dcompute VS zig

Shameless plug: LDC (the LLVM based D compiler) can already target CUDA (and OpenCL) and wraps its API and all of the nasty details involved in replicating <<<>>> kernel launches with https://github.com/libmir/dcompute/ with a sane syntax that's type safe. LLVM handles the codegen, and all of the "magic" is done in the library.

As noted elsewhere it seems your experience is somewhat outdated: the releases of the LLVM D Compiler (one of the two compilers worth using for production builds, the other being GDC) are buffered to the bugs introduced in DMD (which is more stable than it used to be although there are still regressions), and there is a fork based GC available for linux, but as the GC will only ever trigger on allocation, don't use it and it won't collect.

> While C++ is not by any means a great meta-language, it's improved considerably since that time.

C++ has also painted itself into a corner multiple times too, which despite being technically an improvement over the status quo are lacking severely in their utility. C++ screwed up "constexpr if" big time by always introducing a scope (which costs you a pair of {}'s in the rare occasion you need one) which means you can't conditionally insert declarations (i.e. variables, structs/classes, functions).

> but beyond the novelty you'd hardly find a mature or reliable codebase written by a team of professionals using hacks like [string manipulation and mixins].

They are a wonderful hack when you need them and nothing else will do what you want. This is not unlike resorting to macros in C++, except that its hygienic, unlike macros.

I'm not claiming the project is mature and I'm only one person, but reliable definitely out there. The most heinous set of string mixins i've ever written[1] has definitely got to be the code for generating wrappers to call the OpenCL object property querying functions (clGetDeviceInfo & friends). You need to pass a size and a void pointer to the address of the return object that you have to call once, twice or more (depending on the type of the queried property) to figure out how much memory you need to allocate to call it again.

The important thing is that the interface[2] you use to drive this code generation is very clean and return on investment for getting the generic case correct is large.

[1]: https://github.com/libmir/dcompute/blob/master/source/dcompu...

There is an issue proposing this approach: https://github.com/ziglang/zig/issues/4284

https://github.com/ziglang/zig/issues/224

e.g.:

> > When debugging/prototyping, it's useful to comment out a line without having to refactor, e.g.

When writing code in a scripting language, sometimes you need that extra bit of performance (or maybe an async feature from Zig).

NodeJS is by no means a slow runtime, it wouldn’t be so popular if it was. But compared to Bun, it’s slow. Bun was built from the ground up with speed in mind, using both JavascriptCore and Zig. The Bun team spent an enormous amount of time and energy trying to make Bun fast, including lots of profiling, benchmarking, and optimizations.

ntdll.dll!RtlUserThreadStart()

There are valid reasons to use APIs from NTDLL. Where I disagree with zig#1840 is the idea that it is always better to use NTDLL versions of API. Every other software ecosystem uses the standard Win32 APIs and diverging from that without a good reason seems like a good way to have unexpected behavior. One concrete example is most users and programmers expect Windows to redirect some file system paths when running on WOW64. But this is implemented in Kernel32, not ntdll.

https://github.com/ziglang/zig/issues/11894

Zig also has an IRC channel on libera (#zig) that is moderated by Andrew Kelley.[1]

[1] https://github.com/ziglang/zig/wiki/Community