BuildingLLVM.md

In duncantl/Rllvm: Interface to llvm for dynamically compiling native code.

To install LLVM, you can download a binary distribution from http://llvm.org/releases/download.html. These are pre-built binaries that mostly work. Of course, you have a great deal of more flexibility (and more work) by creating your own version of LLVM.

Using the downloadable .tar.xz file is very simple. You extract the file with, e.g.,

tar xf clang+llvm-3.8.0-x86_64-apple-darwin.tar.xz

Then we can use the binaries directly within the sub-directory, i.e.,

export LLVM_CONFIG=$HOME/Downloads/clang+llvm-3.8.0-x86_64-apple-darwin/bin/llvm-config

Building From Source

We can build LLVM and clang from source. The instructions are available at, for example, http://clang.llvm.org/get_started.html

Update

tar xzf ~/Downloads/llvm-project-llvmorg-15.0.6.tar.gz

From the top-level directory of the untarred archive

cmake -S llvm -B build -G "Unix Makefiles" -DCMAKE_INSTALL_PREFIX=$HOME/LLVM/local -DCMAKE_BUILD_TYPE=RelWithDebInfo -DLLVM_ENABLE_PROJECTS="clang;lldb;lld;flang;clang-tools-extra" -DLLVM_ENABLE_RUNTIMES="libcxx;libcxxabi" -DLLDB_USE_SYSTEM_DEBUGSERVER=ON

Then compile from within the build/ directory:

cd build
make -j 8

• Change the target directory (CMAKE_INSTALL_PREFIX) if we want to have multiple parallel installations, e.g. local_v18

• And then make install

Older Guidelines

It is often useful to compile clang at the same time (to ensure the same version). For this, see below.

On some machines, you will need to have a more recent version of GCC to build LLVM and/or clang. You can install this in your own directory. You install GCC from source with, e.g., configure --prefix=$HOME/local to install the resulting software into $HOME/local/{bin,lib}

When building LLVM with this new version of GCC, you should add $HOME/local/bin to your PATH environment variable and also include $HOME/local/lib in your LD_LIBRARY_PATH variable.

If the configuration for LLVM still fails with an error message about the version of libstdc++ being insufficient, you can explicitly set the environment variables CC and CXX to point to the fully qualified paths to gcc and g++, respectively, e.g.

export CC=$HOME/local/bin/gcc export CXX=$HOME/local/bin/g++

Then re-run the configuration script for LLVM.

cmake -DCMAKE_INSTALL_PREFIX=/usr  -DLLVM_ENABLE_FFI=ON -DLLVM_BUILD_LLVM_DYLIB=ON \
-DCMAKE_BUILD_TYPE=Release -DLLVM_TARGETS_TO_BUILD="host;X86;NVPTX;CppBackend" ..

(Don't forget the .. at the end.)

Compiling Clang with LLVM

See https://quuxplusone.github.io/blog/2019/11/09/llvm-from-scratch/ or in the README.md file for monorepo source code.

The build takes 32G on my OSX machine

cmake -G Ninja 
   -DLLVM_ENABLE_PROJECTS="clang;libcxx;libcxxabi;clang-tools-extra;lldb;compiler-rt" 
   -DDEFAULT_SYSROOT="$(xcrun --show-sdk-path)"
   -DCMAKE_BUILD_TYPE=RelWithDebInfo 
   -DCMAKE_INSTALL_PREFIX=$HOME/local ../llvm

Download the clang source code, named cfe, extra clang tools, compiler-rt (runtime), libcxx. You extract these source archives into the following sub-directories of the top-level llvm source tree:

|code|directory|Notes| |------|------|-----| |cfe|tools/clang|rename the cfe- directory to clang| |clang-tools-extra|tools/clang/tools/|| |compiler-rt|projects| |libcxx|projects|

The extract top-level directories such as libcxx-4.0.1-src should be renamed to, e.g., libcxx.

Alternative Instructions

An alternative description of how to build LLVM is available at http://www.linuxfromscratch.org/blfs/view/svn/general/llvm.html

On Linux

On an older linux machine with gcc 4.?, etc. I had to install + ninja + swig (4.0.2) + openssl (1.1.1i) + cmake (3.19.3 as LLVM 12.0 will require cmake 3.13) + gcc 10

before installing LLVM from source.

Since I installed these into a local directory rather than replacing the system versions, I needed to update PATH and LD_LIBRARY_PATH to, e.g., $HOME/local/bin and $HOME/local/lib64 respectively (if I used $HOME/local as the --prefix argument during the installations.)

Calling external routines and the Foreign Function Interface

In order to be able to invoke existing routines that we do not create within our own modules, LLVM needs to be compiled with support for libffi - the foreign function interface.

Install the libffi package (downloading it from https://sourceware.org/libffi/) and then follow the installation instructions in INSTALL.md

Coincidentally, we have an R package that provides libffi in R - Rffi - that allows one to invoke arbitrary compiled (C) routines. There is also the rdyncall package that provides another facility for this with some additional features. The Rllvm package uses the Rffi package to invoke the routines we generate with LLVM.

duncantl/Rllvm documentation built on April 23, 2024, 6:14 p.m.