blob: b631ad27885967c7aa5cc00cb17c663438e2fa60 [file] [log] [blame]
Paul Bakker5121ce52009-01-03 21:22:43 +00001/**
2 * \file bn_mul.h
Paul Bakkere0ccd0a2009-01-04 16:27:10 +00003 *
Darryl Greena40a1012018-01-05 15:33:17 +00004 * \brief Multi-precision integer library
5 */
6/*
Manuel Pégourié-Gonnard6fb81872015-07-27 11:11:48 +02007 * Copyright (C) 2006-2015, ARM Limited, All Rights Reserved
Manuel Pégourié-Gonnard37ff1402015-09-04 14:21:07 +02008 * SPDX-License-Identifier: Apache-2.0
9 *
10 * Licensed under the Apache License, Version 2.0 (the "License"); you may
11 * not use this file except in compliance with the License.
12 * You may obtain a copy of the License at
13 *
14 * http://www.apache.org/licenses/LICENSE-2.0
15 *
16 * Unless required by applicable law or agreed to in writing, software
17 * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
18 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
19 * See the License for the specific language governing permissions and
20 * limitations under the License.
Paul Bakkerb96f1542010-07-18 20:36:00 +000021 *
Manuel Pégourié-Gonnardfe446432015-03-06 13:17:10 +000022 * This file is part of mbed TLS (https://tls.mbed.org)
Paul Bakker5121ce52009-01-03 21:22:43 +000023 */
24/*
25 * Multiply source vector [s] with b, add result
26 * to destination vector [d] and set carry c.
27 *
28 * Currently supports:
29 *
30 * . IA-32 (386+) . AMD64 / EM64T
31 * . IA-32 (SSE2) . Motorola 68000
32 * . PowerPC, 32-bit . MicroBlaze
33 * . PowerPC, 64-bit . TriCore
34 * . SPARC v8 . ARM v3+
35 * . Alpha . MIPS32
36 * . C, longlong . C, generic
37 */
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +020038#ifndef MBEDTLS_BN_MUL_H
39#define MBEDTLS_BN_MUL_H
Paul Bakker5121ce52009-01-03 21:22:43 +000040
Paul Bakker66219872012-01-22 20:38:13 +000041#include "bignum.h"
Paul Bakker5121ce52009-01-03 21:22:43 +000042
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +020043#if defined(MBEDTLS_HAVE_ASM)
Paul Bakker5121ce52009-01-03 21:22:43 +000044
Manuel Pégourié-Gonnardba194322015-05-29 09:47:57 +020045#ifndef asm
46#define asm __asm
47#endif
48
Manuel Pégourié-Gonnard854dab92015-08-10 12:08:34 +020049/* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */
50#if defined(__GNUC__) && \
51 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
Paul Bakker5121ce52009-01-03 21:22:43 +000052#if defined(__i386__)
53
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +020054#define MULADDC_INIT \
55 asm( \
56 "movl %%ebx, %0 \n\t" \
57 "movl %5, %%esi \n\t" \
58 "movl %6, %%edi \n\t" \
59 "movl %7, %%ecx \n\t" \
60 "movl %8, %%ebx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +000061
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +020062#define MULADDC_CORE \
63 "lodsl \n\t" \
64 "mull %%ebx \n\t" \
65 "addl %%ecx, %%eax \n\t" \
66 "adcl $0, %%edx \n\t" \
67 "addl (%%edi), %%eax \n\t" \
68 "adcl $0, %%edx \n\t" \
69 "movl %%edx, %%ecx \n\t" \
70 "stosl \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +000071
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +020072#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +000073
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +020074#define MULADDC_HUIT \
75 "movd %%ecx, %%mm1 \n\t" \
76 "movd %%ebx, %%mm0 \n\t" \
77 "movd (%%edi), %%mm3 \n\t" \
78 "paddq %%mm3, %%mm1 \n\t" \
79 "movd (%%esi), %%mm2 \n\t" \
80 "pmuludq %%mm0, %%mm2 \n\t" \
81 "movd 4(%%esi), %%mm4 \n\t" \
82 "pmuludq %%mm0, %%mm4 \n\t" \
83 "movd 8(%%esi), %%mm6 \n\t" \
84 "pmuludq %%mm0, %%mm6 \n\t" \
85 "movd 12(%%esi), %%mm7 \n\t" \
86 "pmuludq %%mm0, %%mm7 \n\t" \
87 "paddq %%mm2, %%mm1 \n\t" \
88 "movd 4(%%edi), %%mm3 \n\t" \
89 "paddq %%mm4, %%mm3 \n\t" \
90 "movd 8(%%edi), %%mm5 \n\t" \
91 "paddq %%mm6, %%mm5 \n\t" \
92 "movd 12(%%edi), %%mm4 \n\t" \
93 "paddq %%mm4, %%mm7 \n\t" \
94 "movd %%mm1, (%%edi) \n\t" \
