blob: 30d445e63b8fb3282b4d5a52534c23083ae79a23 [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
Ron Eldor8b0cf2e2018-02-14 16:02:41 +020039#if !defined(MBEDTLS_CONFIG_FILE)
Jaeden Ameroc49fbbf2019-07-04 20:01:14 +010040#include "mbedtls/config.h"
Ron Eldor8b0cf2e2018-02-14 16:02:41 +020041#else
42#include MBEDTLS_CONFIG_FILE
43#endif
44
Jaeden Ameroc49fbbf2019-07-04 20:01:14 +010045#include "mbedtls/bignum.h"
Paul Bakker5121ce52009-01-03 21:22:43 +000046
Janos Follathbc96a792021-06-24 14:48:38 +010047
48/*
49 * Conversion macros for embedded constants:
50 * build lists of mbedtls_mpi_uint's from lists of unsigned char's grouped by 8, 4 or 2
51 */
52#if defined(MBEDTLS_HAVE_INT32)
53
Janos Follathbc589022021-06-25 12:43:26 +010054#define MBEDTLS_BYTES_TO_T_UINT_4( a, b, c, d ) \
Janos Follathbc96a792021-06-24 14:48:38 +010055 ( (mbedtls_mpi_uint) (a) << 0 ) | \
56 ( (mbedtls_mpi_uint) (b) << 8 ) | \
57 ( (mbedtls_mpi_uint) (c) << 16 ) | \
58 ( (mbedtls_mpi_uint) (d) << 24 )
59
Janos Follathbc589022021-06-25 12:43:26 +010060#define MBEDTLS_BYTES_TO_T_UINT_2( a, b ) \
61 MBEDTLS_BYTES_TO_T_UINT_4( a, b, 0, 0 )
Janos Follathbc96a792021-06-24 14:48:38 +010062
Janos Follathbc589022021-06-25 12:43:26 +010063#define MBEDTLS_BYTES_TO_T_UINT_8( a, b, c, d, e, f, g, h ) \
64 MBEDTLS_BYTES_TO_T_UINT_4( a, b, c, d ), \
65 MBEDTLS_BYTES_TO_T_UINT_4( e, f, g, h )
Janos Follathbc96a792021-06-24 14:48:38 +010066
67#else /* 64-bits */
68
Janos Follathbc589022021-06-25 12:43:26 +010069#define MBEDTLS_BYTES_TO_T_UINT_8( a, b, c, d, e, f, g, h ) \
Janos Follathbc96a792021-06-24 14:48:38 +010070 ( (mbedtls_mpi_uint) (a) << 0 ) | \
71 ( (mbedtls_mpi_uint) (b) << 8 ) | \
72 ( (mbedtls_mpi_uint) (c) << 16 ) | \
73 ( (mbedtls_mpi_uint) (d) << 24 ) | \
74 ( (mbedtls_mpi_uint) (e) << 32 ) | \
75 ( (mbedtls_mpi_uint) (f) << 40 ) | \
76 ( (mbedtls_mpi_uint) (g) << 48 ) | \
77 ( (mbedtls_mpi_uint) (h) << 56 )
78
Janos Follathbc589022021-06-25 12:43:26 +010079#define MBEDTLS_BYTES_TO_T_UINT_4( a, b, c, d ) \
80 MBEDTLS_BYTES_TO_T_UINT_8( a, b, c, d, 0, 0, 0, 0 )
Janos Follathbc96a792021-06-24 14:48:38 +010081
Janos Follathbc589022021-06-25 12:43:26 +010082#define MBEDTLS_BYTES_TO_T_UINT_2( a, b ) \
83 MBEDTLS_BYTES_TO_T_UINT_8( a, b, 0, 0, 0, 0, 0, 0 )
Janos Follathbc96a792021-06-24 14:48:38 +010084
85#endif /* bits in mbedtls_mpi_uint */
86
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +020087#if defined(MBEDTLS_HAVE_ASM)
Paul Bakker5121ce52009-01-03 21:22:43 +000088
David Horstmannb6bf5f52023-01-03 11:07:09 +000089/* *INDENT-OFF* */
Manuel Pégourié-Gonnardba194322015-05-29 09:47:57 +020090#ifndef asm
91#define asm __asm
92#endif
David Horstmannb6bf5f52023-01-03 11:07:09 +000093/* *INDENT-ON* */
Manuel Pégourié-Gonnardba194322015-05-29 09:47:57 +020094
Manuel Pégourié-Gonnard854dab92015-08-10 12:08:34 +020095/* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */
96#if defined(__GNUC__) && \
97 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
Simon Butcher4b9a3ad2018-07-10 20:18:29 +010098
99/*
Peter Korsgaard9d9d45c2018-08-27 22:50:57 +0200100 * GCC < 5.0 treated the x86 ebx (which is used for the GOT) as a
101 * fixed reserved register when building as PIC, leading to errors
102 * like: bn_mul.h:46:13: error: PIC register clobbered by 'ebx' in 'asm'
103 *
104 * This is fixed by an improved register allocator in GCC 5+. From the
105 * release notes:
106 * Register allocation improvements: Reuse of the PIC hard register,
107 * instead of using a fixed register, was implemented on x86/x86-64
108 * targets. This improves generated PIC code performance as more hard
109 * registers can be used.
