[PATCH AutoFDO]Restoring indirect call value profile transformation

[Bin.Cheng]

On Tue, Jan 15, 2019 at 1:10 AM Andi Kleen <ak@linux.intel.com> wrote:
>
> On Mon, Jan 14, 2019 at 04:15:20PM +0800, Bin.Cheng wrote:
> > On Mon, Jan 14, 2019 at 4:07 PM Andi Kleen <ak@linux.intel.com> wrote:
> > >
> > > Bin Cheng,
> > >
> > > I did some testing on this now. The attached patch automatically increases the iterations
> > > for autofdo profiles.
> > Hi Andi, thanks very much for tuning these.
> > >
> > > But even with even more iterations I still have stable failures in
> > >
> > > FAIL: gcc.dg/tree-prof/cold_partition_label.c scan-assembler foo[._]+cold
> > > FAIL: gcc.dg/tree-prof/cold_partition_label.c scan-assembler size[ \ta-zA-Z0-0]+foo[._]+cold
> > I think these two are supposed to fail with current code base.
>
>
Sorry for being late.
> We should mark it as XFAIL then I guess.
Yeah
>
> Is it understood why it doesn't work?
I didn't dig into it, but I think the reason is autofdo::zero count is
not considered as cold as profile count.  It's kind of reasonable,
otherwise too much code would be categorized as cold.
>
> > > FAIL: gcc.dg/tree-prof/indir-call-prof.c scan-ipa-dump afdo "Indirect call -> direct call.* a1 transformation on insn"
> > I also got unstable pass/fail for indirect call optimization when
> > tuning iterations, and haven't got an iteration number which passes
> > all the time.  I guess we need to combine decreasing of sampling count
> > here.
>
> Okay I will look into that.
>
> Could also try if prime sample after values help, this sometimes fixes
> problems with systematically missing some code in sampling.
>
> > > FAIL: gcc.dg/tree-prof/peel-1.c scan-tree-dump cunroll "Peeled loop ., 1 times"
> > This one should fail too.
>
> Same.
Maybe I shouldn't use words "should fail", but the reason is we don't
consider autofdo profile count as reliable in niters analysis.  Maybe
this can be improved somehow.

Thanks,
bin
>
> -Andi