TrustedFirmware Git Browser
Code Review Sign In
review.trustedfirmware.org / mirror / mbed-tls / 84f12b76fc6d5b9dc3ec4eff3dc0f1e6b03a259a / . / include / polarssl / bn_mul.h
blob: 2470fc7b50874d159b8b0e7f5009d67a00769afc [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 Bakker84f12b72010-07-18 10:13:04 +0000[diff] [blame^]4 * Copyright (C) 2006-2010, Brainspark B.V.
5 * Lead Maintainer: Paul Bakker <polarssl_maintainer at polarssl.org>
Paul Bakker77b385e2009-07-28 17:23:11 +0000[diff] [blame]6 * All rights reserved.
Paul Bakkere0ccd0a2009-01-04 16:27:10 +0000[diff] [blame]7 *
Paul Bakkere0ccd0a2009-01-04 16:27:10 +0000[diff] [blame]8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
12 *
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License along
19 * with this program; if not, write to the Free Software Foundation, Inc.,
20 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]21 */
22/*
23 * Multiply source vector [s] with b, add result
24 * to destination vector [d] and set carry c.
25 *
26 * Currently supports:
27 *
28 * . IA-32 (386+) . AMD64 / EM64T
29 * . IA-32 (SSE2) . Motorola 68000
30 * . PowerPC, 32-bit . MicroBlaze
31 * . PowerPC, 64-bit . TriCore
32 * . SPARC v8 . ARM v3+
33 * . Alpha . MIPS32
34 * . C, longlong . C, generic
35 */
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]36#ifndef POLARSSL_BN_MUL_H
37#define POLARSSL_BN_MUL_H
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]38
Paul Bakker8e831ed2009-01-03 21:24:11 +0000[diff] [blame]39#include "polarssl/config.h"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]40
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]41#if defined(POLARSSL_HAVE_ASM)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]42
43#if defined(__GNUC__)
44#if defined(__i386__)
45
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]46#define MULADDC_INIT \
47 asm( " \
48 movl %%ebx, %0; \
49 movl %5, %%esi; \
50 movl %6, %%edi; \
51 movl %7, %%ecx; \
52 movl %8, %%ebx; \
53 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]54
Paul Bakker40fe2b02009-07-08 19:49:01 +0000[diff] [blame]55#define MULADDC_CORE \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]56 " \
Paul Bakker40fe2b02009-07-08 19:49:01 +0000[diff] [blame]57 lodsl; \
58 mull %%ebx; \
59 addl %%ecx, %%eax; \
60 adcl $0, %%edx; \
61 addl (%%edi), %%eax; \
62 adcl $0, %%edx; \
63 movl %%edx, %%ecx; \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]64 stosl; \
65 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]66
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]67#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]68
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]69#define MULADDC_HUIT \
70 " \
71 movd %%ecx, %%mm1; \
72 movd %%ebx, %%mm0; \
73 movd (%%edi), %%mm3; \
74 paddq %%mm3, %%mm1; \
75 movd (%%esi), %%mm2; \
76 pmuludq %%mm0, %%mm2; \
77 movd 4(%%esi), %%mm4; \
78 pmuludq %%mm0, %%mm4; \
79 movd 8(%%esi), %%mm6; \
80 pmuludq %%mm0, %%mm6; \
81 movd 12(%%esi), %%mm7; \
82 pmuludq %%mm0, %%mm7; \
83 paddq %%mm2, %%mm1; \
84 movd 4(%%edi), %%mm3; \
85 paddq %%mm4, %%mm3; \
86 movd 8(%%edi), %%mm5; \
87 paddq %%mm6, %%mm5; \
88 movd 12(%%edi), %%mm4; \
89 paddq %%mm4, %%mm7; \
90 movd %%mm1, (%%edi); \
91 movd 16(%%esi), %%mm2; \
92 pmuludq %%mm0, %%mm2; \
93 psrlq $32, %%mm1; \
94 movd 20(%%esi), %%mm4; \
