Re: [draft patch] <unavailable> unwinder for btrace [Re: [rfc 3/5] record: make it build again]

[Jan Kratochvil <jan dot kratochvil at redhat dot com>]

On Thu, 28 Mar 2013 07:09:23 +0100, Markus Metzger wrote:
> On Mar 27, 2013, at 16:22 , "Metzger, Markus T" <markus.t.metzger@intel.com> wrote:
> > I've been experimenting with this a bit.  It looks like there will always be
> > a sentinel frame at the very bottom that is reading the registers directly
> > from the inferior.  I can only hook in at the second frame.
> > 
> > In order to fake the back trace for btrace replay, I would also need to
> > replace the sentinel frame, since otherwise, the first frame will always
> > point to the current location.
> > 
> > Am I doing something wrong?
> 
> Yes, I am.  I need to also provide the target registers.  Then the
> sentinel frame should do exactly what I want.

It looks right to me.


> > On a related but different topic, I added a btrace frame type and
> > prologue cache.  The cache holds a pointer to some btrace data
> > structure that is used to compute the fake back trace.  In order to
> > unwind a btrace frame, I would need to access the next frame's
> > location in this btrace data structure.
> > 
> > The easiest would be to check for the next frame's type and then
> > access it's cache - which doesn't work since struct frame_info is
> > opaque.  I ended up encoding the pointer into the special_addr
> > of a btrace frame's frame_id - which is somewhat ugly.  Any better
> > idea?

special_addr is really not right.

dwarf2-frame-tailcall.c uses for a similar problem 'htab_t cache_htab' which
is indexed by 'struct frame_info *' which you can iterate in any direction so
you even do not need a new cache entry for every 'struct frame_info *'.


> > Also what's the lifetime of a frame_info and frame_id object?
> > When the branch trace is cleared, any pointers to it will become
> > stale.

There is frame_unwind->dealloc_cache, any reinit_frame_cache() call inside GDB
will clear the prologue cache which is very common.

I see now btrace_thread_info->btrace may change more often - such as during
each "info record" command.  So call reinit_frame_cache() in the cases btrace
cache may get rebuilt.


Thanks,
Jan