Go to the documentation of this file.
38 #ifndef MBEDTLS_BN_MUL_H
39 #define MBEDTLS_BN_MUL_H
43 #if defined(MBEDTLS_HAVE_ASM)
50 #if defined(__GNUC__) && \
51 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
65 #if defined(__GNUC__) && __GNUC__ < 5 && defined(__PIC__)
66 #define MULADDC_CANNOT_USE_EBX
75 #if defined(__i386__) && defined(__OPTIMIZE__) && !defined(MULADDC_CANNOT_USE_EBX)
77 #define MULADDC_INIT \
79 "movl %%ebx, %0 \n\t" \
80 "movl %5, %%esi \n\t" \
81 "movl %6, %%edi \n\t" \
82 "movl %7, %%ecx \n\t" \
85 #define MULADDC_CORE \
88 "addl %%ecx, %%eax \n\t" \
89 "adcl $0, %%edx \n\t" \
90 "addl (%%edi), %%eax \n\t" \
91 "adcl $0, %%edx \n\t" \
92 "movl %%edx, %%ecx \n\t" \
95 #if defined(MBEDTLS_HAVE_SSE2)
97 #define MULADDC_HUIT \
98 "movd %%ecx, %%mm1 \n\t" \
99 "movd %%ebx, %%mm0 \n\t" \
100 "movd (%%edi), %%mm3 \n\t" \
101 "paddq %%mm3, %%mm1 \n\t" \
102 "movd (%%esi), %%mm2 \n\t" \
103 "pmuludq %%mm0, %%mm2 \n\t" \
104 "movd 4(%%esi), %%mm4 \n\t" \
105 "pmuludq %%mm0, %%mm4 \n\t" \
106 "movd 8(%%esi), %%mm6 \n\t" \
107 "pmuludq %%mm0, %%mm6 \n\t" \
108 "movd 12(%%esi), %%mm7 \n\t" \
109 "pmuludq %%mm0, %%mm7 \n\t" \
110 "paddq %%mm2, %%mm1 \n\t" \
111 "movd 4(%%edi), %%mm3 \n\t" \
112 "paddq %%mm4, %%mm3 \n\t" \
113 "movd 8(%%edi), %%mm5 \n\t" \
114 "paddq %%mm6, %%mm5 \n\t" \
115 "movd 12(%%edi), %%mm4 \n\t" \
116 "paddq %%mm4, %%mm7 \n\t" \
117 "movd %%mm1, (%%edi) \n\t" \
118 "movd 16(%%esi), %%mm2 \n\t" \
119 "pmuludq %%mm0, %%mm2 \n\t" \
120 "psrlq $32, %%mm1 \n\t" \
121 "movd 20(%%esi), %%mm4 \n\t" \
122 "pmuludq %%mm0, %%mm4 \n\t" \
123 "paddq %%mm3, %%mm1 \n\t" \
124 "movd 24(%%esi), %%mm6 \n\t" \
125 "pmuludq %%mm0, %%mm6 \n\t" \
126 "movd %%mm1, 4(%%edi) \n\t" \
127 "psrlq $32, %%mm1 \n\t" \
128 "movd 28(%%esi), %%mm3 \n\t" \
129 "pmuludq %%mm0, %%mm3 \n\t" \
130 "paddq %%mm5, %%mm1 \n\t" \
131 "movd 16(%%edi), %%mm5 \n\t" \
132 "paddq %%mm5, %%mm2 \n\t" \
133 "movd %%mm1, 8(%%edi) \n\t" \
134 "psrlq $32, %%mm1 \n\t" \
135 "paddq %%mm7, %%mm1 \n\t" \
136 "movd 20(%%edi), %%mm5 \n\t" \
137 "paddq %%mm5, %%mm4 \n\t" \
138 "movd %%mm1, 12(%%edi) \n\t" \
139 "psrlq $32, %%mm1 \n\t" \
140 "paddq %%mm2, %%mm1 \n\t" \
141 "movd 24(%%edi), %%mm5 \n\t" \
142 "paddq %%mm5, %%mm6 \n\t" \
143 "movd %%mm1, 16(%%edi) \n\t" \
144 "psrlq $32, %%mm1 \n\t" \
145 "paddq %%mm4, %%mm1 \n\t" \
146 "movd 28(%%edi), %%mm5 \n\t" \
147 "paddq %%mm5, %%mm3 \n\t" \
148 "movd %%mm1, 20(%%edi) \n\t" \
149 "psrlq $32, %%mm1 \n\t" \
150 "paddq %%mm6, %%mm1 \n\t" \
151 "movd %%mm1, 24(%%edi) \n\t" \
152 "psrlq $32, %%mm1 \n\t" \
153 "paddq %%mm3, %%mm1 \n\t" \
154 "movd %%mm1, 28(%%edi) \n\t" \
155 "addl $32, %%edi \n\t" \
156 "addl $32, %%esi \n\t" \
157 "psrlq $32, %%mm1 \n\t" \
158 "movd %%mm1, %%ecx \n\t"
160 #define MULADDC_STOP \
