TrustedFirmware Git Browser
Code Review Sign In
review.trustedfirmware.org / mirror / mbed-tls / 77b385e91a0e48c4c263b2db1baa4887c7f61860 / . / include / polarssl / bn_mul.h
blob: 3eca5cc33070797ec4d99f41c37943f73290e573 [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 Bakker77b385e2009-07-28 17:23:11 +0000[diff] [blame^]4 * Copyright (C) 2006-2009, Paul Bakker <polarssl_maintainer at polarssl.org>
5 * All rights reserved.
Paul Bakkere0ccd0a2009-01-04 16:27:10 +0000[diff] [blame]6 *
Paul Bakker77b385e2009-07-28 17:23:11 +0000[diff] [blame^]7 * Joined copyright on original XySSL code with: Christophe Devine
Paul Bakkere0ccd0a2009-01-04 16:27:10 +0000[diff] [blame]8 *
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License as published by
11 * the Free Software Foundation; either version 2 of the License, or
12 * (at your option) any later version.
13 *
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
18 *
19 * You should have received a copy of the GNU General Public License along
20 * with this program; if not, write to the Free Software Foundation, Inc.,
21 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]22 */
23/*
24 * Multiply source vector [s] with b, add result
25 * to destination vector [d] and set carry c.
26 *
27 * Currently supports:
28 *
29 * . IA-32 (386+) . AMD64 / EM64T
30 * . IA-32 (SSE2) . Motorola 68000
31 * . PowerPC, 32-bit . MicroBlaze
32 * . PowerPC, 64-bit . TriCore
33 * . SPARC v8 . ARM v3+
34 * . Alpha . MIPS32
35 * . C, longlong . C, generic
36 */
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]37#ifndef POLARSSL_BN_MUL_H
38#define POLARSSL_BN_MUL_H
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]39
Paul Bakker8e831ed2009-01-03 21:24:11 +0000[diff] [blame]40#include "polarssl/config.h"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]41
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]42#if defined(POLARSSL_HAVE_ASM)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]43
44#if defined(__GNUC__)
45#if defined(__i386__)
46
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]47#define MULADDC_INIT \
48 asm( " \
49 movl %%ebx, %0; \
50 movl %5, %%esi; \
51 movl %6, %%edi; \
52 movl %7, %%ecx; \
53 movl %8, %%ebx; \
54 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]55
Paul Bakker40fe2b02009-07-08 19:49:01 +0000[diff] [blame]56#define MULADDC_CORE \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]57 " \
Paul Bakker40fe2b02009-07-08 19:49:01 +0000[diff] [blame]58 lodsl; \
59 mull %%ebx; \
60 addl %%ecx, %%eax; \
61 adcl $0, %%edx; \
62 addl (%%edi), %%eax; \
63 adcl $0, %%edx; \
64 movl %%edx, %%ecx; \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]65 stosl; \
66 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]67
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]68#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]69
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]70#define MULADDC_HUIT \
71 " \
72 movd %%ecx, %%mm1; \
73 movd %%ebx, %%mm0; \
74 movd (%%edi), %%mm3; \
75 paddq %%mm3, %%mm1; \
76 movd (%%esi), %%mm2; \
77 pmuludq %%mm0, %%mm2; \
78 movd 4(%%esi), %%mm4; \
79 pmuludq %%mm0, %%mm4; \
80 movd 8(%%esi), %%mm6; \
81 pmuludq %%mm0, %%mm6; \
82 movd 12(%%esi), %%mm7; \
83 pmuludq %%mm0, %%mm7; \
84 paddq %%mm2, %%mm1; \
85 movd 4(%%edi), %%mm3; \
86 paddq %%mm4, %%mm3; \
87 movd 8(%%edi), %%mm5; \