95 pmuludq %%mm0, %%mm4; \
96 paddq %%mm3, %%mm1; \
97 movd 24(%%esi), %%mm6; \
98 pmuludq %%mm0, %%mm6; \
99 movd %%mm1, 4(%%edi); \
100 psrlq $32, %%mm1; \
101 movd 28(%%esi), %%mm3; \
102 pmuludq %%mm0, %%mm3; \
103 paddq %%mm5, %%mm1; \
104 movd 16(%%edi), %%mm5; \
105 paddq %%mm5, %%mm2; \
106 movd %%mm1, 8(%%edi); \
107 psrlq $32, %%mm1; \
108 paddq %%mm7, %%mm1; \
109 movd 20(%%edi), %%mm5; \
110 paddq %%mm5, %%mm4; \
111 movd %%mm1, 12(%%edi); \
112 psrlq $32, %%mm1; \
113 paddq %%mm2, %%mm1; \
114 movd 24(%%edi), %%mm5; \
115 paddq %%mm5, %%mm6; \
116 movd %%mm1, 16(%%edi); \
117 psrlq $32, %%mm1; \
118 paddq %%mm4, %%mm1; \
119 movd 28(%%edi), %%mm5; \
120 paddq %%mm5, %%mm3; \
121 movd %%mm1, 20(%%edi); \
122 psrlq $32, %%mm1; \
123 paddq %%mm6, %%mm1; \
124 movd %%mm1, 24(%%edi); \
125 psrlq $32, %%mm1; \
126 paddq %%mm3, %%mm1; \
127 movd %%mm1, 28(%%edi); \
128 addl $32, %%edi; \
129 addl $32, %%esi; \
130 psrlq $32, %%mm1; \
131 movd %%mm1, %%ecx; \
132 "
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]133
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]134#define MULADDC_STOP \
135 " \
136 emms; \
137 movl %4, %%ebx; \
138 movl %%ecx, %1; \
139 movl %%edi, %2; \
140 movl %%esi, %3; \
141 " \
142 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
143 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
144 : "eax", "ecx", "edx", "esi", "edi" \
145 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]146
147#else
148
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]149#define MULADDC_STOP \
150 " \
151 movl %4, %%ebx; \
152 movl %%ecx, %1; \
153 movl %%edi, %2; \
154 movl %%esi, %3; \
155 " \
156 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
157 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
158 : "eax", "ecx", "edx", "esi", "edi" \
159 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]160#endif /* SSE2 */
161#endif /* i386 */
162
163#if defined(__amd64__) || defined (__x86_64__)
164
165#define MULADDC_INIT \
166 asm( "movq %0, %%rsi " :: "m" (s)); \
167 asm( "movq %0, %%rdi " :: "m" (d)); \
168 asm( "movq %0, %%rcx " :: "m" (c)); \
169 asm( "movq %0, %%rbx " :: "m" (b)); \
170 asm( "xorq %r8, %r8 " );
171
172#define MULADDC_CORE \
173 asm( "movq (%rsi),%rax " ); \
174 asm( "mulq %rbx " ); \
175 asm( "addq $8, %rsi " ); \
176 asm( "addq %rcx, %rax " ); \
177 asm( "movq %r8, %rcx " ); \
178 asm( "adcq $0, %rdx " ); \
179 asm( "nop " ); \
180 asm( "addq %rax, (%rdi) " ); \
181 asm( "adcq %rdx, %rcx " ); \
182 asm( "addq $8, %rdi " );
183
184#define MULADDC_STOP \
185 asm( "movq %%rcx, %0 " : "=m" (c)); \
186 asm( "movq %%rdi, %0 " : "=m" (d)); \
187 asm( "movq %%rsi, %0 " : "=m" (s) :: \
188 "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" );
189
190#endif /* AMD64 */
191
192#if defined(__mc68020__) || defined(__mcpu32__)
193
194#define MULADDC_INIT \
195 asm( "movl %0, %%a2 " :: "m" (s)); \
196 asm( "movl %0, %%a3 " :: "m" (d)); \
197 asm( "movl %0, %%d3 " :: "m" (c)); \
198 asm( "movl %0, %%d2 " :: "m" (b)); \
199 asm( "moveq #0, %d0 " );
200
201#define MULADDC_CORE \
202 asm( "movel %a2@+, %d1 " ); \
203 asm( "mulul %d2, %d4:%d1 " ); \
204 asm( "addl %d3, %d1 " ); \
205 asm( "addxl %d0, %d4 " ); \
206 asm( "moveq #0, %d3 " ); \
207 asm( "addl %d1, %a3@+ " ); \
208 asm( "addxl %d4, %d3 " );
209
210#define MULADDC_STOP \
