mbed TLS v2.28.2
bn_mul.h
Go to the documentation of this file.
1
6/*
7 * Copyright The Mbed TLS Contributors
8 * SPDX-License-Identifier: Apache-2.0
9 *
10 * Licensed under the Apache License, Version 2.0 (the "License"); you may
11 * not use this file except in compliance with the License.
12 * You may obtain a copy of the License at
13 *
14 * http://www.apache.org/licenses/LICENSE-2.0
15 *
16 * Unless required by applicable law or agreed to in writing, software
17 * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
18 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
19 * See the License for the specific language governing permissions and
20 * limitations under the License.
21 */
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 */
36#ifndef MBEDTLS_BN_MUL_H
37#define MBEDTLS_BN_MUL_H
38
39#if !defined(MBEDTLS_CONFIG_FILE)
40#include "mbedtls/config.h"
41#else
42#include MBEDTLS_CONFIG_FILE
43#endif
44
45#include "mbedtls/bignum.h"
46
47
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
54#define MBEDTLS_BYTES_TO_T_UINT_4( a, b, c, d ) \
55 ( (mbedtls_mpi_uint) (a) << 0 ) | \
56 ( (mbedtls_mpi_uint) (b) << 8 ) | \
57 ( (mbedtls_mpi_uint) (c) << 16 ) | \
58 ( (mbedtls_mpi_uint) (d) << 24 )
59
60#define MBEDTLS_BYTES_TO_T_UINT_2( a, b ) \
61 MBEDTLS_BYTES_TO_T_UINT_4( a, b, 0, 0 )
62
63#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 )
66
67#else /* 64-bits */
68
69#define MBEDTLS_BYTES_TO_T_UINT_8( a, b, c, d, e, f, g, h ) \
70 ( (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
79#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 )
81
82#define MBEDTLS_BYTES_TO_T_UINT_2( a, b ) \
83 MBEDTLS_BYTES_TO_T_UINT_8( a, b, 0, 0, 0, 0, 0, 0 )
84
85#endif /* bits in mbedtls_mpi_uint */
86
87#if defined(MBEDTLS_HAVE_ASM)
88
89#ifndef asm
90#define asm __asm
91#endif
92
93/* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */
94#if defined(__GNUC__) && \
95 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
96
97/*
98 * 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/*
114 * 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 */
119#if defined(__i386__) && defined(__OPTIMIZE__) && !defined(MULADDC_CANNOT_USE_EBX)
120
121#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"
128
129#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"
138
139#if defined(MBEDTLS_HAVE_SSE2)
140
141#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"
203
204#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" \
210 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
211 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
212 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
213 );
214
215#else
216
217#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" \
222 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
223 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
224 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
225 );
226#endif /* SSE2 */
227#endif /* i386 */
228
229#if defined(__amd64__) || defined (__x86_64__)
230
231#define MULADDC_INIT \
232 asm( \
233 "xorq %%r8, %%r8\n"
234
235#define MULADDC_CORE \
236 "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"
246
247#define MULADDC_STOP \
248 : "+c" (c), "+D" (d), "+S" (s), "+m" (*(uint64_t (*)[16]) d) \
249 : "b" (b), "m" (*(const uint64_t (*)[16]) s) \
250 : "rax", "rdx", "r8" \
