TrustedFirmware Git Browser
Code Review Sign In
review.trustedfirmware.org / mirror / mbed-tls / c49fbbf3eb2d7caa9c9f8edb44f857435d37347f / . / include / mbedtls / bn_mul.h
blob: db03ba2c054b10763f33e376ef59b725827241c4 [file] [log] [blame]
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]1/**
2 * \file bn_mul.h
Paul Bakkere0ccd0a2009-01-04 16:27:10 +0000[diff] [blame]3 *
Darryl Greena40a1012018-01-05 15:33:17 +0000[diff] [blame]4 * \brief Multi-precision integer library
5 */
6/*
Manuel Pégourié-Gonnard6fb81872015-07-27 11:11:48 +0200[diff] [blame]7 * Copyright (C) 2006-2015, ARM Limited, All Rights Reserved
Manuel Pégourié-Gonnard37ff1402015-09-04 14:21:07 +0200[diff] [blame]8 * 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 +0000[diff] [blame]21 *
Manuel Pégourié-Gonnardfe446432015-03-06 13:17:10 +0000[diff] [blame]22 * This file is part of mbed TLS (https://tls.mbed.org)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]23 */
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 +0200[diff] [blame]38#ifndef MBEDTLS_BN_MUL_H
39#define MBEDTLS_BN_MUL_H
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]40
Ron Eldor8b0cf2e2018-02-14 16:02:41 +0200[diff] [blame]41#if !defined(MBEDTLS_CONFIG_FILE)
Jaeden Ameroc49fbbf2019-07-04 20:01:14 +0100[diff] [blame^]42#include "mbedtls/config.h"
Ron Eldor8b0cf2e2018-02-14 16:02:41 +0200[diff] [blame]43#else
44#include MBEDTLS_CONFIG_FILE
45#endif
46
Jaeden Ameroc49fbbf2019-07-04 20:01:14 +0100[diff] [blame^]47#include "mbedtls/bignum.h"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]48
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200[diff] [blame]49#if defined(MBEDTLS_HAVE_ASM)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]50
Manuel Pégourié-Gonnardba194322015-05-29 09:47:57 +0200[diff] [blame]51#ifndef asm
52#define asm __asm
53#endif
54
Manuel Pégourié-Gonnard854dab92015-08-10 12:08:34 +0200[diff] [blame]55/* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */
56#if defined(__GNUC__) && \
57 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
Simon Butcher4b9a3ad2018-07-10 20:18:29 +0100[diff] [blame]58
59/*
60 * Disable use of the i386 assembly code below if option -O0, to disable all
61 * compiler optimisations, is passed, detected with __OPTIMIZE__
62 * This is done as the number of registers used in the assembly code doesn't
63 * work with the -O0 option.
64 */
Simon Butchere459f072018-07-23 09:44:42 +0100[diff] [blame]65#if defined(__i386__) && defined(__OPTIMIZE__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]66
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]67#define MULADDC_INIT \
68 asm( \
69 "movl %%ebx, %0 \n\t" \
70 "movl %5, %%esi \n\t" \
71 "movl %6, %%edi \n\t" \
72 "movl %7, %%ecx \n\t" \
73 "movl %8, %%ebx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]74
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]75#define MULADDC_CORE \
76 "lodsl \n\t" \
77 "mull %%ebx \n\t" \
78 "addl %%ecx, %%eax \n\t" \
79 "adcl $0, %%edx \n\t" \
80 "addl (%%edi), %%eax \n\t" \
81 "adcl $0, %%edx \n\t" \
82 "movl %%edx, %%ecx \n\t" \
83 "stosl \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]84
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200[diff] [blame]85#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]86
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]87#define MULADDC_HUIT \
88 "movd %%ecx, %%mm1 \n\t" \
89 "movd %%ebx, %%mm0 \n\t" \
90 "movd (%%edi), %%mm3 \n\t" \
91 "paddq %%mm3, %%mm1 \n\t" \
92 "movd (%%esi), %%mm2 \n\t" \
93 "pmuludq %%mm0, %%mm2 \n\t" \
94 "movd 4(%%esi), %%mm4 \n\t" \
95 "pmuludq %%mm0, %%mm4 \n\t" \
96 "movd 8(%%esi), %%mm6 \n\t" \
97 "pmuludq %%mm0, %%mm6 \n\t" \
98 "movd 12(%%esi), %%mm7 \n\t" \
99 "pmuludq %%mm0, %%mm7 \n\t" \
100 "paddq %%mm2, %%mm1 \n\t" \
101 "movd 4(%%edi), %%mm3 \n\t" \
102 "paddq %%mm4, %%mm3 \n\t" \
103 "movd 8(%%edi), %%mm5 \n\t" \
104 "paddq %%mm6, %%mm5 \n\t" \
105 "movd 12(%%edi), %%mm4 \n\t" \
106 "paddq %%mm4, %%mm7 \n\t" \
107 "movd %%mm1, (%%edi) \n\t" \
