Using a perpetual timer to poll anything is a waste of your users’ battery life. Even if the actual computation you’re doing seems insignificant, the repeated wakeups prevent the processor from reaching deeper sleep states.
While a perfectly reasonable point, and one which might benefit other readers, I think it’s worth adding that the article seems fairly clear that it’s demonstrating a proof of concept that can be built upon. I certainly didn’t take it to suggest that I should go write a convoluted setInterval to implement GC in WASM.
Amusingly (to me), I’ve been working on a variety of performance-focused efforts—professionally on directly deliverable tasks, professionally on exploratory spikes for more holistic future improvements, and in pet personal projects. I find myself deliberately writing throwaway code that’s just as painfully suboptimal quite a lot. At worst it helps me quickly iterate on the actual bottleneck without focusing on the things I’m not trying to optimize.
It can be more helpful than that though, when it reveals sometimes unexpected JIT perf characteristics. Sometimes it reveals that a hot loop optimization isn’t going to bear much fruit because the JIT is going to optimize the baseline code better than anything I have in mind. Other times it reveals cases where I see dramatic benefits from an initial JIT pass that vanish because other factors cause a deoptimization.
Granted I think it’s still good to slap a big warning like this on obviously suboptimal code! But I also think people working on performance sensitive code in a dynamic JIT/GC environment should be encouraged to write weird suboptimal throwaway glue code around their perf focal point. You can learn a lot if you’re profiling!
One thing is lack of performance profiling, another thing is a design flaw (can this performance issue be really tackled or is it a limitation derived from the platform?). Which one are we looking at?
We’re looking at neither. I even double checked the article and it’s very explicitly an example demonstrating very clearly a concept. It’s a proof of concept. It’s not a design so it won’t suffer from a flaw, it’s not expected to be profiled because it’s not a solution.
My comment was an aside that such code lacking design can aid profiling other code. Intentionally writing poorly performing code has benefits in understanding other performance details in some scenarios.
You didn’t seem to address this either, just reiterated a grievance with the article. I expect this is going to be a typical HN topic where strong opinions fly by, but I sincerely wonder if you might also benefit from slowing down and reading more closely?
How does this handle cycles? It seems to me that if JS holds on to a WebAssembly object, and that WebAssembly object holds on to another JS object that in turn has a reference to the same WebAssembly object, this solution will leak.
This may seem like a contrived scenario, but it happens a lot with callback patterns. You might have code like:
// MyNativeButton is a native WebAssembly object
// Perhaps it's part of a WebAssembly GUI framework that renders to canvas
this.button = new MyNativeButton();
this.button.onclick = () => this.clicked = true;
This will leak the "this" object forever because there's a cycle: "this" -> MyNativeButton -> JS closure -> "this".
The “WebAssembly object” doesn’t exist as you describe. You probably have an object that wraps WebAssembly function calls and primitive WebAssembly data (individual numbers).
WebAssembly functions can only return numbers and can only take numbers as arguments. Everything else is a wrapper that lives in JavaScript.
MyNativeButton works exactly as you’d expect. If the GC can clean it up, it might. Otherwise you have a leak that you’d have had in normal JavaScript.
I’d assume the baseline advice would be the same as I’d give without WASM: if you’re treating `this` as mutable outside of a constructor you’re probably already screwed. Maybe it gets more complicated if you can’t see clearly whether your instance is a value type in WASM, but like your example shows… people generally don’t build value types with this kind of semantics in JS anyway. They’re already incurring those costs. My hope would be that WASM library authors are focused enough on resource usage that they’ll be more mindful of that than the baseline.
The only closed reference to `this` in the example is a mutation, in an event handler. This is supposed retain the reference and it’s not even a leak. The simplest way to solve that would be to remove the event handler, but even simpler would be to treat `this` as a value type and never change it after construction. The other reference is easily rewritten as a WeakMap key and would not retain a reference, but easier still could be written as a local binding with no reference to `this` at all.
This is a bit odd. It's polling WeakRefs to simulate FinalizationRegistry. I don't know what problem the author had with FinalizationRegistry, but WeakRef and FinalizationRegistry use the same logic (and mostly the same code) to determine when the target has been collected. (Source: I reviewed the implementations of WeakRef and FinalizationRegistry in Firefox -- https://bugzilla.mozilla.org/show_bug.cgi?id=1587096 for one of the bugs.)
