TrustedFirmware Git Browser
Code Review Sign In
review.trustedfirmware.org / mirror / mbed-tls / 5c2364c2bac4c0c0c269235aa701ec7e07a1e00e / . / include / polarssl / bn_mul.h
blob: 06324df78bdab1b5253915f6819c45d2ac937589 [file] [log] [blame]
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]1/**
2 * \file bn_mul.h
Paul Bakkere0ccd0a2009-01-04 16:27:10 +0000[diff] [blame]3 *
Paul Bakker37ca75d2011-01-06 12:28:03 +0000[diff] [blame]4 * \brief Multi-precision integer library
5 *
Paul Bakker84f12b72010-07-18 10:13:04 +0000[diff] [blame]6 * Copyright (C) 2006-2010, Brainspark B.V.
Paul Bakkerb96f1542010-07-18 20:36:00 +0000[diff] [blame]7 *
8 * This file is part of PolarSSL (http://www.polarssl.org)
Paul Bakker84f12b72010-07-18 10:13:04 +0000[diff] [blame]9 * Lead Maintainer: Paul Bakker <polarssl_maintainer at polarssl.org>
Paul Bakkerb96f1542010-07-18 20:36:00 +0000[diff] [blame]10 *
Paul Bakker77b385e2009-07-28 17:23:11 +0000[diff] [blame]11 * All rights reserved.
Paul Bakkere0ccd0a2009-01-04 16:27:10 +0000[diff] [blame]12 *
Paul Bakkere0ccd0a2009-01-04 16:27:10 +0000[diff] [blame]13 * 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 +0000[diff] [blame]26 */
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 +0000[diff] [blame]41#ifndef POLARSSL_BN_MUL_H
42#define POLARSSL_BN_MUL_H
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]43
Paul Bakker66219872012-01-22 20:38:13 +0000[diff] [blame]44#include "bignum.h"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]45
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]46#if defined(POLARSSL_HAVE_ASM)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]47
48#if defined(__GNUC__)
49#if defined(__i386__)
50
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]51#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 +0000[diff] [blame]59
Paul Bakker40fe2b02009-07-08 19:49:01 +0000[diff] [blame]60#define MULADDC_CORE \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]61 " \
Paul Bakker40fe2b02009-07-08 19:49:01 +0000[diff] [blame]62 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 +0000[diff] [blame]69 stosl; \
70 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]71
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]72#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]73
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]74#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 +0000[diff] [blame]138
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]139#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 +0000[diff] [blame]151
152#else
153
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]154#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 +0000[diff] [blame]165#endif /* SSE2 */
166#endif /* i386 */
167
168#if defined(__amd64__) || defined (__x86_64__)
169
170#define MULADDC_INIT \
171 asm( "movq %0, %%rsi " :: "m" (s)); \
172 asm( "movq %0, %%rdi " :: "m" (d)); \
173 asm( "movq %0, %%rcx " :: "m" (c)); \
174 asm( "movq %0, %%rbx " :: "m" (b)); \
175 asm( "xorq %r8, %r8 " );
176
177#define MULADDC_CORE \
178 asm( "movq (%rsi),%rax " ); \
179 asm( "mulq %rbx " ); \
180 asm( "addq $8, %rsi " ); \
181 asm( "addq %rcx, %rax " ); \
182 asm( "movq %r8, %rcx " ); \
183 asm( "adcq $0, %rdx " ); \
184 asm( "nop " ); \
