blob: b6b3b9e4ca42dc18a9e1ff9d72262a7159554cb2 [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/*
Bence Szépkúti1e148272020-08-07 13:07:28 +02007 * Copyright The Mbed TLS Contributors
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 Bakker5121ce52009-01-03 21:22:43 +000021 */
22/*
23 * Multiply source vector [s] with b, add result
24 * to destination vector [d] and set carry c.
25 *
26 * Currently supports:
27 *
28 * . IA-32 (386+) . AMD64 / EM64T
29 * . IA-32 (SSE2) . Motorola 68000
30 * . PowerPC, 32-bit . MicroBlaze
31 * . PowerPC, 64-bit . TriCore
32 * . SPARC v8 . ARM v3+
33 * . Alpha . MIPS32
34 * . C, longlong . C, generic
35 */
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +020036#ifndef MBEDTLS_BN_MUL_H
37#define MBEDTLS_BN_MUL_H
Paul Bakker5121ce52009-01-03 21:22:43 +000038
Bence Szépkútic662b362021-05-27 11:25:03 +020039#include "mbedtls/build_info.h"
Ron Eldor8b0cf2e2018-02-14 16:02:41 +020040
Jaeden Ameroc49fbbf2019-07-04 20:01:14 +010041#include "mbedtls/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 Butcher4b9a3ad2018-07-10 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 */
Simon Butchere459f072018-07-23 09:44:42 +010059#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 Butcher53571642018-06-24 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 Butcher53571642018-06-24 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( \
Simon Butchera86de142018-09-30 12:09:47 +0100173 "xorq %%r8, %%r8\n"
Paul Bakker5121ce52009-01-03 21:22:43 +0000174
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200175#define MULADDC_CORE \
Simon Butchera86de142018-09-30 12:09:47 +0100176 "movq (%%rsi), %%rax\n" \
177 "mulq %%rbx\n" \
178 "addq $8, %%rsi\n" \
179 "addq %%rcx, %%rax\n" \
180 "movq %%r8, %%rcx\n" \
181 "adcq $0, %%rdx\n" \
182 "nop \n" \
183 "addq %%rax, (%%rdi)\n" \
184 "adcq %%rdx, %%rcx\n" \
185 "addq $8, %%rdi\n"
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
Ko-cc1871e2018-08-16 02:01:57 -0700195#if defined(__aarch64__)
196
197#define MULADDC_INIT \
198 asm(
199
200#define MULADDC_CORE \
Ko-cb260bb2018-08-20 13:59:53 +0100201 "ldr x4, [%2], #8 \n\t" \
202 "ldr x5, [%1] \n\t" \
203 "mul x6, x4, %3 \n\t" \
204 "umulh x7, x4, %3 \n\t" \
Ko-cc1871e2018-08-16 02:01:57 -0700205 "adds x5, x5, x6 \n\t" \
206 "adc x7, x7, xzr \n\t" \
Ko-cb260bb2018-08-20 13:59:53 +0100207 "adds x5, x5, %0 \n\t" \
Ko-cc1871e2018-08-16 02:01:57 -0700208 "adc %0, x7, xzr \n\t" \
209 "str x5, [%1], #8 \n\t"
210
Ko-cb260bb2018-08-20 13:59:53 +0100211#define MULADDC_STOP \
212 : "+r" (c), "+r" (d), "+r" (s) \
213 : "r" (b) \
214 : "x4", "x5", "x6", "x7", "cc" \
Ko-cc1871e2018-08-16 02:01:57 -0700215 );
216
217#endif /* Aarch64 */
218
Paul Bakker5121ce52009-01-03 21:22:43 +0000219#if defined(__mc68020__) || defined(__mcpu32__)
220
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200221#define MULADDC_INIT \
222 asm( \
223 "movl %3, %%a2 \n\t" \
224 "movl %4, %%a3 \n\t" \
225 "movl %5, %%d3 \n\t" \
226 "movl %6, %%d2 \n\t" \
227 "moveq #0, %%d0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000228
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200229#define MULADDC_CORE \
230 "movel %%a2@+, %%d1 \n\t" \
231 "mulul %%d2, %%d4:%%d1 \n\t" \
232 "addl %%d3, %%d1 \n\t" \
233 "addxl %%d0, %%d4 \n\t" \
234 "moveq #0, %%d3 \n\t" \
235 "addl %%d1, %%a3@+ \n\t" \
236 "addxl %%d4, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000237
