mbed TLS v2.0.0
bn_mul.h
Go to the documentation of this file.
1 
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 #if defined(__GNUC__)
50 #if defined(__i386__)
51 
52 #define MULADDC_INIT \
53  asm( \
54  "movl %%ebx, %0 \n\t" \
55  "movl %5, %%esi \n\t" \
56  "movl %6, %%edi \n\t" \
57  "movl %7, %%ecx \n\t" \
58  "movl %8, %%ebx \n\t"
59 
60 #define MULADDC_CORE \
61  "lodsl \n\t" \
62  "mull %%ebx \n\t" \
63  "addl %%ecx, %%eax \n\t" \
64  "adcl $0, %%edx \n\t" \
65  "addl (%%edi), %%eax \n\t" \
66  "adcl $0, %%edx \n\t" \
67  "movl %%edx, %%ecx \n\t" \
68  "stosl \n\t"
69 
70 #if defined(MBEDTLS_HAVE_SSE2)
71 
72 #define MULADDC_HUIT \
73  "movd %%ecx, %%mm1 \n\t" \
74  "movd %%ebx, %%mm0 \n\t" \
75  "movd (%%edi), %%mm3 \n\t" \
76  "paddq %%mm3, %%mm1 \n\t" \
77  "movd (%%esi), %%mm2 \n\t" \
78  "pmuludq %%mm0, %%mm2 \n\t" \
79  "movd 4(%%esi), %%mm4 \n\t" \
80  "pmuludq %%mm0, %%mm4 \n\t" \
81  "movd 8(%%esi), %%mm6 \n\t" \
82  "pmuludq %%mm0, %%mm6 \n\t" \
83  "movd 12(%%esi), %%mm7 \n\t" \
84  "pmuludq %%mm0, %%mm7 \n\t" \
85  "paddq %%mm2, %%mm1 \n\t" \
86  "movd 4(%%edi), %%mm3 \n\t" \
87  "paddq %%mm4, %%mm3 \n\t" \
88  "movd 8(%%edi), %%mm5 \n\t" \
89  "paddq %%mm6, %%mm5 \n\t" \
90  "movd 12(%%edi), %%mm4 \n\t" \
91  "paddq %%mm4, %%mm7 \n\t" \
92  "movd %%mm1, (%%edi) \n\t" \
93  "movd 16(%%esi), %%mm2 \n\t" \
94  "pmuludq %%mm0, %%mm2 \n\t" \
95  "psrlq $32, %%mm1 \n\t" \
96  "movd 20(%%esi), %%mm4 \n\t" \
97  "pmuludq %%mm0, %%mm4 \n\t" \
98  "paddq %%mm3, %%mm1 \n\t" \
99  "movd 24(%%esi), %%mm6 \n\t" \
100  "pmuludq %%mm0, %%mm6 \n\t" \
101  "movd %%mm1, 4(%%edi) \n\t" \
102  "psrlq $32, %%mm1 \n\t" \
103  "movd 28(%%esi), %%mm3 \n\t" \
104  "pmuludq %%mm0, %%mm3 \n\t" \
105  "paddq %%mm5, %%mm1 \n\t" \
106  "movd 16(%%edi), %%mm5 \n\t" \
107  "paddq %%mm5, %%mm2 \n\t" \
108  "movd %%mm1, 8(%%edi) \n\t" \
109  "psrlq $32, %%mm1 \n\t" \
110  "paddq %%mm7, %%mm1 \n\t" \
111  "movd 20(%%edi), %%mm5 \n\t" \
112  "paddq %%mm5, %%mm4 \n\t" \
113  "movd %%mm1, 12(%%edi) \n\t" \
114  "psrlq $32, %%mm1 \n\t" \
115  "paddq %%mm2, %%mm1 \n\t" \
116  "movd 24(%%edi), %%mm5 \n\t" \
117  "paddq %%mm5, %%mm6 \n\t" \
118  "movd %%mm1, 16(%%edi) \n\t" \
119  "psrlq $32, %%mm1 \n\t" \
120  "paddq %%mm4, %%mm1 \n\t" \
121  "movd 28(%%edi), %%mm5 \n\t" \
122  "paddq %%mm5, %%mm3 \n\t" \
123  "movd %%mm1, 20(%%edi) \n\t" \
124  "psrlq $32, %%mm1 \n\t" \
125  "paddq %%mm6, %%mm1 \n\t" \
126  "movd %%mm1, 24(%%edi) \n\t" \
127  "psrlq $32, %%mm1 \n\t" \
128  "paddq %%mm3, %%mm1 \n\t" \
129  "movd %%mm1, 28(%%edi) \n\t" \
130  "addl $32, %%edi \n\t" \
131  "addl $32, %%esi \n\t" \
