Looking for a deeper understanding of PhantomData

In the dark nomicon: https://doc.rust-lang.org/nomicon/phantom-data.html

To me, I just use phantomdata whenever I need to include a lifetime that doesn't happen to be present within a struct's fields, but may in its surrounding context be needed.

Who has a good understanding of PhantomData, its purpose, and what the table in the nomicon (variance, contravariance, invariance, etc) means?

This other chapter describes variance:
https://doc.rust-lang.org/nomicon/subtyping.html

PhantomData is also useful for influencing auto-traits ("OIBITs"), like Send and Sync. You can manually unsafe impl !Send etc. for each one, or you can get them all in one stroke with a phantom type. You can see this in action in the Void type used by fmt::Arguments.

3 Likes

Let's say you have a generic type, Foo<T>. How do you figure out if Foo<T> is covariant (or is Sync)?

There can be two approaches:

  • The language can require that the users explicitly specify properties like variance. For example, in Scala you write +T or -T
  • The compiler can infer variance from definition, so that Foo(fn(T)) is contrvariant and Foo(T) is covariant.

Rust is in the second category.

But what if the type parameter is unused, like in struct Foo<T> {}? The Rust's solution is to error out if the type parameter is unused. And PhatomData is a way to add a fake use of type parameter, for the purposes of inference of type's properties.

3 Likes

What is PhantomData<T>

The two defining properties of PhantomData<T> are the following:

  • At runtime, PhantomData<T> is exactly like (), i.e., nothing.

    • This gives the peace of mind of it being (a) zero-cost (abstraction)
  • For compile-time analysis, PhantomData<T> is exactly like T

    • The utility of this may not be obvious at first glance, but it turns out that types have properties (or more precisely, they can lack properties), which in turn leads to the containing struct transitively lacking these properties too.

(Some) Type Properties

1) auto traits

The most intuitive of these "properties" are auto traits (a.k.a., OIBITs, as @cuviper said). These are (marker) traits that any type implements, unless the type contains / is made of some specific type that does not implement it. So instead of "white listing" the implementors of the trait, as usual, one goes and "black lists" the root non-implementors of the trait.

Example

Imagine wanting to have a (marker) trait for types of data that is truly immutable when behind a shared reference:

trait ImmutableIfAliased {}

In Rust, data behind a shared reference is immutable unless it contains somewhere an UnsafeCell<_>, the root element behind all the constructions offering aliased mutability / interior mutabilty (Cell, RefCell, Mutex, RwLock). So this makes for a nice candidate of a black list approach:

/// If a type does not contain any `!ImmutableIfAliased` stuff,
/// and if it is not  excplicitely marked `!ImmutableIfAliased`
/// itself, then it automagically becomes `ImmutableIfAliased`
auto
trait ImmutableIfAliased {}

struct Foo {
    x: i32,
}

const_assert_clauses!( Foo : ImmutableIfAliased );

And now let's "black list" UnsafeCell:

impl<T : ?Sized> !ImmutableIfAliased for UnsafeCell<T> {}
  • UnsafeCell<T> : ImmutableIfAliased scratched successfully

Now, if the trait was made unsafe (to implement), then this could be used in a generic function <T : ImmutableIfAliased> to know that the data pointed to by a &T is truly immutable (at least for the duration of the reference).

2) 'static lifetime vs. genericity

It turns out that being 'static is also one such property: a struct is 'static as longs as it does not contain a non-'static element:

const_assert_clauses!( Foo : 'static );

As soon as a struct contains an element with a lifetime parameter that is not 'static (i.e., a generic lifetime parameter), then the whole struct becomes generic and tied to this lifetime parameter:

struct FancyReferenceToAnInteger<'lifetime> (
    &'lifetime i32,
);

In other words, a type genericity "infects" the containing type, which, in turn, becomes generic too:

enum Option<T> {
    None,
    Some(T),
}

Applications to PhantomData<_>

1) impl !Trait

In stable Rust (at least as of 1.37), one cannot "unimplement" / blacklist an auto trait:

struct CannotBeSharedAcrossThreads {
    // fields
}
impl !Sync for CannotBeSharedAcrossThreads {}

yields

error[E0658]: negative trait bounds are not yet fully implemented; use marker types for now

