Cross-compilation fixes for Apple M1/x86

[kornelski]

Cross-compilation was assuming a setup typical for Linux. Cross-compilation from ARM macOS to Intel macOS has its own quirky requirements, because it uses the same system-global compiler, but needs different flags instead.

Also fixes a bug where cfg!(target_os) was used. build.rs runs on the host OS, so compilation target of build.rs is always the same as the host.

[kornelski]

@zmwangx Hello. Can you have a look at this?

[Polochon-street]

Hi! Thanks for your submission :)
Could you fix the linting errors so we can see whether it builds successfully?

[kornelski]

This is still broken in 5.0.0.

[Polochon-street]

Is there really no way to use rust's target_os! or something more standard? I'm afraid this will bite us if other architectures start to pop up

[kornelski]

Rust's target definitions are Rust's, and have no guarantee of compatibility with anything, e.g. Rust uses macos, but the rest of the world uses darwin.

So yes, it is already broken, and will keep breaking on every new OS.

[Polochon-street]

would it be possible to have something that just checks if target_os! is macos, and replace it by darwin if it is? And maybe that does that for other OSes, if it happens for others? Maybe just a fn target_os() that wraps target_os! and does the changes or something

[kornelski]

I suspect it's needed for more OSes, but I don't have them to test. Whatever you do, it's going to be an incomplete list that may break on OSes that haven't been tested.

[Polochon-street]

Quick question, but what's the value of CARGO_CFG_TARGET_OS on macos? Can't you just use a wrapper that gets CARGO_CFG_TARGET_OS and if it's macos, replace it by darwin?

[kornelski]

CARGO_CFG_TARGET_* variables are derived from rustc's cfg values, same as cfg!(target_*), so they use the same naming scheme.

I can hardcode the result to be correct only on darwin/macos, and leave it incorrect everywhere else, but why should I leave it incorrect everywhere else?

[Polochon-street]

my question is just, why do this triple splitting magic (and explode if it doesn't work) instead of just using CARGO_CFG_TARGET and use a hashmap or something like that to substitute the wrong stuff with whatever you know is right? Or maybe I'm missing the point here?

[kornelski]

Split doesn't work, because Rust's triples are inconsistently named, and sometimes have the "vendor" and sometimes they don't, so you don't know where the OS part starts.

HashMap would work for a closed set. I'm trying to make it work for an open set.

[Polochon-street]

my point was that maybe we could do a hashmap like {'macos': 'darwin'} and then just have a custom target_os function that would just return either the string itself, if it's not in the hashmap? Like

fn target_os(target: &str) -> &str {
    let mut os_replacements = HashMap::new();

    os_replacements.insert(
        "macos".to_string(),
         "darwin".to_string(),
     );
     os_replacements.get(&target).unwrap_or("target")
}

But I feel like we're talking about different things though, so I could be completely wrong 😄

https://github.com/rustsec/rustsec/blob/main/platforms/src/target/os.rs#L112-L134 seems to have a pretty consistent list of OS if we wanna list more stuff here?

[emfrosztovis]

Any updates on this? Cross-compilation to x86_64 seems still broken on macOS 🥲