build-llvm.md

Building LLVM Yourself

First, you need to clone the LLVM repository and check out the llvmorg-19.1.6 tag. For example, the following commands would check out the right commit (I would strongly recommend running these outside this repository):

$ git clone https://github.com/llvm/llvm-project.git llvm
$ cd llvm
$ git checkout llvmorg-19.1.6

There are some scripts for building LLVM on different operating systems under ci/, but the instructions below guide you through building LLVM by hand.

macOS/Linux

Try the following. Say Release instead of RelWithDebInfo if you don't need debug symbols and want to save some disk space. You should also set -DLLVM_PARALLEL_LINK_JOBS=1 instead of 2 if you have less than ~32 GB of RAM. You may want to leave out enabling assertions too, if you care about the performance impact, but this may cause you other problems — see the Debugging guide of the README). Also, on macOS, you don't need to bother specifying -DLLVM_USE_LINKER=gold since lld is fast enough (the GNU ld installed on a typical GNU/Linux system is the slow one). Also, -DLLVM_INSTALL_UTILS=TRUE is important if you want to run the test suite, since it installs the LLVM FileCheck tool we use in testing.

$ mkdir build && cd build
$ cmake -G Ninja -DLLVM_ENABLE_PROJECTS=mlir -DCMAKE_BUILD_TYPE=RelWithDebInfo -DLLVM_ENABLE_ASSERTIONS=TRUE -DLLVM_PARALLEL_LINK_JOBS=2 -DCMAKE_INSTALL_PREFIX=/path_to_llvm_install_dir/ -DLLVM_TARGETS_TO_BUILD=Native -DLLVM_USE_LINKER=gold -DLLVM_INSTALL_UTILS=TRUE ../llvm
$ ninja install

You'll need to do the following as well:

$ export MLIR_DIR=/path_to_llvm_install_dir/lib/cmake/mlir/
$ export PATH=$PATH:/path_to_llvm_install_dir/bin/

You will probably want to put both exports in your ~/.bashrc.

Windows

I had luck following the LLVM guide for LLVM in Visual Studio. Here is some Qwerty-specific advice to complement their guide:

1. Use the git clone command above with C:\qwerty as your current directory (that is, clone LLVM at C:\qwerty\llvm-project). Check out the tag llvmorg-19.1.6.
2. When you install Python, if you want to do a debug build of the Qwerty compiler, you should check "Download debugging symbols" and "Downloaded debug binaries" under "Advanced Options" in the installer. This will install python312_d.lib (as opposed to only python312.lib ), without which you may see some linker errors when you do a debug build. (Note, however, that debug builds are currently broken on Windows due to other linker problems.)
3. I used the following cmake command line inside an administrator "x64 Native Tools Command Prompt for VS 2022":

   cmake -S llvm\llvm -B build -DLLVM_ENABLE_PROJECTS=mlir -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=TRUE -DCMAKE_INSTALL_PREFIX=C:\qwerty\llvm19 -DLLVM_TARGETS_TO_BUILD=Native -DLLVM_ENABLE_DIA_SDK=OFF -Thost=x64 -DLLVM_INSTALL_UTILS=TRUE
   

   (Turning off DIA with -DLLVM_ENABLE_DIA_SDK=OFF is more important than it looks; without it, the generated LLVM CMake scripts end up making the Qwerty CMake scripts attempt to link with C:/Program Files/Microsoft Visual Studio/2022/Community/DIA SDK/lib/amd64/diaguids.lib, which may not exist on some systems, namely GitHub Actions hosts.)
4. The guide mentions this, but make sure that when you open C:\qwerty\llvm\, you choose "Release" as your configuration here:
   
5. To build, you can right-click "ALL_BUILD" under "CMakePredefinedTargets" in Solution Explorer and choose "Build":
   
   This took about an hour on my machine, during which it was otherwise unusable.
6. To install, you can right-click "INSTALL" in the same place and again choose "Build". If you skip the previous step, this fails (in my experience).
7. Press Start and type in "environment variables" (I would recommend editing system variables, not user variables). Set MLIR_DIR to C:\qwerty\llvm19\lib\cmake\mlir and add C:\qwerty\llvm19\bin to Path