blob: 438aa8cea192576e8f5971be2290e90603ece09f [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 )
Simon Butcherd064b5c2018-07-10 20:18:29 +010052
53/*
54 * Disable use of the i386 assembly code below if option -O0, to disable all
55 * compiler optimisations, is passed, detected with __OPTIMIZE__
56 * This is done as the number of registers used in the assembly code doesn't
57 * work with the -O0 option.
58 */
59#if defined(__i386__) && !defined(__OPTIMIZE__)
Paul Bakker5121ce52009-01-03 21:22:43 +000060
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +020061#define MULADDC_INIT \
62 asm( \
63 "movl %%ebx, %0 \n\t" \
64 "movl %5, %%esi \n\t" \
65 "movl %6, %%edi \n\t" \
66 "movl %7, %%ecx \n\t" \
67 "movl %8, %%ebx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +000068
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +020069#define MULADDC_CORE \
70 "lodsl \n\t" \
71 "mull %%ebx \n\t" \
72 "addl %%ecx, %%eax \n\t" \
73 "adcl $0, %%edx \n\t" \
74 "addl (%%edi), %%eax \n\t" \
75 "adcl $0, %%edx \n\t" \
76 "movl %%edx, %%ecx \n\t" \
77 "stosl \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +000078
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +020079#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +000080
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +020081#define MULADDC_HUIT \
82 "movd %%ecx, %%mm1 \n\t" \
83 "movd %%ebx, %%mm0 \n\t" \
84 "movd (%%edi), %%mm3 \n\t" \
85 "paddq %%mm3, %%mm1 \n\t" \
86 "movd (%%esi), %%mm2 \n\t" \
87 "pmuludq %%mm0, %%mm2 \n\t" \
88 "movd 4(%%esi), %%mm4 \n\t" \
89 "pmuludq %%mm0, %%mm4 \n\t" \
90 "movd 8(%%esi), %%mm6 \n\t" \
91 "pmuludq %%mm0, %%mm6 \n\t" \
92 "movd 12(%%esi), %%mm7 \n\t" \
93 "pmuludq %%mm0, %%mm7 \n\t" \
94 "paddq %%mm2, %%mm1 \n\t" \
95 "movd 4(%%edi), %%mm3 \n\t" \
96 "paddq %%mm4, %%mm3 \n\t" \
97 "movd 8(%%edi), %%mm5 \n\t" \
98 "paddq %%mm6, %%mm5 \n\t" \
99 "movd 12(%%edi), %%mm4 \n\t" \
100 "paddq %%mm4, %%mm7 \n\t" \
101 "movd %%mm1, (%%edi) \n\t" \
102 "movd 16(%%esi), %%mm2 \n\t" \
103 "pmuludq %%mm0, %%mm2 \n\t" \
104 "psrlq $32, %%mm1 \n\t" \
105 "movd 20(%%esi), %%mm4 \n\t" \
106 "pmuludq %%mm0, %%mm4 \n\t" \
107 "paddq %%mm3, %%mm1 \n\t" \
108 "movd 24(%%esi), %%mm6 \n\t" \
109 "pmuludq %%mm0, %%mm6 \n\t" \
110 "movd %%mm1, 4(%%edi) \n\t" \
111 "psrlq $32, %%mm1 \n\t" \
112 "movd 28(%%esi), %%mm3 \n\t" \
113 "pmuludq %%mm0, %%mm3 \n\t" \
114 "paddq %%mm5, %%mm1 \n\t" \
115 "movd 16(%%edi), %%mm5 \n\t" \
116 "paddq %%mm5, %%mm2 \n\t" \
117 "movd %%mm1, 8(%%edi) \n\t" \
118 "psrlq $32, %%mm1 \n\t" \
119 "paddq %%mm7, %%mm1 \n\t" \
120 "movd 20(%%edi), %%mm5 \n\t" \
121 "paddq %%mm5, %%mm4 \n\t" \
122 "movd %%mm1, 12(%%edi) \n\t" \
123 "psrlq $32, %%mm1 \n\t" \
124 "paddq %%mm2, %%mm1 \n\t" \
125 "movd 24(%%edi), %%mm5 \n\t" \
126 "paddq %%mm5, %%mm6 \n\t" \
127 "movd %%mm1, 16(%%edi) \n\t" \
128 "psrlq $32, %%mm1 \n\t" \
129 "paddq %%mm4, %%mm1 \n\t" \
130 "movd 28(%%edi), %%mm5 \n\t" \
131 "paddq %%mm5, %%mm3 \n\t" \
132 "movd %%mm1, 20(%%edi) \n\t" \
133 "psrlq $32, %%mm1 \n\t" \
134 "paddq %%mm6, %%mm1 \n\t" \
135 "movd %%mm1, 24(%%edi) \n\t" \
136 "psrlq $32, %%mm1 \n\t" \
137 "paddq %%mm3, %%mm1 \n\t" \
138 "movd %%mm1, 28(%%edi) \n\t" \
139 "addl $32, %%edi \n\t" \
140 "addl $32, %%esi \n\t" \
141 "psrlq $32, %%mm1 \n\t" \
142 "movd %%mm1, %%ecx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000143
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200144#define MULADDC_STOP \
145 "emms \n\t" \
