Aaron Turon     Archive     Feed

Shipping specialization: a story of soundness

Rust’s impl specialization is a major language feature that appeared after Rust 1.0, but has yet to be stabilized, despite strong demand.

Historically, there have been three big blockers to stabilization:

  • The interplay between specialization rules and coherence, which I resovled in an earlier blog post.

  • The precise ways in which specialization employs negative reasoning, which will be resolved by incorporating ideas from Chalk into the compiler.

  • The soundness of specialization’s interactions with lifetimes. The RFC talks about this issue and proposes a way to address it, but it has never been implemented, and early attempts to implement it in Chalk have revealed serious problems.

I’ve been wrestling, together with nmatsakis, withoutboats and others, with these soundness issues.

Spoiler alert: we have not fully solved them yet. But we see a viable way to ship a sound, useful subset of specialization in the meantime. Feel free to jump to “A modest proposal” if you just want to hear about that.

This blog post is an attempt to write up what we’ve learned so far, with the hopes that it will clarify that thinking, and maybe open the door to you cracking the nut!

The problem

In stable Rust, it is not possible for lifetimes to influence runtime behavior. This is partly an architectural issue, and partly a design issue:

  • Architecture: the compiler erases lifetime information prior to monomorphization and code generation, meaning that the generated code simply has no way to depend on lifetimes. That could be changed, but we’d have to work hard to avoid code blowup by generating separate copies of code for each lifetime it was used within, assuming that the behavior didn’t change.

  • Design: lifetime inference generally chooses the smallest lifetime that fits the constraints at any given moment. That means that you can have a piece of data that is valid for the 'static lifetime, yet is viewed as having a shorter lifetime. Having runtime behavior depend on these choices seems bound to result in confusion and bugs.

Unfortunately, specialization makes the story more difficult:

trait Bad1 {
    fn bad1(&self);
}

impl<T> Bad1 for T {
    default fn bad1(&self) {
        println!("generic");
    }
}

// Specialization cannot work: trans doesn't know if T: 'static
impl<T: 'static> Bad1 for T {
    fn bad1(&self) {
        println!("specialized");
    }
}

fn main() {
    "test".bad1()
}

What does this program print? Since the string literal "test" has type &'static str, you might expect the second, specialized impl to be used (and hence to get specialized as the output). But, as explained above, from the perspective of trans this type will look like &'erased str, making it impossible to know whether the more specialized impl can safely be used.

Here’s another, less obvious example:

trait Bad2<U> {}

impl<T, U> Bad2<U> for T {}

// Specialization cannot work: trans doesn't know if two refs have equal lifetimes
impl<'a, T, U> Bad2<&'a U> for &'a T {}

Here, the second impl is requiring that two lifetimes are the same, and once more for trans we can’t tell whether the impl safely applies.

On the other hand, simply naming a lifetime that must exist, without constraining it, is fine:

trait Good {}

impl<T> Good for T {}

// Fine: specializes based on being *any* reference, regardless of lifetime
impl<'a, T> Good for &'a T {}

In addition, it’s in principle okay for lifetime constraints to show up as long as they don’t influence specialization:

trait MustBeStatic {}

impl<T: 'static> MustBeStatic for T {}

// Potentially fine: *all* impls impose the 'static requirement; the dispatch is
// happening purely based on `Clone`
impl<T: 'static + Clone> MustBeStatic for T {}

Why does this lead to unsoundness?

So far, it might seem like we can just be conservative in trans, which could lead to confusing behavior but is otherwise alright.

Sadly, it’s not, at least given the original design of specialization:

trait Bomb {
    type Assoc: Default;
}

impl<T> Bomb for T {
    default type Assoc = ();
}

impl Bomb for &'static str {
    type Assoc = String;
}

fn build<T: Bomb>(t: T) -> T::Assoc {
    T::Assoc::default()
}

fn main() {
    let s: String = build("Uh oh");
    drop(s) // typeck and trans disagree about the type of `s`
}

The problem here: specialization as originally designed will allow the typechecker to conclude that T::Assoc is String if it knows that T is &'static str. That’s because the impl for &'static str does not use the default keyword when defining its associated type, meaning that no further specialization is allowed (so the type checker knows everything there is to know).

