blob: 0e9385beffdffdaaefcf56c7763ed3854c3f4509 [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
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +020051#define MULADDC_INIT \
52 asm( \
53 "movl %%ebx, %0 \n\t" \
54 "movl %5, %%esi \n\t" \
55 "movl %6, %%edi \n\t" \
56 "movl %7, %%ecx \n\t" \
57 "movl %8, %%ebx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +000058
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +020059#define MULADDC_CORE \
60 "lodsl \n\t" \
61 "mull %%ebx \n\t" \
62 "addl %%ecx, %%eax \n\t" \
63 "adcl $0, %%edx \n\t" \
64 "addl (%%edi), %%eax \n\t" \
65 "adcl $0, %%edx \n\t" \
66 "movl %%edx, %%ecx \n\t" \
67 "stosl \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +000068
Paul Bakker40e46942009-01-03 21:51:57 +000069#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +000070
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +020071#define MULADDC_HUIT \
72 "movd %%ecx, %%mm1 \n\t" \
73 "movd %%ebx, %%mm0 \n\t" \
74 "movd (%%edi), %%mm3 \n\t" \
75 "paddq %%mm3, %%mm1 \n\t" \
76 "movd (%%esi), %%mm2 \n\t" \
77 "pmuludq %%mm0, %%mm2 \n\t" \
78 "movd 4(%%esi), %%mm4 \n\t" \
79 "pmuludq %%mm0, %%mm4 \n\t" \
80 "movd 8(%%esi), %%mm6 \n\t" \
81 "pmuludq %%mm0, %%mm6 \n\t" \
82 "movd 12(%%esi), %%mm7 \n\t" \
83 "pmuludq %%mm0, %%mm7 \n\t" \
84 "paddq %%mm2, %%mm1 \n\t" \
85 "movd 4(%%edi), %%mm3 \n\t" \
86 "paddq %%mm4, %%mm3 \n\t" \
87 "movd 8(%%edi), %%mm5 \n\t" \
88 "paddq %%mm6, %%mm5 \n\t" \
89 "movd 12(%%edi), %%mm4 \n\t" \
90 "paddq %%mm4, %%mm7 \n\t" \
91 "movd %%mm1, (%%edi) \n\t" \
92 "movd 16(%%esi), %%mm2 \n\t" \
93 "pmuludq %%mm0, %%mm2 \n\t" \
94 "psrlq $32, %%mm1 \n\t" \
95 "movd 20(%%esi), %%mm4 \n\t" \
96 "pmuludq %%mm0, %%mm4 \n\t" \
97 "paddq %%mm3, %%mm1 \n\t" \
98 "movd 24(%%esi), %%mm6 \n\t" \
99 "pmuludq %%mm0, %%mm6 \n\t" \
100 "movd %%mm1, 4(%%edi) \n\t" \
101 "psrlq $32, %%mm1 \n\t" \
102 "movd 28(%%esi), %%mm3 \n\t" \
103 "pmuludq %%mm0, %%mm3 \n\t" \
104 "paddq %%mm5, %%mm1 \n\t" \
105 "movd 16(%%edi), %%mm5 \n\t" \
106 "paddq %%mm5, %%mm2 \n\t" \
107 "movd %%mm1, 8(%%edi) \n\t" \
108 "psrlq $32, %%mm1 \n\t" \
109 "paddq %%mm7, %%mm1 \n\t" \
110 "movd 20(%%edi), %%mm5 \n\t" \
111 "paddq %%mm5, %%mm4 \n\t" \
112 "movd %%mm1, 12(%%edi) \n\t" \
113 "psrlq $32, %%mm1 \n\t" \
114 "paddq %%mm2, %%mm1 \n\t" \
115 "movd 24(%%edi), %%mm5 \n\t" \
116 "paddq %%mm5, %%mm6 \n\t" \
117 "movd %%mm1, 16(%%edi) \n\t" \
118 "psrlq $32, %%mm1 \n\t" \
119 "paddq %%mm4, %%mm1 \n\t" \
120 "movd 28(%%edi), %%mm5 \n\t" \
121 "paddq %%mm5, %%mm3 \n\t" \
122 "movd %%mm1, 20(%%edi) \n\t" \
123 "psrlq $32, %%mm1 \n\t" \
124 "paddq %%mm6, %%mm1 \n\t" \
125 "movd %%mm1, 24(%%edi) \n\t" \
126 "psrlq $32, %%mm1 \n\t" \
127 "paddq %%mm3, %%mm1 \n\t" \
128 "movd %%mm1, 28(%%edi) \n\t" \
129 "addl $32, %%edi \n\t" \
130 "addl $32, %%esi \n\t" \
131 "psrlq $32, %%mm1 \n\t" \
132 "movd %%mm1, %%ecx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000133
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200134#define MULADDC_STOP \
135 "emms \n\t" \
136 "movl %4, %%ebx \n\t" \
137 "movl %%ecx, %1 \n\t" \
138 "movl %%edi, %2 \n\t" \
139 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000140 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
141 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
142 : "eax", "ecx", "edx", "esi", "edi" \
143 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000144
145#else
146
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200147#define MULADDC_STOP \
148 "movl %4, %%ebx \n\t" \
149 "movl %%ecx, %1 \n\t" \
150 "movl %%edi, %2 \n\t" \
151 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000152 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
153 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
