About this guide

Status: Stub

This guide is for contributors and reviewers to Rust's standard library.

Other places to find information

You might also find the following sites useful:

  • std API docs -- rustdoc documentation for the standard library itself
  • Forge -- contains documentation about rust infrastructure, team procedures, and more
  • libs-team -- the home-base for the rust Library Team, with description of the team procedures, active working groups, and the team calendar.

Getting started

Status: Stub

Welcome to the standard library!

This guide is an effort to capture some of the context needed to develop and maintain the Rust standard library. Its goal is to help members of the Libs team share the process and experience they bring to working on the standard library so other members can benefit. It’ll probably accumulate a lot of trivia that might also be interesting to members of the wider Rust community.

Where to get help

Maintaining the standard library can feel like a daunting responsibility!

Ping the @rust-lang/libs-impl or @rust-lang/libs teams on GitHub anytime.

You can also reach out in the t-libs stream on Zulip.

A tour of the standard library

Status: Stub

The standard library codebase lives in the rust-lang/rust repository under the /library directory.

The standard library is made up of three crates that exist in a loose hierarchy:

  • core: dependency free and makes minimal assumptions about the runtime environment.
  • alloc: depends on core, assumes allocator support. alloc doesn't re-export core's public API, so it's not strictly above it in the layering.
  • std: depends on core and alloc and re-exports both of their public APIs.

Reviewer checklist

Status: Stub

Check the getting started guide for an introduction to developing in the standard library.

If you'd like to reassign the PR, you can:

r? @user

Before you review

  • Is this a stabilization or deprecation?
    • Make sure there's a completed FCP somewhere for it.
    • Ping @rust-lang/libs for input.

As you review

Look out for code considerations:

Before you merge

  • Is the commit log tidy? Avoid merge commits, these can be squashed down.
  • Can this change be rolled up?
  • Is this a new unstable feature?
    • Create a tracking issue.
    • Update the #[unstable] attributes to point to it.

When you're ready

Use @bors to merge the pull request.

To roll up:

@bors r+ rollup

or just:

@bors r+

The feature lifecycle

Status: Stub

Landing new features

Status: Stub

New unstable features can be added and approved without going through a Libs FCP. There should be some buy-in from Libs that a feature is desirable and likely to be stabilized at some point before landing though.

If you're not sure, open an issue against rust-lang/rust first suggesting the feature before developing it.

All public items in the standard library need a #[stable] or #[unstable] attribute on them. When a feature is first added, it gets a #[unstable] attribute.

Before a new feature is merged, those #[unstable] attributes need to be linked to a tracking issue.

Using tracking issues

Status: Stub

Tracking issues are used to facilitate discussion and report on the status of standard library features. All public APIs need a dedicated tracking issue. Some larger internal units of work may also use them.

Creating a tracking issue

There's a template that can be used to fill out the initial tracking issue. The Libs team also maintains a Cargo tool that can be used to quickly dump the public API of an unstable feature.

Working on an unstable feature

The current state of an unstable feature should be outlined in its tracking issue.

If there's a change you'd like to make to an unstable feature, it can be discussed on the tracking issue first.

Stabilizing features

Status: Stub

Feature stabilization involves adding #[stable] attributes. They may be introduced alongside new trait impls or replace existing #[unstable] attributes.

Stabilization goes through the Libs FCP process, which occurs on the tracking issue for the feature.

Before writing a PR to stabilize a feature

Check to see if a FCP has completed first. If not, either ping @rust-lang/libs or leave a comment asking about the status of the feature.

This will save you from opening a stabilization PR and having it need regular rebasing while the FCP process runs its course.

Writing a stabilization PR

  • Replace any #[unstable] attributes for the given feature with stable ones. The value of the since field is usually the current nightly version.
  • Remove any #![feature()] attributes that were previously required.
  • Submit a PR with a stabilization report.

When there's const involved

Const functions can be stabilized in a PR that replaces #[rustc_const_unstable] attributes with #[rustc_const_stable] ones. The Constant Evaluation WG should be pinged for input on whether or not the const-ness is something we want to commit to. If it is an intrinsic being exposed that is const-stabilized then @rust-lang/lang should also be included in the FCP.

Check whether the function internally depends on other unstable const functions through #[allow_internal_unstable] attributes and consider how the function could be implemented if its internally unstable calls were removed. See the Stability attributes page for more details on #[allow_internal_unstable].

Where unsafe and const is involved, e.g., for operations which are "unconst", that the const safety argument for the usage also be documented. That is, a const fn has additional determinism (e.g. run-time/compile-time results must correspond and the function's output only depends on its inputs...) restrictions that must be preserved, and those should be argued when unsafe is used.

