Liberalising IncoherentInstances

[AntC]

> Axel Simon <Axel.Simon <at> in.tum.de> writes:

> ...
> 
> In recent years I acquired the view that 
> type inferences should have some sort of semantic completeness property 
> ...

If you want 'semantic completeness', don't use IncoherentInstances.

(As several have said, it's not a popular extension;
more of a 'necessary evil' to use with caution.
That's why it's a good suggestion to limit it's 'reach'
to particular instances.)

The price to pay for avoiding IncoherentInstances might be:
- adding extra instances to avoid partial overlap
- adding type annotations
- 'helper' classes to resolve type-ambivalent instances
  (see Oleg's work for examples of these)
- fancy uses of FunDeps
- ultimately, programs that fail to compile

> ... With "semantic completeness" I mean "the best possible type
> that that can be expressed by the type language". ...

If there is a (unique) 'best possible type',
then you don't need IncoherentInstances, IMO.


AntC