[llvm] [AMDGPU] Make S_WAIT_EVENT a scheduling boundary (PR #129032)

[Jay Foad via llvm-commits]

jayfoad wrote:

> > @kerbowa your suggestion to use `sched_barrier(0)` in the front end seems to be working well for our use case, so I may abandon this patch.
> 
> I think the scheduler should try to do reasonable things without these type of markers

I agee. I would prefer to schedule s_wait_event as late as possible by default, maybe as-if it is dependent on another instruction with infinite latency. But I can't find a nice way to implement that.

https://github.com/llvm/llvm-project/pull/129032