This tutorial will guide you through the process of making a change to LLVM, and contributing it back to the LLVM project. We’ll be making a change to Clang, but the steps for other parts of LLVM are the same. Even though the change we’ll be making is simple, we’re going to cover steps like building LLVM, running the tests, and code review. This is good practice, and you’ll be prepared for making larger changes.

We’ll assume you:

  • know how to use an editor,
  • have basic C++ knowledge,
  • know how to install software on your system,
  • are comfortable with the command line,
  • have basic knowledge of git.

The change we’re making

Clang has a warning for infinite recursion:

$ echo "void foo() { foo(); }" > ~/
$ clang -c -Wall ~/ warning: all paths through this function will call
itself [-Winfinite-recursion]

This is clear enough, but not exactly catchy. Let’s improve the wording a little: warning: to understand recursion, you must first
understand recursion [-Winfinite-recursion]


We’re going to need some tools:

  • git: to check out the LLVM source code,
  • a C++ compiler: to compile LLVM source code. You’ll want a recent version of Clang, GCC, or Visual Studio.
  • CMake: used to configure how LLVM should be built on your system,
  • ninja: runs the C++ compiler to (re)build specific parts of LLVM,
  • python: to run the LLVM tests,
  • arcanist: for uploading changes for review,

As an example, on Ubuntu:

$ sudo apt-get install git clang cmake ninja-build python arcanist

Building LLVM


The source code is stored on Github in one large repository (“the monorepo”).

It may take a while to download!

$ git clone

This will create a directory “llvm-project” with all of the source code.(Checking out anonymously is OK - pushing commits uses a different mechanism, as we’ll see later)

Configure your workspace

Before we can build the code, we must configure exactly how to build it by running CMake. CMake combines information from three sources:

  • explicit choices you make (is this a debug build?)
  • settings detected from your system (where are libraries installed?)
  • project structure (which files are part of ‘clang’?)

First, create a directory to build in. Usually, this is llvm-project/build.

$ mkdir llvm-project/build
$ cd llvm-project/build

Now, run CMake:

$ cmake -G Ninja ../llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_PROJECTS=clang

If all goes well, you’ll see a lot of “performing test” lines, and finally:

Configuring done
Generating done
Build files have been written to: /path/llvm-project/build

And you should see a file.

Let’s break down that last command a little:

  • -G Ninja: we’re going to use ninja to build; please create

  • ../llvm: this is the path to the source of the “main” LLVM project

  • The two -D flags set CMake variables, which override CMake/project defaults:

  • CMAKEBUILDTYPE=Release: build in optimized mode, which is (surprisingly) the fastest option.

    If you want to run under a debugger, you should use the default Debug (which is totally unoptimized, and will lead to >10x slower test runs) or RelWithDebInfo which is a halfway point. CMAKEBUILDTYPE affects code generation only, assertions are on by default regardless! LLVMENABLEASSERTIONS=Off disables them.

  • LLVMENABLEPROJECTS=clang : this lists the LLVM subprojects you are interested in building, in addition to LLVM itself. Multiple projects can be listed, separated by semicolons, such as “clang; lldb”.In this example, we’ll be making a change to Clang, so we should build it.

Finally, create a symlink (or a copy) of llvm-project/build/compile-commands.json into llvm-project/:

$ ln -s build/compile_commands.json ../

(This isn’t strictly necessary for building and testing, but allows tools like clang-tidy, clang-query, and clangd to work in your source tree).

Build and test

Finally, we can build the code! It’s important to do this first, to ensure we’re in a good state before making changes. But what to build? In ninja, you specify a target. If we just want to build the clang binary, our target name is “clang” and we run:

$ ninja clang

The first time we build will be very slow - Clang + LLVM is a lot of code. But incremental builds are fast: ninja will only rebuild the parts that have changed. When it finally finishes you should have a working clang binary. Try running:

$ bin/clang --version

There’s also a target for building and running all the clang tests:

$ ninja check-clang

This is a common pattern in LLVM: check-llvm is all the checks for core, other projects have targets like check-lldb.

