
On 03/03/12 01:40, Johan Tibell wrote:
Hi all,
These ideas are still in very early stages. I present them here in hope of starting a discussion. (We discussed this quite a bit at last year's ICFP, I hope this slightly different take on the problem might lead to new ideas.)
I think the next big step in Haskell performance is going to come from using better data representation in common types such as list, sets, and maps. Today these polymorphic data structures use both more memory and have more indirections than necessary, due to boxing of values. This boxing is due to the values being stored in fields of polymorphic type.
First idea: instead of rejecting unpack pragmas on polymorphic fields, have them require a class constraint on the field types. Example:
data UnboxPair a b = (Unbox a, Unbox b) => UP {-# UNPACK #-} !a {-# UNPACK #-} !b
I expect that this will not be easy to implement, because it requires interaction with things like the garbage collector. For example, UnboxPair will need a different info table for different a and b. It might be possible to essentially specialize UnboxPair for each different a and b for which it is used, but that gets tricky with generic code.
The Unbox type class would be similar in spirit to the class with the same name in the vector package, but be implemented internally by GHC. To a first approximation instances would only exist for fields that unpack to non-pointer types (e.g. Int.)
Second idea: Introduce a new pragma, that has similar effect on representations as DPH's [::] vector type. This new pragma does deep unpacking, allowing for more types to be instances of the Unbox type.
Could this be handled by just having/deriving an Unbox instance for (a,b)? I imagine the Unbox type class would have to contain essentially the same things as Storable, maybe something like type UnboxedRepr :: Int -> # class Unbox a where type Repr a :: # -- gives size and alignment unbox :: a -> Repr a box :: Repr a -> a A problem with an instance (Unboxed a, Unboxed b) => Unboxed (a,b) is that it allows arbitrarily large unboxed values to be created at runtime. That doesn't work when you use specialization to create the needed info tables. Twan