LLVM-based cross compilation

[Ericson2314]

LLVM is a much nicer for cross compilation. It has been multi-target for the vast majority, and its associated runtime libraries (e.g. compiler-rt, libcxx) are readily built separately from the compiler itself.

Unfortunately, our current packaging doesn't take advantage of either of these. We build compiler-rt in the same derivation as LLVM, and we pull everything from the same boostrappping stage, precluding cross compilation infra from working its magic.

I took a stab at this in obsidiansystems@acf2ed5 (the top commit in https://github.com/obsidiansystems/nixpkgs/tree/split-compiler-rt). I should get around to at least rebasing that, but even in its current state that serves as a basic reference to what should be accomplished.

---

CC @angerman @dtzWill

[Mic92]

Too bad it still cannot build the kernel: https://lwn.net/Articles/734071/

[eternaleye]

@Mic92: Um, to quote the article:

This work has shown that the FUD surrounding using a new toolchain for the kernel is unfounded, Desaulniers said. It is working now, but there are a few asterisks. Clang, the front end, can compile the kernel, but the assembler and the linker from GNU Binutils are needed to complete the build process.

(emphasis mine)

Reducing the use of gcc down to using gas for the kernel (rather than the full stack for everything) is pretty big in its own right, and using binutils' ld will likely be needed in various cases for a while anyway as lld gets up to speed.

In addition, this mail from one month later claims that the gas requirement boils down to this bug affecting x86, and works on other architectures (such as arm64).

Further improvement has likely occurred in the four months since then.

[Ericson2314]

Anyways, the point of this is to change how we package LLVM. Once we do that, we'll be able to attempt cross compiling any package we want, including Linux. Getting Linux to build with LLVM is an orthogonal concern.

[wmertens]

Does this mean your ultimate goal is to only build Nixpkgs with LLVM? Because that seems untenable :-/

[eternaleye]

@wmertens I'd argue that using clang, libc++, etc. would count as a distinct cross target - one potentially easier/simpler/cleaner to bootstrap.

[Ericson2314]

@wmertens No. I want to support as many combinations as we can without a a soup of conditionals. What we choose to support via CI resources I'll happily leave to others. Compare to NixOS/rfcs#23 not being about ditching glibc.

[Ericson2314]

@eternaleye Today, compiler and libc++ choices don't require a different platform name, but they could. libc ones do with -musl vs -gnu, for example, and in any event, the hashes will reflect the change like any other dependency change.

[matthewbauer]

@Ericson2314 Can this be closed?