Bug 2120752

Description of problem:

When examining the performance of linux perf doing a report I found that excessive time (88% of total samples) being spent in the bfd function lookup_func_by_offset on Fedora 36. Nick Clifton suggested trying rawhide as there were some efficiency patches in rawhide's version of binutils.  Nick identified "Reduce-O-n2-performance-overhead-when-parsing-DWARF" as the patch that will improve performance.


Version-Release number of selected component (if applicable):
binutils-2.37-27.fc36.x86_64

How reproducible:

Every time.


Steps to Reproduce:

dnf download --source kernel-tools
dnf download --source systemtap
sudo dnf builddep ./kernel-tools*.src.rpm -y
sudo dnf builddep ./systemtap*.src.rpm -y
rpm -Uvh ./kernel-tools*.src.rpm
rpm -Uvh ./systemtap*.src.rpm
cd ~/rpmbuild/SPECS
rpmbuild -bp kernel-tools.spec
rpmbuild -bc systemtap.spec
cd ~/rpmbuild/BUILD/kernel*/linux-*/tools/perf
make
mkdir ~/bin
make install
cd ~/rpmbuild/BUILD/systemtap*
make
sudo make install
cd

# warm things up with and avoid measuring some systemtap setup done on the first run:
stap --example -p2 sleeptime.stp

#actual data collection
~/bin/perf record -e cpu-clock --output=perf_record.data --call-graph=dwarf -- ~/bin/perf record -e cpu-clock --call-graph=dwarf --output=perf_stap.data stap --example -p2 sleeptime.stp
~/bin/perf record -e cpu-clock --output=perf_report.data --call-graph=dwarf -- ~/bin/perf report --input=perf_stap.data --stdio > junk

# See where time is in perf processing the 

~/bin/perf report --input=perf_report.data --dso=perf --stdio > where_perf_report_spends_time.log

# measure how long it takes to generate the report
 /usr/bin/time ~/bin/perf report --input=perf_report.data --dso=perf --stdio  > /dev/null

Actual results:

Fedora 36 is about 4 times slower than the rawhide version.

f36:
$ /usr/bin/time ~/bin/perf report --input=perf_report.data --dso=perf --stdio  > /dev/null
3.02user 1.23system 0:04.29elapsed 99%CPU (0avgtext+0avgdata 532680maxresident)k
0inputs+0outputs (0major+373420minor)pagefaults 0swaps

rawhide:
$ /usr/bin/time ~/bin/perf report --input=perf_report.data --dso=perf --stdio  > /dev/null
0.66user 0.55system 0:01.25elapsed 97%CPU (0avgtext+0avgdata 351120maxresident)k
0inputs+0outputs (0major+124043minor)pagefaults 0swaps



Expected results:

On Fedora 36 the "perf report" shouldn't be spending 85%+ of the time in a single function, lookup_func_by_offset.  It should have a time comparable to rawhide.

Additional info:

May need to rebuild kernel-tools as lookup_func_by_offset is being inlined.