pcwalton

> I really don’t think Rust is a good match for game dev. Both because of the borrow checker which requires a lot of handles instead of pointers and because compile times are just not great.

I completely disagree, having been doing game dev in Rust for well over a year at this point. I've been extremely productive in Bevy, because of the ECS. And Unity compile times are pretty much just as bad (it's true, if you actually measure how long that dreaded "Reloading Domain" screen takes).

> More than anything else, this sounds like a good lesson in why commercial game engines have taken over most of game dev. There are so many things you have to do to make a game, but they're mostly quite common and have lots of off-the-shelf solutions.

> That is, any sufficiently mature indie game project will end up implementing an informally specified, ad hoc, bug-ridden implementation of Unity (... or just use the informally specified, ad hoc and bug-ridden game engine called "Unity")

But using Bevy isn't writing your own game engine. Bevy is 400k lines of code that does quite a lot. Using Bevy right now is more like taking a game engine and filling in some missing bits. While this is significantly more effort than using Unity, it's an order of magnitude less work than writing your own game engine from scratch.

But Unity game objects are the same way: you allocate them when they spawn into the scene, and you deallocate them when they despawn. Accessing them after you destroyed them throws an exception. This is exactly the same as entity IDs! The GC doesn't buy you much, other than memory safety, which you can get in other ways (e.g. generational indices, like Bevy does).

> We were writing an animation system in Bevy and were hit by the painful upgrade cycle twice.

I definitely sympathize with the frustration around the churn--I feel it too and regularly complain upstream--but I should mention that Bevy didn't really have anything production-quality for animation until I landed the animation graph in Bevy 0.15. So sticking with a compatible API wasn't really an option: if you don't have arbitrary blending between animations and opt-in additive blending then you can't really ship most 3D games.

Most game engines other than the latest in-house AAA engines are leaving comparable levels of performance on the table on scenes that really benefit from GPU-driven rendering (that's not to say all scenes, of course). A Google search for [Unity drawcall optimization] will show how important it is. GPU-driven rendering allows developers to avoid having to do all that optimization manually, which is a huge benefit.

> Just a mode to automatically coerce between numeric types would make Rust so much more ergonomic for gamedev.

C# is stricter about float vs. double for literals than Rust is, and the default in C# (double) is the opposite of the one you want for gamedev. That hasn't stopped Unity from gaining enormous market share. I don't think this is remotely near the top issue.

> Nobody has really pushed the performance issues.

This is clearly false. The Bevy performance improvements that I and the rest of the team landed in 0.16 speak for themselves [1]: 3x faster rendering on our test scenes and excellent performance compared to other popular engines. It may be true that little work is being done on rend3, but please don't claim that there isn't work being done in other parts of the ecosystem.

[1]: https://bevyengine.org/news/bevy-0-16/

> Do people actually travel between SF and LA?

Yes. The most obvious example of a company that's built around that corridor is Netflix: tech in the Bay Area, film production in L.A.

> I do not understand how people expect to write convincingly that tools that reliably turn slapdash prose into median-grade idiomatic working code "provide little value".

Honestly, I'm curious why your experience is so different from mine. Approximately 50% of the time for me, LLMs hallucinate APIs, which is deeply frustrating and sometimes costs me more time than it would have taken to just look up the API. I still use them regularly, and the net value they've imparted has been overall greater than zero, but in general, my experience has been decidedly mixed.

It might be simply that my code tends to be in specialized areas in which the LLM has little training data. Still, I get regular frustrating API hallucinations even in areas you'd think would be perfect use cases, like writing Blender plugins, where the documentation is poor (so the LLM has a relatively higher advantage over reading the documentation) and examples are plentiful.

Edit: Specifically, the frustrating pattern is: (1) the LLM produces some code that contains hallucinated APIs; (2) in order to test (or even compile) that code, I need to write some extra supporting code to integrate it into my project; (3) I discover that the APIs were hallucinated because the code doesn't work; (4) now I not only have to rewrite the LLM's code, but I also have to rewrite all the supporting code I wrote, because it was based around a pattern that didn't work. Overall, this adds up to more time than if I had just written the code from scratch.

> The OP cites these (https://github.com/devkitPro/libogc/blob/52c525a13fd1762c103... and https://github.com/atgreen/RTEMS/blob/2f200c7e642c214accb7cc...), but that's hardly a smoking gun. The function is trivial, just filling in some struct fields.

