mbed TLS v2.14.1
bn_mul.h
Go to the documentation of this file.
1 
6 /*
7  * Copyright (C) 2006-2015, ARM Limited, All Rights Reserved
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  * This file is part of mbed TLS (https://tls.mbed.org)
23  */
24 /*
25  * Multiply source vector [s] with b, add result
26  * to destination vector [d] and set carry c.
27  *
28  * Currently supports:
29  *
30  * . IA-32 (386+) . AMD64 / EM64T
31  * . IA-32 (SSE2) . Motorola 68000
32  * . PowerPC, 32-bit . MicroBlaze
33  * . PowerPC, 64-bit . TriCore
34  * . SPARC v8 . ARM v3+
35  * . Alpha . MIPS32
36  * . C, longlong . C, generic
37  */
38 #ifndef MBEDTLS_BN_MUL_H
39 #define MBEDTLS_BN_MUL_H
40 
41 #include "bignum.h"
42 
43 #if defined(MBEDTLS_HAVE_ASM)
44 
45 #ifndef asm
46 #define asm __asm
47 #endif
48 
49 /* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */
50 #if defined(__GNUC__) && \
51  ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
52 
53 /*
54  * GCC < 5.0 treated the x86 ebx (which is used for the GOT) as a
55  * fixed reserved register when building as PIC, leading to errors
56  * like: bn_mul.h:46:13: error: PIC register clobbered by ‘ebx’ in ‘asm’
57  *
58  * This is fixed by an improved register allocator in GCC 5+. From the
59  * release notes:
60  * Register allocation improvements: Reuse of the PIC hard register,
61  * instead of using a fixed register, was implemented on x86/x86-64
62  * targets. This improves generated PIC code performance as more hard
63  * registers can be used.
64  */
65 #if defined(__GNUC__) && __GNUC__ < 5 && defined(__PIC__)
66 #define MULADDC_CANNOT_USE_EBX
67 #endif
68 
69 /*
70  * Disable use of the i386 assembly code below if option -O0, to disable all
71  * compiler optimisations, is passed, detected with __OPTIMIZE__
72  * This is done as the number of registers used in the assembly code doesn't
73  * work with the -O0 option.
74  */
75 #if defined(__i386__) && defined(__OPTIMIZE__) && !defined(MULADDC_CANNOT_USE_EBX)
76 
77 #define MULADDC_INIT \
78  asm( \
79  "movl %%ebx, %0 \n\t" \
80  "movl %5, %%esi \n\t" \
81  "movl %6, %%edi \n\t" \
82  "movl %7, %%ecx \n\t" \
83  "movl %8, %%ebx \n\t"
84 
85 #define MULADDC_CORE \
86  "lodsl \n\t" \
87  "mull %%ebx \n\t" \
88  "addl %%ecx, %%eax \n\t" \
89  "adcl $0, %%edx \n\t" \
90  "addl (%%edi), %%eax \n\t" \
91  "adcl $0, %%edx \n\t" \
92  "movl %%edx, %%ecx \n\t" \
93  "stosl \n\t"
94 
95 #if defined(MBEDTLS_HAVE_SSE2)
96 
97 #define MULADDC_HUIT \
98  "movd %%ecx, %%mm1 \n\t" \
99  "movd %%ebx, %%mm0 \n\t" \
100  "movd (%%edi), %%mm3 \n\t" \
101  "paddq %%mm3, %%mm1 \n\t" \
102  "movd (%%esi), %%mm2 \n\t" \
103  "pmuludq %%mm0, %%mm2 \n\t" \
104  "movd 4(%%esi), %%mm4 \n\t" \
105  "pmuludq %%mm0, %%mm4 \n\t" \
106  "movd 8(%%esi), %%mm6 \n\t" \
107  "pmuludq %%mm0, %%mm6 \n\t" \
108  "movd 12(%%esi), %%mm7 \n\t" \
109  "pmuludq %%mm0, %%mm7 \n\t" \
110  "paddq %%mm2, %%mm1 \n\t" \
111  "movd 4(%%edi), %%mm3 \n\t" \
112  "paddq %%mm4, %%mm3 \n\t" \