132  "psrlq $32, %%mm1 \n\t" \
133  "movd %%mm1, %%ecx \n\t"
134 
135 #define MULADDC_STOP \
136  "emms \n\t" \
137  "movl %4, %%ebx \n\t" \
138  "movl %%ecx, %1 \n\t" \
139  "movl %%edi, %2 \n\t" \
140  "movl %%esi, %3 \n\t" \
141  : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
142  : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
143  : "eax", "ecx", "edx", "esi", "edi" \
144  );
145 
146 #else
147 
148 #define MULADDC_STOP \
149  "movl %4, %%ebx \n\t" \
150  "movl %%ecx, %1 \n\t" \
151  "movl %%edi, %2 \n\t" \
152  "movl %%esi, %3 \n\t" \
153  : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
154  : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
155  : "eax", "ecx", "edx", "esi", "edi" \
156  );
157 #endif /* SSE2 */
158 #endif /* i386 */
159 
160 #if defined(__amd64__) || defined (__x86_64__)
161 
162 #define MULADDC_INIT \
163  asm( \
164  "movq %3, %%rsi \n\t" \
165  "movq %4, %%rdi \n\t" \
166  "movq %5, %%rcx \n\t" \
167  "movq %6, %%rbx \n\t" \
168  "xorq %%r8, %%r8 \n\t"
169 
170 #define MULADDC_CORE \
171  "movq (%%rsi), %%rax \n\t" \
172  "mulq %%rbx \n\t" \
173  "addq $8, %%rsi \n\t" \
174  "addq %%rcx, %%rax \n\t" \
175  "movq %%r8, %%rcx \n\t" \
176  "adcq $0, %%rdx \n\t" \
177  "nop \n\t" \
178  "addq %%rax, (%%rdi) \n\t" \
179  "adcq %%rdx, %%rcx \n\t" \
180  "addq $8, %%rdi \n\t"
181 
182 #define MULADDC_STOP \
183  "movq %%rcx, %0 \n\t" \
184  "movq %%rdi, %1 \n\t" \
185  "movq %%rsi, %2 \n\t" \
186  : "=m" (c), "=m" (d), "=m" (s) \
187  : "m" (s), "m" (d), "m" (c), "m" (b) \
188  : "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" \
189  );
190 
191 #endif /* AMD64 */
192 
193 #if defined(__mc68020__) || defined(__mcpu32__)
194 
195 #define MULADDC_INIT \
196  asm( \
197  "movl %3, %%a2 \n\t" \
198  "movl %4, %%a3 \n\t" \
199  "movl %5, %%d3 \n\t" \
200  "movl %6, %%d2 \n\t" \
201  "moveq #0, %%d0 \n\t"
202 
203 #define MULADDC_CORE \
204  "movel %%a2@+, %%d1 \n\t" \
205  "mulul %%d2, %%d4:%%d1 \n\t" \
206  "addl %%d3, %%d1 \n\t" \
207  "addxl %%d0, %%d4 \n\t" \
208  "moveq #0, %%d3 \n\t" \
209  "addl %%d1, %%a3@+ \n\t" \
210  "addxl %%d4, %%d3 \n\t"
211 
212 #define MULADDC_STOP \
213  "movl %%d3, %0 \n\t" \
214  "movl %%a3, %1 \n\t" \
215  "movl %%a2, %2 \n\t" \
216  : "=m" (c), "=m" (d), "=m" (s) \
217  : "m" (s), "m" (d), "m" (c), "m" (b) \
218  : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
219  );
220 
221 #define MULADDC_HUIT \
222  "movel %%a2@+, %%d1 \n\t" \
223  "mulul %%d2, %%d4:%%d1 \n\t" \
224  "addxl %%d3, %%d1 \n\t" \
225  "addxl %%d0, %%d4 \n\t" \
226  "addl %%d1, %%a3@+ \n\t" \
