How to make a fast dynamic language interpreter
zef-lang.dev
zef-lang.dev
Unlike the Zef article, which describes implementation techniques, the Wren page also shows ways in which language design can contribute to performance.
In particular, Wren gives up dynamic object shapes, which enables copy-down inheritance and substantially simplifies (and hence accelerates) method lookup. Personally I think that’s a good trade-off - how often have you really needed to add a method to a class after construction?
See experience with Smalltalk and Self, where everything is dynamic dispatch, everything is an object, in a live image that can be monkey patched at any given second.
PyPy and GraalPy, and the oldie IronPython, are much better experiences than where CPython currently stands on.
The JIT would help everyone else more than removing the GIL, I wish PyPy became the reference implementation during 2.7
It is also because of AI, that Intel, AMD and NVidia are now getting serious about Python GPU JITs, that allow writing kernels in a Python subset.
To the point that I bet Mojo will be too late to matter.
Edit: I think what you're alluding to is that tracing JITs can overcome a lot of dynamic language features which make things hopeless for method JITs. Where LuaJIT really shines vs PyPy is outside of JITed loops. (Also memory and compile overheads). I realise this is a bit of a motte and bailey.
On the other side, having a type holding a closed set of applicable functions is somehow questioning.
There are languages out there that allows to define arbitrary functions and then use them as a methods with dot notation on any variable matching the type of the first argument, including Nim (with macros), Scala (with implicit classes and type classes), Kotlin (with extension functions) and Rust (with traits).
"Efficient implementation of the smalltalk-80 system"
Or its maintainability, and this is one of the big reasons why. Methods and variables are dynamically generated at runtime which makes it impossible to even grep for them. If you have a large Ruby codebase (say Gitlab or Asciidoctor), it can be almost impossible to trace through code unless you are familiar with the entire codebase.
Their "answer" is that you run the code and use the debugger, but that's clearly ridiculous.
So I would say dynamically defined classes is not only bad for performance; it's just bad in general.
https://benchmarksgame-team.pages.debian.net/benchmarksgame/...
A general rule of thumb is that if you can assign an expression a static type, then you can compile it fairly efficiently. Complex dynamic languages obviously actively fight this in numerous ways, and so end up being difficult to optimize. Seems obvious in retrospect.
I’m basing that on the 1.6% improvement they got on speeding up sqrt. That surprised me, because, to get such an improvement, the benchmark must spend over 1.6% of its time in there, to start with.
Looking in the git repo, it seems that did happen in the nbody simulation (https://github.com/pizlonator/zef/blob/master/ScriptBench/nb...).
Basically the flow was:
- check if we’re calling a method of an object
- nope, ok, so cascade through 10+ symbol comparisons
- sqrt was towards the bottom of the cascade
I also like how, according to Github, the repo is 99.7% HTML and 0.3% C++. A testament to the interpreter's size, I guess?
But yeah the interpreter is very small
I didn't want any optimisation complexities and just focused on being able to understand my own Rust code. I was surprised by the performance I got simply by using my favourite language and as a bonus, since Rust takes care of all the ownership and lifetimes, I don't need a garbage collector. For sure, right now I'm being super conservative and rely on cloning stuff to avoid lifetime hell in stuff like closures, but the speed and memory profile is still very decent.
For anyone interested in a simple to understand tree-walking interpreter in Rust, which is heavily based in expressive enums where code is data, here's my interpreter:
> as a bonus, since Rust takes care of all the ownership > and lifetimes, I don't need a garbage collector.
I can imagine GluconScript's memory handling comes at a cost, even if the tradeoff of using a borrow checker is well worth it. Was that your experience?
Relatedly, since you commented there has been submission about garbage collectors in Rust ("Garbage Collection Without Unsafe Code"):
As far as I researched, only closures could generate dangling references and therefore need memory cleanup, but only if I allowed closures to access their environment (variables and functions) by reference / mutable reference.
