mbed TLS v2.6.0
bn_mul.h
Go to the documentation of this file.
1 
25 /*
26  * Multiply source vector [s] with b, add result
27  * to destination vector [d] and set carry c.
28  *
29  * Currently supports:
30  *
31  * . IA-32 (386+) . AMD64 / EM64T
32  * . IA-32 (SSE2) . Motorola 68000
33  * . PowerPC, 32-bit . MicroBlaze
34  * . PowerPC, 64-bit . TriCore
35  * . SPARC v8 . ARM v3+
36  * . Alpha . MIPS32
37  * . C, longlong . C, generic
38  */
39 #ifndef MBEDTLS_BN_MUL_H
40 #define MBEDTLS_BN_MUL_H
41 
42 #include "bignum.h"
43 
44 #if defined(MBEDTLS_HAVE_ASM)
45 
46 #ifndef asm
47 #define asm __asm
48 #endif
49 
50 /* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */
51 #if defined(__GNUC__) && \
52  ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
53 #if defined(__i386__)
54 
55 #define MULADDC_INIT \
56  asm( \
57  "movl %%ebx, %0 \n\t" \
58  "movl %5, %%esi \n\t" \
59  "movl %6, %%edi \n\t" \
60  "movl %7, %%ecx \n\t" \
61  "movl %8, %%ebx \n\t"
62 
63 #define MULADDC_CORE \
64  "lodsl \n\t" \
65  "mull %%ebx \n\t" \
66  "addl %%ecx, %%eax \n\t" \
67  "adcl $0, %%edx \n\t" \
68  "addl (%%edi), %%eax \n\t" \
69  "adcl $0, %%edx \n\t" \
70  "movl %%edx, %%ecx \n\t" \
71  "stosl \n\t"
72 
73 #if defined(MBEDTLS_HAVE_SSE2)
74 
75 #define MULADDC_HUIT \
76  "movd %%ecx, %%mm1 \n\t" \
77  "movd %%ebx, %%mm0 \n\t" \
78  "movd (%%edi), %%mm3 \n\t" \
79  "paddq %%mm3, %%mm1 \n\t" \
80  "movd (%%esi), %%mm2 \n\t" \
81  "pmuludq %%mm0, %%mm2 \n\t" \
82  "movd 4(%%esi), %%mm4 \n\t" \
83  "pmuludq %%mm0, %%mm4 \n\t" \
84  "movd 8(%%esi), %%mm6 \n\t" \
85  "pmuludq %%mm0, %%mm6 \n\t" \
86  "movd 12(%%esi), %%mm7 \n\t" \
87  "pmuludq %%mm0, %%mm7 \n\t" \
88  "paddq %%mm2, %%mm1 \n\t" \
89  "movd 4(%%edi), %%mm3 \n\t" \
90  "paddq %%mm4, %%mm3 \n\t" \
91  "movd 8(%%edi), %%mm5 \n\t" \
92  "paddq %%mm6, %%mm5 \n\t" \
93  "movd 12(%%edi), %%mm4 \n\t" \
94  "paddq %%mm4, %%mm7 \n\t" \
95  "movd %%mm1, (%%edi) \n\t" \
96  "movd 16(%%esi), %%mm2 \n\t" \
97  "pmuludq %%mm0, %%mm2 \n\t" \
98  "psrlq $32, %%mm1 \n\t" \
99  "movd 20(%%esi), %%mm4 \n\t" \
100  "pmuludq %%mm0, %%mm4 \n\t" \
101  "paddq %%mm3, %%mm1 \n\t" \
102  "movd 24(%%esi), %%mm6 \n\t" \
103  "pmuludq %%mm0, %%mm6 \n\t" \
104  "movd %%mm1, 4(%%edi) \n\t" \
105  "psrlq $32, %%mm1 \n\t" \
106  "movd 28(%%esi), %%mm3 \n\t" \
107  "pmuludq %%mm0, %%mm3 \n\t" \
108  "paddq %%mm5, %%mm1 \n\t" \
109  "movd 16(%%edi), %%mm5 \n\t" \
110  "paddq %%mm5, %%mm2 \n\t" \
111  "movd %%mm1, 8(%%edi) \n\t" \
112  "psrlq $32, %%mm1 \n\t" \
113  "paddq %%mm7, %%mm1 \n\t" \
114  "movd 20(%%edi), %%mm5 \n\t" \
115  "paddq %%mm5, %%mm4 \n\t" \
116  "movd %%mm1, 12(%%edi) \n\t" \
117  "psrlq $32, %%mm1 \n\t" \
118  "paddq %%mm2, %%mm1 \n\t" \
