blob: fc0c3cf318969add8b5c9b3ee35c3cfb48cd4260 [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
Dave Rodgman7ff79652023-11-03 12:04:52 +00008 * SPDX-License-Identifier: Apache-2.0 OR GPL-2.0-or-later
Paul Bakker5121ce52009-01-03 21:22:43 +00009 */
10/*
11 * Multiply source vector [s] with b, add result
12 * to destination vector [d] and set carry c.
13 *
14 * Currently supports:
15 *
16 * . IA-32 (386+) . AMD64 / EM64T
17 * . IA-32 (SSE2) . Motorola 68000
18 * . PowerPC, 32-bit . MicroBlaze
19 * . PowerPC, 64-bit . TriCore
20 * . SPARC v8 . ARM v3+
21 * . Alpha . MIPS32
22 * . C, longlong . C, generic
23 */
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +020024#ifndef MBEDTLS_BN_MUL_H
25#define MBEDTLS_BN_MUL_H
Paul Bakker5121ce52009-01-03 21:22:43 +000026
Ron Eldor8b0cf2e2018-02-14 16:02:41 +020027#if !defined(MBEDTLS_CONFIG_FILE)
Jaeden Ameroc49fbbf2019-07-04 20:01:14 +010028#include "mbedtls/config.h"
Ron Eldor8b0cf2e2018-02-14 16:02:41 +020029#else
30#include MBEDTLS_CONFIG_FILE
31#endif
32
Jaeden Ameroc49fbbf2019-07-04 20:01:14 +010033#include "mbedtls/bignum.h"
Paul Bakker5121ce52009-01-03 21:22:43 +000034
Janos Follathbc96a792021-06-24 14:48:38 +010035
36/*
37 * Conversion macros for embedded constants:
38 * build lists of mbedtls_mpi_uint's from lists of unsigned char's grouped by 8, 4 or 2
39 */
40#if defined(MBEDTLS_HAVE_INT32)
41
Gilles Peskine1b6c09a2023-01-11 14:52:35 +010042#define MBEDTLS_BYTES_TO_T_UINT_4(a, b, c, d) \
43 ((mbedtls_mpi_uint) (a) << 0) | \
44 ((mbedtls_mpi_uint) (b) << 8) | \
45 ((mbedtls_mpi_uint) (c) << 16) | \
46 ((mbedtls_mpi_uint) (d) << 24)
Janos Follathbc96a792021-06-24 14:48:38 +010047
Gilles Peskine1b6c09a2023-01-11 14:52:35 +010048#define MBEDTLS_BYTES_TO_T_UINT_2(a, b) \
49 MBEDTLS_BYTES_TO_T_UINT_4(a, b, 0, 0)
Janos Follathbc96a792021-06-24 14:48:38 +010050
Gilles Peskine1b6c09a2023-01-11 14:52:35 +010051#define MBEDTLS_BYTES_TO_T_UINT_8(a, b, c, d, e, f, g, h) \
52 MBEDTLS_BYTES_TO_T_UINT_4(a, b, c, d), \
53 MBEDTLS_BYTES_TO_T_UINT_4(e, f, g, h)
Janos Follathbc96a792021-06-24 14:48:38 +010054
55#else /* 64-bits */
56
Gilles Peskine1b6c09a2023-01-11 14:52:35 +010057#define MBEDTLS_BYTES_TO_T_UINT_8(a, b, c, d, e, f, g, h) \
58 ((mbedtls_mpi_uint) (a) << 0) | \
59 ((mbedtls_mpi_uint) (b) << 8) | \
60 ((mbedtls_mpi_uint) (c) << 16) | \
61 ((mbedtls_mpi_uint) (d) << 24) | \
62 ((mbedtls_mpi_uint) (e) << 32) | \
63 ((mbedtls_mpi_uint) (f) << 40) | \
64 ((mbedtls_mpi_uint) (g) << 48) | \
65 ((mbedtls_mpi_uint) (h) << 56)
Janos Follathbc96a792021-06-24 14:48:38 +010066
Gilles Peskine1b6c09a2023-01-11 14:52:35 +010067#define MBEDTLS_BYTES_TO_T_UINT_4(a, b, c, d) \
68 MBEDTLS_BYTES_TO_T_UINT_8(a, b, c, d, 0, 0, 0, 0)
Janos Follathbc96a792021-06-24 14:48:38 +010069
Gilles Peskine1b6c09a2023-01-11 14:52:35 +010070#define MBEDTLS_BYTES_TO_T_UINT_2(a, b) \
71 MBEDTLS_BYTES_TO_T_UINT_8(a, b, 0, 0, 0, 0, 0, 0)
Janos Follathbc96a792021-06-24 14:48:38 +010072
73#endif /* bits in mbedtls_mpi_uint */
74
David Horstmannd6818e32023-01-04 17:50:08 +000075/* *INDENT-OFF* */
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +020076#if defined(MBEDTLS_HAVE_ASM)
Paul Bakker5121ce52009-01-03 21:22:43 +000077
Manuel Pégourié-Gonnardba194322015-05-29 09:47:57 +020078#ifndef asm
79#define asm __asm
80#endif
81
Manuel Pégourié-Gonnard854dab92015-08-10 12:08:34 +020082/* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */
83#if defined(__GNUC__) && \
84 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
Simon Butcher4b9a3ad2018-07-10 20:18:29 +010085
86/*
Peter Korsgaard9d9d45c2018-08-27 22:50:57 +020087 * GCC < 5.0 treated the x86 ebx (which is used for the GOT) as a
88 * fixed reserved register when building as PIC, leading to errors
89 * like: bn_mul.h:46:13: error: PIC register clobbered by 'ebx' in 'asm'
90 *
91 * This is fixed by an improved register allocator in GCC 5+. From the
92 * release notes:
93 * Register allocation improvements: Reuse of the PIC hard register,
94 * instead of using a fixed register, was implemented on x86/x86-64
95 * targets. This improves generated PIC code performance as more hard
96 * registers can be used.
