Getting Started

---
local:
---

Building flang

There are two ways to build flang. The first method is to build it at the same time that you build all of the projects on which it depends. This is called building in tree. The second method is to first do an in tree build to create all of the projects on which flang depends. Then, after creating this base build, only build the flang code itself. This is called building standalone. Building standalone has the advantage of being smaller and faster. Once you create the base build and base install areas, you can create multiple standalone builds using them.

Note that instructions for building LLVM can be found at https://llvm.org/docs/GettingStarted.html.

All of the examples below use GCC as the C/C++ compilers and ninja as the build tool.

Building flang in tree

Building flang in tree means building flang along with all of the projects on which it depends. These projects include mlir, clang, flang, openmp, and compiler-rt. Note that compiler-rt is only needed to access libraries that support 16 bit floating point numbers. It's not needed to run the automated tests. You can use several different C++ compilers for most of the build, includig GNU and clang. But building compiler-rt requres using the clang compiler built in the initial part of the build.

Here's a directory structure that works. Create a root directory for the cloned and built files. Under that root directory, clone the source code into a directory called llvm-project. The build will also create subdirectories under the root directory called build (holds most of the built files), install (holds the installed files, and compiler-rt (holds the result of building compiler-rt).

Here's a complete set of commands to clone all of the necessary source and do the build.

First, create the root directory and cd into it.

mkdir root
cd root

Now clone the source:

git clone https://github.com/llvm/llvm-project.git

Once the clone is complete, execute the following commands:

rm -rf build
mkdir build
rm -rf install
mkdir install
ROOTDIR=`pwd`
INSTALLDIR=$ROOTDIR/install

cd build

cmake \
  -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_PREFIX=$INSTALLDIR \
  -DCMAKE_CXX_STANDARD=17 \
  -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
  -DCMAKE_CXX_LINK_FLAGS="-Wl,-rpath,$LD_LIBRARY_PATH" \
  -DFLANG_ENABLE_WERROR=ON \
  -DLLVM_ENABLE_ASSERTIONS=ON \
  -DLLVM_TARGETS_TO_BUILD=host \
  -DLLVM_LIT_ARGS=-v \
  -DLLVM_ENABLE_PROJECTS="clang;mlir;flang;openmp" \
  -DLLVM_ENABLE_RUNTIMES="compiler-rt" \
  ../llvm-project/llvm

ninja

On Darwin, to make flang able to link binaries with the default sysroot without having to specify additional flags, use the DEFAULT_SYSROOT CMake flag, e.g. -DDEFAULT_SYSROOT="$(xcrun --show-sdk-path)".

By default flang tests that do not specify an explicit --target flag use LLVM's default target triple. For these tests, if there is a need to test on a different triple by overriding the default, the following needs to be added to the cmake command above: -DLLVM_TARGET_TRIPLE_ENV="<some string>" -DFLANG_TEST_TARGET_TRIPLE="<your triple>".

To run the flang tests on this build, execute the command in the “build” directory:

ninja check-flang

To create the installed files:

ninja install

echo "latest" > $INSTALLDIR/bin/versionrc

To build compiler-rt:

cd $ROOTDIR
rm -rf compiler-rt
mkdir compiler-rt
cd compiler-rt
CC=$INSTALLDIR/bin/clang \
CXX=$INSTALLDIR/bin/clang++ \
cmake \
  -G Ninja \
  ../llvm-project/compiler-rt \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_PREFIX=$INSTALLDIR \
  -DCMAKE_CXX_STANDARD=11 \
  -DCMAKE_C_CFLAGS=-mlong-double-128 \
  -DCMAKE_CXX_CFLAGS=-mlong-double-128 \
  -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
  -DCOMPILER_RT_BUILD_ORC=OFF \
  -DCOMPILER_RT_BUILD_XRAY=OFF \
  -DCOMPILER_RT_BUILD_MEMPROF=OFF \
  -DCOMPILER_RT_BUILD_LIBFUZZER=OFF \
  -DCOMPILER_RT_BUILD_SANITIZERS=OFF \
  -DLLVM_CONFIG_PATH=$INSTALLDIR/bin/llvm-config

ninja
ninja install

Note that these instructions specify flang as one of the projects to build in the in tree build. This is not strictly necessary for subsequent standalone builds, but doing so lets you run the flang tests to verify that the source code is in good shape.