119  "movd 24(%%edi), %%mm5 \n\t" \
120  "paddq %%mm5, %%mm6 \n\t" \
121  "movd %%mm1, 16(%%edi) \n\t" \
122  "psrlq $32, %%mm1 \n\t" \
123  "paddq %%mm4, %%mm1 \n\t" \
124  "movd 28(%%edi), %%mm5 \n\t" \
125  "paddq %%mm5, %%mm3 \n\t" \
126  "movd %%mm1, 20(%%edi) \n\t" \
127  "psrlq $32, %%mm1 \n\t" \
128  "paddq %%mm6, %%mm1 \n\t" \
129  "movd %%mm1, 24(%%edi) \n\t" \
130  "psrlq $32, %%mm1 \n\t" \
131  "paddq %%mm3, %%mm1 \n\t" \
132  "movd %%mm1, 28(%%edi) \n\t" \
133  "addl $32, %%edi \n\t" \
134  "addl $32, %%esi \n\t" \
135  "psrlq $32, %%mm1 \n\t" \
136  "movd %%mm1, %%ecx \n\t"
137 
138 #define MULADDC_STOP \
139  "emms \n\t" \
140  "movl %4, %%ebx \n\t" \
141  "movl %%ecx, %1 \n\t" \
142  "movl %%edi, %2 \n\t" \
143  "movl %%esi, %3 \n\t" \
144  : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
145  : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
146  : "eax", "ecx", "edx", "esi", "edi" \
147  );
148 
149 #else
150 
151 #define MULADDC_STOP \
152  "movl %4, %%ebx \n\t" \
153  "movl %%ecx, %1 \n\t" \
154  "movl %%edi, %2 \n\t" \
155  "movl %%esi, %3 \n\t" \
156  : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
157  : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
158  : "eax", "ecx", "edx", "esi", "edi" \
159  );
160 #endif /* SSE2 */
161 #endif /* i386 */
162 
163 #if defined(__amd64__) || defined (__x86_64__)
164 
165 #define MULADDC_INIT \
166  asm( \
167  "xorq %%r8, %%r8 \n\t"
168 
169 #define MULADDC_CORE \
170  "movq (%%rsi), %%rax \n\t" \
171  "mulq %%rbx \n\t" \
172  "addq $8, %%rsi \n\t" \
173  "addq %%rcx, %%rax \n\t" \
174  "movq %%r8, %%rcx \n\t" \
175  "adcq $0, %%rdx \n\t" \
176  "nop \n\t" \
177  "addq %%rax, (%%rdi) \n\t" \
178  "adcq %%rdx, %%rcx \n\t" \
179  "addq $8, %%rdi \n\t"
180 
181 #define MULADDC_STOP \
182  : "+c" (c), "+D" (d), "+S" (s) \
183  : "b" (b) \
184  : "rax", "rdx", "r8" \
185  );
186 
187 #endif /* AMD64 */
188 
189 #if defined(__mc68020__) || defined(__mcpu32__)
190 
191 #define MULADDC_INIT \
192  asm( \
193  "movl %3, %%a2 \n\t" \
194  "movl %4, %%a3 \n\t" \
195  "movl %5, %%d3 \n\t" \
196  "movl %6, %%d2 \n\t" \
197  "moveq #0, %%d0 \n\t"
198 
199 #define MULADDC_CORE \
200  "movel %%a2@+, %%d1 \n\t" \
201  "mulul %%d2, %%d4:%%d1 \n\t" \
202  "addl %%d3, %%d1 \n\t" \
203  "addxl %%d0, %%d4 \n\t" \
204  "moveq #0, %%d3 \n\t" \
205  "addl %%d1, %%a3@+ \n\t" \
206  "addxl %%d4, %%d3 \n\t"
207 
208 #define MULADDC_STOP \
209  "movl %%d3, %0 \n\t" \
210  "movl %%a3, %1 \n\t" \
211  "movl %%a2, %2 \n\t" \
212  : "=m" (c), "=m" (d), "=m" (s) \
213  : "m" (s), "m" (d), "m" (c), "m" (b) \
214  : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
215  );
216 
217 #define MULADDC_HUIT \
