[llvm] [amdgpu] Add llvm.amdgcn.init.whole.wave intrinsic (PR #105822)

[Nicolai Hähnle via llvm-commits]

https://github.com/nhaehnle commented:

Thanks! We've just had a quick offline discussion. To summarize, I think the overall flow of where stuff happens is good, but the optimization of the initial branch is subtle. I suggest splitting this up into multiple changes, where the first one implements init.whole.wave in a simple but hopefully obviously correct manner. The optimization of the initial branch can then be done as a separate change.

I also think we should try to find an alternative to using V_SET_INACTIVE, since the whole WWM infrastructure is awfully fragile. The closer we can get to "normal looking" IR between isel and register allocation, the better.

https://github.com/llvm/llvm-project/pull/105822