Deprecating features

Status: Stub

Public APIs aren't deleted from the standard library. If something shouldn't be used anymore it gets deprecated by adding a #[rustc_deprecated] attribute. Deprecating need to go through a Libs FCP, just like stabilizations do.

To try reduce noise in the docs from deprecated items, they should be moved to the bottom of the module or impl block so they're rendered at the bottom of the docs page. The docs should then be cut down to focus on why the item is deprecated rather than how you might use it.

Code considerations

Code considerations capture our experiences working on the standard library for all contributors. If you come across something new or unexpected then a code consideration is a great place to record it. Then other contributors and reviewers can find it by searching the guide.

How to write a code consideration

Code considerations are a bit like guidelines. They should try make concrete recommendations that reviewers and contributors can refer to in discussions. A link to a real case where this was discussed or tripped us up is good to include.

Code considerations should also try include a For reviewers section. These can call out specific things to look out for in reviews that could suggest the consideration applies. They can also include advice on how to apply it.

It's more important that we capture these experiences somehow though, so don't be afraid to drop some sketchy notes in and debate the details later!


Status: Stub

Most of the considerations in this guide are quality in some sense. This section has some general advice on maintaining code quality in the standard library.

For reviewers

Think about how you would implement a feature and whether your approach would differ from what's being proposed. What trade-offs are being made? Is the weighting of those trade-offs the most appropriate?

Public API design

Status: Stub

Standard library APIs typically follow the API Guidelines, which were originally spawned from the standard library itself.

For reviewers

For new unstable features, look for any prior discussion of the proposed API to see what options and tradeoffs have already been considered. If in doubt, ping @rust-lang/libs for input.

When to add #[must_use]

The #[must_use] attribute can be applied to types or functions when failing to explicitly consider them or their output is almost certainly a bug.

As an example, Result is #[must_use] because failing to consider it may indicate a caller didn't realise a method was fallible:

// Is `check_status` infallible? Or did we forget to look at its `Result`?

Operators like saturating_add are also #[must_use] because failing to consider their output might indicate a caller didn't realise they don't mutate the left-hand-side:

// A caller might assume this method mutates `a`

Combinators produced by the Iterator trait are #[must_use] because failing to use them might indicate a caller didn't realize Iterators are lazy and won't actually do anything unless you drive them:

// A caller might not realise this code won't do anything
// unless they call `collect`, `count`, etc.
v.iter().map(|x| println!("{}", x));

On the other hand, thread::JoinHandle isn't #[must_use] because spawning fire-and-forget work is a legitimate pattern and forcing callers to explicitly ignore handles could be a nuisance rather than an indication of a bug:

thread::spawn(|| {
    // this background work isn't waited on

For reviewers

Look for any legitimate use-cases where #[must_use] will cause callers to explicitly ignore values. If these are common then #[must_use] probably isn't appropriate.

The #[must_use] attribute only produces warnings, so it can technically be introduced at any time. To avoid accumulating nuisance warnings though ping @rust-lang/libs for input before adding new #[must_use] attributes to existing types and functions.

Breaking changes

Breaking changes should be avoided when possible. RFC 1105 lays the foundations for what constitutes a breaking change. Breakage may be deemed acceptable or not based on its actual impact, which can be approximated with a crater run.

There are strategies for mitigating breakage depending on the impact.

For changes where the value is high and the impact is high too:

  • Using compiler lints to try phase out broken behavior.

If the impact isn't too high:

  • Looping in maintainers of broken crates and submitting PRs to fix them.

For reviewers

Look out for changes to documented behavior and new trait impls for existing stable traits.

Breakage from changing behavior

Breaking changes aren't just limited to compilation failures. Behavioral changes to stable functions generally can't be accepted. See the home_dir issue for an example.

An exception is when a behavior is specified in an RFC (such as IETF specifications for IP addresses). If a behavioral change fixes non-conformance then it can be considered a bug fix. In these cases, @rust-lang/libs should still be pinged for input.

For reviewers

Look out for changes in existing implementations for stable functions, especially if assertions in test cases have been changed.

Breakage from new trait impls

A lot of PRs to the standard library are adding new impls for already stable traits, which can break consumers in many weird and wonderful ways. The following sections gives some examples of breakage from new trait impls that may not be obvious just from the change made to the standard library.

Also see #[fundamental] types for special considerations for types like &T, &mut T, Box<T>, and other core smart pointers.