218  "movel %%a2@+, %%d1 \n\t" \
219  "mulul %%d2, %%d4:%%d1 \n\t" \
220  "addxl %%d3, %%d1 \n\t" \
221  "addxl %%d0, %%d4 \n\t" \
222  "addl %%d1, %%a3@+ \n\t" \
223  "movel %%a2@+, %%d1 \n\t" \
224  "mulul %%d2, %%d3:%%d1 \n\t" \
225  "addxl %%d4, %%d1 \n\t" \
226  "addxl %%d0, %%d3 \n\t" \
227  "addl %%d1, %%a3@+ \n\t" \
228  "movel %%a2@+, %%d1 \n\t" \
229  "mulul %%d2, %%d4:%%d1 \n\t" \
230  "addxl %%d3, %%d1 \n\t" \
231  "addxl %%d0, %%d4 \n\t" \
232  "addl %%d1, %%a3@+ \n\t" \
233  "movel %%a2@+, %%d1 \n\t" \
234  "mulul %%d2, %%d3:%%d1 \n\t" \
235  "addxl %%d4, %%d1 \n\t" \
236  "addxl %%d0, %%d3 \n\t" \
237  "addl %%d1, %%a3@+ \n\t" \
238  "movel %%a2@+, %%d1 \n\t" \
239  "mulul %%d2, %%d4:%%d1 \n\t" \
240  "addxl %%d3, %%d1 \n\t" \
241  "addxl %%d0, %%d4 \n\t" \
242  "addl %%d1, %%a3@+ \n\t" \
243  "movel %%a2@+, %%d1 \n\t" \
244  "mulul %%d2, %%d3:%%d1 \n\t" \
245  "addxl %%d4, %%d1 \n\t" \
246  "addxl %%d0, %%d3 \n\t" \
247  "addl %%d1, %%a3@+ \n\t" \
248  "movel %%a2@+, %%d1 \n\t" \
249  "mulul %%d2, %%d4:%%d1 \n\t" \
250  "addxl %%d3, %%d1 \n\t" \
251  "addxl %%d0, %%d4 \n\t" \
252  "addl %%d1, %%a3@+ \n\t" \
253  "movel %%a2@+, %%d1 \n\t" \
254  "mulul %%d2, %%d3:%%d1 \n\t" \
255  "addxl %%d4, %%d1 \n\t" \
256  "addxl %%d0, %%d3 \n\t" \
257  "addl %%d1, %%a3@+ \n\t" \
258  "addxl %%d0, %%d3 \n\t"
259 
260 #endif /* MC68000 */
261 
262 #if defined(__powerpc64__) || defined(__ppc64__)
263 
264 #if defined(__MACH__) && defined(__APPLE__)
265 
266 #define MULADDC_INIT \
267  asm( \
268  "ld r3, %3 \n\t" \
269  "ld r4, %4 \n\t" \
270  "ld r5, %5 \n\t" \
271  "ld r6, %6 \n\t" \
272  "addi r3, r3, -8 \n\t" \
273  "addi r4, r4, -8 \n\t" \
274  "addic r5, r5, 0 \n\t"
275 
276 #define MULADDC_CORE \
277  "ldu r7, 8(r3) \n\t" \
278  "mulld r8, r7, r6 \n\t" \
279  "mulhdu r9, r7, r6 \n\t" \
280  "adde r8, r8, r5 \n\t" \
281  "ld r7, 8(r4) \n\t" \
282  "addze r5, r9 \n\t" \
283  "addc r8, r8, r7 \n\t" \
284  "stdu r8, 8(r4) \n\t"
285 
286 #define MULADDC_STOP \
287  "addze r5, r5 \n\t" \
288  "addi r4, r4, 8 \n\t" \
289  "addi r3, r3, 8 \n\t" \
290  "std r5, %0 \n\t" \
291  "std r4, %1 \n\t" \
292  "std r3, %2 \n\t" \
293  : "=m" (c), "=m" (d), "=m" (s) \
294  : "m" (s), "m" (d), "m" (c), "m" (b) \
295  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
296  );
297 
298 
299 #else /* __MACH__ && __APPLE__ */
300 
301 #define MULADDC_INIT \
302  asm( \
303  "ld %%r3, %3 \n\t" \
304  "ld %%r4, %4 \n\t" \
305  "ld %%r5, %5 \n\t" \
306  "ld %%r6, %6 \n\t" \
307  "addi %%r3, %%r3, -8 \n\t" \
308  "addi %%r4, %%r4, -8 \n\t" \
309  "addic %%r5, %%r5, 0 \n\t"
310 
311 #define MULADDC_CORE \
312  "ldu %%r7, 8(%%r3) \n\t" \
313  "mulld %%r8, %%r7, %%r6 \n\t" \
