GHC 7.11 errors on dictionary casting tricks

I am not saying this is a bug, it just seems to mess with techniques that assume white-box knowledge of GHC's dictionary-passing internals.

So, I have this function (in the gdiff fork I am fooling around with):

iFeelDirty :: (forall ts . List f ts => f t ts -> Con f t) -> (forall ts . IsList f ts -> f t ts -> Con f t)
iFeelDirty = unsafeCoerce

It compiles perfectly before 7.11, but now I get:

src/Data/Generic/Diff/Effectful.hs:139:14:
    No instance for (List f ts3)
    The type variable `ts3' is ambiguous
    Relevant bindings include
      iFeelDirty :: (forall ts. List f ts => f t ts -> Con f t)
                    -> forall ts. IsList f ts -> f t ts -> Con f t
        (bound at src/Data/Generic/Diff/Effectful.hs:139:1)
    In the expression: unsafeCoerce
    In an equation for `iFeelDirty': iFeelDirty = unsafeCoerce

For reference, this is the file I am looking at: https://github.com/ggreif/gdiff/blob/97b13e2dca0c35c2c53d5430f6f0700dc89938ec/src/Data/Generic/Diff/Effectful.hs#L138

Looks like the old trick (also used by ekmett's reflection lib?) fails to deliver.

Trac metadata

Trac field	Value
Version	7.11
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system
Architecture

changed weight to 5

added Tbug Trac import labels

I don't like that there's been a typechecker regression here, but this doesn't seem like an outright bug to me. The one bit I'm quite unsure of is the eta-contraction... and my reasoning-about-higher-rank-eta-contracted-functions neuron seems to be enjoying the snow day here. So I'll eta-expand:

iFeelDirty x = unsafeCoerce x

Now, I can see why your error appears. When you say x, GHC will try to instantiate the implicit parameters of x, namely ts and List f ts. But, GHC has no guidance as to what these should be, because it can't glean anything useful from unsafeCoerce. Some more type annotations might be helpful, as might be newtype wrappers to accurately control what you're unsafeCoerceing.

So, I have two conclusions:

• I'm happy enough with an error in the eta-expanded version, but I agree that the eta-contracted one should probably work.
• I don't think this bug shows up a problem with the dirty dictionary-casting trick. That should still be valid. It's just a typechecker behavior change, which (I believe) could be overcome with more types.

A recent (somewhat obscure) change is that GHC actually does eta-expansion to make higher-rank types match up. That is often useful (i.e. makes more programs typecheck). And really it is confusing if eta-expansion can make a difference to type checking.

So I'm disinclined to "fix" this. Or at least I'd require more motivation!

Simon

added Pnormal label