Metadata missing from inlined frames in backtraces

[c42f]

Having the parent Module and other method metadata available for each frame in a backtrace is quite useful for pretty printing, but we're currently missing it for inlined frames. For example:

julia> bar() = backtrace()
julia> foo() = bar()

julia> st = stacktrace(foo())
15-element Vector{Base.StackTraces.StackFrame}:
 bar at REPL[10]:1 [inlined]
 foo() at REPL[11]:1
 top-level scope at REPL[13]:1
...

For the foo frame, the linfo field contains the MethodInstance which in turn can be traced back to the Method using st[2].linfo.def to get the Module or other desired metadata. However, for the bar frame we just get nothing here.

julia> dump(st[1])
Base.StackTraces.StackFrame
  func: Symbol bar
  file: Symbol ./REPL[10]
  line: Int64 1
  linfo: Nothing nothing
  from_c: Bool false
  inlined: Bool true
  pointer: UInt64 0x00007f832223b37c

julia> dump(st[2])
Base.StackTraces.StackFrame
  func: Symbol foo
  file: Symbol ./REPL[11]
  line: Int64 1
  linfo: Core.MethodInstance
    def: Method
      name: Symbol foo
      module: Module Main
      file: Symbol REPL[11]
      line: Int32 1
    [.....]
  from_c: Bool false
  inlined: Bool false
  pointer: UInt64 0x00007f832223b37c

A place to start looking is inside the lookup code at

julia/src/debuginfo.cpp

Line 542 in 1a2285b

frame->linfo = NULL; // TODO: if (new_frames[n_frames - 1].linfo) frame->linfo = lookup(func_name in linfo)?

As suggested there, we should have the MethodInstance (or a CodeInfo thunk for toplevel code... IIRC) for the parent non-inlined frame in new_frames[n_frames-1]. So if we made sure that contained sufficient information about inlined methods we could fill in linfo in some way from there.

See also #33065 (comment)

[BioTurboNick]

I poked around a little bit. I noticed there's a linetable field that contains stack frame information; and inlining.jl places information on inlined frames into a vector linetable. I'm not sure if they aren't removed at some point.

Is there a point at which the linetable is finalized and could be inspected on the Julia side?

jl_print_bt_entry_codeloc uses it to obtain information on each frame from its CodeInfo block (unwrapping MethodInstances to get it), and seems to include inlined stack frames. But not sure where this code is used, suggests in an interpreter? The function used when Julia has a segfault is jl_print_native_codeloc, but it just uses the information provided by jl_getFunctionInfo, which is what we're trying to change.

I'd like to determine if this linetable contains inlined frame information, and then figure out how to access it if so. Perhaps linfo for the parent non-inlined frame could just contain the pointers to do that lookup?

Update: Found out I can inspect linetable at the end of run_passes. It does contain module information for the inlined frames. At a minimum, could scan the table for the matching function name, file name, and line number?

[c42f]

At a minimum, could scan the table for the matching function name, file name, and line number?

Sounds great if it's reliable. Getting module information would be useful, but it would be even better if we could trace back to the Method. as this would allow things like #35371 to work reliably for inline frames (not that I'm sure that PR will be merged. It seemed controversial...)

[BioTurboNick]

I have a partial solution.

Now I just need to get it not to segfault when trying to display the function name and signature.

[c42f]

Awesome to see this fixed. Thanks so much @BioTurboNick :-)