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200238#define MULADDC_STOP \
239 "movl %%d3, %0 \n\t" \
240 "movl %%a3, %1 \n\t" \
241 "movl %%a2, %2 \n\t" \
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100242 : "=m" (c), "=m" (d), "=m" (s) \
243 : "m" (s), "m" (d), "m" (c), "m" (b) \
244 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
245 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000246
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200247#define MULADDC_HUIT \
248 "movel %%a2@+, %%d1 \n\t" \
249 "mulul %%d2, %%d4:%%d1 \n\t" \
250 "addxl %%d3, %%d1 \n\t" \
251 "addxl %%d0, %%d4 \n\t" \
252 "addl %%d1, %%a3@+ \n\t" \
253 "movel %%a2@+, %%d1 \n\t" \
254 "mulul %%d2, %%d3:%%d1 \n\t" \
255 "addxl %%d4, %%d1 \n\t" \
256 "addxl %%d0, %%d3 \n\t" \
257 "addl %%d1, %%a3@+ \n\t" \
258 "movel %%a2@+, %%d1 \n\t" \
259 "mulul %%d2, %%d4:%%d1 \n\t" \
260 "addxl %%d3, %%d1 \n\t" \
261 "addxl %%d0, %%d4 \n\t" \
262 "addl %%d1, %%a3@+ \n\t" \
263 "movel %%a2@+, %%d1 \n\t" \
264 "mulul %%d2, %%d3:%%d1 \n\t" \
265 "addxl %%d4, %%d1 \n\t" \
266 "addxl %%d0, %%d3 \n\t" \
267 "addl %%d1, %%a3@+ \n\t" \
268 "movel %%a2@+, %%d1 \n\t" \
269 "mulul %%d2, %%d4:%%d1 \n\t" \
270 "addxl %%d3, %%d1 \n\t" \
271 "addxl %%d0, %%d4 \n\t" \
272 "addl %%d1, %%a3@+ \n\t" \
273 "movel %%a2@+, %%d1 \n\t" \
274 "mulul %%d2, %%d3:%%d1 \n\t" \
275 "addxl %%d4, %%d1 \n\t" \
276 "addxl %%d0, %%d3 \n\t" \
277 "addl %%d1, %%a3@+ \n\t" \
278 "movel %%a2@+, %%d1 \n\t" \
279 "mulul %%d2, %%d4:%%d1 \n\t" \
280 "addxl %%d3, %%d1 \n\t" \
281 "addxl %%d0, %%d4 \n\t" \
282 "addl %%d1, %%a3@+ \n\t" \
283 "movel %%a2@+, %%d1 \n\t" \
284 "mulul %%d2, %%d3:%%d1 \n\t" \
285 "addxl %%d4, %%d1 \n\t" \
286 "addxl %%d0, %%d3 \n\t" \
287 "addl %%d1, %%a3@+ \n\t" \
288 "addxl %%d0, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000289
290#endif /* MC68000 */
291
Paul Bakker5121ce52009-01-03 21:22:43 +0000292#if defined(__powerpc64__) || defined(__ppc64__)
293
294#if defined(__MACH__) && defined(__APPLE__)
295
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200296#define MULADDC_INIT \
297 asm( \
298 "ld r3, %3 \n\t" \
299 "ld r4, %4 \n\t" \
300 "ld r5, %5 \n\t" \
301 "ld r6, %6 \n\t" \
302 "addi r3, r3, -8 \n\t" \
303 "addi r4, r4, -8 \n\t" \
304 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000305
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200306#define MULADDC_CORE \
307 "ldu r7, 8(r3) \n\t" \
308 "mulld r8, r7, r6 \n\t" \
309 "mulhdu r9, r7, r6 \n\t" \
310 "adde r8, r8, r5 \n\t" \
311 "ld r7, 8(r4) \n\t" \
312 "addze r5, r9 \n\t" \
313 "addc r8, r8, r7 \n\t" \
314 "stdu r8, 8(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000315
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200316#define MULADDC_STOP \
317 "addze r5, r5 \n\t" \
318 "addi r4, r4, 8 \n\t" \
319 "addi r3, r3, 8 \n\t" \
320 "std r5, %0 \n\t" \
321 "std r4, %1 \n\t" \
322 "std r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100323 : "=m" (c), "=m" (d), "=m" (s) \
324 : "m" (s), "m" (d), "m" (c), "m" (b) \
325 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
326 );
327
Paul Bakker5121ce52009-01-03 21:22:43 +0000328
Paul Bakker9af723c2014-05-01 13:03:14 +0200329#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000330
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200331#define MULADDC_INIT \
332 asm( \
