Re: backtrace/1767: Backtrace in naked functions

[Daniel Jacobowitz <drow at false dot org>]

The following reply was made to PR backtrace/1767; it has been noted by GNATS.

From: Daniel Jacobowitz <drow@false.org>
To: gdb-gnats@sourceware.org
Cc:  
Subject: Re: backtrace/1767: Backtrace in naked functions
Date: Tue, 27 Sep 2005 10:29:56 -0400

 On Sat, Sep 24, 2005 at 10:55:46AM +0200, Marius Groeger wrote:
 > Thanks for responding. I'm using GDB 6.3.50.20050810 which I think is
 > fairly recent. Acutally I only noticed a /comparable/ symptom on i386
 > and ppc hosts, and was hoping the solutition to PR#1767 could help me
 > too. We've also discussed this on the crossgcc list. On embedded
 > systems, you often run into having to debug code which does not hava
 > main(), like an exception handler or a self-made threading
 > environment.
 
 No, sorry, this is a problem that has nothing to do with the naked
 functions bug.  Without main in the backtrace GDB should work just
 fine; it will go a few functions past the end, wander into garbage,
 catch itself, and stop.
 
 > I received the following promising hint regarding this, maybe you can
 > add some thoughts: by using the N_MAIN stab flag, you can make any
 > function appear to GDB as if was main(). Unfortunatly, while the GDB
 > source (as far as I understood it :-) somehow seems to indicate this
 > might work, it doesn't. I also tried -g options of CC up to -ggdb but
 > had no look. OBJDUMP correctly lists my toplevel wrapper function as
 > MAIN stab.
 > 
 > Reading the frame.c:get_prev_frame(), there are a lot of tests which
 > could lead to terminating stack frame. It is my impression, though,
 > that most systems use the main() thing. How else can I reliably
 > terminate a stack backtrace, ideally platform neutrally?
 
 Rather than messing around with the name of main, you need to do a
 generally architecture-specific trick to indicate the end of the call
 stack, e.g. setting the frame pointer to zero.  Current versions of GDB
 allow you to express this using DWARF-2 unwind information, by defining
 the return address column as undefined (see the gdb@ list archives for
 more details, a couple of months ago).
 
 -- 
 Daniel Jacobowitz
 CodeSourcery, LLC