Bonus: recent talk from Ralf Jung on his group's efforts to precisely specify Rust's operational semantics in executable form in a dialect of Rust: https://youtube.com/watch?v=yoeuW_dSe0o
> The problem with unsafe code is that it can do things like this:
fn main() {
let mut x = 42;
let ptr = &mut x as *mut i32;
let val = unsafe { write_both(&mut *ptr, &mut *ptr) };
println!("{val}");
}
"Unsafe code allows to express the following, which is UB:"
- Anything accepted by the borrow checker is legal
- Unsafe can express illegal / undefined behavior
- There's some set of rules, broader than what the borrow checker can check, that is still legal / defined behavior
The goal of this line of work is to precisely specify that set of rules. The outlines are clear (basically, no writable pointers should alias) but the details (interior pointers, invalidation of iterators, is it creating or using bad pointers that's bad, etc) are really hard. The previous paper in this series, on Stacked Borrows, was simpler but more restrictive, and real-world unsafe code often failed its rules (while still seeming correct). Tree Borrows is broader and allows more while still being provably safe.
Note that we have not yet proven this. :) I hope to one day prove that every program accepted by the borrow checker is compatible with TB, but right now, that is only a (very well-tested) conjecture.
> Given that aliasing optimizations are something that the Rust compiler developers clearly want to support, we need some way of “ruling out” counterexamples like the one above from consideration.
Yes, but which exact rule does it violate? What is the exact definition that says that it is UB? Tree Borrows is a proposal for exactly such a definition.
"code can do things like this" here means "you can write this code and compile it and run it and it will do something, and unless we have something like Tree Borrows we have no argument for why there would be anything wrong with this code".
You seem to have already accepted that we need something like Tree Borrows (i.e., we should say code like this is UB). This part of the paper is arguing why we need something like Tree Borrows. :)
You misunderstand the word "can". Yes, you can, in unsafe code, do that. And yes, that is undefined behaviour ;)
https://play.rust-lang.org/?version=stable&mode=debug&editio...
And one could say that they borrow from the tree some of their qualities. Sorry, couldn't resist.
That's before he went into politics, though.
All have different costs and capabilities across implementation, performance and developer experience.
Then we have what everyone else besides Rust is actually going for, the productivity of automatic resource management (regardless of how), coupled with one of the type systems above, only for performance critical code paths.
Doesn't matter if it's purely "syntaxical" because the language is garbage collected, just the fact of specifying what owns what and be explicit about multiple references is great imo.
Some sort of effects systems can already be simulated with Kotlin features too.
Programming language theory is so interesting!
I'd just like to interject for a moment. What you’re referring to as "affine types", is in fact, Uniqueness Types. The difference has to do with how they interact with unrestricted types. In Rust, these "unrestricted types" are references (which can be used multiple times due to implementing Copy).
Uniqueness types allow functions to place a constraint on the caller ("this argument cannot be aliased when you pass it to me"), but places no restriction on the callee. This is useful for Rust, because (among other reasons) if a value is not aliased you can free it and be sure that you're not leaving behind references to freed data.
Affine types are the opposite - they allow the caller to place a restriction on the callee ("I'm passing you this value, but you may use it at most once"), which is not something possible to express in Rust's type system, because the callee is always free to create a reference from its argument and pass that reference to multiple functions..
This is often explained via the "do not use more than once rule", but that's not the actual definition, and as your example shows, following that simplified explanation to the letter can cause confusion.
> because the callee is always free to create a reference from its argument and pass that reference to multiple functions..
Passing a reference is not the same thing as passing the actual value, so this does not contradict affinity.
I agree that passing a reference is not the same thing as passing the actual value. If it were, there would really be no point to references. However, it does contradict affinity. Specifically, the fact that multiple references can be created from the same value, combined with the properties of references, contradicts affinity.
> At its core, "affine" means that the type system has exchange and weakening but not contraction, and that exactly characterizes Rust's type system.
Well, the rust type system certainly does support contraction, as I can use a reference multiple times. So what is that if not contraction? It seems like rust at least does support contraction for references.
But in practice, having absolutely no contraction is not a very useful definition of affine, because no practical programming language would ever satisfy it. It prohibits too much and the language would not even be turing complete. Instead, there is usually an "affine world" and an "exponential world". (Exponential meaning "unrestricted" values that you can do whatever you want with). And the convention is that values can go from the exponential world to the affine world, but not back. So a function taking an affine value can be passed any value, but must use in in an affine way, and meanwhile but a function taking an exponential (unrestricted) value can only be passed exponential and not an affine value.
If you don't believe me, you can try using linear haskell, and notice that a function taking a linear argument can be passed a non-linear argument, but not the other way around.
