Re: [PATCH 06/10] perf tools: Collect other metrics in struct egroup

From: Ian Rogers
Date: Mon Jun 29 2020 - 15:34:25 EST


On Sun, Jun 28, 2020 at 3:06 PM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
>
> On Fri, Jun 26, 2020 at 02:48:02PM -0700, Ian Rogers wrote:
> > On Fri, Jun 26, 2020 at 12:47 PM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> > >
> > > Collecting other metrics in struct egroup object,
> > > so we can process them later on.
> > >
> > > The change will parse or 'other' metric names out of
> > > expression and 'resolve' them.
> > >
> > > Every used expression needs to have 'metric:' prefix,
> > > like:
> > > cache_miss_cycles = metric:dcache_miss_cpi + metric:icache_miss_cycles
> > >
> > > All 'other' metrics are disolved into one context,
> > > meaning all 'other' metrics events and addded to
> > > the parent context.
> > >
> > > Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> > > ---
> > > .../arch/x86/skylake/skl-metrics.json | 2 +-
> > > tools/perf/util/expr.c | 11 ++
> > > tools/perf/util/expr.h | 1 +
> > > tools/perf/util/metricgroup.c | 158 ++++++++++++++++--
> > > 4 files changed, 157 insertions(+), 15 deletions(-)
> > >
> > > diff --git a/tools/perf/pmu-events/arch/x86/skylake/skl-metrics.json b/tools/perf/pmu-events/arch/x86/skylake/skl-metrics.json
> > > index 8704efeb8d31..71e5a2b471ac 100644
> > > --- a/tools/perf/pmu-events/arch/x86/skylake/skl-metrics.json
> > > +++ b/tools/perf/pmu-events/arch/x86/skylake/skl-metrics.json
> > > @@ -57,7 +57,7 @@
> > > },
> > > {
> > > "BriefDescription": "Instructions Per Cycle (per Logical Processor)",
> > > - "MetricExpr": "INST_RETIRED.ANY / CPU_CLK_UNHALTED.THREAD",
> > > + "MetricExpr": "1/metric:CPI",
> > > "MetricGroup": "TopDownL1",
> > > "MetricName": "IPC"
> > > },
> > > diff --git a/tools/perf/util/expr.c b/tools/perf/util/expr.c
> > > index aa14c7111ecc..cd73dae4588c 100644
> > > --- a/tools/perf/util/expr.c
> > > +++ b/tools/perf/util/expr.c
> > > @@ -150,3 +150,14 @@ int expr__find_other(const char *expr, const char *one,
> > >
> > > return ret;
> > > }
> > > +
> > > +#define METRIC "metric:"
> > > +
> > > +bool expr__is_metric(const char *name, const char **metric)
> > > +{
> > > + int ret = !strncmp(name, METRIC, sizeof(METRIC) - 1);
> > > +
> > > + if (ret && metric)
> > > + *metric = name + sizeof(METRIC) - 1;
> > > + return ret;
> > > +}
> >
> > Should expr.l recognize metric:... as a different kind of token rather
> > than an ID?
>
> hm, we still want it to be returned as ID token, and the processing
> code needs a way to distinguish between event and metric, so I'd think
> we need to keep it, but I'll double check

Thanks, the struct names sound good. I suggested using a token as it
is a little strange that we have layers of parsing and this would be a
chance to avoid one layer. It isn't a big deal, the event parsing is
far more complex :-)

Ian

> thanks,
> jirka
>