211 asm( "movl %%d3, %0 " : "=m" (c)); \
212 asm( "movl %%a3, %0 " : "=m" (d)); \
213 asm( "movl %%a2, %0 " : "=m" (s) :: \
214 "d0", "d1", "d2", "d3", "d4", "a2", "a3" );
215
216#define MULADDC_HUIT \
217 asm( "movel %a2@+, %d1 " ); \
218 asm( "mulul %d2, %d4:%d1 " ); \
219 asm( "addxl %d3, %d1 " ); \
220 asm( "addxl %d0, %d4 " ); \
221 asm( "addl %d1, %a3@+ " ); \
222 asm( "movel %a2@+, %d1 " ); \
223 asm( "mulul %d2, %d3:%d1 " ); \
224 asm( "addxl %d4, %d1 " ); \
225 asm( "addxl %d0, %d3 " ); \
226 asm( "addl %d1, %a3@+ " ); \
227 asm( "movel %a2@+, %d1 " ); \
228 asm( "mulul %d2, %d4:%d1 " ); \
229 asm( "addxl %d3, %d1 " ); \
230 asm( "addxl %d0, %d4 " ); \
231 asm( "addl %d1, %a3@+ " ); \
232 asm( "movel %a2@+, %d1 " ); \
233 asm( "mulul %d2, %d3:%d1 " ); \
234 asm( "addxl %d4, %d1 " ); \
235 asm( "addxl %d0, %d3 " ); \
236 asm( "addl %d1, %a3@+ " ); \
237 asm( "movel %a2@+, %d1 " ); \
238 asm( "mulul %d2, %d4:%d1 " ); \
239 asm( "addxl %d3, %d1 " ); \
240 asm( "addxl %d0, %d4 " ); \
241 asm( "addl %d1, %a3@+ " ); \
242 asm( "movel %a2@+, %d1 " ); \
243 asm( "mulul %d2, %d3:%d1 " ); \
244 asm( "addxl %d4, %d1 " ); \
245 asm( "addxl %d0, %d3 " ); \
246 asm( "addl %d1, %a3@+ " ); \
247 asm( "movel %a2@+, %d1 " ); \
248 asm( "mulul %d2, %d4:%d1 " ); \
249 asm( "addxl %d3, %d1 " ); \
250 asm( "addxl %d0, %d4 " ); \
251 asm( "addl %d1, %a3@+ " ); \
252 asm( "movel %a2@+, %d1 " ); \
253 asm( "mulul %d2, %d3:%d1 " ); \
254 asm( "addxl %d4, %d1 " ); \
255 asm( "addxl %d0, %d3 " ); \
256 asm( "addl %d1, %a3@+ " ); \
257 asm( "addxl %d0, %d3 " );
258
259#endif /* MC68000 */
260
261#if defined(__powerpc__) || defined(__ppc__)
262#if defined(__powerpc64__) || defined(__ppc64__)
263
264#if defined(__MACH__) && defined(__APPLE__)
265
266#define MULADDC_INIT \
267 asm( "ld r3, %0 " :: "m" (s)); \
268 asm( "ld r4, %0 " :: "m" (d)); \
269 asm( "ld r5, %0 " :: "m" (c)); \
270 asm( "ld r6, %0 " :: "m" (b)); \
271 asm( "addi r3, r3, -8 " ); \
272 asm( "addi r4, r4, -8 " ); \
273 asm( "addic r5, r5, 0 " );
274
275#define MULADDC_CORE \
276 asm( "ldu r7, 8(r3) " ); \
277 asm( "mulld r8, r7, r6 " ); \
278 asm( "mulhdu r9, r7, r6 " ); \
279 asm( "adde r8, r8, r5 " ); \
280 asm( "ld r7, 8(r4) " ); \
281 asm( "addze r5, r9 " ); \
282 asm( "addc r8, r8, r7 " ); \
283 asm( "stdu r8, 8(r4) " );
284
285#define MULADDC_STOP \
286 asm( "addze r5, r5 " ); \
287 asm( "addi r4, r4, 8 " ); \
288 asm( "addi r3, r3, 8 " ); \
289 asm( "std r5, %0 " : "=m" (c)); \
290 asm( "std r4, %0 " : "=m" (d)); \
291 asm( "std r3, %0 " : "=m" (s) :: \
292 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
293
294#else
295
296#define MULADDC_INIT \
297 asm( "ld %%r3, %0 " :: "m" (s)); \
298 asm( "ld %%r4, %0 " :: "m" (d)); \
299 asm( "ld %%r5, %0 " :: "m" (c)); \
300 asm( "ld %%r6, %0 " :: "m" (b)); \
301 asm( "addi %r3, %r3, -8 " ); \
302 asm( "addi %r4, %r4, -8 " ); \
303 asm( "addic %r5, %r5, 0 " );
304
305#define MULADDC_CORE \
306 asm( "ldu %r7, 8(%r3) " ); \
307 asm( "mulld %r8, %r7, %r6 " ); \
308 asm( "mulhdu %r9, %r7, %r6 " ); \
309 asm( "adde %r8, %r8, %r5 " ); \
310 asm( "ld %r7, 8(%r4) " ); \
311 asm( "addze %r5, %r9 " ); \
312 asm( "addc %r8, %r8, %r7 " ); \
313 asm( "stdu %r8, 8(%r4) " );
314
315#define MULADDC_STOP \
316 asm( "addze %r5, %r5 " ); \
317 asm( "addi %r4, %r4, 8 " ); \
