notramo (Oct 12 2026 at 13:43): notramo (Oct 12 2026 at 13:43):I found this interesting blog post by the author of the Cosmopolitan libc:
notramo (Oct 12 2026 at 13:44):I wonder if this idea could be implemented into Roc, to create cross-platform binaries.
Richard Feldman (Oct 12 2026 at 14:15):I think so! I haven't looked deeply into it but I'm pretty sure we could do it. Seems like a cool idea
notramo (Oct 12 2026 at 17:49):@Richard Feldman, should I open an issue for tracking?
Richard Feldman (Oct 12 2026 at 17:53):I'd rather wait until someone tries actually building something with it :smile:
Richard Feldman (Oct 12 2026 at 17:53):I don't think we need an issue for it
Richard Feldman (Oct 12 2026 at 17:53):if anyone wants to try and gets blocked by Roc compiler blockers, lmk!
notramo (Oct 12 2026 at 17:54):So, it would not need anything from Roc, just create a C platform and use Cosmopolitan libc?
Richard Feldman (Oct 12 2026 at 17:58):it's been years since I looked into it, but my vague memory was that you needed to make the output binary in a specific way so that each of the OSes look at the same bytes and think "this is in the right format for me"
Richard Feldman (Oct 12 2026 at 18:00):if that's correct, I suspect we might need a --target=x64ape or similar, but I'm not sure if llvm supports that output format. I seem to recall they have a reference c compiler, which might use llvm but I'm not sure haha
Richard Feldman (Oct 12 2026 at 18:02):there's also a question of whether lld knows how to link that format
Richard Feldman (Oct 12 2026 at 18:03):part of what I think needs a proof-of-concept is building an APE host and then linking it together with a fake "Roc program compiled to APE" and making sure it all comes together
Richard Feldman (Oct 12 2026 at 18:05):using the same linking settings we always use
Richard Feldman (Oct 12 2026 at 18:06):getting all the crti.o and similar stuff figured out would likely be tricky there
notramo (Oct 12 2026 at 19:18):I prompted an LLM agent to do some research, and it said the task is even more complex than what you described here.
notramo (Oct 12 2026 at 19:20):So it would not worth the time and effort with the current state of the compiler. Maybe after 1.0 or later?
Peter Marreck (Oct 13 2026 at 15:01):@notramo
I like APE, but I found out that it has a few warts- for example, you cannot build for x86_64 from macOS aarch64, you need to cross-compile to aarch64 from x86_64 Linux.
And there's an issue with a fundamental thing that makes it work, IIRC (which might actually be related to the above)- Justine literally rewrote libc (cosmopolitan-libc) to make it all work. And... That may cause OTHER issues.
It's quite an attractive idea, though- a true universal binary!
Ah, another thing I remembered- last I checked, the Mac builds it made were x86_64 only and not aarch64, and that won't be possible to run on macOS soon
Peter Marreck (Oct 14 2026 at 00:15):uh, I got it to work. Across roc, zig and rust. :laughing:
A single executable artifact for everything. Basically... Better than APE, @notramo :sweat_smile:
Catch: On some platforms, it writes a new executable to a cache location and that's what it actually runs, since only Linux can basically run a binary from a subsection of another binary.
Peter Marreck (Oct 14 2026 at 00:18):I'm testing it out on my "pinterm" roc terminal game thing from the "show and tell" forum
notramo (Oct 14 2026 at 10:57):So, you compile normal Roc targets, then bundle them into an archive-like file with an external tool, which simply unpacks them when it's started?
Peter Marreck (Oct 14 2026 at 12:06):oh boy. you'll see in a few. it's... good. ;) Doesn't require APE anymore, either, but uses a technique inspired by it (but better... this one works on all the BSD's as well)
Last updated: Oct 15 2026 at 17:56 UTC