Build using llvm-config instead of $prefix/lib/cmake/llvm

[jedbrown]

When the path in -DLLVM_DIR=/path/to/llvm does not exist or does not contain the necessary CMake files, Enzyme silently falls back to a system version (if it exists). This is the case when building for Rust, in which $prefix/lib contains only libLLVM-17-rust-1.77.0-nightly.so (no cmake/llvm subdirectory). It would be a bit of a chore to ensure these CMake files are available in all contexts because sysroot is distributed multiple ways.

libLLVM.so, the shared object file for the LLVM project. In-tree this is either built from source or downloaded from CI; out-of-tree, you need to install llvm-tools-preview with rustup.

However, llvm-config is always available. How hard would it be for Enzyme's build to configure itself entirely from llvm-config, without the CMake files for LLVM?

[wsmoses]

I'm okay with that, though I'm not sure how difficult it would be a-priori.

I'm sure someone has done this in the past for LLVM, so perhaps we can find something?

[jedbrown]

CHIP-SPV has a start, though it looks like it only needs command-line tools. I see Enzyme's CMakeLists.txt has include(AddLLVM), which is a big file. If we only needed the variables like LLVM_DEFINITIONS, LLVM_INCLUDE_DIRS, etc., then it would be pretty easy. There is src/llvm-project/llvm/cmake/modules/AddLLVM.cmake inside the rust repo so perhaps it can be called, but I've never seen this sort of CMake task not snowball.

[vchuravy]

However, llvm-config is always available

That is not the case for cross-compilation llvm-config is only available on the host and my be wrong for the target system.

[jedbrown]

Interesting, I'm trying to work out whether that is important for Rust. During bootstrap, stage 1 is always for the host. I don't think it's documented how a target libLLVM is produced when cross-compiling, but we must eventually need it.

I don't think we would ever need to execute llvm-config on the target. When executed on the host, is llvm-config unable to specify how to build for the target?

[vchuravy]

When executed on the host, is llvm-config unable to specify how to build for the target?

Sadly. llvm-config only knows about the llvm build it comes from, not the one you are creating. There might be ways around this with a staged build, but those have been very fragile for us and for binarybuilder/Yggdrasil we went with relying on the cmake files.