Posted on February 3, 2021

Generic Coerce

Limits of coerce

Someone asked a question on stack overflow that got me thinking about Generics. The question itself was about why Coercible is so limited. A very simple version of the question goes like this:

I have a data type that is isomorphic to Bool, like so:

Why is there no instance for Coercible Bool Foo? How can I make one?

It’s pretty obvious that these two types are isomorphic, and indeed, GHC has the same memory representation for them under the hood, but there is no way to generate a Coercible instance for them, which means that coerce is not allowed.

Of course, one could write the obvious coercion functions manually, but this is tedious and error-prone (especially with more complex types), and it may have poor performance too. Another option is to use unsafeCoerce, as in:

This has the advantage of being high performance, but there’s absolutely no type safety. If someone goes ahead and changes Foo so that it’s no longer isomorphic to Bool, this function will still type check, even if behavior becomes … suspect.

Generic Coerce

This is where Generic comes in. Using Generic, we can define a type-level isomorphism between two types, up to meta data. I’ll hide the Haskell header behind this dropdown, but then consider the following definitions:

Haskell extensions and imports

If we have two types that are both Generic, and we can coerce between their Reps, then we can coerce between them. What’s left is to define the instances of GenericCoerce. As we must for any Generic class, we need to handle the six cases: void, unit, sum, product, container, and wrapper. These all have quite natural definitions:

For instance, unit is obviously coercible to unit. Furthermore, if we can individually coerce the lefts and rights of two sum types, then we can coerce those sum types too.

Indeed, this seems to solve our problem and give us a whole new way to coerce between two types. In action, we have:

Non-Generic types

There are some obvious problems with genericCoerce. For one, it only works on types that are instances of Generic, so, for example, genericCoerce @Int @Int 3 doesn’t work at all. But in fact, because of the way we’ve written the K1 (recursive) case, even something like genericCoerce @[Int] @[Int] [3] doesn’t work.

In a perfect world, we’d be able to tell GHC to do something like “Use coerce if there’s a Coercible instance, and otherwise, try genericCoerce”, but this isn’t possible. The best I’ve been able to come up with is to use incoherent instances. Incoherent instances are typically considered dangerous, and I’m not sure this is really a good enough use for them (is there one?), but let’s have some fun! Let’s replace the instance we have with:

Here we’ve used a stupid trick to get around the fact that these two really should be duplicate instances. In one of them, we use K1 t x for the first parameter and Rec0 y for the second, and in the other, we flip that. Remember that type Rec0 = K1 R, and Generics doesn’t use any other type with K1 besides R. That means that these are functionally the same, but since GHC doesn’t know that, it gets around the duplicate instance issue. Yay!

We still can’t write genericCoerce @Int @Int 3, but now

works just fine, and we even get


Higher Kinded Newtype Wrappers

That stackoverflow question I mentioned that put me onto this topic actually demanded even more. The poster’s data types were actually built using the “higher-kinded types” design principle, and they look like:

The question is: Can we coerce between Typ and HKTyp Id v? Alas, genericCoerce breaks down again. This time, it’s because of that pesky use of f within the definition of HKListT. The types are not generically coercible because the Generic representation of a newtype wrapped data type and the data type itself are actually different, but they’re also not data coercible because Typ and HKTyp Id v are not Coercible so obviously Typ and Identity (HKType Id v) are not Coercible either. It turns out we can cheat a bit to get around this hiccup too, once again using some scary looking incoherent instances:

These instances allow us to unwrap (or rewrap if we’re going in the other direction) a newtype wrapper that has already been converted to its Rep. It’s a little janky, but it definitely does the trick. For instance, we can now define:

and then see them in action:


One really nice thing about coerce is that, like newtype wrappers, it has no runtime cost. On the other hand, genericCoerce definitely has a cost. But why should it? The whole idea behind genericCoerce is that the structure of the input type and the output type are the same, either because their representations are the same (and we’re assuming the Rep from Generic is a good proxy for the memory representation) or because they’re equivalent up to newtype wrappers. Therefore, if we trust that our types are doing the right thing, we can simply take a shortcut on the implementation. Indeed, we can just write:

The result will always be the same as the old genericCoerce, but the performance will be much better. Looked at another way, this is much safer than unsafeCoerce because it’s guarded by GenericCoerce.


As far as I’m aware, there are no guarantees from GHC about how memory is utilized, which means that our use of unsafeCoerce is still, to some extent, unsafe. That said, excluding the various incoherent instances we added, the original version of genericCoerce converts a data type with phantom type parameters into the same datatype just with different phantom parameters. Technically, I can provide no guarantee that GHC will store multiple instances of the same runtime data the same way, but it seems to me like a pretty easy assumption to make. Once we add in the incoherent instance shenanigans, we’re standing on less solid grounds, but the fact that it all works is some consolation.


Unfortunately, we were not able to build a truly more generic version of coerce (e.g., genericCoerce still can’t coerce between Int and anything else since Int is not an instance of Generic), but genericCoerce does allow us a surprising amount of freedom over Generic types. For instance, we get type-safe coercions between any two Generic data types that have the same structure, which could be useful for users of higher-kinded data types.

Another downside of our implementation is in the use of incoherent instances, which, as mentioned above are considered harmful. Incoherent instances have a bad habit of seeming pretty good until you stretch them to their limits, at which point they spontaneously do something unexpected. I’m not sure if that danger would rear its ugly hear here, but it is rather frustrating that there’s no other way to write the K1/Rec0 instance(s) that we were after.