227  "movel %%a2@+, %%d1 \n\t" \
228  "mulul %%d2, %%d3:%%d1 \n\t" \
229  "addxl %%d4, %%d1 \n\t" \
230  "addxl %%d0, %%d3 \n\t" \
231  "addl %%d1, %%a3@+ \n\t" \
232  "movel %%a2@+, %%d1 \n\t" \
233  "mulul %%d2, %%d4:%%d1 \n\t" \
234  "addxl %%d3, %%d1 \n\t" \
235  "addxl %%d0, %%d4 \n\t" \
236  "addl %%d1, %%a3@+ \n\t" \
237  "movel %%a2@+, %%d1 \n\t" \
238  "mulul %%d2, %%d3:%%d1 \n\t" \
239  "addxl %%d4, %%d1 \n\t" \
240  "addxl %%d0, %%d3 \n\t" \
241  "addl %%d1, %%a3@+ \n\t" \
242  "movel %%a2@+, %%d1 \n\t" \
243  "mulul %%d2, %%d4:%%d1 \n\t" \
244  "addxl %%d3, %%d1 \n\t" \
245  "addxl %%d0, %%d4 \n\t" \
246  "addl %%d1, %%a3@+ \n\t" \
247  "movel %%a2@+, %%d1 \n\t" \
248  "mulul %%d2, %%d3:%%d1 \n\t" \
249  "addxl %%d4, %%d1 \n\t" \
250  "addxl %%d0, %%d3 \n\t" \
251  "addl %%d1, %%a3@+ \n\t" \
252  "movel %%a2@+, %%d1 \n\t" \
253  "mulul %%d2, %%d4:%%d1 \n\t" \
254  "addxl %%d3, %%d1 \n\t" \
255  "addxl %%d0, %%d4 \n\t" \
256  "addl %%d1, %%a3@+ \n\t" \
257  "movel %%a2@+, %%d1 \n\t" \
258  "mulul %%d2, %%d3:%%d1 \n\t" \
259  "addxl %%d4, %%d1 \n\t" \
260  "addxl %%d0, %%d3 \n\t" \
261  "addl %%d1, %%a3@+ \n\t" \
262  "addxl %%d0, %%d3 \n\t"
263 
264 #endif /* MC68000 */
265 
266 #if defined(__powerpc64__) || defined(__ppc64__)
267 
268 #if defined(__MACH__) && defined(__APPLE__)
269 
270 #define MULADDC_INIT \
271  asm( \
272  "ld r3, %3 \n\t" \
273  "ld r4, %4 \n\t" \
274  "ld r5, %5 \n\t" \
275  "ld r6, %6 \n\t" \
276  "addi r3, r3, -8 \n\t" \
277  "addi r4, r4, -8 \n\t" \
278  "addic r5, r5, 0 \n\t"
279 
280 #define MULADDC_CORE \
281  "ldu r7, 8(r3) \n\t" \
282  "mulld r8, r7, r6 \n\t" \
283  "mulhdu r9, r7, r6 \n\t" \
284  "adde r8, r8, r5 \n\t" \
285  "ld r7, 8(r4) \n\t" \
286  "addze r5, r9 \n\t" \
287  "addc r8, r8, r7 \n\t" \
288  "stdu r8, 8(r4) \n\t"
289 
290 #define MULADDC_STOP \
291  "addze r5, r5 \n\t" \
292  "addi r4, r4, 8 \n\t" \
293  "addi r3, r3, 8 \n\t" \
294  "std r5, %0 \n\t" \
295  "std r4, %1 \n\t" \
296  "std r3, %2 \n\t" \
297  : "=m" (c), "=m" (d), "=m" (s) \
298  : "m" (s), "m" (d), "m" (c), "m" (b) \
299  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
300  );
301 
302 
303 #else /* __MACH__ && __APPLE__ */
304 
305 #define MULADDC_INIT \
306  asm( \
307  "ld %%r3, %3 \n\t" \
308  "ld %%r4, %4 \n\t" \
309  "ld %%r5, %5 \n\t" \
310  "ld %%r6, %6 \n\t" \
311  "addi %%r3, %%r3, -8 \n\t" \
312  "addi %%r4, %%r4, -8 \n\t" \
313  "addic %%r5, %%r5, 0 \n\t"
314 
315 #define MULADDC_CORE \
316  "ldu %%r7, 8(%%r3) \n\t" \
317  "mulld %%r8, %%r7, %%r6 \n\t" \
318  "mulhdu %%r9, %%r7, %%r6 \n\t" \
319  "adde %%r8, %%r8, %%r5 \n\t" \
320  "ld %%r7, 8(%%r4) \n\t" \
