[Hs-Generics] A Lightweight Implementation of Generics and Dynamics (LIGD)

[Bruno Oliveira]

Hello all,

Johan mentioned in his last email that, starting from October 2nd (yesterday), we could
begin to discuss the different approaches to Generic Programming (GP). 
Since I have not seen any email concerning this I'll open the hostilities, which will 
hopefully sparkle some discussion. 

I provide my summary for LIGD in the end of this email, but before that I would like to 
discuss a few things:

1) While discussing the different approaches, we should also develop a "template"
for reporting our conclusions. I think that having a base template will make our life 
easier. At the moment I am using the following:

- Approach (Just the name of the approach)
- Required features (What do we need)
- Usage (how to use it, what different users do we have and what knowledge do they need)
- Extensibility (is it extensible, if not can we do anything about it?)
- Performance considerations (what are the performance impacts, is it possible to improve?)
- Expressibility (what kind of generic functions can we write? producer and consumers? different 
                       arities? local redefinition?)
- Boilerplate (what is the boilerplate involved, is it possible to remove it?)
- Helpful extra features (what could Haskell have that would make the approach easier to use?
                                     Are those features realistically feasible?)
- Discussion (General discussion, advantages/disadvantages, comparasion with other 
                     approaches)

but note that this is not, by any means, supposed to be a proposal for the final template.
It is just my initial template. We probably should discuss the template as we go along 
and tune it (if you think it is a good idea to have it in the first place).

2) Johan also mentioned a darcs/svn repository for the resulting paper/document. I think this 
is a great idea. Is anyone taking care of this?
It would also be a good idea to use the repository to store some code for the different 
approaches. This way people who would like to experiment with code, would not have to repeat 
the task individually. What do you think?

3) Related to 2), we could start by collecting code for the approaches from the authors 
of the papers. I guess most of them are in this mailing list anyway. For example, we could 
now ask Ralf Hinze or James Cheney (either one of you two reading this email ?) if they some code 
available for LIGD that they could submit. We could then add this code to darcs/svn and do some 
experiments with it: organize it like a library, define generic functions, etc. If no code is available, 
perhaps we should find a volunteer that implements the code (if that's the case for LIGD I can 
volunteer for it).

Here is my summary of LIGD, which is not complete but I think it can be a starting point. 

=======================================================================
Approach: A Lightweight Implementation of Generics and Dynamics

Required features: 
   - Haskell 98 + existential types

Usage:
   - Library Writer: Responsible for the sums-of-products machinery,
     the representation(s) data type(s) to be defined in the library and 
     defining the class and standard instances for "Representable". 
     The library writer needs to be have a deep knowledge about generic 
     programming.
   - Power User: Defines the boilerplate for new data types. Needs to be fairly 
     familiar with the implementation.
   - User: defines generic functions by pattern matching on the structure of "Rep".
     This requires that he is comfortable with sums of products and isomorphisms.
     For convenience, the user should also provide a definition that uses the 
     type class "Representable", which avoids the representations to be passed
     explicitly.
   - End User: If "Representable" is used, then the end can just use generic functions 
     in the same way as any other ad-hoc (that is, type class overloaded) functions.

Extensibility: 
   - Data types are non-extensible, however we can try to combine the 
     approach with type classes to get some form of extensibility (RepLib 
     proposes one solution like this). 

Performance considerations: 
   - Sums of products based and therefore involving considerable 
     performance penalties.

Expressibility:
   - Can do both producer and consumer functions.
   - Multiple representations needed for generic functions of different 
     arities;
   - No local redefinitions.
   
Boilerplate:
   - Isomorphisms between sums of products and data types
   - Instances of Representable
   - Smart constructors for representation of data types

Helpful extra features:
   - GADTs (already implemented in GHC) allow more direct definitions 
     of generic functions, since there is no need to apply the type 
     coercions explicitly;
   - A boilerplate generation mechanism. This would effectively mean 
     that power users would not be necessary.
   - Open data types and functions in order for extensibility.
   - Kind polymorphism, for eliminating the need for multiple 
     representations;

Discussion: 
    An interesting approach which is quite lightweight and that is fairly easy to 
use (specially if we would not need to  handle any of the boilerplate). It is a bit 
outdated because with GADTs, the use can be slightly simplified. However, this also 
takes us further away from Haskell 98 or even Haskell' (since GADTs will not be there).
I think that LIGD is still a reference for a generic programming library that uses 
a data type for type representations and sums of products as a family of data types. 
   A drawback may be performance, since we need to convert a data type into its 
isomorphic sum of products representation. 
=============================================================

Cheers, 

Bruno