95 "movd 16(%%esi), %%mm2 \n\t" \
96 "pmuludq %%mm0, %%mm2 \n\t" \
97 "psrlq $32, %%mm1 \n\t" \
98 "movd 20(%%esi), %%mm4 \n\t" \
99 "pmuludq %%mm0, %%mm4 \n\t" \
100 "paddq %%mm3, %%mm1 \n\t" \
101 "movd 24(%%esi), %%mm6 \n\t" \
102 "pmuludq %%mm0, %%mm6 \n\t" \
103 "movd %%mm1, 4(%%edi) \n\t" \
104 "psrlq $32, %%mm1 \n\t" \
105 "movd 28(%%esi), %%mm3 \n\t" \
106 "pmuludq %%mm0, %%mm3 \n\t" \
107 "paddq %%mm5, %%mm1 \n\t" \
108 "movd 16(%%edi), %%mm5 \n\t" \
109 "paddq %%mm5, %%mm2 \n\t" \
110 "movd %%mm1, 8(%%edi) \n\t" \
111 "psrlq $32, %%mm1 \n\t" \
112 "paddq %%mm7, %%mm1 \n\t" \
113 "movd 20(%%edi), %%mm5 \n\t" \
114 "paddq %%mm5, %%mm4 \n\t" \
115 "movd %%mm1, 12(%%edi) \n\t" \
116 "psrlq $32, %%mm1 \n\t" \
117 "paddq %%mm2, %%mm1 \n\t" \
118 "movd 24(%%edi), %%mm5 \n\t" \
119 "paddq %%mm5, %%mm6 \n\t" \
120 "movd %%mm1, 16(%%edi) \n\t" \
121 "psrlq $32, %%mm1 \n\t" \
122 "paddq %%mm4, %%mm1 \n\t" \
123 "movd 28(%%edi), %%mm5 \n\t" \
124 "paddq %%mm5, %%mm3 \n\t" \
125 "movd %%mm1, 20(%%edi) \n\t" \
126 "psrlq $32, %%mm1 \n\t" \
127 "paddq %%mm6, %%mm1 \n\t" \
128 "movd %%mm1, 24(%%edi) \n\t" \
129 "psrlq $32, %%mm1 \n\t" \
130 "paddq %%mm3, %%mm1 \n\t" \
131 "movd %%mm1, 28(%%edi) \n\t" \
132 "addl $32, %%edi \n\t" \
133 "addl $32, %%esi \n\t" \
134 "psrlq $32, %%mm1 \n\t" \
135 "movd %%mm1, %%ecx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000136
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200137#define MULADDC_STOP \
138 "emms \n\t" \
139 "movl %4, %%ebx \n\t" \
140 "movl %%ecx, %1 \n\t" \
141 "movl %%edi, %2 \n\t" \
142 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000143 : "=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 +0000147
148#else
149
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200150#define MULADDC_STOP \
151 "movl %4, %%ebx \n\t" \
152 "movl %%ecx, %1 \n\t" \
153 "movl %%edi, %2 \n\t" \
154 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000155 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
156 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
157 : "eax", "ecx", "edx", "esi", "edi" \
158 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000159#endif /* SSE2 */
160#endif /* i386 */
161
162#if defined(__amd64__) || defined (__x86_64__)
163
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200164#define MULADDC_INIT \
165 asm( \
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200166 "xorq %%r8, %%r8 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000167
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200168#define MULADDC_CORE \
169 "movq (%%rsi), %%rax \n\t" \
170 "mulq %%rbx \n\t" \
171 "addq $8, %%rsi \n\t" \
172 "addq %%rcx, %%rax \n\t" \
173 "movq %%r8, %%rcx \n\t" \
174 "adcq $0, %%rdx \n\t" \
175 "nop \n\t" \
176 "addq %%rax, (%%rdi) \n\t" \
177 "adcq %%rdx, %%rcx \n\t" \
178 "addq $8, %%rdi \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000179
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200180#define MULADDC_STOP \
James Cowgill21e402a2015-12-17 01:51:09 +0000181 : "+c" (c), "+D" (d), "+S" (s) \
182 : "b" (b) \
183 : "rax", "rdx", "r8" \
Manuel Pégourié-Gonnarddef018d2014-01-07 17:50:46 +0100184 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000185
186#endif /* AMD64 */
187
188#if defined(__mc68020__) || defined(__mcpu32__)
189
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200190#define MULADDC_INIT \
191 asm( \
192 "movl %3, %%a2 \n\t" \
193 "movl %4, %%a3 \n\t" \
194 "movl %5, %%d3 \n\t" \
195 "movl %6, %%d2 \n\t" \
196 "moveq #0, %%d0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000197
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200198#define MULADDC_CORE \
199 "movel %%a2@+, %%d1 \n\t" \
200 "mulul %%d2, %%d4:%%d1 \n\t" \
201 "addl %%d3, %%d1 \n\t" \
202 "addxl %%d0, %%d4 \n\t" \
203 "moveq #0, %%d3 \n\t" \
