Re: perf report with newest perf not showing jump targets on "most" C++ code

Re: perf report with newest perf not showing jump targets on "most" C++ code

[Travis Downs]

Just keeping the dream alive that this will be fixed one day. It's
still making perf report hard to use in C++.

As noted above it was introduced in 751b1783da784299b0509adb6a9cd3024cc4f837.


On Tue, Jun 25, 2019 at 1:18 PM Travis Downs <travis.downs@gmail.com> wrote:
>
> Hi,
>
> Just wondering if there is anything I can do to help move this along.
>
> I guess it hasn't been reported elsewhere yet because this perf
> version hasn't hit the mainsteam, but I guess it will break in a
> widespread way for C++ perf report users when it starts to get
> deployed by the distros.
>
> Cheers,
> Travis
>
>
> On Tue, May 21, 2019 at 8:38 PM Travis Downs <travis.downs@gmail.com> wrote:
> >
> > Hi Arnaldo,
> >
> > Yes, --stdio2 worked well (but a bit tricky because --stdio2 was
> > introduced fairly close in the commit stream to this issue)!
> >
> > I bisected it to:
> >
> > 751b1783da784299b0509adb6a9cd3024cc4f837
> > perf annotate: Mark jumps to outher functions with the call arrow
> >
> > Indeed, in the detailed description we have:
> >
> > "... [f]or now lets just make some progress by marking jumps to outside the
> >     current function as call like ..."
> >
> > And empirically I see most jumps being marked as call-like after this
> > change. Note however: with this change I see a middle-ground behavior
> > compared to what I described earlier: all the arrows on the jumps look
> > call-like, but when you select them in report tui it still correctly
> > shows the jump arrow. However, later on (by v4.20 or earlier) this
> > jump indicator disappears too.
> >
> > Here's some reproducing code:
> >
> > ---
> > #include <algorithm>
> > #include <memory>
> > #include <random>
> >
> > const size_t size = 10000000;
> >
> > int main() {
> >
> > auto buf = std::unique_ptr<uint64_t[]>(new uint64_t[size]);
> >
> > auto b = buf.get(), e = b + size;
> > std::iota(b, e, 0);
> > std::shuffle(b, e, std::default_random_engine{});
> >
> > clock_t before, after;
> >
> > before = clock();
> > std::sort(b, e);
> > after = clock();
> >
> > printf("Soring took %5.2f ns/elem\n", (after - before) * 1000000000. /
> > size / CLOCKS_PER_SEC);
> > }
> > ---
> >
> > I compile it with:
> >
> > g++ -O3 -g -march=haswell -Wall -Wextra -std=c++11  -c bench.cpp
> >
> > If you run `perf annotate --stdio2 --no-source -Mintel` you see something like:
> >
> >              Disassembly of section .text:
> >
> >              0000000000400dc0 <void std::__introsort_loop<unsigned
> > long*, long, __gnu_cxx::__ops::_Iter_less_iter>(unsigned long*,
> > unsigned long*, long, __gnu_cxx::__ops::_Iter_less_iter) [clone
> > .isra.20]
> >              _ZSt16__introsort_loopIPmlN9__gnu_cxx5__ops15_Iter_less_iterEEvT_S4_T0_T1_():
> >   0.03         mov    rax,rsi
> >                sub    rax,rdi
> >                cmp    rax,0x87
> >   0.03       → jle    400f5f <void std::__introsort_loop<unsigned
> > long*, long, __gnu_cxx::__ops::_Iter_less_iter>(unsigned long*,
> > unsigned long*, long, __gnu_cxx::__ops::_Iter_less_iter) [clone
> > .isra.20]+
> >   0.18         push   r14
> >                mov    r14,rdx
> >                push   r13
> >                mov    r13,rdi
> >   0.05         push   r12
> >   0.05         push   rbp
> >                push   rbx
> >                test   rdx,rdx
> >              → je     400eeb <void std::__introsort_loop<unsigned
> > long*, long, __gnu_cxx::__ops::_Iter_less_iter>(unsigned long*,
> > unsigned long*, long, __gnu_cxx::__ops::_Iter_less_iter) [clone
> > .isra.20]+
> >   0.03         lea    rbp,[rdi+0x10]
> >              _ZSt27__unguarded_partition_pivotIPmN9__gnu_cxx5__ops15_Iter_less_iterEET_S4_S4_T0_():
> >                sar    rax,0x4
> >   0.15         mov    rcx,QWORD PTR [r13+0x8]
