Re: [PATCH 5/8] perf, tool: Add pmu event parse support for branch_sample_type values

[Jiri Olsa <jolsa@redhat.com>]

On Wed, Apr 04, 2012 at 09:50:12PM -0700, Stephane Eranian wrote:
> On Wed, Apr 4, 2012 at 1:21 PM, Jiri Olsa <jolsa@redhat.com> wrote:
> > Adding a support to specify branch_type as a hardcoded term
> > inside the pmu event definition.
> >
> > It is possible to specify pmu event like:
> >  "cpu/config=1,branch_type=hv|any_ret|1,config2=2/u"
> >
> > Following string values could be used as value for branch_type:
> >  u (PERF_SAMPLE_BRANCH_USER)
> >  k (PERF_SAMPLE_BRANCH_KERNEL)
> >  hv (PERF_SAMPLE_BRANCH_HV)
> >  any (PERF_SAMPLE_BRANCH_ANY)
> >  any_call (PERF_SAMPLE_BRANCH_ANY_CALL)
> >  any_ret (PERF_SAMPLE_BRANCH_ANY_RETURN)
> >  ind_call (PERF_SAMPLE_BRANCH_IND_CALL)
> >
> > Also a number could be specified as value.
> >
> Although it would be nice to have, the current kernel sampling
> buffer layout + perf would not be able to parse the RECORD_SAMPLE
> if you were to sample different things for different events:
> 
>  perf record -e
> cpu/event=0xc0,umask=1:branch_type=any/,cpu/event=0x3c,umask=0x1/
hi,
for some reason I thought there might be sense to have multiple branch
types enabled in a single event.. I misread the code probably

> 
> Perf report/annotate would not be able to distinguish a RECORD_SAMPLE
> generated by the first or the second event. That's because the RECORD_SAMPLE
> fixed header does not contain enough info to determine which event caused the
> record to be generated. You need to event ID to decode the sample. The event ID
> gives you the attr struct which gives you the attr->sample_type which drives the
> layout of the RECORD_SAMPLE variable size body. The event ID is currently
> saved "somewhere" in the variable-size body of the sample. You have  a chicken
> and egg problem here.

yep, I'm aware of this.. cool you're preparing fix

thanks,
jirka