blob: 029eff2185b8d57ab8f950e7f4b7535928e78ceb [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
Manuel Pégourié-Gonnardba194322015-05-29 09:47:57 +020089#ifndef asm
90#define asm __asm
91#endif
92
Manuel Pégourié-Gonnard854dab92015-08-10 12:08:34 +020093/* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */
94#if defined(__GNUC__) && \
95 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
Simon Butcher4b9a3ad2018-07-10 20:18:29 +010096
97/*
Peter Korsgaard9d9d45c2018-08-27 22:50:57 +020098 * GCC < 5.0 treated the x86 ebx (which is used for the GOT) as a
99 * fixed reserved register when building as PIC, leading to errors
100 * like: bn_mul.h:46:13: error: PIC register clobbered by 'ebx' in 'asm'
101 *
102 * This is fixed by an improved register allocator in GCC 5+. From the
103 * release notes:
104 * Register allocation improvements: Reuse of the PIC hard register,
105 * instead of using a fixed register, was implemented on x86/x86-64
106 * targets. This improves generated PIC code performance as more hard
107 * registers can be used.
108 */
109#if defined(__GNUC__) && __GNUC__ < 5 && defined(__PIC__)
110#define MULADDC_CANNOT_USE_EBX
111#endif
112
113/*
Simon Butcher4b9a3ad2018-07-10 20:18:29 +0100114 * Disable use of the i386 assembly code below if option -O0, to disable all
115 * compiler optimisations, is passed, detected with __OPTIMIZE__
116 * This is done as the number of registers used in the assembly code doesn't
117 * work with the -O0 option.
118 */
Peter Korsgaard9d9d45c2018-08-27 22:50:57 +0200119#if defined(__i386__) && defined(__OPTIMIZE__) && !defined(MULADDC_CANNOT_USE_EBX)
Paul Bakker5121ce52009-01-03 21:22:43 +0000120
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200121#define MULADDC_INIT \
122 asm( \
123 "movl %%ebx, %0 \n\t" \
124 "movl %5, %%esi \n\t" \
125 "movl %6, %%edi \n\t" \
126 "movl %7, %%ecx \n\t" \
127 "movl %8, %%ebx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000128
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200129#define MULADDC_CORE \
130 "lodsl \n\t" \
131 "mull %%ebx \n\t" \
132 "addl %%ecx, %%eax \n\t" \
133 "adcl $0, %%edx \n\t" \
134 "addl (%%edi), %%eax \n\t" \
135 "adcl $0, %%edx \n\t" \
136 "movl %%edx, %%ecx \n\t" \
137 "stosl \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000138
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200139#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000140
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200141#define MULADDC_HUIT \
142 "movd %%ecx, %%mm1 \n\t" \
143 "movd %%ebx, %%mm0 \n\t" \
144 "movd (%%edi), %%mm3 \n\t" \
145 "paddq %%mm3, %%mm1 \n\t" \
146 "movd (%%esi), %%mm2 \n\t" \
147 "pmuludq %%mm0, %%mm2 \n\t" \
148 "movd 4(%%esi), %%mm4 \n\t" \
149 "pmuludq %%mm0, %%mm4 \n\t" \
150 "movd 8(%%esi), %%mm6 \n\t" \
151 "pmuludq %%mm0, %%mm6 \n\t" \
152 "movd 12(%%esi), %%mm7 \n\t" \
153 "pmuludq %%mm0, %%mm7 \n\t" \
154 "paddq %%mm2, %%mm1 \n\t" \
155 "movd 4(%%edi), %%mm3 \n\t" \
156 "paddq %%mm4, %%mm3 \n\t" \
157 "movd 8(%%edi), %%mm5 \n\t" \
158 "paddq %%mm6, %%mm5 \n\t" \
159 "movd 12(%%edi), %%mm4 \n\t" \
160 "paddq %%mm4, %%mm7 \n\t" \
161 "movd %%mm1, (%%edi) \n\t" \
162 "movd 16(%%esi), %%mm2 \n\t" \
163 "pmuludq %%mm0, %%mm2 \n\t" \
164 "psrlq $32, %%mm1 \n\t" \
165 "movd 20(%%esi), %%mm4 \n\t" \
166 "pmuludq %%mm0, %%mm4 \n\t" \
167 "paddq %%mm3, %%mm1 \n\t" \
168 "movd 24(%%esi), %%mm6 \n\t" \
169 "pmuludq %%mm0, %%mm6 \n\t" \
170 "movd %%mm1, 4(%%edi) \n\t" \
171 "psrlq $32, %%mm1 \n\t" \
172 "movd 28(%%esi), %%mm3 \n\t" \
173 "pmuludq %%mm0, %%mm3 \n\t" \
174 "paddq %%mm5, %%mm1 \n\t" \
175 "movd 16(%%edi), %%mm5 \n\t" \
176 "paddq %%mm5, %%mm2 \n\t" \
177 "movd %%mm1, 8(%%edi) \n\t" \
178 "psrlq $32, %%mm1 \n\t" \
179 "paddq %%mm7, %%mm1 \n\t" \
180 "movd 20(%%edi), %%mm5 \n\t" \
181 "paddq %%mm5, %%mm4 \n\t" \
182 "movd %%mm1, 12(%%edi) \n\t" \
183 "psrlq $32, %%mm1 \n\t" \
184 "paddq %%mm2, %%mm1 \n\t" \
185 "movd 24(%%edi), %%mm5 \n\t" \
186 "paddq %%mm5, %%mm6 \n\t" \
