asm( "andi r7, r6, 0xffff" ); \
548 asm( "bsrli r6, r6, 16 " );
549
550#define MULADDC_CORE \
551 asm( "lhui r8, r3, 0 " ); \
552 asm( "addi r3, r3, 2 " ); \
553 asm( "lhui r9, r3, 0 " ); \
554 asm( "addi r3, r3, 2 " ); \
555 asm( "mul r10, r9, r6 " ); \
556 asm( "mul r11, r8, r7 " ); \
557 asm( "mul r12, r9, r7 " ); \
558 asm( "mul r13, r8, r6 " ); \
559 asm( "bsrli r8, r10, 16 " ); \
560 asm( "bsrli r9, r11, 16 " ); \
561 asm( "add r13, r13, r8 " ); \
562 asm( "add r13, r13, r9 " ); \
563 asm( "bslli r10, r10, 16 " ); \
564 asm( "bslli r11, r11, 16 " ); \
565 asm( "add r12, r12, r10 " ); \
566 asm( "addc r13, r13, r0 " ); \
567 asm( "add r12, r12, r11 " ); \
568 asm( "addc r13, r13, r0 " ); \
569 asm( "lwi r10, r4, 0 " ); \
570 asm( "add r12, r12, r10 " ); \
571 asm( "addc r13, r13, r0 " ); \
572 asm( "add r12, r12, r5 " ); \
573 asm( "addc r5, r13, r0 " ); \
574 asm( "swi r12, r4, 0 " ); \
575 asm( "addi r4, r4, 4 " );
576
577#define MULADDC_STOP \
578 asm( "swi r5, %0 " : "=m" (c)); \
579 asm( "swi r4, %0 " : "=m" (d)); \
580 asm( "swi r3, %0 " : "=m" (s) :: \
581 "r3", "r4" , "r5" , "r6" , "r7" , "r8" , \
582 "r9", "r10", "r11", "r12", "r13" );
583
584#endif /* MicroBlaze */
585
586#if defined(__tricore__)
587
588#define MULADDC_INIT \
589 asm( "ld.a %%a2, %0 " :: "m" (s)); \
590 asm( "ld.a %%a3, %0 " :: "m" (d)); \
591 asm( "ld.w %%d4, %0 " :: "m" (c)); \
592 asm( "ld.w %%d1, %0 " :: "m" (b)); \
593 asm( "xor %d5, %d5 " );
594
595#define MULADDC_CORE \
596 asm( "ld.w %d0, [%a2+] " ); \
597 asm( "madd.u %e2, %e4, %d0, %d1 " ); \
598 asm( "ld.w %d0, [%a3] " ); \
599 asm( "addx %d2, %d2, %d0 " ); \
600 asm( "addc %d3, %d3, 0 " ); \
601 asm( "mov %d4, %d3 " ); \
602 asm( "st.w [%a3+], %d2 " );
603
604#define MULADDC_STOP \
605 asm( "st.w %0, %%d4 " : "=m" (c)); \
606 asm( "st.a %0, %%a3 " : "=m" (d)); \
607 asm( "st.a %0, %%a2 " : "=m" (s) :: \
608 "d0", "d1", "e2", "d4", "a2", "a3" );
609
610#endif /* TriCore */
611
612#if defined(__arm__)
613
Paul Bakkerfc4f46f2013-06-24 19:23:56 +0200614#if defined(__thumb__) && !defined(__thumb2__)
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000615
616#define MULADDC_INIT \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100617 asm( \
618 " \
619 ldr r0, %3; \
620 ldr r1, %4; \
621 ldr r2, %5; \
622 ldr r3, %6; \
623 lsr r7, r3, #16; \
624 mov r9, r7; \
625 lsl r7, r3, #16; \
626 lsr r7, r7, #16; \
627 mov r8, r7; \
628 "
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000629
630#define MULADDC_CORE \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100631 " \
632 ldmia r0!, {r6}; \
633 lsr r7, r6, #16; \
634 lsl r6, r6, #16; \
635 lsr r6, r6, #16; \
636 mov r4, r8; \
637 mul r4, r6; \
638 mov r3, r9; \
639 mul r6, r3; \
640 mov r5, r9; \
641 mul r5, r7; \
642 mov r3, r8; \
643 mul r7, r3; \
644 lsr r3, r6, #16; \
645 add r5, r5, r3; \
646 lsr r3, r7, #16; \
647 add r5, r5, r3; \
648 add r4, r4, r2; \
649 mov r2, #0; \
650 adc r5, r2; \
651 lsl r3, r6, #16; \
652 add r4, r4, r3; \
653 adc r5, r2; \
654 lsl r3, r7, #16; \
655 add r4, r4, r3; \
656 adc r5, r2; \
657 ldr r3, [r1]; \
658 add r4, r4, r3; \
659 adc r2, r5; \
660 stmia r1!, {r4}; \
661 "
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000662
663#define MULADDC_STOP \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100664 " \
665 str r2, %0; \
666 str r1, %1; \
667 str r0, %2; \
668 " \
669 : "=m" (c), "=m" (d), "=m" (s) \
670 : "m" (s), "m" (d), "m" (c), "m" (b) \
671 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200672 "r6", "r7", "r8", "r9", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100673 );
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000674
675#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000676
Paul Bakker5121ce52009-01-03 21:22:43 +0000677#define MULADDC_INIT \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100678 asm( \
679 " \
680 ldr r0, %3; \
681 ldr r1, %4; \
682 ldr r2, %5; \
683 ldr r3, %6; \
684 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000685
686#define MULADDC_CORE \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100687 " \
688 ldr r4, [r0], #4; \
689 mov r5, #0; \
690 ldr r6, [r1]; \
691 umlal r2, r5, r3, r4; \
692 adds r7, r6, r2; \
693 adc r2, r5, #0; \
694 str r7, [r1], #4; \
