Getting Started
Thank you for your interest in contributing to Rust! There are many ways to contribute, and we appreciate all of them.
If this is your first time contributing, the walkthrough chapter can give you a good example of how a typical contribution would go.
This documentation is not intended to be comprehensive; it is meant to be a quick guide for the most useful things. For more information, see this chapter on how to build and run the compiler.
Asking Questions
If you have questions, please make a post on the Rust Zulip server or
internals.rust-lang.org. If you are contributing to Rustup, be aware they are not on
Zulip - you can ask questions in #wg-rustup
on Discord.
See the list of teams and working groups and the Community page on the
official website for more resources.
As a reminder, all contributors are expected to follow our Code of Conduct.
The compiler team (or t-compiler
) usually hangs out in Zulip in this
"stream"; it will be easiest to get questions answered there.
Please ask questions! A lot of people report feeling that they are "wasting
expert time", but nobody on t-compiler
feels this way. Contributors are
important to us.
Also, if you feel comfortable, prefer public topics, as this means others can see the questions and answers, and perhaps even integrate them back into this guide :)
Experts
Not all t-compiler
members are experts on all parts of rustc
; it's a
pretty large project. To find out who could have some expertise on
different parts of the compiler, consult triagebot assign groups.
The sections that start with [assign*
in triagebot.toml
file.
But also, feel free to ask questions even if you can't figure out who to ping.
Another way to find experts for a given part of the compiler is to see who has made recent commits.
For example, to find people who have recently worked on name resolution since the 1.68.2 release,
you could run git shortlog -n 1.68.2.. compiler/rustc_resolve/
. Ignore any commits starting with
"Rollup merge" or commits by @bors
(see CI contribution procedures for
more information about these commits).
Etiquette
We do ask that you be mindful to include as much useful information as you can in your question, but we recognize this can be hard if you are unfamiliar with contributing to Rust.
Just pinging someone without providing any context can be a bit annoying and
just create noise, so we ask that you be mindful of the fact that the
t-compiler
folks get a lot of pings in a day.
What should I work on?
The Rust project is quite large and it can be difficult to know which parts of the project need help, or are a good starting place for beginners. Here are some suggested starting places.
Easy or mentored issues
If you're looking for somewhere to start, check out the following issue search. See the Triage for an explanation of these labels. You can also try filtering the search to areas you're interested in. For example:
repo:rust-lang/rust-clippy
will only show clippy issueslabel:T-compiler
will only show issues related to the compilerlabel:A-diagnostics
will only show diagnostic issues
Not all important or beginner work has issue labels. See below for how to find work that isn't labelled.
Recurring work
Some work is too large to be done by a single person. In this case, it's common to have "Tracking issues" to co-ordinate the work between contributors. Here are some example tracking issues where it's easy to pick up work without a large time commitment:
- Rustdoc Askama Migration
- Diagnostic Translation
- Move UI tests to subdirectories
- Port run-make tests from Make to Rust
If you find more recurring work, please feel free to add it here!
Clippy issues
The Clippy project has spent a long time making its contribution process as friendly to newcomers as possible. Consider working on it first to get familiar with the process and the compiler internals.
See the Clippy contribution guide for instructions on getting started.
Diagnostic issues
Many diagnostic issues are self-contained and don't need detailed background knowledge of the compiler. You can see a list of diagnostic issues here.
Picking up abandoned pull requests
Sometimes, contributors send a pull request, but later find out that they don't have enough
time to work on it, or they simply are not interested in it anymore. Such PRs are often
eventually closed and they receive the S-inactive
label. You could try to examine some of
these PRs and pick up the work. You can find the list of such PRs here.
If the PR has been implemented in some other way in the meantime, the S-inactive
label
should be removed from it. If not, and it seems that there is still interest in the change,
you can try to rebase the pull request on top of the latest master
branch and send a new
pull request, continuing the work on the feature.
Contributing to std (standard library)
See std-dev-guide.
Contributing code to other Rust projects
There are a bunch of other projects that you can contribute to outside of the
rust-lang/rust
repo, including cargo
, miri
, rustup
, and many others.
These repos might have their own contributing guidelines and procedures. Many of them are owned by working groups. For more info, see the documentation in those repos' READMEs.
Other ways to contribute
There are a bunch of other ways you can contribute, especially if you don't
feel comfortable jumping straight into the large rust-lang/rust
codebase.
The following tasks are doable without much background knowledge but are incredibly helpful:
- Cleanup crew: find minimal reproductions of ICEs, bisect regressions, etc. This is a way of helping that saves a ton of time for others to fix an error later.
- Writing documentation: if you are feeling a bit more intrepid, you could try to read a part of the code and write doc comments for it. This will help you to learn some part of the compiler while also producing a useful artifact!
- Triaging issues: categorizing, replicating, and minimizing issues is very helpful to the Rust maintainers.
- Working groups: there are a bunch of working groups on a wide variety of rust-related things.
- Answer questions in the Get Help! channels on the Rust Discord server, on users.rust-lang.org, or on StackOverflow.
- Participate in the RFC process.
- Find a requested community library, build it, and publish it to Crates.io. Easier said than done, but very, very valuable!
Cloning and Building
See "How to build and run the compiler".
Contributor Procedures
This section has moved to the "Contribution Procedures" chapter.
Other Resources
This section has moved to the "About this guide" chapter.
About this guide
This guide is meant to help document how rustc – the Rust compiler – works, as well as to help new contributors get involved in rustc development.
There are seven parts to this guide:
- Building
rustc
: Contains information that should be useful no matter how you are contributing, about building, debugging, profiling, etc. - Contributing to
rustc
: Contains information that should be useful no matter how you are contributing, about procedures for contribution, using git and Github, stabilizing features, etc. - High-Level Compiler Architecture: Discusses the high-level architecture of the compiler and stages of the compile process.
- Source Code Representation: Describes the process of taking raw source code from the user and transforming it into various forms that the compiler can work with easily.
- Analysis: discusses the analyses that the compiler uses to check various properties of the code and inform later stages of the compile process (e.g., type checking).
- From MIR to Binaries: How linked executable machine code is generated.
- Appendices at the end with useful reference information. There are a few of these with different information, including a glossary.
Constant change
Keep in mind that rustc
is a real production-quality product,
being worked upon continuously by a sizeable set of contributors.
As such, it has its fair share of codebase churn and technical debt.
In addition, many of the ideas discussed throughout this guide are idealized designs
that are not fully realized yet.
All this makes keeping this guide completely up to date on everything very hard!
The Guide itself is of course open-source as well, and the sources can be found at the GitHub repository. If you find any mistakes in the guide, please file an issue about it. Even better, open a PR with a correction!
If you do contribute to the guide, please see the corresponding subsection on writing documentation in this guide.
“‘All conditioned things are impermanent’ — when one sees this with wisdom, one turns away from suffering.” The Dhammapada, verse 277
Other places to find information
You might also find the following sites useful:
- This guide contains information about how various parts of the compiler work and how to contribute to the compiler.
- rustc API docs -- rustdoc documentation for the compiler, devtools, and internal tools
- Forge -- contains documentation about Rust infrastructure, team procedures, and more
- compiler-team -- the home-base for the Rust compiler team, with description of the team procedures, active working groups, and the team calendar.
- std-dev-guide -- a similar guide for developing the standard library.
- The t-compiler zulip
#contribute
and#wg-rustup
on Discord.- The Rust Internals forum, a place to ask questions and discuss Rust's internals
- The Rust reference, even though it doesn't specifically talk about Rust's internals, is a great resource nonetheless
- Although out of date, Tom Lee's great blog article is very helpful
- rustaceans.org is helpful, but mostly dedicated to IRC
- The Rust Compiler Testing Docs
- For @bors, this cheat sheet is helpful
- Google is always helpful when programming. You can search all Rust documentation (the standard library, the compiler, the books, the references, and the guides) to quickly find information about the language and compiler.
- You can also use Rustdoc's built-in search feature to find documentation on
types and functions within the crates you're looking at. You can also search
by type signature! For example, searching for
* -> vec
should find all functions that return aVec<T>
. Hint: Find more tips and keyboard shortcuts by typing?
on any Rustdoc page!
How to build and run the compiler
- Quick Start
- Get the source code
- What is
x.py
? - Create a
config.toml
- Common
x
commands - Creating a rustup toolchain
- Building targets for cross-compilation
- Other
x
commands - Remarks on disk space
The compiler is built using a tool called x.py
. You will need to
have Python installed to run it.
Quick Start
For a less in-depth quick-start of getting the compiler running, see quickstart.
Get the source code
The main repository is rust-lang/rust
. This contains the compiler,
the standard library (including core
, alloc
, test
, proc_macro
, etc),
and a bunch of tools (e.g. rustdoc
, the bootstrapping infrastructure, etc).
The very first step to work on rustc
is to clone the repository:
git clone https://github.com/rust-lang/rust.git
cd rust
Partial clone the repository
Due to the size of the repository, cloning on a slower internet connection can take a long time, and requires disk space to store the full history of every file and directory. Instead, it is possible to tell git to perform a partial clone, which will only fully retrieve the current file contents, but will automatically retrieve further file contents when you, e.g., jump back in the history. All git commands will continue to work as usual, at the price of requiring an internet connection to visit not-yet-loaded points in history.
git clone --filter='blob:none' https://github.com/rust-lang/rust.git
cd rust
NOTE: This link describes this type of checkout in more detail, and also compares it to other modes, such as shallow cloning.
Shallow clone the repository
An older alternative to partial clones is to use shallow clone the repository instead.
To do so, you can use the --depth N
option with the git clone
command.
This instructs git
to perform a "shallow clone", cloning the repository but truncating it to
the last N
commits.
Passing --depth 1
tells git
to clone the repository but truncate the history to the latest
commit that is on the master
branch, which is usually fine for browsing the source code or
building the compiler.
git clone --depth 1 https://github.com/rust-lang/rust.git
cd rust
NOTE: A shallow clone limits which
git
commands can be run. If you intend to work on and contribute to the compiler, it is generally recommended to fully clone the repository as shown above, or to perform a partial clone instead.For example,
git bisect
andgit blame
require access to the commit history, so they don't work if the repository was cloned with--depth 1
.
What is x.py
?
x.py
is the build tool for the rust
repository. It can build docs, run tests, and compile the
compiler and standard library.
This chapter focuses on the basics to be productive, but
if you want to learn more about x.py
, read this chapter.
Also, using x
rather than x.py
is recommended as:
./x
is the most likely to work on every system (on Unix it runs the shell script that does python version detection, on Windows it will probably run the powershell script - certainly less likely to break than./x.py
which often just opens the file in an editor).1
(You can find the platform related scripts around the x.py
, like x.ps1
)
Notice that this is not absolute. For instance, using Nushell in VSCode on Win10,
typing x
or ./x
still opens x.py
in an editor rather than invoking the program. :)
In the rest of this guide, we use x
rather than x.py
directly. The following
command:
./x check
could be replaced by:
./x.py check
Running x.py
The x.py
command can be run directly on most Unix systems in the following format:
./x <subcommand> [flags]
This is how the documentation and examples assume you are running x.py
.
Some alternative ways are:
# On a Unix shell if you don't have the necessary `python3` command
./x <subcommand> [flags]
# In Windows Powershell (if powershell is configured to run scripts)
./x <subcommand> [flags]
./x.ps1 <subcommand> [flags]
# On the Windows Command Prompt (if .py files are configured to run Python)
x.py <subcommand> [flags]
# You can also run Python yourself, e.g.:
python x.py <subcommand> [flags]
On Windows, the Powershell commands may give you an error that looks like this:
PS C:\Users\vboxuser\rust> ./x
./x : File C:\Users\vboxuser\rust\x.ps1 cannot be loaded because running scripts is disabled on this system. For more
information, see about_Execution_Policies at https:/go.microsoft.com/fwlink/?LinkID=135170.
At line:1 char:1
+ ./x
+ ~~~
+ CategoryInfo : SecurityError: (:) [], PSSecurityException
+ FullyQualifiedErrorId : UnauthorizedAccess
You can avoid this error by allowing powershell to run local scripts:
Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser
Running x.py
slightly more conveniently
There is a binary that wraps x.py
called x
in src/tools/x
. All it does is
run x.py
, but it can be installed system-wide and run from any subdirectory
of a checkout. It also looks up the appropriate version of python
to use.
You can install it with cargo install --path src/tools/x
.
To clarify that this is another global installed binary util, which is
similar to the one declared in section What is x.py
, but
it works as an independent process to execute the x.py
rather than calling the
shell to run the platform related scripts.
Create a config.toml
To start, run ./x setup
and select the compiler
defaults. This will do some initialization
and create a config.toml
for you with reasonable defaults. If you use a different default (which
you'll likely want to do if you want to contribute to an area of rust other than the compiler, such
as rustdoc), make sure to read information about that default (located in src/bootstrap/defaults
)
as the build process may be different for other defaults.
Alternatively, you can write config.toml
by hand. See config.example.toml
for all the available
settings and explanations of them. See src/bootstrap/defaults
for common settings to change.
If you have already built rustc
and you change settings related to LLVM, then you may have to
execute rm -rf build
for subsequent configuration changes to take effect. Note that ./x clean
will not cause a rebuild of LLVM.
Common x
commands
Here are the basic invocations of the x
commands most commonly used when
working on rustc
, std
, rustdoc
, and other tools.
Command | When to use it |
---|---|
./x check | Quick check to see if most things compile; rust-analyzer can run this automatically for you |
./x build | Builds rustc , std , and rustdoc |
./x test | Runs all tests |
./x fmt | Formats all code |
As written, these commands are reasonable starting points. However, there are
additional options and arguments for each of them that are worth learning for
serious development work. In particular, ./x build
and ./x test
provide many ways to compile or test a subset of the code, which can save a lot
of time.
Also, note that x
supports all kinds of path suffixes for compiler
, library
,
and src/tools
directories. So, you can simply run x test tidy
instead of
x test src/tools/tidy
. Or, x build std
instead of x build library/std
.
See the chapters on testing and rustdoc for more details.
Building the compiler
Note that building will require a relatively large amount of storage space. You may want to have upwards of 10 or 15 gigabytes available to build the compiler.
Once you've created a config.toml
, you are now ready to run
x
. There are a lot of options here, but let's start with what is
probably the best "go to" command for building a local compiler:
./x build library
This may look like it only builds the standard library, but that is not the case. What this command does is the following:
- Build
std
using the stage0 compiler - Build
rustc
using the stage0 compiler- This produces the stage1 compiler
- Build
std
using the stage1 compiler
This final product (stage1 compiler + libs built using that compiler)
is what you need to build other Rust programs (unless you use #![no_std]
or
#![no_core]
).
You will probably find that building the stage1 std
is a bottleneck for you,
but fear not, there is a (hacky) workaround...
see the section on avoiding rebuilds for std.
Sometimes you don't need a full build. When doing some kind of
"type-based refactoring", like renaming a method, or changing the
signature of some function, you can use ./x check
instead for a much faster build.
Note that this whole command just gives you a subset of the full rustc
build. The full rustc
build (what you get with ./x build --stage 2 compiler/rustc
) has quite a few more steps:
- Build
rustc
with the stage1 compiler.- The resulting compiler here is called the "stage2" compiler.
- Build
std
with stage2 compiler. - Build
librustdoc
and a bunch of other things with the stage2 compiler.
You almost never need to do this.
Build specific components
If you are working on the standard library, you probably don't need to build the compiler unless you are planning to use a recently added nightly feature. Instead, you can just build using the bootstrap compiler.
./x build --stage 0 library
If you choose the library
profile when running x setup
, you can omit --stage 0
(it's the
default).
Creating a rustup toolchain
Once you have successfully built rustc
, you will have created a bunch
of files in your build
directory. In order to actually run the
resulting rustc
, we recommend creating rustup toolchains. The first
one will run the stage1 compiler (which we built above). The second
will execute the stage2 compiler (which we did not build, but which
you will likely need to build at some point; for example, if you want
to run the entire test suite).
rustup toolchain link stage0 build/host/stage0-sysroot # beta compiler + stage0 std
rustup toolchain link stage1 build/host/stage1
rustup toolchain link stage2 build/host/stage2
Now you can run the rustc
you built with. If you run with -vV
, you
should see a version number ending in -dev
, indicating a build from
your local environment:
$ rustc +stage1 -vV
rustc 1.48.0-dev
binary: rustc
commit-hash: unknown
commit-date: unknown
host: x86_64-unknown-linux-gnu
release: 1.48.0-dev
LLVM version: 11.0
The rustup toolchain points to the specified toolchain compiled in your build
directory,
so the rustup toolchain will be updated whenever x build
or x test
are run for
that toolchain/stage.
Note: the toolchain we've built does not include cargo
. In this case, rustup
will
fall back to using cargo
from the installed nightly
, beta
, or stable
toolchain
(in that order). If you need to use unstable cargo
flags, be sure to run
rustup install nightly
if you haven't already. See the
rustup documentation on custom toolchains.
Note: rust-analyzer and IntelliJ Rust plugin use a component called
rust-analyzer-proc-macro-srv
to work with proc macros. If you intend to use a
custom toolchain for a project (e.g. via rustup override set stage1
) you may
want to build this component:
./x build proc-macro-srv-cli
Building targets for cross-compilation
To produce a compiler that can cross-compile for other targets,
pass any number of target
flags to x build
.
For example, if your host platform is x86_64-unknown-linux-gnu
and your cross-compilation target is wasm32-wasip1
, you can build with:
./x build --target x86_64-unknown-linux-gnu,wasm32-wasip1
Note that if you want the resulting compiler to be able to build crates that
involve proc macros or build scripts, you must be sure to explicitly build target support for the
host platform (in this case, x86_64-unknown-linux-gnu
).
If you want to always build for other targets without needing to pass flags to x build
,
you can configure this in the [build]
section of your config.toml
like so:
[build]
target = ["x86_64-unknown-linux-gnu", "wasm32-wasip1"]
Note that building for some targets requires having external dependencies installed
(e.g. building musl targets requires a local copy of musl).
Any target-specific configuration (e.g. the path to a local copy of musl)
will need to be provided by your config.toml
.
Please see config.example.toml
for information on target-specific configuration keys.
For examples of the complete configuration necessary to build a target, please visit the rustc book, select any target under the "Platform Support" heading on the left, and see the section related to building a compiler for that target. For targets without a corresponding page in the rustc book, it may be useful to inspect the Dockerfiles that the Rust infrastructure itself uses to set up and configure cross-compilation.
If you have followed the directions from the prior section on creating a rustup toolchain, then once you have built your compiler you will be able to use it to cross-compile like so:
cargo +stage1 build --target wasm32-wasip1
Other x
commands
Here are a few other useful x
commands. We'll cover some of them in detail
in other sections:
- Building things:
./x build
– builds everything using the stage 1 compiler, not just up tostd
./x build --stage 2
– builds everything with the stage 2 compiler includingrustdoc
- Running tests (see the section on running tests for
more details):
./x test library/std
– runs the unit tests and integration tests fromstd
./x test tests/ui
– runs theui
test suite./x test tests/ui/const-generics
- runs all the tests in theconst-generics/
subdirectory of theui
test suite./x test tests/ui/const-generics/const-types.rs
- runs the single testconst-types.rs
from theui
test suite
Cleaning out build directories
Sometimes you need to start fresh, but this is normally not the case. If you need to run this then bootstrap is most likely not acting right and you should file a bug as to what is going wrong. If you do need to clean everything up then you only need to run one command!
./x clean
rm -rf build
works too, but then you have to rebuild LLVM, which can take
a long time even on fast computers.
Remarks on disk space
Building the compiler (especially if beyond stage 1) can require significant amounts of free disk
space, possibly around 100GB. This is compounded if you have a separate build directory for
rust-analyzer (e.g. build-rust-analyzer
). This is easy to hit with dev-desktops which have a set
disk
quota
for each user, but this also applies to local development as well. Occassionally, you may need to:
- Remove
build/
directory. - Remove
build-rust-analyzer/
directory (if you have a separate rust-analyzer build directory). - Uninstall unnecessary toolchains if you use
cargo-bisect-rustc
. You can check which toolchains are installed withrustup toolchain list
.
issue#1707
Quickstart
This is a quickstart guide about getting the compiler running. For more information on the individual steps, see the other pages in this chapter.
First, clone the repository:
git clone https://github.com/rust-lang/rust.git
cd rust
When building the compiler, we don't use cargo
directly, instead we use a
wrapper called "x". It is invoked with ./x
.
We need to create a configuration for the build. Use ./x setup
to create a
good default.
./x setup
Then, we can build the compiler. Use ./x build
to build the compiler, standard
library and a few tools. You can also ./x check
to just check it. All these
commands can take specific components/paths as arguments, for example ./x check compiler
to just check the compiler.
./x build
When doing a change to the compiler that does not affect the way it compiles the standard library (so for example, a change to an error message), use
--keep-stage-std 1
to avoid recompiling it.
After building the compiler and standard library, you now have a working compiler toolchain. You can use it with rustup by linking it.
rustup toolchain link stage1 build/host/stage1
Now you have a toolchain called stage1
linked to your build. You can use it to
test the compiler.
rustc +stage1 testfile.rs
After doing a change, you can run the compiler test suite with ./x test
.
./x test
runs the full test suite, which is slow and rarely what you want.
Usually, ./x test tests/ui
is what you want after a compiler change, testing
all UI tests that invoke the compiler on a specific test file
and check the output.
./x test tests/ui
Use --bless
if you've made a change and want to update the .stderr
files
with the new output.
./x suggest
can also be helpful for suggesting which tests to run after a change.
Congrats, you are now ready to make a change to the compiler! If you have more questions, the full chapter might contain the answers, and if it doesn't, feel free to ask for help on Zulip.
If you use VSCode, Vim, Emacs or Helix, ./x setup
will ask you if you want to
set up the editor config. For more information, check out suggested
workflows.
Prerequisites
Dependencies
See the rust-lang/rust
INSTALL.
Hardware
You will need an internet connection to build. The bootstrapping process involves updating git submodules and downloading a beta compiler. It doesn't need to be super fast, but that can help.
There are no strict hardware requirements, but building the compiler is computationally expensive, so a beefier machine will help, and I wouldn't recommend trying to build on a Raspberry Pi! We recommend the following.
- 30GB+ of free disk space. Otherwise, you will have to keep clearing incremental caches. More space is better, the compiler is a bit of a hog; it's a problem we are aware of.
- 8GB+ RAM
- 2+ cores. Having more cores really helps. 10 or 20 or more is not too many!
Beefier machines will lead to much faster builds. If your machine is not very
powerful, a common strategy is to only use ./x check
on your local machine
and let the CI build test your changes when you push to a PR branch.
Building the compiler takes more than half an hour on my moderately powerful laptop. We suggest downloading LLVM from CI so you don't have to build it from source (see here).
Like cargo
, the build system will use as many cores as possible. Sometimes
this can cause you to run low on memory. You can use -j
to adjust the number
of concurrent jobs. If a full build takes more than ~45 minutes to an hour, you
are probably spending most of the time swapping memory in and out; try using
-j1
.
If you don't have too much free disk space, you may want to turn off incremental compilation (see here). This will make compilation take longer (especially after a rebase), but will save a ton of space from the incremental caches.
Suggested Workflows
The full bootstrapping process takes quite a while. Here are some suggestions to make your life easier.
- Installing a pre-push hook
- Configuring
rust-analyzer
forrustc
- Check, check, and check again
x suggest
- Configuring
rustup
to use nightly - Faster builds with
--keep-stage
. - Using incremental compilation
- Fine-tuning optimizations
- Working on multiple branches at the same time
- Using nix-shell
- Shell Completions
Installing a pre-push hook
CI will automatically fail your build if it doesn't pass tidy
, our internal
tool for ensuring code quality. If you'd like, you can install a Git
hook that will
automatically run ./x test tidy
on each push, to ensure your code is up to
par. If the hook fails then run ./x test tidy --bless
and commit the changes.
If you decide later that the pre-push behavior is undesirable, you can delete
the pre-push
file in .git/hooks
.
A prebuilt git hook lives at src/etc/pre-push.sh
. It can be copied into
your .git/hooks
folder as pre-push
(without the .sh
extension!).
You can also install the hook as a step of running ./x setup
!
Configuring rust-analyzer
for rustc
Project-local rust-analyzer setup
rust-analyzer
can help you check and format your code whenever you save a
file. By default, rust-analyzer
runs the cargo check
and rustfmt
commands,
but you can override these commands to use more adapted versions of these tools
when hacking on rustc
. With custom setup, rust-analyzer
can use ./x check
to check the sources, and the stage 0 rustfmt to format them.
The default rust-analyzer.check.overrideCommand
command line will check all
the crates and tools in the repository. If you are working on a specific part,
you can override the command to only check the part you are working on to save
checking time. For example, if you are working on the compiler, you can override
the command to x check compiler --json-output
to only check the compiler part.
You can run x check --help --verbose
to see the available parts.
Running ./x setup editor
will prompt you to create a project-local LSP config
file for one of the supported editors. You can also create the config file as a
step of running ./x setup
.
Using a separate build directory for rust-analyzer
By default, when rust-analyzer runs a check or format command, it will share the same build directory as manual command-line builds. This can be inconvenient for two reasons:
- Each build will lock the build directory and force the other to wait, so it becomes impossible to run command-line builds while rust-analyzer is running commands in the background.
- There is an increased risk of one of the builds deleting previously-built artifacts due to conflicting compiler flags or other settings, forcing additional rebuilds in some cases.
To avoid these problems:
- Add
--build-dir=build-rust-analyzer
to all of the customx
commands in your editor's rust-analyzer configuration. (Feel free to choose a different directory name if desired.) - Modify the
rust-analyzer.rustfmt.overrideCommand
setting so that it points to the copy ofrustfmt
in that other build directory. - Modify the
rust-analyzer.procMacro.server
setting so that it points to the copy ofrust-analyzer-proc-macro-srv
in that other build directory.
Using separate build directories for command-line builds and rust-analyzer
requires extra disk space, and also means that running ./x clean
on the
command-line will not clean out the separate build directory. To clean the
separate build directory, run ./x clean --build-dir=build-rust-analyzer
instead.
Visual Studio Code
Selecting vscode
in ./x setup editor
will prompt you to create a
.vscode/settings.json
file which will configure Visual Studio code. The
recommended rust-analyzer
settings live at
src/etc/rust_analyzer_settings.json
.
If running ./x check
on save is inconvenient, in VS Code you can use a Build
Task instead:
// .vscode/tasks.json
{
"version": "2.0.0",
"tasks": [
{
"label": "./x check",
"command": "./x check",
"type": "shell",
"problemMatcher": "$rustc",
"presentation": { "clear": true },
"group": { "kind": "build", "isDefault": true }
}
]
}
Neovim
For Neovim users there are several options for configuring for rustc. The easiest way is by using neoconf.nvim, which allows for project-local configuration files with the native LSP. The steps for how to use it are below. Note that they require rust-analyzer to already be configured with Neovim. Steps for this can be found here.
- First install the plugin. This can be done by following the steps in the README.
- Run
./x setup editor
, and selectvscode
to create a.vscode/settings.json
file.neoconf
is able to read and update rust-analyzer settings automatically when the project is opened when this file is detected.
If you're using coc.nvim
, you can run ./x setup editor
and select vim
to
create a .vim/coc-settings.json
. The settings can be edited with
:CocLocalConfig
. The recommended settings live at
src/etc/rust_analyzer_settings.json
.
Another way is without a plugin, and creating your own logic in your
configuration. To do this you must translate the JSON to Lua yourself. The
translation is 1:1 and fairly straight-forward. It must be put in the
["rust-analyzer"]
key of the setup table, which is shown
here.
If you would like to use the build task that is described above, you may either
make your own command in your config, or you can install a plugin such as
overseer.nvim that can read
VSCode's task.json
files,
and follow the same instructions as above.
Emacs
Emacs provides support for rust-analyzer with project-local configuration
through Eglot.
Steps for setting up Eglot with rust-analyzer can be found
here.
Having set up Emacs & Eglot for Rust development in general, you can run
./x setup editor
and select emacs
, which will prompt you to create
.dir-locals.el
with the recommended configuration for Eglot.
The recommended settings live at src/etc/rust_analyzer_eglot.el
.
For more information on project-specific Eglot configuration, consult the
manual.
Helix
Helix comes with built-in LSP and rust-analyzer support.
It can be configured through languages.toml
, as described
here.
You can run ./x setup editor
and select helix
, which will prompt you to
create languages.toml
with the recommended configuration for Helix. The
recommended settings live at src/etc/rust_analyzer_helix.toml
.
Check, check, and check again
When doing simple refactoring, it can be useful to run ./x check
continuously. If you set up rust-analyzer
as described above, this will be
done for you every time you save a file. Here you are just checking that the
compiler can build, but often that is all you need (e.g., when renaming a
method). You can then run ./x build
when you actually need to run tests.
In fact, it is sometimes useful to put off tests even when you are not 100% sure
the code will work. You can then keep building up refactoring commits and only
run the tests at some later time. You can then use git bisect
to track down
precisely which commit caused the problem. A nice side-effect of this style
is that you are left with a fairly fine-grained set of commits at the end, all
of which build and pass tests. This often helps reviewing.
x suggest
The x suggest
subcommand suggests (and runs) a subset of the extensive
rust-lang/rust
tests based on files you have changed. This is especially
useful for new contributors who have not mastered the arcane x
flags yet and
more experienced contributors as a shorthand for reducing mental effort. In all
cases it is useful not to run the full tests (which can take on the order of
tens of minutes) and just run a subset which are relevant to your changes. For
example, running tidy
and linkchecker
is useful when editing Markdown files,
whereas UI tests are much less likely to be helpful. While x suggest
is a
useful tool, it does not guarantee perfect coverage (just as PR CI isn't a
substitute for bors). See the dedicated chapter for
more information and contribution instructions.
Please note that x suggest
is in a beta state currently and the tests that it
will suggest are limited.
Configuring rustup
to use nightly
Some parts of the bootstrap process uses pinned, nightly versions of tools like
rustfmt. To make things like cargo fmt
work correctly in your repo, run
cd <path to rustc repo>
rustup override set nightly
after installing a nightly toolchain with rustup
. Don't forget to do this
for all directories you have setup a worktree for. You may need to use the
pinned nightly version from src/stage0
, but often the normal nightly
channel
will work.
Note see the section on vscode for how to configure it with this real
rustfmt x
uses, and the section on rustup for how to setup rustup
toolchain for your bootstrapped compiler
Note This does not allow you to build rustc
with cargo directly. You
still have to use x
to work on the compiler or standard library, this just
lets you use cargo fmt
.
Faster builds with --keep-stage
.
Sometimes just checking whether the compiler builds is not enough. A common
example is that you need to add a debug!
statement to inspect the value of
some state or better understand the problem. In that case, you don't really need
a full build. By bypassing bootstrap's cache invalidation, you can often get
these builds to complete very fast (e.g., around 30 seconds). The only catch is
this requires a bit of fudging and may produce compilers that don't work (but
that is easily detected and fixed).
The sequence of commands you want is as follows:
- Initial build:
./x build library
- As documented previously, this will build a functional stage1 compiler as
part of running all stage0 commands (which include building a
std
compatible with the stage1 compiler) as well as the first few steps of the "stage 1 actions" up to "stage1 (sysroot stage1) builds std".
- As documented previously, this will build a functional stage1 compiler as
part of running all stage0 commands (which include building a
- Subsequent builds:
./x build library --keep-stage 1
- Note that we added the
--keep-stage 1
flag here
- Note that we added the
As mentioned, the effect of --keep-stage 1
is that we just assume that the
old standard library can be re-used. If you are editing the compiler, this is
almost always true: you haven't changed the standard library, after all. But
sometimes, it's not true: for example, if you are editing the "metadata" part of
the compiler, which controls how the compiler encodes types and other states
into the rlib
files, or if you are editing things that wind up in the metadata
(such as the definition of the MIR).
The TL;DR is that you might get weird behavior from a compile when using
--keep-stage 1
-- for example, strange ICEs
or other panics. In that case, you should simply remove the --keep-stage 1
from the command and rebuild. That ought to fix the problem.
You can also use --keep-stage 1
when running tests. Something like this:
- Initial test run:
./x test tests/ui
- Subsequent test run:
./x test tests/ui --keep-stage 1
Using incremental compilation
You can further enable the --incremental
flag to save additional time in
subsequent rebuilds:
./x test tests/ui --incremental --test-args issue-1234
If you don't want to include the flag with every command, you can enable it in
the config.toml
:
[rust]
incremental = true
Note that incremental compilation will use more disk space than usual. If disk
space is a concern for you, you might want to check the size of the build
directory from time to time.
Fine-tuning optimizations
Setting optimize = false
makes the compiler too slow for tests. However, to
improve the test cycle, you can disable optimizations selectively only for the
crates you'll have to rebuild
(source).
For example, when working on rustc_mir_build
, the rustc_mir_build
and
rustc_driver
crates take the most time to incrementally rebuild. You could
therefore set the following in the root Cargo.toml
:
[profile.release.package.rustc_mir_build]
opt-level = 0
[profile.release.package.rustc_driver]
opt-level = 0
Working on multiple branches at the same time
Working on multiple branches in parallel can be a little annoying, since building the compiler on one branch will cause the old build and the incremental compilation cache to be overwritten. One solution would be to have multiple clones of the repository, but that would mean storing the Git metadata multiple times, and having to update each clone individually.
Fortunately, Git has a better solution called worktrees. This lets you create multiple "working trees", which all share the same Git database. Moreover, because all of the worktrees share the same object database, if you update a branch (e.g. master) in any of them, you can use the new commits from any of the worktrees. One caveat, though, is that submodules do not get shared. They will still be cloned multiple times.
Given you are inside the root directory for your Rust repository, you can create a "linked working tree" in a new "rust2" directory by running the following command:
git worktree add ../rust2
Creating a new worktree for a new branch based on master
looks like:
git worktree add -b my-feature ../rust2 master
You can then use that rust2 folder as a separate workspace for modifying and
building rustc
!
Using nix-shell
If you're using nix, you can use the following nix-shell to work on Rust:
{ pkgs ? import <nixpkgs> {} }:
pkgs.mkShell {
name = "rustc";
nativeBuildInputs = with pkgs; [
binutils cmake ninja pkg-config python3 git curl cacert patchelf nix
];
buildInputs = with pkgs; [
openssl glibc.out glibc.static
];
# Avoid creating text files for ICEs.
RUSTC_ICE = "0";
# Provide `libstdc++.so.6` for the self-contained lld.
LD_LIBRARY_PATH = "${with pkgs; lib.makeLibraryPath [
stdenv.cc.cc.lib
]}";
}
Note that when using nix on a not-NixOS distribution, it may be necessary to set
patch-binaries-for-nix = true
in config.toml
. Bootstrap tries to detect
whether it's running in nix and enable patching automatically, but this
detection can have false negatives.
You can also use your nix shell to manage config.toml
:
let
config = pkgs.writeText "rustc-config" ''
# Your config.toml content goes here
''
pkgs.mkShell {
/* ... */
# This environment variable tells bootstrap where our config.toml is.
RUST_BOOTSTRAP_CONFIG = config;
}
Shell Completions
If you use Bash, Fish or PowerShell, you can find automatically-generated shell
completion scripts for x.py
in
src/etc/completions
.
Zsh support will also be included once issues with
clap_complete
have been resolved.
You can use source ./src/etc/completions/x.py.<extension>
to load completions
for your shell of choice, or & .\src\etc\completions\x.py.ps1
for PowerShell.
Adding this to your shell's startup script (e.g. .bashrc
) will automatically
load this completion.
Build distribution artifacts
You might want to build and package up the compiler for distribution. You’ll want to run this command to do it:
./x dist
Install from source
You might want to prefer installing Rust (and tools configured in your configuration) by building from source. If so, you want to run this command:
./x install
Note: If you are testing out a modification to a compiler, you might
want to build the compiler (with ./x build
) then create a toolchain as
discussed in here.
For example, if the toolchain you created is called "foo", you would then
invoke it with rustc +foo ...
(where ... represents the rest of the arguments).
Instead of installing Rust (and tools in your config file) globally, you can set DESTDIR
environment variable to change the installation path. If you want to set installation paths
more dynamically, you should prefer install options in your config file to achieve that.
Building documentation
This chapter describes how to build documentation of toolchain components, like the standard library (std) or the compiler (rustc).
-
Document everything
This uses
rustdoc
from the beta toolchain, so will produce (slightly) different output to stage 1 rustdoc, as rustdoc is under active development:./x doc
If you want to be sure the documentation looks the same as on CI:
./x doc --stage 1
This ensures that (current) rustdoc gets built, then that is used to document the components.
-
Much like running individual tests or building specific components, you can build just the documentation you want:
./x doc src/doc/book ./x doc src/doc/nomicon ./x doc compiler library
See the nightly docs index page for a full list of books.
-
Document internal rustc items
Compiler documentation is not built by default. To create it by default with
x doc
, modifyconfig.toml
:[build] compiler-docs = true
Note that when enabled, documentation for internal compiler items will also be built.
NOTE: The documentation for the compiler is found at this link.
Rustdoc overview
rustdoc
lives in-tree with the
compiler and standard library. This chapter is about how it works.
For information about Rustdoc's features and how to use them, see
the Rustdoc book.
For more details about how rustdoc works, see the
"Rustdoc internals" chapter.
rustdoc
uses rustc
internals (and, of course, the standard library), so you
will have to build the compiler and std
once before you can build rustdoc
.
Rustdoc is implemented entirely within the crate librustdoc
. It runs
the compiler up to the point where we have an internal representation of a
crate (HIR) and the ability to run some queries about the types of items. HIR
and queries are discussed in the linked chapters.
librustdoc
performs two major steps after that to render a set of
documentation:
- "Clean" the AST into a form that's more suited to creating documentation (and slightly more resistant to churn in the compiler).
- Use this cleaned AST to render a crate's documentation, one page at a time.
Naturally, there's more than just this, and those descriptions simplify out lots of details, but that's the high-level overview.
(Side note: librustdoc
is a library crate! The rustdoc
binary is created
using the project in src/tools/rustdoc
. Note that literally all that
does is call the main()
that's in this crate's lib.rs
, though.)
Cheat sheet
- Run
./x setup tools
before getting started. This will configurex
with nice settings for developing rustdoc and other tools, including downloading a copy of rustc rather than building it. - Use
./x check rustdoc
to quickly check for compile errors. - Use
./x build library rustdoc
to make a usable rustdoc you can run on other projects.- Add
library/test
to be able to userustdoc --test
. - Run
rustup toolchain link stage2 build/host/stage2
to add a custom toolchain calledstage2
to your rustup environment. After running that,cargo +stage2 doc
in any directory will build with your locally-compiled rustdoc.
- Add
- Use
./x doc library
to use this rustdoc to generate the standard library docs.- The completed docs will be available in
build/host/doc
(undercore
,alloc
, andstd
). - If you want to copy those docs to a webserver, copy all of
build/host/doc
, since that's where the CSS, JS, fonts, and landing page are.
- The completed docs will be available in
- Use
./x test tests/rustdoc*
to run the tests using a stage1 rustdoc.- See Rustdoc internals for more information about tests.
Code structure
- All paths in this section are relative to
src/librustdoc
in the rust-lang/rust repository. - Most of the HTML printing code is in
html/format.rs
andhtml/render/mod.rs
. It's in a bunch offmt::Display
implementations and supplementary functions. - The types that got
Display
impls above are defined inclean/mod.rs
, right next to the customClean
trait used to process them out of the rustc HIR. - The bits specific to using rustdoc as a test harness are in
doctest.rs
. - The Markdown renderer is loaded up in
html/markdown.rs
, including functions for extracting doctests from a given block of Markdown. - The tests on the structure of rustdoc HTML output are located in
tests/rustdoc
, where they're handled by the test runner of bootstrap and the supplementary scriptsrc/etc/htmldocck.py
.
Tests
- All paths in this section are relative to
tests
in the rust-lang/rust repository. - Tests on search engine and index are located in
rustdoc-js
andrustdoc-js-std
. The format is specified in the search guide. - Tests on the "UI" of rustdoc (the terminal output it produces when run) are in
rustdoc-ui
- Tests on the "GUI" of rustdoc (the HTML, JS, and CSS as rendered in a browser)
are in
rustdoc-gui
. These use a NodeJS tool called browser-UI-test that uses puppeteer to run tests in a headless browser and check rendering and interactivity.
Constraints
We try to make rustdoc work reasonably well with JavaScript disabled, and when browsing local files. We support these browsers.
Supporting local files (file:///
URLs) brings some surprising restrictions.
Certain browser features that require secure origins, like localStorage
and
Service Workers, don't work reliably. We can still use such features but we
should make sure pages are still usable without them.
Multiple runs, same output directory
Rustdoc can be run multiple times for varying inputs, with its output set to the same directory. That's how cargo produces documentation for dependencies of the current crate. It can also be done manually if a user wants a big documentation bundle with all of the docs they care about.
HTML is generated independently for each crate, but there is some cross-crate information that we update as we add crates to the output directory:
crates<SUFFIX>.js
holds a list of all crates in the output directory.search-index<SUFFIX>.js
holds a list of all searchable items.- For each trait, there is a file under
implementors/.../trait.TraitName.js
containing a list of implementors of that trait. The implementors may be in different crates than the trait, and the JS file is updated as we discover new ones.
Use cases
There are a few major use cases for rustdoc that you should keep in mind when working on it:
Standard library docs
These are published at https://doc.rust-lang.org/std as part of the Rust release process. Stable releases are also uploaded to specific versioned URLs like https://doc.rust-lang.org/1.57.0/std/. Beta and nightly docs are published to https://doc.rust-lang.org/beta/std/ and https://doc.rust-lang.org/nightly/std/. The docs are uploaded with the promote-release tool and served from S3 with CloudFront.
The standard library docs contain five crates: alloc, core, proc_macro, std, and test.
docs.rs
When crates are published to crates.io, docs.rs automatically builds and publishes their documentation, for instance at https://docs.rs/serde/latest/serde/. It always builds with the current nightly rustdoc, so any changes you land in rustdoc are "insta-stable" in that they will have an immediate public effect on docs.rs. Old documentation is not rebuilt, so you will see some variation in UI when browsing old releases in docs.rs. Crate authors can request rebuilds, which will be run with the latest rustdoc.
Docs.rs performs some transformations on rustdoc's output in order to save
storage and display a navigation bar at the top. In particular, certain static
files, like main.js and rustdoc.css, may be shared across multiple invocations
of the same version of rustdoc. Others, like crates.js and sidebar-items.js, are
different for different invocations. Still others, like fonts, will never
change. These categories are distinguished using the SharedResource
enum in
src/librustdoc/html/render/write_shared.rs
Documentation on docs.rs is always generated for a single crate at a time, so the search and sidebar functionality don't include dependencies of the current crate.
Locally generated docs
Crate authors can run cargo doc --open
in crates they have checked
out locally to see the docs. This is useful to check that the docs they
are writing are useful and display correctly. It can also be useful for
people to view documentation on crates they aren't authors of, but want to
use. In both cases, people may use --document-private-items
Cargo flag to
see private methods, fields, and so on, which are normally not displayed.
By default cargo doc
will generate documentation for a crate and all of its
dependencies. That can result in a very large documentation bundle, with a large
(and slow) search corpus. The Cargo flag --no-deps
inhibits that behavior and
generates docs for just the crate.
Self-hosted project docs
Some projects like to host their own documentation. For example:
https://docs.serde.rs/. This is easy to do by locally generating docs, and
simply copying them to a web server. Rustdoc's HTML output can be extensively
customized by flags. Users can add a theme, set the default theme, and inject
arbitrary HTML. See rustdoc --help
for details.
Adding a new target
These are a set of steps to add support for a new target. There are numerous end states and paths to get there, so not all sections may be relevant to your desired goal.
- Specifying a new LLVM
- Creating a target specification
- Patching crates
- Cross-compiling
- Promoting a target from tier 2 (target) to tier 2 (host)
Specifying a new LLVM
For very new targets, you may need to use a different fork of LLVM
than what is currently shipped with Rust. In that case, navigate to
the src/llvm-project
git submodule (you might need to run ./x check
at least once so the submodule is updated), check out the
appropriate commit for your fork, then commit that new submodule
reference in the main Rust repository.
An example would be:
cd src/llvm-project
git remote add my-target-llvm some-llvm-repository
git checkout my-target-llvm/my-branch
cd ..
git add llvm-project
git commit -m 'Use my custom LLVM'
Using pre-built LLVM
If you have a local LLVM checkout that is already built, you may be able to configure Rust to treat your build as the system LLVM to avoid redundant builds.
You can tell Rust to use a pre-built version of LLVM using the target
section
of config.toml
:
[target.x86_64-unknown-linux-gnu]
llvm-config = "/path/to/llvm/llvm-7.0.1/bin/llvm-config"
If you are attempting to use a system LLVM, we have observed the following paths before, though they may be different from your system:
/usr/bin/llvm-config-8
/usr/lib/llvm-8/bin/llvm-config
Note that you need to have the LLVM FileCheck
tool installed, which is used
for codegen tests. This tool is normally built with LLVM, but if you use your
own preinstalled LLVM, you will need to provide FileCheck
in some other way.
On Debian-based systems, you can install the llvm-N-tools
package (where N
is the LLVM version number, e.g. llvm-8-tools
). Alternately, you can specify
the path to FileCheck
with the llvm-filecheck
config item in config.toml
or you can disable codegen test with the codegen-tests
item in config.toml
.
Creating a target specification
You should start with a target JSON file. You can see the specification
for an existing target using --print target-spec-json
:
rustc -Z unstable-options --target=wasm32-unknown-unknown --print target-spec-json
Save that JSON to a file and modify it as appropriate for your target.
Adding a target specification
Once you have filled out a JSON specification and been able to compile somewhat successfully, you can copy the specification into the compiler itself.
You will need to add a line to the big table inside of the
supported_targets
macro in the rustc_target::spec
module. You
will then add a corresponding file for your new target containing a
target
function.
Look for existing targets to use as examples.
After adding your target to the rustc_target
crate you may want to add
core
, std
, ... with support for your new target. In that case you will
probably need access to some target_*
cfg. Unfortunately when building with
stage0 (the beta compiler), you'll get an error that the target cfg is
unexpected because stage0 doesn't know about the new target specification and
we pass --check-cfg
in order to tell it to check.
To fix the errors you will need to manually add the unexpected value to the
different Cargo.toml
in library/{std,alloc,core}/Cargo.toml
. Here is an
example for adding NEW_TARGET_ARCH
as target_arch
:
library/std/Cargo.toml
:
[lints.rust.unexpected_cfgs]
level = "warn"
check-cfg = [
'cfg(bootstrap)',
- 'cfg(target_arch, values("xtensa"))',
+ # #[cfg(bootstrap)] NEW_TARGET_ARCH
+ 'cfg(target_arch, values("xtensa", "NEW_TARGET_ARCH"))',
To use this target in bootstrap, we need to explicitly add the target triple to the STAGE0_MISSING_TARGETS
list in src/bootstrap/src/core/sanity.rs
. This is necessary because the default compiler bootstrap uses does
not recognize the new target we just added. Therefore, it should be added to STAGE0_MISSING_TARGETS
so that the
bootstrap is aware that this target is not yet supported by the stage0 compiler.
const STAGE0_MISSING_TARGETS: &[&str] = &[
+ "NEW_TARGET_TRIPLE"
];
Patching crates
You may need to make changes to crates that the compiler depends on,
such as libc
or cc
. If so, you can use Cargo's
[patch]
ability. For example, if you want to use an
unreleased version of libc
, you can add it to the top-level
Cargo.toml
file:
diff --git a/Cargo.toml b/Cargo.toml
index 1e83f05e0ca..4d0172071c1 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -113,6 +113,8 @@ cargo-util = { path = "src/tools/cargo/crates/cargo-util" }
[patch.crates-io]
+libc = { git = "https://github.com/rust-lang/libc", rev = "0bf7ce340699dcbacabdf5f16a242d2219a49ee0" }
# See comments in `src/tools/rustc-workspace-hack/README.md` for what's going on
# here
rustc-workspace-hack = { path = 'src/tools/rustc-workspace-hack' }
After this, run cargo update -p libc
to update the lockfiles.
Beware that if you patch to a local path
dependency, this will enable
warnings for that dependency. Some dependencies are not warning-free, and due
to the deny-warnings
setting in config.toml
, the build may suddenly start
to fail.
To work around warnings, you may want to:
- Modify the dependency to remove the warnings
- Or for local development purposes, suppress the warnings by setting deny-warnings = false in config.toml.
# config.toml
[rust]
deny-warnings = false
Cross-compiling
Once you have a target specification in JSON and in the code, you can
cross-compile rustc
:
DESTDIR=/path/to/install/in \
./x install -i --stage 1 --host aarch64-apple-darwin.json --target aarch64-apple-darwin \
compiler/rustc library/std
If your target specification is already available in the bootstrap compiler, you can use it instead of the JSON file for both arguments.
Promoting a target from tier 2 (target) to tier 2 (host)
There are two levels of tier 2 targets:
a) Targets that are only cross-compiled (rustup target add
)
b) Targets that have a native toolchain (rustup toolchain install
)
For an example of promoting a target from cross-compiled to native, see #75914.
Optimized build of the compiler
There are multiple additional build configuration options and techniques that can be used to compile a
build of rustc
that is as optimized as possible (for example when building rustc
for a Linux
distribution). The status of these configuration options for various Rust targets is tracked here.
This page describes how you can use these approaches when building rustc
yourself.
Link-time optimization
Link-time optimization is a powerful compiler technique that can increase program performance. To
enable (Thin-)LTO when building rustc
, set the rust.lto
config option to "thin"
in config.toml
:
[rust]
lto = "thin"
Note that LTO for
rustc
is currently supported and tested only for thex86_64-unknown-linux-gnu
target. Other targets may work, but no guarantees are provided. Notably, LTO-optimizedrustc
currently produces miscompilations on Windows.
Enabling LTO on Linux has produced speed-ups by up to 10%.
Memory allocator
Using a different memory allocator for rustc
can provide significant performance benefits. If you
want to enable the jemalloc
allocator, you can set the rust.jemalloc
option to true
in config.toml
:
[rust]
jemalloc = true
Note that this option is currently only supported for Linux and macOS targets.
Codegen units
Reducing the amount of codegen units per rustc
crate can produce a faster build of the compiler.
You can modify the number of codegen units for rustc
and libstd
in config.toml
with the
following options:
[rust]
codegen-units = 1
codegen-units-std = 1
Instruction set
By default, rustc
is compiled for a generic (and conservative) instruction set architecture
(depending on the selected target), to make it support as many CPUs as possible. If you want to
compile rustc
for a specific instruction set architecture, you can set the target_cpu
compiler
option in RUSTFLAGS
:
RUSTFLAGS="-C target_cpu=x86-64-v3" ./x build ...
If you also want to compile LLVM for a specific instruction set, you can set llvm
flags
in config.toml
:
[llvm]
cxxflags = "-march=x86-64-v3"
cflags = "-march=x86-64-v3"
Profile-guided optimization
Applying profile-guided optimizations (or more generally, feedback-directed optimizations) can
produce a large increase to rustc
performance, by up to 15% (1, 2). However, these techniques
are not simply enabled by a configuration option, but rather they require a complex build workflow
that compiles rustc
multiple times and profiles it on selected benchmarks.
There is a tool called opt-dist
that is used to optimize rustc
with PGO (profile-guided
optimizations) and BOLT (a post-link binary optimizer) for builds distributed to end users. You
can examine the tool, which is located in src/tools/opt-dist
, and build a custom PGO build
workflow based on it, or try to use it directly. Note that the tool is currently quite hardcoded to
the way we use it in Rust's continuous integration workflows, and it might require some custom
changes to make it work in a different environment.
To use the tool, you will need to provide some external dependencies:
- A Python3 interpreter (for executing
x.py
). - Compiled LLVM toolchain, with the
llvm-profdata
binary. Optionally, if you want to use BOLT, thellvm-bolt
andmerge-fdata
binaries have to be available in the toolchain.
These dependencies are provided to opt-dist
by an implementation of the Environment
struct.
It specifies directories where will the PGO/BOLT pipeline take place, and also external dependencies
like Python or LLVM.
Here is an example of how can opt-dist
be used locally (outside of CI):
- Build the tool with the following command:
./x build tools/opt-dist
- Run the tool with the
local
mode and provide necessary parameters:
You can run./build/host/stage0-tools-bin/opt-dist local \ --target-triple <target> \ # select target, e.g. "x86_64-unknown-linux-gnu" --checkout-dir <path> \ # path to rust checkout, e.g. "." --llvm-dir <path> \ # path to built LLVM toolchain, e.g. "/foo/bar/llvm/install" -- python3 x.py dist # pass the actual build command
--help
to see further parameters that you can modify.
Note: if you want to run the actual CI pipeline, instead of running
opt-dist
locally, you can executeDEPLOY=1 src/ci/docker/run.sh dist-x86_64-linux
.
Testing the compiler
The Rust project runs a wide variety of different tests, orchestrated by the
build system (./x test
). This section gives a brief overview of the different
testing tools. Subsequent chapters dive into running tests and
adding new tests.
Kinds of tests
There are several kinds of tests to exercise things in the Rust distribution.
Almost all of them are driven by ./x test
, with some exceptions noted below.
Compiletest
The main test harness for testing the compiler itself is a tool called compiletest.
compiletest supports running different styles of tests, organized into test
suites. A test mode may provide common presets/behavior for a set of test
suites. compiletest-supported tests are located in the tests
directory.
The Compiletest chapter goes into detail on how to use this tool.
Example:
./x test tests/ui
Package tests
The standard library and many of the compiler packages include typical Rust
#[test]
unit tests, integration tests, and documentation tests. You can pass a
path to ./x test
for almost any package in the library/
or compiler/
directory, and x
will essentially run cargo test
on that package.
Examples:
Command | Description |
---|---|
./x test library/std | Runs tests on std only |
./x test library/core | Runs tests on core only |
./x test compiler/rustc_data_structures | Runs tests on rustc_data_structures |
The standard library relies very heavily on documentation tests to cover its functionality. However, unit tests and integration tests can also be used as needed. Almost all of the compiler packages have doctests disabled.
All standard library and compiler unit tests are placed in separate tests
file
(which is enforced in tidy). This ensures that when the test
file is changed, the crate does not need to be recompiled. For example:
#[cfg(test)]
mod tests;
If it wasn't done this way, and you were working on something like core
, that
would require recompiling the entire standard library, and the entirety of
rustc
.
./x test
includes some CLI options for controlling the behavior with these
package tests:
--doc
— Only runs documentation tests in the package.--no-doc
— Run all tests except documentation tests.
Tidy
Tidy is a custom tool used for validating source code style and formatting conventions, such as rejecting long lines. There is more information in the section on coding conventions.
Examples:
./x test tidy
Formatting
Rustfmt is integrated with the build system to enforce uniform style across the compiler. The formatting check is automatically run by the Tidy tool mentioned above.
Examples:
Command | Description |
---|---|
./x fmt --check | Checks formatting and exits with an error if formatting is needed. |
./x fmt | Runs rustfmt across the entire codebase. |
./x test tidy --bless | First runs rustfmt to format the codebase, then runs tidy checks. |
Book documentation tests
All of the books that are published have their own tests, primarily for
validating that the Rust code examples pass. Under the hood, these are
essentially using rustdoc --test
on the markdown files. The tests can be run
by passing a path to a book to ./x test
.
Example:
./x test src/doc/book
Documentation link checker
Links across all documentation is validated with a link checker tool.
Example:
./x test src/tools/linkchecker
Example:
./x test linkchecker
This requires building all of the documentation, which might take a while.
Dist check
distcheck
verifies that the source distribution tarball created by the build
system will unpack, build, and run all tests.
Example:
./x test distcheck
Tool tests
Packages that are included with Rust have all of their tests run as well. This includes things such as cargo, clippy, rustfmt, miri, bootstrap (testing the Rust build system itself), etc.
Most of the tools are located in the src/tools
directory. To run the tool's
tests, just pass its path to ./x test
.
Example:
./x test src/tools/cargo
Usually these tools involve running cargo test
within the tool's directory.
If you want to run only a specified set of tests, append --test-args FILTER_NAME
to the command.
Example:
./x test src/tools/miri --test-args padding
In CI, some tools are allowed to fail. Failures send notifications to the corresponding teams, and is tracked on the toolstate website. More information can be found in the toolstate documentation.
Ecosystem testing
Rust tests integration with real-world code to catch regressions and make informed decisions about the evolution of the language. There are several kinds of ecosystem tests, including Crater. See the Ecosystem testing chapter for more details.
Performance testing
A separate infrastructure is used for testing and tracking performance of the compiler. See the Performance testing chapter for more details.
Miscellaneous information
There are some other useful testing-related info at Misc info.
Further reading
The following blog posts may also be of interest:
- brson's classic "How Rust is tested"
Running tests
- Running a subset of the test suites
- Run unit tests on the compiler/library
- Running an individual test
- Passing arguments to
rustc
when running tests - Editing and updating the reference files
- Configuring test running
- Passing
--pass $mode
- Running tests with different "compare modes"
- Running tests manually
- Running
run-make
tests - Running tests on a remote machine
- Testing on emulators
- Running rustc_codegen_gcc tests
You can run the entire test collection using x
. But note that running the
entire test collection is almost never what you want to do during local
development because it takes a really long time. For local development, see the
subsection after on how to run a subset of tests.
You usually only want to run a subset of the test suites (or even a smaller set of tests than that) which you expect will exercise your changes. PR CI exercises a subset of test collections, and merge queue CI will exercise all of the test collection.
./x test
The test results are cached and previously successful tests are ignored
during
testing. The stdout/stderr contents as well as a timestamp file for every test
can be found under build/<target-triple>/test/
for the given
<target-triple>
. To force-rerun a test (e.g. in case the test runner fails to
notice a change) you can use the --force-rerun
CLI option.
Note on requirements of external dependencies
Some test suites may require external dependecies. This is especially true of debuginfo tests. Some debuginfo tests require a Python-enabled gdb. You can test if your gdb install supports Python by using the
python
command from within gdb. Once invoked you can type some Python code (e.g.print("hi")
) followed by return and thenCTRL+D
to execute it. If you are building gdb from source, you will need to configure with--with-python=<path-to-python-binary>
.
Running a subset of the test suites
When working on a specific PR, you will usually want to run a smaller set of
tests. For example, a good "smoke test" that can be used after modifying rustc
to see if things are generally working correctly would be to exercise the ui
test suite (tests/ui
):
./x test tests/ui
This will run the ui
test suite. Of course, the choice of test suites is
somewhat arbitrary, and may not suit the task you are doing. For example, if you
are hacking on debuginfo, you may be better off with the debuginfo test suite:
./x test tests/debuginfo
If you only need to test a specific subdirectory of tests for any given test
suite, you can pass that directory as a filter to ./x test
:
./x test tests/ui/const-generics
Note for MSYS2
On MSYS2 the paths seem to be strange and
./x test
neither recognizestests/ui/const-generics
nortests\ui\const-generics
. In that case, you can workaround it by using e.g../x test ui --test-args="tests/ui/const-generics"
.
Likewise, you can test a single file by passing its path:
./x test tests/ui/const-generics/const-test.rs
x
doesn't support running a single tool test by passing its path yet. You'll
have to use the --test-args
argument as describled
below.
./x test src/tools/miri --test-args tests/fail/uninit/padding-enum.rs
Run only the tidy script
./x test tidy
Run tests on the standard library
./x test --stage 0 library/std
Note that this only runs tests on std
; if you want to test core
or other
crates, you have to specify those explicitly.
Run the tidy script and tests on the standard library
./x test --stage 0 tidy library/std
Run tests on the standard library using a stage 1 compiler
./x test --stage 1 library/std
By listing which test suites you want to run you avoid having to run tests for components you did not change at all.
Run all tests using a stage 2 compiler
./x test --stage 2
Run unit tests on the compiler/library
You may want to run unit tests on a specific file with following:
./x test compiler/rustc_data_structures/src/thin_vec/tests.rs
But unfortunately, it's impossible. You should invoke the following instead:
./x test compiler/rustc_data_structures/ --test-args thin_vec
Running an individual test
Another common thing that people want to do is to run an individual test,
often the test they are trying to fix. As mentioned earlier, you may pass the
full file path to achieve this, or alternatively one may invoke x
with the
--test-args
option:
./x test tests/ui --test-args issue-1234
Under the hood, the test runner invokes the standard Rust test runner (the same
one you get with #[test]
), so this command would wind up filtering for tests
that include "issue-1234" in the name. Thus, --test-args
is a good way to run
a collection of related tests.
Passing arguments to rustc
when running tests
It can sometimes be useful to run some tests with specific compiler arguments,
without using RUSTFLAGS
(during development of unstable features, with -Z
flags, for example).
This can be done with ./x test
's --compiletest-rustc-args
option, to pass
additional arguments to the compiler when building the tests.
Editing and updating the reference files
If you have changed the compiler's output intentionally, or you are making a new
test, you can pass --bless
to the test subcommand. E.g. if some tests in
tests/ui
are failing, you can run
./x test tests/ui --bless
to automatically adjust the .stderr
, .stdout
or .fixed
files of
all tests. Of course you can also target just specific tests with the
--test-args your_test_name
flag, just like when running the tests.
Configuring test running
There are a few options for running tests:
config.toml
has therust.verbose-tests
option. Iffalse
, each test will print a single dot (the default). Iftrue
, the name of every test will be printed. This is equivalent to the--quiet
option in the Rust test harness- The environment variable
RUST_TEST_THREADS
can be set to the number of concurrent threads to use for testing.
Passing --pass $mode
Pass UI tests now have three modes, check-pass
, build-pass
and run-pass
.
When --pass $mode
is passed, these tests will be forced to run under the given
$mode
unless the directive //@ ignore-pass
exists in the test file. For
example, you can run all the tests in tests/ui
as check-pass
:
./x test tests/ui --pass check
By passing --pass $mode
, you can reduce the testing time. For each mode,
please see Controlling pass/fail
expectations.
Running tests with different "compare modes"
UI tests may have different output depending on certain "modes" that the
compiler is in. For example, when using the Polonius mode, a test foo.rs
will
first look for expected output in foo.polonius.stderr
, falling back to the
usual foo.stderr
if not found. The following will run the UI test suite in
Polonius mode:
./x test tests/ui --compare-mode=polonius
See Compare modes for more details.
Running tests manually
Sometimes it's easier and faster to just run the test by hand. Most tests are
just .rs
files, so after creating a rustup
toolchain, you
can do something like:
rustc +stage1 tests/ui/issue-1234.rs
This is much faster, but doesn't always work. For example, some tests include directives that specify specific compiler flags, or which rely on other crates, and they may not run the same without those options.
Running run-make
tests
Windows
Running the run-make
test suite on Windows is a currently bit more involved.
There are numerous prerequisites and environmental requirements:
- Install msys2: https://www.msys2.org/
- Specify
MSYS2_PATH_TYPE=inherit
inmsys2.ini
in the msys2 installation directory, run the following withMSYS2 MSYS
:pacman -Syuu
pacman -S make
pacman -S diffutils
pacman -S binutils
./x test run-make
(./x test tests/run-make
doesn't work)
There is on-going work to not rely on Makefile
s in the
run-make test suite. Once this work is completed, you can run the entire
run-make
test suite on native Windows inside cmd
or PowerShell
without
needing to install and use MSYS2. As of Oct 2024, it is
already possible to run the vast majority of the run-make
test suite outside
of MSYS2, but there will be failures for the tests that still use Makefile
s
due to not finding make
.
Running tests on a remote machine
Tests may be run on a remote machine (e.g. to test builds for a different
architecture). This is done using remote-test-client
on the build machine to
send test programs to remote-test-server
running on the remote machine.
remote-test-server
executes the test programs and sends the results back to
the build machine. remote-test-server
provides unauthenticated remote code
execution so be careful where it is used.
To do this, first build remote-test-server
for the remote machine, e.g. for
RISC-V
./x build src/tools/remote-test-server --target riscv64gc-unknown-linux-gnu
The binary will be created at
./build/host/stage2-tools/$TARGET_ARCH/release/remote-test-server
. Copy this
over to the remote machine.
On the remote machine, run the remote-test-server
with the --bind 0.0.0.0:12345
flag (and optionally -v
for verbose output). Output should look
like this:
$ ./remote-test-server -v --bind 0.0.0.0:12345
starting test server
listening on 0.0.0.0:12345!
Note that binding the server to 0.0.0.0 will allow all hosts able to reach your machine to execute arbitrary code on your machine. We strongly recommend either setting up a firewall to block external access to port 12345, or to use a more restrictive IP address when binding.
You can test if the remote-test-server
is working by connecting to it and
sending ping\n
. It should reply pong
:
$ nc $REMOTE_IP 12345
ping
pong
To run tests using the remote runner, set the TEST_DEVICE_ADDR
environment
variable then use x
as usual. For example, to run ui
tests for a RISC-V
machine with the IP address 1.2.3.4
use
export TEST_DEVICE_ADDR="1.2.3.4:12345"
./x test tests/ui --target riscv64gc-unknown-linux-gnu
If remote-test-server
was run with the verbose flag, output on the test
machine may look something like
[...]
run "/tmp/work/test1007/a"
run "/tmp/work/test1008/a"
run "/tmp/work/test1009/a"
run "/tmp/work/test1010/a"
run "/tmp/work/test1011/a"
run "/tmp/work/test1012/a"
run "/tmp/work/test1013/a"
run "/tmp/work/test1014/a"
run "/tmp/work/test1015/a"
run "/tmp/work/test1016/a"
run "/tmp/work/test1017/a"
run "/tmp/work/test1018/a"
[...]
Tests are built on the machine running x
not on the remote machine. Tests
which fail to build unexpectedly (or ui
tests producing incorrect build
output) may fail without ever running on the remote machine.
Testing on emulators
Some platforms are tested via an emulator for architectures that aren't readily available. For architectures where the standard library is well supported and the host operating system supports TCP/IP networking, see the above instructions for testing on a remote machine (in this case the remote machine is emulated).
There is also a set of tools for orchestrating running the tests within the
emulator. Platforms such as arm-android
and arm-unknown-linux-gnueabihf
are
set up to automatically run the tests under emulation on GitHub Actions. The
following will take a look at how a target's tests are run under emulation.
The Docker image for armhf-gnu includes QEMU to emulate the ARM CPU
architecture. Included in the Rust tree are the tools remote-test-client and
remote-test-server which are programs for sending test programs and libraries
to the emulator, and running the tests within the emulator, and reading the
results. The Docker image is set up to launch remote-test-server
and the
build tools use remote-test-client
to communicate with the server to
coordinate running tests (see src/bootstrap/src/core/build_steps/test.rs).
TODO
- Is there any support for using an iOS emulator?
- It's also unclear to me how the wasm or asm.js tests are run.
Running rustc_codegen_gcc tests
First thing to know is that it only supports linux x86_64 at the moment. We will extend its support later on.
You need to update codegen-backends
value in your config.toml
file in the
[rust]
section and add "gcc" in the array:
codegen-backends = ["llvm", "gcc"]
Then you need to install libgccjit 12. For example with apt
:
$ apt install libgccjit-12-dev
Now you can run the following command:
$ ./x test compiler/rustc_codegen_gcc/
If it cannot find the .so
library (if you installed it with apt
for example), you
need to pass the library file path with LIBRARY_PATH
:
$ LIBRARY_PATH=/usr/lib/gcc/x86_64-linux-gnu/12/ ./x test compiler/rustc_codegen_gcc/
If you encounter bugs or problems, don't hesitate to open issues on the
rustc_codegen_gcc
repository.
Testing with Docker
The Rust tree includes Docker image definitions for the platforms used on
GitHub Actions in src/ci/docker
.
The script src/ci/docker/run.sh
is used to build the Docker image, run it,
build Rust within the image, and run the tests.
