shared_ptr's increment/decrement use sequential consistency.

[horenmar]

Bugzilla Link	22803
Resolution	FIXED
Resolved on	Jul 08, 2016 03:03
Version	unspecified
OS	All
CC	@mclow

Extended Description

Current shared_ptr implementation uses sequential consistency for reference counting, even though less strong atomics are good enough. This leads to suboptimal performance for platforms other than x86 (x64).

Relaxed and acq/rel atomics (for increment and decrement respectively) should be sufficient for thread safety. The reason it is currently implemented like this was that Clang at the time didn't support full range of C++11 atomics and noone ever got back to cleaning it up.

[horenmar]

assigned to @mclow

[mclow]

Reference: http://stackoverflow.com/questions/28199212/why-does-libcs-implementation-of-shared-ptr-use-full-memory-barriers-instead

[llvmbot]

Related Patch (that I should probably change to fix this):

http://reviews.llvm.org/D10406

[llvmbot]

Fixed in r241532.

[llvmbot]

Current shared_ptr implementation uses sequential consistency for reference
counting, even though less strong atomics are good enough. This leads to
suboptimal performance for platforms other than x86 (x64).

Relaxed and acq/rel atomics (for increment and decrement respectively)
should be sufficient for thread safety. The reason it is currently
implemented like this was that Clang at the time didn't support full range
of C++11 atomics and noone ever got back to cleaning it up.

Hi;candecrement() use relaxed atomics?