41 #ifndef POLARSSL_BN_MUL_H
42 #define POLARSSL_BN_MUL_H
46 #if defined(POLARSSL_HAVE_ASM)
51 #define MULADDC_INIT \
60 #define MULADDC_CORE \
66 addl (%%edi), %%eax; \
72 #if defined(POLARSSL_HAVE_SSE2)
74 #define MULADDC_HUIT \
78 movd (%%edi), %%mm3; \
80 movd (%%esi), %%mm2; \
81 pmuludq %%mm0, %%mm2; \
82 movd 4(%%esi), %%mm4; \
83 pmuludq %%mm0, %%mm4; \
84 movd 8(%%esi), %%mm6; \
85 pmuludq %%mm0, %%mm6; \
86 movd 12(%%esi), %%mm7; \
87 pmuludq %%mm0, %%mm7; \
89 movd 4(%%edi), %%mm3; \
91 movd 8(%%edi), %%mm5; \
93 movd 12(%%edi), %%mm4; \
95 movd %%mm1, (%%edi); \
96 movd 16(%%esi), %%mm2; \
97 pmuludq %%mm0, %%mm2; \
99 movd 20(%%esi), %%mm4; \
100 pmuludq %%mm0, %%mm4; \
101 paddq %%mm3, %%mm1; \
102 movd 24(%%esi), %%mm6; \
103 pmuludq %%mm0, %%mm6; \
104 movd %%mm1, 4(%%edi); \
106 movd 28(%%esi), %%mm3; \
107 pmuludq %%mm0, %%mm3; \
108 paddq %%mm5, %%mm1; \
109 movd 16(%%edi), %%mm5; \
110 paddq %%mm5, %%mm2; \
111 movd %%mm1, 8(%%edi); \
113 paddq %%mm7, %%mm1; \
114 movd 20(%%edi), %%mm5; \
115 paddq %%mm5, %%mm4; \
116 movd %%mm1, 12(%%edi); \
118 paddq %%mm2, %%mm1; \
119 movd 24(%%edi), %%mm5; \
120 paddq %%mm5, %%mm6; \
121 movd %%mm1, 16(%%edi); \
123 paddq %%mm4, %%mm1; \
124 movd 28(%%edi), %%mm5; \
125 paddq %%mm5, %%mm3; \
126 movd %%mm1, 20(%%edi); \
128 paddq %%mm6, %%mm1; \
129 movd %%mm1, 24(%%edi); \
131 paddq %%mm3, %%mm1; \
132 movd %%mm1, 28(%%edi); \
139 #define MULADDC_STOP \
147 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
148 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
149 : "eax", "ecx", "edx", "esi", "edi" \
154 #define MULADDC_STOP \
161 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
162 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
163 : "eax", "ecx", "edx", "esi", "edi" \
168 #if defined(__amd64__) || defined (__x86_64__)
170 #define MULADDC_INIT \
180 #define MULADDC_CORE \
182 movq (%%rsi), %%rax; \
189 addq %%rax, (%%rdi); \
194 #define MULADDC_STOP \
200 : "=m" (c), "=m" (d), "=m" (s) \
201 : "m" (s), "m" (d), "m" (c), "m" (b) \
202 : "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" \
207 #if defined(__mc68020__) || defined(__mcpu32__)
209 #define MULADDC_INIT \
219 #define MULADDC_CORE \
221 movel %%a2@+, %%d1; \
222 mulul %%d2, %%d4:%%d1; \
230 #define MULADDC_STOP \
236 : "=m" (c), "=m" (d), "=m" (s) \
237 : "m" (s), "m" (d), "m" (c), "m" (b) \
238 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
241 #define MULADDC_HUIT \
243 movel %%a2@+, %%d1; \
244 mulul %%d2, %%d4:%%d1; \
248 movel %%a2@+, %%d1; \
249 mulul %%d2, %%d3:%%d1; \
253 movel %%a2@+, %%d1; \
254 mulul %%d2, %%d4:%%d1; \
258 movel %%a2@+, %%d1; \
259 mulul %%d2, %%d3:%%d1; \
263 movel %%a2@+, %%d1; \
264 mulul %%d2, %%d4:%%d1; \
268 movel %%a2@+, %%d1; \
269 mulul %%d2, %%d3:%%d1; \
273 movel %%a2@+, %%d1; \
274 mulul %%d2, %%d4:%%d1; \
278 movel %%a2@+, %%d1; \
279 mulul %%d2, %%d3:%%d1; \
288 #if defined(__powerpc__) || defined(__ppc__)
289 #if defined(__powerpc64__) || defined(__ppc64__)
291 #if defined(__MACH__) && defined(__APPLE__)
293 #define MULADDC_INIT \
305 #define MULADDC_CORE \
317 #define MULADDC_STOP \
326 : "=m" (c), "=m" (d), "=m" (s) \
327 : "m" (s), "m" (d), "m" (c), "m" (b) \
328 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
334 #define MULADDC_INIT \
341 addi %%r3, %%r3, -8; \
342 addi %%r4, %%r4, -8; \
343 addic %%r5, %%r5, 0; \
346 #define MULADDC_CORE \
349 mulld %%r8, %%r7, %%r6; \
350 mulhdu %%r9, %%r7, %%r6; \
351 adde %%r8, %%r8, %%r5; \
354 addc %%r8, %%r8, %%r7; \
355 stdu %%r8, 8(%%r4); \
358 #define MULADDC_STOP \
361 addi %%r4, %%r4, 8; \
362 addi %%r3, %%r3, 8; \
367 : "=m" (c), "=m" (d), "=m" (s) \
368 : "m" (s), "m" (d), "m" (c), "m" (b) \
369 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
376 #if defined(__MACH__) && defined(__APPLE__)
378 #define MULADDC_INIT \
390 #define MULADDC_CORE \
402 #define MULADDC_STOP \
411 : "=m" (c), "=m" (d), "=m" (s) \
412 : "m" (s), "m" (d), "m" (c), "m" (b) \
