+## The reason why
+
+Why is it the case that a safe function can break `Vec`?
+How can we even say that it is the safe function which is at fault, rather than some piece of `unsafe` code elsewhere in `Vec`?
+
+The intuitive answer is that `Vec` has *additional invariants* on its fields.
+For example, `cap` must be greater than or equal to `len`.
+More precisely speaking, `ptr` points to an array of type `T` and size `cap`, of which the first `len` elements have already been initialized.
+The function `evil` above violates this invariant, while all the functions actually provided by `Vec` (including the ones that are implemented unsafely) preserve the invariant.
+That's why `evil` is the bad guy. (The name kind of gave it away, didn't it?)
+
+This may seem obvious in hindsight, but I think it is actually fairly subtle.
+There used to be claims on the interwebs that "if a Rust program crashes, the bug must be in some `unsafe` block". (And there probably still are.)
+Even academic researchers working on Rust got this wrong, arguing that in order to detect bugs in data structures like `Vec` it suffices to check functions involving unsafe code.
+That's why I think it's worth dedicating an entire blog post to this point.
+But we are not done yet, we can actually use this observation to learn more about types and type systems.
+
+## The semantic perspective
+
+There is another way to phrase the intuition of types having additional invariants:
+
+> *There is more to types like `Vec` than their definition gives away*.
+
+Imagine we define another type in our Rust program:
+{% highlight rust %}
+pub struct MyType<T> {
+ ptr: *mut T,
+ cap: usize,
+ len: usize,
+}
+{% endhighlight %}
+We will define only one function for this type:
+{% highlight rust %}
+impl MyType<T> {
+ fn evil(&mut self) {
+ self.len += 2;
+ }
+}
+{% endhighlight %}
+This type *looks* exactly like `Vec`, doesn't it?
+The two types are *syntactically* equal, and the same goes for the two `evil` functions.
+Still, `MyType::evil` is a perfeclty benign function (despite its name).
+How can this be?
+
+Remember that in a [previous blog post]({{ site.baseurl }}{% post_url 2015-10-12-formalizing-rust %}), I argued that types have a *semantic* aspect.
+For example, a function is semantically well-typed if it *behaves* properly on all valid arguments, independently of how, *syntactically*, the function body has been written down.
+
+The reason for `MyType::evil` being fine, while `Vec::evil` is bad, is that semantically speaking, `Vec` is very different from `MyType` -- even though they look so similar.
+If I have a `Vec`, I actually know that `ptr` is valid, that `len <= cap`, and so on.
+I know nothing like that about an arbitrary instance of `MyType`: All I have here is that `cap` and `len` are valid elements of `usize`.
+In other words, the additional invariants that we associate with `Vec` actually make this an *entirely different type*.
+In order to formally describe what it means to be a `Vec`, we can't just say "well, that's a `struct` with these three fields".
+That would be *wrong*.
+But still, that's actually all the Rust compiler knows, and it is all that it actually checks:
+When type checking `Vec::evil`, the compiler *thinks* that `Vec` is like `MyType`, and under that assumption, it deems `evil` a good function.
+(Clearly, the compiler did not watch enough movies with serious villains.)
+
+In other words, we have to look beyond the mere syntactic appearance of a type, and into its deeper, semantic meaning.
+As part of my work on formalizing Rust, I will eventually want to prove the soundness of types like `Vec` or `RefCell` and their associated operations.
+In order to do so, I will first have to figure out what exactly the semantics of these types are. (This will be fairly easy for `Vec`, but really hard for `RefCell`. Can you guess why?)
+Then I will have to check *every* function operating on these types and prove that they are actually well-typed.
+Even if the Rust compiler says "yup, that function over there is fine, I checked every single statement of it", that's not enough because the Rust compiler has no idea about which types these functions are *actually* about.
+All it sees is the syntactic surface.
+
+While proving soundness of `Vec`, it would then turn out that `Vec::evil` is actually not a semantically well-typed function. It may leave an invalid `Vec` in `self`.
+This is just as bad as a function assigning `*f = 42u32` to a `f: &mut f32`.
+The difference between `Vec` and `MyType` is no less significant than the difference between `f32` and `u32`.
+It's just that the compiler has been specifically taught about `f32`, while it doesn't know enough about the semantics of `Vec`.
+
+## The actual scope of unsafe
+
+At this point, you may be slightly worried about the safety of the Rust ecosystem.
+I just spent two sections arguing that the Rust compiler actually dosn't know what it is doing when it comes to checking functions that work on `Vec`.
+How does it come that people carelessly use `Vec` in their programs without causing havoc all the time?
+Or, to put it slightly differently: If the scope of `unsafe` grows beyond the syntactc `unsafe` blocks, then how far does it reach?
+Does it sprawl through all our code, silently infecting everything we write -- or is there some limit to its effect?
+
+As you probably imagined, of course there *is* a limit. Rust would not be a useful language otherwise.
+The scope of `unsafe` ends at the next *abstraction boundary*.
+This means that everything outside of the `std::vec` module does not have to worry about `Vec`.
+Due to the privacy rules enforced by the compiler, code outside of that module cannot access the private fields of `Vec`, and hence it cannot tell the difference between the syntactic appearance of `Vec` and its actual, semantic meaning.
+Of course, this also means that *everything* inside `std::vec` is potentially dangerous and needs to be proven to respect the semantics of `Vec`.
+
+## Abstraction Safety
+
+This nicely brings us to another important point, which I can only glimpse at here:
+
+> *The purpose of type systems goes far beyond type safety: They (also) serve to establish safe abstractions.*
+
+If the type system of Rust lacked a mechanism to establish abstraction (i.e., if there were no private fields), type safety would not be affected.
+However, it would be very dangerous to write a type like `Vec` that has a semantic meaning beyond its syntactic appearance.
+Since users of `Vec` can accidentally perform invalid operations, there would actually be *no bound to the scope of `unsafe`*.
+To formally establish safety, one would have to literally go over the entire program and prove that it doesn't misuse `Vec`.
+The safety promise of Rust would be pretty much useless.
+
+This should not be entirely surprising if you read the aforementioned [post about formalizing Rust's type system]({{ site.baseurl }}{% post_url 2015-10-12-formalizing-rust %}), where I already argued that a proof of (syntactic) type safety does not help to justify safety of most Rust programs out there.
+I am now making a similar point, coming from a different angle.
+
+The fact that Rust programmers *can* use `Vec` and many other types without much care is a property of the type system that is independent of type safety.
+We rely on the fact that the type system lets us build abstraction boundaries that cannot be violated by safe code.
+We may call this *abstraction safety*.
+Many type systems have this property, just think of `private` fields in Java or C++, or opaque signatures in ML.
+However, few languages embrace using unsafe primitives and encapsulating them behind a layer of abstraction as much as Rust does.
+In some sense, Rust relies way more on abstraction safety than most languages -- though in practice, programmers of all languages rely on abstraction safety a lot, if it is available.
+(Most scripting languages, I am looking at you.)
+
+Abstraction safety is a very deep point that deserves its own blog post, I only skimmed the surface.
+Watch this space, a link to such a post will eventually appear. (And I hope I don't have to write it myself.)
+
projects / web.git / blobdiff