[llvm] [AMDGPU][Passes] Invoke CGSCCOptimizerLateEP callbacks in LTO pipeline (PR #203262)

[Joseph Huber via llvm-commits]

jhuber6 wrote:

> I think that wasn't invoked, probably by design. That EP is only used for per-TU compilation (and in ThinLTO it's still per-"TU"), so it might not make sense to run those passes on a monolithic module.
> 
> Instead, maybe we should insert some AMDGPU passes into other EPs?

Full LTO is still a TU in the strictest sense, is the concern compile times?

https://github.com/llvm/llvm-project/pull/203262