113  "movd 8(%%edi), %%mm5 \n\t" \
114  "paddq %%mm6, %%mm5 \n\t" \
115  "movd 12(%%edi), %%mm4 \n\t" \
116  "paddq %%mm4, %%mm7 \n\t" \
117  "movd %%mm1, (%%edi) \n\t" \
118  "movd 16(%%esi), %%mm2 \n\t" \
119  "pmuludq %%mm0, %%mm2 \n\t" \
120  "psrlq $32, %%mm1 \n\t" \
121  "movd 20(%%esi), %%mm4 \n\t" \
122  "pmuludq %%mm0, %%mm4 \n\t" \
123  "paddq %%mm3, %%mm1 \n\t" \
124  "movd 24(%%esi), %%mm6 \n\t" \
125  "pmuludq %%mm0, %%mm6 \n\t" \
126  "movd %%mm1, 4(%%edi) \n\t" \
127  "psrlq $32, %%mm1 \n\t" \
128  "movd 28(%%esi), %%mm3 \n\t" \
129  "pmuludq %%mm0, %%mm3 \n\t" \
130  "paddq %%mm5, %%mm1 \n\t" \
131  "movd 16(%%edi), %%mm5 \n\t" \
132  "paddq %%mm5, %%mm2 \n\t" \
133  "movd %%mm1, 8(%%edi) \n\t" \
134  "psrlq $32, %%mm1 \n\t" \
135  "paddq %%mm7, %%mm1 \n\t" \
136  "movd 20(%%edi), %%mm5 \n\t" \
137  "paddq %%mm5, %%mm4 \n\t" \
138  "movd %%mm1, 12(%%edi) \n\t" \
139  "psrlq $32, %%mm1 \n\t" \
140  "paddq %%mm2, %%mm1 \n\t" \
141  "movd 24(%%edi), %%mm5 \n\t" \
142  "paddq %%mm5, %%mm6 \n\t" \
143  "movd %%mm1, 16(%%edi) \n\t" \
144  "psrlq $32, %%mm1 \n\t" \
145  "paddq %%mm4, %%mm1 \n\t" \
146  "movd 28(%%edi), %%mm5 \n\t" \
147  "paddq %%mm5, %%mm3 \n\t" \
148  "movd %%mm1, 20(%%edi) \n\t" \
149  "psrlq $32, %%mm1 \n\t" \
150  "paddq %%mm6, %%mm1 \n\t" \
151  "movd %%mm1, 24(%%edi) \n\t" \
152  "psrlq $32, %%mm1 \n\t" \
153  "paddq %%mm3, %%mm1 \n\t" \
154  "movd %%mm1, 28(%%edi) \n\t" \
155  "addl $32, %%edi \n\t" \
156  "addl $32, %%esi \n\t" \
157  "psrlq $32, %%mm1 \n\t" \
158  "movd %%mm1, %%ecx \n\t"
159 
160 #define MULADDC_STOP \
161  "emms \n\t" \
162  "movl %4, %%ebx \n\t" \
163  "movl %%ecx, %1 \n\t" \
164  "movl %%edi, %2 \n\t" \
165  "movl %%esi, %3 \n\t" \
166  : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
167  : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
168  : "eax", "ebx", "ecx", "edx", "esi", "edi" \
169  );
170 
171 #else
172 
173 #define MULADDC_STOP \
174  "movl %4, %%ebx \n\t" \
175  "movl %%ecx, %1 \n\t" \
176  "movl %%edi, %2 \n\t" \
177  "movl %%esi, %3 \n\t" \
178  : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
179  : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
180  : "eax", "ebx", "ecx", "edx", "esi", "edi" \
181  );
182 #endif /* SSE2 */
183 #endif /* i386 */
184 
185 #if defined(__amd64__) || defined (__x86_64__)
186 
187 #define MULADDC_INIT \
188  asm( \
189  "xorq %%r8, %%r8 \n\t"
190 
191 #define MULADDC_CORE \
192  "movq (%%rsi), %%rax \n\t" \
193  "mulq %%rbx \n\t" \
194  "addq $8, %%rsi \n\t" \
195  "addq %%rcx, %%rax \n\t" \
196  "movq %%r8, %%rcx \n\t" \
197  "adcq $0, %%rdx \n\t" \
198  "nop \n\t" \
199  "addq %%rax, (%%rdi) \n\t" \
200  "adcq %%rdx, %%rcx \n\t" \
201  "addq $8, %%rdi \n\t"
202 
203 #define MULADDC_STOP \
204  : "+c" (c), "+D" (d), "+S" (s) \
205  : "b" (b) \
206  : "rax", "rdx", "r8" \
207  );
208 
209 #endif /* AMD64 */
210 
211 #if defined(__mc68020__) || defined(__mcpu32__)
212 
213 #define MULADDC_INIT \
214  asm( \
215  "movl %3, %%a2 \n\t" \
