https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85759
Martin Liška <marxin at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|WAITING |ASSIGNED
--- Comment #14 from Martin Liška <marxin at gcc dot gnu.org> ---
Got it, it's really happening. Funny how that happens.
It's related to a known issue with -fprofile-generate=directory. All *.gcda
files are placed directly in the directory, object files subfolder location is
ignored. I have a patch draft here:
https://gcc.gnu.org/ml/gcc-patches/2017-12/msg01429.html
I'll finish it for GCC 9.1.
So when ./bash is invoked, it does a fork. When doing a fork, we flush counters
and mark that runs count has been already saved. Then a child runs, and it
writes alias.gcda for a different TU:
marxin@marxinbox:/tmp/bash-4.4.18> find . -name alias.o
./builtins/alias.o
./alias.o
And when parent process terminates it attempts to open alias.gcda and realizes
it's from a different TU. Thus it's rewritten, but with runs=1 because it was
already streamed. Then we see the fishy gcda file.
Nice example, thanks for providing a KVM image.
