[clang] [RFC][Clang][AMDGPU] Emit only delta target-features to reduce IR bloat (PR #176533)

[Matt Arsenault via cfe-commits]

arsenm wrote:

> This may not be easy to apply to all targets, especially for CPU, because of LTO where you can link IR for one CPU against IR for another. I don’t have enough knowledge on that though. If that is safe, I'd be happy to do.

Linking isn't a problem, the problem is only with the contents vs. attributes of a single function. This is only a problem if you are dropping the target-cpu attribute, in which case code doing so would need to be responsible for filling out the appropriate feature set with whatever subset is valid for the new merged pseudo-cpu. I don't think anything attempts to do that 

https://github.com/llvm/llvm-project/pull/176533