[Bug target/96898] [nvptx] libatomic support

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96898

--- Comment #2 from Tom de Vries <vries at gcc dot gnu.org> ---
Hmm, I found this difference: 
- AFAIU, GOMP_atomic_start/end have barrier semantics
- libatomics protect_start/end are always paired with explicit barriers, so
  presumably these don't have barrier semantics

So, using GOMP_atomic_start for protect_start in libatomics will have the
effect of issuing the barrier twice, which might be a performance problem.