This trick is very useful on Nvidia GPUs for calculating mins and maxes in some cases, e.g. atomic mins (better u32 support than f32) or warp-wide mins with `redux.sync` (only supports u32, not f32).