The type class implementaion in jhc is based on pattern matching on GADTs rather than dictionary passing. This is an attempt to document how this works theoreticaly and practically as well as discuss the advantages/disadvantages of implementing classes this way rather than via dictionary passing. Another related feature in jhc is auto-reification of types to values and that is also discussed. Though some of this is new (the use of GADTs to implement Haskell-compatable type classes and the optimizations that specifically take advantage of it) a whole lot of it is just re-hashing of Tim Sheard and SPJ's work and discussions about the future of GHC. (writing about things helps me think about them). It is also serving as a sort of gameplan for jhc development.
I have not tried to document my type class system before because when I implemented this system initially, I had not heard of GADTs and independently implemented them as many have done before. So, in addition to "first-class phantom types", "guarded recursive data type", and "generalized algebraic data types", you can add "that type class case typing hack" or TTCCTHs to the list of things they have been called. Before I heard about GADTs I had not done much formal work proving soundness of my extensions so it was a great relief to find out not only other people have done the work, but there is a rich lexicon of terms as well as a concrete syntax implemented in GHC available to express them. It made me feel like a real theorist to be able to change comments like
-- XXX hack to get class method case branches to typecheck properly
-- perform GADT based type refinement
I will use GHC syntax rather than jhc internal syntax throughout.
The jhc intermediate language is based on the lambda-cube rather than system F. what this means is that types are values and are bound by lambda expressions and normal application just like any other value. The upshot of this is that it allows you to perform 'case' matching on _types_ since the bound types scope over the entire body of the lambda at the term and type level.
What this effectively means is that there is an inherent extensible GADT defined that reifies all types which we can think of as being defined as below: (and is perfectly implementable in GHC-haskell too)
data Type a where Prelude.Int :: Type Prelude.Int Prelude.Char :: Type Prelude.Char IO.IO :: Type a -> Type (IO.IO a) (:->) :: Type a -> Type b -> Type (a -> b) ....
Note that while in ghc this declaration will create data constructors which mirror the names of the type constructors, in jhc you pattern match and create values with the type constructors themselves. so the same a that is passed into a function can both be pattern matched on and used in type signatures to type things properly.
Implementing classes in this fashion opens up a number of interesting optimization opprotunites.
rather than dictionaries of higher order functions, a simple algebraic datatype is passed. For each type variable in a function only a single value is passed, no matter how many class constraints it occurs in.
f :: (Ord a, Num a, MonadIO m, StateMonad a m) => a -> m a ==> f :: Type a -> Type m -> a -> m a
whereas a traditional implementaion would require 4 extra arguments each containing as many values as there are methods in the class.
case statements are also much faster than indirect function calls in many cases on modern CPUs so time is saved there too.
case coalescing becomes a very powerful optimization in this framework. imagine we had the following.
class Num a where a + b :: a -> a -> a a - b :: a -> a -> a a `div` b :: a -> a -> a class Ord a where a <= b :: a -> a -> Bool instance Num Int where (I# a) + (I# b) = a +# b -- unboxed, very fast. (I# a) - (I# b) = a -# b -- unboxed, very fast. (I# a) `div` (I# b) = div# a b -- unboxed, very fast. instance Ord Int where (I# a) <= (I# b) = a <=# b -- unboxed, very fast. instance Num Integer where a + b = addIntegers a b -- slow a - b = subtractIntegars a b a `div` b = divIntegers a b instance Ord Integer where a <= b = lteInteger a b
now take the following function
f :: (Ord a, Num a) => a -> a -> Bool f x y = (x + y) <= (x - y)
now, first desugaring goes to
f a x y = case a of Prelude.Int -> case (case a of Prelude.Int -> case (x,y) of (I# x',I# y') -> I# (x' +# y') Prelude.Integer -> ... ) of (I# xpy') -> xpy' <=# ... Prelude.Integer -> ....
but you get the idea, now if we apply the very standard transformations described in the ghc unboxing paper, along with alternative elimination based on type refinement we end up with the following.
f :: (Ord a, Num a) => a -> a -> Bool f a x y = case a of Prelude.Int -> case (x,y) of (I# x',I# y') -> (x' +# y') <=# (x' -# y') Prelude.Integer -> addIntegers x y `lteInteger` subtractIntegers x y .....
notice we have done gotten fully unboxed versions of the routines without ever losing polymorphism! f is still polymorphic in a yet is able to take advantage of type dependent optimizations.
the key thing that allows this is that this class implementation captures the invarient that figuring out any particular method on a type statically determines all the methods from all classes on that type.
Of course, willy-nilly inlining of classes like this can lead to code explosion so there are many different pragmas in jhc to carefully control its behavior.
one interesting possibility is run-time specialization. traditionally you can only call a specialized version of a routine if you can determine its type at compile-time, but when a function is declared SUPERSPECIALIZED it does a single case scrutination of the type immediatly and then calls the appropriate specialized routine. you can SUPERSPECIALIZE at just certain types letting the rest fall through to the generic version, superspecializing complicated math routines with 'Int' can be very advantagous.
an interesting side effect is that fundeps are not only useful for refining classes at compile time, but they have signifigant run-time benefits.
class StateMonad m a | m -> a where ... f :: (... , StateMonad m a, ...) => ...
now, a single scrutination of 'm' will not only tell you statically method of every class on 'm', but also every method of every class on 'a' too.