You can run these images on your local development machine. This can be
helpful to test environments different from your local system. First you will
need to install Docker on a Linux, Windows, or macOS system (typically Linux
will be much faster than Windows or macOS because the latter use virtual
machines to emulate a Linux environment). To enter interactive mode which will
start a bash shell in the container, run src/ci/docker/run.sh --dev <IMAGE>
where <IMAGE>
is one of the directory names in src/ci/docker
(for example
x86_64-gnu
is a fairly standard Ubuntu environment).
The docker script will mount your local Rust source tree in read-only mode,
and an obj
directory in read-write mode. All of the compiler artifacts will
be stored in the obj
directory. The shell will start out in the obj
directory. From there, you can run ../src/ci/run.sh
which will run the build
as defined by the image.
Alternatively, you can run individual commands to do specific tasks. For
example, you can run ../x test tests/ui
to just run UI tests.
Note that there is some configuration in the src/ci/run.sh
script that you
may need to recreate. Particularly, set submodules = false
in your
config.toml
so that it doesn't attempt to modify the read-only directory.
Some additional notes about using the Docker images:
- Some of the std tests require IPv6 support. Docker on Linux seems to have it
disabled by default. Run the commands in
enable-docker-ipv6.sh
to enable IPv6 before creating the container. This only needs to be done once. - The container will be deleted automatically when you exit the shell, however
the build artifacts persist in the
obj
directory. If you are switching between different Docker images, the artifacts from previous environments stored in theobj
directory may confuse the build system. Sometimes you will need to delete parts or all of theobj
directory before building inside the container. - The container is bare-bones, with only a minimal set of packages. You may
want to install some things like
apt install less vim
. - You can open multiple shells in the container. First you need the container
name (a short hash), which is displayed in the shell prompt, or you can run
docker container ls
outside of the container to list the available containers. With the container name, rundocker exec -it <CONTAINER> /bin/bash
where<CONTAINER>
is the container name like4ba195e95cef
.
Testing with CI
The primary goal of our CI system is to ensure that the master
branch of
rust-lang/rust
is always in a valid state and passes our test suite.
From a high-level point of view, when you open a pull request at
rust-lang/rust
, the following will happen:
- A small subset of tests and checks are run after each push to the PR. This should help catching common errors.
- When the PR is approved, the bors bot enqueues the PR into a merge queue.
- Once the PR gets to the front of the queue, bors will create a merge commit and run the full test suite on it. The merge commit either contains only one specific PR or it can be a "rollup" which combines multiple PRs together, to save CI costs.
- Once the whole test suite finishes, two things can happen. Either CI fails
with an error that needs to be addressed by the developer, or CI succeeds and
the merge commit is then pushed to the
master
branch.
If you want to modify what gets executed on CI, see Modifying CI jobs.
CI workflow
Our CI is primarily executed on GitHub Actions, with a single workflow defined
in .github/workflows/ci.yml
, which contains a bunch of steps that are
unified for all CI jobs that we execute. When a commit is pushed to a
corresponding branch or a PR, the workflow executes the
calculate-job-matrix.py
script, which dynamically generates the specific CI
jobs that should be executed. This script uses the jobs.yml
file as an
input, which contains a declarative configuration of all our CI jobs.
Almost all build steps shell out to separate scripts. This keeps the CI fairly platform independent (i.e., we are not overly reliant on GitHub Actions). GitHub Actions is only relied on for bootstrapping the CI process and for orchestrating the scripts that drive the process.
In essence, all CI jobs run ./x test
, ./x dist
or some other command with
different configurations, across various operating systems, targets and
platforms. There are two broad categories of jobs that are executed, dist
and
non-dist
jobs.
- Dist jobs build a full release of the compiler for a specific platform,
including all the tools we ship through rustup; Those builds are then uploaded
to the
rust-lang-ci2
S3 bucket and are available to be locally installed with the rustup-toolchain-install-master tool. The same builds are also used for actual releases: our release process basically consists of copying those artifacts fromrust-lang-ci2
to the production endpoint and signing them. - Non-dist jobs run our full test suite on the platform, and the test suite of all the tools we ship through rustup; The amount of stuff we test depends on the platform (for example some tests are run only on Tier 1 platforms), and some quicker platforms are grouped together on the same builder to avoid wasting CI resources.
Based on an input event (usually a push to a branch), we execute one of three kinds of builds (sets of jobs).
- PR builds
- Auto builds
- Try builds
Pull Request builds
After each push to a pull request, a set of pr
jobs are executed. Currently,
these execute the x86_64-gnu-llvm-X
, x86_64-gnu-tools
, mingw-check
and
mingw-check-tidy
jobs, all running on Linux. These execute a relatively short
(~30 minutes) and lightweight test suite that should catch common issues. More
specifically, they run a set of lints, they try to perform a cross-compile check
build to Windows mingw (without producing any artifacts) and they test the
compiler using a system version of LLVM. Unfortunately, it would take too many
resources to run the full test suite for each commit on every PR.
Note on doc comments
Note that PR CI as of Oct 2024 by default does not try to run
./x doc xxx
. This means that if you have any broken intradoc links that would lead to./x doc xxx
failing, it will happen very late into the full merge queue CI pipeline.Thus, it is a good idea to run
./x doc xxx
locally for any doc comment changes to help catch these early.
PR jobs are defined in the pr
section of jobs.yml
. They run under the
rust-lang/rust
repository, and their results can be observed directly on the
PR, in the "CI checks" section at the bottom of the PR page.
Auto builds
Before a commit can be merged into the master
branch, it needs to pass our
complete test suite. We call this an auto
build. This build runs tens of CI
jobs that exercise various tests across operating systems and targets. The full
test suite is quite slow; it can take two hours or more until all the auto
CI
jobs finish.
Most platforms only run the build steps, some run a restricted set of tests, only a subset run the full suite of tests (see Rust's platform tiers).
Auto jobs are defined in the auto
section of jobs.yml
. They are executed
on the auto
branch under the rust-lang-ci/rust
repository1 and
their results can be seen here,
although usually you will be notified of the result by a comment made by bors on
the corresponding PR.
At any given time, at most a single auto
build is being executed. Find out
more here.
The auto
and try
jobs run under the rust-lang-ci
fork for
historical reasons. This may change in the future.
Try builds
Sometimes we want to run a subset of the test suite on CI for a given PR, or
build a set of compiler artifacts from that PR, without attempting to merge it.
We call this a "try build". A try build is started after a user with the proper
permissions posts a PR comment with the @bors try
command.
There are several use-cases for try builds:
- Run a set of performance benchmarks using our rustc-perf benchmark suite.
For this, a working compiler build is needed, which can be generated with a
try build that runs the dist-x86_64-linux CI job, which builds an optimized
version of the compiler on Linux (this job is currently executed by default
when you start a try build). To create a try build and schedule it for a
performance benchmark, you can use the
@bors try @rust-timer queue
command combination. - Check the impact of the PR across the Rust ecosystem, using a crater run. Again, a working compiler build is needed for this, which can be produced by the dist-x86_64-linux CI job.
- Run a specific CI job (e.g. Windows tests) on a PR, to quickly test if it
passes the test suite executed by that job. You can select which CI jobs will
be executed in the try build by adding up to 10 lines containing
try-job: <name of job>
to the PR description. All such specified jobs will be executed in the try build once the@bors try
command is used on the PR. If no try jobs are specified in this way, the jobs defined in thetry
section ofjobs.yml
will be executed by default.
Using
try-job
PR description directives
Identify which set of try-jobs (max 10) you would like to exercise. You can find the name of the CI jobs in
jobs.yml
.Amend PR description to include (usually at the end of the PR description) e.g.
This PR fixes #123456. try-job: x86_64-msvc try-job: test-various
Each
try-job
directive must be on its own line.Run the prescribed try jobs with
@bors try
. As aforementioned, this requires the user to either (1) havetry
permissions or (2) be delegated withtry
permissions by@bors delegate
by someone who hastry
permissions.Note that this is usually easier to do than manually edit
jobs.yml
. However, it can be less flexible because you cannot adjust the set of tests that are exercised this way.
Try jobs are defined in the try
section of jobs.yml
. They are executed on
the try
branch under the rust-lang-ci/rust
repository1 and
their results can be seen here,
although usually you will be notified of the result by a comment made by bors on
the corresponding PR.
Multiple try builds can execute concurrently across different PRs.
Modifying CI jobs
If you want to modify what gets executed on our CI, you can simply modify the
pr
, auto
or try
sections of the jobs.yml
file.
You can also modify what gets executed temporarily, for example to test a particular platform or configuration that is challenging to test locally (for example, if a Windows build fails, but you don't have access to a Windows machine). Don't hesitate to use CI resources in such situations to try out a fix!
You can perform an arbitrary CI job in two ways:
- Use the try build functionality, and specify the CI jobs that you want to be executed in try builds in your PR description.
- Modify the
pr
section ofjobs.yml
to specify which CI jobs should be executed after each push to your PR. This might be faster than repeatedly starting try builds.
To modify the jobs executed after each push to a PR, you can simply copy one of
the job definitions from the auto
section to the pr
section. For example,
the x86_64-msvc
job is responsible for running the 64-bit MSVC tests. You can
copy it to the pr
section to cause it to be executed after a commit is pushed
to your PR, like this:
pr:
...
- image: x86_64-gnu-tools
<<: *job-linux-16c
# this item was copied from the `auto` section
# vvvvvvvvvvvvvvvvvv
- image: x86_64-msvc
env:
RUST_CONFIGURE_ARGS: --build=x86_64-pc-windows-msvc --enable-profiler
SCRIPT: make ci-msvc
<<: *job-windows-8c
Then you can commit the file and push it to your PR branch on GitHub. GitHub Actions should then execute this CI job after each push to your PR.
After you have finished your experiments, don't forget to remove any changes
you have made to jobs.yml
, if they were supposed to be temporary!
A good practice is to prefix [WIP]
in PR title while still running try jobs
and [DO NOT MERGE]
in the commit that modifies the CI jobs for testing
purposes.
Although you are welcome to use CI, just be conscious that this is a shared resource with limited concurrency. Try not to enable too many jobs at once (one or two should be sufficient in most cases).
Merging PRs serially with bors
CI services usually test the last commit of a branch merged with the last commit
in master
, and while that’s great to check if the feature works in isolation,
it doesn’t provide any guarantee the code is going to work once it’s merged.
Breakages like these usually happen when another, incompatible PR is merged
after the build happened.
To ensure a master
branch that works all the time, we forbid manual merges.
Instead, all PRs have to be approved through our bot, bors (the software
behind it is called homu). All the approved PRs are put in a merge queue
(sorted by priority and creation date) and are automatically tested one at the
time. If all the builders are green, the PR is merged, otherwise the failure is
recorded and the PR will have to be re-approved again.
Bors doesn’t interact with CI services directly, but it works by pushing the
merge commit it wants to test to specific branches (like auto
or try
), which
are configured to execute CI checks. Bors then detects the outcome of the build
by listening for either Commit Statuses or Check Runs. Since the merge commit is
based on the latest master
and only one can be tested at the same time, when
the results are green, master
is fast-forwarded to that merge commit.
Unfortunately testing a single PR at the time, combined with our long CI (~2 hours for a full run), means we can’t merge too many PRs in a single day, and a single failure greatly impacts our throughput for the day. The maximum number of PRs we can merge in a day is around ~10.
The large CI run times and requirement for a large builder pool is largely due
to the fact that full release artifacts are built in the dist-
builders. This
is worth it because these release artifacts:
- Allow perf testing even at a later date.
- Allow bisection when bugs are discovered later.
- Ensure release quality since if we're always releasing, we can catch problems early.
Rollups
Some PRs don’t need the full test suite to be executed: trivial changes like typo fixes or README improvements shouldn’t break the build, and testing every single one of them for 2+ hours is a big waste of time. To solve this, we regularly create a "rollup", a PR where we merge several pending trivial PRs so they can be tested together. Rollups are created manually by a team member using the "create a rollup" button on the merge queue. The team member uses their judgment to decide if a PR is risky or not, and are the best tool we have at the moment to keep the queue in a manageable state.
Docker
All CI jobs, except those on macOS and Windows, are executed inside that platform’s custom Docker container. This has a lot of advantages for us:
- The build environment is consistent regardless of the changes of the underlying image (switching from the trusty image to xenial was painless for us).
- We can use ancient build environments to ensure maximum binary compatibility, for example using older CentOS releases on our Linux builders.
- We can avoid reinstalling tools (like QEMU or the Android emulator) every time thanks to Docker image caching.
- Users can run the same tests in the same environment locally by just running
src/ci/docker/run.sh image-name
, which is awesome to debug failures. Note that there are only linux docker images available locally due to licensing and other restrictions.
The docker images prefixed with dist-
are used for building artifacts while
those without that prefix run tests and checks.
We also run tests for less common architectures (mainly Tier 2 and Tier 3 platforms) in CI. Since those platforms are not x86 we either run everything inside QEMU or just cross-compile if we don’t want to run the tests for that platform.
These builders are running on a special pool of builders set up and maintained for us by GitHub.
Caching
Our CI workflow uses various caching mechanisms, mainly for two things:
Docker images caching
The Docker images we use to run most of the Linux-based builders take a long time to fully build. To speed up the build, we cache it using Docker registry caching, with the intermediate artifacts being stored on ghcr.io. We also push the built Docker images to ghcr, so that they can be reused by other tools (rustup) or by developers running the Docker build locally (to speed up their build).
Since we test multiple, diverged branches (master
, beta
and stable
), we
can’t rely on a single cache for the images, otherwise builds on a branch would
override the cache for the others. Instead, we store the images under different
tags, identifying them with a custom hash made from the contents of all the
Dockerfiles and related scripts.
LLVM caching with sccache
We build some C/C++ stuff in various CI jobs, and we rely on sccache to cache
the intermediate LLVM artifacts. Sccache is a distributed ccache developed by
Mozilla, which can use an object storage bucket as the storage backend. In our
case, the artefacts are uploaded to an S3 bucket that we control
(rust-lang-ci-sccache2
).
Custom tooling around CI
During the years we developed some custom tooling to improve our CI experience.
Rust Log Analyzer to show the error message in PRs
The build logs for rust-lang/rust
are huge, and it’s not practical to find
what caused the build to fail by looking at the logs. To improve the developers’
experience we developed a bot called Rust Log Analyzer (RLA) that
receives the build logs on failure and extracts the error message automatically,
posting it on the PR.
The bot is not hardcoded to look for error strings, but was trained with a bunch of build failures to recognize which lines are common between builds and which are not. While the generated snippets can be weird sometimes, the bot is pretty good at identifying the relevant lines even if it’s an error we've never seen before.
Toolstate to support allowed failures
The rust-lang/rust
repo doesn’t only test the compiler on its CI, but also a
variety of tools and documentation. Some documentation is pulled in via git
submodules. If we blocked merging rustc PRs on the documentation being fixed, we
would be stuck in a chicken-and-egg problem, because the documentation's CI
would not pass since updating it would need the not-yet-merged version of rustc
to test against (and we usually require CI to be passing).
To avoid the problem, submodules are allowed to fail, and their status is recorded in rust-toolstate. When a submodule breaks, a bot automatically pings the maintainers so they know about the breakage, and it records the failure on the toolstate repository. The release process will then ignore broken tools on nightly, removing them from the shipped nightlies.
While tool failures are allowed most of the time, they’re automatically forbidden a week before a release: we don’t care if tools are broken on nightly but they must work on beta and stable, so they also need to work on nightly a few days before we promote nightly to beta.
More information is available in the toolstate documentation.
Adding new tests
In general, we expect every PR that fixes a bug in rustc to come accompanied by a regression test of some kind. This test should fail in master but pass after the PR. These tests are really useful for preventing us from repeating the mistakes of the past.
The first thing to decide is which kind of test to add. This will depend on the nature of the change and what you want to exercise. Here are some rough guidelines:
- The majority of compiler tests are done with compiletest.
- Changes to the standard library are usually tested within the standard library
itself.
- The majority of standard library tests are written as doctests, which illustrate and exercise typical API behavior.
- Additional unit tests should go in
library/${crate}/tests
(where${crate}
is usuallycore
,alloc
, orstd
).
- If the code is part of an isolated system, and you are not testing compiler output, consider using a unit or integration test.
- Need to run rustdoc? Prefer a
rustdoc
orrustdoc-ui
test. Occasionally you'll needrustdoc-js
as well. - Other compiletest test suites are generally used for special purposes:
- Need to run gdb or lldb? Use the
debuginfo
test suite. - Need to inspect LLVM IR or MIR IR? Use the
codegen
ormir-opt
test suites. - Need to inspect the resulting binary in some way? Or if all the other test
suites are too limited for your purposes? Then use
run-make
. - Check out the compiletest chapter for more specialized test suites.
- Need to run gdb or lldb? Use the
After deciding on which kind of test to add, see best practices for guidance on how to author tests that are easy to work with that stand the test of time (i.e. if a test fails or need to be modified several years later, how can we make it easier for them?).
UI test walkthrough
The following is a basic guide for creating a UI test, which is one of the most common compiler tests. For this tutorial, we'll be adding a test for an async error message.
Step 1: Add a test file
The first step is to create a Rust source file somewhere in the tests/ui
tree. When creating a test, do your best to find a good location and name (see
Test organization for more). Since naming is the
hardest part of development, everything should be downhill from here!
Let's place our async test at tests/ui/async-await/await-without-async.rs
:
// Provide diagnostics when the user writes `await` in a non-`async` function.
//@ edition:2018
async fn foo() {}
fn bar() {
foo().await
}
fn main() {}
A few things to notice about our test:
- The top should start with a short comment that explains what the test is for.
- The
//@ edition:2018
comment is called a directive which provides instructions to compiletest on how to build the test. Here we need to set the edition forasync
to work (the default is edition 2015). - Following that is the source of the test. Try to keep it succinct and to the point. This may require some effort if you are trying to minimize an example from a bug report.
- We end this test with an empty
fn main
function. This is because the default for UI tests is abin
crate-type, and we don't want the "main not found" error in our test. Alternatively, you could add#![crate_type="lib"]
.
Step 2: Generate the expected output
The next step is to create the expected output snapshots from the compiler. This
can be done with the --bless
option:
./x test tests/ui/async-await/await-without-async.rs --bless
This will build the compiler (if it hasn't already been built), compile the
test, and place the output of the compiler in a file called
tests/ui/async-await/await-without-async.stderr
.
However, this step will fail! You should see an error message, something like this:
error: /rust/tests/ui/async-await/await-without-async.rs:7: unexpected error: '7:10: 7:16:
await
is only allowed insideasync
functions and blocks E0728'
This is because the stderr contains errors which were not matched by error annotations in the source file.
Step 3: Add error annotations
Every error needs to be annotated with a comment in the source with the text of the error. In this case, we can add the following comment to our test file:
fn bar() {
foo().await
//~^ ERROR `await` is only allowed inside `async` functions and blocks
}
The //~^
squiggle caret comment tells compiletest that the error belongs to
the previous line (more on this in the Error
annotations section).
Save that, and run the test again:
./x test tests/ui/async-await/await-without-async.rs
It should now pass, yay!
Step 4: Review the output
Somewhat hand-in-hand with the previous step, you should inspect the .stderr
file that was created to see if it looks like how you expect. If you are adding
a new diagnostic message, now would be a good time to also consider how readable
the message looks overall, particularly for people new to Rust.
Our example tests/ui/async-await/await-without-async.stderr
file should look
like this:
error[E0728]: `await` is only allowed inside `async` functions and blocks
--> $DIR/await-without-async.rs:7:10
|
LL | fn bar() {
| --- this is not `async`
LL | foo().await
| ^^^^^^ only allowed inside `async` functions and blocks
error: aborting due to previous error
For more information about this error, try `rustc --explain E0728`.
You may notice some things look a little different than the regular compiler output.
- The
$DIR
removes the path information which will differ between systems. - The
LL
values replace the line numbers. That helps avoid small changes in the source from triggering large diffs. See the Normalization section for more.
Around this stage, you may need to iterate over the last few steps a few times to tweak your test, re-bless the test, and re-review the output.
Step 5: Check other tests
Sometimes when adding or changing a diagnostic message, this will affect other tests in the test suite. The final step before posting a PR is to check if you have affected anything else. Running the UI suite is usually a good start:
./x test tests/ui
If other tests start failing, you may need to investigate what has changed and if the new output makes sense.
You may also need to re-bless the output with the --bless
flag.
Comment explaining what the test is about
The first comment of a test file should summarize the point of the test, and highlight what is important about it. If there is an issue number associated with the test, include the issue number.
This comment doesn't have to be super extensive. Just something like "Regression test for #18060: match arms were matching in the wrong order." might already be enough.
These comments are very useful to others later on when your test breaks, since they often can highlight what the problem is. They are also useful if for some reason the tests need to be refactored, since they let others know which parts of the test were important. Often a test must be rewritten because it no longer tests what is was meant to test, and then it's useful to know what it was meant to test exactly.
Best practices for writing tests
This chapter describes best practices related to authoring and modifying tests. We want to make sure the tests we author are easy to understand and modify, even several years later, without needing to consult the original author and perform a bunch of git archeology.
It's good practice to review the test that you authored by pretending that you are a different contributor who is looking at the test that failed several years later without much context (this also helps yourself even a few days or months later!). Then ask yourself: how can I make my life and their lives easier?
To help put this into perspective, let's start with an aside on how to write a test that makes the life of another contributor as hard as possible.
Aside: Simple Test Sabotage Field Manual
To make the life of another contributor as hard as possible, one might:
- Name the test after an issue number alone without any other context, e.g.
issue-123456.rs
.- Have no comments at all on what the test is trying to exercise, no links to relevant context.
- Include a test that is massive (that can otherwise be minimized) and contains non-essential pieces which distracts from the core thing the test is actually trying to test.
- Include a bunch of unrelated syntax errors and other errors which are not critical to what the test is trying to check.
- Weirdly format the snippets.
- Include a bunch of unused and unrelated features.
- Have e.g.
ignore-windows
compiletest directives but don't offer any explanation as to why they are needed.
Test naming
Make it easy for the reader to immediately understand what the test is
exercising, instead of having to type in the issue number and dig through github
search for what the test is trying to exercise. This has an additional benefit
of making the test possible to be filtered via --test-args
as a collection of
related tests.
- Name the test after what it's trying to exercise or prevent regressions of.
- Keep it concise.
- Avoid using issue numbers alone as test names.
- Avoid starting the test name with
issue-xxxxx
prefix as it degrades auto-completion.
Avoid using only issue numbers as test names
Prefer including them as links or
#123456
in test comments instead. Or if it makes sense to include the issue number, also include brief keywords likemacro-external-span-ice-123956.rs
.tests/ui/typeck/issue-123456.rs // bad tests/ui/typeck/issue-123456-asm-macro-external-span-ice.rs // bad (for tab completion) tests/ui/typeck/asm-macro-external-span-ice-123456.rs // good tests/ui/typeck/asm-macro-external-span-ice.rs // good
issue-123456.rs
does not tell you immediately anything about what the test is actually exercising meaning you need to do additional searching. Including the issue number in the test name as a prefix makes tab completion less useful (if youls
a test directory and get a bunch ofissue-xxxxx
prefixes). We can link to the issue in a test comment.//! Check that `asm!` macro including nested macros that come from external //! crates do not lead to a codepoint boundary assertion ICE. //! //! Regression test for <https://github.com/rust-lang/rust/issues/123456>.
Test organization
- For most test suites, try to find a semantically meaningful subdirectory to
home the test.
- E.g. for an implementation of RFC 2093 specifically, we can group a
collection of tests under
tests/ui/rfc-2093-infer-outlives/
. For the directory name, include what the RFC is about.
- E.g. for an implementation of RFC 2093 specifically, we can group a
collection of tests under
- For the
run-make
test suite, eachrmake.rs
must be contained within an immediate subdirectory undertests/run-make/
. Further nesting is not presently supported. Avoid including issue number in the directory name too, include that info in a comment insidermake.rs
.
Test descriptions
To help other contributors understand what the test is about if their changes lead to the test failing, we should make sure a test has sufficient docs about its intent/purpose, links to relevant context (incl. issue numbers or other discussions) and possibly relevant resources (e.g. can be helpful to link to Win32 APIs for specific behavior).
Synopsis of a test with good comments
//! Brief summary of what the test is exercising.
//! Example: Regression test for #123456: make sure coverage attribute don't ICE
//! when applied to non-items.
//!
//! Optional: Remarks on related tests/issues, external APIs/tools, crash
//! mechanism, how it's fixed, FIXMEs, limitations, etc.
//! Example: This test is like `tests/attrs/linkage.rs`, but this test is
//! specifically checking `#[coverage]` which exercises a different code
//! path. The ICE was triggered during attribute validation when we tried
//! to construct a `def_path_str` but only emitted the diagnostic when the
//! platform is windows, causing an ICE on unix.
//!
//! Links to relevant issues and discussions. Examples below:
//! Regression test for <https://github.com/rust-lang/rust/issues/123456>.
//! See also <https://github.com/rust-lang/rust/issues/101345>.
//! See discussion at <https://rust-lang.zulipchat.com/#narrow/stream/131828-t-compiler/topic/123456-example-topic>.
//! See [`clone(2)`].
//!
//! [`clone(2)`]: https://man7.org/linux/man-pages/man2/clone.2.html
//@ ignore-windows
// Reason: (why is this test ignored for windows? why not specifically
// windows-gnu or windows-msvc?)
// Optional: Summary of test cases: What positive cases are checked?
// What negative cases are checked? Any specific quirks?
fn main() {
#[coverage]
//~^ ERROR coverage attribute can only be applied to function items.
let _ = {
// Comment highlighting something that deserves reader attention.
fn foo() {}
};
}
For how much context/explanation is needed, it is up to the author and reviewer's discretion. A good rule of thumb is non-trivial things exercised in the test deserves some explanation to help other contributors to understand. This may include remarks on:
- How an ICE can get triggered if it's quite elaborate.
- Related issues and tests (e.g. this test is like another test but is kept separate because...).
- Platform-specific behaviors.
- Behavior of external dependencies and APIs: syscalls, linkers, tools, environments and the likes.
Test content
- Try to make sure the test is as minimal as possible.
- Minimize non-critical code and especially minimize unnecessary syntax and type errors which can clutter stderr snapshots.
- Where possible, use semantically meaningful names (e.g.
fn bare_coverage_attributes() {}
).
Flaky tests
All tests need to strive to be reproducible and reliable. Flaky tests are the worst kind of tests, arguably even worse than not having the test in the first place.
- Flaky tests can fail in completely unrelated PRs which can confuse other contributors and waste their time trying to figure out if test failure is related.
- Flaky tests provide no useful information from its test results other than it's flaky and not reliable: if a test passed but it's flakey, did I just get lucky? if a test is flakey but it failed, was it just spurious?
- Flaky tests degrade confidence in the whole test suite. If a test suite can randomly spuriously fail due to flaky tests, did the whole test suite pass or did I just get lucky/unlucky?
- Flaky tests can randomly fail in full CI, wasting previous full CI resources.
Compiletest directives
See compiletest directives for a listing of directives.
- For
ignore-*
/needs-*
/only-*
directives, unless extremely obvious, provide a brief remark on why the directive is needed. E.g."//@ ignore-wasi (wasi codegens the main symbol differently)"
.
FileCheck best practices
See LLVM FileCheck guide for details.
- Avoid matching on specific register numbers or basic block numbers unless they're special or critical for the test. Consider using patterns to match them where suitable.
TODO
Pending concrete advice.
Compiletest
- Introduction
- Test suites
- Building auxiliary crates
- Revisions
- Compare modes
Introduction
compiletest
is the main test harness of the Rust test suite. It allows test
authors to organize large numbers of tests (the Rust compiler has many
thousands), efficient test execution (parallel execution is supported), and
allows the test author to configure behavior and expected results of both
individual and groups of tests.
Note for macOS users
For macOS users,
SIP
(System Integrity Protection) may consistently check the compiled binary by sending network requests to Apple, so you may get a huge performance degradation when running tests.You can resolve it by tweaking the following settings:
Privacy & Security -> Developer Tools -> Add Terminal (Or VsCode, etc.)
.
compiletest
may check test code for compile-time or run-time success/failure.
Tests are typically organized as a Rust source file with annotations in comments
before and/or within the test code. These comments serve to direct compiletest
on if or how to run the test, what behavior to expect, and more. See
directives and the test suite documentation below for more details
on these annotations.
See the Adding new tests and Best practies chapters for a tutorial on creating a new test and advice on writing a good test, and the Running tests chapter on how to run the test suite.
Compiletest itself tries to avoid running tests when the artifacts that are
involved (mainly the compiler) haven't changed. You can use x test --test-args --force-rerun
to rerun a test even when none of the inputs have changed.
Test suites
All of the tests are in the tests
directory. The tests are organized into
"suites", with each suite in a separate subdirectory. Each test suite behaves a
little differently, with different compiler behavior and different checks for
correctness. For example, the tests/incremental
directory contains tests for
incremental compilation. The various suites are defined in
src/tools/compiletest/src/common.rs
in the pub enum Mode
declaration.
The following test suites are available, with links for more information:
Compiler-specific test suites
Test suite | Purpose |
---|---|
ui | Check the stdout/stderr snapshots from the compilation and/or running the resulting executable |
ui-fulldeps | ui tests which require a linkable build of rustc (such as using extern crate rustc_span; or used as a plugin) |
pretty | Check pretty printing |
incremental | Check incremental compilation behavior |
debuginfo | Check debuginfo generation running debuggers |
codegen | Check code generation |
codegen-units | Check codegen unit partitioning |
assembly | Check assembly output |
mir-opt | Check MIR generation and optimizations |
coverage | Check coverage instrumentation |
coverage-run-rustdoc | coverage tests that also run instrumented doctests |
General purpose test suite
run-make
are general purpose tests using Rust programs (or
Makefiles (legacy)).
Rustdoc test suites
See Rustdoc tests for more details.
Test suite | Purpose |
---|---|
rustdoc | Check rustdoc generated files contain the expected documentation |
rustdoc-gui | Check rustdoc 's GUI using a web browser |
rustdoc-js | Check rustdoc search is working as expected |
rustdoc-js-std | Check rustdoc search is working as expected specifically on the std docs |
rustdoc-json | Check JSON output of rustdoc |
rustdoc-ui | Check terminal output of rustdoc |
Pretty-printer tests
The tests in tests/pretty
exercise the "pretty-printing" functionality of
rustc
. The -Z unpretty
CLI option for rustc
causes it to translate the
input source into various different formats, such as the Rust source after macro
expansion.
The pretty-printer tests have several directives described below. These commands can significantly change the behavior of the test, but the default behavior without any commands is to:
- Run
rustc -Zunpretty=normal
on the source file. - Run
rustc -Zunpretty=normal
on the output of the previous step. - The output of the previous two steps should be the same.
- Run
rustc -Zno-codegen
on the output to make sure that it can type check (this is similar to runningcargo check
).
If any of the commands above fail, then the test fails.
The directives for pretty-printing tests are:
pretty-mode
specifies the mode pretty-print tests should run in (that is, the argument to-Zunpretty
). The default isnormal
if not specified.pretty-compare-only
causes a pretty test to only compare the pretty-printed output (stopping after step 3 from above). It will not try to compile the expanded output to type check it. This is needed for a pretty-mode that does not expand to valid Rust, or for other situations where the expanded output cannot be compiled.pp-exact
is used to ensure a pretty-print test results in specific output. If specified without a value, then it means the pretty-print output should match the original source. If specified with a value, as in//@ pp-exact:foo.pp
, it will ensure that the pretty-printed output matches the contents of the given file. Otherwise, ifpp-exact
is not specified, then the pretty-printed output will be pretty-printed one more time, and the output of the two pretty-printing rounds will be compared to ensure that the pretty-printed output converges to a steady state.
Incremental tests
The tests in tests/incremental
exercise incremental compilation. They use
revisions
directive to tell compiletest to run the compiler in a
series of steps.
Compiletest starts with an empty directory with the -C incremental
flag, and
then runs the compiler for each revision, reusing the incremental results from
previous steps.
The revisions should start with:
rpass
— the test should compile and run successfullyrfail
— the test should compile successfully, but the executable should fail to runcfail
— the test should fail to compile
To make the revisions unique, you should add a suffix like rpass1
and
rpass2
.
To simulate changing the source, compiletest also passes a --cfg
flag with the
current revision name.
For example, this will run twice, simulating changing a function:
//@ revisions: rpass1 rpass2
#[cfg(rpass1)]
fn foo() {
println!("one");
}
#[cfg(rpass2)]
fn foo() {
println!("two");
}
fn main() { foo(); }
cfail
tests support the forbid-output
directive to specify that a certain
substring must not appear anywhere in the compiler output. This can be useful to
ensure certain errors do not appear, but this can be fragile as error messages
change over time, and a test may no longer be checking the right thing but will
still pass.
cfail
tests support the should-ice
directive to specify that a test should
cause an Internal Compiler Error (ICE). This is a highly specialized directive
to check that the incremental cache continues to work after an ICE.
Debuginfo tests
The tests in tests/debuginfo
test debuginfo generation. They build a
program, launch a debugger, and issue commands to the debugger. A single test
can work with cdb, gdb, and lldb.
Most tests should have the //@ compile-flags: -g
directive or something
similar to generate the appropriate debuginfo.
To set a breakpoint on a line, add a // #break
comment on the line.
The debuginfo tests consist of a series of debugger commands along with "check" lines which specify output that is expected from the debugger.
The commands are comments of the form // $DEBUGGER-command:$COMMAND
where
$DEBUGGER
is the debugger being used and $COMMAND
is the debugger command
to execute.
The debugger values can be:
cdb
gdb
gdbg
— GDB without Rust support (versions older than 7.11)gdbr
— GDB with Rust supportlldb
lldbg
— LLDB without Rust supportlldbr
— LLDB with Rust support (this no longer exists)
The command to check the output are of the form // $DEBUGGER-check:$OUTPUT
where $OUTPUT
is the output to expect.
For example, the following will build the test, start the debugger, set a breakpoint, launch the program, inspect a value, and check what the debugger prints:
//@ compile-flags: -g
//@ lldb-command: run
//@ lldb-command: print foo
//@ lldb-check: $0 = 123
fn main() {
let foo = 123;
b(); // #break
}
fn b() {}
The following directives are available to disable a test based on the debugger currently being used:
min-cdb-version: 10.0.18317.1001
— ignores the test if the version of cdb is below the given versionmin-gdb-version: 8.2
— ignores the test if the version of gdb is below the given versionignore-gdb-version: 9.2
— ignores the test if the version of gdb is equal to the given versionignore-gdb-version: 7.11.90 - 8.0.9
— ignores the test if the version of gdb is in a range (inclusive)min-lldb-version: 310
— ignores the test if the version of lldb is below the given versionrust-lldb
— ignores the test if lldb is not contain the Rust plugin. NOTE: The "Rust" version of LLDB doesn't exist anymore, so this will always be ignored. This should probably be removed.
Note on running lldb debuginfo tests locally
If you want to run lldb debuginfo tests locally, then currently on Windows it is required that:
- You have Python 3.10 installed.
- You have
python310.dll
available in yourPATH
env var. This is not provided by the standard Python installer you obtain frompython.org
; you need to add this toPATH
manually.Otherwise the lldb debuginfo tests can produce crashes in mysterious ways.
Note on acquiring
cdb.exe
on Windows 11
cdb.exe
is acquired alongside a suitable "Windows 11 SDK" which is part of the "Desktop Development with C++" workload profile in a Visual Studio installer (e.g. Visual Studio 2022 installer).HOWEVER this is not sufficient by default alone. If you need
cdb.exe
, you must go to Installed Apps, find the newest "Windows Software Development Kit" (and yes, this can still sayWindows 10.0.22161.3233
even though the OS is called Windows 11). You must then click "Modify" -> "Change" and then selected "Debugging Tools for Windows" in order to acquirecdb.exe
.
Codegen tests
The tests in tests/codegen
test LLVM code generation. They compile the test
with the --emit=llvm-ir
flag to emit LLVM IR. They then run the LLVM
FileCheck tool. The test is annotated with various // CHECK
comments to
check the generated code. See the FileCheck documentation for a tutorial and
more information.
See also the assembly tests for a similar set of tests.
If you need to work with #![no_std]
cross-compiling tests, consult the
minicore
test auxiliary chapter.
Assembly tests
The tests in tests/assembly
test LLVM assembly output. They compile the test
with the --emit=asm
flag to emit a .s
file with the assembly output. They
then run the LLVM FileCheck tool.
Each test should be annotated with the //@ assembly-output:
directive with a
value of either emit-asm
or ptx-linker
to indicate the type of assembly
output.
Then, they should be annotated with various // CHECK
comments to check the
assembly output. See the FileCheck documentation for a tutorial and more
information.
See also the codegen tests for a similar set of tests.
If you need to work with #![no_std]
cross-compiling tests, consult the
minicore
test auxiliary chapter.
Codegen-units tests
The tests in tests/codegen-units
test the
monomorphization collector and CGU partitioning.
These tests work by running rustc
with a flag to print the result of the
monomorphization collection pass, and then special annotations in the file are
used to compare against that.
Each test should be annotated with the //@ compile-flags:-Zprint-mono-items=VAL
directive with the appropriate VAL
to
instruct rustc
to print the monomorphization information.
Then, the test should be annotated with comments of the form //~ MONO_ITEM name
where name
is the monomorphized string printed by rustc like fn <u32 as Trait>::foo
.
To check for CGU partitioning, a comment of the form //~ MONO_ITEM name @@ cgu
where cgu
is a space separated list of the CGU names and the linkage
information in brackets. For example: //~ MONO_ITEM static function::FOO @@ statics[Internal]
Mir-opt tests
The tests in tests/mir-opt
check parts of the generated MIR to make sure it
is generated correctly and is doing the expected optimizations. Check out the
MIR Optimizations chapter for more.
Compiletest will build the test with several flags to dump the MIR output and set a baseline for optimizations:
-Copt-level=1
-Zdump-mir=all
-Zmir-opt-level=4
-Zvalidate-mir
-Zdump-mir-exclude-pass-number
The test should be annotated with // EMIT_MIR
comments that specify files that
will contain the expected MIR output. You can use x test --bless
to create the
initial expected files.
There are several forms the EMIT_MIR
comment can take:
-
// EMIT_MIR $MIR_PATH.mir
— This will check that the given filename matches the exact output from the MIR dump. For example,my_test.main.SimplifyCfg-elaborate-drops.after.mir
will load that file from the test directory, and compare it against the dump from rustc.Checking the "after" file (which is after optimization) is useful if you are interested in the final state after an optimization. Some rare cases may want to use the "before" file for completeness.
-
// EMIT_MIR $MIR_PATH.diff
— where$MIR_PATH
is the filename of the MIR dump, such asmy_test_name.my_function.EarlyOtherwiseBranch
. Compiletest will diff the.before.mir
and.after.mir
files, and compare the diff output to the expected.diff
file from theEMIT_MIR
comment.This is useful if you want to see how an optimization changes the MIR.
-
// EMIT_MIR $MIR_PATH.dot
— When using specific flags that dump additional MIR data (e.g.-Z dump-mir-graphviz
to produce.dot
files), this will check that the output matches the given file.
By default 32 bit and 64 bit targets use the same dump files, which can be
problematic in the presence of pointers in constants or other bit width
dependent things. In that case you can add // EMIT_MIR_FOR_EACH_BIT_WIDTH
to
your test, causing separate files to be generated for 32bit and 64bit systems.
run-make
tests
Note on phasing out
Makefile
sWe are planning to migrate all existing Makefile-based
run-make
tests to Rust programs. You should not be adding new Makefile-basedrun-make
tests.
The tests in tests/run-make
are general-purpose tests using Rust recipes,
which are small programs (rmake.rs
) allowing arbitrary Rust code such as
rustc
invocations, and is supported by a run_make_support
library. Using
Rust recipes provide the ultimate in flexibility.
run-make
tests should be used if no other test suites better suit your needs.
Using Rust recipes
Each test should be in a separate directory with a rmake.rs
Rust program,
called the recipe. A recipe will be compiled and executed by compiletest with
the run_make_support
library linked in.
If you need new utilities or functionality, consider extending and improving the
run_make_support
library.
Compiletest directives like //@ only-<target>
or //@ ignore-<target>
are
supported in rmake.rs
, like in UI tests. However, revisions or building
auxiliary via directives are not currently supported.
Two run-make
tests are ported over to Rust recipes as examples:
- https://github.com/rust-lang/rust/tree/master/tests/run-make/CURRENT_RUSTC_VERSION
- https://github.com/rust-lang/rust/tree/master/tests/run-make/a-b-a-linker-guard
Quickly check if rmake.rs
tests can be compiled
You can quickly check if rmake.rs
tests can be compiled without having to
build stage1 rustc by forcing rmake.rs
to be compiled with the stage0
compiler:
$ COMPILETEST_FORCE_STAGE0=1 x test --stage 0 tests/run-make/<test-name>
Of course, some tests will not successfully run in this way.
Using Makefiles (legacy)
Each test should be in a separate directory with a Makefile
indicating the
commands to run.
There is a tools.mk
Makefile which you can include which provides a bunch of
utilities to make it easier to run commands and compare outputs. Take a look at
some of the other tests for some examples on how to get started.
Coverage tests
The tests in tests/coverage
are shared by multiple test modes that test
coverage instrumentation in different ways. Running the coverage
test suite
will automatically run each test in all of the different coverage modes.
Each mode also has an alias to run the coverage tests in just that mode:
./x test coverage # runs all of tests/coverage in all coverage modes
./x test tests/coverage # same as above
./x test tests/coverage/if.rs # runs the specified test in all coverage modes
./x test coverage-map # runs all of tests/coverage in "coverage-map" mode only
./x test coverage-run # runs all of tests/coverage in "coverage-run" mode only
./x test coverage-map -- tests/coverage/if.rs # runs the specified test in "coverage-map" mode only
If a particular test should not be run in one of the coverage test modes for
some reason, use the //@ ignore-coverage-map
or //@ ignore-coverage-run
directives.
coverage-map
suite
In coverage-map
mode, these tests verify the mappings between source code
regions and coverage counters that are emitted by LLVM. They compile the test
with --emit=llvm-ir
, then use a custom tool (src/tools/coverage-dump
) to
extract and pretty-print the coverage mappings embedded in the IR. These tests
don't require the profiler runtime, so they run in PR CI jobs and are easy to
run/bless locally.
These coverage map tests can be sensitive to changes in MIR lowering or MIR optimizations, producing mappings that are different but produce identical coverage reports.
As a rule of thumb, any PR that doesn't change coverage-specific code should
feel free to re-bless the coverage-map
tests as necessary, without
worrying about the actual changes, as long as the coverage-run
tests still
pass.
coverage-run
suite
In coverage-run
mode, these tests perform an end-to-end test of coverage
reporting. They compile a test program with coverage instrumentation, run that
program to produce raw coverage data, and then use LLVM tools to process that
data into a human-readable code coverage report.
Instrumented binaries need to be linked against the LLVM profiler runtime, so
coverage-run
tests are automatically skipped unless the profiler runtime
is enabled in config.toml
:
# config.toml
[build]
profiler = true
This also means that they typically don't run in PR CI jobs, though they do run as part of the full set of CI jobs used for merging.
coverage-run-rustdoc
suite
The tests in tests/coverage-run-rustdoc
also run instrumented doctests and
include them in the coverage report. This avoids having to build rustdoc when
only running the main coverage
suite.
Crashes tests
tests/crashes
serve as a collection of tests that are expected to cause the
compiler to ICE, panic or crash in some other way, so that accidental fixes are
tracked. This was formally done at https://github.com/rust-lang/glacier but
doing it inside the rust-lang/rust testsuite is more convenient.
It is imperative that a test in the suite causes rustc to ICE, panic or crash crash in some other way. A test will "pass" if rustc exits with an exit status other than 1 or 0.
If you want to see verbose stdout/stderr, you need to set
COMPILETEST_VERBOSE_CRASHES=1
, e.g.
$ COMPILETEST_VERBOSE_CRASHES=1 ./x test tests/crashes/999999.rs --stage 1
When adding crashes from https://github.com/rust-lang/rust/issues, the issue
number should be noted in the file name (12345.rs
should suffice) and also
inside the file include a //@ known-bug: #4321
directive.
If you happen to fix one of the crashes, please move it to a fitting
subdirectory in tests/ui
and give it a meaningful name. Please add a doc
comment at the top of the file explaining why this test exists, even better if
you can briefly explain how the example causes rustc to crash previously and
what was done to prevent rustc to ICE/panic/crash.
Adding
Fixes #NNNNN
Fixes #MMMMM
to the description of your pull request will ensure the corresponding tickets be closed automatically upon merge.
Make sure that your fix actually fixes the root cause of the issue and not just
a subset first. The issue numbers can be found in the file name or the //@ known-bug
directive inside the test file.
Building auxiliary crates
It is common that some tests require additional auxiliary crates to be compiled. There are multiple directives to assist with that:
aux-build
aux-crate
aux-bin
aux-codegen-backend
proc-macro
aux-build
will build a separate crate from the named source file. The source
file should be in a directory called auxiliary
beside the test file.
//@ aux-build: my-helper.rs
extern crate my_helper;
// ... You can use my_helper.
The aux crate will be built as a dylib if possible (unless on a platform that
does not support them, or the no-prefer-dynamic
header is specified in the aux
file). The -L
flag is used to find the extern crates.
aux-crate
is very similar to aux-build
. However, it uses the --extern
flag
to link to the extern crate to make the crate be available as an extern prelude.
That allows you to specify the additional syntax of the --extern
flag, such as
renaming a dependency. For example, // aux-crate:foo=bar.rs
will compile
auxiliary/bar.rs
and make it available under then name foo
within the test.
This is similar to how Cargo does dependency renaming.
aux-bin
is similar to aux-build
but will build a binary instead of a
library. The binary will be available in auxiliary/bin
relative to the working
directory of the test.
aux-codegen-backend
is similar to aux-build
, but will then pass the compiled
dylib to -Zcodegen-backend
when building the main file. This will only work
for tests in tests/ui-fulldeps
, since it requires the use of compiler crates.
Auxiliary proc-macro
If you want a proc-macro dependency, then you can use the proc-macro
directive. This directive behaves just like aux-build
, i.e. that you should
place the proc-macro test auxiliary file under a auxiliary
folder under the
same parent folder as the main test file. However, it also has four additional
preset behavior compared to aux-build
for the proc-macro test auxiliary:
- The aux test file is built with
--crate-type=proc-macro
. - The aux test file is built without
-C prefer-dynamic
, i.e. it will not try to produce a dylib for the aux crate. - The aux crate is made available to the test file via extern prelude with
--extern <aux_crate_name>
. Note that since UI tests default to edition 2015, you still need to specifyextern <aux_crate_name>
unless the main test file is using an edition that is 2018 or newer if you want to use the aux crate name in ause
import. - The
proc_macro
crate is made available as an extern prelude module. Same edition 2015 vs newer edition distinction forextern proc_macro;
applies.
For example, you might have a test tests/ui/cat/meow.rs
and proc-macro
auxiliary tests/ui/cat/auxiliary/whiskers.rs
:
tests/ui/cat/
meow.rs # main test file
auxiliary/whiskers.rs # auxiliary
// tests/ui/cat/meow.rs
//@ proc-macro: whiskers.rs
extern crate whiskers; // needed as ui test defaults to edition 2015
fn main() {
whiskers::identity!();
}
// tests/ui/cat/auxiliary/whiskers.rs
extern crate proc_macro;
use proc_macro::*;
#[proc_macro]
pub fn identity(ts: TokenStream) -> TokenStream {
ts
}
Note: The
proc-macro
header currently does not work with thebuild-aux-doc
header for rustdoc tests. In that case, you will need to use theaux-build
header, and use#![crate_type="proc_macro"]
, and//@ force-host
and//@ no-prefer-dynamic
headers in the proc-macro.
Revisions
Revisions allow a single test file to be used for multiple tests. This is done by adding a special directive at the top of the file:
//@ revisions: foo bar baz
This will result in the test being compiled (and tested) three times, once with
--cfg foo
, once with --cfg bar
, and once with --cfg baz
. You can therefore
use #[cfg(foo)]
etc within the test to tweak each of these results.
You can also customize directives and expected error messages to a particular
revision. To do this, add [revision-name]
after the //@
for directives, and
after //
for UI error annotations, like so:
// A flag to pass in only for cfg `foo`:
//@[foo]compile-flags: -Z verbose-internals
#[cfg(foo)]
fn test_foo() {
let x: usize = 32_u32; //[foo]~ ERROR mismatched types
}
Multiple revisions can be specified in a comma-separated list, such as
//[foo,bar,baz]~^
.
In test suites that use the LLVM FileCheck tool, the current revision name is also registered as an additional prefix for FileCheck directives:
//@ revisions: NORMAL COVERAGE
//@[COVERAGE] compile-flags: -Cinstrument-coverage
//@[COVERAGE] needs-profiler-runtime
// COVERAGE: @__llvm_coverage_mapping
// NORMAL-NOT: @__llvm_coverage_mapping
// CHECK: main
fn main() {}
Note that not all directives have meaning when customized to a revision. For
example, the ignore-test
directives (and all "ignore" directives) currently
only apply to the test as a whole, not to particular revisions. The only
directives that are intended to really work when customized to a revision are
error patterns and compiler flags.
The following test suites support revisions:
- ui
- assembly
- codegen
- coverage
- debuginfo
- rustdoc UI tests
- incremental (these are special in that they inherently cannot be run in parallel)
Ignoring unused revision names
Normally, revision names mentioned in other directives and error annotations
must correspond to an actual revision declared in a revisions
directive. This is
enforced by an ./x test tidy
check.
If a revision name needs to be temporarily removed from the revision list for
some reason, the above check can be suppressed by adding the revision name to an
//@ unused-revision-names:
header instead.
Specifying an unused name of *
(i.e. //@ unused-revision-names: *
) will
permit any unused revision name to be mentioned.
Compare modes
Compiletest can be run in different modes, called compare modes, which can be used to compare the behavior of all tests with different compiler flags enabled. This can help highlight what differences might appear with certain flags, and check for any problems that might arise.
To run the tests in a different mode, you need to pass the --compare-mode
CLI
flag:
./x test tests/ui --compare-mode=chalk
The possible compare modes are:
polonius
— Runs with Polonius with-Zpolonius
.chalk
— Runs with Chalk with-Zchalk
.split-dwarf
— Runs with unpacked split-DWARF with-Csplit-debuginfo=unpacked
.split-dwarf-single
— Runs with packed split-DWARF with-Csplit-debuginfo=packed
.
See UI compare modes for more information about how UI tests support different output for different modes.
In CI, compare modes are only used in one Linux builder, and only with the following settings:
tests/debuginfo
: Usessplit-dwarf
mode. This helps ensure that none of the debuginfo tests are affected when enabling split-DWARF.
Note that compare modes are separate to revisions. All revisions
are tested when running ./x test tests/ui
, however compare-modes must be
manually run individually via the --compare-mode
flag.
UI tests
- Introduction
- General structure of a test
- Output comparison
- Error annotations
- Controlling pass/fail expectations
- Known bugs
- Test organization
- Rustfix tests
- Compare modes
rustc_*
TEST attributes
UI tests are a particular test suite of compiletest.
Introduction
The tests in tests/ui
are a collection of general-purpose tests which
primarily focus on validating the console output of the compiler, but can be
used for many other purposes. For example, tests can also be configured to run
the resulting program to verify its
behavior.
If you need to work with #![no_std]
cross-compiling tests, consult the
minicore
test auxiliary chapter.
General structure of a test
A test consists of a Rust source file located anywhere in the tests/ui
directory, but they should be placed in a suitable sub-directory. For example,
tests/ui/hello.rs
is a basic hello-world test.
Compiletest will use rustc
to compile the test, and compare the output against
the expected output which is stored in a .stdout
or .stderr
file located
next to the test. See Output comparison for more.
Additionally, errors and warnings should be annotated with comments within the source file. See Error annotations for more.
Compiletest directives in the form of special comments prefixed
with //@
control how the test is compiled and what the expected behavior is.
Tests are expected to fail to compile, since most tests are testing compiler errors. You can change that behavior with a directive, see Controlling pass/fail expectations.
By default, a test is built as an executable binary. If you need a different
crate type, you can use the #![crate_type]
attribute to set it as needed.
Output comparison
UI tests store the expected output from the compiler in .stderr
and .stdout
snapshots next to the test. You normally generate these files with the --bless
CLI option, and then inspect them manually to verify they contain what you
expect.
The output is normalized to ignore unwanted differences, see the Normalization section. If the file is missing, then compiletest expects the corresponding output to be empty.
There can be multiple stdout/stderr files. The general form is:
*test-name*`.`*revision*`.`*compare_mode*`.`*extension*
- test-name cannot contain dots. This is so that the general form of test output filenames have a predictable form we can pattern match on in order to track stray test output files.
- revision is the revision name. This is not included when not using revisions.
- compare_mode is the compare mode. This will only be checked when the given compare mode is active. If the file does not exist, then compiletest will check for a file without the compare mode.
- extension is the kind of output being checked:
stderr
— compiler stderrstdout
— compiler stdoutrun.stderr
— stderr when running the testrun.stdout
— stdout when running the test64bit.stderr
— compiler stderr withstderr-per-bitwidth
directive on a 64-bit target32bit.stderr
— compiler stderr withstderr-per-bitwidth
directive on a 32-bit target
A simple example would be foo.stderr
next to a foo.rs
test.
A more complex example would be foo.my-revision.polonius.stderr
.
There are several directives which will change how compiletest will check for output files:
stderr-per-bitwidth
— checks separate output files based on the target pointer width. Consider using thenormalize-stderr
directive instead (see Normalization).dont-check-compiler-stderr
— Ignores stderr from the compiler.dont-check-compiler-stdout
— Ignores stdout from the compiler.
UI tests run with -Zdeduplicate-diagnostics=no
flag which disables rustc's
built-in diagnostic deduplication mechanism. This means you may see some
duplicate messages in the output. This helps illuminate situations where
duplicate diagnostics are being generated.
Normalization
The compiler output is normalized to eliminate output difference between platforms, mainly about filenames.
Compiletest makes the following replacements on the compiler output:
- The directory where the test is defined is replaced with
$DIR
. Example:/path/to/rust/tests/ui/error-codes
- The directory to the standard library source is replaced with
$SRC_DIR
. Example:/path/to/rust/library
- Line and column numbers for paths in
$SRC_DIR
are replaced withLL:COL
. This helps ensure that changes to the layout of the standard library do not cause widespread changes to the.stderr
files. Example:$SRC_DIR/alloc/src/sync.rs:53:46
- The base directory where the test's output goes is replaced with
$TEST_BUILD_DIR
. This only comes up in a few rare circumstances. Example:/path/to/rust/build/x86_64-unknown-linux-gnu/test/ui
- Tabs are replaced with
\t
. - Backslashes (
\
) are converted to forward slashes (/
) within paths (using a heuristic). This helps normalize differences with Windows-style paths. - CRLF newlines are converted to LF.
- Error line annotations like
//~ ERROR some message
are removed. - Various v0 and legacy symbol hashes are replaced with placeholders like
[HASH]
or<SYMBOL_HASH>
.
Additionally, the compiler is run with the -Z ui-testing
flag which causes
the compiler itself to apply some changes to the diagnostic output to make it
more suitable for UI testing.
For example, it will anonymize line numbers in the output (line numbers
prefixing each source line are replaced with LL
). In extremely rare
situations, this mode can be disabled with the directive //@ compile-flags: -Z ui-testing=no
.
Note: The line and column numbers for -->
lines pointing to the test are not
normalized, and left as-is. This ensures that the compiler continues to point to
the correct location, and keeps the stderr files readable. Ideally all
line/column information would be retained, but small changes to the source
causes large diffs, and more frequent merge conflicts and test errors.
Sometimes these built-in normalizations are not enough. In such cases, you may
provide custom normalization rules using normalize-*
directives, e.g.
//@ normalize-stdout-test: "foo" -> "bar"
//@ normalize-stderr-32bit: "fn\(\) \(32 bits\)" -> "fn\(\) \($$PTR bits\)"
//@ normalize-stderr-64bit: "fn\(\) \(64 bits\)" -> "fn\(\) \($$PTR bits\)"
This tells the test, on 32-bit platforms, whenever the compiler writes fn() (32 bits)
to stderr, it should be normalized to read fn() ($PTR bits)
instead.
Similar for 64-bit. The replacement is performed by regexes using default regex
flavor provided by regex
crate.
The corresponding reference file will use the normalized output to test both 32-bit and 64-bit platforms:
...
|
= note: source type: fn() ($PTR bits)
= note: target type: u16 (16 bits)
...
Please see ui/transmute/main.rs
and main.stderr
for a concrete
usage example.
Besides normalize-stderr-32bit
and -64bit
, one may use any target
information or stage supported by ignore-X
here
as well (e.g. normalize-stderr-windows
or simply normalize-stderr-test
for
unconditional replacement).
Error annotations
Error annotations specify the errors that the compiler is expected to emit. They are "attached" to the line in source where the error is located.
fn main() {
boom //~ ERROR cannot find value `boom` in this scope [E0425]
}
Although UI tests have a .stderr
file which contains the entire compiler
output, UI tests require that errors are also annotated within the source. This
redundancy helps avoid mistakes since the .stderr
files are usually
auto-generated. It also helps to directly see where the error spans are expected
to point to by looking at one file instead of having to compare the .stderr
file with the source. Finally, they ensure that no additional unexpected errors
are generated.
They have several forms, but generally are a comment with the diagnostic level
(such as ERROR
) and a substring of the expected error output. You don't have
to write out the entire message, just make sure to include the important part of
the message to make it self-documenting.
The error annotation needs to match with the line of the diagnostic. There are several ways to match the message with the line (see the examples below):
~
: Associates the error level and message with the current line~^
: Associates the error level and message with the previous error annotation line. Each caret (^
) that you add adds a line to this, so~^^^
is three lines above the error annotation line.~|
: Associates the error level and message with the same line as the previous comment. This is more convenient than using multiple carets when there are multiple messages associated with the same line.
Example:
let _ = same_line; //~ ERROR undeclared variable
fn meow(_: [u8]) {}
//~^ ERROR unsized
//~| ERROR anonymous parameters
The space character between //~
(or other variants) and the subsequent text is
negligible (i.e. there is no semantic difference between //~ ERROR
and
//~ERROR
although the former is more common in the codebase).
Error annotation examples
Here are examples of error annotations on different lines of UI test source.
Positioned on error line
Use the //~ ERROR
idiom:
fn main() {
let x = (1, 2, 3);
match x {
(_a, _x @ ..) => {} //~ ERROR `_x @` is not allowed in a tuple
_ => {}
}
}
Positioned below error line
Use the //~^
idiom with number of carets in the string to indicate the number
of lines above. In the example below, the error line is four lines above the
error annotation line so four carets are included in the annotation.
fn main() {
let x = (1, 2, 3);
match x {
(_a, _x @ ..) => {} // <- the error is on this line
_ => {}
}
}
//~^^^^ ERROR `_x @` is not allowed in a tuple
Use same error line as defined on error annotation line above
Use the //~|
idiom to define the same error line as the error annotation
line above:
struct Binder(i32, i32, i32);
fn main() {
let x = Binder(1, 2, 3);
match x {
Binder(_a, _x @ ..) => {} // <- the error is on this line
_ => {}
}
}
//~^^^^ ERROR `_x @` is not allowed in a tuple struct
//~| ERROR this pattern has 1 field, but the corresponding tuple struct has 3 fields [E0023]
error-pattern
The error-pattern
directive can be used for messages that don't
have a specific span.
Let's think about this test:
fn main() {
let a: *const [_] = &[1, 2, 3];
unsafe {
let _b = (*a)[3];
}
}
We want to ensure this shows "index out of bounds" but we cannot use the ERROR
annotation since the error doesn't have any span. Then it's time to use the
error-pattern
directive:
//@ error-pattern: index out of bounds
fn main() {
let a: *const [_] = &[1, 2, 3];
unsafe {
let _b = (*a)[3];
}
}
But for strict testing, try to use the ERROR
annotation as much as possible.
Error levels
The error levels that you can have are:
ERROR
WARN
orWARNING
NOTE
HELP
andSUGGESTION
You are allowed to not include a level, but you should include it at least for the primary message.
The SUGGESTION
level is used for specifying what the expected replacement text
should be for a diagnostic suggestion.
UI tests use the -A unused
flag by default to ignore all unused warnings, as
unused warnings are usually not the focus of a test. However, simple code
samples often have unused warnings. If the test is specifically testing an
unused warning, just add the appropriate #![warn(unused)]
attribute as needed.
cfg
revisions
When using revisions, different messages can be conditionally checked based on the current revision. This is done by placing the revision cfg name in brackets like this:
//@ edition:2018
//@ revisions: mir thir
//@[thir] compile-flags: -Z thir-unsafeck
async unsafe fn f() {}
async fn g() {
f(); //~ ERROR call to unsafe function is unsafe
}
fn main() {
f(); //[mir]~ ERROR call to unsafe function is unsafe
}
In this example, the second error message is only emitted in the mir
revision.
The thir
revision only emits the first error.
If the cfg
causes the compiler to emit different output, then a test can have
multiple .stderr
files for the different outputs. In the example above, there
would be a .mir.stderr
and .thir.stderr
file with the different outputs of
the different revisions.
Note: cfg revisions also work inside the source code with
#[cfg]
attributes.By convention, the
FALSE
cfg is used to have an always-false config.
Controlling pass/fail expectations
By default, a UI test is expected to generate a compile error because most of the tests are checking for invalid input and error diagnostics. However, you can also make UI tests where compilation is expected to succeed, and you can even run the resulting program. Just add one of the following directives:
- Pass directives:
//@ check-pass
— compilation should succeed but skip codegen (which is expensive and isn't supposed to fail in most cases).//@ build-pass
— compilation and linking should succeed but do not run the resulting binary.//@ run-pass
— compilation should succeed and running the resulting binary should also succeed.
- Fail directives:
//@ check-fail
— compilation should fail (the codegen phase is skipped). This is the default for UI tests.//@ build-fail
— compilation should fail during the codegen phase. This will runrustc
twice, once to verify that it compiles successfully without the codegen phase, then a second time the full compile should fail.//@ run-fail
— compilation should succeed, but running the resulting binary should fail.
For run-pass
and run-fail
tests, by default the output of the program itself
is not checked.
If you want to check the output of running the program, include the
check-run-results
directive. This will check for a .run.stderr
and
.run.stdout
files to compare against the actual output of the program.
Tests with the *-pass
directives can be overridden with the --pass
command-line option:
./x test tests/ui --pass check
The --pass
option only affects UI tests. Using --pass check
can run the UI
test suite much faster (roughly twice as fast on my system), though obviously
not exercising as much.
The ignore-pass
directive can be used to ignore the --pass
CLI flag if the
test won't work properly with that override.
Known bugs
The known-bug
directive may be used for tests that demonstrate a known bug
that has not yet been fixed. Adding tests for known bugs is helpful for several
reasons, including:
- Maintaining a functional test that can be conveniently reused when the bug is fixed.
- Providing a sentinel that will fail if the bug is incidentally fixed. This can alert the developer so they know that the associated issue has been fixed and can possibly be closed.
Do not include error annotations in a test with
known-bug
. The test should still include other normal directives and
stdout/stderr files.
Test organization
When deciding where to place a test file, please try to find a subdirectory that best matches what you are trying to exercise. Do your best to keep things organized. Admittedly it can be difficult as some tests can overlap different categories, and the existing layout may not fit well.
Name the test by a concise description of what the test is checking. Avoid including the issue number in the test name. See best practices for a more in-depth discussion of this.
Ideally, the test should be added to a directory that helps identify what piece
of code is being tested here (e.g.,
tests/ui/borrowck/reject-move-out-of-borrow-via-pat.rs
)
When writing a new feature, you may want to create a subdirectory to store
your tests. For example, if you are implementing RFC 1234 ("Widgets"), then it
might make sense to put the tests in a directory like
tests/ui/rfc1234-widgets/
.
In other cases, there may already be a suitable directory.
Over time, the tests/ui
directory has grown very fast. There is a check in
tidy that will ensure none of the subdirectories has more than
1000 entries. Having too many files causes problems because it isn't editor/IDE
friendly and the GitHub UI won't show more than 1000 entries. However, since
tests/ui
(UI test root directory) and tests/ui/issues
directories have more
than 1000 entries, we set a different limit for those directories. So, please
avoid putting a new test there and try to find a more relevant place.
For example, if your test is related to closures, you should put it in
tests/ui/closures
. When you reach the limit, you could increase it by tweaking
here.
Rustfix tests
UI tests can validate that diagnostic suggestions apply correctly and that the
resulting changes compile correctly. This can be done with the run-rustfix
directive:
//@ run-rustfix
//@ check-pass
#![crate_type = "lib"]
pub struct not_camel_case {}
//~^ WARN `not_camel_case` should have an upper camel case name
//~| HELP convert the identifier to upper camel case
//~| SUGGESTION NotCamelCase
Rustfix tests should have a file with the .fixed
extension which contains the
source file after the suggestion has been applied.
- When the test is run, compiletest first checks that the correct lint/warning is generated.
- Then, it applies the suggestion and compares against
.fixed
(they must match). - Finally, the fixed source is compiled, and this compilation is required to succeed.
Usually when creating a rustfix test you will generate the .fixed
file
automatically with the x test --bless
option.
The run-rustfix
directive will cause all suggestions to be applied, even if
they are not MachineApplicable
. If this is a
problem, then you can add the rustfix-only-machine-applicable
directive in
addition to run-rustfix
. This should be used if there is a mixture of
different suggestion levels, and some of the non-machine-applicable ones do not
apply cleanly.
Compare modes
Compare modes can be used to run all tests with different flags from what they are normally compiled with. In some cases, this might result in different output from the compiler. To support this, different output files can be saved which contain the output based on the compare mode.
For example, when using the Polonius mode, a test foo.rs
will first look for
expected output in foo.polonius.stderr
, falling back to the usual foo.stderr
if not found. This is useful as different modes can sometimes result in
different diagnostics and behavior. This can help track which tests have
differences between the modes, and to visually inspect those diagnostic
differences.
If in the rare case you encounter a test that has different behavior, you can run something like the following to generate the alternate stderr file:
./x test tests/ui --compare-mode=polonius --bless
Currently none of the compare modes are checked in CI for UI tests.
rustc_*
TEST attributes
The compiler defines several perma-unstable #[rustc_*]
attributes gated behind
the internal feature rustc_attrs
that dump extra compiler-internal
information. See the corresponding subsection in compiler debugging for more
details.
They can be used in tests to more precisely, legibly and easily test internal compiler state in cases where it would otherwise be very hard to do the same with "user-facing" Rust alone. Indeed, one could say that this slightly abuses the term "UI" (user interface) and turns such UI tests from black-box tests into white-box ones. Use them carefully and sparingly.
Compiletest directives
FIXME(jieyouxu) completely revise this chapter.
Directives are special comments that tell compiletest how to build and interpret
a test. They must appear before the Rust source in the test. They may also
appear in rmake.rs
or legacy Makefiles for run-make
tests.
They are normally put after the short comment that explains the point of this
test. Compiletest test suites use //@
to signal that a comment is a directive.
For example, this test uses the //@ compile-flags
command to specify a custom
flag to give to rustc when the test is compiled:
// Test the behavior of `0 - 1` when overflow checks are disabled.
//@ compile-flags: -C overflow-checks=off
fn main() {
let x = 0 - 1;
...
}
Directives can be standalone (like //@ run-pass
) or take a value (like //@ compile-flags: -C overflow-checks=off
).
Directives are written one directive per line: you cannot write multiple
directives on the same line. For example, if you write //@ only-x86 only-windows
then only-windows
is interpreted as a comment, not a separate
directive.
Listing of compiletest directives
The following is a list of compiletest directives. Directives are linked to
sections that describe the command in more detail if available. This list may
not be exhaustive. Directives can generally be found by browsing the
TestProps
structure found in header.rs
from the compiletest source.
