On 4/5/11 11:22 PM, Ivan Lazar Miljenovic wrote:

On 6 April 2011 13:13, wren ng thornton wrote:

...

On 4/4/11 4:42 PM, Jason Dagit wrote:

...

Is this something people would be interested in having as an extension in GHC? Or is it just too fluffy for anyone to really care?

I'd much rather have _ on the RHS of equations be a way of specifying terms that the compiler should infer. This is pretty standard for dependently typed languages, and more closely mimics the behavior of _ on the LHS as a gensym-named pattern variable.

Can you provide an example of this behaviour?

Which behavior, the LHS or RHS? A canonical example for inferring RHS terms is handling the passing of implicit arguments, e.g. the types passed in System F. Haskell already does this inference for type passing--- though Haskell lacks the explicit form that System F has for doing so. Another example is the inferences for passing around type-class dictionaries. In Haskell-as-specified this isn't how it works, but it would be easy to imagine a variant of Haskell where type-class constraints are just there for inferring the implicit dictionary argument--- which would allow us to unify the implementations of functions which take a parameter (e.g., an equivalence predicate) and those which take a type-class from which they extract the parameter (e.g., taking an Eq dictionary in order to use (==)). Languages like Coq, Agda, Epigram, etc combine these ideas about implicit argument passing and inference into a single system and extend it to allow inferring all terms, not just types and type-class dictionaries.

...

For Haskell we could perhaps use Djinn to infer the term and have compilation fail if there isn't a unique total function/value that can be inferred for the missing term.

Doesn't Djinn also return a function in cases where there is more than one possible value?

I was just pointing out Djinn as prior art. In principle it shouldn't be too hard to adjust it to complain when there are multiple solutions. Of course, we'd have to define "multiple solutions" in order to avoid counting multiple terms which all normalize to the same thing; which is undecidable in general since Haskell allows non-terminating functions. But, in practice, we only need a semi-decidable inference mechanism; it's fine for the compiler to give up and force the programmer to supply the term. -- Live well, ~wren