[Haskell-cafe] Re: [Haskell] rigid variables

[Robert Dockins]

[moved to cafe]

On Jul 20, 2006, at 12:48 PM, Rodney D Price wrote:

> I've gotten this sort of error several times, which mysteriously  
> disappears
> when I add more functions to the code:
>
> storeError.hs:13:38:
>     Couldn't match expected type `a' (a rigid variable)
>            against inferred type `String'
>       `a' is bound by the type signature for `throwError'
>         at <no location info>
>       Expected type: a
>       Inferred type: String
>     In the first argument of `return', namely `msg'
>     In the call (return msg)
>
> (This is GHCi.)  The code is below.  The type variable a can't be  
> bound to
> String, obviously, but a relative novice like myself has no idea  
> why.  Can
> someone tell me?
>
> Thanks,
>
> -Rod
>
> --
> module Store where
>
> import Control.Monad.Error
> import Control.Concurrent.STM
>
> data StoreError = Default String
>
> instance Error StoreError where
>     noMsg  = Default "Store error"
>     strMsg = Default
>
> instance MonadError StoreError STM where
>     throwError (Default msg) = return msg


Lets take a look here at the definition of MonadError from  
Control.Monad.Error:

class Monad m => MonadError e m | m -> e where
     throwError :: e -> m a
     catchError :: m a -> (e -> m a) -> m a


In the signature for 'throwError' there are three type variables: e,  
m and a.  e and m are bound by the instance declaration, but a is  
free.  In Haskell the rule is that free variables are implicitly  
bound with a universal quantifier.  So, the type for throwError can  
be regarded as,

     throwError :: forall a. e -> m a,  for some concrete choices of  
e and m which are determined by the instance.

The forall means that a user of this function can put any type there  
that he likes.  In other words, the monadic action created by  
'throwError', when executed, evaluates to _a value of any type at  
all_ (yes, I know the terminology is a little loose here).  That  
means you can't just 'return msg', which has type 'm String' because  
a user might have used throwError to create an action of type 'm  
Int', for example.  In fact, you won't really be able to return  
anything at all, because there isn't any way to write a program that  
can generate a value of any unknown type.  This should hopefully  
correspond to your intuition about what throwing an exception does.

The error generated by the typechecker basically tells you that the  
function you have written is not polymorphic enough.  It has type  
'StoreError -> STM String' rather than 'forall a. StoreError -> STM  
a' as it ought.

I think perhaps you have misunderstood how MonadError is used.  The  
idea is to expose to users a particular non-local control flow  
construct (throw/catch style exceptions) by hiding all the stuff  
necessary for that inside the monad plumbing.  Usually, whoever  
writes the monad itself will provide the necessary instances.  It's  
often not possible to write instances like this by using the external  
API of the monad.  This is particularly the case for the abstract  
monads available in GHC (IO, ST, and STM).

In short, I don't think you'll be successful in writing a  
'MonadError' instance for STM that has the customary semantics.  What  
you may be looking for is the ErrorT monad transformer, which will  
let you layer error handling over STM.  It's hard to know with what  
info you've provided here.  If you give a few more details on what  
you're trying to accomplish, someone may be able to give you a push  
in the correct direction.



Rob Dockins

Speak softly and drive a Sherman tank.
Laugh hard; it's a long way to the bank.
           -- TMBG