This is an update of my previous post “How to build LLVM from source” (2018-04-16), which is now out-of-date. LLVM has moved to a “monorepo” design instead of a collection of smaller tool-specific repositories. So this new post has fewer steps!
UPDATE, 2020-03-22: Patch D69221 seems to have made the procedure much simpler, eliminating the need for Jens Jorgensen’s patch, at least on OSX 10.14.6. Therefore I have shortened this post considerably. You can find the old version in the blog’s git history or on the Wayback Machine.
The LLVM codebase’s official home is
(However, to submit patches to LLVM projects, you must use
the official Phabricator;
don’t submit GitHub pull requests against
At least not as of December 2020.)
Step 1: Fork!
Go to your GitHub account and fork the following repository:
Step 2: Get the code!
Locally clone the repo to the right place.
cd $ROOT git clone email@example.com:llvm/llvm-project
This is a good time to set up the
.git/config for the repo
you just cloned (for example,
I set it up this way:
[remote "origin"] url = firstname.lastname@example.org:Quuxplusone/llvm-project.git fetch = +refs/heads/*:refs/remotes/origin/* [remote "upstream"] url = email@example.com:llvm/llvm-project.git fetch = +refs/heads/*:refs/remotes/upstream/* [branch "master"] remote = upstream merge = refs/heads/master
This gives me two remotes: one named
upstream whence I can pull,
and one named
origin whither I can push. My local
will both track
upstream/master. Anything I do in my local repo, I will do in a
feature branch; my feature branches will track
Step 3: Build!
mkdir $ROOT/llvm-project/build cd $ROOT/llvm-project/build cmake -G Ninja \ -DDEFAULT_SYSROOT="$(xcrun --show-sdk-path)" \ -DLLVM_ENABLE_PROJECTS="clang;libcxx;libcxxabi" \ -DCMAKE_BUILD_TYPE=Release ../llvm ninja clang ninja cxx
clang will build both
If you omit
-DCMAKE_BUILD_TYPE=Release (or at least
this first part will still work, but you’ll produce
a “debug-build” version of
clang that is super slow, and then the “bootstrap” step below will
take days instead of minutes. So watch out for that.
On my laptop,
cmake takes about 42 seconds;
ninja clang takes about 96 minutes.
Troubleshooting step 3
If something goes wrong, you can usually recover via
and, absolute worst case, you can
rm -rf $ROOT/llvm-project/build and start over.
If you succeed in building
clang, but then when you run it you get errors about
the standard C-language headers, like this,
$ bin/clang++ test.cpp test.cpp:1:10: fatal error: 'stdio.h' file not found #include <stdio.h> ^~~~~~~~~ 1 error generated.
then you may have set
On 10.14.6, when I run
xcrun --show-sdk-path, I get
Your results may vary.
If you get errors about the C++ headers, such as
<vector>, it’s because you still
need to build libc++: run
Step 4: Bootstrap
check-clang and libc++.
Here we will instruct CMake to build Clang again, using the Clang we just built. There is apparently an official way to bootstrap Clang (probably out-of-date). However, I use an approach inspired by the CMake FAQ. Note that we will not be installing Clang over top of the system compiler; that would be super dangerous and you should never do it!
mkdir $ROOT/llvm-project/build2 cd $ROOT/llvm-project/build2 CXX="$ROOT/llvm-project/build/bin/clang++" \ cmake -G Ninja \ -DDEFAULT_SYSROOT="$(xcrun --show-sdk-path)" \ -DLLVM_ENABLE_PROJECTS="clang;compiler-rt;libcxx;libcxxabi" \ -DCMAKE_BUILD_TYPE=RelWithDebInfo ../llvm ninja clang cxx ninja check-clang check-cxx
Now you have built two versions of
$ROOT/llvm-project/build/bin/clang++ is the version
built with your system compiler, and
$ROOT/llvm-project/build2/bin/clang++ is the version built with
that version. You can extend this to
cxx will build
cxxabi will build
check-cxx will build
compiler-rt will build
which are needed by Clang’s
cmake takes about 58 seconds;
ninja clang takes about 153 minutes.
ninja check-clang takes another 53 minutes:
36 minutes to build a bunch of additional tools, and then 17 minutes to run the actual tests.
ninja cxx takes about 84 seconds.
ninja check-cxx takes about 50 minutes (but see the caveat below about
You can invoke CMake with
-DCMAKE_BUILD_TYPE=Debug to produce a
clang binary with assertions enabled.
This takes only about 60 minutes to build, but you wouldn’t want to use the resulting binary for
anything heavy-duty because it’s so slow.
Step 5: Run specific tests.
Running a specific test or directory-of-tests for any product is easy:
cd $ROOT/llvm-project/build2 ./bin/llvm-lit -sv ../llvm/test/Analysis ./bin/llvm-lit -sv ../clang/test/ARCMT ./bin/llvm-lit -sv --param std=c++17 ../libcxx/test/std/re
However, before you can successfully run one of these lines,
you must have run the corresponding one of
at least once, to initialize the right stuff under the
(Thanks to Brian Cain for documenting this recipe.)
But watch out — both
make check-cxx and
llvm-lit will by default use your system compiler
to run the libc++ tests! This is not what you want! Tell
llvm-lit to use your newly built Clang
by passing the
cxx_under_test parameter, like this:
./bin/llvm-lit -sv --param std=c++17 --param cxx_under_test=`pwd`/bin/clang ../libcxx/test/
On my laptop, this command line again takes about 50 minutes to run all the libc++ tests, but this time it correctly uses the bootstrapped compiler.