187 "movd %%mm1, 16(%%edi) \n\t" \
188 "psrlq $32, %%mm1 \n\t" \
189 "paddq %%mm4, %%mm1 \n\t" \
190 "movd 28(%%edi), %%mm5 \n\t" \
191 "paddq %%mm5, %%mm3 \n\t" \
192 "movd %%mm1, 20(%%edi) \n\t" \
193 "psrlq $32, %%mm1 \n\t" \
194 "paddq %%mm6, %%mm1 \n\t" \
195 "movd %%mm1, 24(%%edi) \n\t" \
196 "psrlq $32, %%mm1 \n\t" \
197 "paddq %%mm3, %%mm1 \n\t" \
198 "movd %%mm1, 28(%%edi) \n\t" \
199 "addl $32, %%edi \n\t" \
200 "addl $32, %%esi \n\t" \
201 "psrlq $32, %%mm1 \n\t" \
202 "movd %%mm1, %%ecx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000203
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200204#define MULADDC_STOP \
205 "emms \n\t" \
206 "movl %4, %%ebx \n\t" \
207 "movl %%ecx, %1 \n\t" \
208 "movl %%edi, %2 \n\t" \
209 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000210 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
211 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
Simon Butcher53571642018-06-24 12:58:31 +0100212 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000213 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000214
215#else
216
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200217#define MULADDC_STOP \
218 "movl %4, %%ebx \n\t" \
219 "movl %%ecx, %1 \n\t" \
220 "movl %%edi, %2 \n\t" \
221 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000222 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
223 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
Simon Butcher53571642018-06-24 12:58:31 +0100224 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000225 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000226#endif /* SSE2 */
227#endif /* i386 */
228
229#if defined(__amd64__) || defined (__x86_64__)
230
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200231#define MULADDC_INIT \
232 asm( \
Simon Butchera86de142018-09-30 12:09:47 +0100233 "xorq %%r8, %%r8\n"
Paul Bakker5121ce52009-01-03 21:22:43 +0000234
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200235#define MULADDC_CORE \
Simon Butchera86de142018-09-30 12:09:47 +0100236 "movq (%%rsi), %%rax\n" \
237 "mulq %%rbx\n" \
238 "addq $8, %%rsi\n" \
239 "addq %%rcx, %%rax\n" \
240 "movq %%r8, %%rcx\n" \
241 "adcq $0, %%rdx\n" \
242 "nop \n" \
243 "addq %%rax, (%%rdi)\n" \
244 "adcq %%rdx, %%rcx\n" \
245 "addq $8, %%rdi\n"
Paul Bakker5121ce52009-01-03 21:22:43 +0000246
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200247#define MULADDC_STOP \
Gilles Peskinec6d977e2021-09-14 00:13:05 +0200248 : "+c" (c), "+D" (d), "+S" (s), "+m" (*(uint64_t (*)[16]) d) \
249 : "b" (b), "m" (*(const uint64_t (*)[16]) s) \
250 : "rax", "rdx", "r8" \
Manuel Pégourié-Gonnarddef018d2014-01-07 17:50:46 +0100251 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000252
253#endif /* AMD64 */
254
Ko-cc1871e2018-08-16 02:01:57 -0700255#if defined(__aarch64__)
256
257#define MULADDC_INIT \
258 asm(
259
260#define MULADDC_CORE \
Ko-cb260bb2018-08-20 13:59:53 +0100261 "ldr x4, [%2], #8 \n\t" \
262 "ldr x5, [%1] \n\t" \
David Horstmanna23be222021-09-22 18:15:51 +0100263 "mul x6, x4, %4 \n\t" \
264 "umulh x7, x4, %4 \n\t" \
Ko-cc1871e2018-08-16 02:01:57 -0700265 "adds x5, x5, x6 \n\t" \
266 "adc x7, x7, xzr \n\t" \
Ko-cb260bb2018-08-20 13:59:53 +0100267 "adds x5, x5, %0 \n\t" \
Ko-cc1871e2018-08-16 02:01:57 -0700268 "adc %0, x7, xzr \n\t" \
269 "str x5, [%1], #8 \n\t"
270
David Horstmanna23be222021-09-22 18:15:51 +0100271#define MULADDC_STOP \
272 : "+r" (c), "+r" (d), "+r" (s), "+m" (*(uint64_t (*)[16]) d) \
273 : "r" (b), "m" (*(const uint64_t (*)[16]) s) \
274 : "x4", "x5", "x6", "x7", "cc" \
Ko-cc1871e2018-08-16 02:01:57 -0700275 );
276
277#endif /* Aarch64 */
278
Paul Bakker5121ce52009-01-03 21:22:43 +0000279#if defined(__mc68020__) || defined(__mcpu32__)
280
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200281#define MULADDC_INIT \
282 asm( \
283 "movl %3, %%a2 \n\t" \
284 "movl %4, %%a3 \n\t" \
285 "movl %5, %%d3 \n\t" \
286 "movl %6, %%d2 \n\t" \