333 "ld %%r3, %3 \n\t" \
334 "ld %%r4, %4 \n\t" \
335 "ld %%r5, %5 \n\t" \
336 "ld %%r6, %6 \n\t" \
337 "addi %%r3, %%r3, -8 \n\t" \
338 "addi %%r4, %%r4, -8 \n\t" \
339 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000340
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200341#define MULADDC_CORE \
342 "ldu %%r7, 8(%%r3) \n\t" \
343 "mulld %%r8, %%r7, %%r6 \n\t" \
344 "mulhdu %%r9, %%r7, %%r6 \n\t" \
345 "adde %%r8, %%r8, %%r5 \n\t" \
346 "ld %%r7, 8(%%r4) \n\t" \
347 "addze %%r5, %%r9 \n\t" \
348 "addc %%r8, %%r8, %%r7 \n\t" \
349 "stdu %%r8, 8(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000350
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200351#define MULADDC_STOP \
352 "addze %%r5, %%r5 \n\t" \
353 "addi %%r4, %%r4, 8 \n\t" \
354 "addi %%r3, %%r3, 8 \n\t" \
355 "std %%r5, %0 \n\t" \
356 "std %%r4, %1 \n\t" \
357 "std %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100358 : "=m" (c), "=m" (d), "=m" (s) \
359 : "m" (s), "m" (d), "m" (c), "m" (b) \
360 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
361 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000362
Paul Bakker9af723c2014-05-01 13:03:14 +0200363#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000364
Barry K. Nathan35e7cb92014-05-05 23:26:13 -0700365#elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000366
367#if defined(__MACH__) && defined(__APPLE__)
368
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200369#define MULADDC_INIT \
370 asm( \
371 "lwz r3, %3 \n\t" \
372 "lwz r4, %4 \n\t" \
373 "lwz r5, %5 \n\t" \
374 "lwz r6, %6 \n\t" \
375 "addi r3, r3, -4 \n\t" \
376 "addi r4, r4, -4 \n\t" \
377 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000378
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200379#define MULADDC_CORE \
380 "lwzu r7, 4(r3) \n\t" \
381 "mullw r8, r7, r6 \n\t" \
382 "mulhwu r9, r7, r6 \n\t" \
383 "adde r8, r8, r5 \n\t" \
384 "lwz r7, 4(r4) \n\t" \
385 "addze r5, r9 \n\t" \
386 "addc r8, r8, r7 \n\t" \
387 "stwu r8, 4(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000388
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200389#define MULADDC_STOP \
390 "addze r5, r5 \n\t" \
391 "addi r4, r4, 4 \n\t" \
392 "addi r3, r3, 4 \n\t" \
393 "stw r5, %0 \n\t" \
394 "stw r4, %1 \n\t" \
395 "stw r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100396 : "=m" (c), "=m" (d), "=m" (s) \
397 : "m" (s), "m" (d), "m" (c), "m" (b) \
398 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
399 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000400
Paul Bakker9af723c2014-05-01 13:03:14 +0200401#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000402
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200403#define MULADDC_INIT \
404 asm( \
405 "lwz %%r3, %3 \n\t" \
406 "lwz %%r4, %4 \n\t" \
407 "lwz %%r5, %5 \n\t" \
408 "lwz %%r6, %6 \n\t" \
409 "addi %%r3, %%r3, -4 \n\t" \
410 "addi %%r4, %%r4, -4 \n\t" \
411 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000412
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200413#define MULADDC_CORE \
414 "lwzu %%r7, 4(%%r3) \n\t" \
415 "mullw %%r8, %%r7, %%r6 \n\t" \
416 "mulhwu %%r9, %%r7, %%r6 \n\t" \
417 "adde %%r8, %%r8, %%r5 \n\t" \
418 "lwz %%r7, 4(%%r4) \n\t" \
419 "addze %%r5, %%r9 \n\t" \
420 "addc %%r8, %%r8, %%r7 \n\t" \
421 "stwu %%r8, 4(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000422
