1. When a literal is typed, and is not a constant, we don't need its
elements for type inference. Moreover, in this case elements are
allowed to include nodes that are not immediately-evident expressions,
such as instance creations without all type arguments, closures with
block bodies, etc.
2. When a literal is a constant, so we still have to serialize all its
elements, it is possible that we will encounter something not allowed
for immediately-evident expressions, and fail. But that's OK, because
such expressions are also not allowed in constants, so we will have
a compile time error anyway, and IMHO it is OK to degrade a little
in presence of errors.
3. The same is true for method invocations and instance creations,
we need to serialize arguments only for constants, because top-level
inference never uses arguments to infer returned types.
R=brianwilkerson@google.com
BUG=
Review-Url: https://codereview.chromium.org/2779993002 .
Konstantin Shcheglov
84a4b01081