GHC hackers,
Finally! I have committed the new representation for coercions! This is a
pretty big patch touching a lot of files. If you are working on anything that
intersects the type checker, it'll affect you.
GHC now implements, rather closely, the system described in our paper
"Practical aspects of evidence-based compilation in System FC" here
http://research.microsoft.com/en-us/um/people/simonpj/papers/ext-f/
Treat this paper as the new main statement of the type system implemented in
GHC.
Key changes are these:
* There is a new data type, Coercion (defined in module types/Coercion),
completely
distinct from Type.Type. The Coercion module is the heart of all the changes.
* Equality evidence is handled *completely uniformly* with class evidence
and implicit-parameter evidence. All the beastly special cases for EqPred
have
gone away. In particular, the evidence for an equality is a *value* (not a
type-level
thingy), but it is a value that has zero width (like the State# token), and
hence
costs nothing to allocate or pass around. See the paper.
* A term can be a coercion, so there is a new constructor in CoreSyn.Expr that
lifts Coercions
into Exprs.
All this should make it trivial to implement equality superclasses, but I have
not quite taken that step yet.
Interface-file formats may have changed, so the safest thing is to recompile
everything. There are some minor wibble to haddock, so you'll need to pull
there too. No changes to the libraries though.
Many thanks to Brent Yorgey for doing the first pass of all this work, and to
Stephanie, Steve, and Dimitrios for lots of the design work.
Simon
_______________________________________________
Cvs-ghc mailing list
Cvs-ghc@haskell.org
http://www.haskell.org/mailman/listinfo/cvs-ghc