108 "movd 16(%%esi), %%mm2 \n\t" \
109 "pmuludq %%mm0, %%mm2 \n\t" \
110 "psrlq $32, %%mm1 \n\t" \
111 "movd 20(%%esi), %%mm4 \n\t" \
112 "pmuludq %%mm0, %%mm4 \n\t" \
113 "paddq %%mm3, %%mm1 \n\t" \
114 "movd 24(%%esi), %%mm6 \n\t" \
115 "pmuludq %%mm0, %%mm6 \n\t" \
116 "movd %%mm1, 4(%%edi) \n\t" \
117 "psrlq $32, %%mm1 \n\t" \
118 "movd 28(%%esi), %%mm3 \n\t" \
119 "pmuludq %%mm0, %%mm3 \n\t" \
120 "paddq %%mm5, %%mm1 \n\t" \
121 "movd 16(%%edi), %%mm5 \n\t" \
122 "paddq %%mm5, %%mm2 \n\t" \
123 "movd %%mm1, 8(%%edi) \n\t" \
124 "psrlq $32, %%mm1 \n\t" \
125 "paddq %%mm7, %%mm1 \n\t" \
126 "movd 20(%%edi), %%mm5 \n\t" \
127 "paddq %%mm5, %%mm4 \n\t" \
128 "movd %%mm1, 12(%%edi) \n\t" \
129 "psrlq $32, %%mm1 \n\t" \
130 "paddq %%mm2, %%mm1 \n\t" \
131 "movd 24(%%edi), %%mm5 \n\t" \
132 "paddq %%mm5, %%mm6 \n\t" \
133 "movd %%mm1, 16(%%edi) \n\t" \
134 "psrlq $32, %%mm1 \n\t" \
135 "paddq %%mm4, %%mm1 \n\t" \
136 "movd 28(%%edi), %%mm5 \n\t" \
137 "paddq %%mm5, %%mm3 \n\t" \
138 "movd %%mm1, 20(%%edi) \n\t" \
139 "psrlq $32, %%mm1 \n\t" \
140 "paddq %%mm6, %%mm1 \n\t" \
141 "movd %%mm1, 24(%%edi) \n\t" \
142 "psrlq $32, %%mm1 \n\t" \
143 "paddq %%mm3, %%mm1 \n\t" \
144 "movd %%mm1, 28(%%edi) \n\t" \
145 "addl $32, %%edi \n\t" \
146 "addl $32, %%esi \n\t" \
147 "psrlq $32, %%mm1 \n\t" \
148 "movd %%mm1, %%ecx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]149
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]150#define MULADDC_STOP \
151 "emms \n\t" \
152 "movl %4, %%ebx \n\t" \
153 "movl %%ecx, %1 \n\t" \
154 "movl %%edi, %2 \n\t" \
155 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]156 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
157 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
Simon Butcher53571642018-06-24 12:58:31 +0100[diff] [blame]158 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]159 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]160
161#else
162
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]163#define MULADDC_STOP \
164 "movl %4, %%ebx \n\t" \
165 "movl %%ecx, %1 \n\t" \
166 "movl %%edi, %2 \n\t" \
167 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]168 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
169 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
Simon Butcher53571642018-06-24 12:58:31 +0100[diff] [blame]170 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000[diff] [blame]171 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]172#endif /* SSE2 */
173#endif /* i386 */
174
175#if defined(__amd64__) || defined (__x86_64__)
176
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]177#define MULADDC_INIT \
178 asm( \
Simon Butchera86de142018-09-30 12:09:47 +0100[diff] [blame]179 "xorq %%r8, %%r8\n"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]180
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]181#define MULADDC_CORE \
Simon Butchera86de142018-09-30 12:09:47 +0100[diff] [blame]182 "movq (%%rsi), %%rax\n" \
183 "mulq %%rbx\n" \
184 "addq $8, %%rsi\n" \
185 "addq %%rcx, %%rax\n" \
186 "movq %%r8, %%rcx\n" \
187 "adcq $0, %%rdx\n" \
188 "nop \n" \
189 "addq %%rax, (%%rdi)\n" \
190 "adcq %%rdx, %%rcx\n" \
191 "addq $8, %%rdi\n"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]192
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]193#define MULADDC_STOP \
James Cowgill21e402a2015-12-17 01:51:09 +0000[diff] [blame]194 : "+c" (c), "+D" (d), "+S" (s) \
195 : "b" (b) \
196 : "rax", "rdx", "r8" \
Manuel Pégourié-Gonnarddef018d2014-01-07 17:50:46 +0100[diff] [blame]197 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]198
199#endif /* AMD64 */
200
201#if defined(__mc68020__) || defined(__mcpu32__)
202
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]203#define MULADDC_INIT \
204 asm( \
205 "movl %3, %%a2 \n\t" \
206 "movl %4, %%a3 \n\t" \
207 "movl %5, %%d3 \n\t" \
208 "movl %6, %%d2 \n\t" \
209 "moveq #0, %%d0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]210