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200423#define MULADDC_STOP \
424 "addze %%r5, %%r5 \n\t" \
425 "addi %%r4, %%r4, 4 \n\t" \
426 "addi %%r3, %%r3, 4 \n\t" \
427 "stw %%r5, %0 \n\t" \
428 "stw %%r4, %1 \n\t" \
429 "stw %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100430 : "=m" (c), "=m" (d), "=m" (s) \
431 : "m" (s), "m" (d), "m" (c), "m" (b) \
432 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
433 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000434
Paul Bakker9af723c2014-05-01 13:03:14 +0200435#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000436
437#endif /* PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000438
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200439/*
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200440 * The Sparc(64) assembly is reported to be broken.
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200441 * Disable it for now, until we're able to fix it.
442 */
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200443#if 0 && defined(__sparc__)
444#if defined(__sparc64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000445
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200446#define MULADDC_INIT \
447 asm( \
448 "ldx %3, %%o0 \n\t" \
449 "ldx %4, %%o1 \n\t" \
450 "ld %5, %%o2 \n\t" \
451 "ld %6, %%o3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000452
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200453#define MULADDC_CORE \
454 "ld [%%o0], %%o4 \n\t" \
455 "inc 4, %%o0 \n\t" \
456 "ld [%%o1], %%o5 \n\t" \
457 "umul %%o3, %%o4, %%o4 \n\t" \
458 "addcc %%o4, %%o2, %%o4 \n\t" \
459 "rd %%y, %%g1 \n\t" \
460 "addx %%g1, 0, %%g1 \n\t" \
461 "addcc %%o4, %%o5, %%o4 \n\t" \
462 "st %%o4, [%%o1] \n\t" \
463 "addx %%g1, 0, %%o2 \n\t" \
464 "inc 4, %%o1 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000465
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200466 #define MULADDC_STOP \
467 "st %%o2, %0 \n\t" \
468 "stx %%o1, %1 \n\t" \
469 "stx %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000470 : "=m" (c), "=m" (d), "=m" (s) \
471 : "m" (s), "m" (d), "m" (c), "m" (b) \
472 : "g1", "o0", "o1", "o2", "o3", "o4", \
473 "o5" \
474 );
Paul Bakker4f024b72012-10-30 07:29:57 +0000475
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200476#else /* __sparc64__ */
Paul Bakker4f024b72012-10-30 07:29:57 +0000477
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200478#define MULADDC_INIT \
479 asm( \
480 "ld %3, %%o0 \n\t" \
481 "ld %4, %%o1 \n\t" \
482 "ld %5, %%o2 \n\t" \
483 "ld %6, %%o3 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000484
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200485#define MULADDC_CORE \
486 "ld [%%o0], %%o4 \n\t" \
487 "inc 4, %%o0 \n\t" \
488 "ld [%%o1], %%o5 \n\t" \
489 "umul %%o3, %%o4, %%o4 \n\t" \
490 "addcc %%o4, %%o2, %%o4 \n\t" \
491 "rd %%y, %%g1 \n\t" \
492 "addx %%g1, 0, %%g1 \n\t" \
493 "addcc %%o4, %%o5, %%o4 \n\t" \
494 "st %%o4, [%%o1] \n\t" \
495 "addx %%g1, 0, %%o2 \n\t" \
496 "inc 4, %%o1 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000497
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200498#define MULADDC_STOP \
499 "st %%o2, %0 \n\t" \
500 "st %%o1, %1 \n\t" \
501 "st %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000502 : "=m" (c), "=m" (d), "=m" (s) \
503 : "m" (s), "m" (d), "m" (c), "m" (b) \
504 : "g1", "o0", "o1", "o2", "o3", "o4", \
505 "o5" \
