#include <tfm.h>

#if defined(TFM_X86)

#define COMBA_START

/* clear the chaining variables */
#define COMBA_CLEAR \
   c0 = c1 = c2 = 0;

/* forward the carry to the next digit */
#define COMBA_FORWARD \
   do { c0 = c1; c1 = c2; c2 = 0; } while (0);

/* store the first sum */
#define COMBA_STORE(x) \
   x = c0;

/* store the second sum [carry] */
#define COMBA_STORE2(x) \
   x = c1;

/* anything you need at the end */
#define COMBA_FINI

/* this should multiply i and j  */
#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");

#endif


/* generic PxQ multiplier */
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;

   COMBA_START;
   COMBA_CLEAR;
   
   /* get size of output and trim */
   pa = A->used + B->used;
   if (pa >= FP_SIZE) {
      pa = FP_SIZE-1;
   }

   if (A == C || B == C) {
      fp_zero(&tmp);
      dst = &tmp;
   } else {
      fp_zero(C);
      dst = C;
   }

   for (ix = 0; ix < pa; ix++) {
      /* get offsets into the two bignums */
      ty = MIN(ix, B->used-1);
      tx = ix - ty;

      /* setup temp aliases */
      tmpx = A->dp + tx;
      tmpy = B->dp + ty;

      /* this is the number of times the loop will iterrate, essentially its 
         while (tx++ < a->used && ty-- >= 0) { ... }
       */
      iy = MIN(A->used-tx, ty+1);

      /* execute loop */
      COMBA_FORWARD;
      for (iz = 0; iz < iy; ++iz) {
          MULADD(*tmpx++, *tmpy--);
      }

      /* store term */
      COMBA_STORE(dst->dp[ix]);
  }
  /* store final carry */
  COMBA_STORE2(dst->dp[ix]);
  COMBA_FINI;

  dst->used = pa;
  fp_clamp(dst);
  dst->sign = dst->used ? A->sign ^ B->sign : FP_ZPOS;
  fp_copy(dst, C);
}