What’s wrong with ImplicitParams

Implicit parameters, enabled by the ImplicitParams extension in GHC, often come up in Haskell discussions,¹ with the statement that they’re so convenient, yet nobody² seems to use them. Why is that?

The problem is incoherence. Type class resolution works in a very binary way: either it finds one, singular instance with one behavior and we compile, or it finds none at all and we don’t compile. It’s never³ the case that changing our code will change a resolved instance. I can change my import list, or update my type-signatures.

Incoherence breaks this guarantee. Implicit parameters rely on incoherence to work, because they piggyback on the type-class machinery with magical instances that appear as needed.

Let’s see an example of how this leads to surprising results. The two declarations below produce different results.

> let ?myparam = 123 in terror
(123,123)
> let ?myparam = 123 in horror
(123,456)

Check out the code and find the reason why.

{-# LANGUAGE ImplicitParams #-}
horror :: (?myparam :: Int) => (Int, Int)
horror =
  let result :: (?myparam :: Int) => Int
      result = ?myparam
   in ( result
      , let ?myparam = 456
         in result)

terror :: (?myparam :: Int) => (Int, Int)
terror =
  let result :: Int
      result = ?myparam
   in ( result
      , let ?myparam = 456
         in result)

You found it; it’s that we added a type signature to result. That changed the behavior! If you find this surprising, it’s because you’re used to coherence. In normal Haskell, an inferred type and an explicit type don’t produce different code.

Here’s another, more insidious, way to evoke the problem:

{-# LANGUAGE ImplicitParams, NoMonomorphismRestriction #-}
terror :: (?myparam :: Int) => (Int, Int)
terror =
  let result = ?myparam
   in ( result
      , let ?myparam = 456
         in result)

Try removing NoMonomorphismRestriction. The output is different depending on whether this extension is on or not. Be afraid, be very afraid.

If this constraint was Show, I wouldn’t be able to reproduce this difference, because there’s only one Show Int instance.

Implicit parameters break fundamental expectations about one of Haskell’s most basic features: its type inference. For me, this is reason enough to avoid it.

Thanks Andrew Martin for insightful discussion on this topic a few years ago.