> Hmm, if full CFI for the explicit register saving code is emitted, I'd
> expect backtraces to work.

Exceptions work well so the CFIs are very likely correct, at least in eh_frame.
A quick look didn't reveal differences between eh_frame and debug_frame so this 
is a little strange then.  Is GDB supposed to fall back to "manual" unwinding 
only when the DWARF unwinder gives up?

> I'd like to get a chance to run a regression test on OpenBSD/sparc64.
> Can you give me a couple of days and perhaps ping me towards the end
> of this week to remind me if I haven't responded by then?

Sure.

> For consistency that probably should be (cache->copied_regs_mask & 0x04).

Fixed.

> Elsewhere in GDB, or at least in the code written by me, like most of
> the existing sparc/sparc64 code places the '=' sign on the line
> before.  I can see how you can interpret the GNU coding style to say
> that you should put the '=' sign on the next line, but for now it'd be
> better to be consistent with the rest of the code.  Can you change
> this?

Sure, done.  This was the standard GCC style.

> This change isn't quite right.  The wcookie stuff is there to support
> StackGhost[1], which is unsupportable for the "flat" model.  So what
> needs to be checked here is whether we executed a save instruction or
> not.  As far as I know OpenBSD is the only OS that does StackGhost.
> On everything else wcookie will be 0.  So you can probably leave this
> as is right now, since the existing frameless_p check is more correct
> than what you have in your diff.

OK, reverted.

> StackGhost again.  Better leave this alone as well.

Yes, there is a bit of code duplication in these files. ;-)

> [1] http://projects.cerias.purdue.edu/stackghost/stackghost/index.html

Thanks for the pointer.


Revised patch attached, will retest if approved.

-- 
Eric Botcazou