[llvm] [AMDGPU] Added Downward Propagation for the attribute "amdgpu-agpr-alloc" (PR #208847)

[via llvm-commits]

JoshuaGrindstaff wrote:

> The RA is per function, so I'm not sure how the caller's allocation would affect the callee's, unless the kernel eventually reports fewer than needed, which should not happen here.

You're right, RA is independent for each function. I'm not saying the register allocation of the caller is affecting the callee, I'm saying that the vgpr and agpr maximums given to RA can be mismatched between callee and caller functions.

https://github.com/llvm/llvm-project/pull/208847