318 asm( "addi %r3, %r3, 8 " ); \
319 asm( "std %%r5, %0 " : "=m" (c)); \
320 asm( "std %%r4, %0 " : "=m" (d)); \
321 asm( "std %%r3, %0 " : "=m" (s) :: \
322 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
323
324#endif
325
326#else /* PPC32 */
327
328#if defined(__MACH__) && defined(__APPLE__)
329
330#define MULADDC_INIT \
331 asm( "lwz r3, %0 " :: "m" (s)); \
332 asm( "lwz r4, %0 " :: "m" (d)); \
333 asm( "lwz r5, %0 " :: "m" (c)); \
334 asm( "lwz r6, %0 " :: "m" (b)); \
335 asm( "addi r3, r3, -4 " ); \
336 asm( "addi r4, r4, -4 " ); \
337 asm( "addic r5, r5, 0 " );
338
339#define MULADDC_CORE \
340 asm( "lwzu r7, 4(r3) " ); \
341 asm( "mullw r8, r7, r6 " ); \
342 asm( "mulhwu r9, r7, r6 " ); \
343 asm( "adde r8, r8, r5 " ); \
344 asm( "lwz r7, 4(r4) " ); \
345 asm( "addze r5, r9 " ); \
346 asm( "addc r8, r8, r7 " ); \
347 asm( "stwu r8, 4(r4) " );
348
349#define MULADDC_STOP \
350 asm( "addze r5, r5 " ); \
351 asm( "addi r4, r4, 4 " ); \
352 asm( "addi r3, r3, 4 " ); \
353 asm( "stw r5, %0 " : "=m" (c)); \
354 asm( "stw r4, %0 " : "=m" (d)); \
355 asm( "stw r3, %0 " : "=m" (s) :: \
356 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
357
358#else
359
360#define MULADDC_INIT \
361 asm( "lwz %%r3, %0 " :: "m" (s)); \
362 asm( "lwz %%r4, %0 " :: "m" (d)); \
363 asm( "lwz %%r5, %0 " :: "m" (c)); \
364 asm( "lwz %%r6, %0 " :: "m" (b)); \
365 asm( "addi %r3, %r3, -4 " ); \
366 asm( "addi %r4, %r4, -4 " ); \
367 asm( "addic %r5, %r5, 0 " );
368
369#define MULADDC_CORE \
370 asm( "lwzu %r7, 4(%r3) " ); \
371 asm( "mullw %r8, %r7, %r6 " ); \
372 asm( "mulhwu %r9, %r7, %r6 " ); \
373 asm( "adde %r8, %r8, %r5 " ); \
374 asm( "lwz %r7, 4(%r4) " ); \
375 asm( "addze %r5, %r9 " ); \
376 asm( "addc %r8, %r8, %r7 " ); \
377 asm( "stwu %r8, 4(%r4) " );
378
379#define MULADDC_STOP \
380 asm( "addze %r5, %r5 " ); \
381 asm( "addi %r4, %r4, 4 " ); \
382 asm( "addi %r3, %r3, 4 " ); \
383 asm( "stw %%r5, %0 " : "=m" (c)); \
384 asm( "stw %%r4, %0 " : "=m" (d)); \
385 asm( "stw %%r3, %0 " : "=m" (s) :: \
386 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
387
388#endif
389
390#endif /* PPC32 */
391#endif /* PPC64 */
392
393#if defined(__sparc__)
394
395#define MULADDC_INIT \
396 asm( "ld %0, %%o0 " :: "m" (s)); \
397 asm( "ld %0, %%o1 " :: "m" (d)); \
398 asm( "ld %0, %%o2 " :: "m" (c)); \
399 asm( "ld %0, %%o3 " :: "m" (b));
400
401#define MULADDC_CORE \
402 asm( "ld [%o0], %o4 " ); \
403 asm( "inc 4, %o0 " ); \
404 asm( "ld [%o1], %o5 " ); \
405 asm( "umul %o3, %o4, %o4 " ); \
406 asm( "addcc %o4, %o2, %o4 " ); \
407 asm( "rd %y, %g1 " ); \
408 asm( "addx %g1, 0, %g1 " ); \
409 asm( "addcc %o4, %o5, %o4 " ); \
410 asm( "st %o4, [%o1] " ); \
411 asm( "addx %g1, 0, %o2 " ); \
412 asm( "inc 4, %o1 " );
413
414#define MULADDC_STOP \
415 asm( "st %%o2, %0 " : "=m" (c)); \
416 asm( "st %%o1, %0 " : "=m" (d)); \
417 asm( "st %%o0, %0 " : "=m" (s) :: \
418 "g1", "o0", "o1", "o2", "o3", "o4", "o5" );
419
420#endif /* SPARCv8 */
421
422#if defined(__microblaze__) || defined(microblaze)
423
424#define MULADDC_INIT \
425 asm( "lwi r3, %0 " :: "m" (s)); \
426 asm( "lwi r4, %0 " :: "m" (d)); \
427 asm( "lwi r5, %0 " :: "m" (c)); \
428 asm( "lwi r6, %0 " :: "m" (b)); \
429 asm( "andi r7, r6, 0xffff" ); \
430 asm( "bsrli r6, r6, 16 " );
431
432#define MULADDC_CORE \
433 asm( "lhui r8, r3, 0 " ); \
