when suggesting RUST_BACKTRACE=1, add a special note for Miri's env var isolation

[RalfJung]

Fixes rust-lang/miri#2855

[rustbot]

r? @Nilstrieb

rustbot has assigned @Nilstrieb.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[rustbot]

The Miri subtree was changed

cc @rust-lang/miri

[Nilstrieb]

@bors r+ rollup

[bors]

📌 Commit 3e63398 has been approved by Nilstrieb

It is now in the queue for this repository.

[workingjubilee]

Huh.

@RalfJung Would you rather we always print a backtrace when we panic while running in Miri? That was one of the solutions discussed in rust-lang/miri#2855

Note this is different from "setting RUST_BACKTRACE=1", it won't affect user library behavior.

[workingjubilee]

Well, approximately-equal-to one of the solutions.

[RalfJung]

@RalfJung Would you rather we always print a backtrace when we panic while running in Miri? That was one of the solutions discussed in rust-lang/miri#2855

I don't think that would be a good default for Miri, just like it is not a good default for regular execution.

[workingjubilee]

🤔 It would seem to me you usually would want a backtrace? But okay!

[saethlin]

just like it is not a good default for regular execution.

Funny you should say that. Every codebase at my employer has this at the top of main:

    if std::env::var("RUST_BACKTRACE").is_err() {
        std::env::set_var("RUST_BACKTRACE", "1");
    }    

And I know from a discussion on Mastodon a few days ago that we are not the only ones doing this.

[workingjubilee]

Yes. I think we would want to default to it on, except we may wish to get better at printing shorter backtraces first, so that RUST_BACKTRACE=1 strikes the right balance of output and informative (or maybe RUST_BACKTRACE=0 just becomes a very short backtrace by default and =1 is longer).

[RalfJung]

Okay, let me rephrase: I don't think Miri's default should be different from that of regular Rust.

[workingjubilee]

Ah, I see! I only figured that since Miri strongly functions as an error-reporting tool, that you might consider deviating on this point acceptable. But I do understand preferring to avoid differences in execution, considering it mostly is good at error-reporting by being good at simulation.

[RalfJung]

Almost every time we made Miri divert from regular execution, we eventually decided that was a bad idea. Maybe this is one of the few exceptions, but it doesn't seem obvious to me.

[saethlin]

Yeah; on balance I think the extra note is the best approach.