[llvm] [AMDGPU] Add alignment attr & propagate alignment through make.buffer.rsrc inst (PR #145278)

[via llvm-commits]

Shoreshen wrote:

> Description should be more clear about what this is. This isn't adding anything to the intrinsic itself.
> 
> I think the only thing special about this case is it's an intrinsic that propagates a pointer operand, similar to llvm.ptrmask. It looks like AAAlign also fails to propagate the alignment (which may have been improved by the ptrmask). Can you start by adding support for llvm.ptrmask, and then generalizing the support to other pointer intrinsics?

Hi @arsenm , to support llvm.ptrmask need to update AAAlign's ability. 

Currently the only thing we can do is to tell AAAlign that the alignment of the call is the same to some operand of the call.

By my understanding `%ret = call ptr addrspace(7) @llvm.ptrmask.p7.i32(ptr addrspace(7) %p, i32 -4)` means `ret = p & (-4)`, this need calculation based on the 2 operand......

https://github.com/llvm/llvm-project/pull/145278