321  "addze %%r5, %%r9 \n\t" \
322  "addc %%r8, %%r8, %%r7 \n\t" \
323  "stdu %%r8, 8(%%r4) \n\t"
324 
325 #define MULADDC_STOP \
326  "addze %%r5, %%r5 \n\t" \
327  "addi %%r4, %%r4, 8 \n\t" \
328  "addi %%r3, %%r3, 8 \n\t" \
329  "std %%r5, %0 \n\t" \
330  "std %%r4, %1 \n\t" \
331  "std %%r3, %2 \n\t" \
332  : "=m" (c), "=m" (d), "=m" (s) \
333  : "m" (s), "m" (d), "m" (c), "m" (b) \
334  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
335  );
336 
337 #endif /* __MACH__ && __APPLE__ */
338 
339 #elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
340 
341 #if defined(__MACH__) && defined(__APPLE__)
342 
343 #define MULADDC_INIT \
344  asm( \
345  "lwz r3, %3 \n\t" \
346  "lwz r4, %4 \n\t" \
347  "lwz r5, %5 \n\t" \
348  "lwz r6, %6 \n\t" \
349  "addi r3, r3, -4 \n\t" \
350  "addi r4, r4, -4 \n\t" \
351  "addic r5, r5, 0 \n\t"
352 
353 #define MULADDC_CORE \
354  "lwzu r7, 4(r3) \n\t" \
355  "mullw r8, r7, r6 \n\t" \
356  "mulhwu r9, r7, r6 \n\t" \
357  "adde r8, r8, r5 \n\t" \
358  "lwz r7, 4(r4) \n\t" \
359  "addze r5, r9 \n\t" \
360  "addc r8, r8, r7 \n\t" \
361  "stwu r8, 4(r4) \n\t"
362 
363 #define MULADDC_STOP \
364  "addze r5, r5 \n\t" \
365  "addi r4, r4, 4 \n\t" \
366  "addi r3, r3, 4 \n\t" \
367  "stw r5, %0 \n\t" \
368  "stw r4, %1 \n\t" \
369  "stw r3, %2 \n\t" \
370  : "=m" (c), "=m" (d), "=m" (s) \
371  : "m" (s), "m" (d), "m" (c), "m" (b) \
372  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
373  );
374 
375 #else /* __MACH__ && __APPLE__ */
376 
377 #define MULADDC_INIT \
378  asm( \
379  "lwz %%r3, %3 \n\t" \
380  "lwz %%r4, %4 \n\t" \
381  "lwz %%r5, %5 \n\t" \
382  "lwz %%r6, %6 \n\t" \
383  "addi %%r3, %%r3, -4 \n\t" \
384  "addi %%r4, %%r4, -4 \n\t" \
385  "addic %%r5, %%r5, 0 \n\t"
386 
387 #define MULADDC_CORE \
388  "lwzu %%r7, 4(%%r3) \n\t" \
389  "mullw %%r8, %%r7, %%r6 \n\t" \
390  "mulhwu %%r9, %%r7, %%r6 \n\t" \
391  "adde %%r8, %%r8, %%r5 \n\t" \
392  "lwz %%r7, 4(%%r4) \n\t" \
393  "addze %%r5, %%r9 \n\t" \
394  "addc %%r8, %%r8, %%r7 \n\t" \
395  "stwu %%r8, 4(%%r4) \n\t"
396 
397 #define MULADDC_STOP \
398  "addze %%r5, %%r5 \n\t" \
399  "addi %%r4, %%r4, 4 \n\t" \
400  "addi %%r3, %%r3, 4 \n\t" \
401  "stw %%r5, %0 \n\t" \
402  "stw %%r4, %1 \n\t" \
403  "stw %%r3, %2 \n\t" \
404  : "=m" (c), "=m" (d), "=m" (s) \
405  : "m" (s), "m" (d), "m" (c), "m" (b) \
406  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
407  );
408 
409 #endif /* __MACH__ && __APPLE__ */
410 
411 #endif /* PPC32 */
412 
413 /*
414  * The Sparc64 assembly is reported to be broken.
415  * Disable it for now, until we're able to fix it.
