Re: patches to frysk-imports/lib/dw

[Andrew Cagney <cagney at redhat dot com>]

Adam Jocksch wrote:
> I've wrapped the necessary functionality needed to find inlined 
> function calls in the libdw wrapper (see attached patch). I'm not 100% 
> certain where this code should be called from though. Ideally while 
> libunwind is unwinding (i.e. in one of the callbacks so we open as few 
> Dwfl objects as possible), but for a first implementation maybe it 
> would be worth doing it after libunwind has created the stack trace.
Yes.  Even long term.

They are separate operations.  libunwind provides a list of ABI frame 
instances; given an ABI frame instance, libdw can return a list of 
inlined functions.

Further, in general, the code needs to head in the direction of being 
on-demand.  While the initial cut of libunwind, for instance, can 
generate a full backtrace up front, it will longer term need to act more 
on demand - only performing an unwind an actually required.  For instance:

- when single stepping, the code will need to do a single frame unwind 
to determine if/where a step-into or step-out-of function occured - no 
other frames are needed and typically the test is immediatly followed by 
an unblock (i.e., continue)

- as a performance tweak, when backtracing, the current backtrace can be 
spliced onto a previous cached backtrace - avoiding unwind overhead.

Andrew