But trans, of course, sees &'erased str instead, and so cannot safely use the specialized impl. That means that trans will make the call to build return (), but the rest of the code assumed that a String was returned.

Oops.

(Spoiler alert: the “as originally designed” bit above is a give-away of where we’re ultimately going to end up…)

Some “solutions” that don’t work

Before giving my proposed way forward, let me explain why some of the solution that are probably coming to mind don’t work out.

Can’t we just rule out “bad” specializations?

It’s very tempting to blame the specialized impls for Bad1 and Bad2 above, since they clearly impose lifetime constraints. Maybe we could just make it an error to do so.

Unfortunately, the trait system is very powerful, and you can “hide” lifetime constraints within other trait impls that don’t involve specialization. Worse still: the problem can arise from two independent crates, each of which is doing something seemingly reasonable.

////////////////////////////////////////////////////////////////////////////////
// Crate marker
////////////////////////////////////////////////////////////////////////////////

trait Marker {}
impl Marker for u32 {}

////////////////////////////////////////////////////////////////////////////////
// Crate foo
////////////////////////////////////////////////////////////////////////////////

extern crate marker;

trait Foo {
    fn foo(&self);
}

impl<T> Foo for T {
    default fn foo(&self) {
        println!("Default impl");
    }
}

impl<T: marker::Marker> Foo for T {
    fn foo(&self) {
        println!("Marker impl");
    }
}

////////////////////////////////////////////////////////////////////////////////
// Crate bar
////////////////////////////////////////////////////////////////////////////////

extern crate marker;

pub struct Bar<T>(T);
impl<T: 'static> marker::Marker for Bar<T> {}

////////////////////////////////////////////////////////////////////////////////
// Crate client
////////////////////////////////////////////////////////////////////////////////

extern crate foo;
extern crate bar;

fn main() {
    // prints: Marker impl
    0u32.foo();

    // prints: ???
    // the relevant specialization depends on the 'static lifetime
    bar::Bar("Activate the marker!").foo();
}

The problem here is that all of the crates in isolation look perfectly innocent. The code in marker, bar and client is accepted today. It’s only when these crates are plugged together that a problem arises – you end up with a specialization based on a 'static lifetime. And the client crate may not even be aware of the existence of the marker crate.

If we make this kind of situation a hard error, we could easily end up with a scenario in which plugging together otherwise-unrelated crates is impossible. Or where a minor version bump in one dependency could irrevocably break your code.

Can we make a knob: “lifetime-dependent” vs “specializable”?

Thinking more about the previous example, you might imagine the problem is that the Marker trait ends up being used in two incompatible ways:

  • It’s used in a specialization, the second Foo impl.
  • It’s used in impls that constrain lifetimes (the Bar impl).

It’s the combination of these things that gets us into trouble. And each one arises from a different crate. So you might be tempted to add an attribute, say #[lifetime_sensitive], which allows for impls that constrain lifetimes but prevents use in specialization.

In other words, the Marker trait could say, in advance, whether the Foo impls or the Bar impl are acceptable.

There are several downsides to this idea, but the real death-knell is that “constraining lifetimes” is a surprisingly easy thing to do. To wit:

trait Sneaky {
    fn sneaky(self);
}

impl<T> Sneaky for T {
    default fn sneaky(self) {
        println!("generic");
    }
}

impl<T> Sneaky for (T, T) {
    fn sneaky(self) {
        println!("specialized");
    }
}

fn main() {
    // what does this print?
    ("hello", "world").sneaky()
}

Here we have a specialized impl that doesn’t mention any lifetimes or any other traits; it just talks about the type (T, T). The problem is that it’s asking for the two tuple components to have the same type, which means that if a lifetime appears, it must be the same in both.

Once more, when we go to trans the main function, we’ll be invoking sneaky on the type (&'erased str, &'erased str), and we can’t tell for sure whether the more specialized impl applies.

