ptr::copy_nonoverlapping slower then manual bytewise copy

[tower120]

The following code:

#[inline]
unsafe fn copy_bytes(src: *const u8, dst: *mut u8, count: usize){
    for i in 0..count{
        *dst.add(i) = *src.add(i);
    }
}

is significantly FASTER then ptr::copy_nonoverlapping, for small counts, and cases where count is not known at compile time. (2-3 times for 8 byte copy).

People suggest that this happens, because compiler does not inline copy_nonoverlapping code:
https://users.rust-lang.org/t/ptr-copy-nonoverlapping-slower-then-manual-per-byte-copy/75588/2?u=tower120

I'm not sure if this is actually a bug though...

[saethlin]

Similar to #92993

[mati865]

Related to #83785

[zeroflaw]

fyi, this now compiles down to the same assembly as copy_nonoverlapping.