Missed enum layout optimization with a NonZeroU64 + more space in an enum

[RalfJung]

The Provenance type defined here is 24 bytes in size:

use std::num::NonZeroU64;

pub enum Provenance {
    Concrete {
        alloc_id: NonZeroU64,
        sb: NonZeroU64,
    },
    Wildcard,
    None,
}

However, it should be possible to encode Provenance in 16 bytes: e.g. (0usize, 0suize) could encode None and (0usize, 1usize) could encode Wildcard.

In Miri, a slight variant of this would help reduce the size of a fairly common type from 32 bytes to 24 bytes.

Unfortunately even #94075 does not help here. The entire concept of reading a single field to determine the discriminant is not flexible enough to represent this layout.

[oxalica]

Wouldn't it cause severe performance issue for enum with a number of variants? It makes a single-variant check matches!(enum, Variant) load and compare multiple fields (up to O(number_of_niches)). It also disables jump table optimization.

It's really a trade-off instead of an optimization.

enum Enum {
    _0000, // .0 == 0 && .1 == 0 && .2 == 0 && .3 == 0
    _0001, // .0 == 0 && .1 == 0 && .2 == 0 && .3 != 0
    _0010,
    _0011,
    // ...
    _1110,
    _1111(Box<()>, Box<()>, Box<()>, Box<()>),
}

[RalfJung]

It's really a trade-off instead of an optimization.

It optimizes space usage, but overall I agree it's a tradeoff. We might want attributes like repr(compact) or so.

[hkalbasi]

It makes a single-variant check matches!(enum, Variant) load and compare multiple fields (up to O(number_of_niches)).

Why O(number_of_niches)? Isn't two field enough? Every value for the second field is a niche, so the example can be:

enum Enum {
    _0000, // .0 == 0 && .1 == 0
    _0001, // .0 == 0 && .1 == 1
    _0010, // .0 == 0 && .1 == 2
    _0011, // .0 == 0 && .1 == 3
    // ...
    _1110,  // .0 == 0 && .1 == 15
    _1111(Box<()>, Box<()>, Box<()>, Box<()>),
}

[mikebenfield]

There's even more generality. You don't even need the second field to have any "invalid values" at all. Even if the original enum had only one NonZeroU64 it would work:

use std::num::NonZeroU64;

pub enum Provenance2 {
    Concrete {
        alloc_id: NonZeroU64,
        sb: u64,
    },
    Wildcard, // .0 = 0, .1 = 0
    None, // .0 = 0, .1 = 1
}

[cuviper]

Related, if you have:

pub enum Provenance {
    Concrete {
        alloc_id: NonZeroU64,
        sb: NonZeroU64,
    },
    Wildcard,
}

This does niche to 16 bytes, but Option<Provenance> is still 24 bytes -- one could imagine finding the other niche for None. However, the compiler only ever tracks one largest niche AFAIK, and forgets about others.

[RalfJung]

Yeah, that is the case we need for Miri. But I figured we would start with the one where it is all a single enum as that should be easier?

[cuviper]

Hmm, actually the Wildcard variant would consider the unused niche to be just padding, so all bits are valid -- no longer niche for an Option wrapper.

[BGR360]

@rustbot label T-compiler A-layout

[mikebenfield]

I'm working on this. I'm introducing the concept of a Flag. A Flag is an extra scalar field that may be associated with a Niche. The flag must be set to a particular value to indicate we should find the discriminat in the Niche. Maybe I'll come up with a better name than Flag.

I've also got a commit to improve the flexibility of niche-optimized enums by allowing the fields of a tagged variant to go around the niche, with some appearing before and some after.

And I've got a commit to allow layouts to store more than just the largest_niche; this way multiple niches can be tried.

I should have a PR pretty soon.