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]211#define MULADDC_CORE \
212 "movel %%a2@+, %%d1 \n\t" \
213 "mulul %%d2, %%d4:%%d1 \n\t" \
214 "addl %%d3, %%d1 \n\t" \
215 "addxl %%d0, %%d4 \n\t" \
216 "moveq #0, %%d3 \n\t" \
217 "addl %%d1, %%a3@+ \n\t" \
218 "addxl %%d4, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]219
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]220#define MULADDC_STOP \
221 "movl %%d3, %0 \n\t" \
222 "movl %%a3, %1 \n\t" \
223 "movl %%a2, %2 \n\t" \
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100[diff] [blame]224 : "=m" (c), "=m" (d), "=m" (s) \
225 : "m" (s), "m" (d), "m" (c), "m" (b) \
226 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
227 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]228
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]229#define MULADDC_HUIT \
230 "movel %%a2@+, %%d1 \n\t" \
231 "mulul %%d2, %%d4:%%d1 \n\t" \
232 "addxl %%d3, %%d1 \n\t" \
233 "addxl %%d0, %%d4 \n\t" \
234 "addl %%d1, %%a3@+ \n\t" \
235 "movel %%a2@+, %%d1 \n\t" \
236 "mulul %%d2, %%d3:%%d1 \n\t" \
237 "addxl %%d4, %%d1 \n\t" \
238 "addxl %%d0, %%d3 \n\t" \
239 "addl %%d1, %%a3@+ \n\t" \
240 "movel %%a2@+, %%d1 \n\t" \
241 "mulul %%d2, %%d4:%%d1 \n\t" \
242 "addxl %%d3, %%d1 \n\t" \
243 "addxl %%d0, %%d4 \n\t" \
244 "addl %%d1, %%a3@+ \n\t" \
245 "movel %%a2@+, %%d1 \n\t" \
246 "mulul %%d2, %%d3:%%d1 \n\t" \
247 "addxl %%d4, %%d1 \n\t" \
248 "addxl %%d0, %%d3 \n\t" \
249 "addl %%d1, %%a3@+ \n\t" \
250 "movel %%a2@+, %%d1 \n\t" \
251 "mulul %%d2, %%d4:%%d1 \n\t" \
252 "addxl %%d3, %%d1 \n\t" \
253 "addxl %%d0, %%d4 \n\t" \
254 "addl %%d1, %%a3@+ \n\t" \
255 "movel %%a2@+, %%d1 \n\t" \
256 "mulul %%d2, %%d3:%%d1 \n\t" \
257 "addxl %%d4, %%d1 \n\t" \
258 "addxl %%d0, %%d3 \n\t" \
259 "addl %%d1, %%a3@+ \n\t" \
260 "movel %%a2@+, %%d1 \n\t" \
261 "mulul %%d2, %%d4:%%d1 \n\t" \
262 "addxl %%d3, %%d1 \n\t" \
263 "addxl %%d0, %%d4 \n\t" \
264 "addl %%d1, %%a3@+ \n\t" \
265 "movel %%a2@+, %%d1 \n\t" \
266 "mulul %%d2, %%d3:%%d1 \n\t" \
267 "addxl %%d4, %%d1 \n\t" \
268 "addxl %%d0, %%d3 \n\t" \
269 "addl %%d1, %%a3@+ \n\t" \
270 "addxl %%d0, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]271
272#endif /* MC68000 */
273
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]274#if defined(__powerpc64__) || defined(__ppc64__)
275
276#if defined(__MACH__) && defined(__APPLE__)
277
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]278#define MULADDC_INIT \
279 asm( \
280 "ld r3, %3 \n\t" \
281 "ld r4, %4 \n\t" \
282 "ld r5, %5 \n\t" \
283 "ld r6, %6 \n\t" \
284 "addi r3, r3, -8 \n\t" \
285 "addi r4, r4, -8 \n\t" \
286 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]287
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]288#define MULADDC_CORE \
289 "ldu r7, 8(r3) \n\t" \
290 "mulld r8, r7, r6 \n\t" \
291 "mulhdu r9, r7, r6 \n\t" \
292 "adde r8, r8, r5 \n\t" \
293 "ld r7, 8(r4) \n\t" \
294 "addze r5, r9 \n\t" \
295 "addc r8, r8, r7 \n\t" \
296 "stdu r8, 8(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]297
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]298#define MULADDC_STOP \
299 "addze r5, r5 \n\t" \
300 "addi r4, r4, 8 \n\t" \
301 "addi r3, r3, 8 \n\t" \
302 "std r5, %0 \n\t" \
303 "std r4, %1 \n\t" \
304 "std r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100[diff] [blame]305 : "=m" (c), "=m" (d), "=m" (s) \
306 : "m" (s), "m" (d), "m" (c), "m" (b) \
307 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
308 );
309
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]310
Paul Bakker9af723c2014-05-01 13:03:14 +0200[diff] [blame]311#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]312
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]313#define MULADDC_INIT \
314 asm( \
315 "ld %%r3, %3 \n\t" \
316 "ld %%r4, %4 \n\t" \
317 "ld %%r5, %5 \n\t" \
318 "ld %%r6, %6 \n\t" \
319 "addi %%r3, %%r3, -8 \n\t" \
320 "addi %%r4, %%r4, -8 \n\t" \
321 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]322
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]323#define MULADDC_CORE \
