vld3* alignment requirement

[awxkee]

Current NEON vld3* loading implementation inserts an alignment requirement where none is required.

Consider following code:

use std::arch::aarch64::*;
use std::mem::MaybeUninit;

fn main() {
    let data: Vec<u8> = vec![0u8; 200];

    let base = data.as_ptr() as usize;
    let misaligned_offset = if base % 16 == 0 { 2 } else { 0 };

    let ptr = unsafe {
        data.as_ptr().add(misaligned_offset) as *const f32
    };
    unsafe {
        type W = [f32; 16]; 
        let mut mem = MaybeUninit::<W>::uninit();
        std::ptr::copy_nonoverlapping(
            ptr,                    
            mem.as_mut_ptr().cast::<f32>(),
            12,           
        );
        let _ = mem.assume_init();
    }

    println!("we're done");
}

Since copy_nonoverlapping requires both pointers to be aligned to at least align_of::<f32>(), passing a pointer that is not sufficiently aligned is an UB. I have not checked the release build, but this is immediately detected by the fuzzer

[sayantn]

Does the ARM spec say that the pointer can be unaligned? Because the ARM signature takes a float const*, so the natural assumption is that it should be mem::align_of::<f32>()-aligned (typically 32, but can be lower)

[awxkee]

I'm not aware of such requirements, intrinsics guide also doesn't enforce alignment. Even if it were required, vld2* and vld4* would be wrong since they don't require alignment.

[sayantn]

Oh I see, it is more of a consistency thing. cc @adamgemmell to confirm what is the correct behavior.

[adamgemmell]

Yep, LD3 has no requirements on alignment (which in the aarch64 architecture usually refers to the element size, not the register size anyway)