88 paddq %%mm6, %%mm5; \
89 movd 12(%%edi), %%mm4; \
90 paddq %%mm4, %%mm7; \
91 movd %%mm1, (%%edi); \
92 movd 16(%%esi), %%mm2; \
93 pmuludq %%mm0, %%mm2; \
94 psrlq $32, %%mm1; \
95 movd 20(%%esi), %%mm4; \
96 pmuludq %%mm0, %%mm4; \
97 paddq %%mm3, %%mm1; \
98 movd 24(%%esi), %%mm6; \
99 pmuludq %%mm0, %%mm6; \
100 movd %%mm1, 4(%%edi); \
101 psrlq $32, %%mm1; \
102 movd 28(%%esi), %%mm3; \
103 pmuludq %%mm0, %%mm3; \
104 paddq %%mm5, %%mm1; \
105 movd 16(%%edi), %%mm5; \
106 paddq %%mm5, %%mm2; \
107 movd %%mm1, 8(%%edi); \
108 psrlq $32, %%mm1; \
109 paddq %%mm7, %%mm1; \
110 movd 20(%%edi), %%mm5; \
111 paddq %%mm5, %%mm4; \
112 movd %%mm1, 12(%%edi); \
113 psrlq $32, %%mm1; \
114 paddq %%mm2, %%mm1; \
115 movd 24(%%edi), %%mm5; \
116 paddq %%mm5, %%mm6; \
117 movd %%mm1, 16(%%edi); \
118 psrlq $32, %%mm1; \
119 paddq %%mm4, %%mm1; \
120 movd 28(%%edi), %%mm5; \
121 paddq %%mm5, %%mm3; \
122 movd %%mm1, 20(%%edi); \
123 psrlq $32, %%mm1; \
124 paddq %%mm6, %%mm1; \
125 movd %%mm1, 24(%%edi); \
126 psrlq $32, %%mm1; \
127 paddq %%mm3, %%mm1; \
128 movd %%mm1, 28(%%edi); \
129 addl $32, %%edi; \
130 addl $32, %%esi; \
131 psrlq $32, %%mm1; \
132 movd %%mm1, %%ecx; \
133 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]134
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]135#define MULADDC_STOP \
136 " \
137 emms; \
138 movl %4, %%ebx; \
139 movl %%ecx, %1; \
140 movl %%edi, %2; \
141 movl %%esi, %3; \
142 " \
143 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
144 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
145 : "eax", "ecx", "edx", "esi", "edi" \
146 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]147
148#else
149
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]150#define MULADDC_STOP \
151 " \
152 movl %4, %%ebx; \
153 movl %%ecx, %1; \
154 movl %%edi, %2; \
155 movl %%esi, %3; \
156 " \
157 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
158 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
159 : "eax", "ecx", "edx", "esi", "edi" \
160 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]161#endif /* SSE2 */
162#endif /* i386 */
163
164#if defined(__amd64__) || defined (__x86_64__)
165
166#define MULADDC_INIT \
167 asm( "movq %0, %%rsi " :: "m" (s)); \
168 asm( "movq %0, %%rdi " :: "m" (d)); \
169 asm( "movq %0, %%rcx " :: "m" (c)); \
170 asm( "movq %0, %%rbx " :: "m" (b)); \
171 asm( "xorq %r8, %r8 " );
172
173#define MULADDC_CORE \
174 asm( "movq (%rsi),%rax " ); \
175 asm( "mulq %rbx " ); \
176 asm( "addq $8, %rsi " ); \
177 asm( "addq %rcx, %rax " ); \
178 asm( "movq %r8, %rcx " ); \
179 asm( "adcq $0, %rdx " ); \
180 asm( "nop " ); \
181 asm( "addq %rax, (%rdi) " ); \
182 asm( "adcq %rdx, %rcx " ); \
183 asm( "addq $8, %rdi " );
184
185#define MULADDC_STOP \
186 asm( "movq %%rcx, %0 " : "=m" (c)); \
187 asm( "movq %%rdi, %0 " : "=m" (d)); \
188 asm( "movq %%rsi, %0 " : "=m" (s) :: \
189 "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" );
190
191#endif /* AMD64 */
192
193#if defined(__mc68020__) || defined(__mcpu32__)
194
195#define MULADDC_INIT \
196 asm( "movl %0, %%a2 " :: "m" (s)); \
197 asm( "movl %0, %%a3 " :: "m" (d)); \
198 asm( "movl %0, %%d3 " :: "m" (c)); \
199 asm( "movl %0, %%d2 " :: "m" (b)); \
200 asm( "moveq #0, %d0 " );
201
202#define MULADDC_CORE \
203 asm( "movel %a2@+, %d1 " ); \
204 asm( "mulul %d2, %d4:%d1 " ); \