506 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000507
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200508#endif /* __sparc64__ */
509#endif /* __sparc__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000510
511#if defined(__microblaze__) || defined(microblaze)
512
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200513#define MULADDC_INIT \
514 asm( \
515 "lwi r3, %3 \n\t" \
516 "lwi r4, %4 \n\t" \
517 "lwi r5, %5 \n\t" \
518 "lwi r6, %6 \n\t" \
519 "andi r7, r6, 0xffff \n\t" \
520 "bsrli r6, r6, 16 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000521
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200522#define MULADDC_CORE \
523 "lhui r8, r3, 0 \n\t" \
524 "addi r3, r3, 2 \n\t" \
525 "lhui r9, r3, 0 \n\t" \
526 "addi r3, r3, 2 \n\t" \
527 "mul r10, r9, r6 \n\t" \
528 "mul r11, r8, r7 \n\t" \
529 "mul r12, r9, r7 \n\t" \
530 "mul r13, r8, r6 \n\t" \
531 "bsrli r8, r10, 16 \n\t" \
532 "bsrli r9, r11, 16 \n\t" \
533 "add r13, r13, r8 \n\t" \
534 "add r13, r13, r9 \n\t" \
535 "bslli r10, r10, 16 \n\t" \
536 "bslli r11, r11, 16 \n\t" \
537 "add r12, r12, r10 \n\t" \
538 "addc r13, r13, r0 \n\t" \
539 "add r12, r12, r11 \n\t" \
540 "addc r13, r13, r0 \n\t" \
541 "lwi r10, r4, 0 \n\t" \
542 "add r12, r12, r10 \n\t" \
543 "addc r13, r13, r0 \n\t" \
544 "add r12, r12, r5 \n\t" \
545 "addc r5, r13, r0 \n\t" \
546 "swi r12, r4, 0 \n\t" \
547 "addi r4, r4, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000548
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200549#define MULADDC_STOP \
550 "swi r5, %0 \n\t" \
551 "swi r4, %1 \n\t" \
552 "swi r3, %2 \n\t" \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100553 : "=m" (c), "=m" (d), "=m" (s) \
554 : "m" (s), "m" (d), "m" (c), "m" (b) \
Zach van Rijne7d3f8e2018-05-21 10:52:34 -0400555 : "r3", "r4", "r5", "r6", "r7", "r8", \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100556 "r9", "r10", "r11", "r12", "r13" \
557 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000558
559#endif /* MicroBlaze */
560
561#if defined(__tricore__)
562
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200563#define MULADDC_INIT \
564 asm( \
565 "ld.a %%a2, %3 \n\t" \
566 "ld.a %%a3, %4 \n\t" \
567 "ld.w %%d4, %5 \n\t" \
568 "ld.w %%d1, %6 \n\t" \
569 "xor %%d5, %%d5 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000570
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200571#define MULADDC_CORE \
572 "ld.w %%d0, [%%a2+] \n\t" \
573 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
574 "ld.w %%d0, [%%a3] \n\t" \
575 "addx %%d2, %%d2, %%d0 \n\t" \
576 "addc %%d3, %%d3, 0 \n\t" \
577 "mov %%d4, %%d3 \n\t" \
578 "st.w [%%a3+], %%d2 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000579
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200580#define MULADDC_STOP \
581 "st.w %0, %%d4 \n\t" \
582 "st.a %1, %%a3 \n\t" \
583 "st.a %2, %%a2 \n\t" \
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100584 : "=m" (c), "=m" (d), "=m" (s) \
585 : "m" (s), "m" (d), "m" (c), "m" (b) \
586 : "d0", "d1", "e2", "d4", "a2", "a3" \
587 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000588
589#endif /* TriCore */
590
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100591/*
Simon Butcher3ad2efd2018-05-02 14:49:38 +0100592 * Note, gcc -O0 by default uses r7 for the frame pointer, so it complains about
593 * our use of r7 below, unless -fomit-frame-pointer is passed.