Assembly
Directive | Explanation | Supported test suites | Possible values |
---|---|---|---|
assembly-output | Assembly output kind to check | assembly | emit-asm , bpf-linker , ptx-linker |
Auxiliary builds
Directive | Explanation | Supported test suites | Possible values |
---|---|---|---|
aux-bin | Build a aux binary, made available in auxiliary/bin relative to test directory | All except run-make | Path to auxiliary .rs file |
aux-build | Build a separate crate from the named source file | All except run-make | Path to auxiliary .rs file |
aux-crate | Like aux-build but makes available as extern prelude | All except run-make | <extern_prelude_name>=<path/to/aux/file.rs> |
aux-codegen-backend | Similar to aux-build but pass the compiled dylib to -Zcodegen-backend when building the main file | ui-fulldeps | Path to codegen backend file |
proc-macro | Similar to aux-build , but for aux forces host and don't use -Cprefer-dynamic 1. | All except run-make | Path to auxiliary proc-macro .rs file |
build_aux_docs | Build docs for auxiliaries as well | All except run-make | N/A |
please see the Auxiliary proc-macro section in the compiletest chapter for specifics.
Controlling outcome expectations
See Controlling pass/fail expectations.
Directive | Explanation | Supported test suites | Possible values |
---|---|---|---|
check-pass | Building (no codegen) should pass | ui , crashes , incremental | N/A |
check-fail | Building (no codegen) should fail | ui , crashes | N/A |
build-pass | Building should pass | ui , crashes , codegen , incremental | N/A |
build-fail | Building should fail | ui , crashes | N/A |
run-pass | Running the test binary should pass | ui , crashes , incremental | N/A |
run-fail | Running the test binary should fail | ui , crashes | N/A |
ignore-pass | Ignore --pass flag | ui , crashes , codegen , incremental | N/A |
dont-check-failure-status | Don't check exact failure status (i.e. 1 ) | ui , incremental | N/A |
failure-status | Check | ui , crashes | Any u16 |
should-ice | Check failure status is 101 | coverage , incremental | N/A |
should-fail | Compiletest self-test | All | N/A |
Controlling output snapshots and normalizations
See Normalization, Output comparison and Rustfix tests for more details.
Directive | Explanation | Supported test suites | Possible values |
---|---|---|---|
check-run-results | Check run test binary run-{pass,fail} output snapshot | ui , crashes , incremental if run-pass | N/A |
error-pattern | Check that output contains a regex pattern | ui , crashes , incremental if run-pass | Regex |
check-stdout | Check stdout against error-pattern s from running test binary2 | ui , crashes , incremental | N/A |
normalize-stderr-32bit | Normalize actual stderr (for 32-bit platforms) with a rule "<raw>" -> "<normalized>" before comparing against snapshot | ui , incremental | "<RAW>" -> "<NORMALIZED>" , <RAW> /<NORMALIZED> is regex capture and replace syntax |
normalize-stderr-64bit | Normalize actual stderr (for 64-bit platforms) with a rule "<raw>" -> "<normalized>" before comparing against snapshot | ui , incremental | "<RAW>" -> "<NORMALIZED>" , <RAW> /<NORMALIZED> is regex capture and replace syntax |
normalize-stderr-test | Normalize actual stderr with a rule "<raw>" -> "<normalized>" before comparing against snapshot | ui , incremental | "<RAW>" -> "<NORMALIZED>" , <RAW> /<NORMALIZED> is regex capture and replace syntax |
normalize-stdout-test | Normalize actual stdout with a rule "<raw>" -> "<normalized>" before comparing against snapshot | ui , incremental | "<RAW>" -> "<NORMALIZED>" , <RAW> /<NORMALIZED> is regex capture and replace syntax |
dont-check-compiler-stderr | Don't check actual compiler stderr vs stderr snapshot | ui | N/A |
dont-check-compiler-stdout | Don't check actual compiler stdout vs stdout snapshot | ui | N/A |
run-rustfix | Apply all suggestions via rustfix , snapshot fixed output, and check fixed output builds | ui | N/A |
rustfix-only-machine-applicable | run-rustfix but only machine-applicable suggestions | ui | N/A |
exec-env | Env var to set when executing a test | ui , crashes | <KEY>=<VALUE> |
unset-exec-env | Env var to unset when executing a test | ui , crashes | Any env var name |
stderr-per-bitwidth | Generate a stderr snapshot for each bitwidth | ui | N/A |
forbid-output | A pattern which must not appear in cfail output | incremental | Regex pattern |
run-flags | Flags passed to the test executable | ui | Arbitrary flags |
known-bug | No error annotation needed due to known bug | ui , crashes , incremental | Issue number #123456 |
presently this has a weird quirk
where the test binary's stdout and stderr gets concatenated and then
error-pattern
s are matched on this combined output, which is ??? slightly
questionable to say the least.
Controlling when tests are run
These directives are used to ignore the test in some situations, which means the test won't be compiled or run.
ignore-X
whereX
is a target detail or stage will ignore the test accordingly (see below)only-X
is likeignore-X
, but will only run the test on that target or stageignore-test
always ignores the test. This can be used to temporarily disable a test if it is currently not working, but you want to keep it in tree to re-enable it later.
Some examples of X
in ignore-X
or only-X
:
- A full target triple:
aarch64-apple-ios
- Architecture:
aarch64
,arm
,mips
,wasm32
,x86_64
,x86
, ... - OS:
android
,emscripten
,freebsd
,ios
,linux
,macos
,windows
, ... - Environment (fourth word of the target triple):
gnu
,msvc
,musl
- WASM:
wasm32-bare
matcheswasm32-unknown-unknown
.emscripten
also matches that target as well as the emscripten targets. - Pointer width:
32bit
,64bit
- Endianness:
endian-big
- Stage:
stage0
,stage1
,stage2
- Channel:
stable
,beta
- When cross compiling:
cross-compile
- When remote testing is used:
remote
- When particular debuggers are being tested:
cdb
,gdb
,lldb
- When particular debugger versions are matched:
ignore-gdb-version
- Specific compare modes:
compare-mode-polonius
,compare-mode-chalk
,compare-mode-split-dwarf
,compare-mode-split-dwarf-single
- The two different test modes used by coverage tests:
ignore-coverage-map
,ignore-coverage-run
The following directives will check rustc build settings and target settings:
needs-asm-support
— ignores if it is running on a target that doesn't have stable support forasm!
needs-profiler-runtime
— ignores the test if the profiler runtime was not enabled for the target (build.profiler = true
in rustc'sconfig.toml
)needs-sanitizer-support
— ignores if the sanitizer support was not enabled for the target (sanitizers = true
in rustc'sconfig.toml
)needs-sanitizer-{address,hwaddress,leak,memory,thread}
— ignores if the corresponding sanitizer is not enabled for the target (AddressSanitizer, hardware-assisted AddressSanitizer, LeakSanitizer, MemorySanitizer or ThreadSanitizer respectively)needs-run-enabled
— ignores if it is a test that gets executed, and running has been disabled. Running tests can be disabled with thex test --run=never
flag, or running on fuchsia.needs-unwind
— ignores if the target does not support unwindingneeds-rust-lld
— ignores if the rust lld support is not enabled (rust.lld = true
inconfig.toml
)needs-threads
— ignores if the target does not have threading supportneeds-symlink
— ignores if the target does not support symlinks. This can be the case on Windows if the developer did not enable privileged symlink permissions.ignore-std-debug-assertions
— ignores if std was built with debug assertions.needs-std-debug-assertions
— ignores if std was not built with debug assertions.ignore-rustc-debug-assertions
— ignores if rustc was built with debug assertions.needs-rustc-debug-assertions
— ignores if rustc was not built with debug assertions.needs-target-has-atomic
— ignores if target does not have support for all specified atomic widths, e.g. the test with//@ needs-target-has-atomic: 8, 16, ptr
will only run if it supports the comma-separated list of atomic widths.
The following directives will check LLVM support:
no-system-llvm
— ignores if the system llvm is usedexact-llvm-major-version: 19
— ignores if the llvm major version does not match the specified llvm major version.min-llvm-version: 13.0
— ignored if the LLVM version is less than the given valuemin-system-llvm-version: 12.0
— ignored if using a system LLVM and its version is less than the given valuemax-llvm-major-version: 19
— ignored if the LLVM major version is higher than the given major versionignore-llvm-version: 9.0
— ignores a specific LLVM versionignore-llvm-version: 7.0 - 9.9.9
— ignores LLVM versions in a range (inclusive)needs-llvm-components: powerpc
— ignores if the specific LLVM component was not built. Note: The test will fail on CI (whenCOMPILETEST_REQUIRE_ALL_LLVM_COMPONENTS
is set) if the component does not exist.needs-forced-clang-based-tests
— test is ignored unless the environment variableRUSTBUILD_FORCE_CLANG_BASED_TESTS
is set, which enables building clang alongside LLVM- This is only set in two CI jobs (
x86_64-gnu-debug
andaarch64-gnu-debug
), which only runs a subset ofrun-make
tests. Other tests with this directive will not run at all, which is usually not what you want. - Notably, the
aarch64-gnu-debug
CI job currently only runsrun-make
tests which additionally containclang
in their test name.
- This is only set in two CI jobs (
See also Debuginfo tests for directives for ignoring debuggers.
Affecting how tests are built
Directive | Explanation | Supported test suites | Possible values |
---|---|---|---|
compile-flags | Flags passed to rustc when building the test or aux file | All except for run-make | Any valid rustc flags, e.g. -Awarnings -Dfoo . Cannot be -Cincremental . |
edition | Alias for compile-flags: --edition=xxx | All except for run-make | Any valid --edition value |
rustc-env | Env var to set when running rustc | All except for run-make | <KEY>=<VALUE> |
unset-rustc-env | Env var to unset when running rustc | All except for run-make | Any env var name |
incremental | Proper incremental support for tests outside of incremental test suite | ui , crashes | N/A |
no-prefer-dynamic | Don't use -C prefer-dynamic , don't build as a dylib via a --crate-type=dylib preset flag | ui , crashes | N/A |
Consider writing the test as a proper incremental test instead.
Rustdoc
Directive | Explanation | Supported test suites | Possible values |
---|---|---|---|
doc-flags | Flags passed to rustdoc when building the test or aux file | rustdoc , js-doc-test , rustdoc-json | Any valid rustdoc flags |
FIXME(rustdoc): what does
check-test-line-numbers-match
do?
Pretty printing
See Pretty-printer.
Misc directives
no-auto-check-cfg
— disable auto check-cfg (only for--check-cfg
tests)revisions
— compile multiple timesunused-revision-names
- suppress tidy checks for mentioning unknown revision names -forbid-output
— incremental cfail rejects output patternshould-ice
— incremental cfail should ICEreference
— an annotation linking to a rule in the reference
Tool-specific directives
The following directives affect how certain command-line tools are invoked, in test suites that use those tools:
filecheck-flags
adds extra flags when running LLVM'sFileCheck
tool.- Used by codegen tests, assembly tests, and MIR-opt tests.
llvm-cov-flags
adds extra flags when running LLVM'sllvm-cov
tool.- Used by coverage tests in
coverage-run
mode.
- Used by coverage tests in
Substitutions
Directive values support substituting a few variables which will be replaced with their corresponding value. For example, if you need to pass a compiler flag with a path to a specific file, something like the following could work:
//@ compile-flags: --remap-path-prefix={{src-base}}=/the/src
Where the sentinel {{src-base}}
will be replaced with the appropriate path
described below:
{{cwd}}
: The directory where compiletest is run from. This may not be the root of the checkout, so you should avoid using it where possible.- Examples:
/path/to/rust
,/path/to/build/root
- Examples:
{{src-base}}
: The directory where the test is defined. This is equivalent to$DIR
for output normalization.- Example:
/path/to/rust/tests/ui/error-codes
- Example:
{{build-base}}
: The base directory where the test's output goes. This is equivalent to$TEST_BUILD_DIR
for output normalization.- Example:
/path/to/rust/build/x86_64-unknown-linux-gnu/test/ui
- Example:
{{rust-src-base}}
: The sysroot directory where libstd/libcore/... are located{{sysroot-base}}
: Path of the sysroot directory used to build the test.- Mainly intended for
ui-fulldeps
tests that run the compiler via API.
- Mainly intended for
{{target-linker}}
: Linker that would be passed to-Clinker
for this test, or blank if no linker override is active.- Mainly intended for
ui-fulldeps
tests that run the compiler via API.
- Mainly intended for
{{target}}
: The target the test is compiling for- Example:
x86_64-unknown-linux-gnu
- Example:
See
tests/ui/commandline-argfile.rs
for an example of a test that uses this substitution.
Adding a directive
One would add a new directive if there is a need to define some test property or behavior on an individual, test-by-test basis. A directive property serves as the directive's backing store (holds the command's current value) at runtime.
To add a new directive property:
- Look for the
pub struct TestProps
declaration insrc/tools/compiletest/src/header.rs
and add the new public property to the end of the declaration. - Look for the
impl TestProps
implementation block immediately following the struct declaration and initialize the new property to its default value.
Adding a new directive parser
When compiletest
encounters a test file, it parses the file a line at a time
by calling every parser defined in the Config
struct's implementation block,
also in src/tools/compiletest/src/header.rs
(note that the Config
struct's
declaration block is found in src/tools/compiletest/src/common.rs
).
TestProps
's load_from()
method will try passing the current line of text to
each parser, which, in turn typically checks to see if the line begins with a
particular commented (//@
) directive such as //@ must-compile-successfully
or //@ failure-status
. Whitespace after the comment marker is optional.
Parsers will override a given directive property's default value merely by being specified in the test file as a directive or by having a parameter value specified in the test file, depending on the directive.
Parsers defined in impl Config
are typically named parse_<directive-name>
(note kebab-case <directive-command>
transformed to snake-case
<directive_command>
). impl Config
also defines several 'low-level' parsers
which make it simple to parse common patterns like simple presence or not
(parse_name_directive()
), directive:parameter(s)
(parse_name_value_directive()
), optional parsing only if a particular cfg
attribute is defined (has_cfg_prefix()
) and many more. The low-level parsers
are found near the end of the impl Config
block; be sure to look through them
and their associated parsers immediately above to see how they are used to avoid
writing additional parsing code unnecessarily.
As a concrete example, here is the implementation for the
parse_failure_status()
parser, in src/tools/compiletest/src/header.rs
:
@@ -232,6 +232,7 @@ pub struct TestProps {
// customized normalization rules
pub normalize_stdout: Vec<(String, String)>,
pub normalize_stderr: Vec<(String, String)>,
+ pub failure_status: i32,
}
impl TestProps {
@@ -260,6 +261,7 @@ impl TestProps {
run_pass: false,
normalize_stdout: vec![],
normalize_stderr: vec![],
+ failure_status: 101,
}
}
@@ -383,6 +385,10 @@ impl TestProps {
if let Some(rule) = config.parse_custom_normalization(ln, "normalize-stderr") {
self.normalize_stderr.push(rule);
}
+
+ if let Some(code) = config.parse_failure_status(ln) {
+ self.failure_status = code;
+ }
});
for key in &["RUST_TEST_NOCAPTURE", "RUST_TEST_THREADS"] {
@@ -488,6 +494,13 @@ impl Config {
self.parse_name_directive(line, "pretty-compare-only")
}
+ fn parse_failure_status(&self, line: &str) -> Option<i32> {
+ match self.parse_name_value_directive(line, "failure-status") {
+ Some(code) => code.trim().parse::<i32>().ok(),
+ _ => None,
+ }
+ }
Implementing the behavior change
When a test invokes a particular directive, it is expected that some behavior
will change as a result. What behavior, obviously, will depend on the purpose of
the directive. In the case of failure-status
, the behavior that changes is
that compiletest
expects the failure code defined by the directive invoked in
the test, rather than the default value.
Although specific to failure-status
(as every directive will have a different
implementation in order to invoke behavior change) perhaps it is helpful to see
the behavior change implementation of one case, simply as an example. To
implement failure-status
, the check_correct_failure_status()
function found
in the TestCx
implementation block, located in
src/tools/compiletest/src/runtest.rs
, was modified as per below:
@@ -295,11 +295,14 @@ impl<'test> TestCx<'test> {
}
fn check_correct_failure_status(&self, proc_res: &ProcRes) {
- // The value the Rust runtime returns on failure
- const RUST_ERR: i32 = 101;
- if proc_res.status.code() != Some(RUST_ERR) {
+ let expected_status = Some(self.props.failure_status);
+ let received_status = proc_res.status.code();
+
+ if expected_status != received_status {
self.fatal_proc_rec(
- &format!("failure produced the wrong error: {}", proc_res.status),
+ &format!("Error: expected failure status ({:?}) but received status {:?}.",
+ expected_status,
+ received_status),
proc_res,
);
}
@@ -320,7 +323,6 @@ impl<'test> TestCx<'test> {
);
let proc_res = self.exec_compiled_test();
-
if !proc_res.status.success() {
self.fatal_proc_rec("test run failed!", &proc_res);
}
@@ -499,7 +501,6 @@ impl<'test> TestCx<'test> {
expected,
actual
);
- panic!();
}
}
Note the use of self.props.failure_status
to access the directive property. In
tests which do not specify the failure status directive,
self.props.failure_status
will evaluate to the default value of 101 at the
time of this writing. But for a test which specifies a directive of, for
example, //@ failure-status: 1
, self.props.failure_status
will evaluate to
1, as parse_failure_status()
will have overridden the TestProps
default
value, for that test specifically.
minicore
test auxiliary: using core
stubs
tests/auxiliary/minicore.rs
is a test auxiliary for
ui/codegen/assembly test suites. It provides core
stubs for tests that need to
build for cross-compiled targets but do not need/want to run.
A test can use minicore
by specifying the //@ add-core-stubs
directive.
Then, mark the test with #![feature(no_core)]
+ #![no_std]
+ #![no_core]
.
Due to Edition 2015 extern prelude rules, you will probably need to declare
minicore
as an extern crate.
Due to the no_std
+ no_core
nature of these tests, //@ add-core-stubs
implies and requires that the test will be built with -C panic=abort
.
Unwinding panics are not supported.
If you find a core
item to be missing from the minicore
stub, consider
adding it to the test auxiliary if it's likely to be used or is already needed
by more than one test.
Example codegen test that uses minicore
#![allow(unused)] fn main() { //@ add-core-stubs //@ revisions: meow bark //@[meow] compile-flags: --target=x86_64-unknown-linux-gnu //@[meow] needs-llvm-components: x86 //@[bark] compile-flags: --target=wasm32-unknown-unknown //@[bark] needs-llvm-components: webassembly #![crate_type = "lib"] #![feature(no_core)] #![no_std] #![no_core] extern crate minicore; use minicore::*; struct Meow; impl Copy for Meow {} // `Copy` here is provided by `minicore` // CHECK-LABEL: meow #[unsafe(no_mangle)] fn meow() {} }
Ecosystem testing
Rust tests integration with real-world code in the ecosystem to catch regressions and make informed decisions about the evolution of the language.
Testing methods
Crater
Crater is a tool which runs tests on many thousands of public projects. This tool has its own separate infrastructure for running, and is not run as part of CI. See the Crater chapter for more details.
cargotest
cargotest
is a small tool which runs cargo test
on a few sample projects
(such as servo
, ripgrep
, tokei
, etc.). This runs as part of CI and ensures
there aren't any significant regressions.
Example:
./x test src/tools/cargotest
Large OSS Project builders
We have CI jobs that build large open-source Rust projects that are used as regression tests in CI. Our integration jobs build the following projects:
Crater
Crater is a tool for compiling and running tests for every crate on crates.io (and a few on GitHub). It is mainly used for checking the extent of breakage when implementing potentially breaking changes and ensuring lack of breakage by running beta vs stable compiler versions.
When to run Crater
You should request a crater run if your PR makes large changes to the compiler or could cause breakage. If you are unsure, feel free to ask your PR's reviewer.
Requesting Crater Runs
The rust team maintains a few machines that can be used for running crater runs on the changes introduced by a PR. If your PR needs a crater run, leave a comment for the triage team in the PR thread. Please inform the team whether you require a "check-only" crater run, a "build only" crater run, or a "build-and-test" crater run. The difference is primarily in time; the conservative (if you're not sure) option is to go for the build-and-test run. If making changes that will only have an effect at compile-time (e.g., implementing a new trait) then you only need a check run.
Your PR will be enqueued by the triage team and the results will be posted when they are ready. Check runs will take around ~3-4 days, with the other two taking 5-6 days on average.
While crater is really useful, it is also important to be aware of a few caveats:
-
Not all code is on crates.io! There is a lot of code in repos on GitHub and elsewhere. Also, companies may not wish to publish their code. Thus, a successful crater run is not a magically green light that there will be no breakage; you still need to be careful.
-
Crater only runs Linux builds on x86_64. Thus, other architectures and platforms are not tested. Critically, this includes Windows.
-
Many crates are not tested. This could be for a lot of reasons, including that the crate doesn't compile any more (e.g. used old nightly features), has broken or flaky tests, requires network access, or other reasons.
-
Before crater can be run,
@bors try
needs to succeed in building artifacts. This means that if your code doesn't compile, you cannot run crater.
Fuchsia integration tests
Fuchsia is an open-source operating system with about 2 million lines of Rust code.1 It has caught a large number of regressions in the past and was subsequently included in CI.
Building Fuchsia in CI
Fuchsia builds as part of the suite of bors tests that run before a pull request is merged.
If you are worried that a pull request might break the Fuchsia builder and want to test it out before submitting it to the bors queue, simply add this line to your PR description:
try-job: x86_64-fuchsia
Then when you @bors try
it will pick the job that builds Fuchsia.
Building Fuchsia locally
Because Fuchsia uses languages other than Rust, it does not use Cargo as a build system. It also requires the toolchain build to be configured in a certain way.
The recommended way to build Fuchsia is to use the Docker scripts that check out and run a Fuchsia build for you. If you've run Docker tests before, you can simply run this command from your Rust checkout to download and build Fuchsia using your local Rust toolchain.
src/ci/docker/run.sh x86_64-fuchsia
See the Testing with Docker chapter for more details on how to run and debug jobs with Docker.
Note that a Fuchsia checkout is large – as of this writing, a checkout and build takes 46G of space – and as you might imagine, it takes a while to complete.
Modifying the Fuchsia checkout
The main reason you would want to build Fuchsia locally is because you need to
investigate a regression. After running a Docker build, you'll find the Fuchsia
checkout inside the obj/fuchsia
directory of your Rust checkout. If you
modify the KEEP_CHECKOUT
line in the build-fuchsia.sh script to
KEEP_CHECKOUT=1
, you can change the checkout as needed and rerun the build
command above. This will reuse all the build results from before.
You can find more options to customize the Fuchsia checkout in the build-fuchsia.sh script.
Customizing the Fuchsia build
You can find more info about the options used to build Fuchsia in Rust CI in the build_fuchsia_from_rust_ci.sh script invoked by build-fuchsia.sh.
The Fuchsia build system uses GN, a metabuild system that generates Ninja files and then hands off the work of running the build to Ninja.
Fuchsia developers use fx
to run builds and perform other development tasks.
This tool is located in .jiri_root/bin
of the Fuchsia checkout; you may need
to add this to your $PATH
for some workflows.
There are a few fx
subcommands that are relevant, including:
fx set
accepts build arguments, writes them toout/default/args.gn
, and runs GN.fx build
builds the Fuchsia project using Ninja. It will automatically pick up changes to build arguments and rerun GN. By default it builds everything, but it also accepts target paths to build specific targets (see below).fx clippy
runs Clippy on specific Rust targets (or all of them). We use this in the Rust CI build to avoid running codegen on most Rust targets. Underneath it invokes Ninja, just likefx build
. The clippy results are saved in json files inside the build output directory before being printed.
Target paths
GN uses paths like the following to identify build targets:
//src/starnix/kernel:starnix_core
The initial //
means the root of the checkout, and the remaining slashes are
directory names. The string after :
is the target name of a target defined
in the BUILD.gn
file of that directory.
The target name can be omitted if it is the same as the directory name. In other
words, //src/starnix/kernel
is the same as //src/starnix/kernel:kernel
.
These target paths are used inside BUILD.gn
files to reference dependencies,
and can also be used in fx build
.
Modifying compiler flags
You can put custom compiler flags inside a GN config
that is added to a
target. As a simple example:
config("everybody_loops") {
rustflags = [ "-Zeverybody-loops" ]
}
rustc_binary("example") {
crate_root = "src/bin.rs"
# ...existing keys here...
configs += [ ":everybody_loops" ]
}
This will add the flag -Zeverybody-loops
to rustc when building the example
target. Note that you can also use public_configs
for a config to be added
to every target that depends on that target.
If you want to add a flag to every Rust target in the build, you can add
rustflags to the //build/config:compiler
config or to the OS-specific
configs referenced in that file. Note that cflags
and ldflags
are ignored on
Rust targets.
Running ninja and rustc commands directly
Going down one layer, fx build
invokes ninja
, which in turn eventually
invokes rustc
. All build actions are run inside the out directory, which is
usually out/default
inside the Fuchsia checkout.
You can get ninja to print the actual command it invokes by forcing that command to fail, e.g. by adding a syntax error to one of the source files of the target. Once you have the command, you can run it from inside the output directory.
After changing the toolchain itself, the build setting rustc_version_string
in
out/default/args.gn
needs to be changed so that fx build
or ninja
will
rebuild all the Rust targets. This can be done in a text editor and the contents
of the string do not matter, as long as it changes from one build to the next.
build_fuchsia_from_rust_ci.sh does this for you by hashing the toolchain
directory.
The Fuchsia website has more detailed documentation of the build system.
Other tips and tricks
When using build_fuchsia_from_rust_ci.sh
you can comment out the fx set
command after the initial run so it won't rerun GN each time. If you do this you
can also comment out the version_string line to save a couple seconds.
export NINJA_PERSISTENT_MODE=1
to get faster ninja startup times after the
initial build.
Fuchsia target support
To learn more about Fuchsia target support, see the Fuchsia chapter in the rustc book.
As of June 2024, Fuchsia had about 2 million lines of first-party Rust code and a roughly equal amount of third-party code, as counted by tokei (excluding comments and blanks).
Rust for Linux integration tests
Rust for Linux (RfL) is an effort for adding support for the Rust programming language into the Linux kernel.
Building Rust for Linux in CI
Rust for Linux builds as part of the suite of bors tests that run before a pull request is merged.
The workflow builds a stage1 sysroot of the Rust compiler, downloads the Linux kernel, and tries to compile several Rust for Linux drivers and examples using this sysroot. RfL uses several unstable compiler/language features, therefore this workflow notifies us if a given compiler change would break it.
If you are worried that a pull request might break the Rust for Linux builder and want to test it out before submitting it to the bors queue, simply add this line to your PR description:
try-job: x86_64-rust-for-linux
Then when you @bors try
it will pick the job that builds the Rust for Linux
integration.
What to do in case of failure
If a PR breaks the Rust for Linux CI job, then:
- If the breakage was unintentional and seems spurious, then let RfL
know and retry.
- If the PR is urgent and retrying doesn't fix it, then disable the CI job
temporarily (comment out the
image: x86_64-rust-for-linux
job insrc/ci/github-actions/jobs.yml
).
- If the PR is urgent and retrying doesn't fix it, then disable the CI job
temporarily (comment out the
- If the breakage was unintentional, then change the PR to resolve the breakage.
- If the breakage was intentional, then let RfL know and discuss
what will the kernel need to change.
- If the PR is urgent, then disable the CI job temporarily (comment out
the
image: x86_64-rust-for-linux
job insrc/ci/github-actions/jobs.yml
). - If the PR can wait a few days, then wait for RfL maintainers to provide a
new Linux kernel commit hash with the needed changes done, and apply it to
the PR, which would confirm the changes work (update the
LINUX_VERSION
environment variable insrc/ci/docker/scripts/rfl-build.sh
).
- If the PR is urgent, then disable the CI job temporarily (comment out
the
If you need to contact the RfL developers, you can ping the Rust for Linux ping group to ask for help:
@rustbot ping rfl
Performance testing
rustc-perf
A lot of work is put into improving the performance of the compiler and preventing performance regressions.
The rustc-perf project provides
several services for testing and tracking performance. It provides hosted
infrastructure for running benchmarks as a service. At this time, only
x86_64-unknown-linux-gnu
builds are tracked.
A "perf run" is used to compare the performance of the compiler in different configurations for a large collection of popular crates. Different configurations include "fresh builds", builds with incremental compilation, etc.
The result of a perf run is a comparison between two versions of the compiler (by their commit hashes).
You can also use rustc-perf
to manually benchmark and profile the compiler
locally.
Automatic perf runs
After every PR is merged, a suite of benchmarks are run against the compiler. The results are tracked over time on the https://perf.rust-lang.org/ website. Any changes are noted in a comment on the PR.
Manual perf runs
Additionally, performance tests can be ran before a PR is merged on an as-needed basis. You should request a perf run if your PR may affect performance, especially if it can affect performance adversely.
To evaluate the performance impact of a PR, write this comment on the PR:
@bors try @rust-timer queue
Note: Only users authorized to do perf runs are allowed to post this comment. Teams that are allowed to use it are tracked in the Teams repository with the
perf = true
value in the[permissions]
section (and bors permissions are also required). If you are not on one of those teams, feel free to ask for someone to post it for you (either on Zulip or ask the assigned reviewer).
This will first tell bors to do a "try" build which do a full release build for
x86_64-unknown-linux-gnu
. After the build finishes, it will place it in the
queue to run the performance suite against it. After the performance tests
finish, the bot will post a comment on the PR with a summary and a link to a
full report.
If you want to do a perf run for an already built artifact (e.g. for a previous try build that wasn't benchmarked yet), you can run this instead:
@rust-timer build <commit-sha>
You cannot benchmark the same artifact twice though.
More information about the available perf bot commands can be found here.
More details about the benchmarking process itself are available in the perf collector documentation.
Suggest tests tool
This chapter is about the internals of and contribution instructions for the
suggest-tests
tool. For a high-level overview of the tool, see this
section. This tool is currently in a beta
state and is tracked by this
issue on Github. Currently the number of tests it will suggest are very limited
in scope, we are looking to expand this (contributions welcome!).
Internals
The tool is defined in a separate crate
(src/tools/suggest-tests
)
which outputs suggestions which are parsed by a shim in bootstrap
(src/bootstrap/src/core/build_steps/suggest.rs
).
The only notable thing the bootstrap shim does is (when invoked with the --run
flag) use bootstrap's internal mechanisms to create a new Builder
and uses it
to invoke the suggested commands. The suggest-tests
crate is where the fun
happens, two kinds of suggestions are defined: "static" and "dynamic"
suggestions.
Static suggestions
Defined
here.
Static suggestions are simple: they are just
globs which map to a x
command. In
suggest-tests
, this is implemented with a simple macro_rules
macro.
Dynamic suggestions
Defined
here.
These are more complicated than static suggestions and are implemented as
functions with the following signature: fn(&Path) -> Vec<Suggestion>
. In other
words, each suggestion takes a path to a modified file and (after running
arbitrary Rust code) can return any number of suggestions, or none. Dynamic
suggestions are useful for situations where fine-grained control over
suggestions is needed. For example, modifications to the compiler/xyz/
path
should trigger the x test compiler/xyz
suggestion. In the future, dynamic
suggestions might even read file contents to determine if (what) tests should
run.
Adding a suggestion
The following steps should serve as a rough guide to add suggestions to
suggest-tests
(very welcome!):
- Determine the rules for your suggestion. Is it simple and operates only on a single path or does it match globs? Does it need fine-grained control over the resulting command or does "one size fit all"?
- Based on the previous step, decide if your suggestion should be implemented as either static or dynamic.
- Implement the suggestion. If it is dynamic then a test is highly recommended, to verify that your logic is correct and to give an example of the suggestion. See the tests.rs file.
- Open a PR implementing your suggestion. (TODO: add example PR)
Miscellaneous testing-related info
RUSTC_BOOTSTRAP
and stability
This is a bootstrap/compiler implementation detail, but it can also be useful for testing:
RUSTC_BOOTSTRAP=1
will "cheat" and bypass usual stability checking, allowing you to use unstable features and cli flags on a stablerustc
.RUSTC_BOOTSTRAP=-1
will force a givenrustc
to pretend that is a stable compiler, even if it's actually a nightlyrustc
. This is useful because some behaviors of the compiler (e.g. diagnostics) can differ depending on whether the compiler is nightly or not.
In ui
tests and other test suites that support //@ rustc-env
, you can specify
// Force unstable features to be usable on stable rustc
//@ rustc-env:RUSTC_BOOTSTRAP=1
// Or force nightly rustc to pretend it is a stable rustc
//@ rustc-env:RUSTC_BOOTSTRAP=-1
For run-make
tests, //@ rustc-env
is not supported. You can do something
like the following for individual rustc
invocations.
use run_make_support::rustc;
fn main() {
rustc()
// Pretend that I am very stable
.env("RUSTC_BOOTSTRAP", "-1")
//...
.run();
}
Debugging the compiler
- Configuring the compiler
- Suppressing the ICE file
- Getting a backtrace
-Z
flags- Getting logging output
- Narrowing (Bisecting) Regressions
- Downloading Artifacts from Rust's CI
#[rustc_*]
TEST attributes- Configuring CodeLLDB for debugging
rustc
This chapter contains a few tips to debug the compiler. These tips aim to be useful no matter what you are working on. Some of the other chapters have advice about specific parts of the compiler (e.g. the Queries Debugging and Testing chapter or the LLVM Debugging chapter).
Configuring the compiler
By default, rustc is built without most debug information. To enable debug info,
set debug = true
in your config.toml.
Setting debug = true
turns on many different debug options (e.g., debug-assertions
,
debug-logging
, etc.) which can be individually tweaked if you want to, but many people
simply set debug = true
.
If you want to use GDB to debug rustc, please set config.toml
with options:
[rust]
debug = true
debuginfo-level = 2
NOTE: This will use a lot of disk space (upwards of 35GB), and will take a lot more compile time. With
debuginfo-level = 1
(the default whendebug = true
), you will be able to track the execution path, but will lose the symbol information for debugging.
The default configuration will enable symbol-mangling-version
v0.
This requires at least GDB v10.2,
otherwise you need to disable new symbol-mangling-version in config.toml
.
[rust]
new-symbol-mangling = false
See the comments in
config.example.toml
for more info.
You will need to rebuild the compiler after changing any configuration option.
Suppressing the ICE file
By default, if rustc encounters an Internal Compiler Error (ICE) it will dump the ICE contents to an
ICE file within the current working directory named rustc-ice-<timestamp>-<pid>.txt
. If this is
not desirable, you can prevent the ICE file from being created with RUSTC_ICE=0
.
Getting a backtrace
When you have an ICE (panic in the compiler), you can set
RUST_BACKTRACE=1
to get the stack trace of the panic!
like in
normal Rust programs. IIRC backtraces don't work on MinGW,
sorry. If you have trouble or the backtraces are full of unknown
,
you might want to find some way to use Linux, Mac, or MSVC on Windows.
In the default configuration (without debug
set to true
), you don't have line numbers
enabled, so the backtrace looks like this:
stack backtrace:
0: std::sys::imp::backtrace::tracing::imp::unwind_backtrace
1: std::sys_common::backtrace::_print
2: std::panicking::default_hook::{{closure}}
3: std::panicking::default_hook
4: std::panicking::rust_panic_with_hook
5: std::panicking::begin_panic
(~~~~ LINES REMOVED BY ME FOR BREVITY ~~~~)
32: rustc_typeck::check_crate
33: <std::thread::local::LocalKey<T>>::with
34: <std::thread::local::LocalKey<T>>::with
35: rustc::ty::context::TyCtxt::create_and_enter
36: rustc_driver::driver::compile_input
37: rustc_driver::run_compiler
If you set debug = true
, you will get line numbers for the stack trace.
Then the backtrace will look like this:
stack backtrace:
(~~~~ LINES REMOVED BY ME FOR BREVITY ~~~~)
at /home/user/rust/compiler/rustc_typeck/src/check/cast.rs:110
7: rustc_typeck::check::cast::CastCheck::check
at /home/user/rust/compiler/rustc_typeck/src/check/cast.rs:572
at /home/user/rust/compiler/rustc_typeck/src/check/cast.rs:460
at /home/user/rust/compiler/rustc_typeck/src/check/cast.rs:370
(~~~~ LINES REMOVED BY ME FOR BREVITY ~~~~)
33: rustc_driver::driver::compile_input
at /home/user/rust/compiler/rustc_driver/src/driver.rs:1010
at /home/user/rust/compiler/rustc_driver/src/driver.rs:212
34: rustc_driver::run_compiler
at /home/user/rust/compiler/rustc_driver/src/lib.rs:253
-Z
flags
The compiler has a bunch of -Z *
flags. These are unstable flags that are only
enabled on nightly. Many of them are useful for debugging. To get a full listing
of -Z
flags, use -Z help
.
One useful flag is -Z verbose-internals
, which generally enables printing more
info that could be useful for debugging.
Right below you can find elaborate explainers on a selected few.
Getting a backtrace for errors
If you want to get a backtrace to the point where the compiler emits an
error message, you can pass the -Z treat-err-as-bug=n
, which will make
the compiler panic on the nth
error. If you leave off =n
, the compiler will
assume 1
for n
and thus panic on the first error it encounters.
For example:
$ cat error.rs
fn main() { 1 + (); }
$ rustc +stage1 error.rs
error[E0277]: cannot add `()` to `{integer}`
--> error.rs:2:7
|
2 | 1 + ();
| ^ no implementation for `{integer} + ()`
|
= help: the trait `Add<()>` is not implemented for `{integer}`
error: aborting due to previous error
Now, where does the error above come from?
$ RUST_BACKTRACE=1 rustc +stage1 error.rs -Z treat-err-as-bug
error[E0277]: the trait bound `{integer}: std::ops::Add<()>` is not satisfied
--> error.rs:2:7
|
2 | 1 + ();
| ^ no implementation for `{integer} + ()`
|
= help: the trait `std::ops::Add<()>` is not implemented for `{integer}`
error: internal compiler error: unexpected panic
note: the compiler unexpectedly panicked. this is a bug.
note: we would appreciate a bug report: https://github.com/rust-lang/rust/blob/master/CONTRIBUTING.md#bug-reports
note: rustc 1.24.0-dev running on x86_64-unknown-linux-gnu
note: run with `RUST_BACKTRACE=1` for a backtrace
thread 'rustc' panicked at 'encountered error with `-Z treat_err_as_bug',
/home/user/rust/compiler/rustc_errors/src/lib.rs:411:12
note: Some details are omitted, run with `RUST_BACKTRACE=full` for a verbose
backtrace.
stack backtrace:
(~~~ IRRELEVANT PART OF BACKTRACE REMOVED BY ME ~~~)
7: rustc::traits::error_reporting::<impl rustc::infer::InferCtxt<'a, 'tcx>>
::report_selection_error
at /home/user/rust/compiler/rustc_middle/src/traits/error_reporting.rs:823
8: rustc::traits::error_reporting::<impl rustc::infer::InferCtxt<'a, 'tcx>>
::report_fulfillment_errors
at /home/user/rust/compiler/rustc_middle/src/traits/error_reporting.rs:160
at /home/user/rust/compiler/rustc_middle/src/traits/error_reporting.rs:112
9: rustc_typeck::check::FnCtxt::select_obligations_where_possible
at /home/user/rust/compiler/rustc_typeck/src/check/mod.rs:2192
(~~~ IRRELEVANT PART OF BACKTRACE REMOVED BY ME ~~~)
36: rustc_driver::run_compiler
at /home/user/rust/compiler/rustc_driver/src/lib.rs:253
Cool, now I have a backtrace for the error!
Debugging delayed bugs
The -Z eagerly-emit-delayed-bugs
option makes it easy to debug delayed bugs.
It turns them into normal errors, i.e. makes them visible. This can be used in
combination with -Z treat-err-as-bug
to stop at a particular delayed bug and
get a backtrace.
Getting the error creation location
-Z track-diagnostics
can help figure out where errors are emitted. It uses #[track_caller]
for this and prints its location alongside the error:
$ RUST_BACKTRACE=1 rustc +stage1 error.rs -Z track-diagnostics
error[E0277]: cannot add `()` to `{integer}`
--> src\error.rs:2:7
|
2 | 1 + ();
| ^ no implementation for `{integer} + ()`
-Ztrack-diagnostics: created at compiler/rustc_trait_selection/src/traits/error_reporting/mod.rs:638:39
|
= help: the trait `Add<()>` is not implemented for `{integer}`
= help: the following other types implement trait `Add<Rhs>`:
<&'a f32 as Add<f32>>
<&'a f64 as Add<f64>>
<&'a i128 as Add<i128>>
<&'a i16 as Add<i16>>
<&'a i32 as Add<i32>>
<&'a i64 as Add<i64>>
<&'a i8 as Add<i8>>
<&'a isize as Add<isize>>
and 48 others
For more information about this error, try `rustc --explain E0277`.
This is similar but different to -Z treat-err-as-bug
:
- it will print the locations for all errors emitted
- it does not require a compiler built with debug symbols
- you don't have to read through a big stack trace.
Getting logging output
The compiler uses the tracing
crate for logging.
For details see the guide section on tracing
Narrowing (Bisecting) Regressions
The cargo-bisect-rustc tool can be used as a quick and easy way to
find exactly which PR caused a change in rustc
behavior. It automatically
downloads rustc
PR artifacts and tests them against a project you provide
until it finds the regression. You can then look at the PR to get more context
on why it was changed. See this tutorial on how to use
it.
Downloading Artifacts from Rust's CI
The rustup-toolchain-install-master tool by kennytm can be used to
download the artifacts produced by Rust's CI for a specific SHA1 -- this
basically corresponds to the successful landing of some PR -- and then sets
them up for your local use. This also works for artifacts produced by @bors try
. This is helpful when you want to examine the resulting build of a PR
without doing the build yourself.
#[rustc_*]
TEST attributes
The compiler defines a whole lot of internal (perma-unstable) attributes some of which are useful
for debugging by dumping extra compiler-internal information. These are prefixed with rustc_
and
are gated behind the internal feature rustc_attrs
(enabled via e.g. #![feature(rustc_attrs)]
).
For a complete and up to date list, see builtin_attrs
. More specifically, the ones marked TEST
.
Here are some notable ones:
Attribute | Description |
---|---|
rustc_def_path | Dumps the def_path_str of an item. |
rustc_dump_def_parents | Dumps the chain of DefId parents of certain definitions. |
rustc_dump_item_bounds | Dumps the item_bounds of an item. |
rustc_dump_predicates | Dumps the predicates_of an item. |
rustc_dump_vtable | |
rustc_hidden_type_of_opaques | Dumps the hidden type of each opaque types in the crate. |
rustc_layout | See this section. |
rustc_object_lifetime_default | Dumps the object lifetime defaults of an item. |
rustc_outlives | Dumps implied bounds of an item. More precisely, the inferred_outlives_of an item. |
rustc_regions | Dumps NLL closure region requirements. |
rustc_symbol_name | Dumps the mangled & demangled symbol_name of an item. |
rustc_variances | Dumps the variances of an item. |
Right below you can find elaborate explainers on a selected few.
Formatting Graphviz output (.dot files)
Some compiler options for debugging specific features yield graphviz graphs -
e.g. the #[rustc_mir(borrowck_graphviz_postflow="suffix.dot")]
attribute
dumps various borrow-checker dataflow graphs.
These all produce .dot
files. To view these files, install graphviz (e.g.
apt-get install graphviz
) and then run the following commands:
$ dot -T pdf maybe_init_suffix.dot > maybe_init_suffix.pdf
$ firefox maybe_init_suffix.pdf # Or your favorite pdf viewer
Debugging type layouts
The internal attribute #[rustc_layout]
can be used to dump the Layout
of
the type it is attached to. For example:
#![allow(unused)] #![feature(rustc_attrs)] fn main() { #[rustc_layout(debug)] type T<'a> = &'a u32; }
Will emit the following:
error: layout_of(&'a u32) = Layout {
fields: Primitive,
variants: Single {
index: 0,
},
abi: Scalar(
Scalar {
value: Pointer,
valid_range: 1..=18446744073709551615,
},
),
largest_niche: Some(
Niche {
offset: Size {
raw: 0,
},
scalar: Scalar {
value: Pointer,
valid_range: 1..=18446744073709551615,
},
},
),
align: AbiAndPrefAlign {
abi: Align {
pow2: 3,
},
pref: Align {
pow2: 3,
},
},
size: Size {
raw: 8,
},
}
--> src/lib.rs:4:1
|
4 | type T<'a> = &'a u32;
| ^^^^^^^^^^^^^^^^^^^^^
error: aborting due to previous error
Configuring CodeLLDB for debugging rustc
If you are using VSCode, and have edited your config.toml
to request debugging
level 1 or 2 for the parts of the code you're interested in, then you should be
able to use the CodeLLDB extension in VSCode to debug it.
Here is a sample launch.json
file, being used to run a stage 1 compiler direct
from the directory where it is built (does not have to be "installed"):
// .vscode/launch.json
{
"version": "0.2.0",
"configurations": [
{
"type": "lldb",
"request": "launch",
"name": "Launch",
"args": [], // array of string command-line arguments to pass to compiler
"program": "${workspaceFolder}/build/host/stage1/bin/rustc",
"windows": { // applicable if using windows
"program": "${workspaceFolder}/build/host/stage1/bin/rustc.exe"
},
"cwd": "${workspaceFolder}", // current working directory at program start
"stopOnEntry": false,
"sourceLanguages": ["rust"]
}
]
}
Using tracing to debug the compiler
- Function level filters
- Query level filters
- Broad module level filters
- Log colors
- How to keep or remove
debug!
andtrace!
calls from the resulting binary - Logging etiquette and conventions
The compiler has a lot of debug!
(or trace!
) calls, which print out logging information
at many points. These are very useful to at least narrow down the location of
a bug if not to find it entirely, or just to orient yourself as to why the
compiler is doing a particular thing.
To see the logs, you need to set the RUSTC_LOG
environment variable to your
log filter. The full syntax of the log filters can be found in the rustdoc
of tracing-subscriber
.
Function level filters
Lots of functions in rustc are annotated with
#[instrument(level = "debug", skip(self))]
fn foo(&self, bar: Type) {}
which allows you to use
RUSTC_LOG=[foo]
to do the following all at once
- log all function calls to
foo
- log the arguments (except for those in the
skip
list) - log everything (from anywhere else in the compiler) until the function returns
I don't want everything
Depending on the scope of the function, you may not want to log everything in its body.
As an example: the do_mir_borrowck
function will dump hundreds of lines even for trivial
code being borrowchecked.
Since you can combine all filters, you can add a crate/module path, e.g.
RUSTC_LOG=rustc_borrowck[do_mir_borrowck]
I don't want all calls
If you are compiling libcore, you likely don't want all borrowck dumps, but only one for a specific function. You can filter function calls by their arguments by regexing them.
RUSTC_LOG=[do_mir_borrowck{id=\.\*from_utf8_unchecked\.\*}]
will only give you the logs of borrowchecking from_utf8_unchecked
. Note that you will
still get a short message per ignored do_mir_borrowck
, but none of the things inside those
calls. This helps you in looking through the calls that are happening and helps you adjust
your regex if you mistyped it.
Query level filters
Every query is automatically tagged with a logging span so that you can display all log messages during the execution of the query. For example, if you want to log everything during type checking:
RUSTC_LOG=[typeck]
The query arguments are included as a tracing field which means that you can
filter on the debug display of the arguments. For example, the typeck
query
has an argument key: LocalDefId
of what is being checked. You can use a
regex to match on that LocalDefId
to log type checking for a specific
function:
RUSTC_LOG=[typeck{key=.*name_of_item.*}]
Different queries have different arguments. You can find a list of queries and
their arguments in
rustc_middle/src/query/mod.rs
.
Broad module level filters
You can also use filters similar to the log
crate's filters, which will enable
everything within a specific module. This is often too verbose and too unstructured,
so it is recommended to use function level filters.
Your log filter can be just debug
to get all debug!
output and
higher (e.g., it will also include info!
), or path::to::module
to get all
output (which will include trace!
) from a particular module, or
path::to::module=debug
to get debug!
output and higher from a particular
module.
For example, to get the debug!
output and higher for a specific module, you
can run the compiler with RUSTC_LOG=path::to::module=debug rustc my-file.rs
.
All debug!
output will then appear in standard error.
Note that you can use a partial path and the filter will still work. For
example, if you want to see info!
output from only
rustdoc::passes::collect_intra_doc_links
, you could use
RUSTDOC_LOG=rustdoc::passes::collect_intra_doc_links=info
or you could use
RUSTDOC_LOG=rustdoc::passes::collect_intra=info
.
If you are developing rustdoc, use RUSTDOC_LOG
instead. If you are developing
Miri, use MIRI_LOG
instead. You get the idea :)
See the tracing
crate's docs, and specifically the docs for debug!
to
see the full syntax you can use. (Note: unlike the compiler, the tracing
crate and its examples use the RUST_LOG
environment variable. rustc, rustdoc,
and other tools set custom environment variables.)
Note that unless you use a very strict filter, the logger will emit a lot of output, so use the most specific module(s) you can (comma-separated if multiple). It's typically a good idea to pipe standard error to a file and look at the log output with a text editor.
So, to put it together:
# This puts the output of all debug calls in `rustc_middle/src/traits` into
# standard error, which might fill your console backscroll.
$ RUSTC_LOG=rustc_middle::traits=debug rustc +stage1 my-file.rs
# This puts the output of all debug calls in `rustc_middle/src/traits` in
# `traits-log`, so you can then see it with a text editor.
$ RUSTC_LOG=rustc_middle::traits=debug rustc +stage1 my-file.rs 2>traits-log
# Not recommended! This will show the output of all `debug!` calls
# in the Rust compiler, and there are a *lot* of them, so it will be
# hard to find anything.
$ RUSTC_LOG=debug rustc +stage1 my-file.rs 2>all-log
# This will show the output of all `info!` calls in `rustc_codegen_ssa`.
#
# There's an `info!` statement in `codegen_instance` that outputs
# every function that is codegen'd. This is useful to find out
# which function triggers an LLVM assertion, and this is an `info!`
# log rather than a `debug!` log so it will work on the official
# compilers.
$ RUSTC_LOG=rustc_codegen_ssa=info rustc +stage1 my-file.rs
# This will show all logs in `rustc_codegen_ssa` and `rustc_resolve`.
$ RUSTC_LOG=rustc_codegen_ssa,rustc_resolve rustc +stage1 my-file.rs
# This will show the output of all `info!` calls made by rustdoc
# or any rustc library it calls.
$ RUSTDOC_LOG=info rustdoc +stage1 my-file.rs
# This will only show `debug!` calls made by rustdoc directly,
# not any `rustc*` crate.
$ RUSTDOC_LOG=rustdoc=debug rustdoc +stage1 my-file.rs
Log colors
By default, rustc (and other tools, like rustdoc and Miri) will be smart about
when to use ANSI colors in the log output. If they are outputting to a terminal,
they will use colors, and if they are outputting to a file or being piped
somewhere else, they will not. However, it's hard to read log output in your
terminal unless you have a very strict filter, so you may want to pipe the
output to a pager like less
. But then there won't be any colors, which makes
it hard to pick out what you're looking for!
You can override whether to have colors in log output with the RUSTC_LOG_COLOR
environment variable (or RUSTDOC_LOG_COLOR
for rustdoc, or MIRI_LOG_COLOR
for Miri, etc.). There are three options: auto
(the default), always
, and
never
. So, if you want to enable colors when piping to less
, use something
similar to this command:
# The `-R` switch tells less to print ANSI colors without escaping them.
$ RUSTC_LOG=debug RUSTC_LOG_COLOR=always rustc +stage1 ... | less -R
Note that MIRI_LOG_COLOR
will only color logs that come from Miri, not logs
from rustc functions that Miri calls. Use RUSTC_LOG_COLOR
to color logs from
rustc.
How to keep or remove debug!
and trace!
calls from the resulting binary
While calls to error!
, warn!
and info!
are included in every build of the compiler,
calls to debug!
and trace!
are only included in the program if
debug-logging=true
is turned on in config.toml (it is
turned off by default), so if you don't see DEBUG
logs, especially
if you run the compiler with RUSTC_LOG=rustc rustc some.rs
and only see
INFO
logs, make sure that debug-logging=true
is turned on in your
config.toml.
Logging etiquette and conventions
Because calls to debug!
are removed by default, in most cases, don't worry
about the performance of adding "unnecessary" calls to debug!
and leaving them in code you
commit - they won't slow down the performance of what we ship.
That said, there can also be excessive tracing calls, especially
when they are redundant with other calls nearby or in functions called from
here. There is no perfect balance to hit here, and is left to the reviewer's
discretion to decide whether to let you leave debug!
statements in or whether to ask
you to remove them before merging.
It may be preferable to use trace!
over debug!
for very noisy logs.
A loosely followed convention is to use #[instrument(level = "debug")]
(also see the attribute's documentation)
in favour of debug!("foo(...)")
at the start of a function foo
.
Within functions, prefer debug!(?variable.field)
over debug!("xyz = {:?}", variable.field)
and debug!(bar = ?var.method(arg))
over debug!("bar = {:?}", var.method(arg))
.
The documentation for this syntax can be found here.
One thing to be careful of is expensive operations in logs.
If in the module rustc::foo
you have a statement
debug!(x = ?random_operation(tcx));
Then if someone runs a debug rustc
with RUSTC_LOG=rustc::foo
, then
random_operation()
will run. RUSTC_LOG
filters that do not enable this
debug statement will not execute random_operation
.
This means that you should not put anything too expensive or likely to crash there - that would annoy anyone who wants to use logging for that module. No-one will know it until someone tries to use logging to find another bug.
Profiling the compiler
This section talks about how to profile the compiler and find out where it spends its time.
Depending on what you're trying to measure, there are several different approaches:
-
If you want to see if a PR improves or regresses compiler performance, see the rustc-perf chapter for requesting a benchmarking run.
-
If you want a medium-to-high level overview of where
rustc
is spending its time:- The
-Z self-profile
flag and measureme tools offer a query-based approach to profiling. See their docs for more information.
- The
-
If you want function level performance data or even just more details than the above approaches:
-
If you want a nice visual representation of the compile times of your crate graph, you can use cargo's
--timings
flag, e.g.cargo build --timings
. You can use this flag on the compiler itself withCARGOFLAGS="--timings" ./x build
-
If you want to profile memory usage, you can use various tools depending on what operating system you are using.
- For Windows, read our WPA guide.
Optimizing rustc's bootstrap times with cargo-llvm-lines
Using cargo-llvm-lines you can count the number of lines of LLVM IR across all instantiations of a generic function. Since most of the time compiling rustc is spent in LLVM, the idea is that by reducing the amount of code passed to LLVM, compiling rustc gets faster.
To use cargo-llvm-lines
together with somewhat custom rustc build process, you can use
-C save-temps
to obtain required LLVM IR. The option preserves temporary work products
created during compilation. Among those is LLVM IR that represents an input to the
optimization pipeline; ideal for our purposes. It is stored in files with *.no-opt.bc
extension in LLVM bitcode format.
Example usage:
cargo install cargo-llvm-lines
# On a normal crate you could now run `cargo llvm-lines`, but `x` isn't normal :P
# Do a clean before every run, to not mix in the results from previous runs.
./x clean
env RUSTFLAGS=-Csave-temps ./x build --stage 0 compiler/rustc
# Single crate, e.g., rustc_middle. (Relies on the glob support of your shell.)
# Convert unoptimized LLVM bitcode into a human readable LLVM assembly accepted by cargo-llvm-lines.
for f in build/x86_64-unknown-linux-gnu/stage0-rustc/x86_64-unknown-linux-gnu/release/deps/rustc_middle-*.no-opt.bc; do
./build/x86_64-unknown-linux-gnu/llvm/bin/llvm-dis "$f"
done
cargo llvm-lines --files ./build/x86_64-unknown-linux-gnu/stage0-rustc/x86_64-unknown-linux-gnu/release/deps/rustc_middle-*.ll > llvm-lines-middle.txt
# Specify all crates of the compiler.
for f in build/x86_64-unknown-linux-gnu/stage0-rustc/x86_64-unknown-linux-gnu/release/deps/*.no-opt.bc; do
./build/x86_64-unknown-linux-gnu/llvm/bin/llvm-dis "$f"
done
cargo llvm-lines --files ./build/x86_64-unknown-linux-gnu/stage0-rustc/x86_64-unknown-linux-gnu/release/deps/*.ll > llvm-lines.txt
Example output for the compiler:
Lines Copies Function name
----- ------ -------------
45207720 (100%) 1583774 (100%) (TOTAL)
2102350 (4.7%) 146650 (9.3%) core::ptr::drop_in_place
615080 (1.4%) 8392 (0.5%) std::thread::local::LocalKey<T>::try_with
594296 (1.3%) 1780 (0.1%) hashbrown::raw::RawTable<T>::rehash_in_place
592071 (1.3%) 9691 (0.6%) core::option::Option<T>::map
528172 (1.2%) 5741 (0.4%) core::alloc::layout::Layout::array
466854 (1.0%) 8863 (0.6%) core::ptr::swap_nonoverlapping_one
412736 (0.9%) 1780 (0.1%) hashbrown::raw::RawTable<T>::resize
367776 (0.8%) 2554 (0.2%) alloc::raw_vec::RawVec<T,A>::grow_amortized
367507 (0.8%) 643 (0.0%) rustc_query_system::dep_graph::graph::DepGraph<K>::with_task_impl
355882 (0.8%) 6332 (0.4%) alloc::alloc::box_free
354556 (0.8%) 14213 (0.9%) core::ptr::write
354361 (0.8%) 3590 (0.2%) core::iter::traits::iterator::Iterator::fold
347761 (0.8%) 3873 (0.2%) rustc_middle::ty::context::tls::set_tlv
337534 (0.7%) 2377 (0.2%) alloc::raw_vec::RawVec<T,A>::allocate_in
331690 (0.7%) 3192 (0.2%) hashbrown::raw::RawTable<T>::find
328756 (0.7%) 3978 (0.3%) rustc_middle::ty::context::tls::with_context_opt
326903 (0.7%) 642 (0.0%) rustc_query_system::query::plumbing::try_execute_query
Since this doesn't seem to work with incremental compilation or ./x check
,
you will be compiling rustc a lot.
I recommend changing a few settings in config.toml
to make it bearable:
[rust]
# A debug build takes _a third_ as long on my machine,
# but compiling more than stage0 rustc becomes unbearably slow.
optimize = false
# We can't use incremental anyway, so we disable it for a little speed boost.
incremental = false
# We won't be running it, so no point in compiling debug checks.
debug = false
# Using a single codegen unit gives less output, but is slower to compile.
codegen-units = 0 # num_cpus
The llvm-lines output is affected by several options.
optimize = false
increases it from 2.1GB to 3.5GB and codegen-units = 0
to 4.1GB.
MIR optimizations have little impact. Compared to the default RUSTFLAGS="-Z mir-opt-level=1"
, level 0 adds 0.3GB and level 2 removes 0.2GB.
As of July 2022,
inlining happens in LLVM and GCC codegen backends,
missing only in the Cranelift one.
Profiling with perf
This is a guide for how to profile rustc with perf.
Initial steps
- Get a clean checkout of rust-lang/master, or whatever it is you want to profile.
- Set the following settings in your
config.toml
:debuginfo-level = 1
- enables line debuginfojemalloc = false
- lets you do memory use profiling with valgrind- leave everything else the defaults
- Run
./x build
to get a full build - Make a rustup toolchain pointing to that result
Gathering a perf profile
perf is an excellent tool on linux that can be used to gather and analyze all kinds of information. Mostly it is used to figure out where a program spends its time. It can also be used for other sorts of events, though, like cache misses and so forth.
The basics
The basic perf
command is this:
perf record -F99 --call-graph dwarf XXX
The -F99
tells perf to sample at 99 Hz, which avoids generating too
much data for longer runs (why 99 Hz you ask? It is often chosen
because it is unlikely to be in lockstep with other periodic
activity). The --call-graph dwarf
tells perf to get call-graph
information from debuginfo, which is accurate. The XXX
is the
command you want to profile. So, for example, you might do:
perf record -F99 --call-graph dwarf cargo +<toolchain> rustc
to run cargo
-- here <toolchain>
should be the name of the toolchain
you made in the beginning. But there are some things to be aware of:
- You probably don't want to profile the time spend building
dependencies. So something like
cargo build; cargo clean -p $C
may be helpful (where$C
is the crate name)- Though usually I just do
touch src/lib.rs
and rebuild instead. =)
- Though usually I just do
- You probably don't want incremental messing about with your
profile. So something like
CARGO_INCREMENTAL=0
can be helpful.
Gathering a perf profile from a perf.rust-lang.org
test
Often we want to analyze a specific test from perf.rust-lang.org
.
The easiest way to do that is to use the rustc-perf
benchmarking suite, this approach is described here.
Instead of using the benchmark suite CLI, you can also profile the benchmarks manually. First, you need to clone the rustc-perf repository:
$ git clone https://github.com/rust-lang/rustc-perf
and then find the source code of the test that you want to profile. Sources for the tests
are found in the collector/compile-benchmarks
directory
and the collector/runtime-benchmarks
directory. So let's
go into the directory of a specific test; we'll use clap-rs
as an example:
cd collector/compile-benchmarks/clap-3.1.6
In this case, let's say we want to profile the cargo check
performance. In that case, I would first run some basic commands to
build the dependencies:
# Setup: first clean out any old results and build the dependencies:
cargo +<toolchain> clean
CARGO_INCREMENTAL=0 cargo +<toolchain> check
(Again, <toolchain>
should be replaced with the name of the
toolchain we made in the first step.)
Next: we want record the execution time for just the clap-rs crate,
running cargo check. I tend to use cargo rustc
for this, since it
also allows me to add explicit flags, which we'll do later on.
touch src/lib.rs
CARGO_INCREMENTAL=0 perf record -F99 --call-graph dwarf cargo rustc --profile check --lib
Note that final command: it's a doozy! It uses the cargo rustc
command, which executes rustc with (potentially) additional options;
the --profile check
and --lib
options specify that we are doing a
cargo check
execution, and that this is a library (not a binary).
At this point, we can use perf
tooling to analyze the results. For example:
perf report
will open up an interactive TUI program. In simple cases, that can be
helpful. For more detailed examination, the perf-focus
tool
can be helpful; it is covered below.
A note of caution. Each of the rustc-perf tests is its own special
snowflake. In particular, some of them are not libraries, in which
case you would want to do touch src/main.rs
and avoid passing
--lib
. I'm not sure how best to tell which test is which to be
honest.
Gathering NLL data
If you want to profile an NLL run, you can just pass extra options to
the cargo rustc
command, like so:
touch src/lib.rs
CARGO_INCREMENTAL=0 perf record -F99 --call-graph dwarf cargo rustc --profile check --lib -- -Z borrowck=mir
Analyzing a perf profile with perf focus
Once you've gathered a perf profile, we want to get some information about it. For this, I personally use perf focus. It's a kind of simple but useful tool that lets you answer queries like:
- "how much time was spent in function F" (no matter where it was called from)
- "how much time was spent in function F when it was called from G"
- "how much time was spent in function F excluding time spent in G"
- "what functions does F call and how much time does it spend in them"
To understand how it works, you have to know just a bit about
perf. Basically, perf works by sampling your process on a regular
basis (or whenever some event occurs). For each sample, perf gathers a
backtrace. perf focus
lets you write a regular expression that tests
which functions appear in that backtrace, and then tells you which
percentage of samples had a backtrace that met the regular
expression. It's probably easiest to explain by walking through how I
would analyze NLL performance.
Installing perf-focus
You can install perf-focus using cargo install
:
cargo install perf-focus
Example: How much time is spent in MIR borrowck?
Let's say we've gathered the NLL data for a test. We'd like to know
how much time it is spending in the MIR borrow-checker. The "main"
function of the MIR borrowck is called do_mir_borrowck
, so we can do
this command:
$ perf focus '{do_mir_borrowck}'
Matcher : {do_mir_borrowck}
Matches : 228
Not Matches: 542
Percentage : 29%
The '{do_mir_borrowck}'
argument is called the matcher. It
specifies the test to be applied on the backtrace. In this case, the
{X}
indicates that there must be some function on the backtrace
that meets the regular expression X
. In this case, that regex is
just the name of the function we want (in fact, it's a subset of the name;
the full name includes a bunch of other stuff, like the module
path). In this mode, perf-focus just prints out the percentage of
samples where do_mir_borrowck
was on the stack: in this case, 29%.
A note about c++filt. To get the data from perf
, perf focus
currently executes perf script
(perhaps there is a better
way...). I've sometimes found that perf script
outputs C++ mangled
names. This is annoying. You can tell by running perf script | head
yourself — if you see names like 5rustc6middle
instead of
rustc::middle
, then you have the same problem. You can solve this
by doing:
perf script | c++filt | perf focus --from-stdin ...
This will pipe the output from perf script
through c++filt
and
should mostly convert those names into a more friendly format. The
--from-stdin
flag to perf focus
tells it to get its data from
stdin, rather than executing perf focus
. We should make this more
convenient (at worst, maybe add a c++filt
option to perf focus
, or
just always use it — it's pretty harmless).
Example: How much time does MIR borrowck spend solving traits?
Perhaps we'd like to know how much time MIR borrowck spends in the trait checker. We can ask this using a more complex regex:
$ perf focus '{do_mir_borrowck}..{^rustc::traits}'
Matcher : {do_mir_borrowck},..{^rustc::traits}
Matches : 12
Not Matches: 1311
Percentage : 0%
Here we used the ..
operator to ask "how often do we have
do_mir_borrowck
on the stack and then, later, some function whose
name begins with rustc::traits
?" (basically, code in that module). It
turns out the answer is "almost never" — only 12 samples fit that
description (if you ever see no samples, that often indicates your
query is messed up).
If you're curious, you can find out exactly which samples by using the
--print-match
option. This will print out the full backtrace for
each sample. The |
at the front of the line indicates the part that
the regular expression matched.
Example: Where does MIR borrowck spend its time?
Often we want to do more "explorational" queries. Like, we know that
MIR borrowck is 29% of the time, but where does that time get spent?
For that, the --tree-callees
option is often the best tool. You
usually also want to give --tree-min-percent
or
--tree-max-depth
. The result looks like this:
$ perf focus '{do_mir_borrowck}' --tree-callees --tree-min-percent 3
Matcher : {do_mir_borrowck}
Matches : 577
Not Matches: 746
Percentage : 43%
Tree
| matched `{do_mir_borrowck}` (43% total, 0% self)
: | rustc_borrowck::nll::compute_regions (20% total, 0% self)
: : | rustc_borrowck::nll::type_check::type_check_internal (13% total, 0% self)
: : : | core::ops::function::FnOnce::call_once (5% total, 0% self)
: : : : | rustc_borrowck::nll::type_check::liveness::generate (5% total, 3% self)
: : : | <rustc_borrowck::nll::type_check::TypeVerifier<'a, 'b, 'tcx> as rustc::mir::visit::Visitor<'tcx>>::visit_mir (3% total, 0% self)
: | rustc::mir::visit::Visitor::visit_mir (8% total, 6% self)
: | <rustc_borrowck::MirBorrowckCtxt<'cx, 'tcx> as rustc_mir_dataflow::DataflowResultsConsumer<'cx, 'tcx>>::visit_statement_entry (5% total, 0% self)
: | rustc_mir_dataflow::do_dataflow (3% total, 0% self)
What happens with --tree-callees
is that
- we find each sample matching the regular expression
- we look at the code that occurs after the regex match and try to build up a call tree
The --tree-min-percent 3
option says "only show me things that take
more than 3% of the time. Without this, the tree often gets really
noisy and includes random stuff like the innards of
malloc. --tree-max-depth
can be useful too, it just limits how many
levels we print.
For each line, we display the percent of time in that function altogether ("total") and the percent of time spent in just that function and not some callee of that function (self). Usually "total" is the more interesting number, but not always.
Relative percentages
By default, all in perf-focus are relative to the total program execution. This is useful to help you keep perspective — often as we drill down to find hot spots, we can lose sight of the fact that, in terms of overall program execution, this "hot spot" is actually not important. It also ensures that percentages between different queries are easily compared against one another.