185 asm( "addq %rax, (%rdi) " ); \
186 asm( "adcq %rdx, %rcx " ); \
187 asm( "addq $8, %rdi " );
188
189#define MULADDC_STOP \
190 asm( "movq %%rcx, %0 " : "=m" (c)); \
191 asm( "movq %%rdi, %0 " : "=m" (d)); \
192 asm( "movq %%rsi, %0 " : "=m" (s) :: \
193 "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" );
194
195#endif /* AMD64 */
196
197#if defined(__mc68020__) || defined(__mcpu32__)
198
199#define MULADDC_INIT \
200 asm( "movl %0, %%a2 " :: "m" (s)); \
201 asm( "movl %0, %%a3 " :: "m" (d)); \
202 asm( "movl %0, %%d3 " :: "m" (c)); \
203 asm( "movl %0, %%d2 " :: "m" (b)); \
204 asm( "moveq #0, %d0 " );
205
206#define MULADDC_CORE \
207 asm( "movel %a2@+, %d1 " ); \
208 asm( "mulul %d2, %d4:%d1 " ); \
209 asm( "addl %d3, %d1 " ); \
210 asm( "addxl %d0, %d4 " ); \
211 asm( "moveq #0, %d3 " ); \
212 asm( "addl %d1, %a3@+ " ); \
213 asm( "addxl %d4, %d3 " );
214
215#define MULADDC_STOP \
216 asm( "movl %%d3, %0 " : "=m" (c)); \
217 asm( "movl %%a3, %0 " : "=m" (d)); \
218 asm( "movl %%a2, %0 " : "=m" (s) :: \
219 "d0", "d1", "d2", "d3", "d4", "a2", "a3" );
220
221#define MULADDC_HUIT \
222 asm( "movel %a2@+, %d1 " ); \
223 asm( "mulul %d2, %d4:%d1 " ); \
224 asm( "addxl %d3, %d1 " ); \
225 asm( "addxl %d0, %d4 " ); \
226 asm( "addl %d1, %a3@+ " ); \
227 asm( "movel %a2@+, %d1 " ); \
228 asm( "mulul %d2, %d3:%d1 " ); \
229 asm( "addxl %d4, %d1 " ); \
230 asm( "addxl %d0, %d3 " ); \
231 asm( "addl %d1, %a3@+ " ); \
232 asm( "movel %a2@+, %d1 " ); \
233 asm( "mulul %d2, %d4:%d1 " ); \
234 asm( "addxl %d3, %d1 " ); \
235 asm( "addxl %d0, %d4 " ); \
236 asm( "addl %d1, %a3@+ " ); \
237 asm( "movel %a2@+, %d1 " ); \
238 asm( "mulul %d2, %d3:%d1 " ); \
239 asm( "addxl %d4, %d1 " ); \
240 asm( "addxl %d0, %d3 " ); \
241 asm( "addl %d1, %a3@+ " ); \
242 asm( "movel %a2@+, %d1 " ); \
243 asm( "mulul %d2, %d4:%d1 " ); \
244 asm( "addxl %d3, %d1 " ); \
245 asm( "addxl %d0, %d4 " ); \
246 asm( "addl %d1, %a3@+ " ); \
247 asm( "movel %a2@+, %d1 " ); \
248 asm( "mulul %d2, %d3:%d1 " ); \
249 asm( "addxl %d4, %d1 " ); \
250 asm( "addxl %d0, %d3 " ); \
251 asm( "addl %d1, %a3@+ " ); \
252 asm( "movel %a2@+, %d1 " ); \
253 asm( "mulul %d2, %d4:%d1 " ); \
254 asm( "addxl %d3, %d1 " ); \
255 asm( "addxl %d0, %d4 " ); \
256 asm( "addl %d1, %a3@+ " ); \
257 asm( "movel %a2@+, %d1 " ); \
258 asm( "mulul %d2, %d3:%d1 " ); \
259 asm( "addxl %d4, %d1 " ); \
260 asm( "addxl %d0, %d3 " ); \
261 asm( "addl %d1, %a3@+ " ); \
262 asm( "addxl %d0, %d3 " );
263
264#endif /* MC68000 */
265
266#if defined(__powerpc__) || defined(__ppc__)
267#if defined(__powerpc64__) || defined(__ppc64__)
268
269#if defined(__MACH__) && defined(__APPLE__)
270
271#define MULADDC_INIT \
272 asm( "ld r3, %0 " :: "m" (s)); \
273 asm( "ld r4, %0 " :: "m" (d)); \
274 asm( "ld r5, %0 " :: "m" (c)); \
275 asm( "ld r6, %0 " :: "m" (b)); \
276 asm( "addi r3, r3, -8 " ); \
277 asm( "addi r4, r4, -8 " ); \
278 asm( "addic r5, r5, 0 " );
279
280#define MULADDC_CORE \
281 asm( "ldu r7, 8(r3) " ); \
282 asm( "mulld r8, r7, r6 " ); \
283 asm( "mulhdu r9, r7, r6 " ); \
284 asm( "adde r8, r8, r5 " ); \
285 asm( "ld r7, 8(r4) " ); \
286 asm( "addze r5, r9 " ); \
287 asm( "addc r8, r8, r7 " ); \
288 asm( "stdu r8, 8(r4) " );
289
290#define MULADDC_STOP \
291 asm( "addze r5, r5 " ); \
