I have this crazy idea that you could build Clojure on Rust the same way Clojure sits on top of JVM right now and it could provide you with some interesting benefits.
First of all Rust doesn't implement a GC but it has very precise memory management semantics which means integrating in to a GC should be possible and safe.
I believe Clojure immutable data structures would allow you to (theoretically) build a fast/simplified GC because immutable data has some very nice properties :
* no cycles except trough reference types which are explicitly marked - this means you never need to scan old generation to collect the new generation
* data is immutable which means copying can be done in parallel
* mutable references are explicitly marked and isolated which means you could have a very minimal pause to update references
And Rust could safely encode all those requirements in it's type system to make the interop safe. No need to sprinkle write barriers all over the place, localized and isolated mutable references (you can modify the update semantics and track them easily), minimal pause times, parallel collection, etc.
I think there are other interesting things you could do, for example Clojure is pretty static for a dynamically typed language and protocols map nicely to traits I wonder if you could somehow remove the dynamic typing and just have compile time type inference figure out the types (requiring explicit types when the system can't infer the types) - this would allow you to send Clojure down to LLVM along with Rust. Or maybe you could have an interpreter for the truly dynamic stuff - but then you'd need to figure out how to export metadata.
Also I'm sure that the only sane/safe way to do it would be to restrict the kinds of objects and that would require serious thinking too.
I have this crazy idea that you could build Clojure on Rust the same way Clojure sits on top of JVM right now and it could provide you with some interesting benefits.
If you read some of the design docs and articles, you'll find that Clojure is designed to be a "hosted" language from the ground up. It would be a challenge to implement some of Clojure's dynamic features. You could very well do a Clojure-like language which has a different approach to being functional. Clojure is very Lisp-y in that it's actually a dynamic runtime fronted by a language that makes it look functional.
I had similar thoughts but wanted to take it a step further and just have a GC less Clojure built on Rust. You would have to do things like nest functions inside of (ns), you couldn't have (def) but this is all before I've spent any real hammock time on it. If you get started it would be fun to follow / contribute to.
It's an interesting idea, but the fact that the Rust compiler is an AOT compiler with an extremely minimal runtime makes it seem like a poor fit for Clojure which is very much an interpreter/REPL-driven language.
> this means you never need to scan old generation to collect the new generation
Even in HotSpot you don't need to scan the old generation if references in it don't change. You only scan (roughly) those old objects who hold references that have been mutated since the last collection.
> data is immutable which means copying can be done in parallel
This is also possible when mutation is allowed (see Red Hat's Shenandoah[1][2], A new GC for HotSpot), and immutability doesn't make the challenge significantly easier (there is added overhead for mutation only).
> mutable references are explicitly marked and isolated which means you could have a very minimal pause to update references... No need to sprinkle write barriers all over the place
Again, mutating reference fields has a special barrier in HotSpot's GCs, but they're not "sprinkled all over the place" -- they're as good as being specially marked.
To summarize, current (and future) HotSpot GCs punish you for mutation, but not for the possibility of mutation; you get all the benefits of immutability automatically when you simply don't mutate references. Immutability -- as you note -- might indeed simplify the GC (though not in the case of Clojure, which does allow quite a lot of mutability -- not everywhere, but enough that you have to account for it not as an outlier), but it doesn't make it faster.
----
> for example Clojure is pretty static for a dynamically typed language and protocols map nicely to traits I wonder if you could somehow remove the dynamic typing and just have compile time type inference figure out the types (requiring explicit types when the system can't infer the types) - this would allow you to send Clojure down to LLVM along with Rust.
A good JIT (like HotSpot) does this a lot better. It is true that Clojure is nowhere near as crazily-dispatched as, say Ruby, but it does rely a lot on Java-like interface-based dynamic dispatch. If your type inference is really, really good, you might be able to identify a great deal of the monomorphic call-sites, but a good JIT finds all of them and makes those calls virtually free, something an AOT compiler can never hope to achieve. Besides, HotSpot's optimizing JIT is a state-of-the-art compiler. You won't do any better with LLVM.
There are approaches that are good for some languages, and others that are good for others. Language's that rely on dynamic dispatch are better off using a JIT than an AOT compiler (actually, pretty much every language would get better code generation with a JIT -- except C which hardly does any dynamic dispatch -- but a JIT has some runtime overheads in warmup, RAM and energy, and languages that don't do too much dynamic dispatch can get excellent machine code generated by an AOT compiler, so that a JIT isn't worth it).
----
In short, while you could (maybe) compile Clojure down to Rust, you'll see no benefit, and probably quite a few disadvantages to that approach. If you want Clojure-Rust interoperability, you can get it today with RustJNI[3].
>Even in HotSpot you don't need to scan the old generation if references in it don't change. You only scan (roughly) those old objects who hold references that have been mutated since the last collection.
>This is also possible when mutation is allowed (see Red Hat's Shenandoah[1][2], A new GC for HotSpot), and immutability doesn't make the challenge significantly easier (there is added overhead for mutation only).
You pay for this with barriers and implementation complexity.
>To summarize, current (and future) HotSpot GCs punish you for mutation, but not for the possibility of mutation; you get all the benefits of immutability automatically when you simply don't mutate references. Immutability -- as you note -- might indeed simplify the GC (though not in the case of Clojure, which does allow quite a lot of mutability -- not everywhere, but enough that you have to account for it not as an outlier), but it doesn't make it faster.