110 */
111#if defined(__GNUC__) && __GNUC__ < 5 && defined(__PIC__)
112#define MULADDC_CANNOT_USE_EBX
113#endif
114
115/*
Simon Butcher4b9a3ad2018-07-10 20:18:29 +0100116 * Disable use of the i386 assembly code below if option -O0, to disable all
117 * compiler optimisations, is passed, detected with __OPTIMIZE__
118 * This is done as the number of registers used in the assembly code doesn't
119 * work with the -O0 option.
120 */
Peter Korsgaard9d9d45c2018-08-27 22:50:57 +0200121#if defined(__i386__) && defined(__OPTIMIZE__) && !defined(MULADDC_CANNOT_USE_EBX)
Paul Bakker5121ce52009-01-03 21:22:43 +0000122
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200123#define MULADDC_INIT \
124 asm( \
125 "movl %%ebx, %0 \n\t" \
126 "movl %5, %%esi \n\t" \
127 "movl %6, %%edi \n\t" \
128 "movl %7, %%ecx \n\t" \
129 "movl %8, %%ebx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000130
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200131#define MULADDC_CORE \
132 "lodsl \n\t" \
133 "mull %%ebx \n\t" \
134 "addl %%ecx, %%eax \n\t" \
135 "adcl $0, %%edx \n\t" \
136 "addl (%%edi), %%eax \n\t" \
137 "adcl $0, %%edx \n\t" \
138 "movl %%edx, %%ecx \n\t" \
139 "stosl \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000140
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200141#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000142
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200143#define MULADDC_HUIT \
144 "movd %%ecx, %%mm1 \n\t" \
145 "movd %%ebx, %%mm0 \n\t" \
146 "movd (%%edi), %%mm3 \n\t" \
147 "paddq %%mm3, %%mm1 \n\t" \
148 "movd (%%esi), %%mm2 \n\t" \
149 "pmuludq %%mm0, %%mm2 \n\t" \
150 "movd 4(%%esi), %%mm4 \n\t" \
151 "pmuludq %%mm0, %%mm4 \n\t" \
152 "movd 8(%%esi), %%mm6 \n\t" \
153 "pmuludq %%mm0, %%mm6 \n\t" \
154 "movd 12(%%esi), %%mm7 \n\t" \
155 "pmuludq %%mm0, %%mm7 \n\t" \
156 "paddq %%mm2, %%mm1 \n\t" \
157 "movd 4(%%edi), %%mm3 \n\t" \
158 "paddq %%mm4, %%mm3 \n\t" \
159 "movd 8(%%edi), %%mm5 \n\t" \
160 "paddq %%mm6, %%mm5 \n\t" \
161 "movd 12(%%edi), %%mm4 \n\t" \
162 "paddq %%mm4, %%mm7 \n\t" \
163 "movd %%mm1, (%%edi) \n\t" \
164 "movd 16(%%esi), %%mm2 \n\t" \
165 "pmuludq %%mm0, %%mm2 \n\t" \
166 "psrlq $32, %%mm1 \n\t" \
167 "movd 20(%%esi), %%mm4 \n\t" \
168 "pmuludq %%mm0, %%mm4 \n\t" \
169 "paddq %%mm3, %%mm1 \n\t" \
170 "movd 24(%%esi), %%mm6 \n\t" \
171 "pmuludq %%mm0, %%mm6 \n\t" \
172 "movd %%mm1, 4(%%edi) \n\t" \
173 "psrlq $32, %%mm1 \n\t" \
174 "movd 28(%%esi), %%mm3 \n\t" \
175 "pmuludq %%mm0, %%mm3 \n\t" \
176 "paddq %%mm5, %%mm1 \n\t" \
177 "movd 16(%%edi), %%mm5 \n\t" \
178 "paddq %%mm5, %%mm2 \n\t" \
179 "movd %%mm1, 8(%%edi) \n\t" \
180 "psrlq $32, %%mm1 \n\t" \
181 "paddq %%mm7, %%mm1 \n\t" \
182 "movd 20(%%edi), %%mm5 \n\t" \
183 "paddq %%mm5, %%mm4 \n\t" \
184 "movd %%mm1, 12(%%edi) \n\t" \
185 "psrlq $32, %%mm1 \n\t" \
186 "paddq %%mm2, %%mm1 \n\t" \
187 "movd 24(%%edi), %%mm5 \n\t" \
188 "paddq %%mm5, %%mm6 \n\t" \
189 "movd %%mm1, 16(%%edi) \n\t" \
190 "psrlq $32, %%mm1 \n\t" \
191 "paddq %%mm4, %%mm1 \n\t" \
192 "movd 28(%%edi), %%mm5 \n\t" \
193 "paddq %%mm5, %%mm3 \n\t" \
194 "movd %%mm1, 20(%%edi) \n\t" \
195 "psrlq $32, %%mm1 \n\t" \
196 "paddq %%mm6, %%mm1 \n\t" \
197 "movd %%mm1, 24(%%edi) \n\t" \
198 "psrlq $32, %%mm1 \n\t" \
199 "paddq %%mm3, %%mm1 \n\t" \
200 "movd %%mm1, 28(%%edi) \n\t" \
201 "addl $32, %%edi \n\t" \
202 "addl $32, %%esi \n\t" \
203 "psrlq $32, %%mm1 \n\t" \
204 "movd %%mm1, %%ecx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000205
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200206#define MULADDC_STOP \
207 "emms \n\t" \
208 "movl %4, %%ebx \n\t" \
209 "movl %%ecx, %1 \n\t" \
210 "movl %%edi, %2 \n\t" \
211 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000212 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
213 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
Simon Butcher53571642018-06-24 12:58:31 +0100214 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000215 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000216
217#else
218
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200219#define MULADDC_STOP \
220 "movl %4, %%ebx \n\t" \
221 "movl %%ecx, %1 \n\t" \
222 "movl %%edi, %2 \n\t" \
223 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000224 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