146 "movl %4, %%ebx \n\t" \
147 "movl %%ecx, %1 \n\t" \
148 "movl %%edi, %2 \n\t" \
149 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000150 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
151 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
Simon Butcher83868a02018-06-24 12:58:31 +0100152 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000153 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000154
155#else
156
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200157#define MULADDC_STOP \
158 "movl %4, %%ebx \n\t" \
159 "movl %%ecx, %1 \n\t" \
160 "movl %%edi, %2 \n\t" \
161 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000162 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
163 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
Simon Butcher83868a02018-06-24 12:58:31 +0100164 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000165 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000166#endif /* SSE2 */
167#endif /* i386 */
168
169#if defined(__amd64__) || defined (__x86_64__)
170
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200171#define MULADDC_INIT \
172 asm( \
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200173 "xorq %%r8, %%r8 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000174
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200175#define MULADDC_CORE \
176 "movq (%%rsi), %%rax \n\t" \
177 "mulq %%rbx \n\t" \
178 "addq $8, %%rsi \n\t" \
179 "addq %%rcx, %%rax \n\t" \
180 "movq %%r8, %%rcx \n\t" \
181 "adcq $0, %%rdx \n\t" \
182 "nop \n\t" \
183 "addq %%rax, (%%rdi) \n\t" \
184 "adcq %%rdx, %%rcx \n\t" \
185 "addq $8, %%rdi \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000186
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200187#define MULADDC_STOP \
James Cowgill21e402a2015-12-17 01:51:09 +0000188 : "+c" (c), "+D" (d), "+S" (s) \
189 : "b" (b) \
190 : "rax", "rdx", "r8" \
Manuel Pégourié-Gonnarddef018d2014-01-07 17:50:46 +0100191 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000192
193#endif /* AMD64 */
194
195#if defined(__mc68020__) || defined(__mcpu32__)
196
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200197#define MULADDC_INIT \
198 asm( \
199 "movl %3, %%a2 \n\t" \
200 "movl %4, %%a3 \n\t" \
201 "movl %5, %%d3 \n\t" \
202 "movl %6, %%d2 \n\t" \
203 "moveq #0, %%d0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000204
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200205#define MULADDC_CORE \
206 "movel %%a2@+, %%d1 \n\t" \
207 "mulul %%d2, %%d4:%%d1 \n\t" \
208 "addl %%d3, %%d1 \n\t" \
209 "addxl %%d0, %%d4 \n\t" \
210 "moveq #0, %%d3 \n\t" \
211 "addl %%d1, %%a3@+ \n\t" \
212 "addxl %%d4, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000213
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200214#define MULADDC_STOP \
215 "movl %%d3, %0 \n\t" \
216 "movl %%a3, %1 \n\t" \
217 "movl %%a2, %2 \n\t" \
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100218 : "=m" (c), "=m" (d), "=m" (s) \
219 : "m" (s), "m" (d), "m" (c), "m" (b) \
220 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
221 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000222
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200223#define MULADDC_HUIT \
224 "movel %%a2@+, %%d1 \n\t" \
225 "mulul %%d2, %%d4:%%d1 \n\t" \
226 "addxl %%d3, %%d1 \n\t" \
227 "addxl %%d0, %%d4 \n\t" \
228 "addl %%d1, %%a3@+ \n\t" \
229 "movel %%a2@+, %%d1 \n\t" \
230 "mulul %%d2, %%d3:%%d1 \n\t" \
