A tail-call interpreter in (nightly) Rust(mattkeeter.com) |
A tail-call interpreter in (nightly) Rust(mattkeeter.com) |
it's always surprising for me how absurdly efficient "highly specialized VM/instruction interpreters" are
like e.g. two independent research projects into how to have better (fast, more compact) serialization in rust ended up with something like a VM/interpreter for serialization instructions leading to both higher performance and more compact code size while still being cable of supporting similar feature sets as serde(1)
(in general monomorphisation and double dispatch (e.g. serde) can bring you very far, but the best approach is like always not the extrem. Neither allays monomorphisation nor dynamic dispatch but a balance between taking advantage of the strength of both. And specialized mini VMs are in a certain way an extra flexible form of dynamic dispatch.)
---
(1): More compact code size on normal to large project, not necessary on micro projects as the "fixed overhead" is often slightly larger while the per serialization type/protocol overhead can be smaller.
(1b): They have been experimental research project, not sure if any of them got published to GitHub, non are suited for usage in production or similar.
You're adding a layer of abstraction and indirection, so how is it possible that a more indirect solution can have better performance?
This seems counterintuitive, so I googled it. Apparently, it boils down to instruction cache efficiency and branch prediction, largely. The best content I could find was this post, as well as some scattered comments from Mike Pall of LuaJIT fame:
https://sillycross.github.io/2022/11/22/2022-11-22/
Interestingly, this is also discussed on a similar blogpost about using Clang's recent-ish [[musttail]] tailcall attribute to improve C++ JSON parsing performance:
https://blog.reverberate.org/2021/04/21/musttail-efficient-i...
It is funny, but (like I’ve already mentioned[1] a few months ago) for serialization(-adjacent) formats in particular the preferential position of bytecode interpreters has been rediscovered again and again.
The earliest example I know about is Microsoft’s MIDL, which started off generating C code for NDR un/marshalling but very soon (ca. 1995) switched to bytecode programs (which Microsoft for some reason called “format strings”; these days there’s also typelib marshalling and WinRT metadata-driven marshalling, the latter completely undocumented, but both data-driven). Bellard’s nonfree ffasn1 also (seemingly) uses bytecode, unlike the main FOSS implementations of ASN.1. Protocol Buffers started off with codegen (burying Google user in de/serialization code) but UPB uses “table-driven”, i.e. bytecode, parsing[2].
The most interesting chapter in this long history is in my opinion Swift’s bytecode-based value witnesses[3,4]. Swift (uniquely) has support for ABI compatibility with polymorphic value types, so e.g. you can have a field in the middle of your struct whose size and alignment only become known at dynamic linking time. It does this in pretty much the way you expect[5] (and the same way IBM’s SOM did inheritance across ABI boundaries decades ago): each type has a vtable (“value witness”) full of compiler-generated methods like size, alignment, copy, move, etc., which for polymorphic type instances will call the type arguments’ witness methods and compute on the results. Anyways, here too the story is that they started with native codegen, got buried under the generated code, and switched to bytecode instead. (I wonder—are they going to PGO and JIT next, like hyperpb[6] for Protobuf? Also, bytecode-based serde when?)
[1] https://news.ycombinator.com/item?id=44665671, I’m too lazy to copy over the links so refer there for the missing references.
[2] https://news.ycombinator.com/item?id=44664592 and parent’s second link.
[3] https://forums.swift.org/t/sr-14273-byte-code-based-value-wi...
[4] Rexin, “Compact value witnesses in Swift”, 2023 LLVM Dev. Mtg., https://www.youtube.com/watch?v=hjgDwdGJIhI
[5] Pestov, McCall, “Implementing Swift generics”, 2017 LLVM Dev. Mtg., https://www.youtube.com/watch?v=ctS8FzqcRug
Tail recursion opens up for people to write really really neat looping facilities using macros.
Last year I was working on a tail-call interpreter (https://github.com/anematode/b-jvm/blob/main/vm/interpreter2...) and found a similar regression on WASM when transforming it from a switch-dispatch loop to tail calls. SpiderMonkey did the best with almost no regression, while V8 and JSC totally crapped out – same finding as the blog post. Because I was targeting both native and WASM I wrote a convoluted macro system that would do a switch-dispatch on WASM and tail calls on native.
Ultimately, because V8's register allocation couldn't handle the switch-loop and was spilling everything, I basically manually outlined all the bytecodes whose implementations were too bloated. But V8 would still inline those implementations and shoot itself in the foot, so I wrote a wasm-opt pass to indirect them through a __funcref table, which prevented inlining.