225 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
Simon Butcher53571642018-06-24 12:58:31 +0100226 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000227 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000228#endif /* SSE2 */
229#endif /* i386 */
230
231#if defined(__amd64__) || defined (__x86_64__)
232
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200233#define MULADDC_INIT \
234 asm( \
Simon Butchera86de142018-09-30 12:09:47 +0100235 "xorq %%r8, %%r8\n"
Paul Bakker5121ce52009-01-03 21:22:43 +0000236
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200237#define MULADDC_CORE \
Simon Butchera86de142018-09-30 12:09:47 +0100238 "movq (%%rsi), %%rax\n" \
239 "mulq %%rbx\n" \
240 "addq $8, %%rsi\n" \
241 "addq %%rcx, %%rax\n" \
242 "movq %%r8, %%rcx\n" \
243 "adcq $0, %%rdx\n" \
244 "nop \n" \
245 "addq %%rax, (%%rdi)\n" \
246 "adcq %%rdx, %%rcx\n" \
247 "addq $8, %%rdi\n"
Paul Bakker5121ce52009-01-03 21:22:43 +0000248
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200249#define MULADDC_STOP \
Gilles Peskinec6d977e2021-09-14 00:13:05 +0200250 : "+c" (c), "+D" (d), "+S" (s), "+m" (*(uint64_t (*)[16]) d) \
251 : "b" (b), "m" (*(const uint64_t (*)[16]) s) \
252 : "rax", "rdx", "r8" \
Manuel Pégourié-Gonnarddef018d2014-01-07 17:50:46 +0100253 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000254
255#endif /* AMD64 */
256
Ko-cc1871e2018-08-16 02:01:57 -0700257#if defined(__aarch64__)
258
259#define MULADDC_INIT \
260 asm(
261
262#define MULADDC_CORE \
Ko-cb260bb2018-08-20 13:59:53 +0100263 "ldr x4, [%2], #8 \n\t" \
264 "ldr x5, [%1] \n\t" \
David Horstmanna23be222021-09-22 18:15:51 +0100265 "mul x6, x4, %4 \n\t" \
266 "umulh x7, x4, %4 \n\t" \
Ko-cc1871e2018-08-16 02:01:57 -0700267 "adds x5, x5, x6 \n\t" \
268 "adc x7, x7, xzr \n\t" \
Ko-cb260bb2018-08-20 13:59:53 +0100269 "adds x5, x5, %0 \n\t" \
Ko-cc1871e2018-08-16 02:01:57 -0700270 "adc %0, x7, xzr \n\t" \
271 "str x5, [%1], #8 \n\t"
272
David Horstmanna23be222021-09-22 18:15:51 +0100273#define MULADDC_STOP \
274 : "+r" (c), "+r" (d), "+r" (s), "+m" (*(uint64_t (*)[16]) d) \
275 : "r" (b), "m" (*(const uint64_t (*)[16]) s) \
276 : "x4", "x5", "x6", "x7", "cc" \
Ko-cc1871e2018-08-16 02:01:57 -0700277 );
278
279#endif /* Aarch64 */
280
Paul Bakker5121ce52009-01-03 21:22:43 +0000281#if defined(__mc68020__) || defined(__mcpu32__)
282
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200283#define MULADDC_INIT \
284 asm( \
285 "movl %3, %%a2 \n\t" \
286 "movl %4, %%a3 \n\t" \
287 "movl %5, %%d3 \n\t" \
288 "movl %6, %%d2 \n\t" \
289 "moveq #0, %%d0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000290
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200291#define MULADDC_CORE \
292 "movel %%a2@+, %%d1 \n\t" \
293 "mulul %%d2, %%d4:%%d1 \n\t" \
294 "addl %%d3, %%d1 \n\t" \
295 "addxl %%d0, %%d4 \n\t" \
296 "moveq #0, %%d3 \n\t" \
297 "addl %%d1, %%a3@+ \n\t" \
298 "addxl %%d4, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000299
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200300#define MULADDC_STOP \
301 "movl %%d3, %0 \n\t" \
302 "movl %%a3, %1 \n\t" \
303 "movl %%a2, %2 \n\t" \
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100304 : "=m" (c), "=m" (d), "=m" (s) \
305 : "m" (s), "m" (d), "m" (c), "m" (b) \
306 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
307 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000308
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200309#define MULADDC_HUIT \
310 "movel %%a2@+, %%d1 \n\t" \
311 "mulul %%d2, %%d4:%%d1 \n\t" \
312 "addxl %%d3, %%d1 \n\t" \
313 "addxl %%d0, %%d4 \n\t" \
314 "addl %%d1, %%a3@+ \n\t" \
315 "movel %%a2@+, %%d1 \n\t" \
316 "mulul %%d2, %%d3:%%d1 \n\t" \
317 "addxl %%d4, %%d1 \n\t" \
318 "addxl %%d0, %%d3 \n\t" \
319 "addl %%d1, %%a3@+ \n\t" \
320 "movel %%a2@+, %%d1 \n\t" \
321 "mulul %%d2, %%d4:%%d1 \n\t" \
322 "addxl %%d3, %%d1 \n\t" \
323 "addxl %%d0, %%d4 \n\t" \
324 "addl %%d1, %%a3@+ \n\t" \
325 "movel %%a2@+, %%d1 \n\t" \
326 "mulul %%d2, %%d3:%%d1 \n\t" \
327 "addxl %%d4, %%d1 \n\t" \
328 "addxl %%d0, %%d3 \n\t" \
329 "addl %%d1, %%a3@+ \n\t" \
330 "movel %%a2@+, %%d1 \n\t" \
331 "mulul %%d2, %%d4:%%d1 \n\t" \
332 "addxl %%d3, %%d1 \n\t" \
333 "addxl %%d0, %%d4 \n\t" \
334 "addl %%d1, %%a3@+ \n\t" \
335 "movel %%a2@+, %%d1 \n\t" \
336 "mulul %%d2, %%d3:%%d1 \n\t" \