97 */
98#if defined(__GNUC__) && __GNUC__ < 5 && defined(__PIC__)
99#define MULADDC_CANNOT_USE_EBX
100#endif
101
102/*
Simon Butcher4b9a3ad2018-07-10 20:18:29 +0100103 * Disable use of the i386 assembly code below if option -O0, to disable all
104 * compiler optimisations, is passed, detected with __OPTIMIZE__
105 * This is done as the number of registers used in the assembly code doesn't
106 * work with the -O0 option.
107 */
Peter Korsgaard9d9d45c2018-08-27 22:50:57 +0200108#if defined(__i386__) && defined(__OPTIMIZE__) && !defined(MULADDC_CANNOT_USE_EBX)
Paul Bakker5121ce52009-01-03 21:22:43 +0000109
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200110#define MULADDC_INIT \
111 asm( \
112 "movl %%ebx, %0 \n\t" \
113 "movl %5, %%esi \n\t" \
114 "movl %6, %%edi \n\t" \
115 "movl %7, %%ecx \n\t" \
116 "movl %8, %%ebx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000117
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200118#define MULADDC_CORE \
119 "lodsl \n\t" \
120 "mull %%ebx \n\t" \
121 "addl %%ecx, %%eax \n\t" \
122 "adcl $0, %%edx \n\t" \
123 "addl (%%edi), %%eax \n\t" \
124 "adcl $0, %%edx \n\t" \
125 "movl %%edx, %%ecx \n\t" \
126 "stosl \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000127
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200128#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000129
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200130#define MULADDC_HUIT \
131 "movd %%ecx, %%mm1 \n\t" \
132 "movd %%ebx, %%mm0 \n\t" \
133 "movd (%%edi), %%mm3 \n\t" \
134 "paddq %%mm3, %%mm1 \n\t" \
135 "movd (%%esi), %%mm2 \n\t" \
136 "pmuludq %%mm0, %%mm2 \n\t" \
137 "movd 4(%%esi), %%mm4 \n\t" \
138 "pmuludq %%mm0, %%mm4 \n\t" \
139 "movd 8(%%esi), %%mm6 \n\t" \
140 "pmuludq %%mm0, %%mm6 \n\t" \
141 "movd 12(%%esi), %%mm7 \n\t" \
142 "pmuludq %%mm0, %%mm7 \n\t" \
143 "paddq %%mm2, %%mm1 \n\t" \
144 "movd 4(%%edi), %%mm3 \n\t" \
145 "paddq %%mm4, %%mm3 \n\t" \
146 "movd 8(%%edi), %%mm5 \n\t" \
147 "paddq %%mm6, %%mm5 \n\t" \
148 "movd 12(%%edi), %%mm4 \n\t" \
149 "paddq %%mm4, %%mm7 \n\t" \
150 "movd %%mm1, (%%edi) \n\t" \
151 "movd 16(%%esi), %%mm2 \n\t" \
152 "pmuludq %%mm0, %%mm2 \n\t" \
153 "psrlq $32, %%mm1 \n\t" \
154 "movd 20(%%esi), %%mm4 \n\t" \
155 "pmuludq %%mm0, %%mm4 \n\t" \
156 "paddq %%mm3, %%mm1 \n\t" \
157 "movd 24(%%esi), %%mm6 \n\t" \
158 "pmuludq %%mm0, %%mm6 \n\t" \
159 "movd %%mm1, 4(%%edi) \n\t" \
160 "psrlq $32, %%mm1 \n\t" \
161 "movd 28(%%esi), %%mm3 \n\t" \
162 "pmuludq %%mm0, %%mm3 \n\t" \
163 "paddq %%mm5, %%mm1 \n\t" \
164 "movd 16(%%edi), %%mm5 \n\t" \
165 "paddq %%mm5, %%mm2 \n\t" \
166 "movd %%mm1, 8(%%edi) \n\t" \
167 "psrlq $32, %%mm1 \n\t" \
168 "paddq %%mm7, %%mm1 \n\t" \
169 "movd 20(%%edi), %%mm5 \n\t" \
170 "paddq %%mm5, %%mm4 \n\t" \
171 "movd %%mm1, 12(%%edi) \n\t" \
172 "psrlq $32, %%mm1 \n\t" \
173 "paddq %%mm2, %%mm1 \n\t" \
174 "movd 24(%%edi), %%mm5 \n\t" \
175 "paddq %%mm5, %%mm6 \n\t" \
176 "movd %%mm1, 16(%%edi) \n\t" \
177 "psrlq $32, %%mm1 \n\t" \
178 "paddq %%mm4, %%mm1 \n\t" \
179 "movd 28(%%edi), %%mm5 \n\t" \
180 "paddq %%mm5, %%mm3 \n\t" \
181 "movd %%mm1, 20(%%edi) \n\t" \
182 "psrlq $32, %%mm1 \n\t" \
183 "paddq %%mm6, %%mm1 \n\t" \
184 "movd %%mm1, 24(%%edi) \n\t" \
185 "psrlq $32, %%mm1 \n\t" \
186 "paddq %%mm3, %%mm1 \n\t" \
187 "movd %%mm1, 28(%%edi) \n\t" \
188 "addl $32, %%edi \n\t" \
189 "addl $32, %%esi \n\t" \
190 "psrlq $32, %%mm1 \n\t" \
191 "movd %%mm1, %%ecx \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000192
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200193#define MULADDC_STOP \
194 "emms \n\t" \
195 "movl %4, %%ebx \n\t" \
196 "movl %%ecx, %1 \n\t" \
197 "movl %%edi, %2 \n\t" \
198 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000199 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
200 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
Simon Butcher53571642018-06-24 12:58:31 +0100201 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000202 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000203
204#else
205
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200206#define MULADDC_STOP \
207 "movl %4, %%ebx \n\t" \
208 "movl %%ecx, %1 \n\t" \
209 "movl %%edi, %2 \n\t" \
210 "movl %%esi, %3 \n\t" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000211 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
212 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
Simon Butcher53571642018-06-24 12:58:31 +0100213 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
