[cousin_it]

GHC's implementation of IO isn't the only one possible. For example, if Haskell had only two available IO operations (read a character or print a character), the following would be a workable definition of the IO type:

    data IO a = PutChar (Char, IO a)
              | GetChar (Char -> IO a)
              | Return a

Programs can work by constructing a value of type IO () and defining it as main, just like in regular Haskell. Note that IO is no longer a wrapped impure function, and there are no magical tokens in sight. In fact, the above implementation doesn't even have to be opaque, it can be completely exposed to programmers. Nevertheless, it's just as easy for the runtime to interpret in a sequential way, and there won't be any problems with accidentally optimizing away or memoizing anything.