Restructure Try node.

[DarkDimius]

Now Try node always has cases as handlers. In case handler is an Ident of
type Throwable => T than it's desugared to a CaseDef during parsing.

[DarkDimius]

@odersky please review

[retronym]

In Scala2, handlers of Throwable => T are not allowed:

scala> try { ??? } catch { case F }
<console>:1: error: '=>' expected but '}' found.
       try { ??? } catch { case F }
                                  ^

scala> try { ??? } catch F
<console>:9: error: value isDefinedAt is not a member of Throwable => Unit
              try { ??? } catch F
                                ^

Only PartialFunction-s are supported:

scala> val F: PartialFunction[Throwable, Unit] = { case _ => }
F: PartialFunction[Throwable,Unit] = <function1>

scala> try { ??? } catch F

Perhaps this is a conscious generalisation of the existing facility, though.

[retronym]

I also notice that we allow an expression that is implicit convertible to that type.

scala> try { ??? } catch ""

scala> try { ??? } catch "".reverse

The spec says:

The handler h is expected conform to type scala.PartialFunction[scala.Throwable, pt]

The implementation wrongly appears to assume s/conform to/be compatible with/.

I have lodged this as https://issues.scala-lang.org/browse/SI-8945.

[retronym]

Isn't the dotty parser designed as a desugaring-free zone?

[odersky]

I proposed the representation in the PR, but Jason is right. It is not good that the parser does this desugaring. The parser represents naked case blocks

 { case P1 => E1 ... case Pn => En }

as Match nodes with EmptyTree as the selector. How about we use the same representation for catch blocks? I.e. a Try would be represented as follows:

 case class Try(body: Tree, handler: Tree, finalizer: Tree)

and after desugaring we would make sure that the handler is always of the form

 Match(EmptyTree, cases)

What do people think about this?

[retronym]

SGTM

[DarkDimius]

@odersky In case Try after desugaring(and typeing) should always have handlers as cases, I believe it would be better to have this enforced by types.
We already have a untyped trees that are parsed but are gone after typechecking&desugaring. I propose to have a different tree representing Try nodes before and after desugaring. See my last commit.

What do you think?

[DarkDimius]

@odersky should be ready for review. #206 Conflicts with this as both change typer and I would prefer to get this in first.

[odersky]

Typo

[odersky]

Otherwise LGTM.