https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99613
--- Comment #8 from Michal Zientkiewicz <michalz at nvidia dot com> ---
Jakub: You read coorectly, I was checking for global construction/destruction
order of many variables. I agree that a global lock is a heavy-handed solution
- and likely the only one that would always guarantee destruction order.
The example I posted is not very useful - the variables are independent. There
are cases where moving atexit would help:
Thread A Thread B
acquire lock on S1
initialize S1
release lock on S1
acquire lock on S2
initialize S2, which depends on S1
S1 is seen as initialized -> proceed
release lock on S2
atexit destroy S2
atexit destroy S1
Now we have reversed destruction order on _dependent_ variables and moving
atexit inside the per-variable lock scope would fix the issue.