416  */
417 #if 0 && defined(__sparc__) && defined(__sparc64__)
418 
419 #define MULADDC_INIT \
420  asm( \
421  "ldx %3, %%o0 \n\t" \
422  "ldx %4, %%o1 \n\t" \
423  "ld %5, %%o2 \n\t" \
424  "ld %6, %%o3 \n\t"
425 
426 #define MULADDC_CORE \
427  "ld [%%o0], %%o4 \n\t" \
428  "inc 4, %%o0 \n\t" \
429  "ld [%%o1], %%o5 \n\t" \
430  "umul %%o3, %%o4, %%o4 \n\t" \
431  "addcc %%o4, %%o2, %%o4 \n\t" \
432  "rd %%y, %%g1 \n\t" \
433  "addx %%g1, 0, %%g1 \n\t" \
434  "addcc %%o4, %%o5, %%o4 \n\t" \
435  "st %%o4, [%%o1] \n\t" \
436  "addx %%g1, 0, %%o2 \n\t" \
437  "inc 4, %%o1 \n\t"
438 
439  #define MULADDC_STOP \
440  "st %%o2, %0 \n\t" \
441  "stx %%o1, %1 \n\t" \
442  "stx %%o0, %2 \n\t" \
443  : "=m" (c), "=m" (d), "=m" (s) \
444  : "m" (s), "m" (d), "m" (c), "m" (b) \
445  : "g1", "o0", "o1", "o2", "o3", "o4", \
446  "o5" \
447  );
448 #endif /* SPARCv9 */
449 
450 #if defined(__sparc__) && !defined(__sparc64__)
451 
452 #define MULADDC_INIT \
453  asm( \
454  "ld %3, %%o0 \n\t" \
455  "ld %4, %%o1 \n\t" \
456  "ld %5, %%o2 \n\t" \
457  "ld %6, %%o3 \n\t"
458 
459 #define MULADDC_CORE \
460  "ld [%%o0], %%o4 \n\t" \
461  "inc 4, %%o0 \n\t" \
462  "ld [%%o1], %%o5 \n\t" \
463  "umul %%o3, %%o4, %%o4 \n\t" \
464  "addcc %%o4, %%o2, %%o4 \n\t" \
465  "rd %%y, %%g1 \n\t" \
466  "addx %%g1, 0, %%g1 \n\t" \
467  "addcc %%o4, %%o5, %%o4 \n\t" \
468  "st %%o4, [%%o1] \n\t" \
469  "addx %%g1, 0, %%o2 \n\t" \
470  "inc 4, %%o1 \n\t"
471 
472 #define MULADDC_STOP \
473  "st %%o2, %0 \n\t" \
474  "st %%o1, %1 \n\t" \
475  "st %%o0, %2 \n\t" \
476  : "=m" (c), "=m" (d), "=m" (s) \
477  : "m" (s), "m" (d), "m" (c), "m" (b) \
478  : "g1", "o0", "o1", "o2", "o3", "o4", \
479  "o5" \
480  );
481 
482 #endif /* SPARCv8 */
483 
484 #if defined(__microblaze__) || defined(microblaze)
485 
486 #define MULADDC_INIT \
487  asm( \
488  "lwi r3, %3 \n\t" \
489  "lwi r4, %4 \n\t" \
490  "lwi r5, %5 \n\t" \
491  "lwi r6, %6 \n\t" \
492  "andi r7, r6, 0xffff \n\t" \
493  "bsrli r6, r6, 16 \n\t"
494 
495 #define MULADDC_CORE \
496  "lhui r8, r3, 0 \n\t" \
497  "addi r3, r3, 2 \n\t" \
498  "lhui r9, r3, 0 \n\t" \
499  "addi r3, r3, 2 \n\t" \
500  "mul r10, r9, r6 \n\t" \
501  "mul r11, r8, r7 \n\t" \
502  "mul r12, r9, r7 \n\t" \
503  "mul r13, r8, r6 \n\t" \
504  "bsrli r8, r10, 16 \n\t" \
505  "bsrli r9, r11, 16 \n\t" \
506  "add r13, r13, r8 \n\t" \
507  "add r13, r13, r9 \n\t" \
508  "bslli r10, r10, 16 \n\t" \
509  "bslli r11, r11, 16 \n\t" \
510  "add r12, r12, r10 \n\t" \
511  "addc r13, r13, r0 \n\t" \
512  "add r12, r12, r11 \n\t" \
513  "addc r13, r13, r0 \n\t" \
514  "lwi r10, r4, 0 \n\t" \
515  "add r12, r12, r10 \n\t" \
516  "addc r13, r13, r0 \n\t" \
517  "add r12, r12, r5 \n\t" \
518  "addc r5, r13, r0 \n\t" \
519  "swi r12, r4, 0 \n\t" \
520  "addi r4, r4, 4 \n\t"
521 
522 #define MULADDC_STOP \
523  "swi r5, %0 \n\t" \
524  "swi r4, %1 \n\t" \
525  "swi r3, %2 \n\t" \
526  : "=m" (c), "=m" (d), "=m" (s) \
527  : "m" (s), "m" (d), "m" (c), "m" (b) \
528  : "r3", "r4" "r5", "r6", "r7", "r8", \
529  "r9", "r10", "r11", "r12", "r13" \
530  );
531 
532 #endif /* MicroBlaze */
533 
534 #if defined(__tricore__)
535 
536 #define MULADDC_INIT \
537  asm( \
538  "ld.a %%a2, %3 \n\t" \
539  "ld.a %%a3, %4 \n\t" \
540  "ld.w %%d4, %5 \n\t" \
541  "ld.w %%d1, %6 \n\t" \
542  "xor %%d5, %%d5 \n\t"
543 
544 #define MULADDC_CORE \
