[clang] [llvm] [openmp] [offload][OpenMP] Add atomic cross-team reductions (PR #209298)

[Robert Imschweiler via cfe-commits]

ro-i wrote:

> Are you planning a follow up to turn this on depending on the target? [...]  I would like us to turn this on if we have reason to believe it's the faster path

Yes.

> you have good data for some HPC devices

Let me see if I can collect some more data about when/where atomic instructions are favorable to the default path. At the moment, I only have data for gfx90a and gfx942, which is not bad, but maybe I can get my hands on some more.

https://github.com/llvm/llvm-project/pull/209298