[llvm] [AMDGPU] SelectionDAG divergence tracking should take into account Target divergency. (PR #144947)

[via llvm-commits]

alex-t wrote:

> I think you would need a test containing something (probably an intrinsic) that is a source of divergence, but with workgroup size set to 1. Then subsequent operations that depend on the result of that intrinsic will actually be uniform, so the compiler can insert readfirstlane and use SALU instructions for them.

Ah yes. It is easy.

https://github.com/llvm/llvm-project/pull/144947