205 asm( "addl %d3, %d1 " ); \
206 asm( "addxl %d0, %d4 " ); \
207 asm( "moveq #0, %d3 " ); \
208 asm( "addl %d1, %a3@+ " ); \
209 asm( "addxl %d4, %d3 " );
210
211#define MULADDC_STOP \
212 asm( "movl %%d3, %0 " : "=m" (c)); \
213 asm( "movl %%a3, %0 " : "=m" (d)); \
214 asm( "movl %%a2, %0 " : "=m" (s) :: \
215 "d0", "d1", "d2", "d3", "d4", "a2", "a3" );
216
217#define MULADDC_HUIT \
218 asm( "movel %a2@+, %d1 " ); \
219 asm( "mulul %d2, %d4:%d1 " ); \
220 asm( "addxl %d3, %d1 " ); \
221 asm( "addxl %d0, %d4 " ); \
222 asm( "addl %d1, %a3@+ " ); \
223 asm( "movel %a2@+, %d1 " ); \
224 asm( "mulul %d2, %d3:%d1 " ); \
225 asm( "addxl %d4, %d1 " ); \
226 asm( "addxl %d0, %d3 " ); \
227 asm( "addl %d1, %a3@+ " ); \
228 asm( "movel %a2@+, %d1 " ); \
229 asm( "mulul %d2, %d4:%d1 " ); \
230 asm( "addxl %d3, %d1 " ); \
231 asm( "addxl %d0, %d4 " ); \
232 asm( "addl %d1, %a3@+ " ); \
233 asm( "movel %a2@+, %d1 " ); \
234 asm( "mulul %d2, %d3:%d1 " ); \
235 asm( "addxl %d4, %d1 " ); \
236 asm( "addxl %d0, %d3 " ); \
237 asm( "addl %d1, %a3@+ " ); \
238 asm( "movel %a2@+, %d1 " ); \
239 asm( "mulul %d2, %d4:%d1 " ); \
240 asm( "addxl %d3, %d1 " ); \
241 asm( "addxl %d0, %d4 " ); \
242 asm( "addl %d1, %a3@+ " ); \
243 asm( "movel %a2@+, %d1 " ); \
244 asm( "mulul %d2, %d3:%d1 " ); \
245 asm( "addxl %d4, %d1 " ); \
246 asm( "addxl %d0, %d3 " ); \
247 asm( "addl %d1, %a3@+ " ); \
248 asm( "movel %a2@+, %d1 " ); \
249 asm( "mulul %d2, %d4:%d1 " ); \
250 asm( "addxl %d3, %d1 " ); \
251 asm( "addxl %d0, %d4 " ); \
252 asm( "addl %d1, %a3@+ " ); \
253 asm( "movel %a2@+, %d1 " ); \
254 asm( "mulul %d2, %d3:%d1 " ); \
255 asm( "addxl %d4, %d1 " ); \
256 asm( "addxl %d0, %d3 " ); \
257 asm( "addl %d1, %a3@+ " ); \
258 asm( "addxl %d0, %d3 " );
259
260#endif /* MC68000 */
261
262#if defined(__powerpc__) || defined(__ppc__)
263#if defined(__powerpc64__) || defined(__ppc64__)
264
265#if defined(__MACH__) && defined(__APPLE__)
266
267#define MULADDC_INIT \
268 asm( "ld r3, %0 " :: "m" (s)); \
269 asm( "ld r4, %0 " :: "m" (d)); \
270 asm( "ld r5, %0 " :: "m" (c)); \
271 asm( "ld r6, %0 " :: "m" (b)); \
272 asm( "addi r3, r3, -8 " ); \
273 asm( "addi r4, r4, -8 " ); \
274 asm( "addic r5, r5, 0 " );
275
276#define MULADDC_CORE \
277 asm( "ldu r7, 8(r3) " ); \
278 asm( "mulld r8, r7, r6 " ); \
279 asm( "mulhdu r9, r7, r6 " ); \
280 asm( "adde r8, r8, r5 " ); \
281 asm( "ld r7, 8(r4) " ); \
282 asm( "addze r5, r9 " ); \
283 asm( "addc r8, r8, r7 " ); \
284 asm( "stdu r8, 8(r4) " );
285
286#define MULADDC_STOP \
287 asm( "addze r5, r5 " ); \
288 asm( "addi r4, r4, 8 " ); \
289 asm( "addi r3, r3, 8 " ); \
290 asm( "std r5, %0 " : "=m" (c)); \
291 asm( "std r4, %0 " : "=m" (d)); \
292 asm( "std r3, %0 " : "=m" (s) :: \
293 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
294
295#else
296
297#define MULADDC_INIT \
298 asm( "ld %%r3, %0 " :: "m" (s)); \
299 asm( "ld %%r4, %0 " :: "m" (d)); \
300 asm( "ld %%r5, %0 " :: "m" (c)); \
301 asm( "ld %%r6, %0 " :: "m" (b)); \
302 asm( "addi %r3, %r3, -8 " ); \
303 asm( "addi %r4, %r4, -8 " ); \
304 asm( "addic %r5, %r5, 0 " );
305
306#define MULADDC_CORE \
307 asm( "ldu %r7, 8(%r3) " ); \
308 asm( "mulld %r8, %r7, %r6 " ); \
309 asm( "mulhdu %r9, %r7, %r6 " ); \
310 asm( "adde %r8, %r8, %r5 " ); \
311 asm( "ld %r7, 8(%r4) " ); \
312 asm( "addze %r5, %r9 " ); \
313 asm( "addc %r8, %r8, %r7 " ); \
