Alternative fix for #1857: Consider tvars in bounds of uninstantiated tvars

[smarter]

This is an alternative to #1859, review by @odersky

---

Given:

  val f: Foo = new Foo
  def wrap2a[A, W <: Wrapper[A]](host: A)(implicit w: W): W = ???
  wrap2a(f)

We need to instantiate A before doing the implicit search for w
because the type of w indirectly depends on A and A was constrained by the first parameter list. This was not
done before because we did not traverse the constraint bounds of an
uninstantiated type variable.

[odersky]

I agree it's better to do it this way compared to #1859, because interpolateUndetVars is really more an optimization and we should not hang semantics on it. There's still an open question, though: IsFullyDefinedAccumulator is called in several situations and only one is instantiateSelected. Are we sure it's the right behavior in the other calling contexts as well? We'd need to see an argument why that would be the case.

Or else we should do this only when called from instantiateSelected.

Also, this new behavior has to be documented, in isFullyDefinedAccumulator and all callers.

[smarter]

IsFullyDefinedAccumulator is called in several situations and only one is instantiateSelected. Are we sure it's the right behavior in the other calling contexts as well? We'd need to see an argument why that would be the case.

I agree, we should try to come up with some examples where this makes a difference.

[allanrenucci]

@smarter #1857 is fixed. Can we close this?

[odersky]

I am closing for now.