Get rid of CheckedType::UnknownType

[flip1995]

otherwise, to get exhaustive enum matches, we need to take care of this, wherever we match on CheckedType

See #141 (comment)

[joshuabach]

If I understood @hediet correctly, UnknownType is constructed if an undefined class is used (an error is obviously also generated). This needs to be represented somehow because we want to continue checking even in case of this error. I don't think we can easily get rid of this. This case can be handled as unreachable!() or sth similar in the IR generation, since that is only called if no error occured

[hediet]

Better having an explicit UnknownType than having TypeRefs which point to nothing. Either way, you need to panic if something is wrong (calling expect on the result of a type lookup or asserting unreachable for UnknownTypes). I strongly prefer the UnknownType variant.

[flip1995]

While looking through the code I couldn't figure out a way to get rid of this. A TypeRef pointing to nothing is indeed not a good solution. But having to return None or panic isn't either. The later is for now better, but still not optimal.

[hediet]

What would an "optimal" solution look like? As long as we construct CheckedTypes for possibly invalid types, we need to consider that these CheckedTypes actually may represent invalid types.

[flip1995]

Removing the UnknownType. Maybe replacing it with an Option in checked_type_from_basic_ty. I don't really know. But this is low prio, since we can always return None in the firm generation when matching CheckedType (currently we even have to do this for CheckedType::Void anyway, since the void type is forbidden everywhere except return types).

[reiner-dolp]

In my opinion there should be just a second typesystem struct that is "into()"ed at the end of the semantic analysis phase that has all this impossible states removed.

[hediet]

In my opinion there should also be a second ast for that.

[reiner-dolp]

You are actually right. Most of the impossible states are in the ast. @hediet what are you thinking of exactly? A reshaped ast with typesystem information on each node?

[hediet]

@reiner-dolp yes, exactly.