413 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
418 #define MULADDC_INIT \
425 addi %%r3, %%r3, -4; \
426 addi %%r4, %%r4, -4; \
427 addic %%r5, %%r5, 0; \
430 #define MULADDC_CORE \
432 lwzu %%r7, 4(%%r3); \
433 mullw %%r8, %%r7, %%r6; \
434 mulhwu %%r9, %%r7, %%r6; \
435 adde %%r8, %%r8, %%r5; \
438 addc %%r8, %%r8, %%r7; \
439 stwu %%r8, 4(%%r4); \
442 #define MULADDC_STOP \
445 addi %%r4, %%r4, 4; \
446 addi %%r3, %%r3, 4; \
451 : "=m" (c), "=m" (d), "=m" (s) \
452 : "m" (s), "m" (d), "m" (c), "m" (b) \
453 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
461 #if defined(__sparc__) && defined(__sparc64__)
463 #define MULADDC_INIT \
472 #define MULADDC_CORE \
477 umul %%o3, %%o4, %%o4; \
478 addcc %%o4, %%o2, %%o4; \
480 addx %%g1, 0, %%g1; \
481 addcc %%o4, %%o5, %%o4; \
483 addx %%g1, 0, %%o2; \
487 #define MULADDC_STOP \
493 : "=m" (c), "=m" (d), "=m" (s) \
494 : "m" (s), "m" (d), "m" (c), "m" (b) \
495 : "g1", "o0", "o1", "o2", "o3", "o4", \
500 #if defined(__sparc__) && !defined(__sparc64__)
502 #define MULADDC_INIT \
511 #define MULADDC_CORE \
516 umul %%o3, %%o4, %%o4; \
517 addcc %%o4, %%o2, %%o4; \
519 addx %%g1, 0, %%g1; \
520 addcc %%o4, %%o5, %%o4; \
522 addx %%g1, 0, %%o2; \
526 #define MULADDC_STOP \
532 : "=m" (c), "=m" (d), "=m" (s) \
533 : "m" (s), "m" (d), "m" (c), "m" (b) \
534 : "g1", "o0", "o1", "o2", "o3", "o4", \
540 #if defined(__microblaze__) || defined(microblaze)
542 #define MULADDC_INIT \
549 andi r7, r6, 0xffff; \
553 #define MULADDC_CORE \
567 bslli r10, r10, 16; \
568 bslli r11, r11, 16; \
582 #define MULADDC_STOP \
588 : "=m" (c), "=m" (d), "=m" (s) \
589 : "m" (s), "m" (d), "m" (c), "m" (b) \
590 : "r3", "r4" "r5", "r6", "r7", "r8", \
591 "r9", "r10", "r11", "r12", "r13" \
596 #if defined(__tricore__)
598 #define MULADDC_INIT \
608 #define MULADDC_CORE \
610 ld.w %%d0, [%%a2+]; \
611 madd.u %%e2, %%e4, %%d0, %%d1; \
613 addx %%d2, %%d2, %%d0; \
614 addc %%d3, %%d3, 0; \
616 st.w [%%a3+], %%d2; \
619 #define MULADDC_STOP \
625 : "=m" (c), "=m" (d), "=m" (s) \
626 : "m" (s), "m" (d), "m" (c), "m" (b) \
627 : "d0", "d1", "e2", "d4", "a2", "a3" \
634 #if defined(__thumb__) && !defined(__thumb2__)
636 #define MULADDC_INIT \
650 #define MULADDC_CORE \
683 #define MULADDC_STOP \
689 : "=m" (c), "=m" (d), "=m" (s) \
690 : "m" (s), "m" (d), "m" (c), "m" (b) \
691 : "r0", "r1", "r2", "r3", "r4", "r5", \
692 "r6", "r7", "r8", "r9", "cc" \
697 #define MULADDC_INIT \
706 #define MULADDC_CORE \
711 umlal r2, r5, r3, r4; \
717 #define MULADDC_STOP \
723 : "=m" (c), "=m" (d), "=m" (s) \
724 : "m" (s), "m" (d), "m" (c), "m" (b) \
725 : "r0", "r1", "r2", "r3", "r4", "r5", \
733 #if defined(__alpha__)
735 #define MULADDC_INIT \
744 #define MULADDC_CORE \
761 #define MULADDC_STOP \
767 : "=m" (c), "=m" (d), "=m" (s) \
768 : "m" (s), "m" (d), "m" (c), "m" (b) \
769 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
773 #if defined(__mips__)
775 #define MULADDC_INIT \
784 #define MULADDC_CORE \
791 addu $14, $12, $14; \
793 sltu $12, $14, $12; \
794 addu $15, $14, $15; \
795 sltu $14, $15, $14; \
798 addu $12, $12, $14; \
802 #define MULADDC_STOP \
808 : "=m" (c), "=m" (d), "=m" (s) \
809 : "m" (s), "m" (d), "m" (c), "m" (b) \
810 : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
816 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
818 #define MULADDC_INIT \
824 #define MULADDC_CORE \
829 __asm add eax, [edi] \
834 #if defined(POLARSSL_HAVE_SSE2)
836 #define EMIT __asm _emit
838 #define MULADDC_HUIT \
839 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
840 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
841 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
842 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
843 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
844 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
845 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
