I still think that getting other authors to use it would be the
biggest difficulty. Another concern of mine is that RULEs-based
fusion can be fragile; if the type classes prevent fusion from
occurring you'll never approach the performance of monomorphic code.
That said, I think this is worth pursuing further because of the
benefits you describe. I have spent a great deal of time on exactly
this issue with the iteratee package, although my needs there are
relatively simple.
As a general question to the Haskell community: have you ever
attempted to write container-polymorphic code? I'd like to hear about
either successes or stumbling blocks.
On Wed, Mar 24, 2010 at 12:02 PM, Alberto G. Corona
Once we have a tree of type classes suitable for all containers, as you said, theoretically it shouldn't very difficult to incorporate the testing of different instances for each class used in a program, besides testing different compilation flags in a genetic algoritm. This latter has already been done. http://donsbot.wordpress.com/2010/03/01/evolving-faster-haskell-programs-now... To find automatically the best combination of class implementations and compilation flags in a single process would be very useful and would save a lot of manual testing.
2010/3/24 John Lato
Hi Alberto,
To some extent this already exists, it's just that nobody uses it. I believe it's the approach taken by the Edison libraries. Also the ListLike package provides the type classes ListLike, StringLike, and a few others. Neither seems to have become very popular despite having well-respected authors (Okasaki and Goerzon, respectively).
Some container functions are already provided by other classes, namely Foldable, Traversable, and Monoid.
The first bit, creating a tree of type classes suitable for all containers, is probably a few hours work.
An automated system to determine the best implementation is significantly more difficult; I can't say if the scope would be appropriate for SoC.
Best, John
From: "Alberto G. Corona "
Just a dream: -separate interface and implementation for all containers, via type classes -develop, by genetic programming techniques + quickcheck, a system that find the best container implementation for a particular program.
Is that suitable for a Google Summer of Code project?
2010/3/23 Alberto G. Corona
The question can be generalized via type classes: Is there any common set of
primitives encapsulated into a single type class that has instances for Strings (Data.List) ByteStrings, Data.Text, Lazy bytestrings, Arrays, vectors and wathever container that can store an boxed, unboxed, packed unpacked sequence of wathever including chars? All of them have folds, heads, tails and a lot of common functions with the same name, but since there is not a single type class, the library programmer can not abstract his code when it is possible, so the library user can chose the particular instance for his particular problem.