line numbers don't match with autofdo output for llvm

[harubyy]

Hello,
I'm working on a simple bubblesort code. I'm on Ubuntu 16.04, I compile it with LLVM5 and generate the profile output using autofdo as shown below.

clang -O2  -g sort.c -o sort_g_O0
sudo perf record -e cpu/event=0xc4,umask=0x20,name=br_inst_retired_near_taken,period=400009/pp -b ./sort_g_O0
sudo ../../autofdo/create_llvm_prof --binary=./sort_g_O0 --out=sort_g_O0.prof

Then I use llvm-prof to read this sampled profile. The output gives me some weird line numbers that do not match with the source code.

The (partial) output of
llvm-profdata show -sample sort_g_O0.prof
is this:
2: inlined callee: sort_array: 583159, 0, 4 sampled lines
Samples collected in the function's body {
1: 0
3: 1
4: 1
65501: 1
}
Samples collected in inlined callsites {
6: inlined callee: bubble_sort: 583142, 0, 6 sampled lines
Samples collected in the function's body {
2: 3
4: 12608
5: 16740
7: 4209
8: 4209
65516: 12673
}
No inlined callsites in this function

I tried to figure it out myself,

• I changed the optimization level to 0 and I still get the weird line number in sort_array function.
• I printed the offset and line numbers in llvm_profile_writer.cc and indeed something is wrong. But I'm lost tracking it back to dwarf info.

Can somebody help me on this? Am I doing something wrong, if not, why am I seeing these line numbers?

I attached the files.
llvm_autofdo.zip

Thanks!

[danielcdh]

Sorry for the late reply.

The "line number" is not line number but "line offset to the start of the function". If you see a very large number like 65516, it's normally because the sampled instruction has incorrect debug info or no debug info, thus it is mapped to the line 0 of the file, giving it an negative "line offset to the start of the function". The profile you provided looks reasonable and should be good to use.

Dehao

[harubyy]

Ah, I see, then yes, there's no problems. Thank you very much for the information. And sorry for my late reply as well.