To avoid this and simplify both my code as well as the mental model for the users of GluonScript, as of now, closures capture their environment by cloning it immutably. There's an increased memory usage with all the copying of the environment but there are never references to something that isn't being used anymore and therefore no need for a GC. At the end of the day all values captured by closures are owned Rust values that are dropped by Rust when no longer in scope.
So this can lead to high memory usage in hot loops but it can't lead to memory leaks.
I've gone through something similar, but for a more functional language (a Scheme). It's interesting how here the biggest wins are from optimizing the objects, while the biggest wins in my case were optimizing closures. The optimizations were very similar.
"Three implementation models for scheme" gives all the answers to make a fast enough scheme, though it has something of a compilation step, so it's not interpreting the original AST.
And the fact that having outline calls to methods of value objects is so expensive
Is this tied to unions? Or otherwise, when does this happen? I don't see the connection w/ invisicaps or &c
“Escape” is defined very loosely; it currently means: some function other than the one that owns the stack allocation needs a pointer to that allocation.
For example even if you could prove that `bar(Value* p)` never stashes p anywhere, the fil-C compiler will currently heap allocate that value anytime bar is called. The one exception is if bar had already been inlined, and so from the FilPizlonator’s perspective there isn’t even a call.
This is clearly dumb and fixable. It’s dumb because lots of functions aren’t worth inlining but their body is analyzable. Slow paths are like that. It’s fixable because those slow paths - and lots of code like them - takes ptrs as arguments and then obviously just uses them for loads and stores but doesn’t escape them any further.
You’ll sometimes hear me say that Fil-C is nowhere near as optimal as it could be. This is just one example of that
It was materially useful in this project.
- Caught multiple memory safety issues in a nice deterministic way, so designing the object model was easier than it would have been otherwise.
- C++ with accurate GC is a really great programming model. I feel like it speeds me up by 1.5x relative to normal C++, and maybe like 1.2x relative to other GC’d languages (because C++’s APIs are so rich and the lambdas/templates and class system is so mature).
But I’m biased in multiple ways
- I made Fil-C++
- I’ve been programming in C++ for like 35ish years now
This will greatly reduce coordination bugs in parallel programs and may even speed things up.
- STM interacts badly with any interesting effects (like IO)
- STM interacts badly with locks (C and C++ use locks implicitly, because they’re in libc and libc++)
- STM performs badly
- STM is harder to use than locks
It doesn't seem like that is necessarily a performance win, especially since you could always use a smart pointer's raw pointer (preferably const) in a performance critical path.
Smart pointers give you reference counting at best. Language implementations tend to have a lot of reference cycles, so you need actual GC (or a very sophisticated kind of manual memory management)
> happen to know C++ really well
That’s my bias yeah. But C++ is good for more than just perf. If you need access to low level APIs, or libraries that happen to be exposed as C/C++ API, or you need good support for dynamic linking and separate compilation - then C++ (or C) are a great choice
The syntax and ownership rules can take some getting used to but after doing it I start to wonder how I ever enjoyed the masochism of the rule of 5 magic incantation that no one else ever followed and writing the class definition twice. + the language gaining complexity constantly without ever paying back tech debt or solving real problems.
There are many runtimes that I could have included but didn’t.
Also, it’s quite impressive how much faster PUC Lua is than QuickJS and Python
(I suppose the quick in QuickJS means "quick for a pure interpreter without JIT compilation or something...)
So like that’s wild
Python's execution time is mostly spent looking up stuff. I don't think lua is quite as dynamic.
local t = setmetatable({}, {
__index = pcall, __newindex = rawset,
__call = function(t, i) t[i] = 42 end,
})
for i=1,100 do assert(t[i] == true and rawget(t, i) == 42) end
A more nuanced take is that Lua has many happy fast paths, whereas Python has some unfortunate semantic baggage that complicates those. Another key issue is the over-reliance on C modules with bindings that expose way too many internals.
This is a good way to describe it. Most of the semantic baggage doesn't make some speed improvements, up to and including JITing, impossible, but it certainly complicates them.
And of course, any semantic baggage will be useful to someone.