314 asm( "stdu %r8, 8(%r4) " );
315
316#define MULADDC_STOP \
317 asm( "addze %r5, %r5 " ); \
318 asm( "addi %r4, %r4, 8 " ); \
319 asm( "addi %r3, %r3, 8 " ); \
320 asm( "std %%r5, %0 " : "=m" (c)); \
321 asm( "std %%r4, %0 " : "=m" (d)); \
322 asm( "std %%r3, %0 " : "=m" (s) :: \
323 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
324
325#endif
326
327#else /* PPC32 */
328
329#if defined(__MACH__) && defined(__APPLE__)
330
331#define MULADDC_INIT \
332 asm( "lwz r3, %0 " :: "m" (s)); \
333 asm( "lwz r4, %0 " :: "m" (d)); \
334 asm( "lwz r5, %0 " :: "m" (c)); \
335 asm( "lwz r6, %0 " :: "m" (b)); \
336 asm( "addi r3, r3, -4 " ); \
337 asm( "addi r4, r4, -4 " ); \
338 asm( "addic r5, r5, 0 " );
339
340#define MULADDC_CORE \
341 asm( "lwzu r7, 4(r3) " ); \
342 asm( "mullw r8, r7, r6 " ); \
343 asm( "mulhwu r9, r7, r6 " ); \
344 asm( "adde r8, r8, r5 " ); \
345 asm( "lwz r7, 4(r4) " ); \
346 asm( "addze r5, r9 " ); \
347 asm( "addc r8, r8, r7 " ); \
348 asm( "stwu r8, 4(r4) " );
349
350#define MULADDC_STOP \
351 asm( "addze r5, r5 " ); \
352 asm( "addi r4, r4, 4 " ); \
353 asm( "addi r3, r3, 4 " ); \
354 asm( "stw r5, %0 " : "=m" (c)); \
355 asm( "stw r4, %0 " : "=m" (d)); \
356 asm( "stw r3, %0 " : "=m" (s) :: \
357 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
358
359#else
360
361#define MULADDC_INIT \
362 asm( "lwz %%r3, %0 " :: "m" (s)); \
363 asm( "lwz %%r4, %0 " :: "m" (d)); \
364 asm( "lwz %%r5, %0 " :: "m" (c)); \
365 asm( "lwz %%r6, %0 " :: "m" (b)); \
366 asm( "addi %r3, %r3, -4 " ); \
367 asm( "addi %r4, %r4, -4 " ); \
368 asm( "addic %r5, %r5, 0 " );
369
370#define MULADDC_CORE \
371 asm( "lwzu %r7, 4(%r3) " ); \
372 asm( "mullw %r8, %r7, %r6 " ); \
373 asm( "mulhwu %r9, %r7, %r6 " ); \
374 asm( "adde %r8, %r8, %r5 " ); \
375 asm( "lwz %r7, 4(%r4) " ); \
376 asm( "addze %r5, %r9 " ); \
377 asm( "addc %r8, %r8, %r7 " ); \
378 asm( "stwu %r8, 4(%r4) " );
379
380#define MULADDC_STOP \
381 asm( "addze %r5, %r5 " ); \
382 asm( "addi %r4, %r4, 4 " ); \
383 asm( "addi %r3, %r3, 4 " ); \
384 asm( "stw %%r5, %0 " : "=m" (c)); \
385 asm( "stw %%r4, %0 " : "=m" (d)); \
386 asm( "stw %%r3, %0 " : "=m" (s) :: \
387 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
388
389#endif
390
391#endif /* PPC32 */
392#endif /* PPC64 */
393
394#if defined(__sparc__)
395
396#define MULADDC_INIT \
397 asm( "ld %0, %%o0 " :: "m" (s)); \
398 asm( "ld %0, %%o1 " :: "m" (d)); \
399 asm( "ld %0, %%o2 " :: "m" (c)); \
400 asm( "ld %0, %%o3 " :: "m" (b));
401
402#define MULADDC_CORE \
403 asm( "ld [%o0], %o4 " ); \
404 asm( "inc 4, %o0 " ); \
405 asm( "ld [%o1], %o5 " ); \
406 asm( "umul %o3, %o4, %o4 " ); \
407 asm( "addcc %o4, %o2, %o4 " ); \
408 asm( "rd %y, %g1 " ); \
409 asm( "addx %g1, 0, %g1 " ); \
410 asm( "addcc %o4, %o5, %o4 " ); \
411 asm( "st %o4, [%o1] " ); \
412 asm( "addx %g1, 0, %o2 " ); \
413 asm( "inc 4, %o1 " );
414
415#define MULADDC_STOP \
416 asm( "st %%o2, %0 " : "=m" (c)); \
417 asm( "st %%o1, %0 " : "=m" (d)); \
418 asm( "st %%o0, %0 " : "=m" (s) :: \
419 "g1", "o0", "o1", "o2", "o3", "o4", "o5" );
420
421#endif /* SPARCv8 */
422
423#if defined(__microblaze__) || defined(microblaze)
424
425#define MULADDC_INIT \
426 asm( "lwi r3, %0 " :: "m" (s)); \
427 asm( "lwi r4, %0 " :: "m" (d)); \
428 asm( "lwi r5, %0 " :: "m" (c)); \
429 asm( "lwi r6, %0 " :: "m" (b)); \
430 asm( "andi r7, r6, 0xffff" ); \
