[llvm] [AMDGPU] Add llvm.amdgcn.dead intrinsic (PR #123190)

[Jay Foad via llvm-commits]

jayfoad wrote:

> Unfortunately, this kind of phi node will get optimized away and the backend won't be able to figure out that it's ok to use the active lanes of `%inactive.vgpr*` inside `shader`.

Does InstSimplify/InstCombine optimize away a phi if all but one of its inputs is undefined? I can see why it might be tempting to do that to simplify the IR, but it also increases liveness (and hence register pressure) in a way that is impossible to undo later. @nikic do you have any thoughts/comments about this?

https://github.com/llvm/llvm-project/pull/123190