Re: [PATCH 4/8] gdb/s390: Fill gen_return_address hook.

[Pedro Alves <palves@redhat.com>]

On 03/11/2016 05:19 PM, Marcin Kościelnicki wrote:
> 
> 11 mar 2016 6:02 PM Pedro Alves <palves@redhat.com> napisał(a):
>  >
>  > On 03/11/2016 04:45 PM, Andreas Arnez wrote:
>  > > On Fri, Mar 11 2016, Pedro Alves wrote:
>  > >
>  > >> On 03/11/2016 03:31 PM, Andreas Arnez wrote:
>  > >>> So I'm OK with the patch.  Please add a small comment stating 
> that this
>  > >>> is a best-can-do approach that usually works near function entry 
> and may
>  > >>> yield wrong results otherwise.
>  > >>
>  > >> I think that should be put in the manual, even.  Users will also 
> trip on
>  > >> this, not just our tests.
>  > >
>  > > Right, I thought about this as well.  How about this?
>  > >
>  > > -- >8 --
>  > > Subject: [PATCH] Document possible unreliability of `$_ret'
>  > >
>  > > diff --git a/gdb/doc/gdb.texinfo b/gdb/doc/gdb.texinfo
>  > > index 4ec0ec1..a14fe19 100644
>  > > --- a/gdb/doc/gdb.texinfo
>  > > +++ b/gdb/doc/gdb.texinfo
>  > > @@ -12863,7 +12863,9 @@ Collect all local variables.
>  > >
>  > >   @item $_ret
>  > >   Collect the return address.  This is helpful if you want to see more
>  > > -of a backtrace.
>  > > +of a backtrace.  Note that the return address can not always be
>  > > +determined reliably, and a wrong address may be collected instead.
>  > > +The reliability is usually higher for tracepoints at function entry.
>  >
>  > Hmm, this reads a bit as if the backtrace will be incorrect/bogus
>  > later on, which is not true.
>  >
>  > How about a merge of your suggestion with Marcin's previous reply,
>  > and some extras on top:
>  >
>  > @item $_ret
>  > Collect the set of memory addresses and/or registers necessary to 
> compute
>  > the frame's return address.  This is helpful if you want to see
>  > more of a backtrace.
>  >
>  > @emph{Note:} The necessary set can not always be reliability 
> determined up
>  > front, and the wrong address / registers may end up collected instead.
>  > The reliability is usually higher for tracepoints at function entry.
>  > When this happens, backtracing will stop because the return address
>  > is found unavailable (unless another collect rule happened to match it).
> 
> Note that this is arch-dependent: powerpc/s390/aarch64 will work at 
> entry (since they dump LR), while x86 has the opposite problem: it uses 
> the frame pointer and will never work at entry (or with 
> -fomit-frame-pointer for that matter).

OK.  I wouldn't want to go too deep into details here, I think just
pointing in the direction of experimenting with different tracepoint
addresses, if important, may be sufficient.

@item $_ret
Collect the set of memory addresses and/or registers necessary to compute
the frame's return address.  This is helpful if you want to see 
more of a backtrace.

@emph{Note:} The necessary set can not always be reliability determined up
front, and the wrong address / registers may end up collected instead.
On some architectures the reliability is higher for tracepoints
at function entry, while on others it's the opposite.
When this happens, backtracing will stop because the return address
is found unavailable (unless another collect rule happened to match it).

WDYT?

Thanks,
Pedro Alves