162 "movl %4, %%ebx \n\t" \
163 "movl %%ecx, %1 \n\t" \
164 "movl %%edi, %2 \n\t" \
165 "movl %%esi, %3 \n\t" \
166 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
167 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
168 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
173 #define MULADDC_STOP \
174 "movl %4, %%ebx \n\t" \
175 "movl %%ecx, %1 \n\t" \
176 "movl %%edi, %2 \n\t" \
177 "movl %%esi, %3 \n\t" \
178 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
179 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
180 : "eax", "ebx", "ecx", "edx", "esi", "edi" \
185 #if defined(__amd64__) || defined (__x86_64__)
187 #define MULADDC_INIT \
189 "xorq %%r8, %%r8 \n\t"
191 #define MULADDC_CORE \
192 "movq (%%rsi), %%rax \n\t" \
194 "addq $8, %%rsi \n\t" \
195 "addq %%rcx, %%rax \n\t" \
196 "movq %%r8, %%rcx \n\t" \
197 "adcq $0, %%rdx \n\t" \
199 "addq %%rax, (%%rdi) \n\t" \
200 "adcq %%rdx, %%rcx \n\t" \
201 "addq $8, %%rdi \n\t"
203 #define MULADDC_STOP \
204 : "+c" (c), "+D" (d), "+S" (s) \
206 : "rax", "rdx", "r8" \
211 #if defined(__mc68020__) || defined(__mcpu32__)
213 #define MULADDC_INIT \
215 "movl %3, %%a2 \n\t" \
216 "movl %4, %%a3 \n\t" \
217 "movl %5, %%d3 \n\t" \
218 "movl %6, %%d2 \n\t" \
219 "moveq #0, %%d0 \n\t"
221 #define MULADDC_CORE \
222 "movel %%a2@+, %%d1 \n\t" \
223 "mulul %%d2, %%d4:%%d1 \n\t" \
224 "addl %%d3, %%d1 \n\t" \
225 "addxl %%d0, %%d4 \n\t" \
226 "moveq #0, %%d3 \n\t" \
227 "addl %%d1, %%a3@+ \n\t" \
228 "addxl %%d4, %%d3 \n\t"
230 #define MULADDC_STOP \
231 "movl %%d3, %0 \n\t" \
232 "movl %%a3, %1 \n\t" \
233 "movl %%a2, %2 \n\t" \
234 : "=m" (c), "=m" (d), "=m" (s) \
235 : "m" (s), "m" (d), "m" (c), "m" (b) \
236 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
239 #define MULADDC_HUIT \
240 "movel %%a2@+, %%d1 \n\t" \
241 "mulul %%d2, %%d4:%%d1 \n\t" \
242 "addxl %%d3, %%d1 \n\t" \
243 "addxl %%d0, %%d4 \n\t" \
244 "addl %%d1, %%a3@+ \n\t" \
245 "movel %%a2@+, %%d1 \n\t" \
246 "mulul %%d2, %%d3:%%d1 \n\t" \
247 "addxl %%d4, %%d1 \n\t" \
248 "addxl %%d0, %%d3 \n\t" \
249 "addl %%d1, %%a3@+ \n\t" \
250 "movel %%a2@+, %%d1 \n\t" \
251 "mulul %%d2, %%d4:%%d1 \n\t" \
252 "addxl %%d3, %%d1 \n\t" \
253 "addxl %%d0, %%d4 \n\t" \
254 "addl %%d1, %%a3@+ \n\t" \
255 "movel %%a2@+, %%d1 \n\t" \
256 "mulul %%d2, %%d3:%%d1 \n\t" \
257 "addxl %%d4, %%d1 \n\t" \
258 "addxl %%d0, %%d3 \n\t" \
259 "addl %%d1, %%a3@+ \n\t" \
260 "movel %%a2@+, %%d1 \n\t" \
261 "mulul %%d2, %%d4:%%d1 \n\t" \
262 "addxl %%d3, %%d1 \n\t" \
263 "addxl %%d0, %%d4 \n\t" \
264 "addl %%d1, %%a3@+ \n\t" \
265 "movel %%a2@+, %%d1 \n\t" \
266 "mulul %%d2, %%d3:%%d1 \n\t" \
267 "addxl %%d4, %%d1 \n\t" \
268 "addxl %%d0, %%d3 \n\t" \
269 "addl %%d1, %%a3@+ \n\t" \
270 "movel %%a2@+, %%d1 \n\t" \
271 "mulul %%d2, %%d4:%%d1 \n\t" \
