Recommend MonadFailDesugaring?

[snoyberg]

I haven't used this extension myself previously, so my understanding may not be perfect. Consider this script:

#!/usr/bin/env stack
-- stack --resolver nightly-2018-04-15 script
{-# LANGUAGE NoImplicitPrelude #-}
-- {-# LANGUAGE MonadFailDesugaring #-}
{-# OPTIONS_GHC -Wall -Werror #-}
import RIO

main :: IO ()
main = runRIO () $ do
  Just x <- return Nothing
  return x

As is, it will crash at runtime with:

Main.hs: user error (Pattern match failure in do expression at /Users/michael/Desktop/Main.hs:10:3-8)

Note that even with -Wall -Werror turned on, the compiler does nothing to prevent us from the partial pattern match Just x <- return Nothing. By contrast, if we uncomment the {-# LANGUAGE MonadFailDesugaring #-} line, we get a compiler time error due to a missing MonadFail instance instead:

/Users/michael/Desktop/Main.hs:10:3: error:
    • No instance for (Control.Monad.Fail.MonadFail (RIO ()))
        arising from a do statement
        with the failable pattern ‘Just x’
    • In a stmt of a 'do' block: Just x <- return Nothing
      In the second argument of ‘($)’, namely
        ‘do Just x <- return Nothing
            return x’
      In the expression:
        runRIO ()
          $ do Just x <- return Nothing
               return x
   |
10 |   Just x <- return Nothing

It seems preferable to me to totally block partial pattern matches via this language extension and lack of MonadFail instance for RIO.

[chrisdone]

By the same token we should recommend -Wincomplete-uni-patterns which is not enabled by -Wall.

stack --resolver ghc-8.4.1 exec ghci
GHCi, version 8.4.1: http://www.haskell.org/ghc/  :? for help
Loaded GHCi configuration from /Users/chris/.ghci
Prelude> :set -Wall
Prelude> let f = x where Just x = Nothing
Prelude> :set -fwarn-incomplete-uni-patterns
Prelude> let g = x where Just x = Nothing

<interactive>:5:17: warning: [-Wincomplete-uni-patterns]
    Pattern match(es) are non-exhaustive
    In a pattern binding: Patterns not matched: Nothing
Prelude> 

[akhra]

+1 to both of these.

It's unfortunate that -Wincomplete-uni-patterns doesn't catch the do-notation case, or I'd prefer just that by default. I think you're far more likely to have situations where you're stuck in Maybe but legitimately know it's a Just value in monadic code, due to the shape of other peoples' APIs; and the MonadFail error may not be clear to everyone, while the warning is nicely succinct. But it's better than nothing, and trivial to work around.