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100594 *
595 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
596 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
597 * clang and armcc5 under the same conditions).
598 *
599 * So, only use the optimized assembly below for optimized build, which avoids
600 * the build error and is pretty reasonable anyway.
601 */
602#if defined(__GNUC__) && !defined(__OPTIMIZE__)
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100603#define MULADDC_CANNOT_USE_R7
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100604#endif
605
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100606#if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
Paul Bakker5121ce52009-01-03 21:22:43 +0000607
Paul Bakkerfc4f46f2013-06-24 19:23:56 +0200608#if defined(__thumb__) && !defined(__thumb2__)
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000609
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200610#define MULADDC_INIT \
611 asm( \
612 "ldr r0, %3 \n\t" \
613 "ldr r1, %4 \n\t" \
614 "ldr r2, %5 \n\t" \
615 "ldr r3, %6 \n\t" \
616 "lsr r7, r3, #16 \n\t" \
617 "mov r9, r7 \n\t" \
618 "lsl r7, r3, #16 \n\t" \
619 "lsr r7, r7, #16 \n\t" \
620 "mov r8, r7 \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000621
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200622#define MULADDC_CORE \
623 "ldmia r0!, {r6} \n\t" \
624 "lsr r7, r6, #16 \n\t" \
625 "lsl r6, r6, #16 \n\t" \
626 "lsr r6, r6, #16 \n\t" \
627 "mov r4, r8 \n\t" \
628 "mul r4, r6 \n\t" \
629 "mov r3, r9 \n\t" \
630 "mul r6, r3 \n\t" \
631 "mov r5, r9 \n\t" \
632 "mul r5, r7 \n\t" \
633 "mov r3, r8 \n\t" \
634 "mul r7, r3 \n\t" \
635 "lsr r3, r6, #16 \n\t" \
636 "add r5, r5, r3 \n\t" \
637 "lsr r3, r7, #16 \n\t" \
638 "add r5, r5, r3 \n\t" \
639 "add r4, r4, r2 \n\t" \
640 "mov r2, #0 \n\t" \
641 "adc r5, r2 \n\t" \
642 "lsl r3, r6, #16 \n\t" \
643 "add r4, r4, r3 \n\t" \
644 "adc r5, r2 \n\t" \
645 "lsl r3, r7, #16 \n\t" \
646 "add r4, r4, r3 \n\t" \
647 "adc r5, r2 \n\t" \
648 "ldr r3, [r1] \n\t" \
649 "add r4, r4, r3 \n\t" \
650 "adc r2, r5 \n\t" \
651 "stmia r1!, {r4} \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000652
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200653#define MULADDC_STOP \
654 "str r2, %0 \n\t" \
655 "str r1, %1 \n\t" \
656 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100657 : "=m" (c), "=m" (d), "=m" (s) \
658 : "m" (s), "m" (d), "m" (c), "m" (b) \
659 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200660 "r6", "r7", "r8", "r9", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100661 );
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000662
Aurelien Jarno5daa34f2018-11-03 00:46:06 +0100663#elif (__ARM_ARCH >= 6) && \
664 defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)
Aurelien Jarno16b1bd82018-05-21 22:01:21 +0200665
666#define MULADDC_INIT \
667 asm(
668
669#define MULADDC_CORE \
670 "ldr r0, [%0], #4 \n\t" \
671 "ldr r1, [%1] \n\t" \
672 "umaal r1, %2, %3, r0 \n\t" \
673 "str r1, [%1], #4 \n\t"
674
675#define MULADDC_STOP \
676 : "=r" (s), "=r" (d), "=r" (c) \
677 : "r" (b), "0" (s), "1" (d), "2" (c) \
678 : "r0", "r1", "memory" \
679 );
680
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000681#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000682
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200683#define MULADDC_INIT \
684 asm( \
685 "ldr r0, %3 \n\t" \
686 "ldr r1, %4 \n\t" \
687 "ldr r2, %5 \n\t" \
688 "ldr r3, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000689
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200690#define MULADDC_CORE \
691 "ldr r4, [r0], #4 \n\t" \
692 "mov r5, #0 \n\t" \
693 "ldr r6, [r1] \n\t" \
694 "umlal r2, r5, r3, r4 \n\t" \
695 "adds r7, r6, r2 \n\t" \
696 "adc r2, r5, #0 \n\t" \
697 "str r7, [r1], #4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000698
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200699#define MULADDC_STOP \
700 "str r2, %0 \n\t" \
701 "str r1, %1 \n\t" \
702 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100703 : "=m" (c), "=m" (d), "=m" (s) \
704 : "m" (s), "m" (d), "m" (c), "m" (b) \
705 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200706 "r6", "r7", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100707 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000708
Paul Bakkera2713a32011-11-18 12:47:23 +0000709#endif /* Thumb */
