miri no longer builds after rust-lang/rust#73441

[rust-highfive]

Hello, this is your friendly neighborhood mergebot.
After merging PR #73441, I observed that the tool miri has failing tests.
A follow-up PR to the repository https://github.com/rust-lang/miri is needed to fix the fallout.

cc @alexcrichton, do you think you would have time to do the follow-up work?
If so, that would be great!

And nominating for compiler team prioritization.

[oli-obk]

I don't think we can resolve this until #74506 is fixed, as we can't build a libstd anymore :(

[RalfJung]

However, that cannot be the only problem -- Miri also fails to build in-tree:

2020-07-18T16:41:13.8191493Z error[E0432]: unresolved import `rustc_demangle`
2020-07-18T16:41:13.8192822Z   --> /checkout/src/libstd/../backtrace/src/symbolize/mod.rs:13:5
2020-07-18T16:41:13.8193112Z    |
2020-07-18T16:41:13.8193595Z 13 | use rustc_demangle::{try_demangle, Demangle};
2020-07-18T16:41:13.8193892Z    |     ^^^^^^^^^^^^^^ use of undeclared type or module `rustc_demangle`
2020-07-18T16:41:13.8197541Z 
2020-07-18T16:41:15.8938727Z error: aborting due to previous error

@alexcrichton any idea why those might be missing?

[RalfJung]

rustc-demangle is an optional dependency of libstd, and Miri does not enable that option. Looks like the backtrace code incorrectly expects it to be present anyway?

[alexcrichton]

Ah this is my mistake, I didn't check that libstd still compiled without the backtrace dependency. I believe that unconditionally depending on rustc-demangle is indeed the right thing to do. I can send a PR for this tomorrow if no one else gets to it.

[RalfJung]

Just please make sure to also include it in rust-src then :)

[alexcrichton]

I'm hoping this will be fixed by #74540

[RalfJung]

You can test that via 'x.py test --stage 0 src/tools/miri'

[RalfJung]

That indeed solved the libstd build issue, at least for in-tree builds. Now the problem is

2020-07-20T06:48:53.8607834Z error: unsupported operation: can't call foreign function: _Unwind_Backtrace
2020-07-20T06:48:53.8608239Z   --> /checkout/src/libstd/../backtrace/src/backtrace/libunwind.rs:96:5
2020-07-20T06:48:53.8608410Z    |
2020-07-20T06:48:53.8608769Z 96 |     uw::_Unwind_Backtrace(trace_fn, &mut cb as *mut _ as *mut _);
2020-07-20T06:48:53.8609220Z    |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ can't call foreign function: _Unwind_Backtrace

Previously, disabling the backtrace feature meant we did not have to do anything. Now it seems like some C FFI (_Unwind_Backtrace) is mandatory? That's odd though, elsewhere @alexcrichton said this is now Rust-only.

[RalfJung]

Interestingly, _Unwind_Backtrace is only called on a double-panic.

[RalfJung]

rust-lang/backtrace-rs#360 should help (and that's way easier than finding all the methods we have to stub, in particular on Windows).

[alexcrichton]

Backtraces are unconditionally generated on double panics, and some reshuffling of the crate itself probably meant that #[cfg(miri)] happened to go to the noop unwinder before and now it went to the libunwind unwinder. In any case rust-lang/backtrace-rs#360 should fix that issue.