846 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
847 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
848 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
849 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
850 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
851 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
852 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
853 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
854 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
855 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
856 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
857 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
858 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
859 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
860 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
861 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
862 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
863 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
864 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
865 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
866 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
867 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
868 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
869 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
870 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
871 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
872 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
873 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
874 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
875 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
876 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
877 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
878 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
879 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
880 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
881 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
882 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
883 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
884 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
885 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
886 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
887 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
888 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
889 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
890 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
891 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
892 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
893 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
894 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
895 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
896 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
897 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
898 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
899 EMIT 0x0F EMIT 0x7E EMIT 0xC9
901 #define MULADDC_STOP \
902 EMIT 0x0F EMIT 0x77 \
909 #define MULADDC_STOP \
919 #if !defined(MULADDC_CORE)
920 #if defined(POLARSSL_HAVE_UDBL)
922 #define MULADDC_INIT \
927 #define MULADDC_CORE \
928 r = *(s++) * (t_udbl) b; \
931 r0 += c; r1 += (r0 < c); \
932 r0 += *d; r1 += (r0 < *d); \
935 #define MULADDC_STOP \
939 #define MULADDC_INIT \
941 t_uint s0, s1, b0, b1; \
942 t_uint r0, r1, rx, ry; \
943 b0 = ( b << biH ) >> biH; \
946 #define MULADDC_CORE \
947 s0 = ( *s << biH ) >> biH; \
948 s1 = ( *s >> biH ); s++; \
949 rx = s0 * b1; r0 = s0 * b0; \
950 ry = s1 * b0; r1 = s1 * b1; \
951 r1 += ( rx >> biH ); \
952 r1 += ( ry >> biH ); \
953 rx <<= biH; ry <<= biH; \
954 r0 += rx; r1 += (r0 < rx); \
955 r0 += ry; r1 += (r0 < ry); \
956 r0 += c; r1 += (r0 < c); \
957 r0 += *d; r1 += (r0 < *d); \
960 #define MULADDC_STOP \
Multi-precision integer library.