Use size_hint when using read_to_end on something that impl's Read

[v1gnesh]

In the default_read_to_end implementation, a parameter for size_hint is available.
However, when doing the actual read_to_end, it's initialized with None.

When using the the below pattern to implement a custom iterator over a file, when the max possible line length is known, it would be helpful to make use of the size_hint parameter, assuming it's a slightly faster path (?).

This will be, in spirit, a continuation of #89165, #89582, and #110655

impl<R: Read> Iterator for LineIter<R> {
    type Item = io::Result<Vec<u8>>;

    fn next(&mut self) -> Option<Self::Item> {

        let mut data = Vec::with_capacity(MAX_LINE_LENGTH);
        file.by_ref().take(x).read_to_end(&mut data)

In short, I'm asking if the ability to pass a size_hint of x (from the take) into the read_to_end function will be useful, and if it'd be slightly better performance wise?

[Jules-Bertholet]

@rustbot label T-libs I-slow

[chenyukang]

If you means fs file, seems we added size_hint at here.

rust/library/std/src/fs.rs

Lines 766 to 769 in 1c44af9

	fn read_to_end(&mut self, buf: &mut Vec<u8>) -> io::Result<usize> {
	let size = buffer_capacity_required(self);
	buf.reserve(size.unwrap_or(0));
	io::default_read_to_end(self, buf, size)

For the default read_to_end , because we don't have a default method to get the guessing size?
I don't have the whole code snippet you are showing, if we want to passing the x from take, seems need to add an API such as read_to_end_with_hint(self, &mut data, x).

[v1gnesh]

@chenyukang, yes I noticed the addition to fs has read_to_end.
I wanted to ask if a similar thing can be provided for use when reading larger than memory files (through an iter in this case).

x is an unsigned int that is calculated in the next() function. It is found at run-time as we iterate through a file, and contains the length of each Item.

It may be as simple as surfacing the default_read_to_end function, as that'll allow us to use Option<len> in place of None for the size_hint.

As a side note, I wonder if the ability to define functions with optional parameters will simplify things in general.
Certainly don't know anything about Rust's language design, etc. ... just curious.

[saethlin]

I am very confused by this issue, maybe you can clear things up.

In short, I'm asking if the ability to pass a size_hint of x (from the take) into the read_to_end function will be useful, and if it'd be slightly better performance wise?

My basic response to this is: Why would we do this? Isn't it just as much work to call buf.reserve(capacity) as it is to call read_to_end(&mut buf, capacity)? std::fs::read has very good reason to handle the reserving itself, there's no way for a caller to reserve for it. But I don't see why the change described here needs a library change.

I wanted to ask if a similar thing can be provided for use when reading larger than memory files (through an iter in this case).

If the file is larger than memory, surely the size hint would also be larger than memory, and thus acting on it would just OOM faster. So I don't see how a size hint helps here. If you want to iterate through a file larger than memory, use BufReader, and you can use that type to pick a capacity that suits your use case.

As a side note, I wonder if the ability to define functions with optional parameters will simplify things in general.
Certainly don't know anything about Rust's language design, etc. ... just curious.

You're looking for https://github.com/rust-lang/rfcs, you may find some interesting reading material by searching the PRs on that repo. I did a quick skim and I didn't find anything about optional arguments specifically, but I did find a lot of discussion on an RFC about keyword arguments.

[v1gnesh]

If the file is larger than memory, surely the size hint would also be larger than memory, and thus acting on it would just OOM faster. So I don't see how a size hint helps here.

Ah, sorry, I could have been clearer.
The file can be larger than memory, but it is structured in such a way that creating a custom iterator works, because the file is like this

| item | item | item |
u8 u8 u8 u8 ............ u8 u8 u8 u8 ............ u8 u8 u8 u8 ............

... where those u8s represent the length (including itself) of each Item in the file.
This is the x I refer to in the first comment. File size is huge, but the size of each line/item is easy to get.

Does that help?

Because read_to_end writes only to a Vec, and I'm looking to avoid the chance of growing the Vec because the Item sizes are known, either using a size_hint, or writing to a sized array instead of a Vec (I'm assuming) would be better?
Or is there a better way that you'd recommend?

[saethlin]

It sounds to me like you don't want read_to_end at all. You probably want the feature called read_buf so you can create a large uninit buffer, then call Read::read_buf_exact to fill the buffer.

But more generally, this sounds like you're suggesting a new API, in which case you should follow https://std-dev-guide.rust-lang.org/development/feature-lifecycle.html or try to start a conversation about your idea on https://internals.rust-lang.org/. This repo's issue tracker is not the right place for the kind of discussion you're trying to have; there's just not the right audience here. You can find the right audience for developing a new feature at those links.

[v1gnesh]

Okay, thanks for the guidance! I'll close this off then.