This can be circumvented thanks to the infectious nature of the !Sync property: since for all types T, Cell<T> is not Sync, we could go and add an unused Cell<T> field to CannotBeSharedAcrossThreads:

struct CannotBeSharedAcrossThreads {
    // fields
    _unused: Cell<i32>,
}

Now our type is soundly not Sync, but it does have to carry an extra data field, integer-sized. It can be made smaller, as Cell<u8>, or even actually zero-sized / non-existent at runtime, with Cell<()>:

struct CannotBeSharedAcrossThreads {
    // fields
    _unused: Cell<()>, // weird: are we intending to mutate `()` ?
}

but this only works because we could choose any T for our Cell<T> (plus it would be sound for Cell<()> to be Sync, since ZSTs cannot be mutated, and in that case our abstraction would break). In some cases, we may need to use a type that has no zero-sized-version of its properties.

That's when PhantomData shines: PhantomData<T> has all the properties (and lack thereof) that T has, while also being guaranteed to be zero-sized:

type PhantomNotSync = PhantomData<Cell<u8>>;

struct CannotBeSharedAcrossThreads {
    // fields
    _not_sync: PhantomNotSync, // => impl !Sync for CannotBeSharedAcrossThreads 
}

we can also have a PhantomNotSendNorSync:

type PhantomNotSendNorSync = PhantomData<&'static Cell<u8>>;

with which we can create PhantomNotSend:

#[derive(Default)]
pub
struct PhantomNotSend (
    PhantomNotSendNorSync,
);
// Protect the type from being infected by `!Sync`
unsafe impl Sync for PhantomNotSend {}

2) Adding type/lifetime parameters: artificially generic structs

In a similar fashion to unimplementing auto traits, one may want to have a struct have a(n additional) type or lifetime parameter, despite it not really having a generic type. One such case is when trying to define pointer types less restrictive / with weaker semantics than Rust references, but not as raw as plain pointers.

Example: out references

Let's imagine we want to have an API working with so-called "out pointers" for the sake of performance (or to enable more complex stack pinning abstractions):

// instead of
fn get_ft ()
  -> i32
{
    42
}

// do 
fn write_ft (out: &'_ mut i32)
  -> &'_ mut i32
{
    *out = 42;
    out
}

Now, write_ft works fine, but it requires that the caller have already initialized the memory where write_ft writes (having a &mut i32 point to uninitialised memory is currently UB).

It would be nice to have an API like the following:

struct OutRef<'a, T : 'a> { ... }

impl<'a, T : 'a> OutRef<'a, T> {
    fn write (self: OutRef<'a, T>, value: T)
      -> &'a mut T
    { ... }
}

impl<T> MaybeUninit<T> {
    fn as_out_ref(self: &'_ mut MaybeUninit<T>)
      -> Out<'_, T>
    { ... }
}

impl<'a, T : 'a> From<&'a mut T> for OutRef<'a, T> { ... }

This way we could write:

fn write_ft (out: OutRef<'_, i32>)
  -> &'_ mut i32
{
    out.write(42)
}

fn main ()
{
    let mut x = MaybeUninit::<i32>::uninit();
    let at_x = x.as_out_ref();
    // drop(x);
    let x: &mut i32 = write_ft(at_x);
}

The fact that OutRef includes a lifetime parameter, that is then used in the as_out_ref method to force the borrow on the MaybeUninit to live as long the the OutRef lives, makes it so that if the drop(x) line is uncommented, the program no longer compiles.
Whereas if we had used MaybeUninit's .as_mut_ptr() as our "out pointer", then the program (with the uncommented drop) would have compiled despite it being unsound.

