PolarSSL v1.2.11
bn_mul.h
Go to the documentation of this file.
1 
27 /*
28  * Multiply source vector [s] with b, add result
29  * to destination vector [d] and set carry c.
30  *
31  * Currently supports:
32  *
33  * . IA-32 (386+) . AMD64 / EM64T
34  * . IA-32 (SSE2) . Motorola 68000
35  * . PowerPC, 32-bit . MicroBlaze
36  * . PowerPC, 64-bit . TriCore
37  * . SPARC v8 . ARM v3+
38  * . Alpha . MIPS32
39  * . C, longlong . C, generic
40  */
41 #ifndef POLARSSL_BN_MUL_H
42 #define POLARSSL_BN_MUL_H
43 
44 #include "bignum.h"
45 
46 #if defined(POLARSSL_HAVE_ASM)
47 
48 #if defined(__GNUC__)
49 #if defined(__i386__)
50 
51 #define MULADDC_INIT \
52  asm( \
53  "movl %%ebx, %0 \n\t" \
54  "movl %5, %%esi \n\t" \
55  "movl %6, %%edi \n\t" \
56  "movl %7, %%ecx \n\t" \
57  "movl %8, %%ebx \n\t"
58 
59 #define MULADDC_CORE \
60  "lodsl \n\t" \
61  "mull %%ebx \n\t" \
62  "addl %%ecx, %%eax \n\t" \
63  "adcl $0, %%edx \n\t" \
64  "addl (%%edi), %%eax \n\t" \
65  "adcl $0, %%edx \n\t" \
66  "movl %%edx, %%ecx \n\t" \
67  "stosl \n\t"
68 
69 #if defined(POLARSSL_HAVE_SSE2)
70 
71 #define MULADDC_HUIT \
72  "movd %%ecx, %%mm1 \n\t" \
73  "movd %%ebx, %%mm0 \n\t" \
74  "movd (%%edi), %%mm3 \n\t" \
75  "paddq %%mm3, %%mm1 \n\t" \
76  "movd (%%esi), %%mm2 \n\t" \
77  "pmuludq %%mm0, %%mm2 \n\t" \
78  "movd 4(%%esi), %%mm4 \n\t" \
79  "pmuludq %%mm0, %%mm4 \n\t" \
80  "movd 8(%%esi), %%mm6 \n\t" \
81  "pmuludq %%mm0, %%mm6 \n\t" \
82  "movd 12(%%esi), %%mm7 \n\t" \
83  "pmuludq %%mm0, %%mm7 \n\t" \
84  "paddq %%mm2, %%mm1 \n\t" \
85  "movd 4(%%edi), %%mm3 \n\t" \
86  "paddq %%mm4, %%mm3 \n\t" \
87  "movd 8(%%edi), %%mm5 \n\t" \
88  "paddq %%mm6, %%mm5 \n\t" \
89  "movd 12(%%edi), %%mm4 \n\t" \
90  "paddq %%mm4, %%mm7 \n\t" \
91  "movd %%mm1, (%%edi) \n\t" \
92  "movd 16(%%esi), %%mm2 \n\t" \
93  "pmuludq %%mm0, %%mm2 \n\t" \
94  "psrlq $32, %%mm1 \n\t" \
95  "movd 20(%%esi), %%mm4 \n\t" \
96  "pmuludq %%mm0, %%mm4 \n\t" \
97  "paddq %%mm3, %%mm1 \n\t" \
98  "movd 24(%%esi), %%mm6 \n\t" \
99  "pmuludq %%mm0, %%mm6 \n\t" \
100  "movd %%mm1, 4(%%edi) \n\t" \
101  "psrlq $32, %%mm1 \n\t" \
102  "movd 28(%%esi), %%mm3 \n\t" \
103  "pmuludq %%mm0, %%mm3 \n\t" \
104  "paddq %%mm5, %%mm1 \n\t" \
105  "movd 16(%%edi), %%mm5 \n\t" \
106  "paddq %%mm5, %%mm2 \n\t" \
107  "movd %%mm1, 8(%%edi) \n\t" \
108  "psrlq $32, %%mm1 \n\t" \
109  "paddq %%mm7, %%mm1 \n\t" \
110  "movd 20(%%edi), %%mm5 \n\t" \
111  "paddq %%mm5, %%mm4 \n\t" \
112  "movd %%mm1, 12(%%edi) \n\t" \
113  "psrlq $32, %%mm1 \n\t" \
114  "paddq %%mm2, %%mm1 \n\t" \
115  "movd 24(%%edi), %%mm5 \n\t" \
116  "paddq %%mm5, %%mm6 \n\t" \
