We have not done any scheduling for what comes after 1.0. I've been doing a huge amount of triage to move over issues that used to be in rust-lang/rust to be under that label, too, and to make it more detailed than just 'wishlist.' We'll figure out exactly what's next shortly before the first post-1.0 cycle actually begins.
Since last July or so, Servo has been tracking all of the major features we'd like to see that have been postponed to post-1.0:
https://github.com/servo/servo/issues/2854
A feature that has already seen many proposed rfcs and long discussions is "Efficient code reuse" (a.k.a some kind of inheritance), summarized here: https://github.com/rust-lang/rfcs/issues/349 It was explicitly postponed until after 1.0.
I just wrote a comment on that issue with some half-baked ideas, but I really think that this is one of those "line in the sand" features that will determine (at least for me) whether rust is really staying true to its emerging identity or whether it's on the road to becoming another opinionated kitchen sink language.
The thing about "efficient code reuse" is it probably requires dynamic dispatch. Once you have dynamic dispatch, you suddenly have vtables. But who decides what those vtables look like? Where do they reside in memory? What's the layout of that? If a struct suddenly has an is-a pointer, where is that mentioned in the code? Now my struct isn't just a struct.
I love the rust idea that tons of modern language design still allows for zero-cost abstraction. Inheritance in dispatch starts getting into the land of "putting a lot more stuff in my binary than I asked you to", and I would argue that it's this property more than anything else that keeps embedded programmers and kernel guys safely in the minimalistic land of C.
It would be nice to have a new C, finally. But the more a language has an opinion on runtime layout, behavior, and symbol names, the less C-like it becomes.
Rust got this right when it decided that GC was NOT the correct default behavior for a language. The reason you see people playing with OS kernels in rust and not as much in D is, I think, mainly due to this decision. I think rust should continue carrying this torch.
If not for the ability to have the best of both worlds (a modern language and access to to-the-metal programming with a controllable runtime layout and deterministic performance profile), where exactly is the value in learning how to use the borrow checker?
> The thing about "efficient code reuse" is it probably requires dynamic dispatch. Once you have dynamic dispatch, you suddenly have vtables. But who decides what those vtables look like? Where do they reside in memory? What's the layout of that? If a struct suddenly has an is-a pointer, where is that mentioned in the code? Now my struct isn't just a struct.
1. We already have vtables through trait objects (though not for structs), so this would be nothing new. It's important that we have them, because otherwise common dynamic dispatch would be very annoying to write.
2. Structure layout is already not defined. The compiler is permitted to reorder structure fields as it likes. However, you can force it to adopt your specified in-memory order with the `#[repr(C)]` annotation.
> It would be nice to have a new C, finally. But the more a language has an opinion on runtime layout, behavior, and symbol names, the less C-like it becomes.
The language already has an opinion on runtime layout and symbol names. However, you can specify the layout and symbol names manually if you like (through `#[repr(C)]` in the former case and `#[no_mangle]` in the latter case).
> Rust got this right when it decided that GC was NOT the correct default behavior for a language. The reason you see people playing with OS kernels in rust and not as much in D is, I think, mainly due to this decision. I think rust should continue carrying this torch.
GC has performance costs, while trait objects and symbol names do not, as long as they're opt-in. Garbage collection and virtual dispatch are completely different things; having one in no way moves us closer to the other.
GC has performance costs, while trait objects and symbol names do not, as long as they're opt-in.
Not exactly.
Any program even in a GC'd language can allocate non-GCd data. Even in Java, there is the Unsafe class that can do manual mallocs/frees. C# integrates it with the language. If you do a big pile of work on the non-GC heap then no GC would be triggered and GC is effectively "zero cost" for this code.
And vtable dispatch has a cost for any code that calls a virtual method. Yes, it's "opt in" but this can be misleading. You pay the cost of the virtual method call every time it's invoked. Static languages like Rust and C++ require the programmer to explicitly state which methods can be virtual to try and control this cost, but that isn't the only way.
For example the JVM is capable of eliminating the virtual method dispatch overhead in almost all cases without requiring the programmer to manually specify which methods are virtual. In fact, the JVM can eliminate the vtable overhead for calls that could be virtual, but in fact at that specific call site are not, and even for calls which are only slightly virtual (e.g. there's only really two destinations). So it's possible that in a JIT compiled program you have way more virtual dispatch in theory, but less than the equivalent C++ program would in practice once the code is warmed up.
So vtable and GC performance are very complex topics which no longer reduce neatly down to our intuitions.
>And vtable dispatch has a cost for any code that calls a virtual method. Yes, it's "opt in" but this can be misleading. You pay the cost of the virtual method call every time it's invoked.
So you're basically repeating what he said -- I don't see how the "Not really" you begin with is justified.
Of course you "pay the cost of the virtual method call every time it's invoked".
In languages like C++ or Rust you pay the cost every time the method is invoked. With other types of compiler you may not pay the cost even if the method is marked as virtual, even if it's actually used virtually in other parts of the codebase, due to call site specialisation.
> Garbage collection and virtual dispatch are completely different things; having one in no way moves us closer to the other.
Right, of course not. The comparison was philosophical rather than technical.
My point was that I believe there's a "sweet spot" for a language that is expressive and convenient and modern, but also tries hard not to stray too far from C's spartan abstract machine model (and when it does, it exposes that complexity in a composed pluggable fashion).
I'm beginning to believe rust really has a shot at replacing C and needs to court "bare metal" programmers as well as higher-level programmers to do it; I'm just preemptively registering my wish that rust continue to head down that path.
Rust will not replace c, rust is a more refined c++. If you want something to replace c have: a better type system, raii, better macros, better lifetime management, etc but keep it simple and don't layer it.
The best way to make that wish come true is to use Rust in a project where you would normally use C, and report your experience to help us discover the best ways to support that use case. :)
I think what personally got me excited about this direction were the many "OS kernel in rust" hobby projects [1, 2]. For some reason these strike me as a sort of reverse canary-in-a-coal-mine to judge whether a language is seen as a potential C replacement.
Prior to rust the hobby OS dev community was primarily C / asm with some honorable mentions for other languages. It's also something of a stand-in for the requirements of the professional embedded community.
For these use cases, it's just really cool to be able to use more and more "layers" of the language as you implement more of the underlying abstract machine model.
Rust already supports dynamic dispatch[1], and the 'code re-use' proposals may not look like traditional inheritance. [2] Furthermore, they're being driven _by_ those low-level requirements: it's about being able to efficiently represent things that truly need it, like the DOM.
In the book, you mention that "the std::raw modules contains structs with layouts that are the same as the complicated built-in types". I suppose what I'm asking for is some way for std::raw (plus some contract-like traits and other things) to somehow be the complicated built-in types.
I know that's a hard ask, but the upside would be that anything in the language that's represented by a complex layout + behavior could in principle be replaced by another implementation that preserves the size and behavior contracts. If an implementation is not provided, those features of the language are unavailable. This kind of takes the "you can't use x and y until you give me an allocator" approach and turns it up to 11.
Rust has had dynamic dispatch via since time immemorial, because there are a few things that static dispatch can't express (the original motivating factor was having an array full of different types that all implement the same interface). It is achieved by leveraging a feature of the traits system known as "trait objects": http://doc.rust-lang.org/book/static-and-dynamic-dispatch.ht...
But you are correct in that one of the loudest objections to the potential inclusion of inheritance in Rust is that people don't want two ways of achieving dynamic dispatch, because suddenly then you're in the same boat as C++ where you must decide which incompatible subset of the language to use in your codebase.
Is there a "post 1.0 wishlist" somewhere?