#define FP_SIZE \
((4096+(4*(int)((8) * sizeof(fp_digit))))/(int)((8) * sizeof(fp_digit)))

typedef unsigned long fp_digit;
typedef struct {
    fp_digit dp[FP_SIZE];
    int      used, 
             sign;
} fp_int;

#define MULADD(i, j)                                      \
asm(                                                     \
     "movl  %6,%%eax     \n\t"                            \
     "mull  %7           \n\t"                            \
     "addl  %%eax,%0     \n\t"                            \
     "adcl  %%edx,%1     \n\t"                            \
     "adcl  $0,%2        \n\t"                            \
     :"=r"(c0), "=r"(c1), "=r"(c2): "0"(c0), "1"(c1), "2"(c2), "m"(i), "m"(j)  :"%eax","%edx","%cc");

void fp_mul_comba(fp_int *A, fp_int *B, fp_int *C)
{
   int       ix, iy, iz, tx, ty, pa;
   fp_digit  c0, c1, c2, *tmpx, *tmpy;
   fp_int    tmp, *dst;
   for (ix = 0; ix < pa; ix++)
      for (iz = 0; iz < iy; ++iz)
          MULADD(*tmpx++, *tmpy--);
}