Re: Newtype wrappers

On 1/14/13 2:42 PM, Johan Tibell wrote:

On Mon, Jan 14, 2013 at 11:14 AM, Andrea Vezzosi wrote:

...

Have you considered the effect on types like Data.Set that use the uniqueness of typeclass instances to maintain invariants? e.g. even when we have "newtype X = X Y" coercing "Set X" to "Set Y" can produce a tree with the wrong shape for the Ord instance of Y. I was just going to say that. Changing newtypes changes instances, which isn't safe in the general case.

Perhaps it would be useful for data structures that need to remain opaque/abstract to be allowed to declare such explicitly, either with special syntax, or a distinguished pragma? Also, I'm fond of Roman's "coerce" proposal, because I can imagine cases where explicit declaration of wrap/unwrap functions might not necessarily make sense. My understanding of the "lens" library, for example, is that it builds up chains of coercions compositionally. In such a case, even if we've eliminated the eta issue for a *single* coercion, we'd still have it across a chain of them? Meanwhile, a single "coerce" whose semantics were like unsafeCoerce (but only when it's safe!) would do the job just fine at any level. That said, I think the general direction of this proposal is great, and I hope we can work out the kinks and get it implemented. --Gershom