subdividing octree

I am currently using this function to subdivide an AABB in an octree. I am dissatisfied with it, as it uses 3 loops and std::pow(), maybe there’s some trick to simplify this? Please show me some C++ power. The AABBs are all center-radius AABBs.

template <class U1, class U2>
void subdivide_aabb(AABB<U1, U2> const& aabb,
  std::vector<AABB<U1, U2> >& aabbs)
{
  BOOST_ASSERT(aabbs.empty());

  AABB<U1, U2> naabb;
  naabb.b = 0.5 * aabb.b;

  for (unsigned char i(0); i != 2; ++i)
  {
    for (unsigned char j(0); j != 2; ++j)
    {
      for (unsigned char k(0); k != 2; ++k)
      {
        naabb.a = aabb.a;

        naabb.a += U1(std::pow(-1, i) * naabb.b(0),
          std::pow(-1, j) * naabb.b(1),
          std::pow(-1, k) * naabb.b(2));

        aabbs.push_back(naabb);
      }
    }
  }

  BOOST_ASSERT(8 == aabbs.size());
}

I got this idea:

        naabb.a += U1((2 * i - 1) * naabb.b(0),
          (2 * j - 1) * naabb.b(1),
          (2 * k - 1) * naabb.b(2));

I prefer to keep things simple and coherent, for things that don’t need optimization.
Index with bit 1 set: positive x, index with bit 2 set: positive y, bit 3: z

for(int i=0;i<8;i++){
	AABB aabb;
	aabb.minBB.z = minBB.z;
	aabb.maxBB.z = avgBB.z;
	aabb.minBB.y = minBB.y;
	aabb.maxBB.y = avgBB.y;
	aabb.minBB.x = minBB.x;
	aabb.maxBB.x = avgBB.x;

	if(i & 4){ // greater z
		aabb.minBB.z = avgBB.z;
		aabb.maxBB.z = maxBB.z;	
	}
	if(i & 2){ // greater y
		aabb.minBB.y = avgBB.y;
		aabb.maxBB.y = maxBB.y;
	}
	if(i & 1){ // greater x
		aabb.minBB.x = avgBB.x;
		aabb.maxBB.x = maxBB.x;
	}
	
	
	child = new OCTREE;
	parent->nodes[i]=child;
	child->parent = parent;
	child->aabb = aabb;
}

If speed is an issue, I’d use CMOVxx and SSE intrinsics, inline asm or external asm; and of course get around the allocalization.