[llvm] [AMDGPU] Reenable BackOffBarrier on GFX11/12 (PR #155370)

[Pierre van Houtryve via llvm-commits]

Pierre-vh wrote:

> > Re-enable it by adding a wait on vm_vsrc before every barrier "start" instruction.
> > This is a less strong wait than what we do without BackOffBarrier,
> 
> Description should mention that we only add the wait in CU mode? Is that why you describe it as "less strong" than without BackOffBarrier?

Yes, and yes. I updated the description.

The wait is less strong because it only waits on the "coherency point" for a workgroup in CU mode, instead of waiting on the operation to complete entirely

https://github.com/llvm/llvm-project/pull/155370