154 : "eax", "ecx", "edx", "esi", "edi" \
155 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000156#endif /* SSE2 */
157#endif /* i386 */
158
159#if defined(__amd64__) || defined (__x86_64__)
160
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200161#define MULADDC_INIT \
162 asm( \
163 "movq %3, %%rsi \n\t" \
164 "movq %4, %%rdi \n\t" \
165 "movq %5, %%rcx \n\t" \
166 "movq %6, %%rbx \n\t" \
167 "xorq %%r8, %%r8 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000168
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200169#define MULADDC_CORE \
170 "movq (%%rsi), %%rax \n\t" \
171 "mulq %%rbx \n\t" \
172 "addq $8, %%rsi \n\t" \
173 "addq %%rcx, %%rax \n\t" \
174 "movq %%r8, %%rcx \n\t" \
175 "adcq $0, %%rdx \n\t" \
176 "nop \n\t" \
177 "addq %%rax, (%%rdi) \n\t" \
178 "adcq %%rdx, %%rcx \n\t" \
179 "addq $8, %%rdi \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000180
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200181#define MULADDC_STOP \
182 "movq %%rcx, %0 \n\t" \
183 "movq %%rdi, %1 \n\t" \
184 "movq %%rsi, %2 \n\t" \
Manuel Pégourié-Gonnarddef018d2014-01-07 17:50:46 +0100185 : "=m" (c), "=m" (d), "=m" (s) \
186 : "m" (s), "m" (d), "m" (c), "m" (b) \
187 : "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" \
188 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000189
190#endif /* AMD64 */
191
192#if defined(__mc68020__) || defined(__mcpu32__)
193
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200194#define MULADDC_INIT \
195 asm( \
196 "movl %3, %%a2 \n\t" \
197 "movl %4, %%a3 \n\t" \
198 "movl %5, %%d3 \n\t" \
199 "movl %6, %%d2 \n\t" \
200 "moveq #0, %%d0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000201
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200202#define MULADDC_CORE \
203 "movel %%a2@+, %%d1 \n\t" \
204 "mulul %%d2, %%d4:%%d1 \n\t" \
205 "addl %%d3, %%d1 \n\t" \
206 "addxl %%d0, %%d4 \n\t" \
207 "moveq #0, %%d3 \n\t" \
208 "addl %%d1, %%a3@+ \n\t" \
209 "addxl %%d4, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000210
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200211#define MULADDC_STOP \
212 "movl %%d3, %0 \n\t" \
213 "movl %%a3, %1 \n\t" \
214 "movl %%a2, %2 \n\t" \
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100215 : "=m" (c), "=m" (d), "=m" (s) \
216 : "m" (s), "m" (d), "m" (c), "m" (b) \
217 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
218 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000219
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200220#define MULADDC_HUIT \
221 "movel %%a2@+, %%d1 \n\t" \
222 "mulul %%d2, %%d4:%%d1 \n\t" \
223 "addxl %%d3, %%d1 \n\t" \
224 "addxl %%d0, %%d4 \n\t" \
225 "addl %%d1, %%a3@+ \n\t" \
226 "movel %%a2@+, %%d1 \n\t" \
227 "mulul %%d2, %%d3:%%d1 \n\t" \
228 "addxl %%d4, %%d1 \n\t" \
229 "addxl %%d0, %%d3 \n\t" \
230 "addl %%d1, %%a3@+ \n\t" \
231 "movel %%a2@+, %%d1 \n\t" \
232 "mulul %%d2, %%d4:%%d1 \n\t" \
233 "addxl %%d3, %%d1 \n\t" \
234 "addxl %%d0, %%d4 \n\t" \
235 "addl %%d1, %%a3@+ \n\t" \
236 "movel %%a2@+, %%d1 \n\t" \
237 "mulul %%d2, %%d3:%%d1 \n\t" \
238 "addxl %%d4, %%d1 \n\t" \
239 "addxl %%d0, %%d3 \n\t" \
240 "addl %%d1, %%a3@+ \n\t" \
241 "movel %%a2@+, %%d1 \n\t" \
242 "mulul %%d2, %%d4:%%d1 \n\t" \
243 "addxl %%d3, %%d1 \n\t" \
244 "addxl %%d0, %%d4 \n\t" \
245 "addl %%d1, %%a3@+ \n\t" \
246 "movel %%a2@+, %%d1 \n\t" \
247 "mulul %%d2, %%d3:%%d1 \n\t" \
248 "addxl %%d4, %%d1 \n\t" \
249 "addxl %%d0, %%d3 \n\t" \
250 "addl %%d1, %%a3@+ \n\t" \
251 "movel %%a2@+, %%d1 \n\t" \
252 "mulul %%d2, %%d4:%%d1 \n\t" \
253 "addxl %%d3, %%d1 \n\t" \
254 "addxl %%d0, %%d4 \n\t" \
255 "addl %%d1, %%a3@+ \n\t" \
256 "movel %%a2@+, %%d1 \n\t" \
257 "mulul %%d2, %%d3:%%d1 \n\t" \
258 "addxl %%d4, %%d1 \n\t" \
259 "addxl %%d0, %%d3 \n\t" \
