[llvm] [AMDGPU] Use table strategy for LowerModuleLDSPass at O0 (PR #160181)

[Jon Chesterfield via llvm-commits]

JonChesterfield wrote:

There is an underlying error here. We are too eager to promote variables to the structure that is allocated at address zero and that can lead to allocating variables in kernels that should not do so.

There's an easy correctness fix, I'm still considering whether there's a reasonable way to get a better result. Forcing table mode for everything would get you correct behaviour I believe, but you should get performance regressions from it relative to fixing the underlying error

https://github.com/llvm/llvm-project/pull/160181