fix eof issues and add tests

[nhz2]

This fixes the issues with eof discussed in #168 (comment)

If eof returns false it is safe to read a byte. This means when the stream is in :write mode, eof must return true.

Also, if the stream is stopped when writing (stop_on_end=true) eof should not suddenly change from true to false. To fix this I added a :done state.

[vtjnash]

This should be implied by eof

Suggested change

ready_to_read!(stream)

[nhz2]

This change means an EOFError is thrown when reading a byte from a closed stream.
This seems to match what happens with IOStream

julia> p, io = mktemp()
("/tmp/jl_GKa3GO", IOStream(<fd 21>))

julia> close(io)

julia> read(io, UInt8)
ERROR: EOFError: read end of file
Stacktrace:
 [1] read(s::IOStream, ::Type{UInt8})
   @ Base ./iostream.jl:400
 [2] top-level scope
   @ REPL[43]:1

But doesn't match with what IOBuffer does:

julia> stream = IOBuffer();

julia> close(stream)

julia> read(stream, UInt8)
ERROR: ArgumentError: read failed, IOBuffer is not readable
Stacktrace:
 [1] _throw_not_readable()
   @ Base ./iobuffer.jl:165
 [2] read(from::IOBuffer, ::Type{UInt8})
   @ Base ./iobuffer.jl:218
 [3] top-level scope
   @ REPL[49]:1

[vtjnash]

True, though I think eof could optionally throw a state-exception in that case also sometimes. The ambiguity occurs when a stream is ambiguously readable or writable, which sometimes cannot be avoided. So seekable files usually close both ends at once (and eof is a transient condition, which changes to an error after calling close), whereas (non-seekable) streams the eof is a property of reading and isopen is a property of writing

[nhz2]

I'm undoing the change because I like how IOBuffer is handling this better than IOStream.

For example, I think read(io, String) after close should error instead of returning an empty string:

julia> p, io = mktemp()

julia> close(io)

julia> read(io, String)
""

julia> stream = IOBuffer();

julia> close(stream)

julia> read(stream, String)
ERROR: ArgumentError: read failed, IOBuffer is not readable

[vtjnash]

Yes, but that is already eof’s job to throw when there is a state violation, so it should not be necessary for read to also redundantly check

[nhz2]

This would break this test that was added in #32:

TranscodingStreams.jl/test/codecnoop.jl

Lines 283 to 287 in 92adcd8

	stream = NoopStream(IOBuffer(""))
	unsafe_write(stream, C_NULL, 0)
	@test eof(stream) # write
	close(stream)
	@test eof(stream) # close

[nhz2]

I agree that eof should throw if the stream is not in a readable state.

However this seems like a breaking change, so I will split this PR into smaller parts, and add the eof changes in v0.11 for v0.11 I would also like to fix #93, #109 and remove the seekend implementation.