216  "movl %4, %%a3 \n\t" \
217  "movl %5, %%d3 \n\t" \
218  "movl %6, %%d2 \n\t" \
219  "moveq #0, %%d0 \n\t"
220 
221 #define MULADDC_CORE \
222  "movel %%a2@+, %%d1 \n\t" \
223  "mulul %%d2, %%d4:%%d1 \n\t" \
224  "addl %%d3, %%d1 \n\t" \
225  "addxl %%d0, %%d4 \n\t" \
226  "moveq #0, %%d3 \n\t" \
227  "addl %%d1, %%a3@+ \n\t" \
228  "addxl %%d4, %%d3 \n\t"
229 
230 #define MULADDC_STOP \
231  "movl %%d3, %0 \n\t" \
232  "movl %%a3, %1 \n\t" \
233  "movl %%a2, %2 \n\t" \
234  : "=m" (c), "=m" (d), "=m" (s) \
235  : "m" (s), "m" (d), "m" (c), "m" (b) \
236  : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
237  );
238 
239 #define MULADDC_HUIT \
240  "movel %%a2@+, %%d1 \n\t" \
241  "mulul %%d2, %%d4:%%d1 \n\t" \
242  "addxl %%d3, %%d1 \n\t" \
243  "addxl %%d0, %%d4 \n\t" \
244  "addl %%d1, %%a3@+ \n\t" \
245  "movel %%a2@+, %%d1 \n\t" \
246  "mulul %%d2, %%d3:%%d1 \n\t" \
247  "addxl %%d4, %%d1 \n\t" \
248  "addxl %%d0, %%d3 \n\t" \
249  "addl %%d1, %%a3@+ \n\t" \
250  "movel %%a2@+, %%d1 \n\t" \
251  "mulul %%d2, %%d4:%%d1 \n\t" \
252  "addxl %%d3, %%d1 \n\t" \
253  "addxl %%d0, %%d4 \n\t" \
254  "addl %%d1, %%a3@+ \n\t" \
255  "movel %%a2@+, %%d1 \n\t" \
256  "mulul %%d2, %%d3:%%d1 \n\t" \
257  "addxl %%d4, %%d1 \n\t" \
258  "addxl %%d0, %%d3 \n\t" \
259  "addl %%d1, %%a3@+ \n\t" \
260  "movel %%a2@+, %%d1 \n\t" \
261  "mulul %%d2, %%d4:%%d1 \n\t" \
262  "addxl %%d3, %%d1 \n\t" \
263  "addxl %%d0, %%d4 \n\t" \
264  "addl %%d1, %%a3@+ \n\t" \
265  "movel %%a2@+, %%d1 \n\t" \
266  "mulul %%d2, %%d3:%%d1 \n\t" \
267  "addxl %%d4, %%d1 \n\t" \
268  "addxl %%d0, %%d3 \n\t" \
269  "addl %%d1, %%a3@+ \n\t" \
270  "movel %%a2@+, %%d1 \n\t" \
271  "mulul %%d2, %%d4:%%d1 \n\t" \
272  "addxl %%d3, %%d1 \n\t" \
273  "addxl %%d0, %%d4 \n\t" \
274  "addl %%d1, %%a3@+ \n\t" \
275  "movel %%a2@+, %%d1 \n\t" \
276  "mulul %%d2, %%d3:%%d1 \n\t" \
277  "addxl %%d4, %%d1 \n\t" \
278  "addxl %%d0, %%d3 \n\t" \
279  "addl %%d1, %%a3@+ \n\t" \
280  "addxl %%d0, %%d3 \n\t"
281 
282 #endif /* MC68000 */
283 
284 #if defined(__powerpc64__) || defined(__ppc64__)
285 
286 #if defined(__MACH__) && defined(__APPLE__)
287 
288 #define MULADDC_INIT \
289  asm( \
290  "ld r3, %3 \n\t" \
291  "ld r4, %4 \n\t" \
292  "ld r5, %5 \n\t" \
293  "ld r6, %6 \n\t" \
294  "addi r3, r3, -8 \n\t" \
295  "addi r4, r4, -8 \n\t" \
296  "addic r5, r5, 0 \n\t"
297 
298 #define MULADDC_CORE \
299  "ldu r7, 8(r3) \n\t" \
300  "mulld r8, r7, r6 \n\t" \
301  "mulhdu r9, r7, r6 \n\t" \
302  "adde r8, r8, r5 \n\t" \
303  "ld r7, 8(r4) \n\t" \
304  "addze r5, r9 \n\t" \
305  "addc r8, r8, r7 \n\t" \
306  "stdu r8, 8(r4) \n\t"
307 
308 #define MULADDC_STOP \
309  "addze r5, r5 \n\t" \
310  "addi r4, r4, 8 \n\t" \
311  "addi r3, r3, 8 \n\t" \
312  "std r5, %0 \n\t" \
313  "std r4, %1 \n\t" \
314  "std r3, %2 \n\t" \
315  : "=m" (c), "=m" (d), "=m" (s) \
316  : "m" (s), "m" (d), "m" (c), "m" (b) \
317  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
318  );
319 
320 
321 #else /* __MACH__ && __APPLE__ */
322 
