On Thu, May 11, 2006 at 01:46:43PM +0100, Ben Rudiak-Gould wrote:

Otakar Smrz wrote:

...

data ... = ... | forall b . FMap (b -> a) (Mapper s b)

... where FMap qf qc = stripFMap f q

the GHC compiler as well as GHCi (6.4.2 and earlier) issue an error

My brain just exploded. I can't handle pattern bindings for existentially-quantified constructors.

The problem here is a tricky interaction between irrefutable pattern matching and existential tuples. In Core, the FMap constructor has a third field which stores the type b, and when you match against the pattern (FMap qf qc) you're really matching against (FMap b qf qc). (stripFMap f q) might evaluate to _|_, in which case, according to the rules for irrefutable matching, all of the pattern variables have to be bound to _|_. But type variables (like b) can't be bound to _|_.

From an operational standpoint, the problem is that the (fully-evaluated) value of b has to be available in the body of the let statement, which means that (stripFMap f q) must be evaluated before the body, and the let statement must diverge without reaching the body if (stripFMap f q) diverges, since no value can be assigned to b in that case. But the semantics of let clearly require that execution always proceed to the body no matter what (stripFMap f q) evaluates to.

So I'm not convinced that your program is well-typed, even though it looks fine at first. I'm not sure what happens to Haskell's semantics when you allow type variables to be bound to _|_. The fact that Hugs allows it may be a bug.

Why would a type variable be present at runtime, let alone bound to _|_? I believe the Hugs behaviour was intentional. It's particularly handy with single-constructor data types, e.g. representing objects. It does complicate the formal specification of pattern matching a bit, but I don't think it's unsound in any way.