337 "addxl %%d4, %%d1 \n\t" \
338 "addxl %%d0, %%d3 \n\t" \
339 "addl %%d1, %%a3@+ \n\t" \
340 "movel %%a2@+, %%d1 \n\t" \
341 "mulul %%d2, %%d4:%%d1 \n\t" \
342 "addxl %%d3, %%d1 \n\t" \
343 "addxl %%d0, %%d4 \n\t" \
344 "addl %%d1, %%a3@+ \n\t" \
345 "movel %%a2@+, %%d1 \n\t" \
346 "mulul %%d2, %%d3:%%d1 \n\t" \
347 "addxl %%d4, %%d1 \n\t" \
348 "addxl %%d0, %%d3 \n\t" \
349 "addl %%d1, %%a3@+ \n\t" \
350 "addxl %%d0, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000351
352#endif /* MC68000 */
353
Paul Bakker5121ce52009-01-03 21:22:43 +0000354#if defined(__powerpc64__) || defined(__ppc64__)
355
356#if defined(__MACH__) && defined(__APPLE__)
357
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200358#define MULADDC_INIT \
359 asm( \
360 "ld r3, %3 \n\t" \
361 "ld r4, %4 \n\t" \
362 "ld r5, %5 \n\t" \
363 "ld r6, %6 \n\t" \
364 "addi r3, r3, -8 \n\t" \
365 "addi r4, r4, -8 \n\t" \
366 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000367
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200368#define MULADDC_CORE \
369 "ldu r7, 8(r3) \n\t" \
370 "mulld r8, r7, r6 \n\t" \
371 "mulhdu r9, r7, r6 \n\t" \
372 "adde r8, r8, r5 \n\t" \
373 "ld r7, 8(r4) \n\t" \
374 "addze r5, r9 \n\t" \
375 "addc r8, r8, r7 \n\t" \
376 "stdu r8, 8(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000377
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200378#define MULADDC_STOP \
379 "addze r5, r5 \n\t" \
380 "addi r4, r4, 8 \n\t" \
381 "addi r3, r3, 8 \n\t" \
382 "std r5, %0 \n\t" \
383 "std r4, %1 \n\t" \
384 "std r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100385 : "=m" (c), "=m" (d), "=m" (s) \
386 : "m" (s), "m" (d), "m" (c), "m" (b) \
387 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
388 );
389
Paul Bakker5121ce52009-01-03 21:22:43 +0000390
Paul Bakker9af723c2014-05-01 13:03:14 +0200391#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000392
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200393#define MULADDC_INIT \
394 asm( \
395 "ld %%r3, %3 \n\t" \
396 "ld %%r4, %4 \n\t" \
397 "ld %%r5, %5 \n\t" \
398 "ld %%r6, %6 \n\t" \
399 "addi %%r3, %%r3, -8 \n\t" \
400 "addi %%r4, %%r4, -8 \n\t" \
401 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000402
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200403#define MULADDC_CORE \
404 "ldu %%r7, 8(%%r3) \n\t" \
405 "mulld %%r8, %%r7, %%r6 \n\t" \
406 "mulhdu %%r9, %%r7, %%r6 \n\t" \
407 "adde %%r8, %%r8, %%r5 \n\t" \
408 "ld %%r7, 8(%%r4) \n\t" \
409 "addze %%r5, %%r9 \n\t" \
410 "addc %%r8, %%r8, %%r7 \n\t" \
411 "stdu %%r8, 8(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000412
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200413#define MULADDC_STOP \
414 "addze %%r5, %%r5 \n\t" \
415 "addi %%r4, %%r4, 8 \n\t" \
416 "addi %%r3, %%r3, 8 \n\t" \
417 "std %%r5, %0 \n\t" \
418 "std %%r4, %1 \n\t" \
419 "std %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100420 : "=m" (c), "=m" (d), "=m" (s) \
421 : "m" (s), "m" (d), "m" (c), "m" (b) \
422 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
423 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000424
Paul Bakker9af723c2014-05-01 13:03:14 +0200425#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000426
Barry K. Nathan35e7cb92014-05-05 23:26:13 -0700427#elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000428
429#if defined(__MACH__) && defined(__APPLE__)
430
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200431#define MULADDC_INIT \
432 asm( \
433 "lwz r3, %3 \n\t" \
434 "lwz r4, %4 \n\t" \
435 "lwz r5, %5 \n\t" \
436 "lwz r6, %6 \n\t" \
437 "addi r3, r3, -4 \n\t" \
438 "addi r4, r4, -4 \n\t" \
439 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000440
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200441#define MULADDC_CORE \
442 "lwzu r7, 4(r3) \n\t" \
443 "mullw r8, r7, r6 \n\t" \
444 "mulhwu r9, r7, r6 \n\t" \
445 "adde r8, r8, r5 \n\t" \
446 "lwz r7, 4(r4) \n\t" \
447 "addze r5, r9 \n\t" \
448 "addc r8, r8, r7 \n\t" \
449 "stwu r8, 4(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000450
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200451#define MULADDC_STOP \
452 "addze r5, r5 \n\t" \
453 "addi r4, r4, 4 \n\t" \
454 "addi r3, r3, 4 \n\t" \
455 "stw r5, %0 \n\t" \
456 "stw r4, %1 \n\t" \