323 #define MULADDC_INIT \
324  asm( \
325  "ld %%r3, %3 \n\t" \
326  "ld %%r4, %4 \n\t" \
327  "ld %%r5, %5 \n\t" \
328  "ld %%r6, %6 \n\t" \
329  "addi %%r3, %%r3, -8 \n\t" \
330  "addi %%r4, %%r4, -8 \n\t" \
331  "addic %%r5, %%r5, 0 \n\t"
332 
333 #define MULADDC_CORE \
334  "ldu %%r7, 8(%%r3) \n\t" \
335  "mulld %%r8, %%r7, %%r6 \n\t" \
336  "mulhdu %%r9, %%r7, %%r6 \n\t" \
337  "adde %%r8, %%r8, %%r5 \n\t" \
338  "ld %%r7, 8(%%r4) \n\t" \
339  "addze %%r5, %%r9 \n\t" \
340  "addc %%r8, %%r8, %%r7 \n\t" \
341  "stdu %%r8, 8(%%r4) \n\t"
342 
343 #define MULADDC_STOP \
344  "addze %%r5, %%r5 \n\t" \
345  "addi %%r4, %%r4, 8 \n\t" \
346  "addi %%r3, %%r3, 8 \n\t" \
347  "std %%r5, %0 \n\t" \
348  "std %%r4, %1 \n\t" \
349  "std %%r3, %2 \n\t" \
350  : "=m" (c), "=m" (d), "=m" (s) \
351  : "m" (s), "m" (d), "m" (c), "m" (b) \
352  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
353  );
354 
355 #endif /* __MACH__ && __APPLE__ */
356 
357 #elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
358 
359 #if defined(__MACH__) && defined(__APPLE__)
360 
361 #define MULADDC_INIT \
362  asm( \
363  "lwz r3, %3 \n\t" \
364  "lwz r4, %4 \n\t" \
365  "lwz r5, %5 \n\t" \
366  "lwz r6, %6 \n\t" \
367  "addi r3, r3, -4 \n\t" \
368  "addi r4, r4, -4 \n\t" \
369  "addic r5, r5, 0 \n\t"
370 
371 #define MULADDC_CORE \
372  "lwzu r7, 4(r3) \n\t" \
373  "mullw r8, r7, r6 \n\t" \
374  "mulhwu r9, r7, r6 \n\t" \
375  "adde r8, r8, r5 \n\t" \
376  "lwz r7, 4(r4) \n\t" \
377  "addze r5, r9 \n\t" \
378  "addc r8, r8, r7 \n\t" \
379  "stwu r8, 4(r4) \n\t"
380 
381 #define MULADDC_STOP \
382  "addze r5, r5 \n\t" \
383  "addi r4, r4, 4 \n\t" \
384  "addi r3, r3, 4 \n\t" \
385  "stw r5, %0 \n\t" \
386  "stw r4, %1 \n\t" \
387  "stw r3, %2 \n\t" \
388  : "=m" (c), "=m" (d), "=m" (s) \
389  : "m" (s), "m" (d), "m" (c), "m" (b) \
390  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
391  );
392 
393 #else /* __MACH__ && __APPLE__ */
394 
395 #define MULADDC_INIT \
396  asm( \
397  "lwz %%r3, %3 \n\t" \
398  "lwz %%r4, %4 \n\t" \
399  "lwz %%r5, %5 \n\t" \
400  "lwz %%r6, %6 \n\t" \
401  "addi %%r3, %%r3, -4 \n\t" \
402  "addi %%r4, %%r4, -4 \n\t" \
403  "addic %%r5, %%r5, 0 \n\t"
404 
405 #define MULADDC_CORE \
406  "lwzu %%r7, 4(%%r3) \n\t" \
407  "mullw %%r8, %%r7, %%r6 \n\t" \
408  "mulhwu %%r9, %%r7, %%r6 \n\t" \
409  "adde %%r8, %%r8, %%r5 \n\t" \
410  "lwz %%r7, 4(%%r4) \n\t" \
411  "addze %%r5, %%r9 \n\t" \
412  "addc %%r8, %%r8, %%r7 \n\t" \
413  "stwu %%r8, 4(%%r4) \n\t"
414 
415 #define MULADDC_STOP \
416  "addze %%r5, %%r5 \n\t" \
417  "addi %%r4, %%r4, 4 \n\t" \
418  "addi %%r3, %%r3, 4 \n\t" \
419  "stw %%r5, %0 \n\t" \
420  "stw %%r4, %1 \n\t" \
421  "stw %%r3, %2 \n\t" \
422  : "=m" (c), "=m" (d), "=m" (s) \
423  : "m" (s), "m" (d), "m" (c), "m" (b) \
424  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
425  );
426 
427 #endif /* __MACH__ && __APPLE__ */
428 
429 #endif /* PPC32 */
430 
431 /*
432  * The Sparc(64) assembly is reported to be broken.
