Some weird interactions between Monomorphism Restriction and Template Haskell

Today I’m going to look at a weird issue I encountered this past weekend while working on a DSL in Haskell.

I’ll start with the code. As this example uses Template Haskell, we need to have the source broken up into two files:


        TemplateHaskell #-}
module Testa where

import Language.Haskell.TH

someth = [| () |]

unit = ()

mrIssue :: (Monad m) => b -> m b
mrIssue = return


        TemplateHaskell #-}
module Testb where

import Testa

g = $(someth)
foo1 = mrIssue $(someth)

f = ()
foo2 = mrIssue ()

h = unit
foo3 = mrIssue unit

Now, if we fire up ghci Testb.hs, we get the following:

$ ghci Testb.hs 
GHCi, version 6.10.4:  :? for help
Loading package ghc-prim ... linking ... done.
Loading package integer ... linking ... done.
Loading package base ... linking ... done.
[1 of 2] Compiling Testa            ( Testa.hs, interpreted )
[2 of 2] Compiling Testb            ( Testb.hs, interpreted )
Loading package syb ... linking ... done.
Loading package array- ... linking ... done.
Loading package packedstring- ... linking ... done.
Loading package containers- ... linking ... done.
Loading package pretty- ... linking ... done.
Loading package template-haskell ... linking ... done.
Ok, modules loaded: Testb, Testa.
*Testb> :t foo1
foo1 :: (Monad m) => m ()
*Testb> :t g

    Ambiguous type variable `m' in the constraint:
      `Monad m' arising from a use of `g' at :1:0
    Probable fix: add a type signature that fixes these type variable(s)

Notice that, while we’re able to get the type for foo1, for some reason g is ill-typed with an ambiguous type variable in the constraint Monad m. The only problem, of course, is that when we look at our source we don’t see any way in which the type for g should have this constraint!

So maybe the problem is that g needs a type signature. But if we go in and modify Testb.hs, giving it

g :: ()
g = $(someth)

then we get the following from ghci:

    Could not deduce (Monad m) from the context ()
      arising from a use of `mrIssue' at Testb.hs:10:7-23
    Possible fix:
      add (Monad m) to the context of the type signature for `g'
    In the expression: mrIssue ($someth)
    In the definition of `foo1': foo1 = mrIssue ($someth)
Failed, modules loaded: Testa.

So now it's upset about the type for foo1. Fine. Let's give it a signature as well:

g :: ()
g = $(someth)
foo1 :: (Monad m) => m ()
foo1 = mrIssue $(someth)

Now we get a new error from ghci:

    Contexts differ in length
      (Use -XRelaxedPolyRec to allow this)
    When matching the contexts of the signatures for
      g :: ()
      foo1 :: forall (m :: * -> *). (Monad m) => m ()
    The signature contexts in a mutually recursive group should all be identical
    When generalising the type(s) for g, foo1
Failed, modules loaded: Testa.

If we add RelaxedPolyRec to our list of LANGUAGE extensions, the problem does, indeed, go away. In this case, we can even remove our type signature for g or for foo1, but not both -- we need to have at least one of them present.

Lastly, if we go back to our original source given above, but replace the signature for mrIssue with

mrIssue :: b -> IO b
mrIssue = return

then we can remove the NoMonomorphismRestriction, and everything works just fine (we can :t g and :t foo1 without any problems).

I'm entirely unsure of what's going on here. Any theories?