287 "moveq #0, %%d0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000288
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200289#define MULADDC_CORE \
290 "movel %%a2@+, %%d1 \n\t" \
291 "mulul %%d2, %%d4:%%d1 \n\t" \
292 "addl %%d3, %%d1 \n\t" \
293 "addxl %%d0, %%d4 \n\t" \
294 "moveq #0, %%d3 \n\t" \
295 "addl %%d1, %%a3@+ \n\t" \
296 "addxl %%d4, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000297
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200298#define MULADDC_STOP \
299 "movl %%d3, %0 \n\t" \
300 "movl %%a3, %1 \n\t" \
301 "movl %%a2, %2 \n\t" \
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100302 : "=m" (c), "=m" (d), "=m" (s) \
303 : "m" (s), "m" (d), "m" (c), "m" (b) \
304 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
305 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000306
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200307#define MULADDC_HUIT \
308 "movel %%a2@+, %%d1 \n\t" \
309 "mulul %%d2, %%d4:%%d1 \n\t" \
310 "addxl %%d3, %%d1 \n\t" \
311 "addxl %%d0, %%d4 \n\t" \
312 "addl %%d1, %%a3@+ \n\t" \
313 "movel %%a2@+, %%d1 \n\t" \
314 "mulul %%d2, %%d3:%%d1 \n\t" \
315 "addxl %%d4, %%d1 \n\t" \
316 "addxl %%d0, %%d3 \n\t" \
317 "addl %%d1, %%a3@+ \n\t" \
318 "movel %%a2@+, %%d1 \n\t" \
319 "mulul %%d2, %%d4:%%d1 \n\t" \
320 "addxl %%d3, %%d1 \n\t" \
321 "addxl %%d0, %%d4 \n\t" \
322 "addl %%d1, %%a3@+ \n\t" \
323 "movel %%a2@+, %%d1 \n\t" \
324 "mulul %%d2, %%d3:%%d1 \n\t" \
325 "addxl %%d4, %%d1 \n\t" \
326 "addxl %%d0, %%d3 \n\t" \
327 "addl %%d1, %%a3@+ \n\t" \
328 "movel %%a2@+, %%d1 \n\t" \
329 "mulul %%d2, %%d4:%%d1 \n\t" \
330 "addxl %%d3, %%d1 \n\t" \
331 "addxl %%d0, %%d4 \n\t" \
332 "addl %%d1, %%a3@+ \n\t" \
333 "movel %%a2@+, %%d1 \n\t" \
334 "mulul %%d2, %%d3:%%d1 \n\t" \
335 "addxl %%d4, %%d1 \n\t" \
336 "addxl %%d0, %%d3 \n\t" \
337 "addl %%d1, %%a3@+ \n\t" \
338 "movel %%a2@+, %%d1 \n\t" \
339 "mulul %%d2, %%d4:%%d1 \n\t" \
340 "addxl %%d3, %%d1 \n\t" \
341 "addxl %%d0, %%d4 \n\t" \
342 "addl %%d1, %%a3@+ \n\t" \
343 "movel %%a2@+, %%d1 \n\t" \
344 "mulul %%d2, %%d3:%%d1 \n\t" \
345 "addxl %%d4, %%d1 \n\t" \
346 "addxl %%d0, %%d3 \n\t" \
347 "addl %%d1, %%a3@+ \n\t" \
348 "addxl %%d0, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000349
350#endif /* MC68000 */
351
Paul Bakker5121ce52009-01-03 21:22:43 +0000352#if defined(__powerpc64__) || defined(__ppc64__)
353
354#if defined(__MACH__) && defined(__APPLE__)
355
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200356#define MULADDC_INIT \
357 asm( \
358 "ld r3, %3 \n\t" \
359 "ld r4, %4 \n\t" \
360 "ld r5, %5 \n\t" \
361 "ld r6, %6 \n\t" \
362 "addi r3, r3, -8 \n\t" \
363 "addi r4, r4, -8 \n\t" \
364 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000365
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200366#define MULADDC_CORE \
367 "ldu r7, 8(r3) \n\t" \
368 "mulld r8, r7, r6 \n\t" \
369 "mulhdu r9, r7, r6 \n\t" \
370 "adde r8, r8, r5 \n\t" \
371 "ld r7, 8(r4) \n\t" \
372 "addze r5, r9 \n\t" \
373 "addc r8, r8, r7 \n\t" \
374 "stdu r8, 8(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000375
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200376#define MULADDC_STOP \
377 "addze r5, r5 \n\t" \
378 "addi r4, r4, 8 \n\t" \
379 "addi r3, r3, 8 \n\t" \
380 "std r5, %0 \n\t" \
381 "std r4, %1 \n\t" \
382 "std r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100383 : "=m" (c), "=m" (d), "=m" (s) \
384 : "m" (s), "m" (d), "m" (c), "m" (b) \
385 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
386 );
387
Paul Bakker5121ce52009-01-03 21:22:43 +0000388
Paul Bakker9af723c2014-05-01 13:03:14 +0200389#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000390
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200391#define MULADDC_INIT \
392 asm( \
393 "ld %%r3, %3 \n\t" \
394 "ld %%r4, %4 \n\t" \
395 "ld %%r5, %5 \n\t" \
396 "ld %%r6, %6 \n\t" \
397 "addi %%r3, %%r3, -8 \n\t" \
398 "addi %%r4, %%r4, -8 \n\t" \
399 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000400