231 "addxl %%d4, %%d1 \n\t" \
232 "addxl %%d0, %%d3 \n\t" \
233 "addl %%d1, %%a3@+ \n\t" \
234 "movel %%a2@+, %%d1 \n\t" \
235 "mulul %%d2, %%d4:%%d1 \n\t" \
236 "addxl %%d3, %%d1 \n\t" \
237 "addxl %%d0, %%d4 \n\t" \
238 "addl %%d1, %%a3@+ \n\t" \
239 "movel %%a2@+, %%d1 \n\t" \
240 "mulul %%d2, %%d3:%%d1 \n\t" \
241 "addxl %%d4, %%d1 \n\t" \
242 "addxl %%d0, %%d3 \n\t" \
243 "addl %%d1, %%a3@+ \n\t" \
244 "movel %%a2@+, %%d1 \n\t" \
245 "mulul %%d2, %%d4:%%d1 \n\t" \
246 "addxl %%d3, %%d1 \n\t" \
247 "addxl %%d0, %%d4 \n\t" \
248 "addl %%d1, %%a3@+ \n\t" \
249 "movel %%a2@+, %%d1 \n\t" \
250 "mulul %%d2, %%d3:%%d1 \n\t" \
251 "addxl %%d4, %%d1 \n\t" \
252 "addxl %%d0, %%d3 \n\t" \
253 "addl %%d1, %%a3@+ \n\t" \
254 "movel %%a2@+, %%d1 \n\t" \
255 "mulul %%d2, %%d4:%%d1 \n\t" \
256 "addxl %%d3, %%d1 \n\t" \
257 "addxl %%d0, %%d4 \n\t" \
258 "addl %%d1, %%a3@+ \n\t" \
259 "movel %%a2@+, %%d1 \n\t" \
260 "mulul %%d2, %%d3:%%d1 \n\t" \
261 "addxl %%d4, %%d1 \n\t" \
262 "addxl %%d0, %%d3 \n\t" \
263 "addl %%d1, %%a3@+ \n\t" \
264 "addxl %%d0, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000265
266#endif /* MC68000 */
267
Paul Bakker5121ce52009-01-03 21:22:43 +0000268#if defined(__powerpc64__) || defined(__ppc64__)
269
270#if defined(__MACH__) && defined(__APPLE__)
271
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200272#define MULADDC_INIT \
273 asm( \
274 "ld r3, %3 \n\t" \
275 "ld r4, %4 \n\t" \
276 "ld r5, %5 \n\t" \
277 "ld r6, %6 \n\t" \
278 "addi r3, r3, -8 \n\t" \
279 "addi r4, r4, -8 \n\t" \
280 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000281
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200282#define MULADDC_CORE \
283 "ldu r7, 8(r3) \n\t" \
284 "mulld r8, r7, r6 \n\t" \
285 "mulhdu r9, r7, r6 \n\t" \
286 "adde r8, r8, r5 \n\t" \
287 "ld r7, 8(r4) \n\t" \
288 "addze r5, r9 \n\t" \
289 "addc r8, r8, r7 \n\t" \
290 "stdu r8, 8(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000291
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200292#define MULADDC_STOP \
293 "addze r5, r5 \n\t" \
294 "addi r4, r4, 8 \n\t" \
295 "addi r3, r3, 8 \n\t" \
296 "std r5, %0 \n\t" \
297 "std r4, %1 \n\t" \
298 "std r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100299 : "=m" (c), "=m" (d), "=m" (s) \
300 : "m" (s), "m" (d), "m" (c), "m" (b) \
301 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
302 );
303
Paul Bakker5121ce52009-01-03 21:22:43 +0000304
Paul Bakker9af723c2014-05-01 13:03:14 +0200305#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000306
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200307#define MULADDC_INIT \
308 asm( \
309 "ld %%r3, %3 \n\t" \
310 "ld %%r4, %4 \n\t" \
311 "ld %%r5, %5 \n\t" \
312 "ld %%r6, %6 \n\t" \
313 "addi %%r3, %%r3, -8 \n\t" \
314 "addi %%r4, %%r4, -8 \n\t" \
315 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000316
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200317#define MULADDC_CORE \
318 "ldu %%r7, 8(%%r3) \n\t" \
319 "mulld %%r8, %%r7, %%r6 \n\t" \
320 "mulhdu %%r9, %%r7, %%r6 \n\t" \
321 "adde %%r8, %%r8, %%r5 \n\t" \
322 "ld %%r7, 8(%%r4) \n\t" \
323 "addze %%r5, %%r9 \n\t" \
324 "addc %%r8, %%r8, %%r7 \n\t" \
325 "stdu %%r8, 8(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000326
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200327#define MULADDC_STOP \
328 "addze %%r5, %%r5 \n\t" \
329 "addi %%r4, %%r4, 8 \n\t" \
330 "addi %%r3, %%r3, 8 \n\t" \
331 "std %%r5, %0 \n\t" \
332 "std %%r4, %1 \n\t" \