260 "addl %%d1, %%a3@+ \n\t" \
261 "addxl %%d0, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000262
263#endif /* MC68000 */
264
Paul Bakker5121ce52009-01-03 21:22:43 +0000265#if defined(__powerpc64__) || defined(__ppc64__)
266
267#if defined(__MACH__) && defined(__APPLE__)
268
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200269#define MULADDC_INIT \
270 asm( \
271 "ld r3, %3 \n\t" \
272 "ld r4, %4 \n\t" \
273 "ld r5, %5 \n\t" \
274 "ld r6, %6 \n\t" \
275 "addi r3, r3, -8 \n\t" \
276 "addi r4, r4, -8 \n\t" \
277 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000278
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200279#define MULADDC_CORE \
280 "ldu r7, 8(r3) \n\t" \
281 "mulld r8, r7, r6 \n\t" \
282 "mulhdu r9, r7, r6 \n\t" \
283 "adde r8, r8, r5 \n\t" \
284 "ld r7, 8(r4) \n\t" \
285 "addze r5, r9 \n\t" \
286 "addc r8, r8, r7 \n\t" \
287 "stdu r8, 8(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000288
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200289#define MULADDC_STOP \
290 "addze r5, r5 \n\t" \
291 "addi r4, r4, 8 \n\t" \
292 "addi r3, r3, 8 \n\t" \
293 "std r5, %0 \n\t" \
294 "std r4, %1 \n\t" \
295 "std r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100296 : "=m" (c), "=m" (d), "=m" (s) \
297 : "m" (s), "m" (d), "m" (c), "m" (b) \
298 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
299 );
300
Paul Bakker5121ce52009-01-03 21:22:43 +0000301
Paul Bakker9af723c2014-05-01 13:03:14 +0200302#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000303
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200304#define MULADDC_INIT \
305 asm( \
306 "ld %%r3, %3 \n\t" \
307 "ld %%r4, %4 \n\t" \
308 "ld %%r5, %5 \n\t" \
309 "ld %%r6, %6 \n\t" \
310 "addi %%r3, %%r3, -8 \n\t" \
311 "addi %%r4, %%r4, -8 \n\t" \
312 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000313
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200314#define MULADDC_CORE \
315 "ldu %%r7, 8(%%r3) \n\t" \
316 "mulld %%r8, %%r7, %%r6 \n\t" \
317 "mulhdu %%r9, %%r7, %%r6 \n\t" \
318 "adde %%r8, %%r8, %%r5 \n\t" \
319 "ld %%r7, 8(%%r4) \n\t" \
320 "addze %%r5, %%r9 \n\t" \
321 "addc %%r8, %%r8, %%r7 \n\t" \
322 "stdu %%r8, 8(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000323
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200324#define MULADDC_STOP \
325 "addze %%r5, %%r5 \n\t" \
326 "addi %%r4, %%r4, 8 \n\t" \
327 "addi %%r3, %%r3, 8 \n\t" \
328 "std %%r5, %0 \n\t" \
329 "std %%r4, %1 \n\t" \
330 "std %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100331 : "=m" (c), "=m" (d), "=m" (s) \
332 : "m" (s), "m" (d), "m" (c), "m" (b) \
333 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
334 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000335
Paul Bakker9af723c2014-05-01 13:03:14 +0200336#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000337
Barry K. Nathan35e7cb92014-05-05 23:26:13 -0700338#elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000339
340#if defined(__MACH__) && defined(__APPLE__)
341
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200342#define MULADDC_INIT \
343 asm( \
344 "lwz r3, %3 \n\t" \
345 "lwz r4, %4 \n\t" \
346 "lwz r5, %5 \n\t" \
347 "lwz r6, %6 \n\t" \
348 "addi r3, r3, -4 \n\t" \
349 "addi r4, r4, -4 \n\t" \
350 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000351
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200352#define MULADDC_CORE \
353 "lwzu r7, 4(r3) \n\t" \
354 "mullw r8, r7, r6 \n\t" \
355 "mulhwu r9, r7, r6 \n\t" \
356 "adde r8, r8, r5 \n\t" \
357 "lwz r7, 4(r4) \n\t" \
358 "addze r5, r9 \n\t" \
359 "addc r8, r8, r7 \n\t" \
360 "stwu r8, 4(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000361
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200362#define MULADDC_STOP \