204 "addl %%d1, %%a3@+ \n\t" \
205 "addxl %%d4, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000206
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200207#define MULADDC_STOP \
208 "movl %%d3, %0 \n\t" \
209 "movl %%a3, %1 \n\t" \
210 "movl %%a2, %2 \n\t" \
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100211 : "=m" (c), "=m" (d), "=m" (s) \
212 : "m" (s), "m" (d), "m" (c), "m" (b) \
213 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
214 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000215
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200216#define MULADDC_HUIT \
217 "movel %%a2@+, %%d1 \n\t" \
218 "mulul %%d2, %%d4:%%d1 \n\t" \
219 "addxl %%d3, %%d1 \n\t" \
220 "addxl %%d0, %%d4 \n\t" \
221 "addl %%d1, %%a3@+ \n\t" \
222 "movel %%a2@+, %%d1 \n\t" \
223 "mulul %%d2, %%d3:%%d1 \n\t" \
224 "addxl %%d4, %%d1 \n\t" \
225 "addxl %%d0, %%d3 \n\t" \
226 "addl %%d1, %%a3@+ \n\t" \
227 "movel %%a2@+, %%d1 \n\t" \
228 "mulul %%d2, %%d4:%%d1 \n\t" \
229 "addxl %%d3, %%d1 \n\t" \
230 "addxl %%d0, %%d4 \n\t" \
231 "addl %%d1, %%a3@+ \n\t" \
232 "movel %%a2@+, %%d1 \n\t" \
233 "mulul %%d2, %%d3:%%d1 \n\t" \
234 "addxl %%d4, %%d1 \n\t" \
235 "addxl %%d0, %%d3 \n\t" \
236 "addl %%d1, %%a3@+ \n\t" \
237 "movel %%a2@+, %%d1 \n\t" \
238 "mulul %%d2, %%d4:%%d1 \n\t" \
239 "addxl %%d3, %%d1 \n\t" \
240 "addxl %%d0, %%d4 \n\t" \
241 "addl %%d1, %%a3@+ \n\t" \
242 "movel %%a2@+, %%d1 \n\t" \
243 "mulul %%d2, %%d3:%%d1 \n\t" \
244 "addxl %%d4, %%d1 \n\t" \
245 "addxl %%d0, %%d3 \n\t" \
246 "addl %%d1, %%a3@+ \n\t" \
247 "movel %%a2@+, %%d1 \n\t" \
248 "mulul %%d2, %%d4:%%d1 \n\t" \
249 "addxl %%d3, %%d1 \n\t" \
250 "addxl %%d0, %%d4 \n\t" \
251 "addl %%d1, %%a3@+ \n\t" \
252 "movel %%a2@+, %%d1 \n\t" \
253 "mulul %%d2, %%d3:%%d1 \n\t" \
254 "addxl %%d4, %%d1 \n\t" \
255 "addxl %%d0, %%d3 \n\t" \
256 "addl %%d1, %%a3@+ \n\t" \
257 "addxl %%d0, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000258
259#endif /* MC68000 */
260
Paul Bakker5121ce52009-01-03 21:22:43 +0000261#if defined(__powerpc64__) || defined(__ppc64__)
262
263#if defined(__MACH__) && defined(__APPLE__)
264
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200265#define MULADDC_INIT \
266 asm( \
267 "ld r3, %3 \n\t" \
268 "ld r4, %4 \n\t" \
269 "ld r5, %5 \n\t" \
270 "ld r6, %6 \n\t" \
271 "addi r3, r3, -8 \n\t" \
272 "addi r4, r4, -8 \n\t" \
273 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000274
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200275#define MULADDC_CORE \
276 "ldu r7, 8(r3) \n\t" \
277 "mulld r8, r7, r6 \n\t" \
278 "mulhdu r9, r7, r6 \n\t" \
279 "adde r8, r8, r5 \n\t" \
280 "ld r7, 8(r4) \n\t" \
281 "addze r5, r9 \n\t" \
282 "addc r8, r8, r7 \n\t" \
283 "stdu r8, 8(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000284
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200285#define MULADDC_STOP \
286 "addze r5, r5 \n\t" \
287 "addi r4, r4, 8 \n\t" \
288 "addi r3, r3, 8 \n\t" \
289 "std r5, %0 \n\t" \
290 "std r4, %1 \n\t" \
291 "std r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100292 : "=m" (c), "=m" (d), "=m" (s) \
293 : "m" (s), "m" (d), "m" (c), "m" (b) \
294 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
295 );
296
Paul Bakker5121ce52009-01-03 21:22:43 +0000297
Paul Bakker9af723c2014-05-01 13:03:14 +0200298#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000299
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200300#define MULADDC_INIT \
301 asm( \
302 "ld %%r3, %3 \n\t" \
303 "ld %%r4, %4 \n\t" \
304 "ld %%r5, %5 \n\t" \
305 "ld %%r6, %6 \n\t" \
306 "addi %%r3, %%r3, -8 \n\t" \
307 "addi %%r4, %%r4, -8 \n\t" \
308 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000309
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200310#define MULADDC_CORE \
311 "ldu %%r7, 8(%%r3) \n\t" \
312 "mulld %%r8, %%r7, %%r6 \n\t" \