695 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000696
697#define MULADDC_STOP \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100698 " \
699 str r2, %0; \
700 str r1, %1; \
701 str r0, %2; \
702 " \
703 : "=m" (c), "=m" (d), "=m" (s) \
704 : "m" (s), "m" (d), "m" (c), "m" (b) \
705 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200706 "r6", "r7", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100707 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000708
Paul Bakkera2713a32011-11-18 12:47:23 +0000709#endif /* Thumb */
710
Paul Bakker5121ce52009-01-03 21:22:43 +0000711#endif /* ARMv3 */
712
713#if defined(__alpha__)
714
715#define MULADDC_INIT \
716 asm( "ldq $1, %0 " :: "m" (s)); \
717 asm( "ldq $2, %0 " :: "m" (d)); \
718 asm( "ldq $3, %0 " :: "m" (c)); \
719 asm( "ldq $4, %0 " :: "m" (b));
720
721#define MULADDC_CORE \
722 asm( "ldq $6, 0($1) " ); \
723 asm( "addq $1, 8, $1 " ); \
724 asm( "mulq $6, $4, $7 " ); \
725 asm( "umulh $6, $4, $6 " ); \
726 asm( "addq $7, $3, $7 " ); \
727 asm( "cmpult $7, $3, $3 " ); \
728 asm( "ldq $5, 0($2) " ); \
729 asm( "addq $7, $5, $7 " ); \
730 asm( "cmpult $7, $5, $5 " ); \
731 asm( "stq $7, 0($2) " ); \
732 asm( "addq $2, 8, $2 " ); \
733 asm( "addq $6, $3, $3 " ); \
734 asm( "addq $5, $3, $3 " );
735
736#define MULADDC_STOP \
737 asm( "stq $3, %0 " : "=m" (c)); \
738 asm( "stq $2, %0 " : "=m" (d)); \
739 asm( "stq $1, %0 " : "=m" (s) :: \
740 "$1", "$2", "$3", "$4", "$5", "$6", "$7" );
741
742#endif /* Alpha */
743
744#if defined(__mips__)
745
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100746#define MULADDC_INIT \
747 asm( \
748 " \
749 lw $10, %3; \
750 lw $11, %4; \
751 lw $12, %5; \
752 lw $13, %6; \
753 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000754
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100755#define MULADDC_CORE \
756 " \
757 lw $14, 0($10); \
758 multu $13, $14; \
759 addi $10, $10, 4; \
760 mflo $14; \
761 mfhi $9; \
762 addu $14, $12, $14; \
763 lw $15, 0($11); \
764 sltu $12, $14, $12; \
765 addu $15, $14, $15; \
766 sltu $14, $15, $14; \
767 addu $12, $12, $9; \
768 sw $15, 0($11); \
769 addu $12, $12, $14; \
770 addi $11, $11, 4; \
771 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000772
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100773#define MULADDC_STOP \
774 " \
775 sw $12, %0; \
776 sw $11, %1; \
777 sw $10, %2; \
778 " \
779 : "=m" (c), "=m" (d), "=m" (s) \
780 : "m" (s), "m" (d), "m" (c), "m" (b) \
781 : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
782 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000783
784#endif /* MIPS */
785#endif /* GNUC */
786
787#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
788
789#define MULADDC_INIT \
790 __asm mov esi, s \
791 __asm mov edi, d \
792 __asm mov ecx, c \
793 __asm mov ebx, b
794
795#define MULADDC_CORE \
796 __asm lodsd \
797 __asm mul ebx \
798 __asm add eax, ecx \
799 __asm adc edx, 0 \
800 __asm add eax, [edi] \
801 __asm adc edx, 0 \
802 __asm mov ecx, edx \
803 __asm stosd
804
Paul Bakker40e46942009-01-03 21:51:57 +0000805#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000806
807#define EMIT __asm _emit
808
809#define MULADDC_HUIT \
810 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
811 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
812 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
813 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
814 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
815 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
816 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
817 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
818 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
819 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
820 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
821 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
822 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
823 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
824 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
825 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
826 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
827 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