If you interpret Rust's type system this way, it's natural to interpret references as exponentials. But references have the opposite convention. You can go from owned values to references, but not the other way around, which is precisely the opposite situation as the convention around linear/affine type systems. Because these systems feel very different to use and enforce very different properties, I do think it's important that we have separate names for them rather than referring to both as "affine". And the usual name for the rust-like system is "uniqueness types", see https://docs.idris-lang.org/en/latest/reference/uniqueness-t... or https://en.wikipedia.org/wiki/Uniqueness_type .
Good question! For shared references, the answer is that they are `Copy`, so they indeed have contraction. Affinity just means that contraction is not a universal property, but some types/propositions may still have contraction. For mutable references, you can't actually use them multiple times. However, there is a desugaring phase going on before affinity gets checked, so uses of mutable references `r` get replaced by `&mut *r` everywhere. That's not using contraction, it's not literally passing `r` somewhere, it is calling a particular (and interesting) operating on `r` ("reborrowing").
Rust is not just an affine system, it is an affine system extended with borrowing. But I think it is still entirely fair to call it an affine system, for the simple fact that the language will prevent you from "using" a variable twice. "reborrowing" is just not a case of "using", it is its own special case with its own rules.
> But in practice, having absolutely no contraction is not a very useful definition of affine,
Obviously Rust has a class of "duplicable" types, called `Copy`. That's besides the point though.
> If you interpret Rust's type system this way, it's natural to interpret references as exponentials.
Why would that be natural? Mutable references are not even duplicable, so what you say makes little sense for references in general. Maybe you mean shared references -- those are just an example of a duplicable type.
Rust doesn't have a modality in its type system that would make every type duplicable, so there is no equivalent to exponentials. (In particular, `&T` isn't a modality around `T`. It's a different type, with a different representation. And as you noted, even if it were a modality, it wouldn't correspond to exponentials.)
But a type system can be affine/linear without having exponentials so I don't understand the point of this remark.
Uniqueness types seem to be all about how many references there are to a value. You can use linear/affine types to enforce such a uniqueness property (and that is indeed what Rust does), but that doesn't take away from the fact that you have a linear/affine type system.
> Because these systems feel very different to use and enforce very different properties,
I can't talk about the "feel" as I never programmed in an affine language (other than Rust ;), but in terms of the properties, what Rust does is extremely closely related to affine logics: the core property being enforced is that things do not get duplicated. My model of Rust, RustBelt, uses an affine separation logic to encode the properties of the Rust type system, and there's a lot of overlap between separation logic and linear logic. So we have further strong evidence here that it makes perfect sense to call Rust an affine language.
Also, the claims about Curry-Howard correspondence are wrong. It does not prove that rust is an affine language: https://liamoc.net/forest/loc-000S/index.xml
And some people like to claim that the Curry-Howard correspondence proves something about their type system, but this is only true for dependently typed languages.
And the proofs aren't about program behavior.
Maybe a dumb question but couldn't you just run multiple implementations in parallel threads and whichever finishes first with a positive result wins?
It has migrated from a scope-based borrow checker to non-lexical borrow checker, and has next experimental Polonius implementation as an option. However, once the new implementation becomes production-ready, the old one gets discarded, because there's no reason to choose it. Borrow checking is fast, and the newer ones accept strictly more (correct) programs.
You also have Rc and RefCell types which give you greater flexibility at cost of some runtime checks.
"Rust", in this context, is "merely" "the usual invariants that people want" and "a suite of optimizations that assume those usual invariants, but not more or less".
[1] https://github.com/Voultapher/sort-research-rs/blob/main/wri... Miri column
[2] https://github.com/rust-lang/rust/blob/6b3ae3f6e45a33c2d95fa...
And it seams to not be the case on the stable compiler version?
fn write(x: &mut i32) {*x = 10}
fn main() {
let x = &mut 0;
let y = x as *mut i32;
//write(x); // this should use the mention implicit twophase borrow
*x = 10; // this should not and therefore be rejected by the compiler
unsafe {*y = 15 };
}rustc itself has no reason to reject either version, because y is a *mut and thus has no borrow/lifetime relation to the &mut that x is, from a compile-time/typesystem perspective.
How true is this really?
Torvalds has argued for a long time that strict aliasing rules in C are more trouble than they're worth, I find his arguments compelling. Here's one of many examples: https://lore.kernel.org/all/CAHk-=wgq1DvgNVoodk7JKc6BuU1m9Un... (the entire thread worth reading if you find this sort of thing interesting)
Is Rust somehow fundamentally different? Based on limited experience, it seems not (at least, when unsafe is involved...).
But in the end, it's a trade-off, like everything in language design. (In life, really. ;) We think that in Rust we may have found a new sweet spot for this kind of optimizations. Time will tell whether we are right.
When you're working with anything below the application level, C's confusing and underspecified rules about UB are almost impossible to keep track of, especially when it comes to aliasing and volatile/MMIO. The spec is so difficult to read and full of complicated cross-references that to actually get a practical answer you have to look for a random Stack Overflow post that may or may not have a correct interpretation of the spec, and may or may not address your specific problem.
