Fix TLS dtor + memory leaks on macOS

[RalfJung]

We currently ignore memory leaks on macOS. Likely, this is due to not properly executing TLS destructors in _tlv_atexit. Debugging this requires a libstd with full MIR for macOS, which currently we do not know how to do cross-platform -- so, someone who's actually using macOS should take this.

[mtak-]

IIUC this was fixed here: rust-lang/rust#57534 rust-lang/rust#57655

[RalfJung]

The issue here is about Miri entirely ignoring _tlv_atexit. This is a Miri issue, not a rustc issue, and I don't think these PRs fix it.

[kungfukennyg]

I'd like to take a crack at this issue. I have a macOS laptop, just need some guidance on where to get started. 😄

[RalfJung]

@kennethbgoodin great! Not knowing the macOS APIs in detail, I can only guess here. The first step is to make sure that you can locally build and run Miri; please let me know if you need any help with that.

Then remove macOS from this line to enable leak checking on macOS. ./miri test should fail now. Note which test(s) failed so that you can run them directly to check your changes!

My guess for what you will need to implement is this stub. Unfortunately I found no docs for _tlv_atexit, but it is instructive to see how Rust calls that function, which happens here. You will have to add a new field to Miri's Machine to remember the fn ptr that is to be called on exit, and then call that function around here.

These are high-level instructions; I can go into more detail once you have concrete questions. :)

[kungfukennyg]

Okay I've done the first steps and have which tests are failing (168!).

From looking at what rustc does it sounds like our _tlv_atexit needs to accept a thread local list of destructors for each thread's globals. I'm not sure how I would go about constructing that list though.

Adding the field to Machine and calling that function ptr sounds straightforward enough.

[RalfJung]

Okay I've done the first steps and have which tests are failing (168!).

That's likely the same cause for all of them, just pick one (let's say ./miri run tests/run-pass/bools.rs) and work on that for now.

From looking at what rustc does it sounds like our _tlv_atexit needs to accept a thread local list of destructors for each thread's globals. I'm not sure how I would go about constructing that list though.

Why a list? All I can see is a single function pointer, and a pointer that is passed to that function as an argument.