314  "mulhdu %%r9, %%r7, %%r6 \n\t" \
315  "adde %%r8, %%r8, %%r5 \n\t" \
316  "ld %%r7, 8(%%r4) \n\t" \
317  "addze %%r5, %%r9 \n\t" \
318  "addc %%r8, %%r8, %%r7 \n\t" \
319  "stdu %%r8, 8(%%r4) \n\t"
320 
321 #define MULADDC_STOP \
322  "addze %%r5, %%r5 \n\t" \
323  "addi %%r4, %%r4, 8 \n\t" \
324  "addi %%r3, %%r3, 8 \n\t" \
325  "std %%r5, %0 \n\t" \
326  "std %%r4, %1 \n\t" \
327  "std %%r3, %2 \n\t" \
328  : "=m" (c), "=m" (d), "=m" (s) \
329  : "m" (s), "m" (d), "m" (c), "m" (b) \
330  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
331  );
332 
333 #endif /* __MACH__ && __APPLE__ */
334 
335 #elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
336 
337 #if defined(__MACH__) && defined(__APPLE__)
338 
339 #define MULADDC_INIT \
340  asm( \
341  "lwz r3, %3 \n\t" \
342  "lwz r4, %4 \n\t" \
343  "lwz r5, %5 \n\t" \
344  "lwz r6, %6 \n\t" \
345  "addi r3, r3, -4 \n\t" \
346  "addi r4, r4, -4 \n\t" \
347  "addic r5, r5, 0 \n\t"
348 
349 #define MULADDC_CORE \
350  "lwzu r7, 4(r3) \n\t" \
351  "mullw r8, r7, r6 \n\t" \
352  "mulhwu r9, r7, r6 \n\t" \
353  "adde r8, r8, r5 \n\t" \
354  "lwz r7, 4(r4) \n\t" \
355  "addze r5, r9 \n\t" \
356  "addc r8, r8, r7 \n\t" \
357  "stwu r8, 4(r4) \n\t"
358 
359 #define MULADDC_STOP \
360  "addze r5, r5 \n\t" \
361  "addi r4, r4, 4 \n\t" \
362  "addi r3, r3, 4 \n\t" \
363  "stw r5, %0 \n\t" \
364  "stw r4, %1 \n\t" \
365  "stw r3, %2 \n\t" \
366  : "=m" (c), "=m" (d), "=m" (s) \
367  : "m" (s), "m" (d), "m" (c), "m" (b) \
368  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
369  );
370 
371 #else /* __MACH__ && __APPLE__ */
372 
373 #define MULADDC_INIT \
374  asm( \
375  "lwz %%r3, %3 \n\t" \
376  "lwz %%r4, %4 \n\t" \
377  "lwz %%r5, %5 \n\t" \
378  "lwz %%r6, %6 \n\t" \
379  "addi %%r3, %%r3, -4 \n\t" \
380  "addi %%r4, %%r4, -4 \n\t" \
381  "addic %%r5, %%r5, 0 \n\t"
382 
383 #define MULADDC_CORE \
384  "lwzu %%r7, 4(%%r3) \n\t" \
385  "mullw %%r8, %%r7, %%r6 \n\t" \
386  "mulhwu %%r9, %%r7, %%r6 \n\t" \
387  "adde %%r8, %%r8, %%r5 \n\t" \
388  "lwz %%r7, 4(%%r4) \n\t" \
389  "addze %%r5, %%r9 \n\t" \
390  "addc %%r8, %%r8, %%r7 \n\t" \
391  "stwu %%r8, 4(%%r4) \n\t"
392 
393 #define MULADDC_STOP \
394  "addze %%r5, %%r5 \n\t" \
395  "addi %%r4, %%r4, 4 \n\t" \
396  "addi %%r3, %%r3, 4 \n\t" \
397  "stw %%r5, %0 \n\t" \
398  "stw %%r4, %1 \n\t" \
399  "stw %%r3, %2 \n\t" \
400  : "=m" (c), "=m" (d), "=m" (s) \
401  : "m" (s), "m" (d), "m" (c), "m" (b) \
402  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
403  );
404 
405 #endif /* __MACH__ && __APPLE__ */
406 
407 #endif /* PPC32 */
408 
409 /*
410  * The Sparc(64) assembly is reported to be broken.
411  * Disable it for now, until we're able to fix it.
