What? You can easily escape from it if there are better alternatives you can use.

So there is no general escape hatch.

Pointing at one language and saying it is not easy to code like it is another language is a pointless argument.

I’m not arguing that it is easier to code in C# than in Rust, just that this particular escape hatch is possible in C# and not in Rust. It’s just an observation.

They all differ for good reasons and as long as you can solve similar problems in both, even if in different ways then what does it matter that you cannot do it in the same way?

It does not really matter, but does it have to?

This is a kind of stupid example, but imagine you have to implement some external trait (e.g. in order to work with some dependency) with the following shape:

trait Foo {
    fn foo(&self, i: usize) -> &Bar;
}

Which is not too unreasonable, for example it’s very similar to the stdlib’s Index. In order to implement this trait you must return a reference, you can’t return e.g. a Cow or an Arc. The fact that it takes a parameter means there might not even be one single value it has to return, so you can’t even cache that inside of self with e.g. LazyLock.

Of course I’m not saying I would try to reach for an escape hatch if I had to do something like this. I would first try to see if this is an intrinsic problem in my code, or if maybe I can change the crate I’m working with to be more permissible. That is, I would try to look for proper solutions first, though Cow might not always work for that.

You dont write code like this in rust.

I perfectly agree, that would be horrible code! I would generally try to restructure my code, making it better fit the actual lifetimes of the data I’m working with. The point in the article is that you can’t really escape from this. I’m not arguing this is a real problem, and I don’t think the article is neither, just pointing out that this is something you can easily do in C# and not in Rust. It’s just a difference between the two languages.

I can agree that the example function is not the best usecase. But the point still stand that there’s no realistic escape hatch from lifetimes and memory management in Rust.

Cow does not work when you are actually required to return a reference, e.g. if you’re working with some other crate that requires that. Cow also has some more strict requirements on reborrows (i.e. you can reborrow a &'short &'long T to a &'long T, but you can only reborrow a &'short Cow<'long, T> to a &'short T).

LazyLock can solve very specific issues like static, but is not a general escape hatch. Again, the example is not the best to showcase this, but imagine if you have to perform this operation for an unknown amount of runtime values. LazyLock will only work for the very first one.

It’s a very different kind of borrow checking than Rust’s. For example there’s no mutability xor sharing checking, because mutability cannot invalidate memory (there’s still a GC in C# after all!). As such, Rust’s NLL (which are available in the 2015 edition too btw) don’t really make sense in C#.

(Note that I’m not the article author)

In this example, you could have just made a constant with value 0 and returned a reference to that. It would also have a 'static lifetime and there would be no leaking.

I believe the intention was to demonstrate something that works with runtime values too, and a constant 0 does not.

Btw you can just write &0 to do what you proposed, there’s no need for an explicit constant/static.

No, macros can see only the tokens you give them. They have no notion of the fact that crate::functions is a module, that PluginFunction is a trait and that functions_map is a variable. Not even the compiler may know those informations when the macro is expanded.

If you really really want to do this, you can use something like inventory. Note that inventory uses a special linker section to do this, which some consider a hack. This is also not supported on WASM if you want to target it.

Would it makes sense to provide a “share nothing” runtime implementation that can be injected at startup?

Isn’t this tokio::task::spawn_local?

compile rust 0.7 in scala

Not sure if there was another rewrite, but AFAIK (and the article agrees with me) rustc was originally written in Ocaml

Have you actually measured a performance impact from RefCell checks?

High or low level doesn’t matter. Mathematically it just makes more sense to use 0-based indexing https://www.cs.utexas.edu/users/EWD/transcriptions/EWD08xx/EWD831.html

It’s mentioned in footnote 6:

As an example, to make this work I’m assuming some kind of “true deref” trait that indicates that Deref yields a reference that remains valid even as the value being deref’d moves from place to place. We need a trait much like this for other reasons too.

It would only work for references that are stable after the value they reference is moved. Think for example of a &str you get from a String.

They tested the same strings on that implementation

The code they were looking at was used for writing the table, but they were testing the one that read it (which is instead correct).

though judging by the recent comments someone’s found something.

Yeah that’s me :)The translation using an associated const also works when the const block uses generic parameters. For example:

fn require_zst<T>() {
    const { assert!(std::mem::size_of::<T>() == 0) }
}

This can be written as:

fn require_zst<T>() {
    struct Foo<T>(PhantomData<T>);
    impl<T> Foo<T> {
        const FOO: () = assert!(std::mem::size_of::<T>() == 0);
    }
    Foo::<T>::FOO
}

However it cannot be written as:

fn require_zst<T>() {
    const FOO: () = assert!(std::mem::size_of::<T>() == 0);
    FOO
}

Because const FOO: () is an item, thus it is only lexically scoped (i.e. visible) inside require_zst, but does not inherit its generics (thus it cannot use T).

They tested the same strings on that implementation

The strings were the same, but not the implementation. They were testing the decoding of the strings, but the C function they were looking at was the one for encoding them. The decoding function was correct but what it read didn’t match the encoding one.

though judging by the recent comments someone’s found something.

Yeah, that’s me :)

while a similar C implementation does not need this fix

No, that implementation also needs the fix. It’s just that it was never properly tested, so they thought it was working correctly.

It’s also extremely unlikely that you’d be running a bat script with untrusted arguments on Windows.

It happens in yt-dl, which is where this was first reported https://github.com/yt-dlp/yt-dlp/security/advisories/GHSA-hjq6-52gw-2g7p

Very likely yes

Loop unrolling is not really the speedup, autovectorization is. Loop unrolling does often help with autovectorization, but is not enough, especially with floating point numbers. In fact the accumulation operation you’re doing needs to be associative, and floating point numbers addition is not associative (i.e. (x + y) + z is not always equal to (x + (y + z)). Hence autovectorizing the code would change the semantics and the compiler is not allowed to do that.

If your goal is to just .await some future that require the Tokio runtime you can use the async-compat crate.

Another option is to compile dependencies with LLVM and optimizations, which will likely be done only once in the first clean build, and then compile your main binary with Cranelift, thus getting the juicy fast compile times without having to worry about the slow dependencies.