All current (2001) S/390 and z/Architecture machines are single-issue,
but some newer machines have a deep pipeline.  Software-pipelining is
therefore beneficial.

* mpn_add_n, mpn_sub_n: Use code along the lines below.  Two-way unrolling
  would be adequate.

  mp_limb_t
  mpn_add_n (mp_ptr rp, mp_srcptr up, mp_srcptr vp, mp_size_t n)
  {
    mp_limb_t a, b, r, cy;
    mp_size_t i;
    mp_limb_t mm = -1;

    cy = 0;
    up += n;
    vp += n;
    rp += n;
    i = -n;
    do
      {
	a = up[i];
	b = vp[i];
	r = a + b + cy;
	rp[i] = r;
	cy = (((a & b) | ((a | b) & (r ^ mm)))) >> 31;
	i++;
      }
    while (i < 0);
    return cy;
  }

* mpn_lshift, mpn_rshift: Use SLDL/SRDL, and two-way unrolling.

* mpn_mul_1, mpn_addmul_1, mpn_submul_1: For machines with just signed
  multiply (MR), use two loops, similar to the corresponding VAX or
  POWER functions.  Handle carry like for mpn_add_n.