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200401#define MULADDC_CORE \
402 "ldu %%r7, 8(%%r3) \n\t" \
403 "mulld %%r8, %%r7, %%r6 \n\t" \
404 "mulhdu %%r9, %%r7, %%r6 \n\t" \
405 "adde %%r8, %%r8, %%r5 \n\t" \
406 "ld %%r7, 8(%%r4) \n\t" \
407 "addze %%r5, %%r9 \n\t" \
408 "addc %%r8, %%r8, %%r7 \n\t" \
409 "stdu %%r8, 8(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000410
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200411#define MULADDC_STOP \
412 "addze %%r5, %%r5 \n\t" \
413 "addi %%r4, %%r4, 8 \n\t" \
414 "addi %%r3, %%r3, 8 \n\t" \
415 "std %%r5, %0 \n\t" \
416 "std %%r4, %1 \n\t" \
417 "std %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100418 : "=m" (c), "=m" (d), "=m" (s) \
419 : "m" (s), "m" (d), "m" (c), "m" (b) \
420 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
421 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000422
Paul Bakker9af723c2014-05-01 13:03:14 +0200423#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000424
Barry K. Nathan35e7cb92014-05-05 23:26:13 -0700425#elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000426
427#if defined(__MACH__) && defined(__APPLE__)
428
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200429#define MULADDC_INIT \
430 asm( \
431 "lwz r3, %3 \n\t" \
432 "lwz r4, %4 \n\t" \
433 "lwz r5, %5 \n\t" \
434 "lwz r6, %6 \n\t" \
435 "addi r3, r3, -4 \n\t" \
436 "addi r4, r4, -4 \n\t" \
437 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000438
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200439#define MULADDC_CORE \
440 "lwzu r7, 4(r3) \n\t" \
441 "mullw r8, r7, r6 \n\t" \
442 "mulhwu r9, r7, r6 \n\t" \
443 "adde r8, r8, r5 \n\t" \
444 "lwz r7, 4(r4) \n\t" \
445 "addze r5, r9 \n\t" \
446 "addc r8, r8, r7 \n\t" \
447 "stwu r8, 4(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000448
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200449#define MULADDC_STOP \
450 "addze r5, r5 \n\t" \
451 "addi r4, r4, 4 \n\t" \
452 "addi r3, r3, 4 \n\t" \
453 "stw r5, %0 \n\t" \
454 "stw r4, %1 \n\t" \
455 "stw r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100456 : "=m" (c), "=m" (d), "=m" (s) \
457 : "m" (s), "m" (d), "m" (c), "m" (b) \
458 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
459 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000460
Paul Bakker9af723c2014-05-01 13:03:14 +0200461#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000462
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200463#define MULADDC_INIT \
464 asm( \
465 "lwz %%r3, %3 \n\t" \
466 "lwz %%r4, %4 \n\t" \
467 "lwz %%r5, %5 \n\t" \
468 "lwz %%r6, %6 \n\t" \
469 "addi %%r3, %%r3, -4 \n\t" \
470 "addi %%r4, %%r4, -4 \n\t" \
471 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000472
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200473#define MULADDC_CORE \
474 "lwzu %%r7, 4(%%r3) \n\t" \
475 "mullw %%r8, %%r7, %%r6 \n\t" \
476 "mulhwu %%r9, %%r7, %%r6 \n\t" \
477 "adde %%r8, %%r8, %%r5 \n\t" \
478 "lwz %%r7, 4(%%r4) \n\t" \
479 "addze %%r5, %%r9 \n\t" \
480 "addc %%r8, %%r8, %%r7 \n\t" \
481 "stwu %%r8, 4(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000482
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200483#define MULADDC_STOP \
484 "addze %%r5, %%r5 \n\t" \
485 "addi %%r4, %%r4, 4 \n\t" \
486 "addi %%r3, %%r3, 4 \n\t" \
487 "stw %%r5, %0 \n\t" \
488 "stw %%r4, %1 \n\t" \
489 "stw %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100490 : "=m" (c), "=m" (d), "=m" (s) \
491 : "m" (s), "m" (d), "m" (c), "m" (b) \
492 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
493 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000494
Paul Bakker9af723c2014-05-01 13:03:14 +0200495#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000496
497#endif /* PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000498
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200499/*
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200500 * The Sparc(64) assembly is reported to be broken.