Paul Bakkerc89cf7c2009-07-19 21:37:39 +0000214 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000215#endif /* SSE2 */
216#endif /* i386 */
217
218#if defined(__amd64__) || defined (__x86_64__)
219
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200220#define MULADDC_INIT \
221 asm( \
Simon Butchera86de142018-09-30 12:09:47 +0100222 "xorq %%r8, %%r8\n"
Paul Bakker5121ce52009-01-03 21:22:43 +0000223
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200224#define MULADDC_CORE \
Simon Butchera86de142018-09-30 12:09:47 +0100225 "movq (%%rsi), %%rax\n" \
226 "mulq %%rbx\n" \
227 "addq $8, %%rsi\n" \
228 "addq %%rcx, %%rax\n" \
229 "movq %%r8, %%rcx\n" \
230 "adcq $0, %%rdx\n" \
231 "nop \n" \
232 "addq %%rax, (%%rdi)\n" \
233 "adcq %%rdx, %%rcx\n" \
234 "addq $8, %%rdi\n"
Paul Bakker5121ce52009-01-03 21:22:43 +0000235
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200236#define MULADDC_STOP \
Gilles Peskinec6d977e2021-09-14 00:13:05 +0200237 : "+c" (c), "+D" (d), "+S" (s), "+m" (*(uint64_t (*)[16]) d) \
238 : "b" (b), "m" (*(const uint64_t (*)[16]) s) \
239 : "rax", "rdx", "r8" \
Manuel Pégourié-Gonnarddef018d2014-01-07 17:50:46 +0100240 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000241
242#endif /* AMD64 */
243
Ko-cc1871e2018-08-16 02:01:57 -0700244#if defined(__aarch64__)
245
246#define MULADDC_INIT \
247 asm(
248
249#define MULADDC_CORE \
Ko-cb260bb2018-08-20 13:59:53 +0100250 "ldr x4, [%2], #8 \n\t" \
251 "ldr x5, [%1] \n\t" \
David Horstmanna23be222021-09-22 18:15:51 +0100252 "mul x6, x4, %4 \n\t" \
253 "umulh x7, x4, %4 \n\t" \
Ko-cc1871e2018-08-16 02:01:57 -0700254 "adds x5, x5, x6 \n\t" \
255 "adc x7, x7, xzr \n\t" \
Ko-cb260bb2018-08-20 13:59:53 +0100256 "adds x5, x5, %0 \n\t" \
Ko-cc1871e2018-08-16 02:01:57 -0700257 "adc %0, x7, xzr \n\t" \
258 "str x5, [%1], #8 \n\t"
259
David Horstmanna23be222021-09-22 18:15:51 +0100260#define MULADDC_STOP \
261 : "+r" (c), "+r" (d), "+r" (s), "+m" (*(uint64_t (*)[16]) d) \
262 : "r" (b), "m" (*(const uint64_t (*)[16]) s) \
263 : "x4", "x5", "x6", "x7", "cc" \
Ko-cc1871e2018-08-16 02:01:57 -0700264 );
265
266#endif /* Aarch64 */
267
Paul Bakker5121ce52009-01-03 21:22:43 +0000268#if defined(__mc68020__) || defined(__mcpu32__)
269
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200270#define MULADDC_INIT \
271 asm( \
272 "movl %3, %%a2 \n\t" \
273 "movl %4, %%a3 \n\t" \
274 "movl %5, %%d3 \n\t" \
275 "movl %6, %%d2 \n\t" \
276 "moveq #0, %%d0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000277
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200278#define MULADDC_CORE \
279 "movel %%a2@+, %%d1 \n\t" \
280 "mulul %%d2, %%d4:%%d1 \n\t" \
281 "addl %%d3, %%d1 \n\t" \
282 "addxl %%d0, %%d4 \n\t" \
283 "moveq #0, %%d3 \n\t" \
284 "addl %%d1, %%a3@+ \n\t" \
285 "addxl %%d4, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000286
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200287#define MULADDC_STOP \
288 "movl %%d3, %0 \n\t" \
289 "movl %%a3, %1 \n\t" \
290 "movl %%a2, %2 \n\t" \
Manuel Pégourié-Gonnard3b05e4c2014-01-10 15:30:23 +0100291 : "=m" (c), "=m" (d), "=m" (s) \
292 : "m" (s), "m" (d), "m" (c), "m" (b) \
293 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
294 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000295
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200296#define MULADDC_HUIT \
297 "movel %%a2@+, %%d1 \n\t" \
298 "mulul %%d2, %%d4:%%d1 \n\t" \
299 "addxl %%d3, %%d1 \n\t" \
300 "addxl %%d0, %%d4 \n\t" \
301 "addl %%d1, %%a3@+ \n\t" \
302 "movel %%a2@+, %%d1 \n\t" \
303 "mulul %%d2, %%d3:%%d1 \n\t" \
304 "addxl %%d4, %%d1 \n\t" \
305 "addxl %%d0, %%d3 \n\t" \
306 "addl %%d1, %%a3@+ \n\t" \
307 "movel %%a2@+, %%d1 \n\t" \
308 "mulul %%d2, %%d4:%%d1 \n\t" \
309 "addxl %%d3, %%d1 \n\t" \
310 "addxl %%d0, %%d4 \n\t" \
311 "addl %%d1, %%a3@+ \n\t" \
312 "movel %%a2@+, %%d1 \n\t" \
313 "mulul %%d2, %%d3:%%d1 \n\t" \
314 "addxl %%d4, %%d1 \n\t" \
315 "addxl %%d0, %%d3 \n\t" \
316 "addl %%d1, %%a3@+ \n\t" \
317 "movel %%a2@+, %%d1 \n\t" \
318 "mulul %%d2, %%d4:%%d1 \n\t" \
319 "addxl %%d3, %%d1 \n\t" \
320 "addxl %%d0, %%d4 \n\t" \
321 "addl %%d1, %%a3@+ \n\t" \
322 "movel %%a2@+, %%d1 \n\t" \
323 "mulul %%d2, %%d3:%%d1 \n\t" \
324 "addxl %%d4, %%d1 \n\t" \
325 "addxl %%d0, %%d3 \n\t" \
326 "addl %%d1, %%a3@+ \n\t" \
327 "movel %%a2@+, %%d1 \n\t" \
328 "mulul %%d2, %%d4:%%d1 \n\t" \
329 "addxl %%d3, %%d1 \n\t" \
330 "addxl %%d0, %%d4 \n\t" \
331 "addl %%d1, %%a3@+ \n\t" \
332 "movel %%a2@+, %%d1 \n\t" \
333 "mulul %%d2, %%d3:%%d1 \n\t" \
334 "addxl %%d4, %%d1 \n\t" \
