On Wed, Jun 22, 2011 at 09:36, <oleg@okmij.org> wrote:

> the need to define a duplicate of the class (MonadState' in your
> example) bloats the code significantly.

I'm quite puzzled at the statement. Is there really significant
bloat? Let us count.

As the first example, let's take
class C a
with N special instances, mutually non-overlapping instances, and one
catch-all instance "instance C a". We have N+1 instances total.

With the technique in the earlier message, we define an auxiliary
class C' a flag
with N special and one catch-all instance, all non-overlapping. We
then define one instance of the original class C, performing the
dispatch. The additional cost: one class declaration and one
instance. Is one extra class and one instance considered a bloat?
The extra class is systematically derived from the original one
(so, one could get TH to do the deriving).

My problem is exactly with this unnecessary repetition. If the extra class can be systematically derived from the original code, then it shouldn't have to be specified. All I'm saying is that if we would want to have such a mechanism for replacing OverlappingInstances, then we should also come up with a way (possibly even involving new syntax) to make it intuitive to use.

(We've done something similar for the new generics stuff, by introducing DefaultSignatures.)

> But I am still doubtful of the runtime performance of your code

I believe discussing performance may be a bit premature;

Perhaps, but this was prompted by your suggestion of considering deprecating OverlappingInstances. I have lots of code that relies on it, and I wouldn't want it to start triggering warnings just yet :-)

Cheers,
Pedro