Making changes


We need to find the file containing the error message.

$ git grep "all paths through this function" ..
../clang/include/clang/Basic/  "all paths through this function will call itself">,

The string that appears in is the one that is printed by Clang. *.td files define tables - in this case it’s a list of warnings and errors clang can emit and their messages. Let’s update the message in your favorite editor:

$ vi ../clang/include/clang/Basic/

Find the message (it should be under warninfiniterecursive_function)Change the message to “in order to understand recursion, you must first understand recursion”.

Test again

To verify our change, we can build clang and manually check that it works.

$ ninja clang
$ bin/clang -Wall ~/

**/path/** **warning****: in order to understand recursion, you must
first understand recursion [-Winfinite-recursion]**

We should also run the tests to make sure we didn’t break something.

$ ninja check-clang

Notice that it is much faster to build this time, but the tests take just as long to run. Ninja doesn’t know which tests might be affected, so it runs them all.

Testing Time: 408.84s
Failing Tests (1):
    Clang :: SemaCXX/warn-infinite-recursion.cpp

Well, that makes sense… and the test output suggests it’s looking for the old string “call itself” and finding our new message instead. Note that more tests may fail in a similar way as new tests are added time to time.

Let’s fix it by updating the expectation in the test.

$ vi ../clang/test/SemaCXX/warn-infinite-recursion.cpp

Everywhere we see // expected-warning{{call itself}} (or something similar from the original warning text), let’s replace it with // expected-warning{{to understand recursion}}.

Now we could run all the tests again, but this is a slow way to iterate on a change! Instead, let’s find a way to re-run just the specific test. There are two main types of tests in LLVM:

  • lit tests (e.g. SemaCXX/warn-infinite-recursion.cpp).

These are fancy shell scripts that run command-line tools and verify the output. They live in files like clang/test/FixIt/dereference-addressof.c. Re-run like this:

$ bin/llvm-lit -v ../clang/test/SemaCXX/warn-infinite-recursion.cpp
  • unit tests (e.g. ToolingTests/ReplacementTest.CanDeleteAllText)

These are C++ programs that call LLVM functions and verify the results. They live in suites like ToolingTests. Re-run like this:

$ ninja ToolingTests && tools/clang/unittests/Tooling/ToolingTests

Commit locally

We’ll save the change to a local git branch. This lets us work on other things while the change is being reviewed. Changes should have a description, to explain to reviewers and future readers of the code why the change was made.

$ git checkout -b myfirstpatch
$ git commit -am "[Diagnostic] Clarify -Winfinite-recursion message"

Now we’re ready to send this change out into the world! By the way, There is a unwritten convention of using tag for your commit. Tags usually represent modules that you intend to modify. If you don’t know the tags for your modules, you can look at the commit history :

Code review

Finding a reviewer

Changes can be reviewed by anyone in the LLVM community who has commit access.For larger and more complicated changes, it’s important that the reviewer has experience with the area of LLVM and knows the design goals well. The author of a change will often assign a specific reviewer (git blame and git log can be useful to find one).

As our change is fairly simple, we’ll add the cfe-commits mailing list as a subscriber; anyone who works on clang can likely pick up the review. (For changes outside clang, llvm-commits is the usual list. See for all the *-commits mailing lists).

Uploading a change for review

LLVM code reviews happen at The web interface is called Phabricator, and the code review part is Differential. You should create a user account there for reviews (click “Log In” and then “Register new account”).

Now you can upload your change for review:

$ arc diff HEAD^

This creates a review for your change, comparing your current commit with the previous commit. You will be prompted to fill in the review details. Your commit message is already there, so just add cfe-commits under the “subscribers” section. It should print a code review URL: You can always find your active reviews on Phabricator under “My activity”.