333 "std %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100334 : "=m" (c), "=m" (d), "=m" (s) \
335 : "m" (s), "m" (d), "m" (c), "m" (b) \
336 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
337 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000338
Paul Bakker9af723c2014-05-01 13:03:14 +0200339#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000340
Barry K. Nathan35e7cb92014-05-05 23:26:13 -0700341#elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000342
343#if defined(__MACH__) && defined(__APPLE__)
344
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200345#define MULADDC_INIT \
346 asm( \
347 "lwz r3, %3 \n\t" \
348 "lwz r4, %4 \n\t" \
349 "lwz r5, %5 \n\t" \
350 "lwz r6, %6 \n\t" \
351 "addi r3, r3, -4 \n\t" \
352 "addi r4, r4, -4 \n\t" \
353 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000354
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200355#define MULADDC_CORE \
356 "lwzu r7, 4(r3) \n\t" \
357 "mullw r8, r7, r6 \n\t" \
358 "mulhwu r9, r7, r6 \n\t" \
359 "adde r8, r8, r5 \n\t" \
360 "lwz r7, 4(r4) \n\t" \
361 "addze r5, r9 \n\t" \
362 "addc r8, r8, r7 \n\t" \
363 "stwu r8, 4(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000364
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200365#define MULADDC_STOP \
366 "addze r5, r5 \n\t" \
367 "addi r4, r4, 4 \n\t" \
368 "addi r3, r3, 4 \n\t" \
369 "stw r5, %0 \n\t" \
370 "stw r4, %1 \n\t" \
371 "stw r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100372 : "=m" (c), "=m" (d), "=m" (s) \
373 : "m" (s), "m" (d), "m" (c), "m" (b) \
374 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
375 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000376
Paul Bakker9af723c2014-05-01 13:03:14 +0200377#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000378
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200379#define MULADDC_INIT \
380 asm( \
381 "lwz %%r3, %3 \n\t" \
382 "lwz %%r4, %4 \n\t" \
383 "lwz %%r5, %5 \n\t" \
384 "lwz %%r6, %6 \n\t" \
385 "addi %%r3, %%r3, -4 \n\t" \
386 "addi %%r4, %%r4, -4 \n\t" \
387 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000388
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200389#define MULADDC_CORE \
390 "lwzu %%r7, 4(%%r3) \n\t" \
391 "mullw %%r8, %%r7, %%r6 \n\t" \
392 "mulhwu %%r9, %%r7, %%r6 \n\t" \
393 "adde %%r8, %%r8, %%r5 \n\t" \
394 "lwz %%r7, 4(%%r4) \n\t" \
395 "addze %%r5, %%r9 \n\t" \
396 "addc %%r8, %%r8, %%r7 \n\t" \
397 "stwu %%r8, 4(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000398
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200399#define MULADDC_STOP \
400 "addze %%r5, %%r5 \n\t" \
401 "addi %%r4, %%r4, 4 \n\t" \
402 "addi %%r3, %%r3, 4 \n\t" \
403 "stw %%r5, %0 \n\t" \
404 "stw %%r4, %1 \n\t" \
405 "stw %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100406 : "=m" (c), "=m" (d), "=m" (s) \
407 : "m" (s), "m" (d), "m" (c), "m" (b) \
408 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
409 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000410
Paul Bakker9af723c2014-05-01 13:03:14 +0200411#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000412
413#endif /* PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000414
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200415/*
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200416 * The Sparc(64) assembly is reported to be broken.
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200417 * Disable it for now, until we're able to fix it.
