[llvm] [AMDGPU] Fix wrong uniformity for ballot NOT in UniformIntrinsicCombine (PR #220009)

[Jay Foad via llvm-commits]

================
@@ -86,7 +86,10 @@ static bool optimizeUniformIntrinsic(IntrinsicInst &II, UniformityInfo &UI,
           // Case: (icmp eq %ballot, 0) -> xor %ballot_arg, 1
           Instruction *NotOp =
               BinaryOperator::CreateNot(Src, "", ICmp->getIterator());
----------------
jayfoad wrote:

It would be better to go back to a conservatively correct implementation that considers all newly created values to be divergent.

https://github.com/llvm/llvm-project/pull/220009