251 );
252
253#endif /* AMD64 */
254
255#if defined(__aarch64__)
256
257#define MULADDC_INIT \
258 asm(
259
260#define MULADDC_CORE \
261 "ldr x4, [%2], #8 \n\t" \
262 "ldr x5, [%1] \n\t" \
263 "mul x6, x4, %4 \n\t" \
264 "umulh x7, x4, %4 \n\t" \
265 "adds x5, x5, x6 \n\t" \
266 "adc x7, x7, xzr \n\t" \
267 "adds x5, x5, %0 \n\t" \
268 "adc %0, x7, xzr \n\t" \
269 "str x5, [%1], #8 \n\t"
270
271#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" \
275 );
276
277#endif /* Aarch64 */
278
279#if defined(__mc68020__) || defined(__mcpu32__)
280
281#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"
288
289#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"
297
298#define MULADDC_STOP \
299 "movl %%d3, %0 \n\t" \
300 "movl %%a3, %1 \n\t" \
301 "movl %%a2, %2 \n\t" \
302 : "=m" (c), "=m" (d), "=m" (s) \
303 : "m" (s), "m" (d), "m" (c), "m" (b) \
304 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
305 );
306
307#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"
349
350#endif /* MC68000 */
351
352#if defined(__powerpc64__) || defined(__ppc64__)
353
354#if defined(__MACH__) && defined(__APPLE__)
355
356#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"
365
366#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"
375
376#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" \
383 : "=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
388
389#else /* __MACH__ && __APPLE__ */
390
391#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"
400
401#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"
410
411#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" \
418 : "=m" (c), "=m" (d), "=m" (s) \
419 : "m" (s), "m" (d), "m" (c), "m" (b) \
420 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
421 );
422
423#endif /* __MACH__ && __APPLE__ */
424
425#elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
426
427#if defined(__MACH__) && defined(__APPLE__)
428
429#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"
438
439#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"
448
449#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" \
456 : "=m" (c), "=m" (d), "=m" (s) \
457 : "m" (s), "m" (d), "m" (c), "m" (b) \
458 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
459 );
460
461#else /* __MACH__ && __APPLE__ */
462
463#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"
472
473#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"
482
483#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" \
490 : "=m" (c), "=m" (d), "=m" (s) \
491 : "m" (s), "m" (d), "m" (c), "m" (b) \
492 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
493 );
494
495#endif /* __MACH__ && __APPLE__ */
496
497#endif /* PPC32 */
498
499/*
500 * The Sparc(64) assembly is reported to be broken.
501 * Disable it for now, until we're able to fix it.
502 */
503#if 0 && defined(__sparc__)
504#if defined(__sparc64__)
505
506#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"
512
513#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"
525
526 #define MULADDC_STOP \
527 "st %%o2, %0 \n\t" \
528 "stx %%o1, %1 \n\t" \
529 "stx %%o0, %2 \n\t" \
530 : "=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 );
535
536#else /* __sparc64__ */
537
538#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"
544
545#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"