457 "stw r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100458 : "=m" (c), "=m" (d), "=m" (s) \
459 : "m" (s), "m" (d), "m" (c), "m" (b) \
460 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
461 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000462
Paul Bakker9af723c2014-05-01 13:03:14 +0200463#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000464
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200465#define MULADDC_INIT \
466 asm( \
467 "lwz %%r3, %3 \n\t" \
468 "lwz %%r4, %4 \n\t" \
469 "lwz %%r5, %5 \n\t" \
470 "lwz %%r6, %6 \n\t" \
471 "addi %%r3, %%r3, -4 \n\t" \
472 "addi %%r4, %%r4, -4 \n\t" \
473 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000474
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200475#define MULADDC_CORE \
476 "lwzu %%r7, 4(%%r3) \n\t" \
477 "mullw %%r8, %%r7, %%r6 \n\t" \
478 "mulhwu %%r9, %%r7, %%r6 \n\t" \
479 "adde %%r8, %%r8, %%r5 \n\t" \
480 "lwz %%r7, 4(%%r4) \n\t" \
481 "addze %%r5, %%r9 \n\t" \
482 "addc %%r8, %%r8, %%r7 \n\t" \
483 "stwu %%r8, 4(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000484
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200485#define MULADDC_STOP \
486 "addze %%r5, %%r5 \n\t" \
487 "addi %%r4, %%r4, 4 \n\t" \
488 "addi %%r3, %%r3, 4 \n\t" \
489 "stw %%r5, %0 \n\t" \
490 "stw %%r4, %1 \n\t" \
491 "stw %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100492 : "=m" (c), "=m" (d), "=m" (s) \
493 : "m" (s), "m" (d), "m" (c), "m" (b) \
494 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
495 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000496
Paul Bakker9af723c2014-05-01 13:03:14 +0200497#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000498
499#endif /* PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000500
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200501/*
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200502 * The Sparc(64) assembly is reported to be broken.
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200503 * Disable it for now, until we're able to fix it.
504 */
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200505#if 0 && defined(__sparc__)
506#if defined(__sparc64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000507
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200508#define MULADDC_INIT \
509 asm( \
510 "ldx %3, %%o0 \n\t" \
511 "ldx %4, %%o1 \n\t" \
512 "ld %5, %%o2 \n\t" \
513 "ld %6, %%o3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000514
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200515#define MULADDC_CORE \
516 "ld [%%o0], %%o4 \n\t" \
517 "inc 4, %%o0 \n\t" \
518 "ld [%%o1], %%o5 \n\t" \
519 "umul %%o3, %%o4, %%o4 \n\t" \
520 "addcc %%o4, %%o2, %%o4 \n\t" \
521 "rd %%y, %%g1 \n\t" \
522 "addx %%g1, 0, %%g1 \n\t" \
523 "addcc %%o4, %%o5, %%o4 \n\t" \
524 "st %%o4, [%%o1] \n\t" \
525 "addx %%g1, 0, %%o2 \n\t" \
526 "inc 4, %%o1 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000527
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200528 #define MULADDC_STOP \
529 "st %%o2, %0 \n\t" \
530 "stx %%o1, %1 \n\t" \
531 "stx %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000532 : "=m" (c), "=m" (d), "=m" (s) \
533 : "m" (s), "m" (d), "m" (c), "m" (b) \
534 : "g1", "o0", "o1", "o2", "o3", "o4", \
535 "o5" \
536 );
Paul Bakker4f024b72012-10-30 07:29:57 +0000537
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200538#else /* __sparc64__ */
Paul Bakker4f024b72012-10-30 07:29:57 +0000539
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200540#define MULADDC_INIT \
541 asm( \
542 "ld %3, %%o0 \n\t" \
543 "ld %4, %%o1 \n\t" \
544 "ld %5, %%o2 \n\t" \
545 "ld %6, %%o3 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000546
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200547#define MULADDC_CORE \
548 "ld [%%o0], %%o4 \n\t" \
549 "inc 4, %%o0 \n\t" \
550 "ld [%%o1], %%o5 \n\t" \
551 "umul %%o3, %%o4, %%o4 \n\t" \
552 "addcc %%o4, %%o2, %%o4 \n\t" \
553 "rd %%y, %%g1 \n\t" \
554 "addx %%g1, 0, %%g1 \n\t" \
555 "addcc %%o4, %%o5, %%o4 \n\t" \
556 "st %%o4, [%%o1] \n\t" \
557 "addx %%g1, 0, %%o2 \n\t" \
558 "inc 4, %%o1 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000559
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200560#define MULADDC_STOP \
561 "st %%o2, %0 \n\t" \
562 "st %%o1, %1 \n\t" \
563 "st %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000564 : "=m" (c), "=m" (d), "=m" (s) \