That said, sometimes it's useful to get relative percentages, so perf focus
offers a --relative
option. In this case, the percentages are
listed only for samples that match (vs all samples). So for example we
could get our percentages relative to the borrowck itself
like so:
$ perf focus '{do_mir_borrowck}' --tree-callees --relative --tree-max-depth 1 --tree-min-percent 5
Matcher : {do_mir_borrowck}
Matches : 577
Not Matches: 746
Percentage : 100%
Tree
| matched `{do_mir_borrowck}` (100% total, 0% self)
: | rustc_borrowck::nll::compute_regions (47% total, 0% self) [...]
: | rustc::mir::visit::Visitor::visit_mir (19% total, 15% self) [...]
: | <rustc_borrowck::MirBorrowckCtxt<'cx, 'tcx> as rustc_mir_dataflow::DataflowResultsConsumer<'cx, 'tcx>>::visit_statement_entry (13% total, 0% self) [...]
: | rustc_mir_dataflow::do_dataflow (8% total, 1% self) [...]
Here you see that compute_regions
came up as "47% total" — that
means that 47% of do_mir_borrowck
is spent in that function. Before,
we saw 20% — that's because do_mir_borrowck
itself is only 43% of
the total time (and .47 * .43 = .20
).
Profiling on Windows
Introducing WPR and WPA
High-level performance analysis (including memory usage) can be performed with the Windows Performance Recorder (WPR) and Windows Performance Analyzer (WPA). As the names suggest, WPR is for recording system statistics (in the form of event trace log a.k.a. ETL files), while WPA is for analyzing these ETL files.
WPR collects system wide statistics, so it won't just record things relevant to rustc but also everything else that's running on the machine. During analysis, we can filter to just the things we find interesting.
These tools are quite powerful but also require a bit of learning before we can successfully profile the Rust compiler.
Here we will explore how to use WPR and WPA for analyzing the Rust compiler as well as provide links to useful "profiles" (i.e., settings files that tweak the defaults for WPR and WPA) that are specifically designed to make analyzing rustc easier.
Installing WPR and WPA
You can install WPR and WPA as part of the Windows Performance Toolkit which itself is an option as part of downloading the Windows Assessment and Deployment Kit (ADK). You can download the ADK installer here. Make sure to select the Windows Performance Toolkit (you don't need to select anything else).
Recording
In order to perform system analysis, you'll first need to record your system with WPR. Open WPR and at the bottom of the window select the "profiles" of the things you want to record. For looking into memory usage of the rustc bootstrap process, we'll want to select the following items:
- CPU usage
- VirtualAlloc usage
You might be tempted to record "Heap usage" as well, but this records every single heap allocation and can be very, very expensive. For high-level analysis, it might be best to leave that turned off.
Now we need to get our setup ready to record. For memory usage analysis, it is best to record the stage 2 compiler build with a stage 1 compiler build with debug symbols. Having symbols in the compiler we're using to build rustc will aid our analysis greatly by allowing WPA to resolve Rust symbols correctly. Unfortunately, the stage 0 compiler does not have symbols turned on which is why we'll need to build a stage 1 compiler and then a stage 2 compiler ourselves.
To do this, make sure you have set debuginfo-level = 1
in your config.toml
file. This tells
rustc to generate debug information which includes stack frames when bootstrapping.
Now you can build the stage 1 compiler: x build --stage 1 -i library
or however
else you want to build the stage 1 compiler.
Now that the stage 1 compiler is built, we can record the stage 2 build. Go back to WPR, click the
"start" button and build the stage 2 compiler (e.g., x build --stage=2 -i library
).
When this process finishes, stop the recording.
Click the Save button and once that process is complete, click the "Open in WPA" button which appears.
Note: The trace file is fairly large so it can take WPA some time to finish opening the file.
Analysis
Now that our ETL file is open in WPA, we can analyze the results. First, we'll want to apply the pre-made "profile" which will put WPA into a state conducive to analyzing rustc bootstrap. Download the profile here. Select the "Profiles" menu at the top, then "apply" and then choose the downloaded profile.
You should see something resembling the following:
Next, we will need to tell WPA to load and process debug symbols so that it can properly demangle the Rust stack traces. To do this, click "Trace" and then choose "Load Symbols". This step can take a while.
Once WPA has loaded symbols for rustc, we can expand the rustc.exe node and begin drilling down into the stack with the largest allocations.
To do that, we'll expand the [Root]
node in the "Commit Stack" column and continue expanding
until we find interesting stack frames.
Tip: After selecting the node you want to expand, press the right arrow key. This will expand the node and put the selection on the next largest node in the expanded set. You can continue pressing the right arrow key until you reach an interesting frame.
In this sample, you can see calls through codegen are allocating ~30gb of memory in total throughout this profile.
Other Analysis Tabs
The profile also includes a few other tabs which can be helpful:
- System Configuration
- General information about the system the capture was recorded on.
- rustc Build Processes
- A flat list of relevant processes such as rustc.exe, cargo.exe, link.exe etc.
- Each process lists its command line arguments.
- Useful for figuring out what a specific rustc process was working on.
- rustc Build Process Tree
- Timeline showing when processes started and exited.
- rustc CPU Analysis
- Contains charts preconfigured to show hotspots in rustc.
- These charts are designed to support analyzing where rustc is spending its time.
- rustc Memory Analysis
- Contains charts preconfigured to show where rustc is allocating memory.
Profiling with rustc-perf
The Rust benchmark suite provides a comprehensive way of profiling and benchmarking the Rust compiler. You can find instructions on how to use the suite in its manual.
However, using the suite manually can be a bit cumbersome. To make this easier for rustc
contributors,
the compiler build system (bootstrap
) also provides built-in integration with the benchmarking suite,
which will download and build the suite for you, build a local compiler toolchain and let you profile it using a simplified command-line interface.
You can use the ./x perf -- <command> [options]
command to use this integration.
Note that you need to specify arguments after
--
in thex perf
command! You will not be able to pass arguments without the double dashes.
You can use normal bootstrap flags for this command, such as --stage 1
or --stage 2
, for example to modify the stage of the created sysroot. It might also be useful to configure config.toml
to better support profiling, e.g. set rust.debuginfo-level = 1
to add source line information to the built compiler.
x perf
currently supports the following commands:
benchmark <id>
: Benchmark the compiler and store the results under the passedid
.compare <baseline> <modified>
: Compare the benchmark results of two compilers with the two passedid
s.eprintln
: Just run the compiler and capture itsstderr
output. Note that the compiler normally does not print anything tostderr
, you might want to add someeprintln!
calls to get any output.samply
: Profile the compiler using the samply sampling profiler.cachegrind
: Use Cachegrind to generate a detailed simulated trace of the compiler's execution.
You can find a more detailed description of the profilers in the
rustc-perf
manual.
You can use the following options for the x perf
command, which mirror the corresponding options of the
profile_local
and bench_local
commands that you can use in the suite:
--include
: Select benchmarks which should be profiled/benchmarked.--profiles
: Select profiles (Check
,Debug
,Opt
,Doc
) which should be profiled/benchmarked.--scenarios
: Select scenarios (Full
,IncrFull
,IncrPatched
,IncrUnchanged
) which should be profiled/benchmarked.
crates.io Dependencies
The Rust compiler supports building with some dependencies from crates.io
.
Examples are log
and env_logger
.
In general, you should avoid adding dependencies to the compiler for several reasons:
- The dependency may not be of high quality or well-maintained.
- The dependency may not be using a compatible license.
- The dependency may have transitive dependencies that have one of the above problems.
Note that there is no official policy for vetting new dependencies to the compiler. Decisions are made on a case-by-case basis, during code review.
Permitted dependencies
The tidy
tool has a list of crates that are allowed. To add a
dependency that is not already in the compiler, you will need to add it to the list.
Contribution Procedures
- Bug reports
- Bug fixes or "normal" code changes
- New features
- Pull requests
- External dependencies
- Writing documentation
- Issue triage
- Helpful links and information
Bug reports
While bugs are unfortunate, they're a reality in software. We can't fix what we don't know about, so please report liberally. If you're not sure if something is a bug or not, feel free to file a bug anyway.
If you believe reporting your bug publicly represents a security risk to Rust users, please follow our instructions for reporting security vulnerabilities.
If you're using the nightly channel, please check if the bug exists in the latest toolchain before filing your bug. It might be fixed already.
If you have the chance, before reporting a bug, please search existing issues, as it's possible that someone else has already reported your error. This doesn't always work, and sometimes it's hard to know what to search for, so consider this extra credit. We won't mind if you accidentally file a duplicate report.
Similarly, to help others who encountered the bug find your issue, consider filing an issue with a descriptive title, which contains information that might be unique to it. This can be the language or compiler feature used, the conditions that trigger the bug, or part of the error message if there is any. An example could be: "impossible case reached" on lifetime inference for impl Trait in return position.
Opening an issue is as easy as following this link and filling out the fields in the appropriate provided template.
Bug fixes or "normal" code changes
For most PRs, no special procedures are needed. You can just open a PR, and it will be reviewed, approved, and merged. This includes most bug fixes, refactorings, and other user-invisible changes. The next few sections talk about exceptions to this rule.
Also, note that it is perfectly acceptable to open WIP PRs or GitHub Draft PRs. Some people prefer to do this so they can get feedback along the way or share their code with a collaborator. Others do this so they can utilize the CI to build and test their PR (e.g. when developing on a slow machine).
New features
Rust has strong backwards-compatibility guarantees. Thus, new features can't just be implemented directly in stable Rust. Instead, we have 3 release channels: stable, beta, and nightly.
- Stable: this is the latest stable release for general usage.
- Beta: this is the next release (will be stable within 6 weeks).
- Nightly: follows the
master
branch of the repo. This is the only channel where unstable, incomplete, or experimental features are usable with feature gates.
See this chapter on implementing new features for more information.
Breaking changes
Breaking changes have a dedicated section in the dev-guide.
Major changes
The compiler team has a special process for large changes, whether or not they cause breakage. This process is called a Major Change Proposal (MCP). MCP is a relatively lightweight mechanism for getting feedback on large changes to the compiler (as opposed to a full RFC or a design meeting with the team).
Example of things that might require MCPs include major refactorings, changes to important types, or important changes to how the compiler does something, or smaller user-facing changes.
When in doubt, ask on zulip. It would be a shame to put a lot of work into a PR that ends up not getting merged! See this document for more info on MCPs.
Performance
Compiler performance is important. We have put a lot of effort over the last few years into gradually improving it.
If you suspect that your change may cause a performance regression (or improvement), you can request a "perf run" (and your reviewer may also request one before approving). This is yet another bot that will compile a collection of benchmarks on a compiler with your changes. The numbers are reported here, and you can see a comparison of your changes against the latest master.
For an introduction to the performance of Rust code in general which would also be useful in rustc development, see The Rust Performance Book.
Pull requests
Pull requests (or PRs for short) are the primary mechanism we use to change Rust. GitHub itself has some great documentation on using the Pull Request feature. We use the "fork and pull" model described here, where contributors push changes to their personal fork and create pull requests to bring those changes into the source repository. We have more info about how to use git when contributing to Rust under the git section.
Advice for potentially large, complex, cross-cutting and/or very domain-specific changes
The compiler reviewers on rotation usually each have areas of the compiler that they know well, but also have areas that they are not very familiar with. If your PR contains changes that are large, complex, cross-cutting and/or highly domain-specific, it becomes very difficult to find a suitable reviewer who is comfortable in reviewing all of the changes in such a PR. This is also true if the changes are not only compiler-specific but also contains changes which fall under the purview of reviewers from other teams, like the standard library team. There's a bot which notifies the relevant teams and pings people who have setup specific alerts based on the files modified.
Before making such changes, you are strongly encouraged to discuss your proposed changes with the compiler team beforehand (and with other teams that the changes would require approval from), and work with the compiler team to see if we can help you break down a large potentially unreviewable PR into a series of smaller more individually reviewable PRs.
You can communicate with the compiler team by creating a #t-compiler thread on zulip to discuss your proposed changes.
Communicating with the compiler team beforehand helps in several ways:
- It increases the likelihood of your PRs being reviewed in a timely manner.
- We can help you identify suitable reviewers before you open actual PRs, or help find advisors and liaisons to help you navigate the change procedures, or help with running try-jobs, perf runs and crater runs as suitable.
- It helps the compiler team track your changes.
- The compiler team can perform vibe checks on your changes early and often, to see if the direction of the changes align with what the compiler team prefers to see.
- Helps to avoid situations where you may have invested significant time and effort into large changes that the compiler team might not be willing to accept, or finding out very late that the changes are in a direction that the compiler team disagrees with.
r?
All pull requests are reviewed by another person. We have a bot, @rustbot, that will automatically assign a random person to review your request based on which files you changed.
If you want to request that a specific person reviews your pull request, you
can add an r?
to the pull request description or in a comment. For example,
if you want to ask a review to @awesome-reviewer, add
r? @awesome-reviewer
to the end of the pull request description, and @rustbot will assign them instead of a random person. This is entirely optional.
You can also assign a random reviewer from a specific team by writing r? rust-lang/groupname
.
As an example,
if you were making a diagnostics change,
then you could get a reviewer from the diagnostics team by adding:
r? rust-lang/diagnostics
For a full list of possible groupname
s,
check the adhoc_groups
section at the triagebot.toml config file,
or the list of teams in the rust-lang teams database.
Waiting for reviews
NOTE
Pull request reviewers are often working at capacity, and many of them are contributing on a volunteer basis. In order to minimize review delays, pull request authors and assigned reviewers should ensure that the review label (
S-waiting-on-review
andS-waiting-on-author
) stays updated, invoking these commands when appropriate:
@rustbot author
: the review is finished, and PR author should check the comments and take action accordingly.
@rustbot review
: the author is ready for a review, and this PR will be queued again in the reviewer's queue.
Please note that the reviewers are humans, who for the most part work on rustc
in their free time. This means that they can take some time to respond and review
your PR. It also means that reviewers can miss some PRs that are assigned to them.
To try to move PRs forward, the Triage WG regularly goes through all PRs that are waiting for review and haven't been discussed for at least 2 weeks. If you don't get a review within 2 weeks, feel free to ask the Triage WG on Zulip (#t-release/triage). They have knowledge of when to ping, who might be on vacation, etc.
The reviewer may request some changes using the GitHub code review interface. They may also request special procedures for some PRs. See Crater and Breaking Changes chapters for some examples of such procedures.
CI
In addition to being reviewed by a human, pull requests are automatically tested, thanks to continuous integration (CI). Basically, every time you open and update a pull request, CI builds the compiler and tests it against the compiler test suite, and also performs other tests such as checking that your pull request is in compliance with Rust's style guidelines.
Running continuous integration tests allows PR authors to catch mistakes early without going through a first review cycle, and also helps reviewers stay aware of the status of a particular pull request.
Rust has plenty of CI capacity, and you should never have to worry about wasting
computational resources each time you push a change. It is also perfectly fine
(and even encouraged!) to use the CI to test your changes if it can help your
productivity. In particular, we don't recommend running the full ./x test
suite locally,
since it takes a very long time to execute.
r+
After someone has reviewed your pull request, they will leave an annotation
on the pull request with an r+
. It will look something like this:
@bors r+
This tells @bors, our lovable integration bot, that your pull request has
been approved. The PR then enters the merge queue, where @bors
will run all the tests on every platform we support. If it all works out,
@bors will merge your code into master
and close the pull request.
Depending on the scale of the change, you may see a slightly different form of r+
:
@bors r+ rollup
The additional rollup
tells @bors that this change should always be "rolled up".
Changes that are rolled up are tested and merged alongside other PRs, to
speed the process up. Typically only small changes that are expected not to conflict
with one another are marked as "always roll up".
Be patient; this can take a while and the queue can sometimes be long. PRs are never merged by hand.
Opening a PR
You are now ready to file a pull request? Great! Here are a few points you should be aware of.
All pull requests should be filed against the master
branch,
unless you know for sure that you should target a different branch.
Make sure your pull request is in compliance with Rust's style guidelines by running
$ ./x test tidy --bless
We recommend to make this check before every pull request (and every new commit in a pull request); you can add git hooks before every push to make sure you never forget to make this check. The CI will also run tidy and will fail if tidy fails.
Rust follows a no merge-commit policy, meaning, when you encounter merge
conflicts you are expected to always rebase instead of merging. E.g. always use
rebase when bringing the latest changes from the master branch to your feature
branch. If your PR contains merge commits, it will get marked as has-merge-commits
.
Once you have removed the merge commits, e.g., through an interactive rebase, you
should remove the label again:
@rustbot label -has-merge-commits
See this chapter for more details.
If you encounter merge conflicts or when a reviewer asks you to perform some
changes, your PR will get marked as S-waiting-on-author
. When you resolve
them, you should use @rustbot
to mark it as S-waiting-on-review
:
@rustbot ready
GitHub allows closing issues using keywords. This feature should be used to keep the issue tracker tidy. However, it is generally preferred to put the "closes #123" text in the PR description rather than the issue commit; particularly during rebasing, citing the issue number in the commit can "spam" the issue in question.
However, if your PR fixes a stable-to-beta or stable-to-stable regression and has
been accepted for a beta and/or stable backport (i.e., it is marked beta-accepted
and/or stable-accepted
), please do not use any such keywords since we don't
want the corresponding issue to get auto-closed once the fix lands on master.
Please update the PR description while still mentioning the issue somewhere.
For example, you could write Fixes (after beta backport) #NNN.
.
As for further actions, please keep a sharp look-out for a PR whose title begins with
[beta]
or [stable]
and which backports the PR in question. When that one gets
merged, the relevant issue can be closed. The closing comment should mention all
PRs that were involved. If you don't have the permissions to close the issue, please
leave a comment on the original PR asking the reviewer to close it for you.
Reverting a PR
When a PR leads to miscompile, significant performance regressions, or other critical issues, we may want to revert that PR with a regression test case. You can also check out the revert policy on Forge docs (which is mainly targeted for reviewers, but contains useful info for PR authors too).
If the PR contains huge changes, it can be challenging to revert, making it harder to review incremental fixes in subsequent updates. Or if certain code in that PR is heavily depended upon by subsequent PRs, reverting it can become difficult.
In such cases, we can identify the problematic code and disable it for some input, as shown in #128271.
For MIR optimizations, we can also use the -Zunsound-mir-opt
option to gate the mir-opt, as shown
in #132356.
External dependencies
This section has moved to "Using External Repositories".
Writing documentation
Documentation improvements are very welcome. The source of doc.rust-lang.org
is located in src/doc
in the tree, and standard API documentation is generated
from the source code itself (e.g. library/std/src/lib.rs
). Documentation pull requests
function in the same way as other pull requests.
To find documentation-related issues, sort by the A-docs label.
You can find documentation style guidelines in RFC 1574.
To build the standard library documentation, use x doc --stage 0 library --open
.
To build the documentation for a book (e.g. the unstable book), use x doc src/doc/unstable-book.
Results should appear in build/host/doc
, as well as automatically open in your default browser.
See Building Documentation for more
information.
You can also use rustdoc
directly to check small fixes. For example,
rustdoc src/doc/reference.md
will render reference to doc/reference.html
.
The CSS might be messed up, but you can verify that the HTML is right.
Contributing to rustc-dev-guide
Contributions to the rustc-dev-guide are always welcome, and can be made directly at the rust-lang/rustc-dev-guide repo. The issue tracker in that repo is also a great way to find things that need doing. There are issues for beginners and advanced compiler devs alike!
Just a few things to keep in mind:
-
Please try to avoid overly long lines and use semantic line breaks (where you break the line after each sentence). There is no strict limit on line lengths; let the sentence or part of the sentence flow to its proper end on the same line.
-
When contributing text to the guide, please contextualize the information with some time period and/or a reason so that the reader knows how much to trust or mistrust the information. Aim to provide a reasonable amount of context, possibly including but not limited to:
-
A reason for why the data may be out of date other than "change", as change is a constant across the project.
-
The date the comment was added, e.g. instead of writing "Currently, ..." or "As of now, ...", consider adding the date, in one of the following formats:
- Jan 2021
- January 2021
- jan 2021
- january 2021
There is a CI action (in
~/.github/workflows/date-check.yml
) that generates a monthly showing those that are over 6 months old (example).For the action to pick the date, add a special annotation before specifying the date:
<!-- date-check --> Sep 2024
Example:
As of <!-- date-check --> Sep 2024, the foo did the bar.
For cases where the date should not be part of the visible rendered output, use the following instead:
<!-- date-check: Sep 2024 -->
-
A link to a relevant WG, tracking issue,
rustc
rustdoc page, or similar, that may provide further explanation for the change process or a way to verify that the information is not outdated.
-
-
If a text grows rather long (more than a few page scrolls) or complicated (more than four subsections), it might benefit from having a Table of Contents at the beginning, which you can auto-generate by including the
<!-- toc -->
marker at the top.
Issue triage
Sometimes, an issue will stay open, even though the bug has been fixed. And sometimes, the original bug may go stale because something has changed in the meantime.
It can be helpful to go through older bug reports and make sure that they are still valid. Load up an older issue, double check that it's still true, and leave a comment letting us know if it is or is not. The least recently updated sort is good for finding issues like this.
Thanks to @rustbot
, anyone can help triage issues by adding
appropriate labels to issues that haven't been triaged yet:
Labels | Color | Description |
---|---|---|
A- | Yellow | The area of the project an issue relates to. |
B- | Magenta | Issues which are blockers. |
beta- | Dark Blue | Tracks changes which need to be backported to beta |
C- | Light Purple | The category of an issue. |
D- | Mossy Green | Issues for diagnostics. |
E- | Green | The experience level necessary to fix an issue. |
F- | Peach | Issues for nightly features. |
I- | Red | The importance of the issue. |
I-*-nominated | Red | The issue has been nominated for discussion at the next meeting of the corresponding team. |
I-prioritize | Red | The issue has been nominated for prioritization by the team tagged with a T-prefixed label. |
L- | Teal | The relevant lint. |
metabug | Purple | Bugs that collect other bugs. |
O- | Purple Grey | The operating system or platform that the issue is specific to. |
P- | Orange | The issue priority. These labels can be assigned by anyone that understand the issue and is able to prioritize it, and remove the I-prioritize label. |
regression- | Pink | Tracks regressions from a stable release. |
relnotes | Light Orange | Changes that should be documented in the release notes of the next release. |
S- | Gray | Tracks the status of pull requests. |
S-tracking- | Steel Blue | Tracks the status of tracking issues. |
stable- | Dark Blue | Tracks changes which need to be backported to stable in anticipation of a point release. |
T- | Blue | Denotes which team the issue belongs to. |
WG- | Green | Denotes which working group the issue belongs to. |
Rfcbot labels
rfcbot uses its own labels for tracking the process of coordinating asynchronous decisions, such as approving or rejecting a change. This is used for RFCs, issues, and pull requests.
Labels | Color | Description |
---|---|---|
proposed-final-comment-period | Gray | Currently awaiting signoff of all team members in order to enter the final comment period. |
disposition-merge | Green | Indicates the intent is to merge the change. |
disposition-close | Red | Indicates the intent is to not accept the change and close it. |
disposition-postpone | Gray | Indicates the intent is to not accept the change at this time and postpone it to a later date. |
final-comment-period | Blue | Currently soliciting final comments before merging or closing. |
finished-final-comment-period | Light Yellow | The final comment period has concluded, and the issue will be merged or closed. |
postponed | Yellow | The issue has been postponed. |
closed | Red | The issue has been rejected. |
to-announce | Gray | Issues that have finished their final-comment-period and should be publicly announced. Note: the rust-lang/rust repository uses this label differently, to announce issues at the triage meetings. |
Helpful links and information
This section has moved to the "About this guide" chapter.
About the compiler team
rustc is maintained by the Rust compiler team. The people who belong to this team collectively work to track regressions and implement new features. Members of the Rust compiler team are people who have made significant contributions to rustc and its design.
Discussion
Currently the compiler team chats in Zulip:
- Team chat occurs in the
t-compiler
stream on the Zulip instance - There are also a number of other associated Zulip streams,
such as
t-compiler/help
, where people can ask for help with rustc development, ort-compiler/meetings
, where the team holds their weekly triage and steering meetings.
Reviewers
If you're interested in figuring out who can answer questions about a particular part of the compiler, or you'd just like to know who works on what, check out triagebot.toml's assign section. It contains a listing of the various parts of the compiler and a list of people who are reviewers of each part.
Rust compiler meeting
The compiler team has a weekly meeting where we do triage and try to generally stay on top of new bugs, regressions, and discuss important things in general. They are held on Zulip. It works roughly as follows:
- Announcements, MCPs/FCPs, and WG-check-ins: We share some announcements with the rest of the team about important things we want everyone to be aware of. We also share the status of MCPs and FCPs and we use the opportunity to have a couple of WGs giving us an update about their work.
- Check for beta and stable nominations: These are nominations of things to backport to beta and stable respectively. We then look for new cases where the compiler broke previously working code in the wild. Regressions are important issues to fix, so it's likely that they are tagged as P-critical or P-high; the major exception would be bug fixes (though even there we often aim to give warnings first).
- Review P-critical and P-high bugs: P-critical and P-high bugs are those that are sufficiently important for us to actively track progress. P-critical and P-high bugs should ideally always have an assignee.
- Check S-waiting-on-team and I-nominated issues: These are issues where feedback from the team is desired.
- Look over the performance triage report: We check for PRs that made the performance worse and try to decide if it's worth reverting the performance regression or if the regression can be addressed in a future PR.
The meeting currently takes place on Thursdays at 10am Boston time (UTC-4 typically, but daylight savings time sometimes makes things complicated).
Team membership
Membership in the Rust team is typically offered when someone has been making significant contributions to the compiler for some time. Membership is both a recognition but also an obligation: compiler team members are generally expected to help with upkeep as well as doing reviews and other work.
If you are interested in becoming a compiler team member, the first thing to do is to start fixing some bugs, or get involved in a working group. One good way to find bugs is to look for open issues tagged with E-easy or E-mentor.
You can also dig through the graveyard of PRs that were closed due to inactivity, some of them may contain work that is still useful - refer to the associated issues, if any - and only needs some finishing touches for which the original author didn't have time.
r+ rights
Once you have made a number of individual PRs to rustc, we will often offer r+ privileges. This means that you have the right to instruct "bors" (the robot that manages which PRs get landed into rustc) to merge a PR (here are some instructions for how to talk to bors).
The guidelines for reviewers are as follows:
- You are always welcome to review any PR, regardless of who it is
assigned to. However, do not r+ PRs unless:
- You are confident in that part of the code.
- You are confident that nobody else wants to review it first.
- For example, sometimes people will express a desire to review a PR before it lands, perhaps because it touches a particularly sensitive part of the code.
- Always be polite when reviewing: you are a representative of the Rust project, so it is expected that you will go above and beyond when it comes to the Code of Conduct.
Reviewer rotation
Once you have r+ rights, you can also be added to the reviewer rotation.
triagebot is the bot that automatically assigns incoming PRs to reviewers.
If you are added, you will be randomly selected to review
PRs. If you find you are assigned a PR that you don't feel comfortable
reviewing, you can also leave a comment like r? @so-and-so
to assign
to someone else — if you don't know who to request, just write r? @nikomatsakis for reassignment
and @nikomatsakis will pick someone
for you.
Getting on the reviewer rotation is much appreciated as it lowers the review burden for all of us! However, if you don't have time to give people timely feedback on their PRs, it may be better that you don't get on the list.
Full team membership
Full team membership is typically extended once someone made many contributions to the Rust compiler over time, ideally (but not necessarily) to multiple areas. Sometimes this might be implementing a new feature, but it is also important — perhaps more important! — to have time and willingness to help out with general upkeep such as bugfixes, tracking regressions, and other less glamorous work.
Using Git
- Prerequisites
- Standard Process
- Troubleshooting git issues
- I made a merge commit by accident.
- I deleted my fork on GitHub!
- I changed a submodule by accident
- I see "error: cannot rebase" when I try to rebase
- I see 'Untracked Files: src/stdarch'?
- I see
<<< HEAD
? - failed to push some refs
- Git is trying to rebase commits I didn't write?
- Quick note about submodules
- Rebasing and Conflicts
- Advanced Rebasing
- No-Merge Policy
- Tips for reviewing
- Git submodules
- Ignoring commits during
git blame
The Rust project uses Git to manage its source code. In order to contribute, you'll need some familiarity with its features so that your changes can be incorporated into the compiler.
The goal of this page is to cover some of the more common questions and problems new contributors face. Although some Git basics will be covered here, if you find that this is still a little too fast for you, it might make sense to first read some introductions to Git, such as the Beginner and Getting started sections of this tutorial from Atlassian. GitHub also provides documentation and guides for beginners, or you can consult the more in depth book from Git.
This guide is incomplete. If you run into trouble with git that this page doesn't help with, please open an issue so we can document how to fix it.
Prerequisites
We'll assume that you've installed Git, forked rust-lang/rust, and cloned the forked repo to your PC. We'll use the command line interface to interact with Git; there are also a number of GUIs and IDE integrations that can generally do the same things.
If you've cloned your fork, then you will be able to reference it with origin
in your local repo. It may be helpful to also set up a remote for the official
rust-lang/rust repo via
git remote add upstream https://github.com/rust-lang/rust.git
if you're using HTTPS, or
git remote add upstream git@github.com:rust-lang/rust.git
if you're using SSH.
NOTE: This page is dedicated to workflows for rust-lang/rust
, but will likely be
useful when contributing to other repositories in the Rust project.
Standard Process
Below is the normal procedure that you're likely to use for most minor changes and PRs:
- Ensure that you're making your changes on top of master:
git checkout master
. - Get the latest changes from the Rust repo:
git pull upstream master --ff-only
. (see No-Merge Policy for more info about this). - Make a new branch for your change:
git checkout -b issue-12345-fix
. - Make some changes to the repo and test them.
- Stage your changes via
git add src/changed/file.rs src/another/change.rs
and then commit them withgit commit
. Of course, making intermediate commits may be a good idea as well. Avoidgit add .
, as it makes it too easy to unintentionally commit changes that should not be committed, such as submodule updates. You can usegit status
to check if there are any files you forgot to stage. - Push your changes to your fork:
git push --set-upstream origin issue-12345-fix
(After adding commits, you can usegit push
and after rebasing or pulling-and-rebasing, you can usegit push --force-with-lease
). - Open a PR from your fork to
rust-lang/rust
's master branch.
If you end up needing to rebase and are hitting conflicts, see Rebasing. If you want to track upstream while working on long-running feature/issue, see Keeping things up to date.
If your reviewer requests changes, the procedure for those changes looks much the same, with some steps skipped:
- Ensure that you're making changes to the most recent version of your code:
git checkout issue-12345-fix
. - Make, stage, and commit your additional changes just like before.
- Push those changes to your fork:
git push
.
Troubleshooting git issues
You don't need to clone rust-lang/rust
from scratch if it's out of date!
Even if you think you've messed it up beyond repair, there are ways to fix
the git state that don't require downloading the whole repository again.
Here are some common issues you might run into:
I made a merge commit by accident.
Git has two ways to update your branch with the newest changes: merging and rebasing.
Rust uses rebasing. If you make a merge commit, it's not too hard to fix:
git rebase -i upstream/master
.
See Rebasing for more about rebasing.
I deleted my fork on GitHub!
This is not a problem from git's perspective. If you run git remote -v
,
it will say something like this:
$ git remote -v
origin git@github.com:jyn514/rust.git (fetch)
origin git@github.com:jyn514/rust.git (push)
upstream https://github.com/rust-lang/rust (fetch)
upstream https://github.com/rust-lang/rust (fetch)
If you renamed your fork, you can change the URL like this:
git remote set-url origin <URL>
where the <URL>
is your new fork.
I changed a submodule by accident
Usually people notice this when rustbot posts a comment on github that cargo
has been modified:
You might also notice conflicts in the web UI:
The most common cause is that you rebased after a change and ran git add .
without first running
x
to update the submodules. Alternatively, you might have run cargo fmt
instead of x fmt
and modified files in a submodule, then committed the changes.
To fix it, do the following things:
- See which commit has the accidental changes:
git log --stat -n1 src/tools/cargo
- Revert the changes to that commit:
git checkout <my-commit>~ src/tools/cargo
. Type~
literally but replace<my-commit>
with the output from step 1. - Tell git to commit the changes:
git commit --fixup <my-commit>
- Repeat steps 1-3 for all the submodules you modified.
- If you modified the submodule in several different commits, you will need to repeat steps 1-3
for each commit you modified. You'll know when to stop when the
git log
command shows a commit that's not authored by you.
- If you modified the submodule in several different commits, you will need to repeat steps 1-3
for each commit you modified. You'll know when to stop when the
- Squash your changes into the existing commits:
git rebase --autosquash -i upstream/master
- Push your changes.
I see "error: cannot rebase" when I try to rebase
These are two common errors to see when rebasing:
error: cannot rebase: Your index contains uncommitted changes.
error: Please commit or stash them.
error: cannot rebase: You have unstaged changes.
error: Please commit or stash them.
(See https://git-scm.com/book/en/v2/Getting-Started-What-is-Git%3F#_the_three_states for the difference between the two.)
This means you have made changes since the last time you made a commit. To be able to rebase, either commit your changes, or make a temporary commit called a "stash" to have them still not be committed when you finish rebasing. You may want to configure git to make this "stash" automatically, which will prevent the "cannot rebase" error in nearly all cases:
git config --global rebase.autostash true
See https://git-scm.com/book/en/v2/Git-Tools-Stashing-and-Cleaning for more info about stashing.
I see 'Untracked Files: src/stdarch'?
This is left over from the move to the library/
directory.
Unfortunately, git rebase
does not follow renames for submodules, so you
have to delete the directory yourself:
rm -r src/stdarch
I see <<< HEAD
?
You were probably in the middle of a rebase or merge conflict. See
Conflicts for how to fix the conflict. If you don't care about the changes
and just want to get a clean copy of the repository back, you can use git reset
:
# WARNING: this throws out any local changes you've made! Consider resolving the conflicts instead.
git reset --hard master
failed to push some refs
git push
will not work properly and say something like this:
! [rejected] issue-xxxxx -> issue-xxxxx (non-fast-forward)
error: failed to push some refs to 'https://github.com/username/rust.git'
hint: Updates were rejected because the tip of your current branch is behind
hint: its remote counterpart. Integrate the remote changes (e.g.
hint: 'git pull ...') before pushing again.
hint: See the 'Note about fast-forwards' in 'git push --help' for details.
The advice this gives is incorrect! Because of Rust's
"no-merge" policy the merge commit created by git pull
will not be allowed in the final PR, in addition to defeating the point of the
rebase! Use git push --force-with-lease
instead.
Git is trying to rebase commits I didn't write?
If you see many commits in your rebase list, or merge commits, or commits by other people that you
didn't write, it likely means you're trying to rebase over the wrong branch. For example, you may
have a rust-lang/rust
remote upstream
, but ran git rebase origin/master
instead of git rebase upstream/master
. The fix is to abort the rebase and use the correct branch instead:
git rebase --abort
git rebase -i upstream/master
Click here to see an example of rebasing over the wrong branch
Quick note about submodules
When updating your local repository with git pull
, you may notice that sometimes
Git says you have modified some files that you have never edited. For example,
running git status
gives you something like (note the new commits
mention):
On branch master
Your branch is up to date with 'origin/master'.
Changes not staged for commit:
(use "git add <file>..." to update what will be committed)
(use "git restore <file>..." to discard changes in working directory)
modified: src/llvm-project (new commits)
modified: src/tools/cargo (new commits)
no changes added to commit (use "git add" and/or "git commit -a")
These changes are not changes to files: they are changes to submodules (more on this
later). To get rid of those, run ./x --help
, which will automatically update
the submodules.
Some submodules are not actually needed; for example, src/llvm-project
doesn't need to be checked
out if you're using download-ci-llvm
. To avoid having to keep fetching its history, you can use
git submodule deinit -f src/llvm-project
, which will also avoid it showing as modified again.
Rebasing and Conflicts
When you edit your code locally, you are making changes to the version of rust-lang/rust that existed when you created your feature branch. As such, when you submit your PR it is possible that some of the changes that have been made to rust-lang/rust since then are in conflict with the changes you've made. When this happens, you need to resolve the conflicts before your changes can be merged. To do that, you need to rebase your work on top of rust-lang/rust.
Rebasing
To rebase your feature branch on top of the newest version of the master branch of rust-lang/rust, checkout your branch, and then run this command:
git pull --rebase https://github.com/rust-lang/rust.git master
If you are met with the following error:
error: cannot pull with rebase: Your index contains uncommitted changes. error: please commit or stash them.
it means that you have some uncommitted work in your working tree. In that case, run
git stash
before rebasing, and thengit stash pop
after you have rebased and fixed all conflicts.
When you rebase a branch on master, all the changes on your branch are reapplied to the most recent version of master. In other words, Git tries to pretend that the changes you made to the old version of master were instead made to the new version of master. During this process, you should expect to encounter at least one "rebase conflict." This happens when Git's attempt to reapply the changes fails because your changes conflicted with other changes that have been made. You can tell that this happened because you'll see lines in the output that look like
CONFLICT (content): Merge conflict in file.rs
When you open these files, you'll see sections of the form
<<<<<<< HEAD
Original code
=======
Your code
>>>>>>> 8fbf656... Commit fixes 12345
This represents the lines in the file that Git could not figure out how to
rebase. The section between <<<<<<< HEAD
and =======
has the code from
master, while the other side has your version of the code. You'll need to
decide how to deal with the conflict. You may want to keep your changes,
keep the changes on master, or combine the two.
Generally, resolving the conflict consists of two steps: First, fix the
particular conflict. Edit the file to make the changes you want and remove the
<<<<<<<
, =======
and >>>>>>>
lines in the process. Second, check the
surrounding code. If there was a conflict, its likely there are some logical
errors lying around too! It's a good idea to run x check
here to make sure
there are no glaring errors.
Once you're all done fixing the conflicts, you need to stage the files that had
conflicts in them via git add
. Afterwards, run git rebase --continue
to let
Git know that you've resolved the conflicts and it should finish the rebase.
Once the rebase has succeeded, you'll want to update the associated branch on
your fork with git push --force-with-lease
.
Keeping things up to date
The above section on Rebasing is a specific guide on rebasing work and dealing with merge conflicts. Here is some general advice about how to keep your local repo up-to-date with upstream changes:
Using git pull upstream master
while on your local master branch regularly
will keep it up-to-date. You will also want to rebase your feature branches
up-to-date as well. After pulling, you can checkout the feature branches
and rebase them:
git checkout master
git pull upstream master --ff-only # to make certain there are no merge commits
git rebase master feature_branch
git push --force-with-lease # (set origin to be the same as local)
To avoid merges as per the No-Merge Policy, you may want to use
git config pull.ff only
(this will apply the config only to the local repo)
to ensure that Git doesn't create merge commits when git pull
ing, without
needing to pass --ff-only
or --rebase
every time.
You can also git push --force-with-lease
from master to double-check that your
feature branches are in sync with their state on the Github side.
Advanced Rebasing
Squash your commits
"Squashing" commits into each other causes them to be merged into a single commit. Both the upside and downside of this is that it simplifies the history. On the one hand, you lose track of the steps in which changes were made, but the history becomes easier to work with.
If there are no conflicts and you are just squashing to clean up the history,
use git rebase --interactive --keep-base master
. This keeps the fork point
of your PR the same, making it easier to review the diff of what happened
across your rebases.
Squashing can also be useful as part of conflict resolution.
If your branch contains multiple consecutive rewrites of the same code, or if
the rebase conflicts are extremely severe, you can use
git rebase --interactive master
to gain more control over the process. This
allows you to choose to skip commits, edit the commits that you do not skip,
change the order in which they are applied, or "squash" them into each other.
Alternatively, you can sacrifice the commit history like this:
# squash all the changes into one commit so you only have to worry about conflicts once
git rebase -i --keep-base master # and squash all changes along the way
git rebase master
# fix all merge conflicts
git rebase --continue
You also may want to squash just the last few commits together, possibly
because they only represent "fixups" and not real changes. For example,
git rebase --interactive HEAD~2
will allow you to edit the two commits only.
git range-diff
After completing a rebase, and before pushing up your changes, you may want to
review the changes between your old branch and your new one. You can do that
with git range-diff master @{upstream} HEAD
.
The first argument to range-diff
, master
in this case, is the base revision
that you're comparing your old and new branch against. The second argument is
the old version of your branch; in this case, @upstream
means the version that
you've pushed to GitHub, which is the same as what people will see in your pull
request. Finally, the third argument to range-diff
is the new version of
your branch; in this case, it is HEAD
, which is the commit that is currently
checked-out in your local repo.
Note that you can also use the equivalent, abbreviated form git range-diff master @{u} HEAD
.
Unlike in regular Git diffs, you'll see a -
or +
next to another -
or +
in the range-diff output. The marker on the left indicates a change between the
old branch and the new branch, and the marker on the right indicates a change
you've committed. So, you can think of a range-diff as a "diff of diffs" since
it shows you the differences between your old diff and your new diff.
Here's an example of git range-diff
output (taken from Git's
docs):
-: ------- > 1: 0ddba11 Prepare for the inevitable!
1: c0debee = 2: cab005e Add a helpful message at the start
2: f00dbal ! 3: decafe1 Describe a bug
@@ -1,3 +1,3 @@
Author: A U Thor <author@example.com>
-TODO: Describe a bug
+Describe a bug
@@ -324,5 +324,6
This is expected.
-+What is unexpected is that it will also crash.
++Unexpectedly, it also crashes. This is a bug, and the jury is
++still out there how to fix it best. See ticket #314 for details.
Contact
3: bedead < -: ------- TO-UNDO
(Note that git range-diff
output in your terminal will probably be easier to
read than in this example because it will have colors.)
Another feature of git range-diff
is that, unlike git diff
, it will also
diff commit messages. This feature can be useful when amending several commit
messages so you can make sure you changed the right parts.
git range-diff
is a very useful command, but note that it can take some time
to get used to its output format. You may also find Git's documentation on the
command useful, especially their "Examples" section.
No-Merge Policy
The rust-lang/rust repo uses what is known as a "rebase workflow." This means
that merge commits in PRs are not accepted. As a result, if you are running
git merge
locally, chances are good that you should be rebasing instead. Of
course, this is not always true; if your merge will just be a fast-forward,
like the merges that git pull
usually performs, then no merge commit is
created and you have nothing to worry about. Running git config merge.ff only
(this will apply the config to the local repo)
once will ensure that all the merges you perform are of this type, so that you
cannot make a mistake.
There are a number of reasons for this decision and like all others, it is a tradeoff. The main advantage is the generally linear commit history. This greatly simplifies bisecting and makes the history and commit log much easier to follow and understand.
Tips for reviewing
NOTE: This section is for reviewing PRs, not authoring them.
Hiding whitespace
Github has a button for disabling whitespace changes that may be useful.
You can also use git diff -w origin/master
to view changes locally.
Fetching PRs
To checkout PRs locally, you can use git fetch upstream pull/NNNNN/head && git checkout FETCH_HEAD
.
You can also use github's cli tool. Github shows a button on PRs where you can copy-paste the command to check it out locally. See https://cli.github.com/ for more info.
Moving large sections of code
Git and Github's default diff view for large moves within a file is quite poor; it will show each line as deleted and each line as added, forcing you to compare each line yourself. Git has an option to show moved lines in a different color:
git log -p --color-moved=dimmed-zebra --color-moved-ws=allow-indentation-change
See the docs for --color-moved
for more info.
range-diff
See the relevant section for PR authors. This can be useful for comparing code that was force-pushed to make sure there are no unexpected changes.
Ignoring changes to specific files
Many large files in the repo are autogenerated. To view a diff that ignores changes to those files, you can use the following syntax (e.g. Cargo.lock):
git log -p ':!Cargo.lock'
Arbitrary patterns are supported (e.g. :!compiler/*
). Patterns use the same syntax as
.gitignore
, with :
prepended to indicate a pattern.
Git submodules
NOTE: submodules are a nice thing to know about, but it isn't an absolute
prerequisite to contribute to rustc
. If you are using Git for the first time,
you might want to get used to the main concepts of Git before reading this section.
The rust-lang/rust
repository uses Git submodules as a way to use other
Rust projects from within the rust
repo. Examples include Rust's fork of
llvm-project
, cargo
and libraries like stdarch
and backtrace
.
Those projects are developed and maintained in an separate Git (and GitHub)
repository, and they have their own Git history/commits, issue tracker and PRs.
Submodules allow us to create some sort of embedded sub-repository inside the
rust
repository and use them like they were directories in the rust
repository.
Take llvm-project
for example. llvm-project
is maintained in the rust-lang/llvm-project
repository, but it is used in rust-lang/rust
by the compiler for code generation and
optimization. We bring it in rust
as a submodule, in the src/llvm-project
folder.
The contents of submodules are ignored by Git: submodules are in some sense isolated
from the rest of the repository. However, if you try to cd src/llvm-project
and then
run git status
:
HEAD detached at 9567f08afc943
nothing to commit, working tree clean
As far as git is concerned, you are no longer in the rust
repo, but in the llvm-project
repo.
You will notice that we are in "detached HEAD" state, i.e. not on a branch but on a
particular commit.
This is because, like any dependency, we want to be able to control which version to use.
Submodules allow us to do just that: every submodule is "pinned" to a certain
commit, which doesn't change unless modified manually. If you use git checkout <commit>
in the llvm-project
directory and go back to the rust
directory, you can stage this
change like any other, e.g. by running git add src/llvm-project
. (Note that if
you don't stage the change to commit, then you run the risk that running
x
will just undo your change by switching back to the previous commit when
it automatically "updates" the submodules.)
This version selection is usually done by the maintainers of the project, and looks like this.
Git submodules take some time to get used to, so don't worry if it isn't perfectly clear yet. You will rarely have to use them directly and, again, you don't need to know everything about submodules to contribute to Rust. Just know that they exist and that they correspond to some sort of embedded subrepository dependency that Git can nicely and fairly conveniently handle for us.
Hard-resetting submodules
Sometimes you might run into (when you run git status
)
Changes not staged for commit:
(use "git add <file>..." to update what will be committed)
(use "git restore <file>..." to discard changes in working directory)
(commit or discard the untracked or modified content in submodules)
modified: src/llvm-project (new commits, modified content)
and when you try to run git submodule update
it breaks horribly with errors like
error: RPC failed; curl 92 HTTP/2 stream 7 was not closed cleanly: CANCEL (err 8)
error: 2782 bytes of body are still expected
fetch-pack: unexpected disconnect while reading sideband packet
fatal: early EOF
fatal: fetch-pack: invalid index-pack output
fatal: Fetched in submodule path 'src/llvm-project', but it did not contain 5a5152f653959d14d68613a3a8a033fb65eec021. Direct fetching of that commit failed.
If you see (new commits, modified content)
you can run
$ git submodule foreach git reset --hard
and then try git submodule update
again.
Deinit git submodules
If that doesn't work, you can try to deinit all git submodules...
git submodule deinit -f --all
Unfortunately sometimes your local git submodules configuration can become completely messed up for some reason.
Overcoming fatal: not a git repository: <submodule>/../../.git/modules/<submodule>
Sometimes, for some forsaken reason, you might run into
fatal: not a git repository: src/gcc/../../.git/modules/src/gcc
In this situation, for the given submodule path, i.e. <submodule_path> = src/gcc
in this example, you need to:
rm -rf <submodule_path>/.git
rm -rf .git/modules/<submodule_path>/config
rm -rf .gitconfig.lock
if somehow the.gitconfig
lock is orphaned.
Then do something like ./x fmt
to have bootstrap manage the submodule
checkouts for you.
Ignoring commits during git blame
Some commits contain large reformatting changes that don't otherwise change functionality. They can
be instructed to be ignored by git blame
through
.git-blame-ignore-revs
:
- Configure
git blame
to use.git-blame-ignore-revs
as the list of commits to ignore:git config blame.ignorerevsfile .git-blame-ignore-revs
- Add suitable commits that you wish to be ignored by
git blame
.
Please include a comment for the commit that you add to .git-blame-ignore-revs
so people can
easily figure out why a commit is ignored.
Mastering @rustbot
@rustbot
(also known as triagebot
) is a utility robot that is mostly used to
allow any contributor to achieve certain tasks that would normally require GitHub
membership to the rust-lang
organization. Its most interesting features for
contributors to rustc
are issue claiming and relabeling.
Issue claiming
@rustbot
exposes a command that allows anyone to assign an issue to themselves.
If you see an issue you want to work on, you can send the following message as a
comment on the issue at hand:
@rustbot claim
This will tell @rustbot
to assign the issue to you if it has no assignee yet.
Note that because of some GitHub restrictions, you may be assigned indirectly,
i.e. @rustbot
will assign itself as a placeholder and edit the top comment to
reflect the fact that the issue is now assigned to you.
If you want to unassign from an issue, @rustbot
has a different command:
@rustbot release-assignment
Issue relabeling
Changing labels for an issue or PR is also normally reserved for members of the
organization. However, @rustbot
allows you to relabel an issue yourself, only
with a few restrictions. This is mostly useful in two cases:
Helping with issue triage: Rust's issue tracker has more than 5,000 open issues at the time of this writing, so labels are the most powerful tool that we have to keep it as tidy as possible. You don't need to spend hours in the issue tracker to triage issues, but if you open an issue, you should feel free to label it if you are comfortable with doing it yourself.
Updating the status of a PR: We use "status labels" to reflect the status of
PRs. For example, if your PR has merge conflicts, it will automatically be assigned
the S-waiting-on-author
, and reviewers might not review it until you rebase your
PR. Once you do rebase your branch, you should change the labels yourself to remove
the S-waiting-on-author
label and add back S-waiting-on-review
. In this case,
the @rustbot
command will look like this:
@rustbot label -S-waiting-on-author +S-waiting-on-review
The syntax for this command is pretty loose, so there are other variants of this
command invocation. There are also some shortcuts to update labels,
for instance @rustbot ready
will do the same thing with above command.
For more details, see the docs page about labeling and shortcuts.
Other commands
If you are interested in seeing what @rustbot
is capable of, check out its documentation,
which is meant as a reference for the bot and should be kept up to date every time the
bot gets an upgrade.
@rustbot
is maintained by the Release team. If you have any feedback regarding
existing commands or suggestions for new commands, feel free to reach out
on Zulip or file an issue in the triagebot repository
Walkthrough: a typical contribution
There are a lot of ways to contribute to the Rust compiler, including fixing bugs, improving performance, helping design features, providing feedback on existing features, etc. This chapter does not claim to scratch the surface. Instead, it walks through the design and implementation of a new feature. Not all of the steps and processes described here are needed for every contribution, and I will try to point those out as they arise.
In general, if you are interested in making a contribution and aren't sure where to start, please feel free to ask!
Overview
The feature I will discuss in this chapter is the ?
Kleene operator for
macros. Basically, we want to be able to write something like this:
macro_rules! foo {
($arg:ident $(, $optional_arg:ident)?) => {
println!("{}", $arg);
$(
println!("{}", $optional_arg);
)?
}
}
fn main() {
let x = 0;
foo!(x); // ok! prints "0"
foo!(x, x); // ok! prints "0 0"
}
So basically, the $(pat)?
matcher in the macro means "this pattern can occur
0 or 1 times", similar to other regex syntaxes.
There were a number of steps to go from an idea to stable Rust feature. Here is a quick list. We will go through each of these in order below. As I mentioned before, not all of these are needed for every type of contribution.
- Idea discussion/Pre-RFC A Pre-RFC is an early draft or design discussion of a feature. This stage is intended to flesh out the design space a bit and get a grasp on the different merits and problems with an idea. It's a great way to get early feedback on your idea before presenting it to the wider audience. You can find the original discussion here.
- RFC This is when you formally present your idea to the community for consideration. You can find the RFC here.
- Implementation Implement your idea unstably in the compiler. You can find the original implementation here.
- Possibly iterate/refine As the community gets experience with your
feature on the nightly compiler and in
std
, there may be additional feedback about design choice that might be adjusted. This particular feature went through a number of iterations. - Stabilization When your feature has baked enough, a Rust team member may propose to stabilize it. If there is consensus, this is done.
- Relax Your feature is now a stable Rust feature!
Pre-RFC and RFC
NOTE: In general, if you are not proposing a new feature or substantial change to Rust or the ecosystem, you don't need to follow the RFC process. Instead, you can just jump to implementation.
You can find the official guidelines for when to open an RFC here.
An RFC is a document that describes the feature or change you are proposing in detail. Anyone can write an RFC; the process is the same for everyone, including Rust team members.
To open an RFC, open a PR on the rust-lang/rfcs repo on GitHub. You can find detailed instructions in the README.
Before opening an RFC, you should do the research to "flesh out" your idea. Hastily-proposed RFCs tend not to be accepted. You should generally have a good description of the motivation, impact, disadvantages, and potential interactions with other features.
If that sounds like a lot of work, it's because it is. But no fear! Even if you're not a compiler hacker, you can get great feedback by doing a pre-RFC. This is an informal discussion of the idea. The best place to do this is internals.rust-lang.org. Your post doesn't have to follow any particular structure. It doesn't even need to be a cohesive idea. Generally, you will get tons of feedback that you can integrate back to produce a good RFC.
(Another pro-tip: try searching the RFCs repo and internals for prior related ideas. A lot of times an idea has already been considered and was either rejected or postponed to be tried again later. This can save you and everybody else some time)
In the case of our example, a participant in the pre-RFC thread pointed out a syntax ambiguity and a potential resolution. Also, the overall feedback seemed positive. In this case, the discussion converged pretty quickly, but for some ideas, a lot more discussion can happen (e.g. see this RFC which received a whopping 684 comments!). If that happens, don't be discouraged; it means the community is interested in your idea, but it perhaps needs some adjustments.
The RFC for our ?
macro feature did receive some discussion on the RFC thread
too. As with most RFCs, there were a few questions that we couldn't answer by
discussion: we needed experience using the feature to decide. Such questions
are listed in the "Unresolved Questions" section of the RFC. Also, over the
course of the RFC discussion, you will probably want to update the RFC document
itself to reflect the course of the discussion (e.g. new alternatives or prior
work may be added or you may decide to change parts of the proposal itself).
In the end, when the discussion seems to reach a consensus and die down a bit, a Rust team member may propose to move to "final comment period" (FCP) with one of three possible dispositions. This means that they want the other members of the appropriate teams to review and comment on the RFC. More discussion may ensue, which may result in more changes or unresolved questions being added. At some point, when everyone is satisfied, the RFC enters the FCP, which is the last chance for people to bring up objections. When the FCP is over, the disposition is adopted. Here are the three possible dispositions:
- Merge: accept the feature. Here is the proposal to merge for our
?
macro feature. - Close: this feature in its current form is not a good fit for rust. Don't be discouraged if this happens to your RFC, and don't take it personally. This is not a reflection on you, but rather a community decision that rust will go a different direction.
- Postpone: there is interest in going this direction but not at the moment. This happens most often because the appropriate Rust team doesn't have the bandwidth to shepherd the feature through the process to stabilization. Often this is the case when the feature doesn't fit into the team's roadmap. Postponed ideas may be revisited later.
When an RFC is merged, the PR is merged into the RFCs repo. A new tracking
issue is created in the rust-lang/rust repo to track progress on the feature
and discuss unresolved questions, implementation progress and blockers, etc.
Here is the tracking issue on for our ?
macro feature.
Implementation
To make a change to the compiler, open a PR against the rust-lang/rust repo.
Depending on the feature/change/bug fix/improvement, implementation may be relatively-straightforward or it may be a major undertaking. You can always ask for help or mentorship from more experienced compiler devs. Also, you don't have to be the one to implement your feature; but keep in mind that if you don't, it might be a while before someone else does.
For the ?
macro feature, I needed to go understand the relevant parts of
macro expansion in the compiler. Personally, I find that improving the
comments in the code is a helpful way of making sure I understand
it, but you don't have to do that if you don't want to.
I then implemented the original feature, as described in the RFC. When
a new feature is implemented, it goes behind a feature gate, which means that
you have to use #![feature(my_feature_name)]
to use the feature. The feature
gate is removed when the feature is stabilized.
Most bug fixes and improvements don't require a feature gate. You can just make your changes/improvements.
When you open a PR on the rust-lang/rust, a bot will assign your PR to a
reviewer. If there is a particular Rust team member you are working with, you can
request that reviewer by leaving a comment on the thread with r? @reviewer-github-id
(e.g. r? @eddyb
). If you don't know who to request,
don't request anyone; the bot will assign someone automatically based on which files you changed.
The reviewer may request changes before they approve your PR, they may mark the PR with label
"S-waiting-on-author" after leaving comments, this means that the PR is blocked on you to make
some requested changes. When you finished iterating on the changes, you can mark the PR as
S-waiting-on-review
again by leaving a comment with @rustbot ready
, this will remove the
S-waiting-on-author
label and add the S-waiting-on-review
label.
Feel free to ask questions or discuss things you don't understand or disagree with. However,
recognize that the PR won't be merged unless someone on the Rust team approves
it. If a reviewer leave a comment like r=me after fixing ...
, that means they approve the PR and
you can merge it with comment with @bors r=reviewer-github-id
(e.g. @bors r=eddyb
) to merge it
after fixing trivial issues. Note that r=someone
requires permission and bors could say
something like "🔑 Insufficient privileges..." when commenting r=someone
. In that case,
you have to ask the reviewer to revisit your PR.
When your reviewer approves the PR, it will go into a queue for yet another bot
called @bors
. @bors
manages the CI build/merge queue. When your PR reaches
the head of the @bors
queue, @bors
will test out the merge by running all
tests against your PR on GitHub Actions. This takes a lot of time to
finish. If all tests pass, the PR is merged and becomes part of the next
nightly compiler!
There are a couple of things that may happen for some PRs during the review process
- If the change is substantial enough, the reviewer may request an FCP on the PR. This gives all members of the appropriate team a chance to review the changes.
- If the change may cause breakage, the reviewer may request a crater run. This compiles the compiler with your changes and then attempts to compile all crates on crates.io with your modified compiler. This is a great smoke test to check if you introduced a change to compiler behavior that affects a large portion of the ecosystem.
- If the diff of your PR is large or the reviewer is busy, your PR may have some merge conflicts with other PRs that happen to get merged first. You should fix these merge conflicts using the normal git procedures.
If you are not doing a new feature or something like that (e.g. if you are fixing a bug), then that's it! Thanks for your contribution :)
Refining your implementation
As people get experience with your new feature on nightly, slight changes may
be proposed and unresolved questions may become resolved. Updates/changes go
through the same process for implementing any other changes, as described
above (i.e. submit a PR, go through review, wait for @bors
, etc).
Some changes may be major enough to require an FCP and some review by Rust team members.
For the ?
macro feature, we went through a few different iterations after the
original implementation: 1, 2, 3.
Along the way, we decided that ?
should not take a separator, which was
previously an unresolved question listed in the RFC. We also changed the
disambiguation strategy: we decided to remove the ability to use ?
as a
separator token for other repetition operators (e.g. +
or *
). However,
since this was a breaking change, we decided to do it over an edition boundary.
Thus, the new feature can be enabled only in edition 2018. These deviations
from the original RFC required another
FCP.
Stabilization
Finally, after the feature had baked for a while on nightly, a language team member moved to stabilize it.
A stabilization report needs to be written that includes
- brief description of the behavior and any deviations from the RFC
- which edition(s) are affected and how
- links to a few tests to show the interesting aspects
The stabilization report for our feature is here.
After this, a PR is made to remove the feature gate, enabling the feature by default (on the 2018 edition). A note is added to the Release notes about the feature.
Steps to stabilize the feature can be found at Stabilizing Features.
Implementing new language features
When you want to implement a new significant feature in the compiler, you need to go through this process to make sure everything goes smoothly.
NOTE: this section is for language features, not library features, which use a different process.
The @rfcbot FCP process
When the change is small and uncontroversial, then it can be done with just writing a PR and getting an r+ from someone who knows that part of the code. However, if the change is potentially controversial, it would be a bad idea to push it without consensus from the rest of the team (both in the "distributed system" sense to make sure you don't break anything you don't know about, and in the social sense to avoid PR fights).
If such a change seems to be too small to require a full formal RFC process (e.g., a small standard library addition, a big refactoring of the code, a "technically-breaking" change, or a "big bugfix" that basically amounts to a small feature) but is still too controversial or big to get by with a single r+, you can propose a final comment period (FCP). Or, if you're not on the relevant team (and thus don't have @rfcbot permissions), ask someone who is to start one; unless they have a concern themselves, they should.
Again, the FCP process is only needed if you need consensus – if you don't think anyone would have a problem with your change, it's OK to get by with only an r+. For example, it is OK to add or modify unstable command-line flags or attributes without an FCP for compiler development or standard library use, as long as you don't expect them to be in wide use in the nightly ecosystem. Some teams have lighter weight processes that they use in scenarios like this; for example, the compiler team recommends filing a Major Change Proposal (MCP) as a lightweight way to garner support and feedback without requiring full consensus.
You don't need to have the implementation fully ready for r+ to propose an FCP, but it is generally a good idea to have at least a proof of concept so that people can see what you are talking about.
When an FCP is proposed, it requires all members of the team to sign off the FCP. After they all do so, there's a 10-day-long "final comment period" (hence the name) where everybody can comment, and if no concerns are raised, the PR/issue gets FCP approval.
The logistics of writing features
There are a few "logistic" hoops you might need to go through in order to implement a feature in a working way.
Warning Cycles
In some cases, a feature or bugfix might break some existing programs in some edge cases. In that case, you might want to do a crater run to assess the impact and possibly add a future-compatibility lint, similar to those used for edition-gated lints.
Stability
We value the stability of Rust. Code that works and runs on stable should (mostly) not break. Because of that, we don't want to release a feature to the world with only team consensus and code review - we want to gain real-world experience on using that feature on nightly, and we might want to change the feature based on that experience.
To allow for that, we must make sure users don't accidentally depend on that new feature - otherwise, especially if experimentation takes time or is delayed and the feature takes the trains to stable, it would end up de facto stable and we'll not be able to make changes in it without breaking people's code.
The way we do that is that we make sure all new features are feature
gated - they can't be used without enabling a feature gate
(#[feature(foo)]
), which can't be done in a stable/beta compiler.
See the stability in code section for the technical details.
Eventually, after we gain enough experience using the feature, make the necessary changes, and are satisfied, we expose it to the world using the stabilization process described here. Until then, the feature is not set in stone: every part of the feature can be changed, or the feature might be completely rewritten or removed. Features are not supposed to gain tenure by being unstable and unchanged for a year.
Tracking Issues
To keep track of the status of an unstable feature, the experience we get while using it on nightly, and of the concerns that block its stabilization, every feature-gate needs a tracking issue. General discussions about the feature should be done on the tracking issue.
For features that have an RFC, you should use the RFC's tracking issue for the feature.
For other features, you'll have to make a tracking issue for that feature. The issue title should be "Tracking issue for YOUR FEATURE". Use the "Tracking Issue" issue template.
Stability in code
The below steps needs to be followed in order to implement a new unstable feature:
-
Open a tracking issue - if you have an RFC, you can use the tracking issue for the RFC.
The tracking issue should be labeled with at least
C-tracking-issue
. For a language feature, a labelF-feature_name
should be added as well. -
Pick a name for the feature gate (for RFCs, use the name in the RFC).
-
Add the feature name to
rustc_span/src/symbol.rs
in theSymbols {...}
block.Note that this block must be in alphabetical order.
-
Add a feature gate declaration to
rustc_feature/src/unstable.rs
in the unstabledeclare_features
block./// description of feature (unstable, $feature_name, "CURRENT_RUSTC_VERSION", Some($tracking_issue_number))
If you haven't yet opened a tracking issue (e.g. because you want initial feedback on whether the feature is likely to be accepted), you can temporarily use
None
- but make sure to update it before the PR is merged!For example:
/// Allows defining identifiers beyond ASCII. (unstable, non_ascii_idents, "CURRENT_RUSTC_VERSION", Some(55467), None),
Features can be marked as incomplete, and trigger the warn-by-default
incomplete_features
lint by setting their type toincomplete
:/// Allows unsized rvalues at arguments and parameters. (incomplete, unsized_locals, "CURRENT_RUSTC_VERSION", Some(48055), None),
To avoid semantic merge conflicts, please use
CURRENT_RUSTC_VERSION
instead of1.70
or another explicit version number. -
Prevent usage of the new feature unless the feature gate is set. You can check it in most places in the compiler using the expression
tcx.features().$feature_name
(orsess.features_untracked().$feature_name
if the tcx is unavailable)If the feature gate is not set, you should either maintain the pre-feature behavior or raise an error, depending on what makes sense. Errors should generally use
rustc_session::parse::feature_err
. For an example of adding an error, see #81015.For features introducing new syntax, pre-expansion gating should be used instead. During parsing, when the new syntax is parsed, the symbol must be inserted to the current crate's
GatedSpans
viaself.sess.gated_span.gate(sym::my_feature, span)
.After being inserted to the gated spans, the span must be checked in the
rustc_ast_passes::feature_gate::check_crate
function, which actually denies features. Exactly how it is gated depends on the exact type of feature, but most likely will use thegate_all!()
macro. -
Add a test to ensure the feature cannot be used without a feature gate, by creating
tests/ui/feature-gates/feature-gate-$feature_name.rs
. You can generate the corresponding.stderr
file by running./x test tests/ui/feature-gates/ --bless
. -
Add a section to the unstable book, in
src/doc/unstable-book/src/language-features/$feature_name.md
. -
Write a lot of tests for the new feature, preferably in
tests/ui/$feature_name/
. PRs without tests will not be accepted! -
Get your PR reviewed and land it. You have now successfully implemented a feature in Rust!
Stability attributes
This section is about the stability attributes and schemes that allow stable APIs to use unstable APIs internally in the rustc standard library.
NOTE: this section is for library features, not language features. For instructions on stabilizing a language feature see Stabilizing Features.
- unstable
- stable
- rustc_const_unstable
- rustc_const_stable
- rustc_const_stable_indirect
- rustc_intrinsic_const_stable_indirect
- rustc_default_body_unstable
- Stabilizing a library feature
- allow_internal_unstable
- rustc_allow_const_fn_unstable
- staged_api
- deprecated
unstable
The #[unstable(feature = "foo", issue = "1234", reason = "lorem ipsum")]
attribute explicitly marks an item as unstable. Items that are marked as
"unstable" cannot be used without a corresponding #![feature]
attribute on
the crate, even on a nightly compiler. This restriction only applies across
crate boundaries, unstable items may be used within the crate that defines
them.
The issue
field specifies the associated GitHub issue number. This field is
required and all unstable features should have an associated tracking issue. In
rare cases where there is no sensible value issue = "none"
is used.
The unstable
attribute infects all sub-items, where the attribute doesn't
have to be reapplied. So if you apply this to a module, all items in the module
will be unstable.
You can make specific sub-items stable by using the #[stable]
attribute on
them. The stability scheme works similarly to how pub
works. You can have
public functions of nonpublic modules and you can have stable functions in
unstable modules or vice versa.
Previously, due to a rustc bug, stable items inside unstable modules were
available to stable code in that location.
As of September 2024, items with accidentally stabilized
paths are marked with the #[rustc_allowed_through_unstable_modules]
attribute
to prevent code dependent on those paths from breaking.
The unstable
attribute may also have the soft
value, which makes it a
future-incompatible deny-by-default lint instead of a hard error. This is used
by the bench
attribute which was accidentally accepted in the past. This
prevents breaking dependencies by leveraging Cargo's lint capping.
stable
The #[stable(feature = "foo", since = "1.420.69")]
attribute explicitly
marks an item as stabilized. Note that stable functions may use unstable things in their body.
rustc_const_unstable
The #[rustc_const_unstable(feature = "foo", issue = "1234", reason = "lorem ipsum")]
has the same interface as the unstable
attribute. It is used to mark
const fn
as having their constness be unstable. This is only needed in rare cases:
- If a
const fn
makes use of unstable language features or intrinsics. (The compiler will tell you to add the attribute if you run into this.) - If a
const fn
is#[stable]
but not yet intended to be const-stable. - To change the feature gate that is required to call a const-unstable intrinsic.
