[libcxx-commits] [libcxx] [libc++] Speed up classic locale (PR #70631)

[via libcxx-commits]

philnik777 wrote:

> > I'm a bit confused about this. The statements here seem contradictory. Could you try to reword this?
> 
> This change does 2 optimizations:
> 
>     1. Allow inlining of the common locale ctor/dtor.
> 
>     2. Avoid atomic reference counting for the classic locale.
> 
> 
> (1) has small fixed-size impact (shaves off few cycles). (2) may have very profound effect on highly threaded apps that use locale/streams concurrently (these numbers like -98.22%).
> 
> If inlining creates problems that are hard to resolve, then I am ready to drop it from this change and leave only optimization (2).

Ah, OK. I think that is the way to go right now. That should simplify this patch quite a bit. We can still look into inlining the functions later if we want to. Dropping this part will most likely also resolve any ABI issues.

https://github.com/llvm/llvm-project/pull/70631