[llvm] [AMDGPU] Cgscc amdgpu attributor (PR #179719)

[Shilei Tian via llvm-commits]

shiltian wrote:

Some attributes are implemented with assumption that it can see all callsites (not necessarily a closed-world but similar at least based on the programming model we support), especially for those to propagate from leaves to kernels, which is not the case for CGSCC. I'm not sure what we are trying to achieve here. What is the benefit of running the full attributor but just at different scale? I'd expect to split the pass into something that can run earlier (in any form necessary), and something that we want to run later in the pipeline. Simply just doing this doesn't look right.

https://github.com/llvm/llvm-project/pull/179719