117  "movd %%mm1, 16(%%edi) \n\t" \
118  "psrlq $32, %%mm1 \n\t" \
119  "paddq %%mm4, %%mm1 \n\t" \
120  "movd 28(%%edi), %%mm5 \n\t" \
121  "paddq %%mm5, %%mm3 \n\t" \
122  "movd %%mm1, 20(%%edi) \n\t" \
123  "psrlq $32, %%mm1 \n\t" \
124  "paddq %%mm6, %%mm1 \n\t" \
125  "movd %%mm1, 24(%%edi) \n\t" \
126  "psrlq $32, %%mm1 \n\t" \
127  "paddq %%mm3, %%mm1 \n\t" \
128  "movd %%mm1, 28(%%edi) \n\t" \
129  "addl $32, %%edi \n\t" \
130  "addl $32, %%esi \n\t" \
131  "psrlq $32, %%mm1 \n\t" \
132  "movd %%mm1, %%ecx \n\t"
133 
134 #define MULADDC_STOP \
135  "emms \n\t" \
136  "movl %4, %%ebx \n\t" \
137  "movl %%ecx, %1 \n\t" \
138  "movl %%edi, %2 \n\t" \
139  "movl %%esi, %3 \n\t" \
140  : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
141  : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
142  : "eax", "ecx", "edx", "esi", "edi" \
143  );
144 
145 #else
146 
147 #define MULADDC_STOP \
148  "movl %4, %%ebx \n\t" \
149  "movl %%ecx, %1 \n\t" \
150  "movl %%edi, %2 \n\t" \
151  "movl %%esi, %3 \n\t" \
152  : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
153  : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
154  : "eax", "ecx", "edx", "esi", "edi" \
155  );
156 #endif /* SSE2 */
157 #endif /* i386 */
158 
159 #if defined(__amd64__) || defined (__x86_64__)
160 
161 #define MULADDC_INIT \
162  asm( \
163  "movq %3, %%rsi \n\t" \
164  "movq %4, %%rdi \n\t" \
165  "movq %5, %%rcx \n\t" \
166  "movq %6, %%rbx \n\t" \
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  "movq %%rcx, %0 \n\t" \
183  "movq %%rdi, %1 \n\t" \
184  "movq %%rsi, %2 \n\t" \
185  : "=m" (c), "=m" (d), "=m" (s) \
186  : "m" (s), "m" (d), "m" (c), "m" (b) \
187  : "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" \
188  );
189 
190 #endif /* AMD64 */
191 
192 #if defined(__mc68020__) || defined(__mcpu32__)
193 
194 #define MULADDC_INIT \
195  asm( \
196  "movl %3, %%a2 \n\t" \
197  "movl %4, %%a3 \n\t" \
198  "movl %5, %%d3 \n\t" \
199  "movl %6, %%d2 \n\t" \
200  "moveq #0, %%d0 \n\t"
201 
202 #define MULADDC_CORE \
203  "movel %%a2@+, %%d1 \n\t" \
204  "mulul %%d2, %%d4:%%d1 \n\t" \
205  "addl %%d3, %%d1 \n\t" \
206  "addxl %%d0, %%d4 \n\t" \
207  "moveq #0, %%d3 \n\t" \
208  "addl %%d1, %%a3@+ \n\t" \
209  "addxl %%d4, %%d3 \n\t"
210 
211 #define MULADDC_STOP \
212  "movl %%d3, %0 \n\t" \
213  "movl %%a3, %1 \n\t" \
214  "movl %%a2, %2 \n\t" \
215  : "=m" (c), "=m" (d), "=m" (s) \
216  : "m" (s), "m" (d), "m" (c), "m" (b) \
217  : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
218  );
219 
220 #define MULADDC_HUIT \
221  "movel %%a2@+, %%d1 \n\t" \
