Re: [PATCH] kernel32: make it possible to create a backtrace from a RaiseException exception in gdb

[Alexandre Julliard]

Bernhard Loos  writes:
> @@ -60,6 +60,17 @@ static PTOP_LEVEL_EXCEPTION_FILTER top_filter;
>  typedef INT (WINAPI *MessageBoxA_funcptr)(HWND,LPCSTR,LPCSTR,UINT);
>  typedef INT (WINAPI *MessageBoxW_funcptr)(HWND,LPCWSTR,LPCWSTR,UINT);
>
> +#ifdef __i386__
> +/* without this, it's impossible to get a backtrace from the raised
> -
> - exception in winedbg in gdb mode */
>
> +void RaiseException_helper(EXCEPTION_RECORD *r, void *f) DECLSPEC_HIDDEN;
> +__ASM_GLOBAL_FUNC(RaiseException_helper,
> -
>                "mov 0x4(%esp),%eax\n\t"
> -
>                "push %eax\n\t"
> -
>                "mov 0xc(%esp),%eax\n\t"
> -
>                "call *%eax\n\t"
> -
>                "ret")
>
> +#endif

You shouldn't need anything like that.

[PATCH] kernel32: make it possible to create a backtrace from a RaiseException exception in gdb

[Bernhard Loos]

On Tue, Sep 27, 2011 at 4:34 PM, Alexandre Julliard  wrote:
> Bernhard Loos  writes:
> > @@ -60,6 +60,17 @@ static PTOP_LEVEL_EXCEPTION_FILTER top_filter;
> >  typedef INT (WINAPI *MessageBoxA_funcptr)(HWND,LPCSTR,LPCSTR,UINT);
> >  typedef INT (WINAPI *MessageBoxW_funcptr)(HWND,LPCWSTR,LPCWSTR,UINT);
> >
> > +#ifdef __i386__
> > +/* without this, it's impossible to get a backtrace from the raised
> > -
> > - exception in winedbg in gdb mode */
> >
> > +void RaiseException_helper(EXCEPTION_RECORD *r, void *f) DECLSPEC_HIDDEN;
> > +__ASM_GLOBAL_FUNC(RaiseException_helper,
> > - "mov 0x4(%esp),%eax\n\t"
> > - "push %eax\n\t"
> > - "mov 0xc(%esp),%eax\n\t"
> > - "call *%eax\n\t"
> > - "ret")
> >
> > +#endif
>
> You shouldn't need anything like that.

I don't think it's really possible to avoid this.
The problem goes like that:
gcc adjusts esp once at the function start and then never touches it
again until the function returns. This works quite well, if only
__cdecl functions are called, but as __stdcall functions clean up
their stack, gcc produces something like this:

0x7b839cf1 <+81>:    mov    %eax,(%esp)
   0x7b839cf4 <+84>:    call   0x7b839c90 <RtlRaiseException_helper2>
   0x7b839cf9 <+89>:    sub    $0x4,%esp
   0x7b839cfc <+92>:    add    $0x64,%esp
   0x7b839cff <+95>:    pop    %ebx
   0x7b839d00 <+96>:    pop    %esi
   0x7b839d01 <+97>:    ret    $0x10

RtlRaiseException_helper2 has one argument in this case and gcc
retores the original esp value after the call. The problem is, that
gcc doesn't emit DWARF frame adjustment ops for this whole exercise,
so for the single instruction at 0x7b839cf9, gdb assumes the wrong Esp
value and the result are broken backtraces. Usually, this doesn't hurt
much but in this case the Eip value for the exception context points
there which makes this pretty annoying.
> --
> Alexandre Julliard
> [e-mail]

[PATCH] kernel32: make it possible to create a backtrace from a RaiseException exception in gdb

[Alexandre Julliard]

Bernhard Loos  writes:
> RtlRaiseException_helper2 has one argument in this case and gcc
> retores the original esp value after the call. The problem is, that
> gcc doesn't emit DWARF frame adjustment ops for this whole exercise,
> so for the single instruction at 0x7b839cf9, gdb assumes the wrong Esp
> value and the result are broken backtraces. Usually, this doesn't hurt
> much but in this case the Eip value for the exception context points
> there which makes this pretty annoying.

I suggest filing a bug with gcc then.

[PATCH] kernel32: make it possible to create a backtrace from a RaiseException exception in gdb

[Bernhard Loos]

On Tue, Sep 27, 2011 at 6:06 PM, Alexandre Julliard  wrote:
> Bernhard Loos  writes:
> > RtlRaiseException_helper2 has one argument in this case and gcc
> > retores the original esp value after the call. The problem is, that
> > gcc doesn't emit DWARF frame adjustment ops for this whole exercise,
> > so for the single instruction at 0x7b839cf9, gdb assumes the wrong Esp
> > value and the result are broken backtraces. Usually, this doesn't hurt
> > much but in this case the Eip value for the exception context points
> > there which makes this pretty annoying.
>
> I suggest filing a bug with gcc then.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50544
Let's see how it goes.
> --
> Alexandre Julliard
> [e-mail]