Implementation

OutRef is nothing more than a raw pointer:

struct OutRef<'a, T : 'a> {
    ptr: *mut T,
}

which leads to a compilation error due to the parameter 'a not being used (if this were accepted, then the struct would be bivariant over 'a, thus allowing to cast any OutRef<'a, T> to OutRef<'static, T>, and we would be back to just having a raw pointer (i.e., it is not the outer genericity that matters, but how the generic parameters are used within a struct)):

struct OutRef<'a, T : 'a> {
    ptr: *mut T,
    _lifetime: PhantomData<&'a ()>,
}

So, much like in the !Sync case and Cell, we have used here the type &'a () since it is generic over 'a, and then wrapped it in a PhantomData to make it be a zero-cost abstraction.

Then we implement the rest of our imagined API:

// Invariant: ptr points to a valid `MaybeUninit<T>`
pub
struct OutRef<'a, T : 'a> {
    ptr: *mut T,
    _lifetime: PhantomData<&'a ()>,
}

impl<'a, T : 'a> OutRef<'a, T> {
    pub
    fn write (self: OutRef<'a, T>, value: T)
      -> &'a mut T
    {
        unsafe {
            // # Safety
            //
            //   - this is `MaybeUninit<T>::write`
            self.ptr.write(value);
            &mut *self.ptr
        }
    }
}

pub
trait MaybeUninitExt<T> {
    fn as_out_ref (self: &'_ mut Self)
      -> OutRef<'_, T>
    ;
}
impl<T> MaybeUninitExt<T> for MaybeUninit<T> {
    fn as_out_ref(self: &'_ mut MaybeUninit<T>)
      -> OutRef<'_, T>
    {
        #[allow(unused_unsafe)]
        unsafe {
            // # Safety
            //
            //   - Respects the invariant
            OutRef {
                ptr: self.as_mut_ptr(),
                _lifetime: PhantomData,
            }
        }
    }
}

impl<'a, T : 'a> From<&'a mut T> for OutRef<'a, T> {
    #[inline]
    fn from (p: &'a mut T)
      -> OutRef<'a, T>
    {
        #[allow(unused_unsafe)]
        unsafe {
            // # Safety
            //
            //   - Respects the invariant, since something
            //     init can be seen as maybeuninit
            OutRef {
                ptr: p,
                _lifetime: PhantomData,
            }
        }
    }
}

Non-Zero optimisation

So, everything is actually quite fine, but our struct is not as optimal as it could be. More precisely,

Option<OutRef<'_, T>>

is not pointer-sized, which is a pity, since given our invariant, we know that our pointer is not null. Luckily, the ::core library offers a handy structure for non-nullable raw pointers: ptr::NonNull, with which we hurry and update our struct definition:

struct OutRef<'a, T : 'a> {
    ptr: ptr::NonNull<T>,
    _lifetime: PhantomData<&'a ()>,
}

Hooray, all is good, in our brave new code!

Except when someome comes and posts that our library is unsound, since it allows the following code to compile:

fn main ()
{
    let mut s: &'static str = "Static str"; // s pointer to static str
    let at_s = OutRef::<&'static str>::from(&mut s); // pointer to s
    {
        let s = String::from("Short-lived str");
        at_s.write(&s); // override s with a pointer to the short-lived str
    }
    dbg!(s); // Use after free!
}

Now, you may be wondering, why did such obviously wrong code compile? Was the compiler not paying attention?

Well, let me tell you: unless your code is cooler than ICE, the compiler does not miss anything. Rust accompanies us at each step of our path, very gently pulling our hand when we are too close to falling onto a (safety) hole, and also very gently letting us fall all the way down the hole, as soon we spell the forbidden incantation: unsafe.

And we have indeed used unsafe, more precisely, we have written:

struct OutRef<'a, T : 'a> {
    ptr: ptr::NonNull<T>,
    _lifetime: PhantomData<&'a ()>,
}

impl<'a, T : 'a> OutRef<'a, T> {
    pub
    fn write (self: OutRef<'a, T>, value: T)
      -> &'a mut T
    {
        unsafe {
            // # Safety
            //
            //   - this is `MaybeUninit<T>::write`
            self.ptr.write(value);
            &mut *self.ptr
        }
    }

Now, the problem with this code can be incredibly tricky to spot, since it involves one of the most arcane mechanics of the language:

Variance (informally)

Variance is the property involved when trying to sub-type. Sub-typing is a process whereby T1 is said to be a subtype of a type T2 (I will write it as T1 : T2) if the compiler accepts any expression of type T1 where the type T2 was expected.

