| 21 Nov 2025 |
 Sergei Zimmerman (xokdvium) | I did research a bit the distribution of list sizes when evaling nixpkgs. Small lists are quite often used, but lists just pale in comparison with attrsets overall | 21:17:12 |
 mzero | name value pairs are so common because of listToAttrs, etc, I wonder if a pair type would actually make a significant perf difference | 21:19:00 |
| Sergei Zimmerman (xokdvium) | A more generic approach could be to have "attrset shapes". Very similar to how v8 optimizes classes with the same layout | 21:20:09 |
| Sergei Zimmerman (xokdvium) | Then names wouldn't need to be stored in the attribute set and function calls could reuse the same memory and not copy anything to a new Env.
(I and Robert Hensing discussed this a bit at some point) | 21:21:57 |
 Qyriad | In reply to @xokdvium:matrix.org
A more generic approach could be to have "attrset shapes". Very similar to how v8 optimizes classes with the same layout benefits of a JIT especially | 21:22:04 |
| Sergei Zimmerman (xokdvium) | Not exactly tho. jit/memory layout are tangential somewhat | 21:22:32 |
| Qyriad | yeah but you can optimize hot shapes | 21:22:56 |
| Sergei Zimmerman (xokdvium) | Yeah. I think python has slots for this kind of thing | 21:23:13 |
| Sergei Zimmerman (xokdvium) | * Yeah. I think python has __slots__ for this kind of thing | 21:23:28 |
| mzero | do you have something I could read about that? I'm curious how it would work. | 21:25:02 |
| Sergei Zimmerman (xokdvium) | Probably staring at prior art in V8? I haven't done that myself, but I'd be curious how they do tracing to find hot shapes | 21:26:04 |
| Sergei Zimmerman (xokdvium) | All of those optimizations are pretty inaccessible without tracing infra :( | 21:27:26 |
| Sergei Zimmerman (xokdvium) | But tracing attrset shapes doesn't exactly depend on function call tracing. Computing a shape fingerprint and having a small hashmap of that could give a rough picture of the frequency. | 21:29:23 |