545  "ld.w %%d0, [%%a2+] \n\t" \
546  "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
547  "ld.w %%d0, [%%a3] \n\t" \
548  "addx %%d2, %%d2, %%d0 \n\t" \
549  "addc %%d3, %%d3, 0 \n\t" \
550  "mov %%d4, %%d3 \n\t" \
551  "st.w [%%a3+], %%d2 \n\t"
552 
553 #define MULADDC_STOP \
554  "st.w %0, %%d4 \n\t" \
555  "st.a %1, %%a3 \n\t" \
556  "st.a %2, %%a2 \n\t" \
557  : "=m" (c), "=m" (d), "=m" (s) \
558  : "m" (s), "m" (d), "m" (c), "m" (b) \
559  : "d0", "d1", "e2", "d4", "a2", "a3" \
560  );
561 
562 #endif /* TriCore */
563 
564 #if defined(__arm__)
565 
566 #if defined(__thumb__) && !defined(__thumb2__)
567 
568 #define MULADDC_INIT \
569  asm( \
570  "ldr r0, %3 \n\t" \
571  "ldr r1, %4 \n\t" \
572  "ldr r2, %5 \n\t" \
573  "ldr r3, %6 \n\t" \
574  "lsr r7, r3, #16 \n\t" \
575  "mov r9, r7 \n\t" \
576  "lsl r7, r3, #16 \n\t" \
577  "lsr r7, r7, #16 \n\t" \
578  "mov r8, r7 \n\t"
579 
580 #define MULADDC_CORE \
581  "ldmia r0!, {r6} \n\t" \
582  "lsr r7, r6, #16 \n\t" \
583  "lsl r6, r6, #16 \n\t" \
584  "lsr r6, r6, #16 \n\t" \
585  "mov r4, r8 \n\t" \
586  "mul r4, r6 \n\t" \
587  "mov r3, r9 \n\t" \
588  "mul r6, r3 \n\t" \
589  "mov r5, r9 \n\t" \
590  "mul r5, r7 \n\t" \
591  "mov r3, r8 \n\t" \
592  "mul r7, r3 \n\t" \
593  "lsr r3, r6, #16 \n\t" \
594  "add r5, r5, r3 \n\t" \
595  "lsr r3, r7, #16 \n\t" \
596  "add r5, r5, r3 \n\t" \
597  "add r4, r4, r2 \n\t" \
598  "mov r2, #0 \n\t" \
599  "adc r5, r2 \n\t" \
600  "lsl r3, r6, #16 \n\t" \
601  "add r4, r4, r3 \n\t" \
602  "adc r5, r2 \n\t" \
603  "lsl r3, r7, #16 \n\t" \
604  "add r4, r4, r3 \n\t" \
605  "adc r5, r2 \n\t" \
606  "ldr r3, [r1] \n\t" \
607  "add r4, r4, r3 \n\t" \
608  "adc r2, r5 \n\t" \
609  "stmia r1!, {r4} \n\t"
610 
611 #define MULADDC_STOP \
612  "str r2, %0 \n\t" \
613  "str r1, %1 \n\t" \
614  "str r0, %2 \n\t" \
615  : "=m" (c), "=m" (d), "=m" (s) \
616  : "m" (s), "m" (d), "m" (c), "m" (b) \
617  : "r0", "r1", "r2", "r3", "r4", "r5", \
618  "r6", "r7", "r8", "r9", "cc" \
619  );
620 
621 #else
622 
623 #define MULADDC_INIT \
624  asm( \
625  "ldr r0, %3 \n\t" \
626  "ldr r1, %4 \n\t" \
627  "ldr r2, %5 \n\t" \
628  "ldr r3, %6 \n\t"
629 
630 #define MULADDC_CORE \
631  "ldr r4, [r0], #4 \n\t" \
632  "mov r5, #0 \n\t" \
633  "ldr r6, [r1] \n\t" \
634  "umlal r2, r5, r3, r4 \n\t" \
635  "adds r7, r6, r2 \n\t" \
636  "adc r2, r5, #0 \n\t" \
637  "str r7, [r1], #4 \n\t"
638 
639 #define MULADDC_STOP \
640  "str r2, %0 \n\t" \
641  "str r1, %1 \n\t" \
642  "str r0, %2 \n\t" \
643  : "=m" (c), "=m" (d), "=m" (s) \
644  : "m" (s), "m" (d), "m" (c), "m" (b) \
645  : "r0", "r1", "r2", "r3", "r4", "r5", \
646  "r6", "r7", "cc" \
647  );
648 
649 #endif /* Thumb */
650 
651 #endif /* ARMv3 */
652 
653 #if defined(__alpha__)
654 
655 #define MULADDC_INIT \
656  asm( \
657  "ldq $1, %3 \n\t" \
658  "ldq $2, %4 \n\t" \
659  "ldq $3, %5 \n\t" \
660  "ldq $4, %6 \n\t"
661 
662 #define MULADDC_CORE \
663  "ldq $6, 0($1) \n\t" \
664  "addq $1, 8, $1 \n\t" \
665  "mulq $6, $4, $7 \n\t" \
666  "umulh $6, $4, $6 \n\t" \
667  "addq $7, $3, $7 \n\t" \
668  "cmpult $7, $3, $3 \n\t" \
669  "ldq $5, 0($2) \n\t" \
670  "addq $7, $5, $7 \n\t" \
671  "cmpult $7, $5, $5 \n\t" \
672  "stq $7, 0($2) \n\t" \
673  "addq $2, 8, $2 \n\t" \
674  "addq $6, $3, $3 \n\t" \
675  "addq $5, $3, $3 \n\t"
676 
677 #define MULADDC_STOP \
678  "stq $3, %0 \n\t" \
679  "stq $2, %1 \n\t" \