Oh, like when Rust accepts any reference to an array &[T; N] where a reference to a slice &[T] is expected? Or when Rust accepts a Box<T : Trait> where a Box<dyn Trait> is expected? Or when a &String becomes a &str?

Nope, nope, and nope. All these cases are (implicit) coercions, meaning that a transformation takes place:

  • &[T; N] == *const T becomes &[T] == (*const T, usize),

  • Box<T> becomes Box<dyn Trait> = (Box<TypeErased>, *const Trait::VTable)

  • &String becomes &str by Dereference of String: it's actually &*String

Subtyping, on the other hand, does not involve tranformation, just compiler type-level tolerance.

One good way to check for subtyping is with the following macro (AFAIK, no (implicit) coercions behind a raw pointer):

macro_rules! const_assert_subtypes {(
    for [<$($generics:tt)*]
    $T1:ty : $($T2:tt)*
) => (
    const _: () = {
        fn foo<$($generics)* (x: *const $T1)
          -> *const $($T2)*
        {
             x
        }
    };
)}

const_assert_subtypes!(
    for [<T>]
    T : T
);

const_assert_subtypes!(
    for [<'a, T>]
    &'static T : &'a T
);

const_assert_subtypes!(
    for [<'a, 'b, T>]
    &'b T : &'a T
    where 'b : 'a
);

const_assert_subtypes!(
    for [<'a, 'b : 'a, T>]
    fn(&'a T) : fn(&'b T)
);

const_assert_subtypes!(
    for [<'a, 'b : 'a, T>]
    &'b mut T : &'a mut T
);

// Errors
// const_assert_subtypes!(
//     for [<'a, 'b : 'a>]
//     &'a mut &'b () : &'a mut &'a ()
// );

So now that we know some examples of subtyping, let's try to resume them:

  • Rust only subtypes regarding lifetimes: if 'a : 'b (read "if the lifetime 'a lives at least as long as 'b"), then &'a T is a subtype of &'b T: hence the famous "lifetimes can shrink".

  • Subtyping builds up by using compound types.

Variance is then the property of a generic type (that is, asking for the variance of a non-generic type makes no sense) w.r.t. how different monomorphisations of the type may subtype each other.

Example: given the type Ref<'a> = &'a (), we know that Ref<'a> : Ref<'b> when / where 'a : 'b. In that case, Ref is said to be covariant.

  • F<T> is said to be covariant iff for all T1 : T2 => F<T1> : F<T2>

    • in other words: F preserves subtyping

    Examples:

    • type F<T> = T;

    • Thus, we also have:
      type F<T> = PhantomData<T>;

    • When T<'a> = &'a [mut] _, this gives:

      type F<'a> = &'a ();
      type F<'a> = &'a u8;
      type F<'a> = &'a Whatever;

      type F<'a> = &'a mut ();
      type F<'a> = &'a mut u8;
      type F<'a> = &'a mut Whatever;

    • type F<T> = Box<T>;
      type F<T> = [A]Rc<T>;
      type F<T> = *const T;
      type F<T> = ptr::NonNull<T>;

    • More generally, covariance is the usual behavior, and the exceptions are the examples in the other categories of variance.

  • F<T> is said to be contravariant iff for all T1 : T2 => F<T2> : F<T1>

    • in other words: F reverses subtyping

    Example:

    • type F<T> = fn(T, ...) -> Whatever
  • F<T> is said to be bivariant iff it is both covariant and contravariant.

    • This could be theoretically be possible for the function fn id<T> (x: T) -> T { x }, but in practice there is no bivariance in Rust
  • F<T> is said to be invariant iff it is neither covariant nor contravariant.

    Examples:

    • The main rule is the following: mutation usually leads to invariance;
      More specifically: non owning-mutation has to be invariant:

    • type F<T> = &'whatever mut T being the best known case,

      • With T<'a> = &'a Whatever, it gives:
        type F<'a> = &'whatever mut &'a Whatever

      • Careful:
        type F<'a, T> = &'a mut T is invariant in T but is covariant in 'a

    • type F<T> = *mut T;

    • UnsafeCell<T> is also invariant (it is the mutability wrapper by definition).

