Extend the "nullable pointer optimization" to two different size variants

[ordian]

Play:

#[allow(dead_code)]
enum StrOrChar<'a> {
    Str(&'a str),
    Char(char),
}

fn main() {
    println!("{}", std::mem::size_of::<&str>());
    println!("{}", std::mem::size_of::<Option<&str>>()); // current optimization
    println!("{}", std::mem::size_of::<StrOrChar>());   // proposed optimization
}

currently prints (on x86_64) 16, 16 and 24. But it could potentially be 16, 16 and 16.

This optimization should be possible if we have an enum with two variants a and b, the a variant contains non-nullable field and the size of a is greater than the size of b (probably size of b should be a divisor of size of a or 0 for alignment).

The current "nullable pointer optimization" is a special case with size of b equal to 0.

src link

[oli-obk]

So basically the char would just occupy the same memory as the length field of the slice? I think the example should use a u32, because a char enables so many other optimizations it could cause confusion.

Also I think this needs an RFC. Also note the related RFC issue: rust-lang/rfcs#1230

[ordian]

So basically the char would just occupy the same memory as the length field of the slice? I think the example should use a u32, because a char enables so many other optimizations it could cause confusion.

Right.

Also note the related RFC issue: rust-lang/rfcs#1230

Thanks for the link. Should I close this issue in favor of rust-lang/rfcs#1230? This optimization was mentioned in the comment.

[Mark-Simulacrum]

Yes, I think this is best tracked there -- we need an RFC to make changes in this area, I believe.