412  */
413 #if 0 && defined(__sparc__)
414 #if defined(__sparc64__)
415 
416 #define MULADDC_INIT \
417  asm( \
418  "ldx %3, %%o0 \n\t" \
419  "ldx %4, %%o1 \n\t" \
420  "ld %5, %%o2 \n\t" \
421  "ld %6, %%o3 \n\t"
422 
423 #define MULADDC_CORE \
424  "ld [%%o0], %%o4 \n\t" \
425  "inc 4, %%o0 \n\t" \
426  "ld [%%o1], %%o5 \n\t" \
427  "umul %%o3, %%o4, %%o4 \n\t" \
428  "addcc %%o4, %%o2, %%o4 \n\t" \
429  "rd %%y, %%g1 \n\t" \
430  "addx %%g1, 0, %%g1 \n\t" \
431  "addcc %%o4, %%o5, %%o4 \n\t" \
432  "st %%o4, [%%o1] \n\t" \
433  "addx %%g1, 0, %%o2 \n\t" \
434  "inc 4, %%o1 \n\t"
435 
436  #define MULADDC_STOP \
437  "st %%o2, %0 \n\t" \
438  "stx %%o1, %1 \n\t" \
439  "stx %%o0, %2 \n\t" \
440  : "=m" (c), "=m" (d), "=m" (s) \
441  : "m" (s), "m" (d), "m" (c), "m" (b) \
442  : "g1", "o0", "o1", "o2", "o3", "o4", \
443  "o5" \
444  );
445 
446 #else /* __sparc64__ */
447 
448 #define MULADDC_INIT \
449  asm( \
450  "ld %3, %%o0 \n\t" \
451  "ld %4, %%o1 \n\t" \
452  "ld %5, %%o2 \n\t" \
453  "ld %6, %%o3 \n\t"
454 
455 #define MULADDC_CORE \
456  "ld [%%o0], %%o4 \n\t" \
457  "inc 4, %%o0 \n\t" \
458  "ld [%%o1], %%o5 \n\t" \
459  "umul %%o3, %%o4, %%o4 \n\t" \
460  "addcc %%o4, %%o2, %%o4 \n\t" \
461  "rd %%y, %%g1 \n\t" \
462  "addx %%g1, 0, %%g1 \n\t" \
463  "addcc %%o4, %%o5, %%o4 \n\t" \
464  "st %%o4, [%%o1] \n\t" \
465  "addx %%g1, 0, %%o2 \n\t" \
466  "inc 4, %%o1 \n\t"
467 
468 #define MULADDC_STOP \
469  "st %%o2, %0 \n\t" \
470  "st %%o1, %1 \n\t" \
471  "st %%o0, %2 \n\t" \
472  : "=m" (c), "=m" (d), "=m" (s) \
473  : "m" (s), "m" (d), "m" (c), "m" (b) \
474  : "g1", "o0", "o1", "o2", "o3", "o4", \
475  "o5" \
476  );
477 
478 #endif /* __sparc64__ */
479 #endif /* __sparc__ */
480 
481 #if defined(__microblaze__) || defined(microblaze)
482 
483 #define MULADDC_INIT \
484  asm( \
485  "lwi r3, %3 \n\t" \
486  "lwi r4, %4 \n\t" \
487  "lwi r5, %5 \n\t" \
488  "lwi r6, %6 \n\t" \
489  "andi r7, r6, 0xffff \n\t" \
490  "bsrli r6, r6, 16 \n\t"
491 
492 #define MULADDC_CORE \
493  "lhui r8, r3, 0 \n\t" \
494  "addi r3, r3, 2 \n\t" \
495  "lhui r9, r3, 0 \n\t" \
496  "addi r3, r3, 2 \n\t" \
497  "mul r10, r9, r6 \n\t" \
498  "mul r11, r8, r7 \n\t" \
499  "mul r12, r9, r7 \n\t" \
500  "mul r13, r8, r6 \n\t" \
501  "bsrli r8, r10, 16 \n\t" \
502  "bsrli r9, r11, 16 \n\t" \
503  "add r13, r13, r8 \n\t" \
504  "add r13, r13, r9 \n\t" \
505  "bslli r10, r10, 16 \n\t" \
506  "bslli r11, r11, 16 \n\t" \
507  "add r12, r12, r10 \n\t" \
508  "addc r13, r13, r0 \n\t" \
509  "add r12, r12, r11 \n\t" \
510  "addc r13, r13, r0 \n\t" \
511  "lwi r10, r4, 0 \n\t" \
512  "add r12, r12, r10 \n\t" \
513  "addc r13, r13, r0 \n\t" \
514  "add r12, r12, r5 \n\t" \
515  "addc r5, r13, r0 \n\t" \
516  "swi r12, r4, 0 \n\t" \
517  "addi r4, r4, 4 \n\t"
518 
519 #define MULADDC_STOP \
520  "swi r5, %0 \n\t" \
521  "swi r4, %1 \n\t" \
522  "swi r3, %2 \n\t" \
523  : "=m" (c), "=m" (d), "=m" (s) \
524  : "m" (s), "m" (d), "m" (c), "m" (b) \
525  : "r3", "r4" "r5", "r6", "r7", "r8", \
526  "r9", "r10", "r11", "r12", "r13" \
527  );
528 
529 #endif /* MicroBlaze */
530 
531 #if defined(__tricore__)
532 
533 #define MULADDC_INIT \
534  asm( \
535  "ld.a %%a2, %3 \n\t" \
536  "ld.a %%a3, %4 \n\t" \
537  "ld.w %%d4, %5 \n\t" \
538  "ld.w %%d1, %6 \n\t" \
539  "xor %%d5, %%d5 \n\t"
540 
541 #define MULADDC_CORE \
542  "ld.w %%d0, [%%a2+] \n\t" \
543  "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
544  "ld.w %%d0, [%%a3] \n\t" \
545  "addx %%d2, %%d2, %%d0 \n\t" \
546  "addc %%d3, %%d3, 0 \n\t" \