222  "mulul %%d2, %%d4:%%d1 \n\t" \
223  "addxl %%d3, %%d1 \n\t" \
224  "addxl %%d0, %%d4 \n\t" \
225  "addl %%d1, %%a3@+ \n\t" \
226  "movel %%a2@+, %%d1 \n\t" \
227  "mulul %%d2, %%d3:%%d1 \n\t" \
228  "addxl %%d4, %%d1 \n\t" \
229  "addxl %%d0, %%d3 \n\t" \
230  "addl %%d1, %%a3@+ \n\t" \
231  "movel %%a2@+, %%d1 \n\t" \
232  "mulul %%d2, %%d4:%%d1 \n\t" \
233  "addxl %%d3, %%d1 \n\t" \
234  "addxl %%d0, %%d4 \n\t" \
235  "addl %%d1, %%a3@+ \n\t" \
236  "movel %%a2@+, %%d1 \n\t" \
237  "mulul %%d2, %%d3:%%d1 \n\t" \
238  "addxl %%d4, %%d1 \n\t" \
239  "addxl %%d0, %%d3 \n\t" \
240  "addl %%d1, %%a3@+ \n\t" \
241  "movel %%a2@+, %%d1 \n\t" \
242  "mulul %%d2, %%d4:%%d1 \n\t" \
243  "addxl %%d3, %%d1 \n\t" \
244  "addxl %%d0, %%d4 \n\t" \
245  "addl %%d1, %%a3@+ \n\t" \
246  "movel %%a2@+, %%d1 \n\t" \
247  "mulul %%d2, %%d3:%%d1 \n\t" \
248  "addxl %%d4, %%d1 \n\t" \
249  "addxl %%d0, %%d3 \n\t" \
250  "addl %%d1, %%a3@+ \n\t" \
251  "movel %%a2@+, %%d1 \n\t" \
252  "mulul %%d2, %%d4:%%d1 \n\t" \
253  "addxl %%d3, %%d1 \n\t" \
254  "addxl %%d0, %%d4 \n\t" \
255  "addl %%d1, %%a3@+ \n\t" \
256  "movel %%a2@+, %%d1 \n\t" \
257  "mulul %%d2, %%d3:%%d1 \n\t" \
258  "addxl %%d4, %%d1 \n\t" \
259  "addxl %%d0, %%d3 \n\t" \
260  "addl %%d1, %%a3@+ \n\t" \
261  "addxl %%d0, %%d3 \n\t"
262 
263 #endif /* MC68000 */
264 
265 #if defined(__powerpc64__) || defined(__ppc64__)
266 
267 #if defined(__MACH__) && defined(__APPLE__)
268 
269 #define MULADDC_INIT \
270  asm( \
271  "ld r3, %3 \n\t" \
272  "ld r4, %4 \n\t" \
273  "ld r5, %5 \n\t" \
274  "ld r6, %6 \n\t" \
275  "addi r3, r3, -8 \n\t" \
276  "addi r4, r4, -8 \n\t" \
277  "addic r5, r5, 0 \n\t"
278 
279 #define MULADDC_CORE \
280  "ldu r7, 8(r3) \n\t" \
281  "mulld r8, r7, r6 \n\t" \
282  "mulhdu r9, r7, r6 \n\t" \
283  "adde r8, r8, r5 \n\t" \
284  "ld r7, 8(r4) \n\t" \
285  "addze r5, r9 \n\t" \
286  "addc r8, r8, r7 \n\t" \
287  "stdu r8, 8(r4) \n\t"
288 
289 #define MULADDC_STOP \
290  "addze r5, r5 \n\t" \
291  "addi r4, r4, 8 \n\t" \
292  "addi r3, r3, 8 \n\t" \
293  "std r5, %0 \n\t" \
294  "std r4, %1 \n\t" \
295  "std r3, %2 \n\t" \
296  : "=m" (c), "=m" (d), "=m" (s) \
297  : "m" (s), "m" (d), "m" (c), "m" (b) \
298  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
299  );
300 
301 
302 #else
303 
304 #define MULADDC_INIT \
305  asm( \
306  "ld %%r3, %3 \n\t" \
307  "ld %%r4, %4 \n\t" \
308  "ld %%r5, %5 \n\t" \
309  "ld %%r6, %6 \n\t" \
310  "addi %%r3, %%r3, -8 \n\t" \
311  "addi %%r4, %%r4, -8 \n\t" \
312  "addic %%r5, %%r5, 0 \n\t"
313 
314 #define MULADDC_CORE \
315  "ldu %%r7, 8(%%r3) \n\t" \