335 "addxl %%d0, %%d3 \n\t" \
336 "addl %%d1, %%a3@+ \n\t" \
337 "addxl %%d0, %%d3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000338
339#endif /* MC68000 */
340
Paul Bakker5121ce52009-01-03 21:22:43 +0000341#if defined(__powerpc64__) || defined(__ppc64__)
342
343#if defined(__MACH__) && defined(__APPLE__)
344
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200345#define MULADDC_INIT \
346 asm( \
347 "ld r3, %3 \n\t" \
348 "ld r4, %4 \n\t" \
349 "ld r5, %5 \n\t" \
350 "ld r6, %6 \n\t" \
351 "addi r3, r3, -8 \n\t" \
352 "addi r4, r4, -8 \n\t" \
353 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000354
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200355#define MULADDC_CORE \
356 "ldu r7, 8(r3) \n\t" \
357 "mulld r8, r7, r6 \n\t" \
358 "mulhdu r9, r7, r6 \n\t" \
359 "adde r8, r8, r5 \n\t" \
360 "ld r7, 8(r4) \n\t" \
361 "addze r5, r9 \n\t" \
362 "addc r8, r8, r7 \n\t" \
363 "stdu r8, 8(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000364
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200365#define MULADDC_STOP \
366 "addze r5, r5 \n\t" \
367 "addi r4, r4, 8 \n\t" \
368 "addi r3, r3, 8 \n\t" \
369 "std r5, %0 \n\t" \
370 "std r4, %1 \n\t" \
371 "std r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100372 : "=m" (c), "=m" (d), "=m" (s) \
373 : "m" (s), "m" (d), "m" (c), "m" (b) \
374 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
375 );
376
Paul Bakker5121ce52009-01-03 21:22:43 +0000377
Paul Bakker9af723c2014-05-01 13:03:14 +0200378#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000379
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200380#define MULADDC_INIT \
381 asm( \
382 "ld %%r3, %3 \n\t" \
383 "ld %%r4, %4 \n\t" \
384 "ld %%r5, %5 \n\t" \
385 "ld %%r6, %6 \n\t" \
386 "addi %%r3, %%r3, -8 \n\t" \
387 "addi %%r4, %%r4, -8 \n\t" \
388 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000389
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200390#define MULADDC_CORE \
391 "ldu %%r7, 8(%%r3) \n\t" \
392 "mulld %%r8, %%r7, %%r6 \n\t" \
393 "mulhdu %%r9, %%r7, %%r6 \n\t" \
394 "adde %%r8, %%r8, %%r5 \n\t" \
395 "ld %%r7, 8(%%r4) \n\t" \
396 "addze %%r5, %%r9 \n\t" \
397 "addc %%r8, %%r8, %%r7 \n\t" \
398 "stdu %%r8, 8(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000399
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200400#define MULADDC_STOP \
401 "addze %%r5, %%r5 \n\t" \
402 "addi %%r4, %%r4, 8 \n\t" \
403 "addi %%r3, %%r3, 8 \n\t" \
404 "std %%r5, %0 \n\t" \
405 "std %%r4, %1 \n\t" \
406 "std %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100407 : "=m" (c), "=m" (d), "=m" (s) \
408 : "m" (s), "m" (d), "m" (c), "m" (b) \
409 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
410 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000411
Paul Bakker9af723c2014-05-01 13:03:14 +0200412#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000413
Barry K. Nathan35e7cb92014-05-05 23:26:13 -0700414#elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000415
416#if defined(__MACH__) && defined(__APPLE__)
417
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200418#define MULADDC_INIT \
419 asm( \
420 "lwz r3, %3 \n\t" \
421 "lwz r4, %4 \n\t" \
422 "lwz r5, %5 \n\t" \
423 "lwz r6, %6 \n\t" \
424 "addi r3, r3, -4 \n\t" \
425 "addi r4, r4, -4 \n\t" \
426 "addic r5, r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000427
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200428#define MULADDC_CORE \
429 "lwzu r7, 4(r3) \n\t" \
430 "mullw r8, r7, r6 \n\t" \
431 "mulhwu r9, r7, r6 \n\t" \
432 "adde r8, r8, r5 \n\t" \
433 "lwz r7, 4(r4) \n\t" \
434 "addze r5, r9 \n\t" \
435 "addc r8, r8, r7 \n\t" \
436 "stwu r8, 4(r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000437
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200438#define MULADDC_STOP \
439 "addze r5, r5 \n\t" \
440 "addi r4, r4, 4 \n\t" \
441 "addi r3, r3, 4 \n\t" \
442 "stw r5, %0 \n\t" \
443 "stw r4, %1 \n\t" \
444 "stw r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100445 : "=m" (c), "=m" (d), "=m" (s) \
446 : "m" (s), "m" (d), "m" (c), "m" (b) \
447 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
448 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000449
Paul Bakker9af723c2014-05-01 13:03:14 +0200450#else /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000451
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200452#define MULADDC_INIT \
453 asm( \
454 "lwz %%r3, %3 \n\t" \
455 "lwz %%r4, %4 \n\t" \
456 "lwz %%r5, %5 \n\t" \
457 "lwz %%r6, %6 \n\t" \
