MaxGraey wrote: If we implement per-word `x == ~y`, we need to make sure to mask (exclude) the most significant cleared bits from most significant (or single) word. I'm afraid this won't be much more efficient than before. https://github.com/llvm/llvm-project/pull/191919