316  "mulld %%r8, %%r7, %%r6 \n\t" \
317  "mulhdu %%r9, %%r7, %%r6 \n\t" \
318  "adde %%r8, %%r8, %%r5 \n\t" \
319  "ld %%r7, 8(%%r4) \n\t" \
320  "addze %%r5, %%r9 \n\t" \
321  "addc %%r8, %%r8, %%r7 \n\t" \
322  "stdu %%r8, 8(%%r4) \n\t"
323 
324 #define MULADDC_STOP \
325  "addze %%r5, %%r5 \n\t" \
326  "addi %%r4, %%r4, 8 \n\t" \
327  "addi %%r3, %%r3, 8 \n\t" \
328  "std %%r5, %0 \n\t" \
329  "std %%r4, %1 \n\t" \
330  "std %%r3, %2 \n\t" \
331  : "=m" (c), "=m" (d), "=m" (s) \
332  : "m" (s), "m" (d), "m" (c), "m" (b) \
333  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
334  );
335 
336 #endif
337 
338 #elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
339 
340 #if defined(__MACH__) && defined(__APPLE__)
341 
342 #define MULADDC_INIT \
343  asm( \
344  "lwz r3, %3 \n\t" \
345  "lwz r4, %4 \n\t" \
346  "lwz r5, %5 \n\t" \
347  "lwz r6, %6 \n\t" \
348  "addi r3, r3, -4 \n\t" \
349  "addi r4, r4, -4 \n\t" \
350  "addic r5, r5, 0 \n\t"
351 
352 #define MULADDC_CORE \
353  "lwzu r7, 4(r3) \n\t" \
354  "mullw r8, r7, r6 \n\t" \
355  "mulhwu r9, r7, r6 \n\t" \
356  "adde r8, r8, r5 \n\t" \
357  "lwz r7, 4(r4) \n\t" \
358  "addze r5, r9 \n\t" \
359  "addc r8, r8, r7 \n\t" \
360  "stwu r8, 4(r4) \n\t"
361 
362 #define MULADDC_STOP \
363  "addze r5, r5 \n\t" \
364  "addi r4, r4, 4 \n\t" \
365  "addi r3, r3, 4 \n\t" \
366  "stw r5, %0 \n\t" \
367  "stw r4, %1 \n\t" \
368  "stw r3, %2 \n\t" \
369  : "=m" (c), "=m" (d), "=m" (s) \
370  : "m" (s), "m" (d), "m" (c), "m" (b) \
371  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
372  );
373 
374 #else
375 
376 #define MULADDC_INIT \
377  asm( \
378  "lwz %%r3, %3 \n\t" \
379  "lwz %%r4, %4 \n\t" \
380  "lwz %%r5, %5 \n\t" \
381  "lwz %%r6, %6 \n\t" \
382  "addi %%r3, %%r3, -4 \n\t" \
383  "addi %%r4, %%r4, -4 \n\t" \
384  "addic %%r5, %%r5, 0 \n\t"
385 
386 #define MULADDC_CORE \
387  "lwzu %%r7, 4(%%r3) \n\t" \
388  "mullw %%r8, %%r7, %%r6 \n\t" \
389  "mulhwu %%r9, %%r7, %%r6 \n\t" \
390  "adde %%r8, %%r8, %%r5 \n\t" \
391  "lwz %%r7, 4(%%r4) \n\t" \
392  "addze %%r5, %%r9 \n\t" \
393  "addc %%r8, %%r8, %%r7 \n\t" \
394  "stwu %%r8, 4(%%r4) \n\t"
395 
396 #define MULADDC_STOP \
397  "addze %%r5, %%r5 \n\t" \
398  "addi %%r4, %%r4, 4 \n\t" \
399  "addi %%r3, %%r3, 4 \n\t" \
400  "stw %%r5, %0 \n\t" \
401  "stw %%r4, %1 \n\t" \
402  "stw %%r3, %2 \n\t" \
403  : "=m" (c), "=m" (d), "=m" (s) \
404  : "m" (s), "m" (d), "m" (c), "m" (b) \
405  : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
406  );
407 
408 #endif
409 
410 #endif /* PPC32 */
411 
412 /*
413  * The Sparc64 assembly is reported to be broken.