363 "addze r5, r5 \n\t" \
364 "addi r4, r4, 4 \n\t" \
365 "addi r3, r3, 4 \n\t" \
366 "stw r5, %0 \n\t" \
367 "stw r4, %1 \n\t" \
368 "stw r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100369 : "=m" (c), "=m" (d), "=m" (s) \
370 : "m" (s), "m" (d), "m" (c), "m" (b) \
371 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
372 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000373
Paul Bakker9af723c2014-05-01 13:03:14 +0200374#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000375
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200376#define MULADDC_INIT \
377 asm( \
378 "lwz %%r3, %3 \n\t" \
379 "lwz %%r4, %4 \n\t" \
380 "lwz %%r5, %5 \n\t" \
381 "lwz %%r6, %6 \n\t" \
382 "addi %%r3, %%r3, -4 \n\t" \
383 "addi %%r4, %%r4, -4 \n\t" \
384 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000385
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200386#define MULADDC_CORE \
387 "lwzu %%r7, 4(%%r3) \n\t" \
388 "mullw %%r8, %%r7, %%r6 \n\t" \
389 "mulhwu %%r9, %%r7, %%r6 \n\t" \
390 "adde %%r8, %%r8, %%r5 \n\t" \
391 "lwz %%r7, 4(%%r4) \n\t" \
392 "addze %%r5, %%r9 \n\t" \
393 "addc %%r8, %%r8, %%r7 \n\t" \
394 "stwu %%r8, 4(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000395
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200396#define MULADDC_STOP \
397 "addze %%r5, %%r5 \n\t" \
398 "addi %%r4, %%r4, 4 \n\t" \
399 "addi %%r3, %%r3, 4 \n\t" \
400 "stw %%r5, %0 \n\t" \
401 "stw %%r4, %1 \n\t" \
402 "stw %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100403 : "=m" (c), "=m" (d), "=m" (s) \
404 : "m" (s), "m" (d), "m" (c), "m" (b) \
405 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
406 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000407
Paul Bakker9af723c2014-05-01 13:03:14 +0200408#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000409
410#endif /* PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000411
Paul Bakker4f024b72012-10-30 07:29:57 +0000412#if defined(__sparc__) && defined(__sparc64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000413
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200414#define MULADDC_INIT \
415 asm( \
416 "ldx %3, %%o0 \n\t" \
417 "ldx %4, %%o1 \n\t" \
418 "ld %5, %%o2 \n\t" \
419 "ld %6, %%o3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000420
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200421#define MULADDC_CORE \
422 "ld [%%o0], %%o4 \n\t" \
423 "inc 4, %%o0 \n\t" \
424 "ld [%%o1], %%o5 \n\t" \
425 "umul %%o3, %%o4, %%o4 \n\t" \
426 "addcc %%o4, %%o2, %%o4 \n\t" \
427 "rd %%y, %%g1 \n\t" \
428 "addx %%g1, 0, %%g1 \n\t" \
429 "addcc %%o4, %%o5, %%o4 \n\t" \
430 "st %%o4, [%%o1] \n\t" \
431 "addx %%g1, 0, %%o2 \n\t" \
432 "inc 4, %%o1 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000433
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200434 #define MULADDC_STOP \
435 "st %%o2, %0 \n\t" \
436 "stx %%o1, %1 \n\t" \
437 "stx %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000438 : "=m" (c), "=m" (d), "=m" (s) \
439 : "m" (s), "m" (d), "m" (c), "m" (b) \
440 : "g1", "o0", "o1", "o2", "o3", "o4", \
441 "o5" \
442 );
443#endif /* SPARCv9 */
444
445#if defined(__sparc__) && !defined(__sparc64__)
446
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200447#define MULADDC_INIT \
448 asm( \
449 "ld %3, %%o0 \n\t" \
450 "ld %4, %%o1 \n\t" \
451 "ld %5, %%o2 \n\t" \
452 "ld %6, %%o3 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000453
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200454#define MULADDC_CORE \
455 "ld [%%o0], %%o4 \n\t" \
456 "inc 4, %%o0 \n\t" \
457 "ld [%%o1], %%o5 \n\t" \
458 "umul %%o3, %%o4, %%o4 \n\t" \
459 "addcc %%o4, %%o2, %%o4 \n\t" \
460 "rd %%y, %%g1 \n\t" \
461 "addx %%g1, 0, %%g1 \n\t" \
462 "addcc %%o4, %%o5, %%o4 \n\t" \