272 "addxl %%d3, %%d1 \n\t" \
273 "addxl %%d0, %%d4 \n\t" \
274 "addl %%d1, %%a3@+ \n\t" \
275 "movel %%a2@+, %%d1 \n\t" \
276 "mulul %%d2, %%d3:%%d1 \n\t" \
277 "addxl %%d4, %%d1 \n\t" \
278 "addxl %%d0, %%d3 \n\t" \
279 "addl %%d1, %%a3@+ \n\t" \
280 "addxl %%d0, %%d3 \n\t"
284 #if defined(__powerpc64__) || defined(__ppc64__)
286 #if defined(__MACH__) && defined(__APPLE__)
288 #define MULADDC_INIT \
294 "addi r3, r3, -8 \n\t" \
295 "addi r4, r4, -8 \n\t" \
296 "addic r5, r5, 0 \n\t"
298 #define MULADDC_CORE \
299 "ldu r7, 8(r3) \n\t" \
300 "mulld r8, r7, r6 \n\t" \
301 "mulhdu r9, r7, r6 \n\t" \
302 "adde r8, r8, r5 \n\t" \
303 "ld r7, 8(r4) \n\t" \
304 "addze r5, r9 \n\t" \
305 "addc r8, r8, r7 \n\t" \
306 "stdu r8, 8(r4) \n\t"
308 #define MULADDC_STOP \
309 "addze r5, r5 \n\t" \
310 "addi r4, r4, 8 \n\t" \
311 "addi r3, r3, 8 \n\t" \
315 : "=m" (c), "=m" (d), "=m" (s) \
316 : "m" (s), "m" (d), "m" (c), "m" (b) \
317 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
323 #define MULADDC_INIT \
329 "addi %%r3, %%r3, -8 \n\t" \
330 "addi %%r4, %%r4, -8 \n\t" \
331 "addic %%r5, %%r5, 0 \n\t"
333 #define MULADDC_CORE \
334 "ldu %%r7, 8(%%r3) \n\t" \
335 "mulld %%r8, %%r7, %%r6 \n\t" \
336 "mulhdu %%r9, %%r7, %%r6 \n\t" \
337 "adde %%r8, %%r8, %%r5 \n\t" \
338 "ld %%r7, 8(%%r4) \n\t" \
339 "addze %%r5, %%r9 \n\t" \
340 "addc %%r8, %%r8, %%r7 \n\t" \
341 "stdu %%r8, 8(%%r4) \n\t"
343 #define MULADDC_STOP \
344 "addze %%r5, %%r5 \n\t" \
345 "addi %%r4, %%r4, 8 \n\t" \
346 "addi %%r3, %%r3, 8 \n\t" \
347 "std %%r5, %0 \n\t" \
348 "std %%r4, %1 \n\t" \
349 "std %%r3, %2 \n\t" \
350 : "=m" (c), "=m" (d), "=m" (s) \
351 : "m" (s), "m" (d), "m" (c), "m" (b) \
352 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
357 #elif defined(__powerpc__) || defined(__ppc__)
359 #if defined(__MACH__) && defined(__APPLE__)
361 #define MULADDC_INIT \
367 "addi r3, r3, -4 \n\t" \
368 "addi r4, r4, -4 \n\t" \
369 "addic r5, r5, 0 \n\t"
371 #define MULADDC_CORE \
372 "lwzu r7, 4(r3) \n\t" \
373 "mullw r8, r7, r6 \n\t" \
374 "mulhwu r9, r7, r6 \n\t" \
375 "adde r8, r8, r5 \n\t" \
376 "lwz r7, 4(r4) \n\t" \
377 "addze r5, r9 \n\t" \
378 "addc r8, r8, r7 \n\t" \
379 "stwu r8, 4(r4) \n\t"
381 #define MULADDC_STOP \
382 "addze r5, r5 \n\t" \
383 "addi r4, r4, 4 \n\t" \
384 "addi r3, r3, 4 \n\t" \
388 : "=m" (c), "=m" (d), "=m" (s) \
389 : "m" (s), "m" (d), "m" (c), "m" (b) \
390 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
395 #define MULADDC_INIT \
397 "lwz %%r3, %3 \n\t" \
398 "lwz %%r4, %4 \n\t" \
399 "lwz %%r5, %5 \n\t" \
400 "lwz %%r6, %6 \n\t" \
401 "addi %%r3, %%r3, -4 \n\t" \
402 "addi %%r4, %%r4, -4 \n\t" \
403 "addic %%r5, %%r5, 0 \n\t"
405 #define MULADDC_CORE \
406 "lwzu %%r7, 4(%%r3) \n\t" \
407 "mullw %%r8, %%r7, %%r6 \n\t" \
408 "mulhwu %%r9, %%r7, %%r6 \n\t" \