680  "stq $1, %2 \n\t" \
681  : "=m" (c), "=m" (d), "=m" (s) \
682  : "m" (s), "m" (d), "m" (c), "m" (b) \
683  : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
684  );
685 #endif /* Alpha */
686 
687 #if defined(__mips__) && !defined(__mips64)
688 
689 #define MULADDC_INIT \
690  asm( \
691  "lw $10, %3 \n\t" \
692  "lw $11, %4 \n\t" \
693  "lw $12, %5 \n\t" \
694  "lw $13, %6 \n\t"
695 
696 #define MULADDC_CORE \
697  "lw $14, 0($10) \n\t" \
698  "multu $13, $14 \n\t" \
699  "addi $10, $10, 4 \n\t" \
700  "mflo $14 \n\t" \
701  "mfhi $9 \n\t" \
702  "addu $14, $12, $14 \n\t" \
703  "lw $15, 0($11) \n\t" \
704  "sltu $12, $14, $12 \n\t" \
705  "addu $15, $14, $15 \n\t" \
706  "sltu $14, $15, $14 \n\t" \
707  "addu $12, $12, $9 \n\t" \
708  "sw $15, 0($11) \n\t" \
709  "addu $12, $12, $14 \n\t" \
710  "addi $11, $11, 4 \n\t"
711 
712 #define MULADDC_STOP \
713  "sw $12, %0 \n\t" \
714  "sw $11, %1 \n\t" \
715  "sw $10, %2 \n\t" \
716  : "=m" (c), "=m" (d), "=m" (s) \
717  : "m" (s), "m" (d), "m" (c), "m" (b) \
718  : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
719  );
720 
721 #endif /* MIPS */
722 #endif /* GNUC */
723 
724 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
725 
726 #define MULADDC_INIT \
727  __asm mov esi, s \
728  __asm mov edi, d \
729  __asm mov ecx, c \
730  __asm mov ebx, b
731 
732 #define MULADDC_CORE \
733  __asm lodsd \
734  __asm mul ebx \
735  __asm add eax, ecx \
736  __asm adc edx, 0 \
737  __asm add eax, [edi] \
738  __asm adc edx, 0 \
739  __asm mov ecx, edx \
740  __asm stosd
741 
742 #if defined(MBEDTLS_HAVE_SSE2)
743 
744 #define EMIT __asm _emit
745 
746 #define MULADDC_HUIT \
747  EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
748  EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
749  EMIT 0x0F EMIT 0x6E EMIT 0x1F \
750  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
751  EMIT 0x0F EMIT 0x6E EMIT 0x16 \
752  EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
753  EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
754  EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
755  EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
756  EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
757  EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
758  EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
759  EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
760  EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
761  EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
762  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
763  EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
764  EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
765  EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
766  EMIT 0x0F EMIT 0x7E EMIT 0x0F \
767  EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
768  EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
769  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
770  EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
771  EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
772  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
773  EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
774  EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
775  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
776  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