324 "ldu %%r7, 8(%%r3) \n\t" \
325 "mulld %%r8, %%r7, %%r6 \n\t" \
326 "mulhdu %%r9, %%r7, %%r6 \n\t" \
327 "adde %%r8, %%r8, %%r5 \n\t" \
328 "ld %%r7, 8(%%r4) \n\t" \
329 "addze %%r5, %%r9 \n\t" \
330 "addc %%r8, %%r8, %%r7 \n\t" \
331 "stdu %%r8, 8(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]332
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]333#define MULADDC_STOP \
334 "addze %%r5, %%r5 \n\t" \
335 "addi %%r4, %%r4, 8 \n\t" \
336 "addi %%r3, %%r3, 8 \n\t" \
337 "std %%r5, %0 \n\t" \
338 "std %%r4, %1 \n\t" \
339 "std %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100[diff] [blame]340 : "=m" (c), "=m" (d), "=m" (s) \
341 : "m" (s), "m" (d), "m" (c), "m" (b) \
342 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
343 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]344
Paul Bakker9af723c2014-05-01 13:03:14 +0200[diff] [blame]345#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]346
Barry K. Nathan35e7cb92014-05-05 23:26:13 -0700[diff] [blame]347#elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]348
349#if defined(__MACH__) && defined(__APPLE__)
350
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]351#define MULADDC_INIT \
352 asm( \
353 "lwz r3, %3 \n\t" \
354 "lwz r4, %4 \n\t" \
355 "lwz r5, %5 \n\t" \
356 "lwz r6, %6 \n\t" \
357 "addi r3, r3, -4 \n\t" \
358 "addi r4, r4, -4 \n\t" \
359 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]360
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]361#define MULADDC_CORE \
362 "lwzu r7, 4(r3) \n\t" \
363 "mullw r8, r7, r6 \n\t" \
364 "mulhwu r9, r7, r6 \n\t" \
365 "adde r8, r8, r5 \n\t" \
366 "lwz r7, 4(r4) \n\t" \
367 "addze r5, r9 \n\t" \
368 "addc r8, r8, r7 \n\t" \
369 "stwu r8, 4(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]370
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]371#define MULADDC_STOP \
372 "addze r5, r5 \n\t" \
373 "addi r4, r4, 4 \n\t" \
374 "addi r3, r3, 4 \n\t" \
375 "stw r5, %0 \n\t" \
376 "stw r4, %1 \n\t" \
377 "stw r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100[diff] [blame]378 : "=m" (c), "=m" (d), "=m" (s) \
379 : "m" (s), "m" (d), "m" (c), "m" (b) \
380 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
381 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]382
Paul Bakker9af723c2014-05-01 13:03:14 +0200[diff] [blame]383#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]384
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]385#define MULADDC_INIT \
386 asm( \
387 "lwz %%r3, %3 \n\t" \
388 "lwz %%r4, %4 \n\t" \
389 "lwz %%r5, %5 \n\t" \
390 "lwz %%r6, %6 \n\t" \
391 "addi %%r3, %%r3, -4 \n\t" \
392 "addi %%r4, %%r4, -4 \n\t" \
393 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]394
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]395#define MULADDC_CORE \
396 "lwzu %%r7, 4(%%r3) \n\t" \
397 "mullw %%r8, %%r7, %%r6 \n\t" \
398 "mulhwu %%r9, %%r7, %%r6 \n\t" \
399 "adde %%r8, %%r8, %%r5 \n\t" \
400 "lwz %%r7, 4(%%r4) \n\t" \
401 "addze %%r5, %%r9 \n\t" \
402 "addc %%r8, %%r8, %%r7 \n\t" \
403 "stwu %%r8, 4(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]404
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]405#define MULADDC_STOP \
406 "addze %%r5, %%r5 \n\t" \
407 "addi %%r4, %%r4, 4 \n\t" \
408 "addi %%r3, %%r3, 4 \n\t" \
409 "stw %%r5, %0 \n\t" \
410 "stw %%r4, %1 \n\t" \
411 "stw %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100[diff] [blame]412 : "=m" (c), "=m" (d), "=m" (s) \
413 : "m" (s), "m" (d), "m" (c), "m" (b) \
414 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
415 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]416
Paul Bakker9af723c2014-05-01 13:03:14 +0200[diff] [blame]417#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]418
419#endif /* PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]420
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200[diff] [blame]421/*
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200[diff] [blame]422 * The Sparc(64) assembly is reported to be broken.
