Re: [RFC/PATCHSET 00/15] perf report: Add support to accumulate hist periods

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, Sep 28, 2012 at 09:07:57AM +0200, Stephane Eranian wrote:
> On Fri, Sep 28, 2012 at 7:49 AM, Namhyung Kim <namhyung@kernel.org> wrote:
> > Hi Frederic,
> >
> > On Fri, 28 Sep 2012 01:01:48 +0200, Frederic Weisbecker wrote:
> >> When Arun was working on this, I asked him to explore if it could make sense to reuse
> >> the "-b, --branch-stack"  perf report option. Because after all, this feature is doing
> >> about the same than "-b" except it's using callchains instead of full branch tracing.
> >> But callchains are branches. Just a limited subset of all branches taken on excecution.
> >> So you can probably reuse some interface and even ground code there.
> >>
> >> What do you think?
> >
> > Umm.. first of all, I'm not familiar with the branch stack thing.  It's
> > intel-specific, right?
> >
> The kernel API is NOT specific to Intel. It is abstracted to be portable
> across architecture. The implementation only exists on certain Intel
> X86 processors.
> 
> > Also I don't understand what exactly you want here.  What kind of
> > interface did you say?  Can you elaborate it bit more?
> >
> Not clear to me either.
> 
> > And AFAIK branch stack can collect much more branch information than
> > just callstacks.  Can we differentiate which is which easily?  Is there
> > any limitation on using it?  What if callstacks are not sync'ed with
> > branch stacks - is it possible though?
> >
> First of all branch stack is not a branch tracing mechanism. This is a
> branch sampling mechanism. Not all branches are captured. Only the
> last N consecutive branches leading to a PMU interrupt are captured
> in each sample.
> 
> Yes, the branch stack mechanism as it exists on Intel processors
> can capture more then call branches. It is HW based and provides
> a branch type filter. Filtering capability is exposed at the API level
> in a generic fashion. The hw filter is based on opcodes. Call branches
> all cover call, syscall instructions. As such, the branch stack mechanism
> cannot be used to capture callstacks to shared libraries, simply because
> there a a non call instruction in the trampoline. To obtain a better quality
> callstack you have instead to sample return branches. So yes, callstacks
> are not sync'ed with branch stack even if limited to call branches.
> 

You're right. One doesn't simply sample callchains on top of branch tracing. Not easily at least.
But that's not what we want here. We want the other way round: use callchains as branch sampling.
And a callchain _is_ a branch sampling. Just a specialized one.

PERF_SAMPLE_BRANCH_STACK either records only calls, only ret, or everything, or....
You can define the filter with "-j" option. Now callchains can be considered as the result
of a specific "-j" filter option. It's just a high level filtering. ie: not just based on opcode
types but on semantic post-processing. As if we applied a specific filter on a pure branch tracing
that cancelled calls that had matching ret.

But in the end, what we have is just branches. Some branch layout that is biased, that already passed
through a semantic wheel, still it's just _branches_.

Note I'm not arguing about adding a "-j callchain" option, just trying to show you that callchains
are not really different from other filtered source of branch sampling.


> > But I think it'd be good if the branch stack can be changed to call
> > stack in general.  Did you mean this?
> >
> That's not going to happen. The mechanism is much more generic than
> that.
> 
> Quite frankly, I don't understand Frederic's motivation here. The mechanism
> are not quite the same.

So, considering that callchains are just "branches", why can't we use them as
a branch source, just like PERF_SAMPLE_BRANCH_STACK data samples, that we
can reuse in "perf report -b".

Look at commit b50311dc2ac1c04ad19163c2359910b25e16caf6
"perf report: Add support for taken branch sampling". It's doing (except for a few details
like the period weight of branch samples) the same than in Namhyung patch, just with
PERF_SAMPLE_BRANCH_STACK instead of callchains.

I don't understand what justifies this duplication.