mbed TLS v1.3.11
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 POLARSSL_BN_MUL_H
39 #define POLARSSL_BN_MUL_H
40 
41 #include "bignum.h"
42 
43 #if defined(POLARSSL_HAVE_ASM)
44 
45 #if defined(__GNUC__)
46 #if defined(__i386__)
47 
48 #define MULADDC_INIT \
49  asm( \
50  "movl %%ebx, %0 \n\t" \
51  "movl %5, %%esi \n\t" \
52  "movl %6, %%edi \n\t" \
53  "movl %7, %%ecx \n\t" \
54  "movl %8, %%ebx \n\t"
55 
56 #define MULADDC_CORE \
57  "lodsl \n\t" \
58  "mull %%ebx \n\t" \
59  "addl %%ecx, %%eax \n\t" \
60  "adcl $0, %%edx \n\t" \
61  "addl (%%edi), %%eax \n\t" \
62  "adcl $0, %%edx \n\t" \
63  "movl %%edx, %%ecx \n\t" \
64  "stosl \n\t"
65 
66 #if defined(POLARSSL_HAVE_SSE2)
67 
68 #define MULADDC_HUIT \
69  "movd %%ecx, %%mm1 \n\t" \
70  "movd %%ebx, %%mm0 \n\t" \
71  "movd (%%edi), %%mm3 \n\t" \
72  "paddq %%mm3, %%mm1 \n\t" \
73  "movd (%%esi), %%mm2 \n\t" \
74  "pmuludq %%mm0, %%mm2 \n\t" \
75  "movd 4(%%esi), %%mm4 \n\t" \
76  "pmuludq %%mm0, %%mm4 \n\t" \
77  "movd 8(%%esi), %%mm6 \n\t" \
78  "pmuludq %%mm0, %%mm6 \n\t" \
79  "movd 12(%%esi), %%mm7 \n\t" \
80  "pmuludq %%mm0, %%mm7 \n\t" \
81  "paddq %%mm2, %%mm1 \n\t" \
82  "movd 4(%%edi), %%mm3 \n\t" \
83  "paddq %%mm4, %%mm3 \n\t" \
84  "movd 8(%%edi), %%mm5 \n\t" \
85  "paddq %%mm6, %%mm5 \n\t" \
86  "movd 12(%%edi), %%mm4 \n\t" \
87  "paddq %%mm4, %%mm7 \n\t" \
88  "movd %%mm1, (%%edi) \n\t" \
89  "movd 16(%%esi), %%mm2 \n\t" \
90  "pmuludq %%mm0, %%mm2 \n\t" \
91  "psrlq $32, %%mm1 \n\t" \
92  "movd 20(%%esi), %%mm4 \n\t" \
93  "pmuludq %%mm0, %%mm4 \n\t" \
94  "paddq %%mm3, %%mm1 \n\t" \
95  "movd 24(%%esi), %%mm6 \n\t" \
96  "pmuludq %%mm0, %%mm6 \n\t" \
97  "movd %%mm1, 4(%%edi) \n\t" \
98  "psrlq $32, %%mm1 \n\t" \
99  "movd 28(%%esi), %%mm3 \n\t" \
100  "pmuludq %%mm0, %%mm3 \n\t" \
101  "paddq %%mm5, %%mm1 \n\t" \
102  "movd 16(%%edi), %%mm5 \n\t" \
103  "paddq %%mm5, %%mm2 \n\t" \
104  "movd %%mm1, 8(%%edi) \n\t" \
105  "psrlq $32, %%mm1 \n\t" \
106  "paddq %%mm7, %%mm1 \n\t" \
107  "movd 20(%%edi), %%mm5 \n\t" \
108  "paddq %%mm5, %%mm4 \n\t" \
109  "movd %%mm1, 12(%%edi) \n\t" \
110  "psrlq $32, %%mm1 \n\t" \
111  "paddq %%mm2, %%mm1 \n\t" \
112  "movd 24(%%edi), %%mm5 \n\t" \
113  "paddq %%mm5, %%mm6 \n\t" \
114  "movd %%mm1, 16(%%edi) \n\t" \
115  "psrlq $32, %%mm1 \n\t" \
116  "paddq %%mm4, %%mm1 \n\t" \
117  "movd 28(%%edi), %%mm5 \n\t" \
118  "paddq %%mm5, %%mm3 \n\t" \
119  "movd %%mm1, 20(%%edi) \n\t" \
120  "psrlq $32, %%mm1 \n\t" \
121  "paddq %%mm6, %%mm1 \n\t" \
122  "movd %%mm1, 24(%%edi) \n\t" \
123  "psrlq $32, %%mm1 \n\t" \
124  "paddq %%mm3, %%mm1 \n\t" \
125  "movd %%mm1, 28(%%edi) \n\t" \
126  "addl $32, %%edi \n\t" \
127  "addl $32, %%esi \n\t" \
128  "psrlq $32, %%mm1 \n\t" \
129  "movd %%mm1, %%ecx \n\t"
130 
131 #define MULADDC_STOP \
132  "emms \n\t" \
133  "movl %4, %%ebx \n\t" \
134  "movl %%ecx, %1 \n\t" \
135  "movl %%edi, %2 \n\t" \
136  "movl %%esi, %3 \n\t" \
137  : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
138  : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
139  : "eax", "ecx", "edx", "esi", "edi" \
140  );
141 
142 #else
143 
144 #define MULADDC_STOP \
145  "movl %4, %%ebx \n\t" \
146  "movl %%ecx, %1 \n\t" \
147  "movl %%edi, %2 \n\t" \
148  "movl %%esi, %3 \n\t" \
149  : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
150  : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
151  : "eax", "ecx", "edx", "esi", "edi" \
152  );