> >              _ZSt16__introsort_loopIPmlN9__gnu_cxx5__ops15_Iter_less_iterEEvT_S4_T0_T1_():
> >   0.03         sub    r14,0x1
> >              _ZSt27__unguarded_partition_pivotIPmN9__gnu_cxx5__ops15_Iter_less_iterEET_S4_S4_T0_():
> >                lea    rdx,[r13+rax*8+0x0]
> >                mov    rdi,QWORD PTR [rsi-0x8]
> >   0.03         mov    rax,QWORD PTR [rdx]
> >              _ZSt22__move_median_to_firstIPmN9__gnu_cxx5__ops15_Iter_less_iterEEvT_S4_S4_S4_T0_():
> >                cmp    rcx,rax
> >              → jae    400eb0 <void std::__introsort_loop<unsigned
> > long*, long, __gnu_cxx::__ops::_Iter_less_iter>(unsigned long*,
> > unsigned long*, long, __gnu_cxx::__ops::_Iter_less_iter) [clone
> > .isra.20]+
> >   0.08         cmp    rax,rdi
> >   0.18       → jb     400ebe <void std::__introsort_loop<unsigned
> > long*, long, __gnu_cxx::__ops::_Iter_less_iter>(unsigned long*,
> > unsigned long*, long, __gnu_cxx::__ops::_Iter_less_iter) [clone
> > .isra.20]+
> >   0.03         cmp    rcx,rdi
> >              → jae    400ed6 <void std::__introsort_loop<unsigned
> > long*, long, __gnu_cxx::__ops::_Iter_less_iter>(unsigned long*,
> > unsigned long*, long, __gnu_cxx::__ops::_Iter_less_iter) [clone
> > .isra.20]+
> >              _ZSt4swapImEvRT_S1_():
> >   0.10         mov    rdx,QWORD PTR [r13+0x0]
> >
> >
> > Note all the "call-like arrows", even though most of those jumps are
> > nearby in the same function. This function is [clone .isra] so a
> > cloned function with certain values inlined. Maybe annotate/report is
> > having trouble in those cases, thinking the jump is to another
> > function?
> >
> > Cheers,
> > Travis
> >
> >
> > On Wed, May 15, 2019 at 9:38 PM Arnaldo Carvalho de Melo
> > <arnaldo.melo@gmail.com> wrote:
> > >
> > > On May 15, 2019 10:04:05 PM GMT-03:00, Travis Downs <travis.downs@gmail.com> wrote:
> > > >Hi Arnaldo,
> > > >
> > > >I'll work on a repro case. One issue though: the arrows show up in the
> > > >interactive tui mode, but to git bisect I'd like to script it, so have
> > > >some
> > > >arrow indication in --stdio mode: any way to do that?
> > >
> > > Try --stdio2, maybe it's enough.
> > >
> > > - Arnaldo
> > > >
> > > >Cheers,
> > > >Travis
> > > >
> > > >On Tue, May 14, 2019, 5:01 PM Arnaldo Carvalho de Melo <
> > > >arnaldo.melo@gmail.com> wrote:
> > > >
> > > >> Em Sun, May 05, 2019 at 06:58:51PM -0500, Travis Downs escreveu:
> > > >> > For a few months the newest perf (build from Linus' tip) doens't
> > > >show
> > > >> > jump targets in most C++ code for me.
> > > >> >
> > > >> > The arrays on jcc instructions always point to the left, like for
> > > >> > jumps that can't be resolved, even when the jumps are only a few
> > > >bytes
> > > >> > away.
> > > >> >
> > > >> > I tried to reproduce it on a "simple" C++ looping process, with the
> > > >> > same compile flags as my project, but it worked there! So something
> > > >is
> > > >> > triggering failure to resolve jump targets.
> > > >> >
> > > >> > Any advice on how to bisect or diagnose this issue would be
> > > >> > appreciated, e.g,. some way to turn on logging to see why perf
> > > >report
> > > >> > fails to resolve jump targets.
> > > >> >
> > > >> > FWIW perf as included in my distribution (perf version 4.15.18)
> > > >works
> > > >> > OK on the same files (it's perf report that matters, not perf
> > > >record)
> > > >> > but doesn't demangle the C++ symbols.
> > > >>
> > > >> Adding this to my backlog, but you could just try to use normal 'git
> > > >> bisect', i.e. find a version that works for a workload you have, then
> > > >to
> > > >> on rebuilding it till you find the version that introduce the
> > > >problem.
> > > >>
> > > >> If you can send some simple C++ source code that exhibits the
> > > >problem,
> > > >> please do, I'll try to work on this as soon as I process my
> > > >backlog...
> > > >>
> > > >> I just came back from vacation.
> > > >>
> > > >> - Arnaldo
> > > >>
> > >