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200501 * Disable it for now, until we're able to fix it.
502 */
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200503#if 0 && defined(__sparc__)
504#if defined(__sparc64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000505
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200506#define MULADDC_INIT \
507 asm( \
508 "ldx %3, %%o0 \n\t" \
509 "ldx %4, %%o1 \n\t" \
510 "ld %5, %%o2 \n\t" \
511 "ld %6, %%o3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000512
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200513#define MULADDC_CORE \
514 "ld [%%o0], %%o4 \n\t" \
515 "inc 4, %%o0 \n\t" \
516 "ld [%%o1], %%o5 \n\t" \
517 "umul %%o3, %%o4, %%o4 \n\t" \
518 "addcc %%o4, %%o2, %%o4 \n\t" \
519 "rd %%y, %%g1 \n\t" \
520 "addx %%g1, 0, %%g1 \n\t" \
521 "addcc %%o4, %%o5, %%o4 \n\t" \
522 "st %%o4, [%%o1] \n\t" \
523 "addx %%g1, 0, %%o2 \n\t" \
524 "inc 4, %%o1 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000525
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200526 #define MULADDC_STOP \
527 "st %%o2, %0 \n\t" \
528 "stx %%o1, %1 \n\t" \
529 "stx %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000530 : "=m" (c), "=m" (d), "=m" (s) \
531 : "m" (s), "m" (d), "m" (c), "m" (b) \
532 : "g1", "o0", "o1", "o2", "o3", "o4", \
533 "o5" \
534 );
Paul Bakker4f024b72012-10-30 07:29:57 +0000535
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200536#else /* __sparc64__ */
Paul Bakker4f024b72012-10-30 07:29:57 +0000537
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200538#define MULADDC_INIT \
539 asm( \
540 "ld %3, %%o0 \n\t" \
541 "ld %4, %%o1 \n\t" \
542 "ld %5, %%o2 \n\t" \
543 "ld %6, %%o3 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000544
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200545#define MULADDC_CORE \
546 "ld [%%o0], %%o4 \n\t" \
547 "inc 4, %%o0 \n\t" \
548 "ld [%%o1], %%o5 \n\t" \
549 "umul %%o3, %%o4, %%o4 \n\t" \
550 "addcc %%o4, %%o2, %%o4 \n\t" \
551 "rd %%y, %%g1 \n\t" \
552 "addx %%g1, 0, %%g1 \n\t" \
553 "addcc %%o4, %%o5, %%o4 \n\t" \
554 "st %%o4, [%%o1] \n\t" \
555 "addx %%g1, 0, %%o2 \n\t" \
556 "inc 4, %%o1 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000557
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200558#define MULADDC_STOP \
559 "st %%o2, %0 \n\t" \
560 "st %%o1, %1 \n\t" \
561 "st %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000562 : "=m" (c), "=m" (d), "=m" (s) \
563 : "m" (s), "m" (d), "m" (c), "m" (b) \
564 : "g1", "o0", "o1", "o2", "o3", "o4", \
565 "o5" \
566 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000567
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200568#endif /* __sparc64__ */
569#endif /* __sparc__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000570
571#if defined(__microblaze__) || defined(microblaze)
572
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200573#define MULADDC_INIT \
574 asm( \
575 "lwi r3, %3 \n\t" \
576 "lwi r4, %4 \n\t" \
577 "lwi r5, %5 \n\t" \
578 "lwi r6, %6 \n\t" \
579 "andi r7, r6, 0xffff \n\t" \
580 "bsrli r6, r6, 16 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000581
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200582#define MULADDC_CORE \
583 "lhui r8, r3, 0 \n\t" \
584 "addi r3, r3, 2 \n\t" \
585 "lhui r9, r3, 0 \n\t" \
586 "addi r3, r3, 2 \n\t" \
587 "mul r10, r9, r6 \n\t" \
588 "mul r11, r8, r7 \n\t" \
589 "mul r12, r9, r7 \n\t" \
590 "mul r13, r8, r6 \n\t" \
591 "bsrli r8, r10, 16 \n\t" \
592 "bsrli r9, r11, 16 \n\t" \
593 "add r13, r13, r8 \n\t" \
594 "add r13, r13, r9 \n\t" \
595 "bslli r10, r10, 16 \n\t" \
596 "bslli r11, r11, 16 \n\t" \
597 "add r12, r12, r10 \n\t" \
598 "addc r13, r13, r0 \n\t" \
599 "add r12, r12, r11 \n\t" \
600 "addc r13, r13, r0 \n\t" \
601 "lwi r10, r4, 0 \n\t" \
602 "add r12, r12, r10 \n\t" \
603 "addc r13, r13, r0 \n\t" \
604 "add r12, r12, r5 \n\t" \
605 "addc r5, r13, r0 \n\t" \
606 "swi r12, r4, 0 \n\t" \
607 "addi r4, r4, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000608
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200609#define MULADDC_STOP \
610 "swi r5, %0 \n\t" \
611 "swi r4, %1 \n\t" \
612 "swi r3, %2 \n\t" \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100613 : "=m" (c), "=m" (d), "=m" (s) \
614 : "m" (s), "m" (d), "m" (c), "m" (b) \
Zach van Rijne7d3f8e2018-05-21 10:52:34 -0400615 : "r3", "r4", "r5", "r6", "r7", "r8", \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100616 "r9", "r10", "r11", "r12", "r13" \
617 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000618
619#endif /* MicroBlaze */
620
621#if defined(__tricore__)
622
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200623#define MULADDC_INIT \
624 asm( \
625 "ld.a %%a2, %3 \n\t" \
626 "ld.a %%a3, %4 \n\t" \