557
558#define MULADDC_STOP \
559 "st %%o2, %0 \n\t" \
560 "st %%o1, %1 \n\t" \
561 "st %%o0, %2 \n\t" \
562 : "=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 );
567
568#endif /* __sparc64__ */
569#endif /* __sparc__ */
570
571#if defined(__microblaze__) || defined(microblaze)
572
573#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"
581
582#if(__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
583#define MULADDC_LHUI \
584 "lhui r9, r3, 0 \n\t" \
585 "addi r3, r3, 2 \n\t" \
586 "lhui r8, r3, 0 \n\t"
587#else
588#define MULADDC_LHUI \
589 "lhui r8, r3, 0 \n\t" \
590 "addi r3, r3, 2 \n\t" \
591 "lhui r9, r3, 0 \n\t"
592#endif
593
594#define MULADDC_CORE \
595 MULADDC_LHUI \
596 "addi r3, r3, 2 \n\t" \
597 "mul r10, r9, r6 \n\t" \
598 "mul r11, r8, r7 \n\t" \
599 "mul r12, r9, r7 \n\t" \
600 "mul r13, r8, r6 \n\t" \
601 "bsrli r8, r10, 16 \n\t" \
602 "bsrli r9, r11, 16 \n\t" \
603 "add r13, r13, r8 \n\t" \
604 "add r13, r13, r9 \n\t" \
605 "bslli r10, r10, 16 \n\t" \
606 "bslli r11, r11, 16 \n\t" \
607 "add r12, r12, r10 \n\t" \
608 "addc r13, r13, r0 \n\t" \
609 "add r12, r12, r11 \n\t" \
610 "addc r13, r13, r0 \n\t" \
611 "lwi r10, r4, 0 \n\t" \
612 "add r12, r12, r10 \n\t" \
613 "addc r13, r13, r0 \n\t" \
614 "add r12, r12, r5 \n\t" \
615 "addc r5, r13, r0 \n\t" \
616 "swi r12, r4, 0 \n\t" \
617 "addi r4, r4, 4 \n\t"
618
619#define MULADDC_STOP \
620 "swi r5, %0 \n\t" \
621 "swi r4, %1 \n\t" \
622 "swi r3, %2 \n\t" \
623 : "=m" (c), "=m" (d), "=m" (s) \
624 : "m" (s), "m" (d), "m" (c), "m" (b) \
625 : "r3", "r4", "r5", "r6", "r7", "r8", \
626 "r9", "r10", "r11", "r12", "r13" \
627 );
628
629#endif /* MicroBlaze */
630
631#if defined(__tricore__)
632
633#define MULADDC_INIT \
634 asm( \
635 "ld.a %%a2, %3 \n\t" \
636 "ld.a %%a3, %4 \n\t" \
637 "ld.w %%d4, %5 \n\t" \
638 "ld.w %%d1, %6 \n\t" \
639 "xor %%d5, %%d5 \n\t"
640
641#define MULADDC_CORE \
642 "ld.w %%d0, [%%a2+] \n\t" \
643 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
644 "ld.w %%d0, [%%a3] \n\t" \
645 "addx %%d2, %%d2, %%d0 \n\t" \
646 "addc %%d3, %%d3, 0 \n\t" \
647 "mov %%d4, %%d3 \n\t" \
648 "st.w [%%a3+], %%d2 \n\t"
649
650#define MULADDC_STOP \
651 "st.w %0, %%d4 \n\t" \
652 "st.a %1, %%a3 \n\t" \
653 "st.a %2, %%a2 \n\t" \
654 : "=m" (c), "=m" (d), "=m" (s) \
655 : "m" (s), "m" (d), "m" (c), "m" (b) \
656 : "d0", "d1", "e2", "d4", "a2", "a3" \
657 );
658
659#endif /* TriCore */
660
661/*
662 * Note, gcc -O0 by default uses r7 for the frame pointer, so it complains about
663 * our use of r7 below, unless -fomit-frame-pointer is passed.
664 *
665 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
666 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
667 * clang and armcc5 under the same conditions).
668 *
669 * So, only use the optimized assembly below for optimized build, which avoids
670 * the build error and is pretty reasonable anyway.
671 */
672#if defined(__GNUC__) && !defined(__OPTIMIZE__)
673#define MULADDC_CANNOT_USE_R7
674#endif
675
676#if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
677
678#if defined(__thumb__) && !defined(__thumb2__)
679
680#define MULADDC_INIT \
681 asm( \
682 "ldr r0, %3 \n\t" \
683 "ldr r1, %4 \n\t" \
684 "ldr r2, %5 \n\t" \
685 "ldr r3, %6 \n\t" \
686 "lsr r7, r3, #16 \n\t" \
687 "mov r9, r7 \n\t" \
688 "lsl r7, r3, #16 \n\t" \
