io::Error is UnwindSafe, but only on 64-bit platforms

[KamilaBorowska]

Code

I tried this code:

use std::io;
use std::panic::UnwindSafe;

pub fn is_unwind_safe<T: UnwindSafe>() {}

fn main() {
    is_unwind_safe::<io::Error>();
}

I expected to see this happen: either it to compile or not compile.

Instead, this happened: it compiled but only on 64-bit platforms.

Version it worked on

It most recently worked on: Rust 1.59 (where it did not compile no matter what platform)

Version with regression

rustc --version --verbose:

rustc 1.60.0-beta.5 (6ee5a408a 2022-03-18)
binary: rustc
commit-hash: 6ee5a408a3bbc76598637d50450165917f8756bb
commit-date: 2022-03-18
host: x86_64-unknown-linux-gnu
release: 1.60.0-beta.5
LLVM version: 14.0.0

Backtrace

Backtrace

<backtrace>

@rustbot modify labels: +regression-from-stable-to-beta -regression-untriaged

[compiler-errors]

Not sure what the best approach is here.

We could just impl !UnwindSafe for io::Error to restore consistency from #87869 which made io::Error auto-implement UnwindSafe.

cc #94791

[apiraino]

@xfix which non-64bit platform does reproduce the issue (not sure if this question matters a lot)?

[KamilaBorowska]

@xfix which non-64bit platform does reproduce the issue (not sure if this question matters a lot)?

All platforms that have std and not(target_pointer_width = "64"). For example, i686-unknown-linux-gnu. This is caused by this code:

rust/library/std/src/io/error.rs

Lines 4 to 12 in 547369d

	#[cfg(target_pointer_width = "64")]
	mod repr_bitpacked;
	#[cfg(target_pointer_width = "64")]
	use repr_bitpacked::Repr;
	#[cfg(not(target_pointer_width = "64"))]
	mod repr_unpacked;
	#[cfg(not(target_pointer_width = "64"))]
	use repr_unpacked::Repr;

[thomcc]

Ah, repr_bitpacked::Repr should have PhantomData<ErrorData<Box<Custom>>> or something along those lines. I'll get a fix for this up later this evening or tomorrow.

[thomcc]

Hmm, this should probably not be closed, unless prioritization will still see it if it is closed.

[compiler-errors]

Yeah, for backports I think the issue stays open until the backport is landed or decided against. Reopened for tracking.