[llvm] [AMDGPU] ISel & PEI for whole wave functions (PR #145858)

[Diana Picus via llvm-commits]

rovka wrote:

> What is the fundamental motivation of doing this?

The fundamental motivation for doing this is that the WWM intrinsics are fragile and don't handle control flow very well. We have an increasing number of use cases where we need to have some control flow inside a WWM region, and that isn't supported that well. For one of these use cases we've introduced the `init.whole.wave` intrinsic, but that's kind of hacky. Whole wave functions are meant to replace all that with something more robust (kind of like a MachineIR region).

https://github.com/llvm/llvm-project/pull/145858