[Lldb-commits] [lldb] [lldb] Extend frame recognizers to hide frames from backtraces (PR #104523)

[via lldb-commits]

jimingham wrote:

> > When we were talking about this originally I thought that StackFrames would get an "ImplementationFrame" property, and we would use that to determine whether to hide the frame in bt (when not passing --unfiltered). Then the frame recognizers when they are run on the frame in the ordinary course of things (we do this to determine recognized args anyway) they would set the property on the StackFrame. Then StackFrameList could just check that property. RecognizeFrame might be expensive, and we don't currently have any machinery in place to make sure it caches the results per stop. Most of the RecognizeFrame's just do the work every time. I also think for structural reasons we shouldn't force FrameRecognizers to be the only entity that can mark frames as "should be hidden."
> 
> I understand that not recomputing them might be desirable, but this is not how recognizers currently work. The only place frame recognizers currently get called is by `Thread::GetStopDescription()` (for frame #0) and by `Thread::GetCurrentException()` (also frame #0). So it's not true that we are computing them anyway and thus have an opportunity to precompute the isHidden attribute. I also think that from an architectural perspective it would be wrong for a frame recognizer to be allowed to modify the state of a StackFrame.

This is entirely tangential to this patch, which looks fine to me, but one of the jobs of a Frame recognizer is to add arguments to a StackFrame that doesn't have them.  That's way more intrusive than just setting a "ShouldHide" type flag...

> 
> When a frame recognizer is registered it takes a module regex a function regex and a first_instruction flag. All three of these must match before the recognizer is executed. That's num_recognizers*num_frames regexes. That's not expensive in the grans scheme of things, but you're right that it's not necessary.
> 
> I'll see if I can hide the recognizer invocation inside StackFrame::shouldHide() and let StackFrame cache the result instead of having the Recognizer modify the StackFrame. Maybe I can come up with some cache invalidation strategy that makes sense.



https://github.com/llvm/llvm-project/pull/104523