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200[diff] [blame]423 * Disable it for now, until we're able to fix it.
424 */
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200[diff] [blame]425#if 0 && defined(__sparc__)
426#if defined(__sparc64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]427
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]428#define MULADDC_INIT \
429 asm( \
430 "ldx %3, %%o0 \n\t" \
431 "ldx %4, %%o1 \n\t" \
432 "ld %5, %%o2 \n\t" \
433 "ld %6, %%o3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]434
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]435#define MULADDC_CORE \
436 "ld [%%o0], %%o4 \n\t" \
437 "inc 4, %%o0 \n\t" \
438 "ld [%%o1], %%o5 \n\t" \
439 "umul %%o3, %%o4, %%o4 \n\t" \
440 "addcc %%o4, %%o2, %%o4 \n\t" \
441 "rd %%y, %%g1 \n\t" \
442 "addx %%g1, 0, %%g1 \n\t" \
443 "addcc %%o4, %%o5, %%o4 \n\t" \
444 "st %%o4, [%%o1] \n\t" \
445 "addx %%g1, 0, %%o2 \n\t" \
446 "inc 4, %%o1 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]447
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]448 #define MULADDC_STOP \
449 "st %%o2, %0 \n\t" \
450 "stx %%o1, %1 \n\t" \
451 "stx %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000[diff] [blame]452 : "=m" (c), "=m" (d), "=m" (s) \
453 : "m" (s), "m" (d), "m" (c), "m" (b) \
454 : "g1", "o0", "o1", "o2", "o3", "o4", \
455 "o5" \
456 );
Paul Bakker4f024b72012-10-30 07:29:57 +0000[diff] [blame]457
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200[diff] [blame]458#else /* __sparc64__ */
Paul Bakker4f024b72012-10-30 07:29:57 +0000[diff] [blame]459
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]460#define MULADDC_INIT \
461 asm( \
462 "ld %3, %%o0 \n\t" \
463 "ld %4, %%o1 \n\t" \
464 "ld %5, %%o2 \n\t" \
465 "ld %6, %%o3 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000[diff] [blame]466
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]467#define MULADDC_CORE \
468 "ld [%%o0], %%o4 \n\t" \
469 "inc 4, %%o0 \n\t" \
470 "ld [%%o1], %%o5 \n\t" \
471 "umul %%o3, %%o4, %%o4 \n\t" \
472 "addcc %%o4, %%o2, %%o4 \n\t" \
473 "rd %%y, %%g1 \n\t" \
474 "addx %%g1, 0, %%g1 \n\t" \
475 "addcc %%o4, %%o5, %%o4 \n\t" \
476 "st %%o4, [%%o1] \n\t" \
477 "addx %%g1, 0, %%o2 \n\t" \
478 "inc 4, %%o1 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000[diff] [blame]479
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]480#define MULADDC_STOP \
481 "st %%o2, %0 \n\t" \
482 "st %%o1, %1 \n\t" \
483 "st %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000[diff] [blame]484 : "=m" (c), "=m" (d), "=m" (s) \
485 : "m" (s), "m" (d), "m" (c), "m" (b) \
486 : "g1", "o0", "o1", "o2", "o3", "o4", \
487 "o5" \
488 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]489
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200[diff] [blame]490#endif /* __sparc64__ */
491#endif /* __sparc__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]492
493#if defined(__microblaze__) || defined(microblaze)
494
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]495#define MULADDC_INIT \
496 asm( \
497 "lwi r3, %3 \n\t" \
498 "lwi r4, %4 \n\t" \
499 "lwi r5, %5 \n\t" \
500 "lwi r6, %6 \n\t" \
501 "andi r7, r6, 0xffff \n\t" \
502 "bsrli r6, r6, 16 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]503
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]504#define MULADDC_CORE \
505 "lhui r8, r3, 0 \n\t" \
506 "addi r3, r3, 2 \n\t" \
507 "lhui r9, r3, 0 \n\t" \
508 "addi r3, r3, 2 \n\t" \
509 "mul r10, r9, r6 \n\t" \
510 "mul r11, r8, r7 \n\t" \
511 "mul r12, r9, r7 \n\t" \
512 "mul r13, r8, r6 \n\t" \
513 "bsrli r8, r10, 16 \n\t" \
514 "bsrli r9, r11, 16 \n\t" \
515 "add r13, r13, r8 \n\t" \
516 "add r13, r13, r9 \n\t" \
517 "bslli r10, r10, 16 \n\t" \
518 "bslli r11, r11, 16 \n\t" \
519 "add r12, r12, r10 \n\t" \
520 "addc r13, r13, r0 \n\t" \
521 "add r12, r12, r11 \n\t" \
522 "addc r13, r13, r0 \n\t" \
523 "lwi r10, r4, 0 \n\t" \
524 "add r12, r12, r10 \n\t" \
525 "addc r13, r13, r0 \n\t" \
526 "add r12, r12, r5 \n\t" \
527 "addc r5, r13, r0 \n\t" \
528 "swi r12, r4, 0 \n\t" \
529 "addi r4, r4, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]530
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]531#define MULADDC_STOP \
532 "swi r5, %0 \n\t" \
533 "swi r4, %1 \n\t" \
534 "swi r3, %2 \n\t" \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100[diff] [blame]535 : "=m" (c), "=m" (d), "=m" (s) \
536 : "m" (s), "m" (d), "m" (c), "m" (b) \
Zach van Rijne7d3f8e2018-05-21 10:52:34 -0400[diff] [blame]537 : "r3", "r4", "r5", "r6", "r7", "r8", \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100[diff] [blame]538 "r9", "r10", "r11", "r12", "r13" \
539 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]540
541#endif /* MicroBlaze */
542
543#if defined(__tricore__)
544
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]545#define MULADDC_INIT \
546 asm( \
547 "ld.a %%a2, %3 \n\t" \
548 "ld.a %%a3, %4 \n\t" \
