Re: [patch/rfc] Add a sentinel frame

[Daniel Jacobowitz <drow@mvista.com>]

On Wed, Feb 19, 2003 at 03:27:36PM -0500, Andrew Cagney wrote:
> >You've lost the call to inside_entry_func.  Why?  You've changed the
> >inside_entry_file check to check the PC for the next frame instead of
> >the forthcoming frame, which is not at all the same thing.  Why?
> 
> Dig up old notes.
> 
> This is the test you added.  It stops the unwind past main:
> 
>   if (next_frame->level >= 0
>       && !backtrace_below_main
>       && inside_main_func (get_frame_pc (next_frame)))
>     /* Don't unwind past main(), bug always unwind the sentinel frame.
>        Note, this is done _before_ the frame has been marked as
>        previously unwound.  That way if the user later decides to
>        allow unwinds past main(), that just happens.  */
>     return NULL;
> 
> It occures first (as it should).  It occures before any 
> frame_id_unwind() as needed by frame_chain_valid.  It also occures 
> before the test:
> 
>   /* Only try to do the unwind once.  */
>   if (next_frame->prev_p)
>     return next_frame->prev;
>   next_frame->prev_p = 1;
> 
> so that frame flush code was eliminated (ya!).

OK.

> On the other hand, if GDB is to unwind past main (presumably, if 
> s/backtrace_below_main/unwind_past_main/ is false) it does the test:
> 
>   if (inside_entry_file (get_frame_pc (next_frame)))
> 
> (note the comments about how, if this becomes optional, it should also 
> be moved to before `next_frame->prev_p = 1').
> 
> Anyway, now that missing test.  frame_chain_valid() also contained:
> 
>   /* If we're inside the entry file, it isn't valid.  */
>   if (inside_entry_file (frame_pc_unwind (fi)))
>       return 0;
> 
> Note the frame_pc_unwind().  This test is looking one level along the 
> stack frame to determine if it should unwind to that level.  That is, 
> when FI->prev->pc is in the entry_file, don't unwind to FI->prev.  The 
> problem is, FI->prev->pc is in entry_file when FI->pc is in main.
> 
> Even when unwind-past-main is disabled, GDB refuses to unwind past main! 
>  Consequently, on the branch, I dropped the test.
> 
> (It also unwinds the PC when we're probably not ready).

Oh, but you're misunderstanding.  There's more than one frame in there. 
The call stack in glibc looks like:
  _start
  calls __libc_start_main
  calls main

_start is written in assembly; it generally doesn't have a frame worth
talking about.  Even if we want to show __libc_start_main, we can't
safely backtrace into _start.  That's what the inside_entry_file
(frame_pc_unwind (fi)) is for.

Now, if we want to do this anyway, that's different.  But it's a
change, not a redundancy.


The missing test I mentioned above inside_entry_func, not
inside_entry_file.  Where'd that go?

> As things progress, and more targets switch to the new code, the tests 
> in get_prev_frame will most likely evolve.  However, I don't know that 
> we want to be adding tests without hard evidence that they are needed :-/
> 
> Having said that, sanity checks that the frame didn't go backwards:
> 	!frame_id_inner (frame_id, get_frame_id (next_frame))?

Yes.

> and that they changed:
> 	!frame_id_eq (frame_id, get_frame_id (next_frame))?

Can we do that?  Hmm, we probably can.  A frame ID has a PC in it and a
stack pointer, and if neither has changed we're probably stuck in a
rut.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer