winch corrupts rlang backtraces about winch HOT 9 CLOSED

Maybe winch should provide a function that returns a list of backtraces and locations of insertions and let rlang include it.

from winch.

I think the error_frame column in the rlang traceback data frame isn't being added in the winch traceback. Is that a new part of the rlang traceback?

If I follow the path of the original error through then I get to winch:::insert_native_chunk() where it actually does the modification of the rlang traceback object (which is itself a data frame). Everything looks correct except for the fact that the rlang traceback object has an extra error_frame column.

i.e. here is the original trace data frame and the new trace data frame that winch builds

Browse[5]> as.data.frame(trace)
                                                                                                call parent visible namespace scope error_frame
1                                 vctrs::num_as_location(-5, 4, oob = "extend", negative = "ignore")      0    TRUE     vctrs    ::        TRUE
2 stop_internal_c_lib(file = file, line = line, call = call, message = message, ,     frame = frame)      0   FALSE     rlang   :::       FALSE
3                                      abort(message, call = call, .internal = TRUE, .frame = frame)      2   FALSE     rlang    ::       FALSE
Browse[5]> new
                                                                                                call parent visible namespace scope
1                                 vctrs::num_as_location(-5, 4, oob = "extend", negative = "ignore")      0    TRUE     vctrs    ::
2                                                                                  ffi_as_location()      1    TRUE /vctrs.so    ::
3                                                                             vec_as_location_opts()      2    TRUE /vctrs.so    ::
4                                                                                  int_as_location()      3    TRUE /vctrs.so    ::
5                                                                             rlang_stop_internal2()      4    TRUE /rlang.so    ::
6                                                                                 stop_internal_cb()      5    TRUE /rlang.so    ::
7                                                                                    r_exec_mask_n()      6    TRUE /rlang.so    ::
8 stop_internal_c_lib(file = file, line = line, call = call, message = message, ,     frame = frame)      7   FALSE     rlang   :::
9                                      abort(message, call = call, .internal = TRUE, .frame = frame)      8   FALSE     rlang    ::

If I just hack the result of insert_native_chunk() with new$error_frame <- FALSE then I get a traceback again

from winch.

I think it'd be better if winch did not modify rlang backtraces.

from winch.

I agree with your suggestion about winch returning backtraces and locations and then letting rlang handle updating the traceback. I was just trying to narrow down exactly what the problem was 😛. If rlang handles the backtrace insertion then it is free to tweak the internal structure of the backtrace object without worrying about winch issues like this one.

from winch.

How do we handle this in rlang for now? I've been a bit out of touch with winch, but I need to do a release today and can work on improvements while I'm at it.

from winch.

How do we handle this in rlang for now?

Sorry I'm not sure what you mean.

from winch.

Missed the PR link, I now see it's been dealt with already. I'll fix here and PR in rlang then.

from winch.

Thanks. IIUC, you'd rather reimplement this functionality in rlang? I have tweaked winch_trace_back(), it now contains all the data necessary to do the interweaving of the stack traces. Would you like to move winch_add_trace_back() to rlang and adapt, do you need help with that?

from winch.

yup we could do that. On the other hand I'd like to improve extensibility / backtrace interop. But we could start from rlang for now and move back to winch later on. If you could draft the hook, that'd be great, thanks!

from winch.