414  * Disable it for now, until we're able to fix it.
415  */
416 #if 0 && defined(__sparc__) && defined(__sparc64__)
417 
418 #define MULADDC_INIT \
419  asm( \
420  "ldx %3, %%o0 \n\t" \
421  "ldx %4, %%o1 \n\t" \
422  "ld %5, %%o2 \n\t" \
423  "ld %6, %%o3 \n\t"
424 
425 #define MULADDC_CORE \
426  "ld [%%o0], %%o4 \n\t" \
427  "inc 4, %%o0 \n\t" \
428  "ld [%%o1], %%o5 \n\t" \
429  "umul %%o3, %%o4, %%o4 \n\t" \
430  "addcc %%o4, %%o2, %%o4 \n\t" \
431  "rd %%y, %%g1 \n\t" \
432  "addx %%g1, 0, %%g1 \n\t" \
433  "addcc %%o4, %%o5, %%o4 \n\t" \
434  "st %%o4, [%%o1] \n\t" \
435  "addx %%g1, 0, %%o2 \n\t" \
436  "inc 4, %%o1 \n\t"
437 
438  #define MULADDC_STOP \
439  "st %%o2, %0 \n\t" \
440  "stx %%o1, %1 \n\t" \
441  "stx %%o0, %2 \n\t" \
442  : "=m" (c), "=m" (d), "=m" (s) \
443  : "m" (s), "m" (d), "m" (c), "m" (b) \
444  : "g1", "o0", "o1", "o2", "o3", "o4", \
445  "o5" \
446  );
447 #endif /* SPARCv9 */
448 
449 #if defined(__sparc__) && !defined(__sparc64__)
450 
451 #define MULADDC_INIT \
452  asm( \
453  "ld %3, %%o0 \n\t" \
454  "ld %4, %%o1 \n\t" \
455  "ld %5, %%o2 \n\t" \
456  "ld %6, %%o3 \n\t"
457 
458 #define MULADDC_CORE \
459  "ld [%%o0], %%o4 \n\t" \
460  "inc 4, %%o0 \n\t" \
461  "ld [%%o1], %%o5 \n\t" \
462  "umul %%o3, %%o4, %%o4 \n\t" \
463  "addcc %%o4, %%o2, %%o4 \n\t" \
464  "rd %%y, %%g1 \n\t" \
465  "addx %%g1, 0, %%g1 \n\t" \
466  "addcc %%o4, %%o5, %%o4 \n\t" \
467  "st %%o4, [%%o1] \n\t" \
468  "addx %%g1, 0, %%o2 \n\t" \
469  "inc 4, %%o1 \n\t"
470 
471 #define MULADDC_STOP \
472  "st %%o2, %0 \n\t" \
473  "st %%o1, %1 \n\t" \
474  "st %%o0, %2 \n\t" \
475  : "=m" (c), "=m" (d), "=m" (s) \
476  : "m" (s), "m" (d), "m" (c), "m" (b) \
477  : "g1", "o0", "o1", "o2", "o3", "o4", \
478  "o5" \
479  );
480 
481 #endif /* SPARCv8 */
482 
483 #if defined(__microblaze__) || defined(microblaze)
484 
485 #define MULADDC_INIT \
486  asm( \
487  "lwi r3, %3 \n\t" \
488  "lwi r4, %4 \n\t" \
489  "lwi r5, %5 \n\t" \
490  "lwi r6, %6 \n\t" \
491  "andi r7, r6, 0xffff \n\t" \
492  "bsrli r6, r6, 16 \n\t"
493 
494 #define MULADDC_CORE \
495  "lhui r8, r3, 0 \n\t" \
496  "addi r3, r3, 2 \n\t" \
497  "lhui r9, r3, 0 \n\t" \
498  "addi r3, r3, 2 \n\t" \
499  "mul r10, r9, r6 \n\t" \
500  "mul r11, r8, r7 \n\t" \
501  "mul r12, r9, r7 \n\t" \
502  "mul r13, r8, r6 \n\t" \
503  "bsrli r8, r10, 16 \n\t" \
504  "bsrli r9, r11, 16 \n\t" \
505  "add r13, r13, r8 \n\t" \
506  "add r13, r13, r9 \n\t" \
507  "bslli r10, r10, 16 \n\t" \
508  "bslli r11, r11, 16 \n\t" \
509  "add r12, r12, r10 \n\t" \
510  "addc r13, r13, r0 \n\t" \
511  "add r12, r12, r11 \n\t" \
512  "addc r13, r13, r0 \n\t" \
513  "lwi r10, r4, 0 \n\t" \
514  "add r12, r12, r10 \n\t" \
515  "addc r13, r13, r0 \n\t" \
516  "add r12, r12, r5 \n\t" \
517  "addc r5, r13, r0 \n\t" \
518  "swi r12, r4, 0 \n\t" \
519  "addi r4, r4, 4 \n\t"
520 
521 #define MULADDC_STOP \
522  "swi r5, %0 \n\t" \
523  "swi r4, %1 \n\t" \
524  "swi r3, %2 \n\t" \
525  : "=m" (c), "=m" (d), "=m" (s) \
526  : "m" (s), "m" (d), "m" (c), "m" (b) \
527  : "r3", "r4" "r5", "r6", "r7", "r8", \
528  "r9", "r10", "r11", "r12", "r13" \
529  );
530 
531 #endif /* MicroBlaze */
532 
533 #if defined(__tricore__)
534 
535 #define MULADDC_INIT \
536  asm( \
537  "ld.a %%a2, %3 \n\t" \
538  "ld.a %%a3, %4 \n\t" \
539  "ld.w %%d4, %5 \n\t" \
540  "ld.w %%d1, %6 \n\t" \