Inference breaks when a second generic impl is introduced

Rust will use the fact that there's only a single impl for a generic trait during inference. This breaks once a second impl makes the type of that generic ambiguous. Say we have:

// in `std`
impl From<&str> for Arc<str> { .. }
// in an external `lib`
let b = Arc::from("a");

then we add:

impl From<&str> for Arc<str> { .. }
+ impl From<&str> for Arc<String> { .. }


let b = Arc::from("a");

will no longer compile, because we've previously been relying on inference to figure out the T in Box<T>.

This kind of breakage can be ok, but a crater run should estimate the scope.

Deref coercion breaks when a new impl is introduced

Rust will use deref coercion to find a valid trait impl if the arguments don't type check directly. This only seems to occur if there's a single impl so introducing a new one may break consumers relying on deref coercion. Say we have:

// in `std`
impl Add<&str> for String { .. }

impl Deref for String { type Target = str; .. }
// in an external `lib`
let a = String::from("a");
let b = String::from("b");

let c = a + &b;

then we add:

impl Add<&str> for String { .. }
+ impl Add<char> for String { .. }


let c = a + &b;

will no longer compile, because we won't attempt to use deref to coerce the &String into &str.

This kind of breakage can be ok, but a crater run should estimate the scope.

For reviewers

Look out for new #[stable] trait implementations for existing stable traits.

#[fundamental] types

Status: Stub

Type annotated with the #[fundamental] attribute have different coherence rules. See RFC 1023 for details. That includes:

  • &T
  • &mut T
  • Box<T>
  • Pin<T>

Typically, the scope of breakage in new trait impls is limited to inference and deref-coercion. New trait impls on #[fundamental] types may overlap with downstream impls and cause other kinds of breakage.

For reviewers

Look out for blanket trait implementations for fundamental types, like:

impl<'a, T> PublicTrait for &'a T
    T: SomeBound,


unless the blanket implementation is being stabilized along with PublicTrait. In cases where we really want to do this, a crater run can help estimate the scope of the breakage.

Breaking changes to the prelude

Making changes to the prelude can easily cause breakage because it impacts all Rust code. In most cases the impact is limited since prelude items have the lowest priority in name lookup (lower than glob imports), but there are two cases where this doesn't work.


Adding a new trait to the prelude causes new methods to become available for existing types. This can cause name resolution errors in user code if a method with the same name is also available from a different trait.

For this reason, TryFrom and TryInto were only added to the prelude for the 2021 edition despite being stabilized in 2019.


Unlike other item types, rustc's name resolution for macros does not support giving prelude macros a lower priority than other macros, even if the macro is unstable. As a general rule, avoid adding macros to the prelude except at edition boundaries.

This issues was encoutered when trying to land the assert_matches! macro.

Safety and soundness

Status: Stub

Unsafe code blocks in the standard library need a comment explaining why they're ok. There's a lint that checks this. The unsafe code also needs to actually be ok.

The rules around what's sound and what's not can be subtle. See the Unsafe Code Guidelines WG for current thinking, and consider pinging @rust-lang/libs-impl, @rust-lang/lang, and/or somebody from the WG if you're in any doubt. We love debating the soundness of unsafe code, and the more eyes on it the better!

For reviewers

Look out for any unsafe blocks. If they're optimizations consider whether they're actually necessary. If the unsafe code is necessary then always feel free to ping somebody to help review it.

Look at the level of test coverage for the new unsafe code. Tests do catch bugs!

Generics and unsafe

Be careful of generic types that interact with unsafe code. Unless the generic type is bounded by an unsafe trait that specifies its contract, we can't rely on the results of generic types being reliable or correct.

A place where this commonly comes up is with the RangeBounds trait. You might assume that the start and end bounds given by a RangeBounds implementation will remain the same since it works through shared references. That's not necessarily the case though, an adversarial implementation may change the bounds between calls:

struct EvilRange(Cell<bool>);

impl RangeBounds<usize> for EvilRange {
    fn start_bound(&self) -> Bound<&usize> {
        Bound::Included(if self.0.get() {
        } else {
    fn end_bound(&self) -> Bound<&usize> {

This has caused problems in the past for code making safety assumptions based on bounds without asserting they stay the same.

Code using generic types to interact with unsafe should try convert them into known types first, then work with those instead of the generic. For our example with RangeBounds, this may mean converting into a concrete Range, or a tuple of (Bound, Bound).

For reviewers

Look out for generic functions that also contain unsafe blocks and consider how adversarial implementations of those generics could violate safety.

Drop and #[may_dangle]

A generic Type<T> that manually implements Drop should consider whether a #[may_dangle] attribute is appropriate on T. The Nomicon has some details on what #[may_dangle] is all about.

