Don't do (x - (y - z)) --> (x + (z - y)) on floating-point types, because