Summary

This is a conventions RFC for settling the location of unsafe APIs relative to the types they work with, as well as the use of raw submodules.

The brief summary is:

Unsafe APIs should be made into methods or static functions in the same cases that safe APIs would be.
raw submodules should be used only to define explicit low-level representations.

Motivation

Many data structures provide unsafe APIs either for avoiding checks or working directly with their (otherwise private) representation. For example, string provides:

An as_mut_vec method on String that provides a Vec<u8> view of the string. This method makes it easy to work with the byte-based representation of the string, but thereby also allows violation of the utf8 guarantee.
A raw submodule with a number of free functions, like from_parts, that constructs a String instances from a raw-pointer-based representation, a from_utf8 variant that does not actually check for utf8 validity, and so on. The unifying theme is that all of these functions avoid checking some key invariant.

The problem is that currently, there is no clear/consistent guideline about which of these APIs should live as methods/static functions associated with a type, and which should live in a raw submodule. Both forms appear throughout the standard library.

Detailed design

The proposed convention is:

When an unsafe function/method is clearly “about” a certain type (as a way of constructing, destructuring, or modifying values of that type), it should be a method or static function on that type. This is the same as the convention for placement of safe functions/methods. So functions like string::raw::from_parts would become static functions on String.
raw submodules should only be used to define low-level types/representations (and methods/functions on them). Methods for converting to/from such low-level types should be available directly on the high-level types. Examples: core::raw, sync::raw.

The benefits are:

Ergonomics. You can gain easy access to unsafe APIs merely by having a value of the type (or, for static functions, importing the type).
Consistency and simplicity. The rules for placement of unsafe APIs are the same as those for safe APIs.

The perspective here is that marking APIs unsafe is enough to deter their use in ordinary situations; they don’t need to be further distinguished by placement into a separate module.

There are also some naming conventions to go along with unsafe static functions and methods:

When an unsafe function/method is an unchecked variant of an otherwise safe API, it should be marked using an _unchecked suffix.

For example, the String module should provide both from_utf8 and from_utf8_unchecked constructors, where the latter does not actually check the utf8 encoding. The string::raw::slice_bytes and string::raw::slice_unchecked functions should be merged into a single slice_unchecked method on strings that checks neither bounds nor utf8 boundaries.
When an unsafe function/method produces or consumes a low-level representation of a data structure, the API should use raw in its name. Specifically, from_raw_parts is the typical name used for constructing a value from e.g. a pointer-based representation.
Otherwise, consider using a name that suggests why the API is unsafe. In some cases, like String::as_mut_vec, other stronger conventions apply, and the unsafe qualifier on the signature (together with API documentation) is enough.

The unsafe methods and static functions for a given type should be placed in their own impl block, at the end of the module defining the type; this will ensure that they are grouped together in rustdoc. (Thanks @lilyball for the suggestion.)

Drawbacks

One potential drawback of these conventions is that the documentation for a module will be cluttered with rarely-used unsafe APIs, whereas the raw submodule approach neatly groups these APIs. But rustdoc could easily be changed to either hide or separate out unsafe APIs by default, and in the meantime the impl block grouping should help.

More specifically, the convention of placing unsafe constructors in raw makes them very easy to find. But the usual from_ convention, together with the naming conventions suggested above, should make it fairly easy to discover such constructors even when they’re supplied directly as static functions.

More generally, these conventions give unsafe APIs more equal status with safe APIs. Whether this is a drawback depends on your philosophy about the status of unsafe programming. But on a technical level, the key point is that the APIs are marked unsafe, so users still have to opt-in to using them. Ed note: from my perspective, low-level/unsafe programming is important to support, and there is no reason to penalize its ergonomics given that it’s opt-in anyway.

Alternatives

There are a few alternatives:

Rather than providing unsafe APIs directly as methods/static functions, they could be grouped into a single extension trait. For example, the String type could be accompanied by a StringRaw extension trait providing APIs for working with raw string representations. This would allow a clear grouping of unsafe APIs, while still providing them as methods/static functions and allowing them to easily be imported with e.g. use std::string::StringRaw. On the other hand, it still further penalizes the raw APIs (beyond marking them unsafe), and given that rustdoc could easily provide API grouping, it’s unclear exactly what the benefit is.
(Suggested by @lilyball):

Use raw for functions that construct a value of the type without checking for one or more invariants.

The advantage is that it’s easy to find such invariant-ignoring functions. The disadvantage is that their ergonomics is worsened, since they much be separately imported or referenced through a lengthy path:
```
// Compare the ergonomics:
string::raw::slice_unchecked(some_string, start, end)
some_string.slice_unchecked(start, end)
```
Another suggestion by @lilyball is to keep the basic structure of raw submodules, but use associated types to improve the ergonomics. Details (and discussions of pros/cons) are in this comment.
Use raw submodules to group together all manipulation of low-level representations. No module in std currently does this; existing modules provide some free functions in raw, and some unsafe methods, without a clear driving principle. The ergonomics of moving everything into free functions in a raw submodule are quite poor.

Unresolved questions