313 "mulhdu %%r9, %%r7, %%r6 \n\t" \
314 "adde %%r8, %%r8, %%r5 \n\t" \
315 "ld %%r7, 8(%%r4) \n\t" \
316 "addze %%r5, %%r9 \n\t" \
317 "addc %%r8, %%r8, %%r7 \n\t" \
318 "stdu %%r8, 8(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000319
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200320#define MULADDC_STOP \
321 "addze %%r5, %%r5 \n\t" \
322 "addi %%r4, %%r4, 8 \n\t" \
323 "addi %%r3, %%r3, 8 \n\t" \
324 "std %%r5, %0 \n\t" \
325 "std %%r4, %1 \n\t" \
326 "std %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100327 : "=m" (c), "=m" (d), "=m" (s) \
328 : "m" (s), "m" (d), "m" (c), "m" (b) \
329 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
330 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000331
Paul Bakker9af723c2014-05-01 13:03:14 +0200332#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000333
Barry K. Nathan35e7cb92014-05-05 23:26:13 -0700334#elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000335
336#if defined(__MACH__) && defined(__APPLE__)
337
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200338#define MULADDC_INIT \
339 asm( \
340 "lwz r3, %3 \n\t" \
341 "lwz r4, %4 \n\t" \
342 "lwz r5, %5 \n\t" \
343 "lwz r6, %6 \n\t" \
344 "addi r3, r3, -4 \n\t" \
345 "addi r4, r4, -4 \n\t" \
346 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000347
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200348#define MULADDC_CORE \
349 "lwzu r7, 4(r3) \n\t" \
350 "mullw r8, r7, r6 \n\t" \
351 "mulhwu r9, r7, r6 \n\t" \
352 "adde r8, r8, r5 \n\t" \
353 "lwz r7, 4(r4) \n\t" \
354 "addze r5, r9 \n\t" \
355 "addc r8, r8, r7 \n\t" \
356 "stwu r8, 4(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000357
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200358#define MULADDC_STOP \
359 "addze r5, r5 \n\t" \
360 "addi r4, r4, 4 \n\t" \
361 "addi r3, r3, 4 \n\t" \
362 "stw r5, %0 \n\t" \
363 "stw r4, %1 \n\t" \
364 "stw r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100365 : "=m" (c), "=m" (d), "=m" (s) \
366 : "m" (s), "m" (d), "m" (c), "m" (b) \
367 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
368 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000369
Paul Bakker9af723c2014-05-01 13:03:14 +0200370#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000371
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200372#define MULADDC_INIT \
373 asm( \
374 "lwz %%r3, %3 \n\t" \
375 "lwz %%r4, %4 \n\t" \
376 "lwz %%r5, %5 \n\t" \
377 "lwz %%r6, %6 \n\t" \
378 "addi %%r3, %%r3, -4 \n\t" \
379 "addi %%r4, %%r4, -4 \n\t" \
380 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000381
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200382#define MULADDC_CORE \
383 "lwzu %%r7, 4(%%r3) \n\t" \
384 "mullw %%r8, %%r7, %%r6 \n\t" \
385 "mulhwu %%r9, %%r7, %%r6 \n\t" \
386 "adde %%r8, %%r8, %%r5 \n\t" \
387 "lwz %%r7, 4(%%r4) \n\t" \
388 "addze %%r5, %%r9 \n\t" \
389 "addc %%r8, %%r8, %%r7 \n\t" \
390 "stwu %%r8, 4(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000391
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200392#define MULADDC_STOP \
393 "addze %%r5, %%r5 \n\t" \
394 "addi %%r4, %%r4, 4 \n\t" \
395 "addi %%r3, %%r3, 4 \n\t" \
396 "stw %%r5, %0 \n\t" \
397 "stw %%r4, %1 \n\t" \
398 "stw %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100399 : "=m" (c), "=m" (d), "=m" (s) \
400 : "m" (s), "m" (d), "m" (c), "m" (b) \
401 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
402 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000403
Paul Bakker9af723c2014-05-01 13:03:14 +0200404#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000405
406#endif /* PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000407
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200408/*
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200409 * The Sparc(64) assembly is reported to be broken.