541  "xor %%d5, %%d5 \n\t"
542 
543 #define MULADDC_CORE \
544  "ld.w %%d0, [%%a2+] \n\t" \
545  "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
546  "ld.w %%d0, [%%a3] \n\t" \
547  "addx %%d2, %%d2, %%d0 \n\t" \
548  "addc %%d3, %%d3, 0 \n\t" \
549  "mov %%d4, %%d3 \n\t" \
550  "st.w [%%a3+], %%d2 \n\t"
551 
552 #define MULADDC_STOP \
553  "st.w %0, %%d4 \n\t" \
554  "st.a %1, %%a3 \n\t" \
555  "st.a %2, %%a2 \n\t" \
556  : "=m" (c), "=m" (d), "=m" (s) \
557  : "m" (s), "m" (d), "m" (c), "m" (b) \
558  : "d0", "d1", "e2", "d4", "a2", "a3" \
559  );
560 
561 #endif /* TriCore */
562 
563 #if defined(__arm__)
564 
565 #if defined(__thumb__) && !defined(__thumb2__)
566 
567 #define MULADDC_INIT \
568  asm( \
569  "ldr r0, %3 \n\t" \
570  "ldr r1, %4 \n\t" \
571  "ldr r2, %5 \n\t" \
572  "ldr r3, %6 \n\t" \
573  "lsr r7, r3, #16 \n\t" \
574  "mov r9, r7 \n\t" \
575  "lsl r7, r3, #16 \n\t" \
576  "lsr r7, r7, #16 \n\t" \
577  "mov r8, r7 \n\t"
578 
579 #define MULADDC_CORE \
580  "ldmia r0!, {r6} \n\t" \
581  "lsr r7, r6, #16 \n\t" \
582  "lsl r6, r6, #16 \n\t" \
583  "lsr r6, r6, #16 \n\t" \
584  "mov r4, r8 \n\t" \
585  "mul r4, r6 \n\t" \
586  "mov r3, r9 \n\t" \
587  "mul r6, r3 \n\t" \
588  "mov r5, r9 \n\t" \
589  "mul r5, r7 \n\t" \
590  "mov r3, r8 \n\t" \
591  "mul r7, r3 \n\t" \
592  "lsr r3, r6, #16 \n\t" \
593  "add r5, r5, r3 \n\t" \
594  "lsr r3, r7, #16 \n\t" \
595  "add r5, r5, r3 \n\t" \
596  "add r4, r4, r2 \n\t" \
597  "mov r2, #0 \n\t" \
598  "adc r5, r2 \n\t" \
599  "lsl r3, r6, #16 \n\t" \
600  "add r4, r4, r3 \n\t" \
601  "adc r5, r2 \n\t" \
602  "lsl r3, r7, #16 \n\t" \
603  "add r4, r4, r3 \n\t" \
604  "adc r5, r2 \n\t" \
605  "ldr r3, [r1] \n\t" \
606  "add r4, r4, r3 \n\t" \
607  "adc r2, r5 \n\t" \
608  "stmia r1!, {r4} \n\t"
609 
610 #define MULADDC_STOP \
611  "str r2, %0 \n\t" \
612  "str r1, %1 \n\t" \
613  "str r0, %2 \n\t" \
614  : "=m" (c), "=m" (d), "=m" (s) \
615  : "m" (s), "m" (d), "m" (c), "m" (b) \
616  : "r0", "r1", "r2", "r3", "r4", "r5", \
617  "r6", "r7", "r8", "r9", "cc" \
618  );
619 
620 #else
621 
622 #define MULADDC_INIT \
623  asm( \
624  "ldr r0, %3 \n\t" \
625  "ldr r1, %4 \n\t" \
626  "ldr r2, %5 \n\t" \
627  "ldr r3, %6 \n\t"
628 
629 #define MULADDC_CORE \
630  "ldr r4, [r0], #4 \n\t" \
631  "mov r5, #0 \n\t" \
632  "ldr r6, [r1] \n\t" \
633  "umlal r2, r5, r3, r4 \n\t" \
634  "adds r7, r6, r2 \n\t" \
635  "adc r2, r5, #0 \n\t" \
636  "str r7, [r1], #4 \n\t"
637 
638 #define MULADDC_STOP \
639  "str r2, %0 \n\t" \
640  "str r1, %1 \n\t" \
641  "str r0, %2 \n\t" \
642  : "=m" (c), "=m" (d), "=m" (s) \
643  : "m" (s), "m" (d), "m" (c), "m" (b) \
644  : "r0", "r1", "r2", "r3", "r4", "r5", \
645  "r6", "r7", "cc" \
646  );
647 
648 #endif /* Thumb */
649 
650 #endif /* ARMv3 */
651 
652 #if defined(__alpha__)
653 
654 #define MULADDC_INIT \
655  asm( \
656  "ldq $1, %3 \n\t" \
657  "ldq $2, %4 \n\t" \
658  "ldq $3, %5 \n\t" \
659  "ldq $4, %6 \n\t"
660 
661 #define MULADDC_CORE \
662  "ldq $6, 0($1) \n\t" \
663  "addq $1, 8, $1 \n\t" \
664  "mulq $6, $4, $7 \n\t" \
665  "umulh $6, $4, $6 \n\t" \
666  "addq $7, $3, $7 \n\t" \
667  "cmpult $7, $3, $3 \n\t" \
668  "ldq $5, 0($2) \n\t" \
669  "addq $7, $5, $7 \n\t" \
670  "cmpult $7, $5, $5 \n\t" \
671  "stq $7, 0($2) \n\t" \
672  "addq $2, 8, $2 \n\t" \
673  "addq $6, $3, $3 \n\t" \
674  "addq $5, $3, $3 \n\t"
675 