But saying that you can never repeat a type within a specialization would be very restrictive. And there’s always the worry that we’ve missed other sneaky ways to constrain lifetimes…

Can we make trans smarter?

At this point it becomes tempting to start blaming trans. After all, if we tracked lifetime information all the way through, wouldn’t that solve everything?

It would solve some things: it would make specialization sound. But at a high cost.

As explained at the outset, tracking information through trans would involve a massive overhaul of the compiler, and we’d have to be very smart about coalescing code with different lifetimes but identical behavior. There’s no guarantee we could do this without making the compiler significantly slower and/or creating more code bloat.

More fundamentally, though, it would lead to highly unpredictable behavior:

trait Print() {
    fn print(self);
}

impl<'a, T> Print for &'a str {
    fn print(self) {
        println!("Arbitrary str: {}", self);
    }
}

impl<T> Print for &'static str {
    fn print(self) {
        println!("'static str: {}", self);
    }
}

fn print_str(s: &str) {
    s.print()
}

fn main() {
    let s = "hello, world!";
    s.print();
    print_str(s);
}

Does this program print 'static str: hello, world! twice?

No! Because the call to print_str will reborrow the string slice at a shorter lifetime, and so trans will monomorphize it differently.

Making program behavior sensitive to the exact rules around lifetime inference and reborrowing seems extremely risky.

A modest proposal

Hopefully the above gives you some taste of the challenge here. Later in this post we’ll look at some more promising, clever solutions. But none of them have worked out completely, so I want to pause here and propose an incremental step forward.

First off, we add a new feature gate, assoc_specialization, which is needed whenever you use default type in an impl. We then focus on stabilizing just the core specialization feature, i.e. without being able to specialize associated types. That immediately means we can stop worrying about making type checking and trans agree, since type checking will essentially no longer care about specialization.

Many uses of specialization, including most of the original motivating examples, do not need to be able to specialize associated types.

With that out of the way, we still have work to do at the trans level. In particular, we must ensure that trans is conservative when it comes to lifetime constraints. The proposal here is twofold:

  • Any time a specialized impl imposes any lifetime constraints not present in the more general impl, trans uses the more general impl instead.

  • However, in these cases, we trigger an error-by-default lint to warn that a possibly-applicable specialization is not being used. (This needs to be a lint, not a hard error, because the relevant impls aren’t always under your crate’s control.)

Let’s revisit some of the earlier examples in this light:

// Specialization cannot work: trans doesn't know if T: 'static:
trait Bad1 {
    fn bad1(&self);
}

impl<T> Bad1 for T {
    default fn bad1(&self) {
        println!("generic");
    }
}

impl<T: 'static> Bad1 for T {
    fn bad1(&self) {
        println!("specialized");
    }
}

fn main() {
    // prints `generic`, but also generates a warning
    "test".bad1()
}

For this example, trans would pick the generic implementation, but issue a warning that a specialization might have applied. You could imagine going further and detecting simple cases like this where a given impl will never be used (as in the second impl of Bad1) and issuing errors. But as explained above, we cannot catch them all.

On the other hand, consider this case:

trait MustBeStatic {}
impl<T: 'static> MustBeStatic for T {}
impl<T: 'static + Clone> MustBeStatic for T {}

Here, both impls impose 'static constraints, so the second impl doesn’t impose any new lifetime constraints, and trans can choose it.

To make this work, in trans, when we query the trait system we replace each instance of 'erased with a distinct, fresh lifetime variable, which is a simple way to encode that anything we deduce in the query must be valid for all sets of unerased lifetimes. The Chalk approach will make this quite easy to do.

Even for the cases we’re covering, though, it’s possible to do better (we’ll see more on that later). That means we might want to “improve” the behavior of trans after stabilizing the core of specialization. Fortunately, we should be able to statically detect all cases where the behavior of trans would change, and issue a different warning that the behavior will improve. That gives us leverage to use something like epochs to make trans smarter over time, while still shipping some version of specialization relatively soon.

