IoResult bloats return values

[mneumann]

IoResult adds 3 words to every function that returns it:

pub type IoResult<T> = Result<T, IoError>;

pub struct IoError {
    kind: IoErrorKind,
    desc: &'static str,
    detail: Option<~str>,
}

Wouldn't it be better to retrieve the exact error information out of the underlying reader/writer. I know that this introduces state which has some other problems. My idea is to simply return an Ok(T) or Err in the same way as Go handles this.

[adrientetar]

@mneumann Are you looking for Result?
http://static.rust-lang.org/doc/master/std/result/enum.Result.html

[zkamsler]

I am doubtful about this. Moving the error into the underlying reader or writer makes it very difficult to bubble up errors. The ultimate caller may not even have direct access to the original source of the error. In addition, there are some methods that return an IoResult that do not take &mut self, such as Seek::tell.

[sfackler]

Have you seen a measurable performance impact from use of IoResult?

[mneumann]

Returning an IoResult makes sense for unbuffered I/O operations, but when byte-wise iterating a BufferedStream for example, this probably has some significant overhead. I think the current implementation is not optimal yet, or maybe I just have to implement my own buffering.

[alexcrichton]

It would be helpful for you to provide concrete test cases and performance numbers in order to debug this. Stating that IoResult<T> adds 3 words is a factual statement, but stating that it cause significant overhead is tough to understand without any data supporting it.

[mneumann]

Ok, I'll try to come up with a benchmark.

[mneumann]

I benchmarked my msgpack library, once with rustc 0.9 and then with rustc-0.10-pre as of today on a 120 MB msgpack file. I couldn't notice any difference in performance.

I have looked at the assembly output. The only difference is that it allocates more stack space for the bigger error result.