292 asm( "addi r4, r4, 8 " ); \
293 asm( "addi r3, r3, 8 " ); \
294 asm( "std r5, %0 " : "=m" (c)); \
295 asm( "std r4, %0 " : "=m" (d)); \
296 asm( "std r3, %0 " : "=m" (s) :: \
297 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
298
299#else
300
301#define MULADDC_INIT \
302 asm( "ld %%r3, %0 " :: "m" (s)); \
303 asm( "ld %%r4, %0 " :: "m" (d)); \
304 asm( "ld %%r5, %0 " :: "m" (c)); \
305 asm( "ld %%r6, %0 " :: "m" (b)); \
306 asm( "addi %r3, %r3, -8 " ); \
307 asm( "addi %r4, %r4, -8 " ); \
308 asm( "addic %r5, %r5, 0 " );
309
310#define MULADDC_CORE \
311 asm( "ldu %r7, 8(%r3) " ); \
312 asm( "mulld %r8, %r7, %r6 " ); \
313 asm( "mulhdu %r9, %r7, %r6 " ); \
314 asm( "adde %r8, %r8, %r5 " ); \
315 asm( "ld %r7, 8(%r4) " ); \
316 asm( "addze %r5, %r9 " ); \
317 asm( "addc %r8, %r8, %r7 " ); \
318 asm( "stdu %r8, 8(%r4) " );
319
320#define MULADDC_STOP \
321 asm( "addze %r5, %r5 " ); \
322 asm( "addi %r4, %r4, 8 " ); \
323 asm( "addi %r3, %r3, 8 " ); \
324 asm( "std %%r5, %0 " : "=m" (c)); \
325 asm( "std %%r4, %0 " : "=m" (d)); \
326 asm( "std %%r3, %0 " : "=m" (s) :: \
327 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
328
329#endif
330
331#else /* PPC32 */
332
333#if defined(__MACH__) && defined(__APPLE__)
334
335#define MULADDC_INIT \
336 asm( "lwz r3, %0 " :: "m" (s)); \
337 asm( "lwz r4, %0 " :: "m" (d)); \
338 asm( "lwz r5, %0 " :: "m" (c)); \
339 asm( "lwz r6, %0 " :: "m" (b)); \
340 asm( "addi r3, r3, -4 " ); \
341 asm( "addi r4, r4, -4 " ); \
342 asm( "addic r5, r5, 0 " );
343
344#define MULADDC_CORE \
345 asm( "lwzu r7, 4(r3) " ); \
346 asm( "mullw r8, r7, r6 " ); \
347 asm( "mulhwu r9, r7, r6 " ); \
348 asm( "adde r8, r8, r5 " ); \
349 asm( "lwz r7, 4(r4) " ); \
350 asm( "addze r5, r9 " ); \
351 asm( "addc r8, r8, r7 " ); \
352 asm( "stwu r8, 4(r4) " );
353
354#define MULADDC_STOP \
355 asm( "addze r5, r5 " ); \
356 asm( "addi r4, r4, 4 " ); \
357 asm( "addi r3, r3, 4 " ); \
358 asm( "stw r5, %0 " : "=m" (c)); \
359 asm( "stw r4, %0 " : "=m" (d)); \
360 asm( "stw r3, %0 " : "=m" (s) :: \
361 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
362
363#else
364
365#define MULADDC_INIT \
366 asm( "lwz %%r3, %0 " :: "m" (s)); \
367 asm( "lwz %%r4, %0 " :: "m" (d)); \
368 asm( "lwz %%r5, %0 " :: "m" (c)); \
369 asm( "lwz %%r6, %0 " :: "m" (b)); \
370 asm( "addi %r3, %r3, -4 " ); \
371 asm( "addi %r4, %r4, -4 " ); \
372 asm( "addic %r5, %r5, 0 " );
373
374#define MULADDC_CORE \
375 asm( "lwzu %r7, 4(%r3) " ); \
376 asm( "mullw %r8, %r7, %r6 " ); \
377 asm( "mulhwu %r9, %r7, %r6 " ); \
378 asm( "adde %r8, %r8, %r5 " ); \
379 asm( "lwz %r7, 4(%r4) " ); \
380 asm( "addze %r5, %r9 " ); \
381 asm( "addc %r8, %r8, %r7 " ); \
382 asm( "stwu %r8, 4(%r4) " );
383
384#define MULADDC_STOP \
385 asm( "addze %r5, %r5 " ); \
386 asm( "addi %r4, %r4, 4 " ); \
387 asm( "addi %r3, %r3, 4 " ); \
388 asm( "stw %%r5, %0 " : "=m" (c)); \
389 asm( "stw %%r4, %0 " : "=m" (d)); \
390 asm( "stw %%r3, %0 " : "=m" (s) :: \
391 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
392
393#endif
394
395#endif /* PPC32 */
396#endif /* PPC64 */
397
398#if defined(__sparc__)
399
400#define MULADDC_INIT \
401 asm( "ld %0, %%o0 " :: "m" (s)); \
402 asm( "ld %0, %%o1 " :: "m" (d)); \
403 asm( "ld %0, %%o2 " :: "m" (c)); \
