Feature request: use the Artifacts system to provide a C compiler on Linux and macOS

[DilumAluthge]

On Windows, PackageCompiler now uses the Artifacts system to provide a C compiler.

Would it be possible to use the Artifacts system to also provide a C compiler on Linux and macOS?

That would make it possible to build custom Julia sysimages on any of the three major operating systems without needing any additional dependencies, which I think would be really cool!

[KristofferC]

It is so super easy to get a compiler there so bit sure the extra download is worth it. For people making containers with PackageCompilerX it might make them significantly larger?

[DilumAluthge]

That's a good point. It would be nice to have as an optional feature maybe?

Is there a way to download Artifacts on-demand? As in, by default we don't download a C compiler on Linux or macOS, but if the user runs PackageCompiler.download_compiler(), then it downloads the artifact.

[KristofferC]

Yes, if the artifact is lazy=true in the Artifact file (this is actually the case already on Windows) so if you have JULIA_CC set it just uses that one (not documented right now though).

[DilumAluthge]

Yes, if the artifact is lazy=true in the Artifact file (this is actually the case already on Windows) so if you have JULIA_CC set it just uses that one (not documented right now though).

So perhaps we could add compilers for Linux and macOS to the Artifact file with lazy=true. And then we would have an exported function PackageCompiler.download_compiler() that calls the Pkg.Artifacts.artifact"..." macro and sets ENV["JULIA_CC"].

Then the workflow (for people that want to download the compiler on Linux and macOS) would be:

julia> using PackageCompiler

julia> download_compiler()

julia> create_sysimage(...)

[DilumAluthge]

In addition to setting ENV["JULIA_CC"], we'd probably need to set ENV["C_INCLUDE_PATH"] to make sure that the gcc we provide via Artifacts is able to find the correct header files.

[DilumAluthge]

I tried to get this working by providing Clang binaries (#326), but I kept running into strange issues, such as Clang trying (and failing) to find ncurses.

[KristofferC]

There hasn't been a single issue/post about this so I think the current situation is fine the way it is.