Should `Option<u128>` use a smaller (maybe usize) tag?

[scottmcm]

Looking at some examples in #137278 (comment) shows that Option<u128> is using a full u128 for its stored tag. That tag, of course, can only be 0 or 1.

As a result things like returning Some(…) need to write two qwords on x64 into the return value, even though of course one of them is only ever all-zeroes.

Could we maybe give it a layout equivalent to (usize, u128) instead?

Or, in general, can we cap the size of tags to a native machine word, given a small discriminant that could fit in just about any type?

(Aside: It's good for Option<u32> to still be (u32, MaybeUninit<u32>), since that has more niches than (u8, MaybeUninit<u32>). But going so big that it takes multiple instructions to write the tag seems like overkill, and for unconstrained repr(Rust) we should probably stop at something reasonable for the target.)

[lolbinarycat]

Unfortunately, we're constrained by this line from the reference:

The size of a value is always a multiple of its alignment

[lolbinarycat]

It might be possible to give Option a custom calling convention though, similar to how slices are passed as 2 arguments.

[hanna-kruppe]

This wouldn’t change alignment, just (possibly) introduce padding in the memory layout, just like how moving (u8, u128) only needs to load/store 9 bytes despite the tuple being 8 + align_of::<u128>() bytes.

[scottmcm]

It might be possible to give Option a custom calling convention though

We already do that -- Option<u32> uses ScalarPair layout. (Even Option<u128> does, it's just a pair of u128 and MaybeUninit<u128>.)

As does (usize, std::mem::MaybeUninit<u128>) https://rust.godbolt.org/z/qGavbhvoo, so this is just a request to make Option<u128> use the same layout as that tuple.

(Sure, it won't make it smaller in memory on any platform where alignof(u128) is 16, but that's ok.)