Build Clang toolchain

Fuchsia is using Clang as the official compiler.

Prerequisites

You need CMake version 3.13.4 or newer to execute these commands. This is the minimum required version to build LLVM.

While CMake supports different build systems, it is recommended to use Ninja.

Both should be present in your Fuchsia checkout as prebuilts. The commands below assume that cmake and ninja are in your PATH:

export PATH=${FUCHSIA}/prebuilt/third_party/cmake/${platform}/bin:${PATH}
export PATH=${FUCHSIA}/prebuilt/third_party/ninja/${platform}/bin:${PATH}

Getting Source

The example commands below use ${LLVM_SRCDIR} to refer to the root of your LLVM source tree checkout. You can use the official monorepo https://github.com/llvm/llvm-project maintained by the LLVM community:

LLVM_SRCDIR=${HOME}/llvm/llvm-project
git clone https://github.com/llvm/llvm-project ${LLVM_SRCDIR}

Note: It is recommended checking out to the revision that's currently used for Fuchsia. The latest upstream revision may be broken or fail to build Fuchsia, whereas it is guaranteed that the prebuilt revision can always build Fuchsia. This revision can be found in [//integration/toolchain]. Search for the package fuchsia/third_party/clang/${platform}, and checkout the git_revision associated with it.

cd ${LLVM_SRCDIR}
git checkout ${REVISON_NUMBER}

Fuchsia IDK

Before building the runtime libraries that are built along with the toolchain, you need a Fuchsia IDK (formerly known as the SDK). The IDK must be located in the directory pointed to by the ${IDK_DIR} variable:

IDK_DIR=${HOME}/fuchsia-idk

To download the latest IDK, you can use the following:

# For Linux
cipd install fuchsia/sdk/core/linux-amd64 latest -root ${IDK_DIR}

# For macOS
cipd install fuchsia/sdk/core/mac-amd64 latest -root ${IDK_DIR}

Sysroot for Linux

To include compiler runtimes and C++ library for Linux, download the sysroot. It must be located in the directory pointed by the ${SYSROOT_DIR} variable.

SYSROOT_DIR=${HOME}/fuchsia-sysroot/

To download the latest sysroot, you can use the following:

cipd install fuchsia/third_party/sysroot/linux latest -root ${SYSROOT_DIR}

{% dynamic if user.is_googler %}

[Googlers only] Goma

Goma is a service for accelerating builds by distributing compilations across many machines. Googlers should ensure Goma is installed on your machine for faster builds. If you have Goma installed in ${GOMA_DIR} (which should be provided in //prebuilt/third_party/goma/${platform}), you can enable Goma by adding these extra CMake flags to your CMake invocation:

  -DCMAKE_C_COMPILER_LAUNCHER=${GOMA_DIR}/gomacc \
  -DCMAKE_CXX_COMPILER_LAUNCHER=${GOMA_DIR}/gomacc \

Then you can take advantage of Goma by allowing multiple ninja jobs to run in parallel:

ninja -j1000

Use -j100 for Goma on macOS and -j1000 for Goma on Linux. You may need to tune the job count to suit your particular machine and workload.

Warning: The examples below assume you can use Goma. If you cannot use Goma, do not add the provided CMake flags or use an absurdly high number of jobs.

Note: In order to use Goma, you need a host compiler that is supported by Goma such as the Fuchsia Clang installation. To verify your compiler is available on Goma, you can set GOMA_USE_LOCAL=0 GOMA_FALLBACK=0 environment variables. If the compiler is not available, you will see an error.

{% dynamic endif %}

Building a Clang Toolchain for Fuchsia

The Clang CMake build system supports bootstrap (aka multi-stage) builds. Fuchsia uses two-stage bootstrap build for the Clang compiler. However, for toolchain related development it is recommended to use the single-stage build.

If your goal is to experiment with clang, the single-stage build is likely what you are looking for. The first stage compiler is a host-only compiler with some options set needed for the second stage. The second stage compiler is the fully optimized compiler intended to ship to users.

Setting up these compilers requires a lot of options. To simplify the configuration the Fuchsia Clang build settings are contained in CMake cache files, which are part of the Clang codebase (Fuchsia.cmake and Fuchsia-stage2.cmake).