565 : "m" (s), "m" (d), "m" (c), "m" (b) \
566 : "g1", "o0", "o1", "o2", "o3", "o4", \
567 "o5" \
568 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000569
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200570#endif /* __sparc64__ */
571#endif /* __sparc__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000572
573#if defined(__microblaze__) || defined(microblaze)
574
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200575#define MULADDC_INIT \
576 asm( \
577 "lwi r3, %3 \n\t" \
578 "lwi r4, %4 \n\t" \
579 "lwi r5, %5 \n\t" \
580 "lwi r6, %6 \n\t" \
581 "andi r7, r6, 0xffff \n\t" \
582 "bsrli r6, r6, 16 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000583
Kazuyuki Kimura90b99bf2021-05-31 17:07:28 +0900584#if(__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
585#define MULADDC_LHUI \
586 "lhui r9, r3, 0 \n\t" \
587 "addi r3, r3, 2 \n\t" \
588 "lhui r8, r3, 0 \n\t"
589#else
590#define MULADDC_LHUI \
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200591 "lhui r8, r3, 0 \n\t" \
592 "addi r3, r3, 2 \n\t" \
Kazuyuki Kimura90b99bf2021-05-31 17:07:28 +0900593 "lhui r9, r3, 0 \n\t"
594#endif
595
596#define MULADDC_CORE \
597 MULADDC_LHUI \
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200598 "addi r3, r3, 2 \n\t" \
599 "mul r10, r9, r6 \n\t" \
600 "mul r11, r8, r7 \n\t" \
601 "mul r12, r9, r7 \n\t" \
602 "mul r13, r8, r6 \n\t" \
603 "bsrli r8, r10, 16 \n\t" \
604 "bsrli r9, r11, 16 \n\t" \
605 "add r13, r13, r8 \n\t" \
606 "add r13, r13, r9 \n\t" \
607 "bslli r10, r10, 16 \n\t" \
608 "bslli r11, r11, 16 \n\t" \
609 "add r12, r12, r10 \n\t" \
610 "addc r13, r13, r0 \n\t" \
611 "add r12, r12, r11 \n\t" \
612 "addc r13, r13, r0 \n\t" \
613 "lwi r10, r4, 0 \n\t" \
614 "add r12, r12, r10 \n\t" \
615 "addc r13, r13, r0 \n\t" \
616 "add r12, r12, r5 \n\t" \
617 "addc r5, r13, r0 \n\t" \
618 "swi r12, r4, 0 \n\t" \
619 "addi r4, r4, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000620
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200621#define MULADDC_STOP \
622 "swi r5, %0 \n\t" \
623 "swi r4, %1 \n\t" \
624 "swi r3, %2 \n\t" \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100625 : "=m" (c), "=m" (d), "=m" (s) \
626 : "m" (s), "m" (d), "m" (c), "m" (b) \
Zach van Rijne7d3f8e2018-05-21 10:52:34 -0400627 : "r3", "r4", "r5", "r6", "r7", "r8", \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100628 "r9", "r10", "r11", "r12", "r13" \
629 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000630
631#endif /* MicroBlaze */
632
633#if defined(__tricore__)
634
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200635#define MULADDC_INIT \
636 asm( \
637 "ld.a %%a2, %3 \n\t" \
638 "ld.a %%a3, %4 \n\t" \
639 "ld.w %%d4, %5 \n\t" \
640 "ld.w %%d1, %6 \n\t" \
641 "xor %%d5, %%d5 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000642
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200643#define MULADDC_CORE \
644 "ld.w %%d0, [%%a2+] \n\t" \
645 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
646 "ld.w %%d0, [%%a3] \n\t" \
647 "addx %%d2, %%d2, %%d0 \n\t" \
648 "addc %%d3, %%d3, 0 \n\t" \
649 "mov %%d4, %%d3 \n\t" \
650 "st.w [%%a3+], %%d2 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000651
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200652#define MULADDC_STOP \
653 "st.w %0, %%d4 \n\t" \
654 "st.a %1, %%a3 \n\t" \
655 "st.a %2, %%a2 \n\t" \
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100656 : "=m" (c), "=m" (d), "=m" (s) \
657 : "m" (s), "m" (d), "m" (c), "m" (b) \
658 : "d0", "d1", "e2", "d4", "a2", "a3" \
659 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000660
661#endif /* TriCore */
662
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100663/*
Simon Butcher3ad2efd2018-05-02 14:49:38 +0100664 * Note, gcc -O0 by default uses r7 for the frame pointer, so it complains about
665 * our use of r7 below, unless -fomit-frame-pointer is passed.
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100666 *
667 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
668 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
669 * clang and armcc5 under the same conditions).
670 *
671 * So, only use the optimized assembly below for optimized build, which avoids
672 * the build error and is pretty reasonable anyway.
