[llvm] AMDGPU: Fix temporal divergence introduced by machine-sink and performance regression introduced by D155343 (PR #67456)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> I agree. Do we actually have a proposal for how this should look? I've seen some ideas and had some myself, but haven't seen anything fully coherent yet.

No not really. I was most recently thinking of splitting isConvergent into some notion of read and write convergent, and then adding a corresponding flag to basic blocks that acts sort of like IsEHPad. You could then compose read/write convergent to approximately represent vector vs. scalar vs. wwm ops with different convergence types 

https://github.com/llvm/llvm-project/pull/67456