777  EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
778  EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
779  EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
780  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
781  EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
782  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
783  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
784  EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
785  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
786  EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
787  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
788  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
789  EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
790  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
791  EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
792  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
793  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
794  EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
795  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
796  EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
797  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
798  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
799  EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
800  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
801  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
802  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
803  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
804  EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
805  EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
806  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
807  EMIT 0x0F EMIT 0x7E EMIT 0xC9
808 
809 #define MULADDC_STOP \
810  EMIT 0x0F EMIT 0x77 \
811  __asm mov c, ecx \
812  __asm mov d, edi \
813  __asm mov s, esi \
814 
815 #else
816 
817 #define MULADDC_STOP \
818  __asm mov c, ecx \
819  __asm mov d, edi \
820  __asm mov s, esi \
821 
822 #endif /* SSE2 */
823 #endif /* MSVC */
824 
825 #endif /* MBEDTLS_HAVE_ASM */
826 
827 #if !defined(MULADDC_CORE)
828 #if defined(MBEDTLS_HAVE_UDBL)
829 
830 #define MULADDC_INIT \
831 { \
832  mbedtls_t_udbl r; \
833  mbedtls_mpi_uint r0, r1;
834 
835 #define MULADDC_CORE \
836  r = *(s++) * (mbedtls_t_udbl) b; \
837  r0 = (mbedtls_mpi_uint) r; \
838  r1 = (mbedtls_mpi_uint)( r >> biL ); \
839  r0 += c; r1 += (r0 < c); \
840  r0 += *d; r1 += (r0 < *d); \
841  c = r1; *(d++) = r0;
842 
843 #define MULADDC_STOP \
844 }
845 
846 #else
847 #define MULADDC_INIT \
848 { \
849  mbedtls_mpi_uint s0, s1, b0, b1; \
850  mbedtls_mpi_uint r0, r1, rx, ry; \
851  b0 = ( b << biH ) >> biH; \
852  b1 = ( b >> biH );
853 
854 #define MULADDC_CORE \
855  s0 = ( *s << biH ) >> biH; \
856  s1 = ( *s >> biH ); s++; \
857  rx = s0 * b1; r0 = s0 * b0; \
858  ry = s1 * b0; r1 = s1 * b1; \
859  r1 += ( rx >> biH ); \
860  r1 += ( ry >> biH ); \
861  rx <<= biH; ry <<= biH; \
862  r0 += rx; r1 += (r0 < rx); \
863  r0 += ry; r1 += (r0 < ry); \
864  r0 += c; r1 += (r0 < c); \
865  r0 += *d; r1 += (r0 < *d); \
866  c = r1; *(d++) = r0;
867 
868 #define MULADDC_STOP \
869 }
870 
871 #endif /* C (generic) */
872 #endif /* C (longlong) */
873 
874 #endif /* bn_mul.h */
Multi-precision integer library.