GHC/Stand-alone deriving declarations
Bjorn Bringert has recently implemented "stand-alone deriving" declarations, documented briefly here [1]. There are a few loose ends which I summarise here:
Syntax
The current syntax is
deriving Show for T
There seems to be a consensus that this would be better:
derive instance Show T
so that it looks more like a regular instance declaration. Here derive
is not a new keyword; it's a "special-id", distinguished by the following instance
keyword. That means that derive
can still be used freely as a regular varid.
(Perhaps derived
would sound more declarative than derive
.)
Context on the declaration
Because it looks like a regular instance declaration, it would arguably be reasonable to require the programmer to supply the context. It seems odd to say:
derive instance Show (T a)
and perhaps cleaner to say
derive instance Show a => Show (T a)
(At the moment, the compiler figures out the appropriate context, but at some point that automation may run out of steam.)
Alternative Syntax
Alternatively the syntax could be:
instance deriving Show a => Show (T a)
with the advantage that we don't need a complicated rule to determine whether or not derive
is an identifier or a keyword (all other reserved ids always occur inside the construct not before it).
There was a suggestion to allow just
instance Show a => Show (T a)
where the fact that the compiler is automatically deriving the content of the instance declaration is implicit in the absence of a where
clause. The rationale was that the
where
clause is already optional - if a class does not have a body then we can instead say that all instances of it are derived. However this is to misunderstand the current meaning of a missing where
clause. In Haskell'98 it means that all the methods of the class are undefined for this instance (except for defaults). That is the opposite meaning to this proposal, which would say that all methods are fully defined.
Interaction with "newtype-deriving"
GHC's "newtype deriving mechanism" (see [2]) should obviously work in a standalone deriving setting too. But perhaps it can be generalised a little. Currently you can only say
deriving C a for Foo
(where Foo is the newtype), and get an instance for (C a Foo)
. But what if you want and instance for C Foo a
, where the new type is not the last parameter. You can't do that at the moment. However, even with the new instance-like syntax, it's not clear to me how to signal the type to be derived. Consider
newtype Foo = F Int
newtype Bar = B Bool
derive instance C Foo Bar
Which of these thee instances do we want?
instance C Foo Bool => C Foo Bar
instance C Int Bar => C Foo Bar
instance C Int Bool => C Foo Bar
The obvious way to signal this is to give the instance context (just as above). This is perhaps another reason for having an explicit instance context in a standalone deriving declaration.
Incidentally, notice that the third of the alternatives in the previous bullet unwraps two newtypes simultaneously. John Meacham suggested this example:
class SetLike m k where
instance SetLike IntSet Int where
newtype Id = Id Int
newtype IdSet = IdSet IntSet
derive instance SetLike IntSet Int => SetLike IdSet Id
Duplicate instances
Suppose two modules, M1 and M2 both contain an identical standalone deriving declaration
derive Show T
Then, can you import M1 and M2 into another module X and use show on values of type T, or will you get an overlapping instance error? Since both instances are derived in the very same way, their code must be identical, so arguably we can choose either. (There is some duplicated code of course.)
This situation is expected to be common, as the main use of the standalone feature is to obtain derived instances that were omitted when the data type was defined.
But, that means whether or not an instance was derived is now part of the module's. Programs would be able to use this (mis)feature to perform a compile-time check and execute code differently depending on whether any given instance is derived or hand-coded:
module MA(A) where
data A = A deriving Show
module MB(B) where
data B = B deriving Show
module MC where
import MA
import MB
-- verify that the A and B Show instances were derived
-- (they need to be derived to ensure the output can
-- be parsed in our non-Haskell code).
derive instance Show A
derive instance Show B
The writer of MC already knows that MA and MB defined instances of Show for A and B. He just wants to ensure that nobody changes either module to use a non-derived instance; if someone does try to use a non-derived instance:
module MA(A) where
data A = A
instance Show A where
show _ = "a"
then they will get an overlapping instance error in MC.
The result is that programs would be able to require, for any Class, not just that an instance of the class was defined for a type, but that a /derived/ instance was defined. Is this good?