MSVC: Implement 32-bit Unwinding

[alexcrichton]

Updated report

• 64-bit unwinding
• 32-bit unwinding (see comment below)

Original Report

Right now the MSVC target turns all calls to panic! into a process-wide abort as unwinding is not implemented.

• Update LLVM (rustc: Update LLVM #26025)
• https://llvm.org/bugs/show_bug.cgi?id=23884

WIP branch: https://github.com/alexcrichton/rust/tree/msvc-llvm-update (now merged)

[alexcrichton]

LLVM has been upgrade, this is now blocked on https://llvm.org/bugs/show_bug.cgi?id=23884 I believe.

[briansmith]

A question about this: Last time we talked, Alex mentioned that this was blocked on getting SEH-compatible unwinding to work on 32-bit Windows. According to Microsoft's documentation, in MSVC C++ mode, they recommend not to use the SEH-compatible unwinding for reasons of correctness and performance. See https://msdn.microsoft.com/en-us/library/1deeycx5.aspx.

Which MSVC /EH option is most similar to Rust's (proposed) handling of unwinding? And, is SEH-compatible unwinding really necessary?

[retep998]

@briansmith That article is recommending to not use a flag which would let C++ exceptions catch arbitrary SEH exceptions, things like stack overflow and divide by zero. Internally all the exceptions and unwinding in the msvc world are using SEH. Basically you're confused between the class of exceptions referred to as SEH exceptions and SEH as the unwinding/exception mechanism.

[alexcrichton]

Ah thanks for the ping @briansmith, I should give a status update on this. Currently unwinding on 64-bit MSVC is fully implemented using DWARF side tables like the GNU triple uses and we implement the necessary runtime support to parse these side tables. Unwinding on 32-bit MSVC is unimplemented, however. I have not investigated too closely what it would mean to implement DWARF unwinding for 32-bit like we have for 64-bit (I believe @vadimcn knows more about this than I do).

My current idea for implementing 32-bit is to wait for LLVM's exception handling intrinsics to be fully implemented. I'm under the impression they're relatively experimental for now and are unlikely to work for us (but I haven't checked in awhile). Once those exist we can switch to those for 32 and 64 bit MSVC, tweak some runtime support, and in theory "Just Work".

In any case, however, unwinding from C++ to Rust is undefined behavior no matter what happens (and vice versa), so exceptions/panics must always be caught at FFI boundaries. In that sense SEH compatible unwinding isn't necessary, but we need a solution for 32-bit still.

[briansmith]

SEH compatible unwinding isn't necessary, but we need a solution for 32-bit still.

Right. Maybe I misunderstood, but my impression is that people felt that resolving this was blocked on SEH-compatible unwinding. My point is that a lot of people targetting -msvc don't need or want SEH-compatible unwinding, especially if it has negative perf/size implications. That is, if this bug were resolved with (only) a SEH-compatible unwinder, then there should be a follow-up issue to track more efficient unwinding.

I believe the LLVM work is intended to support mixing MSVC-compiled C++ code with clang-compiled C++ code with exceptions that traverse stacks of code implemented with each compiler; i.e. 100% interop with the MSVC C++ ABI. That's quite a different problem from what Rust needs, AFAICT.

AFAICT, the requirements for the unwinder are just that implementations of Drop have their drop method called, and that a panic terminates the thread in a way that the thing that spawned the thread can recover from. Relative to the other things mentioned above, these seem like much simpler things to address.

[vadimcn]

@briansmith: The code that unwinds stack using DWARF frame info lives in libgcc. Furthermore, because of the way module registration with the unwinder is implemented, we must link to the dynamic version of libgcc. I am guessing that @alexcrichton did not want to have this dependency in *-pc-windows-msvc executables.
We could re-implement the unwinder ourselves and thus get rid of the libgcc dependency, but given that LLVM will soon have proper 32-bit SEH support, is it a good investment of time?

[briansmith]

We could re-implement the unwinder ourselves and thus get rid of the libgcc dependency, but given that LLVM will soon have proper 32-bit SEH support, is it a good investment of time?

Yes, from my perspective. 32-bit Windows is the most important target in that it is the target that covers the most users. In particular, 32-bit Windows apps work on 64-bit Windows AND 32-bit Windows. From the end-user-of-Rust perspective, it seems very risky to wait for LLVM to eventually fix this, especially when Alex said that LLVM's implementation is likely to not even work for Rust even when it is finished at some unknown point in time in the future. That said, I'm biased because I've invested a lot of time to create infrastructure software in Rust, and in order to convince people to use my software, I have to convince them that Rust isn't super risky, and this is one of the issues that people have mentioned as a concern.

[alexcrichton]

Ah to be clear I just mean that LLVM's implementation right this red hot second is unlikely to work (e.g. the WIP variant), I expect it is certainly aimed at working quite well in the very near future, e.g. the next 3.8 release most likely.

I haven't checked up on the progress in awhile though, but if there's good pressure to get this working robustly on 32-bit MSVC I can certainly shift some time into investigating!