547  "mov %%d4, %%d3 \n\t" \
548  "st.w [%%a3+], %%d2 \n\t"
549 
550 #define MULADDC_STOP \
551  "st.w %0, %%d4 \n\t" \
552  "st.a %1, %%a3 \n\t" \
553  "st.a %2, %%a2 \n\t" \
554  : "=m" (c), "=m" (d), "=m" (s) \
555  : "m" (s), "m" (d), "m" (c), "m" (b) \
556  : "d0", "d1", "e2", "d4", "a2", "a3" \
557  );
558 
559 #endif /* TriCore */
560 
561 /*
562  * gcc -O0 by default uses r7 for the frame pointer, so it complains about our
563  * use of r7 below, unless -fomit-frame-pointer is passed. Unfortunately,
564  * passing that option is not easy when building with yotta.
565  *
566  * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
567  * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
568  * clang and armcc5 under the same conditions).
569  *
570  * So, only use the optimized assembly below for optimized build, which avoids
571  * the build error and is pretty reasonable anyway.
572  */
573 #if defined(__GNUC__) && !defined(__OPTIMIZE__)
574 #define MULADDC_CANNOT_USE_R7
575 #endif
576 
577 #if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
578 
579 #if defined(__thumb__) && !defined(__thumb2__)
580 
581 #define MULADDC_INIT \
582  asm( \
583  "ldr r0, %3 \n\t" \
584  "ldr r1, %4 \n\t" \
585  "ldr r2, %5 \n\t" \
586  "ldr r3, %6 \n\t" \
587  "lsr r7, r3, #16 \n\t" \
588  "mov r9, r7 \n\t" \
589  "lsl r7, r3, #16 \n\t" \
590  "lsr r7, r7, #16 \n\t" \
591  "mov r8, r7 \n\t"
592 
593 #define MULADDC_CORE \
594  "ldmia r0!, {r6} \n\t" \
595  "lsr r7, r6, #16 \n\t" \
596  "lsl r6, r6, #16 \n\t" \
597  "lsr r6, r6, #16 \n\t" \
598  "mov r4, r8 \n\t" \
599  "mul r4, r6 \n\t" \
600  "mov r3, r9 \n\t" \
601  "mul r6, r3 \n\t" \
602  "mov r5, r9 \n\t" \
603  "mul r5, r7 \n\t" \
604  "mov r3, r8 \n\t" \
605  "mul r7, r3 \n\t" \
606  "lsr r3, r6, #16 \n\t" \
607  "add r5, r5, r3 \n\t" \
608  "lsr r3, r7, #16 \n\t" \
609  "add r5, r5, r3 \n\t" \
610  "add r4, r4, r2 \n\t" \
611  "mov r2, #0 \n\t" \
612  "adc r5, r2 \n\t" \
613  "lsl r3, r6, #16 \n\t" \
614  "add r4, r4, r3 \n\t" \
615  "adc r5, r2 \n\t" \
616  "lsl r3, r7, #16 \n\t" \
617  "add r4, r4, r3 \n\t" \
618  "adc r5, r2 \n\t" \
619  "ldr r3, [r1] \n\t" \
620  "add r4, r4, r3 \n\t" \
621  "adc r2, r5 \n\t" \
622  "stmia r1!, {r4} \n\t"
623 
624 #define MULADDC_STOP \
625  "str r2, %0 \n\t" \
626  "str r1, %1 \n\t" \
627  "str r0, %2 \n\t" \
628  : "=m" (c), "=m" (d), "=m" (s) \
629  : "m" (s), "m" (d), "m" (c), "m" (b) \
630  : "r0", "r1", "r2", "r3", "r4", "r5", \
631  "r6", "r7", "r8", "r9", "cc" \
632  );
633 
634 #else
635 
636 #define MULADDC_INIT \
637  asm( \
638  "ldr r0, %3 \n\t" \
639  "ldr r1, %4 \n\t" \
640  "ldr r2, %5 \n\t" \
641  "ldr r3, %6 \n\t"
642 
643 #define MULADDC_CORE \
644  "ldr r4, [r0], #4 \n\t" \
645  "mov r5, #0 \n\t" \
646  "ldr r6, [r1] \n\t" \
647  "umlal r2, r5, r3, r4 \n\t" \
648  "adds r7, r6, r2 \n\t" \
649  "adc r2, r5, #0 \n\t" \
650  "str r7, [r1], #4 \n\t"
651 
652 #define MULADDC_STOP \
653  "str r2, %0 \n\t" \
654  "str r1, %1 \n\t" \
655  "str r0, %2 \n\t" \
656  : "=m" (c), "=m" (d), "=m" (s) \
657  : "m" (s), "m" (d), "m" (c), "m" (b) \
658  : "r0", "r1", "r2", "r3", "r4", "r5", \
659  "r6", "r7", "cc" \
660  );
661 
662 #endif /* Thumb */
663 
664 #endif /* ARMv3 */
665 
666 #if defined(__alpha__)
667 
668 #define MULADDC_INIT \
669  asm( \
670  "ldq $1, %3 \n\t" \
671  "ldq $2, %4 \n\t" \
672  "ldq $3, %5 \n\t" \
673  "ldq $4, %6 \n\t"
674 
675 #define MULADDC_CORE \
676  "ldq $6, 0($1) \n\t" \
677  "addq $1, 8, $1 \n\t" \
678  "mulq $6, $4, $7 \n\t" \
679  "umulh $6, $4, $6 \n\t" \