153 #endif /* SSE2 */
154 #endif /* i386 */
155 
156 #if defined(__amd64__) || defined (__x86_64__)
157 
158 #define MULADDC_INIT \
159  asm( \
160  "movq %3, %%rsi \n\t" \
161  "movq %4, %%rdi \n\t" \
162  "movq %5, %%rcx \n\t" \
163  "movq %6, %%rbx \n\t" \
164  "xorq %%r8, %%r8 \n\t"
165 
166 #define MULADDC_CORE \
167  "movq (%%rsi), %%rax \n\t" \
168  "mulq %%rbx \n\t" \
169  "addq $8, %%rsi \n\t" \
170  "addq %%rcx, %%rax \n\t" \
171  "movq %%r8, %%rcx \n\t" \
172  "adcq $0, %%rdx \n\t" \
173  "nop \n\t" \
174  "addq %%rax, (%%rdi) \n\t" \
175  "adcq %%rdx, %%rcx \n\t" \
176  "addq $8, %%rdi \n\t"
177 
178 #define MULADDC_STOP \
179  "movq %%rcx, %0 \n\t" \
180  "movq %%rdi, %1 \n\t" \
181  "movq %%rsi, %2 \n\t" \
182  : "=m" (c), "=m" (d), "=m" (s) \
183  : "m" (s), "m" (d), "m" (c), "m" (b) \
184  : "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "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 Sparc64 assembly is reported to be broken.
411  * Disable it for now, until we're able to fix it.
412  */
413 #if 0 && defined(__sparc__) && defined(__sparc64__)
414 
415 #define MULADDC_INIT \
416  asm( \
417  "ldx %3, %%o0 \n\t" \
418  "ldx %4, %%o1 \n\t" \
419  "ld %5, %%o2 \n\t" \
420  "ld %6, %%o3 \n\t"
421 
422 #define MULADDC_CORE \
423  "ld [%%o0], %%o4 \n\t" \
424  "inc 4, %%o0 \n\t" \
425  "ld [%%o1], %%o5 \n\t" \
426  "umul %%o3, %%o4, %%o4 \n\t" \
427  "addcc %%o4, %%o2, %%o4 \n\t" \
428  "rd %%y, %%g1 \n\t" \
429  "addx %%g1, 0, %%g1 \n\t" \
430  "addcc %%o4, %%o5, %%o4 \n\t" \
431  "st %%o4, [%%o1] \n\t" \
432  "addx %%g1, 0, %%o2 \n\t" \
433  "inc 4, %%o1 \n\t"
434 
435  #define MULADDC_STOP \
436  "st %%o2, %0 \n\t" \
437  "stx %%o1, %1 \n\t" \
438  "stx %%o0, %2 \n\t" \
439  : "=m" (c), "=m" (d), "=m" (s) \
440  : "m" (s), "m" (d), "m" (c), "m" (b) \
441  : "g1", "o0", "o1", "o2", "o3", "o4", \
442  "o5" \
443  );
444 #endif /* SPARCv9 */
445 
446 #if defined(__sparc__) && !defined(__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 /* SPARCv8 */
479 
480 #if defined(__microblaze__) || defined(microblaze)
481 
482 #define MULADDC_INIT \
483  asm( \
484  "lwi r3, %3 \n\t" \
485  "lwi r4, %4 \n\t" \
486  "lwi r5, %5 \n\t" \
487  "lwi r6, %6 \n\t" \
488  "andi r7, r6, 0xffff \n\t" \
489  "bsrli r6, r6, 16 \n\t"
490 
491 #define MULADDC_CORE \
492  "lhui r8, r3, 0 \n\t" \
493  "addi r3, r3, 2 \n\t" \
494  "lhui r9, r3, 0 \n\t" \
495  "addi r3, r3, 2 \n\t" \
496  "mul r10, r9, r6 \n\t" \
497  "mul r11, r8, r7 \n\t" \
498  "mul r12, r9, r7 \n\t" \
499  "mul r13, r8, r6 \n\t" \
500  "bsrli r8, r10, 16 \n\t" \
501  "bsrli r9, r11, 16 \n\t" \
502  "add r13, r13, r8 \n\t" \
503  "add r13, r13, r9 \n\t" \
504  "bslli r10, r10, 16 \n\t" \
505  "bslli r11, r11, 16 \n\t" \
506  "add r12, r12, r10 \n\t" \
507  "addc r13, r13, r0 \n\t" \
508  "add r12, r12, r11 \n\t" \
509  "addc r13, r13, r0 \n\t" \
510  "lwi r10, r4, 0 \n\t" \
511  "add r12, r12, r10 \n\t" \
512  "addc r13, r13, r0 \n\t" \
513  "add r12, r12, r5 \n\t" \
514  "addc r5, r13, r0 \n\t" \
515  "swi r12, r4, 0 \n\t" \
516  "addi r4, r4, 4 \n\t"
517 
518 #define MULADDC_STOP \
519  "swi r5, %0 \n\t" \
520  "swi r4, %1 \n\t" \
521  "swi r3, %2 \n\t" \
522  : "=m" (c), "=m" (d), "=m" (s) \
523  : "m" (s), "m" (d), "m" (c), "m" (b) \
524  : "r3", "r4" "r5", "r6", "r7", "r8", \
525  "r9", "r10", "r11", "r12", "r13" \
526  );
527 
528 #endif /* MicroBlaze */
529 
530 #if defined(__tricore__)
531 