In the following CMake invocations, ${CLANG_TOOLCHAIN_PREFIX} refers to the directory of binaries from a previous Clang toolchain. Normally, this refers to the current toolchain shipped with Fuchsia, but any references to binaries from this directory could theoretically be replaced with one's own binaries.

# FUCHSIA_SRCDIR refers to the root directory of your Fuchsia source tree
CLANG_TOOLCHAIN_PREFIX=${FUCHSIA_SRCDIR}/prebuilt/third_party/clang/linux-x64/bin/

Note: Clang must be built in a separate build directory. The directory itself can be a subdirectory or in a whole other path.

mkdir llvm-build
mkdir llvm-install  # For placing stripped binaries here
INSTALL_DIR=${pwd}/llvm-install
cd llvm-build

Single Stage Build Fuchsia Configuration

When developing Clang for Fuchsia, you can use the cache file to test the Fuchsia configuration, but run only the second stage, with LTO disabled, which gives you a faster build time suitable even for incremental development, without having to manually specify all options:

cmake -G Ninja -DCMAKE_BUILD_TYPE=Debug \
  -DCMAKE_C_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang \
  -DCMAKE_CXX_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang++ \
  -DCMAKE_ASM_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang \
  -DCMAKE_C_COMPILER_LAUNCHER=${GOMA_DIR}/gomacc \
  -DCMAKE_CXX_COMPILER_LAUNCHER=${GOMA_DIR}/gomacc \
  -DCMAKE_ASM_COMPILER_LAUNCHER=${GOMA_DIR}/gomacc \
  -DLLVM_ENABLE_LTO=OFF \
  -DLINUX_x86_64-unknown-linux-gnu_SYSROOT=${SYSROOT_DIR} \
  -DLINUX_aarch64-unknown-linux-gnu_SYSROOT=${SYSROOT_DIR} \
  -DFUCHSIA_SDK=${IDK_DIR} \
  -DCMAKE_INSTALL_PREFIX= \
  -C ${LLVM_SRCDIR}/clang/cmake/caches/Fuchsia-stage2.cmake \
  ${LLVM_SRCDIR}/llvm
ninja distribution  -j1000  # Build the distribution

If the above fails with an error related to Ninja, then you may need to add ninja to your PATH. You can find the prebuilt executable at //prebuilt/third_party/ninja/${platform}/bin.

ninja distribution should be enough for building all binaries, but the Fuchsia build assumes some libraries are stripped so ninja install-distribution-stripped is necessary.

Caution: Due to a bug in Clang, builds with assertions enabled might crash while building Fuchsia. As a workaround, you can disable Clang assertions by setting -DLLVM_ENABLE_ASSERTIONS=OFF or using a release build (-DCMAKE_BUILD_TYPE=Release).

Two-Stage Build Fuchsia Configuration

This is roughly equivalent to what is run on the prod builders and used to build a toolchain that Fuchsia ships to users.

cmake -GNinja \
  -DCMAKE_C_COMPILER=${CLANG_TOOLCHAIN_PREFIX}/clang \
  -DCMAKE_CXX_COMPILER=${CLANG_TOOLCHAIN_PREFIX}/clang++ \
  -DCMAKE_ASM_COMPILER=${CLANG_TOOLCHAIN_PREFIX}/clang \
  -DCMAKE_C_COMPILER_LAUNCHER=${GOMA_DIR}/gomacc \
  -DCMAKE_CXX_COMPILER_LAUNCHER=${GOMA_DIR}/gomacc \
  -DCMAKE_ASM_COMPILER_LAUNCHER=${GOMA_DIR}/gomacc \
  -DCMAKE_INSTALL_PREFIX= \
  -DSTAGE2_LINUX_aarch64-unknown-linux-gnu_SYSROOT=${SYSROOT_DIR} \
  -DSTAGE2_LINUX_x86_64-unknown-linux-gnu_SYSROOT=${SYSROOT_DIR} \
  -DSTAGE2_FUCHSIA_SDK=${IDK_DIR} \
  -C ${LLVM_SRCDIR}/clang/cmake/caches/Fuchsia.cmake \
  ${LLVM_SRCDIR}/llvm
ninja stage2-distribution -j1000
DESTDIR=${INSTALL_DIR} ninja stage2-install-distribution-stripped -j1000