458 "addi %%r3, %%r3, -4 \n\t" \
459 "addi %%r4, %%r4, -4 \n\t" \
460 "addic %%r5, %%r5, 0 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000461
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200462#define MULADDC_CORE \
463 "lwzu %%r7, 4(%%r3) \n\t" \
464 "mullw %%r8, %%r7, %%r6 \n\t" \
465 "mulhwu %%r9, %%r7, %%r6 \n\t" \
466 "adde %%r8, %%r8, %%r5 \n\t" \
467 "lwz %%r7, 4(%%r4) \n\t" \
468 "addze %%r5, %%r9 \n\t" \
469 "addc %%r8, %%r8, %%r7 \n\t" \
470 "stwu %%r8, 4(%%r4) \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000471
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200472#define MULADDC_STOP \
473 "addze %%r5, %%r5 \n\t" \
474 "addi %%r4, %%r4, 4 \n\t" \
475 "addi %%r3, %%r3, 4 \n\t" \
476 "stw %%r5, %0 \n\t" \
477 "stw %%r4, %1 \n\t" \
478 "stw %%r3, %2 \n\t" \
Manuel Pégourié-Gonnard02d800c2014-01-07 19:16:48 +0100479 : "=m" (c), "=m" (d), "=m" (s) \
480 : "m" (s), "m" (d), "m" (c), "m" (b) \
481 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
482 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000483
Paul Bakker9af723c2014-05-01 13:03:14 +0200484#endif /* __MACH__ && __APPLE__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000485
486#endif /* PPC32 */
Paul Bakker5121ce52009-01-03 21:22:43 +0000487
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200488/*
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200489 * The Sparc(64) assembly is reported to be broken.
Manuel Pégourié-Gonnard31357252014-06-24 17:57:57 +0200490 * Disable it for now, until we're able to fix it.
491 */
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200492#if 0 && defined(__sparc__)
493#if defined(__sparc64__)
Paul Bakker5121ce52009-01-03 21:22:43 +0000494
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200495#define MULADDC_INIT \
496 asm( \
497 "ldx %3, %%o0 \n\t" \
498 "ldx %4, %%o1 \n\t" \
499 "ld %5, %%o2 \n\t" \
500 "ld %6, %%o3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000501
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200502#define MULADDC_CORE \
503 "ld [%%o0], %%o4 \n\t" \
504 "inc 4, %%o0 \n\t" \
505 "ld [%%o1], %%o5 \n\t" \
506 "umul %%o3, %%o4, %%o4 \n\t" \
507 "addcc %%o4, %%o2, %%o4 \n\t" \
508 "rd %%y, %%g1 \n\t" \
509 "addx %%g1, 0, %%g1 \n\t" \
510 "addcc %%o4, %%o5, %%o4 \n\t" \
511 "st %%o4, [%%o1] \n\t" \
512 "addx %%g1, 0, %%o2 \n\t" \
513 "inc 4, %%o1 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000514
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200515 #define MULADDC_STOP \
516 "st %%o2, %0 \n\t" \
517 "stx %%o1, %1 \n\t" \
518 "stx %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000519 : "=m" (c), "=m" (d), "=m" (s) \
520 : "m" (s), "m" (d), "m" (c), "m" (b) \
521 : "g1", "o0", "o1", "o2", "o3", "o4", \
522 "o5" \
523 );
Paul Bakker4f024b72012-10-30 07:29:57 +0000524
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200525#else /* __sparc64__ */
Paul Bakker4f024b72012-10-30 07:29:57 +0000526
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200527#define MULADDC_INIT \
528 asm( \
529 "ld %3, %%o0 \n\t" \
530 "ld %4, %%o1 \n\t" \
531 "ld %5, %%o2 \n\t" \
532 "ld %6, %%o3 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000533
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200534#define MULADDC_CORE \
535 "ld [%%o0], %%o4 \n\t" \
536 "inc 4, %%o0 \n\t" \
537 "ld [%%o1], %%o5 \n\t" \
538 "umul %%o3, %%o4, %%o4 \n\t" \
539 "addcc %%o4, %%o2, %%o4 \n\t" \
540 "rd %%y, %%g1 \n\t" \
541 "addx %%g1, 0, %%g1 \n\t" \
542 "addcc %%o4, %%o5, %%o4 \n\t" \
543 "st %%o4, [%%o1] \n\t" \
544 "addx %%g1, 0, %%o2 \n\t" \
545 "inc 4, %%o1 \n\t"
Paul Bakker4f024b72012-10-30 07:29:57 +0000546
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200547#define MULADDC_STOP \
548 "st %%o2, %0 \n\t" \
549 "st %%o1, %1 \n\t" \
550 "st %%o0, %2 \n\t" \
Paul Bakker4f024b72012-10-30 07:29:57 +0000551 : "=m" (c), "=m" (d), "=m" (s) \
552 : "m" (s), "m" (d), "m" (c), "m" (b) \
553 : "g1", "o0", "o1", "o2", "o3", "o4", \
554 "o5" \
555 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000556
Manuel Pégourié-Gonnard7c5fcdc2015-10-21 14:52:24 +0200557#endif /* __sparc64__ */
558#endif /* __sparc__ */
Paul Bakker5121ce52009-01-03 21:22:43 +0000559
560#if defined(__microblaze__) || defined(microblaze)
561
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200562#define MULADDC_INIT \
563 asm( \
564 "lwi r3, %3 \n\t" \
565 "lwi r4, %4 \n\t" \
566 "lwi r5, %5 \n\t" \
567 "lwi r6, %6 \n\t" \
568 "andi r7, r6, 0xffff \n\t" \