One trick, to get a little more perf out of the WASM tail-call version, is to use a typed __funcref table. This was really horrible to set up and I actually had to write a wasm-opt pass for this, but basically, if you just naively do a tail call of a "function pointer" (which in WASM is usually an index into some global table), the VM has to check for the validity of the pointer as well as a matching signature. With a __funcref table you can guarantee that the function is valid, avoiding all these annoying checks.
Do you have any idea which operations regress the most?
The "become" keyword allows us to express our meaning, we want the tail call, and, duh, of course the compiler will optimize that if it can be a tail call but also now the compiler is authorized to say "Sorry Dave, that's not possible" rather than grow the stack. Most often you wrote something silly. "Oh, the debug logging happens after the call, that's never going to work, I will shuffle things around".
> Tail calls can be implemented without adding a new stack frame to the call stack. Most of the frame of the current procedure is no longer needed, and can be replaced by the frame of the tail call, modified as appropriate (similar to overlay for processes, but for function calls). The program can then jump to the called subroutine. Producing such code instead of a standard call sequence is called tail-call elimination or tail-call optimization. (https://en.wikipedia.org/wiki/Tail_call)
Not really, a trampoline could emulate them effectively where the stack won't keep growing at the cost of a function call for every opcode dispatch. Tail calls just optimize out this dispatch loop (or tail call back to the trampoline, however you want to set it up).
The current RFC generally does not allow `become` to be used with `async` for now [0]:
> Tail calling from async functions or async blocks is not allowed. This is due to the high implementation effort as it requires special handling for the async state machine. This restriction can be relaxed by a future RFC.
> Using `become` on a `.await` expression, such as `become f().await`, is also not allowed. This is because `become` requires a function call and `.await` is not a function call, but is a special construct.
> Note that tail calling async functions from sync code is possible but the return type for async functions is `impl Future`, which is unlikely to be interesting.
[0]: https://github.com/phi-go/rfcs/blob/guaranteed-tco/text/0000...
I wonder why they went with a new keyword; I assumed the compiler would opportunistically do TCO when it thinks it's possible, and I figured that the simplest way to require TCO (or else fail compilation) could be done with an attribute.
(Not sure if the article addressed that... I only skimmed it.)
Rust has been really good at providing ergonomic support for features we're too used to seeing provided as "Experts only" features with correspondingly poor UX.
> Last week, I wrote a tail-call interpreter using the become keyword, which was recently added to nightly Rust (seven months ago is recent, right?).
Maybe the same idea can be used in Rust where some constructs are easier to write in recursive form instead of a loop?
In any case, here's a silly example of a `for-loop` macro in Scheme using a tail call:
(define-syntax for-loop
(syntax-rules ()
((for-loop var start end body ...)
(letrec ((loop (lambda (var)
(unless (>= var end)
body ...
(loop (+ var 1)))))) ; <-- tail call
(loop start)))))
(for-loop i 0 10
(display i)
(newline))
(Note that you'll probably never see a `letrec` used like this: people would use a named `let`, which is syntax sugar for that exact usage of `letrec`. I wrote it the `letrec` way to make the function explicit).
This is just sugar over tail calls.
The compiler is free to do something clever (eg., two versions of fn Foo(...), one guaranteeing TCE), but as a developer merely using Foo, I don't want to have to decorate it (or, in a library, have it already decorated) to make TCE (not just TCO) possible. That requirement would constitute delving into the internals.
The first RFC for guaranteed tail calls stated an attribute on `return` was a possible alternative "if and when it becomes possible to attach attributes to expressions" [0]. That was from pre-1.0, though; I believe Rust now supports attributes on at least some expressions, but I don't know when that was added.
The second RFC [1] doesn't seem to discuss keyword vs. attribute, but it does mention that the proof-of-concept implementation "parses `become` exactly how it parses the `return` keyword. The difference in semantics is handled later", so perhaps a keyword is actually simpler implementation-wise?
There's some more discussion on attribute vs. keyword starting here [2], though the attribute being discussed there is a function-level attribute rather than something that effectively replaces a `return`. The consensus seems to be that a function-level attribute is not expressive enough to support the desired semantics, at least. There's also a brief mention of `become` vs. `return` (i.e., new keyword because different semantics).
[0]: https://github.com/rust-lang/rfcs/pull/81/changes
[1]: https://github.com/DemiMarie/rfcs/blob/become/0000-proper-ta...
[2]: https://github.com/rust-lang/rfcs/pull/1888#issuecomment-278...
It does have some more discussion on keywords vs attributes and other options.
