Re: GDB 6.2 for ARM: bug in backtrace when in Thumb mode

[Richard Earnshaw <rearnsha@gcc.gnu.org>]

On Tue, 2004-09-07 at 13:02, Thierry wrote:
> Hi there,
>  
> I'm trying to debug an application compiled for the
> ARM in Thumb mode. The 'backtrace' command seems buggy
> as soon as there are more than 4 levels in the stack.
> Besides, with the same code compiled in ARM mode, it
> works perfectly.
>  
> Here's an example:
>  
> void dummy2(int a)
> {
>    int b;
>    b = a;
> }
>  
> void dummy1(void)
> {
>    dummy2(4);
> }
>  
> void dummy(void)
> {
>    dummy1();
> }
>  
> In my 'main', I call 'dummy()'.
> 
> When I go step-by-step until I step into 'dummy2', the
> 'bt' command gives the following result:
> #0  dummy2 (a=4) at MainModule/src/dummy.c:4
> #1  0xc000039c in dummy1 () at
> MainModule/src/dummy.c:9
> #2  0x00000000 in ?? ()
>  
> which is wrong!
>  
> The files have been compiled with arm-elf-gcc with the
> following options:
> -ggdb -gdwarf-2 -c -mapcs -fomit-frame-pointer -mthumb
> -mthumb-interwork -mlong-calls
> -fsigned-char -mstructure-size-boundary=8
> -fshort-enums
> 
> When I remove the option '-fomit-frame-pointer', it
> works.
>  
> Any idea?
>  

Most (all?) Released versions of gcc don't generate unwind information
for Thumb functions, so the unwinder gets confused if there's no frame
pointer.

It will be fixed in gcc-3.5, but I don't think it was fixed in time for
gcc-3.4.

R.