Note: The second stage build uses LTO (Link Time Optimization) to achieve better runtime performance of the final compiler. LTO often requires a large amount of memory and is very slow. Therefore it may not be very practical for day-to-day development.

runtime.json

If the Fuchsia build fails due to a missing runtime.json file, you must generate a new runtime.json file by running the following command:

python3 ${FUCHSIA_SRCDIR}/scripts/clang/generate_runtimes.py  \
  --clang-prefix ${INSTALL_DIR} --sdk-dir ${IDK_DIR}          \
  --build-id-dir ${INSTALL_DIR}/lib/.build-id > ${INSTALL_DIR}/lib/runtime.json

The generated file contains relative paths used by the Fuchsia build to know where various libraries from the toolchain are located.

Putting it All Together

Copy-paste code for building a single-stage toolchain. This code can be run from inside your LLVM build directory and assumes a linux environment.

cd ${LLVM_BUILD_DIR}  # The directory your toolchain will be installed in

# Environment setup
FUCHSIA_SRCDIR=${HOME}/fuchsia/  # Replace with wherever Fuchsia lives
LLVM_SRCDIR=${HOME}/llvm/llvm-project  # Replace with wherever llvm-project lives
IDK_DIR=${HOME}/fuchsia-idk/
SYSROOT_DIR=${HOME}/fuchsia-sysroot/
CLANG_TOOLCHAIN_PREFIX=${FUCHSIA_SRCDIR}/prebuilt/third_party/clang/linux-x64/bin/
GOMA_DIR=${FUCHSIA_SRCDIR}/prebuilt/third_party/goma/linux-x64/

# Download necessary dependencies
cipd install fuchsia/sdk/core/linux-amd64 latest -root ${IDK_DIR}
cipd install fuchsia/third_party/sysroot/linux latest -root ${SYSROOT_DIR}

# CMake invocation
cmake -G Ninja -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_C_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang \
  -DCMAKE_CXX_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang++ \
  -DCMAKE_C_COMPILER_LAUNCHER=${GOMA_DIR}/gomacc \
  -DCMAKE_CXX_COMPILER_LAUNCHER=${GOMA_DIR}/gomacc \
  -DLLVM_ENABLE_LTO=OFF \
  -DLINUX_x86_64-unknown-linux-gnu_SYSROOT=${SYSROOT_DIR} \
  -DLINUX_aarch64-unknown-linux-gnu_SYSROOT=${SYSROOT_DIR} \
  -DFUCHSIA_SDK=${IDK_DIR} \
  -DCMAKE_INSTALL_PREFIX= \
  -C ${LLVM_SRCDIR}/clang/cmake/caches/Fuchsia-stage2.cmake \
  ${LLVM_SRCDIR}/llvm

# Build and strip binaries and place them in the install directory
ninja distribution -j1000
DESTDIR=${INSTALL_DIR} ninja install-distribution-stripped -j1000

# Generate runtime.json

python3 ${FUCHSIA_SRCDIR}/scripts/clang/generate_runtimes.py    \
  --clang-prefix ${INSTALL_DIR} --sdk-dir ${IDK_DIR}            \
  --build-id-dir ${INSTALL_DIR}/lib/.build-id > ${INSTALL_DIR}/lib/runtime.json

Building Fuchsia with a Custom Clang

To specify a custom clang toolchain for building Fuchsia, pass --args clang_prefix=\"${INSTALL_DIR}/bin\" --no-goma to fx set command and run fx build.

fx set core.x64 --args=clang_prefix=\"${INSTALL_DIR}/bin\" --no-goma
fx build

This file contains relative paths used by the Fuchsia build to know where various libraries from the toolchain are located.

Note: If you make another change to Clang after building Fuchsia with a previous version of Clang, re-running fx build may not always guarantee that all necessary targets will be built with the new Clang. For this case, should instead run fx clean-build, which will rebuild everything but definitely use the new Clang.

Developing Clang

When developing Clang, you may want to use a setup that is more suitable for incremental development and fast turnaround time.