My guess is that it ran into the restrictions on collection visibility. If you look at the target at all (or a WeakRef to it) while running JS, then it will not do the finalization callback until JS code completes running even if the object is no longer reachable. (It's artificially kept alive until JS finishes running.) The example code handles this with the setTimeout call, which allows the JS code to return before polling all the WeakRefs.
It could have been a testing problem? Or maybe it's simpler: there wasn't enough allocation pressure to trigger a GC. JS objects won't get collected until there's a GC. (I guess that's kind of a tautology...)
Although allowed by the spec, it would be very surprising for an implementation to clear a WeakRef at a different time than it calls the FinalizationRegistry callback. (Or did we end up requiring it to be in the same tick? I can't remember. There's some restriction, maybe because the finalizer could check whether the WeakRef target is still there and you wouldn't want them to disagree?)
But in general, it's true that you can't rely on any particular timing of finalization. It'll happen eventually in all browsers if there is continued allocation. (I think all browsers might eventually do an idle GC as well?) Things can also be kept alive for weird internal reasons -- what if something gets JITted and has an IC (inline cache) that refers to an object you expect to be dead? The JS engine might keep that jit code alive for longer than you'd expect, even if it collects other things.
And of course this won't handle cycles. But it is exactly the thing that FinalizationRegistry made it into the language for, for better or worse.
> Although allowed by the spec, it would be very surprising for an implementation to clear a WeakRef at a different time than it calls the FinalizationRegistry callback. (Or did we end up requiring it to be in the same tick? I can't remember. There's some restriction, maybe because the finalizer could check whether the WeakRef target is still there and you wouldn't want them to disagree?)
The spec says that both clearing WeakRefs and scheduling finalization callbacks should happen at the same time: <https://tc39.es/ecma262/#sec-weakref-execution>. Of course, since that is merely ‘scheduling’, this means the callback can actually execute in some indefinite future, including never; the callback might as well be scheduled for AD 2101, with the runtime shut down long before then. It has to be done that way, because collection may happen in the middle of execution, but finalization callbacks cannot possibly interrupt it.
The bigger problem with the spec is that when an object becomes eligible for finalization in the first place is… not at all obvious: <https://github.com/tc39/ecma262/issues/2650>.
Wow, that issue is kind of mindblowing. To me, it proves that making GC observable rips a large hole in the semantics. The damage was ameliorated somewhat with [[KeptAlive]], but the destruction is not contained in any understandable or bounded way, and we'll be bumping into it for years to come.
It feels like we managed to add something a lot like C and C++'s undefined behavior into JS. I thought it was just an issue of nondeterminism, but that issue proves that it could be a lot more problematic than that. I don't find the suggested optimizations particularly far-fetched, especially in a trace-compiled or otherwise specialized subset of the code. Something like inlining in a way that looks for all of the things provable locally even if they may not be true globally; it could totally do things like build alias sets of values and optimize down `w.deref() === obj` to `w.deref() !== undefined`.
Basically, looking at the target of a WeakRef makes it ineligible for collection until at least the next await, which the spec mandates by way of the [[KeptAlive]] list. This is meant to prevent TOCTOU bugs by foolish programmers doing this:
if (weakref.deref()) {
/* garbage collection here */
use(weakref.deref());
}
I don't know about this proposal in particular, but it is common for garbage-collected languages to use the highest bit to distinguish pointers and integers; which means integers cannot use it.
I like webassembly mostly because it doesn’t have garbage collection.
I know the article wasn’t calling for anything like this, but Please keep wasm GC free. Golang, for example, compiles in its own GC (which is why it’s wasm is so large).
You always have to be careful of memory issues when mixing languages - see using Lua or mixing C and Go.
I’d rather deal with memory leaks than fighting garbage collection. If you want GC, maybe just stay in Javascript?
Don't think your comment makes a lot of sense in the context of the article. Author explicitly wants to create/use/drop some WASM instances in the JavaScript world without leaking memory. Piggybacking on the existing JS GC is perfectly reasonable approach here.
Well, I said my comment wasn't super related to the article.
I suppose my somewhat hidden comment was that if one wants GC you should build it into the wasm side like golang does. I don't think creating something in one system, and trying to memory manage it in another system is good idea in the first place - at least not something you should do on a real system.
I think playing around with stuff is fun, and I didn't want to come across negative because while I think it's a bad idea, I love playing with bad ideas.
https://wiki.ubuntu.com/Kernel/PowerManagement/IdentifyingIs...