680  "addq $7, $3, $7 \n\t" \
681  "cmpult $7, $3, $3 \n\t" \
682  "ldq $5, 0($2) \n\t" \
683  "addq $7, $5, $7 \n\t" \
684  "cmpult $7, $5, $5 \n\t" \
685  "stq $7, 0($2) \n\t" \
686  "addq $2, 8, $2 \n\t" \
687  "addq $6, $3, $3 \n\t" \
688  "addq $5, $3, $3 \n\t"
689 
690 #define MULADDC_STOP \
691  "stq $3, %0 \n\t" \
692  "stq $2, %1 \n\t" \
693  "stq $1, %2 \n\t" \
694  : "=m" (c), "=m" (d), "=m" (s) \
695  : "m" (s), "m" (d), "m" (c), "m" (b) \
696  : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
697  );
698 #endif /* Alpha */
699 
700 #if defined(__mips__) && !defined(__mips64)
701 
702 #define MULADDC_INIT \
703  asm( \
704  "lw $10, %3 \n\t" \
705  "lw $11, %4 \n\t" \
706  "lw $12, %5 \n\t" \
707  "lw $13, %6 \n\t"
708 
709 #define MULADDC_CORE \
710  "lw $14, 0($10) \n\t" \
711  "multu $13, $14 \n\t" \
712  "addi $10, $10, 4 \n\t" \
713  "mflo $14 \n\t" \
714  "mfhi $9 \n\t" \
715  "addu $14, $12, $14 \n\t" \
716  "lw $15, 0($11) \n\t" \
717  "sltu $12, $14, $12 \n\t" \
718  "addu $15, $14, $15 \n\t" \
719  "sltu $14, $15, $14 \n\t" \
720  "addu $12, $12, $9 \n\t" \
721  "sw $15, 0($11) \n\t" \
722  "addu $12, $12, $14 \n\t" \
723  "addi $11, $11, 4 \n\t"
724 
725 #define MULADDC_STOP \
726  "sw $12, %0 \n\t" \
727  "sw $11, %1 \n\t" \
728  "sw $10, %2 \n\t" \
729  : "=m" (c), "=m" (d), "=m" (s) \
730  : "m" (s), "m" (d), "m" (c), "m" (b) \
731  : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
732  );
733 
734 #endif /* MIPS */
735 #endif /* GNUC */
736 
737 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
738 
739 #define MULADDC_INIT \
740  __asm mov esi, s \
741  __asm mov edi, d \
742  __asm mov ecx, c \
743  __asm mov ebx, b
744 
745 #define MULADDC_CORE \
746  __asm lodsd \
747  __asm mul ebx \
748  __asm add eax, ecx \
749  __asm adc edx, 0 \
750  __asm add eax, [edi] \
751  __asm adc edx, 0 \
752  __asm mov ecx, edx \
753  __asm stosd
754 
755 #if defined(MBEDTLS_HAVE_SSE2)
756 
757 #define EMIT __asm _emit
758 
759 #define MULADDC_HUIT \
760  EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
761  EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
762  EMIT 0x0F EMIT 0x6E EMIT 0x1F \
763  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
764  EMIT 0x0F EMIT 0x6E EMIT 0x16 \
765  EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
766  EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
767  EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
768  EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
769  EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
770  EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
771  EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
772  EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
773  EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
774  EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
775  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
776  EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
777  EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
778  EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
779  EMIT 0x0F EMIT 0x7E EMIT 0x0F \
780  EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
781  EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
782  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
783  EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
784  EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
785  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
786  EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