>A good JIT (like HotSpot) does this a lot better. It is true that Clojure is nowhere near as crazily-dispatched as, say Ruby, but it does rely a lot on Java-like interface-based dynamic dispatch. If your type inference is really, really good, you might be able to identify a great deal of the monomorphic call-sites, but a good JIT finds all of them and makes those calls virtually free, something an AOT compiler can never hope to achieve. Besides, HotSpot's optimizing JIT is a state-of-the-art compiler. You won't do any better with LLVM.
Rust code doesn't pay for write barriers at all - my point isn't that Clojure on Rust would be faster than Clojure on JVM, it's that Rust code allows much more deterministic performance, better tools for manual optimizations and native interop than JVM and you'd write the performance sensitive code in Rust. Clojure is then a high level tool to consume Rust code.
I think you're looking at this the wrong way and missing the obvious benefits - you would get a simple almost pauseless parallel GC and a high level language with trivial native/safe rust interop (which JNI isn't).
Because of Rust type system you could even isolate threads that have no access to the shared heap and not collect/pause on those - so no chance at pauses at all. And AoT compilation also gives you deterministic performance.
Anyway it would be a lot of work for sure, I doubt it would be a straight port of clojure (would need to make it more static to fit to AOT nicely) and you would have to modify Rust to generate metadata so that the dynamics parts could run (or just ditch the dynamic part completely and force everything to be resolved at compile time with a simplified/constrained type system that would be simpler to infer without explicit typing).
> you would get a simple almost pauseless parallel GC
I don't see why you think you'd do any better than HotSpot's GCs in terms of pauses. You haven't described anything any of the current three production-quality HotSpot GCs don't already do (except for concurrent copying, which Shenandoah does). They also don't use write barriers -- unless you mutate references -- and you'd need those barriers when you mutate in Clojure anyway. I just don't see where the GC you've described differs from Clojure's current GCs.
Also, I'm not sure why you think Clojure data structures contain no loops except when relying on mutable references. Clojure sequences are just an interface, and it's easy to construct a lazy seq that contains loops.
> my point isn't that Clojure on Rust would be faster than Clojure on JVM, it's that Rust code allows much more deterministic performance, better tools for manual optimizations
But Clojure would take away from that deterministic performance and from your ability to manually optimize! Rust is a language designed to achieve good performance -- with safety -- in resource-constrained environments. You then want to take away that advantage by compiling a language that is both too wasteful for such environments and not deterministic.
The JVM has its use-cases and Rust has its own, and Clojure is by far more appropriate for the JVM use-case, so I still don't see the point. For performance sensitive code in Clojure today, you drop down to Java (or Kotlin).
Also, it is extremely hypothetical that you could get a higher-level, less hand-optimized language than Rust to enjoy Rust's benefits (that are bought precisely by paying for them with a more complicated language). I don't see how that could be possible with Clojure. You could, of course, create a more Lispy syntax for Rust -- with Rust semantics and a GC -- but how useful that language would be over, say, Common Lisp is unclear to me.
Think of Rust as a language that makes C++ safer, not as one that makes Java faster/cheaper. Rust can't offer the same abstractions Java/JVM provide for the same low cost -- it offers the same abstractions C++ does, with much better safety. I think that mixing the two will just reduce their advantages.
> which JNI isn't
RustJNI is very nice[1]. Again, I don't see how you think to do better than the current Clojure runtime, which is written in C++. Sure, it might be easier to write in Rust, but the C++ code already exists, and man-centuries of effort have been put into it. Clojure's heavy reliance on interface polymorphism and garbage collection takes advantage of pretty much every optimization provided by that runtime (and more are coming! see later).
> Because of Rust type system you could even isolate threads that have no access to the shared heap and not collect/pause on those - so no chance at pauses at all.
If that's so important, you can do the same with real-time JVMs (RTSJ). They have zero-pause non-heap threads (in fact, they're called NoHeapRealtimeThread[2]), and they access scoped arenas (not that arenas are a good choice for Clojure, a language that tends to produce a lot of garbage even in intermediate computations). A type system is just one way of achieving this isolation, and probably the wrong way to achieve it in Clojure. Plus, Rust's arena's are either type-specific or rely on reflection.
P.S.
IMO, the coolest thing regarding compiling Clojure is now Truffle/Graal -- HotSpot's next-gen JIT and language-compilation platform -- that will be available for stock OpenJDK builds in Java 9. It will allow much more aggressive optimizations for Clojure code. It will also let you write the Clojure machine-code generation logic in Clojure!
First of all Rust doesn't implement a GC but it has very precise memory management semantics which means integrating in to a GC should be possible and safe.
I believe Clojure immutable data structures would allow you to (theoretically) build a fast/simplified GC because immutable data has some very nice properties :
* no cycles except trough reference types which are explicitly marked - this means you never need to scan old generation to collect the new generation
* data is immutable which means copying can be done in parallel
* mutable references are explicitly marked and isolated which means you could have a very minimal pause to update references
And Rust could safely encode all those requirements in it's type system to make the interop safe. No need to sprinkle write barriers all over the place, localized and isolated mutable references (you can modify the update semantics and track them easily), minimal pause times, parallel collection, etc.
I think there are other interesting things you could do, for example Clojure is pretty static for a dynamically typed language and protocols map nicely to traits I wonder if you could somehow remove the dynamic typing and just have compile time type inference figure out the types (requiring explicit types when the system can't infer the types) - this would allow you to send Clojure down to LLVM along with Rust. Or maybe you could have an interpreter for the truly dynamic stuff - but then you'd need to figure out how to export metadata.
Also I'm sure that the only sane/safe way to do it would be to restrict the kinds of objects and that would require serious thinking too.