> Even in a release build, the compiler has not optimized out the stack. As we execute more and more operations, the stack gets deeper and deeper until it inevitably overflows.
Not in serde itself, but people have been experimenting with serde alternatives that are bytecode based. Nothing stable as far as I know, just experiments.
One early experiment is described in https://sdr-podcast.com/episodes/a-different-serde/ , which used a FORTH inspired stack based VM generated by derive macros at compile time (iirc).
Another experiment is https://lib.rs/crates/facet which is a more general derives to generate compile time introspection to generate metadata tables approach.
I suppose maybe TCE (as distinct from TCO) should be expanded to include any mechanism that doesn't expand the stack / heap / whatever for things that rhyme with recursion (in which case the existing sloppiness may as well stand, but we need a new TLA).
Anyway, here's something more-or-less equivalent in Rust, which will blow the stack if made to loop too many times: https://play.rust-lang.org/?version=stable&mode=debug&editio...
(There may be a way to use a closure instead of a function to avoid hard-coding the type of `$i` in the macro, but I can't find an easy way to write a recursive closure call in Rust).
One of the things I forget to give thanks to the Rust designers for is the default of immutability. I remember my days in a Java environment where policy was "everything needs to be declared final" (so "final" got spattered all over the place) with zero nostalgia. Requiring TCE callees to explicitly represent themselves as TCE-friendly would be the same sort of abomination.
Having a way to guarantee TCO/TCE is essential for some cases, yes. GP's question, though, was why a keyword specifically and not a hypothetical attribute that effectively does the same thing.
My fear is that adding yet another keyword it might get lost in the sea of keywords that a Rust developer needs to remember. And if recursion is not something you do often you might not reach for it when actually needed. Having this signal in the function signature means that people would be exposed to it just by reading the documentation and eventually will learn it exists and (hopefully) how to wield it.
What do you folks think?
`become blah(foo, bar)` is the same thing as `blah(foo, bar)` except that we, the caller are promising that we have nothing further to do and so when blah returns it can return to our caller.
If somebody else calls blah they don't want that behaviour, they might have lots more to do and if they were skipped over that's a disaster.
In some languages it's very obvious when you're going to get TCO anyway, but Rust has what C++ calls RAII, when a function ends all the local variables get destroyed and this may be non-trivial work. Presumably destroying a local i32 is trivial & so is a [u8; 32] but destroying a local String isn't, let alone a HashMap and who knows how complicated it is to destroy a File or a DBQuery or a Foo...
So in a sense "all" become does is try to bring that destruction sooner a little, so it happens before the call, leaving nothing to do afterwards. We weren't using that String any more anyway, lets just destroy it first, and the HashMap? Sure, and oh... no, actually if we destroy that Foo before calling blah which needs the Foo that messes things up... Rust's borrowck comes in clutch here to help us avoid a terrible mistake, our code was nonsense, it doesn't build.
Edited: Improve explanation
As for being in the documentation, I'd argue that might even be an explicit non-goal; it's not clear to me why that should be something considered part of the API. If someone updates their tail recursive function to manually loop instead (or vice-versa), it doesn't seem like that should necessitate the documentation being updated.
const NINE: i32 = // Some arbitrary *constant* expression;
OK, but that's just equivalent to a nested function, nothing special?
Relatedly, I still sometimes get tripped up by the nuances of using `const` versus `static` for top-level constants. Most of the time the difference is entirely opaque to the programmer (because it's not obvious when most things are getting inlined or being referenced from a single place in memory), but it's possible to run into cases where one works and the other won't (e.g. trying to be clever with `OnceCell` rather than `OnceLock`).
Statics can be mutated - though not safely - because they are a single concrete thing so they can be changed, whereas it can't mean anything to mutate a constant, hence the word "constant".
For larger objects you might want a single concrete thing even though it might intuitively not seem important because it impacts performance. For example if we keep talking about FACTOR[n] where FACTOR is an array of a million numbers (maybe computed by scientist colleagues for your application) and n is a variable, if FACTOR is const Rust is going to just put a copy of that enormous array everywhere it needed to do this indexing operation, which gets out of hand really fast, whereas if we use static we get a single concrete thing, named FACTOR and everywhere in the program will use that one single million number array, much tidier and less likely to result in say, running out of RAM on a small computer.
For what it's worth, my rule of thumb is usually to start with `static` and then only swap to `const` if I have a reason to. If I recall correctly, the issue I alluded to above was around picking between `LazyCell` and `LazyLock` and swapping between `&Path` and `PathBuf`, and some combination of them only working with `const` and not `static`.