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200410 * Disable it for now, until we're able to fix it.
411 */
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200412#if 0 && defined(__sparc__)
413#if defined(__sparc64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000414
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200415#define MULADDC_INIT \
416 asm( \
417 "ldx %3, %%o0 \n\t" \
418 "ldx %4, %%o1 \n\t" \
419 "ld %5, %%o2 \n\t" \
420 "ld %6, %%o3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000421
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200422#define MULADDC_CORE \
423 "ld [%%o0], %%o4 \n\t" \
424 "inc 4, %%o0 \n\t" \
425 "ld [%%o1], %%o5 \n\t" \
426 "umul %%o3, %%o4, %%o4 \n\t" \
427 "addcc %%o4, %%o2, %%o4 \n\t" \
428 "rd %%y, %%g1 \n\t" \
429 "addx %%g1, 0, %%g1 \n\t" \
430 "addcc %%o4, %%o5, %%o4 \n\t" \
431 "st %%o4, [%%o1] \n\t" \
432 "addx %%g1, 0, %%o2 \n\t" \
433 "inc 4, %%o1 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000434
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200435 #define MULADDC_STOP \
436 "st %%o2, %0 \n\t" \
437 "stx %%o1, %1 \n\t" \
438 "stx %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000439 : "=m" (c), "=m" (d), "=m" (s) \
440 : "m" (s), "m" (d), "m" (c), "m" (b) \
441 : "g1", "o0", "o1", "o2", "o3", "o4", \
442 "o5" \
443 );
Paul Bakker4f024b72012-10-30 07:29:57 +0000444
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200445#else /* __sparc64__ */
Paul Bakker4f024b72012-10-30 07:29:57 +0000446
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
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200477#endif /* __sparc64__ */
478#endif /* __sparc__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000479
480#if defined(__microblaze__) || defined(microblaze)
481
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200482#define MULADDC_INIT \
483 asm( \
484 "lwi r3, %3 \n\t" \
485 "lwi r4, %4 \n\t" \
486 "lwi r5, %5 \n\t" \
487 "lwi r6, %6 \n\t" \
488 "andi r7, r6, 0xffff \n\t" \
489 "bsrli r6, r6, 16 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000490
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200491#define MULADDC_CORE \
492 "lhui r8, r3, 0 \n\t" \
493 "addi r3, r3, 2 \n\t" \
494 "lhui r9, r3, 0 \n\t" \
495 "addi r3, r3, 2 \n\t" \
496 "mul r10, r9, r6 \n\t" \
497 "mul r11, r8, r7 \n\t" \
498 "mul r12, r9, r7 \n\t" \
499 "mul r13, r8, r6 \n\t" \
500 "bsrli r8, r10, 16 \n\t" \
501 "bsrli r9, r11, 16 \n\t" \
502 "add r13, r13, r8 \n\t" \
503 "add r13, r13, r9 \n\t" \
504 "bslli r10, r10, 16 \n\t" \
505 "bslli r11, r11, 16 \n\t" \
506 "add r12, r12, r10 \n\t" \
507 "addc r13, r13, r0 \n\t" \
508 "add r12, r12, r11 \n\t" \
509 "addc r13, r13, r0 \n\t" \
510 "lwi r10, r4, 0 \n\t" \
511 "add r12, r12, r10 \n\t" \
512 "addc r13, r13, r0 \n\t" \
513 "add r12, r12, r5 \n\t" \
514 "addc r5, r13, r0 \n\t" \
515 "swi r12, r4, 0 \n\t" \
516 "addi r4, r4, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000517
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200518#define MULADDC_STOP \
519 "swi r5, %0 \n\t" \
520 "swi r4, %1 \n\t" \
521 "swi r3, %2 \n\t" \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100522 : "=m" (c), "=m" (d), "=m" (s) \
523 : "m" (s), "m" (d), "m" (c), "m" (b) \
524 : "r3", "r4" "r5", "r6", "r7", "r8", \
525 "r9", "r10", "r11", "r12", "r13" \
526 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000527
528#endif /* MicroBlaze */
529
530#if defined(__tricore__)
531
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200532#define MULADDC_INIT \
533 asm( \
534 "ld.a %%a2, %3 \n\t" \
535 "ld.a %%a3, %4 \n\t" \
536 "ld.w %%d4, %5 \n\t" \
537 "ld.w %%d1, %6 \n\t" \
538 "xor %%d5, %%d5 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000539
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200540#define MULADDC_CORE \
541 "ld.w %%d0, [%%a2+] \n\t" \
542 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
543 "ld.w %%d0, [%%a3] \n\t" \
544 "addx %%d2, %%d2, %%d0 \n\t" \
545 "addc %%d3, %%d3, 0 \n\t" \
546 "mov %%d4, %%d3 \n\t" \
547 "st.w [%%a3+], %%d2 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000548
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200549#define MULADDC_STOP \
550 "st.w %0, %%d4 \n\t" \
551 "st.a %1, %%a3 \n\t" \
552 "st.a %2, %%a2 \n\t" \
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100553 : "=m" (c), "=m" (d), "=m" (s) \
554 : "m" (s), "m" (d), "m" (c), "m" (b) \
