remove use of `Ref{Void}` as a struct field type

[juliamath]

remove use of `Ref{Void}` as a struct field type

[vtjnash]

This strikes me as a bug in the type safety of Ref.

true, this may be worth a special case for Ref. In C, the void type is more nearly Union{} in Julia. The common mapping in Julia to Void has some slightly different behaviors (such as this one).

[JeffBezanson]

I don't fully agree with that; Union{} is closer to a noreturn declaration. void in C is pretty close to being a normal observable value.

This problem would occur with any type; assigning a Ptr{Int} to a Ref{Int} field would do the same thing. The only fix I can think of so far is to throw an error for converting a Ptr to a Ref. Ptr could still be a subtype of Ref. That seems to handle the contexts where Ptr can vs. can't be substituted as a Ref.

[vtjnash]

If you declare the field type wrong, the code will be wrong. There's nothing unique to Ref there, since the same applies to any other error in field declaration. otoh, if you declare the fields correctly, then you can't substitute Ref for Ptr (since that requires an unsafe_convert), which is the intended direction for usage-safety.

void is not constructible, cannot be used as a field, and is convertible to/from all other types but cannot itself be constructed: http://programmers.stackexchange.com/a/254191
In C++, void &x would be an error because void cannot be constructed, which is more similar to how Union{} behaves in Julia. In particular, the C standard specifies that the presence of the void type is special in that it does not have a value (usage of it would be an error), but the code must still be executed for side-effects.

[JeffBezanson]

I think this problem reflects some unclarity about when to use Ref instead of Ptr. The docs basically say "use Ref". Maybe all we need to do is update the docs to say "never use Ref as a struct field"? However this is a bit complicated by the fact that if you use Ref{Int} as a field type, assigning an Int to it will actually work. Assigning a Ptr to a Ref field may be wrong, but it's wrong in a very confusing way that could benefit from an error message.

[jiahao]

As I understood from discussion with @JeffBezanson and @vtjnash offline, the general rule of thumb is to use Ref{T} only in covariant context to replace the x = Array(T, 1); ccall(... &x) idiom in old Julia. In other scenarios like the return value of a ccall and a Julia type mimicking a C struct, Ptr{T} should be used instead as the proper representation of a C pointer (an integer representing a potentially useful address, lacking the address validation and GC rooting created by Ref to a Julia object).

[jiahao]

I think all reasonable actions pertaining to this issue have been taken. The Julia manual has been updated and Refs excised.