When i added clamp i thought it should be consistent with the standard library and panic when min > max, max.is_nan or min.is_nan but i realized now that panics make it impossible to do branching via blend.

If this is some scalar code:

fn example(value: f32, min: f32, max: f32, should_clamp: bool) {
    if should_clamp {
        value.clamp(min, max)
    } else {
        value
    }
}

In that example it is ok to call example with NaNs when should_clamp is false, but for a SIMD version this will fail because clamp is still ran for all lanes:

fn example(value: f32x4, min: f32x4, max: f32x4, should_clamp: f32x4) {
    should_clamp.blend(
        value.clamp(min, max),
        value,
    )
}

The SIMD version will panic if any lane of min or max has NaNs even if should_clamp is false for that lane.

Too bad i did not think of this earlier.