433  * Disable it for now, until we're able to fix it.
434  */
435 #if 0 && defined(__sparc__)
436 #if defined(__sparc64__)
437 
438 #define MULADDC_INIT \
439  asm( \
440  "ldx %3, %%o0 \n\t" \
441  "ldx %4, %%o1 \n\t" \
442  "ld %5, %%o2 \n\t" \
443  "ld %6, %%o3 \n\t"
444 
445 #define MULADDC_CORE \
446  "ld [%%o0], %%o4 \n\t" \
447  "inc 4, %%o0 \n\t" \
448  "ld [%%o1], %%o5 \n\t" \
449  "umul %%o3, %%o4, %%o4 \n\t" \
450  "addcc %%o4, %%o2, %%o4 \n\t" \
451  "rd %%y, %%g1 \n\t" \
452  "addx %%g1, 0, %%g1 \n\t" \
453  "addcc %%o4, %%o5, %%o4 \n\t" \
454  "st %%o4, [%%o1] \n\t" \
455  "addx %%g1, 0, %%o2 \n\t" \
456  "inc 4, %%o1 \n\t"
457 
458  #define MULADDC_STOP \
459  "st %%o2, %0 \n\t" \
460  "stx %%o1, %1 \n\t" \
461  "stx %%o0, %2 \n\t" \
462  : "=m" (c), "=m" (d), "=m" (s) \
463  : "m" (s), "m" (d), "m" (c), "m" (b) \
464  : "g1", "o0", "o1", "o2", "o3", "o4", \
465  "o5" \
466  );
467 
468 #else /* __sparc64__ */
469 
470 #define MULADDC_INIT \
471  asm( \
472  "ld %3, %%o0 \n\t" \
473  "ld %4, %%o1 \n\t" \
474  "ld %5, %%o2 \n\t" \
475  "ld %6, %%o3 \n\t"
476 
477 #define MULADDC_CORE \
478  "ld [%%o0], %%o4 \n\t" \
479  "inc 4, %%o0 \n\t" \
480  "ld [%%o1], %%o5 \n\t" \
481  "umul %%o3, %%o4, %%o4 \n\t" \
482  "addcc %%o4, %%o2, %%o4 \n\t" \
483  "rd %%y, %%g1 \n\t" \
484  "addx %%g1, 0, %%g1 \n\t" \
485  "addcc %%o4, %%o5, %%o4 \n\t" \
486  "st %%o4, [%%o1] \n\t" \
487  "addx %%g1, 0, %%o2 \n\t" \
488  "inc 4, %%o1 \n\t"
489 
490 #define MULADDC_STOP \
491  "st %%o2, %0 \n\t" \
492  "st %%o1, %1 \n\t" \
493  "st %%o0, %2 \n\t" \
494  : "=m" (c), "=m" (d), "=m" (s) \
495  : "m" (s), "m" (d), "m" (c), "m" (b) \
496  : "g1", "o0", "o1", "o2", "o3", "o4", \
497  "o5" \
498  );
499 
500 #endif /* __sparc64__ */
501 #endif /* __sparc__ */
502 
503 #if defined(__microblaze__) || defined(microblaze)
504 
505 #define MULADDC_INIT \
506  asm( \
507  "lwi r3, %3 \n\t" \
508  "lwi r4, %4 \n\t" \
509  "lwi r5, %5 \n\t" \
510  "lwi r6, %6 \n\t" \
511  "andi r7, r6, 0xffff \n\t" \
512  "bsrli r6, r6, 16 \n\t"
513 
514 #define MULADDC_CORE \
515  "lhui r8, r3, 0 \n\t" \
516  "addi r3, r3, 2 \n\t" \
517  "lhui r9, r3, 0 \n\t" \
518  "addi r3, r3, 2 \n\t" \
519  "mul r10, r9, r6 \n\t" \
520  "mul r11, r8, r7 \n\t" \
521  "mul r12, r9, r7 \n\t" \
522  "mul r13, r8, r6 \n\t" \
523  "bsrli r8, r10, 16 \n\t" \
524  "bsrli r9, r11, 16 \n\t" \
525  "add r13, r13, r8 \n\t" \
526  "add r13, r13, r9 \n\t" \
527  "bslli r10, r10, 16 \n\t" \
528  "bslli r11, r11, 16 \n\t" \
529  "add r12, r12, r10 \n\t" \
530  "addc r13, r13, r0 \n\t" \
531  "add r12, r12, r11 \n\t" \
532  "addc r13, r13, r0 \n\t" \
533  "lwi r10, r4, 0 \n\t" \
534  "add r12, r12, r10 \n\t" \
535  "addc r13, r13, r0 \n\t" \
536  "add r12, r12, r5 \n\t" \
537  "addc r5, r13, r0 \n\t" \
538  "swi r12, r4, 0 \n\t" \
539  "addi r4, r4, 4 \n\t"
540 
541 #define MULADDC_STOP \
542  "swi r5, %0 \n\t" \
543  "swi r4, %1 \n\t" \
544  "swi r3, %2 \n\t" \
545  : "=m" (c), "=m" (d), "=m" (s) \
546  : "m" (s), "m" (d), "m" (c), "m" (b) \
547  : "r3", "r4", "r5", "r6", "r7", "r8", \
548  "r9", "r10", "r11", "r12", "r13" \
549  );
550 
551 #endif /* MicroBlaze */
552 
553 #if defined(__tricore__)
554 
555 #define MULADDC_INIT \
556  asm( \
557  "ld.a %%a2, %3 \n\t" \
558  "ld.a %%a3, %4 \n\t" \
559  "ld.w %%d4, %5 \n\t" \
560  "ld.w %%d1, %6 \n\t" \
561  "xor %%d5, %%d5 \n\t"
562 
563 #define MULADDC_CORE \
564  "ld.w %%d0, [%%a2+] \n\t" \
565  "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
566  "ld.w %%d0, [%%a3] \n\t" \
567  "addx %%d2, %%d2, %%d0 \n\t" \
568  "addc %%d3, %%d3, 0 \n\t" \
569  "mov %%d4, %%d3 \n\t" \
570  "st.w [%%a3+], %%d2 \n\t"