555 : "d0", "d1", "e2", "d4", "a2", "a3" \
556 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000557
558#endif /* TriCore */
559
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100560/*
561 * gcc -O0 by default uses r7 for the frame pointer, so it complains about our
562 * use of r7 below, unless -fomit-frame-pointer is passed. Unfortunately,
563 * passing that option is not easy when building with yotta.
564 *
565 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
566 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
567 * clang and armcc5 under the same conditions).
568 *
569 * So, only use the optimized assembly below for optimized build, which avoids
570 * the build error and is pretty reasonable anyway.
571 */
572#if defined(__GNUC__) && !defined(__OPTIMIZE__)
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100573#define MULADDC_CANNOT_USE_R7
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100574#endif
575
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100576#if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
Paul Bakker5121ce52009-01-03 21:22:43 +0000577
Paul Bakkerfc4f46f2013-06-24 19:23:56 +0200578#if defined(__thumb__) && !defined(__thumb2__)
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000579
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200580#define MULADDC_INIT \
581 asm( \
582 "ldr r0, %3 \n\t" \
583 "ldr r1, %4 \n\t" \
584 "ldr r2, %5 \n\t" \
585 "ldr r3, %6 \n\t" \
586 "lsr r7, r3, #16 \n\t" \
587 "mov r9, r7 \n\t" \
588 "lsl r7, r3, #16 \n\t" \
589 "lsr r7, r7, #16 \n\t" \
590 "mov r8, r7 \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000591
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200592#define MULADDC_CORE \
593 "ldmia r0!, {r6} \n\t" \
594 "lsr r7, r6, #16 \n\t" \
595 "lsl r6, r6, #16 \n\t" \
596 "lsr r6, r6, #16 \n\t" \
597 "mov r4, r8 \n\t" \
598 "mul r4, r6 \n\t" \
599 "mov r3, r9 \n\t" \
600 "mul r6, r3 \n\t" \
601 "mov r5, r9 \n\t" \
602 "mul r5, r7 \n\t" \
603 "mov r3, r8 \n\t" \
604 "mul r7, r3 \n\t" \
605 "lsr r3, r6, #16 \n\t" \
606 "add r5, r5, r3 \n\t" \
607 "lsr r3, r7, #16 \n\t" \
608 "add r5, r5, r3 \n\t" \
609 "add r4, r4, r2 \n\t" \
610 "mov r2, #0 \n\t" \
611 "adc r5, r2 \n\t" \
612 "lsl r3, r6, #16 \n\t" \
613 "add r4, r4, r3 \n\t" \
614 "adc r5, r2 \n\t" \
615 "lsl r3, r7, #16 \n\t" \
616 "add r4, r4, r3 \n\t" \
617 "adc r5, r2 \n\t" \
618 "ldr r3, [r1] \n\t" \
619 "add r4, r4, r3 \n\t" \
620 "adc r2, r5 \n\t" \
621 "stmia r1!, {r4} \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000622
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200623#define MULADDC_STOP \
624 "str r2, %0 \n\t" \
625 "str r1, %1 \n\t" \
626 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100627 : "=m" (c), "=m" (d), "=m" (s) \
628 : "m" (s), "m" (d), "m" (c), "m" (b) \
629 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200630 "r6", "r7", "r8", "r9", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100631 );
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000632
Aurelien Jarno16b1bd82018-05-21 22:01:21 +0200633#elif defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)
634
635#define MULADDC_INIT \
636 asm(
637
638#define MULADDC_CORE \
639 "ldr r0, [%0], #4 \n\t" \
640 "ldr r1, [%1] \n\t" \
641 "umaal r1, %2, %3, r0 \n\t" \
642 "str r1, [%1], #4 \n\t"
643
644#define MULADDC_STOP \
645 : "=r" (s), "=r" (d), "=r" (c) \
646 : "r" (b), "0" (s), "1" (d), "2" (c) \
647 : "r0", "r1", "memory" \
648 );
649
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000650#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000651
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200652#define MULADDC_INIT \
653 asm( \
654 "ldr r0, %3 \n\t" \
655 "ldr r1, %4 \n\t" \
656 "ldr r2, %5 \n\t" \
657 "ldr r3, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000658
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200659#define MULADDC_CORE \
660 "ldr r4, [r0], #4 \n\t" \
661 "mov r5, #0 \n\t" \
662 "ldr r6, [r1] \n\t" \
663 "umlal r2, r5, r3, r4 \n\t" \
664 "adds r7, r6, r2 \n\t" \
665 "adc r2, r5, #0 \n\t" \
666 "str r7, [r1], #4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000667
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200668#define MULADDC_STOP \
669 "str r2, %0 \n\t" \
670 "str r1, %1 \n\t" \
671 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100672 : "=m" (c), "=m" (d), "=m" (s) \
673 : "m" (s), "m" (d), "m" (c), "m" (b) \
