ptr::swap_nonoverlapping can no longer copy pointers bytewise in latest miri

[tower120]

After updating miri from pre-1.60 to 1.62, I start getting those errors:
https://github.com/tower120/any_vec/runs/6425015915?check_suite_focus=true
(that branch runs OK with older MIRI)

Previously, miri had troubles with any raw pointer deference operations, for me.
But I "fixed" that with ptr::copy_nonoverlapping - and it worked for miri.
It looks like there was some support for std::ptr::* operations on miri side... Which is now gone...

[RalfJung]

Looks like you are using the 2022-04-04 nightly, which is quite old. We had some trouble with copy_nonoverlapping a few weeks ago, but that has since been fixed.

Can you try again with the latest nightly?

[RalfJung]

Or maybe I looked into the wrong parts of the log? But this looks a lot like rust-lang/rust#94371...

Do you have a self-contained example to reproduce the problem? That would help with debugging.

copy_nonoverlapping is definitely still supported, but you cannot use it to copy a part of a pointer -- you have to copy the entire thing in one go. But that has always been like that, so not sure how the upgrade could have changed anything.

[tower120]

Can you try again with the latest nightly?

I'm not very good with rustup, but aren't it says 2022-05-11?

nightly-x86_64-unknown-linux-gnu installed - rustc 1.62.0-nightly (6dd68402c 2022-05-11)

https://github.com/tower120/any_vec/runs/6425015915?check_suite_focus=true#step:3:35

On my local Windows machine I can't install latest nightly, because it missing MIRI.

[tower120]

copy_nonoverlapping is definitely still supported, but you cannot use it to copy a part of a pointer -- you have to copy the entire thing in one go...

It does copy entire element with alignment, but represented as u8 byte array.

[tower120]

Here is working branch https://github.com/tower120/any_vec/blob/exp_drop_after_remove

MIRI problem occurs here:
https://github.com/tower120/any_vec/blob/exp_drop_after_remove/src/lib.rs#L462

I don't have smaller example at hand now.

[RalfJung]

That's a lot of code, I won't really have time to dig into that I am afraid. :/

(FWIW, it is a bad idea to represent pointers as a u8 array. u8 as an integer type does not carry provenance. The recommended type to represent data of arbitrary type is MaybeUninit<_>. But that will probably not make a difference here.)

[tower120]

Hmm, I tried to replace swap with 3 ptr::copy_nonoverlapping and intermediate vec - and it worked!

// same as copy_bytes but for swap_nonoverlapping.
#[inline]
unsafe fn swap_bytes(src: *mut u8, dst: *mut u8, count: usize){
    // MIRI hack
    if cfg!(miri) {
        let mut tmp = Vec::<u8>::new();
        tmp.resize(count, 0);
    
        // src -> tmp
        ptr::copy_nonoverlapping(src, tmp.as_mut_ptr(), count);
        // dst -> src
        ptr::copy_nonoverlapping(dst, src, count);
        // tmp -> dst
        ptr::copy_nonoverlapping(tmp.as_ptr(), dst, count);
    
        return;
    }

    // this does not work with MIRI
    // "unable to turn pointer into raw bytes"
    for i in 0..count{
        let src_pos = src.add(i);
        let dst_pos = dst.add(i);

        let tmp = *src_pos;
        *src_pos = *dst_pos;
        *dst_pos = tmp;
    }
}

The thing is

ptr::swap_nonoverlapping(src, dst, count);

does not work...

[RalfJung]

Btw:

// This is faster then ptr::copy_nonoverlapping,
// when count is runtime value, and count is small.

That sounds like a bug worth reporting against rustc; a naive manual impl should not be faster than the stdlib method.

[tower120]

The recommended type to represent data of arbitrary type is MaybeUninit<_>.

What about type-erased data?

[RalfJung]

That's what I mean -- the correct type to store type-erased data in Rust is an array of MaybeUninit<u8>.

A u8 (a) has to be initialized, and (b) cannot be a part of a pointer ("byte with provenance").

[RalfJung]

Looking at the implementation of swap_nonoverlapping, this is indeed not surprising: it just copies the data one T at a time. Your T is u8, so this copies the pointers byte-per-byte, and Miri does not support this.

swap_nonoverlapping used to be implemented in a different way, that's why you didn't see these problems before.

I wonder if this is even a correct implementation, since it means this is a typed copy...

[tower120]

If you found the problem - you don't need that branch anymore?

[RalfJung]

I added a comment in rust-lang/rust#63159 (comment) to further discuss this.

For now, you can work around the problem by using swap rather than swap_nonoverlapping.

[RalfJung]

Oh, but swap doesn't take a parameter to do this N times... your work-around with 3 ptr::copy_nonoverlapping is probably the best option for now then.

[RalfJung]

Closing this in favor of #2181.