418 */
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200419#if 0 && defined(__sparc__)
420#if defined(__sparc64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000421
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200422#define MULADDC_INIT \
423 asm( \
424 "ldx %3, %%o0 \n\t" \
425 "ldx %4, %%o1 \n\t" \
426 "ld %5, %%o2 \n\t" \
427 "ld %6, %%o3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000428
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200429#define MULADDC_CORE \
430 "ld [%%o0], %%o4 \n\t" \
431 "inc 4, %%o0 \n\t" \
432 "ld [%%o1], %%o5 \n\t" \
433 "umul %%o3, %%o4, %%o4 \n\t" \
434 "addcc %%o4, %%o2, %%o4 \n\t" \
435 "rd %%y, %%g1 \n\t" \
436 "addx %%g1, 0, %%g1 \n\t" \
437 "addcc %%o4, %%o5, %%o4 \n\t" \
438 "st %%o4, [%%o1] \n\t" \
439 "addx %%g1, 0, %%o2 \n\t" \
440 "inc 4, %%o1 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000441
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200442 #define MULADDC_STOP \
443 "st %%o2, %0 \n\t" \
444 "stx %%o1, %1 \n\t" \
445 "stx %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000446 : "=m" (c), "=m" (d), "=m" (s) \
447 : "m" (s), "m" (d), "m" (c), "m" (b) \
448 : "g1", "o0", "o1", "o2", "o3", "o4", \
449 "o5" \
450 );
Paul Bakker4f024b72012-10-30 07:29:57 +0000451
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200452#else /* __sparc64__ */
Paul Bakker4f024b72012-10-30 07:29:57 +0000453
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200454#define MULADDC_INIT \
455 asm( \
456 "ld %3, %%o0 \n\t" \
457 "ld %4, %%o1 \n\t" \
458 "ld %5, %%o2 \n\t" \
459 "ld %6, %%o3 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000460
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200461#define MULADDC_CORE \
462 "ld [%%o0], %%o4 \n\t" \
463 "inc 4, %%o0 \n\t" \
464 "ld [%%o1], %%o5 \n\t" \
465 "umul %%o3, %%o4, %%o4 \n\t" \
466 "addcc %%o4, %%o2, %%o4 \n\t" \
467 "rd %%y, %%g1 \n\t" \
468 "addx %%g1, 0, %%g1 \n\t" \
469 "addcc %%o4, %%o5, %%o4 \n\t" \
470 "st %%o4, [%%o1] \n\t" \
471 "addx %%g1, 0, %%o2 \n\t" \
472 "inc 4, %%o1 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000473
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200474#define MULADDC_STOP \
475 "st %%o2, %0 \n\t" \
476 "st %%o1, %1 \n\t" \
477 "st %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000478 : "=m" (c), "=m" (d), "=m" (s) \
479 : "m" (s), "m" (d), "m" (c), "m" (b) \
480 : "g1", "o0", "o1", "o2", "o3", "o4", \
481 "o5" \
482 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000483
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200484#endif /* __sparc64__ */
485#endif /* __sparc__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000486
487#if defined(__microblaze__) || defined(microblaze)
488
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200489#define MULADDC_INIT \
490 asm( \
491 "lwi r3, %3 \n\t" \
492 "lwi r4, %4 \n\t" \
493 "lwi r5, %5 \n\t" \
494 "lwi r6, %6 \n\t" \
495 "andi r7, r6, 0xffff \n\t" \
496 "bsrli r6, r6, 16 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000497
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200498#define MULADDC_CORE \
499 "lhui r8, r3, 0 \n\t" \
500 "addi r3, r3, 2 \n\t" \
501 "lhui r9, r3, 0 \n\t" \
502 "addi r3, r3, 2 \n\t" \
503 "mul r10, r9, r6 \n\t" \
504 "mul r11, r8, r7 \n\t" \
505 "mul r12, r9, r7 \n\t" \
506 "mul r13, r8, r6 \n\t" \
507 "bsrli r8, r10, 16 \n\t" \
508 "bsrli r9, r11, 16 \n\t" \
509 "add r13, r13, r8 \n\t" \
510 "add r13, r13, r9 \n\t" \
511 "bslli r10, r10, 16 \n\t" \
512 "bslli r11, r11, 16 \n\t" \
513 "add r12, r12, r10 \n\t" \
514 "addc r13, r13, r0 \n\t" \
515 "add r12, r12, r11 \n\t" \
516 "addc r13, r13, r0 \n\t" \
517 "lwi r10, r4, 0 \n\t" \
518 "add r12, r12, r10 \n\t" \
519 "addc r13, r13, r0 \n\t" \
520 "add r12, r12, r5 \n\t" \
521 "addc r5, r13, r0 \n\t" \
522 "swi r12, r4, 0 \n\t" \
523 "addi r4, r4, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000524
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200525#define MULADDC_STOP \
526 "swi r5, %0 \n\t" \
527 "swi r4, %1 \n\t" \
528 "swi r3, %2 \n\t" \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100529 : "=m" (c), "=m" (d), "=m" (s) \
530 : "m" (s), "m" (d), "m" (c), "m" (b) \
Zach van Rijndf484d62018-05-21 10:52:34 -0400531 : "r3", "r4", "r5", "r6", "r7", "r8", \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100532 "r9", "r10", "r11", "r12", "r13" \
533 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000534
535#endif /* MicroBlaze */
536
537#if defined(__tricore__)
538
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200539#define MULADDC_INIT \
