Sub-optimal layout of Option<(u16, [u8; 15])> (missed niche optimization)

[gendx]

I tried this code:

use std::mem::size_of;

fn main() {
    println!("(u16, [u8; 14]) = {} bytes", size_of::<(u16, [u8; 14])>());
    println!("(u16, [u8; 15]) = {} bytes", size_of::<(u16, [u8; 15])>());
    
    println!("Option<(u16, [u8; 14])> = {} bytes", size_of::<Option<(u16, [u8; 14])>>());
    println!("Option<(u16, [u8; 15])> = {} bytes", size_of::<Option<(u16, [u8; 15])>>());
    
    let mut x: Option<(u16, [u8; 15])> = None;
    let bytes: [u8; 20] = unsafe { std::mem::transmute(x) };
    println!("None = {:02x?}", bytes);

    x = Some((0xffff, [0xee; 15]));
    let bytes: [u8; 20] = unsafe { std::mem::transmute(x) };
    println!("Some = {:02x?}", bytes);
}

I expected to see this happen:

• (u16, [u8; 15]) has a size of 18 bytes, because there are 17 bytes of useful data and it is rounded up due to the alignment of 2 bytes.
• I expected Option<(u16, [u8; 15])> to also take 18 bytes due to a niche optimization.

Instead, this happened:

• Option<(u16, [u8; 15])> takes 20 bytes, and appears to be layout as:

struct Foo {
    discriminant: u16,
    x0: u16,
    x1: [u8; 15],
    padding: u8,
}

I expected the Option to leverage the byte of padding at the end of the struct to decide if it's None or Some instead of pre-pending a discriminant.

(u16, [u8; 14]) = 16 bytes
(u16, [u8; 15]) = 18 bytes
Option<(u16, [u8; 14])> = 18 bytes
Option<(u16, [u8; 15])> = 20 bytes
None = [00, 00, 00, 00, 54, d1, 5c, fc, b6, 7f, 00, 00, 05, 00, 00, 00, 00, 00, 00, 00]
Some = [01, 00, ff, ff, ee, ee, ee, ee, ee, ee, ee, ee, ee, ee, ee, ee, ee, ee, ee, f7]

Meta

Tested on the playground: https://play.rust-lang.org/?version=stable&mode=release&edition=2021

---

This is inspired by the tinyvec crate, which is roughly layout as follows (https://docs.rs/tinyvec/latest/src/tinyvec/tinyvec.rs.html#97-103, https://docs.rs/tinyvec/latest/src/tinyvec/arrayvec.rs.html#103-107) - although the niche wouldn't apply in that case (and the smallvec crate has a better layout):

pub enum TinyVec<T, const N: usize> {
    Inline((u16, [T; N])),
    Heap(Vec<T>),
}

[cuviper]

Implicit padding is not a niche, because any bits are valid there in memory -- e.g. for Option, there's no value that we could read in the padding and say that this must be a None. (even setting aside whether padding bytes are undef)

Another way to think about this: Given an Option<Foo(u16, u8)> that is Some, you can get &mut Foo and pass that to arbitrary code that has no idea about the original Option layout. Any Foo value can be written there that clobbers the padding byte as well, writing the full size_of::<Foo>().

[cuviper]

See also #70230.

[ryanavella]

I have a crate struct-pad that I wrote for exactly this purpose. The implementation is simple enough that you could write it yourself.

Changing your Option<(u16, [u8; 15])> to Option<(u16, [u8; 15], PadU8)> shrinks the size from 20 bytes to 18 bytes.