ACP: `std::ffi::c` and `std::ffi::os` submodules

[clarfonthey]

Proposal

Note: this ACP will reference the std crate and submodules of it, although in cases where types are also available in core, equivalent modules should also be added there as well.

Problem statement + Motivation, use cases

(It's difficult to separate these two, so, I combined them into one section for this ACP.)

As part of rust-lang/rust#104353 (ACP: #135), I mention that adding a Bytes iterator type to std::ffi could potentially be ambiguous, since both C strings and OS strings could theoretically offer one. This problem exists for any iterators or other types added adjacent to C strings and OS strings that could be potentially ambiguous.

There is precedent for this among other standard library types, since essentially all collection and sequence types have their own submodule so they can be paired with their iterators. Even std::option and std::result exist for this purpose, since otherwise these modules would only contain the types Option and Result respectively.

Solution sketches

The recommendation of this ACP is to add std::ffi::c and std::ffi::os modules, although std::c_str and std::os_str could technically be added as an alternative.

Proposed moves:

• std::ffi::CStr => std::ffi::c::CStr
• std::ffi::CString => std::ffi::c::CString
• std::ffi::CStrBytes (Tracking Issue for CStr::bytes rust#112115) => std::ffi::c::Bytes
• std::ffi::OsStr => std::ffi::os::OsStr
• std::ffi::OsString => std::ffi::os::OsString

To ensure backwards-compatibility, use self::{c::{CStr, CString}, os::{OsStr, OsString}} would be added into std::ffi to re-export the moved items.

This would be similar to how, for example, std::collections::BinaryHeap is just a re-export of std::collections::binary_heap::BinaryHeap, although there are still types only included in std::collections::binary_heap like std::collections::binary_heap::Iter.

Links and related work

• CStr::bytes PR: Add CStr::bytes iterator rust#104353
• CStr::bytes ACP: ACP: CStr::bytes iterator #135

What happens now?

This issue is part of the libs-api team API change proposal process. Once this issue is filed the libs-api team will review open proposals in its weekly meeting. You should receive feedback within a week or two.

[SUPERCILEX]

I don't think this is worth it due to backwards compat. How would users know to use one import over the other? I guess we could add a clippy lint or deprecate the old imports, but that would break a lot of code. Calling it CStrBytes seems perfectly reasonable to me.

[clarfonthey]

The answer is that both would be valid, and there's no need to deprecate the original aliases. This is the same as how collections work today: you can import std::collections::VecDeque or std::collections::vec_deque::VecDeque and there aren't any specific complaints about using one or the other, and I'd imagine a lot of people would continue to use std::ffi::CStr and friends as well.

[SUPERCILEX]

Hmmm, I see how that would make sense. And ffi presumably wouldn't export the Bytes so you'd have to take it from the c module? I still don't like the churn, but if you're ok with that then this seems like a good idea.

[clarfonthey]

Yup! You would have to explicitly use std::ffi::c::Bytes instead of use std::ffi::CStrBytes, although in most cases you just wouldn't import it at all since I'd imagine most uses won't need to know anything beyond impl Iterator<Item = u8>.

[clarfonthey]

Reworded the original description to include newly approved CStr::bytes iterator, and to say how this would be done via a bulleted list instead of prose for clarity.

[dtolnay]

I am not on board with c and os as module names. I feel like they unnecessarily obscure the point behind this change that's described in the Motivation section, which is to provide grouping for iterator types and error types related to CStr/CString and OsStr/OsString respectively.

For examples does c_int belong in std::ffi::c because it is a C thing / does VaList belong in std::ffi::c because it is an extern "C" thing? According to this API they do not, because the c in std::ffi::c does not refer to C-related APIs, it refers to CStr-/CString-related APIs, despite "str" not appearing in the module path. A module named c might be justifiable in the context of something like std::ffi::str::c or std::str::c or std::ffi::c::str but I don't think those are the ideal organization here.

I don't like the mentioned alternative of std::c_str+std::os_str either. For example we also do not have std::binary_heap which is a type mentioned in the ACP. None of these are ubiquitous enough types for that the way path or vec or option are.

I do see the need for type-specific modules for categorization of iterators and error types inside of std::ffi, for the same reason that std::collections is also done this way.

std::ffi::c_str and std::ffi::os_str would work, I think.

[clarfonthey]

I think that std::c_str and std::os_str were genuinely typos, and I meant to type std::ffi::c_str and std::ffi::os_str like you suggested. I even thought that was what I had wrote initially, and had to double-check.

I'm okay with updating the proposal to use those instead.

[dtolnay]

Marking this ACP-accepted API Change Proposal is accepted (seconded with no objections) specifically in regard to std::ffi::c_str.

There are enough CStr-specific types with nondescript names floating around in std::ffi that really belonged in a c_str module to begin with.

• https://doc.rust-lang.org/1.69.0/std/ffi/struct.FromBytesWithNulError.html
• https://doc.rust-lang.org/1.69.0/std/ffi/struct.FromVecWithNulError.html
• https://doc.rust-lang.org/1.69.0/std/ffi/struct.IntoStringError.html
• https://doc.rust-lang.org/1.69.0/std/ffi/struct.NulError.html

And of course CStr and CString themselves would also go there, and the iterator type from:

• Tracking Issue for CStr::bytes rust#112115

On std::ffi::os_str I am less sold. There is nothing currently proposed to go there beyond OsStr and OsString, right? If so, I don't see the need for now. We can do it when those types begin accumulating more surface area, as might happen following this change:

• Allow limited access to OsStr bytes rust#109698

[dtolnay]

Implementation suggestion: play with #[doc(inline)] / #[doc(no_inline)] to nudge reader toward importing CStr as use std::ffi::CStr but importing e.g. FromBytesWithNulError as use std::ffi::c_str::FromBytesWithNulError.

I see std::collections already uses #[doc(inline)] in places.

[clarfonthey]

Noted! Will open up a PR for the std::ffi::c_str module with the changes you suggested.