409 "adde %%r8, %%r8, %%r5 \n\t" \
410 "lwz %%r7, 4(%%r4) \n\t" \
411 "addze %%r5, %%r9 \n\t" \
412 "addc %%r8, %%r8, %%r7 \n\t" \
413 "stwu %%r8, 4(%%r4) \n\t"
415 #define MULADDC_STOP \
416 "addze %%r5, %%r5 \n\t" \
417 "addi %%r4, %%r4, 4 \n\t" \
418 "addi %%r3, %%r3, 4 \n\t" \
419 "stw %%r5, %0 \n\t" \
420 "stw %%r4, %1 \n\t" \
421 "stw %%r3, %2 \n\t" \
422 : "=m" (c), "=m" (d), "=m" (s) \
423 : "m" (s), "m" (d), "m" (c), "m" (b) \
424 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
435 #if 0 && defined(__sparc__)
436 #if defined(__sparc64__)
438 #define MULADDC_INIT \
440 "ldx %3, %%o0 \n\t" \
441 "ldx %4, %%o1 \n\t" \
445 #define MULADDC_CORE \
446 "ld [%%o0], %%o4 \n\t" \
448 "ld [%%o1], %%o5 \n\t" \
449 "umul %%o3, %%o4, %%o4 \n\t" \
450 "addcc %%o4, %%o2, %%o4 \n\t" \
451 "rd %%y, %%g1 \n\t" \
452 "addx %%g1, 0, %%g1 \n\t" \
453 "addcc %%o4, %%o5, %%o4 \n\t" \
454 "st %%o4, [%%o1] \n\t" \
455 "addx %%g1, 0, %%o2 \n\t" \
458 #define MULADDC_STOP \
460 "stx %%o1, %1 \n\t" \
461 "stx %%o0, %2 \n\t" \
462 : "=m" (c), "=m" (d), "=m" (s) \
463 : "m" (s), "m" (d), "m" (c), "m" (b) \
464 : "g1", "o0", "o1", "o2", "o3", "o4", \
470 #define MULADDC_INIT \
477 #define MULADDC_CORE \
478 "ld [%%o0], %%o4 \n\t" \
480 "ld [%%o1], %%o5 \n\t" \
481 "umul %%o3, %%o4, %%o4 \n\t" \
482 "addcc %%o4, %%o2, %%o4 \n\t" \
483 "rd %%y, %%g1 \n\t" \
484 "addx %%g1, 0, %%g1 \n\t" \
485 "addcc %%o4, %%o5, %%o4 \n\t" \
486 "st %%o4, [%%o1] \n\t" \
487 "addx %%g1, 0, %%o2 \n\t" \
490 #define MULADDC_STOP \
494 : "=m" (c), "=m" (d), "=m" (s) \
495 : "m" (s), "m" (d), "m" (c), "m" (b) \
496 : "g1", "o0", "o1", "o2", "o3", "o4", \
503 #if defined(__microblaze__) || defined(microblaze)
505 #define MULADDC_INIT \
511 "andi r7, r6, 0xffff \n\t" \
512 "bsrli r6, r6, 16 \n\t"
514 #define MULADDC_CORE \
515 "lhui r8, r3, 0 \n\t" \
516 "addi r3, r3, 2 \n\t" \
517 "lhui r9, r3, 0 \n\t" \
518 "addi r3, r3, 2 \n\t" \
519 "mul r10, r9, r6 \n\t" \
520 "mul r11, r8, r7 \n\t" \
521 "mul r12, r9, r7 \n\t" \
522 "mul r13, r8, r6 \n\t" \
523 "bsrli r8, r10, 16 \n\t" \
524 "bsrli r9, r11, 16 \n\t" \
525 "add r13, r13, r8 \n\t" \
526 "add r13, r13, r9 \n\t" \
527 "bslli r10, r10, 16 \n\t" \
528 "bslli r11, r11, 16 \n\t" \
529 "add r12, r12, r10 \n\t" \
530 "addc r13, r13, r0 \n\t" \
531 "add r12, r12, r11 \n\t" \
532 "addc r13, r13, r0 \n\t" \
533 "lwi r10, r4, 0 \n\t" \
534 "add r12, r12, r10 \n\t" \
535 "addc r13, r13, r0 \n\t" \
536 "add r12, r12, r5 \n\t" \
537 "addc r5, r13, r0 \n\t" \
538 "swi r12, r4, 0 \n\t" \
539 "addi r4, r4, 4 \n\t"
541 #define MULADDC_STOP \
545 : "=m" (c), "=m" (d), "=m" (s) \
546 : "m" (s), "m" (d), "m" (c), "m" (b) \
547 : "r3", "r4", "r5", "r6", "r7", "r8", \
548 "r9", "r10", "r11", "r12", "r13" \
553 #if defined(__tricore__)