571 
572 #define MULADDC_STOP \
573  "st.w %0, %%d4 \n\t" \
574  "st.a %1, %%a3 \n\t" \
575  "st.a %2, %%a2 \n\t" \
576  : "=m" (c), "=m" (d), "=m" (s) \
577  : "m" (s), "m" (d), "m" (c), "m" (b) \
578  : "d0", "d1", "e2", "d4", "a2", "a3" \
579  );
580 
581 #endif /* TriCore */
582 
583 /*
584  * Note, gcc -O0 by default uses r7 for the frame pointer, so it complains about
585  * our use of r7 below, unless -fomit-frame-pointer is passed.
586  *
587  * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
588  * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
589  * clang and armcc5 under the same conditions).
590  *
591  * So, only use the optimized assembly below for optimized build, which avoids
592  * the build error and is pretty reasonable anyway.
593  */
594 #if defined(__GNUC__) && !defined(__OPTIMIZE__)
595 #define MULADDC_CANNOT_USE_R7
596 #endif
597 
598 #if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
599 
600 #if defined(__thumb__) && !defined(__thumb2__)
601 
602 #define MULADDC_INIT \
603  asm( \
604  "ldr r0, %3 \n\t" \
605  "ldr r1, %4 \n\t" \
606  "ldr r2, %5 \n\t" \
607  "ldr r3, %6 \n\t" \
608  "lsr r7, r3, #16 \n\t" \
609  "mov r9, r7 \n\t" \
610  "lsl r7, r3, #16 \n\t" \
611  "lsr r7, r7, #16 \n\t" \
612  "mov r8, r7 \n\t"
613 
614 #define MULADDC_CORE \
615  "ldmia r0!, {r6} \n\t" \
616  "lsr r7, r6, #16 \n\t" \
617  "lsl r6, r6, #16 \n\t" \
618  "lsr r6, r6, #16 \n\t" \
619  "mov r4, r8 \n\t" \
620  "mul r4, r6 \n\t" \
621  "mov r3, r9 \n\t" \
622  "mul r6, r3 \n\t" \
623  "mov r5, r9 \n\t" \
624  "mul r5, r7 \n\t" \
625  "mov r3, r8 \n\t" \
626  "mul r7, r3 \n\t" \
627  "lsr r3, r6, #16 \n\t" \
628  "add r5, r5, r3 \n\t" \
629  "lsr r3, r7, #16 \n\t" \
630  "add r5, r5, r3 \n\t" \
631  "add r4, r4, r2 \n\t" \
632  "mov r2, #0 \n\t" \
633  "adc r5, r2 \n\t" \
634  "lsl r3, r6, #16 \n\t" \
635  "add r4, r4, r3 \n\t" \
636  "adc r5, r2 \n\t" \
637  "lsl r3, r7, #16 \n\t" \
638  "add r4, r4, r3 \n\t" \
639  "adc r5, r2 \n\t" \
640  "ldr r3, [r1] \n\t" \
641  "add r4, r4, r3 \n\t" \
642  "adc r2, r5 \n\t" \
643  "stmia r1!, {r4} \n\t"
644 
645 #define MULADDC_STOP \
646  "str r2, %0 \n\t" \
647  "str r1, %1 \n\t" \
648  "str r0, %2 \n\t" \
649  : "=m" (c), "=m" (d), "=m" (s) \
650  : "m" (s), "m" (d), "m" (c), "m" (b) \
651  : "r0", "r1", "r2", "r3", "r4", "r5", \
652  "r6", "r7", "r8", "r9", "cc" \
653  );
654 
655 #elif (__ARM_ARCH >= 6) && \
656  defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)
657 
658 #define MULADDC_INIT \
659  asm(
660 
661 #define MULADDC_CORE \
662  "ldr r0, [%0], #4 \n\t" \
663  "ldr r1, [%1] \n\t" \
664  "umaal r1, %2, %3, r0 \n\t" \
665  "str r1, [%1], #4 \n\t"
666 
667 #define MULADDC_STOP \
668  : "=r" (s), "=r" (d), "=r" (c) \
669  : "r" (b), "0" (s), "1" (d), "2" (c) \
670  : "r0", "r1", "memory" \
671  );
672 
673 #else
674 
675 #define MULADDC_INIT \
676  asm( \
677  "ldr r0, %3 \n\t" \
678  "ldr r1, %4 \n\t" \
679  "ldr r2, %5 \n\t" \
680  "ldr r3, %6 \n\t"
681 
682 #define MULADDC_CORE \
683  "ldr r4, [r0], #4 \n\t" \
684  "mov r5, #0 \n\t" \
685  "ldr r6, [r1] \n\t" \
686  "umlal r2, r5, r3, r4 \n\t" \
687  "adds r7, r6, r2 \n\t" \
688  "adc r2, r5, #0 \n\t" \
689  "str r7, [r1], #4 \n\t"
690 
691 #define MULADDC_STOP \
692  "str r2, %0 \n\t" \
693  "str r1, %1 \n\t" \
694  "str r0, %2 \n\t" \
695  : "=m" (c), "=m" (d), "=m" (s) \
696  : "m" (s), "m" (d), "m" (c), "m" (b) \
697  : "r0", "r1", "r2", "r3", "r4", "r5", \
698  "r6", "r7", "cc" \
699  );
700 
701 #endif /* Thumb */
702 
703 #endif /* ARMv3 */
704 
705 #if defined(__alpha__)
706 
707 #define MULADDC_INIT \
708  asm( \
709  "ldq $1, %3 \n\t" \