431 asm( "bsrli r6, r6, 16 " );
432
433#define MULADDC_CORE \
434 asm( "lhui r8, r3, 0 " ); \
435 asm( "addi r3, r3, 2 " ); \
436 asm( "lhui r9, r3, 0 " ); \
437 asm( "addi r3, r3, 2 " ); \
438 asm( "mul r10, r9, r6 " ); \
439 asm( "mul r11, r8, r7 " ); \
440 asm( "mul r12, r9, r7 " ); \
441 asm( "mul r13, r8, r6 " ); \
442 asm( "bsrli r8, r10, 16 " ); \
443 asm( "bsrli r9, r11, 16 " ); \
444 asm( "add r13, r13, r8 " ); \
445 asm( "add r13, r13, r9 " ); \
446 asm( "bslli r10, r10, 16 " ); \
447 asm( "bslli r11, r11, 16 " ); \
448 asm( "add r12, r12, r10 " ); \
449 asm( "addc r13, r13, r0 " ); \
450 asm( "add r12, r12, r11 " ); \
451 asm( "addc r13, r13, r0 " ); \
452 asm( "lwi r10, r4, 0 " ); \
453 asm( "add r12, r12, r10 " ); \
454 asm( "addc r13, r13, r0 " ); \
455 asm( "add r12, r12, r5 " ); \
456 asm( "addc r5, r13, r0 " ); \
457 asm( "swi r12, r4, 0 " ); \
458 asm( "addi r4, r4, 4 " );
459
460#define MULADDC_STOP \
461 asm( "swi r5, %0 " : "=m" (c)); \
462 asm( "swi r4, %0 " : "=m" (d)); \
463 asm( "swi r3, %0 " : "=m" (s) :: \
464 "r3", "r4" , "r5" , "r6" , "r7" , "r8" , \
465 "r9", "r10", "r11", "r12", "r13" );
466
467#endif /* MicroBlaze */
468
469#if defined(__tricore__)
470
471#define MULADDC_INIT \
472 asm( "ld.a %%a2, %0 " :: "m" (s)); \
473 asm( "ld.a %%a3, %0 " :: "m" (d)); \
474 asm( "ld.w %%d4, %0 " :: "m" (c)); \
475 asm( "ld.w %%d1, %0 " :: "m" (b)); \
476 asm( "xor %d5, %d5 " );
477
478#define MULADDC_CORE \
479 asm( "ld.w %d0, [%a2+] " ); \
480 asm( "madd.u %e2, %e4, %d0, %d1 " ); \
481 asm( "ld.w %d0, [%a3] " ); \
482 asm( "addx %d2, %d2, %d0 " ); \
483 asm( "addc %d3, %d3, 0 " ); \
484 asm( "mov %d4, %d3 " ); \
485 asm( "st.w [%a3+], %d2 " );
486
487#define MULADDC_STOP \
488 asm( "st.w %0, %%d4 " : "=m" (c)); \
489 asm( "st.a %0, %%a3 " : "=m" (d)); \
490 asm( "st.a %0, %%a2 " : "=m" (s) :: \
491 "d0", "d1", "e2", "d4", "a2", "a3" );
492
493#endif /* TriCore */
494
495#if defined(__arm__)
496
497#define MULADDC_INIT \
498 asm( "ldr r0, %0 " :: "m" (s)); \
499 asm( "ldr r1, %0 " :: "m" (d)); \
500 asm( "ldr r2, %0 " :: "m" (c)); \
501 asm( "ldr r3, %0 " :: "m" (b));
502
503#define MULADDC_CORE \
504 asm( "ldr r4, [r0], #4 " ); \
505 asm( "mov r5, #0 " ); \
506 asm( "ldr r6, [r1] " ); \
507 asm( "umlal r2, r5, r3, r4 " ); \
508 asm( "adds r7, r6, r2 " ); \
509 asm( "adc r2, r5, #0 " ); \
510 asm( "str r7, [r1], #4 " );
511
512#define MULADDC_STOP \
513 asm( "str r2, %0 " : "=m" (c)); \
514 asm( "str r1, %0 " : "=m" (d)); \
515 asm( "str r0, %0 " : "=m" (s) :: \
516 "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7" );
517
518#endif /* ARMv3 */
519
520#if defined(__alpha__)
521
522#define MULADDC_INIT \
523 asm( "ldq $1, %0 " :: "m" (s)); \
524 asm( "ldq $2, %0 " :: "m" (d)); \
525 asm( "ldq $3, %0 " :: "m" (c)); \
526 asm( "ldq $4, %0 " :: "m" (b));
527
528#define MULADDC_CORE \
529 asm( "ldq $6, 0($1) " ); \
530 asm( "addq $1, 8, $1 " ); \
531 asm( "mulq $6, $4, $7 " ); \
532 asm( "umulh $6, $4, $6 " ); \
533 asm( "addq $7, $3, $7 " ); \
534 asm( "cmpult $7, $3, $3 " ); \
535 asm( "ldq $5, 0($2) " ); \
536 asm( "addq $7, $5, $7 " ); \
537 asm( "cmpult $7, $5, $5 " ); \
538 asm( "stq $7, 0($2) " ); \
539 asm( "addq $2, 8, $2 " ); \
540 asm( "addq $6, $3, $3 " ); \
541 asm( "addq $5, $3, $3 " );
542
543#define MULADDC_STOP \
544 asm( "stq $3, %0 " : "=m" (c)); \