555 #define MULADDC_INIT \
557 "ld.a %%a2, %3 \n\t" \
558 "ld.a %%a3, %4 \n\t" \
559 "ld.w %%d4, %5 \n\t" \
560 "ld.w %%d1, %6 \n\t" \
561 "xor %%d5, %%d5 \n\t"
563 #define MULADDC_CORE \
564 "ld.w %%d0, [%%a2+] \n\t" \
565 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
566 "ld.w %%d0, [%%a3] \n\t" \
567 "addx %%d2, %%d2, %%d0 \n\t" \
568 "addc %%d3, %%d3, 0 \n\t" \
569 "mov %%d4, %%d3 \n\t" \
570 "st.w [%%a3+], %%d2 \n\t"
572 #define MULADDC_STOP \
573 "st.w %0, %%d4 \n\t" \
574 "st.a %1, %%a3 \n\t" \
575 "st.a %2, %%a2 \n\t" \
576 : "=m" (c), "=m" (d), "=m" (s) \
577 : "m" (s), "m" (d), "m" (c), "m" (b) \
578 : "d0", "d1", "e2", "d4", "a2", "a3" \
594 #if defined(__GNUC__) && !defined(__OPTIMIZE__)
595 #define MULADDC_CANNOT_USE_R7
598 #if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
600 #if defined(__thumb__) && !defined(__thumb2__)
602 #define MULADDC_INIT \
608 "lsr r7, r3, #16 \n\t" \
610 "lsl r7, r3, #16 \n\t" \
611 "lsr r7, r7, #16 \n\t" \
614 #define MULADDC_CORE \
615 "ldmia r0!, {r6} \n\t" \
616 "lsr r7, r6, #16 \n\t" \
617 "lsl r6, r6, #16 \n\t" \
618 "lsr r6, r6, #16 \n\t" \
627 "lsr r3, r6, #16 \n\t" \
628 "add r5, r5, r3 \n\t" \
629 "lsr r3, r7, #16 \n\t" \
630 "add r5, r5, r3 \n\t" \
631 "add r4, r4, r2 \n\t" \
634 "lsl r3, r6, #16 \n\t" \
635 "add r4, r4, r3 \n\t" \
637 "lsl r3, r7, #16 \n\t" \
638 "add r4, r4, r3 \n\t" \
640 "ldr r3, [r1] \n\t" \
641 "add r4, r4, r3 \n\t" \
643 "stmia r1!, {r4} \n\t"
645 #define MULADDC_STOP \
649 : "=m" (c), "=m" (d), "=m" (s) \
650 : "m" (s), "m" (d), "m" (c), "m" (b) \
651 : "r0", "r1", "r2", "r3", "r4", "r5", \
652 "r6", "r7", "r8", "r9", "cc" \
655 #elif (__ARM_ARCH >= 6) && \
656 defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)
658 #define MULADDC_INIT \
661 #define MULADDC_CORE \
662 "ldr r0, [%0], #4 \n\t" \
663 "ldr r1, [%1] \n\t" \
664 "umaal r1, %2, %3, r0 \n\t" \
665 "str r1, [%1], #4 \n\t"
667 #define MULADDC_STOP \
668 : "=r" (s), "=r" (d), "=r" (c) \
669 : "r" (b), "0" (s), "1" (d), "2" (c) \
670 : "r0", "r1", "memory" \
675 #define MULADDC_INIT \
682 #define MULADDC_CORE \
683 "ldr r4, [r0], #4 \n\t" \
685 "ldr r6, [r1] \n\t" \
686 "umlal r2, r5, r3, r4 \n\t" \
687 "adds r7, r6, r2 \n\t" \
688 "adc r2, r5, #0 \n\t" \
689 "str r7, [r1], #4 \n\t"
691 #define MULADDC_STOP \
695 : "=m" (c), "=m" (d), "=m" (s) \
696 : "m" (s), "m" (d), "m" (c), "m" (b) \
697 : "r0", "r1", "r2", "r3", "r4", "r5", \
705 #if defined(__alpha__)
707 #define MULADDC_INIT \
714 #define MULADDC_CORE \
715 "ldq $6, 0($1) \n\t" \
716 "addq $1, 8, $1 \n\t" \
717 "mulq $6, $4, $7 \n\t" \
718 "umulh $6, $4, $6 \n\t" \
719 "addq $7, $3, $7 \n\t" \
720 "cmpult $7, $3, $3 \n\t" \
721 "ldq $5, 0($2) \n\t" \
722 "addq $7, $5, $7 \n\t" \
723 "cmpult $7, $5, $5 \n\t" \
724 "stq $7, 0($2) \n\t" \