569 "bsrli r6, r6, 16 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000570
Kazuyuki Kimura90b99bf2021-05-31 17:07:28 +0900571#if(__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
572#define MULADDC_LHUI \
573 "lhui r9, r3, 0 \n\t" \
574 "addi r3, r3, 2 \n\t" \
575 "lhui r8, r3, 0 \n\t"
576#else
577#define MULADDC_LHUI \
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200578 "lhui r8, r3, 0 \n\t" \
579 "addi r3, r3, 2 \n\t" \
Kazuyuki Kimura90b99bf2021-05-31 17:07:28 +0900580 "lhui r9, r3, 0 \n\t"
581#endif
582
583#define MULADDC_CORE \
584 MULADDC_LHUI \
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200585 "addi r3, r3, 2 \n\t" \
586 "mul r10, r9, r6 \n\t" \
587 "mul r11, r8, r7 \n\t" \
588 "mul r12, r9, r7 \n\t" \
589 "mul r13, r8, r6 \n\t" \
590 "bsrli r8, r10, 16 \n\t" \
591 "bsrli r9, r11, 16 \n\t" \
592 "add r13, r13, r8 \n\t" \
593 "add r13, r13, r9 \n\t" \
594 "bslli r10, r10, 16 \n\t" \
595 "bslli r11, r11, 16 \n\t" \
596 "add r12, r12, r10 \n\t" \
597 "addc r13, r13, r0 \n\t" \
598 "add r12, r12, r11 \n\t" \
599 "addc r13, r13, r0 \n\t" \
600 "lwi r10, r4, 0 \n\t" \
601 "add r12, r12, r10 \n\t" \
602 "addc r13, r13, r0 \n\t" \
603 "add r12, r12, r5 \n\t" \
604 "addc r5, r13, r0 \n\t" \
605 "swi r12, r4, 0 \n\t" \
606 "addi r4, r4, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000607
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200608#define MULADDC_STOP \
609 "swi r5, %0 \n\t" \
610 "swi r4, %1 \n\t" \
611 "swi r3, %2 \n\t" \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100612 : "=m" (c), "=m" (d), "=m" (s) \
613 : "m" (s), "m" (d), "m" (c), "m" (b) \
Zach van Rijne7d3f8e2018-05-21 10:52:34 -0400614 : "r3", "r4", "r5", "r6", "r7", "r8", \
Manuel Pégourié-Gonnard1753e2f2014-01-10 15:35:41 +0100615 "r9", "r10", "r11", "r12", "r13" \
616 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000617
618#endif /* MicroBlaze */
619
620#if defined(__tricore__)
621
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200622#define MULADDC_INIT \
623 asm( \
624 "ld.a %%a2, %3 \n\t" \
625 "ld.a %%a3, %4 \n\t" \
626 "ld.w %%d4, %5 \n\t" \
627 "ld.w %%d1, %6 \n\t" \
628 "xor %%d5, %%d5 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000629
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200630#define MULADDC_CORE \
631 "ld.w %%d0, [%%a2+] \n\t" \
632 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
633 "ld.w %%d0, [%%a3] \n\t" \
634 "addx %%d2, %%d2, %%d0 \n\t" \
635 "addc %%d3, %%d3, 0 \n\t" \
636 "mov %%d4, %%d3 \n\t" \
637 "st.w [%%a3+], %%d2 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000638
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200639#define MULADDC_STOP \
640 "st.w %0, %%d4 \n\t" \
641 "st.a %1, %%a3 \n\t" \
642 "st.a %2, %%a2 \n\t" \
Manuel Pégourié-Gonnard3f687ad2014-01-10 15:47:50 +0100643 : "=m" (c), "=m" (d), "=m" (s) \
644 : "m" (s), "m" (d), "m" (c), "m" (b) \
645 : "d0", "d1", "e2", "d4", "a2", "a3" \
646 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000647
648#endif /* TriCore */
649
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100650/*
Simon Butcher3ad2efd2018-05-02 14:49:38 +0100651 * Note, gcc -O0 by default uses r7 for the frame pointer, so it complains about
652 * our use of r7 below, unless -fomit-frame-pointer is passed.
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100653 *
654 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
655 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
656 * clang and armcc5 under the same conditions).
657 *
658 * So, only use the optimized assembly below for optimized build, which avoids
659 * the build error and is pretty reasonable anyway.
660 */
661#if defined(__GNUC__) && !defined(__OPTIMIZE__)
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100662#define MULADDC_CANNOT_USE_R7
Manuel Pégourié-Gonnard25caaf32016-01-08 14:29:11 +0100663#endif
664
Manuel Pégourié-Gonnard365f3252016-01-08 14:58:45 +0100665#if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
Paul Bakker5121ce52009-01-03 21:22:43 +0000666
Paul Bakkerfc4f46f2013-06-24 19:23:56 +0200667#if defined(__thumb__) && !defined(__thumb2__)
Dave Rodgman763c8b92023-06-05 21:24:57 -0400668#if !defined(__ARMCC_VERSION) && !defined(__clang__) \
669 && !defined(__llvm__) && !defined(__INTEL_COMPILER)
670/*
671 * Thumb 1 ISA. This code path has only been tested successfully on gcc;
672 * it does not compile on clang or armclang.
673 *
674 * Other compilers which define __GNUC__ may not work. The above macro
675 * attempts to exclude these untested compilers.