673 */
674#if defined(__GNUC__) && !defined(__OPTIMIZE__)
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100675#define MULADDC_CANNOT_USE_R7
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100676#endif
677
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100678#if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
Paul Bakker5121ce52009-01-03 21:22:43 +0000679
Paul Bakkerfc4f46f2013-06-24 19:23:56 +0200680#if defined(__thumb__) && !defined(__thumb2__)
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000681
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200682#define MULADDC_INIT \
683 asm( \
684 "ldr r0, %3 \n\t" \
685 "ldr r1, %4 \n\t" \
686 "ldr r2, %5 \n\t" \
687 "ldr r3, %6 \n\t" \
688 "lsr r7, r3, #16 \n\t" \
689 "mov r9, r7 \n\t" \
690 "lsl r7, r3, #16 \n\t" \
691 "lsr r7, r7, #16 \n\t" \
692 "mov r8, r7 \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000693
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200694#define MULADDC_CORE \
695 "ldmia r0!, {r6} \n\t" \
696 "lsr r7, r6, #16 \n\t" \
697 "lsl r6, r6, #16 \n\t" \
698 "lsr r6, r6, #16 \n\t" \
699 "mov r4, r8 \n\t" \
700 "mul r4, r6 \n\t" \
701 "mov r3, r9 \n\t" \
702 "mul r6, r3 \n\t" \
703 "mov r5, r9 \n\t" \
704 "mul r5, r7 \n\t" \
705 "mov r3, r8 \n\t" \
706 "mul r7, r3 \n\t" \
707 "lsr r3, r6, #16 \n\t" \
708 "add r5, r5, r3 \n\t" \
709 "lsr r3, r7, #16 \n\t" \
710 "add r5, r5, r3 \n\t" \
711 "add r4, r4, r2 \n\t" \
712 "mov r2, #0 \n\t" \
713 "adc r5, r2 \n\t" \
714 "lsl r3, r6, #16 \n\t" \
715 "add r4, r4, r3 \n\t" \
716 "adc r5, r2 \n\t" \
717 "lsl r3, r7, #16 \n\t" \
718 "add r4, r4, r3 \n\t" \
719 "adc r5, r2 \n\t" \
720 "ldr r3, [r1] \n\t" \
721 "add r4, r4, r3 \n\t" \
722 "adc r2, r5 \n\t" \
723 "stmia r1!, {r4} \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000724
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200725#define MULADDC_STOP \
726 "str r2, %0 \n\t" \
727 "str r1, %1 \n\t" \
728 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100729 : "=m" (c), "=m" (d), "=m" (s) \
730 : "m" (s), "m" (d), "m" (c), "m" (b) \
731 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200732 "r6", "r7", "r8", "r9", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100733 );
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000734
Aurelien Jarno5daa34f2018-11-03 00:46:06 +0100735#elif (__ARM_ARCH >= 6) && \
736 defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)
Aurelien Jarno16b1bd82018-05-21 22:01:21 +0200737
738#define MULADDC_INIT \
739 asm(
740
741#define MULADDC_CORE \
742 "ldr r0, [%0], #4 \n\t" \
743 "ldr r1, [%1] \n\t" \
744 "umaal r1, %2, %3, r0 \n\t" \
745 "str r1, [%1], #4 \n\t"
746
747#define MULADDC_STOP \
748 : "=r" (s), "=r" (d), "=r" (c) \
749 : "r" (b), "0" (s), "1" (d), "2" (c) \
750 : "r0", "r1", "memory" \
751 );
752
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000753#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000754
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200755#define MULADDC_INIT \
756 asm( \
757 "ldr r0, %3 \n\t" \
758 "ldr r1, %4 \n\t" \
759 "ldr r2, %5 \n\t" \
760 "ldr r3, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000761
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200762#define MULADDC_CORE \
763 "ldr r4, [r0], #4 \n\t" \
764 "mov r5, #0 \n\t" \
765 "ldr r6, [r1] \n\t" \
766 "umlal r2, r5, r3, r4 \n\t" \
767 "adds r7, r6, r2 \n\t" \
768 "adc r2, r5, #0 \n\t" \
769 "str r7, [r1], #4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000770
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200771#define MULADDC_STOP \
772 "str r2, %0 \n\t" \
773 "str r1, %1 \n\t" \
774 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100775 : "=m" (c), "=m" (d), "=m" (s) \
776 : "m" (s), "m" (d), "m" (c), "m" (b) \
777 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200778 "r6", "r7", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100779 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000780
Paul Bakkera2713a32011-11-18 12:47:23 +0000781#endif /* Thumb */
782
Paul Bakker5121ce52009-01-03 21:22:43 +0000783#endif /* ARMv3 */
784
785#if defined(__alpha__)
786
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200787#define MULADDC_INIT \
788 asm( \
789 "ldq $1, %3 \n\t" \
790 "ldq $2, %4 \n\t" \
791 "ldq $3, %5 \n\t" \
792 "ldq $4, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000793
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200794#define MULADDC_CORE \
795 "ldq $6, 0($1) \n\t" \
796 "addq $1, 8, $1 \n\t" \
797 "mulq $6, $4, $7 \n\t" \
798 "umulh $6, $4, $6 \n\t" \
799 "addq $7, $3, $7 \n\t" \
800 "cmpult $7, $3, $3 \n\t" \
801 "ldq $5, 0($2) \n\t" \
802 "addq $7, $5, $7 \n\t" \
803 "cmpult $7, $5, $5 \n\t" \
804 "stq $7, 0($2) \n\t" \
805 "addq $2, 8, $2 \n\t" \
806 "addq $6, $3, $3 \n\t" \
807 "addq $5, $3, $3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000808
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200809#define MULADDC_STOP \
810 "stq $3, %0 \n\t" \
811 "stq $2, %1 \n\t" \
812 "stq $1, %2 \n\t" \
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100813 : "=m" (c), "=m" (d), "=m" (s) \
814 : "m" (s), "m" (d), "m" (c), "m" (b) \
815 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
816 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000817#endif /* Alpha */
818
James Cowgilld1e7e8b2014-12-16 15:24:06 +0000819#if defined(__mips__) && !defined(__mips64)
Paul Bakker5121ce52009-01-03 21:22:43 +0000820
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200821#define MULADDC_INIT \
822 asm( \
823 "lw $10, %3 \n\t" \
824 "lw $11, %4 \n\t" \
825 "lw $12, %5 \n\t" \
826 "lw $13, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000827
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200828#define MULADDC_CORE \
829 "lw $14, 0($10) \n\t" \
830 "multu $13, $14 \n\t" \
831 "addi $10, $10, 4 \n\t" \
832 "mflo $14 \n\t" \
833 "mfhi $9 \n\t" \
834 "addu $14, $12, $14 \n\t" \
835 "lw $15, 0($11) \n\t" \
836 "sltu $12, $14, $12 \n\t" \
837 "addu $15, $14, $15 \n\t" \
838 "sltu $14, $15, $14 \n\t" \
839 "addu $12, $12, $9 \n\t" \
840 "sw $15, 0($11) \n\t" \
841 "addu $12, $12, $14 \n\t" \
842 "addi $11, $11, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000843
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200844#define MULADDC_STOP \
845 "sw $12, %0 \n\t" \
846 "sw $11, %1 \n\t" \