Const-stability differs from regular stability in that it is recursive: a
#[rustc_const_unstable(...)]
function cannot even be indirectly called from stable code. This is
to avoid accidentally leaking unstable compiler implementation artifacts to stable code or locking
us into the accidental quirks of an incomplete implementation. See the rustc_const_stable_indirect
and rustc_allow_const_fn_unstable attributes below for how to fine-tune this check.
rustc_const_stable
The #[rustc_const_stable(feature = "foo", since = "1.420.69")]
attribute explicitly marks
a const fn
as having its constness be stable
.
rustc_const_stable_indirect
The #[rustc_const_stable_indirect]
attribute can be added to a #[rustc_const_unstable(...)]
function to make it callable from #[rustc_const_stable(...)]
functions. This indicates that the
function is ready for stable in terms of its implementation (i.e., it doesn't use any unstable
compiler features); the only reason it is not const-stable yet are API concerns.
This should also be added to lang items for which const-calls are synthesized in the compiler, to ensure those calls do not bypass recursive const stability rules.
rustc_intrinsic_const_stable_indirect
On an intrinsic, this attribute marks the intrinsic as "ready to be used by public stable functions".
If the intrinsic has a rustc_const_unstable
attribute, it should be removed.
Adding this attribute to an intrinsic requires t-lang and wg-const-eval approval!
rustc_default_body_unstable
The #[rustc_default_body_unstable(feature = "foo", issue = "1234", reason = "lorem ipsum")]
attribute has the same interface as the unstable
attribute.
It is used to mark the default implementation for an item within a trait as
unstable.
A trait with a default-body-unstable item can be implemented stably by providing
an explicit body for any such item, or the default body can be used by enabling
its corresponding #![feature]
.
Stabilizing a library feature
To stabilize a feature, follow these steps:
- Ask a @T-libs-api member to start an FCP on the tracking issue and wait for
the FCP to complete (with
disposition-merge
). - Change
#[unstable(...)]
to#[stable(since = "CURRENT_RUSTC_VERSION")]
. - Remove
#![feature(...)]
from any test or doc-test for this API. If the feature is used in the compiler or tools, remove it from there as well. - If this is a
const fn
, add#[rustc_const_stable(since = "CURRENT_RUSTC_VERSION")]
. Alternatively, if this is not supposed to be const-stabilized yet, add#[rustc_const_unstable(...)]
for some new feature gate (with a new tracking issue). - Open a PR against
rust-lang/rust
.- Add the appropriate labels:
@rustbot modify labels: +T-libs-api
. - Link to the tracking issue and say "Closes #XXXXX".
- Add the appropriate labels:
You can see an example of stabilizing a feature with tracking issue #81656 with FCP and the associated implementation PR #84642.
allow_internal_unstable
Macros and compiler desugarings expose their bodies to the call
site. To work around not being able to use unstable things in the standard
library's macros, there's the #[allow_internal_unstable(feature1, feature2)]
attribute that allows the given features to be used in stable macros.
Note that if a macro is used in const context and generates a call to a
#[rustc_const_unstable(...)]
function, that will still be rejected even with
allow_internal_unstable
. Add #[rustc_const_stable_indirect]
to the function to ensure the macro
cannot accidentally bypass the recursive const stability checks.
rustc_allow_const_fn_unstable
As explained above, no unstable const features are allowed inside stable const fn
, not even
indirectly.
However, sometimes we do know that a feature will get stabilized, just not when, or there is a
stable (but e.g. runtime-slow) workaround, so we could always fall back to some stable version if we
scrapped the unstable feature. In those cases, the [rustc_allow_const_fn_unstable(feature1, feature2)]
attribute can be used to allow some unstable features in the body of a stable (or
indirectly stable) const fn
.
You also need to take care to uphold the const fn
invariant that calling it at runtime and
compile-time needs to behave the same (see also this blog post). This means that you
may not create a const fn
that e.g. transmutes a memory address to an integer,
because the addresses of things are nondeterministic and often unknown at
compile-time.
Always ping @rust-lang/wg-const-eval if you are adding more
rustc_allow_const_fn_unstable
attributes to any const fn
.
staged_api
Any crate that uses the stable
or unstable
attributes must include the
#![feature(staged_api)]
attribute on the crate.
deprecated
Deprecations in the standard library are nearly identical to deprecations in
user code. When #[deprecated]
is used on an item, it must also have a stable
or unstable
attribute.
deprecated
has the following form:
#[deprecated(
since = "1.38.0",
note = "explanation for deprecation",
suggestion = "other_function"
)]
The suggestion
field is optional. If given, it should be a string that can be
used as a machine-applicable suggestion to correct the warning. This is
typically used when the identifier is renamed, but no other significant changes
are necessary. When the suggestion
field is used, you need to have
#![feature(deprecated_suggestion)]
at the crate root.
Another difference from user code is that the since
field is actually checked
against the current version of rustc
. If since
is in a future version, then
the deprecated_in_future
lint is triggered which is default allow
, but most
of the standard library raises it to a warning with
#![warn(deprecated_in_future)]
.
Request for stabilization
NOTE: this page is about stabilizing language features. For stabilizing library features, see Stabilizing a library feature.
Once an unstable feature has been well-tested with no outstanding concern, anyone may push for its stabilization. It involves the following steps:
Documentation PRs
If any documentation for this feature exists, it should be
in the Unstable Book
, located at src/doc/unstable-book
.
If it exists, the page for the feature gate should be removed.
If there was documentation there, integrating it into the existing documentation is needed.
If there wasn't documentation there, it needs to be added.
Places that may need updated documentation:
- The Reference: This must be updated, in full detail.
- The Book: This may or may not need updating, depends. If you're not sure, please open an issue on this repository and it can be discussed.
- standard library documentation: As needed. Language features
often don't need this, but if it's a feature that changes
how good examples are written, such as when
?
was added to the language, updating examples is important. - Rust by Example: As needed.
Prepare PRs to update documentation involving this new feature for repositories mentioned above. Maintainers of these repositories will keep these PRs open until the whole stabilization process has completed. Meanwhile, we can proceed to the next step.
Write a stabilization report
Find the tracking issue of the feature, and create a short stabilization report. Essentially this would be a brief summary of the feature plus some links to test cases showing it works as expected, along with a list of edge cases that came up and were considered. This is a minimal "due diligence" that we do before stabilizing.
The report should contain:
- A summary, showing examples (e.g. code snippets) what is enabled by this feature.
- Links to test cases in our test suite regarding this feature and describe the feature's behavior on encountering edge cases.
- Links to the documentations (the PRs we have made in the previous steps).
- Any other relevant information.
- The resolutions of any unresolved questions if the stabilization is for an RFC.
Examples of stabilization reports can be found in rust-lang/rust#44494 and rust-lang/rust#28237 (these links will bring you directly to the comment containing the stabilization report).
FCP
If any member of the team responsible for tracking this feature agrees with stabilizing this feature, they will start the FCP (final-comment-period) process by commenting
@rfcbot fcp merge
The rest of the team members will review the proposal. If the final decision is to stabilize, we proceed to do the actual code modification.
Stabilization PR
This is for stabilizing language features. If you are stabilizing a library feature, see the stabilization chapter of the std dev guide instead.
Once we have decided to stabilize a feature, we need to have a PR that actually makes that stabilization happen. These kinds of PRs are a great way to get involved in Rust, as they take you on a little tour through the source code.
Here is a general guide to how to stabilize a feature -- every feature is different, of course, so some features may require steps beyond what this guide talks about.
Note: Before we stabilize any feature, it's the rule that it should appear in the documentation.
Updating the feature-gate listing
There is a central listing of unstable feature-gates in
compiler/rustc_feature/src/unstable.rs
. Search for the declare_features!
macro. There should be an entry for the feature you are aiming
to stabilize, something like (this example is taken from
rust-lang/rust#32409:
// pub(restricted) visibilities (RFC 1422)
(unstable, pub_restricted, "CURRENT_RUSTC_VERSION", Some(32409)),
The above line should be moved to compiler/rustc_feature/src/accepted.rs
.
Entries in the declare_features!
call are sorted, so find the correct place.
When it is done, it should look like:
// pub(restricted) visibilities (RFC 1422)
(accepted, pub_restricted, "CURRENT_RUSTC_VERSION", Some(32409)),
// note that we changed this
(Even though you will encounter version numbers in the file of past changes,
you should not put the rustc version you expect your stabilization to happen in,
but instead CURRENT_RUSTC_VERSION
)
Removing existing uses of the feature-gate
Next search for the feature string (in this case, pub_restricted
)
in the codebase to find where it appears. Change uses of
#![feature(XXX)]
from the std
and any rustc crates (this includes test folders
under library/
and compiler/
but not the toplevel tests/
one) to be
#![cfg_attr(bootstrap, feature(XXX))]
. This includes the feature-gate
only for stage0, which is built using the current beta (this is
needed because the feature is still unstable in the current beta).
Also, remove those strings from any tests (e.g. under tests/
). If there are tests
specifically targeting the feature-gate (i.e., testing that the
feature-gate is required to use the feature, but nothing else),
simply remove the test.
Do not require the feature-gate to use the feature
Most importantly, remove the code which flags an error if the
feature-gate is not present (since the feature is now considered
stable). If the feature can be detected because it employs some
new syntax, then a common place for that code to be is in the
same compiler/rustc_ast_passes/src/feature_gate.rs
.
For example, you might see code like this:
gate_feature_post!(&self, pub_restricted, span,
"`pub(restricted)` syntax is experimental");
This gate_feature_post!
macro prints an error if the
pub_restricted
feature is not enabled. It is not needed
now that #[pub_restricted]
is stable.
For more subtle features, you may find code like this:
if self.tcx.sess.features.borrow().pub_restricted { /* XXX */ }
This pub_restricted
field (obviously named after the feature)
would ordinarily be false if the feature flag is not present
and true if it is. So transform the code to assume that the field
is true. In this case, that would mean removing the if
and
leaving just the /* XXX */
.
if self.tcx.sess.features.borrow().pub_restricted { /* XXX */ }
becomes
/* XXX */
if self.tcx.sess.features.borrow().pub_restricted && something { /* XXX */ }
becomes
if something { /* XXX */ }
Feature Gates
This chapter is intended to provide basic help for adding, removing, and modifying feature gates.
Note that this is specific to language feature gates; library feature gates use a different mechanism.
Adding a feature gate
See "Stability in code" in the "Implementing new features" section for instructions.
Removing a feature gate
To remove a feature gate, follow these steps:
-
Remove the feature gate declaration in
rustc_feature/src/unstable.rs
. It will look like this:/// description of feature (unstable, $feature_name, "$version", Some($tracking_issue_number))
-
Add a modified version of the feature gate declaration that you just removed to
rustc_feature/src/removed.rs
:/// description of feature (removed, $old_feature_name, "$version", Some($tracking_issue_number), Some("$why_it_was_removed"))
Renaming a feature gate
To rename a feature gate, follow these steps (the first two are the same steps to follow when removing a feature gate):
-
Remove the old feature gate declaration in
rustc_feature/src/unstable.rs
. It will look like this:/// description of feature (unstable, $old_feature_name, "$version", Some($tracking_issue_number))
-
Add a modified version of the old feature gate declaration that you just removed to
rustc_feature/src/removed.rs
:/// description of feature /// Renamed to `$new_feature_name` (removed, $old_feature_name, "$version", Some($tracking_issue_number), Some("renamed to `$new_feature_name`"))
-
Add a feature gate declaration with the new name to
rustc_feature/src/unstable.rs
. It should look very similar to the old declaration:/// description of feature (unstable, $new_feature_name, "$version", Some($tracking_issue_number))
Stabilizing a feature
See "Updating the feature-gate listing" in the "Stabilizing Features" chapter for instructions. There are additional steps you will need to take beyond just updating the declaration!
This file offers some tips on the coding conventions for rustc. This chapter covers formatting, coding for correctness, using crates from crates.io, and some tips on structuring your PR for easy review.
Formatting and the tidy script
rustc is moving towards the Rust standard coding style.
However, for now we don't use stable rustfmt
; we use a pinned version with a
special config, so this may result in different style from normal rustfmt
.
Therefore, formatting this repository using cargo fmt
is not recommended.
Instead, formatting should be done using ./x fmt
. It's a good habit to run
./x fmt
before every commit, as this reduces conflicts later.
Formatting is checked by the tidy
script. It runs automatically when you do
./x test
and can be run in isolation with ./x fmt --check
.
If you want to use format-on-save in your editor, the pinned version of
rustfmt
is built under build/<target>/stage0/bin/rustfmt
. You'll have to
pass the --edition=2021
argument yourself when calling
rustfmt
directly.
Formatting C++ code
The compiler contains some C++ code for interfacing with parts of LLVM that don't have a stable C API. When modifying that code, use this command to format it:
./x test tidy --extra-checks=cpp:fmt --bless
This uses a pinned version of clang-format
, to avoid relying on the local
environment.
Copyright notice
In the past, files began with a copyright and license notice. Please omit this notice for new files licensed under the standard terms (dual MIT/Apache-2.0).
All of the copyright notices should be gone by now, but if you come across one in the rust-lang/rust repo, feel free to open a PR to remove it.
Line length
Lines should be at most 100 characters. It's even better if you can keep things to 80.
Ignoring the line length limit. Sometimes – in particular for tests – it can be necessary to exempt yourself from this limit. In that case, you can add a comment towards the top of the file like so:
#![allow(unused)] fn main() { // ignore-tidy-linelength }
Tabs vs spaces
Prefer 4-space indent.
Coding for correctness
Beyond formatting, there are a few other tips that are worth following.
Prefer exhaustive matches
Using _
in a match is convenient, but it means that when new
variants are added to the enum, they may not get handled correctly.
Ask yourself: if a new variant were added to this enum, what's the
chance that it would want to use the _
code, versus having some
other treatment? Unless the answer is "low", then prefer an
exhaustive match. (The same advice applies to if let
and while let
, which are effectively tests for a single variant.)
Use "TODO" comments for things you don't want to forget
As a useful tool to yourself, you can insert a // TODO
comment
for something that you want to get back to before you land your PR:
fn do_something() {
if something_else {
unimplemented!(); // TODO write this
}
}
The tidy script will report an error for a // TODO
comment, so this
code would not be able to land until the TODO is fixed (or removed).
This can also be useful in a PR as a way to signal from one commit that you are leaving a bug that a later commit will fix:
if foo {
return true; // TODO wrong, but will be fixed in a later commit
}
Using crates from crates.io
See the crates.io dependencies section.
How to structure your PR
How you prepare the commits in your PR can make a big difference for the reviewer. Here are some tips.
Isolate "pure refactorings" into their own commit. For example, if you rename a method, then put that rename into its own commit, along with the renames of all the uses.
More commits is usually better. If you are doing a large change, it's almost always better to break it up into smaller steps that can be independently understood. The one thing to be aware of is that if you introduce some code following one strategy, then change it dramatically (versus adding to it) in a later commit, that 'back-and-forth' can be confusing.
Format liberally. While only the final commit of a PR must be correctly
formatted, it is both easier to review and less noisy to format each commit
individually using ./x fmt
.
No merges. We do not allow merge commits into our history, other
than those by bors. If you get a merge conflict, rebase instead via a
command like git rebase -i rust-lang/master
(presuming you use the
name rust-lang
for your remote).
Individual commits do not have to build (but it's nice). We do not require that every intermediate commit successfully builds – we only expect to be able to bisect at a PR level. However, if you can make individual commits build, that is always helpful.
Naming conventions
Apart from normal Rust style/naming conventions, there are also some specific to the compiler.
-
cx
tends to be short for "context" and is often used as a suffix. For example,tcx
is a common name for the Typing Context. -
'tcx
is used as the lifetime name for the Typing Context. -
Because
crate
is a keyword, if you need a variable to represent something crate-related, often the spelling is changed tokrate
.
Procedures for Breaking Changes
- Motivation
- Detailed design
- Tracking issue
- Tracking issue template
- Issuing future compatibility warnings
- Helpful techniques
- Crater and crates.io
- Is it ever acceptable to go directly to issuing errors?
- Stabilization
- Removing a lint
- Remove the lint.
- Add the lint to the list of removed lints.
- Update the places that issue the lint
- Update tests
- All done!
This page defines the best practices procedure for making bug fixes or soundness corrections in the compiler that can cause existing code to stop compiling. This text is based on RFC 1589.
Motivation
From time to time, we encounter the need to make a bug fix, soundness correction, or other change in the compiler which will cause existing code to stop compiling. When this happens, it is important that we handle the change in a way that gives users of Rust a smooth transition. What we want to avoid is that existing programs suddenly stop compiling with opaque error messages: we would prefer to have a gradual period of warnings, with clear guidance as to what the problem is, how to fix it, and why the change was made. This RFC describes the procedure that we have been developing for handling breaking changes that aims to achieve that kind of smooth transition.
One of the key points of this policy is that (a) warnings should be issued initially rather than hard errors if at all possible and (b) every change that causes existing code to stop compiling will have an associated tracking issue. This issue provides a point to collect feedback on the results of that change. Sometimes changes have unexpectedly large consequences or there may be a way to avoid the change that was not considered. In those cases, we may decide to change course and roll back the change, or find another solution (if warnings are being used, this is particularly easy to do).
What qualifies as a bug fix?
Note that this RFC does not try to define when a breaking change is permitted. That is already covered under RFC 1122. This document assumes that the change being made is in accordance with those policies. Here is a summary of the conditions from RFC 1122:
- Soundness changes: Fixes to holes uncovered in the type system.
- Compiler bugs: Places where the compiler is not implementing the specified semantics found in an RFC or lang-team decision.
- Underspecified language semantics: Clarifications to grey areas where the compiler behaves inconsistently and no formal behavior had been previously decided.
Please see the RFC for full details!
Detailed design
The procedure for making a breaking change is as follows (each of these steps is described in more detail below):
- Do a crater run to assess the impact of the change.
- Make a special tracking issue dedicated to the change.
- Do not report an error right away. Instead, issue forwards-compatibility
lint warnings.
- Sometimes this is not straightforward. See the text below for suggestions on different techniques we have employed in the past.
- For cases where warnings are infeasible:
- Report errors, but make every effort to give a targeted error message that directs users to the tracking issue
- Submit PRs to all known affected crates that fix the issue
- or, at minimum, alert the owners of those crates to the problem and direct them to the tracking issue
- Once the change has been in the wild for at least one cycle, we can stabilize the change, converting those warnings into errors.
Finally, for changes to rustc_ast
that will affect plugins, the general policy
is to batch these changes. That is discussed below in more detail.
Tracking issue
Every breaking change should be accompanied by a dedicated tracking issue for that change. The main text of this issue should describe the change being made, with a focus on what users must do to fix their code. The issue should be approachable and practical; it may make sense to direct users to an RFC or some other issue for the full details. The issue also serves as a place where users can comment with questions or other concerns.
A template for these breaking-change tracking issues can be found below. An example of how such an issue should look can be found here.
The issue should be tagged with (at least) B-unstable
and T-compiler
.
Tracking issue template
This is a template to use for tracking issues:
This is the **summary issue** for the `YOUR_LINT_NAME_HERE`
future-compatibility warning and other related errors. The goal of
this page is describe why this change was made and how you can fix
code that is affected by it. It also provides a place to ask questions
or register a complaint if you feel the change should not be made. For
more information on the policy around future-compatibility warnings,
see our [breaking change policy guidelines][guidelines].
[guidelines]: LINK_TO_THIS_RFC
#### What is the warning for?
*Describe the conditions that trigger the warning and how they can be
fixed. Also explain why the change was made.**
#### When will this warning become a hard error?
At the beginning of each 6-week release cycle, the Rust compiler team
will review the set of outstanding future compatibility warnings and
nominate some of them for **Final Comment Period**. Toward the end of
the cycle, we will review any comments and make a final determination
whether to convert the warning into a hard error or remove it
entirely.
Issuing future compatibility warnings
The best way to handle a breaking change is to begin by issuing future-compatibility warnings. These are a special category of lint warning. Adding a new future-compatibility warning can be done as follows.
#![allow(unused)] fn main() { // 1. Define the lint in `compiler/rustc_middle/src/lint/builtin.rs`: declare_lint! { pub YOUR_ERROR_HERE, Warn, "illegal use of foo bar baz" } // 2. Add to the list of HardwiredLints in the same file: impl LintPass for HardwiredLints { fn get_lints(&self) -> LintArray { lint_array!( .., YOUR_ERROR_HERE ) } } // 3. Register the lint in `compiler/rustc_lint/src/lib.rs`: store.register_future_incompatible(sess, vec![ ..., FutureIncompatibleInfo { id: LintId::of(YOUR_ERROR_HERE), reference: "issue #1234", // your tracking issue here! }, ]); // 4. Report the lint: tcx.lint_node( lint::builtin::YOUR_ERROR_HERE, path_id, binding.span, format!("some helper message here")); }
Helpful techniques
It can often be challenging to filter out new warnings from older, pre-existing errors. One technique that has been used in the past is to run the older code unchanged and collect the errors it would have reported. You can then issue warnings for any errors you would give which do not appear in that original set. Another option is to abort compilation after the original code completes if errors are reported: then you know that your new code will only execute when there were no errors before.
Crater and crates.io
Crater is a bot that will compile all crates.io crates and many public github repos with the compiler with your changes. A report will then be generated with crates that ceased to compile with or began to compile with your changes. Crater runs can take a few days to complete.
We should always do a crater run to assess impact. It is polite and considerate to at least notify the authors of affected crates the breaking change. If we can submit PRs to fix the problem, so much the better.
Is it ever acceptable to go directly to issuing errors?
Changes that are believed to have negligible impact can go directly to issuing
an error. One rule of thumb would be to check against crates.io
: if fewer than
10 total affected projects are found (not root errors), we can move
straight to an error. In such cases, we should still make the "breaking change"
page as before, and we should ensure that the error directs users to this page.
In other words, everything should be the same except that users are getting an
error, and not a warning. Moreover, we should submit PRs to the affected
projects (ideally before the PR implementing the change lands in rustc).
If the impact is not believed to be negligible (e.g., more than 10 crates are affected), then warnings are required (unless the compiler team agrees to grant a special exemption in some particular case). If implementing warnings is not feasible, then we should make an aggressive strategy of migrating crates before we land the change so as to lower the number of affected crates. Here are some techniques for approaching this scenario:
- Issue warnings for subparts of the problem, and reserve the new errors for the smallest set of cases you can.
- Try to give a very precise error message that suggests how to fix the problem and directs users to the tracking issue.
- It may also make sense to layer the fix:
- First, add warnings where possible and let those land before proceeding to issue errors.
- Work with authors of affected crates to ensure that corrected versions are available before the fix lands, so that downstream users can use them.
Stabilization
After a change is made, we will stabilize the change using the same process that we use for unstable features:
-
After a new release is made, we will go through the outstanding tracking issues corresponding to breaking changes and nominate some of them for final comment period (FCP).
-
The FCP for such issues lasts for one cycle. In the final week or two of the cycle, we will review comments and make a final determination:
- Convert to error: the change should be made into a hard error.
- Revert: we should remove the warning and continue to allow the older code to compile.
- Defer: can't decide yet, wait longer, or try other strategies.
Ideally, breaking changes should have landed on the stable branch of the compiler before they are finalized.
Removing a lint
Once we have decided to make a "future warning" into a hard error, we need a PR
that removes the custom lint. As an example, here are the steps required to
remove the overlapping_inherent_impls
compatibility lint. First, convert the
name of the lint to uppercase (OVERLAPPING_INHERENT_IMPLS
) ripgrep through the
source for that string. We will basically by converting each place where this
lint name is mentioned (in the compiler, we use the upper-case name, and a macro
automatically generates the lower-case string; so searching for
overlapping_inherent_impls
would not find much).
NOTE: these exact files don't exist anymore, but the procedure is still the same.
Remove the lint.
The first reference you will likely find is the lint definition in
rustc_session/src/lint/builtin.rs
that resembles this:
#![allow(unused)] fn main() { declare_lint! { pub OVERLAPPING_INHERENT_IMPLS, Deny, // this may also say Warning "two overlapping inherent impls define an item with the same name were erroneously allowed" } }
This declare_lint!
macro creates the relevant data structures. Remove it. You
will also find that there is a mention of OVERLAPPING_INHERENT_IMPLS
later in
the file as part of a lint_array!
; remove it too.
Next, you see a reference to OVERLAPPING_INHERENT_IMPLS
in
rustc_lint/src/lib.rs
. This is defining the lint as a "future
compatibility lint":
#![allow(unused)] fn main() { FutureIncompatibleInfo { id: LintId::of(OVERLAPPING_INHERENT_IMPLS), reference: "issue #36889 <https://github.com/rust-lang/rust/issues/36889>", }, }
Remove this too.
Add the lint to the list of removed lints.
In compiler/rustc_lint/src/lib.rs
there is a list of "renamed and removed lints".
You can add this lint to the list:
#![allow(unused)] fn main() { store.register_removed("overlapping_inherent_impls", "converted into hard error, see #36889"); }
where #36889
is the tracking issue for your lint.
Update the places that issue the lint
Finally, the last class of references you will see are the places that actually
trigger the lint itself (i.e., what causes the warnings to appear). These
you do not want to delete. Instead, you want to convert them into errors. In
this case, the add_lint
call looks like this:
#![allow(unused)] fn main() { self.tcx.sess.add_lint(lint::builtin::OVERLAPPING_INHERENT_IMPLS, node_id, self.tcx.span_of_impl(item1).unwrap(), msg); }
We want to convert this into an error. In some cases, there may be an existing error for this scenario. In others, we will need to allocate a fresh diagnostic code. Instructions for allocating a fresh diagnostic code can be found here. You may want to mention in the extended description that the compiler behavior changed on this point, and include a reference to the tracking issue for the change.
Let's say that we've adopted E0592
as our code. Then we can change the
add_lint()
call above to something like:
#![allow(unused)] fn main() { struct_span_code_err!(self.dcx(), self.tcx.span_of_impl(item1).unwrap(), E0592, msg) .emit(); }
Update tests
Finally, run the test suite. These should be some tests that used to reference
the overlapping_inherent_impls
lint, those will need to be updated. In
general, if the test used to have #[deny(overlapping_inherent_impls)]
, that
can just be removed.
./x test
All done!
Open a PR. =)
Using External Repositories
The rust-lang/rust
git repository depends on several other repos in the rust-lang
organization.
There are three main ways we use dependencies:
- As a Cargo dependency through crates.io (e.g.
rustc-rayon
) - As a git subtree (e.g.
clippy
) - As a git submodule (e.g.
cargo
)
As a general rule, use crates.io for libraries that could be useful for others in the ecosystem; use subtrees for tools that depend on compiler internals and need to be updated if there are breaking changes; and use submodules for tools that are independent of the compiler.
External Dependencies (subtree)
As a developer to this repository, you don't have to treat the following external projects differently from other crates that are directly in this repo:
In contrast to submodule
dependencies
(see below for those), the subtree
dependencies are just regular files and directories which can
be updated in tree. However, if possible, enhancements, bug fixes, etc. specific
to these tools should be filed against the tools directly in their respective
upstream repositories. The exception is that when rustc changes are required to
implement a new tool feature or test, that should happen in one collective rustc PR.
Synchronizing a subtree
Periodically the changes made to subtree based dependencies need to be synchronized between this repository and the upstream tool repositories.
Subtree synchronizations are typically handled by the respective tool maintainers. Other users are welcome to submit synchronization PRs, however, in order to do so you will need to modify your local git installation and follow a very precise set of instructions. These instructions are documented, along with several useful tips and tricks, in the syncing subtree changes section in Clippy's Contributing guide. The instructions are applicable for use with any subtree based tool, just be sure to use the correct corresponding subtree directory and remote repository.
The synchronization process goes in two directions: subtree push
and subtree pull
.
A subtree push
takes all the changes that happened to the copy in this repo and creates commits
on the remote repo that match the local changes. Every local
commit that touched the subtree causes a commit on the remote repo, but
is modified to move the files from the specified directory to the tool repo root.
A subtree pull
takes all changes since the last subtree pull
from the tool repo and adds these commits to the rustc repo along with a merge commit that moves
the tool changes into the specified directory in the Rust repository.
It is recommended that you always do a push first and get that merged to the tool master branch.
Then, when you do a pull, the merge works without conflicts.
While it's definitely possible to resolve conflicts during a pull, you may have to redo the conflict
resolution if your PR doesn't get merged fast enough and there are new conflicts. Do not try to
rebase the result of a git subtree pull
, rebasing merge commits is a bad idea in general.
You always need to specify the -P
prefix to the subtree directory and the corresponding remote
repository. If you specify the wrong directory or repository
you'll get very fun merges that try to push the wrong directory to the wrong remote repository.
Luckily you can just abort this without any consequences by throwing away either the pulled commits
in rustc or the pushed branch on the remote and try again. It is usually fairly obvious
that this is happening because you suddenly get thousands of commits that want to be synchronized.
Creating a new subtree dependency
If you want to create a new subtree dependency from an existing repository, call (from this repository's root directory!)
git subtree add -P src/tools/clippy https://github.com/rust-lang/rust-clippy.git master
This will create a new commit, which you may not rebase under any circumstances! Delete the commit and redo the operation if you need to rebase.
Now you're done, the src/tools/clippy
directory behaves as if Clippy were
part of the rustc monorepo, so no one but you (or others that synchronize
subtrees) actually needs to use git subtree
.
External Dependencies (submodules)
Building Rust will also use external git repositories tracked using git
submodules. The complete list may be found in the .gitmodules
file. Some
of these projects are required (like stdarch
for the standard library) and
some of them are optional (like src/doc/book
).
Usage of submodules is discussed more in the Using Git chapter.
Some of the submodules are allowed to be in a "broken" state where they either don't build or their tests don't pass, e.g. the documentation books like The Rust Reference. Maintainers of these projects will be notified when the project is in a broken state, and they should fix them as soon as possible. The current status is tracked on the toolstate website. More information may be found on the Forge Toolstate chapter. In practice, it is very rare for documentation to have broken toolstate.
Breakage is not allowed in the beta and stable channels, and must be addressed before the PR is merged. They are also not allowed to be broken on master in the week leading up to the beta cut.
Fuzzing
For the purposes of this guide, fuzzing is any testing methodology that involves compiling a wide variety of programs in an attempt to uncover bugs in rustc. Fuzzing is often used to find internal compiler errors (ICEs). Fuzzing can be beneficial, because it can find bugs before users run into them and provide small, self-contained programs that make the bug easier to track down. However, some common mistakes can reduce the helpfulness of fuzzing and end up making contributors' lives harder. To maximize your positive impact on the Rust project, please read this guide before reporting fuzzer-generated bugs!
Guidelines
In a nutshell
Please do:
- Ensure the bug is still present on the latest nightly rustc
- Include a reasonably minimal, standalone example along with any bug report
- Include all of the information requested in the bug report template
- Search for existing reports with the same message and query stack
- Format the test case with
rustfmt
, if it maintains the bug - Indicate that the bug was found by fuzzing
Please don't:
- Don't report lots of bugs that use internal features, including but not
limited to
custom_mir
,lang_items
,no_core
, andrustc_attrs
. - Don't seed your fuzzer with inputs that are known to crash rustc (details below).
Discussion
If you're not sure whether or not an ICE is a duplicate of one that's already been reported, please go ahead and report it and link to issues you think might be related. In general, ICEs on the same line but with different query stacks are usually distinct bugs. For example, #109020 and #109129 had similar error messages:
error: internal compiler error: compiler/rustc_middle/src/ty/normalize_erasing_regions.rs:195:90: Failed to normalize <[closure@src/main.rs:36:25: 36:28] as std::ops::FnOnce<(Emplacable<()>,)>>::Output, maybe try to call `try_normalize_erasing_regions` instead
error: internal compiler error: compiler/rustc_middle/src/ty/normalize_erasing_regions.rs:195:90: Failed to normalize <() as Project>::Assoc, maybe try to call `try_normalize_erasing_regions` instead
but different query stacks:
query stack during panic:
#0 [fn_abi_of_instance] computing call ABI of `<[closure@src/main.rs:36:25: 36:28] as core::ops::function::FnOnce<(Emplacable<()>,)>>::call_once - shim(vtable)`
end of query stack
query stack during panic:
#0 [check_mod_attrs] checking attributes in top-level module
#1 [analysis] running analysis passes on this crate
end of query stack
Building a corpus
When building a corpus, be sure to avoid collecting tests that are already known to crash rustc. A fuzzer that is seeded with such tests is more likely to generate bugs with the same root cause, wasting everyone's time. The simplest way to avoid this is to loop over each file in the corpus, see if it causes an ICE, and remove it if so.
To build a corpus, you may want to use:
- The rustc/rust-analyzer/clippy test suites (or even source code) --- though avoid
tests that are already known to cause failures, which often begin with comments
like
// failure-status: 101
or// known-bug: #NNN
. - The already-fixed ICEs in Glacier --- though avoid the unfixed
ones in
ices/
!
Extra credit
Here are a few things you can do to help the Rust project after filing an ICE.
- Bisect the bug to figure out when it was introduced
- Fix "distractions": problems with the test case that don't contribute to triggering the ICE, such as syntax errors or borrow-checking errors
- Minimize the test case (see below)
- Add the minimal test case to Glacier
Minimization
It is helpful to carefully minimize the fuzzer-generated input. When minimizing, be careful to preserve the original error, and avoid introducing distracting problems such as syntax, type-checking, or borrow-checking errors.
There are some tools that can help with minimization. If you're not sure how
to avoid introducing syntax, type-, and borrow-checking errors while using
these tools, post both the complete and minimized test cases. Generally,
syntax-aware tools give the best results in the least amount of time.
treereduce-rust
and picireny are syntax-aware.
halfempty
is not, but is generally a high-quality tool.
Effective fuzzing
When fuzzing rustc, you may want to avoid generating machine code, since this
is mostly done by LLVM. Try --emit=mir
instead.
A variety of compiler flags can uncover different issues. -Zmir-opt-level=4
will turn on MIR optimization passes that are not run by default, potentially
uncovering interesting bugs. -Zvalidate-mir
can help uncover such bugs.
If you're fuzzing a compiler you built, you may want to build it with -C target-cpu=native
or even PGO/BOLT to squeeze out a few more executions per
second. Of course, it's best to try multiple build configurations and see
what actually results in superior throughput.
You may want to build rustc from source with debug assertions to find
additional bugs, though this is a trade-off: it can slow down fuzzing by
requiring extra work for every execution. To enable debug assertions, add this
to config.toml
when compiling rustc:
[rust]
debug-assertions = true
ICEs that require debug assertions to reproduce should be tagged
requires-debug-assertions
.
Existing projects
- fuzz-rustc demonstrates how to fuzz rustc with libfuzzer
- icemaker runs rustc and other tools on a large number of source files with a variety of flags to catch ICEs
- tree-splicer generates new source files by combining existing ones while maintaining correct syntax
Notification groups
The notification groups are an easy way to help out with rustc in a "piece-meal" fashion, without committing to a larger project. Notification groups are easy to join (just submit a PR!) and joining does not entail any particular commitment.
Once you join a notification group, you will be added to a list that receives pings on github whenever a new issue is found that fits the notification group's criteria. If you are interested, you can then claim the issue and start working on it.
Of course, you don't have to wait for new issues to be tagged! If you prefer, you can use the Github label for a notification group to search for existing issues that haven't been claimed yet.
List of notification groups
Here's the list of the notification groups:
What issues are a good fit for notification groups?
Notification groups tend to get pinged on isolated bugs, particularly those of middle priority:
- By isolated, we mean that we do not expect large-scale refactoring to be required to fix the bug.
- By middle priority, we mean that we'd like to see the bug fixed, but it's not such a burning problem that we are dropping everything else to fix it. The danger with such bugs, of course, is that they can accumulate over time, and the role of the notification group is to try and stop that from happening!
Joining a notification group
To join a notification group, you just have to open a PR adding your Github username to the appropriate file in the Rust team repository. See the "example PRs" below to get a precise idea and to identify the file to edit.
Also, if you are not already a member of a Rust team then -- in addition to adding your name to the file -- you have to checkout the repository and run the following command:
cargo run add-person $your_user_name
Example PRs:
- Example of adding yourself to the Apple group.
- Example of adding yourself to the ARM group.
- Example of adding yourself to the Cleanup Crew.
- Example of adding yourself to the Emscripten group.
- Example of adding yourself to the LLVM group.
- Example of adding yourself to the RISC-V group.
- Example of adding yourself to the WASI group.
- Example of adding yourself to the WebAssembly group.
- Example of adding yourself to the Windows group.
Tagging an issue for a notification group
To tag an issue as appropriate for a notification group, you give
rustbot a ping
command with the name of the notification
group. For example:
@rustbot ping apple
@rustbot ping arm
@rustbot ping cleanup-crew
@rustbot ping emscripten
@rustbot ping llvm
@rustbot ping risc-v
@rustbot ping wasi
@rustbot ping wasm
@rustbot ping windows
To make some commands shorter and easier to remember, there are aliases,
defined in the triagebot.toml
file. For example, all of these commands
are equivalent and will ping the Cleanup Crew:
@rustbot ping cleanup
@rustbot ping bisect
@rustbot ping reduce
Keep in mind that these aliases are meant to make humans' life easier. They might be subject to change. If you need to ensure that a command will always be valid, prefer the full invocations over the aliases.
Note though that this should only be done by compiler team members or contributors, and is typically done as part of compiler team triage.
Apple notification group
Github Labels: O-macos, O-ios, O-tvos, O-watchos and O-visionos
Ping command: @rustbot ping apple
This list will be used to ask for help both in diagnosing and testing Apple-related issues as well as suggestions on how to resolve interesting questions regarding our macOS/iOS/tvOS/watchOS/visionOS support.
To get a better idea for what the group will do, here are some examples of the kinds of questions where we would have reached out to the group for advice in determining the best course of action:
- Raising the minimum supported versions (e.g. #104385)
- Additional Apple targets (e.g. #121419)
- Obscure Xcode linker details (e.g. #121430)
Deployment targets
Apple platforms have a concept of "deployment target", controlled with the
*_DEPLOYMENT_TARGET
environment variables, and specifies the minimum OS
version that a binary runs on.
Using an API from a newer OS version in the standard library than the default
that rustc
uses will result in either a static or a dynamic linker error.
For this reason, try to suggest that people document on extern "C"
APIs
which OS version they were introduced with, and if that's newer than the
current default used by rustc
, suggest to use weak linking.
The App Store and private APIs
Apple are very protective about using undocumented APIs, so it's important that whenever a change uses a new function, that they are verified to actually be public API, as even just mentioning undocumented APIs in the binary (without calling it) can lead to rejections from the App Store.
For example, Darwin / the XNU kernel actually has futex syscalls, but we can't
use them in std
because they are not public API.
In general, for an API to be considered public by Apple, it has to:
- Appear in a public header (i.e. one distributed with Xcode, and found for
the specific platform under
xcrun --show-sdk-path --sdk $SDK
). - Have an availability attribute on it (like
__API_AVAILABLE
,API_AVAILABLE
or similar).
ARM notification group
Github Label: O-ARM
Ping command: @rustbot ping arm
This list will be used to ask for help both in diagnosing and testing ARM-related issues as well as suggestions on how to resolve interesting questions regarding our ARM support.
The group also has an associated Zulip stream (#t-compiler/arm
)
where people can go to pose questions and discuss ARM-specific
topics.
So, if you are interested in participating, please sign up for the ARM group! To do so, open a PR against the rust-lang/team repository. Just follow this example, but change the username to your own!
Cleanup Crew
Github Label: ICEBreaker-Cleanup-Crew
Ping command: @rustbot ping cleanup-crew
The "Cleanup Crew" are focused on improving bug reports. Specifically, the goal is to try to ensure that every bug report has all the information that will be needed for someone to fix it:
- a minimal, standalone example that shows the problem
- links to duplicates or related bugs
- if the bug is a regression (something that used to work, but no longer does), then a bisection to the PR or nightly that caused the regression
This kind of cleanup is invaluable in getting bugs fixed. Better still, it can be done by anybody who knows Rust, without any particularly deep knowledge of the compiler.
Let's look a bit at the workflow for doing "cleanup crew" actions.
Finding a minimal, standalone example
Here the ultimate goal is to produce an example that reproduces the same problem but without relying on any external crates. Such a test ought to contain as little code as possible, as well. This will make it much easier to isolate the problem.
However, even if the "ultimate minimal test" cannot be achieved, it's still useful to post incremental minimizations. For example, if you can eliminate some of the external dependencies, that is helpful, and so forth.
It's particularly useful to reduce to an example that works in the Rust playground, rather than requiring people to checkout a cargo build.
There are many resources for how to produce minimized test cases. Here are a few:
- The rust-reduce tool can try to reduce code automatically.
- pnkfelix's Rust Bug Minimization Patterns blog post
- This post focuses on "heavy bore" techniques, where you are starting with a large, complex cargo project that you wish to narrow down to something standalone.
Links to duplicate or related bugs
If you are on the "Cleanup Crew", you will sometimes see multiple bug reports that seem very similar. You can link one to the other just by mentioning the other bug number in a Github comment. Sometimes it is useful to close duplicate bugs. But if you do so, you should always copy any test case from the bug you are closing to the other bug that remains open, as sometimes duplicate-looking bugs will expose different facets of the same problem.
Bisecting regressions
For regressions (something that used to work, but no longer does), it is super useful if we can figure out precisely when the code stopped working. The gold standard is to be able to identify the precise PR that broke the code, so we can ping the author, but even narrowing it down to a nightly build is helpful, especially as that then gives us a range of PRs. (One other challenge is that we sometimes land "rollup" PRs, which combine multiple PRs into one.)
cargo-bisect-rustc
To help in figuring out the cause of a regression we have a tool called cargo-bisect-rustc. It will automatically download and test various builds of rustc. For recent regressions, it is even able to use the builds from our CI to track down the regression to a specific PR; for older regressions, it will simply identify a nightly.
To learn to use cargo-bisect-rustc, check out this blog post, which
gives a quick introduction to how it works. Additionally, there is a Guide
which goes into more detail on how to use it. You can also ask questions at
the Zulip stream #t-compiler/cargo-bisect-rustc
, or help in
improving the tool.
Emscripten notification group
Github Label: O-emscripten
Ping command: @rustbot ping emscripten
This list will be used to ask for help both in diagnosing and testing Emscripten-related issues as well as suggestions on how to resolve interesting questions regarding our Emscripten support.
The group also has an associated Zulip stream (#t-compiler/wasm
)
where people can go to pose questions and discuss Emscripten-specific
topics.
So, if you are interested in participating, please sign up for the Emscripten group! To do so, open a PR against the rust-lang/team repository. Just follow this example, but change the username to your own!
LLVM Notification group
Github Label: A-LLVM
Ping command: @rustbot ping llvm
The "LLVM Notification Group" are focused on bugs that center around LLVM. These bugs often arise because of LLVM optimizations gone awry, or as the result of an LLVM upgrade. The goal here is:
- to determine whether the bug is a result of us generating invalid LLVM IR, or LLVM misoptimizing;
- if the former, to fix our IR;
- if the latter, to try and file a bug on LLVM (or identify an existing bug).
The group may also be asked to weigh in on other sorts of LLVM-focused questions.
Helpful tips and options
The "Debugging LLVM" section of the
rustc-dev-guide gives a step-by-step process for how to help debug bugs
caused by LLVM. In particular, it discusses how to emit LLVM IR, run
the LLVM IR optimization pipelines, and so forth. You may also find
it useful to look at the various codegen options listed under -C help
and the internal options under -Z help
-- there are a number that
pertain to LLVM (just search for LLVM).
If you do narrow to an LLVM bug
The "Debugging LLVM" section also describes what to do once you've identified the bug.
RISC-V notification group
Github Label: O-riscv
Ping command: @rustbot ping risc-v
This list will be used to ask for help both in diagnosing and testing RISC-V-related issues as well as suggestions on how to resolve interesting questions regarding our RISC-V support.
The group also has an associated Zulip stream (#t-compiler/risc-v
)
where people can go to pose questions and discuss RISC-V-specific
topics.
So, if you are interested in participating, please sign up for the RISC-V group! To do so, open a PR against the rust-lang/team repository. Just follow this example, but change the username to your own!
WASI notification group
Github Label: O-wasi
Ping command: @rustbot ping wasi
This list will be used to ask for help both in diagnosing and testing WASI-related issues as well as suggestions on how to resolve interesting questions regarding our WASI support.
The group also has an associated Zulip stream (#t-compiler/wasm
)
where people can go to pose questions and discuss WASI-specific
topics.
So, if you are interested in participating, please sign up for the WASI group! To do so, open a PR against the rust-lang/team repository. Just follow this example, but change the username to your own!
WebAssembly (WASM) notification group
Github Label: O-wasm
Ping command: @rustbot ping wasm
This list will be used to ask for help both in diagnosing and testing WebAssembly-related issues as well as suggestions on how to resolve interesting questions regarding our WASM support.
The group also has an associated Zulip stream (#t-compiler/wasm
)
where people can go to pose questions and discuss WASM-specific
topics.
So, if you are interested in participating, please sign up for the WASM group! To do so, open a PR against the rust-lang/team repository. Just follow this example, but change the username to your own!
Windows notification group
Github Label: O-Windows
Ping command: @rustbot ping windows
This list will be used to ask for help both in diagnosing and testing Windows-related issues as well as suggestions on how to resolve interesting questions regarding our Windows support.
The group also has an associated Zulip stream (#t-compiler/windows
)
where people can go to pose questions and discuss Windows-specific
topics.
To get a better idea for what the group will do, here are some examples of the kinds of questions where we would have reached out to the group for advice in determining the best course of action:
- Which versions of MinGW should we support?
- Should we remove the legacy InnoSetup GUI installer? #72569
- What names should we use for static libraries on Windows? #29520
So, if you are interested in participating, please sign up for the Windows group! To do so, open a PR against the rust-lang/team repository. Just follow this example, but change the username to your own!
Rust for Linux notification group
Github Label: O-rfl
Ping command: @rustbot ping rfl
This list will be used to notify Rust for Linux (RfL) maintainers when the compiler or the standard library changes in a way that would break Rust for Linux, since it depends on several unstable flags and features. The RfL maintainers should then ideally provide support for resolving the breakage or decide to temporarily accept the breakage and unblock CI by temporarily removing the RfL CI jobs.
The group also has an associated Zulip stream (#rust-for-linux
)
where people can go to ask questions and discuss topics related to Rust
for Linux.
If you are interested in participating, please sign up for the Rust for Linux group on Zulip!
rust-lang/rust
Licenses
The rustc
compiler source and standard library are dual licensed under the Apache License v2.0 and the MIT License unless otherwise specified.
Detailed licensing information is available in the COPYRIGHT document of the rust-lang/rust
repository.
Guidelines for reviewers
In general, reviewers need to be looking not only for the code quality of contributions but also that they are properly licensed. We have some tips below for things to look out for when reviewing, but if you ever feel uncertain as to whether some code might be properly licensed, err on the safe side — reach out to the Council or Compiler Team Leads for feedback!
Things to watch out for:
- The PR author states that they copied, ported, or adapted the code from some other source.
- There is a comment in the code pointing to a webpage or describing where the algorithm was taken from.
- The algorithm or code pattern seems like it was likely copied from somewhere else.
- When adding new dependencies, double check the dependency's license.
In all of these cases, we will want to check that source to make sure it is licensed in a way that is compatible with Rust’s license.
Examples
- Porting C code from a GPL project, like GNU binutils, is not allowed. That would require Rust itself to be licensed under the GPL.
- Copying code from an algorithms text book may be allowed, but some algorithms are patented.
Porting
Contributions to rustc, especially around platform and compiler intrinsics, often include porting over work from other projects, mainly LLVM and GCC.
Some general rules apply:
- Copying work needs to adhere to the original license
- This applies to direct copy & paste
- This also applies to code you looked at and ported
In general, taking inspiration from other codebases is fine, but please exercise caution when porting code.
Ports of full libraries (e.g. C libraries shipped with LLVM) must keep the license of the original library.
Editions
This chapter gives an overview of how Edition support works in rustc. This assumes that you are familiar with what Editions are (see the Edition Guide).
Edition definition
The --edition
CLI flag specifies the edition to use for a crate.
This can be accessed from Session::edition
.
There are convenience functions like Session::at_least_rust_2021
for checking the crate's
edition, though you should be careful about whether you check the global session or the span, see
Edition hygiene below.
As an alternative to the at_least_rust_20xx
convenience methods, the Edition
type also
supports comparisons for doing range checks, such as span.edition() >= Edition::Edition2021
.
Adding a new edition
Adding a new edition mainly involves adding a variant to the Edition
enum and then fixing
everything that is broken. See #94461 for an
example.
Features and Edition stability
The Edition
enum defines whether or not an edition is stable.
If it is not stable, then the -Zunstable-options
CLI option must be passed to enable it.
When adding a new feature, there are two options you can choose for how to handle stability with a future edition:
- Just check the edition of the span like
span.at_least_rust_20xx()
(see Edition hygiene) or theSession::edition
. This will implicitly depend on the stability of the edition itself to indicate that your feature is available. - Place your new behavior behind a feature gate.
It may be sufficient to only check the current edition for relatively simple changes. However, for larger language changes, you should consider creating a feature gate. There are several benefits to using a feature gate:
- A feature gate makes it easier to work on and experiment with a new feature.
- It makes the intent clear when the
#![feature(…)]
attribute is used that your new feature is being enabled. - It makes testing of editions easier so that features that are not yet complete do not interfere with testing of edition-specific features that are complete and ready.
- It decouples the feature from an edition, which makes it easier for the team to make a deliberate decision of whether or not a feature should be added to the next edition when the feature is ready.
When a feature is complete and ready, the feature gate can be removed (and the code should just
check the span or Session
edition to determine if it is enabled).
There are a few different options for doing feature checks:
-
For highly experimental features, that may or may not be involved in an edition, they can implement regular feature gates like
tcx.features().my_feature
, and ignore editions for the time being. -
For experimental features that might be involved in an edition, they should implement gates with
tcx.features().my_feature && span.at_least_rust_20xx()
. This requires the user to still specify#![feature(my_feature)]
, to avoid disrupting testing of other edition features which are ready and have been accepted within the edition. -
For experimental features that have graduated to definitely be part of an edition, they should implement gates with
tcx.features().my_feature || span.at_least_rust_20xx()
, or just remove the feature check altogether and just checkspan.at_least_rust_20xx()
.
If you need to do the feature gating in multiple places, consider placing the check in a single function so that there will only be a single place to update. For example:
// An example from Edition 2021 disjoint closure captures.
fn enable_precise_capture(tcx: TyCtxt<'_>, span: Span) -> bool {
tcx.features().capture_disjoint_fields || span.rust_2021()
}
See Lints and stability below for more information about how lints handle stability.
Edition parsing
For the most part, the lexer is edition-agnostic.
Within StringReader
, tokens can be modified based on edition-specific behavior.
For example, C-String literals like c"foo"
are split into multiple tokens in editions before 2021.
This is also where things like reserved prefixes are handled for the 2021 edition.
Edition-specific parsing is relatively rare. One example is async fn
which checks the span of the
token to determine if it is the 2015 edition, and emits an error in that case.
This can only be done if the syntax was already invalid.
If you need to do edition checking in the parser, you will normally want to look at the edition of
the token, see Edition hygiene.
In some rare cases you may instead need to check the global edition from ParseSess::edition
.
Most edition-specific parsing behavior is handled with migration lints instead of in the parser.
This is appropriate when there is a change in syntax (as opposed to new syntax).
This allows the old syntax to continue to work on previous editions.
The lint then checks for the change in behavior.
On older editions, the lint pass should emit the migration lint to help with migrating to new
editions.
On newer editions, your code should emit a hard error with emit_err
instead.
For example, the deprecated start...end
pattern syntax emits the
ellipsis_inclusive_range_patterns
lint on editions before 2021, and in 2021 is an hard error via
the emit_err
method.
Keywords
New keywords can be introduced across an edition boundary.
This is implemented by functions like Symbol::is_used_keyword_conditional
, which rely on the
ordering of how the keywords are defined.
When new keywords are introduced, the keyword_idents
lint should be updated so that automatic
migrations can transition code that might be using the keyword as an identifier (see
KeywordIdents
).
An alternative to consider is to implement the keyword as a weak keyword if the position it is used
is sufficient to distinguish it.
An additional option to consider is the k#
prefix which was introduced in RFC 3101.
This allows the use of a keyword in editions before the edition where the keyword is introduced.
This is currently not implemented.
Edition hygiene
Spans are marked with the edition of the crate that the span came from. See Macro hygiene in the Edition Guide for a user-centric description of what this means.
You should normally use the edition from the token span instead of looking at the global Session
edition.
For example, use span.edition().at_least_rust_2021()
instead of sess.at_least_rust_2021()
.
This helps ensure that macros behave correctly when used across crates.
Lints
Lints support a few different options for interacting with editions. Lints can be future incompatible edition migration lints, which are used to support migrations to newer editions. Alternatively, lints can be edition-specific, where they change their default level starting in a specific edition.
Migration lints
Migration lints are used to migrate projects from one edition to the next.
They are implemented with a MachineApplicable
suggestion which
will rewrite code so that it will successfully compile in both the previous and the next
edition.
For example, the keyword_idents
lint will take identifiers that conflict with a new keyword to
use the raw identifier syntax to avoid the conflict (for example changing async
to r#async
).
Migration lints must be declared with the FutureIncompatibilityReason::EditionError
or
FutureIncompatibilityReason::EditionSemanticsChange
future-incompatible
option in the lint declaration:
declare_lint! {
pub KEYWORD_IDENTS,
Allow,
"detects edition keywords being used as an identifier",
@future_incompatible = FutureIncompatibleInfo {
reason: FutureIncompatibilityReason::EditionError(Edition::Edition2018),
reference: "issue #49716 <https://github.com/rust-lang/rust/issues/49716>",
};
}
When declared like this, the lint is automatically added to the appropriate
rust-20xx-compatibility
lint group.
When a user runs cargo fix --edition
, cargo will pass the --force-warn rust-20xx-compatibility
flag to force all of these lints to appear during the edition migration.
Cargo also passes --cap-lints=allow
so that no other lints interfere with the edition migration.
Migration lints can be either Allow
or Warn
by default.
If it is Allow
, users usually won't see this warning unless they are doing an edition migration
manually or there is a problem during the migration.
Most migration lints are Allow
.
If it is Warn
by default, users on all editions will see this warning.
Only use Warn
if you think it is important for everyone to be aware of the change, and to
encourage people to update their code on all editions.
Beware that new warn-by-default lint that hit many projects can be very disruptive and frustrating
for users.
You may consider switching an Allow
to Warn
several years after the edition stabilizes.
This will only show up for the relatively small number of stragglers who have not updated to the new
edition.
Edition-specific lints
Lints can be marked so that they have a different level starting in a specific edition.
In the lint declaration, use the @edition
marker:
declare_lint! {
pub SOME_LINT_NAME,
Allow,
"my lint description",
@edition Edition2024 => Warn;
}
Here, SOME_LINT_NAME
defaults to Allow
on all editions before 2024, and then becomes Warn
afterwards.
This should generally be used sparingly, as there are other options:
-
Small impact stylistic changes unrelated to an edition can just make the lint
Warn
on all editions. If you want people to adopt a different way to write things, then go ahead and commit to having it show up for all projects.Beware that if a new warn-by-default lint hits many projects, it can be very disruptive and frustrating for users.
-
Change the new style to be a hard error in the new edition, and use a migration lint to automatically convert projects to the new style. For example,
ellipsis_inclusive_range_patterns
is a hard error in 2021, and warns in all previous editions.Beware that these cannot be added after the edition stabilizes.
-
Migration lints can also change over time. For example, the migration lint can start out as
Allow
by default. For people performing the migration, they will automatically get updated to the new code. Then, after some years, the lint can be made toWarn
in previous editions.For example
anonymous_parameters
was a 2018 Edition migration lint (and a hard-error in 2018) that wasAllow
by default in previous editions. Then, three years later, it was changed toWarn
for all previous editions, so that all users got a warning that the style was being phased out. If this was a warning from the start, it would have impacted many projects and be very disruptive. By making it part of the edition, most users eventually updated to the new edition and were handled by the migration. Switching toWarn
only impacted a few stragglers who did not update.
Lints and stability
Lints can be marked as being unstable, which can be helpful when developing a new edition feature, and you want to test out a migration lint. The feature gate can be specified in the lint's declaration like this:
declare_lint! {
pub SOME_LINT_NAME,
Allow,
"my cool lint",
@feature_gate = sym::my_feature_name;
}
Then, the lint will only fire if the user has the appropriate #![feature(my_feature_name)]
.
Just beware that when it comes time to do crater runs testing the migration that the feature gate
will need to be removed.
Alternatively, you can implement an allow-by-default migration lint for an upcoming unstable edition without a feature gate. Although users may technically be able to enable the lint before the edition is stabilized, most will not notice the new lint exists, and it should not disrupt anything or cause any breakage.
Idiom lints
In the 2018 edition, there was a concept of "idiom lints" under the rust-2018-idioms
lint group.
The concept was to have new idiomatic styles under a different lint group separate from the forced
migrations under the rust-2018-compatibility
lint group, giving some flexibility as to how people
opt-in to certain edition changes.
Overall this approach did not seem to work very well, and it is unlikely that we will use the idiom groups in the future.
Standard library changes
Preludes
Each edition comes with a specific prelude of the standard library.
These are implemented as regular modules in core::prelude
and std::prelude
.
New items can be added to the prelude, just beware that this can conflict with user's pre-existing
code.
Usually a migration lint should be used to migrate existing code to avoid the conflict.
For example, rust_2021_prelude_collisions
is used to handle the collisions with the new traits
in 2021.
Customized language behavior
Usually it is not possible to make breaking changes to the standard library. In some rare cases, the teams may decide that the behavior change is important enough to break this rule. The downside is that this requires special handling in the compiler to be able to distinguish when the old and new signatures or behaviors should be used.
One example is the change in method resolution for into_iter()
of arrays.
This was implemented with the #[rustc_skip_array_during_method_dispatch]
attribute on the
IntoIterator
trait which then tells the compiler to consider an alternate trait resolution choice
based on the edition.
Another example is the panic!
macro changes.
This required defining multiple panic macros, and having the built-in panic macro implementation
determine the appropriate way to expand it.
This also included the non_fmt_panics
migration lint to adjust old code to the new form, which
required the rustc_diagnostic_item
attribute to detect the usage of the panic macro.
In general it is recommended to avoid these special cases except for very high value situations.
Bootstrapping the compiler
Bootstrapping is the process of using a compiler to compile itself. More accurately, it means using an older compiler to compile a newer version of the same compiler.
This raises a chicken-and-egg paradox: where did the first compiler come from? It must have been written in a different language. In Rust's case it was written in OCaml. However it was abandoned long ago and the only way to build a modern version of rustc is a slightly less modern version.
This is exactly how x.py
works: it downloads the current beta release of
rustc, then uses it to compile the new compiler.
In this section, we give a high-level overview of what Bootstrap does, followed by a high-level introduction to how Bootstrap does it.
What Bootstrapping does
- Stages of bootstrapping
- Complications of bootstrapping
- Understanding stages of bootstrap
- Passing flags to commands invoked by
bootstrap
- Environment Variables
- Clarification of build command's
stdout
Bootstrapping is the process of using a compiler to compile itself. More accurately, it means using an older compiler to compile a newer version of the same compiler.
This raises a chicken-and-egg paradox: where did the first compiler come from?
It must have been written in a different language. In Rust's case it was
written in OCaml. However it was abandoned long ago and the
only way to build a modern version of rustc
is a slightly less modern version.
This is exactly how ./x.py
works: it downloads the current beta release of
rustc
, then uses it to compile the new compiler.
Note that this documentation mostly covers user-facing information. See bootstrap/README.md to read about bootstrap internals.
Stages of bootstrapping
Overview
- Stage 0: the pre-compiled compiler
- Stage 1: from current code, by an earlier compiler
- Stage 2: the truly current compiler
- Stage 3: the same-result test
Compiling rustc
is done in stages. Here's a diagram, adapted from Jynn
Nelson's talk on bootstrapping at RustConf 2022, with
detailed explanations below.
The A
, B
, C
, and D
show the ordering of the stages of bootstrapping.
Blue nodes are
downloaded, yellow
nodes are built with the stage0
compiler, and green nodes are built with the stage1
compiler.
graph TD s0c["stage0 compiler (1.63)"]:::downloaded -->|A| s0l("stage0 std (1.64)"):::with-s0c; s0c & s0l --- stepb[ ]:::empty; stepb -->|B| s0ca["stage0 compiler artifacts (1.64)"]:::with-s0c; s0ca -->|copy| s1c["stage1 compiler (1.64)"]:::with-s0c; s1c -->|C| s1l("stage1 std (1.64)"):::with-s1c; s1c & s1l --- stepd[ ]:::empty; stepd -->|D| s1ca["stage1 compiler artifacts (1.64)"]:::with-s1c; s1ca -->|copy| s2c["stage2 compiler"]:::with-s1c; classDef empty width:0px,height:0px; classDef downloaded fill: lightblue; classDef with-s0c fill: yellow; classDef with-s1c fill: lightgreen;
Stage 0: the pre-compiled compiler
The stage0 compiler is usually the current beta rustc
compiler and its
associated dynamic libraries, which ./x.py
will download for you. (You can
also configure ./x.py
to use something else.)
The stage0 compiler is then used only to compile src/bootstrap
,
library/std
, and compiler/rustc
. When assembling the libraries and
binaries that will become the stage1 rustc
compiler, the freshly compiled
std
and rustc
are used. There are two concepts at play here: a compiler
(with its set of dependencies) and its 'target' or 'object' libraries (std
and
rustc
). Both are staged, but in a staggered manner.
Stage 1: from current code, by an earlier compiler
The rustc source code is then compiled with the stage0
compiler to produce the
stage1
compiler.
Stage 2: the truly current compiler
We then rebuild our stage1
compiler with itself to produce the stage2
compiler.
In theory, the stage1
compiler is functionally identical to the stage2
compiler, but in practice there are subtle differences. In particular, the
stage1
compiler itself was built by stage0
and hence not by the source in
your working directory. This means that the ABI generated by the stage0
compiler may not match the ABI that would have been made by the stage1
compiler, which can cause problems for dynamic libraries, tests, and tools using
rustc_private
.
Note that the proc_macro
crate avoids this issue with a C
FFI layer called
proc_macro::bridge
, allowing it to be used with stage1
.
The stage2
compiler is the one distributed with rustup
and all other install
methods. However, it takes a very long time to build because one must first
build the new compiler with an older compiler and then use that to build the new
compiler with itself. For development, you usually only want the stage1
compiler, which you can build with ./x build library
. See Building the
compiler.
Stage 3: the same-result test
Stage 3 is optional. To sanity check our new compiler we can build the libraries
with the stage2
compiler. The result ought to be identical to before, unless
something has broken.
Building the stages
The script ./x
tries to be helpful and pick the stage you most likely meant
for each subcommand. These defaults are as follows:
check
:--stage 0
doc
:--stage 0
build
:--stage 1
test
:--stage 1
dist
:--stage 2
install
:--stage 2
bench
:--stage 2
You can always override the stage by passing --stage N
explicitly.
For more information about stages, see below.
Complications of bootstrapping
Since the build system uses the current beta compiler to build a stage1
bootstrapping compiler, the compiler source code can't use some features until
they reach beta (because otherwise the beta compiler doesn't support them). On
the other hand, for compiler intrinsics and internal features, the
features have to be used. Additionally, the compiler makes heavy use of
nightly
features (#![feature(...)]
). How can we resolve this problem?
There are two methods used:
- The build system sets
--cfg bootstrap
when building withstage0
, so we can usecfg(not(bootstrap))
to only use features when built withstage1
. Setting--cfg bootstrap
in this way is used for features that were just stabilized, which require#![feature(...)]
when built withstage0
, but not forstage1
. - The build system sets
RUSTC_BOOTSTRAP=1
. This special variable means to break the stability guarantees of Rust: allowing use of#![feature(...)]
with a compiler that's notnightly
. SettingRUSTC_BOOTSTRAP=1
should never be used except when bootstrapping the compiler.
Understanding stages of bootstrap
Overview
This is a detailed look into the separate bootstrap stages.
The convention ./x
uses is that:
- A
--stage N
flag means to run the stage N compiler (stageN/rustc
). - A "stage N artifact" is a build artifact that is produced by the stage N compiler.
- The stage N+1 compiler is assembled from stage N artifacts. This process is called uplifting.
Build artifacts
Anything you can build with ./x
is a build artifact. Build artifacts
include, but are not limited to:
- binaries, like
stage0-rustc/rustc-main
- shared objects, like
stage0-sysroot/rustlib/libstd-6fae108520cf72fe.so
- rlib files, like
stage0-sysroot/rustlib/libstd-6fae108520cf72fe.rlib
- HTML files generated by rustdoc, like
doc/std
Examples
./x test tests/ui
means to build thestage1
compiler and runcompiletest
on it. If you're working on the compiler, this is normally the test command you want../x test --stage 0 library/std
means to run tests on the standard library without buildingrustc
from source ('build withstage0
, then test the artifacts'). If you're working on the standard library, this is normally the test command you want../x build --stage 0
means to build with the betarustc
../x doc --stage 0
means to document using the betarustdoc
.
Examples of what not to do
./x test --stage 0 tests/ui
is not useful: it runs tests on the beta compiler and doesn't buildrustc
from source. Usetest tests/ui
instead, which buildsstage1
from source../x test --stage 0 compiler/rustc
builds the compiler but runs no tests: it's runningcargo test -p rustc
, butcargo
doesn't understand Rust's tests. You shouldn't need to use this, usetest
instead (without arguments)../x build --stage 0 compiler/rustc
builds the compiler, but does not buildlibstd
or evenlibcore
. Most of the time, you'll want./x build library
instead, which allows compiling programs without needing to define lang items.
Building vs. running
Note that build --stage N compiler/rustc
does not build the stage N
compiler: instead it builds the stage N+1 compiler using the stage N compiler.
In short, stage 0 uses the stage0
compiler to create stage0
artifacts which
will later be uplifted to be the stage1 compiler.
In each stage, two major steps are performed:
std
is compiled by the stage N compiler.- That
std
is linked to programs built by the stage N compiler, including the stage N artifacts (stage N+1 compiler).
This is somewhat intuitive if one thinks of the stage N artifacts as "just"
another program we are building with the stage N compiler: build --stage N compiler/rustc
is linking the stage N artifacts to the std
built by the stage
N compiler.
Stages and std
Note that there are two std
libraries in play here:
- The library linked to
stageN/rustc
, which was built by stage N-1 (stage N-1std
) - The library used to compile programs with
stageN/rustc
, which was built by stage N (stage Nstd
).
Stage N std
is pretty much necessary for any useful work with the stage N
compiler. Without it, you can only compile programs with #![no_core]
-- not
terribly useful!
The reason these need to be different is because they aren't necessarily
ABI-compatible: there could be new layout optimizations, changes to MIR
, or
other changes to Rust metadata on nightly
that aren't present in beta.
This is also where --keep-stage 1 library/std
comes into play. Since most
changes to the compiler don't actually change the ABI, once you've produced a
std
in stage1
, you can probably just reuse it with a different compiler. If
the ABI hasn't changed, you're good to go, no need to spend time recompiling
that std
. The flag --keep-stage
simply instructs the build script to assumes
the previous compile is fine and copies those artifacts into the appropriate
place, skipping the cargo
invocation.
Cross-compiling rustc
Cross-compiling is the process of compiling code that will run on another
architecture. For instance, you might want to build an ARM version of rustc
using an x86 machine. Building stage2
std
is different when you are
cross-compiling.
This is because ./x
uses the following logic: if HOST
and TARGET
are the
same, it will reuse stage1
std
for stage2
! This is sound because stage1
std
was compiled with the stage1
compiler, i.e. a compiler using the source
code you currently have checked out. So it should be identical (and therefore
ABI-compatible) to the std
that stage2/rustc
would compile.