If a generic Type<T> has a manual drop implementation that may also involve dropping T then dropck needs to know about it. If Type<T>'s ownership of T is expressed through types that don't drop T themselves such as ManuallyDrop<T>, *mut T, or MaybeUninit<T> then Type<T> also needs a PhantomData<T> field to tell dropck that T may be dropped. Types in the standard library that use the internal Unique<T> pointer type don't need a PhantomData<T> marker field. That's taken care of for them by Unique<T>.

As a real-world example of where this can go wrong, consider an OptionCell<T> that looks something like this:

struct OptionCell<T> {
    is_init: bool,
    value: MaybeUninit<T>,

impl<T> Drop for OptionCell<T> {
    fn drop(&mut self) {
        if self.is_init {
            // Safety: `value` is guaranteed to be fully initialized when `is_init` is true.
            // Safety: The cell is being dropped, so it can't be accessed again.
            unsafe { self.value.assume_init_drop() };

Adding a #[may_dangle] attribute to this OptionCell<T> that didn't have a PhantomData<T> marker field opened up a soundness hole for T's that didn't strictly outlive the OptionCell<T>, and so could be accessed after being dropped in their own Drop implementations. The correct application of #[may_dangle] also required a PhantomData<T> field:

struct OptionCell<T> {
    is_init: bool,
    value: MaybeUninit<T>,
+   _marker: PhantomData<T>,

- impl<T> Drop for OptionCell<T> {
+ unsafe impl<#[may_dangle] T> Drop for OptionCell<T> {

For reviewers

If there's a manual Drop implementation, consider whether #[may_dangle] is appropriate. If it is, make sure there's a PhantomData<T> too either through Unique<T> or as a field directly.

Using mem to break assumptions

mem::replace and mem::swap

Any value behind a &mut reference can be replaced with a new one using mem::replace or mem::swap, so code shouldn't assume any reachable mutable references can't have their internals changed by replacing.


Rust doesn't guarantee destructors will run when a value is leaked (which can be done with mem::forget), so code should avoid relying on them for maintaining safety. Remember, everyone poops.

It's ok not to run a destructor when a value is leaked because its storage isn't deallocated or repurposed. If the storage is initialized and is being deallocated or repurposed then destructors need to be run first, because memory may be pinned. Having said that, there can still be exceptions for skipping destructors when deallocating if you can guarantee there's never pinning involved.

For reviewers

If there's a Drop impl involved, look out for possible soundness issues that could come from that destructor never running.

Using unstable language features

The standard library codebase is a great place to try unstable language features, but we have to be careful about exposing them publicly. The following is a list of unstable language features that are ok to use within the standard library itself along with any caveats:

For reviewers

Look out for any use of unstable language features in PRs, especially if any new #![feature] attributes have been added.

Using const generics

Status: Stub

Complete const generics are currently unstable. You can track their progress here.

Const generics are ok to use in public APIs, so long as they fit in the min_const_generics subset.

For reviewers

Look out for const operations on const generics in public APIs like:

pub fn extend_array<T, const N: usize, const M: usize>(arr: [T; N]) -> [T; N + 1] {

or for const generics that aren't integers, bools, or chars:

pub fn tag<const S: &'static str>() {

Using specialization

Specialization is currently unstable. You can track its progress here.

We try to avoid leaning on specialization too heavily, limiting its use to optimizing specific implementations. These specialized optimizations use a private trait to find the correct implementation, rather than specializing the public method itself. Any use of specialization that changes how methods are dispatched for external callers should be carefully considered.

As an example of how to use specialization in the standard library, consider the case of creating an Rc<[T]> from a &[T]:

impl<T: Clone> From<&[T]> for Rc<[T]> {
    fn from(v: &[T]) -> Rc<[T]> {
        unsafe { Self::from_iter_exact(v.iter().cloned(), v.len()) }

It would be nice to have an optimized implementation for the case where T: Copy:

impl<T: Copy> From<&[T]> for Rc<[T]> {
    fn from(v: &[T]) -> Rc<[T]> {
        unsafe { Self::copy_from_slice(v) }

Unfortunately we couldn't have both of these impls normally, because they'd overlap. This is where private specialization can be used to choose the right implementation internally. In this case, we use a trait called RcFromSlice that switches the implementation:

impl<T: Clone> From<&[T]> for Rc<[T]> {
    fn from(v: &[T]) -> Rc<[T]> {
        <Self as RcFromSlice<T>>::from_slice(v)

/// Specialization trait used for `From<&[T]>`.
trait RcFromSlice<T> {
    fn from_slice(slice: &[T]) -> Self;

impl<T: Clone> RcFromSlice<T> for Rc<[T]> {
    default fn from_slice(v: &[T]) -> Self {
        unsafe { Self::from_iter_exact(v.iter().cloned(), v.len()) }

impl<T: Copy> RcFromSlice<T> for Rc<[T]> {
    fn from_slice(v: &[T]) -> Self {
        unsafe { Self::copy_from_slice(v) }

Only specialization using the min_specialization feature should be used. The full specialization feature is known to be unsound.