676 */
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000677
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200678#define MULADDC_INIT \
679 asm( \
680 "ldr r0, %3 \n\t" \
681 "ldr r1, %4 \n\t" \
682 "ldr r2, %5 \n\t" \
683 "ldr r3, %6 \n\t" \
684 "lsr r7, r3, #16 \n\t" \
685 "mov r9, r7 \n\t" \
686 "lsl r7, r3, #16 \n\t" \
687 "lsr r7, r7, #16 \n\t" \
688 "mov r8, r7 \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000689
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200690#define MULADDC_CORE \
691 "ldmia r0!, {r6} \n\t" \
692 "lsr r7, r6, #16 \n\t" \
693 "lsl r6, r6, #16 \n\t" \
694 "lsr r6, r6, #16 \n\t" \
695 "mov r4, r8 \n\t" \
696 "mul r4, r6 \n\t" \
697 "mov r3, r9 \n\t" \
698 "mul r6, r3 \n\t" \
699 "mov r5, r9 \n\t" \
700 "mul r5, r7 \n\t" \
701 "mov r3, r8 \n\t" \
702 "mul r7, r3 \n\t" \
703 "lsr r3, r6, #16 \n\t" \
704 "add r5, r5, r3 \n\t" \
705 "lsr r3, r7, #16 \n\t" \
706 "add r5, r5, r3 \n\t" \
707 "add r4, r4, r2 \n\t" \
708 "mov r2, #0 \n\t" \
709 "adc r5, r2 \n\t" \
710 "lsl r3, r6, #16 \n\t" \
711 "add r4, r4, r3 \n\t" \
712 "adc r5, r2 \n\t" \
713 "lsl r3, r7, #16 \n\t" \
714 "add r4, r4, r3 \n\t" \
715 "adc r5, r2 \n\t" \
716 "ldr r3, [r1] \n\t" \
717 "add r4, r4, r3 \n\t" \
718 "adc r2, r5 \n\t" \
719 "stmia r1!, {r4} \n\t"
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000720
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200721#define MULADDC_STOP \
722 "str r2, %0 \n\t" \
723 "str r1, %1 \n\t" \
724 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100725 : "=m" (c), "=m" (d), "=m" (s) \
726 : "m" (s), "m" (d), "m" (c), "m" (b) \
727 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200728 "r6", "r7", "r8", "r9", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100729 );
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000730
Dave Rodgman763c8b92023-06-05 21:24:57 -0400731#endif /* Compiler is gcc */
732
Aurelien Jarno5daa34f2018-11-03 00:46:06 +0100733#elif (__ARM_ARCH >= 6) && \
734 defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)
Aurelien Jarno16b1bd82018-05-21 22:01:21 +0200735
736#define MULADDC_INIT \
737 asm(
738
739#define MULADDC_CORE \
740 "ldr r0, [%0], #4 \n\t" \
741 "ldr r1, [%1] \n\t" \
742 "umaal r1, %2, %3, r0 \n\t" \
743 "str r1, [%1], #4 \n\t"
744
745#define MULADDC_STOP \
746 : "=r" (s), "=r" (d), "=r" (c) \
747 : "r" (b), "0" (s), "1" (d), "2" (c) \
748 : "r0", "r1", "memory" \
749 );
750
Paul Bakker4f9a7bb2012-07-02 08:36:36 +0000751#else
Paul Bakkera2713a32011-11-18 12:47:23 +0000752
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200753#define MULADDC_INIT \
754 asm( \
755 "ldr r0, %3 \n\t" \
756 "ldr r1, %4 \n\t" \
757 "ldr r2, %5 \n\t" \
758 "ldr r3, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000759
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200760#define MULADDC_CORE \
761 "ldr r4, [r0], #4 \n\t" \
762 "mov r5, #0 \n\t" \
763 "ldr r6, [r1] \n\t" \
764 "umlal r2, r5, r3, r4 \n\t" \
765 "adds r7, r6, r2 \n\t" \
766 "adc r2, r5, #0 \n\t" \
767 "str r7, [r1], #4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000768
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200769#define MULADDC_STOP \
770 "str r2, %0 \n\t" \
771 "str r1, %1 \n\t" \
772 "str r0, %2 \n\t" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100773 : "=m" (c), "=m" (d), "=m" (s) \
774 : "m" (s), "m" (d), "m" (c), "m" (b) \
775 : "r0", "r1", "r2", "r3", "r4", "r5", \
Paul Bakkereff2e6d2013-04-11 17:13:22 +0200776 "r6", "r7", "cc" \
Paul Bakkerfb1cbd32013-03-06 18:14:52 +0100777 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000778
Paul Bakkera2713a32011-11-18 12:47:23 +0000779#endif /* Thumb */
780
Paul Bakker5121ce52009-01-03 21:22:43 +0000781#endif /* ARMv3 */
782
783#if defined(__alpha__)
784
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200785#define MULADDC_INIT \
786 asm( \
787 "ldq $1, %3 \n\t" \
788 "ldq $2, %4 \n\t" \
789 "ldq $3, %5 \n\t" \
790 "ldq $4, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000791
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200792#define MULADDC_CORE \
793 "ldq $6, 0($1) \n\t" \
794 "addq $1, 8, $1 \n\t" \
795 "mulq $6, $4, $7 \n\t" \
796 "umulh $6, $4, $6 \n\t" \
797 "addq $7, $3, $7 \n\t" \
798 "cmpult $7, $3, $3 \n\t" \
799 "ldq $5, 0($2) \n\t" \
800 "addq $7, $5, $7 \n\t" \
801 "cmpult $7, $5, $5 \n\t" \
802 "stq $7, 0($2) \n\t" \
803 "addq $2, 8, $2 \n\t" \
804 "addq $6, $3, $3 \n\t" \
805 "addq $5, $3, $3 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000806
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200807#define MULADDC_STOP \
808 "stq $3, %0 \n\t" \
809 "stq $2, %1 \n\t" \
810 "stq $1, %2 \n\t" \
Manuel Pégourié-Gonnard5af8e642014-01-10 15:53:41 +0100811 : "=m" (c), "=m" (d), "=m" (s) \
812 : "m" (s), "m" (d), "m" (c), "m" (b) \
813 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
814 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000815#endif /* Alpha */
816
James Cowgilld1e7e8b2014-12-16 15:24:06 +0000817#if defined(__mips__) && !defined(__mips64)
Paul Bakker5121ce52009-01-03 21:22:43 +0000818
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200819#define MULADDC_INIT \
820 asm( \
821 "lw $10, %3 \n\t" \
822 "lw $11, %4 \n\t" \
823 "lw $12, %5 \n\t" \
824 "lw $13, %6 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000825
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200826#define MULADDC_CORE \
827 "lw $14, 0($10) \n\t" \
828 "multu $13, $14 \n\t" \
829 "addi $10, $10, 4 \n\t" \
830 "mflo $14 \n\t" \
831 "mfhi $9 \n\t" \
832 "addu $14, $12, $14 \n\t" \
833 "lw $15, 0($11) \n\t" \
834 "sltu $12, $14, $12 \n\t" \
835 "addu $15, $14, $15 \n\t" \
836 "sltu $14, $15, $14 \n\t" \
837 "addu $12, $12, $9 \n\t" \
838 "sw $15, 0($11) \n\t" \
839 "addu $12, $12, $14 \n\t" \
840 "addi $11, $11, 4 \n\t"
Paul Bakker5121ce52009-01-03 21:22:43 +0000841
Manuel Pégourié-Gonnardacbcbba2014-06-19 17:20:43 +0200842#define MULADDC_STOP \
843 "sw $12, %0 \n\t" \
844 "sw $11, %1 \n\t" \
845 "sw $10, %2 \n\t" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100846 : "=m" (c), "=m" (d), "=m" (s) \
847 : "m" (s), "m" (d), "m" (c), "m" (b) \
Jeffrey Martind25fd8d2019-01-14 18:01:40 -0600848 : "$9", "$10", "$11", "$12", "$13", "$14", "$15", "lo", "hi" \
