[PATCH 3/4] perf tools: Check libunwind for availability of dwarf parsing feature

[Jiri Olsa]

On Wed, Sep 25, 2013 at 08:53:44AM +0200, Jean Pihet wrote:
> Hi Jiri,
> 
> On 24 September 2013 19:43, Jiri Olsa <jolsa at redhat.com> wrote:
> > On Tue, Sep 24, 2013 at 02:03:47PM +0200, Jean Pihet wrote:
> >> Hi Jiri, Will,
> >>
> >> On 24 September 2013 12:06, Will Deacon <will.deacon at arm.com> wrote:
> >> > On Tue, Sep 24, 2013 at 10:34:50AM +0100, Jiri Olsa wrote:
> >> >> On Tue, Sep 24, 2013 at 10:55:32AM +0200, Jean Pihet wrote:
> >> >> > Ping on the series. The two patches above (3/4 and 4/4) are generic
> >> >> > while the two others are impacting ARM only.
> >> >> > Is it possible to get an Ack for the generic ones?
> >> >>
> >> >> I'm fine with those changes.. still I'm sort of worried about
> >> >> current DWARF unwind users (but not sure if there're any),
> >> >> who depends on packaged libunwind compiled without
> >> >> --enable-debug-frame option.
> >> >
> >> > Since x86 is the only architecture using libunwind with perf at the moment,
> >> > and I'd expect it to use .eh_frame for unwinding, I'm also not sure there
> >> > are any existing users to worry about.
> >> Right
> >>
> >> >
> >> >> I've seen your libunwind patch to make it default, but
> >> >> not sure if it was accepted.. if not, maybe we should
> >> >> detect this and build that code conditionaly.
> >> >
> >> > It certainly defaults to "on" for ARM, but other architectures have to
> >> > enable it explicitly afaict.
> >> Yes that is correct.
> >> This patch (3/4) detects if the debug frame code is enabled in
> >> libunwind and uses the lib only if it is the case.
> >
> > My concern is about users (again, not sure if there are any ;-) )
> > that use this with packaged libunwind compiled without
> > --enable-debug-frame option.
> >
> > For them perf will consider libunwind as 'not available' with
> > your changes:
> >
> >       ...
> >       CHK libunwind
> >   config/Makefile:223: No libunwind found, disabling post unwind support.  Please install libunwind-dev[el] >= 1.1
> >       ...
> >
> > and they'll need to compile their own libunwind
> > (thats the case on Fedora).
> >
> > This could be solved by detecting this and make your
> > code conditional as attached below (not much tested).
> Ok that makes sense.
> Let me integrate this in the patch series, test it (on ARM and x86)
> and re-submit. Is that OK?

that'd be great

thanks,
jirka