“Bypassing” specialization in Rust

https://oakchris1955.eu/posts/bypassing_specialization/

by todsacerdoti • 1 year ago

46 21 1 year ago

oakchris1955.eu

In Haskell, we call this problem "overlapping instances," and can be allowed with an optional compiler flag. Nevertheless, I try to avoid using it, because of the potential for unforeseen consequences and reduced clarity in which code is actually being run.

In this use case, my instinct tells me that trait specialization is the wrong tool for the job. The author is trying to dispatch different functions based on a flag set in the caller. I can think of two more elegant ways to do this:

* Make the file system itself a trait, with implementations for read-only and read/write. Pass the concrete implementation as a parameter to the function using it.

* Store the capabilities as a variable in the file system object and query it at runtime. This can be done with an if.

Sometimes the simpler approach is better.

nesarkvechnep3• 1 year ago

Your proposal might work but then the user won’t have the compile-time guarantee that they can’t try to write to a read-only filesystem, since the capability will be queried at run-time.

dwattttt3• 1 year ago

I may be misunderstanding, but

Make the file system itself a trait, with implementations for read-only and read/write

Those are separate concrete implementations between read-only and read-write; a write operation can be implemented on the read-write implementation only.

konstantinua002• 1 year ago

can't it be solved by negative traits?

isn't the problem that rw is still r, so passes checks for both?

can't you make one rw and the other r(-w)?

SkiFire132• 1 year ago

Note that negative traits are not for "this trait is not implemented" (i.e. a missing `impl Trait for Type`) but instead for "this trait is guaranteed to not be implemented" (i.e. `impl !Trait for Type`)

konstantinua002• 1 year ago

aka that's what read-only is for write?

maybevoid2• 1 year ago

With Context-Generic Programming (CGP), there is a way to get around the coherence restriction through a provider trait, and have overlapping implementations in safe, stable Rust. I have a proof of concept implementation of example in the article here: https://gist.github.com/soareschen/d37d74b26ecd0709517a80a3f...

maybevoid2• 1 year ago

For those who are new to the concept, in a nutshell, CGP allows you to bypass the coherence restrictions in traits/typeclasses, and define multiple overlapping and generic implementations for each CGP trait. As a consequence, when you define a new type, you need to specifically choose an implementation to be used with that type. This is called a wiring step, where you choose the provider implementations for your context.

On the surface, this addition doesn't seem significant. However, it opens up a world of possibilities for enabling powerful design patterns that are based on trait implementations that are named, generic, and overlappable. One of the greatest strengths of CGP is to enable safe, zero-cost, compile-time dispatching through the trait system to accomplish the things that in OOP would typically require dynamic dispatch with runtime errors.

ta8645• 1 year ago

For anyone interested in reading about CGP in Rust, here is the website:

https://contextgeneric.dev/

Just a FYI from their main page:

As of 2025, CGP remains in its early stages of development. While promising, it still has several rough edges, particularly in areas such as documentation, tooling, debugging techniques, community support, and ecosystem maturity.

...

At this stage, CGP is best suited for early adopters and potential contributors who are willing to experiment and help shape its future.

quotemstr• 1 year ago

Clever work. I'm glad to see Rust is at the AbstractSingletonProxyFactoryBean stage of its lifecycle.

Look: patterns like this might unblock people locally but they point to a gap between what people want to say and what the base language allows them to express. In the long term, it's better to expand the language than devise increasingly heroic workarounds for syntactic sclerosis.

SkiFire132• 1 year ago

Lastly, the unstable specialization feature also deals with lifetimes, which as I mentioned is the feature's biggest obstacle on the way to stabilization.

The issue is really that `specialization` *has* to deal with lifetimes, even when there is apparently none. The example core with `Read` and `Read + Seek` also suffers from lifetime issues: some type could implement `Seek` depending on some lifetime and this would make your specialization unsound. For this reason `min_specialization` does not accept your specialization.

ozgrakkurt• 1 year ago

Isn’t it possible to just use trait objects for this?

You can have default implementation for some methods and implementors of the trait can specialise.

Specifically, you can have couple functions that takes a type that implements some stuff (different based on the wanted specialisation) and gives you a trait object.

This should be simpler and should have better compile time than doing macros or adding more generics

selfmodruntime2• 1 year ago

The length Rust programmers will go to avoid writing `Box<dyn Trait>` is astonishing. The heap is there for a reason :)

alfiedotwtf• 1 year ago

I’m guilty of this… I’ve doing using Rust for 10 years, and I’ve only ever used `Box<dyn>` maybe 4 times all up.

.. yep, I’ll try everything before introducing virtual functions just out of warm fuzzies rather than anything measurable

ccleve• 1 year ago

The performance hit you take for doing dynamic dispatch is real and measurable. It's a no-go if you're in a performance sensitive part of your app.

“Bypassing” specialization in Rust

Related Stories