Re: [patch] java run-time stack trace

[Bryce McKinlay <bryce at albatross dot co dot nz>]

Andrew Haley wrote:

>  > That implementation is written in C++ style, so surely it makes
>  > sense to demangle it in C++ style. We don't _want_ to hide
>  > implementation details in a stack trace!
>
> I disagree very strongly.  Having a method change its name because it
> happens to be native is unnecessarily confusing and makes gcj look
> like less of a real java system than javac.

Showing native frames in native style is confusing, but showing the pc isn't? hmm.

> Better demangle Java
> style if possible, *especially* for anything that is defined as a part
> of the Standard Java API.  This is IMO an important "look and feel"
> issue.

OK. Do you have any suggestions as to how this can be done? I can't think of a
reasonable and robust way to determine whether a frame is a CNI frame or not, without
putting some sort of hint into the name mangling. That would require changes to both
gcj and c++, and seems like overkill (unless there is some other compelling reason why
it should be done). Another possibility would be to have the stack tracer look up the
types it sees in the names and determine if they are java types or not, but this
strikes me as slow and too potentially error prone for a stack tracer.

regards

  [ bryce ]