Toolchain build scripts

There are times where a tip of tree LLVM build will have some issue fixed and it isn't available to you, maybe because it isn't in a release or it isn't available through your distribution's package management system. At that point, to get that fix, LLVM needs to be compiled, which sounds scary but is rather simple. The build-llvm.py script takes it a step farther by trying to optimize both LLVM's build time by:

• Trimming down a lot of things that kernel developers don't care about:
  • Documentation
  • LLVM tests
  • Ocaml bindings
  • libfuzzer
• Building with the faster tools available (in order of fastest to slowest):
  • clang + lld
  • clang/gcc + ld.gold
  • clang/gcc + ld.bfd

Getting started

These scripts have been tested in a Docker image of the following distributions with the following packages installed. LLVM has minimum host tool version requirements so the latest stable version of the chosen distribution should be used whenever possible to ensure recent versions of the tools are used. Build errors from within LLVM are expected if the tool version is not recent enough, in which case it will need to be built from source or installed through other means.

• Debian/Ubuntu

  apt install bc \
              binutils-dev \
              bison \
              build-essential \
              ca-certificates \
              ccache \
              clang \
              cmake \
              curl \
              file \
              flex \
              git \
              libelf-dev \
              libssl-dev \
              libstdc++-$(apt list libstdc++6 2>/dev/null | grep -Eos '[0-9]+\.[0-9]+\.[0-9]+' | head -1 | cut -d . -f 1)-dev \
              lld \
              make \
              ninja-build \
              python3-dev \
              texinfo \
              u-boot-tools \
              xz-utils \
              zlib1g-dev
  

• Fedora

  dnf install bc \
              binutils-devel \
              bison \
              ccache \
              clang \
              cmake \
              elfutils-libelf-devel \
              flex \
              gcc \
              gcc-c++ \
              git \
              lld \
              make \
              ninja-build \
              openssl-devel \
              python3 \
              texinfo-tex \
              uboot-tools \
              xz \
              zlib-devel
  

• Arch Linux / Manjaro

  pacman -S base-devel \
            bc \
            bison \
            ccache \
            clang \
            cpio \
            cmake \
            flex \
            git \
            libelf \
            lld \
            llvm \
            ninja \
            openssl \
            python3 \
            uboot-tools
  

• Clear Linux

  swupd bundle-add c-basic \
                   ccache \
                   curl \
                   dev-utils \
                   devpkg-elfutils \
                   devpkg-openssl \
                   git \
                   python3-basic \
                   which
  

  Additionally, to build PowerPC kernels, you will need to build the U-Boot tools because there is no distribution package. The U-Boot tarballs can be found here and they can be built and used like so:

  $ curl -LSs https://ftp.denx.de/pub/u-boot/u-boot-2021.01.tar.bz2 | tar -xjf -
  $ cd u-boot-2021.01
  $ make -j"$(nproc)" defconfig tools-all
  ...
  $ sudo install -Dm755 tools/mkimage /usr/local/bin/mkimage
  $ mkimage -V
  mkimage version 2021.01
  

  Lastly, Clear Linux has ${CC}, ${CXX}, ${CFLAGS}, and ${CXXFLAGS} in the environment, which messes with the heuristics of the script for selecting a compiler. By default, the script will attempt to use clang and ld.lld but the environment's value of ${CC} and ${CXX} is respected first so gcc and g++ will be used. Clear Linux has optimized their gcc and g++ so this is fine but if you would like to use clang and clang++ instead, invoke the script like so:

  $ CC=clang CFLAGS= CXX=clang++ CXXFLAGS= ./build-llvm.py ...
  

Python 3.5.3+ is recommended, as that is what the script has been tested against. These scripts should be distribution agnostic. Please feel free to add different distribution install commands here through a pull request.

build-llvm.py

By default, ./build-llvm.py will clone LLVM, grab the latest binutils tarball (for the LLVMgold.so plugin), and build LLVM, clang, and lld, and install them into install.

The script automatically clones and manages the llvm-project. If you would like to do this management yourself, such as downloading a release tarball from releases.llvm.org, doing a more aggressive shallow clone (versus what is done in the script via --shallow-clone), or doing a bisection of LLVM, you just need to make sure that your source is in an llvm-project folder within the root of this repository and pass --no-update into the script. See this comment for an example.

Run ./build-llvm.py -h for more options and information.

build-binutils.py

This script builds a standalone copy of binutils. By default, ./build-binutils.py will download the latest stable version of binutils, build for all architectures we currently care about (see the help text or script for the full list), and install them into install. Run ./build-binutils.py -h for more options.

Building a standalone copy of binutils might be needed because certain distributions like Arch Linux (whose options the script uses) might symlink /usr/lib/LLVMgold.so to /usr/lib/bfd-plugins (source), which can cause issues when using the system's linker for LTO (even with LD_LIBRARY_PATH):

bfd plugin: LLVM gold plugin has failed to create LTO module: Unknown attribute kind (60) (Producer: 'LLVM9.0.0svn' Reader: 'LLVM 7.0.1')

Having a standalone copy of binutils (ideally in the same folder at the LLVM toolchain so that only one PATH modification is needed) works around this without any adverse side effects. Another workaround is bind mounting the new LLVMgold.so to /usr/lib/LLVMgold.so.

Contributing

This repository openly welcomes pull requests! There are a few presubmit checks that run to make sure the code stays consistently formatted and free of bugs.

1. All Python files must be passed through yapf. See the installation section for how to get it (it may also be available through your package manager).

2. All shell files must be passed through shfmt (specifically shfmt -ci -i 4 -w) and emit no shellcheck warnings.

The presubmit checks will do these things for you and fail if the code is not formatted properly or has a shellcheck warning. Running these tools on the command line before submitting will make it easier to get your code merged.

Additionally, please write a detailed commit message about why you are submitting your change.

Getting help

Please open an issue on this repo and include your distribution, shell, the command you ran, and the error output.