404 asm( "ld %0, %%o3 " :: "m" (b));
405
406#define MULADDC_CORE \
407 asm( "ld [%o0], %o4 " ); \
408 asm( "inc 4, %o0 " ); \
409 asm( "ld [%o1], %o5 " ); \
410 asm( "umul %o3, %o4, %o4 " ); \
411 asm( "addcc %o4, %o2, %o4 " ); \
412 asm( "rd %y, %g1 " ); \
413 asm( "addx %g1, 0, %g1 " ); \
414 asm( "addcc %o4, %o5, %o4 " ); \
415 asm( "st %o4, [%o1] " ); \
416 asm( "addx %g1, 0, %o2 " ); \
417 asm( "inc 4, %o1 " );
418
419#define MULADDC_STOP \
420 asm( "st %%o2, %0 " : "=m" (c)); \
421 asm( "st %%o1, %0 " : "=m" (d)); \
422 asm( "st %%o0, %0 " : "=m" (s) :: \
423 "g1", "o0", "o1", "o2", "o3", "o4", "o5" );
424
425#endif /* SPARCv8 */
426
427#if defined(__microblaze__) || defined(microblaze)
428
429#define MULADDC_INIT \
430 asm( "lwi r3, %0 " :: "m" (s)); \
431 asm( "lwi r4, %0 " :: "m" (d)); \
432 asm( "lwi r5, %0 " :: "m" (c)); \
433 asm( "lwi r6, %0 " :: "m" (b)); \
434 asm( "andi r7, r6, 0xffff" ); \
435 asm( "bsrli r6, r6, 16 " );
436
437#define MULADDC_CORE \
438 asm( "lhui r8, r3, 0 " ); \
439 asm( "addi r3, r3, 2 " ); \
440 asm( "lhui r9, r3, 0 " ); \
441 asm( "addi r3, r3, 2 " ); \
442 asm( "mul r10, r9, r6 " ); \
443 asm( "mul r11, r8, r7 " ); \
444 asm( "mul r12, r9, r7 " ); \
445 asm( "mul r13, r8, r6 " ); \
446 asm( "bsrli r8, r10, 16 " ); \
447 asm( "bsrli r9, r11, 16 " ); \
448 asm( "add r13, r13, r8 " ); \
449 asm( "add r13, r13, r9 " ); \
450 asm( "bslli r10, r10, 16 " ); \
451 asm( "bslli r11, r11, 16 " ); \
452 asm( "add r12, r12, r10 " ); \
453 asm( "addc r13, r13, r0 " ); \
454 asm( "add r12, r12, r11 " ); \
455 asm( "addc r13, r13, r0 " ); \
456 asm( "lwi r10, r4, 0 " ); \
457 asm( "add r12, r12, r10 " ); \
458 asm( "addc r13, r13, r0 " ); \
459 asm( "add r12, r12, r5 " ); \
460 asm( "addc r5, r13, r0 " ); \
461 asm( "swi r12, r4, 0 " ); \
462 asm( "addi r4, r4, 4 " );
463
464#define MULADDC_STOP \
465 asm( "swi r5, %0 " : "=m" (c)); \
466 asm( "swi r4, %0 " : "=m" (d)); \
467 asm( "swi r3, %0 " : "=m" (s) :: \
468 "r3", "r4" , "r5" , "r6" , "r7" , "r8" , \
469 "r9", "r10", "r11", "r12", "r13" );
470
471#endif /* MicroBlaze */
472
473#if defined(__tricore__)
474
475#define MULADDC_INIT \
476 asm( "ld.a %%a2, %0 " :: "m" (s)); \
477 asm( "ld.a %%a3, %0 " :: "m" (d)); \
478 asm( "ld.w %%d4, %0 " :: "m" (c)); \
479 asm( "ld.w %%d1, %0 " :: "m" (b)); \
480 asm( "xor %d5, %d5 " );
481
482#define MULADDC_CORE \
483 asm( "ld.w %d0, [%a2+] " ); \
484 asm( "madd.u %e2, %e4, %d0, %d1 " ); \
485 asm( "ld.w %d0, [%a3] " ); \
486 asm( "addx %d2, %d2, %d0 " ); \
487 asm( "addc %d3, %d3, 0 " ); \
488 asm( "mov %d4, %d3 " ); \
489 asm( "st.w [%a3+], %d2 " );
490
491#define MULADDC_STOP \
492 asm( "st.w %0, %%d4 " : "=m" (c)); \
493 asm( "st.a %0, %%a3 " : "=m" (d)); \
494 asm( "st.a %0, %%a2 " : "=m" (s) :: \
495 "d0", "d1", "e2", "d4", "a2", "a3" );
496
497#endif /* TriCore */
498
499#if defined(__arm__)
500
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000[diff] [blame]501#if defined(__thumb__)
502
503#define MULADDC_INIT \
504 asm( "ldr r0, %0 " :: "m" (s)); \
505 asm( "ldr r1, %0 " :: "m" (d)); \
506 asm( "ldr r2, %0 " :: "m" (c)); \
507 asm( "ldr r3, %0 " :: "m" (b)); \
508 asm( "lsr r7, r3, #16 " ); \
509 asm( "mov r9, r7 " ); \
510 asm( "lsl r7, r3, #16 " ); \
511 asm( "lsr r7, r7, #16 " ); \
512 asm( "mov r8, r7 " );
