[llvm] [AMDGPU] introduce S_WAITCNT_FENCE_soft emitted by memory legalizer (PR #150167)

[Sameer Sahasrabuddhe via llvm-commits]

ssahasra wrote:

> If I understand correctly, this does not (in its current form) defer waitcnt insertion entirely to InsertWaitCnt; it just adds a context pseudo before the waitcnts that are usually inserted, right? That means we're one small patch away from moving all of the waitcnt insertion from the legalizer to InsertWaitCnt?

Yes, I would love to submit a patch that moves *all* wait count insertion to the waitcnt inserter (because after all, that is its name). But there is no hurry in fixing what isn't broken right now.

https://github.com/llvm/llvm-project/pull/150167