Manuel Pégourié-Gonnard8b1b1032014-01-07 18:31:06 +0100849 );
Paul Bakker5121ce52009-01-03 21:22:43 +0000850
851#endif /* MIPS */
852#endif /* GNUC */
853
854#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
855
856#define MULADDC_INIT \
857 __asm mov esi, s \
858 __asm mov edi, d \
859 __asm mov ecx, c \
860 __asm mov ebx, b
861
862#define MULADDC_CORE \
863 __asm lodsd \
864 __asm mul ebx \
865 __asm add eax, ecx \
866 __asm adc edx, 0 \
867 __asm add eax, [edi] \
868 __asm adc edx, 0 \
869 __asm mov ecx, edx \
870 __asm stosd
871
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200872#if defined(MBEDTLS_HAVE_SSE2)
Paul Bakker5121ce52009-01-03 21:22:43 +0000873
874#define EMIT __asm _emit
875
876#define MULADDC_HUIT \
877 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
878 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
879 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
880 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
881 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
882 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
883 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
884 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
885 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
886 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
887 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
888 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
889 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
890 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
891 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
892 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
893 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
894 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
895 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
896 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
897 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
898 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
899 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
900 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
901 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
902 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
903 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
904 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
905 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
906 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
907 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
908 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
909 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
910 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
911 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
912 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
913 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
914 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
915 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
916 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
917 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
918 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
919 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
920 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
921 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
922 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
923 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
924 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
925 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
926 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
927 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
928 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
929 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
930 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
931 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
932 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
933 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
934 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
935 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
936 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
937 EMIT 0x0F EMIT 0x7E EMIT 0xC9
938
939#define MULADDC_STOP \
940 EMIT 0x0F EMIT 0x77 \
941 __asm mov c, ecx \
942 __asm mov d, edi \
943 __asm mov s, esi \
944
945#else
946
947#define MULADDC_STOP \
948 __asm mov c, ecx \
949 __asm mov d, edi \
950 __asm mov s, esi \
951
952#endif /* SSE2 */
953#endif /* MSVC */
954
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200955#endif /* MBEDTLS_HAVE_ASM */
Paul Bakker5121ce52009-01-03 21:22:43 +0000956
957#if !defined(MULADDC_CORE)
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200958#if defined(MBEDTLS_HAVE_UDBL)
Paul Bakker5121ce52009-01-03 21:22:43 +0000959
960#define MULADDC_INIT \
961{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200962 mbedtls_t_udbl r; \
963 mbedtls_mpi_uint r0, r1;
Paul Bakker5121ce52009-01-03 21:22:43 +0000964
965#define MULADDC_CORE \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200966 r = *(s++) * (mbedtls_t_udbl) b; \
967 r0 = (mbedtls_mpi_uint) r; \
968 r1 = (mbedtls_mpi_uint)( r >> biL ); \
Paul Bakker5121ce52009-01-03 21:22:43 +0000969 r0 += c; r1 += (r0 < c); \
970 r0 += *d; r1 += (r0 < *d); \
971 c = r1; *(d++) = r0;
972
973#define MULADDC_STOP \
974}
975
976#else
977#define MULADDC_INIT \
978{ \
Manuel Pégourié-Gonnard2cf5a7c2015-04-08 12:49:31 +0200979 mbedtls_mpi_uint s0, s1, b0, b1; \
980 mbedtls_mpi_uint r0, r1, rx, ry; \
Paul Bakker5121ce52009-01-03 21:22:43 +0000981 b0 = ( b << biH ) >> biH; \
982 b1 = ( b >> biH );
983
984#define MULADDC_CORE \
985 s0 = ( *s << biH ) >> biH; \
986 s1 = ( *s >> biH ); s++; \
987 rx = s0 * b1; r0 = s0 * b0; \
988 ry = s1 * b0; r1 = s1 * b1; \
989 r1 += ( rx >> biH ); \
990 r1 += ( ry >> biH ); \
991 rx <<= biH; ry <<= biH; \
992 r0 += rx; r1 += (r0 < rx); \
993 r0 += ry; r1 += (r0 < ry); \
994 r0 += c; r1 += (r0 < c); \
995 r0 += *d; r1 += (r0 < *d); \
996 c = r1; *(d++) = r0;
997
998#define MULADDC_STOP \
999}
1000
1001#endif /* C (generic) */
1002#endif /* C (longlong) */
1003
David Horstmann2ccd77a2023-01-05 09:50:47 +00001004/* *INDENT-ON* */
Paul Bakker5121ce52009-01-03 21:22:43 +00001005#endif /* bn_mul.h */