Llogiq on stuff

Opportunistic Mutations

As you may know, my current mutagen project deals with mutation testing in Rust. However, as I remarked, Rust’s famed flexibility leaves us little room to do mutations while keeping the type checker happy. For example, other mutation testing frameworks can mutate x + y to x - y.

This is an interesting mutation, because it’s so easy to do in languages like Java, which have full type information available at the bytecode level and so hard to do in Rust, because the std::ops traits make everything so hecking flexible.

So my first thought was “it’s generally impossible without trait lookup”. But then I remembered the trick I pulled in overflower, which is to shanghai the type system into doing the lookup for us via specialization: We can write our own ops trait that delegates to the core::ops trait by default while specializing for known types.

(Aside: Yes, specialization is unstable and the rules are going to be stricter than what is currently implemented in nightly, but our usage is in line with the rules as outlined by Niko Matsakis)

However, unlike overflower, we want to still be generic over our types: E.g. for replacing + with - (and vice versa) we need to be sure that the Rhs type parameters are equal and the associated Output types are compatible. So without further ado, here’s our trait (I’ll use a trait to mutate add to sub as an example. Mutating other ops is essentially the same):

pub trait AddSub<Rhs = Self> {
    type Output;
    fn add(self, rhs: Rhs, mutation_count: usize) -> Output;
}

A default impl that unconditionally adds looks like this:

impl<T, Rhs> AddSub<Rhs> for T
where T: Add<Rhs> {
    type Output = <T as Add<Rhs>>::Output;
    default fn add(self, rhs: Rhs, _mutation_count: usize) -> Output {
        self + rhs
    }
}

Now we would like to require the Output types of Add and Sub to be equal. This could be done with the following trickery:

/// this trait should only be implemented for pairs of the same type
trait Same;
impl<T> Same for (T, T) { }

impl<T, Rhs> AddSub<Rhs> for T
where T: Add<Rhs>, T: Sub<Rhs>,
    (<T as Sub<Rhs>>::Output, <T as Add<Rhs>>::Output) : Same {
    ..
}

This looks like it should work, and indeed the type system will accept it. However it turns out we cannot implement this method, because ironically, even knowing the types are one and the same gives us no way to convert between them.

Update: redditor kennytm informs me that Rust has a special syntax for requiring two associated types to be equal in where clauses, notably the surprisingly unsurprising where T1 = T2, and unlike the Same trait actually allows the type system to see through the equality relation. Thanks, kennytm! I’ll still use the following solution, because it’s more flexible.

So let’s take a step back and re-evaluate our requirements:

For our specialized impl, we need a way to turn the Output of the Sub operation into that of the Add operation. They don’t even need to be the same type, they just need to be convertible. Fortunately what Rust’s type system taketh away, it also giveth, in this instance with the Into trait, which allows us to abstract over convertible types:

impl<T, Rhs> AddSub<Rhs for T
where T: Add<Rhs>,
      T: Sub<Rhs>,
     <T as Sub<Rhs>>::Output: Into<T as Add<Rhs>>::Output> {
    fn add(self, rhs: Rhs, mutation_count: usize) -> Output {
        if mutagen::now(mutation_count) {
            (self - rhs).into()
        } else {
            self + rhs
        }
    }
}

That where clause is a mouthful, but reading it carefully makes the intent really clear: T implements addition and subtraction and the Output type of T’s subtraction is convertible into the Output type of its addition. So this specifies exactly what we need and it actually works. Neat!

There is but one niggle in that we don’t want to run the ineffective mutation via the default impl, but this can also easily be fixed by setting up yet another backchannel between the non-mutated test run and the test runner that the default impl can use. The test runner can then avoid the ineffective mutation.

The Clone Wat!?

The same thing goes for cloning a mutable ref (which is similar to an early return, but without actually skipping the whole method). As we don’t know if the type in question implements Clone, we cannot change the code and still compile. But what if we could?

There are two snags here:

  1. We need to decide if we can clone the value
  2. We need somewhere to own the cloned instance

The first one is pretty simple:

trait MayClone<T> {
    /// can we clone the current value?
    fn may_clone(&self) -> bool;
    /// only will be called on cloneables, so we can panic for non-cloneables
    fn clone(&self) -> Self;
}

if we specialize this trait for T: Clone, we can use it to prepend the following for a function with the mutable argument x:

let mut _x_clone;
if MayClone::may_clone(x) {
    let _x_clone = MayClone::clone(x);
    x = &mut _x_clone;
}

Now the may_clone method tells us if we can clone and our _x_clone binding holds the cloned value. Again, we should keep a flag to see if the mutation was effective.

Ineffective mutations vs. types

One small caveat about the “flag” for ineffective mutations: As functions may be generic over some type and tests may call a function with multiple types, thereby exercising multiple implementations, we need to keep track of the mutation for each test and only declare the mutation as ineffective unless it wasn’t effective for a single test.

Note that being ineffective once during a test is not enough, it must also never have been effective during the same test.


Are there other mutations that I have not thought of that we can implement with this technique? Tell me about it!