https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111589
Bug ID: 111589
Summary: Use relaxed atomic increment (but not decrement!) in
shared_ptr
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libstdc++
Assignee: unassigned at gcc dot gnu.org
Reporter: redbeard0531 at gmail dot com
Target Milestone: ---
The atomic increment when copying a shared_ptr can be relaxed because it is
never actually used as a synchronization operation. The current thread must
already have sufficient synchronization to access the memory because it can
already deref the pointer. All synchronization is done either via whatever
program-provided code makes the shared_ptr object available to the thread, or
in the atomic decrement (where the decrements to non-zero are releases that
ensure all uses of the object happen before the final decrement to zero
acquires and destroys the object).
As an argument-from-authority, libc++ already is using relaxed for increments
and acq/rel for decements:
https://github.com/llvm/llvm-project/blob/c649fd34e928ad01951cbff298c5c44853dd41dd/libcxx/include/__memory/shared_ptr.h#L101-L121
This will have no impact on x86 where all atomic RMWs are effectively
sequentially consistent, but it will enable the use of ldadd rather than
ldaddal on aarch64, and similar optimizations on other weaker architectures.
