Rust: ConstProp miscompilation around references

[RalfJung]

https://github.com/rust-lang/rust/pull/71911 introduced a (nightly-to-nightly) regression in mir-opt-level=3: since that landed, this assertion fails when running the test in Miri.

[RalfJung]

https://github.com/rust-lang/rust/issues/73223 is also still reproducible with mir-opt-level=2, but @oli-obk said there's a plan to hide it behind a flag that better expresses "this can and will break valid code".

[RalfJung]

In fact even level 2 is broken.

[RalfJung]

... and level 1, for the union issue (so let's treat the other one separately).

[RalfJung]

This doesn't require nightly after all, we have a stable-to-nightly regression when just running this code:

fn main() {
    #[allow(dead_code)]
    union U {
        f1: u32,
        f2: f32,
    }
    let mut u = U { f1: 1 };
    unsafe {
        let b1 = &mut u.f1;
        *b1 = 5;
    }
    assert_eq!(unsafe { u.f1 }, 5);
}

gives

thread 'main' panicked at 'assertion failed: `(left == right)`
  left: `1`,
 right: `5`', src/main.rs:12:5

[RalfJung]

In fact we don't even need a union, nor an unsafe block:

fn main() {
    #[allow(dead_code)]
    struct U {
        f1: u32,
        f2: f32,
    }
    let mut u = U { f1: 1, f2: 1.0 };
    let b1 = &mut u.f1;
    *b1 = 5;
    assert_eq!( { u.f1 }, 5);
}

[jonas-schievink]

Minimized:

fn main() {
    let mut u = (1,);
    *&mut u.0 = 5;
    assert_eq!( { u.0 }, 5);
}

[oli-obk]

on it

[Dylan-DPC]

giving this p-critical based on our discussion

[wesleywiser]

Why is this regression-from-stable-to-beta when the offending PR only merged last night?

[Aaron1011]

@RalfJung: Why didn't the assertion failure cause a Miri toolstate change?

[RalfJung]

I think this is stable-to-nightly only... @oli-obk why did you change that?

@Aaron1011 rustc's CI currently only tests Miri with mir-opt-level=0.

[oli-obk]

I tested it in the playground and I thought it also happened on beta?

[oli-obk]

Hmm... not sure what happened, must have mis-clicked on the playground