847 "sw $10, %2 \n\t" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100848 : "=m" (c), "=m" (d), "=m" (s) \
849 : "m" (s), "m" (d), "m" (c), "m" (b) \
Jeffrey Martind25fd8d2019-01-14 18:01:40 -0600850 : "$9", "$10", "$11", "$12", "$13", "$14", "$15", "lo", "hi" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100851 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000852
853#endif /* MIPS */
854#endif /* GNUC */
855
856#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
857
858#define MULADDC_INIT \
859 __asm mov esi, s \
860 __asm mov edi, d \
861 __asm mov ecx, c \
862 __asm mov ebx, b
863
864#define MULADDC_CORE \
865 __asm lodsd \
866 __asm mul ebx \
867 __asm add eax, ecx \
868 __asm adc edx, 0 \
869 __asm add eax, [edi] \
870 __asm adc edx, 0 \
871 __asm mov ecx, edx \
872 __asm stosd
873
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200874#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000875
876#define EMIT __asm _emit
877
878#define MULADDC_HUIT \
879 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
880 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
881 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
882 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
883 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
884 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
885 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
886 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
887 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
888 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
889 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
890 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
891 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
892 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
893 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
894 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
895 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
896 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
897 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
898 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
899 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
900 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
901 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
902 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
903 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
904 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
905 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
906 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
907 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
908 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
909 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
910 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
911 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
912 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
913 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
914 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
915 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
916 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
917 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
918 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
919 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
920 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
921 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
922 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
923 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
924 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
925 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
926 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
927 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
928 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
929 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
930 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
931 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
932 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
933 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
934 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
935 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
936 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
937 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
938 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
939 EMIT 0x0F EMIT 0x7E EMIT 0xC9
940
941#define MULADDC_STOP \
942 EMIT 0x0F EMIT 0x77 \
943 __asm mov c, ecx \
944 __asm mov d, edi \
945 __asm mov s, esi \
946
947#else
948
949#define MULADDC_STOP \
950 __asm mov c, ecx \
951 __asm mov d, edi \
952 __asm mov s, esi \
953
954#endif /* SSE2 */
955#endif /* MSVC */
956
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200957#endif /* MBEDTLS_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000958
959#if !defined(MULADDC_CORE)
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200960#if defined(MBEDTLS_HAVE_UDBL)
Paul Bakker5121ce52009-01-03 21:22:43 +0000961
962#define MULADDC_INIT \
963{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200964 mbedtls_t_udbl r; \
965 mbedtls_mpi_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000966
967#define MULADDC_CORE \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200968 r = *(s++) * (mbedtls_t_udbl) b; \
969 r0 = (mbedtls_mpi_uint) r; \
970 r1 = (mbedtls_mpi_uint)( r >> biL ); \
Paul Bakker5121ce52009-01-03 21:22:43 +0000971 r0 += c; r1 += (r0 < c); \
972 r0 += *d; r1 += (r0 < *d); \
973 c = r1; *(d++) = r0;
974
975#define MULADDC_STOP \
976}
977
978#else
979#define MULADDC_INIT \
980{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200981 mbedtls_mpi_uint s0, s1, b0, b1; \
982 mbedtls_mpi_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000983 b0 = ( b << biH ) >> biH; \
984 b1 = ( b >> biH );
985
986#define MULADDC_CORE \
987 s0 = ( *s << biH ) >> biH; \
988 s1 = ( *s >> biH ); s++; \
989 rx = s0 * b1; r0 = s0 * b0; \
990 ry = s1 * b0; r1 = s1 * b1; \
991 r1 += ( rx >> biH ); \
992 r1 += ( ry >> biH ); \
993 rx <<= biH; ry <<= biH; \
994 r0 += rx; r1 += (r0 < rx); \
995 r0 += ry; r1 += (r0 < ry); \
996 r0 += c; r1 += (r0 < c); \
997 r0 += *d; r1 += (r0 < *d); \
998 c = r1; *(d++) = r0;
999
1000#define MULADDC_STOP \
1001}
1002
1003#endif /* C (generic) */
1004#endif /* C (longlong) */
1005
1006#endif /* bn_mul.h */