Unwinding intrinsics use non-unwinding ABI

[tmiasko]

There are intrinsics like assert_inhabited, which can unwind, but use rust-intrinsic ABI that never unwinds:

rust/compiler/rustc_middle/src/ty/layout.rs

Lines 1091 to 1093 in 6d651a2

	| RustIntrinsic
	| PlatformIntrinsic
	| Unadjusted => false,

An example demonstrating the issue. The argument passed to function f is never dropped, when assert_inhabited unwinds:

#![feature(core_intrinsics)]
#![feature(never_type)]

struct NoisyDrop(&'static str);

impl Drop for NoisyDrop {
    fn drop(&mut self) {
        println!("{}", self.0);
    }
}

pub fn f<A, B>(_: A) {
    core::intrinsics::assert_inhabited::<B>()
}

fn main() {
    let _a = NoisyDrop("1");
    f::<NoisyDrop, !>(NoisyDrop("2"));
}

$ rustc a.rs && ./a
thread 'main' panicked at 'attempted to instantiate uninhabited type `!`', a.rs:13:5
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
1

[nbdd0121]

I'll work on this after #102906 is merged.

@rustbot claim

[RalfJung]

Are these intrinsics the only ones that unwind? Maybe we should just make them abort instead by calling the (fairly recent) panic_str_nounwind.

However, AFAIK try is also an intrinsic and that can unwind?

[nbdd0121]

try catches unwind and it does not itself unwind.

[nbdd0121]

My current idea is to change assert_inhabited to is_inhabited intrinsic, and then have a wrapper that checks the value and panic if true. But making them abort sounds like an easier alternative.

[RalfJung]

It should be a nice abort via panic_str_nounwind though.

I am honestly not sure why the intrinsics are panicking rather than returning a boolean... I think I tried that and it didn't work? But yeah bool-returning intrinsics sound better (less magic) ayway.

[RalfJung]

One possible reason is that in CTFE and Miri, these intrinsics will actually abort instead of panic. Getting the if ... { panic_str_nounwind } to work in CTFE will be a bit of work (extending the existing hacks that make regular panicking work in CTFE, or using const_eval_select) and the backtrace will be slightly worse in Miri.

So given that we want something different implementations for Miri, CTFE, and runtime, maybe going with 'call panic_str_nounwind instead' is the better option. (An immediate-abort-panic just didn't exist when I added these intrinsics, otherwise this is probably what I would have done from the start.)

[RalfJung]

With #105997, I think assert_inhabited and friends do now indeed never unwind.

But as determined in rust-lang/miri#2839, we still have const_eval_select, which very much can unwind.

[nbdd0121]

Yeah, I originally thought that we can remove all unwinding intrinsics, but unfortunately that's not true. I am working on a PR to change the ABI to be unwindable and stick #[rustc_nounwind] to most intrinsics.

[RalfJung]

That sounds like the cleanest approach, thanks!