Yeah, I'm with Marcan 90% of the time, and in my view Marcan is more likely than not right that that function is derived from the RTEMS function, but in my view there's still reasonable doubt. That is to say, purely based on the evidence linked, I only agree that it's probable that the code is copied and disagree with Marcan's claim that it's "not possible" for the implementation to be non-infringing.

The fact that some of the identifiers are similar raises the biggest suspicions. "__lwp_stack_isenough" vs. "_Stack_Is_enough" is suspicious because I'd probably call that "sufficient_stack_available" or something like that. "LWP_STATES_DORMANT" vs. "STATES_DORMANT" is also suspicious because the normal OS term would be "sleeping". Still, the function logic is different enough that, even with that evidence, one could plausibly claim that only the headers or interface were copied and the implementation was clean-room, which is non-infringing per Oracle v. Google.

In US legal system terms, I'd say that a preponderance of the evidence shows that the code is a derivative work (i.e. it's more likely than not that the code was copied at some point), but that there's still reasonable doubt (i.e. a reasonable person could plausibly believe otherwise).

That statement seems extremely hyperbolic. Just today I tried automating some very pedestrian code for a system that wasn't particularly well-documented and ChatGPT 4o hallucinated the entire API. It was deeply frustrating and wasted more of my time than it would have taken to just slog through the documentation.

I won't deny that LLMs can be useful--I still use them--but in my experience an LLM's success rate in writing working code is somewhere around 50%. That leads to a productivity boost that, while not negative, isn't anywhere near the wild numbers that are bandied about.

I'm a Rust compiler author, and I'm fully in favor of "UB exploitation". In fact, LLVM should be doing more of it. LLVM shouldn't be holding back optimizations in memory-safe languages for edge cases that don't really matter in practice.

That example is an instance of unspecified vs. undefined behavior, but the correctness of the pointer provenance-based optimization example I gave doesn't depend on whether writing to an out-of-bounds pointer is unspecified or undefined.

Hence why I (and many other compiler developers) are inherently skeptical whenever anyone says "just stop exploiting undefined behavior".

"Exploiting undefined behavior" occurs when a simple semantics (however one defines "simple") results in behavior A, but the compiler chooses behavior B instead based on the actual, more complex, language semantics. The code snippet in question passes that test. If I flip the declaration order of i and arr, then I get this [1] at -O0 (the "simple" semantics):

        push    rbp
        mov     rbp, rsp
        mov     dword ptr [rbp - 4], 0
        mov     dword ptr [rbp - 4], 5
        mov     eax, dword ptr [rbp - 4]
        pop     rbp
        ret

Which indeed returns 5. But at -O2 clang optimizes it to this:

        xor     eax, eax
        ret

Which returns 0. So the simple semantics produces one result, and the complex semantics produces another. Hence, it's exploiting undefined behavior.

[1]: https://godbolt.org/z/df4dhzT5a

Note that if you change the code to this you have the same issue:

    int g(int n) {
        int arr[3], i = 0;
        arr[n] = 5;
        return i;
    }

Without "exploiting UB" it's incorrect to optimize this to "return 0", because of the possibility that i was allocated right after arr and n == 3.

Yeah, I would have liked to see the paper specify whether the LTO they tried is fat LTO or ThinLTO.

At FAANG scale the cost is prohibitive. Hence the investment in ThinLTO.

I notice that the paper doesn't claim to eliminate all reasoning about undefined behavior for optimizations. For example:

    int f() {
        int arr[3], i = 0;
        arr[3] = 5;
        return i;
    }

Optimizing this to "return 0" is relying on UB, because it's assuming that i wasn't laid out directly after arr in the stack frame. I believe this is what the paper calls "non-guardable UB".

I don't agree with the claim in the paper that their semantics offers a "flat memory model". A flat memory model would rule out the optimization above. Rather, the memory model still has the notion of object bounds; it's just simplified in some ways.

Nice to see this finally out! I wrote much of the GPU-driven rendering code for this release, which really helps Bevy's rendering performance with large scenes.

I'm surprised that the article didn't mention that Richard Garfield (creator of Magic) himself has a Ph.D. in combinatorial mathematics. I imagine that, if he's aware of this, that he'd get a kick out of it.

My feeling is that dependent types add complexity to a language that's already well-known for having a complex type system, so I'm nervous about blowing the complexity budget. But I'm bullish on tools like Verus, Prusti, and Creusot because they allow people who need to write low-level unsafe code to prove their code safe, while keeping the complexity of the safety proofs localized to that code, so most Rust programmers don't need to worry about it. This allows verification of Rust code without surfacing complexity to most developers. We can have our cake and eat it too.