627 "ld.w %%d4, %5 \n\t" \
628 "ld.w %%d1, %6 \n\t" \
629 "xor %%d5, %%d5 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000630
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200631#define MULADDC_CORE \
632 "ld.w %%d0, [%%a2+] \n\t" \
633 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
634 "ld.w %%d0, [%%a3] \n\t" \
635 "addx %%d2, %%d2, %%d0 \n\t" \
636 "addc %%d3, %%d3, 0 \n\t" \
637 "mov %%d4, %%d3 \n\t" \
638 "st.w [%%a3+], %%d2 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000639
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200640#define MULADDC_STOP \
641 "st.w %0, %%d4 \n\t" \
642 "st.a %1, %%a3 \n\t" \
643 "st.a %2, %%a2 \n\t" \
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100644 : "=m" (c), "=m" (d), "=m" (s) \
645 : "m" (s), "m" (d), "m" (c), "m" (b) \
646 : "d0", "d1", "e2", "d4", "a2", "a3" \
647 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000648
649#endif /* TriCore */
650
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100651/*
Simon Butcher3ad2efd2018-05-02 14:49:38 +0100652 * Note, gcc -O0 by default uses r7 for the frame pointer, so it complains about
653 * our use of r7 below, unless -fomit-frame-pointer is passed.
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100654 *
655 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
656 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
657 * clang and armcc5 under the same conditions).
658 *
659 * So, only use the optimized assembly below for optimized build, which avoids
660 * the build error and is pretty reasonable anyway.
661 */
662#if defined(__GNUC__) && !defined(__OPTIMIZE__)
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100663#define MULADDC_CANNOT_USE_R7
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100664#endif
665
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100666#if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
Paul Bakker5121ce52009-01-03 21:22:43 +0000667
Paul Bakkerfc4f46f2013-06-24 19:23:56 +0200668#if defined(__thumb__) && !defined(__thumb2__)
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000669
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200670#define MULADDC_INIT \
671 asm( \
672 "ldr r0, %3 \n\t" \
673 "ldr r1, %4 \n\t" \
674 "ldr r2, %5 \n\t" \
675 "ldr r3, %6 \n\t" \
676 "lsr r7, r3, #16 \n\t" \
677 "mov r9, r7 \n\t" \
678 "lsl r7, r3, #16 \n\t" \
679 "lsr r7, r7, #16 \n\t" \
680 "mov r8, r7 \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000681
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200682#define MULADDC_CORE \
683 "ldmia r0!, {r6} \n\t" \
684 "lsr r7, r6, #16 \n\t" \
685 "lsl r6, r6, #16 \n\t" \
686 "lsr r6, r6, #16 \n\t" \
687 "mov r4, r8 \n\t" \
688 "mul r4, r6 \n\t" \
689 "mov r3, r9 \n\t" \
690 "mul r6, r3 \n\t" \
691 "mov r5, r9 \n\t" \
692 "mul r5, r7 \n\t" \
693 "mov r3, r8 \n\t" \
694 "mul r7, r3 \n\t" \
695 "lsr r3, r6, #16 \n\t" \
696 "add r5, r5, r3 \n\t" \
697 "lsr r3, r7, #16 \n\t" \
698 "add r5, r5, r3 \n\t" \
699 "add r4, r4, r2 \n\t" \
700 "mov r2, #0 \n\t" \
701 "adc r5, r2 \n\t" \
702 "lsl r3, r6, #16 \n\t" \
703 "add r4, r4, r3 \n\t" \
704 "adc r5, r2 \n\t" \
705 "lsl r3, r7, #16 \n\t" \
706 "add r4, r4, r3 \n\t" \
707 "adc r5, r2 \n\t" \
708 "ldr r3, [r1] \n\t" \
709 "add r4, r4, r3 \n\t" \
710 "adc r2, r5 \n\t" \
711 "stmia r1!, {r4} \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000712
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200713#define MULADDC_STOP \
714 "str r2, %0 \n\t" \
715 "str r1, %1 \n\t" \
716 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100717 : "=m" (c), "=m" (d), "=m" (s) \
718 : "m" (s), "m" (d), "m" (c), "m" (b) \
719 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200720 "r6", "r7", "r8", "r9", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100721 );
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000722
Aurelien Jarno5daa34f2018-11-03 00:46:06 +0100723#elif (__ARM_ARCH >= 6) && \
724 defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)
Aurelien Jarno16b1bd82018-05-21 22:01:21 +0200725
726#define MULADDC_INIT \
727 asm(
728
729#define MULADDC_CORE \
730 "ldr r0, [%0], #4 \n\t" \
731 "ldr r1, [%1] \n\t" \
732 "umaal r1, %2, %3, r0 \n\t" \
733 "str r1, [%1], #4 \n\t"
734
735#define MULADDC_STOP \
736 : "=r" (s), "=r" (d), "=r" (c) \
737 : "r" (b), "0" (s), "1" (d), "2" (c) \
738 : "r0", "r1", "memory" \
739 );
740
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000741#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000742
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200743#define MULADDC_INIT \
744 asm( \
745 "ldr r0, %3 \n\t" \
746 "ldr r1, %4 \n\t" \
747 "ldr r2, %5 \n\t" \
748 "ldr r3, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000749
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200750#define MULADDC_CORE \