However, when cross-compiling, stage1
std
will only run on the host. So the
stage2
compiler has to recompile std
for the target.
(See in the table how stage2
only builds non-host std
targets).
Why does only libstd use cfg(bootstrap)
?
For docs on cfg(bootstrap)
itself, see Complications of
Bootstrapping.
The rustc
generated by the stage0
compiler is linked to the freshly-built
std
, which means that for the most part only std
needs to be cfg
-gated, so
that rustc
can use features added to std
immediately after their addition,
without need for them to get into the downloaded beta
compiler.
Note this is different from any other Rust program: stage1
rustc
is built by
the beta compiler, but using the master version of libstd
!
The only time rustc
uses cfg(bootstrap)
is when it adds internal lints that
use diagnostic items, or when it uses unstable library features that were
recently changed.
What is a 'sysroot'?
When you build a project with cargo
, the build artifacts for dependencies are
normally stored in target/debug/deps
. This only contains dependencies cargo
knows about; in particular, it doesn't have the standard library. Where do std
or proc_macro
come from? They comes from the sysroot, the root of a number
of directories where the compiler loads build artifacts at runtime. The
sysroot
doesn't just store the standard library, though - it includes anything
that needs to be loaded at runtime. That includes (but is not limited to):
- Libraries
libstd
/libtest
/libproc_macro
. - Compiler crates themselves, when using
rustc_private
. In-tree these are always present; out of tree, you need to installrustc-dev
withrustup
. - Shared object file
libLLVM.so
for the LLVM project. In-tree this is either built from source or downloaded from CI; out-of-tree, you need to installllvm-tools-preview
withrustup
.
All the artifacts listed so far are compiler runtime dependencies. You can see
them with rustc --print sysroot
:
$ ls $(rustc --print sysroot)/lib
libchalk_derive-0685d79833dc9b2b.so libstd-25c6acf8063a3802.so
libLLVM-11-rust-1.50.0-nightly.so libtest-57470d2aa8f7aa83.so
librustc_driver-4f0cc9f50e53f0ba.so libtracing_attributes-e4be92c35ab2a33b.so
librustc_macros-5f0ec4a119c6ac86.so rustlib
There are also runtime dependencies for the standard library! These are in
lib/rustlib/
, not lib/
directly.
$ ls $(rustc --print sysroot)/lib/rustlib/x86_64-unknown-linux-gnu/lib | head -n 5
libaddr2line-6c8e02b8fedc1e5f.rlib
libadler-9ef2480568df55af.rlib
liballoc-9c4002b5f79ba0e1.rlib
libcfg_if-512eb53291f6de7e.rlib
libcompiler_builtins-ef2408da76957905.rlib
Directory lib/rustlib/
includes libraries like hashbrown
and cfg_if
, which
are not part of the public API of the standard library, but are used to
implement it. Also lib/rustlib/
is part of the search path for linkers, but
lib
will never be part of the search path.
-Z force-unstable-if-unmarked
Since lib/rustlib/
is part of the search path we have to be careful about
which crates are included in it. In particular, all crates except for the
standard library are built with the flag -Z force-unstable-if-unmarked
, which
means that you have to use #![feature(rustc_private)]
in order to load it (as
opposed to the standard library, which is always available).
The -Z force-unstable-if-unmarked
flag has a variety of purposes to help
enforce that the correct crates are marked as unstable
. It was introduced
primarily to allow rustc and the standard library to link to arbitrary crates on
crates.io which do not themselves use staged_api
. rustc
also relies on this
flag to mark all of its crates as unstable
with the rustc_private
feature so
that each crate does not need to be carefully marked with unstable
.
This flag is automatically applied to all of rustc
and the standard library by
the bootstrap scripts. This is needed because the compiler and all of its
dependencies are shipped in sysroot
to all users.
This flag has the following effects:
- Marks the crate as "
unstable
" with therustc_private
feature if it is not itself marked asstable
orunstable
. - Allows these crates to access other forced-unstable crates without any need
for attributes. Normally a crate would need a
#![feature(rustc_private)]
attribute to use otherunstable
crates. However, that would make it impossible for a crate from crates.io to access its own dependencies since that crate won't have afeature(rustc_private)
attribute, but everything is compiled with-Z force-unstable-if-unmarked
.
Code which does not use -Z force-unstable-if-unmarked
should include the
#![feature(rustc_private)]
crate attribute to access these forced-unstable
crates. This is needed for things which link rustc
its self, such as MIRI
or
clippy
.
You can find more discussion about sysroots in:
- The rustdoc PR explaining why it uses
extern crate
for dependencies loaded fromsysroot
- Discussions about sysroot on Zulip
- Discussions about building rustdoc out of tree
Passing flags to commands invoked by bootstrap
Conveniently ./x
allows you to pass stage-specific flags to rustc
and
cargo
when bootstrapping. The RUSTFLAGS_BOOTSTRAP
environment variable is
passed as RUSTFLAGS
to the bootstrap stage (stage0
), and
RUSTFLAGS_NOT_BOOTSTRAP
is passed when building artifacts for later stages.
RUSTFLAGS
will work, but also affects the build of bootstrap
itself, so it
will be rare to want to use it. Finally, MAGIC_EXTRA_RUSTFLAGS
bypasses the
cargo
cache to pass flags to rustc without recompiling all dependencies.
RUSTDOCFLAGS
,RUSTDOCFLAGS_BOOTSTRAP
andRUSTDOCFLAGS_NOT_BOOTSTRAP
are analogous toRUSTFLAGS
, but forrustdoc
.CARGOFLAGS
will pass arguments to cargo itself (e.g.--timings
).CARGOFLAGS_BOOTSTRAP
andCARGOFLAGS_NOT_BOOTSTRAP
work analogously toRUSTFLAGS_BOOTSTRAP
.--test-args
will pass arguments through to the test runner. Fortests/ui
, this iscompiletest
. For unit tests and doc tests this is thelibtest
runner.
Most test runner accept --help
, which you can use to find out the options
accepted by the runner.
Environment Variables
During bootstrapping, there are a bunch of compiler-internal environment
variables that are used. If you are trying to run an intermediate version of
rustc
, sometimes you may need to set some of these environment variables
manually. Otherwise, you get an error like the following:
thread 'main' panicked at 'RUSTC_STAGE was not set: NotPresent', library/core/src/result.rs:1165:5
If ./stageN/bin/rustc
gives an error about environment variables, that usually
means something is quite wrong -- such as you're trying to compile rustc
or
std
or something which depends on environment variables. In the unlikely case
that you actually need to invoke rustc
in such a situation, you can tell the
bootstrap shim to print all env
variables by adding -vvv
to your x
command.
Finally, bootstrap makes use of the cc-rs crate which has its own
method of configuring C
compilers and C
flags via environment
variables.
Clarification of build command's stdout
In this part, we will investigate the build command's stdout
in an action
(similar, but more detailed and complete documentation compare to topic above).
When you execute x build --dry-run
command, the build output will be something
like the following:
Building stage0 library artifacts (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu)
Copying stage0 library from stage0 (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu / x86_64-unknown-linux-gnu)
Building stage0 compiler artifacts (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu)
Copying stage0 rustc from stage0 (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu / x86_64-unknown-linux-gnu)
Assembling stage1 compiler (x86_64-unknown-linux-gnu)
Building stage1 library artifacts (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu)
Copying stage1 library from stage1 (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu / x86_64-unknown-linux-gnu)
Building stage1 tool rust-analyzer-proc-macro-srv (x86_64-unknown-linux-gnu)
Building rustdoc for stage1 (x86_64-unknown-linux-gnu)
Building stage0 {std,compiler} artifacts
These steps use the provided (downloaded, usually) compiler to compile the local Rust source into libraries we can use.
Copying stage0 {std,rustc}
This copies the library and compiler artifacts from cargo
into
stage0-sysroot/lib/rustlib/{target-triple}/lib
Assembling stage1 compiler
This copies the libraries we built in "building stage0
... artifacts" into the
stage1
compiler's lib/
directory. These are the host libraries that the
compiler itself uses to run. These aren't actually used by artifacts the new
compiler generates. This step also copies the rustc
and rustdoc
binaries we
generated into build/$HOST/stage/bin
.
The stage1/bin/rustc
is a fully functional compiler, but it doesn't yet have
any libraries to link built binaries or libraries to. The next 3 steps will
provide those libraries for it; they are mostly equivalent to constructing the
stage1/bin
compiler so we don't go through them individually here.
How Bootstrap does it
The core concept in Bootstrap is a build Step
, which are chained together
by Builder::ensure
. Builder::ensure
takes a Step
as input, and runs
the Step
if and only if it has not already been run. Let's take a closer
look at Step
.
Synopsis of Step
A Step
represents a granular collection of actions involved in the process
of producing some artifact. It can be thought of like a rule in Makefiles.
The Step
trait is defined as:
pub trait Step: 'static + Clone + Debug + PartialEq + Eq + Hash {
type Output: Clone;
const DEFAULT: bool = false;
const ONLY_HOSTS: bool = false;
// Required methods
fn run(self, builder: &Builder<'_>) -> Self::Output;
fn should_run(run: ShouldRun<'_>) -> ShouldRun<'_>;
// Provided method
fn make_run(_run: RunConfig<'_>) { ... }
}
run
is the function that is responsible for doing the work.Builder::ensure
invokesrun
.should_run
is the command-line interface, which determines if an invocation such asx build foo
should run a givenStep
. In a "default" context where no paths are provided, thenmake_run
is called directly.make_run
is invoked only for things directly asked via the CLI and not for steps which are dependencies of other steps.
The entry points
There's a couple of preliminary steps before core Bootstrap code is reached:
- Shell script or
make
:./x
or./x.ps1
ormake
- Convenience wrapper script:
x.py
src/bootstrap/bootstrap.py
src/bootstrap/src/bin/main.rs
See src/bootstrap/README.md for a more specific description of the implementation details.
High-Level Compiler Architecture
The remaining parts of this guide discuss how the compiler works. They go through everything from high-level structure of the compiler to how each stage of compilation works. They should be friendly to both readers interested in the end-to-end process of compilation and readers interested in learning about a specific system they wish to contribute to. If anything is unclear, feel free to file an issue on the rustc-dev-guide repo or contact the compiler team, as detailed in this chapter from Part 1.
In this part, we will look at the high-level architecture of the compiler. In particular, we will look at three overarching design choices that impact the whole compiler: the query system, incremental compilation, and interning.
Overview of the compiler
This chapter is about the overall process of compiling a program -- how everything fits together.
The Rust compiler is special in two ways: it does things to your code that other compilers don't do (e.g. borrow-checking) and it has a lot of unconventional implementation choices (e.g. queries). We will talk about these in turn in this chapter, and in the rest of the guide, we will look at the individual pieces in more detail.
What the compiler does to your code
So first, let's look at what the compiler does to your code. For now, we will avoid mentioning how the compiler implements these steps except as needed.
Invocation
Compilation begins when a user writes a Rust source program in text and invokes
the rustc
compiler on it. The work that the compiler needs to perform is
defined by command-line options. For example, it is possible to enable nightly
features (-Z
flags), perform check
-only builds, or emit the LLVM
Intermediate Representation (LLVM-IR
) rather than executable machine code.
The rustc
executable call may be indirect through the use of cargo
.
Command line argument parsing occurs in the rustc_driver
. This crate
defines the compile configuration that is requested by the user and passes it
to the rest of the compilation process as a rustc_interface::Config
.
Lexing and parsing
The raw Rust source text is analyzed by a low-level lexer located in
rustc_lexer
. At this stage, the source text is turned into a stream of
atomic source code units known as tokens. The lexer
supports the
Unicode character encoding.
The token stream passes through a higher-level lexer located in
rustc_parse
to prepare for the next stage of the compile process. The
StringReader
struct
is used at this stage to perform a set of validations
and turn strings into interned symbols (interning is discussed later).
String interning is a way of storing only one immutable
copy of each distinct string value.
The lexer has a small interface and doesn't depend directly on the diagnostic
infrastructure in rustc
. Instead it provides diagnostics as plain data which
are emitted in rustc_parse::lexer
as real diagnostics. The lexer
preserves full fidelity information for both IDEs and procedural macros
(sometimes referred to as "proc-macros").
The parser translates the token stream from the lexer
into an Abstract Syntax
Tree (AST). It uses a recursive descent (top-down) approach to syntax
analysis. The crate entry points for the parser
are the
Parser::parse_crate_mod()
and Parser::parse_mod()
methods found in rustc_parse::parser::Parser
. The external module parsing
entry point is rustc_expand::module::parse_external_mod
.
And the macro-parser
entry point is Parser::parse_nonterminal()
.
Parsing is performed with a set of parser
utility methods including bump
,
check
, eat
, expect
, look_ahead
.
Parsing is organized by semantic construct. Separate
parse_*
methods can be found in the rustc_parse
directory. The source file name follows the construct name. For example, the
following files are found in the parser
:
This naming scheme is used across many compiler stages. You will find either a
file or directory with the same name across the parsing, lowering, type
checking, Typed High-level Intermediate Representation (THIR
) lowering, and
Mid-level Intermediate Representation (MIR
) building sources.
Macro-expansion, AST
-validation, name-resolution, and early linting also take
place during the lexing and parsing stage.
The rustc_ast::ast
::{Crate
, Expr
, Pat
, ...} AST
nodes are
returned from the parser while the standard Diag
API is used
for error handling. Generally Rust's compiler will try to recover from errors
by parsing a superset of Rust's grammar, while also emitting an error type.
AST
lowering
Next the AST
is converted into High-Level Intermediate Representation
(HIR
), a more compiler-friendly representation of the AST
. This process
is called "lowering" and involves a lot of desugaring (the expansion and
formalizing of shortened or abbreviated syntax constructs) of things like loops
and async fn
.
We then use the HIR
to do type inference (the process of automatic
detection of the type of an expression), trait solving (the process of
pairing up an impl with each reference to a trait
), and type checking. Type
checking is the process of converting the types found in the HIR
(hir::Ty
),
which represent what the user wrote, into the internal representation used by
the compiler (Ty<'tcx>
). It's called type checking because the information
is used to verify the type safety, correctness and coherence of the types used
in the program.
MIR
lowering
The HIR
is further lowered to MIR
(used for borrow checking) by constructing the THIR
(an even more desugared HIR
used for
pattern and exhaustiveness checking) to convert into MIR
.
We do many optimizations on the MIR because it is generic and that
improves later code generation and compilation speed. It is easier to do some
optimizations at MIR
level than at LLVM-IR
level. For example LLVM doesn't seem
to be able to optimize the pattern the simplify_try
MIR
-opt looks for.
Rust code is also monomorphized during code generation, which means making
copies of all the generic code with the type parameters replaced by concrete
types. To do this, we need to collect a list of what concrete types to generate
code for. This is called monomorphization collection and it happens at the
MIR
level.
Code generation
We then begin what is simply called code generation or codegen. The code
generation stage is when higher-level representations of source are
turned into an executable binary. Since rustc
uses LLVM for code generation,
the first step is to convert the MIR
to LLVM-IR
. This is where the MIR
is
actually monomorphized. The LLVM-IR
is passed to LLVM, which does a lot more
optimizations on it, emitting machine code which is basically assembly code
with additional low-level types and annotations added (e.g. an ELF object or
WASM
). The different libraries/binaries are then linked together to produce
the final binary.
How it does it
Now that we have a high-level view of what the compiler does to your code, let's take a high-level view of how it does all that stuff. There are a lot of constraints and conflicting goals that the compiler needs to satisfy/optimize for. For example,
- Compilation speed: how fast is it to compile a program? More/better
compile-time analyses often means compilation is slower.
- Also, we want to support incremental compilation, so we need to take that
into account. How can we keep track of what work needs to be redone and
what can be reused if the user modifies their program?
- Also we can't store too much stuff in the incremental cache because it would take a long time to load from disk and it could take a lot of space on the user's system...
- Also, we want to support incremental compilation, so we need to take that
into account. How can we keep track of what work needs to be redone and
what can be reused if the user modifies their program?
- Compiler memory usage: while compiling a program, we don't want to use more memory than we need.
- Program speed: how fast is your compiled program? More/better compile-time analyses often means the compiler can do better optimizations.
- Program size: how large is the compiled binary? Similar to the previous point.
- Compiler compilation speed: how long does it take to compile the compiler? This impacts contributors and compiler maintenance.
- Implementation complexity: building a compiler is one of the hardest things a person/group can do, and Rust is not a very simple language, so how do we make the compiler's code base manageable?
- Compiler correctness: the binaries produced by the compiler should do what the input programs says they do, and should continue to do so despite the tremendous amount of change constantly going on.
- Integration: a number of other tools need to use the compiler in
various ways (e.g.
cargo
,clippy
,MIRI
) that must be supported. - Compiler stability: the compiler should not crash or fail ungracefully on the stable channel.
- Rust stability: the compiler must respect Rust's stability guarantees by not breaking programs that previously compiled despite the many changes that are always going on to its implementation.
- Limitations of other tools:
rustc
uses LLVM in its backend, and LLVM has some strengths we leverage and some aspects we need to work around.
So, as you continue your journey through the rest of the guide, keep these things in mind. They will often inform decisions that we make.
Intermediate representations
As with most compilers, rustc
uses some intermediate representations (IRs) to
facilitate computations. In general, working directly with the source code is
extremely inconvenient and error-prone. Source code is designed to be human-friendly while at
the same time being unambiguous, but it's less convenient for doing something
like, say, type checking.
Instead most compilers, including rustc
, build some sort of IR out of the
source code which is easier to analyze. rustc
has a few IRs, each optimized
for different purposes:
- Token stream: the lexer produces a stream of tokens directly from the source code. This stream of tokens is easier for the parser to deal with than raw text.
- Abstract Syntax Tree (
AST
): the abstract syntax tree is built from the stream of tokens produced by the lexer. It represents pretty much exactly what the user wrote. It helps to do some syntactic sanity checking (e.g. checking that a type is expected where the user wrote one). - High-level IR (HIR): This is a sort of desugared
AST
. It's still close to what the user wrote syntactically, but it includes some implicit things such as some elided lifetimes, etc. This IR is amenable to type checking. - Typed
HIR
(THIR) formerly High-level Abstract IR (HAIR): This is an intermediate betweenHIR
and MIR. It is like theHIR
but it is fully typed and a bit more desugared (e.g. method calls and implicit dereferences are made fully explicit). As a result, it is easier to lower toMIR
fromTHIR
than from HIR. - Middle-level IR (
MIR
): This IR is basically a Control-Flow Graph (CFG). A CFG is a type of diagram that shows the basic blocks of a program and how control flow can go between them. Likewise,MIR
also has a bunch of basic blocks with simple typed statements inside them (e.g. assignment, simple computations, etc) and control flow edges to other basic blocks (e.g., calls, dropping values).MIR
is used for borrow checking and other important dataflow-based checks, such as checking for uninitialized values. It is also used for a series of optimizations and for constant evaluation (viaMIRI
). BecauseMIR
is still generic, we can do a lot of analyses here more efficiently than after monomorphization. LLVM-IR
: This is the standard form of all input to the LLVM compiler.LLVM-IR
is a sort of typed assembly language with lots of annotations. It's a standard format that is used by all compilers that use LLVM (e.g. the clang C compiler also outputsLLVM-IR
).LLVM-IR
is designed to be easy for other compilers to emit and also rich enough for LLVM to run a bunch of optimizations on it.
One other thing to note is that many values in the compiler are interned. This is a performance and memory optimization in which we allocate the values in a special allocator called an arena. Then, we pass around references to the values allocated in the arena. This allows us to make sure that identical values (e.g. types in your program) are only allocated once and can be compared cheaply by comparing pointers. Many of the intermediate representations are interned.
Queries
The first big implementation choice is Rust's use of the query system in its compiler. The Rust compiler is not organized as a series of passes over the code which execute sequentially. The Rust compiler does this to make incremental compilation possible -- that is, if the user makes a change to their program and recompiles, we want to do as little redundant work as possible to output the new binary.
In rustc
, all the major steps above are organized as a bunch of queries that
call each other. For example, there is a query to ask for the type of something
and another to ask for the optimized MIR
of a function. These queries can call
each other and are all tracked through the query system. The results of the
queries are cached on disk so that the compiler can tell which queries' results
changed from the last compilation and only redo those. This is how incremental
compilation works.
In principle, for the query-field steps, we do each of the above for each item
individually. For example, we will take the HIR
for a function and use queries
to ask for the LLVM-IR
for that HIR. This drives the generation of optimized
MIR
, which drives the borrow checker, which drives the generation of MIR
, and
so on.
... except that this is very over-simplified. In fact, some queries are not
cached on disk, and some parts of the compiler have to run for all code anyway
for correctness even if the code is dead code (e.g. the borrow checker). For
example, currently the mir_borrowck
query is first executed on all functions
of a crate. Then the codegen backend invokes the
collect_and_partition_mono_items
query, which first recursively requests the
optimized_mir
for all reachable functions, which in turn runs mir_borrowck
for that function and then creates codegen units. This kind of split will need
to remain to ensure that unreachable functions still have their errors emitted.
Moreover, the compiler wasn't originally built to use a query system; the query
system has been retrofitted into the compiler, so parts of it are not query-field
yet. Also, LLVM isn't our code, so that isn't querified either. The plan is to
eventually query-fy all of the steps listed in the previous section,
but as of November 2022, only the steps between HIR
and
LLVM-IR
are query-field. That is, lexing, parsing, name resolution, and macro
expansion are done all at once for the whole program.
One other thing to mention here is the all-important "typing context",
TyCtxt
, which is a giant struct that is at the center of all things.
(Note that the name is mostly historic. This is not a "typing context" in the
sense of Γ
or Δ
from type theory. The name is retained because that's what
the name of the struct is in the source code.) All
queries are defined as methods on the TyCtxt
type, and the in-memory query
cache is stored there too. In the code, there is usually a variable called
tcx
which is a handle on the typing context. You will also see lifetimes with
the name 'tcx
, which means that something is tied to the lifetime of the
TyCtxt
(usually it is stored or interned there).
ty::Ty
Types are really important in Rust, and they form the core of a lot of compiler
analyses. The main type (in the compiler) that represents types (in the user's
program) is rustc_middle::ty::Ty
. This is so important that we have a whole chapter
on ty::Ty
, but for now, we just want to mention that it exists and is the way
rustc
represents types!
Also note that the rustc_middle::ty
module defines the TyCtxt
struct we mentioned before.
Parallelism
Compiler performance is a problem that we would like to improve on
(and are always working on). One aspect of that is parallelizing
rustc
itself.
Currently, there is only one part of rustc that is parallel by default: code generation.
However, the rest of the compiler is still not yet parallel. There have been
lots of efforts spent on this, but it is generally a hard problem. The current
approach is to turn RefCell
s into Mutex
s -- that is, we
switch to thread-safe internal mutability. However, there are ongoing
challenges with lock contention, maintaining query-system invariants under
concurrency, and the complexity of the code base. One can try out the current
work by enabling parallel compilation in config.toml
. It's still early days,
but there are already some promising performance improvements.
Bootstrapping
rustc
itself is written in Rust. So how do we compile the compiler? We use an
older compiler to compile the newer compiler. This is called bootstrapping.
Bootstrapping has a lot of interesting implications. For example, it means that one of the major users of Rust is the Rust compiler, so we are constantly testing our own software ("eating our own dogfood").
For more details on bootstrapping, see the bootstrapping section of the guide.
References
- Command line parsing
- Guide: The Rustc Driver and Interface
- Driver definition:
rustc_driver
- Main entry point:
rustc_session::config::build_session_options
- Lexical Analysis: Lex the user program to a stream of tokens
- Guide: Lexing and Parsing
- Lexer definition:
rustc_lexer
- Main entry point:
rustc_lexer::cursor::Cursor::advance_token
- Parsing: Parse the stream of tokens to an Abstract Syntax Tree (AST)
- Guide: Lexing and Parsing
- Guide: Macro Expansion
- Guide: Name Resolution
- Parser definition:
rustc_parse
- Main entry points:
AST
definition:rustc_ast
- Feature gating: TODO
- Early linting: TODO
- The High Level Intermediate Representation (HIR)
- Guide: The HIR
- Guide: Identifiers in the HIR
- Guide: The
HIR
Map - Guide: Lowering
AST
toHIR
- How to view
HIR
representation for your codecargo rustc -- -Z unpretty=hir-tree
- Rustc
HIR
definition:rustc_hir
- Main entry point: TODO
- Late linting: TODO
- Type Inference
- Guide: Type Inference
- Guide: The ty Module: Representing Types (semantics)
- Main entry point (type inference):
InferCtxtBuilder::enter
- Main entry point (type checking bodies): the
typeck
query- These two functions can't be decoupled.
- The Mid Level Intermediate Representation (MIR)
- Guide: The
MIR
(Mid level IR) - Definition:
rustc_middle/src/mir
- Definition of sources that manipulates the MIR:
rustc_mir_build
,rustc_mir_dataflow
,rustc_mir_transform
- Guide: The
- The Borrow Checker
- Guide: MIR Borrow Check
- Definition:
rustc_borrowck
- Main entry point:
mir_borrowck
query
MIR
Optimizations- Guide: MIR Optimizations
- Definition:
rustc_mir_transform
- Main entry point:
optimized_mir
query
- Code Generation
- Guide: Code Generation
- Generating Machine Code from
LLVM-IR
with LLVM - TODO: reference? - Main entry point:
rustc_codegen_ssa::base::codegen_crate
- This monomorphizes and produces
LLVM-IR
for one codegen unit. It then starts a background thread to run LLVM, which must be joined later. - Monomorphization happens lazily via
FunctionCx::monomorphize
andrustc_codegen_ssa::base::codegen_instance
- This monomorphizes and produces
High-level overview of the compiler source
Now that we have seen what the compiler does,
let's take a look at the structure of the rust-lang/rust
repository,
where the rustc source code lives.
You may find it helpful to read the "Overview of the compiler" chapter, which introduces how the compiler works, before this one.
Workspace structure
The rust-lang/rust
repository consists of a single large cargo workspace
containing the compiler, the standard libraries (core
, alloc
, std
,
proc_macro
, etc
), and rustdoc
, along with the build system and a
bunch of tools and submodules for building a full Rust distribution.
The repository consists of three main directories:
-
compiler/
contains the source code forrustc
. It consists of many crates that together make up the compiler. -
library/
contains the standard libraries (core
,alloc
,std
,proc_macro
,test
), as well as the Rust runtime (backtrace
,rtstartup
,lang_start
). -
tests/
contains the compiler tests. -
src/
contains the source code forrustdoc
,clippy
,cargo
, the build system, language docs, etc.
Compiler
The compiler is implemented in the various compiler/
crates.
The compiler/
crates all have names starting with rustc_*
. These are a
collection of around 50 interdependent crates ranging in size from tiny to
huge. There is also the rustc
crate which is the actual binary (i.e. the
main
function); it doesn't actually do anything besides calling the
rustc_driver
crate, which drives the various parts of compilation in other
crates.
The dependency structure of these crates is complex, but roughly it is something like this:
rustc
(the binary) callsrustc_driver::main
.rustc_driver
depends on a lot of other crates, but the main one isrustc_interface
.rustc_interface
depends on most of the other compiler crates. It is a fairly generic interface for driving the whole compilation.- Most of the other
rustc_*
crates depend onrustc_middle
, which defines a lot of central data structures in the compiler.rustc_middle
and most of the other crates depend on a handful of crates representing the early parts of the compiler (e.g. the parser), fundamental data structures (e.g.Span
), or error reporting:rustc_data_structures
,rustc_span
,rustc_errors
, etc.
- Most of the other
You can see the exact dependencies by reading the Cargo.toml
for the various
crates, just like a normal Rust crate.
One final thing: src/llvm-project
is a submodule for our fork of LLVM.
During bootstrapping, LLVM is built and the compiler/rustc_llvm
crate
contains Rust wrappers around LLVM (which is written in C++), so that the
compiler can interface with it.
Most of this book is about the compiler, so we won't have any further explanation of these crates here.
Big picture
The dependency structure of the compiler is influenced by two main factors:
- Organization. The compiler is a huge codebase; it would be an impossibly large crate. In part, the dependency structure reflects the code structure of the compiler.
- Compile-time. By breaking the compiler into multiple crates, we can take better advantage of incremental/parallel compilation using cargo. In particular, we try to have as few dependencies between crates as possible so that we don't have to rebuild as many crates if you change one.
At the very bottom of the dependency tree are a handful of crates that are used
by the whole compiler (e.g. rustc_span
). The very early parts of the
compilation process (e.g. parsing and the Abstract Syntax Tree (AST
))
depend on only these.
After the AST
is constructed and other early analysis is done, the
compiler's query system gets set up. The query system is set up in a
clever way using function pointers. This allows us to break dependencies
between crates, allowing more parallel compilation. The query system is defined
in rustc_middle
, so nearly all subsequent parts of the compiler depend on
this crate. It is a really large crate, leading to long compile times. Some
efforts have been made to move stuff out of it with varying success. Another
side-effect is that sometimes related functionality gets scattered across
different crates. For example, linting functionality is found across earlier
parts of the crate, rustc_lint
, rustc_middle
, and other places.
Ideally there would be fewer, more cohesive crates, with incremental and parallel compilation making sure compile times stay reasonable. However, incremental and parallel compilation haven't gotten good enough for that yet, so breaking things into separate crates has been our solution so far.
At the top of the dependency tree is rustc_driver
and rustc_interface
which is an unstable wrapper around the query system helping drive various
stages of compilation. Other consumers of the compiler may use this interface
in different ways (e.g. rustdoc
or maybe eventually rust-analyzer
). The
rustc_driver
crate first parses command line arguments and then uses
rustc_interface
to drive the compilation to completion.
rustdoc
The bulk of rustdoc
is in librustdoc
. However, the rustdoc
binary
itself is src/tools/rustdoc
, which does nothing except call rustdoc::main
.
There is also JavaScript
and CSS
for the docs in src/tools/rustdoc-js
and src/tools/rustdoc-themes
.
You can read more about rustdoc
in this chapter.
Tests
The test suite for all of the above is in tests/
. You can read more
about the test suite in this chapter.
The test harness is in src/tools/compiletest/
.
Build System
There are a number of tools in the repository just for building the compiler,
standard library, rustdoc
, etc, along with testing, building a full Rust
distribution, etc.
One of the primary tools is src/bootstrap/
. You can read more about
bootstrapping in this chapter. The process may also use other tools
from src/tools/
, such as tidy/
or compiletest/
.
Standard library
This code is fairly similar to most other Rust crates except that it must be
built in a special way because it can use unstable (nightly
) features.
The standard library is sometimes referred to as libstd or the "standard facade"
.
Other
There are a lot of other things in the rust-lang/rust
repo that are related
to building a full Rust distribution. Most of the time you don't need to worry about them.
These include:
src/ci
: The CI configuration. This actually quite extensive because we run a lot of tests on a lot of platforms.src/doc
: Various documentation, including submodules for a few books.src/etc
: Miscellaneous utilities.- And more...
Queries: demand-driven compilation
As described in the high-level overview of the compiler, the Rust compiler
is still (as of July 2021) transitioning from a
traditional "pass-based" setup to a "demand-driven" system. The compiler query
system is the key to rustc's demand-driven organization.
The idea is pretty simple. Instead of entirely independent passes
(parsing, type-checking, etc.), a set of function-like queries
compute information about the input source. For example,
there is a query called type_of
that, given the DefId
of
some item, will compute the type of that item and return it to you.
Query execution is memoized. The first time you invoke a query, it will go do the computation, but the next time, the result is returned from a hashtable. Moreover, query execution fits nicely into incremental computation; the idea is roughly that, when you invoke a query, the result may be returned to you by loading stored data from disk.1
Eventually, we want the entire compiler
control-flow to be query driven. There will effectively be one
top-level query (compile
) that will run compilation on a crate; this
will in turn demand information about that crate, starting from the
end. For example:
- The
compile
query might demand to get a list of codegen-units (i.e. modules that need to be compiled by LLVM). - But computing the list of codegen-units would invoke some subquery that returns the list of all modules defined in the Rust source.
- That query in turn would invoke something asking for the HIR.
- This keeps going further and further back until we wind up doing the actual parsing.
Although this vision is not fully realized, large sections of the compiler (for example, generating MIR) currently work exactly like this.
The "Incremental Compilation in Detail chapter gives a more in-depth description of what queries are and how they work. If you intend to write a query of your own, this is a good read.
Invoking queries
Invoking a query is simple. The TyCtxt
("type context") struct offers a method
for each defined query. For example, to invoke the type_of
query, you would just do this:
let ty = tcx.type_of(some_def_id);
How the compiler executes a query
So you may be wondering what happens when you invoke a query
method. The answer is that, for each query, the compiler maintains a
cache – if your query has already been executed, then, the answer is
simple: we clone the return value out of the cache and return it
(therefore, you should try to ensure that the return types of queries
are cheaply cloneable; insert an Rc
if necessary).
Providers
If, however, the query is not in the cache, then the compiler will try to find a suitable provider. A provider is a function that has been defined and linked into the compiler somewhere that contains the code to compute the result of the query.
Providers are defined per-crate. The compiler maintains,
internally, a table of providers for every crate, at least
conceptually. Right now, there are really two sets: the providers for
queries about the local crate (that is, the one being compiled)
and providers for queries about external crates (that is,
dependencies of the local crate). Note that what determines the crate
that a query is targeting is not the kind of query, but the key.
For example, when you invoke tcx.type_of(def_id)
, that could be a
local query or an external query, depending on what crate the def_id
is referring to (see the self::keys::Key
trait for more
information on how that works).
Providers always have the same signature:
fn provider<'tcx>(
tcx: TyCtxt<'tcx>,
key: QUERY_KEY,
) -> QUERY_RESULT {
...
}
Providers take two arguments: the tcx
and the query key.
They return the result of the query.
How providers are setup
When the tcx is created, it is given the providers by its creator using
the Providers
struct. This struct is generated by
the macros here, but it is basically a big list of function pointers:
struct Providers {
type_of: for<'tcx> fn(TyCtxt<'tcx>, DefId) -> Ty<'tcx>,
...
}
At present, we have one copy of the struct for local crates, and one for external crates, though the plan is that we may eventually have one per crate.
These Providers
structs are ultimately created and populated by
rustc_driver
, but it does this by distributing the work
throughout the other rustc_*
crates. This is done by invoking
various provide
functions. These functions tend to look
something like this:
pub fn provide(providers: &mut Providers) {
*providers = Providers {
type_of,
..*providers
};
}
That is, they take an &mut Providers
and mutate it in place. Usually
we use the formulation above just because it looks nice, but you could
as well do providers.type_of = type_of
, which would be equivalent.
(Here, type_of
would be a top-level function, defined as we saw
before.) So, if we want to add a provider for some other query,
let's call it fubar
, into the crate above, we might modify the provide()
function like so:
pub fn provide(providers: &mut Providers) {
*providers = Providers {
type_of,
fubar,
..*providers
};
}
fn fubar<'tcx>(tcx: TyCtxt<'tcx>, key: DefId) -> Fubar<'tcx> { ... }
N.B. Most of the rustc_*
crates only provide local
providers. Almost all extern providers wind up going through the
rustc_metadata
crate, which loads the information
from the crate metadata. But in some cases there are crates that
provide queries for both local and external crates, in which case
they define both a provide
and a provide_extern
function, through
wasm_import_module_map
, that rustc_driver
can invoke.
Adding a new query
How do you add a new query? Defining a query takes place in two steps:
- Declare the query name, its arguments and description.
- Supply query providers where needed.
To declare the query name and arguments, you simply add an entry to
the big macro invocation in compiler/rustc_middle/src/query/mod.rs
.
Then you need to add a documentation comment to it with some internal description.
Then, provide the desc
attribute which contains a user-facing description of the query.
The desc
attribute is shown to the user in query cycles.
This looks something like:
rustc_queries! {
/// Records the type of every item.
query type_of(key: DefId) -> Ty<'tcx> {
cache_on_disk_if { key.is_local() }
desc { |tcx| "computing the type of `{}`", tcx.def_path_str(key) }
}
...
}
A query definition has the following form:
query type_of(key: DefId) -> Ty<'tcx> { ... }
^^^^^ ^^^^^^^ ^^^^^ ^^^^^^^^ ^^^
| | | | |
| | | | query modifiers
| | | result type
| | query key type
| name of query
query keyword
Let's go over these elements one by one:
- Query keyword: indicates a start of a query definition.
- Name of query: the name of the query method
(
tcx.type_of(..)
). Also used as the name of a struct (ty::queries::type_of
) that will be generated to represent this query. - Query key type: the type of the argument to this query.
This type must implement the
ty::query::keys::Key
trait, which defines (for example) how to map it to a crate, and so forth. - Result type of query: the type produced by this query. This type
should (a) not use
RefCell
or other interior mutability and (b) be cheaply cloneable. Interning or usingRc
orArc
is recommended for non-trivial data types.2 - Query modifiers: various flags and options that customize how the query is processed (mostly with respect to incremental compilation).
So, to add a query:
- Add an entry to
rustc_queries!
using the format above. - Link the provider by modifying the appropriate
provide
method; or add a new one if needed and ensure thatrustc_driver
is invoking it.
The one exception to those rules is the ty::steal::Steal
type,
which is used to cheaply modify MIR in place. See the definition
of Steal
for more details. New uses of Steal
should not be
added without alerting @rust-lang/compiler
.
External links
Related design ideas, and tracking issues:
- Design document: On-demand Rustc incremental design doc
- Tracking Issue: "Red/Green" dependency tracking in compiler
More discussion and issues:
The Query Evaluation Model in Detail
- What is a query?
- Caching/Memoization
- Input data
- An example execution trace of some queries
- Cycles
- "Steal" Queries
This chapter provides a deeper dive into the abstract model queries are built on. It does not go into implementation details but tries to explain the underlying logic. The examples here, therefore, have been stripped down and simplified and don't directly reflect the compilers internal APIs.
What is a query?
Abstractly we view the compiler's knowledge about a given crate as a "database" and queries are the way of asking the compiler questions about it, i.e. we "query" the compiler's "database" for facts.
However, there's something special to this compiler database: It starts out empty and is filled on-demand when queries are executed. Consequently, a query must know how to compute its result if the database does not contain it yet. For doing so, it can access other queries and certain input values that the database is pre-filled with on creation.
A query thus consists of the following things:
- A name that identifies the query
- A "key" that specifies what we want to look up
- A result type that specifies what kind of result it yields
- A "provider" which is a function that specifies how the result is to be computed if it isn't already present in the database.
As an example, the name of the type_of
query is type_of
, its query key is a
DefId
identifying the item we want to know the type of, the result type is
Ty<'tcx>
, and the provider is a function that, given the query key and access
to the rest of the database, can compute the type of the item identified by the
key.
So in some sense a query is just a function that maps the query key to the corresponding result. However, we have to apply some restrictions in order for this to be sound:
- The key and result must be immutable values.
- The provider function must be a pure function in the sense that for the same key it must always yield the same result.
- The only parameters a provider function takes are the key and a reference to the "query context" (which provides access to the rest of the "database").
The database is built up lazily by invoking queries. The query providers will invoke other queries, for which the result is either already cached or computed by calling another query provider. These query provider invocations conceptually form a directed acyclic graph (DAG) at the leaves of which are input values that are already known when the query context is created.
Caching/Memoization
Results of query invocations are "memoized" which means that the query context will cache the result in an internal table and, when the query is invoked with the same query key again, will return the result from the cache instead of running the provider again.
This caching is crucial for making the query engine efficient. Without memoization the system would still be sound (that is, it would yield the same results) but the same computations would be done over and over again.
Memoization is one of the main reasons why query providers have to be pure functions. If calling a provider function could yield different results for each invocation (because it accesses some global mutable state) then we could not memoize the result.
Input data
When the query context is created, it is still empty: No queries have been executed, no results are cached. But the context already provides access to "input" data, i.e. pieces of immutable data that were computed before the context was created and that queries can access to do their computations.
As of January 2021, this input data consists mainly of the HIR map, upstream crate metadata, and the command-line options the compiler was invoked with; but in the future inputs will just consist of command-line options and a list of source files -- the HIR map will itself be provided by a query which processes these source files.
Without inputs, queries would live in a void without anything to compute their result from (remember, query providers only have access to other queries and the context but not any other outside state or information).
For a query provider, input data and results of other queries look exactly the same: It just tells the context "give me the value of X". Because input data is immutable, the provider can rely on it being the same across different query invocations, just as is the case for query results.
An example execution trace of some queries
How does this DAG of query invocations come into existence? At some point the compiler driver will create the, as yet empty, query context. It will then, from outside of the query system, invoke the queries it needs to perform its task. This looks something like the following:
fn compile_crate() {
let cli_options = ...;
let hir_map = ...;
// Create the query context `tcx`
let tcx = TyCtxt::new(cli_options, hir_map);
// Do type checking by invoking the type check query
tcx.type_check_crate();
}
The type_check_crate
query provider would look something like the following:
fn type_check_crate_provider(tcx, _key: ()) {
let list_of_hir_items = tcx.hir_map.list_of_items();
for item_def_id in list_of_hir_items {
tcx.type_check_item(item_def_id);
}
}
We see that the type_check_crate
query accesses input data
(tcx.hir_map.list_of_items()
) and invokes other queries
(type_check_item
). The type_check_item
invocations will themselves access input data and/or invoke other queries,
so that in the end the DAG of query invocations will be built up backwards
from the node that was initially executed:
(2) (1)
list_of_all_hir_items <----------------------------- type_check_crate()
|
(5) (4) (3) |
Hir(foo) <--- type_of(foo) <--- type_check_item(foo) <-------+
| |
+-----------------+ |
| |
(7) v (6) (8) |
Hir(bar) <--- type_of(bar) <--- type_check_item(bar) <-------+
// (x) denotes invocation order
We also see that often a query result can be read from the cache:
type_of(bar)
was computed for type_check_item(foo)
so when
type_check_item(bar)
needs it, it is already in the cache.
Query results stay cached in the query context as long as the context lives. So if the compiler driver invoked another query later on, the above graph would still exist and already executed queries would not have to be re-done.
Cycles
Earlier we stated that query invocations form a DAG. However, it would be easy to form a cyclic graph by, for example, having a query provider like the following:
fn cyclic_query_provider(tcx, key) -> u32 {
// Invoke the same query with the same key again
tcx.cyclic_query(key)
}
Since query providers are regular functions, this would behave much as expected: Evaluation would get stuck in an infinite recursion. A query like this would not be very useful either. However, sometimes certain kinds of invalid user input can result in queries being called in a cyclic way. The query engine includes a check for cyclic invocations of queries with the same input arguments. And, because cycles are an irrecoverable error, will abort execution with a "cycle error" message that tries to be human readable.
At some point the compiler had a notion of "cycle recovery", that is, one could "try" to execute a query and if it ended up causing a cycle, proceed in some other fashion. However, this was later removed because it is not entirely clear what the theoretical consequences of this are, especially regarding incremental compilation.
"Steal" Queries
Some queries have their result wrapped in a Steal<T>
struct. These queries
behave exactly the same as regular with one exception: Their result is expected
to be "stolen" out of the cache at some point, meaning some other part of the
program is taking ownership of it and the result cannot be accessed anymore.
This stealing mechanism exists purely as a performance optimization because some result values are too costly to clone (e.g. the MIR of a function). It seems like result stealing would violate the condition that query results must be immutable (after all we are moving the result value out of the cache) but it is OK as long as the mutation is not observable. This is achieved by two things:
- Before a result is stolen, we make sure to eagerly run all queries that might ever need to read that result. This has to be done manually by calling those queries.
- Whenever a query tries to access a stolen result, we make an ICE (Internal Compiler Error) so that such a condition cannot go unnoticed.
This is not an ideal setup because of the manual intervention needed, so it should be used sparingly and only when it is well known which queries might access a given result. In practice, however, stealing has not turned out to be much of a maintenance burden.
To summarize: "Steal queries" break some of the rules in a controlled way. There are checks in place that make sure that nothing can go silently wrong.
Incremental compilation
The incremental compilation scheme is, in essence, a surprisingly simple extension to the overall query system. We'll start by describing a slightly simplified variant of the real thing – the "basic algorithm" – and then describe some possible improvements.
The basic algorithm
The basic algorithm is called the red-green algorithm1. The high-level idea is that, after each run of the compiler, we will save the results of all the queries that we do, as well as the query DAG. The query DAG is a DAG that indexes which queries executed which other queries. So, for example, there would be an edge from a query Q1 to another query Q2 if computing Q1 required computing Q2 (note that because queries cannot depend on themselves, this results in a DAG and not a general graph).
NOTE: You might think of a query as simply the definition of a query. A thing that you can invoke, a bit like a function, and which either returns a cached result or actually executes the code.
If that's the way you think about queries, it's good to know that in the following text, queries will be said to have colours. Keep in mind though, that here the word query also refers to a certain invocation of the query for a certain input. As you will read later, queries are fingerprinted based on their arguments. The result of a query might change when we give it one argument and be coloured red, while it stays the same for another argument and is thus green.
In short, the word query is here not just used to mean the definition of a query, but also for a specific instance of that query with given arguments.
On the next run of the compiler, then, we can sometimes reuse these query results to avoid re-executing a query. We do this by assigning every query a color:
- If a query is colored red, that means that its result during this compilation has changed from the previous compilation.
- If a query is colored green, that means that its result is the same as the previous compilation.
There are two key insights here:
- First, if all the inputs to query Q are colored green, then the query Q must result in the same value as last time and hence need not be re-executed (or else the compiler is not deterministic).
- Second, even if some inputs to a query changes, it may be that it
still produces the same result as the previous compilation. In
particular, the query may only use part of its input.
- Therefore, after executing a query, we always check whether it produced the same result as the previous time. If it did, we can still mark the query as green, and hence avoid re-executing dependent queries.
The try-mark-green algorithm
At the core of incremental compilation is an algorithm called "try-mark-green". It has the job of determining the color of a given query Q (which must not have yet been executed). In cases where Q has red inputs, determining Q's color may involve re-executing Q so that we can compare its output, but if all of Q's inputs are green, then we can conclude that Q must be green without re-executing it or inspecting its value at all. In the compiler, this allows us to avoid deserializing the result from disk when we don't need it, and in fact enables us to sometimes skip serializing the result as well (see the refinements section below).
Try-mark-green works as follows:
- First check if the query Q was executed during the previous compilation.
- If not, we can just re-execute the query as normal, and assign it the color of red.
- If yes, then load the 'dependent queries' of Q.
- If there is a saved result, then we load the
reads(Q)
vector from the query DAG. The "reads" is the set of queries that Q executed during its execution.- For each query R in
reads(Q)
, we recursively demand the color of R using try-mark-green.- Note: it is important that we visit each node in
reads(Q)
in same order as they occurred in the original compilation. See the section on the query DAG below. - If any of the nodes in
reads(Q)
wind up colored red, then Q is dirty.- We re-execute Q and compare the hash of its result to the hash of the result from the previous compilation.
- If the hash has not changed, we can mark Q as green and return.
- Otherwise, all of the nodes in
reads(Q)
must be green. In that case, we can color Q as green and return.
- Note: it is important that we visit each node in
- For each query R in
The query DAG
The query DAG code is stored in
compiler/rustc_middle/src/dep_graph
. Construction of the DAG is done
by instrumenting the query execution.
One key point is that the query DAG also tracks ordering; that is, for each query Q, we not only track the queries that Q reads, we track the order in which they were read. This allows try-mark-green to walk those queries back in the same order. This is important because once a subquery comes back as red, we can no longer be sure that Q will continue along the same path as before. That is, imagine a query like this:
fn main_query(tcx) {
if tcx.subquery1() {
tcx.subquery2()
} else {
tcx.subquery3()
}
}
Now imagine that in the first compilation, main_query
starts by
executing subquery1
, and this returns true. In that case, the next
query main_query
executes will be subquery2
, and subquery3
will
not be executed at all.
But now imagine that in the next compilation, the input has
changed such that subquery1
returns false. In this case, subquery2
would never execute. If try-mark-green were to visit reads(main_query)
out
of order, however, it might visit subquery2
before subquery1
, and hence
execute it.
This can lead to ICEs and other problems in the compiler.
Improvements to the basic algorithm
In the description of the basic algorithm, we said that at the end of compilation we would save the results of all the queries that were performed. In practice, this can be quite wasteful – many of those results are very cheap to recompute, and serializing and deserializing them is not a particular win. In practice, what we would do is to save the hashes of all the subqueries that we performed. Then, in select cases, we also save the results.
This is why the incremental algorithm separates computing the color of a node, which often does not require its value, from computing the result of a node. Computing the result is done via a simple algorithm like so:
- Check if a saved result for Q is available. If so, compute the color of Q. If Q is green, deserialize and return the saved result.
- Otherwise, execute Q.
- We can then compare the hash of the result and color Q as green if it did not change.
Resources
The initial design document can be found here, which expands on the memoization details, provides more high-level overview and motivation for this system.
Footnotes
I have long wanted to rename it to the Salsa algorithm, but it never caught on. -@nikomatsakis
Incremental Compilation In Detail
- A Basic Algorithm For Incremental Query Evaluation
- The Problem With The Basic Algorithm: False Positives
- Improving Accuracy: The red-green Algorithm
- The Real World: How Persistence Makes Everything Complicated
- Incremental Compilation and the Compiler Backend
- Shortcomings of the Current System
The incremental compilation scheme is, in essence, a surprisingly simple extension to the overall query system. It relies on the fact that:
- queries are pure functions -- given the same inputs, a query will always yield the same result, and
- the query model structures compilation in an acyclic graph that makes dependencies between individual computations explicit.
This chapter will explain how we can use these properties for making things incremental and then goes on to discuss version implementation issues.
A Basic Algorithm For Incremental Query Evaluation
As explained in the query evaluation model primer, query invocations form a directed-acyclic graph. Here's the example from the previous chapter again:
list_of_all_hir_items <----------------------------- type_check_crate()
|
|
Hir(foo) <--- type_of(foo) <--- type_check_item(foo) <-------+
| |
+-----------------+ |
| |
v |
Hir(bar) <--- type_of(bar) <--- type_check_item(bar) <-------+
Since every access from one query to another has to go through the query context, we can record these accesses and thus actually build this dependency graph in memory. With dependency tracking enabled, when compilation is done, we know which queries were invoked (the nodes of the graph) and for each invocation, which other queries or input has gone into computing the query's result (the edges of the graph).
Now suppose we change the source code of our program so that
HIR of bar
looks different than before. Our goal is to only recompute
those queries that are actually affected by the change while re-using
the cached results of all the other queries. Given the dependency graph we can
do exactly that. For a given query invocation, the graph tells us exactly
what data has gone into computing its results, we just have to follow the
edges until we reach something that has changed. If we don't encounter
anything that has changed, we know that the query still would evaluate to
the same result we already have in our cache.
Taking the type_of(foo)
invocation from above as an example, we can check
whether the cached result is still valid by following the edges to its
inputs. The only edge leads to Hir(foo)
, an input that has not been affected
by the change. So we know that the cached result for type_of(foo)
is still
valid.
The story is a bit different for type_check_item(foo)
: We again walk the
edges and already know that type_of(foo)
is fine. Then we get to
type_of(bar)
which we have not checked yet, so we walk the edges of
type_of(bar)
and encounter Hir(bar)
which has changed. Consequently
the result of type_of(bar)
might yield a different result than what we
have in the cache and, transitively, the result of type_check_item(foo)
might have changed too. We thus re-run type_check_item(foo)
, which in
turn will re-run type_of(bar)
, which will yield an up-to-date result
because it reads the up-to-date version of Hir(bar)
. Also, we re-run
type_check_item(bar)
because result of type_of(bar)
might have changed.
The Problem With The Basic Algorithm: False Positives
If you read the previous paragraph carefully you'll notice that it says that
type_of(bar)
might have changed because one of its inputs has changed.
There's also the possibility that it might still yield exactly the same
result even though its input has changed. Consider an example with a
simple query that just computes the sign of an integer:
IntValue(x) <---- sign_of(x) <--- some_other_query(x)
Let's say that IntValue(x)
starts out as 1000
and then is set to 2000
.
Even though IntValue(x)
is different in the two cases, sign_of(x)
yields
the result +
in both cases.
If we follow the basic algorithm, however, some_other_query(x)
would have to
(unnecessarily) be re-evaluated because it transitively depends on a changed
input. Change detection yields a "false positive" in this case because it has
to conservatively assume that some_other_query(x)
might be affected by that
changed input.
Unfortunately it turns out that the actual queries in the compiler are full of examples like this and small changes to the input often potentially affect very large parts of the output binaries. As a consequence, we had to make the change detection system smarter and more accurate.
Improving Accuracy: The red-green Algorithm
The "false positives" problem can be solved by interleaving change detection and query re-evaluation. Instead of walking the graph all the way to the inputs when trying to find out if some cached result is still valid, we can check if a result has actually changed after we were forced to re-evaluate it.
We call this algorithm the red-green algorithm because nodes in the dependency graph are assigned the color green if we were able to prove that its cached result is still valid and the color red if the result has turned out to be different after re-evaluating it.
The meat of red-green change tracking is implemented in the try-mark-green algorithm, that, you've guessed it, tries to mark a given node as green:
fn try_mark_green(tcx, current_node) -> bool {
// Fetch the inputs to `current_node`, i.e. get the nodes that the direct
// edges from `node` lead to.
let dependencies = tcx.dep_graph.get_dependencies_of(current_node);
// Now check all the inputs for changes
for dependency in dependencies {
match tcx.dep_graph.get_node_color(dependency) {
Green => {
// This input has already been checked before and it has not
// changed; so we can go on to check the next one
}
Red => {
// We found an input that has changed. We cannot mark
// `current_node` as green without re-running the
// corresponding query.
return false
}
Unknown => {
// This is the first time we look at this node. Let's try
// to mark it green by calling try_mark_green() recursively.
if try_mark_green(tcx, dependency) {
// We successfully marked the input as green, on to the
// next.
} else {
// We could *not* mark the input as green. This means we
// don't know if its value has changed. In order to find
// out, we re-run the corresponding query now!
tcx.run_query_for(dependency);
// Fetch and check the node color again. Running the query
// has forced it to either red (if it yielded a different
// result than we have in the cache) or green (if it
// yielded the same result).
match tcx.dep_graph.get_node_color(dependency) {
Red => {
// The input turned out to be red, so we cannot
// mark `current_node` as green.
return false
}
Green => {
// Re-running the query paid off! The result is the
// same as before, so this particular input does
// not invalidate `current_node`.
}
Unknown => {
// There is no way a node has no color after
// re-running the query.
panic!("unreachable")
}
}
}
}
}
}
// If we have gotten through the entire loop, it means that all inputs
// have turned out to be green. If all inputs are unchanged, it means
// that the query result corresponding to `current_node` cannot have
// changed either.
tcx.dep_graph.mark_green(current_node);
true
}
NOTE: The actual implementation can be found in
compiler/rustc_query_system/src/dep_graph/graph.rs
By using red-green marking we can avoid the devastating cumulative effect of
having false positives during change detection. Whenever a query is executed
in incremental mode, we first check if its already green. If not, we run
try_mark_green()
on it. If it still isn't green after that, then we actually
invoke the query provider to re-compute the result. Re-computing the query might
then itself involve recursively invoking more queries, which can mean we come back
to the try_mark_green()
algorithm for the dependencies recursively.
The Real World: How Persistence Makes Everything Complicated
The sections above described the underlying algorithm for incremental compilation but because the compiler process exits after being finished and takes the query context with its result cache with it into oblivion, we have to persist data to disk, so the next compilation session can make use of it. This comes with a whole new set of implementation challenges:
- The query result cache is stored to disk, so they are not readily available for change comparison.
- A subsequent compilation session will start off with new version of the code
that has arbitrary changes applied to it. All kinds of IDs and indices that
are generated from a global, sequential counter (e.g.
NodeId
,DefId
, etc) might have shifted, making the persisted results on disk not immediately usable anymore because the same numeric IDs and indices might refer to completely new things in the new compilation session. - Persisting things to disk comes at a cost, so not every tiny piece of information should be actually cached in between compilation sessions. Fixed-sized, plain-old-data is preferred to complex things that need to run through an expensive (de-)serialization step.
The following sections describe how the compiler solves these issues.
A Question Of Stability: Bridging The Gap Between Compilation Sessions
As noted before, various IDs (like DefId
) are generated by the compiler in a
way that depends on the contents of the source code being compiled. ID assignment
is usually deterministic, that is, if the exact same code is compiled twice,
the same things will end up with the same IDs. However, if something
changes, e.g. a function is added in the middle of a file, there is no
guarantee that anything will have the same ID as it had before.
As a consequence we cannot represent the data in our on-disk cache the same
way it is represented in memory. For example, if we just stored a piece
of type information like TyKind::FnDef(DefId, &'tcx Substs<'tcx>)
(as we do
in memory) and then the contained DefId
points to a different function in
a new compilation session we'd be in trouble.
The solution to this problem is to find "stable" forms for IDs which remain
valid in between compilation sessions. For the most important case, DefId
s,
these are the so-called DefPath
s. Each DefId
has a
corresponding DefPath
but in place of a numeric ID, a DefPath
is based on
the path to the identified item, e.g. std::collections::HashMap
. The
advantage of an ID like this is that it is not affected by unrelated changes.
For example, one can add a new function to std::collections
but
std::collections::HashMap
would still be std::collections::HashMap
. A
DefPath
is "stable" across changes made to the source code while a DefId
isn't.
There is also the DefPathHash
which is just a 128-bit hash value of the
DefPath
. The two contain the same information and we mostly use the
DefPathHash
because it simpler to handle, being Copy
and self-contained.
This principle of stable identifiers is used to make the data in the on-disk
cache resilient to source code changes. Instead of storing a DefId
, we store
the DefPathHash
and when we deserialize something from the cache, we map the
DefPathHash
to the corresponding DefId
in the current compilation session
(which is just a simple hash table lookup).
The HirId
, used for identifying HIR components that don't have their own
DefId
, is another such stable ID. It is (conceptually) a pair of a DefPath
and a LocalId
, where the LocalId
identifies something (e.g. a hir::Expr
)
locally within its "owner" (e.g. a hir::Item
). If the owner is moved around,
the LocalId
s within it are still the same.
Checking Query Results For Changes: HashStable And Fingerprints
In order to do red-green-marking we often need to check if the result of a query has changed compared to the result it had during the previous compilation session. There are two performance problems with this though:
- We'd like to avoid having to load the previous result from disk just for doing the comparison. We already computed the new result and will use that. Also loading a result from disk will "pollute" the interners with data that is unlikely to ever be used.
- We don't want to store each and every result in the on-disk cache. For example, it would be wasted effort to persist things to disk that are already available in upstream crates.
The compiler avoids these problems by using so-called Fingerprint
s. Each time
a new query result is computed, the query engine will compute a 128 bit hash
value of the result. We call this hash value "the Fingerprint
of the query
result". The hashing is (and has to be) done "in a stable way". This means
that whenever something is hashed that might change in between compilation
sessions (e.g. a DefId
), we instead hash its stable equivalent
(e.g. the corresponding DefPath
). That's what the whole HashStable
infrastructure is for. This way Fingerprint
s computed in two
different compilation sessions are still comparable.
The next step is to store these fingerprints along with the dependency graph. This is cheap since fingerprints are just bytes to be copied. It's also cheap to load the entire set of fingerprints together with the dependency graph.
Now, when red-green-marking reaches the point where it needs to check if a result has changed, it can just compare the (already loaded) previous fingerprint to the fingerprint of the new result.
This approach works rather well but it's not without flaws:
-
There is a small possibility of hash collisions. That is, two different results could have the same fingerprint and the system would erroneously assume that the result hasn't changed, leading to a missed update.
We mitigate this risk by using a high-quality hash function and a 128 bit wide hash value. Due to these measures the practical risk of a hash collision is negligible.
-
Computing fingerprints is quite costly. It is the main reason why incremental compilation can be slower than non-incremental compilation. We are forced to use a good and thus expensive hash function, and we have to map things to their stable equivalents while doing the hashing.
A Tale Of Two DepGraphs: The Old And The New
The initial description of dependency tracking glosses over a few details that quickly become a head scratcher when actually trying to implement things. In particular it's easy to overlook that we are actually dealing with two dependency graphs: The one we built during the previous compilation session and the one that we are building for the current compilation session.
When a compilation session starts, the compiler loads the previous dependency
graph into memory as an immutable piece of data. Then, when a query is invoked,
it will first try to mark the corresponding node in the graph as green. This
means really that we are trying to mark the node in the previous dep-graph
as green that corresponds to the query key in the current session. How do we
do this mapping between current query key and previous DepNode
? The answer
is again Fingerprint
s: Nodes in the dependency graph are identified by a
fingerprint of the query key. Since fingerprints are stable across compilation
sessions, computing one in the current session allows us to find a node
in the dependency graph from the previous session. If we don't find a node with
the given fingerprint, it means that the query key refers to something that
did not yet exist in the previous session.
So, having found the dep-node in the previous dependency graph, we can look up its dependencies (i.e. also dep-nodes in the previous graph) and continue with the rest of the try-mark-green algorithm. The next interesting thing happens when we successfully marked the node as green. At that point we copy the node and the edges to its dependencies from the old graph into the new graph. We have to do this because the new dep-graph cannot acquire the node and edges via the regular dependency tracking. The tracking system can only record edges while actually running a query -- but running the query, although we have the result already cached, is exactly what we want to avoid.
Once the compilation session has finished, all the unchanged parts have been copied over from the old into the new dependency graph, while the changed parts have been added to the new graph by the tracking system. At this point, the new graph is serialized out to disk, alongside the query result cache, and can act as the previous dep-graph in a subsequent compilation session.
Didn't You Forget Something?: Cache Promotion
The system described so far has a somewhat subtle property: If all inputs of a dep-node are green then the dep-node itself can be marked as green without computing or loading the corresponding query result. Applying this property transitively often leads to the situation that some intermediate results are never actually loaded from disk, as in the following example:
input(A) <-- intermediate_query(B) <-- leaf_query(C)
The compiler might need the value of leaf_query(C)
in order to generate some
output artifact. If it can mark leaf_query(C)
as green, it will load the
result from the on-disk cache. The result of intermediate_query(B)
is never
loaded though. As a consequence, when the compiler persists the new result
cache by writing all in-memory query results to disk, intermediate_query(B)
will not be in memory and thus will be missing from the new result cache.
If there subsequently is another compilation session that actually needs the
result of intermediate_query(B)
it will have to be re-computed even though we
had a perfectly valid result for it in the cache just before.
In order to prevent this from happening, the compiler does something called "cache promotion": Before emitting the new result cache it will walk all green dep-nodes and make sure that their query result is loaded into memory. That way the result cache doesn't unnecessarily shrink again.
Incremental Compilation and the Compiler Backend
The compiler backend, the part involving LLVM, is using the query system but it is not implemented in terms of queries itself. As a consequence it does not automatically partake in dependency tracking. However, the manual integration with the tracking system is pretty straight-forward. The compiler simply tracks what queries get invoked when generating the initial LLVM version of each codegen unit, which results in a dep-node for each of them. In subsequent compilation sessions it then tries to mark the dep-node for a CGU as green. If it succeeds it knows that the corresponding object and bitcode files on disk are still valid. If it doesn't succeed, the entire codegen unit has to be recompiled.
This is the same approach that is used for regular queries. The main differences are:
-
that we cannot easily compute a fingerprint for LLVM modules (because they are opaque C++ objects),
-
that the logic for dealing with cached values is rather different from regular queries because here we have bitcode and object files instead of serialized Rust values in the common result cache file, and
-
the operations around LLVM are so expensive in terms of computation time and memory consumption that we need to have tight control over what is executed when and what stays in memory for how long.
The query system could probably be extended with general purpose mechanisms to deal with all of the above but so far that seemed like more trouble than it would save.
Query Modifiers
The query system allows for applying modifiers to queries. These modifiers affect certain aspects of how the system treats the query with respect to incremental compilation:
-
eval_always
- A query with theeval_always
attribute is re-executed unconditionally during incremental compilation. I.e. the system will not even try to mark the query's dep-node as green. This attribute has two use cases:-
eval_always
queries can read inputs (from files, global state, etc). They can also produce side effects like writing to files and changing global state. -
Some queries are very likely to be re-evaluated because their result depends on the entire source code. In this case
eval_always
can be used as an optimization because the system can skip recording dependencies in the first place.
-
-
no_hash
- Applyingno_hash
to a query tells the system to not compute the fingerprint of the query's result. This has two consequences:-
Not computing the fingerprint can save quite a bit of time because fingerprinting is expensive, especially for large, complex values.
-
Without the fingerprint, the system has to unconditionally assume that the result of the query has changed. As a consequence anything depending on a
no_hash
query will always be re-executed.
Using
no_hash
for a query can make sense in two circumstances:-
If the result of the query is very likely to change whenever one of its inputs changes, e.g. a function like
|a, b, c| -> (a * b * c)
. In such a case recomputing the query will always yield a red node if one of the inputs is red so we can spare us the trouble and default to red immediately. A counter example would be a function like|a| -> (a == 42)
where the result does not change for most changes ofa
. -
If the result of a query is a big, monolithic collection (e.g.
index_hir
) and there are "projection queries" reading from that collection (e.g.hir_owner
). In such a case the big collection will likely fulfill the condition above (any changed input means recomputing the whole collection) and the results of the projection queries will be hashed anyway. If we also hashed the collection query it would mean that we effectively hash the same data twice: once when hashing the collection and another time when hashing all the projection query results.no_hash
allows us to avoid that redundancy and the projection queries act as a "firewall", shielding their dependents from the unconditionally redno_hash
node.
-
-
cache_on_disk_if
- This attribute is what determines which query results are persisted in the incremental compilation query result cache. The attribute takes an expression that allows per query invocation decisions. For example, it makes no sense to store values from upstream crates in the cache because they are already available in the upstream crate's metadata. -
anon
- This attribute makes the system use "anonymous" dep-nodes for the given query. An anonymous dep-node is not identified by the corresponding query key, instead its ID is computed from the IDs of its dependencies. This allows the red-green system to do its change detection even if there is no query key available for a given dep-node -- something which is needed for handling trait selection because it is not based on queries.
The Projection Query Pattern
It's interesting to note that eval_always
and no_hash
can be used together
in the so-called "projection query" pattern. It is often the case that there is
one query that depends on the entirety of the compiler's input (e.g. the indexed HIR)
and another query that projects individual values out of this monolithic value
(e.g. a HIR item with a certain DefId
). These projection queries allow for
building change propagation "firewalls" because even if the result of the
monolithic query changes (which it is very likely to do) the small projections
can still mostly be marked as green.
+------------+
| | +---------------+ +--------+
| | <---------| projection(x) | <---------| foo(a) |
| | +---------------+ +--------+
| |
| monolithic | +---------------+ +--------+
| query | <---------| projection(y) | <---------| bar(b) |
| | +---------------+ +--------+
| |
| | +---------------+ +--------+
| | <---------| projection(z) | <---------| baz(c) |
| | +---------------+ +--------+
+------------+
Let's assume that the result monolithic_query
changes so that also the result
of projection(x)
has changed, i.e. both their dep-nodes are being marked as
red. As a consequence foo(a)
needs to be re-executed; but bar(b)
and
baz(c)
can be marked as green. However, if foo
, bar
, and baz
would have
directly depended on monolithic_query
then all of them would have had to be
re-evaluated.
This pattern works even without eval_always
and no_hash
but the two
modifiers can be used to avoid unnecessary overhead. If the monolithic query
is likely to change at any minor modification of the compiler's input it makes
sense to mark it as eval_always
, thus getting rid of its dependency tracking
cost. And it always makes sense to mark the monolithic query as no_hash
because we have the projections to take care of keeping things green as much
as possible.
Shortcomings of the Current System
There are many things that still can be improved.
Incrementality of on-disk data structures
The current system is not able to update on-disk caches and the dependency graph in-place. Instead it has to rewrite each file entirely in each compilation session. The overhead of doing so is a few percent of total compilation time.