434 asm( "addi r3, r3, 2 " ); \
435 asm( "lhui r9, r3, 0 " ); \
436 asm( "addi r3, r3, 2 " ); \
437 asm( "mul r10, r9, r6 " ); \
438 asm( "mul r11, r8, r7 " ); \
439 asm( "mul r12, r9, r7 " ); \
440 asm( "mul r13, r8, r6 " ); \
441 asm( "bsrli r8, r10, 16 " ); \
442 asm( "bsrli r9, r11, 16 " ); \
443 asm( "add r13, r13, r8 " ); \
444 asm( "add r13, r13, r9 " ); \
445 asm( "bslli r10, r10, 16 " ); \
446 asm( "bslli r11, r11, 16 " ); \
447 asm( "add r12, r12, r10 " ); \
448 asm( "addc r13, r13, r0 " ); \
449 asm( "add r12, r12, r11 " ); \
450 asm( "addc r13, r13, r0 " ); \
451 asm( "lwi r10, r4, 0 " ); \
452 asm( "add r12, r12, r10 " ); \
453 asm( "addc r13, r13, r0 " ); \
454 asm( "add r12, r12, r5 " ); \
455 asm( "addc r5, r13, r0 " ); \
456 asm( "swi r12, r4, 0 " ); \
457 asm( "addi r4, r4, 4 " );
458
459#define MULADDC_STOP \
460 asm( "swi r5, %0 " : "=m" (c)); \
461 asm( "swi r4, %0 " : "=m" (d)); \
462 asm( "swi r3, %0 " : "=m" (s) :: \
463 "r3", "r4" , "r5" , "r6" , "r7" , "r8" , \
464 "r9", "r10", "r11", "r12", "r13" );
465
466#endif /* MicroBlaze */
467
468#if defined(__tricore__)
469
470#define MULADDC_INIT \
471 asm( "ld.a %%a2, %0 " :: "m" (s)); \
472 asm( "ld.a %%a3, %0 " :: "m" (d)); \
473 asm( "ld.w %%d4, %0 " :: "m" (c)); \
474 asm( "ld.w %%d1, %0 " :: "m" (b)); \
475 asm( "xor %d5, %d5 " );
476
477#define MULADDC_CORE \
478 asm( "ld.w %d0, [%a2+] " ); \
479 asm( "madd.u %e2, %e4, %d0, %d1 " ); \
480 asm( "ld.w %d0, [%a3] " ); \
481 asm( "addx %d2, %d2, %d0 " ); \
482 asm( "addc %d3, %d3, 0 " ); \
483 asm( "mov %d4, %d3 " ); \
484 asm( "st.w [%a3+], %d2 " );
485
486#define MULADDC_STOP \
487 asm( "st.w %0, %%d4 " : "=m" (c)); \
488 asm( "st.a %0, %%a3 " : "=m" (d)); \
489 asm( "st.a %0, %%a2 " : "=m" (s) :: \
490 "d0", "d1", "e2", "d4", "a2", "a3" );
491
492#endif /* TriCore */
493
494#if defined(__arm__)
495
496#define MULADDC_INIT \
497 asm( "ldr r0, %0 " :: "m" (s)); \
498 asm( "ldr r1, %0 " :: "m" (d)); \
499 asm( "ldr r2, %0 " :: "m" (c)); \
500 asm( "ldr r3, %0 " :: "m" (b));
501
502#define MULADDC_CORE \
503 asm( "ldr r4, [r0], #4 " ); \
504 asm( "mov r5, #0 " ); \
505 asm( "ldr r6, [r1] " ); \
506 asm( "umlal r2, r5, r3, r4 " ); \
507 asm( "adds r7, r6, r2 " ); \
508 asm( "adc r2, r5, #0 " ); \
509 asm( "str r7, [r1], #4 " );
510
511#define MULADDC_STOP \
512 asm( "str r2, %0 " : "=m" (c)); \
513 asm( "str r1, %0 " : "=m" (d)); \
514 asm( "str r0, %0 " : "=m" (s) :: \
515 "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7" );
516
517#endif /* ARMv3 */
518
519#if defined(__alpha__)
520
521#define MULADDC_INIT \
522 asm( "ldq $1, %0 " :: "m" (s)); \
523 asm( "ldq $2, %0 " :: "m" (d)); \
524 asm( "ldq $3, %0 " :: "m" (c)); \
525 asm( "ldq $4, %0 " :: "m" (b));
526
527#define MULADDC_CORE \
528 asm( "ldq $6, 0($1) " ); \
529 asm( "addq $1, 8, $1 " ); \
530 asm( "mulq $6, $4, $7 " ); \
531 asm( "umulh $6, $4, $6 " ); \
532 asm( "addq $7, $3, $7 " ); \
533 asm( "cmpult $7, $3, $3 " ); \
534 asm( "ldq $5, 0($2) " ); \
535 asm( "addq $7, $5, $7 " ); \
536 asm( "cmpult $7, $5, $5 " ); \
537 asm( "stq $7, 0($2) " ); \
538 asm( "addq $2, 8, $2 " ); \
539 asm( "addq $6, $3, $3 " ); \
540 asm( "addq $5, $3, $3 " );
541
542#define MULADDC_STOP \
543 asm( "stq $3, %0 " : "=m" (c)); \
544 asm( "stq $2, %0 " : "=m" (d)); \
545 asm( "stq $1, %0 " : "=m" (s) :: \
