GPU code can now use Rust's standard library. We share the implementation approach and what this unlocks for GPU programming.

This works because of our custom hostcall framework. A hostcall is analogous to a syscall. A hostcall is a structured request from GPU code to the host CPU to perform something it cannot execute itself. You can think of it like a remote procedure call from the GPU to the host to achieve syscall-like functionality.

To end users, Rust's std APIs appear unchanged and act as one would expect.

impl<T> Hash for UnconditionallyEqual<T> {
    fn hash<H: Hasher>(&self, _state: &mut H) {}
}

const hello_world_in_c =
    \\#include <stdio.h>
    \\
    \\int main(int argc, char **argv) {
    \\    printf("hello world\n");
    \\    return 0;
    \\}
;

    let hello_world_in_c = "\
    #include <stdio.h>\
    \n\
    \nint main(int argc, char **argv) {\
    \n    printf(\"hello world\\n\");\
    \n    return 0;\
    \n}";

    let hello_world_in_c = r#"#include <stdio.h>

int main(int argc, char **argv) {
    printf("hello world\n");
    return 0;
}"#;

macro_rules! text {
    (#[doc=$first_line:literal] $(#[doc=$lines:literal])*) => {
        concat!($first_line, $("\n", $lines),*)
    };
}
let hello_world_in_c = text! {
    ///#include <stdio.h>
    ///
    ///int main(int argc, char **argv) {
    ///    printf("hello world\n");
    ///    return 0;
    ///}
};

Revy is a proof-of-concept time-travel debugger for the Bevy game engine, built using Rerun.

Revy works by snapshotting diffs of the Bevy database every frame that are then logged into the Rerun database.
This allows you to inspect and visualize the state of the engine at any point in time, either in real-time or after the fact.
These recordings can then be shared to be replayed or e.g. attached to bug reports.

cargo tree --format='{p} {f}'

fn get<Q>(&self, k: &Q) -> Option<&V>
where
    K: Borrow<Q>,
    Q: Hash + Eq + ?Sized,

trait Borrow<Borrowed>
where
    Borrowed: ?Sized,
{
    fn borrow(&self) -> &Borrowed;
}

trait Equivalent<K>
where
    K: ?Sized,
{
    fn equivalent(&self, key: &K) -> bool;
}

Overall, I reduced the typical runtime on very large crates down to ~2s from ~8s - nearly an 80% speedup - without compromising performance on smaller crates. Along the way I reduced test time from ~7min to ~1min.

You can fill out the survey here.

Filling the survey should take you approximately 5 minutes, and the survey is fully anonymous. We will accept submissions until Friday, March 13th, 2026.

#[allow(dead_code)]
trait Foo {
    const FOO: u32;
}

impl Foo for u32 {
    const FOO: u32 = roundtrip(0);
}

// нет предупреждения о мёртвом коде
const fn roundtrip(x: u32) -> u32 {
    x
}

For example, part of one 2016 Advent of Code puzzle is looking for ABBA patterns: "two different characters followed by the reverse of that pair, such as xyyx or abba." If we assume only ASCII characters, that could be written by sweeping windows of the byte slice like this:

fn has_abba(s: &str) -> bool {
    s.as_bytes()
        .array_windows()
        .any(|[a1, b1, b2, a2]|
            (a1 != b1)
            && (a1 == a2)
            && (b1 == b2)
        )
}

let mut iter = "125 GOTO 10".chars().peekable();
let mut line_num = 0_u32;
while let Some(digit) = iter.next_if_map(|c| c.to_digit(10).ok_or(c)) {
    line_num = line_num * 10 + digit;
}
assert_eq!(line_num, 125);
assert_eq!(iter.collect::<String>(), " GOTO 10");

// \p{utf8} expands to:
// ([\x00-\x7F]
// | [\xC0-\xDF][\x80-\xBF]
// | [\xE0-\xEF][\x80-\xBF]{2}
// | [\xF0-\xF7][\x80-\xBF]{3})*

use resharp::Regex;

// basic matching
let re = Regex::new(r"hello.*world").unwrap();
assert!(re.is_match("hello beautiful world"));

// intersection: contains both "cat" and "dog", 5-15 chars
let re = Regex::new(r"_*cat_*&_*dog_*&_{5,15}").unwrap();

// complement: does not contain "1"
let re = Regex::new(r"~(_*1_*)").unwrap();

I've been writing proc macros for a while now. Derive macros for internal tools, attribute macros for instrumentation. And every time, the same two problems: quote! doesn't compose (you end up passing TokenStream fragments through five layers of helper functions and writing hundreds of let statements), and debugging generated code means cargo expand and then squinting at unformatted token output hoping something jumps out.

Because of this I ended up writing the same helper methods, composite AST parsing and tokenizing types, extractors etc. I would have to copy these from project to project as needed, and eventually just decided to publish a crate so I never have to do it again.

So I built zyn — a proc macro framework with a template language, composable components, and compile-time diagnostics.

I wrote the debug system after spending two days on a bug where a generated impl block was missing a lifetime bound. cargo expand spat out 400 lines of tokens and I couldn't find it, so I built a debug system.

use bnum::prelude::*;
// imports common use items
// including the `Integer` type and the `n!` macro

// say we want to write a polynomial function
// which takes any unsigned or signed integer
// of any bit width and with any overflow behaviour
// for example, the polynomial could be p(x) = 2x^3 + 3x^2 + 5x + 7

fn p<const S: bool, const N: usize, const B: usize, const OM: u8>(x: Integer<S, N, B, OM>) -> Integer<S, N, B, OM> {
    n!(2)*x.pow(3) + n!(3)*x.pow(2) + n!(5)*x + n!(7)
    // type inference means we don't need to specify the width of the integers in the n! macro
}

// 2*10^3 + 3*10^2 + 5*10 + 7 = 2357
assert_eq!(p(n!(10U256)), n!(2357));
// evaluates p(10) as a 256-bit unsigned integer

type U24w = t!(U24w);
// 24-bit unsigned integer with wrapping arithmetic
type I1044s = t!(I1044s);
// 1044-bit signed integer with saturating arithmetic
type U753p = t!(U753p);
// 753-bit unsigned integer that panics on arithmetic overflow

let a = p(U24w::MAX); // result wraps around and doesn't panic
let b = p(I1044s::MAX); // result is too large to be represented by the type, so saturates to I044s::MAX
// let c = p(U753p::MAX); // this would result in panic due to overflow