751 "ldr r4, [r0], #4 \n\t" \
752 "mov r5, #0 \n\t" \
753 "ldr r6, [r1] \n\t" \
754 "umlal r2, r5, r3, r4 \n\t" \
755 "adds r7, r6, r2 \n\t" \
756 "adc r2, r5, #0 \n\t" \
757 "str r7, [r1], #4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000758
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200759#define MULADDC_STOP \
760 "str r2, %0 \n\t" \
761 "str r1, %1 \n\t" \
762 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100763 : "=m" (c), "=m" (d), "=m" (s) \
764 : "m" (s), "m" (d), "m" (c), "m" (b) \
765 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200766 "r6", "r7", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100767 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000768
Paul Bakkera2713a32011-11-18 12:47:23 +0000769#endif /* Thumb */
770
Paul Bakker5121ce52009-01-03 21:22:43 +0000771#endif /* ARMv3 */
772
773#if defined(__alpha__)
774
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200775#define MULADDC_INIT \
776 asm( \
777 "ldq $1, %3 \n\t" \
778 "ldq $2, %4 \n\t" \
779 "ldq $3, %5 \n\t" \
780 "ldq $4, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000781
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200782#define MULADDC_CORE \
783 "ldq $6, 0($1) \n\t" \
784 "addq $1, 8, $1 \n\t" \
785 "mulq $6, $4, $7 \n\t" \
786 "umulh $6, $4, $6 \n\t" \
787 "addq $7, $3, $7 \n\t" \
788 "cmpult $7, $3, $3 \n\t" \
789 "ldq $5, 0($2) \n\t" \
790 "addq $7, $5, $7 \n\t" \
791 "cmpult $7, $5, $5 \n\t" \
792 "stq $7, 0($2) \n\t" \
793 "addq $2, 8, $2 \n\t" \
794 "addq $6, $3, $3 \n\t" \
795 "addq $5, $3, $3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000796
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200797#define MULADDC_STOP \
798 "stq $3, %0 \n\t" \
799 "stq $2, %1 \n\t" \
800 "stq $1, %2 \n\t" \
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100801 : "=m" (c), "=m" (d), "=m" (s) \
802 : "m" (s), "m" (d), "m" (c), "m" (b) \
803 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
804 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000805#endif /* Alpha */
806
James Cowgilld1e7e8b2014-12-16 15:24:06 +0000807#if defined(__mips__) && !defined(__mips64)
Paul Bakker5121ce52009-01-03 21:22:43 +0000808
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200809#define MULADDC_INIT \
810 asm( \
811 "lw $10, %3 \n\t" \
812 "lw $11, %4 \n\t" \
813 "lw $12, %5 \n\t" \
814 "lw $13, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000815
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200816#define MULADDC_CORE \
817 "lw $14, 0($10) \n\t" \
818 "multu $13, $14 \n\t" \
819 "addi $10, $10, 4 \n\t" \
820 "mflo $14 \n\t" \
821 "mfhi $9 \n\t" \
822 "addu $14, $12, $14 \n\t" \
823 "lw $15, 0($11) \n\t" \
824 "sltu $12, $14, $12 \n\t" \
825 "addu $15, $14, $15 \n\t" \
826 "sltu $14, $15, $14 \n\t" \
827 "addu $12, $12, $9 \n\t" \
828 "sw $15, 0($11) \n\t" \
829 "addu $12, $12, $14 \n\t" \
830 "addi $11, $11, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000831
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200832#define MULADDC_STOP \
833 "sw $12, %0 \n\t" \
834 "sw $11, %1 \n\t" \
835 "sw $10, %2 \n\t" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100836 : "=m" (c), "=m" (d), "=m" (s) \
837 : "m" (s), "m" (d), "m" (c), "m" (b) \
Jeffrey Martind25fd8d2019-01-14 18:01:40 -0600838 : "$9", "$10", "$11", "$12", "$13", "$14", "$15", "lo", "hi" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100839 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000840
841#endif /* MIPS */
842#endif /* GNUC */
843
844#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
845
846#define MULADDC_INIT \
847 __asm mov esi, s \
848 __asm mov edi, d \
849 __asm mov ecx, c \
850 __asm mov ebx, b
851
852#define MULADDC_CORE \
853 __asm lodsd \
854 __asm mul ebx \
855 __asm add eax, ecx \
856 __asm adc edx, 0 \
857 __asm add eax, [edi] \
858 __asm adc edx, 0 \
859 __asm mov ecx, edx \
860 __asm stosd
861
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200862#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000863
864#define EMIT __asm _emit
865
866#define MULADDC_HUIT \
867 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
868 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
869 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
870 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
871 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
872 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
873 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
874 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
875 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
876 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
877 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