463 "st %%o4, [%%o1] \n\t" \
464 "addx %%g1, 0, %%o2 \n\t" \
465 "inc 4, %%o1 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000466
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200467#define MULADDC_STOP \
468 "st %%o2, %0 \n\t" \
469 "st %%o1, %1 \n\t" \
470 "st %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000471 : "=m" (c), "=m" (d), "=m" (s) \
472 : "m" (s), "m" (d), "m" (c), "m" (b) \
473 : "g1", "o0", "o1", "o2", "o3", "o4", \
474 "o5" \
475 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000476
477#endif /* SPARCv8 */
478
479#if defined(__microblaze__) || defined(microblaze)
480
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200481#define MULADDC_INIT \
482 asm( \
483 "lwi r3, %3 \n\t" \
484 "lwi r4, %4 \n\t" \
485 "lwi r5, %5 \n\t" \
486 "lwi r6, %6 \n\t" \
487 "andi r7, r6, 0xffff \n\t" \
488 "bsrli r6, r6, 16 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000489
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200490#define MULADDC_CORE \
491 "lhui r8, r3, 0 \n\t" \
492 "addi r3, r3, 2 \n\t" \
493 "lhui r9, r3, 0 \n\t" \
494 "addi r3, r3, 2 \n\t" \
495 "mul r10, r9, r6 \n\t" \
496 "mul r11, r8, r7 \n\t" \
497 "mul r12, r9, r7 \n\t" \
498 "mul r13, r8, r6 \n\t" \
499 "bsrli r8, r10, 16 \n\t" \
500 "bsrli r9, r11, 16 \n\t" \
501 "add r13, r13, r8 \n\t" \
502 "add r13, r13, r9 \n\t" \
503 "bslli r10, r10, 16 \n\t" \
504 "bslli r11, r11, 16 \n\t" \
505 "add r12, r12, r10 \n\t" \
506 "addc r13, r13, r0 \n\t" \
507 "add r12, r12, r11 \n\t" \
508 "addc r13, r13, r0 \n\t" \
509 "lwi r10, r4, 0 \n\t" \
510 "add r12, r12, r10 \n\t" \
511 "addc r13, r13, r0 \n\t" \
512 "add r12, r12, r5 \n\t" \
513 "addc r5, r13, r0 \n\t" \
514 "swi r12, r4, 0 \n\t" \
515 "addi r4, r4, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000516
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200517#define MULADDC_STOP \
518 "swi r5, %0 \n\t" \
519 "swi r4, %1 \n\t" \
520 "swi r3, %2 \n\t" \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100521 : "=m" (c), "=m" (d), "=m" (s) \
522 : "m" (s), "m" (d), "m" (c), "m" (b) \
523 : "r3", "r4" "r5", "r6", "r7", "r8", \
524 "r9", "r10", "r11", "r12", "r13" \
525 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000526
527#endif /* MicroBlaze */
528
529#if defined(__tricore__)
530
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200531#define MULADDC_INIT \
532 asm( \
533 "ld.a %%a2, %3 \n\t" \
534 "ld.a %%a3, %4 \n\t" \
535 "ld.w %%d4, %5 \n\t" \
536 "ld.w %%d1, %6 \n\t" \
537 "xor %%d5, %%d5 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000538
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200539#define MULADDC_CORE \
540 "ld.w %%d0, [%%a2+] \n\t" \
541 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
542 "ld.w %%d0, [%%a3] \n\t" \
543 "addx %%d2, %%d2, %%d0 \n\t" \
544 "addc %%d3, %%d3, 0 \n\t" \
545 "mov %%d4, %%d3 \n\t" \
546 "st.w [%%a3+], %%d2 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000547
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200548#define MULADDC_STOP \
549 "st.w %0, %%d4 \n\t" \
550 "st.a %1, %%a3 \n\t" \
551 "st.a %2, %%a2 \n\t" \
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100552 : "=m" (c), "=m" (d), "=m" (s) \
553 : "m" (s), "m" (d), "m" (c), "m" (b) \
554 : "d0", "d1", "e2", "d4", "a2", "a3" \
555 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000556
557#endif /* TriCore */
558
559#if defined(__arm__)
560
Paul Bakkerfc4f46f2013-06-24 19:23:56 +0200561#if defined(__thumb__) && !defined(__thumb2__)
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000562
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200563#define MULADDC_INIT \
564 asm( \
565 "ldr r0, %3 \n\t" \
566 "ldr r1, %4 \n\t" \
567 "ldr r2, %5 \n\t" \