710
Paul Bakker5121ce52009-01-03 21:22:43 +0000711#endif /* ARMv3 */
712
713#if defined(__alpha__)
714
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200715#define MULADDC_INIT \
716 asm( \
717 "ldq $1, %3 \n\t" \
718 "ldq $2, %4 \n\t" \
719 "ldq $3, %5 \n\t" \
720 "ldq $4, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000721
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200722#define MULADDC_CORE \
723 "ldq $6, 0($1) \n\t" \
724 "addq $1, 8, $1 \n\t" \
725 "mulq $6, $4, $7 \n\t" \
726 "umulh $6, $4, $6 \n\t" \
727 "addq $7, $3, $7 \n\t" \
728 "cmpult $7, $3, $3 \n\t" \
729 "ldq $5, 0($2) \n\t" \
730 "addq $7, $5, $7 \n\t" \
731 "cmpult $7, $5, $5 \n\t" \
732 "stq $7, 0($2) \n\t" \
733 "addq $2, 8, $2 \n\t" \
734 "addq $6, $3, $3 \n\t" \
735 "addq $5, $3, $3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000736
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200737#define MULADDC_STOP \
738 "stq $3, %0 \n\t" \
739 "stq $2, %1 \n\t" \
740 "stq $1, %2 \n\t" \
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100741 : "=m" (c), "=m" (d), "=m" (s) \
742 : "m" (s), "m" (d), "m" (c), "m" (b) \
743 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
744 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000745#endif /* Alpha */
746
James Cowgilld1e7e8b2014-12-16 15:24:06 +0000747#if defined(__mips__) && !defined(__mips64)
Paul Bakker5121ce52009-01-03 21:22:43 +0000748
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200749#define MULADDC_INIT \
750 asm( \
751 "lw $10, %3 \n\t" \
752 "lw $11, %4 \n\t" \
753 "lw $12, %5 \n\t" \
754 "lw $13, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000755
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200756#define MULADDC_CORE \
757 "lw $14, 0($10) \n\t" \
758 "multu $13, $14 \n\t" \
759 "addi $10, $10, 4 \n\t" \
760 "mflo $14 \n\t" \
761 "mfhi $9 \n\t" \
762 "addu $14, $12, $14 \n\t" \
763 "lw $15, 0($11) \n\t" \
764 "sltu $12, $14, $12 \n\t" \
765 "addu $15, $14, $15 \n\t" \
766 "sltu $14, $15, $14 \n\t" \
767 "addu $12, $12, $9 \n\t" \
768 "sw $15, 0($11) \n\t" \
769 "addu $12, $12, $14 \n\t" \
770 "addi $11, $11, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000771
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200772#define MULADDC_STOP \
773 "sw $12, %0 \n\t" \
774 "sw $11, %1 \n\t" \
775 "sw $10, %2 \n\t" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100776 : "=m" (c), "=m" (d), "=m" (s) \
777 : "m" (s), "m" (d), "m" (c), "m" (b) \
Jeffrey Martind25fd8d2019-01-14 18:01:40 -0600778 : "$9", "$10", "$11", "$12", "$13", "$14", "$15", "lo", "hi" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100779 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000780
781#endif /* MIPS */
782#endif /* GNUC */
783
784#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
785
786#define MULADDC_INIT \
787 __asm mov esi, s \
788 __asm mov edi, d \
789 __asm mov ecx, c \
790 __asm mov ebx, b
791
792#define MULADDC_CORE \
793 __asm lodsd \
794 __asm mul ebx \
795 __asm add eax, ecx \
796 __asm adc edx, 0 \
797 __asm add eax, [edi] \
798 __asm adc edx, 0 \
799 __asm mov ecx, edx \
800 __asm stosd
801
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200802#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000803
804#define EMIT __asm _emit
805
806#define MULADDC_HUIT \
807 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
808 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
809 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
810 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
811 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
812 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
813 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
814 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
815 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
816 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
817 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
818 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
819 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
820 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
821 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
822 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
823 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
824 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