532 #define MULADDC_INIT \
533  asm( \
534  "ld.a %%a2, %3 \n\t" \
535  "ld.a %%a3, %4 \n\t" \
536  "ld.w %%d4, %5 \n\t" \
537  "ld.w %%d1, %6 \n\t" \
538  "xor %%d5, %%d5 \n\t"
539 
540 #define MULADDC_CORE \
541  "ld.w %%d0, [%%a2+] \n\t" \
542  "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
543  "ld.w %%d0, [%%a3] \n\t" \
544  "addx %%d2, %%d2, %%d0 \n\t" \
545  "addc %%d3, %%d3, 0 \n\t" \
546  "mov %%d4, %%d3 \n\t" \
547  "st.w [%%a3+], %%d2 \n\t"
548 
549 #define MULADDC_STOP \
550  "st.w %0, %%d4 \n\t" \
551  "st.a %1, %%a3 \n\t" \
552  "st.a %2, %%a2 \n\t" \
553  : "=m" (c), "=m" (d), "=m" (s) \
554  : "m" (s), "m" (d), "m" (c), "m" (b) \
555  : "d0", "d1", "e2", "d4", "a2", "a3" \
556  );
557 
558 #endif /* TriCore */
559 
560 #if defined(__arm__)
561 
562 #if defined(__thumb__) && !defined(__thumb2__)
563 
564 #define MULADDC_INIT \
565  asm( \
566  "ldr r0, %3 \n\t" \
567  "ldr r1, %4 \n\t" \
568  "ldr r2, %5 \n\t" \
569  "ldr r3, %6 \n\t" \
570  "lsr r7, r3, #16 \n\t" \
571  "mov r9, r7 \n\t" \
572  "lsl r7, r3, #16 \n\t" \
573  "lsr r7, r7, #16 \n\t" \
574  "mov r8, r7 \n\t"
575 
576 #define MULADDC_CORE \
577  "ldmia r0!, {r6} \n\t" \
578  "lsr r7, r6, #16 \n\t" \
579  "lsl r6, r6, #16 \n\t" \
580  "lsr r6, r6, #16 \n\t" \
581  "mov r4, r8 \n\t" \
582  "mul r4, r6 \n\t" \
583  "mov r3, r9 \n\t" \
584  "mul r6, r3 \n\t" \
585  "mov r5, r9 \n\t" \
586  "mul r5, r7 \n\t" \
587  "mov r3, r8 \n\t" \
588  "mul r7, r3 \n\t" \
589  "lsr r3, r6, #16 \n\t" \
590  "add r5, r5, r3 \n\t" \
591  "lsr r3, r7, #16 \n\t" \
592  "add r5, r5, r3 \n\t" \
593  "add r4, r4, r2 \n\t" \
594  "mov r2, #0 \n\t" \
595  "adc r5, r2 \n\t" \
596  "lsl r3, r6, #16 \n\t" \
597  "add r4, r4, r3 \n\t" \
598  "adc r5, r2 \n\t" \
599  "lsl r3, r7, #16 \n\t" \
600  "add r4, r4, r3 \n\t" \
601  "adc r5, r2 \n\t" \
602  "ldr r3, [r1] \n\t" \
603  "add r4, r4, r3 \n\t" \
604  "adc r2, r5 \n\t" \
605  "stmia r1!, {r4} \n\t"
606 
607 #define MULADDC_STOP \
608  "str r2, %0 \n\t" \
609  "str r1, %1 \n\t" \
610  "str r0, %2 \n\t" \
611  : "=m" (c), "=m" (d), "=m" (s) \
612  : "m" (s), "m" (d), "m" (c), "m" (b) \
613  : "r0", "r1", "r2", "r3", "r4", "r5", \
614  "r6", "r7", "r8", "r9", "cc" \
615  );
616 
617 #else
618 
619 #define MULADDC_INIT \
620  asm( \
621  "ldr r0, %3 \n\t" \
622  "ldr r1, %4 \n\t" \
623  "ldr r2, %5 \n\t" \
624  "ldr r3, %6 \n\t"
625 
626 #define MULADDC_CORE \
627  "ldr r4, [r0], #4 \n\t" \
628  "mov r5, #0 \n\t" \
629  "ldr r6, [r1] \n\t" \
630  "umlal r2, r5, r3, r4 \n\t" \
631  "adds r7, r6, r2 \n\t" \
632  "adc r2, r5, #0 \n\t" \
633  "str r7, [r1], #4 \n\t"
634 
635 #define MULADDC_STOP \
636  "str r2, %0 \n\t" \
637  "str r1, %1 \n\t" \
638  "str r0, %2 \n\t" \
639  : "=m" (c), "=m" (d), "=m" (s) \
640  : "m" (s), "m" (d), "m" (c), "m" (b) \
641  : "r0", "r1", "r2", "r3", "r4", "r5", \
642  "r6", "r7", "cc" \
643  );
644 
645 #endif /* Thumb */
646 
647 #endif /* ARMv3 */
648 
649 #if defined(__alpha__)
650 
651 #define MULADDC_INIT \
652  asm( \
653  "ldq $1, %3 \n\t" \
654  "ldq $2, %4 \n\t" \
655  "ldq $3, %5 \n\t" \
656  "ldq $4, %6 \n\t"
657 
658 #define MULADDC_CORE \
659  "ldq $6, 0($1) \n\t" \
660  "addq $1, 8, $1 \n\t" \
661  "mulq $6, $4, $7 \n\t" \
662  "umulh $6, $4, $6 \n\t" \
663  "addq $7, $3, $7 \n\t" \
664  "cmpult $7, $3, $3 \n\t" \
665  "ldq $5, 0($2) \n\t" \
666  "addq $7, $5, $7 \n\t" \
667  "cmpult $7, $5, $5 \n\t" \
668  "stq $7, 0($2) \n\t" \
