Flaky CI in `cargo test -p c2rust-pdg` (snapshot)

[kkysen]

The cargo test step of CI is occasionally flaky due to the snapshot test in cargo test -p c2rust-pdg. Usually it can be resolved by just re-running CI.

The error shown is a difference in the snapshot, but I believe it's caused by the instrumented binary being terminated early (and silently), as the difference is always purely in the trace being cut off early.

[kkysen]

I believe this is caused by UB in analysis/test. When I reduced main_0 to just

unsafe fn main_0(mut argc: libc::c_int, mut argv: *mut *mut libc::c_char) -> libc::c_int {
    simple();
    exercise_allocator();
    simple_analysis();
}

it was reliably crashing the snapshot tests on every other run or so, locally, not just in CI.

When I fixed the UB in that (checked with cargo miri run), the crashes went away.

For example, there is a lot of UB in simple.

[kkysen]

The significantly rarer race condition in cargo add is a separate bug. It is much more rare, though, and we should be able to fix it by adding locking at some level (perhaps in the test code, and don't run cargo add in c2rust-instrument).

[kkysen]

Fixed by:

• Fix UB in analysis/test #683
• Flush the Backend buffer #691