Hi David,
2011/6/21
At Tue, 21 Jun 2011 10:01:24 +0200, José Pedro Magalhães wrote:
| One thing you could do to help in this specific case would be to
use a
| different M1 tag--e.g., M1 S ... for selectors and M1 NS ... for | fields without selectors (or K1 NS). I presume you've already | considered this and/or it's too late to make such a change. (Or to | move the distinction up to the constructor with two different | constructor tags, CR and CN for record and no-record.)
I don't think it's too late to make a change. The stuff has only
just
gone in, so it's still very malleable. There may be other
considerations,
but legacy code isn't one of them!
I suppose that could be changed, yes, but what exactly are we trying to
solve
here? One can already specify different behavior for constructors with/without named fields. Are we trying to avoid OverlappingInstances? Then yes, this might help, but I'm not sure this change alone would make all generic programming possible without OverlappingInstances.
Sorry, I wasn't necessarily advocating the change. This came out of a long discussion of UndecidableInstances. Oleg's TYPEOF approach is one way to avoid them (or at least to avoid them in conjunction with OverlappingInstances). Before Oleg posted his examples, Simon asked if it might be possible to use the new Generic class in a similar way. I was pointing out that no you can't, though the specific example I gave could be accommodated by a small change to the Generic interface.
Ah, I see. Oleg's TYPEOF encodes only the type arguments to a datatype, whereas Rep encodes the structure of the datatype. As far as I can tell, TYPEOF never depends on the right-hand side of the datatype. A simple example is some datatype with a phantom type, say `data P a = P`. TYPEOF will encode the `a`, whereas Rep will not. So now it is very clear to me that you cannot get the TYPEOF if you have the Rep, in general. Cheers, Pedro
David