689 "lsr r7, r7, #16 \n\t" \
690 "mov r8, r7 \n\t"
691
692#define MULADDC_CORE \
693 "ldmia r0!, {r6} \n\t" \
694 "lsr r7, r6, #16 \n\t" \
695 "lsl r6, r6, #16 \n\t" \
696 "lsr r6, r6, #16 \n\t" \
697 "mov r4, r8 \n\t" \
698 "mul r4, r6 \n\t" \
699 "mov r3, r9 \n\t" \
700 "mul r6, r3 \n\t" \
701 "mov r5, r9 \n\t" \
702 "mul r5, r7 \n\t" \
703 "mov r3, r8 \n\t" \
704 "mul r7, r3 \n\t" \
705 "lsr r3, r6, #16 \n\t" \
706 "add r5, r5, r3 \n\t" \
707 "lsr r3, r7, #16 \n\t" \
708 "add r5, r5, r3 \n\t" \
709 "add r4, r4, r2 \n\t" \
710 "mov r2, #0 \n\t" \
711 "adc r5, r2 \n\t" \
712 "lsl r3, r6, #16 \n\t" \
713 "add r4, r4, r3 \n\t" \
714 "adc r5, r2 \n\t" \
715 "lsl r3, r7, #16 \n\t" \
716 "add r4, r4, r3 \n\t" \
717 "adc r5, r2 \n\t" \
718 "ldr r3, [r1] \n\t" \
719 "add r4, r4, r3 \n\t" \
720 "adc r2, r5 \n\t" \
721 "stmia r1!, {r4} \n\t"
722
723#define MULADDC_STOP \
724 "str r2, %0 \n\t" \
725 "str r1, %1 \n\t" \
726 "str r0, %2 \n\t" \
727 : "=m" (c), "=m" (d), "=m" (s) \
728 : "m" (s), "m" (d), "m" (c), "m" (b) \
729 : "r0", "r1", "r2", "r3", "r4", "r5", \
730 "r6", "r7", "r8", "r9", "cc" \
731 );
732
733#elif (__ARM_ARCH >= 6) && \
734 defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)
735
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
751#else
752
753#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"
759
760#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"
768
769#define MULADDC_STOP \
770 "str r2, %0 \n\t" \
771 "str r1, %1 \n\t" \
772 "str r0, %2 \n\t" \
773 : "=m" (c), "=m" (d), "=m" (s) \
774 : "m" (s), "m" (d), "m" (c), "m" (b) \
775 : "r0", "r1", "r2", "r3", "r4", "r5", \
776 "r6", "r7", "cc" \
777 );
778
779#endif /* Thumb */
780
781#endif /* ARMv3 */
782
783#if defined(__alpha__)
784
785#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"
791
792#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"
806
807#define MULADDC_STOP \
808 "stq $3, %0 \n\t" \
809 "stq $2, %1 \n\t" \
810 "stq $1, %2 \n\t" \
811 : "=m" (c), "=m" (d), "=m" (s) \
812 : "m" (s), "m" (d), "m" (c), "m" (b) \
813 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
814 );
815#endif /* Alpha */
816
817#if defined(__mips__) && !defined(__mips64)
818
819#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"
825
826#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"
841
842#define MULADDC_STOP \
843 "sw $12, %0 \n\t" \
844 "sw $11, %1 \n\t" \
845 "sw $10, %2 \n\t" \
846 : "=m" (c), "=m" (d), "=m" (s) \
847 : "m" (s), "m" (d), "m" (c), "m" (b) \
848 : "$9", "$10", "$11", "$12", "$13", "$14", "$15", "lo", "hi" \
849 );
850
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
872#if defined(MBEDTLS_HAVE_SSE2)
873
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
955#endif /* MBEDTLS_HAVE_ASM */
956
957#if !defined(MULADDC_CORE)
958#if defined(MBEDTLS_HAVE_UDBL)
959
960#define MULADDC_INIT \
961{ \
962 mbedtls_t_udbl r; \
963 mbedtls_mpi_uint r0, r1;
964
965#define MULADDC_CORE \
966 r = *(s++) * (mbedtls_t_udbl) b; \
967 r0 = (mbedtls_mpi_uint) r; \
968 r1 = (mbedtls_mpi_uint)( r >> biL ); \
969 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{ \
979 mbedtls_mpi_uint s0, s1, b0, b1; \
980 mbedtls_mpi_uint r0, r1, rx, ry; \
981 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
1004#endif /* bn_mul.h */
Multi-precision integer library.
Configuration options (set of defines)