545 asm( "stq $2, %0 " : "=m" (d)); \
546 asm( "stq $1, %0 " : "=m" (s) :: \
547 "$1", "$2", "$3", "$4", "$5", "$6", "$7" );
548
549#endif /* Alpha */
550
551#if defined(__mips__)
552
553#define MULADDC_INIT \
554 asm( "lw $10, %0 " :: "m" (s)); \
555 asm( "lw $11, %0 " :: "m" (d)); \
556 asm( "lw $12, %0 " :: "m" (c)); \
557 asm( "lw $13, %0 " :: "m" (b));
558
559#define MULADDC_CORE \
560 asm( "lw $14, 0($10) " ); \
561 asm( "multu $13, $14 " ); \
562 asm( "addi $10, $10, 4 " ); \
563 asm( "mflo $14 " ); \
564 asm( "mfhi $9 " ); \
565 asm( "addu $14, $12, $14 " ); \
566 asm( "lw $15, 0($11) " ); \
567 asm( "sltu $12, $14, $12 " ); \
568 asm( "addu $15, $14, $15 " ); \
569 asm( "sltu $14, $15, $14 " ); \
570 asm( "addu $12, $12, $9 " ); \
571 asm( "sw $15, 0($11) " ); \
572 asm( "addu $12, $12, $14 " ); \
573 asm( "addi $11, $11, 4 " );
574
575#define MULADDC_STOP \
576 asm( "sw $12, %0 " : "=m" (c)); \
577 asm( "sw $11, %0 " : "=m" (d)); \
578 asm( "sw $10, %0 " : "=m" (s) :: \
579 "$9", "$10", "$11", "$12", "$13", "$14", "$15" );
580
581#endif /* MIPS */
582#endif /* GNUC */
583
584#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
585
586#define MULADDC_INIT \
587 __asm mov esi, s \
588 __asm mov edi, d \
589 __asm mov ecx, c \
590 __asm mov ebx, b
591
592#define MULADDC_CORE \
593 __asm lodsd \
594 __asm mul ebx \
595 __asm add eax, ecx \
596 __asm adc edx, 0 \
597 __asm add eax, [edi] \
598 __asm adc edx, 0 \
599 __asm mov ecx, edx \
600 __asm stosd
601
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]602#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]603
604#define EMIT __asm _emit
605
606#define MULADDC_HUIT \
607 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
608 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
609 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
610 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
611 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
612 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
613 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
614 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
615 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
616 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
617 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
618 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
619 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
620 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
621 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
622 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
623 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
624 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
625 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
626 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
627 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
628 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
629 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
630 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
631 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
632 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
633 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
634 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
635 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
636 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