549 "ld.w %%d4, %5 \n\t" \
550 "ld.w %%d1, %6 \n\t" \
551 "xor %%d5, %%d5 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]552
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]553#define MULADDC_CORE \
554 "ld.w %%d0, [%%a2+] \n\t" \
555 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
556 "ld.w %%d0, [%%a3] \n\t" \
557 "addx %%d2, %%d2, %%d0 \n\t" \
558 "addc %%d3, %%d3, 0 \n\t" \
559 "mov %%d4, %%d3 \n\t" \
560 "st.w [%%a3+], %%d2 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]561
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]562#define MULADDC_STOP \
563 "st.w %0, %%d4 \n\t" \
564 "st.a %1, %%a3 \n\t" \
565 "st.a %2, %%a2 \n\t" \
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100[diff] [blame]566 : "=m" (c), "=m" (d), "=m" (s) \
567 : "m" (s), "m" (d), "m" (c), "m" (b) \
568 : "d0", "d1", "e2", "d4", "a2", "a3" \
569 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]570
571#endif /* TriCore */
572
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100[diff] [blame]573/*
Simon Butcher3ad2efd2018-05-02 14:49:38 +0100[diff] [blame]574 * Note, gcc -O0 by default uses r7 for the frame pointer, so it complains about
575 * our use of r7 below, unless -fomit-frame-pointer is passed.
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100[diff] [blame]576 *
577 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
578 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
579 * clang and armcc5 under the same conditions).
580 *
581 * So, only use the optimized assembly below for optimized build, which avoids
582 * the build error and is pretty reasonable anyway.
583 */
584#if defined(__GNUC__) && !defined(__OPTIMIZE__)
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100[diff] [blame]585#define MULADDC_CANNOT_USE_R7
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100[diff] [blame]586#endif
587
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100[diff] [blame]588#if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]589
Paul Bakkerfc4f46f2013-06-24 19:23:56 +0200[diff] [blame]590#if defined(__thumb__) && !defined(__thumb2__)
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000[diff] [blame]591
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]592#define MULADDC_INIT \
593 asm( \
594 "ldr r0, %3 \n\t" \
595 "ldr r1, %4 \n\t" \
596 "ldr r2, %5 \n\t" \
597 "ldr r3, %6 \n\t" \
598 "lsr r7, r3, #16 \n\t" \
599 "mov r9, r7 \n\t" \
600 "lsl r7, r3, #16 \n\t" \
601 "lsr r7, r7, #16 \n\t" \
602 "mov r8, r7 \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000[diff] [blame]603
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]604#define MULADDC_CORE \
605 "ldmia r0!, {r6} \n\t" \
606 "lsr r7, r6, #16 \n\t" \
607 "lsl r6, r6, #16 \n\t" \
608 "lsr r6, r6, #16 \n\t" \
609 "mov r4, r8 \n\t" \
610 "mul r4, r6 \n\t" \
611 "mov r3, r9 \n\t" \
612 "mul r6, r3 \n\t" \
613 "mov r5, r9 \n\t" \
614 "mul r5, r7 \n\t" \
615 "mov r3, r8 \n\t" \
616 "mul r7, r3 \n\t" \
617 "lsr r3, r6, #16 \n\t" \
618 "add r5, r5, r3 \n\t" \
619 "lsr r3, r7, #16 \n\t" \
620 "add r5, r5, r3 \n\t" \
621 "add r4, r4, r2 \n\t" \
622 "mov r2, #0 \n\t" \
623 "adc r5, r2 \n\t" \
624 "lsl r3, r6, #16 \n\t" \
625 "add r4, r4, r3 \n\t" \
626 "adc r5, r2 \n\t" \
627 "lsl r3, r7, #16 \n\t" \
628 "add r4, r4, r3 \n\t" \
629 "adc r5, r2 \n\t" \
630 "ldr r3, [r1] \n\t" \
631 "add r4, r4, r3 \n\t" \
632 "adc r2, r5 \n\t" \
633 "stmia r1!, {r4} \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000[diff] [blame]634
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]635#define MULADDC_STOP \
636 "str r2, %0 \n\t" \
637 "str r1, %1 \n\t" \
638 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100[diff] [blame]639 : "=m" (c), "=m" (d), "=m" (s) \
640 : "m" (s), "m" (d), "m" (c), "m" (b) \
641 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200[diff] [blame]642 "r6", "r7", "r8", "r9", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100[diff] [blame]643 );
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000[diff] [blame]644
Aurelien Jarno16b1bd82018-05-21 22:01:21 +0200[diff] [blame]645#elif defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)
646
647#define MULADDC_INIT \
648 asm(
649
650#define MULADDC_CORE \
651 "ldr r0, [%0], #4 \n\t" \
652 "ldr r1, [%1] \n\t" \
653 "umaal r1, %2, %3, r0 \n\t" \
654 "str r1, [%1], #4 \n\t"
655
656#define MULADDC_STOP \
657 : "=r" (s), "=r" (d), "=r" (c) \
658 : "r" (b), "0" (s), "1" (d), "2" (c) \
659 : "r0", "r1", "memory" \
660 );
661
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000[diff] [blame]662#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000[diff] [blame]663
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]664#define MULADDC_INIT \
665 asm( \
666 "ldr r0, %3 \n\t" \
667 "ldr r1, %4 \n\t" \
668 "ldr r2, %5 \n\t" \
669 "ldr r3, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]670
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]671#define MULADDC_CORE \