828 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
829 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
830 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
831 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
832 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
833 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
834 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
835 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
836 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
837 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
838 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
839 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
840 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
841 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
842 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
843 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
844 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
845 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
846 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
847 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
848 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
849 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
850 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
851 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
852 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
853 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
854 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
855 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
856 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
857 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
858 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
859 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
860 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
861 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
862 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
863 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
864 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
865 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
866 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
867 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
868 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
869 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
870 EMIT 0x0F EMIT 0x7E EMIT 0xC9
871
872#define MULADDC_STOP \
873 EMIT 0x0F EMIT 0x77 \
874 __asm mov c, ecx \
875 __asm mov d, edi \
876 __asm mov s, esi \
877
878#else
879
880#define MULADDC_STOP \
881 __asm mov c, ecx \
882 __asm mov d, edi \
883 __asm mov s, esi \
884
885#endif /* SSE2 */
886#endif /* MSVC */
887
Paul Bakker40e46942009-01-03 21:51:57 +0000888#endif /* POLARSSL_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000889
890#if !defined(MULADDC_CORE)
Paul Bakker0fd018e2012-10-23 12:44:47 +0000891#if defined(POLARSSL_HAVE_UDBL)
Paul Bakker5121ce52009-01-03 21:22:43 +0000892
893#define MULADDC_INIT \
894{ \
Paul Bakkera755ca12011-04-24 09:11:17 +0000895 t_udbl r; \
896 t_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000897
898#define MULADDC_CORE \
Paul Bakker5c2364c2012-10-01 14:41:15 +0000899 r = *(s++) * (t_udbl) b; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000900 r0 = r; \
901 r1 = r >> biL; \
902 r0 += c; r1 += (r0 < c); \
903 r0 += *d; r1 += (r0 < *d); \
904 c = r1; *(d++) = r0;
905
906#define MULADDC_STOP \
907}
908
909#else
910#define MULADDC_INIT \
911{ \
Paul Bakkera755ca12011-04-24 09:11:17 +0000912 t_uint s0, s1, b0, b1; \
913 t_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000914 b0 = ( b << biH ) >> biH; \
915 b1 = ( b >> biH );
916
917#define MULADDC_CORE \
918 s0 = ( *s << biH ) >> biH; \
919 s1 = ( *s >> biH ); s++; \
920 rx = s0 * b1; r0 = s0 * b0; \
921 ry = s1 * b0; r1 = s1 * b1; \
922 r1 += ( rx >> biH ); \
923 r1 += ( ry >> biH ); \
924 rx <<= biH; ry <<= biH; \
925 r0 += rx; r1 += (r0 < rx); \
926 r0 += ry; r1 += (r0 < ry); \
927 r0 += c; r1 += (r0 < c); \
928 r0 += *d; r1 += (r0 < *d); \
929 c = r1; *(d++) = r0;
930
931#define MULADDC_STOP \
932}
933
934#endif /* C (generic) */
935#endif /* C (longlong) */
936
937#endif /* bn_mul.h */