This, by the way, shows the only difference between *const _ and *mut _: the former is covariant, whereas the latter is not (=> invariant).

Back to the soundness issue

Let's recap the code with the soundness issue:

struct OutRef<'a, T : 'a> {
    ptr: ptr::NonNull<T>,
    _lifetime: PhantomData<&'a ()>,
}

impl<'a, T : 'a> OutRef<'a, T> {
    pub
    fn write (self: OutRef<'a, T>, value: T)
      -> &'a mut T
    {
        unsafe {
            // # Safety
            //
            //   - this is `MaybeUninit<T>::write`
            self.ptr.write(value);
            &mut *self.ptr
        }
    }
}

fn main ()
{
    let mut s: &'static str = "Static str"; // s pointer to static str
    let at_s = OutRef::<&'static str>::from(&mut s); // pointer to s
    {
        let s = String::from("Short-lived str");
        at_s.write(&s); // override s with a pointer to the short-lived str
    }
    dbg!(s); // Use after free!
}

So, the problem is that our OutRef<'a, T>, which is morally a &'a mut T (another implementation, by the way, would have been &'a mut MaybeUninit<T>), and should thus be:

  • covariant in 'a :white_check_mark:

    • &'a () is covariant in 'a
  • invariant in T :x:

    • ptr::NonNull<T> is covariant in T !

So our OutRef<'_, T> was covariant in T, which means that OutRef<'_, &'static str> was a subtype of OutRef<'_, &'short_lived str> since 'static : 'short_lived,
so the call

at_s.write(&s);
// i.e.
<OutRef<T?>>::write(
    // self: OutRef<T?>
    at_s, // : OutRef<&'static str> : OutRef<&'short_lived str>
    // value: T?
    &s, // : &'short_lived str
); // Thus T? = &'short_lived str

was not prevented by Rust, thus allowing to write a pointer to a short lived local where a pointer to a longer lived local was expected; which is unsound, and is the reason why non-owning mutable "references" (of any sort) need to be invariant.

Solution

If you imagine Covariant<{"T"}> as a marker trait for covariance over the generic parameter T, then, to get Invariant<{"T"}> = !Covariant<{"T"}>, it suffices to have one of the fields of our struct not be covariant in T. In our case, the most simple fix is to change our _lifetime: PhantomData<&'a ()> field into

struct OutRef<'a, T : 'a> {
    ptr: ptr::NonNull<T>, // : Covariant<{"T"}>
    _marker: PhantomData<&'a mut T>, // : !Covariant<{"T"}>
}
// => OutRef<T> : !Covariant<{"T"}>

which should have been the case anyways, given how our OutRef was a form of mutable reference (before the NonNull<T> "upgrade", our code was actually sound thanks to having used *mut T, invariant, as the pointer type. Had we used *const T instead we would have had the same problem).

Quick rule of thumb about variance in unsafe code

When in doubt, be invariant! This (potentially conservative) choice is never unsound; it may just hurt some call-site ergonomics for the user of the library / API.

Addendum The "ownership" type property

PhantomData<T> may also be used to express ownership over T when the struct otherwise uses T within a raw pointer, so as to make Rust's drop checker prevent some configurations that could be unsound.

I detailed this usage in this other post.

22 Likes

This was a fantastic explanation!

2 Likes

Absolutely fantastic post!

Minor nit, shouldn't this be &'a mut MaybeUninit<T>, given that ManuallyDrop can never be uninitialized, and OutRef might point to uninitialized memory?

1 Like

Another example of the "artificially generic type" use case: I use it when I want to wrap a dynamically-typed API in a generic, strongly-typed layer, e.g. in this MongoDB driver I wrote, I use PhantomData to give collections a type parameter, making them homogeneous.

1 Like

Indeed, the name got Mangled when writing it down :sweat_smile:

2 Likes

I love it! You have made a great contribution for this community. Thank you!

2 Likes