Re: [PATCH v1 1/4] perf annotate: create a new hists to manage multiple events samples

From: Arnaldo Carvalho de Melo <acme@kernel.org>
To: "Jin, Yao" <yao.jin@linux.intel.com>
Cc: jolsa@kernel.org, peterz@infradead.org, mingo@redhat.com,
	alexander.shishkin@linux.intel.com, Linux-kernel@vger.kernel.org,
	ak@linux.intel.com, kan.liang@intel.com, yao.jin@intel.com
Subject: Re: [PATCH v1 1/4] perf annotate: create a new hists to manage multiple events samples
Date: Thu, 14 Sep 2017 11:05:31 -0300	[thread overview]
Message-ID: <20170914140531.GA10371@kernel.org> (raw)
In-Reply-To: <561eb4d7-ce80-a62d-ade3-e9b95028748f@linux.intel.com>

Em Thu, Sep 14, 2017 at 09:03:55AM +0800, Jin, Yao escreveu:
> On 9/11/2017 9:33 AM, Jin, Yao wrote:
> > On 9/8/2017 9:43 PM, Arnaldo Carvalho de Melo wrote:
> >> Em Wed, Aug 16, 2017 at 06:18:33PM +0800, Jin Yao escreveu:
> >>> An issue is found during using perf annotate.
> >>>
> >>> perf record -e cycles,branches ...
> >>> perf annotate main --stdio
> >>>
> >>> The result only shows cycles. It should show both cycles and
> >>> branches on the left side. It works with "--group", but need
> >>> this to work even without groups.
> >>>
> >>> In current design, the hists is per event. So we need a new
> >>> hists to manage the samples for multiple events and use a new
> >>> hist_event data structure to save the map/symbol information
> >>> for per event.
> >> Humm, why do we need another hists? Don't we have one per evsel, don't
> >> we have a evlist from where to get all of those evsels, can't we just
> >> use that to add one column per evsel?

> > I'm considering a case.

> > Suppose we sample 2 events ("branches" and "cache-misses"). The samples of "branches" are hit in function A and the samples of "cache-misses" are hit in function B.
> > 
> > The branches evsel has one hists and cache-misses evsel has another hists.
> > 
> > The hists of branches evsel has one hist-entry which stands for the function A symbol. The hists of cache-misses evsel has one hist-entry which stands for the function B symbol.
> > 
> > If we start to show the instructions in function B from cache-misses evsel, we will lose the function A.
> > 
> > Because even if we get the branches evsel from the link in cache-misses evsel, but the function A is before function B and function B has been displayed yet, so the function A is lost.
> > 
> > Considering the number of events can be greater than 2, the code will be much more complicated. So using a global hists should be an easy solution.
> 
> Could the solution of using a new hists for multiple events be accepted?
> 
> Or anything I should update in the patches?

I'm not having time at this moment for doing a proper review, wait a bit
more please.

- Arnaldo