Consider emitting C/C++ instead of LLVM IR

[opti-mix]

PythoC currently generates LLVM IR, which is great!

But to make it more portable and available on more exotic target platforms, one could also consider emitting C/C++ instead of LLVM IR. This would make the emitted code more readable, possible to inspect and change if needed and more importantly, possible to compile with a target C/C++ compiler on systems where LLVM toolchains may be unavailable (yet). BTW, IIRC Nim actually generates C/C++ instead of e.g. LLVM IR directly. So, there is definitely a prior art related to this idea.

Is it something technically doable?

Does PythoC use any LLVM IR or LLVM features that cannot be represented in C/C++ form?

Is current IR easy and rich enough to be used as a basis for generating C/C++ or it needs to be extended quite a bit to make it possible?

It would be really interesting to hear your thoughts about this idea.

[1flei]

Will also consider the C backend. However, some features in pythoc (e.g. code generation and meta programming) cannot be generated to C directly without explicitly specification (like C++ template). I would like pythoc to do more things in itself to make use of the meta-programming rather than a simple C code generator.

[jason-s]

Seconded, and I might be able to contribute towards this effort as part of my day job, depending on the constraints PythoC requires to do this. I have taken some crude efforts toward this general topic, but with the intent of making readable C code that mirrors the author's intent, rather than creating any optimization at the Python -> C step.

[1flei]

Seconded, and I might be able to contribute towards this effort as part of my day job, depending on the constraints PythoC requires to do this. I have taken some crude efforts toward this general topic, but with the intent of making readable C code that mirrors the author's intent, rather than creating any optimization at the Python -> C step.

That would be great, and contributions are very welcome. However, although the project already has backend and builder classes, the frontend and backend are not fully separated yet. Some codes still rely on LLVM backend indirectly, so I don't think adding a C backend and codegen will work cleanly at this stage. Also, there are some zero-sized types like proof stuff, which may require some work around in pure C.
I think a reasonable way to contribute could be first identifing the LLVM-specific codes and fully separating the frontend and backend (maybe by a minimum working example) and then add the C backend and codegen.