In reply to @roberthensing:matrix.org
do you also call nix_gc_now()?

In reply to @roberthensing:matrix.org
iirc the cache of parsed files is not released along with the eval state. This is usually not an issue, because it just dies with the process, but this could be improved

In reply to @roberthensing:matrix.org
iirc the cache of parsed files is not released along with the eval state. This is usually not an issue, because it just dies with the process, but this could be improved

This seems to be an omission in both the evaluator implementation and in the C API for it. A virtual destructor does exist for ExternalValueBase, but since Value relies entirely on GC and mkExternal doesn't register a finalizer, it seems that this destructor is never called.

It's worth noting that the GC used by Nix can not guarantee that finalizers are called in the first place, as it is a conservative GC, so it is best not to rely on finalization. (Also GC will not run before the process exits, unless e.g. it's the Nix CLI and NIX_SHOW_STATS=1)

Nonetheless, it should be added, because release a resource most of the time, or even sometimes, is better than never.

I'm a bit wary of adding it without ever calling it, because that will cause untested code to start running in the future, which could be a huge surprise when it starts causing crashes etc.
So either the whole thing needs to be implemented correctly and/or we add a test utility that forces the finalization, just for the purpose of unit testing.

also, I noticed the equality function does not take a value descriptor for other. I guess that's because most programs won't have multiple external types. I also happen to be using Rust, so ultimately my void* v will be something like dyn Any or a wrapper containing an Any; so naturally I can use that to check type identity. but it is also possible to load multiple plugins, right? what if several add external types? then I have no idea what the other value has for a type.

is the value descriptor checked for equality/identity first? that's the only reasonable way this API makes sense, and it's consistent with the documented behavior of not being equal if the other value is of a primitive type. but it's not documented that the equality does any such check before invoking the callback.