Rust right now feels a lot harder to work with, because the spec isn't done. When you have a concrete question about a piece of code, like "is this conversion from an &mut to a *mut and back sound", and you try to look for documentation on it, you get either "Nobody knows, Rust aliasing model isn't defined"; a hand-wavy explanation that is not rigorous or specific; or a model like Stack Borrows or Tree Borrows that's defined a little too formally for easy digestion :)
But when I really started digging, I realized just how much cleaner Rust's semantics are. References aren't actually hard, Tree Borrows basically boils down to "while an &mut reference is live, you can only access the value through pointers or references derived from that reference". Pointer operations have straightforward semantics, there's no confusing notions of typed memory, and no UB "just because" for random things like integer overflow. It's just so much less complicated to understand than C's abstract machine.
I'm really looking forward to things like MiniRust, and to an aliasing model making it into the Reference / other documentation, because at that point I feel like unsafe Rust will be way* easier to write confidently and correctly than C.
Congrats on the publication, and thanks again for the work you all have put into this.
In the safe subset of Rust it's guaranteed in all cases. Even across libraries. Even in multi-threaded code.
fn foo() {
let mut x = 42;
let mut mutable_references = Vec::new();
let test: bool = rand::random();
if test {
mutable_references.push(&mut x);
} else {
mutable_references.push(&mut x);
}
}
fn bar() {
let mut x = 42;
let mut mutable_references = Vec::new();
let test: bool = rand::random();
if test {
mutable_references.push(&mut x);
}
if !test {
mutable_references.push(&mut x); // error: cannot borrow `x` as mutable more than once at a time
}
}
C's aliasing is based on type alone, hence its other name "type based alias analysis" or TBAA.
In C you have a nuclear `restrict` that in my experience does anything only when applied to function arguments across clang & gcc, and type-based aliasing which is both not a generally-usable thing (don't have infinite different copies of the int64_t type (and probably wouldn't want such either)), and annoying (forces using memcpy if you want to reinterpret to a different type).
Whereas with Rust references you have finely-bounded lifetimes and spans and mutability, and it doesn't actually care about the "physical" types, so it is possible to reinterpret memory as both `&mut i32`/`&i32` and `&mut i64`/`&i64` and switch between the two for the same memory, writing/reading halves of/multiple values by the most bog standard Rust safe reads & writes, as long as the unsafe abstraction never gives you overlapping `&mut` references at the same time, or split a `&mut` into multiple non-overlapping `&mut`s.
But like Linus, I've noticed it doesn't seem to make much difference outside of obvious narrow cases.
I’d be interested to see a more thorough analysis, but there is a simple way to gauge this - rip out all the parts of the compiler where aliasing information is propagated to LLVM, and see what happens to performance.
I found a claim that noalias contributes about 5% performance improvement in terms of runtimes[0], though the data is obviously very old.
https://github.com/rust-lang/rust/issues/54878#issuecomment-...
Alias analysis is extremely important for getting good performance these days--but it should also be remembered that the biggest benefits accrue from the simplest heuristics (like two loads that use the same SSA value as the pointer must alias each other). In LLVM terms, that's BasicAA: a collection of very simple heuristics that primarily amounts to "if we can track down the allocation sites of objects, we can definitively resolve these alias queries; otherwise, we don't know."
The real question that you're trying to ask, though, is what is the value of alias analyses that go beyond the most basic, obvious tests. At the point where the alias queries are no longer trivial to solve, then it's generally the case that what you can do as a result of those queries also shrinks dramatically, pretty much to looking for code motion hazards, and the benefits you get from that are much reduced. One of the experiments I would like to do is measure the total speedup you'd get from a theoretically perfect alias analysis, and my guess is that it's somewhere in the 20% range even on non-HPC code like the Linux kernel [1].
[1] This doesn't account for the heroic optimizations, such as data-layout transformations, that you wouldn't attempt to write without a very high-quality alias analysis. But since we already know that alias analysis doesn't exist in practice, we're not going to attempt those optimizations anyways, so it's not worth including such stuff in prospective speed gains.
Both clang/llvm and gcc can do alias checking at runtime if they can't at compile-time, which makes loops vectorizable without alias info, at the cost of a bit of constant overhead for checking aliasing. (there's the exception of gather loads though, where compile-time aliasing info is basically required)
And on the other hand there's good potential for benefit to normal code (esp. code with layers of abstractions) - if you have a `&i32`, or any other immutable reference, it's pretty useful for compiler to be able to deduplicate/CSE loads/computations from it from across the whole function regardless of what intermediate writes to potentially-other things there are.
"I want my fuckin' money back."
"Hoom, hmm, let us not be hasty!"
"You got 48 hours to deliver or the sapling gets it, Treebeard."
pvg•7h ago
https://news.ycombinator.com/item?id=22281205
https://news.ycombinator.com/item?id=17715399