540 asm( \
541 "ld.a %%a2, %3 \n\t" \
542 "ld.a %%a3, %4 \n\t" \
543 "ld.w %%d4, %5 \n\t" \
544 "ld.w %%d1, %6 \n\t" \
545 "xor %%d5, %%d5 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000546
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200547#define MULADDC_CORE \
548 "ld.w %%d0, [%%a2+] \n\t" \
549 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
550 "ld.w %%d0, [%%a3] \n\t" \
551 "addx %%d2, %%d2, %%d0 \n\t" \
552 "addc %%d3, %%d3, 0 \n\t" \
553 "mov %%d4, %%d3 \n\t" \
554 "st.w [%%a3+], %%d2 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000555
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200556#define MULADDC_STOP \
557 "st.w %0, %%d4 \n\t" \
558 "st.a %1, %%a3 \n\t" \
559 "st.a %2, %%a2 \n\t" \
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100560 : "=m" (c), "=m" (d), "=m" (s) \
561 : "m" (s), "m" (d), "m" (c), "m" (b) \
562 : "d0", "d1", "e2", "d4", "a2", "a3" \
563 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000564
565#endif /* TriCore */
566
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100567/*
568 * gcc -O0 by default uses r7 for the frame pointer, so it complains about our
569 * use of r7 below, unless -fomit-frame-pointer is passed. Unfortunately,
570 * passing that option is not easy when building with yotta.
571 *
572 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
573 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
574 * clang and armcc5 under the same conditions).
575 *
576 * So, only use the optimized assembly below for optimized build, which avoids
577 * the build error and is pretty reasonable anyway.
578 */
579#if defined(__GNUC__) && !defined(__OPTIMIZE__)
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100580#define MULADDC_CANNOT_USE_R7
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100581#endif
582
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100583#if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
Paul Bakker5121ce52009-01-03 21:22:43 +0000584
Paul Bakkerfc4f46f2013-06-24 19:23:56 +0200585#if defined(__thumb__) && !defined(__thumb2__)
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000586
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200587#define MULADDC_INIT \
588 asm( \
589 "ldr r0, %3 \n\t" \
590 "ldr r1, %4 \n\t" \
591 "ldr r2, %5 \n\t" \
592 "ldr r3, %6 \n\t" \
593 "lsr r7, r3, #16 \n\t" \
594 "mov r9, r7 \n\t" \
595 "lsl r7, r3, #16 \n\t" \
596 "lsr r7, r7, #16 \n\t" \
597 "mov r8, r7 \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000598
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200599#define MULADDC_CORE \
600 "ldmia r0!, {r6} \n\t" \
601 "lsr r7, r6, #16 \n\t" \
602 "lsl r6, r6, #16 \n\t" \
603 "lsr r6, r6, #16 \n\t" \
604 "mov r4, r8 \n\t" \
605 "mul r4, r6 \n\t" \
606 "mov r3, r9 \n\t" \
607 "mul r6, r3 \n\t" \
608 "mov r5, r9 \n\t" \
609 "mul r5, r7 \n\t" \
610 "mov r3, r8 \n\t" \
611 "mul r7, r3 \n\t" \
612 "lsr r3, r6, #16 \n\t" \
613 "add r5, r5, r3 \n\t" \
614 "lsr r3, r7, #16 \n\t" \
615 "add r5, r5, r3 \n\t" \
616 "add r4, r4, r2 \n\t" \
617 "mov r2, #0 \n\t" \
618 "adc r5, r2 \n\t" \
619 "lsl r3, r6, #16 \n\t" \
620 "add r4, r4, r3 \n\t" \
621 "adc r5, r2 \n\t" \
622 "lsl r3, r7, #16 \n\t" \
623 "add r4, r4, r3 \n\t" \
624 "adc r5, r2 \n\t" \
625 "ldr r3, [r1] \n\t" \
626 "add r4, r4, r3 \n\t" \
627 "adc r2, r5 \n\t" \
628 "stmia r1!, {r4} \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000629
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200630#define MULADDC_STOP \
631 "str r2, %0 \n\t" \
632 "str r1, %1 \n\t" \
633 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100634 : "=m" (c), "=m" (d), "=m" (s) \
635 : "m" (s), "m" (d), "m" (c), "m" (b) \
636 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200637 "r6", "r7", "r8", "r9", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100638 );
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000639
640#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000641
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200642#define MULADDC_INIT \
643 asm( \
644 "ldr r0, %3 \n\t" \
645 "ldr r1, %4 \n\t" \
646 "ldr r2, %5 \n\t" \
647 "ldr r3, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000648
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200649#define MULADDC_CORE \
650 "ldr r4, [r0], #4 \n\t" \
651 "mov r5, #0 \n\t" \
652 "ldr r6, [r1] \n\t" \
653 "umlal r2, r5, r3, r4 \n\t" \
654 "adds r7, r6, r2 \n\t" \
655 "adc r2, r5, #0 \n\t" \
656 "str r7, [r1], #4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000657