825 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
826 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
827 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
828 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
829 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
830 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
831 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
832 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
833 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
834 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
835 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
836 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
837 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
838 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
839 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
840 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
841 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
842 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
843 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
844 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
845 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
846 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
847 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
848 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
849 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
850 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
851 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
852 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
853 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
854 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
855 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
856 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
857 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
858 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
859 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
860 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
861 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
862 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
863 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
864 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
865 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
866 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
867 EMIT 0x0F EMIT 0x7E EMIT 0xC9
868
869#define MULADDC_STOP \
870 EMIT 0x0F EMIT 0x77 \
871 __asm mov c, ecx \
872 __asm mov d, edi \
873 __asm mov s, esi \
874
875#else
876
877#define MULADDC_STOP \
878 __asm mov c, ecx \
879 __asm mov d, edi \
880 __asm mov s, esi \
881
882#endif /* SSE2 */
883#endif /* MSVC */
884
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200885#endif /* MBEDTLS_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000886
887#if !defined(MULADDC_CORE)
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200888#if defined(MBEDTLS_HAVE_UDBL)
Paul Bakker5121ce52009-01-03 21:22:43 +0000889
890#define MULADDC_INIT \
891{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200892 mbedtls_t_udbl r; \
893 mbedtls_mpi_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000894
895#define MULADDC_CORE \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200896 r = *(s++) * (mbedtls_t_udbl) b; \
897 r0 = (mbedtls_mpi_uint) r; \
898 r1 = (mbedtls_mpi_uint)( r >> biL ); \
Paul Bakker5121ce52009-01-03 21:22:43 +0000899 r0 += c; r1 += (r0 < c); \
900 r0 += *d; r1 += (r0 < *d); \
901 c = r1; *(d++) = r0;
902
903#define MULADDC_STOP \
904}
905
906#else
907#define MULADDC_INIT \
908{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200909 mbedtls_mpi_uint s0, s1, b0, b1; \
910 mbedtls_mpi_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000911 b0 = ( b << biH ) >> biH; \
912 b1 = ( b >> biH );
913
914#define MULADDC_CORE \
915 s0 = ( *s << biH ) >> biH; \
916 s1 = ( *s >> biH ); s++; \
917 rx = s0 * b1; r0 = s0 * b0; \
918 ry = s1 * b0; r1 = s1 * b1; \
919 r1 += ( rx >> biH ); \
920 r1 += ( ry >> biH ); \
921 rx <<= biH; ry <<= biH; \
922 r0 += rx; r1 += (r0 < rx); \
923 r0 += ry; r1 += (r0 < ry); \
924 r0 += c; r1 += (r0 < c); \
925 r0 += *d; r1 += (r0 < *d); \
926 c = r1; *(d++) = r0;
927
928#define MULADDC_STOP \
929}
930
931#endif /* C (generic) */
932#endif /* C (longlong) */
933
934#endif /* bn_mul.h */