[llvm] AMDGPU: Replace sqrt OpenCL libcalls with llvm.sqrt (PR #74197)

[Stanislav Mekhanoshin via llvm-commits]

rampitec wrote:

> nsz doesn't do anything for sqrt lowering.

I know it does not. Why not remove it?

I will feel much more confident if this one is fixed from nsz nonsense and extra 'fast' test added, because who knows what fast will mean tomorrow? Anyway, formally to have no regressions fast must resolve to a native op, so this test is formally needed.

https://github.com/llvm/llvm-project/pull/74197