Duncan Coutts wrote:
On Wed, 2005-05-04 at 18:29 -0400, Jacques Carette wrote:
There is also Template Haskell vs MetaOCaml.
For the life of me, I still cannot fathom why Template Haskell is untyped, while MetaOCaml is fully typed. Which is the main reason I write meta-program in MetaOCaml and 'other' programs in Haskell.
The reason for this is that Template Haskell is more powerful and correspondingly harder to type so it is currently untyped.
``more powerful'' meaning that it can do some level of introspection, right? I know introspection can be very powerful (I designed Maple's modern reflection and reification facilities...), but I was under the impression that type systems that can handle 'too much' introspection were unsound [a result of Walid Taha].
The main reason it is more porwerful is that TH allows you to pattern match on the abstract syntax of a quoted expression and perform arbitrary transformations on that. It would be hard to preserve the well-typedness of the AST and still allow arbitrary transformations.
Can't this be regarded as a 'convenience' rather than as actual extra power? Using an abstract interpretation formalism (see papers of Taha and co-authors), it is possible to 'lift' the type of code values into staging-time terms, and pattern-match on that instead. Then well-typedness of transformations is much easier to show. In this way one can even implement some transformations that are equivalent to what are currently 'hints' to GHC as well-typed code [I have some code that does this, but there are only hints of this in a preprint of mine -- see the conclusion section of "*Gaussian Elimination: a case study in efficient genericity with MetaOCaml"* http://www.cas.mcmaster.ca/%7Ecarette/publications/ge.pdfavailable from http://www.cas.mcmaster.ca/~carette/publications.html for some hints on how this is done
A colleague of mine is currently developing a type system for Template Haskell that should catch most errors without restricting too much the programs you can write.
I am eagerly awaiting this! Writing Monadic meta-programs in MetaOCaml is feasible, but the lack of native support for Monads makes it significantly harder. But after spending 12 years writing programs in a fully dynamically typed language (Maple), I am much happier writing *all *parts of my programs in a strongly typed language! Jacques