How does the Haskell compiler "know" that IO cannot be unwrapped?

221 views Asked by At

Obviously, the following function is impossible, because it is impossible to unwrap an IO value permanently (ignoring unsafePerformIO or similar):

unwrapIO :: IO String -> String
unwrapIO (IO str) = str

However, similar functions such as the following are possible:

unwrapJust :: Maybe String -> String
unwrapJust (Just str) = str
unwrapJust Nothing = "ignore this plz"

I fully understand the reasoning behind why #2 is possible but #1 is not, but I do not understand how. Can I also make my own types that are not unwrappable?

3

There are 3 answers

0
Li-yao Xia On BEST ANSWER

Just and Nothing are data constructors for the type Maybe a. IO has no data constructors to speak of (in GHC it actually has constructors but they're really implementation details of GHC, and other implementations might define IO differently).

unwrapIO (IO str) = str doesn't make sense in the same way unwrapMaybe (Maybe str) = str doesn't make sense. IO and Maybe are not data constructors, so you cannot pattern-match on them.

0
snak On

It's because the data constructor of IO is not exported. I mean, you can think it's not exported.

You can prevent your own type from being unwrapped by using the same strategy.

module Test (Test, test) where

data Test a = MkTest a

test :: a -> Test a
test = MkTest

You can create a value of Test using test, but you cannot unwrap it using pattern-match because MkTest is not exported.

0
WorldSEnder On

I believe that while the existing answers are mostly true, there is a deeper reason why IO can not be unwrapped. Conceptually, type IO a = RealWorld -> (a, RealWorld). IO is a function type (in real implementations hidden behind a newtype wrapper or equivalent machinery).

So, how would you go about unwrapping a function? Easy, you just call it!. But how are you going to get an instance of RealWorld? That is the truer primitive: you can not construct a RealWorld, there is only ever one.

The monadic instance of course can just pass RealWorld as a state, kind of like StateT and in the end, the only instance ever constructed is at startup of the program, then it is passed down between the IOs.

Coming back to reality again, this is (again) a lie. You actually can get hold of an instance of RealWorld and you can call an IO a action "ahead of time". There is a function that does exactly what you asked for. It is called System.IO.Unsafe.unsafePerformIO :: IO a -> a though it is in an unsafe package for a reason.