Rewrite a.type & B to Nothing

[odersky]

provided !(a.type <: B). Fixes #864, which turned out
to be a long sequence of comparisons between intersections
of singleton types.

Review by @smarter.

[smarter]

!(tp2 <:< tp2) did you mean !(tp2 <:< tp1) ?

[smarter]

(Also we usually write @pre)

[DarkDimius]

Should we also add such rule:
If tp1 is class and tp2 is class and !((tp1 <:< tp2) || (tp2 <:< tp1)) then tp1& tp2 == Nothing

[odersky]

@DarkDimius That would only work for unparameterized classes, and right now their glb would be Null. But yes, we could try that.

[smarter]

This seems to be equivalent to what you tried before in #843 and it fails in the same way, afterwards you changed your patch to only apply to ConstantType and added a comment that says:

// Make use of the fact that the intersection of two constant types
// types which are not subtypes of each other is known to be empty.
// Note: The same does not apply to singleton types in general.
// E.g. we could have a pattern match against `x.type & y.type`
// which might succeed if `x` and `y` happen to be the same ref
// at run time. It would not work to replace that with `Nothing`.
// However, maybe we can still apply the replacement to
// types which are not explicitly written.

[odersky]

@smarter Thanks for reminding me. Yes this is the same as #843 and it does not work.
I have a new PR that fixes the problem in a different way: #1058.