On Wed, 13 Dec 2006 23:11:11 +0100, Jan Kratochvil wrote:
> On Wed, 13 Dec 2006 22:52:52 +0100, Mark Kettenis wrote:
> ...
> > No, sorry, it is me who is confused.  I didn't actually look at Jan's patch
> > and assumed it did the the right thing of marking the return address as
> > undefined.
> 
> Therefore what should the patch do? Currently `.cfi_undefined' looks as too
> radical.

Read all the past threads and I hope this patch complies with all the parties.

I would like if the `.cfi_undefined' patch
	http://sourceware.org/ml/libc-alpha/2006-11/msg00082.html
could stay in the development glibc with the unwinder patched by Jakub Jelinek
	http://sources.redhat.com/ml/gdb/2006-12/msg00100.html
as with such CFI patch gdb does not show any bogus 0x0 unwind of `clone'.

On the other hand as a legacy compatibility I wrote the attached patch to fix
the GDB excessive non-CFI amd64 `clone' unwind as 0x0.  Other platforms (and
functions?) could be carbon-copied from this one.

I consider it the same approach as the way GDB currently on amd64 detects
signal frames by checking the instructions - the CFI for signal frames has been
also recently checked to the development glibc as was the `clone' CFI.

Tested for regressions by the gdb testsuite, it is the most conservative way so
it should not have any sideeffects.  The 0x0 unwinding of any non-`clone'
functions remains the same so Mark should have no objections, I hope.



Regards,
Jan