[llvm] [AMDGPU] Combine redundant ballot intrinsic calls (PR #218357)

[Anshil Gandhi via llvm-commits]

gandhi56 wrote:

> This should not be looking for "exec writes".
> 
> This kind of optimization needs to be performed in terms of convergence tokens

Sorry, do you mean I should use convergence tokens to figure out if an exec was modified or that I don't need to look for exec writes at all? It seems like a necessary condition to eliminate redundant ballot calls.

https://github.com/llvm/llvm-project/pull/218357