672 "ldr r4, [r0], #4 \n\t" \
673 "mov r5, #0 \n\t" \
674 "ldr r6, [r1] \n\t" \
675 "umlal r2, r5, r3, r4 \n\t" \
676 "adds r7, r6, r2 \n\t" \
677 "adc r2, r5, #0 \n\t" \
678 "str r7, [r1], #4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]679
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]680#define MULADDC_STOP \
681 "str r2, %0 \n\t" \
682 "str r1, %1 \n\t" \
683 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100[diff] [blame]684 : "=m" (c), "=m" (d), "=m" (s) \
685 : "m" (s), "m" (d), "m" (c), "m" (b) \
686 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200[diff] [blame]687 "r6", "r7", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100[diff] [blame]688 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]689
Paul Bakkera2713a32011-11-18 12:47:23 +0000[diff] [blame]690#endif /* Thumb */
691
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]692#endif /* ARMv3 */
693
694#if defined(__alpha__)
695
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]696#define MULADDC_INIT \
697 asm( \
698 "ldq $1, %3 \n\t" \
699 "ldq $2, %4 \n\t" \
700 "ldq $3, %5 \n\t" \
701 "ldq $4, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]702
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]703#define MULADDC_CORE \
704 "ldq $6, 0($1) \n\t" \
705 "addq $1, 8, $1 \n\t" \
706 "mulq $6, $4, $7 \n\t" \
707 "umulh $6, $4, $6 \n\t" \
708 "addq $7, $3, $7 \n\t" \
709 "cmpult $7, $3, $3 \n\t" \
710 "ldq $5, 0($2) \n\t" \
711 "addq $7, $5, $7 \n\t" \
712 "cmpult $7, $5, $5 \n\t" \
713 "stq $7, 0($2) \n\t" \
714 "addq $2, 8, $2 \n\t" \
715 "addq $6, $3, $3 \n\t" \
716 "addq $5, $3, $3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]717
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]718#define MULADDC_STOP \
719 "stq $3, %0 \n\t" \
720 "stq $2, %1 \n\t" \
721 "stq $1, %2 \n\t" \
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100[diff] [blame]722 : "=m" (c), "=m" (d), "=m" (s) \
723 : "m" (s), "m" (d), "m" (c), "m" (b) \
724 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
725 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]726#endif /* Alpha */
727
James Cowgilld1e7e8b2014-12-16 15:24:06 +0000[diff] [blame]728#if defined(__mips__) && !defined(__mips64)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]729
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]730#define MULADDC_INIT \
731 asm( \
732 "lw $10, %3 \n\t" \
733 "lw $11, %4 \n\t" \
734 "lw $12, %5 \n\t" \
735 "lw $13, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]736
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]737#define MULADDC_CORE \
738 "lw $14, 0($10) \n\t" \
739 "multu $13, $14 \n\t" \
740 "addi $10, $10, 4 \n\t" \
741 "mflo $14 \n\t" \
742 "mfhi $9 \n\t" \
743 "addu $14, $12, $14 \n\t" \
744 "lw $15, 0($11) \n\t" \
745 "sltu $12, $14, $12 \n\t" \
746 "addu $15, $14, $15 \n\t" \
747 "sltu $14, $15, $14 \n\t" \
748 "addu $12, $12, $9 \n\t" \
749 "sw $15, 0($11) \n\t" \
750 "addu $12, $12, $14 \n\t" \
751 "addi $11, $11, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]752
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200[diff] [blame]753#define MULADDC_STOP \
754 "sw $12, %0 \n\t" \
755 "sw $11, %1 \n\t" \
756 "sw $10, %2 \n\t" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100[diff] [blame]757 : "=m" (c), "=m" (d), "=m" (s) \
758 : "m" (s), "m" (d), "m" (c), "m" (b) \
Jeffrey Martind25fd8d2019-01-14 18:01:40 -0600[diff] [blame]759 : "$9", "$10", "$11", "$12", "$13", "$14", "$15", "lo", "hi" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100[diff] [blame]760 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]761
762#endif /* MIPS */
763#endif /* GNUC */
764
765#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
766
767#define MULADDC_INIT \
768 __asm mov esi, s \
769 __asm mov edi, d \
770 __asm mov ecx, c \
771 __asm mov ebx, b
772
773#define MULADDC_CORE \
774 __asm lodsd \
775 __asm mul ebx \
776 __asm add eax, ecx \
777 __asm adc edx, 0 \
778 __asm add eax, [edi] \
779 __asm adc edx, 0 \
780 __asm mov ecx, edx \
781 __asm stosd
782
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200[diff] [blame]783#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]784
785#define EMIT __asm _emit
786
787#define MULADDC_HUIT \
788 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
789 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
790 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
791 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
792 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
793 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
794 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
795 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
796 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
797 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
798 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