Building flang standalone

To do the standalone build, start by building flang in tree as described above. This build can be used as the base build for several subsequent standalone builds. Set the environment variable ROOT_DIR to the directory that contains the subdirectory build that was created previously, for example:

export ROOTDIR=/home/user/root

Start each standalone build the same way by cloning the source for llvm-project:

mkdir standalone
cd standalone
git clone https://github.com/llvm/llvm-project.git

Once the clone is complete, execute the following commands:

cd llvm-project/flang
rm -rf build
mkdir build
cd build

cmake \
  -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_CXX_STANDARD=17 \
  -DCMAKE_CXX_LINK_FLAGS="-Wl,-rpath,$LD_LIBRARY_PATH" \
  -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
  -DFLANG_ENABLE_WERROR=ON \
  -DLLVM_TARGETS_TO_BUILD=host \
  -DLLVM_ENABLE_ASSERTIONS=ON \
  -DLLVM_BUILD_MAIN_SRC_DIR=$ROOTDIR/build/lib/cmake/llvm \
  -DLLVM_EXTERNAL_LIT=$ROOTDIR/build/bin/llvm-lit \
  -DLLVM_LIT_ARGS=-v \
  -DLLVM_DIR=$ROOTDIR/build/lib/cmake/llvm \
  -DCLANG_DIR=$ROOTDIR/build/lib/cmake/clang \
  -DMLIR_DIR=$ROOTDIR/build/lib/cmake/mlir \
  ..

ninja

To run the flang tests on this build, execute the command in the flang/build directory:

ninja check-flang

Building flang runtime for accelerators

Flang runtime can be built for accelerators in experimental mode, i.e. complete enabling is WIP. CUDA and OpenMP target offload builds are currently supported.

Building out-of-tree

CUDA build

Clang with NVPTX backend and NVCC compilers are supported.

cd llvm-project/flang
rm -rf build_flang_runtime
mkdir build_flang_runtime
cd build_flang_runtime

cmake \
  -DFLANG_EXPERIMENTAL_CUDA_RUNTIME=ON \
  -DCMAKE_CUDA_ARCHITECTURES=80 \
  -DCMAKE_C_COMPILER=clang \
  -DCMAKE_CXX_COMPILER=clang++ \
  -DCMAKE_CUDA_COMPILER=clang \
  -DCMAKE_CUDA_HOST_COMPILER=clang++ \
  ../runtime/
make -j FortranRuntime

Note that the used version of clang must support CUDA toolkit version installed on the build machine. If there are multiple CUDA toolkit installations, please use -DCUDAToolkit_ROOT=/some/path to specify the compatible version.

cd llvm-project/flang
rm -rf build_flang_runtime
mkdir build_flang_runtime
cd build_flang_runtime

cmake \
  -DFLANG_EXPERIMENTAL_CUDA_RUNTIME=ON \
  -DCMAKE_CUDA_ARCHITECTURES=80 \
  -DCMAKE_C_COMPILER=clang \
  -DCMAKE_CXX_COMPILER=clang++ \
  -DCMAKE_CUDA_COMPILER=nvcc \
  -DCMAKE_CUDA_HOST_COMPILER=clang++ \
  ../runtime/

make -j FortranRuntime

Note that nvcc might limit support to certain versions of CMAKE_CUDA_HOST_COMPILER, so please use compatible versions.

The result of the build is a “fat” library with the host and device code. Note that the packaging of the libraries is different between Clang and NVCC, so the library must be linked using compatible compiler drivers.

Building in-tree

One may build Flang runtime library along with building Flang itself by providing these additional CMake variables on top of the Flang in-tree build config:

For example:

  -DFLANG_EXPERIMENTAL_CUDA_RUNTIME=ON \
  -DCMAKE_CUDA_ARCHITECTURES=80 \
  -DCMAKE_C_COMPILER=clang \
  -DCMAKE_CXX_COMPILER=clang++ \
  -DCMAKE_CUDA_COMPILER=clang \
  -DCMAKE_CUDA_HOST_COMPILER=clang++ \

Or:

  -DFLANG_EXPERIMENTAL_CUDA_RUNTIME=ON \
  -DCMAKE_CUDA_ARCHITECTURES=80 \
  -DCMAKE_C_COMPILER=gcc \
  -DCMAKE_CXX_COMPILER=g++ \
  -DCMAKE_CUDA_COMPILER=nvcc \
  -DCMAKE_CUDA_HOST_COMPILER=g++ \

Normal make -j check-flang will work with such CMake configuration.

OpenMP target offload build

Only Clang compiler is currently supported.

cd llvm-project/flang
rm -rf build_flang_runtime
mkdir build_flang_runtime
cd build_flang_runtime

cmake \
  -DFLANG_EXPERIMENTAL_OMP_OFFLOAD_BUILD="host_device" \
  -DCMAKE_C_COMPILER=clang \
  -DCMAKE_CXX_COMPILER=clang++ \
  -DFLANG_OMP_DEVICE_ARCHITECTURES="all" \
  ../runtime/

make -j FortranRuntime

The result of the build is a “device-only” library, i.e. the host part of the library is just a container for the device code. The resulting library may be linked to user programs using Clang-like device linking pipeline.

The same set of CMake variables works for Flang in-tree build.

Build options

One may provide optional CMake variables to customize the build. Available options:

  • -DFLANG_RUNTIME_F128_MATH_LIB=libquadmath: enables build of FortranFloat128Math library that provides REAL(16) math APIs for intrinsics such as SIN, COS, etc. GCC libquadmath's header file quadmath.h must be available to the build compiler. More details.

Supported C++ compilers

Flang is written in C++17.

