[llvm] [AMDGPU] Improve selection of ballot.i64 intrinsic in wave32 mode in SelectionDAG. (PR #71556)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> I would expect it to fail to select if the type does not match the wave size. Is there a real use case for it?

It's a synthetic operation to begin with and has a pretty clear meaning, just 0 the high bits. It's still valid to read exec/exec_hi in wave32 mode, it's just the high bits are always 0. This mirrors the same behavior (such that we are now using this to implement the read_exec builtins)

https://github.com/llvm/llvm-project/pull/71556