Review process

When you upload a change for review, an email is sent to you, the cfe-commits list, and anyone else subscribed to these kinds of changes. Within a few days, someone should start the review. They may add themselves as a reviewer, or simply start leaving comments. You’ll get another email any time the review is updated. The details are in the


The reviewer can leave comments on the change, and you can reply. Some comments are attached to specific lines, and appear interleaved with the code. You can either reply to these, or address them and mark them as “done”. Note that in-line replies are not sent straight away! They become “draft” comments and you must click “Submit” at the bottom of the page.

Updating your change

If you make changes in response to a reviewer’s comments, simply run

$ arc diff

again to update the change and notify the reviewer. Typically this is a good time to send any draft comments as well.

Accepting a revision

When the reviewer is happy with the change, they will Accept the revision. They may leave some more minor comments that you should address, but at this point the review is complete. It’s time to get it committed!

Commit by proxy

As this is your first change, you won’t have access to commit it yourself yet. The reviewer doesn’t know this, so you need to tell them! Leave a message on the review like:

Thanks @somellvmdev. I don’t have commit access, can you land this patch for me? Please use “My Name my@email” to commit the change.

The review will be updated when the change is committed.

Review expectations

In order to make LLVM a long-term sustainable effort, code needs to be maintainable and well tested. Code reviews help to achieve that goal. Especially for new contributors, that often means many rounds of reviews and push-back on design decisions that do not fit well within the overall architecture of the project.

For your first patches, this means:

  • be kind, and expect reviewers to be kind in return - LLVM has a Code of Conduct;
  • be patient - understanding how a new feature fits into the architecture of the project is often a time consuming effort, and people have to juggle this with other responsibilities in their lives; ping the review once a week when there is no response;
  • if you can’t agree, generally the best way is to do what the reviewer asks; we optimize for readability of the code, which the reviewer is in a better position to judge; if this feels like it’s not the right option, you can contact the cfe-dev mailing list to get more feedback on the direction;

Commit access

Once you’ve contributed a handful of patches to LLVM, start to think about getting commit access yourself. It’s probably a good idea if:

  • you’ve landed 3-5 patches of larger scope than “fix a typo”
  • you’d be willing to review changes that are closely related to yours
  • you’d like to keep contributing to LLVM.

Getting commit access

LLVM uses Git for committing changes. The details are in the developer policy document.

With great power

Actually, this would be a great time to read the rest of the developer policy, too. At minimum, you need to be subscribed to the relevant commits list before landing changes (e.g., as discussion often happens there if a new patch causes problems.


Let’s say you have a change on a local git branch, reviewed and ready to commit. Things to do first:

  • if you used multiple fine-grained commits locally, squash them into a single commit. LLVM prefers commits to match the code that was reviewed. (If you created one commit and then used “arc diff”, you’re fine)
  • rebase your patch against the latest LLVM code. LLVM uses a linear history, so everything should be based on an up-to-date origin/main.
$ git pull --rebase main
  • ensure the patch looks correct.
$ git show
  • run the tests one last time, for good luck

At this point git show should show a single commit on top of origin/main.

Now you can push your commit with

$ git push HEAD:main

You should see your change on GitHub within minutes.

Post-commit errors

Once your change is submitted it will be picked up by automated build bots that will build and test your patch in a variety of configurations.

You can see all configurations and their current state in a waterfall view at The waterfall view is good to get a general overview over the tested configurations and to see which configuration have been broken for a while.

The console view at helps to get a better understanding of the build results of a specific patch. If you want to follow along how your change is affecting the build bots, this should be the first place to look at - the colored bubbles correspond to projects in the waterfall.

If you see a broken build, do not despair - some build bots are continuously broken; if your change broke the build, you will see a red bubble in the console view, while an already broken build will show an orange bubble. Of course, even when the build was already broken, a new change might introduce a hidden new failure.

When you want to see more details how a specific build is broken, click the red bubble.
If post-commit error logs confuse you, do not worry too much - everybody on the project is aware that this is a bit unwieldy, so expect people to jump in and help you understand what’s going on!

buildbots, overview of bots, getting error logs.


if in doubt, revert and re-land.


llvm is a land of contrasts.