Re: Using sysprof with 64bits linux

[Christian Hergert <christian hergert me>]

On 05/09/2017 08:15 AM, Jean-Marc Lasgouttes wrote:
> I have done that (keeping the -O2 optimization level). The perf report
> -g output is reasonable:
>  http://www-rocq.inria.fr/~lasgoutt/perf-record.png
>
> What shall I do from there (other than switching to perf, of course :)?

I've created https://bugzilla.gnome.org/show_bug.cgi?id=782407 to track
the issue.

What would help for me to cut the problem domain down is to know:

  - Was your perf recording "whole system" or "just the target
    application". I've noticed that I sometimes get different data back
    from perf in these situations and I'd like to verify if that is
    where I should spend my time.

  - If you could upload a xz/bz2 of the capture file somewhere, that
    would help me determine if the samples are in the capture and we
    are failing to build a proper callgraph, or if they are simply
    missing from the capture.

    If they are missing from the capture, we either need to try harder
    to setup the perf event stream or are doing something wrong when
    setting the PID/cpuset.

    If they are in the capture stream, then we need to improve our
    callgraph profile builder (or symbol resolvers). Given that there
    are 10,000+ samples in that capture, I'm inclined to believe this
    is our problem right now.

-- Christian