Surprised by the benchmark about iter fold and for loop

​

https:\/\/github.com\/mike-barber\/rust-zero-cost-abstractions

I'm surprised by the benchmark about iter fold and for loop, both of which use zip and mutate a variable:

// 3.66 μs
pub fn calculate_fold(slice_a: &[i32], slice_b: &[i32]) -> i64 {
    slice_a
        .iter()
        .zip(slice_b.iter())
        .fold(0_i64, |acc, (a, b)| match *a > 2 {
            true => acc + (*a as i64 * *b as i64),
            false => acc,
        })
}

// 6.6 μs: 1.83x slower than fold
pub fn calculate_direct(slice_a: &[i32], slice_b: &[i32]) -> i64 {
    let mut res = 0;
    for (a, b) in slice_a.iter().zip(slice_b.iter()) {
        if *a > 2 {
            res += *a as i64 * *b as i64;
        }
    }
    res
}

So what matters here? Is fold method preferred to for loops wrt iterators? I remember for var in iter is a zero-cost-abstraction. Did I misremember that?