Nixpkgs: Support Bazel on macOS

[TravisWhitaker]

In principle it should be possible to build Bazel from source on macOS. Bazel itself is quite a mess to patch, since xcrun is sometimes invoked from the PATH, sometimes absolutely from /usr/bin, and sometimes via an xcrunwrapper.sh script.

What has me stumped is that Bazel grabs a bunch of dependencies whose build systems also rely on xcrun, particularly protobuf and grpc. Many of the calls don't straightforwardly do something like /usr/bin/xcrun clang ..., but instead query various SDK locations and feed them into further wrapper scripts. Unfortunately the only success I've had has been allowing these build systems access to /usr/bin/xcrun. For my use case an impure-but-working Bazel derivation on Darwin is better than no working Bazel derivation on Darwin, especially considering that the impurity is confined to Xcode's CLT and libraries. From what I understand there are two reasonable paths forward.

• Accept this impurity on Darwin, but do so in a principled way. Surely there's some existing idiomatic way of doing this. I've seen mention of a "pure Darwin stdenv" and an "impure Darwin stdenv" in discussions elsewhere, but I can't seem to find any documentation on these.

• Build some heavyweight hacks into the Bazel build system. This might be something like inserting extra machinery to patch the external dependencies that Bazel grabs, or perhaps writing a fake xcrun command. I think this would involve a lot of work for questionable benefit; because of Xcode it seems that larger degree of impurity tends to be acceptable when it comes to Darwin in Nixpkgs at large.

@copumpkin and @abbradar, I'm curious about your thoughts.

[TravisWhitaker]

Regrettably, I'm tied to Bazel because of TensorFlow. Furthermore, I'm calling TF from C++ (already a use case poorly supported upstream), for which none of upstream's binary distributions are suitable given my hardware, so I must build TF from source. It's only due to Nix that deploying this monstrosity works at all. A while ago there were a few community supported alternative build systems for TF, but they've all diverged too much for my use case. A working Bazel derivation on macOS would allow my colleagues to build our application and run tests locally without a virtual machine.

It is indeed a shame that the "state of the art" in machine learning is in such a pitiful state.

[grahamc]

For what it is worth, williammpratt's attitude here is not representative of the NixOS community's attitude toward other software. The user has since been blocked from the NixOS organization.

[jyp]

Probably a duplicate of #30590
Can these issues be merged?

[pSub]

Closing this issue in favor of #30590.