On 10 Jan 2008, at 6:04 AM, Nicholls, Mark wrote:
-----Original Message----- From: Bulat Ziganshin [mailto:bulat.ziganshin@gmail.com] Sent: 10 January 2008 13:36 To: Nicholls, Mark Cc: Luke Palmer; haskell-cafe@haskell.org Subject: Re[2]: [Haskell-cafe] confusion about 'instance'....
Hello Mark,
Thursday, January 10, 2008, 4:25:20 PM, you wrote:
"instance Num a =>> A a"
Mean the same thing as
"instance A (forall a.Num a=>a)"
programmers going from OOP world always forget that classes in Haskell doesn't the same as classes in C++. *implementation* of this instance require to pass dictionary of Num class along with type. now imagine the following code:
My confusion is not between OO classes and Haskell classes, but exactly are the members of a Haskell type class...I'd naively believed them to be types (like it says on the packet!)...but now I'm not so sure.
A type class *is* a set of types. But, in Haskell, types like (forall a. Num a => a) aren't quite first-class feeling. A typical example of an expression of this type might be (3 + 5), but if I say x :: Double x = 3 + 5 the compiler won't complain. Furthermore, if the compiler sees instance A Double where somewhere in the code, when it sees foo (3 + 5), for some method foo of the class, it may decide to take (3 + 5) :: Double, not (3 + 5) :: forall a. Num a => a. In that case, you'll get the wrong methods called: class A a where foo :: a -> String instance A Double where foo x = "Double" instance A (forall a. Num a => a) where foo x = "number" If the compiler sees the first instance but not the second, then it will think that foo (3 + 5) = "Double". Adding the second will give foo (3 + 5) = "number". Haskell 98's rules for type classes are chosen so that legal code never changes its meaning when you add an instance (well, this is a bad example --- but the general point is sound). GHC relaxes these rules in quite a few cases, but in this one it's easy enough (in GHC) to get a type isomorphic to forall a. Num a => a that can be an instance of a type class that GHC hasn't bothered relaxing this particular rule. (And paying the subsequent cost in confusion when working code bitrots because somebody added an instance somewhere). jcc