The simplest way to build LLVM is to use the following commands:

cmake -GNinja \
  -DCMAKE_BUILD_TYPE=Debug \
  -DLLVM_ENABLE_PROJECTS="clang;clang-tools-extra;lld" \
  ${LLVM_SRCDIR}/llvm
ninja

You can enable additional projects using the LLVM_ENABLE_PROJECTS variable. To enable all common projects, you would use:

  -DLLVM_ENABLE_PROJECTS="clang;clang-tools-extra;lld;compiler-rt;libcxx;libcxxabi;libunwind"

Similarly, you can also enable some projects to be built as runtimes which means these projects will be built using the just-built rather than the host compiler:

  -DLLVM_ENABLE_PROJECTS="clang;clang-tools-extra;lld" \
  -DLLVM_ENABLE_RUNTIMES="compiler-rt;libcxx;libcxxabi;libunwind" \

Both LLVM_ENABLE_PROJECTS and LLVM_ENABLE_RUNTIMES are already set in the CMake cache files, so you normally don't need to set these unless you would like to explicitly add more projects or runtimes.

Clang is a large project and compiler performance is absolutely critical. To reduce the build time, it is recommended to use Clang as a host compiler, and if possible, LLD as a host linker. These should be ideally built using LTO and for best possible performance also using Profile-Guided Optimizations (PGO).

To set the host compiler to Clang and the host linker to LLD, you can use the following extra flags:

  -DCMAKE_C_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang \
  -DCMAKE_CXX_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang++ \
  -DLLVM_ENABLE_LLD=ON

This assumes that ${CLANG_TOOLCHAIN_PREFIX} points to the bin directory of a Clang installation, with a trailing slash (as this Make variable is used in the Zircon build). For example, to use the compiler from your Fuchsia checkout (on Linux):

CLANG_TOOLCHAIN_PREFIX=${FUCHSIA}/prebuilt/third_party/clang/linux-x64/bin/

Note: To build Fuchsia, you need a stripped version of the toolchain runtime binaries. Use DESTDIR=${INSTALL_DIR} ninja install-distribution-stripped to get a stripped install and then point your build configuration to ${INSTALL_DIR}/bin as your toolchain.

Sanitizers

Most sanitizers can be used on LLVM tools by adding LLVM_USE_SANITIZER=<sanitizer name> to your cmake invocation. MSan is special however because some LLVM tools trigger false positives. To build with MSan support you first need to build libc++ with MSan support. You can do this in the same build. To set up a build with MSan support first run CMake with LLVM_USE_SANITIZER=Memory and LLVM_ENABLE_LIBCXX=ON.

cmake -GNinja \
  -DCMAKE_BUILD_TYPE=Debug \
  -DCMAKE_C_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang \
  -DCMAKE_CXX_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang++ \
  -DLLVM_ENABLE_PROJECTS="clang;clang-tools-extra;lld;libcxx;libcxxabi;libunwind" \
  -DLLVM_USE_SANITIZER=Memory \
  -DLLVM_ENABLE_LIBCXX=ON \
  -DLLVM_ENABLE_LLD=ON \
  ${LLVM_SRCDIR}/llvm

Normally you would run Ninja at this point but we want to build everything using a sanitized version of libc++ but if we build now it will use libc++ from ${CLANG_TOOLCHAIN_PREFIX}, which isn't sanitized. So first we build just the cxx and cxxabi targets. These will be used in place of the ones from ${CLANG_TOOLCHAIN_PREFIX} when tools dynamically link against libcxx

ninja cxx cxxabi

Now that you have a sanitized version of libc++ you can set your build to use it instead of the one from ${CLANG_TOOLCHAIN_PREFIX} and then build everything.

ninja

Putting that all together:

cmake -GNinja \
  -DCMAKE_BUILD_TYPE=Debug \
  -DCMAKE_C_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang \
  -DCMAKE_CXX_COMPILER=${CLANG_TOOLCHAIN_PREFIX}clang++ \
  -DLLVM_USE_SANITIZER=Address \
  -DLLVM_ENABLE_LIBCXX=ON \
  -DLLVM_ENABLE_LLD=ON \
  ${LLVM_SRCDIR}/llvm
ninja libcxx libcxxabi
ninja

Testing Clang

To run Clang tests, you can use the check-<component> target:

ninja check-llvm check-clang

You can all use check-all to run all tests, but keep in mind that this can take significant amount of time depending on the number of projects you have enabled in your build.