513
514#define MULADDC_CORE \
515 asm( "ldmia r0!, {r6} " ); \
516 asm( "lsr r7, r6, #16 " ); \
517 asm( "lsl r6, r6, #16 " ); \
518 asm( "lsr r6, r6, #16 " ); \
519 asm( "mov r4, r8 " ); \
520 asm( "mul r4, r6 " ); \
521 asm( "mov r3, r9 " ); \
522 asm( "mul r6, r3 " ); \
523 asm( "mov r5, r9 " ); \
524 asm( "mul r5, r7 " ); \
525 asm( "mov r3, r8 " ); \
526 asm( "mul r7, r3 " ); \
527 asm( "lsr r3, r6, #16 " ); \
528 asm( "add r5, r5, r3 " ); \
529 asm( "lsr r3, r7, #16 " ); \
530 asm( "add r5, r5, r3 " ); \
531 asm( "add r4, r4, r2 " ); \
532 asm( "mov r2, #0 " ); \
533 asm( "adc r5, r2 " ); \
534 asm( "lsl r3, r6, #16 " ); \
535 asm( "add r4, r4, r3 " ); \
536 asm( "adc r5, r2 " ); \
537 asm( "lsl r3, r7, #16 " ); \
538 asm( "add r4, r4, r3 " ); \
539 asm( "adc r5, r2 " ); \
540 asm( "ldr r3, [r1] " ); \
541 asm( "add r4, r4, r3 " ); \
542 asm( "adc r2, r5 " ); \
543 asm( "stmia r1!, {r4} " );
544
545#define MULADDC_STOP \
546 asm( "str r2, %0 " : "=m" (c)); \
547 asm( "str r1, %0 " : "=m" (d)); \
548 asm( "str r0, %0 " : "=m" (s) :: \
549 "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
550
551#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000[diff] [blame]552
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]553#define MULADDC_INIT \
554 asm( "ldr r0, %0 " :: "m" (s)); \
555 asm( "ldr r1, %0 " :: "m" (d)); \
556 asm( "ldr r2, %0 " :: "m" (c)); \
557 asm( "ldr r3, %0 " :: "m" (b));
558
559#define MULADDC_CORE \
560 asm( "ldr r4, [r0], #4 " ); \
561 asm( "mov r5, #0 " ); \
562 asm( "ldr r6, [r1] " ); \
563 asm( "umlal r2, r5, r3, r4 " ); \
564 asm( "adds r7, r6, r2 " ); \
565 asm( "adc r2, r5, #0 " ); \
566 asm( "str r7, [r1], #4 " );
567
568#define MULADDC_STOP \
569 asm( "str r2, %0 " : "=m" (c)); \
570 asm( "str r1, %0 " : "=m" (d)); \
571 asm( "str r0, %0 " : "=m" (s) :: \
572 "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7" );
573
Paul Bakkera2713a32011-11-18 12:47:23 +0000[diff] [blame]574#endif /* Thumb */
575
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]576#endif /* ARMv3 */
577
578#if defined(__alpha__)
579
580#define MULADDC_INIT \
581 asm( "ldq $1, %0 " :: "m" (s)); \
582 asm( "ldq $2, %0 " :: "m" (d)); \
583 asm( "ldq $3, %0 " :: "m" (c)); \
584 asm( "ldq $4, %0 " :: "m" (b));
585
586#define MULADDC_CORE \
587 asm( "ldq $6, 0($1) " ); \
588 asm( "addq $1, 8, $1 " ); \
589 asm( "mulq $6, $4, $7 " ); \
590 asm( "umulh $6, $4, $6 " ); \
591 asm( "addq $7, $3, $7 " ); \
592 asm( "cmpult $7, $3, $3 " ); \
593 asm( "ldq $5, 0($2) " ); \
594 asm( "addq $7, $5, $7 " ); \
595 asm( "cmpult $7, $5, $5 " ); \
596 asm( "stq $7, 0($2) " ); \
597 asm( "addq $2, 8, $2 " ); \
598 asm( "addq $6, $3, $3 " ); \
599 asm( "addq $5, $3, $3 " );
600
601#define MULADDC_STOP \
602 asm( "stq $3, %0 " : "=m" (c)); \
603 asm( "stq $2, %0 " : "=m" (d)); \
604 asm( "stq $1, %0 " : "=m" (s) :: \
605 "$1", "$2", "$3", "$4", "$5", "$6", "$7" );
606
607#endif /* Alpha */
608
609#if defined(__mips__)
610
611#define MULADDC_INIT \
612 asm( "lw $10, %0 " :: "m" (s)); \
613 asm( "lw $11, %0 " :: "m" (d)); \
614 asm( "lw $12, %0 " :: "m" (c)); \
615 asm( "lw $13, %0 " :: "m" (b));
616
617#define MULADDC_CORE \
618 asm( "lw $14, 0($10) " ); \
619 asm( "multu $13, $14 " ); \
620 asm( "addi $10, $10, 4 " ); \