546 "$1", "$2", "$3", "$4", "$5", "$6", "$7" );
547
548#endif /* Alpha */
549
550#if defined(__mips__)
551
552#define MULADDC_INIT \
553 asm( "lw $10, %0 " :: "m" (s)); \
554 asm( "lw $11, %0 " :: "m" (d)); \
555 asm( "lw $12, %0 " :: "m" (c)); \
556 asm( "lw $13, %0 " :: "m" (b));
557
558#define MULADDC_CORE \
559 asm( "lw $14, 0($10) " ); \
560 asm( "multu $13, $14 " ); \
561 asm( "addi $10, $10, 4 " ); \
562 asm( "mflo $14 " ); \
563 asm( "mfhi $9 " ); \
564 asm( "addu $14, $12, $14 " ); \
565 asm( "lw $15, 0($11) " ); \
566 asm( "sltu $12, $14, $12 " ); \
567 asm( "addu $15, $14, $15 " ); \
568 asm( "sltu $14, $15, $14 " ); \
569 asm( "addu $12, $12, $9 " ); \
570 asm( "sw $15, 0($11) " ); \
571 asm( "addu $12, $12, $14 " ); \
572 asm( "addi $11, $11, 4 " );
573
574#define MULADDC_STOP \
575 asm( "sw $12, %0 " : "=m" (c)); \
576 asm( "sw $11, %0 " : "=m" (d)); \
577 asm( "sw $10, %0 " : "=m" (s) :: \
578 "$9", "$10", "$11", "$12", "$13", "$14", "$15" );
579
580#endif /* MIPS */
581#endif /* GNUC */
582
583#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
584
585#define MULADDC_INIT \
586 __asm mov esi, s \
587 __asm mov edi, d \
588 __asm mov ecx, c \
589 __asm mov ebx, b
590
591#define MULADDC_CORE \
592 __asm lodsd \
593 __asm mul ebx \
594 __asm add eax, ecx \
595 __asm adc edx, 0 \
596 __asm add eax, [edi] \
597 __asm adc edx, 0 \
598 __asm mov ecx, edx \
599 __asm stosd
600
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]601#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]602
603#define EMIT __asm _emit
604
605#define MULADDC_HUIT \
606 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
607 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
608 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
609 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
610 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
611 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
612 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
613 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
614 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
615 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
616 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
617 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
618 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
619 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
620 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
621 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
622 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
623 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
624 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
625 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
626 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
627 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
628 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
629 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
630 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
631 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
632 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
633 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
634 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
635 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