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200658#define MULADDC_STOP \
659 "str r2, %0 \n\t" \
660 "str r1, %1 \n\t" \
661 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100662 : "=m" (c), "=m" (d), "=m" (s) \
663 : "m" (s), "m" (d), "m" (c), "m" (b) \
664 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200665 "r6", "r7", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100666 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000667
Paul Bakkera2713a32011-11-18 12:47:23 +0000668#endif /* Thumb */
669
Paul Bakker5121ce52009-01-03 21:22:43 +0000670#endif /* ARMv3 */
671
672#if defined(__alpha__)
673
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200674#define MULADDC_INIT \
675 asm( \
676 "ldq $1, %3 \n\t" \
677 "ldq $2, %4 \n\t" \
678 "ldq $3, %5 \n\t" \
679 "ldq $4, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000680
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200681#define MULADDC_CORE \
682 "ldq $6, 0($1) \n\t" \
683 "addq $1, 8, $1 \n\t" \
684 "mulq $6, $4, $7 \n\t" \
685 "umulh $6, $4, $6 \n\t" \
686 "addq $7, $3, $7 \n\t" \
687 "cmpult $7, $3, $3 \n\t" \
688 "ldq $5, 0($2) \n\t" \
689 "addq $7, $5, $7 \n\t" \
690 "cmpult $7, $5, $5 \n\t" \
691 "stq $7, 0($2) \n\t" \
692 "addq $2, 8, $2 \n\t" \
693 "addq $6, $3, $3 \n\t" \
694 "addq $5, $3, $3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000695
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200696#define MULADDC_STOP \
697 "stq $3, %0 \n\t" \
698 "stq $2, %1 \n\t" \
699 "stq $1, %2 \n\t" \
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100700 : "=m" (c), "=m" (d), "=m" (s) \
701 : "m" (s), "m" (d), "m" (c), "m" (b) \
702 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
703 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000704#endif /* Alpha */
705
James Cowgilld1e7e8b2014-12-16 15:24:06 +0000706#if defined(__mips__) && !defined(__mips64)
Paul Bakker5121ce52009-01-03 21:22:43 +0000707
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200708#define MULADDC_INIT \
709 asm( \
710 "lw $10, %3 \n\t" \
711 "lw $11, %4 \n\t" \
712 "lw $12, %5 \n\t" \
713 "lw $13, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000714
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200715#define MULADDC_CORE \
716 "lw $14, 0($10) \n\t" \
717 "multu $13, $14 \n\t" \
718 "addi $10, $10, 4 \n\t" \
719 "mflo $14 \n\t" \
720 "mfhi $9 \n\t" \
721 "addu $14, $12, $14 \n\t" \
722 "lw $15, 0($11) \n\t" \
723 "sltu $12, $14, $12 \n\t" \
724 "addu $15, $14, $15 \n\t" \
725 "sltu $14, $15, $14 \n\t" \
726 "addu $12, $12, $9 \n\t" \
727 "sw $15, 0($11) \n\t" \
728 "addu $12, $12, $14 \n\t" \
729 "addi $11, $11, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000730
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200731#define MULADDC_STOP \
732 "sw $12, %0 \n\t" \
733 "sw $11, %1 \n\t" \
734 "sw $10, %2 \n\t" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100735 : "=m" (c), "=m" (d), "=m" (s) \
736 : "m" (s), "m" (d), "m" (c), "m" (b) \
737 : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
738 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000739
740#endif /* MIPS */
741#endif /* GNUC */
742
743#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
744
745#define MULADDC_INIT \
746 __asm mov esi, s \
747 __asm mov edi, d \
748 __asm mov ecx, c \
749 __asm mov ebx, b
750
751#define MULADDC_CORE \
752 __asm lodsd \
753 __asm mul ebx \
754 __asm add eax, ecx \
755 __asm adc edx, 0 \
756 __asm add eax, [edi] \
757 __asm adc edx, 0 \
758 __asm mov ecx, edx \
759 __asm stosd
760
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200761#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000762
763#define EMIT __asm _emit
764
765#define MULADDC_HUIT \
766 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
767 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
768 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
769 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
770 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
771 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
772 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
773 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
774 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
775 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
776 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
777 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
778 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
779 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
780 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