725 "addq $2, 8, $2 \n\t" \
726 "addq $6, $3, $3 \n\t" \
727 "addq $5, $3, $3 \n\t"
729 #define MULADDC_STOP \
733 : "=m" (c), "=m" (d), "=m" (s) \
734 : "m" (s), "m" (d), "m" (c), "m" (b) \
735 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
739 #if defined(__mips__) && !defined(__mips64)
741 #define MULADDC_INIT \
748 #define MULADDC_CORE \
749 "lw $14, 0($10) \n\t" \
750 "multu $13, $14 \n\t" \
751 "addi $10, $10, 4 \n\t" \
754 "addu $14, $12, $14 \n\t" \
755 "lw $15, 0($11) \n\t" \
756 "sltu $12, $14, $12 \n\t" \
757 "addu $15, $14, $15 \n\t" \
758 "sltu $14, $15, $14 \n\t" \
759 "addu $12, $12, $9 \n\t" \
760 "sw $15, 0($11) \n\t" \
761 "addu $12, $12, $14 \n\t" \
762 "addi $11, $11, 4 \n\t"
764 #define MULADDC_STOP \
768 : "=m" (c), "=m" (d), "=m" (s) \
769 : "m" (s), "m" (d), "m" (c), "m" (b) \
770 : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
776 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
778 #define MULADDC_INIT \
784 #define MULADDC_CORE \
789 __asm add eax, [edi] \
794 #if defined(MBEDTLS_HAVE_SSE2)
796 #define EMIT __asm _emit
798 #define MULADDC_HUIT \
799 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
800 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
801 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
802 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
803 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
804 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
805 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
806 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
807 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
808 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
809 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
810 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
811 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
812 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
813 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
814 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
815 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
816 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
817 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
818 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
819 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
820 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
821 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
822 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
823 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
824 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
825 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
826 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
827 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