636 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
637 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
638 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
639 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
640 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
641 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
642 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
643 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
644 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
645 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
646 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
647 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
648 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
649 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
650 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
651 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
652 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
653 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
654 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
655 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
656 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
657 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
658 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
659 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
660 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
661 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
662 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
663 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
664 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
665 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
666 EMIT 0x0F EMIT 0x7E EMIT 0xC9
667
668#define MULADDC_STOP \
669 EMIT 0x0F EMIT 0x77 \
670 __asm mov c, ecx \
671 __asm mov d, edi \
672 __asm mov s, esi \
673
674#else
675
676#define MULADDC_STOP \
677 __asm mov c, ecx \
678 __asm mov d, edi \
679 __asm mov s, esi \
680
681#endif /* SSE2 */
682#endif /* MSVC */
683
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]684#endif /* POLARSSL_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]685
686#if !defined(MULADDC_CORE)
Paul Bakker40e46942009-01-03 21:51:57 +0000[diff] [blame]687#if defined(POLARSSL_HAVE_LONGLONG)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]688
689#define MULADDC_INIT \
690{ \
691 t_dbl r; \
692 t_int r0, r1;
693
694#define MULADDC_CORE \
695 r = *(s++) * (t_dbl) b; \
696 r0 = r; \
697 r1 = r >> biL; \
698 r0 += c; r1 += (r0 < c); \
699 r0 += *d; r1 += (r0 < *d); \
700 c = r1; *(d++) = r0;
701
702#define MULADDC_STOP \
703}
704
705#else
706#define MULADDC_INIT \
707{ \
708 t_int s0, s1, b0, b1; \
709 t_int r0, r1, rx, ry; \
710 b0 = ( b << biH ) >> biH; \
711 b1 = ( b >> biH );
712
713#define MULADDC_CORE \
714 s0 = ( *s << biH ) >> biH; \
715 s1 = ( *s >> biH ); s++; \
716 rx = s0 * b1; r0 = s0 * b0; \
717 ry = s1 * b0; r1 = s1 * b1; \
718 r1 += ( rx >> biH ); \
719 r1 += ( ry >> biH ); \
720 rx <<= biH; ry <<= biH; \
721 r0 += rx; r1 += (r0 < rx); \
722 r0 += ry; r1 += (r0 < ry); \
723 r0 += c; r1 += (r0 < c); \
724 r0 += *d; r1 += (r0 < *d); \
725 c = r1; *(d++) = r0;
726
727#define MULADDC_STOP \
728}
729
730#endif /* C (generic) */
731#endif /* C (longlong) */
732
733#endif /* bn_mul.h */