637 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
638 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
639 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
640 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
641 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
642 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
643 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
644 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
645 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
646 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
647 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
648 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
649 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
650 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
651 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
652 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
653 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
654 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
655 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
656 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
657 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
658 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
659 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
660 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
661 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
662 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
663 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
664 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
665 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
666 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
667 EMIT 0x0F EMIT 0x7E EMIT 0xC9
668
669#define MULADDC_STOP \
670 EMIT 0x0F EMIT 0x77 \
671 __asm mov c, ecx \
672 __asm mov d, edi \
673 __asm mov s, esi \
674
675#else
676
677#define MULADDC_STOP \
678 __asm mov c, ecx \
679 __asm mov d, edi \
680 __asm mov s, esi \
681
682#endif /* SSE2 */
683#endif /* MSVC */
684
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]685#endif /* POLARSSL_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]686
687#if !defined(MULADDC_CORE)
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]688#if defined(POLARSSL_HAVE_LONGLONG)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]689
690#define MULADDC_INIT \
691{ \
692 t_dbl r; \
693 t_int r0, r1;
694
695#define MULADDC_CORE \
696 r = *(s++) * (t_dbl) b; \
697 r0 = r; \
698 r1 = r >> biL; \
699 r0 += c; r1 += (r0 < c); \
700 r0 += *d; r1 += (r0 < *d); \
701 c = r1; *(d++) = r0;
702
703#define MULADDC_STOP \
704}
705
706#else
707#define MULADDC_INIT \
708{ \
709 t_int s0, s1, b0, b1; \
710 t_int r0, r1, rx, ry; \
711 b0 = ( b << biH ) >> biH; \
712 b1 = ( b >> biH );
713
714#define MULADDC_CORE \
715 s0 = ( *s << biH ) >> biH; \
716 s1 = ( *s >> biH ); s++; \
717 rx = s0 * b1; r0 = s0 * b0; \
718 ry = s1 * b0; r1 = s1 * b1; \
719 r1 += ( rx >> biH ); \
720 r1 += ( ry >> biH ); \
721 rx <<= biH; ry <<= biH; \
722 r0 += rx; r1 += (r0 < rx); \
723 r0 += ry; r1 += (r0 < ry); \
724 r0 += c; r1 += (r0 < c); \
725 r0 += *d; r1 += (r0 < *d); \
726 c = r1; *(d++) = r0;
727
728#define MULADDC_STOP \
729}
730
731#endif /* C (generic) */
732#endif /* C (longlong) */
733
734#endif /* bn_mul.h */