674 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200675 "r6", "r7", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100676 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000677
Paul Bakkera2713a32011-11-18 12:47:23 +0000678#endif /* Thumb */
679
Paul Bakker5121ce52009-01-03 21:22:43 +0000680#endif /* ARMv3 */
681
682#if defined(__alpha__)
683
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200684#define MULADDC_INIT \
685 asm( \
686 "ldq $1, %3 \n\t" \
687 "ldq $2, %4 \n\t" \
688 "ldq $3, %5 \n\t" \
689 "ldq $4, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000690
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200691#define MULADDC_CORE \
692 "ldq $6, 0($1) \n\t" \
693 "addq $1, 8, $1 \n\t" \
694 "mulq $6, $4, $7 \n\t" \
695 "umulh $6, $4, $6 \n\t" \
696 "addq $7, $3, $7 \n\t" \
697 "cmpult $7, $3, $3 \n\t" \
698 "ldq $5, 0($2) \n\t" \
699 "addq $7, $5, $7 \n\t" \
700 "cmpult $7, $5, $5 \n\t" \
701 "stq $7, 0($2) \n\t" \
702 "addq $2, 8, $2 \n\t" \
703 "addq $6, $3, $3 \n\t" \
704 "addq $5, $3, $3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000705
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200706#define MULADDC_STOP \
707 "stq $3, %0 \n\t" \
708 "stq $2, %1 \n\t" \
709 "stq $1, %2 \n\t" \
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100710 : "=m" (c), "=m" (d), "=m" (s) \
711 : "m" (s), "m" (d), "m" (c), "m" (b) \
712 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
713 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000714#endif /* Alpha */
715
James Cowgilld1e7e8b2014-12-16 15:24:06 +0000716#if defined(__mips__) && !defined(__mips64)
Paul Bakker5121ce52009-01-03 21:22:43 +0000717
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200718#define MULADDC_INIT \
719 asm( \
720 "lw $10, %3 \n\t" \
721 "lw $11, %4 \n\t" \
722 "lw $12, %5 \n\t" \
723 "lw $13, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000724
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200725#define MULADDC_CORE \
726 "lw $14, 0($10) \n\t" \
727 "multu $13, $14 \n\t" \
728 "addi $10, $10, 4 \n\t" \
729 "mflo $14 \n\t" \
730 "mfhi $9 \n\t" \
731 "addu $14, $12, $14 \n\t" \
732 "lw $15, 0($11) \n\t" \
733 "sltu $12, $14, $12 \n\t" \
734 "addu $15, $14, $15 \n\t" \
735 "sltu $14, $15, $14 \n\t" \
736 "addu $12, $12, $9 \n\t" \
737 "sw $15, 0($11) \n\t" \
738 "addu $12, $12, $14 \n\t" \
739 "addi $11, $11, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000740
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200741#define MULADDC_STOP \
742 "sw $12, %0 \n\t" \
743 "sw $11, %1 \n\t" \
744 "sw $10, %2 \n\t" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100745 : "=m" (c), "=m" (d), "=m" (s) \
746 : "m" (s), "m" (d), "m" (c), "m" (b) \
747 : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
748 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000749
750#endif /* MIPS */
751#endif /* GNUC */
752
753#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
754
755#define MULADDC_INIT \
756 __asm mov esi, s \
757 __asm mov edi, d \
758 __asm mov ecx, c \
759 __asm mov ebx, b
760
761#define MULADDC_CORE \
762 __asm lodsd \
763 __asm mul ebx \
764 __asm add eax, ecx \
765 __asm adc edx, 0 \
766 __asm add eax, [edi] \
767 __asm adc edx, 0 \
768 __asm mov ecx, edx \
769 __asm stosd
770
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200771#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000772
773#define EMIT __asm _emit
774
775#define MULADDC_HUIT \
776 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
777 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
778 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
779 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
780 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
781 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
782 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
783 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
784 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
785 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
786 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
787 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
788 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
789 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
790 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
791 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
792 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
