Add support for unwinding

[antoyo]

• Requires adding support for try/catch in libgccjit.
• Requires changing rustc_codegen_ssa API.

Some notes to myself to debug the issue in release mode:

1. Attempt to reproduce by creating an example without std.
2. Reduce that to a pure libgccjit example (e.g. without using cg_gcc).
3. Try to reproduce in C++ directly (to see if it's a bug there as well or if the warn about something or if they have a workaround for gotos if that's the issue).
4. Perhaps the fact that gotos and conditions use C labels instead of the special jump labels is the issue (see how Add support for NonNull function attribute #326 is resolved).
5. If gotos are are forbidden, try to create new #[inline(always)] functions.

[sapir]

I'd like to try to help with this.

[antoyo]

This issue is very hard so I suggest you start with one of the good first issues.

If you still wanna try, I'll upload my branch of libgccjit that starts adding support for try/catch and I suggest to start looking at the rustc_codegen_ssa API functions related to that (invoke, landing_pad, …) and think of a way to refactor it to make it easier to support a different API based on try/catch (I don't believe it's possible to implement unwinding without changing this API, even with hacks).

[bjorn3]

Can invoke be codegened as a try + catch with the catch goto'ing to the landingpad?

[antoyo]

Can invoke be codegened as a try + catch with the catch goto'ing to the landingpad?

That's an interesting idea, actually. I didn't think of doing that. Maybe that would work.

[sapir]

Actually I was thinking the entire block should have a try/catch around it, so that the result of the invoke can be saved. I don't understand yet what the result of the llvm landingpad instruction actually is, though, except that it carries some info about the thrown exception?

[bjorn3]

When the personality function runs it searches for the location of the cleanup code in a compiler generated table (.gcc_except_table I believe for rust). When it find it, it tells the unwinder to use it, but it also tells the unwinder to pass some arguments. In case of rust these are two arguments, which if I recall correctly are pointers to the unwind context and the actual exception object. The landingpad instruction returns all arguments passed by the personality function I believe.

When you jump out of the catch to get to the landingpad does gcc insert a function call to indicate that unwinding is finished? If so, my suggestion won't work. If it doesn't do you know how to run the unwind continue/finish functions? (at machine evel a tail call to _Eh_Resume for continuing unwind I believe, but maybe gcc expects a special instruction?)

[sapir]

When you jump out of the catch to get to the landingpad does gcc insert a function call to indicate that unwinding is finished?

I haven't actually tried it at the GENERIC level, but I tried supposedly equivalent C++ code and I think it does :( It calls __cxa_end_catch:

https://compiler-explorer.com/z/Kadvfz81W

So I guess maybe the proper behavior is to have all the blocks that are only reachable from the catch side of invoke statements, copied into the catch expr?

btw, will the __cxa_* functions work with the panic implementation?

[bjorn3]

Rust doesn't use the C++ standard library, so it uses _Unwind_* instead of __cxa_*. Resuming with unwinding requires _Unwind_Resume and ending unwinding doesn't seem to generate any implicit function call at all. Only an explicit call to the cleanup function passed to the try intrinsic.

[sapir]

Are unwind tables necessary? Can exceptions panics be implemented just by calling the appropriate functions?

[bjorn3]

Unwind tables are necessary for DWARF based unwinding. There is also the older SjLj based unwinding which doesn't need unwind tables, but still needs support from the compiler backend AFAIK. Other than that the only option is to change ABI to return a pointer optionally representing the panic and checking this pointer after every call.

[bjorn3]

Looks like at least for GIMPLE the __cxa_* calls are explicit. I am not sure if it allows jumping out of a cleanup block though, but your suggestion of

So I guess maybe the proper behavior is to have all the blocks that are only reachable from the catch side of invoke statements, copied into the catch expr?

should work I think @antoyo.

[sapir]

I think I'll try it with the goto then and see if it works

[antoyo]

I haven't touched that gcc branch in a while, so it might not work correctly (and there's a bit more stuff to implement) but here it is.

It adds support to set the personality function and try/catch.

@sapir I hope that helps!