669  "addq $2, 8, $2 \n\t" \
670  "addq $6, $3, $3 \n\t" \
671  "addq $5, $3, $3 \n\t"
672 
673 #define MULADDC_STOP \
674  "stq $3, %0 \n\t" \
675  "stq $2, %1 \n\t" \
676  "stq $1, %2 \n\t" \
677  : "=m" (c), "=m" (d), "=m" (s) \
678  : "m" (s), "m" (d), "m" (c), "m" (b) \
679  : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
680  );
681 #endif /* Alpha */
682 
683 #if defined(__mips__) && !defined(__mips64)
684 
685 #define MULADDC_INIT \
686  asm( \
687  "lw $10, %3 \n\t" \
688  "lw $11, %4 \n\t" \
689  "lw $12, %5 \n\t" \
690  "lw $13, %6 \n\t"
691 
692 #define MULADDC_CORE \
693  "lw $14, 0($10) \n\t" \
694  "multu $13, $14 \n\t" \
695  "addi $10, $10, 4 \n\t" \
696  "mflo $14 \n\t" \
697  "mfhi $9 \n\t" \
698  "addu $14, $12, $14 \n\t" \
699  "lw $15, 0($11) \n\t" \
700  "sltu $12, $14, $12 \n\t" \
701  "addu $15, $14, $15 \n\t" \
702  "sltu $14, $15, $14 \n\t" \
703  "addu $12, $12, $9 \n\t" \
704  "sw $15, 0($11) \n\t" \
705  "addu $12, $12, $14 \n\t" \
706  "addi $11, $11, 4 \n\t"
707 
708 #define MULADDC_STOP \
709  "sw $12, %0 \n\t" \
710  "sw $11, %1 \n\t" \
711  "sw $10, %2 \n\t" \
712  : "=m" (c), "=m" (d), "=m" (s) \
713  : "m" (s), "m" (d), "m" (c), "m" (b) \
714  : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
715  );
716 
717 #endif /* MIPS */
718 #endif /* GNUC */
719 
720 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
721 
722 #define MULADDC_INIT \
723  __asm mov esi, s \
724  __asm mov edi, d \
725  __asm mov ecx, c \
726  __asm mov ebx, b
727 
728 #define MULADDC_CORE \
729  __asm lodsd \
730  __asm mul ebx \
731  __asm add eax, ecx \
732  __asm adc edx, 0 \
733  __asm add eax, [edi] \
734  __asm adc edx, 0 \
735  __asm mov ecx, edx \
736  __asm stosd
737 
738 #if defined(POLARSSL_HAVE_SSE2)
739 
740 #define EMIT __asm _emit
741 
742 #define MULADDC_HUIT \
743  EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
744  EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
745  EMIT 0x0F EMIT 0x6E EMIT 0x1F \
746  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
747  EMIT 0x0F EMIT 0x6E EMIT 0x16 \
748  EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
749  EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
750  EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
751  EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
752  EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
753  EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
754  EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
755  EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
756  EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
757  EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
758  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
759  EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
760  EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
761  EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
762  EMIT 0x0F EMIT 0x7E EMIT 0x0F \
763  EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
764  EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
765  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
766  EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
767  EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
768  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
769  EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
770  EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