Unnecessary data dependencies
Data structures used as query results could be factored in a way that removes edges from the dependency graph. Especially "span" information is very volatile, so including it in query result will increase the chance that the result won't be reusable. See https://github.com/rust-lang/rust/issues/47389 for more information.
Debugging and Testing Dependencies
Testing the dependency graph
There are various ways to write tests against the dependency graph. The
simplest mechanisms are the #[rustc_if_this_changed]
and
#[rustc_then_this_would_need]
annotations. These are used in ui tests to test
whether the expected set of paths exist in the dependency graph.
As an example, see tests/ui/dep-graph/dep-graph-caller-callee.rs
, or the
tests below.
#[rustc_if_this_changed]
fn foo() { }
#[rustc_then_this_would_need(TypeckTables)] //~ ERROR OK
fn bar() { foo(); }
This should be read as
If this (
foo
) is changed, then this (i.e.bar
)'s TypeckTables would need to be changed.
Technically, what occurs is that the test is expected to emit the string "OK" on stderr, associated to this line.
You could also add the lines
#[rustc_then_this_would_need(TypeckTables)] //~ ERROR no path
fn baz() { }
Whose meaning is
If
foo
is changed, thenbaz
's TypeckTables does not need to be changed. The macro must emit an error, and the error message must contains "no path".
Recall that the //~ ERROR OK
is a comment from the point of view of the Rust
code we test, but is meaningful from the point of view of the test itself.
Debugging the dependency graph
Dumping the graph
The compiler is also capable of dumping the dependency graph for your
debugging pleasure. To do so, pass the -Z dump-dep-graph
flag. The
graph will be dumped to dep_graph.{txt,dot}
in the current
directory. You can override the filename with the RUST_DEP_GRAPH
environment variable.
Frequently, though, the full dep graph is quite overwhelming and not particularly helpful. Therefore, the compiler also allows you to filter the graph. You can filter in three ways:
- All edges originating in a particular set of nodes (usually a single node).
- All edges reaching a particular set of nodes.
- All edges that lie between given start and end nodes.
To filter, use the RUST_DEP_GRAPH_FILTER
environment variable, which should
look like one of the following:
source_filter // nodes originating from source_filter
-> target_filter // nodes that can reach target_filter
source_filter -> target_filter // nodes in between source_filter and target_filter
source_filter
and target_filter
are a &
-separated list of strings.
A node is considered to match a filter if all of those strings appear in its
label. So, for example:
RUST_DEP_GRAPH_FILTER='-> TypeckTables'
would select the predecessors of all TypeckTables
nodes. Usually though you
want the TypeckTables
node for some particular fn, so you might write:
RUST_DEP_GRAPH_FILTER='-> TypeckTables & bar'
This will select only the predecessors of TypeckTables
nodes for functions
with bar
in their name.
Perhaps you are finding that when you change foo
you need to re-type-check
bar
, but you don't think you should have to. In that case, you might do:
RUST_DEP_GRAPH_FILTER='Hir & foo -> TypeckTables & bar'
This will dump out all the nodes that lead from Hir(foo)
to
TypeckTables(bar)
, from which you can (hopefully) see the source
of the erroneous edge.
Tracking down incorrect edges
Sometimes, after you dump the dependency graph, you will find some
path that should not exist, but you will not be quite sure how it came
to be. When the compiler is built with debug assertions, it can
help you track that down. Simply set the RUST_FORBID_DEP_GRAPH_EDGE
environment variable to a filter. Every edge created in the dep-graph
will be tested against that filter – if it matches, a bug!
is
reported, so you can easily see the backtrace (RUST_BACKTRACE=1
).
The syntax for these filters is the same as described in the previous section. However, note that this filter is applied to every edge and doesn't handle longer paths in the graph, unlike the previous section.
Example:
You find that there is a path from the Hir
of foo
to the type
check of bar
and you don't think there should be. You dump the
dep-graph as described in the previous section and open dep-graph.txt
to see something like:
Hir(foo) -> Collect(bar)
Collect(bar) -> TypeckTables(bar)
That first edge looks suspicious to you. So you set
RUST_FORBID_DEP_GRAPH_EDGE
to Hir&foo -> Collect&bar
, re-run, and
then observe the backtrace. Voila, bug fixed!
How Salsa works
This chapter is based on the explanation given by Niko Matsakis in this video about Salsa. To find out more you may want to watch Salsa In More Depth, also by Niko Matsakis.
As of November 2022, although Salsa is inspired by (among other things) rustc's query system, it is not used directly in rustc. It is used in chalk, an implementation of Rust's trait system, and extensively in
rust-analyzer
, the official implementation of the language server protocol for Rust, but there are no medium or long-term concrete plans to integrate it into the compiler.
What is Salsa?
Salsa is a library for incremental recomputation. This means it allows reusing computations that were already done in the past to increase the efficiency of future computations.
The objectives of Salsa are:
- Provide that functionality in an automatic way, so reusing old computations is done automatically by the library.
- Doing so in a "sound", or "correct", way, therefore leading to the same results as if it had been done from scratch.
Salsa's actual model is much richer, allowing many kinds of inputs and many different outputs.
For example, integrating Salsa with an IDE could mean that
the inputs could be manifests (Cargo.toml
, rust-toolchain.toml
), entire
source files (foo.rs
), snippets and so on. The outputs of such an integration
could range from a binary executable, to lints, types (for example, if a user
selects a certain variable and wishes to see its type), completions, etc.
How does it work?
The first thing that Salsa has to do is identify the "base inputs" that are not something computed but given as input.
Then Salsa has to also identify intermediate, "derived" values, which are something that the library produces, but, for each derived value there's a "pure" function that computes the derived value.
For example, there might be a function ast(x: Path) -> AST
. The produced
Abstract Syntax Tree (AST
) isn't a final value, it's an intermediate value
that the library would use for the computation.
This means that when you try to compute with the library, Salsa is going to compute various derived values, and eventually read the input and produce the result for the asked computation.
In the course of computing, Salsa tracks which inputs were accessed and which values are derived. This information is used to determine what's going to happen when the inputs change: are the derived values still valid?
This doesn't necessarily mean that each computation downstream from the input is going to be checked, which could be costly. Salsa only needs to check each downstream computation until it finds one that isn't changed. At that point, it won't check other derived computations since they wouldn't need to change.
It's helpful to think about this as a graph with nodes. Each derived value has a dependency on other values, which could themselves be either base or derived. Base values don't have a dependency.
I <- A <- C ...
|
J <- B <--+
When an input I
changes, the derived value A
could change. The derived
value B
, which does not depend on I
, A
, or any value derived from A
or
I
, is not subject to change. Therefore, Salsa can reuse the computation done
for B
in the past, without having to compute it again.
The computation could also terminate early. Keeping the same graph as before,
say that input I
has changed in some way (and input J
hasn't), but when
computing A
again, it's found that A
hasn't changed from the previous
computation. This leads to an "early termination", because there's no need to
check if C
needs to change, since both C
direct inputs, A
and B
,
haven't changed.
Key Salsa concepts
Query
A query is some value that Salsa can access in the course of computation. Each
query can have a number of keys (from 0 to many), and all queries have a
result, akin to functions. 0-key
queries are called "input" queries.
Database
The database is basically the context for the entire computation, it's meant to store Salsa's internal state, all intermediate values for each query, and anything else that the computation might need. The database must know all the queries the library is going to do before it can be built, but they don't need to be specified in the same place.
After the database is formed, it can be accessed with queries that are very
similar to functions. Since each query's result is stored in the database, when
a query is invoked N
-times, it will return N
-cloned results, without having
to recompute the query (unless the input has changed in such a way that it
warrants recomputation).
For each input query (0-key
), a "set" method is generated, allowing the user to
change the output of such query, and trigger previous memoized values to be
potentially invalidated.
Query Groups
A query group is a set of queries which have been defined together as a unit. The database is formed by combining query groups. Query groups are akin to "Salsa modules".
A set of queries in a query group are just a set of methods in a trait.
To create a query group a trait annotated with a specific attribute
(#[salsa::query_group(...)]
) has to be created.
An argument must also be provided to said attribute as it will be used by Salsa
to create a struct
to be used later when the database is created.
Example input query group:
/// This attribute will process this tree, produce this tree as output, and produce
/// a bunch of intermediate stuff that Salsa also uses. One of these things is a
/// "StorageStruct", whose name we have specified in the attribute.
///
/// This query group is a bunch of **input** queries, that do not rely on any
/// derived input.
#[salsa::query_group(InputsStorage)]
pub trait Inputs {
/// This attribute (`#[salsa::input]`) indicates that this query is a base
/// input, therefore `set_manifest` is going to be auto-generated
#[salsa::input]
fn manifest(&self) -> Manifest;
#[salsa::input]
fn source_text(&self, name: String) -> String;
}
To create a derived query group, one must specify which other query groups this one depends on by specifying them as supertraits, as seen in the following example:
/// This query group is going to contain queries that depend on derived values.
/// A query group can access another query group's queries by specifying the
/// dependency as a supertrait. Query groups can be stacked as much as needed using
/// that pattern.
#[salsa::query_group(ParserStorage)]
pub trait Parser: Inputs {
/// This query `ast` is not an input query, it's a derived query this means
/// that a definition is necessary.
fn ast(&self, name: String) -> String;
}
When creating a derived query the implementation of said query must be defined
outside the trait. The definition must take a database parameter as an impl Trait
(or dyn Trait
), where trait is the query group that the definition
belongs to, in addition to the other keys.
/// This is going to be the definition of the `ast` query in the `Parser` trait.
/// So, when the query `ast` is invoked, and it needs to be recomputed, Salsa is
/// going to call this function and it's going to give it the database as `impl
/// Parser`. The function doesn't need to be aware of all the queries of all the
/// query groups
fn ast(db: &impl Parser, name: String) -> String {
//! Note, `impl Parser` is used here but `dyn Parser` works just as well
/* code */
///By passing an `impl Parser`, this is allowed
let source_text = db.input_file(name);
/* do the actual parsing */
return ast;
}
Eventually, after all the query groups have been defined, the database can be
created by declaring a struct
.
To specify which query groups are going to be part of the database an attribute
(#[salsa::database(...)]
) must be added. The argument of said attribute
is a
list of identifiers
, specifying the query groups storages.
///This attribute specifies which query groups are going to be in the database
#[salsa::database(InputsStorage, ParserStorage)]
#[derive(Default)] //optional!
struct MyDatabase {
///You also need this one field
runtime : salsa::Runtime<MyDatabase>,
}
///And this trait has to be implemented
impl salsa::Database for MyDatabase {
fn salsa_runtime(&self) -> &salsa::Runtime<MyDatabase> {
&self.runtime
}
}
Example usage:
fn main() {
let db = MyDatabase::default();
db.set_manifest(...);
db.set_source_text(...);
loop {
db.ast(...); //will reuse results
db.set_source_text(...);
}
}
Memory Management in Rustc
Generally rustc tries to be pretty careful how it manages memory. The compiler allocates a lot of data structures throughout compilation, and if we are not careful, it will take a lot of time and space to do so.
One of the main way the compiler manages this is using arenas and interning.
Arenas and Interning
Since A LOT of data structures are created during compilation, for performance
reasons, we allocate them from a global memory pool.
Each are allocated once from a long-lived arena.
This is called arena allocation.
This system reduces allocations/deallocations of memory.
It also allows for easy comparison of types (more on types here) for equality:
for each interned type X
, we implemented PartialEq
for X,
so we can just compare pointers.
The CtxtInterners
type contains a bunch of maps of interned types and the arena itself.
Example: ty::TyKind
Taking the example of ty::TyKind
which represents a type in the compiler (you
can read more here). Each time we want to construct a type, the
compiler doesn’t naively allocate from the buffer. Instead, we check if that
type was already constructed. If it was, we just get the same pointer we had
before, otherwise we make a fresh pointer. With this schema if we want to know
if two types are the same, all we need to do is compare the pointers which is
efficient. ty::TyKind
should never be constructed on the stack, and it would be unusable
if done so.
You always allocate them from this arena and you always intern them so they are
unique.
At the beginning of the compilation we make a buffer and each time we need to allocate a type we use
some of this memory buffer. If we run out of space we get another one. The lifetime of that buffer
is 'tcx
. Our types are tied to that lifetime, so when compilation finishes all the memory related
to that buffer is freed and our 'tcx
references would be invalid.
In addition to types, there are a number of other arena-allocated data structures that you can allocate, and which are found in this module. Here are a few examples:
GenericArgs
, allocated withmk_args
– this will intern a slice of types, often used to specify the values to be substituted for generics args (e.g.HashMap<i32, u32>
would be represented as a slice&'tcx [tcx.types.i32, tcx.types.u32]
).TraitRef
, typically passed by value – a trait reference consists of a reference to a trait along with its various type parameters (includingSelf
), likei32: Display
(here, the def-id would reference theDisplay
trait, and the args would containi32
). Note thatdef-id
is defined and discussed in depth in theAdtDef and DefId
section.Predicate
defines something the trait system has to prove (see traits module).
The tcx
and how it uses lifetimes
The typing context (tcx
) is the central data structure in the compiler. It is the context that
you use to perform all manner of queries. The struct
TyCtxt
defines a reference to this shared
context:
tcx: TyCtxt<'tcx>
// ----
// |
// arena lifetime
As you can see, the TyCtxt
type takes a lifetime parameter. When you see a reference with a
lifetime like 'tcx
, you know that it refers to arena-allocated data (or data that lives as long as
the arenas, anyhow).
A Note On Lifetimes
The Rust compiler is a fairly large program containing lots of big data
structures (e.g. the Abstract Syntax Tree (AST), High-Level Intermediate
Representation (HIR
), and the type system) and as such, arenas and
references are heavily relied upon to minimize unnecessary memory use. This
manifests itself in the way people can plug into the compiler (i.e. the
driver), preferring a "push"-style API (callbacks) instead
of the more Rust-ic "pull" style (think the Iterator
trait).
Thread-local storage and interning are used a lot through the compiler to reduce
duplication while also preventing a lot of the ergonomic issues due to many
pervasive lifetimes. The rustc_middle::ty::tls
module is used to access these
thread-locals, although you should rarely need to touch it.
Serialization in Rustc
rustc has to serialize and deserialize various data during compilation. Specifically:
- "Crate metadata", consisting mainly of query outputs, are serialized
from a binary format into
rlib
andrmeta
files that are output when compiling a library crate. Theserlib
andrmeta
files are then deserialized by the crates which depend on that library. - Certain query outputs are serialized in a binary format to persist incremental compilation results.
CrateInfo
is serialized toJSON
when the-Z no-link
flag is used, and deserialized fromJSON
when the-Z link-only
flag is used.
The Encodable
and Decodable
traits
The rustc_serialize
crate defines two traits for types which can be serialized:
pub trait Encodable<S: Encoder> {
fn encode(&self, s: &mut S) -> Result<(), S::Error>;
}
pub trait Decodable<D: Decoder>: Sized {
fn decode(d: &mut D) -> Result<Self, D::Error>;
}
It also defines implementations of these for various common standard library
primitive types such as integer
types, floating point types, bool
, char
, str
, etc.
For types that are constructed from those types, Encodable
and Decodable
are usually implemented by derives. These generate implementations that
forward deserialization to the fields of the struct or enum. For a
struct those impls look something like this:
#![feature(rustc_private)]
extern crate rustc_serialize;
use rustc_serialize::{Decodable, Decoder, Encodable, Encoder};
struct MyStruct {
int: u32,
float: f32,
}
impl<E: Encoder> Encodable<E> for MyStruct {
fn encode(&self, s: &mut E) -> Result<(), E::Error> {
s.emit_struct("MyStruct", 2, |s| {
s.emit_struct_field("int", 0, |s| self.int.encode(s))?;
s.emit_struct_field("float", 1, |s| self.float.encode(s))
})
}
}
impl<D: Decoder> Decodable<D> for MyStruct {
fn decode(s: &mut D) -> Result<MyStruct, D::Error> {
s.read_struct("MyStruct", 2, |d| {
let int = d.read_struct_field("int", 0, Decodable::decode)?;
let float = d.read_struct_field("float", 1, Decodable::decode)?;
Ok(MyStruct { int, float })
})
}
}
Encoding and Decoding arena allocated types
rustc has a lot of arena allocated types.
Deserializing these types isn't possible without access to the arena that they need to be allocated on.
The TyDecoder
and TyEncoder
traits are supertraits of Decoder
and Encoder
that allow access to a TyCtxt
.
Types which contain arena
allocated types can then bound the type parameter of their
Encodable
and Decodable
implementations with these traits.
For example
impl<'tcx, D: TyDecoder<'tcx>> Decodable<D> for MyStruct<'tcx> {
/* ... */
}
The TyEncodable
and TyDecodable
derive macros will expand to such
an implementation.
Decoding the actual arena
allocated type is harder, because some of the
implementations can't be written due to the orphan rules. To work around this,
the RefDecodable
trait is defined in rustc_middle
. This can then be
implemented for any type. The TyDecodable
macro will call RefDecodable
to
decode references, but various generic code needs types to actually be
Decodable
with a specific decoder.
For interned types instead of manually implementing RefDecodable
, using a new
type wrapper, like ty::Predicate
and manually implementing Encodable
and
Decodable
may be simpler.
Derive macros
The rustc_macros
crate defines various derives to help implement Decodable
and Encodable
.
- The
Encodable
andDecodable
macros generate implementations that apply to allEncoders
andDecoders
. These should be used in crates that don't depend onrustc_middle
, or that have to be serialized by a type that does not implementTyEncoder
. MetadataEncodable
andMetadataDecodable
generate implementations that only allow decoding byrustc_metadata::rmeta::encoder::EncodeContext
andrustc_metadata::rmeta::decoder::DecodeContext
. These are used for types that containrustc_metadata::rmeta::
Lazy*
.TyEncodable
andTyDecodable
generate implementation that apply to anyTyEncoder
orTyDecoder
. These should be used for types that are only serialized in crate metadata and/or the incremental cache, which is most serializable types inrustc_middle
.
Shorthands
Ty
can be deeply recursive, if each Ty
was encoded naively then crate
metadata would be very large. To handle this, each TyEncoder
has a cache of
locations in its output where it has serialized types. If a type being encoded
is in the cache, then instead of serializing the type as usual, the byte offset
within the file being written is encoded instead. A similar scheme is used for
ty::Predicate
.
LazyValue<T>
Crate metadata is initially loaded before the TyCtxt<'tcx>
is created, so
some deserialization needs to be deferred from the initial loading of metadata.
The LazyValue<T>
type wraps the (relative) offset in the crate metadata
where a T
has been serialized. There are also some variants, LazyArray<T>
and LazyTable<I, T>
.
The LazyArray<[T]>
and LazyTable<I, T>
types provide some functionality over
Lazy<Vec<T>>
and Lazy<HashMap<I, T>>
:
- It's possible to encode a
LazyArray<T>
directly from anIterator
, without first collecting into aVec<T>
. - Indexing into a
LazyTable<I, T>
does not require decoding entries other than the one being read.
note: LazyValue<T>
does not cache its value after being deserialized the
first time. Instead the query system its self is the main way of caching these
results.
Specialization
A few types, most notably DefId
, need to have different implementations for
different Encoder
s. This is currently handled by ad-hoc specializations, for
example: DefId
has a default
implementation of Encodable<E>
and a
specialized one for Encodable<CacheEncoder>
.
Parallel Compilation
Tracking issue: https://github.com/rust-lang/rust/issues/113349
As of November 2024, most of the rust compiler is now parallelized.
- The codegen part is executed concurrently by default. You can use the
-C codegen-units=n
option to control the number of concurrent tasks. - The parts after HIR lowering to codegen such as type checking, borrowing
checking, and mir optimization are parallelized in the nightly version.
Currently, they are executed in serial by default, and parallelization is
manually enabled by the user using the
-Z threads = n
option. - Other parts, such as lexical parsing, HIR lowering, and macro expansion, are still executed in serial mode.
Code Generation
During monomorphization the compiler splits up all the code to
be generated into smaller chunks called codegen units. These are then generated by
independent instances of LLVM running in parallel. At the end, the linker
is run to combine all the codegen units together into one binary. This process
occurs in the rustc_codegen_ssa::base
module.
Data Structures
The underlying thread-safe data-structures used in the parallel compiler
can be found in the rustc_data_structures::sync
module. These data structures
are implemented differently depending on whether parallel-compiler
is true.
data structure | parallel | non-parallel |
---|---|---|
Lrc | std::sync::Arc | std::rc::Rc |
Weak | std::sync::Weak | std::rc::Weak |
Atomic{Bool}/{Usize}/{U32}/{U64} | std::sync::atomic::Atomic{Bool}/{Usize}/{U32}/{U64} | (std::cell::Cell<bool/usize/u32/u64>) |
OnceCell | std::sync::OnceLock | std::cell::OnceCell |
Lock<T> | (parking_lot::Mutex<T>) | (std::cell::RefCell) |
RwLock<T> | (parking_lot::RwLock<T>) | (std::cell::RefCell) |
MTRef<'a, T> | &'a T | &'a mut T |
MTLock<T> | (Lock<T>) | (T) |
ReadGuard | parking_lot::RwLockReadGuard | std::cell::Ref |
MappedReadGuard | parking_lot::MappedRwLockReadGuard | std::cell::Ref |
WriteGuard | parking_lot::RwLockWriteGuard | std::cell::RefMut |
MappedWriteGuard | parking_lot::MappedRwLockWriteGuard | std::cell::RefMut |
LockGuard | parking_lot::MutexGuard | std::cell::RefMut |
MappedLockGuard | parking_lot::MappedMutexGuard | std::cell::RefMut |
-
These thread-safe data structures are interspersed during compilation which can cause lock contention resulting in degraded performance as the number of threads increases beyond 4. So we audit the use of these data structures which leads to either a refactoring so as to reduce the use of shared state, or the authoring of persistent documentation covering the specific of the invariants, the atomicity, and the lock orderings.
-
On the other hand, we still need to figure out what other invariants during compilation might not hold in parallel compilation.
WorkerLocal
WorkerLocal
is a special data structure implemented for parallel compilers. It
holds worker-locals values for each thread in a thread pool. You can only
access the worker local value through the Deref
impl
on the thread pool it
was constructed on. It panics otherwise.
WorkerLocal
is used to implement the Arena
allocator in the parallel
environment, which is critical in parallel queries. Its implementation is
located in the rustc_data_structures::sync::worker_local
module. However,
in the non-parallel compiler, it is implemented as (OneThread<T>)
, whose T
can be accessed directly through Deref::deref
.
Parallel Iterator
The parallel iterators provided by the rayon
crate are easy ways to
implement parallelism. In the current implementation of the parallel compiler
we use a custom fork of rayon
to run tasks in parallel.
Some iterator functions are implemented to run loops in parallel
when parallel-compiler
is true.
Function(Omit Send and Sync ) | Introduction | Owning Module |
---|---|---|
par_iter<T: IntoParallelIterator>(t: T) -> T::Iter | generate a parallel iterator | rustc_data_structure::sync |
par_for_each_in<T: IntoParallelIterator>(t: T, for_each: impl Fn(T::Item)) | generate a parallel iterator and run for_each on each element | rustc_data_structure::sync |
Map::par_body_owners(self, f: impl Fn(LocalDefId)) | run f on all hir owners in the crate | rustc_middle::hir::map |
Map::par_for_each_module(self, f: impl Fn(LocalDefId)) | run f on all modules and sub modules in the crate | rustc_middle::hir::map |
ModuleItems::par_items(&self, f: impl Fn(ItemId)) | run f on all items in the module | rustc_middle::hir |
ModuleItems::par_trait_items(&self, f: impl Fn(TraitItemId)) | run f on all trait items in the module | rustc_middle::hir |
ModuleItems::par_impl_items(&self, f: impl Fn(ImplItemId)) | run f on all impl items in the module | rustc_middle::hir |
ModuleItems::par_foreign_items(&self, f: impl Fn(ForeignItemId)) | run f on all foreign items in the module | rustc_middle::hir |
There are a lot of loops in the compiler which can possibly be parallelized using these functions. As of August 2022, scenarios where the parallel iterator function has been used are as follows:
caller | scenario | callee |
---|---|---|
rustc_metadata::rmeta::encoder::prefetch_mir | Prefetch queries which will be needed later by metadata encoding | par_iter |
rustc_monomorphize::collector::collect_crate_mono_items | Collect monomorphized items reachable from non-generic items | par_for_each_in |
rustc_interface::passes::analysis | Check the validity of the match statements | Map::par_body_owners |
rustc_interface::passes::analysis | MIR borrow check | Map::par_body_owners |
rustc_typeck::check::typeck_item_bodies | Type check | Map::par_body_owners |
rustc_interface::passes::hir_id_validator::check_crate | Check the validity of hir | Map::par_for_each_module |
rustc_interface::passes::analysis | Check the validity of loops body, attributes, naked functions, unstable abi, const bodys | Map::par_for_each_module |
rustc_interface::passes::analysis | Liveness and intrinsic checking of MIR | Map::par_for_each_module |
rustc_interface::passes::analysis | Deathness checking | Map::par_for_each_module |
rustc_interface::passes::analysis | Privacy checking | Map::par_for_each_module |
rustc_lint::late::check_crate | Run per-module lints | Map::par_for_each_module |
rustc_typeck::check_crate | Well-formedness checking | Map::par_for_each_module |
There are still many loops that have the potential to use parallel iterators.
Query System
The query model has some properties that make it actually feasible to evaluate multiple queries in parallel without too much effort:
- All data a query provider can access is via the query context, so the query context can take care of synchronizing access.
- Query results are required to be immutable so they can safely be used by different threads concurrently.
When a query foo
is evaluated, the cache table for foo
is locked.
- If there already is a result, we can clone it, release the lock and we are done.
- If there is no cache entry and no other active query invocation computing the same result, we mark the key as being "in progress", release the lock and start evaluating.
- If there is another query invocation for the same key in progress, we release the lock, and just block the thread until the other invocation has computed the result we are waiting for. Cycle error detection in the parallel compiler requires more complex logic than in single-threaded mode. When worker threads in parallel queries stop making progress due to interdependence, the compiler uses an extra thread (named deadlock handler) to detect, remove and report the cycle error.
The parallel query feature still has implementation to do, most of which is
related to the previous Data Structures
and Parallel Iterators
. See this
open feature tracking issue.
Rustdoc
As of November 2022, there are still a number of steps to
complete before rustdoc
rendering can be made parallel (see a open discussion
of parallel rustdoc
).
Resources
Here are some resources that can be used to learn more:
- This IRLO thread by alexchricton about performance
- This IRLO thread by Zoxc, one of the pioneers of the effort
- This list of interior mutability in the compiler by nikomatsakis
Rustdoc Internals
- From Crate to Clean
- From Clean To HTML
- Other Tricks Up Its Sleeve
- Dotting i's And Crossing t's
- Testing Locally
- See Also
This page describes rustdoc
's passes and modes. For an overview of rustdoc
,
see the "Rustdoc overview" chapter.
From Crate to Clean
In core.rs
are two central items: the rustdoc::core::DocContext
struct
, and the rustdoc::core::run_global_ctxt
function. The latter is
where rustdoc
calls out to rustc
to compile a crate to the point where
rustdoc
can take over. The former is a state container used when crawling
through a crate to gather its documentation.
The main process of crate crawling is done in clean/mod.rs
through several
functions with names that start with clean_
. Each function accepts an hir
or ty
data structure, and outputs a clean
structure used by rustdoc
. For
example, this function for converting lifetimes:
fn clean_lifetime<'tcx>(lifetime: &hir::Lifetime, cx: &mut DocContext<'tcx>) -> Lifetime {
if let Some(
rbv::ResolvedArg::EarlyBound(did)
| rbv::ResolvedArg::LateBound(_, _, did)
| rbv::ResolvedArg::Free(_, did),
) = cx.tcx.named_bound_var(lifetime.hir_id)
&& let Some(lt) = cx.args.get(&did).and_then(|arg| arg.as_lt())
{
return lt.clone();
}
Lifetime(lifetime.ident.name)
}
Also, clean/mod.rs
defines the types for the "cleaned" Abstract Syntax Tree
(AST
) used later to render documentation pages. Each usually accompanies a
clean_*
function that takes some AST
or High-Level Intermediate
Representation (HIR
) type from rustc
and converts it into the
appropriate "cleaned" type. "Big" items like modules or associated items may
have some extra processing in its clean
function, but for the most part these
impl
s are straightforward conversions. The "entry point" to this module is
clean::utils::krate
, which is called by run_global_ctxt
.
The first step in clean::utils::krate
is to invoke
visit_ast::RustdocVisitor
to process the module tree into an intermediate
visit_ast::Module
. This is the step that actually crawls the
rustc_hir::Crate
, normalizing various aspects of name resolution, such as:
- handling
#[doc(inline)]
and#[doc(no_inline)]
- handling import globs and cycles, so there are no duplicates or infinite directory trees
- inlining public
use
exports of private items, or showing a "Reexport" line in the module page - inlining items with
#[doc(hidden)]
if the base item is hidden but the - showing
#[macro_export]
-ed macros at the crate root, regardless of where they're defined reexport is not
After this step, clean::krate
invokes clean_doc_module
, which actually
converts the HIR
items to the cleaned AST
. This is also the step where cross-
crate inlining is performed, which requires converting rustc_middle
data
structures into the cleaned AST
.
The other major thing that happens in clean/mod.rs
is the collection of doc
comments and #[doc=""]
attributes into a separate field of the Attributes
struct
, present on anything that gets hand-written documentation. This makes it
easier to collect this documentation later in the process.
The primary output of this process is a clean::types::Crate
with a tree of Item
s
which describe the publicly-documentable items in the target crate.
Passes Anything But a Gas Station (or: Hot Potato)
Before moving on to the next major step, a few important "passes" occur over
the cleaned AST
. Several of these passes are lint
s and reports, but some of
them mutate or generate new items.
These are all implemented in the librustdoc/passes
directory, one file per pass.
By default, all of these passes are run on a crate, but the ones
regarding dropping private/hidden items can be bypassed by passing
--document-private-items
to rustdoc
. Note that unlike the previous set of AST
transformations, the passes are run on the cleaned crate.
Here is the list of passes as of March 2023:
-
calculate-doc-coverage
calculates information used for the--show-coverage
flag. -
check-doc-test-visibility
runsdoctest
visibility–relatedlint
s. This pass runs beforestrip-private
, which is why it needs to be separate fromrun-lints
. -
collect-intra-doc-links
resolves intra-doc links. -
collect-trait-impls
collectstrait
impl
s for each item in the crate. For example, if we define astruct
that implements atrait
, this pass will note that thestruct
implements thattrait
. -
propagate-doc-cfg
propagates#[doc(cfg(...))]
to child items. -
run-lints
runs some ofrustdoc
'slint
s, defined inpasses/lint
. This is the last pass to run.-
bare_urls
detects links that are not linkified, e.g., in Markdown such asGo to https://example.com/.
It suggests wrapping the link with angle brackets:Go to <https://example.com/>.
to linkify it. This is the code behind therustdoc::bare_urls
lint
. -
check_code_block_syntax
validates syntax inside Rust code blocks (```rust
) -
html_tags
detects invalidHTML
(like an unclosed<span>
) in doc comments.
-
-
strip-hidden
andstrip-private
strip alldoc(hidden)
and private items from the output.strip-private
impliesstrip-priv-imports
. Basically, the goal is to remove items that are not relevant for public documentation. This pass is skipped when--document-hidden-items
is passed. -
strip-priv-imports
strips all private import statements (use
,extern crate
) from a crate. This is necessary becauserustdoc
will handle public imports by either inlining the item's documentation to the module or creating a "Reexports" section with the import in it. The pass ensures that all of these imports are actually relevant to documentation. It is technically only run when--document-private-items
is passed, butstrip-private
accomplishes the same thing. -
strip-private
strips all private items from a crate which cannot be seen externally. This pass is skipped when--document-private-items
is passed.
There is also a stripper
module in librustdoc/passes
, but it is a
collection of utility functions for the strip-*
passes and is not a pass
itself.
From Clean To HTML
This is where the "second phase" in rustdoc
begins. This phase primarily lives
in the librustdoc/formats
and librustdoc/html
folders, and it all starts with
formats::renderer::run_format
. This code is responsible for setting up a type that
impl FormatRenderer
, which for HTML
is Context
.
This structure contains methods that get called by run_format
to drive the
doc rendering, which includes:
init
generatesstatic.files
, as well as search index andsrc/
item
generates the itemHTML
files themselvesafter_krate
generates other global resources likeall.html
In item
, the "page rendering" occurs, via a mixture of Askama templates
and manual write!()
calls, starting in html/layout.rs
. The parts that have
not been converted to templates occur within a series of std::fmt::Display
implementations and functions that pass around a &mut std::fmt::Formatter
.
The parts that actually generate HTML
from the items and documentation start
with print_item
defined in html/render/print_item.rs
, which switches out
to one of several item_*
functions based on kind of Item
being rendered.
Depending on what kind of rendering code you're looking for, you'll probably
find it either in html/render/mod.rs
for major items like "what sections
should I print for a struct
page" or html/format.rs
for smaller component
pieces like "how should I print a where clause as part of some other item".
Whenever rustdoc
comes across an item that should print hand-written
documentation alongside, it calls out to html/markdown.rs
which interfaces
with the Markdown parser. This is exposed as a series of types that wrap a
string of Markdown, and implement fmt::Display
to emit HTML
text. It takes
special care to enable certain features like footnotes and tables and add
syntax highlighting to Rust code blocks (via html/highlight.rs
) before
running the Markdown parser. There's also a function find_codes
which is
called by find_testable_codes
that specifically scans for Rust code blocks so
the test-runner code can find all the doctest
s in the crate.
From Soup to Nuts (or: "An Unbroken Thread Stretches From Those First Cell
s To Us")
It's important to note that rustdoc
can ask the compiler for type information
directly, even during HTML
generation. This didn't used to be the case, and
a lot of rustdoc
's architecture was designed around not doing that, but a
TyCtxt
is now passed to formats::renderer::run_format
, which is used to
run generation for both HTML
and the
(unstable as of March 2023) JSON format.
This change has allowed other changes to remove data from the "clean" AST
that can be easily derived from TyCtxt
queries, and we'll usually accept
PRs that remove fields from "clean" (it's been soft-deprecated), but this
is complicated from two other constraints that rustdoc
runs under:
- Docs can be generated for crates that don't actually pass type checking.
This is used for generating docs that cover mutually-exclusive platform
configurations, such as
libstd
having a single package of docs that cover all supported operating systems. This meansrustdoc
has to be able to generate docs fromHIR
. - Docs can inline across crates. Since crate metadata doesn't contain
HIR
, it must be possible to generate inlined docs from therustc_middle
data.
The "clean" AST
acts as a common output format for both input formats. There
is also some data in clean that doesn't correspond directly to HIR
, such as
synthetic impl
s for auto traits and blanket impl
s generated by the
collect-trait-impls
pass.
Some additional data is stored in
html::render::context::{Context, SharedContext}
. These two types serve as
ways to segregate rustdoc
's data for an eventual future with multithreaded doc
generation, as well as just keeping things organized:
Context
stores data used for generating the current page, such as its path, a list ofHTML
IDs that have been used (to avoid duplicateid=""
), and the pointer toSharedContext
.SharedContext
stores data that does not vary by page, such as thetcx
pointer, and a list of all types.
Other Tricks Up Its Sleeve
All this describes the process for generating HTML
documentation from a Rust
crate, but there are couple other major modes that rustdoc
runs in. It can also
be run on a standalone Markdown file, or it can run doctest
s on Rust code or
standalone Markdown files. For the former, it shortcuts straight to
html/markdown.rs
, optionally including a mode which inserts a Table of
Contents to the output HTML
.
For the latter, rustdoc
runs a similar partial-compilation to get relevant
documentation in test.rs
, but instead of going through the full clean and
render process, it runs a much simpler crate walk to grab just the
hand-written documentation. Combined with the aforementioned
"find_testable_code
" in html/markdown.rs
, it builds up a collection of
tests to run before handing them off to the test runner. One notable location
in test.rs
is the function make_test
, which is where hand-written
doctest
s get transformed into something that can be executed.
Some extra reading about make_test
can be found
here.
Dotting i's And Crossing t's
So that's rustdoc
's code in a nutshell, but there's more things in the
compiler that deal with it. Since we have the full compiletest
suite at hand,
there's a set of tests in tests/rustdoc
that make sure the final HTML
is
what we expect in various situations. These tests also use a supplementary
script, src/etc/htmldocck.py
, that allows it to look through the final HTML
using XPath
notation to get a precise look at the output. The full
description of all the commands available to rustdoc
tests (e.g. @has
and
@matches
) is in htmldocck.py
.
To use multiple crates in a rustdoc
test, add // aux-build:filename.rs
to the top of the test file. filename.rs
should be placed in an auxiliary
directory relative to the test file with the comment. If you need to build
docs for the auxiliary file, use // build-aux-docs
.
In addition, there are separate tests for the search index and rustdoc
's
ability to query it. The files in tests/rustdoc-js
each contain a
different search query and the expected results, broken out by search tab.
These files are processed by a script in src/tools/rustdoc-js
and the Node.js
runtime. These tests don't have as thorough of a writeup, but a broad example
that features results in all tabs can be found in basic.js
. The basic idea is
that you match a given QUERY
with a set of EXPECTED
results, complete with
the full item path of each item.
Testing Locally
Some features of the generated HTML
documentation might require local
storage to be used across pages, which doesn't work well without an HTTP
server. To test these features locally, you can run a local HTTP
server, like
this:
$ ./x doc library
# The documentation has been generated into `build/[YOUR ARCH]/doc`.
$ python3 -m http.server -d build/[YOUR ARCH]/doc
Now you can browse your documentation just like you would if it was hosted
on the internet. For example, the url for std
will be rust/std/
.
See Also
Rustdoc search
Rustdoc Search is two programs: search_index.rs
and search.js
. The first generates a nasty JSON
file with a full list of items and function signatures
in the crates in the doc bundle, and the second reads
it, turns it into some in-memory structures, and
scans them linearly to search.
Search index format
search.js
calls this Raw, because it turns it into
a more normal object tree after loading it.
For space savings, it's also written without newlines or spaces.
[
[ "crate_name", {
// name
"n": ["function_name", "Data"],
// type
"t": "HF",
// parent module
"q": [[0, "crate_name"]],
// parent type
"i": [2, 0],
// type dictionary
"p": [[1, "i32"], [1, "str"], [5, "Data", 0]],
// function signature
"f": "{{gb}{d}}`", // [[3, 1], [2]]
// impl disambiguator
"b": [],
// deprecated flag
"c": "OjAAAAAAAAA=", // empty bitmap
// empty description flag
"e": "OjAAAAAAAAA=", // empty bitmap
// aliases
"a": [["get_name", 0]],
// description shards
"D": "g", // 3
// inlined re-exports
"r": [],
}]
]
src/librustdoc/html/static/js/externs.js
defines an actual schema in a Closure @typedef
.
Key | Name | Description |
---|---|---|
n | Names | Item names |
t | Item Type | One-char item type code |
q | Parent module | Map<index, path> |
i | Parent type | list of indexes |
f | Function signature | encoded |
b | Impl disambiguator | Map<index, string> |
c | Deprecation flag | roaring bitmap |
e | Description is empty | roaring bitmap |
p | Type dictionary | [[item type, path]] |
a | Alias | Map<string, index> |
D | description shards | encoded |
The above index defines a crate called crate_name
with a free function called function_name
and a struct called Data
,
with the type signature Data, i32 -> str
,
and an alias, get_name
, that equivalently refers to function_name
.
The search index needs to fit the needs of the rustdoc
compiler,
the search.js
frontend,
and also be compact and fast to decode.
It makes a lot of compromises:
- The
rustdoc
compiler runs on one crate at a time, so each crate has an essentially separate search index. It merges them by having each crate on one line and looking at the first quoted string. - Names in the search index are given
in their original case and with underscores.
When the search index is loaded,
search.js
stores the original names for display, but also folds them to lowercase and strips underscores for search. You'll see them callednormalized
. - The
f
array stores types as offsets into thep
array. These types might actually be from another crate, sosearch.js
has to turn the numbers into names and then back into numbers to deduplicate them if multiple crates in the same index mention the same types. - It's a JSON file, but not designed to be human-readable.
Browsers already include an optimized JSON decoder,
so this saves on
search.js
code and performs better for small crates, but instead of using objects like normal JSON formats do, it tries to put data of the same type next to each other so that the sliding window used by DEFLATE can find redundancies. Wheresearch.js
does its own compression, it's designed to save memory when the file is finally loaded, not just size on disk or network transfer.
Parallel arrays and indexed maps
Abstractly, Rustdoc Search data is a table, stored in column-major form. Most data in the index represents a set of parallel arrays (the "columns") which refer to the same data if they're at the same position.
For example, the above search index can be turned into this table:
n | t | d | q | i | f | b | c | |
---|---|---|---|---|---|---|---|---|
0 | crate_name | D | Documentation | NULL | 0 | NULL | NULL | 0 |
1 | function_name | H | This function gets the name of an integer with Data | crate_name | 2 | {{gb}{d}} | NULL | 0 |
2 | Data | F | The data struct | crate_name | 0 | ` | NULL | 0 |
The crate row is implied in most columns, since its type is known (it's a crate),
it can't have a parent (crates form the root of the module tree),
its name is specified as the map key,
and function-specific data like the impl disambiguator can't apply either.
However, it can still have a description and it can still be deprecated.
The crate, therefore, has a primary key of 0
.
The above code doesn't use c
, which holds deprecated indices,
or b
, which maps indices to strings.
If crate_name::function_name
used both, it might look like this.
"b": [[0, "impl-Foo-for-Bar"]],
"c": "OjAAAAEAAAAAAAIAEAAAABUAbgZYCQ==",
This attaches a disambiguator to index 1 and marks it deprecated.
The advantage of this layout is that these APIs often have implicit structure that DEFLATE can take advantage of, but that rustdoc can't assume. Like how names are usually CamelCase or snake_case, but descriptions aren't. It also makes it easier to use a sparse data for things like boolean flags.
q
is a Map from the first applicable ID to a parent module path.
This is a weird trick, but it makes more sense in pseudo-code:
#![allow(unused)] fn main() { let mut parent_module = ""; for (i, entry) in search_index.iter().enumerate() { if q.contains(i) { parent_module = q.get(i); } // ... do other stuff with `entry` ... } }
This is valid because everything has a parent module (even if it's just the crate itself), and is easy to assemble because the rustdoc generator sorts by path before serializing. Doing this allows rustdoc to not only make the search index smaller, but reuse the same string representing the parent path across multiple in-memory items.
Representing sparse columns
VLQ Hex
This format is, as far as I know, used nowhere other than rustdoc. It follows this grammar:
VLQHex = { VHItem | VHBackref }
VHItem = VHNumber | ( '{', {VHItem}, '}' )
VHNumber = { '@' | 'A' | 'B' | 'C' | 'D' | 'E' | 'F' | 'G' | 'H' | 'I' | 'J' | 'K' | 'L' | 'M' | 'N' | 'O' }, ( '`' | 'a' | 'b' | 'c' | 'd' | 'e' | 'f' | 'g' | 'h' | 'i' | 'j' | 'k ' | 'l' | 'm' | 'n' | 'o' )
VHBackref = ( '0' | '1' | '2' | '3' | '4' | '5' | '6' | '7' | '8' | '9' | ':' | ';' | '<' | '=' | '>' | '?' )
A VHNumber is a variable-length, self-terminating base16 number (terminated because the last hexit is lowercase while all others are uppercase). The sign bit is represented using zig-zag encoding.
This alphabet is chosen because the characters can be turned into hexits by masking off the last four bits of the ASCII encoding.
A major feature of this encoding, as with all of the "compression" done in rustdoc,
is that it can remain in its compressed format even in memory at runtime.
This is why HBackref
is only used at the top level,
and why we don't just use Flate for everything: the decoder in search.js
will reuse the entire decoded object whenever a backref is seen,
saving decode work and memory.
Roaring Bitmaps
Flag-style data, such as deprecation and empty descriptions, are stored using the standard Roaring Bitmap serialization format with runs. The data is then base64 encoded when writing it.
As a brief overview: a roaring bitmap is a chunked array of bits, described in this paper. A chunk can either be a list of integers, a bitfield, or a list of runs. In any case, the search engine has to base64 decode it, and read the chunk index itself, but the payload data stays as-is.
All roaring bitmaps in rustdoc currently store a flag for each item index. The crate is item 0, all others start at 1.
How descriptions are stored
The largest amount of data, and the main thing Rustdoc Search deals with that isn't actually used for searching, is descriptions. In a SERP table, this is what appears on the rightmost column.
item type item path description (this part) function my_crate::my_function This function gets the name of an integer with Data
When someone runs a search in rustdoc for the first time, their browser will work through a "sandwich workload" of three steps:
- Download the search-index.js and search.js files (a network bottleneck).
- Perform the actual search (a CPU and memory bandwidth bottleneck).
- Download the description data (another network bottleneck).
Reducing the amount of data downloaded here will almost always increase latency, by delaying the decision of what to download behind other work and/or adding data dependencies where something can't be downloaded without first downloading something else. In this case, we can't start downloading descriptions until after the search is done, because that's what allows it to decide which descriptions to download (it needs to sort the results then truncate to 200).
To do this, two columns are stored in the search index, building on both Roaring Bitmaps and on VLQ Hex.
e
is an index of empty descriptions. It's a roaring bitmap of each item (the crate itself is item 0, the rest start at 1).D
is a shard list, stored in VLQ hex as flat list of integers. Each integer gives you the number of descriptions in the shard. As the decoder walks the index, it checks if the description is empty. if it's not, then it's in the "current" shard. When all items are exhausted, it goes on to the next shard.
Inside each shard is a newline-delimited list of descriptions, wrapped in a JSONP-style function call.
i
, f
, and p
i
and f
both index into p
, the array of parent items.
i
is just a one-indexed number
(not zero-indexed because 0
is used for items that have no parent item).
It's different from q
because q
represents the parent module or crate,
which everything has,
while i
/q
are used for type and trait-associated items like methods.
f
, the function signatures, use a VLQ hex tree.
A number is either a one-indexed reference into p
,
a negative number representing a generic,
or zero for null.
(the internal object representation also uses negative numbers, even after decoding, to represent generics).
For example, {{gb}{d}}
is equivalent to the json [[3, 1], [2]]
.
Because of zigzag encoding, `
is +0, a
is -0 (which is not used),
b
is +1, and c
is -1.
Searching by name
Searching by name works by looping through the search index and running these functions on each:
editDistance
is always used to determine a match (unless quotes are specified, which would use simple equality instead). It computes the number of swaps, inserts, and removes needed to turn the query name into the entry name. For example,foo
has zero distance from itself, but a distance of 1 fromofo
(one swap) andfoob
(one insert). It is checked against an heuristic threshold, and then, if it is within that threshold, the distance is stored for ranking.String.prototype.indexOf
is always used to determine a match. If it returns anything other than -1, the result is added, even ifeditDistance
exceeds its threshold, and the index is stored for ranking.checkPath
is used if, and only if, a parent path is specified in the query. For example,vec
has no parent path, butvec::vec
does. Within checkPath, editDistance and indexOf are used, and the path query has its own heuristic threshold, too. If it's not within the threshold, the entry is rejected, even if the first two pass. If it's within the threshold, the path distance is stored for ranking.checkType
is used only if there's a type filter, like the struct instruct:vec
. If it fails, the entry is rejected.
If all four criteria pass
(plus the crate filter, which isn't technically part of the query),
the results are sorted by sortResults
.
Searching by type
Searching by type can be divided into two phases, and the second phase has two sub-phases.
- Turn names in the query into numbers.
- Loop over each entry in the search index:
- Quick rejection using a bloom filter.
- Slow rejection using a recursive type unification algorithm.
In the names->numbers phase, if the query has only one name in it,
the editDistance function is used to find a near match if the exact match fails,
but if there's multiple items in the query,
non-matching items are treated as generics instead.
This means hahsmap
will match hashmap on its own, but hahsmap, u32
is going to match the same things T, u32
matches
(though rustdoc will detect this particular problem and warn about it).
Then, when actually looping over each item,
the bloom filter will probably reject entries that don't have every
type mentioned in the query.
For example, the bloom query allows a query of i32 -> u32
to match
a function with the type i32, u32 -> bool
,
but unification will reject it later.
The unification filter ensures that:
- Bag semantics are respected. If you query says
i32, i32
, then the function has to mention two i32s, not just one. - Nesting semantics are respected. If your query says
vec<option>
, thenvec<option<i32>>
is fine, butoption<vec<i32>>
is not a match. - The division between return type and parameter is respected.
i32 -> u32
andu32 -> i32
are completely different.
The bloom filter checks none of these things, and, on top of that, can have false positives. But it's fast and uses very little memory, so the bloom filter helps.
Re-exports
Re-export inlining allows the same item to be found by multiple names. Search supports this by giving the same item multiple entries and tracking a canonical path for any items where that differs from the given path.
For example, this sample index has a single struct exported from two paths:
[
[ "crate_name", {
"doc": "Documentation",
"n": ["Data", "Data"],
"t": "FF",
"d": ["The data struct", "The data struct"],
"q": [[0, "crate_name"], [1, "crate_name::submodule"]],
"i": [0, 0],
"p": [],
"f": "``",
"b": [],
"c": [],
"a": [],
"r": [[0, 1]],
}]
]
The important part of this example is the r
array,
which indicates that path entry 1 in the q
array is
the canonical path for item 0.
That is, crate_name::Data
has a canonical path of crate_name::submodule::Data
.
This might sound like a strange design, since it has the duplicate data. It's done that way because inlining can happen across crates, which are compiled separately and might not all be present in the docs.
[
[ "crate_name", ... ],
[ "crate_name_2", { "q": [[0, "crate_name::submodule"], [5, "core::option"]], ... }]
]
In the above example, a canonical path actually comes from a dependency, and another one comes from an inlined standard library item: the canonical path isn't even in the index! The canonical path might also be private. In either case, it's never shown to the user, and is only used for deduplication.
Associated types, like methods, store them differently.
These types are connected with an entry in p
(their "parent")
and each one has an optional third tuple element:
"p": [[5, "Data", 0, 1]]
That's:
- 5: It's a struct
- "Data": Its name
- 0: Its display path, "crate_name"
- 1: Its canonical path, "crate_name::submodule"
In both cases, the canonical path might not be public at all, or it might be from another crate that isn't in the docs, so it's never shown to the user, but is used for deduplication.
Testing the search engine
While the generated UI is tested using rustdoc-gui
tests, the
primary way the search engine is tested is the rustdoc-js
and
rustdoc-js-std
tests. They run in NodeJS.
A rustdoc-js
test has a .rs
and .js
file, with the same name.
The .rs
file specifies the hypothetical library crate to run
the searches on (make sure you mark anything you need to find as pub
).
The .js
file specifies the actual searches.
The rustdoc-js-std
tests are the same, but don't require an .rs
file, since they use the standard library.
The .js
file is like a module (except the loader takes care of
exports
for you). It uses these variables:
Name | Type | Description |
---|---|---|
FILTER_CRATE | string | Only include results from the given crate. In the GUI, this is the "Results in crate" drop-down menu. |
EXPECTED | [ResultsTable]\|ResultsTable | List of tests to run, specifying what the hypothetical user types into the search box and sees in the tabs |
PARSED | [ParsedQuery]\|ParsedQuery | List of parser tests to run, without running an actual search |
FILTER_CRATE
can be left out (equivalent to searching "all crates"), but you
have to specify EXPECTED
or PARSED
.
By default, the test fails if any of the results specified in the test case are not found after running the search, or if the results found after running the search don't appear in the same order that they do in the test. The actual search results may, however, include results that aren't in the test. To override this, specify any of the following magic comments. Put them on their own line, without indenting.
// exact-check
: If search results appear that aren't part of the test case, then fail.// ignore-order
: Allow search results to appear in any order.// should-fail
: Used to write negative tests.
Standard library tests usually shouldn't specify // exact-check
, since we
want the libs team to be able to add new items without causing unrelated
tests to fail, but standalone tests will use it more often.
The ResultsTable
and ParsedQuery
types are specified in
externs.js
.
For example, imagine we needed to fix a bug where a function named
constructor
couldn't be found. To do this, write two files:
#![allow(unused)] fn main() { // tests/rustdoc-js/constructor_search.rs // The test case needs to find this result. pub fn constructor(_input: &str) -> i32 { 1 } }
// tests/rustdoc-js/constructor_search.js
// exact-check
// Since this test runs against its own crate,
// new items should not appear in the search results.
const EXPECTED = [
// This first test targets name-based search.
{
query: "constructor",
others: [
{ path: "constructor_search", name: "constructor" },
],
in_args: [],
returned: [],
},
// This test targets the second tab.
{
query: "str",
others: [],
in_args: [
{ path: "constructor_search", name: "constructor" },
],
returned: [],
},
// This test targets the third tab.
{
query: "i32",
others: [],
in_args: [],
returned: [
{ path: "constructor_search", name: "constructor" },
],
},
// This test targets advanced type-driven search.
{
query: "str -> i32",
others: [
{ path: "constructor_search", name: "constructor" },
],
in_args: [],
returned: [],
},
]
Source Code Representation
This part describes the process of taking raw source code from the user and transforming it into various forms that the compiler can work with easily. These are called intermediate representations (IRs).
This process starts with compiler understanding what the user has asked for: parsing the command line arguments given and determining what it is to compile. After that, the compiler transforms the user input into a series of IRs that look progressively less like what the user wrote.
Syntax and the AST
Working directly with source code is very inconvenient and error-prone. Thus, before we do anything else, we convert raw source code into an Abstract Syntax Tree (AST). It turns out that doing this involves a lot of work, including lexing, parsing, macro expansion, name resolution, conditional compilation, feature-gate checking, and validation of the AST. In this chapter, we take a look at all of these steps.
Notably, there isn't always a clean ordering between these tasks. For example, macro expansion relies on name resolution to resolve the names of macros and imports. And parsing requires macro expansion, which in turn may require parsing the output of the macro.
Lexing and Parsing
The very first thing the compiler does is take the program (in UTF-8 Unicode text) and turn it into a data format the compiler can work with more conveniently than strings. This happens in two stages: Lexing and Parsing.
- Lexing takes strings and turns them into streams of tokens. For
example,
foo.bar + buz
would be turned into the tokensfoo
,.
,bar
,+
, andbuz
. This is implemented inrustc_lexer
.
- Parsing takes streams of tokens and turns them into a structured form which is easier for the compiler to work with, usually called an Abstract Syntax Tree (AST) .
The AST
The AST mirrors the structure of a Rust program in memory, using a Span
to
link a particular AST node back to its source text. The AST is defined in
rustc_ast
, along with some definitions for tokens and token
streams, data structures/traits for mutating ASTs, and shared definitions for
other AST-related parts of the compiler (like the lexer and
macro-expansion).
Every node in the AST has its own NodeId
, including top-level items
such as structs, but also individual statements and expressions. A NodeId
is an identifier number that uniquely identifies an AST node within a crate.
However, because they are absolute within a crate, adding or removing a single
node in the AST causes all the subsequent NodeId
s to change. This renders
NodeId
s pretty much useless for incremental compilation, where you want as
few things as possible to change.
NodeId
s are used in all the rustc
bits that operate directly on the AST,
like macro expansion and name resolution (more on these over the next couple chapters).
Parsing
The parser is defined in rustc_parse
, along with a
high-level interface to the lexer and some validation routines that run after
macro expansion. In particular, the rustc_parse::parser
contains
the parser implementation.
The main entrypoint to the parser is via the various parse_*
functions and others in
rustc_parse. They let you do things like turn a SourceFile
(e.g. the source in a single file) into a token stream, create a parser from
the token stream, and then execute the parser to get a Crate
(the root AST
node).
To minimize the amount of copying that is done,
both StringReader
and Parser
have lifetimes which bind them to the parent ParseSess
.
This contains all the information needed while parsing, as well as the SourceMap
itself.
Note that while parsing, we may encounter macro definitions or invocations. We set these aside to be expanded (see Macro Expansion). Expansion itself may require parsing the output of a macro, which may reveal more macros to be expanded, and so on.
More on Lexical Analysis
Code for lexical analysis is split between two crates:
-
rustc_lexer
crate is responsible for breaking a&str
into chunks constituting tokens. Although it is popular to implement lexers as generated finite state machines, the lexer inrustc_lexer
is hand-written. -
StringReader
integratesrustc_lexer
with data structures specific torustc
. Specifically, it addsSpan
information to tokens returned byrustc_lexer
and interns identifiers.
Macro expansion
- Expansion and AST Integration
- Hygiene and Hierarchies
- Producing Macro Output
- Macros By Example
- Procedural Macros
N.B.
rustc_ast
,rustc_expand
, andrustc_builtin_macros
are all undergoing refactoring, so some of the links in this chapter may be broken.
Rust has a very powerful macro system. In the previous chapter, we saw how the parser sets aside macros to be expanded (using temporary placeholders). This chapter is about the process of expanding those macros iteratively until we have a complete Abstract Syntax Tree (AST) for our crate with no unexpanded macros (or a compile error).
First, we discuss the algorithm that expands and integrates macro output into ASTs. Next, we take a look at how hygiene data is collected. Finally, we look at the specifics of expanding different types of macros.
Many of the algorithms and data structures described below are in rustc_expand
,
with fundamental data structures in rustc_expand::base
.
Also of note, cfg
and cfg_attr
are treated specially from other macros, and are
handled in rustc_expand::config
.
Expansion and AST Integration
Firstly, expansion happens at the crate level. Given a raw source code for
a crate, the compiler will produce a massive AST with all macros expanded, all
modules inlined, etc. The primary entry point for this process is the
MacroExpander::fully_expand_fragment
method. With few exceptions, we
use this method on the whole crate (see "Eager Expansion"
below for more detailed discussion of edge case expansion issues).
At a high level, fully_expand_fragment
works in iterations. We keep a
queue of unresolved macro invocations (i.e. macros we haven't found the
definition of yet). We repeatedly try to pick a macro from the queue, resolve
it, expand it, and integrate it back. If we can't make progress in an
iteration, this represents a compile error. Here is the algorithm:
- Initialize a
queue
of unresolved macros. - Repeat until
queue
is empty (or we make no progress, which is an error):- Resolve imports in our partially built crate as much as possible.
- Collect as many macro
Invocation
s as possible from our partially built crate (fn
-like, attributes, derives) and add them to the queue. - Dequeue the first element and attempt to resolve it.
- If it's resolved:
- Run the macro's expander function that consumes a
TokenStream
or AST and produces aTokenStream
orAstFragment
(depending on the macro kind). (ATokenStream
is a collection ofTokenTree
s, each of which are a token (punctuation, identifier, or literal) or a delimited group (anything inside()
/[]
/{}
)).- At this point, we know everything about the macro itself and can
call
set_expn_data
to fill in its properties in the global data; that is the hygiene data associated withExpnId
(see Hygiene below).
- At this point, we know everything about the macro itself and can
call
- Integrate that piece of AST into the currently-existing though
partially-built AST. This is essentially where the "token-like mass"
becomes a proper set-in-stone AST with side-tables. It happens as
follows:
- If the macro produces tokens (e.g. a proc macro), we parse into an AST, which may produce parse errors.
- During expansion, we create
SyntaxContext
s (hierarchy 2) (see Hygiene below). - These three passes happen one after another on every AST fragment
freshly expanded from a macro:
NodeId
s are assigned byInvocationCollector
. This also collects new macro calls from this new AST piece and adds them to the queue.- "Def paths" are created and
DefId
s are assigned to them byDefCollector
. - Names are put into modules (from the resolver's point of
view) by
BuildReducedGraphVisitor
.
- After expanding a single macro and integrating its output, continue
to the next iteration of
fully_expand_fragment
.
- Run the macro's expander function that consumes a
- If it's not resolved:
- Put the macro back in the queue.
- Continue to next iteration...
Error Recovery
If we make no progress in an iteration we have reached a compilation error
(e.g. an undefined macro). We attempt to recover from failures (i.e.
unresolved macros or imports) with the intent of generating diagnostics.
Failure recovery happens by expanding unresolved macros into
ExprKind::Err
and allows compilation to continue past the first error
so that rustc
can report more errors than just the original failure.
Name Resolution
Notice that name resolution is involved here: we need to resolve imports and
macro names in the above algorithm. This is done in
rustc_resolve::macros
, which resolves macro paths, validates
those resolutions, and reports various errors (e.g. "not found", "found, but
it's unstable", "expected x, found y"). However, we don't try to resolve
other names yet. This happens later, as we will see in the chapter: Name
Resolution.
Eager Expansion
Eager expansion means we expand the arguments of a macro invocation before the macro invocation itself. This is implemented only for a few special built-in macros that expect literals; expanding arguments first for some of these macro results in a smoother user experience. As an example, consider the following:
macro bar($i: ident) { $i }
macro foo($i: ident) { $i }
foo!(bar!(baz));
A lazy-expansion would expand foo!
first. An eager-expansion would expand
bar!
first.
Eager-expansion is not a generally available feature of Rust. Implementing
eager-expansion more generally would be challenging, so we implement it for a
few special built-in macros for the sake of user-experience. The built-in
macros are implemented in rustc_builtin_macros
, along with some other
early code generation facilities like injection of standard library imports or
generation of test harness. There are some additional helpers for building
AST fragments in rustc_expand::build
. Eager-expansion generally
performs a subset of the things that lazy (normal) expansion does. It is done
by invoking fully_expand_fragment
on only part of a crate (as opposed
to the whole crate, like we normally do).
Other Data Structures
Here are some other notable data structures involved in expansion and integration:
ResolverExpand
- atrait
used to break crate dependencies. This allows the resolver services to be used inrustc_ast
, despiterustc_resolve
and pretty much everything else depending onrustc_ast
.ExtCtxt
/ExpansionData
- holds various intermediate expansion infrastructure data.Annotatable
- a piece of AST that can be an attribute target, almost the same thing asAstFragment
except for types and patterns that can be produced by macros but cannot be annotated with attributes.MacResult
- a "polymorphic" AST fragment, something that can turn into a differentAstFragment
depending on itsAstFragmentKind
(i.e. an item, expression, pattern, etc).
Hygiene and Hierarchies
If you have ever used the C/C++ preprocessor macros, you know that there are some annoying and hard-to-debug gotchas! For example, consider the following C code:
#define DEFINE_FOO struct Bar {int x;}; struct Foo {Bar bar;};
// Then, somewhere else
struct Bar {
...
};
DEFINE_FOO
Most people avoid writing C like this – and for good reason: it doesn't
compile. The struct Bar
defined by the macro clashes names with the struct Bar
defined in the code. Consider also the following example:
#define DO_FOO(x) {\
int y = 0;\
foo(x, y);\
}
// Then elsewhere
int y = 22;
DO_FOO(y);
Do you see the problem? We wanted to generate a call foo(22, 0)
, but instead
we got foo(0, 0)
because the macro defined its own y
!
These are both examples of macro hygiene issues. Hygiene relates to how to handle names defined within a macro. In particular, a hygienic macro system prevents errors due to names introduced within a macro. Rust macros are hygienic in that they do not allow one to write the sorts of bugs above.
At a high level, hygiene within the Rust compiler is accomplished by keeping track of the context where a name is introduced and used. We can then disambiguate names based on that context. Future iterations of the macro system will allow greater control to the macro author to use that context. For example, a macro author may want to introduce a new name to the context where the macro was called. Alternately, the macro author may be defining a variable for use only within the macro (i.e. it should not be visible outside the macro).
The context is attached to AST nodes. All AST nodes generated by macros have
context attached. Additionally, there may be other nodes that have context
attached, such as some desugared syntax (non-macro-expanded nodes are
considered to just have the "root" context, as described below).
Throughout the compiler, we use rustc_span::Span
s to refer to code locations.
This struct also has hygiene information attached to it, as we will see later.
Because macros invocations and definitions can be nested, the syntax context of a node must be a hierarchy. For example, if we expand a macro and there is another macro invocation or definition in the generated output, then the syntax context should reflect the nesting.
However, it turns out that there are actually a few types of context we may want to track for different purposes. Thus, there are not just one but three expansion hierarchies that together comprise the hygiene information for a crate.
All of these hierarchies need some sort of "macro ID" to identify individual
elements in the chain of expansions. This ID is ExpnId
. All macros receive
an integer ID, assigned continuously starting from 0 as we discover new macro
calls. All hierarchies start at ExpnId::root
, which is its own
parent.
The rustc_span::hygiene
crate contains all of the hygiene-related algorithms
(with the exception of some hacks in Resolver::resolve_crate_root
)
and structures related to hygiene and expansion that are kept in global data.
The actual hierarchies are stored in HygieneData
. This is a global
piece of data containing hygiene and expansion info that can be accessed from
any Ident
without any context.
The Expansion Order Hierarchy
The first hierarchy tracks the order of expansions, i.e., when a macro invocation is in the output of another macro.
Here, the children in the hierarchy will be the "innermost" tokens. The
ExpnData
struct itself contains a subset of properties from both macro
definition and macro call available through global data.
ExpnData::parent
tracks the child-to-parent link in this hierarchy.
For example:
macro_rules! foo { () => { println!(); } }
fn main() { foo!(); }
In this code, the AST nodes that are finally generated would have hierarchy
root -> id(foo) -> id(println)
.
The Macro Definition Hierarchy
The second hierarchy tracks the order of macro definitions, i.e., when we are expanding one macro another macro definition is revealed in its output. This one is a bit tricky and more complex than the other two hierarchies.
SyntaxContext
represents a whole chain in this hierarchy via an ID.
SyntaxContextData
contains data associated with the given
SyntaxContext
; mostly it is a cache for results of filtering that chain in
different ways. SyntaxContextData::parent
is the child-to-parent
link here, and SyntaxContextData::outer_expns
are individual
elements in the chain. The "chaining-operator" is
SyntaxContext::apply_mark
in compiler code.
A Span
, mentioned above, is actually just a compact representation of
a code location and SyntaxContext
. Likewise, an Ident
is just an interned
Symbol
+ Span
(i.e. an interned string + hygiene data).
For built-in macros, we use the context:
SyntaxContext::empty().apply_mark(expn_id)
, and such macros are
considered to be defined at the hierarchy root. We do the same for proc macro
s because we haven't implemented cross-crate hygiene yet.
If the token had context X
before being produced by a macro then after being
produced by the macro it has context X -> macro_id
. Here are some examples:
Example 0:
macro m() { ident }
m!();
Here ident
which initially has context SyntaxContext::root
has
context ROOT -> id(m)
after it's produced by m
.
Example 1:
macro m() { macro n() { ident } }
m!();
n!();
In this example the ident
has context ROOT
initially, then ROOT -> id(m)
after the first expansion, then ROOT -> id(m) -> id(n)
.
Example 2:
Note that these chains are not entirely determined by their last element, in
other words ExpnId
is not isomorphic to SyntaxContext
.
macro m($i: ident) { macro n() { ($i, bar) } }
m!(foo);
After all expansions, foo
has context ROOT -> id(n)
and bar
has context
ROOT -> id(m) -> id(n)
.
Currently this hierarchy for tracking macro definitions is subject to the so-called "context transplantation hack". Modern (i.e. experimental) macros have stronger hygiene than the legacy "Macros By Example" (MBE) system which can result in weird interactions between the two. The hack is intended to make things "just work" for now.
The Call-site Hierarchy
The third and final hierarchy tracks the location of macro invocations.
In this hierarchy ExpnData::call_site
is the child -> parent
link.
Here is an example:
macro bar($i: ident) { $i }
macro foo($i: ident) { $i }
foo!(bar!(baz));
For the baz
AST node in the final output, the expansion-order hierarchy is
ROOT -> id(foo) -> id(bar) -> baz
, while the call-site hierarchy is ROOT -> baz
.
Macro Backtraces
Macro backtraces are implemented in rustc_span
using the hygiene machinery
in rustc_span::hygiene
.
Producing Macro Output
Above, we saw how the output of a macro is integrated into the AST for a crate, and we also saw how the hygiene data for a crate is generated. But how do we actually produce the output of a macro? It depends on the type of macro.