676 #define MULADDC_STOP \
677  "stq $3, %0 \n\t" \
678  "stq $2, %1 \n\t" \
679  "stq $1, %2 \n\t" \
680  : "=m" (c), "=m" (d), "=m" (s) \
681  : "m" (s), "m" (d), "m" (c), "m" (b) \
682  : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
683  );
684 #endif /* Alpha */
685 
686 #if defined(__mips__) && !defined(__mips64__)
687 
688 #define MULADDC_INIT \
689  asm( \
690  "lw $10, %3 \n\t" \
691  "lw $11, %4 \n\t" \
692  "lw $12, %5 \n\t" \
693  "lw $13, %6 \n\t"
694 
695 #define MULADDC_CORE \
696  "lw $14, 0($10) \n\t" \
697  "multu $13, $14 \n\t" \
698  "addi $10, $10, 4 \n\t" \
699  "mflo $14 \n\t" \
700  "mfhi $9 \n\t" \
701  "addu $14, $12, $14 \n\t" \
702  "lw $15, 0($11) \n\t" \
703  "sltu $12, $14, $12 \n\t" \
704  "addu $15, $14, $15 \n\t" \
705  "sltu $14, $15, $14 \n\t" \
706  "addu $12, $12, $9 \n\t" \
707  "sw $15, 0($11) \n\t" \
708  "addu $12, $12, $14 \n\t" \
709  "addi $11, $11, 4 \n\t"
710 
711 #define MULADDC_STOP \
712  "sw $12, %0 \n\t" \
713  "sw $11, %1 \n\t" \
714  "sw $10, %2 \n\t" \
715  : "=m" (c), "=m" (d), "=m" (s) \
716  : "m" (s), "m" (d), "m" (c), "m" (b) \
717  : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
718  );
719 
720 #endif /* MIPS */
721 #endif /* GNUC */
722 
723 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
724 
725 #define MULADDC_INIT \
726  __asm mov esi, s \
727  __asm mov edi, d \
728  __asm mov ecx, c \
729  __asm mov ebx, b
730 
731 #define MULADDC_CORE \
732  __asm lodsd \
733  __asm mul ebx \
734  __asm add eax, ecx \
735  __asm adc edx, 0 \
736  __asm add eax, [edi] \
737  __asm adc edx, 0 \
738  __asm mov ecx, edx \
739  __asm stosd
740 
741 #if defined(POLARSSL_HAVE_SSE2)
742 
743 #define EMIT __asm _emit
744 
745 #define MULADDC_HUIT \
746  EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
747  EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
748  EMIT 0x0F EMIT 0x6E EMIT 0x1F \
749  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
750  EMIT 0x0F EMIT 0x6E EMIT 0x16 \
751  EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
752  EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
753  EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
754  EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
755  EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
756  EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
757  EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
758  EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
759  EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
760  EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
761  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
762  EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
763  EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
764  EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
765  EMIT 0x0F EMIT 0x7E EMIT 0x0F \
766  EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
767  EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
768  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
769  EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
770  EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
771  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
772  EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
773  EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
774  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
