Does C# 4 optimize away namespaces in a manner that previous C# versions did not?
Solution 1:
A little-appreciated fact is that there is no such thing as a "namespace" from the point of view of the underlying CLR type system. Rather, it's just a convention that we say that a type that contains periods in its name is "a member of a namespace". Logically there is no difference at all between the legal code:
namespace N
{
class C {}
}
and the psuedo-code:
class N.C {}
C# forces you to pretend this pleasant fiction is reality, but it is just a fiction -- from the perspective of the CLR type system, of course. From the perspective of the C# compiler, of course namespaces are "real". They just don't correspond to anything in metadata other than a portion of the name of a type.
In short: if you make an assembly with an "empty" namespace then the "namespace" doesn't exist at all in the compiled binary. A "namespace" only comes into existence when there is a type in the library that has periods in its name.
Now, why you would care about ensuring that an "empty" namespace has some presence in the binary form, I have no idea.
I assume previous C# versions exhibit behaviour that do not require this dummy class
Nope. Every version of C# since 1.0 throws away empty namespaces.
Solution 2:
Given that the namespace doesn't contain any members (without that class), I'm not sure there's even the concept of a namespace at that point... nor would I expect it to be useful anyway.
I've just tried to reproduce this with the C# 2 compiler, and I can't see any trace of an empty namespace within the IL.
Solution 3:
The only semi-related issue I can think of is that when compiling a project in msbuild, indirect references are not always copied to the bin directory of the current app. If library B indirectly references library A and library C references B only, library A's output will not necessarily be copied to the bin folder when compiling library C. In the past, I've used a null field reference on a class to ensure that the dependency is explicit and the output is deployed properly. Maybe the original devs experienced something similar and this was their solution?