There are two types of macros in Rust:
macro_rules!
macros (a.k.a. "Macros By Example" (MBE)), and,- procedural macros (proc macros); including custom derives.
During the parsing phase, the normal Rust parser will set aside the contents of macros and their invocations. Later, macros are expanded using these portions of the code.
Some important data structures/interfaces here:
SyntaxExtension
- a lowered macro representation, contains its expander function, which transforms aTokenStream
or AST into anotherTokenStream
or AST + some additional data like stability, or a list of unstable features allowed inside the macro.SyntaxExtensionKind
- expander functions may have several different signatures (take one token stream, or two, or a piece of AST, etc). This is anenum
that lists them.BangProcMacro
/TTMacroExpander
/AttrProcMacro
/MultiItemModifier
-trait
s representing the expander function signatures.
Macros By Example
MBEs have their own parser distinct from the Rust parser. When macros are
expanded, we may invoke the MBE parser to parse and expand a macro. The
MBE parser, in turn, may call the Rust parser when it needs to bind a
metavariable (e.g. $my_expr
) while parsing the contents of a macro
invocation. The code for macro expansion is in
compiler/rustc_expand/src/mbe/
.
Example
macro_rules! printer {
(print $mvar:ident) => {
println!("{}", $mvar);
};
(print twice $mvar:ident) => {
println!("{}", $mvar);
println!("{}", $mvar);
};
}
Here $mvar
is called a metavariable. Unlike normal variables, rather than
binding to a value at runtime, a metavariable binds at compile time to a
tree of tokens. A token is a single "unit" of the grammar, such as an
identifier (e.g. foo
) or punctuation (e.g. =>
). There are also other
special tokens, such as EOF
, which its self indicates that there are no more
tokens. There are token trees resulting from the paired parentheses-like
characters ((
...)
, [
...]
, and {
...}
) – they include the open and
close and all the tokens in between (Rust requires that parentheses-like
characters be balanced). Having macro expansion operate on token streams
rather than the raw bytes of a source-file abstracts away a lot of complexity.
The macro expander (and much of the rest of the compiler) doesn't consider
the exact line and column of some syntactic construct in the code; it considers
which constructs are used in the code. Using tokens allows us to care about
what without worrying about where. For more information about tokens, see
the Parsing chapter of this book.
printer!(print foo); // `foo` is a variable
The process of expanding the macro invocation into the syntax tree
println!("{}", foo)
and then expanding the syntax tree into a call to
Display::fmt
is one common example of macro expansion.
The MBE parser
There are two parts to MBE expansion done by the macro parser:
- parsing the definition, and,
- parsing the invocations.
We think of the MBE parser as a nondeterministic finite automaton (NFA) based
regex parser since it uses an algorithm similar in spirit to the Earley
parsing algorithm. The macro
parser is defined in
compiler/rustc_expand/src/mbe/macro_parser.rs
.
The interface of the macro parser is as follows (this is slightly simplified):
fn parse_tt(
&mut self,
parser: &mut Cow<'_, Parser<'_>>,
matcher: &[MatcherLoc]
) -> ParseResult
We use these items in macro parser:
- a
parser
variable is a reference to the state of a normal Rust parser, including the token stream and parsing session. The token stream is what we are about to ask the MBE parser to parse. We will consume the raw stream of tokens and output a binding of metavariables to corresponding token trees. The parsing session can be used to report parser errors. - a
matcher
variable is a sequence ofMatcherLoc
s that we want to match the token stream against. They're converted from token trees before matching.
In the analogy of a regex parser, the token stream is the input and we are
matching it against the pattern defined by matcher. Using our examples, the
token stream could be the stream of tokens containing the inside of the example
invocation print foo
, while matcher might be the sequence of token (trees)
print $mvar:ident
.
The output of the parser is a ParseResult
, which indicates which of
three cases has occurred:
- Success: the token stream matches the given matcher and we have produced a binding from metavariables to the corresponding token trees.
- Failure: the token stream does not match matcher and results in an error message such as "No rule expected token ...".
- Error: some fatal error has occurred in the parser. For example, this happens if there is more than one pattern match, since that indicates the macro is ambiguous.
The full interface is defined here.
The macro parser does pretty much exactly the same as a normal regex parser
with one exception: in order to parse different types of metavariables, such as
ident
, block
, expr
, etc., the macro parser must call back to the normal
Rust parser. Both the definition and invocation of macros are parsed using
the parser in a process which is non-intuitively self-referential.
The code to parse macro definitions is in
compiler/rustc_expand/src/mbe/macro_rules.rs
. It defines the
pattern for matching a macro definition as $( $lhs:tt => $rhs:tt );+
. In
other words, a macro_rules
definition should have in its body at least one
occurrence of a token tree followed by =>
followed by another token tree.
When the compiler comes to a macro_rules
definition, it uses this pattern to
match the two token trees per the rules of the definition of the macro, thereby
utilizing the macro parser itself. In our example definition, the
metavariable $lhs
would match the patterns of both arms: (print $mvar:ident)
and (print twice $mvar:ident)
. And $rhs
would match the
bodies of both arms: { println!("{}", $mvar); }
and { println!("{}", $mvar); println!("{}", $mvar); }
. The parser keeps this knowledge around for when it
needs to expand a macro invocation.
When the compiler comes to a macro invocation, it parses that invocation using
a NFA-based macro parser described above. However, the matcher variable
used is the first token tree ($lhs
) extracted from the arms of the macro
definition. Using our example, we would try to match the token stream print foo
from the invocation against the matchers print $mvar:ident
and print twice $mvar:ident
that we previously extracted from the definition. The
algorithm is exactly the same, but when the macro parser comes to a place in the
current matcher where it needs to match a non-terminal (e.g. $mvar:ident
),
it calls back to the normal Rust parser to get the contents of that
non-terminal. In this case, the Rust parser would look for an ident
token,
which it finds (foo
) and returns to the macro parser. Then, the macro parser
proceeds in parsing as normal. Also, note that exactly one of the matchers from
the various arms should match the invocation; if there is more than one match,
the parse is ambiguous, while if there are no matches at all, there is a syntax
error.
For more information about the macro parser's implementation, see the comments
in compiler/rustc_expand/src/mbe/macro_parser.rs
.
Procedural Macros
Procedural macros are also expanded during parsing. However, rather than having a parser in the compiler, proc macros are implemented as custom, third-party crates. The compiler will compile the proc macro crate and specially annotated functions in them (i.e. the proc macro itself), passing them a stream of tokens. A proc macro can then transform the token stream and output a new token stream, which is synthesized into the AST.
The token stream type used by proc macros is stable, so rustc
does not
use it internally. The compiler's (unstable) token stream is defined in
rustc_ast::tokenstream::TokenStream
. This is converted into the
stable proc_macro::TokenStream
and back in
rustc_expand::proc_macro
and rustc_expand::proc_macro_server
.
Since the Rust ABI is currently unstable, we use the C ABI for this conversion.
Custom Derive
Custom derives are a special type of proc macro.
Macros By Example and Macros 2.0
There is an legacy and mostly undocumented effort to improve the MBE system by giving it more hygiene-related features, better scoping and visibility rules, etc. Internally this uses the same machinery as today's MBEs with some additional syntactic sugar and are allowed to be in namespaces.
Name resolution
In the previous chapters, we saw how the Abstract Syntax Tree (AST
)
is built with all macros expanded. We saw how doing that requires doing some
name resolution to resolve imports and macro names. In this chapter, we show
how this is actually done and more.
In fact, we don't do full name resolution during macro expansion -- we only
resolve imports and macros at that time. This is required to know what to even
expand. Later, after we have the whole AST, we do full name resolution to
resolve all names in the crate. This happens in rustc_resolve::late
.
Unlike during macro expansion, in this late expansion, we only need to try to
resolve a name once, since no new names can be added. If we fail to resolve a
name, then it is a compiler error.
Name resolution is complex. There are different namespaces (e.g. macros, values, types, lifetimes), and names may be valid at different (nested) scopes. Also, different types of names can fail resolution differently, and failures can happen differently at different scopes. For example, in a module scope, failure means no unexpanded macros and no unresolved glob imports in that module. On the other hand, in a function body scope, failure requires that a name be absent from the block we are in, all outer scopes, and the global scope.
Basics
In our programs we refer to variables, types, functions, etc, by giving them a name. These names are not always unique. For example, take this valid Rust program:
#![allow(unused)] fn main() { type x = u32; let x: x = 1; let y: x = 2; }
How do we know on line 3 whether x
is a type (u32
) or a value (1)? These
conflicts are resolved during name resolution. In this specific case, name
resolution defines that type names and variable names live in separate
namespaces and therefore can co-exist.
The name resolution in Rust is a two-phase process. In the first phase, which runs
during macro
expansion, we build a tree of modules and resolve imports. Macro
expansion and name resolution communicate with each other via the
ResolverAstLoweringExt
trait.
The input to the second phase is the syntax tree, produced by parsing input
files and expanding macros
. This phase produces links from all the names in the
source to relevant places where the name was introduced. It also generates
helpful error messages, like typo suggestions, traits to import or lints about
unused items.
A successful run of the second phase (Resolver::resolve_crate
) creates kind
of an index the rest of the compilation may use to ask about the present names
(through the hir::lowering::Resolver
interface).
The name resolution lives in the rustc_resolve
crate, with the bulk in
lib.rs
and some helpers or symbol-type specific logic in the other modules.
Namespaces
Different kind of symbols live in different namespaces ‒ e.g. types don't clash with variables. This usually doesn't happen, because variables start with lower-case letter while types with upper-case one, but this is only a convention. This is legal Rust code that will compile (with warnings):
#![allow(unused)] fn main() { type x = u32; let x: x = 1; let y: x = 2; // See? x is still a type here. }
To cope with this, and with slightly different scoping rules for these namespaces, the resolver keeps them separated and builds separate structures for them.
In other words, when the code talks about namespaces, it doesn't mean the module hierarchy, it's types vs. values vs. macros.
Scopes and ribs
A name is visible only in certain area in the source code. This forms a hierarchical structure, but not necessarily a simple one ‒ if one scope is part of another, it doesn't mean a name visible in the outer scope is also visible in the inner scope, or that it refers to the same thing.
To cope with that, the compiler introduces the concept of Rib
s. This is
an abstraction of a scope. Every time the set of visible names potentially changes,
a new Rib
is pushed onto a stack. The places where this can happen include for
example:
- The obvious places ‒ curly braces enclosing a block, function boundaries, modules.
- Introducing a
let
binding ‒ this can shadow another binding with the same name. - Macro expansion border ‒ to cope with macro hygiene.
When searching for a name, the stack of ribs
is traversed from the innermost
outwards. This helps to find the closest meaning of the name (the one not
shadowed by anything else). The transition to outer Rib
may also affect
what names are usable ‒ if there are nested functions (not closures),
the inner one can't access parameters and local bindings of the outer one,
even though they should be visible by ordinary scoping rules. An example:
#![allow(unused)] fn main() { fn do_something<T: Default>(val: T) { // <- New rib in both types and values (1) // `val` is accessible, as is the helper function // `T` is accessible let helper = || { // New rib on `helper` (2) and another on the block (3) // `val` is accessible here }; // End of (3) // `val` is accessible, `helper` variable shadows `helper` function fn helper() { // <- New rib in both types and values (4) // `val` is not accessible here, (4) is not transparent for locals // `T` is not accessible here } // End of (4) let val = T::default(); // New rib (5) // `val` is the variable, not the parameter here } // End of (5), (2) and (1) }
Because the rules for different namespaces are a bit different, each namespace
has its own independent Rib
stack that is constructed in parallel to the others.
In addition, there's also a Rib
stack for local labels (e.g. names of loops or
blocks), which isn't a full namespace in its own right.
Overall strategy
To perform the name resolution of the whole crate, the syntax tree is traversed
top-down and every encountered name is resolved. This works for most kinds of
names, because at the point of use of a name it is already introduced in the Rib
hierarchy.
There are some exceptions to this. Items are bit tricky, because they can be
used even before encountered ‒ therefore every block needs to be first scanned
for items to fill in its Rib
.
Other, even more problematic ones, are imports which need recursive fixed-point resolution and macros, that need to be resolved and expanded before the rest of the code can be processed.
Therefore, the resolution is performed in multiple stages.
Speculative crate loading
To give useful errors, rustc suggests importing paths into scope if they're not found. How does it do this? It looks through every module of every crate and looks for possible matches. This even includes crates that haven't yet been loaded!
Eagerly loading crates to include import suggestions that haven't yet been
loaded is called speculative crate loading, because any errors it encounters
shouldn't be reported: rustc_resolve
decided to load them, not the user. The function
that does this is lookup_import_candidates
and lives in
rustc_resolve::diagnostics
.
To tell the difference between speculative loads and loads initiated by the
user, rustc_resolve
passes around a record_used
parameter, which is false
when
the load is speculative.
TODO: #16
This is a result of the first pass of learning the code. It is definitely incomplete and not detailed enough. It also might be inaccurate in places. Still, it probably provides useful first guidepost to what happens in there.
- What exactly does it link to and how is that published and consumed by following stages of compilation?
- Who calls it and how it is actually used.
- Is it a pass and then the result is only used, or can it be computed incrementally?
- The overall strategy description is a bit vague.
- Where does the name
Rib
come from? - Does this thing have its own tests, or is it tested only as part of some e2e testing?
Attributes
Attributes come in two types: inert (or built-in) and active (non-builtin).
Builtin/inert attributes
These attributes are defined in the compiler itself, in
compiler/rustc_feature/src/builtin_attrs.rs
.
Examples include #[allow]
and #[macro_use]
.
These attributes have several important characteristics:
- They are always in scope, and do not participate in typical path-based resolution.
- They cannot be renamed. For example,
use allow as foo
will compile, but writing#[foo]
will produce an error. - They are 'inert', meaning they are left as-is by the macro expansion code.
As a result, any behavior comes as a result of the compiler explicitly checking for their presence.
For example, lint-related code explicitly checks for
#[allow]
,#[warn]
,#[deny]
, and#[forbid]
, rather than the behavior coming from the expansion of the attributes themselves.
'Non-builtin'/'active' attributes
These attributes are defined by a crate - either the standard library, or a proc-macro crate.
Important: Many non-builtin attributes, such as #[derive]
, are still considered part of the
core Rust language. However, they are not called 'builtin attributes', since they have a
corresponding definition in the standard library.
Definitions of non-builtin attributes take two forms:
- Proc-macro attributes, defined via a function annotated with
#[proc_macro_attribute]
in a proc-macro crate. - AST-based attributes, defined in the standard library. These attributes have special 'stub'
macros defined in places like
library/core/src/macros/mod.rs
.
These definitions exist to allow the macros to participate in typical path-based resolution - they
can be imported, re-exported, and renamed just like any other item definition. However, the body of
the definition is empty. Instead, the macro is annotated with the #[rustc_builtin_macro]
attribute, which tells the compiler to run a corresponding function in rustc_builtin_macros
.
All non-builtin attributes have the following characteristics:
- Like all other definitions (e.g. structs), they must be brought into scope via an import.
Many standard library attributes are included in the prelude - this is why writing
#[derive]
works without an import. - They participate in macro expansion. The implementation of the macro may leave the attribute target unchanged, modify the target, produce new AST nodes, or remove the target entirely.
The #[test]
attribute
- Step 1: Re-Exporting
- Step 2: Harness Generation
- Step 3: Test Object Generation
- Inspecting the generated code
Many Rust programmers rely on a built-in attribute called #[test]
. All
you have to do is mark a function and include some asserts like so:
#[test]
fn my_test() {
assert!(2+2 == 4);
}
When this program is compiled using rustc --test
or cargo test
, it will
produce an executable that can run this, and any other test function. This
method of testing allows tests to live alongside code in an organic way. You
can even put tests inside private modules:
mod my_priv_mod {
fn my_priv_func() -> bool {}
#[test]
fn test_priv_func() {
assert!(my_priv_func());
}
}
Private items can thus be easily tested without worrying about how to expose
them to any sort of external testing apparatus. This is key to the
ergonomics of testing in Rust. Semantically, however, it's rather odd.
How does any sort of main
function invoke these tests if they're not visible?
What exactly is rustc --test
doing?
#[test]
is implemented as a syntactic transformation inside the compiler's
rustc_ast
. Essentially, it's a fancy macro
that
rewrites the crate in 3 steps:
Step 1: Re-Exporting
As mentioned earlier, tests can exist inside private modules, so we need a
way of exposing them to the main function, without breaking any existing
code. To that end, rustc_ast
will create local modules called
__test_reexports
that recursively reexport tests. This expansion translates
the above example into:
mod my_priv_mod {
fn my_priv_func() -> bool {}
pub fn test_priv_func() {
assert!(my_priv_func());
}
pub mod __test_reexports {
pub use super::test_priv_func;
}
}
Now, our test can be accessed as
my_priv_mod::__test_reexports::test_priv_func
. For deeper module
structures, __test_reexports
will reexport modules that contain tests, so a
test at a::b::my_test
becomes
a::__test_reexports::b::__test_reexports::my_test
. While this process seems
pretty safe, what happens if there is an existing __test_reexports
module?
The answer: nothing.
To explain, we need to understand how Rust's Abstract Syntax Tree
represents identifiers. The name of every function, variable, module,
etc. is not stored as a string, but rather as an opaque Symbol which
is essentially an ID number for each identifier. The compiler keeps a separate
hashtable that allows us to recover the human-readable name of a Symbol when
necessary (such as when printing a syntax error). When the compiler generates
the __test_reexports
module, it generates a new Symbol for the
identifier, so while the compiler-generated __test_reexports
may share a name
with your hand-written one, it will not share a Symbol. This
technique prevents name collision during code generation and is the foundation
of Rust's macro
hygiene.
Step 2: Harness Generation
Now that our tests are accessible from the root of our crate, we need to do
something with them using rustc_ast
generates a module like so:
#[main]
pub fn main() {
extern crate test;
test::test_main_static(&[&path::to::test1, /*...*/]);
}
Here path::to::test1
is a constant of type test::TestDescAndFn
.
While this transformation is simple, it gives us a lot of insight into how
tests are actually run. The tests are aggregated into an array and passed to
a test runner called test_main_static
. We'll come back to exactly what
TestDescAndFn
is, but for now, the key takeaway is that there is a crate
called test
that is part of Rust core, that implements all of the
runtime for testing. test
's interface is unstable, so the only stable way
to interact with it is through the #[test]
macro.
Step 3: Test Object Generation
If you've written tests in Rust before, you may be familiar with some of the
optional attributes available on test functions. For example, a test can be
annotated with #[should_panic]
if we expect the test to cause a panic. It
looks something like this:
#[test]
#[should_panic]
fn foo() {
panic!("intentional");
}
This means our tests are more than just simple functions, they have
configuration information as well. test
encodes this configuration data into
a struct
called TestDesc
. For each test function in a crate,
rustc_ast
will parse its attributes and generate a TestDesc
instance. It then combines the TestDesc
and test function into the
predictably named TestDescAndFn
struct
, that test_main_static
operates on.
For a given test, the generated TestDescAndFn
instance looks like so:
self::test::TestDescAndFn{
desc: self::test::TestDesc{
name: self::test::StaticTestName("foo"),
ignore: false,
should_panic: self::test::ShouldPanic::Yes,
allow_fail: false,
},
testfn: self::test::StaticTestFn(||
self::test::assert_test_result(::crate::__test_reexports::foo())),
}
Once we've constructed an array of these test objects, they're passed to the test runner via the harness generated in Step 2.
Inspecting the generated code
On nightly
rustc
, there's an unstable flag called unpretty
that you can use
to print out the module source after macro
expansion:
$ rustc my_mod.rs -Z unpretty=hir
Panicking in rust
Step 1: Invocation of the panic!
macro.
There are actually two panic macros - one defined in core
, and one defined in std
.
This is due to the fact that code in core
can panic. core
is built before std
,
but we want panics to use the same machinery at runtime, whether they originate in core
or std
.
core definition of panic!
The core
panic!
macro eventually makes the following call (in library/core/src/panicking.rs
):
#![allow(unused)] fn main() { // NOTE This function never crosses the FFI boundary; it's a Rust-to-Rust call extern "Rust" { #[lang = "panic_impl"] fn panic_impl(pi: &PanicInfo<'_>) -> !; } let pi = PanicInfo::internal_constructor(Some(&fmt), location); unsafe { panic_impl(&pi) } }
Actually resolving this goes through several layers of indirection:
-
In
compiler/rustc_middle/src/middle/weak_lang_items.rs
,panic_impl
is declared as 'weak lang item', with the symbolrust_begin_unwind
. This is used inrustc_hir_analysis/src/collect.rs
to set the actual symbol name torust_begin_unwind
.Note that
panic_impl
is declared in anextern "Rust"
block, which means that core will attempt to call a foreign symbol calledrust_begin_unwind
(to be resolved at link time) -
In
library/std/src/panicking.rs
, we have this definition:
#![allow(unused)] fn main() { /// Entry point of panic from the core crate. #[cfg(not(test))] #[panic_handler] #[unwind(allowed)] pub fn begin_panic_handler(info: &PanicInfo<'_>) -> ! { ... } }
The special panic_handler
attribute is resolved via compiler/rustc_middle/src/middle/lang_items
.
The extract
function converts the panic_handler
attribute to a panic_impl
lang item.
Now, we have a matching panic_handler
lang item in the std
. This function goes
through the same process as the extern { fn panic_impl }
definition in core
, ending
up with a symbol name of rust_begin_unwind
. At link time, the symbol reference in core
will be resolved to the definition of std
(the function called begin_panic_handler
in the
Rust source).
Thus, control flow will pass from core to std at runtime. This allows panics from core
to go through the same infrastructure that other panics use (panic hooks, unwinding, etc)
std implementation of panic!
This is where the actual panic-related logic begins. In library/std/src/panicking.rs
,
control passes to rust_panic_with_hook
. This method is responsible
for invoking the global panic hook, and checking for double panics. Finally,
we call __rust_start_panic
, which is provided by the panic runtime.
The call to __rust_start_panic
is very weird - it is passed a *mut &mut dyn PanicPayload
,
converted to an usize
. Let's break this type down:
-
PanicPayload
is an internal trait. It is implemented forPanicPayload
(a wrapper around the user-supplied payload type), and has a methodfn take_box(&mut self) -> *mut (dyn Any + Send)
. This method takes the user-provided payload (T: Any + Send
), boxes it, and converts the box to a raw pointer. -
When we call
__rust_start_panic
, we have an&mut dyn PanicPayload
. However, this is a fat pointer (twice the size of ausize
). To pass this to the panic runtime across an FFI boundary, we take a mutable reference to this mutable reference (&mut &mut dyn PanicPayload
), and convert it to a raw pointer (*mut &mut dyn PanicPayload
). The outer raw pointer is a thin pointer, since it points to aSized
type (a mutable reference). Therefore, we can convert this thin pointer into ausize
, which is suitable for passing across an FFI boundary.
Finally, we call __rust_start_panic
with this usize
. We have now entered the panic runtime.
Step 2: The panic runtime
Rust provides two panic runtimes: panic_abort
and panic_unwind
. The user chooses
between them at build time via their Cargo.toml
panic_abort
is extremely simple: its implementation of __rust_start_panic
just aborts,
as you would expect.
panic_unwind
is the more interesting case.
In its implementation of __rust_start_panic
, we take the usize
, convert
it back to a *mut &mut dyn PanicPayload
, dereference it, and call take_box
on the &mut dyn PanicPayload
. At this point, we have a raw pointer to the payload
itself (a *mut (dyn Send + Any)
): that is, a raw pointer to the actual value
provided by the user who called panic!
.
At this point, the platform-independent code ends. We now call into
platform-specific unwinding logic (e.g unwind
). This code is
responsible for unwinding the stack, running any 'landing pads' associated
with each frame (currently, running destructors), and transferring control
to the catch_unwind
frame.
Note that all panics either abort the process or get caught by some call to catch_unwind
.
In particular, in std's runtime service,
the call to the user-provided main
function is wrapped in catch_unwind
.
AST Validation
AST validation is a separate AST pass that visits each item in the tree and performs simple checks. This pass doesn't perform any complex analysis, type checking or name resolution.
Before performing any validation, the compiler first expands the macros. Then this pass performs validations to check that each AST item is in the correct state. And when this pass is done, the compiler runs the crate resolution pass.
Validations
Validations are defined in AstValidator
type, which
itself is located in rustc_ast_passes
crate. This
type implements various simple checks which emit errors
when certain language rules are broken.
In addition, AstValidator
implements Visitor
trait
that defines how to visit AST items (which can be functions,
traits, enums, etc).
For each item, visitor performs specific checks. For
example, when visiting a function declaration,
AstValidator
checks that the function has:
- no more than
u16::MAX
parameters; - c-variadic argument goes the last in the declaration;
- documentation comments aren't applied to function parameters;
- and other validations.
Feature Gate Checking
TODO: this chapter #1158
Lang items
The compiler has certain pluggable operations; that is, functionality that isn't hard-coded into
the language, but is implemented in libraries, with a special marker to tell the compiler it
exists. The marker is the attribute #[lang = "..."]
, and there are various different values of
...
, i.e. various different 'lang items'.
Many such lang items can be implemented only in one sensible way, such as add
(trait core::ops::Add
) or future_trait
(trait core::future::Future
). Others can be overridden to
achieve some specific goals; for example, you can control your binary's entrypoint.
Features provided by lang items include:
- overloadable operators via traits: the traits corresponding to the
==
,<
, dereference (*
),+
, etc. operators are all marked with lang items; those specific four areeq
,ord
,deref
, andadd
respectively. - panicking and stack unwinding; the
eh_personality
,panic
andpanic_bounds_checks
lang items. - the traits in
std::marker
used to indicate properties of types used by the compiler; lang itemssend
,sync
andcopy
. - the special marker types used for variance indicators found in
core::marker
; lang itemphantom_data
.
Lang items are loaded lazily by the compiler; e.g. if one never uses Box
then there is no need to define functions for exchange_malloc
and
box_free
. rustc
will emit an error when an item is needed but not found
in the current crate or any that it depends on.
Most lang items are defined by the core
library, but if you're trying to build an
executable with #![no_std]
, you'll still need to define a few lang items that are
usually provided by std
.
Retrieving a language item
You can retrieve lang items by calling tcx.lang_items()
.
Here's a small example of retrieving the trait Sized {}
language item:
#![allow(unused)] fn main() { // Note that in case of `#![no_core]`, the trait is not available. if let Some(sized_trait_def_id) = tcx.lang_items().sized_trait() { // do something with `sized_trait_def_id` } }
Note that sized_trait()
returns an Option
, not the DefId
itself.
That's because language items are defined in the standard library, so if someone compiles with
#![no_core]
(or for some lang items, #![no_std]
), the lang item may not be present.
You can either:
- Give a hard error if the lang item is necessary to continue (don't panic, since this can happen in user code).
- Proceed with limited functionality, by just omitting whatever you were going to do with the
DefId
.
List of all language items
You can find language items in the following places:
- An exhaustive reference in the compiler documentation:
rustc_hir::LangItem
- An auto-generated list with source locations by using ripgrep:
rg '#\[.*lang =' library/
Note that language items are explicitly unstable and may change in any new release.
The HIR
The HIR – "High-Level Intermediate Representation" – is the primary IR used
in most of rustc. It is a compiler-friendly representation of the abstract
syntax tree (AST) that is generated after parsing, macro expansion, and name
resolution (see Lowering for how the HIR is created).
Many parts of HIR resemble Rust surface syntax quite closely, with
the exception that some of Rust's expression forms have been desugared away.
For example, for
loops are converted into a loop
and do not appear in
the HIR. This makes HIR more amenable to analysis than a normal AST.
This chapter covers the main concepts of the HIR.
You can view the HIR representation of your code by passing the
-Z unpretty=hir-tree
flag to rustc:
cargo rustc -- -Z unpretty=hir-tree
You can also use the -Z unpretty=hir
option to generate a HIR
that is closer to the original source code expression:
cargo rustc -- -Z unpretty=hir
Out-of-band storage and the Crate
type
The top-level data-structure in the HIR is the Crate
, which stores
the contents of the crate currently being compiled (we only ever
construct HIR for the current crate). Whereas in the AST the crate
data structure basically just contains the root module, the HIR
Crate
structure contains a number of maps and other things that
serve to organize the content of the crate for easier access.
For example, the contents of individual items (e.g. modules,
functions, traits, impls, etc) in the HIR are not immediately
accessible in the parents. So, for example, if there is a module item
foo
containing a function bar()
:
#![allow(unused)] fn main() { mod foo { fn bar() { } } }
then in the HIR the representation of module foo
(the Mod
struct) would only have the ItemId
I
of bar()
. To get the
details of the function bar()
, we would lookup I
in the
items
map.
One nice result from this representation is that one can iterate over all items in the crate by iterating over the key-value pairs in these maps (without the need to trawl through the whole HIR). There are similar maps for things like trait items and impl items, as well as "bodies" (explained below).
The other reason to set up the representation this way is for better
integration with incremental compilation. This way, if you gain access
to an &rustc_hir::Item
(e.g. for the mod foo
), you do not immediately
gain access to the contents of the function bar()
. Instead, you only
gain access to the id for bar()
, and you must invoke some
function to lookup the contents of bar()
given its id; this gives
the compiler a chance to observe that you accessed the data for
bar()
, and then record the dependency.
Identifiers in the HIR
The HIR uses a bunch of different identifiers that coexist and serve different purposes.
-
A
DefId
, as the name suggests, identifies a particular definition, or top-level item, in a given crate. It is composed of two parts: aCrateNum
which identifies the crate the definition comes from, and aDefIndex
which identifies the definition within the crate. UnlikeHirId
s, there isn't aDefId
for every expression, which makes them more stable across compilations. -
A
LocalDefId
is basically aDefId
that is known to come from the current crate. This allows us to drop theCrateNum
part, and use the type system to ensure that only local definitions are passed to functions that expect a local definition. -
A
HirId
uniquely identifies a node in the HIR of the current crate. It is composed of two parts: anowner
and alocal_id
that is unique within theowner
. This combination makes for more stable values which are helpful for incremental compilation. UnlikeDefId
s, aHirId
can refer to [fine-grained entities][Node] like expressions, but stays local to the current crate. -
A
BodyId
identifies a HIRBody
in the current crate. It is currently only a wrapper around aHirId
. For more info about HIR bodies, please refer to the HIR chapter.
These identifiers can be converted into one another through the HIR map.
The HIR Map
Most of the time when you are working with the HIR, you will do so via
the HIR Map, accessible in the tcx via tcx.hir()
(and defined in
the hir::map
module). The HIR map contains a number of methods to
convert between IDs of various kinds and to lookup data associated
with a HIR node.
For example, if you have a LocalDefId
, and you would like to convert it
to a HirId
, you can use tcx.hir().local_def_id_to_hir_id(def_id)
.
You need a LocalDefId
, rather than a DefId
, since only local items have HIR nodes.
Similarly, you can use tcx.hir().find(n)
to lookup the node for a
HirId
. This returns a Option<Node<'hir>>
, where Node
is an enum
defined in the map. By matching on this, you can find out what sort of
node the HirId
referred to and also get a pointer to the data
itself. Often, you know what sort of node n
is – e.g. if you know
that n
must be some HIR expression, you can do
tcx.hir().expect_expr(n)
, which will extract and return the
&hir::Expr
, panicking if n
is not in fact an expression.
Finally, you can use the HIR map to find the parents of nodes, via
calls like tcx.hir().get_parent(n)
.
HIR Bodies
A rustc_hir::Body
represents some kind of executable code, such as the body
of a function/closure or the definition of a constant. Bodies are
associated with an owner, which is typically some kind of item
(e.g. an fn()
or const
), but could also be a closure expression
(e.g. |x, y| x + y
). You can use the HIR map to find the body
associated with a given def-id (maybe_body_owned_by
) or to find
the owner of a body (body_owner_def_id
).
AST lowering
The AST lowering step converts AST to HIR. This means many structures are removed if they are irrelevant for type analysis or similar syntax agnostic analyses. Examples of such structures include but are not limited to
- Parenthesis
- Removed without replacement, the tree structure makes order explicit
for
loops andwhile (let)
loops- Converted to
loop
+match
and somelet
bindings
- Converted to
if let
- Converted to
match
- Converted to
- Universal
impl Trait
- Converted to generic arguments (but with some flags, to know that the user didn't write them)
- Existential
impl Trait
- Converted to a virtual
existential type
declaration
- Converted to a virtual
Lowering needs to uphold several invariants in order to not trigger the
sanity checks in compiler/rustc_passes/src/hir_id_validator.rs
:
- A
HirId
must be used if created. So if you use thelower_node_id
, you must use the resultingNodeId
orHirId
(either is fine, since anyNodeId
s in theHIR
are checked for existingHirId
s) - Lowering a
HirId
must be done in the scope of the owning item. This means you need to usewith_hir_id_owner
if you are creating parts of an item other than the one being currently lowered. This happens for example during the lowering of existentialimpl Trait
- A
NodeId
that will be placed into a HIR structure must be lowered, even if itsHirId
is unused. Callinglet _ = self.lower_node_id(node_id);
is perfectly legitimate. - If you are creating new nodes that didn't exist in the
AST
, you must create new ids for them. This is done by calling thenext_id
method, which produces both a newNodeId
as well as automatically lowering it for you so you also get theHirId
.
If you are creating new DefId
s, since each DefId
needs to have a
corresponding NodeId
, it is advisable to add these NodeId
s to the
AST
so you don't have to generate new ones during lowering. This has
the advantage of creating a way to find the DefId
of something via its
NodeId
. If lowering needs this DefId
in multiple places, you can't
generate a new NodeId
in all those places because you'd also get a new
DefId
then. With a NodeId
from the AST
this is not an issue.
Having the NodeId
also allows the DefCollector
to generate the DefId
s
instead of lowering having to do it on the fly. Centralizing the DefId
generation in one place makes it easier to refactor and reason about.
HIR Debugging
Use the -Z unpretty=hir
flag to produce a human-readable representation of the HIR.
For cargo projects this can be done with cargo rustc -- -Z unpretty=hir
.
This output is useful when you need to see at a glance how your code was desugared and transformed
during AST lowering.
For a full Debug
dump of the data in the HIR, use the -Z unpretty=hir-tree
flag.
This may be useful when you need to see the full structure of the HIR from the perspective of the
compiler.
If you are trying to correlate NodeId
s or DefId
s with source code, the
-Z unpretty=expanded,identified
flag may be useful.
TODO: anything else? #1159
The THIR
The THIR ("Typed High-Level Intermediate Representation"), previously called HAIR for "High-Level Abstract IR", is another IR used by rustc that is generated after type checking. It is (as of January 2024) used for MIR construction, exhaustiveness checking, and unsafety checking.
As the name might suggest, the THIR is a lowered version of the HIR where all the types have been filled in, which is possible after type checking has completed. But it has some other interesting features that distinguish it from the HIR:
-
Like the MIR, the THIR only represents bodies, i.e. "executable code"; this includes function bodies, but also
const
initializers, for example. Specifically, all body owners have THIR created. Consequently, the THIR has no representation for items likestruct
s ortrait
s. -
Each body of THIR is only stored temporarily and is dropped as soon as it's no longer needed, as opposed to being stored until the end of the compilation process (which is what is done with the HIR).
-
Besides making the types of all nodes available, the THIR also has additional desugaring compared to the HIR. For example, automatic references and dereferences are made explicit, and method calls and overloaded operators are converted into plain function calls. Destruction scopes are also made explicit.
-
Statements, expressions, and match arms are stored separately. For example, statements in the
stmts
array reference expressions by their index (represented as aExprId
) in theexprs
array.
The THIR lives in rustc_mir_build::thir
. To construct a thir::Expr
,
you can use the thir_body
function, passing in the memory arena where the THIR
will be allocated. Dropping this arena will result in the THIR being destroyed,
which is useful to keep peak memory in check. Having a THIR representation of
all bodies of a crate in memory at the same time would be very heavy.
You can get a debug representation of the THIR by passing the -Zunpretty=thir-tree
flag
to rustc
.
To demonstrate, let's use the following example:
fn main() { let x = 1 + 2; }
Here is how that gets represented in THIR (as of Aug 2022):
#![allow(unused)] fn main() { Thir { // no match arms arms: [], exprs: [ // expression 0, a literal with a value of 1 Expr { ty: i32, temp_lifetime: Some( Node(1), ), span: oneplustwo.rs:2:13: 2:14 (#0), kind: Literal { lit: Spanned { node: Int( 1, Unsuffixed, ), span: oneplustwo.rs:2:13: 2:14 (#0), }, neg: false, }, }, // expression 1, scope surrounding literal 1 Expr { ty: i32, temp_lifetime: Some( Node(1), ), span: oneplustwo.rs:2:13: 2:14 (#0), kind: Scope { // reference to expression 0 above region_scope: Node(3), lint_level: Explicit( HirId { owner: DefId(0:3 ~ oneplustwo[6932]::main), local_id: 3, }, ), value: e0, }, }, // expression 2, literal 2 Expr { ty: i32, temp_lifetime: Some( Node(1), ), span: oneplustwo.rs:2:17: 2:18 (#0), kind: Literal { lit: Spanned { node: Int( 2, Unsuffixed, ), span: oneplustwo.rs:2:17: 2:18 (#0), }, neg: false, }, }, // expression 3, scope surrounding literal 2 Expr { ty: i32, temp_lifetime: Some( Node(1), ), span: oneplustwo.rs:2:17: 2:18 (#0), kind: Scope { region_scope: Node(4), lint_level: Explicit( HirId { owner: DefId(0:3 ~ oneplustwo[6932]::main), local_id: 4, }, ), // reference to expression 2 above value: e2, }, }, // expression 4, represents 1 + 2 Expr { ty: i32, temp_lifetime: Some( Node(1), ), span: oneplustwo.rs:2:13: 2:18 (#0), kind: Binary { op: Add, // references to scopes surrounding literals above lhs: e1, rhs: e3, }, }, // expression 5, scope surrounding expression 4 Expr { ty: i32, temp_lifetime: Some( Node(1), ), span: oneplustwo.rs:2:13: 2:18 (#0), kind: Scope { region_scope: Node(5), lint_level: Explicit( HirId { owner: DefId(0:3 ~ oneplustwo[6932]::main), local_id: 5, }, ), value: e4, }, }, // expression 6, block around statement Expr { ty: (), temp_lifetime: Some( Node(9), ), span: oneplustwo.rs:1:11: 3:2 (#0), kind: Block { body: Block { targeted_by_break: false, region_scope: Node(8), opt_destruction_scope: None, span: oneplustwo.rs:1:11: 3:2 (#0), // reference to statement 0 below stmts: [ s0, ], expr: None, safety_mode: Safe, }, }, }, // expression 7, scope around block in expression 6 Expr { ty: (), temp_lifetime: Some( Node(9), ), span: oneplustwo.rs:1:11: 3:2 (#0), kind: Scope { region_scope: Node(9), lint_level: Explicit( HirId { owner: DefId(0:3 ~ oneplustwo[6932]::main), local_id: 9, }, ), value: e6, }, }, // destruction scope around expression 7 Expr { ty: (), temp_lifetime: Some( Node(9), ), span: oneplustwo.rs:1:11: 3:2 (#0), kind: Scope { region_scope: Destruction(9), lint_level: Inherited, value: e7, }, }, ], stmts: [ // let statement Stmt { kind: Let { remainder_scope: Remainder { block: 8, first_statement_index: 0}, init_scope: Node(1), pattern: Pat { ty: i32, span: oneplustwo.rs:2:9: 2:10 (#0), kind: Binding { mutability: Not, name: "x", mode: ByValue, var: LocalVarId( HirId { owner: DefId(0:3 ~ oneplustwo[6932]::main), local_id: 7, }, ), ty: i32, subpattern: None, is_primary: true, }, }, initializer: Some( e5, ), else_block: None, lint_level: Explicit( HirId { owner: DefId(0:3 ~ oneplustwo[6932]::main), local_id: 6, }, ), }, opt_destruction_scope: Some( Destruction(1), ), }, ], } }
The MIR (Mid-level IR)
MIR is Rust's Mid-level Intermediate Representation. It is constructed from HIR. MIR was introduced in RFC 1211. It is a radically simplified form of Rust that is used for certain flow-sensitive safety checks – notably the borrow checker! – and also for optimization and code generation.
If you'd like a very high-level introduction to MIR, as well as some of the compiler concepts that it relies on (such as control-flow graphs and desugaring), you may enjoy the rust-lang blog post that introduced MIR.
Introduction to MIR
MIR is defined in the compiler/rustc_middle/src/mir/
module, but much of the code
that manipulates it is found in compiler/rustc_mir_build
,
compiler/rustc_mir_transform
, and
compiler/rustc_mir_dataflow
.
Some of the key characteristics of MIR are:
- It is based on a control-flow graph.
- It does not have nested expressions.
- All types in MIR are fully explicit.
Key MIR vocabulary
This section introduces the key concepts of MIR, summarized here:
- Basic blocks: units of the control-flow graph, consisting of:
- statements: actions with one successor
- terminators: actions with potentially multiple successors; always at the end of a block
- (if you're not familiar with the term basic block, see the background chapter)
- Locals: Memory locations allocated on the stack (conceptually, at
least), such as function arguments, local variables, and
temporaries. These are identified by an index, written with a
leading underscore, like
_1
. There is also a special "local" (_0
) allocated to store the return value. - Places: expressions that identify a location in memory, like
_1
or_1.f
. - Rvalues: expressions that produce a value. The "R" stands for
the fact that these are the "right-hand side" of an assignment.
- Operands: the arguments to an rvalue, which can either be a
constant (like
22
) or a place (like_1
).
- Operands: the arguments to an rvalue, which can either be a
constant (like
You can get a feeling for how MIR is constructed by translating simple programs into MIR and reading the pretty printed output. In fact, the playground makes this easy, since it supplies a MIR button that will show you the MIR for your program. Try putting this program into play (or clicking on this link), and then clicking the "MIR" button on the top:
fn main() { let mut vec = Vec::new(); vec.push(1); vec.push(2); }
You should see something like:
// WARNING: This output format is intended for human consumers only
// and is subject to change without notice. Knock yourself out.
fn main() -> () {
...
}
This is the MIR format for the main
function.
MIR shown by above link is optimized.
Some statements like StorageLive
are removed in optimization.
This happens because the compiler notices the value is never accessed in the code.
We can use rustc [filename].rs -Z mir-opt-level=0 --emit mir
to view unoptimized MIR.
This requires the nightly toolchain.
Variable declarations. If we drill in a bit, we'll see it begins with a bunch of variable declarations. They look like this:
let mut _0: (); // return place
let mut _1: std::vec::Vec<i32>; // in scope 0 at src/main.rs:2:9: 2:16
let mut _2: ();
let mut _3: &mut std::vec::Vec<i32>;
let mut _4: ();
let mut _5: &mut std::vec::Vec<i32>;
You can see that variables in MIR don't have names, they have indices,
like _0
or _1
. We also intermingle the user's variables (e.g.,
_1
) with temporary values (e.g., _2
or _3
). You can tell apart
user-defined variables because they have debuginfo associated to them (see below).
User variable debuginfo. Below the variable declarations, we find the only
hint that _1
represents a user variable:
scope 1 {
debug vec => _1; // in scope 1 at src/main.rs:2:9: 2:16
}
Each debug <Name> => <Place>;
annotation describes a named user variable,
and where (i.e. the place) a debugger can find the data of that variable.
Here the mapping is trivial, but optimizations may complicate the place,
or lead to multiple user variables sharing the same place.
Additionally, closure captures are described using the same system, and so
they're complicated even without optimizations, e.g.: debug x => (*((*_1).0: &T));
.
The "scope" blocks (e.g., scope 1 { .. }
) describe the lexical structure of
the source program (which names were in scope when), so any part of the program
annotated with // in scope 0
would be missing vec
, if you were stepping
through the code in a debugger, for example.
Basic blocks. Reading further, we see our first basic block (naturally it may look slightly different when you view it, and I am ignoring some of the comments):
bb0: {
StorageLive(_1);
_1 = const <std::vec::Vec<T>>::new() -> bb2;
}
A basic block is defined by a series of statements and a final terminator. In this case, there is one statement:
StorageLive(_1);
This statement indicates that the variable _1
is "live", meaning
that it may be used later – this will persist until we encounter a
StorageDead(_1)
statement, which indicates that the variable _1
is
done being used. These "storage statements" are used by LLVM to
allocate stack space.
The terminator of the block bb0
is the call to Vec::new
:
_1 = const <std::vec::Vec<T>>::new() -> bb2;
Terminators are different from statements because they can have more
than one successor – that is, control may flow to different
places. Function calls like the call to Vec::new
are always
terminators because of the possibility of unwinding, although in the
case of Vec::new
we are able to see that indeed unwinding is not
possible, and hence we list only one successor block, bb2
.
If we look ahead to bb2
, we will see it looks like this:
bb2: {
StorageLive(_3);
_3 = &mut _1;
_2 = const <std::vec::Vec<T>>::push(move _3, const 1i32) -> [return: bb3, unwind: bb4];
}
Here there are two statements: another StorageLive
, introducing the _3
temporary, and then an assignment:
_3 = &mut _1;
Assignments in general have the form:
<Place> = <Rvalue>
A place is an expression like _3
, _3.f
or *_3
– it denotes a
location in memory. An Rvalue is an expression that creates a
value: in this case, the rvalue is a mutable borrow expression, which
looks like &mut <Place>
. So we can kind of define a grammar for
rvalues like so:
<Rvalue> = & (mut)? <Place>
| <Operand> + <Operand>
| <Operand> - <Operand>
| ...
<Operand> = Constant
| copy Place
| move Place
As you can see from this grammar, rvalues cannot be nested – they can
only reference places and constants. Moreover, when you use a place,
we indicate whether we are copying it (which requires that the
place have a type T
where T: Copy
) or moving it (which works
for a place of any type). So, for example, if we had the expression x = a + b + c
in Rust, that would get compiled to two statements and a
temporary:
TMP1 = a + b
x = TMP1 + c
(Try it and see, though you may want to do release mode to skip over the overflow checks.)
MIR data types
The MIR data types are defined in the compiler/rustc_middle/src/mir/
module. Each of the key concepts mentioned in the previous section
maps in a fairly straightforward way to a Rust type.
The main MIR data type is Body
. It contains the data for a single
function (along with sub-instances of Mir for "promoted constants",
but you can read about those below).
- Basic blocks: The basic blocks are stored in the field
Body::basic_blocks
; this is a vector ofBasicBlockData
structures. Nobody ever references a basic block directly: instead, we pass aroundBasicBlock
values, which are newtype'd indices into this vector. - Statements are represented by the type
Statement
. - Terminators are represented by the
Terminator
. - Locals are represented by a newtype'd index type
Local
. The data for a local variable is found in theBody::local_decls
vector. There is also a special constantRETURN_PLACE
identifying the special "local" representing the return value. - Places are identified by the struct
Place
. There are a few fields:- Local variables like
_1
- Projections, which are fields or other things that "project
out" from a base place. These are represented by the newtype'd type
ProjectionElem
. So e.g. the place_1.f
is a projection, withf
being the "projection element" and_1
being the base path.*_1
is also a projection, with the*
being represented by theProjectionElem::Deref
element.
- Local variables like
- Rvalues are represented by the enum
Rvalue
. - Operands are represented by the enum
Operand
.
Representing constants
When code has reached the MIR stage, constants can generally come in two forms:
MIR constants (mir::Constant
) and type system constants (ty::Const
).
MIR constants are used as operands: in x + CONST
, CONST
is a MIR constant;
similarly, in x + 2
, 2
is a MIR constant. Type system constants are used in
the type system, in particular for array lengths but also for const generics.
Generally, both kinds of constants can be "unevaluated" or "already evaluated".
An unevaluated constant simply stores the DefId
of what needs to be evaluated
to compute this result. An evaluated constant (a "value") has already been
computed; their representation differs between type system constants and MIR
constants: MIR constants evaluate to a mir::ConstValue
; type system constants
evaluate to a ty::ValTree
.
Type system constants have some more variants to support const generics: they
can refer to local const generic parameters, and they are subject to inference.
Furthermore, the mir::Constant::Ty
variant lets us use an arbitrary type
system constant as a MIR constant; this happens whenever a const generic
parameter is used as an operand.
MIR constant values
In general, a MIR constant value (mir::ConstValue
) was computed by evaluating
some constant the user wrote. This const evaluation produces
a very low-level representation of the result in terms of individual bytes. We
call this an "indirect" constant (mir::ConstValue::Indirect
) since the value
is stored in-memory.
However, storing everything in-memory would be awfully inefficient. Hence there
are some other variants in mir::ConstValue
that can represent certain simple
and common values more efficiently. In particular, everything that can be
directly written as a literal in Rust (integers, floats, chars, bools, but also
"string literals"
and b"byte string literals"
) has an optimized variant that
avoids the full overhead of the in-memory representation.
ValTrees
An evaluated type system constant is a "valtree". The ty::ValTree
datastructure
allows us to represent
- arrays,
- many structs,
- tuples,
- enums and,
- most primitives.
The most important rule for
this representation is that every value must be uniquely represented. In other
words: a specific value must only be representable in one specific way. For example: there is only
one way to represent an array of two integers as a ValTree
:
ValTree::Branch(&[ValTree::Leaf(first_int), ValTree::Leaf(second_int)])
.
Even though theoretically a [u32; 2]
could be encoded in a u64
and thus just be a
ValTree::Leaf(bits_of_two_u32)
, that is not a legal construction of ValTree
(and is very complex to do, so it is unlikely anyone is tempted to do so).
These rules also mean that some values are not representable. There can be no union
s in type
level constants, as it is not clear how they should be represented, because their active variant
is unknown. Similarly there is no way to represent raw pointers, as addresses are unknown at
compile-time and thus we cannot make any assumptions about them. References on the other hand
can be represented, as equality for references is defined as equality on their value, so we
ignore their address and just look at the backing value. We must make sure that the pointer values
of the references are not observable at compile time. We thus encode &42
exactly like 42
.
Any conversion from
valtree back to a MIR constant value must reintroduce an actual indirection. At codegen time the
addresses may be deduplicated between multiple uses or not, entirely depending on arbitrary
optimization choices.
As a consequence, all decoding of ValTree
must happen by matching on the type first and making
decisions depending on that. The value itself gives no useful information without the type that
belongs to it.
Promoted constants
See the const-eval WG's docs on promotion.
MIR construction
unpack!
all the things- Lowering expressions into the desired MIR
- Operator lowering
- Method call lowering
- Conditions
- Aggregate construction
The lowering of HIR to MIR occurs for the following (probably incomplete) list of items:
- Function and closure bodies
- Initializers of
static
andconst
items - Initializers of enum discriminants
- Glue and shims of any kind
- Tuple struct initializer functions
- Drop code (the
Drop::drop
function is not called directly) - Drop implementations of types without an explicit
Drop
implementation
The lowering is triggered by calling the mir_built
query. The MIR builder does
not actually use the HIR but operates on the THIR instead, processing THIR
expressions recursively.
The lowering creates local variables for every argument as specified in the signature.
Next, it creates local variables for every binding specified (e.g. (a, b): (i32, String)
)
produces 3 bindings, one for the argument, and two for the bindings. Next, it generates
field accesses that read the fields from the argument and writes the value to the binding
variable.
With this initialization out of the way, the lowering triggers a recursive call
to a function that generates the MIR for the body (a Block
expression) and
writes the result into the RETURN_PLACE
.
unpack!
all the things
Functions that generate MIR tend to fall into one of two patterns. First, if the function generates only statements, then it will take a basic block as argument onto which those statements should be appended. It can then return a result as normal:
fn generate_some_mir(&mut self, block: BasicBlock) -> ResultType {
...
}
But there are other functions that may generate new basic blocks as well.
For example, lowering an expression like if foo { 22 } else { 44 }
requires generating a small "diamond-shaped graph".
In this case, the functions take a basic block where their code starts
and return a (potentially) new basic block where the code generation ends.
The BlockAnd
type is used to represent this:
fn generate_more_mir(&mut self, block: BasicBlock) -> BlockAnd<ResultType> {
...
}
When you invoke these functions, it is common to have a local variable block
that is effectively a "cursor". It represents the point at which we are adding new MIR.
When you invoke generate_more_mir
, you want to update this cursor.
You can do this manually, but it's tedious:
let mut block;
let v = match self.generate_more_mir(..) {
BlockAnd { block: new_block, value: v } => {
block = new_block;
v
}
};
For this reason, we offer a macro that lets you write
let v = unpack!(block = self.generate_more_mir(...))
.
It simply extracts the new block and overwrites the
variable block
that you named in the unpack!
.
Lowering expressions into the desired MIR
There are essentially four kinds of representations one might want of an expression:
Place
refers to a (or part of a) preexisting memory location (local, static, promoted)Rvalue
is something that can be assigned to aPlace
Operand
is an argument to e.g. a+
operation or a function call- a temporary variable containing a copy of the value
The following image depicts a general overview of the interactions between the representations:
Click here for a more detailed view
We start out with lowering the function body to an Rvalue
so we can create an
assignment to RETURN_PLACE
, This Rvalue
lowering will in turn trigger lowering to
Operand
for its arguments (if any). Operand
lowering either produces a const
operand, or moves/copies out of a Place
, thus triggering a Place
lowering. An
expression being lowered to a Place
can in turn trigger a temporary to be created
if the expression being lowered contains operations. This is where the snake bites its
own tail and we need to trigger an Rvalue
lowering for the expression to be written
into the local.
Operator lowering
Operators on builtin types are not lowered to function calls (which would end up being
infinite recursion calls, because the trait impls just contain the operation itself
again). Instead there are Rvalue
s for binary and unary operators and index operations.
These Rvalue
s later get codegened to llvm primitive operations or llvm intrinsics.
Operators on all other types get lowered to a function call to their impl
of the
operator's corresponding trait.
Regardless of the lowering kind, the arguments to the operator are lowered to Operand
s.
This means all arguments are either constants, or refer to an already existing value
somewhere in a local or static.
Method call lowering
Method calls are lowered to the same TerminatorKind
that function calls are.
In MIR there is no difference between method calls and function calls anymore.
Conditions
if
conditions and match
statements for enum
s with variants that have no fields are
lowered to TerminatorKind::SwitchInt
. Each possible value (so 0
and 1
for if
conditions) has a corresponding BasicBlock
to which the code continues.
The argument being branched on is (again) an Operand
representing the value of
the if condition.
Pattern matching
match
statements for enum
s with variants that have fields are lowered to
TerminatorKind::SwitchInt
, too, but the Operand
refers to a Place
where the
discriminant of the value can be found. This often involves reading the discriminant
to a new temporary variable.
Aggregate construction
Aggregate values of any kind (e.g. structs or tuples) are built via Rvalue::Aggregate
.
All fields are
lowered to Operator
s. This is essentially equivalent to one assignment
statement per aggregate field plus an assignment to the discriminant in the
case of enum
s.
MIR visitor
The MIR visitor is a convenient tool for traversing the MIR and either
looking for things or making changes to it. The visitor traits are
defined in the rustc_middle::mir::visit
module – there are two of
them, generated via a single macro: Visitor
(which operates on a
&Mir
and gives back shared references) and MutVisitor
(which
operates on a &mut Mir
and gives back mutable references).
To implement a visitor, you have to create a type that represents your visitor. Typically, this type wants to "hang on" to whatever state you will need while processing MIR:
struct MyVisitor<...> {
tcx: TyCtxt<'tcx>,
...
}
and you then implement the Visitor
or MutVisitor
trait for that type:
impl<'tcx> MutVisitor<'tcx> for MyVisitor {
fn visit_foo(&mut self, ...) {
...
self.super_foo(...);
}
}
As shown above, within the impl, you can override any of the
visit_foo
methods (e.g., visit_terminator
) in order to write some
code that will execute whenever a foo
is found. If you want to
recursively walk the contents of the foo
, you then invoke the
super_foo
method. (NB. You never want to override super_foo
.)
A very simple example of a visitor can be found in LocalFinder
.
By implementing visit_local
method, this visitor identifies local variables that
can be candidates for reordering.
Traversal
In addition the visitor, the rustc_middle::mir::traversal
module
contains useful functions for walking the MIR CFG in
different standard orders (e.g. pre-order, reverse
post-order, and so forth).
MIR queries and passes
If you would like to get the MIR:
- for a function - you can use the
optimized_mir
query (typically used by codegen) or themir_for_ctfe
query (typically used by compile time function evaluation, i.e., CTFE); - for a promoted - you can use the
promoted_mir
query.
These will give you back the final, optimized MIR. For foreign def-ids, we simply read the MIR from the other crate's metadata. But for local def-ids, the query will construct the optimized MIR by requesting a pipeline of upstream queries1. Each query will contain a series of passes. This section describes how those queries and passes work and how you can extend them.
To produce the optimized MIR for a given def-id D
, optimized_mir(D)
goes through several suites of passes, each grouped by a
query. Each suite consists of passes which perform linting, analysis, transformation or
optimization. Each query represent a useful intermediate point
where we can access the MIR dialect for type checking or other purposes:
mir_built(D)
– it gives the initial MIR just after it's built;mir_const(D)
– it applies some simple transformation passes to make MIR ready for const qualification;mir_promoted(D)
- it extracts promotable temps into separate MIR bodies, and also makes MIR ready for borrow checking;mir_drops_elaborated_and_const_checked(D)
- it performs borrow checking, runs major transformation passes (such as drop elaboration) and makes MIR ready for optimization;optimized_mir(D)
– it performs all enabled optimizations and reaches the final state.
See the Queries chapter for the general concept of query.
Implementing and registering a pass
A MirPass
is some bit of code that processes the MIR, typically transforming it along the way
somehow. But it may also do other things like linting (e.g., CheckPackedRef
,
CheckConstItemMutation
, FunctionItemReferences
, which implement MirLint
) or
optimization (e.g., SimplifyCfg
, RemoveUnneededDrops
). While most MIR passes
are defined in the rustc_mir_transform
crate, the MirPass
trait itself is
found in the rustc_middle
crate, and it basically consists of one primary method,
run_pass
, that simply gets an &mut Body
(along with the tcx
).
The MIR is therefore modified in place (which helps to keep things efficient).
A basic example of a MIR pass is RemoveStorageMarkers
, which walks
the MIR and removes all storage marks if they won't be emitted during codegen. As you
can see from its source, a MIR pass is defined by first defining a
dummy type, a struct with no fields:
#![allow(unused)] fn main() { pub struct RemoveStorageMarkers; }
for which we implement the MirPass
trait. We can then insert
this pass into the appropriate list of passes found in a query like
mir_built
, optimized_mir
, etc. (If this is an optimization, it
should go into the optimized_mir
list.)
Another example of a simple MIR pass is CleanupPostBorrowck
, which walks
the MIR and removes all statements that are not relevant to code generation. As you can see from
its source, it is defined by first defining a dummy type, a struct with no
fields:
#![allow(unused)] fn main() { pub struct CleanupPostBorrowck; }
for which we implement the MirPass
trait:
#![allow(unused)] fn main() { impl<'tcx> MirPass<'tcx> for CleanupPostBorrowck { fn run_pass(&self, tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) { ... } } }
We register this pass inside the mir_drops_elaborated_and_const_checked
query.
(If this is an optimization, it should go into the optimized_mir
list.)
If you are writing a pass, there's a good chance that you are going to want to use a MIR visitor. MIR visitors are a handy way to walk all the parts of the MIR, either to search for something or to make small edits.
Stealing
The intermediate queries mir_const()
and mir_promoted()
yield up
a &'tcx Steal<Body<'tcx>>
, allocated using tcx.alloc_steal_mir()
.
This indicates that the result may be stolen by a subsequent query – this is an
optimization to avoid cloning the MIR. Attempting to use a stolen
result will cause a panic in the compiler. Therefore, it is important
that you do not accidentally read from these intermediate queries without
the consideration of the dependency in the MIR processing pipeline.
Because of this stealing mechanism, some care must be taken to ensure that, before the MIR at a particular phase in the processing pipeline is stolen, anyone who may want to read from it has already done so.
Concretely, this means that if you have a query foo(D)
that wants to access the result of mir_promoted(D)
, you need to have foo(D)
calling the mir_const(D)
query first. This will force it
to execute even though you don't directly require its result.
This mechanism is a bit dodgy. There is a discussion of more elegant alternatives in rust-lang/rust#41710.
Overview
Below is an overview of the stealing dependency in the MIR processing pipeline2:
flowchart BT mir_for_ctfe* --borrow--> id40 id5 --steal--> id40 mir_borrowck* --borrow--> id3 id41 --steal part 1--> id3 id40 --steal part 0--> id3 mir_const_qualif* -- borrow --> id2 id3 -- steal --> id2 id2 -- steal --> id1 id1([mir_built]) id2([mir_const]) id3([mir_promoted]) id40([mir_drops_elaborated_and_const_checked]) id41([promoted_mir]) id5([optimized_mir]) style id1 fill:#bbf style id2 fill:#bbf style id3 fill:#bbf style id40 fill:#bbf style id41 fill:#bbf style id5 fill:#bbf
The stadium-shape queries (e.g., mir_built
) with a deep color are the primary queries in the
pipeline, while the rectangle-shape queries (e.g., mir_const_qualif*
3) with a shallow color
are those subsequent queries that need to read the results from &'tcx Steal<Body<'tcx>>
. With the
stealing mechanism, the rectangle-shape queries must be performed before any stadium-shape queries,
that have an equal or larger height in the dependency tree, ever do.
The mir_promoted
query will yield up a tuple
(&'tcx Steal<Body<'tcx>>, &'tcx Steal<IndexVec<Promoted, Body<'tcx>>>)
, promoted_mir
will steal
part 1 (&'tcx Steal<IndexVec<Promoted, Body<'tcx>>>
) and mir_drops_elaborated_and_const_checked
will steal part 0 (&'tcx Steal<Body<'tcx>>
). And their stealing is irrelevant to each other,
i.e., can be performed separately.
Note that the *
suffix in the queries represent a set of queries with the same prefix.
For example, mir_borrowck*
represents mir_borrowck
, mir_borrowck_const_arg
and
mir_borrowck_opt_const_arg
.
Example
As an example, consider MIR const qualification. It wants to read the result produced by the
mir_const
query. However, that result will be stolen by the mir_promoted
query at some
time in the pipeline. Before mir_promoted
is ever queried, calling the mir_const_qualif
query
will succeed since mir_const
will produce (if queried the first time) or cache (if queried
multiple times) the Steal
result and the result is not stolen yet. After mir_promoted
is
queried, the result would be stolen and calling the mir_const_qualif
query to read the result
would cause a panic.
Therefore, with this stealing mechanism, mir_promoted
should guarantee any mir_const_qualif*
queries are called before it actually steals, thus ensuring that the reads have already happened
(remember that queries are memoized, so executing a query twice
simply loads from a cache the second time).
Inline assembly
Overview
Inline assembly in rustc mostly revolves around taking an asm!
macro invocation and plumbing it
through all of the compiler layers down to LLVM codegen. Throughout the various stages, an
InlineAsm
generally consists of 3 components:
-
The template string, which is stored as an array of
InlineAsmTemplatePiece
. Each piece represents either a literal or a placeholder for an operand (just like format strings).#![allow(unused)] fn main() { pub enum InlineAsmTemplatePiece { String(String), Placeholder { operand_idx: usize, modifier: Option<char>, span: Span }, } }
-
The list of operands to the
asm!
(in
,[late]out
,in[late]out
,sym
,const
). These are represented differently at each stage of lowering, but follow a common pattern:in
,out
andinout
all have an associated register class (reg
) or explicit register ("eax"
).inout
has 2 forms: one with a single expression that is both read from and written to, and one with two separate expressions for the input and output parts.out
andinout
have alate
flag (lateout
/inlateout
) to indicate that the register allocator is allowed to reuse an input register for this output.out
and the split variant ofinout
allow_
to be specified for an output, which means that the output is discarded. This is used to allocate scratch registers for assembly code.const
refers to an anonymous constants and generally works like an inline const.sym
is a bit special since it only accepts a path expression, which must point to astatic
or afn
.
-
The options set at the end of the
asm!
macro. The only ones that are of particular interest to rustc areNORETURN
which makesasm!
return!
instead of()
, andRAW
which disables format string parsing. The remaining options are mostly passed through to LLVM with little processing.#![allow(unused)] fn main() { bitflags::bitflags! { pub struct InlineAsmOptions: u16 { const PURE = 1 << 0; const NOMEM = 1 << 1; const READONLY = 1 << 2; const PRESERVES_FLAGS = 1 << 3; const NORETURN = 1 << 4; const NOSTACK = 1 << 5; const ATT_SYNTAX = 1 << 6; const RAW = 1 << 7; const MAY_UNWIND = 1 << 8; } } }
AST
InlineAsm
is represented as an expression in the AST with the ast::InlineAsm
type.
The asm!
macro is implemented in rustc_builtin_macros
and outputs an InlineAsm
AST node. The
template string is parsed using fmt_macros
, positional and named operands are resolved to
explicit operand indices. Since target information is not available to macro invocations,
validation of the registers and register classes is deferred to AST lowering.
HIR
InlineAsm
is represented as an expression in the HIR with the hir::InlineAsm
type.
AST lowering is where InlineAsmRegOrRegClass
is converted from Symbol
s to an actual register or
register class. If any modifiers are specified for a template string placeholder, these are
validated against the set allowed for that operand type. Finally, explicit registers for inputs and
outputs are checked for conflicts (same register used for different operands).
Type checking
Each register class has a whitelist of types that it may be used with. After the types of all
operands have been determined, the intrinsicck
pass will check that these types are in the
whitelist. It also checks that split inout
operands have compatible types and that const
operands are integers or floats. Suggestions are emitted where needed if a template modifier should
be used for an operand based on the type that was passed into it.
THIR
InlineAsm
is represented as an expression in the THIR with the InlineAsmExpr
type.
The only significant change compared to HIR is that Sym
has been lowered to either a SymFn
whose expr
is a Literal
ZST of the fn
, or a SymStatic
which points to the DefId
of a
static
.
MIR
InlineAsm
is represented as a Terminator
in the MIR with the TerminatorKind::InlineAsm
variant
As part of THIR lowering, InOut
and SplitInOut
operands are lowered to a split form with a
separate in_value
and out_place
.
Semantically, the InlineAsm
terminator is similar to the Call
terminator except that it has
multiple output places where a Call
only has a single return place output.
Codegen
Operands are lowered one more time before being passed to LLVM codegen, this is represented by the InlineAsmOperandRef
type from rustc_codegen_ssa
.
The operands are lowered to LLVM operands and constraint codes as follow:
out
and the output part ofinout
operands are added first, as required by LLVM. Late output operands have a=
prefix added to their constraint code, non-late output operands have a=&
prefix added to their constraint code.in
operands are added normally.inout
operands are tied to the matching output operand.sym
operands are passed as function pointers or pointers, using the"s"
constraint.const
operands are formatted to a string and directly inserted in the template string.
The template string is converted to LLVM form:
$
characters are escaped as$$
.const
operands are converted to strings and inserted directly.- Placeholders are formatted as
${X:M}
whereX
is the operand index andM
is the modifier character. Modifiers are converted from the Rust form to the LLVM form.
The various options are converted to clobber constraints or LLVM attributes, refer to the RFC for more details.
Note that LLVM is sometimes rather picky about what types it accepts for certain constraint codes so we sometimes need to insert conversions to/from a supported type. See the target-specific ISelLowering.cpp files in LLVM for details of what types are supported for each register class.
Adding support for new architectures
Adding inline assembly support to an architecture is mostly a matter of defining the registers and
register classes for that architecture. All the definitions for register classes are located in
compiler/rustc_target/asm/
.
Additionally you will need to implement lowering of these register classes to LLVM constraint codes
in compiler/rustc_codegen_llvm/asm.rs
.
When adding a new architecture, make sure to cross-reference with the LLVM source code:
- LLVM has restrictions on which types can be used with a particular constraint code. Refer to the
getRegForInlineAsmConstraint
function inlib/Target/${ARCH}/${ARCH}ISelL