621 asm( "mflo $14 " ); \
622 asm( "mfhi $9 " ); \
623 asm( "addu $14, $12, $14 " ); \
624 asm( "lw $15, 0($11) " ); \
625 asm( "sltu $12, $14, $12 " ); \
626 asm( "addu $15, $14, $15 " ); \
627 asm( "sltu $14, $15, $14 " ); \
628 asm( "addu $12, $12, $9 " ); \
629 asm( "sw $15, 0($11) " ); \
630 asm( "addu $12, $12, $14 " ); \
631 asm( "addi $11, $11, 4 " );
632
633#define MULADDC_STOP \
634 asm( "sw $12, %0 " : "=m" (c)); \
635 asm( "sw $11, %0 " : "=m" (d)); \
636 asm( "sw $10, %0 " : "=m" (s) :: \
637 "$9", "$10", "$11", "$12", "$13", "$14", "$15" );
638
639#endif /* MIPS */
640#endif /* GNUC */
641
642#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
643
644#define MULADDC_INIT \
645 __asm mov esi, s \
646 __asm mov edi, d \
647 __asm mov ecx, c \
648 __asm mov ebx, b
649
650#define MULADDC_CORE \
651 __asm lodsd \
652 __asm mul ebx \
653 __asm add eax, ecx \
654 __asm adc edx, 0 \
655 __asm add eax, [edi] \
656 __asm adc edx, 0 \
657 __asm mov ecx, edx \
658 __asm stosd
659
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]660#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]661
662#define EMIT __asm _emit
663
664#define MULADDC_HUIT \
665 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
666 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
667 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
668 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
669 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
670 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
671 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
672 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
673 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
674 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
675 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
676 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
677 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
678 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
679 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
680 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
681 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
682 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
683 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
684 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
685 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
686 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
687 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
688 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
689 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
690 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
691 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
692 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
693 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
694 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
695 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
696 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
697 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
698 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
699 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
700 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
