Re: [PATCH AutoFDO]Restoring indirect call value profile transformation

[Andi Kleen <ak at linux dot intel dot com>]

On Wed, Dec 19, 2018 at 12:08:35PM +0800, Bin.Cheng wrote:
> On Wed, Dec 19, 2018 at 12:00 PM Andi Kleen <ak@linux.intel.com> wrote:
> >
> > On Wed, Dec 19, 2018 at 10:01:15AM +0800, Bin.Cheng wrote:
> > > On Tue, Dec 18, 2018 at 7:15 PM Bin.Cheng <amker.cheng@gmail.com> wrote:
> > > >
> > > > On Sun, Dec 16, 2018 at 9:11 AM Andi Kleen <ak@linux.intel.com> wrote:
> > > > >
> > > > > "bin.cheng" <bin.cheng@linux.alibaba.com> writes:
> > > > >
> > > > > > Hi,
> > > > > >
> > > > > > Due to ICE and mal-functional bugs, indirect call value profile transformation
> > > > > > is disabled on GCC-7/8/trunk.  This patch restores the transformation.  The
> > > > > > main issue is AutoFDO should store cgraph_node's profile_id of callee func in
> > > > > > the first histogram value's counter, rather than pointer to callee's name string
> > > > > > as it is now.
> > > > > > With the patch, some "Indirect call -> direct call" tests pass with autofdo, while
> > > > > > others are unstable.  I think the instability is caused by poor perf data collected
> > > > > > during regrets run, and can confirm these tests pass if good perf data could be
> > > > > > collected in manual experiments.
> > > > >
> > > > > Would be good to make the tests stable, otherwise we'll just have
> > > > > regressions in the future again.
> > > > >
> > > > > The problem is that the tests don't run long enough and don't get enough samples?
> > > > Yes, take g++.dg/tree-prof/morefunc.C as an example:
> > > > -  int i;
> > > > -  for (i = 0; i < 1000; i++)
> > > > +  int i, j;
> > > > +  for (i = 0; i < 1000000; i++)
> > > > +    for (j = 0; j < 50; j++)
> > > >       g += tc->foo();
> > > >     if (g<100) g++;
> > > >  }
> > > > @@ -27,8 +28,9 @@ void test1 (A *tc)
> > > >  static __attribute__((always_inline))
> > > >  void test2 (B *tc)
> > > >  {
> > > > -  int i;
> > > > +  int i, j;
> > > >    for (i = 0; i < 1000000; i++)
> > > > +    for (j = 0; j < 50; j++)
> > > >
> > > > I have to increase loop count like this to get stable pass on my
> > > > machine.  The original count (1000) is too small to be sampled.
> > > >
> > > > >
> > > > > Could add some loop?
> > > > > Or possibly increase the sampling frequency in perf (-F or -c)?
> > > > Maybe, I will have a try.
> > > Turned out all "Indirect call" test can be resolved by adding -c 100
> > > to perf command line:
> > > diff --git a/gcc/config/i386/gcc-auto-profile b/gcc/config/i386/gcc-auto-profile
> > > ...
> > > -exec perf record -e $E -b "$@"
> > > +exec perf record -e $E -c 100 -b "$@"
> > >
> > > Is 100 too small here?  Or is it fine for all scenarios?
> >
> > -c 100 is risky because it can cause perf throttling, which
> > makes it lose data.
> Right, it looks suspicious to me too.
> 
> >
> > perf has a limiter that if the PMU handler uses too much CPU
> > time it stops measuring for some time. A PMI is 10k+ cycles,
> > so doing one every 100 branches is a lot of CPU time.
> >
> > I wouldn't go down that low. It is better to increase the
> > iteration count.
> We can combine the two together, increasing iteration count and
> decreasing perf count at the same time.  What count would you suggest
> from your experience?

Normally nothing less than 50k for a common event like branches.

But for such a limited test 10k might still work, as long
as the runtime is fair controlled.

We would probably need to ensure at least 10+ samples, so
that would be 100k iterations.

iirc that was what we used originally, until people
complained about the simulator run times.

-Andi