On 8/24/10 5:12 PM, C. McCann wrote:
The problem is that instance selection doesn't work the (obvious, seemingly-sensible) way you thought it did. In short, instance contexts are only examined after the fact; [...] The reason it works this way has to do with the nature of type classes being "open".
Not necessarily. You can implement instance resolution just fine by using a backtracking search. There's nothing wrong with verifying the constraints before committing to an instance that unifies the head; It's a perfectly sensible way to do things. Unfortunately this makes instance resolution undecidable, and makes it so that determining whether instances "overlap" is no longer a lexical property of the code. But, well, you get the power you pay for. Apparently we Haskellers don't like to pay for much since we use a greedy algorithm for instance resolution :) -- Live well, ~wren