The code has been compiled and tested with GCC versions from 7.2.0 to 9.3.0.

The code has been compiled and tested with clang version 7.0, 8.0, 9.0 and 10.0 using either GNU‘s libstdc++ or LLVM’s libc++.

The code has been compiled on AArch64, x86_64 and ppc64le servers with CentOS7, Ubuntu18.04, Rhel, MacOs, Mojave, XCode and Apple Clang version 10.0.1.

Note that flang is not supported on 32 bit CPUs.

Building flang with GCC

By default, cmake will search for g++ on your PATH. The g++ version must be one of the supported versions in order to build flang.

Or, cmake will use the variable CXX to find the C++ compiler. CXX should include the full path to the compiler or a name that will be found on your PATH, e.g. g++-8.3, assuming g++-8.3 is on your PATH.

export CXX=g++-8.3

or

CXX=/opt/gcc-8.3/bin/g++-8.3 cmake ...

Building flang with clang

To build flang with clang, cmake needs to know how to find clang++ and the GCC library and tools that were used to build clang++.

CXX should include the full path to clang++ or clang++ should be found on your PATH.

export CXX=clang++

Installation Directory

To specify a custom install location, add -DCMAKE_INSTALL_PREFIX=<INSTALL_PREFIX> to the cmake command where <INSTALL_PREFIX> is the path where flang should be installed.

Build Types

To create a debug build, add -DCMAKE_BUILD_TYPE=Debug to the cmake command. Debug builds execute slowly.

To create a release build, add -DCMAKE_BUILD_TYPE=Release to the cmake command. Release builds execute quickly.

How to Run Tests

Flang supports 2 different categories of tests

  1. Regression tests (https://www.llvm.org/docs/TestingGuide.html#regression-tests)
  2. Unit tests (https://www.llvm.org/docs/TestingGuide.html#unit-tests)

For standalone builds

To run all tests:

cd ~/flang/build
cmake -DLLVM_DIR=$LLVM -DMLIR_DIR=$MLIR ~/flang/src
ninja check-all

To run individual regression tests llvm-lit needs to know the lit configuration for flang. The parameters in charge of this are: flang_site_config and flang_config. And they can be set as shown below:

<path-to-llvm-lit>/llvm-lit \
 --param flang_site_config=<path-to-flang-build>/test-lit/lit.site.cfg.py \
 --param flang_config=<path-to-flang-build>/test-lit/lit.cfg.py \
  <path-to-fortran-test>

Unit tests:

If flang was built with -DFLANG_INCLUDE_TESTS=ON (ON by default), it is possible to generate unittests. Note: Unit-tests will be skipped for LLVM install for an standalone build as it does not include googletest related headers and libraries.

There are various ways to run unit-tests.


1. ninja check-flang-unit 2. ninja check-all or ninja check-flang 3. <path-to-llvm-lit>/llvm-lit \ test/Unit 4. Invoking tests from <standalone flang build>/unittests/<respective unit test folder>

For in tree builds

If flang was built with -DFLANG_INCLUDE_TESTS=ON (ON by default), it is possible to generate unittests.

To run all of the flang unit tests use the check-flang-unit target:

ninja check-flang-unit

To run all of the flang regression tests use the check-flang target:

ninja check-flang

How to Generate Documentation

Generate FIR Documentation

If flang was built with -DLINK_WITH_FIR=ON (ON by default), it is possible to generate FIR language documentation by running ninja flang-doc. This will create <build-dir>/tools/flang/docs/Dialect/FIRLangRef.md in flang build directory.

Generate Doxygen-based Documentation

To generate doxygen-style documentation from source code

  • Pass -DLLVM_ENABLE_DOXYGEN=ON -DFLANG_INCLUDE_DOCS=ON to the cmake command.
cd ~/llvm-project/build
cmake -G Ninja -DLLVM_ENABLE_PROJECTS="clang;flang" -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_DOXYGEN=ON -DFLANG_INCLUDE_DOCS=ON ../llvm
ninja doxygen-flang

It will generate html in

    <build-dir>/tools/flang/docs/doxygen/html # for flang docs

Generate Sphinx-based Documentation

Flang documentation should preferably be written in markdown(.md) syntax (they can be in reStructuredText(.rst) format as well but markdown is recommended in first place), it is mostly meant to be processed by the Sphinx documentation generation system to create HTML pages which would be hosted on the webpage of flang and updated periodically.

If you would like to generate and view the HTML locally:

  • Install Sphinx, and the required extensions using pip install --user -r ~/llvm-projects/docs/requirements.txt
  • Pass -DLLVM_ENABLE_SPHINX=ON -DSPHINX_WARNINGS_AS_ERRORS=OFF to the cmake command.
cd ~/llvm-project/build
cmake -G Ninja -DLLVM_ENABLE_PROJECTS="clang;flang" -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_SPHINX=ON -DSPHINX_WARNINGS_AS_ERRORS=OFF ../llvm
ninja docs-flang-html

It will generate html in

   $BROWSER <build-dir>/tools/flang/docs/html/