Well, I have not done too well with my goal of reading a research paper a day on the train (actually my initial goal was two papers, but seeing as how I’ve failed so spectacularly, I’ve dialed it back some). However, I’ve decided to give it another go. I’ve bought a printer now so I can print papers out (double-sided, no less!) at home (I had initially planned to buy an iPad or something, but a decent printer is only $100, and paper is still nicer to read and write notes on…you do the math). As additional motivation, I’m working again on the paper on Rust’s new borrowed pointers and so I have to catch up on a lot of related work.

To that end, I downloaded and read A Type System for Borrowing Permissions, published at POPL 2012 by Naden, Bocchino, Aldrich, and Bierhoff. It’s a very good read: pretty easy to follow and certainly directly related to a lot of what we’ve been doing in Rust. Overall, I found it to be both more and less expressive than Rust’s system. It’s hard to compare the complexity of the two: each is complex in its own way.

A brief summary of their work

The basic idea of their system is to label object types with permissions. There are several such permissions:

  • unique — the only pointer to an object
  • immutable — the object can be shared but not modified
  • shared — the object can be freely shared and modified
  • local immutable — the object can be aliased from local variables but not mutated
  • local shared — the object can be aliased from local variables and modified

Each variable is annotated with such a permission. Parameters are annotated not only with a permission but with a “change” permission. For example, something like Rust’s send() function, which consumes its parameter, would be written:

void send<T>(unique >> none T send) { ... }

This indicates that the value must begin as unique but, after the function call, will have the permission “none”. You can also annotate the receiver in the typical, C++-inspired style:

void foo() unique { /* Can only be called if the receiver is unique */ }

Unique objects can be temporarily aliased using the “local” permissions, such as local shared or local immutable. Such objects cannot be used outside of local variables.

It is also possible to temporary alias fields of unique type, so long as they are stored in unique objects. This causes the object to be only partially valid: an aliased unique field is no longer unique, after all. An object in this state is called “unpacked”: it must be packed again before it can be passed around. Unpacking like this is only permitted when the owner of the field is unique.

Comparisons to Rust

This is quite close to Rust’s lifetime and borrowing system in many ways. The notion of packing and unpacking operates much like borrow check’s loans, for example. There are of course some differences.

How it is less expressive than Rust?

Storing borrowing items into data structures

There are two areas where our system is more expressive. The first is that, in their system, borrowed, aliasable things (“local permissions”) are confined to local variables; our borrowed pointers suffer from no such limitations. In fact, lifting this restriction was probably the major goal of the shift to using borrowed pointers in place of reference modes.

An example of where this limitation is problematic is when you have some largish function that needs some context to perform its computation. For example, in our compiler translation pass, you get as input a “type context” (&TypeContext) that contains information about the types of all expressions. Then we create a crate context (&CrateContext) that is local to the compilation pass. Moreover we make a function (&FnContext) for each function and a block (&BlockContext) for each basic block. Each of these embeds a pointer to the context up the chain, so from the block context you can reach the function context, then the crate context, and finally the type context. A system confined to local variables cannot express this sort of pattern without passing each of the various contexts as a parameter. You wind up with a lot of functions all taking some big set of parameters initially—the usual solution is to refactor by making a single object to pass around, but you won’t be able to do that if you are using local aliases.

Regions allow us to move past this problem. We can permit arbitrary aliasing, secure in the knowledge that the type system will prevent these aliases from escaping out into the wild. In other words, we know that even if you package up all those redundant arguments into a struct, that struct itself will never outlive the function where it is created. This sort of reasoning is absolutely crucial to effectively support stack allocation, which is a primary goal of Rust (but which I doubt is a goal for Plaid; nonetheless I am sure they will encounter this problem from time to time).

Inherited mutability

The second is that our notion of inherited mutability lets us freeze an entire family of owned objects at once. I intend to write another tutorial-style blog post on this, so for now I’ll just say that if you own a freezable data structure—and all the various container types in Rust will eventually be freezable, I think—you can convert it between mutable and immutable and back again. They have a similar notion of converting unique to immutable, but as far as I can tell this is a shall immutability that does not itself extend to the unique objects within. It’s possible that I am misreading the type rules on this point though. Also, I think Rust’s notion of inherited mutability could be applied to their system without great difficulty.

How is it more expressive than Rust?

The main thing which they can do that we cannot is express the idea of borrowing a unique pointer but retaining the knowledge that the pointer is unique. For us, borrowing is intrinsically tied to aliasing—for one thing, borrowing a unique value does not make the original inaccessible, it just lessens the things you can do with it. For another, we always allow &T types to be copied.

Basically, in our system, a ~T type is always owned. So if you write:

fn foo(bar: ~Map<K,V>) {
    bar.insert(k, v);
    for bar.each |k, v| {
        ...                 // (bar is immutable here)

This is a function which takes in a map, modifies it some, iterates over it, and then (implicitly) frees it. The same function could, if it wanted to, continue by sending bar to another task, or doing whatever. But, of course, it doesn’t here, and perhaps what I really wanted was just to do some work on behalf of the caller foo().

Typically, the way that one expresses the idea that you want to take a Map but not consume it is to use a borrowed pointer:

fn foo(bar: &Map<K,V>) {
    bar.insert(k, v);            // (Error)
    for bar.each |k, v| {
        ...                 // (bar is immutable here)

Unfortunately, as indicated in the comments, this code will not compile. This is because when you borrow a freezable data structure like a map, you must select if it is immutable (&Map<K,V>) or mutable (&mut Map<K,V>). But this example uses insert(), which requires mutability, and iteration, which requires immutability (modifying the map while iterating is a no-no).

There are ways to solve this problem. One is that, in the caller, you could use a ~Mut<Map<K,V>>: the Mut type basically moves the safety checks from compile-time to runtime. This means that if you try to insert into the map while iterating your task will fail (just as it would in Java, for example, or most other languages). However, using the Mut type adds some (slight) overhead and gives you fewer static guarantees, so it is not ideal. It’s really intended for use with managed data (e.g., @Mut<Map<K,V>>), where there is no hope of tracking the mutability statically since the map can be aliased.

Another option is to move the map into the function foo() and then return it back to the caller when foo() has finished:

fn foo(mut bar: ~Map<K,V>) -> ~Map<K,V> {
    bar.insert(k, v);
    for bar.each |k, v| {
        ...                 // (bar is immutable here)
    return map;

This is a bit annoying, but it works. It’s similar to other affine systems like Alms.

Under the system proposed in this paper, however, a parameter labeled unique Map<K,V> would not be consumed by the callee. Basically it must be unique on entry to the callee and on exit. If the callee wished to consume the map, such as by sending it to another task, it would write unique >> none Map<K,V>.