Locale objects use atomic reference counting, which may be very
expensive in parallel applications. The classic locale is used by
default by all streams and can be very contended. But it's never
destroyed, so the reference counting is also completely pointless on the
classic locale. Currently ~70% of time in the parallel stringstream
benchmarks is spent in locale ctor/dtor. And the execution radically
slows down with more threads.
Avoid reference counting on the classic locale. With this change
parallel benchmarks start to scale with threads.
This is a re-application of f8afc53d64 (aka PR #72112) which was
reverted in 4e0c48b907 because it broke the sanitizer builds due
to an initialization order fiasco. This issue has now been fixed by
ensuring that the locale is constinit'ed.
Co-authored-by: Dmitry Vyukov <dvyukov@google.com>
1a5af34e6f