FDEs previously were loaded into a red/black tree and searched. This is
significantly more expensive than a binary search over a slice. Not sure
what I was thinking using a red/black tree - this binary search
implementation is significantly more efficient.