568 "ldr r3, %6 \n\t" \
569 "lsr r7, r3, #16 \n\t" \
570 "mov r9, r7 \n\t" \
571 "lsl r7, r3, #16 \n\t" \
572 "lsr r7, r7, #16 \n\t" \
573 "mov r8, r7 \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000574
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200575#define MULADDC_CORE \
576 "ldmia r0!, {r6} \n\t" \
577 "lsr r7, r6, #16 \n\t" \
578 "lsl r6, r6, #16 \n\t" \
579 "lsr r6, r6, #16 \n\t" \
580 "mov r4, r8 \n\t" \
581 "mul r4, r6 \n\t" \
582 "mov r3, r9 \n\t" \
583 "mul r6, r3 \n\t" \
584 "mov r5, r9 \n\t" \
585 "mul r5, r7 \n\t" \
586 "mov r3, r8 \n\t" \
587 "mul r7, r3 \n\t" \
588 "lsr r3, r6, #16 \n\t" \
589 "add r5, r5, r3 \n\t" \
590 "lsr r3, r7, #16 \n\t" \
591 "add r5, r5, r3 \n\t" \
592 "add r4, r4, r2 \n\t" \
593 "mov r2, #0 \n\t" \
594 "adc r5, r2 \n\t" \
595 "lsl r3, r6, #16 \n\t" \
596 "add r4, r4, r3 \n\t" \
597 "adc r5, r2 \n\t" \
598 "lsl r3, r7, #16 \n\t" \
599 "add r4, r4, r3 \n\t" \
600 "adc r5, r2 \n\t" \
601 "ldr r3, [r1] \n\t" \
602 "add r4, r4, r3 \n\t" \
603 "adc r2, r5 \n\t" \
604 "stmia r1!, {r4} \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000605
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200606#define MULADDC_STOP \
607 "str r2, %0 \n\t" \
608 "str r1, %1 \n\t" \
609 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100610 : "=m" (c), "=m" (d), "=m" (s) \
611 : "m" (s), "m" (d), "m" (c), "m" (b) \
612 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200613 "r6", "r7", "r8", "r9", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100614 );
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000615
616#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000617
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200618#define MULADDC_INIT \
619 asm( \
620 "ldr r0, %3 \n\t" \
621 "ldr r1, %4 \n\t" \
622 "ldr r2, %5 \n\t" \
623 "ldr r3, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000624
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200625#define MULADDC_CORE \
626 "ldr r4, [r0], #4 \n\t" \
627 "mov r5, #0 \n\t" \
628 "ldr r6, [r1] \n\t" \
629 "umlal r2, r5, r3, r4 \n\t" \
630 "adds r7, r6, r2 \n\t" \
631 "adc r2, r5, #0 \n\t" \
632 "str r7, [r1], #4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000633
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200634#define MULADDC_STOP \
635 "str r2, %0 \n\t" \
636 "str r1, %1 \n\t" \
637 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100638 : "=m" (c), "=m" (d), "=m" (s) \
639 : "m" (s), "m" (d), "m" (c), "m" (b) \
640 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200641 "r6", "r7", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100642 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000643
Paul Bakkera2713a32011-11-18 12:47:23 +0000644#endif /* Thumb */
645
Paul Bakker5121ce52009-01-03 21:22:43 +0000646#endif /* ARMv3 */
647
648#if defined(__alpha__)
649
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200650#define MULADDC_INIT \
651 asm( \
652 "ldq $1, %3 \n\t" \
653 "ldq $2, %4 \n\t" \
654 "ldq $3, %5 \n\t" \
655 "ldq $4, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000656
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200657#define MULADDC_CORE \
658 "ldq $6, 0($1) \n\t" \
659 "addq $1, 8, $1 \n\t" \
660 "mulq $6, $4, $7 \n\t" \
661 "umulh $6, $4, $6 \n\t" \
662 "addq $7, $3, $7 \n\t" \
663 "cmpult $7, $3, $3 \n\t" \
664 "ldq $5, 0($2) \n\t" \
665 "addq $7, $5, $7 \n\t" \
666 "cmpult $7, $5, $5 \n\t" \
667 "stq $7, 0($2) \n\t" \
668 "addq $2, 8, $2 \n\t" \
669 "addq $6, $3, $3 \n\t" \
670 "addq $5, $3, $3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000671
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200672#define MULADDC_STOP \
673 "stq $3, %0 \n\t" \
674 "stq $2, %1 \n\t" \
675 "stq $1, %2 \n\t" \