You can mutate variables at compile time in D. See the compile time Newton's method example: https://tour.dlang.org/tour/en/gems/compile-time-function-ev...

> The important thing is that the feature performs the duty of those other features.

Zig's comptime doesn't do everything that Rust (or Java, or C#, or Swift, etc.) generics do, and I know you know this given your background in type theory. Zig doesn't allow for the inference and type-directed method resolution that Rust or the above languages do, because the "generics" that you create using Zig comptime aren't typechecked until they're instantiated. You can improve the error messages using "comptime if" or whatever Zig calls it (at the cost of a lot of ergonomics), but the compiler still can't reliably typecheck the bodies of generic functions before the compiler does the comptime evaluation.

Now I imagine you think that this feature doesn't matter, or at least doesn't matter enough to be worth the complexity it adds to the compiler. (I disagree, of course, because I find reliable IDE autocomplete and inline error messages to be enormously useful when writing generic Rust functions.) But that's the entire point: Zig comptime is not performing the duty of generics; it's approximating generics in a way that offers a tradeoff.

When I first looked at Zig comptime, it didn't evoke the "magic of Lisp" at all in me (and I do share an appreciation of simplicity in programming languages, though I feel like Scheme offers more of that than Lisp). Rather, my reaction was "oh, this is basically just what D does", having played with D a decent amount in years prior. Nothing I've seen in the intervening years has changed that impression. Zig's metaprogramming features are a spin on metaprogramming facilities that D thoroughly explored over a decade before Zig came on the scene.

Edit: Here's an experiment. Start with D and start removing features: GC, the class system, exceptions, etc. etc. Do you get to something that's more or less Zig modulo syntax? From what I can tell, you do. That's what I mean by "not revolutionary".

I tried ChatGPT 4o to help me reverse engineer some game code with the symbols missing and the results were quite disappointing. To say it had a tendency to hallucinate is an understatement. It didn't have any clue what was going on.

For me, those AI tools are much better at saving me time looking up documentation when doing simple things where it has examples of the exact code pattern I'm looking for in its training set. ChatGPT is great at writing one-off Blender scripts for me to give to artists, for instance.

The point is that comptime isn't dependent types at all. If your types can't depend on runtime values, they aren't dependent types. It's something more like kind polymorphism in GHC (except more dynamically typed), something which GHC explicitly calls out as not dependent types. (Also it's 12 years old [1]).

[1]: https://www.seas.upenn.edu/~sweirich/papers/fckinds.pdf

That's still just a function of type ∀K∀L.K → L with a bound on K. From a type theory perspective, a comptime argument, when the function is used in such a way as to return a type, is not a value, even though it looks like one. Rather, true or false in this context is a type. (Yes, really. This is a good example of why Zig reusing the keyword "comptime" obscures the semantics.) If comptime true or comptime false were actually values, then you could put runtime values in there too.

I think another way to put it is that the fact that Zig reuses the keyword "comptime" to denote type-level parameters and to denote compile-time evaluation doesn't mean that there's only one feature. There are still two features (templates and CTFE), just two features that happen to use the same keyword.

> This is one half of the innovation, dependent-types lite.

But that's not dependent types. Dependent types are types that depend on values. If all the arguments to a function are either types or values, then you don't have dependent types: you have kind polymorphism, as implemented for example in GHC extensions [1].

> The second half is how every other major feature is expressed _directly_ via comptime/partial evaluation, not even syntax sugar is necessary. Generic, macros, and conditional compilation are the three big ones.

I'd argue that not having syntactic sugar is pretty minor, but reasonable people can differ I suppose.

[1]: https://ghc.gitlab.haskell.org/ghc/doc/users_guide/exts/poly...

But Zig's comptime only approximates the features you mentioned; it doesn't fully implement them. Which is what the original article is saying. To use your analogy, using a touchscreen to eliminate a keyboard isn't very impressive if your touchscreen keyboard is missing keys.

If you say that incomplete implementations count, then I could argue that the C preprocessor subsumes generics/templates, interfaces/typeclasses†, macros, and conditional compilation.

†Exercise for the reader: build a generics system in the C preprocessor that #error's out if the wrong type is passed using the trick in [1].

[1]: https://stackoverflow.com/a/45450646