Unnecessary loop unrolling to handle tail when tail length has a smaller known size

In the following code, the first while loop should process 8 bytes at a time and exit early if an invalid byte is found.

The remaining bytes should be known to be bytes.len() % 8, but the auto-vectorization unrolls to test again for 32 bytes and 8 bytes at a time
https://rust.godbolt.org/z/z8hnb9PGY

pub const fn is_ascii(bytes: &[u8]) -> bool {
    const N1: usize = 8;
    let mut i = 0;
    while i + N1 <= bytes.len() {
        let chunk_end = i + N1;
        let mut count = 0;
        while i < chunk_end {
            count += (bytes[i] <= 127) as u8;
            i += 1;
        }

        if count != N1 as u8 {
            return false;
        }
    }

    // Process the remaining `bytes.len() % N` bytes.
    let mut is_ascii = true;
    while i < bytes.len() {
        is_ascii &= bytes[i] <= 127;
        i += 1;
    }
    is_ascii
}

Start by reproducing the Rust example in the linked Godbolt comparison and inspect the generated output for the tail loop. Trace the compiler optimization responsible for the extra unrolling, then verify that the remainder is handled without redundant tests while preserving the example's correctness.