[PATCH] gas: Improve documentation for cfi_remember/restore_state

[Alan Modra]

On Fri, Apr 15, 2016 at 11:21:31AM -0300, Martin Galvan wrote:
> On Thu, Apr 14, 2016 at 8:16 PM, Alan Modra <amodra@gmail.com> wrote:
> >
> > On Thu, Apr 14, 2016 at 11:23:08AM -0300, Martin Galvan wrote:
> > > You mean something like:
> > >
> > >     je label
> > >     popq %rbx
> > >     .cfi_remember_state
> > >     .cfi_restore %rbx
> > >     popq %rbp
> > >     .cfi_restore %rbp
> > >     popq %r12
> > >     .cfi_restore %r12
> > >     ret
> > >
> > > label:
> > >     .cfi_restore_state
> > >     /* Do something else */
> > >
> > > In that case we're using .cfi_restore_state to save us having to use
> > > multiple CFI directives to recreate the original save locations.
> >
> > Yes, exactly.  However the above example shows a gcc bug!
> 
> If you're referring to the last example I sent (with the three pops),
> I wrote that manually. So it's a programmer bug, not gcc's :)

Ah, OK.

> > Hmm, seems like current mainline gcc is buggy in this area on x86_64.
> > I see this sort of thing around a tail call:
> >         je      .L4
> >         popq    %rbp
> >         .cfi_remember_state
> >         .cfi_def_cfa 7, 8
> >         movl    $1, %edi
> >         jmp     *%rax
> > .L4:
> >         .cfi_restore_state
> > So the cfa is set back to rsp on popping rbp, but there ought to be a
> > ".cfi_restore 6".  Otherwise when an async interrupt hits after the
> > pop of rbp, the unwinder will load rbp from the stack, which has just
> > been trashed by the interrupt handler..
> 
> That's probably true, though. I can look into it a bit more if you
> want. I know next to nothing about gcc internals, but a couple guys at
> the office can give me a hand with it.

I suspect there's no need to look into it yourself.  x86_64 doesn't
lack interested and capable developers.

> > It might be better to choose an example from gcc -fomit-frame-pointer
> > -fasynchronous-unwind-tables code.
> 
> Could we keep my 3-pop example if I added the required CFA adjustment?
> I'd like to keep the example as simple as possible for the
> documentation.

Yes.

-- 
Alan Modra
Australia Development Lab, IBM