771  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
772  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
773  EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
774  EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
775  EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
776  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
777  EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
778  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
779  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
780  EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
781  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
782  EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
783  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
784  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
785  EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
786  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
787  EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
788  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
789  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
790  EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
791  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
792  EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
793  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
794  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
795  EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
796  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
797  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
798  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
799  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
800  EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
801  EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
802  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
803  EMIT 0x0F EMIT 0x7E EMIT 0xC9
804 
805 #define MULADDC_STOP \
806  EMIT 0x0F EMIT 0x77 \
807  __asm mov c, ecx \
808  __asm mov d, edi \
809  __asm mov s, esi \
810 
811 #else
812 
813 #define MULADDC_STOP \
814  __asm mov c, ecx \
815  __asm mov d, edi \
816  __asm mov s, esi \
817 
818 #endif /* SSE2 */
819 #endif /* MSVC */
820 
821 #endif /* POLARSSL_HAVE_ASM */
822 
823 #if !defined(MULADDC_CORE)
824 #if defined(POLARSSL_HAVE_UDBL)
825 
826 #define MULADDC_INIT \
827 { \
828  t_udbl r; \
829  t_uint r0, r1;
830 
831 #define MULADDC_CORE \
832  r = *(s++) * (t_udbl) b; \
833  r0 = (t_uint) r; \
834  r1 = (t_uint)( r >> biL ); \
835  r0 += c; r1 += (r0 < c); \
836  r0 += *d; r1 += (r0 < *d); \
837  c = r1; *(d++) = r0;
838 
839 #define MULADDC_STOP \
840 }
841 
842 #else
843 #define MULADDC_INIT \
844 { \
845  t_uint s0, s1, b0, b1; \
846  t_uint r0, r1, rx, ry; \
847  b0 = ( b << biH ) >> biH; \
848  b1 = ( b >> biH );
849 
850 #define MULADDC_CORE \
851  s0 = ( *s << biH ) >> biH; \
852  s1 = ( *s >> biH ); s++; \
853  rx = s0 * b1; r0 = s0 * b0; \
854  ry = s1 * b0; r1 = s1 * b1; \
855  r1 += ( rx >> biH ); \
856  r1 += ( ry >> biH ); \
857  rx <<= biH; ry <<= biH; \
858  r0 += rx; r1 += (r0 < rx); \
859  r0 += ry; r1 += (r0 < ry); \
860  r0 += c; r1 += (r0 < c); \
861  r0 += *d; r1 += (r0 < *d); \
862  c = r1; *(d++) = r0;
863 
864 #define MULADDC_STOP \
865 }
866 
867 #endif /* C (generic) */
868 #endif /* C (longlong) */
869 
870 #endif /* bn_mul.h */
Multi-precision integer library.