828 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
829 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
830 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
831 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
832 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
833 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
834 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
835 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
836 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
837 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
838 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
839 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
840 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
841 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
842 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
843 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
844 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
845 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
846 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
847 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
848 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
849 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
850 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
851 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
852 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
853 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
854 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
855 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
856 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
857 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
858 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
859 EMIT 0x0F EMIT 0x7E EMIT 0xC9
861 #define MULADDC_STOP \
862 EMIT 0x0F EMIT 0x77 \
869 #define MULADDC_STOP \
879 #if !defined(MULADDC_CORE)
880 #if defined(MBEDTLS_HAVE_UDBL)
882 #define MULADDC_INIT \
885 mbedtls_mpi_uint r0, r1;
887 #define MULADDC_CORE \
888 r = *(s++) * (mbedtls_t_udbl) b; \
889 r0 = (mbedtls_mpi_uint) r; \
890 r1 = (mbedtls_mpi_uint)( r >> biL ); \
891 r0 += c; r1 += (r0 < c); \
892 r0 += *d; r1 += (r0 < *d); \
895 #define MULADDC_STOP \
899 #define MULADDC_INIT \
901 mbedtls_mpi_uint s0, s1, b0, b1; \
902 mbedtls_mpi_uint r0, r1, rx, ry; \
903 b0 = ( b << biH ) >> biH; \
906 #define MULADDC_CORE \
907 s0 = ( *s << biH ) >> biH; \
908 s1 = ( *s >> biH ); s++; \
909 rx = s0 * b1; r0 = s0 * b0; \
910 ry = s1 * b0; r1 = s1 * b1; \
911 r1 += ( rx >> biH ); \
912 r1 += ( ry >> biH ); \
913 rx <<= biH; ry <<= biH; \
914 r0 += rx; r1 += (r0 < rx); \
915 r0 += ry; r1 += (r0 < ry); \
916 r0 += c; r1 += (r0 < c); \
917 r0 += *d; r1 += (r0 < *d); \
920 #define MULADDC_STOP \