`thread::scope` clobbers the panic payload

[SludgePhD]

I tried this code:

use std::{panic::catch_unwind, thread};

fn main() {
    let res = catch_unwind(|| {
        thread::scope(|scope| {
            scope.spawn(|| panic!("my panic message with useful information"));
        });
    });
    let payload = res.unwrap_err();
    let message = payload.downcast_ref::<&str>().unwrap();
    dbg!(message);
}

I expected to see this happen: the panic payload from the scoped thread should be forwarded to the main thread using panic::resume_unwind. The payload passed to panic! inside the scope should be retrievable there, and there should not be any additional panics.

Instead, this happened: a second panic with no link to the original is caused by the thread::scope implementation, and the payload is a meaningless "a scoped thread panicked" message.

Meta

rustc --version --verbose:

rustc 1.85.1

beta and nightly also behave like this

[theemathas]

The scope() documentation states:

If you want to handle panics from spawned threads, join them before the end of the scope.

You can get the payload with the following code:

use std::thread;

fn main() {
    let res = thread::scope(|scope| -> thread::Result<()> {
        let join_handle = scope.spawn(|| panic!("my panic message with useful information"));
        join_handle.join()?;
        // Maybe join other threads here
        Ok(())
    });
    let payload = res.unwrap_err();
    let message = payload.downcast_ref::<&str>().unwrap();
    dbg!(message);
}

[SludgePhD]

Sure, but there is no good reason for the behavior in the first place. Structured concurrency primitives like thread::scope have the opportunity to perfectly forward panics across thread boundaries, it would be a waste not to use it.

[hanna-kruppe]

What do you expect to happen the implicit join at the end of the scope finds more than one panicked thread?

[SludgePhD]

Then any of the panics can be forwarded (which is the same behavior as rayon's). The code running in the threads also has to ensure that it doesn't cause any knock-on panics for this strategy to yield a "useful" panic payload, of course.