The only alternative seems to be to continue to pursue increasingly clever solutions before shipping anything—which is a worrying approach to take when it comes to soundness. Better, in my opinion, to ship a sound 80% of the feature now, with some rough edges, and improve it over time.

Going deeper

Before I close out this post, I want to write out some of the further explorations we’ve done, and what we’ve learned.

Here’s an interesting example:

trait Special {
    fn special(&self);
}

impl<T> Special for T {
    default fn special(&self) {
        println!("generic");
    }
}

impl<T> Special for (T, T) {
    fn special(&self) {
        println!("specialized");
    }
}

fn pair<T: Clone>(t: T) {
    (t.clone(), t).special()
}

fn main() {
    pair("hi");
}

Using the strategy outlined above, trans will go from (&'erased str, &'erased str) to (&'a str, &'b str) and hence use the generic implementation (and issue a lint that the more specific impl is being ignored). However, type check could deduce that the more specialized impl always applies when invoking special in pair, and you could imagine communicating that information down to trans.

What’s going on here is that type check sees things before monomorphization, and trans sees them afterward. In this particular case, that ends up making trans more conservative, since it can’t tell that two appearances of 'erased always come from the same, single lifetime.

The story changes if we add one layer of “indirection” around trait dispatch:

trait Special {
    fn special(&self);
}

impl<T> Special for T {
    default fn special(&self) {
        println!("generic");
    }
}

impl<T> Special for (T, T) {
    fn special(&self) {
        println!("specialized");
    }
}

fn use_special<T: Special>(t: T) {
    t.special()
}

fn pair<T: Clone>(t: T) {
    use_special((t.clone(), t))
}

fn main() {
    pair("hi");
}

Now at type checking time, the actual use of special occurs in a context where we don’t know that we’ll always be using the more specialized version.

Why harp on this point? Well, for one, it’s the main issue in allowing for sound specialization of associated types. We can see this in a variant of the Bomb example:

trait Bomb {
    type Assoc: Default;
}

impl<T> Bomb for T {
    default type Assoc = ();
}

impl<T> Bomb for (T, T) {
    type Assoc = String;
}

fn build<T: Bomb>(t: T) -> <(T, T) as Bomb>::Assoc {
    <(T, T) as Bomb>::Assoc::default()
}

fn main() {
    let s: String = build("Uh oh");
    drop(s) // typeck and trans disagree about the type of `s`
}

Here, again, type check knows that the relevant uses of Bomb all involve types of the form (T, T) and therefore can use the specialized version, and that could be communicated to trans. But, once more, adding a layer of indirection makes that much harder:

trait Bomb {
    type Assoc: Default;
}

impl<T> Bomb for T {
    default type Assoc = ();
}

impl<T> Bomb for (T, T) {
    type Assoc = String;
}

fn indirect<T: Bomb>() -> T::Assoc {
    T::Assoc::default()
}

fn build<T: Bomb>(t: T) -> <(T, T) as Bomb>::Assoc {
    indirect::<(T, T)>()
}

fn main() {
    let s: String = build("Uh oh");
    drop(s) // typeck and trans disagree about the type of `s`
}

The problem is that type check can no longer tell trans to use the specialized impl in the call to Assoc::default, but it is still assuming that the specialized impl is used externally (i.e., in the build function).

To sum up, there are two inter-related places where type check and trans differ:

  • Lifetime erasure
  • Monomorphization

We can partly deal with the first of these by introducing fresh lifetime variables for each lifetime that appears in type check, just as we do for trans—basically asking for the trait system to only find answers that would apply for arbitrary lifetime choices.

The monomorphization issue, though, appears much harder to cope with. One possible avenue is to track impl choices in a way that crosses functions, in other words allowing the knowledge from build that the specialized impl of Bomb can be used to be used when monomorphizing and generating code for indirect. Niko tells me that, in ancient times, the compiler used to do something much like this—and that it was incredibly painful and complicated.

In any case, taking these further steps would appear to require substantial additional work, and it seems hard to achieve confidence in their soundness. So dropping associated type specialization for now, where it’s relatively easy to argue for soundness, seems like the right step to take (@arielb1, here’s where you prove me wrong).