710  "ldq $2, %4 \n\t" \
711  "ldq $3, %5 \n\t" \
712  "ldq $4, %6 \n\t"
713 
714 #define MULADDC_CORE \
715  "ldq $6, 0($1) \n\t" \
716  "addq $1, 8, $1 \n\t" \
717  "mulq $6, $4, $7 \n\t" \
718  "umulh $6, $4, $6 \n\t" \
719  "addq $7, $3, $7 \n\t" \
720  "cmpult $7, $3, $3 \n\t" \
721  "ldq $5, 0($2) \n\t" \
722  "addq $7, $5, $7 \n\t" \
723  "cmpult $7, $5, $5 \n\t" \
724  "stq $7, 0($2) \n\t" \
725  "addq $2, 8, $2 \n\t" \
726  "addq $6, $3, $3 \n\t" \
727  "addq $5, $3, $3 \n\t"
728 
729 #define MULADDC_STOP \
730  "stq $3, %0 \n\t" \
731  "stq $2, %1 \n\t" \
732  "stq $1, %2 \n\t" \
733  : "=m" (c), "=m" (d), "=m" (s) \
734  : "m" (s), "m" (d), "m" (c), "m" (b) \
735  : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
736  );
737 #endif /* Alpha */
738 
739 #if defined(__mips__) && !defined(__mips64)
740 
741 #define MULADDC_INIT \
742  asm( \
743  "lw $10, %3 \n\t" \
744  "lw $11, %4 \n\t" \
745  "lw $12, %5 \n\t" \
746  "lw $13, %6 \n\t"
747 
748 #define MULADDC_CORE \
749  "lw $14, 0($10) \n\t" \
750  "multu $13, $14 \n\t" \
751  "addi $10, $10, 4 \n\t" \
752  "mflo $14 \n\t" \
753  "mfhi $9 \n\t" \
754  "addu $14, $12, $14 \n\t" \
755  "lw $15, 0($11) \n\t" \
756  "sltu $12, $14, $12 \n\t" \
757  "addu $15, $14, $15 \n\t" \
758  "sltu $14, $15, $14 \n\t" \
759  "addu $12, $12, $9 \n\t" \
760  "sw $15, 0($11) \n\t" \
761  "addu $12, $12, $14 \n\t" \
762  "addi $11, $11, 4 \n\t"
763 
764 #define MULADDC_STOP \
765  "sw $12, %0 \n\t" \
766  "sw $11, %1 \n\t" \
767  "sw $10, %2 \n\t" \
768  : "=m" (c), "=m" (d), "=m" (s) \
769  : "m" (s), "m" (d), "m" (c), "m" (b) \
770  : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
771  );
772 
773 #endif /* MIPS */
774 #endif /* GNUC */
775 
776 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
777 
778 #define MULADDC_INIT \
779  __asm mov esi, s \
780  __asm mov edi, d \
781  __asm mov ecx, c \
782  __asm mov ebx, b
783 
784 #define MULADDC_CORE \
785  __asm lodsd \
786  __asm mul ebx \
787  __asm add eax, ecx \
788  __asm adc edx, 0 \
789  __asm add eax, [edi] \
790  __asm adc edx, 0 \
791  __asm mov ecx, edx \
792  __asm stosd
793 
794 #if defined(MBEDTLS_HAVE_SSE2)
795 
796 #define EMIT __asm _emit
797 
798 #define MULADDC_HUIT \
799  EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
800  EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
801  EMIT 0x0F EMIT 0x6E EMIT 0x1F \
802  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
803  EMIT 0x0F EMIT 0x6E EMIT 0x16 \
804  EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
805  EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
806  EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
807  EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
808  EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
809  EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
810  EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
811  EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
812  EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
813  EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
814  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
815  EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
816  EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
817  EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
818  EMIT 0x0F EMIT 0x7E EMIT 0x0F \
819  EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
820  EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
821  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
822  EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
823  EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