781 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
782 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
783 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
784 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
785 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
786 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
787 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
788 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
789 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
790 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
791 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
792 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
793 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
794 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
795 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
796 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
797 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
798 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
799 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
800 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
801 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
802 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
803 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
804 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
805 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
806 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
807 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
808 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
809 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
810 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
811 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
812 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
813 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
814 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
815 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
816 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
817 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
818 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
819 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
820 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
821 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
822 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
823 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
824 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
825 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
826 EMIT 0x0F EMIT 0x7E EMIT 0xC9
827
828#define MULADDC_STOP \
829 EMIT 0x0F EMIT 0x77 \
830 __asm mov c, ecx \
831 __asm mov d, edi \
832 __asm mov s, esi \
833
834#else
835
836#define MULADDC_STOP \
837 __asm mov c, ecx \
838 __asm mov d, edi \
839 __asm mov s, esi \
840
841#endif /* SSE2 */
842#endif /* MSVC */
843
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200844#endif /* MBEDTLS_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000845
846#if !defined(MULADDC_CORE)
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200847#if defined(MBEDTLS_HAVE_UDBL)
Paul Bakker5121ce52009-01-03 21:22:43 +0000848
849#define MULADDC_INIT \
850{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200851 mbedtls_t_udbl r; \
852 mbedtls_mpi_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000853
854#define MULADDC_CORE \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200855 r = *(s++) * (mbedtls_t_udbl) b; \
856 r0 = (mbedtls_mpi_uint) r; \
857 r1 = (mbedtls_mpi_uint)( r >> biL ); \
Paul Bakker5121ce52009-01-03 21:22:43 +0000858 r0 += c; r1 += (r0 < c); \
859 r0 += *d; r1 += (r0 < *d); \
860 c = r1; *(d++) = r0;
861
862#define MULADDC_STOP \
863}
864
865#else
866#define MULADDC_INIT \
867{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200868 mbedtls_mpi_uint s0, s1, b0, b1; \
869 mbedtls_mpi_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000870 b0 = ( b << biH ) >> biH; \
871 b1 = ( b >> biH );
872
873#define MULADDC_CORE \
874 s0 = ( *s << biH ) >> biH; \
875 s1 = ( *s >> biH ); s++; \
876 rx = s0 * b1; r0 = s0 * b0; \
877 ry = s1 * b0; r1 = s1 * b1; \
878 r1 += ( rx >> biH ); \
879 r1 += ( ry >> biH ); \
880 rx <<= biH; ry <<= biH; \
881 r0 += rx; r1 += (r0 < rx); \
882 r0 += ry; r1 += (r0 < ry); \
883 r0 += c; r1 += (r0 < c); \
884 r0 += *d; r1 += (r0 < *d); \
885 c = r1; *(d++) = r0;
886
887#define MULADDC_STOP \
888}
889
890#endif /* C (generic) */
891#endif /* C (longlong) */
892
893#endif /* bn_mul.h */