799 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
800 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
801 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
802 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
803 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
804 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
805 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
806 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
807 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
808 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
809 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
810 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
811 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
812 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
813 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
814 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
815 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
816 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
817 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
818 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
819 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
820 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
821 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
822 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
823 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
824 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
825 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
826 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
827 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
828 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
829 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
830 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
831 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
832 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
833 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
834 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
835 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
836 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
837 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
838 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
839 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
840 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
841 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
842 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
843 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
844 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
845 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
846 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
847 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
848 EMIT 0x0F EMIT 0x7E EMIT 0xC9
849
850#define MULADDC_STOP \
851 EMIT 0x0F EMIT 0x77 \
852 __asm mov c, ecx \
853 __asm mov d, edi \
854 __asm mov s, esi \
855
856#else
857
858#define MULADDC_STOP \
859 __asm mov c, ecx \
860 __asm mov d, edi \
861 __asm mov s, esi \
862
863#endif /* SSE2 */
864#endif /* MSVC */
865
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200[diff] [blame]866#endif /* MBEDTLS_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]867
868#if !defined(MULADDC_CORE)
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200[diff] [blame]869#if defined(MBEDTLS_HAVE_UDBL)
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]870
871#define MULADDC_INIT \
872{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200[diff] [blame]873 mbedtls_t_udbl r; \
874 mbedtls_mpi_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]875
876#define MULADDC_CORE \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200[diff] [blame]877 r = *(s++) * (mbedtls_t_udbl) b; \
878 r0 = (mbedtls_mpi_uint) r; \
879 r1 = (mbedtls_mpi_uint)( r >> biL ); \
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]880 r0 += c; r1 += (r0 < c); \
881 r0 += *d; r1 += (r0 < *d); \
882 c = r1; *(d++) = r0;
883
884#define MULADDC_STOP \
885}
886
887#else
888#define MULADDC_INIT \
889{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200[diff] [blame]890 mbedtls_mpi_uint s0, s1, b0, b1; \
891 mbedtls_mpi_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000[diff] [blame]892 b0 = ( b << biH ) >> biH; \
893 b1 = ( b >> biH );
894
895#define MULADDC_CORE \
896 s0 = ( *s << biH ) >> biH; \
897 s1 = ( *s >> biH ); s++; \
898 rx = s0 * b1; r0 = s0 * b0; \
899 ry = s1 * b0; r1 = s1 * b1; \
900 r1 += ( rx >> biH ); \
901 r1 += ( ry >> biH ); \
902 rx <<= biH; ry <<= biH; \
903 r0 += rx; r1 += (r0 < rx); \
904 r0 += ry; r1 += (r0 < ry); \
905 r0 += c; r1 += (r0 < c); \
906 r0 += *d; r1 += (r0 < *d); \
907 c = r1; *(d++) = r0;
908
909#define MULADDC_STOP \
910}
911
912#endif /* C (generic) */
913#endif /* C (longlong) */
914
915#endif /* bn_mul.h */