But in Python, everything is an object, which is why, as I said, it spends much of its time looking things up. And things like bindings for closures are late, so that's more lookups as well.
In lua, many things aren't objects, and, for example, you can add two numbers without looking anything up. Another issue, of course, when you do that, is that you could conceivably overflow an integer, but that can't happen in Python either.
The Python interpreter has some fast paths for specific object types, but it is really limited in the optimizations it can do, because there simply aren't any unboxed types.
This feels more like a party trick than anything. But it does represent a deep commitment to founding the whole language on object orientation, even when it seems silly to folks like me.
It also made it really easy to ingest code, and do meta programming.
It doesn’t have to in the absolute. It just that if some speech seel that a programing language is completely object oriented, it’s fun to check to which point it actually is.
There are many valid reasons why one would not to do that, of course. But if it’s marketed as if implicitly one could expect it should, it seems fair to debunk the myth that it’s actually a fully object language.
>Is space dot class a thing?
Could be, though generally spaces are not considered like terms – but Whitespace shows it’s just about what is conventionally retained.
So, supposing that ` .class` and `.class` express the same value, the most obvious convention that would come to my mind then would be to consider that it’s applied to the implicit narrower "context object" in the current lexical scope.
Raku evaluate `.WHAT` and `(.WHAT)` both as `(Any)` for giving a concrete example of related choice of convention.
>Such syntax is merely for producing an AST and that alone doesn't mean "object" or "not object".
Precisely, if the language is not providing complete reflection facility on every meaningful terms, including syncategorematic ones, then it’s not fully object. Once again, being almost fully object is fine, but it’s not being fully object.
To some extent, sure. And, looking at your implementation of your language, something like the optimizations on passing small numbers of parameters could probably help Python out. It spends an inordinate amount of time packing and unpacking parameter tuples.
But, for example, you can easily create a subclass of an integer and alter a small portion of its behavior, without having to code every single operation, which I don't think you can do in lua.
So, the dynamicity I'm describing is what the language has to do (more work at runtime) to support its own semantics.
Don't get me wrong. There are certainly opportunities to make Python go faster, and the core team is working on some of them (for example, one optimization is similar to your creation of additional subtree nodes for attribute lookup for known cases, but in bytecode instead), but I also think that the semantics of Python make large classes of optimization more difficult than for other languages.
For a major example of this kind of dynamicity, lua doesn't chain metatables when looking up metamethods, but Python will look stuff up in as many tables as you have subclasses, and has the complexity of dealing with MRO. That's not something that couldn't be JITed, but the edge cases of what you need to update if someone decides to add or modify a method in a superclass get pretty hairy pretty quickly.
Whereas, in lua, if you want to modify a metamethod and have it affect a particular object, yes, absolutely, you can do that, but it is up to you to modify the direct metatable of the object, rather than some ancestor, because lua is not going to dynamically follow the chain of references on every lookup.
And, back to to the parameter optimization case, I haven't thought that much about it, but there are a lot of Python edge cases in parameter passing, that might make that difficult.
And, of course, the use of ref counting instead of mark/sweep has a cost, but people don't like, e.g., PyPy, because their __del__ methods aren't guaranteed to be called immediately when the object goes out of scope. Lua is more like PyPy in this respect.
So Python has a lot of legacy decisions that make optimization harder.
Then things that try to be called Python often take shortcuts that make things faster, but don't get any traction, because they aren't 100% compatible.
So cPython is a Schelling point with semantics that are more complicated than some other language Schelling points, with enough momentum that it becomes difficult for other Python implementations to keep up with the standard, while simultaneously having enough inertia to keep people engaged in using it even though the optimizations are coming slowly.
I think the sPy language (discussed here a few weeks ago) has the right idea. "Hey, we're not Python, but if you like Python you might like us." Things that claim to be Python but faster either wither on the vine because of incompatibilities with cPython, or quickly decide they aren't really Python after they've lost their, ahem, Mojo, or both.
(The primary exception to this is microPython, which has a strong following because it literally can go where no other Python can go.)
That’s where for example getter inference happens.