[llvm] [RFC] IR: Define noalias.addrspace metadata (PR #102461)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> I hope we watch out for and resist the temptation to use it to reify backend bugs as programming model "features", i.e., instead of declaring that some backend not supporting some op to some statespace is a backend bug, we end up declaring that performing that op to that statespace is UB in the programming model, and end up using this feature to encode that.

But that's what this is doing? We're making the backend handle the IR operation with fully correct generality, and frontends have to opt-in to get the optimizations in the cases the language specifications say are not required 

https://github.com/llvm/llvm-project/pull/102461