There are really four cases to consider for deriving purposes.

Some type T occurs positively, negative, not at all, or in both.

* In the case it doesn't occur at all or occurs only positively, er... we don't care. We're not case analyzing it =)

* In the case of negative position only, the current behavior is more defined than the stricter behavior Ryan proposes. I personally very much favor keeping Eq, Ord, etc. as defined as possible for Void, V1 and the like with the existing behavior. 'a' occurs in negative position only for these cases.

* In the case of it occurring in both positive and negative position we have the option to 'pass it through' to exchange the bottom. In the case of changing the behavior of Functor and the like we're not actually changing the definedness of the result, we're merely choosing between "equivalent" bottoms. This is a rather exceptional case, but the ability to preserve placed bottoms in a structure in case they have meaning to the user seems quite valuable.

Similarly the existing practice of not doing wasted work and producing more defined results, also seems valuable. Aiming for "consistency" here seems to be pursuing a goal that doesn't actually help anyone and just makes stuff less defined.

-Edward

On Mon, Jan 16, 2017 at 9:25 AM, Ryan Scott <ryan.gl.scott@gmail.com> wrote:

To be clear, if you have:

data V a deriving Functor

David is proposing in #13117 [1] that the derived instance use
-XEmptyCase instead of the current error-based implementation:

instance Functor V where
fmap _ x = case x of {}

This seems fine and well, but given that #13117 is basically a
continuation of another discussion in #7401 [2], I feel like if we're
going to tackle the issue of derived instances for empty data
declaration, we should solve it for all stock classes.

In particular, there are two thorny issues to consider here. What if
we have this instead?

data V2 deriving Eq

What instance should this produce? Reid Barton proposed in #10577 [3]
that it should be:

instance Eq V2 where
a == _ = case a of {}

But we have a choice here, since (==) has multiple arguments! The
definition could also conceivably be:

instance Eq V2 where
_ == b = case b of {}

Is there a uniform policy we can decide for functions with multiple
arguments like this? In fmap, it's unambiguous since there's only one
argument of type f a.

Another issue to consider is that if we adopted this convention for
derived Eq instances for empty datatypes, we'd actually be going
against the convention set for Data.Void. As noted in [4], this is the
current Eq instance for Void:

instance Eq Void where
_ == _ = True

I'm not proposing that we change this definition, since there are many
good reasons to have it this way (see the thread in [5] for Edward
Kmett's convincing argument in favor of the current Eq Void instance).
Rather, I'm asking if we would be OK with having this
discrepancy--that is, deriving Eq for your own Void2 type would
produce a different instance. Personally, I'd be fine with it, but I
think we should ask for the community's input as well.

So sorry to hijack this thread, David, but we really should answer
these two questions as well:

1. What do we do in ambiguous cases like derived (==) implementations
for empty datatypes?
2. Are we fine with derived instances for empty datatypes sometimes
being different than the corresponding instances for Data.Void?

Ryan S.
-----
[1] https://ghc.haskell.org/trac/ghc/ticket/13117
[2] https://ghc.haskell.org/trac/ghc/ticket/7401#comment:46
[3] https://ghc.haskell.org/trac/ghc/ticket/10577
[4] https://mail.haskell.org/pipermail/libraries/2015-July/025959.html
[5] https://mail.haskell.org/pipermail/libraries/2015-July/025965.html

_______________________________________________
Libraries mailing list
Libraries@haskell.org
http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries