-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 Ravi Nanavati wrote:

On 4/17/07, Neil Mitchell wrote:

...

Hi,

...

From Section 10 of the Haskell report, regarding automatic derivation:

to derive Bounded for a type: "the type must be either an enumeration (all constructors must be nullary) or have only one constructor."

This seems a very artificial restriction - since it allows you to be in any one of two camps, but no where in between. It also means that Either doesn't derive Bounded, while it could easily do so:

instance (Bounded a, Bounded b) => Bounded (Either a b) where minBound = Left minBound maxBound = Right maxBound

So I propose that this restriction be lifted, and that the obvious extension be given such that minBound is the lowest constructor with a pile of minBounds, and maxBound is the highest constructor with a pile of maxBound.

In general, I like the idea of of allowing more flexible derivation of Bounded, but I'm worried your specific proposal ends up mandating the derivation of Bounded instances for types that aren't really "bounded" (used in a deliberately loose sense). Consider the following type:

data Foo = A Char | B Integer | C Int

On some level, there's no real problem in creating a Bounded instance as follows (which is how I interpret your proposal):

instance Bounded Foo minBound = A (minBound :: Char) maxBound = C (maxBound :: Int)

On the other hand, there's a real sense in which the type isn't actually "bounded". For instance, if it was also an instance of Enum, enumerating all of the values from minBound to maxBound might not terminate. I'm not sure what to do about the scenario. Should we (unnecessarily) insist that all of the arguments of all of the constructors be Bounded to avoid this? Should Bounded more explicitly document what properties the minBound, maxBound and the type should satisfy? Or something else?

IMO, Bounded only needs to satisfy (if Foo is in Ord) forall a::Foo, a >= minBound && a <= maxBound . I want to be able to define bounded for data ExtendedInteger = NegativeInfinity | PlainInteger Integer | PositiveInfinity . Preferably by deriving, because it's easier. If we require properties of Enum... Enum _already_ has problems with instances like Integer where fromEnum :: a -> Int only has a limited possible output; there is little reasonable meaning for (fromEnum (100000000000000000000000000000000000000000000000000000 :: Integer)) (hugs: Program error: arithmetic overflow) (Float and Double *aren't* in Bounded. Then again, Haskell98 doesn't require them to contain non-_|_ values of +-infinity.) Furthermore, there are bounded things that aren't enumerable anyway (I think) (such as some lattices), so it would be odd to add that restriction just because the type might also be in Prelude.Enum. However there is a good argument for having some sort of bounded-enum class for things that have a finite number of discrete positions. These have log(number of possibilities) information content and can (in theory) be serialized with such a number of bits known from the type. Designing such a class could be interesting... Rather, I would ask "Must any inhabitant of a type in Enum be reachable by pred or succ from an arbitrary inhabitant of the type?" For example, I could declare an instance of Enum that contradicted that: data Something = Some Integer | Another Integer where pred and succ always stayed within the same constructor, and for fromEnum/toEnum I would just find some way to encode some common (i.e. relatively small magnitude, just as the usual instance Enum Integer is limited this way) values of Something into an Int. Or are fromEnum/toEnum supposed to obey some sort of properties, when they are defined, relative to the rest of the methods? I would guess not, given the comment - -- NOTE: these default methods only make sense for types - -- that map injectively into Int using fromEnum - -- and toEnum. (hugs: fromEnum (2.6 :: Double) ---> 2) Cheers, Isaac -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.6 (GNU/Linux) Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org iD8DBQFGJgjWHgcxvIWYTTURAlL/AJ97SilRhmd8B59TAAX+Hcyjly5oHQCff0fa 5B4Y9m0Zb3vQtilZr4lRQs0= =Qn2+ -----END PGP SIGNATURE-----

On 2007-04-18, Isaac Dupree wrote:

(Float and Double *aren't* in Bounded. Then again, Haskell98 doesn't require them to contain non-_|_ values of +-infinity.)

And they're only in Enum to support the [a..b] syntax, whose semantics can't really sanely be supported for Float and Double anyways.

Furthermore, there are bounded things that aren't enumerable anyway (I think) (such as some lattices), so it would be odd to add that restriction just because the type might also be in Prelude.Enum.

I'd really like to see one. Unless you're just talking about a set with a partial order, in which case, yes, many are bounded. e.g. reals in [0, 1], as when being used for probabilities.

Rather, I would ask "Must any inhabitant of a type in Enum be reachable by pred or succ from an arbitrary inhabitant of the type?"

That would make sense to me (when restricted to non-bottom inhabitants), and is essentially the objection that many have to Float and Double being in Enum.

For example, I could declare an instance of Enum that contradicted that: data Something = Some Integer | Another Integer where pred and succ always stayed within the same constructor, and for fromEnum/toEnum I would just find some way to encode some common (i.e. relatively small magnitude, just as the usual instance Enum Integer is limited this way) values of Something into an Int. Or are fromEnum/toEnum supposed to obey some sort of properties, when they are defined, relative to the rest of the methods? I would guess not, given the comment - -- NOTE: these default methods only make sense for types - -- that map injectively into Int using fromEnum - -- and toEnum. (hugs: fromEnum (2.6 :: Double) ---> 2)

Cheers, Isaac

The default implementation for the class assumes fromEnuw (succ x) = 1 + fromEnum x and similar things. That's a pretty strong argument that all types should obey that. -- Aaron Denney -><-