878 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
879 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
880 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
881 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
882 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
883 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
884 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
885 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
886 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
887 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
888 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
889 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
890 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
891 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
892 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
893 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
894 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
895 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
896 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
897 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
898 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
899 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
900 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
901 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
902 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
903 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
904 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
905 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
906 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
907 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
908 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
909 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
910 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
911 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
912 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
913 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
914 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
915 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
916 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
917 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
918 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
919 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
920 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
921 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
922 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
923 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
924 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
925 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
926 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
927 EMIT 0x0F EMIT 0x7E EMIT 0xC9
928
929#define MULADDC_STOP \
930 EMIT 0x0F EMIT 0x77 \
931 __asm mov c, ecx \
932 __asm mov d, edi \
933 __asm mov s, esi \
934
935#else
936
937#define MULADDC_STOP \
938 __asm mov c, ecx \
939 __asm mov d, edi \
940 __asm mov s, esi \
941
942#endif /* SSE2 */
943#endif /* MSVC */
944
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200945#endif /* MBEDTLS_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000946
947#if !defined(MULADDC_CORE)
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200948#if defined(MBEDTLS_HAVE_UDBL)
Paul Bakker5121ce52009-01-03 21:22:43 +0000949
950#define MULADDC_INIT \
951{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200952 mbedtls_t_udbl r; \
953 mbedtls_mpi_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000954
955#define MULADDC_CORE \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200956 r = *(s++) * (mbedtls_t_udbl) b; \
957 r0 = (mbedtls_mpi_uint) r; \
958 r1 = (mbedtls_mpi_uint)( r >> biL ); \
Paul Bakker5121ce52009-01-03 21:22:43 +0000959 r0 += c; r1 += (r0 < c); \
960 r0 += *d; r1 += (r0 < *d); \
961 c = r1; *(d++) = r0;
962
963#define MULADDC_STOP \
964}
965
966#else
967#define MULADDC_INIT \
968{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200969 mbedtls_mpi_uint s0, s1, b0, b1; \
970 mbedtls_mpi_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000971 b0 = ( b << biH ) >> biH; \
972 b1 = ( b >> biH );
973
974#define MULADDC_CORE \
975 s0 = ( *s << biH ) >> biH; \
976 s1 = ( *s >> biH ); s++; \
977 rx = s0 * b1; r0 = s0 * b0; \
978 ry = s1 * b0; r1 = s1 * b1; \
979 r1 += ( rx >> biH ); \
980 r1 += ( ry >> biH ); \
981 rx <<= biH; ry <<= biH; \
982 r0 += rx; r1 += (r0 < rx); \
983 r0 += ry; r1 += (r0 < ry); \
984 r0 += c; r1 += (r0 < c); \
985 r0 += *d; r1 += (r0 < *d); \
986 c = r1; *(d++) = r0;
987
988#define MULADDC_STOP \
989}
990
991#endif /* C (generic) */
992#endif /* C (longlong) */
993
994#endif /* bn_mul.h */