The first obvious issue is what happens to all these types at run-time? The answer is that they are treated just like any other data type so actually may be passed at runtime and in fact grin code makes no distinction between type and data constructors anywhere.
However, the first thing done after conversion to Grin from E is a standard dead-code elimination pass getting rid of arguments to functions that arn't used by either being passed to a primitive or pattern matched on. (calculated via fixpoint iteration). the key result of this is that:
All types end up being eliminated except in the exact places where a dictionary would have been passed in a traditional implementation.
In addition, any unused normal arguments to functions are eliminated too. there is no need for a special type-erasure pass, a standard optimization takes care of it. And since types are only erased if they are not used, one is free to experiment with other methods of using run-time type information without modifying the back end.
A naive implementaion would end up creating huge case statements for each method. which is exactly what a naive implementation of the 'eval' operation in GRIN would do. Fortunately the exact same solution works for both. after the E has been converted to GRIN and simplified, it will still have free references to 'evals' (evaluate a thunk) and class methods. A points to analysis is then performed which annotates every such call with all the possible values of its argument. A specialized version of the routine is then instantiated which contains only those cases that might possibly be passed to that point.
It should be noted that this case actually rarely comes up because optimizations at the lambda-cube level can often narrow down the types something will be called with and specialization will take place there. there are also a lot of PRAGMAS with which you can fine tune how aggresively class methods are inlined at their use sites. Partial inlining is possible too, where you just consider the instances in scope and
While this technique is not ammenable to traditional C style separate compilation, it does allow 'semi-separate compilation' which I hope to support as a faster alternative to whole program analysis in jhc. What this means is that each module is compiled seperately into its own '.o' and '.hi' file as in ghc, jhc is run once more right before linking and it goes through collecting the addresses of all eval-points and class methods and outputs generic eval and method functions to be used when optimizations couldn't create specialized ones locally. Strong local points-to analysis would be a key optimization in this case.
another interesting property is that since we don't have to decide on a dictionary at a functions use point, we can propagate class predicates such as 'Num Char' outwards rather than rejecting them when they appear, and then only reject them if such an instance is never defined. I am not sure how useful this would be though.
Although at the moment (Type a) is not exported in jhc. this will change as soon as my types in the front end become wobbly enough. This section deals with Haskell level programming assuming a suitable (Type a) defined as above is made available to the user. Many others have made these observations before, but they are interesting nonetheless. Most of this is just rehashing Sheards work with the Omega varient of Haskell.
While defining an actual GADT like above would be perfectly fine, and could even be done by a preprocessor with full-program analysis capabilities right now for ghc, it would be nice to have some syntatic sugar since you ideally want to interpret type names in the type namespace.
the special syntax I am thinking about for jhc is:
f :: * -> String -> Int f = ... foo = f (Int :: *) "hello"
where * is just syntactic sugar for (Type a) and (Int :: *) is just sugar for the constructor (Prelude.Int) defined above (or whatever type name 'Int' is in scope). since they are just simple rewriting rules that can be carried out before type checking, they don't affect the type checking algorithms at all.
A special syntax in jhc also has the advantage that the compiler need not actually create the data-type even when users write their own case statements on types since it can know to translate them into appropirate matching on the type constructors directly internally.
it is very tempting to allow things like
f (x :: a) y = g (a :: *) y
where the a is taken from the scoped type variable, but this goes beyond simple syntatic sugar and have to think about it more.
TypeReps in Data.Typeable have a particularly compelling implementation:
newtype TypeRep = exists a . TypeRep (Type a)
which is precicely what we want, a TypeRep is a (Type a) for some value of 'a'. we can recover said value by pattern matching on (Type a) and that is how the Eq and Ord instances could be defined.
Another interesting thing is that Data.Dynamic can be implemented without unsafeCoerce
data Dynamic = exists e . Dynamic (Type e) e fromDyn :: forall a . Dynamic -> Type a -> Maybe a fromDyn (Dynamic e x) a = case a of Prelude.Int -> case e of Prelude.Int -> Just x -- note, both types have been refined to Prelude.Int at this point _ -> Nothing IO y -> case e of IO ye -> fromDyn (Dynamic ye x) y _ -> Nothing ....
I am still working out the details, but every type in jhc will automatically get a (Typeable) instance. however, I will probably keep the Typeable class around since the predicates the type checker collects of the form Typeable a => ... can be useful hints in guiding optimizations. (but are not necessary)
when combined with user created kinds (a planned feature), the auto-reification feature of jhc becomes more interesting.
kind Nat = Z | S Nat
would define the following.
the new kind 'Nat'
the type constructors:
data Z data S (a :: Nat)
and the reified version of said types
data Nat (a :: Nat) where Z :: Nat Z S :: (Nat a) -> Nat (S Nat)
note that the type 'Nat a' is effectivly a closed version of 'Type a' which suggests 'Type' might better be called '*' or 'Star'
The jhc type system is impredicative using the boxy types algorithm of SPJ. so things like (Type (forall a . a -> a)) are allowed. though, I have not explored the implications of this. Although the idea of writing class instances for polymorphic values is intriging, I can't think of any particular use.