787  EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
788  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
789  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
790  EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
791  EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
792  EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
793  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
794  EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
795  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
796  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
797  EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
798  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
799  EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
800  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
801  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
802  EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
803  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
804  EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
805  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
806  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
807  EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
808  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
809  EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
810  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
811  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
812  EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
813  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
814  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
815  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
816  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
817  EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
818  EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
819  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
820  EMIT 0x0F EMIT 0x7E EMIT 0xC9
821 
822 #define MULADDC_STOP \
823  EMIT 0x0F EMIT 0x77 \
824  __asm mov c, ecx \
825  __asm mov d, edi \
826  __asm mov s, esi \
827 
828 #else
829 
830 #define MULADDC_STOP \
831  __asm mov c, ecx \
832  __asm mov d, edi \
833  __asm mov s, esi \
834 
835 #endif /* SSE2 */
836 #endif /* MSVC */
837 
838 #endif /* MBEDTLS_HAVE_ASM */
839 
840 #if !defined(MULADDC_CORE)
841 #if defined(MBEDTLS_HAVE_UDBL)
842 
843 #define MULADDC_INIT \
844 { \
845  mbedtls_t_udbl r; \
846  mbedtls_mpi_uint r0, r1;
847 
848 #define MULADDC_CORE \
849  r = *(s++) * (mbedtls_t_udbl) b; \
850  r0 = (mbedtls_mpi_uint) r; \
851  r1 = (mbedtls_mpi_uint)( r >> biL ); \
852  r0 += c; r1 += (r0 < c); \
853  r0 += *d; r1 += (r0 < *d); \
854  c = r1; *(d++) = r0;
855 
856 #define MULADDC_STOP \
857 }
858 
859 #else
860 #define MULADDC_INIT \
861 { \
862  mbedtls_mpi_uint s0, s1, b0, b1; \
863  mbedtls_mpi_uint r0, r1, rx, ry; \
864  b0 = ( b << biH ) >> biH; \
865  b1 = ( b >> biH );
866 
867 #define MULADDC_CORE \
868  s0 = ( *s << biH ) >> biH; \
869  s1 = ( *s >> biH ); s++; \
870  rx = s0 * b1; r0 = s0 * b0; \
871  ry = s1 * b0; r1 = s1 * b1; \
872  r1 += ( rx >> biH ); \
873  r1 += ( ry >> biH ); \
874  rx <<= biH; ry <<= biH; \
875  r0 += rx; r1 += (r0 < rx); \
876  r0 += ry; r1 += (r0 < ry); \
877  r0 += c; r1 += (r0 < c); \
878  r0 += *d; r1 += (r0 < *d); \
879  c = r1; *(d++) = r0;
880 
881 #define MULADDC_STOP \
882 }
883 
884 #endif /* C (generic) */
885 #endif /* C (longlong) */
886 
887 #endif /* bn_mul.h */
Multi-precision integer library.