world::World doctests sometimes hangs with Miri

[RalfJung]

When running the world::World (line 56) doctest in Miri with many different seeds, it eventually runs into an endless loop. This can be avoided by using -Zmiri-disable-weak-memory-emulation. This probably indicates a bug in bevy or its dependencies, but could also be a bug in Miri.

Weak memory emulation makes it so that an atomic load can return "outdated" values, as is permitted by the concurrent memory model of Rust (and C/C++). So likely, some code somewhere reads an old value, thinks it is current, and goes into some kind of spin loop -- and stays there forever.

See here for further analysis by @cbeuw.

[LegNeato]

Possibly relevant: https://github.com/kanidm/concread

[RalfJung]

The fact that they advise to disable Stacked Borrows does not inspire a ton of confidence though. ;)

FWIW, latest Miri has -Zmiri-track-weak-memory-loads, which should help diagnose this issue. But OTOH @cbeuw pretty much figured it out already I think; someone probably needs to go fix smol-rs/concurrent-queue.

[RalfJung]

Ah looks like this is tracked in concurrent-queue: smol-rs/concurrent-queue#17

[taiki-e]

The problem in concurrent-queue that we agreed on is that it was using x86-specific fences that Miri does not understand. This has been fixed in concurrent-queue 1.2.4 (smol-rs/concurrent-queue@d61005f).

[RalfJung]

that Miri does not understand

To be more precise, using those fences is just wrong when writing Rust code, and Miri understands and detects that correctly. They can only be used correctly when writing inline assembly.