Make bindgen apple-aware

[adincebic]

Use apple_support.run(...) when bindgen executes on macOS.

[dzbarsky]

I think ideally the apple_support toolchain would compute this resource-dir in the discovery repo rule and add that to the toolchain's args explicitly, instead of working around it here.

Failing that, I suppose we can bring back this logic which I wasn't a fan of.

@keith wdyt?

[keith]

I guess it doesn't do that since we happily rely on the default? how does bindgen discover that? not from the active compiler?

[dzbarsky]

Yeah, right now apple_support relies on default sibling. Bindgen in rules_rust had logic above to try to guess it. In rules_rs version I removed it and expect cc toolchain to provide all flags it needs (explicit resource dir or -isystem etc)

[keith]

I don't think it's safe to assume many toolchains will have that? quick check shows toolchains_llvm does pass it, but rules_cc, apple_support, and musl-toolchains don't. It would be a bit annoying for us since we'd have to compute it in a repo rule that we don't otherwise do stuff like this in, and then we'd have to relativize the non-hermetic path to one of the placeholder things variables, which also means then you'd have to be handling those replacements (since otherwise it would contain the absolute path to Xcode)

[dzbarsky]

@keith both hermetic-llvm and toolchains_llvm have it, I think musl-toolchain is not very useful now that hermetic-llvm exists, and I'm not sure how much I care about the rules_cc auto toolchains...

But if you think computing it properly in apple_support is too annoying, I guess we can bring back the old detection method, it's just super janky...

[keith]

I don't think that version works for non-hermetic toolchains since all_files doesn't contain those files either?

[dzbarsky]

so has this never worked with rules_rust? @adincebic do you know?

[adincebic]

I think it did, let me check and report back.

[dzbarsky]

btw i was thinking, what if you configure your own bindgen toolxhain that invokes xcrun in a wrapper and then execs underlying bindgen. could even be a wrapper peocided by apple_support that you compose. feels a bit cleaner to me than having rust learn about this?