[llvm] PHIElimination: add target hook to control reuse. (PR #163604)

[Guy David via llvm-commits]

guy-david wrote:

Sorry, I don't think the presented use case, which argued about value semantics in GPU threads, is enough of a justification here because it's already using the wrong tools and a bigger hammer is not the correct approach.
Regarding the original issue you've mentioned- MLIR, for example, has has a nested scopes concept which would probably fit very well as a solution. LLVM does not have that kind of representation, maybe capturing values using intrinsics is as close as it gets.

https://github.com/llvm/llvm-project/pull/163604