A guide to Dockerfiles for building LLVM


You can find a number of sources to build docker images with LLVM components in llvm/utils/docker. They can be used by anyone who wants to build the docker images for their own use, or as a starting point for someone who wants to write their own Dockerfiles.

We currently provide Dockerfiles with debian8 and nvidia-cuda base images. We also provide an example image, which contains placeholders that one would need to fill out in order to produce Dockerfiles for a new docker image.


Docker images provide a way to produce binary distributions of software inside a controlled environment. Having Dockerfiles to builds docker images inside LLVM repo makes them much more discoverable than putting them into any other place.

Docker basics

If you’ve never heard about Docker before, you might find this section helpful to get a very basic explanation of it. Docker is a popular solution for running programs in an isolated and reproducible environment, especially to maintain releases for software deployed to large distributed fleets. It uses linux kernel namespaces and cgroups to provide a lightweight isolation inside currently running linux kernel. A single active instance of dockerized environment is called a docker container. A snapshot of a docker container filesystem is called a docker image. One can start a container from a prebuilt docker image.

Docker images are built from a so-called Dockerfile, a source file written in a specialized language that defines instructions to be used when build the docker image (see official documentation for more details). A minimal Dockerfile typically contains a base image and a number of RUN commands that have to be executed to build the image. When building a new image, docker will first download your base image, mount its filesystem as read-only and then add a writable overlay on top of it to keep track of all filesystem modifications, performed while building your image. When the build process is finished, a diff between your image’s final filesystem state and the base image’s filesystem is stored in the resulting image.


The llvm/utils/docker folder contains Dockerfiles and simple bash scripts to serve as a basis for anyone who wants to create their own Docker image with LLVM components, compiled from sources. The sources are checked out from the upstream svn repository when building the image.

Inside each subfolder we host Dockerfiles for two images:

  • build/ image is used to compile LLVM, it installs a system compiler and all build dependencies of LLVM. After the build process is finished, the build image will have an archive with compiled components at /tmp/clang.tar.gz.
  • release/ image usually only contains LLVM components, compiled by the build/ image, and also libstdc++ and binutils to make image minimally useful for C++ development. The assumption is that you usually want clang to be one of the provided components.

To build both of those images, use build_docker_image.sh script. It will checkout LLVM sources and build clang in the build container, copy results of the build to the local filesystem and then build the release container using those. The build_docker_image.sh accepts a list of LLVM repositories to checkout, and arguments for CMake invocation.

If you want to write your own docker image, start with an example/ subfolder. It provides incomplete Dockerfiles with (very few) FIXMEs explaining the steps you need to take in order to make your Dockerfiles functional.


The llvm/utils/build_docker_image.sh script provides a rather high degree of control on how to run the build. It allows you to specify the projects to checkout from svn and provide a list of CMake arguments to use during when building LLVM inside docker container.

Here’s a very simple example of getting a docker image with clang binary, compiled by the system compiler in the debian8 image:

./llvm/utils/docker/build_docker_image.sh \
    --source debian8 \
    --docker-repository clang-debian8 --docker-tag "staging" \
    -p clang -i install-clang -i install-clang-headers \
    -- \

Note that a build like that doesn’t use a 2-stage build process that you probably want for clang. Running a 2-stage build is a little more intricate, this command will do that:

# Run a 2-stage build.
#   LLVM_TARGETS_TO_BUILD=Native is to reduce stage1 compile time.
#   Options, starting with BOOTSTRAP_* are passed to stage2 cmake invocation.
./build_docker_image.sh \
    --source debian8 \
    --docker-repository clang-debian8 --docker-tag "staging" \
    -p clang -i stage2-install-clang -i stage2-install-clang-headers \
    -- \
    -DCLANG_ENABLE_BOOTSTRAP=ON -DCLANG_BOOTSTRAP_TARGETS="install-clang;install-clang-headers"

This will produce two images, a release image clang-debian8:staging and a build image clang-debian8-build:staging from the latest upstream revision. After the image is built you can run bash inside a container based on your image like this:

docker run -ti clang-debian8:staging bash

Now you can run bash commands as you normally would:

root@80f351b51825:/# clang -v
clang version 5.0.0 (trunk 305064)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /bin
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/4.8
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/4.8.4
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/4.9
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/4.9.2
Selected GCC installation: /usr/lib/gcc/x86_64-linux-gnu/4.9
Candidate multilib: .;@m64
Selected multilib: .;@m64

Which image should I choose?

We currently provide two images: debian8-based and nvidia-cuda-based. They differ in the base image that they use, i.e. they have a different set of preinstalled binaries. Debian8 is very minimal, nvidia-cuda is larger, but has preinstalled CUDA libraries and allows to access a GPU, installed on your machine.

If you need a minimal linux distribution with only clang and libstdc++ included, you should try debian8-based image.

If you want to use CUDA libraries and have access to a GPU on your machine, you should choose nvidia-cuda-based image and use nvidia-docker to run your docker containers. Note that you don’t need nvidia-docker to build the images, but you need it in order to have an access to GPU from a docker container that is running the built image.

If you have a different use-case, you could create your own image based on example/ folder.

Any docker image can be built and run using only the docker binary, i.e. you can run debian8 build on Fedora or any other Linux distribution. You don’t need to install CMake, compilers or any other clang dependencies. It is all handled during the build process inside Docker’s isolated environment.

Stable build

If you want a somewhat recent and somewhat stable build, use the branches/google/stable branch, i.e. the following command will produce a debian8-based image using the latest google/stable sources for you:

./llvm/utils/docker/build_docker_image.sh \
    -s debian8 --d clang-debian8 -t "staging" \
    --branch branches/google/stable \
    -p clang -i install-clang -i install-clang-headers \
    -- \

Minimizing docker image size

Due to Docker restrictions we use two images (i.e., build and release folders) for the release image to be as small as possible. It’s much easier to achieve that using two images, because Docker would store a filesystem layer for each command in the Dockerfile, i.e. if you install some packages in one command, then remove those in a separate command, the size of the resulting image will still be proportinal to the size of an image with installed packages. Therefore, we strive to provide a very simple release image which only copies compiled clang and does not do anything else.

Docker 1.13 added a --squash flag that allows to flatten the layers of the image, i.e. remove the parts that were actually deleted. That is an easier way to produce the smallest images possible by using just a single image. We do not use it because as of today the flag is in experimental stage and not everyone may have the latest docker version available. When the flag is out of experimental stage, we should investigate replacing two images approach with just a single image, built using --squash flag.