701 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
702 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
703 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
704 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
705 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
706 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
707 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
708 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
709 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
710 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
711 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
712 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
713 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
714 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
715 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
716 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
717 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
718 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
719 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
720 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
721 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
722 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
723 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
724 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
725 EMIT 0x0F EMIT 0x7E EMIT 0xC9
726
727#define MULADDC_STOP \
728 EMIT 0x0F EMIT 0x77 \
729 __asm mov c, ecx \
730 __asm mov d, edi \
731 __asm mov s, esi \
732
733#else
734
735#define MULADDC_STOP \
736 __asm mov c, ecx \
737 __asm mov d, edi \
738 __asm mov s, esi \
739
740#endif /* SSE2 */
741#endif /* MSVC */
742
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]743#endif /* POLARSSL_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]744
745#if !defined(MULADDC_CORE)
Paul Bakker5c2364c2012-10-01 14:41:15 +0000[diff] [blame^]746#if defined(POLARSSL_HAVE_INT64)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]747
748#define MULADDC_INIT \
749{ \
Paul Bakkera755ca12011-04-24 09:11:17 +0000[diff] [blame]750 t_udbl r; \
751 t_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]752
753#define MULADDC_CORE \
Paul Bakker5c2364c2012-10-01 14:41:15 +0000[diff] [blame^]754 r = *(s++) * (t_udbl) b; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]755 r0 = r; \
756 r1 = r >> biL; \
757 r0 += c; r1 += (r0 < c); \
758 r0 += *d; r1 += (r0 < *d); \
759 c = r1; *(d++) = r0;
760
761#define MULADDC_STOP \
762}
763
764#else
765#define MULADDC_INIT \
766{ \
Paul Bakkera755ca12011-04-24 09:11:17 +0000[diff] [blame]767 t_uint s0, s1, b0, b1; \
768 t_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]769 b0 = ( b << biH ) >> biH; \
770 b1 = ( b >> biH );
771
772#define MULADDC_CORE \
773 s0 = ( *s << biH ) >> biH; \
774 s1 = ( *s >> biH ); s++; \
775 rx = s0 * b1; r0 = s0 * b0; \
776 ry = s1 * b0; r1 = s1 * b1; \
777 r1 += ( rx >> biH ); \
778 r1 += ( ry >> biH ); \
779 rx <<= biH; ry <<= biH; \
780 r0 += rx; r1 += (r0 < rx); \
781 r0 += ry; r1 += (r0 < ry); \
782 r0 += c; r1 += (r0 < c); \
783 r0 += *d; r1 += (r0 < *d); \
784 c = r1; *(d++) = r0;
785
786#define MULADDC_STOP \
787}
788
789#endif /* C (generic) */
790#endif /* C (longlong) */
791
792#endif /* bn_mul.h */