824  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
825  EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
826  EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
827  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
828  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
829  EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
830  EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
831  EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
832  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
833  EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
834  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
835  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
836  EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
837  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
838  EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
839  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
840  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
841  EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
842  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
843  EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
844  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
845  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
846  EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
847  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
848  EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
849  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
850  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
851  EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
852  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
853  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
854  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
855  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
856  EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
857  EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
858  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
859  EMIT 0x0F EMIT 0x7E EMIT 0xC9
860 
861 #define MULADDC_STOP \
862  EMIT 0x0F EMIT 0x77 \
863  __asm mov c, ecx \
864  __asm mov d, edi \
865  __asm mov s, esi \
866 
867 #else
868 
869 #define MULADDC_STOP \
870  __asm mov c, ecx \
871  __asm mov d, edi \
872  __asm mov s, esi \
873 
874 #endif /* SSE2 */
875 #endif /* MSVC */
876 
877 #endif /* MBEDTLS_HAVE_ASM */
878 
879 #if !defined(MULADDC_CORE)
880 #if defined(MBEDTLS_HAVE_UDBL)
881 
882 #define MULADDC_INIT \
883 { \
884  mbedtls_t_udbl r; \
885  mbedtls_mpi_uint r0, r1;
886 
887 #define MULADDC_CORE \
888  r = *(s++) * (mbedtls_t_udbl) b; \
889  r0 = (mbedtls_mpi_uint) r; \
890  r1 = (mbedtls_mpi_uint)( r >> biL ); \
891  r0 += c; r1 += (r0 < c); \
892  r0 += *d; r1 += (r0 < *d); \
893  c = r1; *(d++) = r0;
894 
895 #define MULADDC_STOP \
896 }
897 
898 #else
899 #define MULADDC_INIT \
900 { \
901  mbedtls_mpi_uint s0, s1, b0, b1; \
902  mbedtls_mpi_uint r0, r1, rx, ry; \
903  b0 = ( b << biH ) >> biH; \
904  b1 = ( b >> biH );
905 
906 #define MULADDC_CORE \
907  s0 = ( *s << biH ) >> biH; \
908  s1 = ( *s >> biH ); s++; \
909  rx = s0 * b1; r0 = s0 * b0; \
910  ry = s1 * b0; r1 = s1 * b1; \
911  r1 += ( rx >> biH ); \
912  r1 += ( ry >> biH ); \
913  rx <<= biH; ry <<= biH; \
914  r0 += rx; r1 += (r0 < rx); \
915  r0 += ry; r1 += (r0 < ry); \
916  r0 += c; r1 += (r0 < c); \
917  r0 += *d; r1 += (r0 < *d); \
918  c = r1; *(d++) = r0;
919 
920 #define MULADDC_STOP \
921 }
922 
923 #endif /* C (generic) */
924 #endif /* C (longlong) */
925 
926 #endif /* bn_mul.h */
Multi-precision integer library.