[llvm] [AMDGPU] Introduce V_READANYLANE_B32 (PR #178312)

[Stanislav Mekhanoshin via llvm-commits]

rampitec wrote:

> Probably the phrasing is what's implied by the node name "readanylane` can be refined to "readfirstlane" or to "readlane X" for any X and still return the same value - that is, the argument is uniform and could be put in SGPRs

I think what we are discussing here is not the semantics of the proposed operation, but the semantics of the uniformity. At the end it will be created for whatever UniformityAnalysis will deem uniform. My point is if we extend the participating lanes, then the analysis shall not tell it is uniform anymore. It shall be a new divergent value.

https://github.com/llvm/llvm-project/pull/178312