For reviewers

Look out for any default annotations on public trait implementations. These will need to be refactored into a private dispatch trait. Also look out for uses of specialization that do more than pick a more optimized implementation.


Status: Stub

Changes to hot code might impact performance in consumers, for better or for worse. Appropriate benchmarks should give an idea of how performance characteristics change. For changes that affect rustc itself, you can also do a rust-timer run.

For reviewers

If a PR is focused on performance then try get some idea of what the impact is. Also consider marking the PR as rollup=never.

When to #[inline]

Inlining is a trade-off between potential execution speed, compile time and code size. There's some discussion about it in this PR to the hashbrown crate. From the thread:

#[inline] is very different than simply just an inline hint. As I mentioned before, there's no equivalent in C++ for what #[inline] does. In debug mode rustc basically ignores #[inline], pretending you didn't even write it. In release mode the compiler will, by default, codegen an #[inline] function into every single referencing codegen unit, and then it will also add inlinehint. This means that if you have 16 CGUs and they all reference an item, every single one is getting the entire item's implementation inlined into it.

You can add #[inline]:

  • To public, small, non-generic functions.

You shouldn't need #[inline]:

  • On methods that have any generics in scope.
  • On methods on traits that don't have a default implementation.

#[inline] can always be introduced later, so if you're in doubt they can just be removed.

What about #[inline(always)]?

You should just about never need #[inline(always)]. It may be beneficial for private helper methods that are used in a limited number of places or for trivial operators. A micro benchmark should justify the attribute.

For reviewers

#[inline] can always be added later, so if there's any debate about whether it's appropriate feel free to defer it by removing the annotations for a start.

doc alias policy

Rust's documentation supports adding aliases to any declaration (such as a function, type, or constant), using the syntax #[doc(alias = "name")]. We want to use doc aliases to help people find what they're looking for, while keeping those aliases maintainable and high-value. This policy outlines the cases where we add doc aliases, and the cases where we omit those aliases.

  • We must have a reasonable expectation that people might search for the term in the documentation search. Rust's documentation provides a name search, not a full-text search; as such, we expect that people may search for plausible names, but that for more general documentation searches they'll turn to a web search engine.
    • Related: we don't expect that people are currently searching Rust documentation for language-specific names from arbitrary languages they're familiar with, and we don't want to add that as a new documentation search feature; please don't add aliases based on your favorite language. Those mappings should live in separate guides or references. We do expect that people might look for the Rust name of a function they reasonably expect to exist in Rust (e.g. a system function or a C library function), to try to figure out what Rust called that function.
  • The proposed alias must be a name we would plausibly have used for the declaration. For instance, mkdir for create_dir, or rmdir for remove_dir, or popcnt and popcount for count_ones, or umask for mode. This feeds into the reasonable expectation that someone might search for the name and expect to find it ("what did Rust call mkdir").
  • There must be an obvious single target for the alias that is an exact analogue of the aliased name. We will not add the same alias to multiple declarations. (const and non-const versions of the same function are fine.) We will also not add an alias for a function that's only somewhat similar or related.
  • The alias must not conflict with the actual name of any existing declaration.
  • As a special case for stdarch, aliases from exact assembly instruction names to the corresponding intrinsic function are welcome, as long as they don't conflict with other names.

Tools and bots

Status: Stub


Status: Stub

PRs to the standard library aren’t merged manually using GitHub’s UI or by pushing remote branches. Everything goes through @bors.

You can approve a PR with:

@bors r+

Rolling up

For Libs PRs, rolling up is usually fine, in particular if it's only a new unstable addition or if it only touches docs. See the rollup guidelines for more details on when to rollup.


Status: Stub

You can kick off a performance test using @rust-timer:

@bors try @rust-timer queue


Status: Stub

Crater is a tool that can test PRs against a public subset of the Rust ecosystem to estimate the scale of potential breakage.

You can kick off a crater run by first calling:

@bors try

Once that finishes, you can then call:

@craterbot check

to ensure crates compile, or:

@craterbot run mode=build-and-test