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100676 : "=m" (c), "=m" (d), "=m" (s) \
677 : "m" (s), "m" (d), "m" (c), "m" (b) \
678 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
679 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000680#endif /* Alpha */
681
Alex Wilson73491422014-03-06 00:04:09 +1000682#if defined(__mips__) && !defined(__mips64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000683
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200684#define MULADDC_INIT \
685 asm( \
686 "lw $10, %3 \n\t" \
687 "lw $11, %4 \n\t" \
688 "lw $12, %5 \n\t" \
689 "lw $13, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000690
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200691#define MULADDC_CORE \
692 "lw $14, 0($10) \n\t" \
693 "multu $13, $14 \n\t" \
694 "addi $10, $10, 4 \n\t" \
695 "mflo $14 \n\t" \
696 "mfhi $9 \n\t" \
697 "addu $14, $12, $14 \n\t" \
698 "lw $15, 0($11) \n\t" \
699 "sltu $12, $14, $12 \n\t" \
700 "addu $15, $14, $15 \n\t" \
701 "sltu $14, $15, $14 \n\t" \
702 "addu $12, $12, $9 \n\t" \
703 "sw $15, 0($11) \n\t" \
704 "addu $12, $12, $14 \n\t" \
705 "addi $11, $11, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000706
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200707#define MULADDC_STOP \
708 "sw $12, %0 \n\t" \
709 "sw $11, %1 \n\t" \
710 "sw $10, %2 \n\t" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100711 : "=m" (c), "=m" (d), "=m" (s) \
712 : "m" (s), "m" (d), "m" (c), "m" (b) \
713 : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
714 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000715
716#endif /* MIPS */
717#endif /* GNUC */
718
719#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
720
721#define MULADDC_INIT \
722 __asm mov esi, s \
723 __asm mov edi, d \
724 __asm mov ecx, c \
725 __asm mov ebx, b
726
727#define MULADDC_CORE \
728 __asm lodsd \
729 __asm mul ebx \
730 __asm add eax, ecx \
731 __asm adc edx, 0 \
732 __asm add eax, [edi] \
733 __asm adc edx, 0 \
734 __asm mov ecx, edx \
735 __asm stosd
736
Paul Bakker40e46942009-01-03 21:51:57 +0000737#if defined(POLARSSL_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000738
739#define EMIT __asm _emit
740
741#define MULADDC_HUIT \
742 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
743 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
744 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
745 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
746 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
747 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
748 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
749 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
750 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
751 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
752 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
753 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
754 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
755 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
756 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
757 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
758 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
759 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
760 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
761 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
762 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
763 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
764 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
765 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
766 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
767 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
768 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
