[llvm] [AMDGPU] Lazily emit waitcnts on function entry (PR #73122)

[Jay Foad via llvm-commits]

jayfoad wrote:

> Should callWaitsOnFunctionEntry still return true all the time?

Yes. It's really saying whether we should wait before making a call (i.e. waitcnts guaranteed to be zero on entry to the callee) or after entering the callee (i.e. waitcnts undefined on entry to the callee), and this patch does not change that.

https://github.com/llvm/llvm-project/pull/73122