775  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
776  EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
777  EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
778  EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
779  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
780  EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
781  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
782  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
783  EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
784  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
785  EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
786  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
787  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
788  EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
789  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
790  EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
791  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
792  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
793  EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
794  EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
795  EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
796  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
797  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
798  EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
799  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
800  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
801  EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
802  EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
803  EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
804  EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
805  EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
806  EMIT 0x0F EMIT 0x7E EMIT 0xC9
807 
808 #define MULADDC_STOP \
809  EMIT 0x0F EMIT 0x77 \
810  __asm mov c, ecx \
811  __asm mov d, edi \
812  __asm mov s, esi \
813 
814 #else
815 
816 #define MULADDC_STOP \
817  __asm mov c, ecx \
818  __asm mov d, edi \
819  __asm mov s, esi \
820 
821 #endif /* SSE2 */
822 #endif /* MSVC */
823 
824 #endif /* POLARSSL_HAVE_ASM */
825 
826 #if !defined(MULADDC_CORE)
827 #if defined(POLARSSL_HAVE_UDBL)
828 
829 #define MULADDC_INIT \
830 { \
831  t_udbl r; \
832  t_uint r0, r1;
833 
834 #define MULADDC_CORE \
835  r = *(s++) * (t_udbl) b; \
836  r0 = r; \
837  r1 = r >> biL; \
838  r0 += c; r1 += (r0 < c); \
839  r0 += *d; r1 += (r0 < *d); \
840  c = r1; *(d++) = r0;
841 
842 #define MULADDC_STOP \
843 }
844 
845 #else
846 #define MULADDC_INIT \
847 { \
848  t_uint s0, s1, b0, b1; \
849  t_uint r0, r1, rx, ry; \
850  b0 = ( b << biH ) >> biH; \
851  b1 = ( b >> biH );
852 
853 #define MULADDC_CORE \
854  s0 = ( *s << biH ) >> biH; \
855  s1 = ( *s >> biH ); s++; \
856  rx = s0 * b1; r0 = s0 * b0; \
857  ry = s1 * b0; r1 = s1 * b1; \
858  r1 += ( rx >> biH ); \
859  r1 += ( ry >> biH ); \
860  rx <<= biH; ry <<= biH; \
861  r0 += rx; r1 += (r0 < rx); \
862  r0 += ry; r1 += (r0 < ry); \
863  r0 += c; r1 += (r0 < c); \
864  r0 += *d; r1 += (r0 < *d); \
865  c = r1; *(d++) = r0;
866 
867 #define MULADDC_STOP \
868 }
869 
870 #endif /* C (generic) */
871 #endif /* C (longlong) */
872 
873 #endif /* bn_mul.h */
Multi-precision integer library.