793 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
794 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
795 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
796 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
797 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
798 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
799 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
800 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
801 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
802 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
803 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
804 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
805 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
806 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
807 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
808 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
809 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
810 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
811 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
812 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
813 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
814 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
815 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
816 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
817 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
818 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
819 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
820 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
821 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
822 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
823 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
824 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
825 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
826 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
827 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
828 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
829 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
830 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
831 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
832 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
833 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
834 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
835 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
836 EMIT 0x0F EMIT 0x7E EMIT 0xC9
837
838#define MULADDC_STOP \
839 EMIT 0x0F EMIT 0x77 \
840 __asm mov c, ecx \
841 __asm mov d, edi \
842 __asm mov s, esi \
843
844#else
845
846#define MULADDC_STOP \
847 __asm mov c, ecx \
848 __asm mov d, edi \
849 __asm mov s, esi \
850
851#endif /* SSE2 */
852#endif /* MSVC */
853
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200854#endif /* MBEDTLS_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000855
856#if !defined(MULADDC_CORE)
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200857#if defined(MBEDTLS_HAVE_UDBL)
Paul Bakker5121ce52009-01-03 21:22:43 +0000858
859#define MULADDC_INIT \
860{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200861 mbedtls_t_udbl r; \
862 mbedtls_mpi_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000863
864#define MULADDC_CORE \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200865 r = *(s++) * (mbedtls_t_udbl) b; \
866 r0 = (mbedtls_mpi_uint) r; \
867 r1 = (mbedtls_mpi_uint)( r >> biL ); \
Paul Bakker5121ce52009-01-03 21:22:43 +0000868 r0 += c; r1 += (r0 < c); \
869 r0 += *d; r1 += (r0 < *d); \
870 c = r1; *(d++) = r0;
871
872#define MULADDC_STOP \
873}
874
875#else
876#define MULADDC_INIT \
877{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200878 mbedtls_mpi_uint s0, s1, b0, b1; \
879 mbedtls_mpi_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000880 b0 = ( b << biH ) >> biH; \
881 b1 = ( b >> biH );
882
883#define MULADDC_CORE \
884 s0 = ( *s << biH ) >> biH; \
885 s1 = ( *s >> biH ); s++; \
886 rx = s0 * b1; r0 = s0 * b0; \
887 ry = s1 * b0; r1 = s1 * b1; \
888 r1 += ( rx >> biH ); \
889 r1 += ( ry >> biH ); \
890 rx <<= biH; ry <<= biH; \
891 r0 += rx; r1 += (r0 < rx); \
892 r0 += ry; r1 += (r0 < ry); \
893 r0 += c; r1 += (r0 < c); \
894 r0 += *d; r1 += (r0 < *d); \
895 c = r1; *(d++) = r0;
896
897#define MULADDC_STOP \
898}
899
900#endif /* C (generic) */
901#endif /* C (longlong) */
902
903#endif /* bn_mul.h */