769 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
770 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
771 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
772 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
773 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
774 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
775 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
776 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
777 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
778 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
779 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
780 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
781 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
782 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
783 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
784 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
785 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
786 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
787 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
788 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
789 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
790 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
791 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
792 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
793 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
794 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
795 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
796 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
797 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
798 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
799 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
800 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
801 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
802 EMIT 0x0F EMIT 0x7E EMIT 0xC9
803
804#define MULADDC_STOP \
805 EMIT 0x0F EMIT 0x77 \
806 __asm mov c, ecx \
807 __asm mov d, edi \
808 __asm mov s, esi \
809
810#else
811
812#define MULADDC_STOP \
813 __asm mov c, ecx \
814 __asm mov d, edi \
815 __asm mov s, esi \
816
817#endif /* SSE2 */
818#endif /* MSVC */
819
Paul Bakker40e46942009-01-03 21:51:57 +0000820#endif /* POLARSSL_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000821
822#if !defined(MULADDC_CORE)
Paul Bakker0fd018e2012-10-23 12:44:47 +0000823#if defined(POLARSSL_HAVE_UDBL)
Paul Bakker5121ce52009-01-03 21:22:43 +0000824
825#define MULADDC_INIT \
826{ \
Paul Bakkera755ca12011-04-24 09:11:17 +0000827 t_udbl r; \
828 t_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000829
830#define MULADDC_CORE \
Paul Bakker5c2364c2012-10-01 14:41:15 +0000831 r = *(s++) * (t_udbl) b; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000832 r0 = r; \
833 r1 = r >> biL; \
834 r0 += c; r1 += (r0 < c); \
835 r0 += *d; r1 += (r0 < *d); \
836 c = r1; *(d++) = r0;
837
838#define MULADDC_STOP \
839}
840
841#else
842#define MULADDC_INIT \
843{ \
Paul Bakkera755ca12011-04-24 09:11:17 +0000844 t_uint s0, s1, b0, b1; \
845 t_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000846 b0 = ( b << biH ) >> biH; \
847 b1 = ( b >> biH );
848
849#define MULADDC_CORE \
850 s0 = ( *s << biH ) >> biH; \
851 s1 = ( *s >> biH ); s++; \
852 rx = s0 * b1; r0 = s0 * b0; \
853 ry = s1 * b0; r1 = s1 * b1; \
854 r1 += ( rx >> biH ); \
855 r1 += ( ry >> biH ); \
856 rx <<= biH; ry <<= biH; \
857 r0 += rx; r1 += (r0 < rx); \
858 r0 += ry; r1 += (r0 < ry); \
859 r0 += c; r1 += (r0 < c); \
860 r0 += *d; r1 += (r0 < *d); \
861 c = r1; *(d++) = r0;
862
863#define MULADDC_STOP \
864}
865
866#endif /* C (generic) */
867#endif /* C (longlong) */
868
869#endif /* bn_mul.h */