[llvm] [AMDGPU] PromoteAlloca: flatten homogeneous structs to vectors (PR #217055)

[Yaxun Liu via llvm-commits]

yxsamliu wrote:

@arsenm Re: SROA support — I think you mean #165159. That change converts homogeneous structs to vectors for `memcpy`/`memset` flows. It does not handle dynamic-index GEPs because SROA needs constant offsets. This PR handles those dynamic-index cases in AMDGPUPromoteAlloca, while PromoteAlloca does not handle copies between different objects or allocas used through pointer selects/PHIs. The two changes are complementary.

https://github.com/llvm/llvm-project/pull/217055