make `write(io, Any[...])` an error

[mbauman]

I noticed this with IOBuffer, but it looks like IOStream, AsyncStream, and File all have the same issue. In each of their read!{T,N}(::IO, a::Array{T,N}) methods, they assume that the number of bytes to read is length(a)*sizeof(T). This is not true for Char:

julia> buf = IOBuffer("abcdefghijklmnop");

julia> read(buf, Char)
'a'

julia> read(buf, Char, 2) # dispatches to read!(IOBuffer, Array{Char,1})
2-element Array{Char,1}:
 '\U65646362'
 '\U69686766'

For IOBuffer, I think the specialization could simply be removed (albeit perhaps at a small performance loss). The others, however, are trickier — File performance is much more critical and the Async/IOStream ones need to know how many bytes to wait for.

[JeffBezanson]

I think the best solution here is to better separate text and binary I/O. If you want to serialize an Array{Char}, then arguably text encoding is beside the point and it makes sense to just dump the bits.

[kshyatt]

This is reproducible on today's master.

[JeffBezanson]

This has gotten slightly better since we changed read(io, Char, 2) to read!(io, Vector{Char}(2)), which makes it clear you're reading an Array, and that Array determines the representation.

One question is what read(io, Char) should do. Currently it's UTF-8; is that the right thing to do?

Another issue is writing non-isbits arrays, or Any arrays. Currently we handle these by calling write in a loop, but this is a strange thing to do since you'll end up with a basically unusable binary representation. Arguably you should have to write the loop yourself if you really want this. Triage thinks we should make this case an error.

[StefanKarpinski]

Shouldn't it be an error to serialize any pointer array?

[JeffBezanson]

Yes, I was using Any as a shorthand for that.