To test only one specific test, you can use the environment variable LIT_FILTER. If the path to the test is clang/test/subpath/testname.cpp, you can use:

LIT_FILTER=testname.cpp ninja check-clang

The same trick can be applied for running tests in other sub-projects by specifying different a different check-<component>.

{% dynamic if user.is_googler %}

[Googlers only] Building Fuchsia with custom Clang on bots

Fuchsia's infrastructure has support for using a non-default version of Clang to build. Only Clang instances that have been uploaded to CIPD or Isolate are available for this type of build, and so any local changes must land in upstream and be built by the CI or production toolchain bots.

You will need the infra codebase and prebuilts. Directions for checkout are on the infra page.

To trigger a bot build with a specific revision of Clang, you will need the Git revision of the Clang with which you want to build. This is on the CIPD page, or can be retrieved using the CIPD CLI. You can then run the following command:

export FUCHSIA_SOURCE=<path_to_fuchsia>
export BUILDER=<builder_name>
export REVISION=<clang_revision>

export INFRA_PREBUILTS=${FUCHSIA_SOURCE}/fuchsia-infra/prebuilt/tools

cd ${FUCHSIA_SOURCE}/fuchsia-infra/recipes

${INFRA_PREBUILTS}/led get-builder 'luci.fuchsia.ci:${BUILDER}' | \
${INFRA_PREBUILTS}/led edit-recipe-bundle -O | \
jq '.userland.recipe_properties."$infra/fuchsia".clang_toolchain.type="cipd"' | \
jq '.userland.recipe_properties."$infra/fuchsia".clang_toolchain.instance="git_revision:${REVISION}"' | \
${INFRA_PREBUILTS}/led launch

It will provide you with a link to the BuildBucket page to track your build.

You will need to run led auth-login prior to triggering any builds, and may need to file an infra ticket to request access to run led jobs.

[Googlers Only] Downloading Toolchains from CAS

Our Clang Toolchain CI builders upload all build artifacts to Content Addressed Storage (CAS). It provides a convenient way to quickly download a specific toolchain without having to build from scratch. This can greatly speedup investigations into toolchain issues, since you can avoid the long LLVM build times, on top of building Fuchsia.

Below is an example of how download a specific toolchain into a corpus directory using cas:

$ ${INFRA_PREBUILTS}/cas download -cas-instance chromium-swarm -digest \
    ad53e1f315a849955190594fde6b07e11e76b40563db5779fcc69d6a6e04dc71/267 -dir corpus

In the example above the -digest field is passed a unique id, which is used by the cas tool to fetch the correct artifacts. The digest can be obtained from Fuchsia's CI builder by selecting the builder you want the toolchain from, then expanding the clang->cas->archive fields and then clicking on the CAS_UI link. The resulting page will give show you some information about the CAS upload, including the digest.

{% dynamic endif %}

Downloading Toolchains from CAS

Our Clang Toolchain CI builders upload all build artifacts to Content Addressed Storage (CAS). It provides a convenient way to quickly download a specific toolchain without having to build from scratch. This can greatly speedup investigations into toolchain issues, since you can avoid the long LLVM build times, on top of building Fuchsia.

Below is an example of how to install the cas tool from scratch and download a specific toolchain into a corpus directory:

$ cipd install infra/tools/luci/cas/linux-amd64 latest -root luci
$ ./luci/cas download -cas-instance chromium-swarm -digest \
    ad53e1f315a849955190594fde6b07e11e76b40563db5779fcc69d6a6e04dc71/267 -dir corpus

In the example above the -digest field is passed a unique id, which is used by the cas tool to fetch the correct artifacts. The digest can be obtained from Fuchsia's CI builder by selecting the builder you want the toolchain from, then expanding the clang->cas->archive fields and then clicking on the CAS_UI link. The resulting page will give show you some information about the CAS upload, including the digest.

Useful CMake Flags

There are many other CMake flags that are useful for building, but these are some that may be useful for toolchain building.

-DLLVM_PARALLEL_LINK_JOBS

Increase the number of link jobs that can be run in parallel (locally). The number of link jobs is dependent on RAM size. For LTO build you will need at least 10GB for each job.

Additional Resources

Documentation:

Talks: