[llvm] AMDGPU/GlobalISel: Fix tail call target in vgprs (PR #201873)

[Petar Avramovic via llvm-commits]

petar-avramovic wrote:

In theory, I think at least, uniformity analysis should figure out that tail call target is actually uniform, but in tests it is just vgpr register so not sure if there is better way then "readfirstlane just in case"

Regbank-legalize could optimize away readfirstlane for example if it has uniform input, or do we need dedicated intinsic for this?

https://github.com/llvm/llvm-project/pull/201873