RFC: api redesign prototype and notes based on gc-arena by shruti2522 · Pull Request #68 · boa-dev/oscars

shruti2522 · 2026-03-27T12:25:41Z

functional prototype and notes for a new GC API based on the requiremnts fo an ideal Gc mentioned in boa#2631. It explores moving from the current implicit rooting behavior to zero overhead, lifetime branded model inspired by gc-arena.

prototype (oscars/examples/api_prototype): standalone implementation demonstrating the core design (Gc, Root, MutationContext) with examples

… feasibilty of context bound pointer lifetimes

nekevss · 2026-03-27T22:37:59Z

+    }
+
+    pub(crate) fn alloc<'gc, T: Trace + Finalize + 'gc>(&'gc self, value: T) -> Gc<'gc, T> {
+        let boxed = Box::new(GcBox {


I'm highly skeptical of a box usage here. We're just going to end up with the same fragmentation issue we have with the current implementation.

removed box usage here, used PoolAllocator instead

hansl · 2026-03-27T23:18:32Z

+| **Pointer** | `Gc<T>` | `Gc<'gc, T>` |
+| **Lifetime** | `'static` + `extend_lifetime()` | `'gc` branded |
+| **Rooting** | Implicit (inc/dec on clone/drop) | Explicit (`Root<T>`) |
+| **Copy cost** | Cell write | Zero |


Also the drop cost is a required access to a TLS which uses a futex lock in most implementations. It's very expensive for the current GC. It would be free for the proposed one.

Thanks, completely missed that. Added TLS futex drop cost to the table

shruti2522 · 2026-03-28T04:22:57Z

Replaced Vec root tracking with intrusive linked list for O(1) root removal, switched from Box allocation to PoolAllocator, and added compile fail tests to check that the 'gc lifetime correctly prevents pointers from escaping the mutation context.

I had kept the initial implementation simple since I wanted to focus the discussion on the API design first, but you're right that it should reflect the actual target implementation more closely. Hope this looks better

jedel1043 · 2026-03-28T06:04:47Z

+    collector_roots: Rc<RefCell<RootList>>,
+    node: RootListNode,  // Intrusive list node


Slight tweak here. You don't need a collector_roots field, you just need to unlink the previous and next elements of RootListNode from the current node, then link them together.

To collect the roots the GC can just take the base root, the go through the linked list to get all the inner Gcs

Maybe the intrusive_collections crate could be a good inspiration for this

Used intrusive_collections as the inspiration.RootLink holds only Cell<Option<NonNull<RootLink>>> for prev/next, with gc_ptr recovered via offset_of!

jedel1043 · 2026-03-28T10:06:34Z

+        // `T: Sized` ensures `gc_ptr` is always a thin (one-word) pointer, so
+        // `link` sits at the same byte offset in `Root<T>` for every sized `T`.


You could also put the link at the start. right? Such that it doesn't matter if T is sized

Done, though T: Sized is still required to keep gc_ptr as a thin pointer for sound type erased *mut u8 collector reads. Removing it would need a second NonNull<GcBox<()>> field for the GC to read, with Root::get() using the fat pointer. should I do that?

Well I guess it kinda doesn't matter, because in our current GC we do some vtable hacks to erase pointers in a slightly safer way, so we also require T:Sized

Makes sense, I'll leave it as is then

Actually that's false. What we did in our current GC was to introduce the concept of a GcErasedPtr, which is just a NonNull<GcBox<NonTraceable>>, and because the GcBox has a VTable of every T, you don't need to ensure T: Sized

We don't need to introduce that here though, since it's just a prototype. We can address these limitations after iterating on this

jedel1043

The new API looks great! Nice job!

We might have some ways to statically ensure Root is not used with a different realm, like relating them to the GcContext itself by some invariant lifetime, but that's a future improvement that we can do while iterating on this.

shruti2522 · 2026-03-29T15:26:23Z

Thanks! Sounds good, I'll keep those improvements in mind for future iterations.

nekevss

Looks good overall and is super interesting

shruti2522 force-pushed the api-redesign-prototype branch 2 times, most recently from c211cae to fed14b3 Compare March 27, 2026 12:31

shruti2522 changed the title ~~add api redesign prototype and notes based on gc-arena~~ RFC: api redesign prototype and notes based on gc-arena Mar 27, 2026

shruti2522 force-pushed the api-redesign-prototype branch from fed14b3 to 8a0ef34 Compare March 27, 2026 13:22

RFC:add api redesign prototype based on gc-arena

cf29ff4

shruti2522 force-pushed the api-redesign-prototype branch from 8a0ef34 to cf29ff4 Compare March 27, 2026 13:26

refactor: split api_prototype.rs, added Array.prototype.push to prove…

f661622

… feasibilty of context bound pointer lifetimes

shruti2522 force-pushed the api-redesign-prototype branch from eb19a58 to f661622 Compare March 27, 2026 15:23

test: verify memory safety with Miri

2dca871

jedel1043 reviewed Mar 27, 2026

View reviewed changes

Comment thread notes/api-redesign-prototype/prototype_findings.md Outdated

jedel1043 reviewed Mar 27, 2026

View reviewed changes

Comment thread oscars/examples/api_prototype/main.rs

nekevss reviewed Mar 27, 2026

View reviewed changes

jedel1043 reviewed Mar 27, 2026

View reviewed changes

Comment thread notes/api-redesign-prototype/prototype_findings.md Outdated

hansl reviewed Mar 27, 2026

View reviewed changes

add O(1) pinned roots, pool allocation and lifetime safety tests

142d613

shruti2522 force-pushed the api-redesign-prototype branch from 61bed6b to 142d613 Compare March 28, 2026 04:18

jedel1043 reviewed Mar 28, 2026

View reviewed changes

feat(gc): implement intrusive_collections inspired RootLink

8310612

shruti2522 force-pushed the api-redesign-prototype branch from eb0b461 to 8310612 Compare March 28, 2026 08:58

jedel1043 reviewed Mar 28, 2026

View reviewed changes

place link before gc_ptr in Root<T>

bcde269

jedel1043 approved these changes Mar 29, 2026

View reviewed changes

nekevss approved these changes Apr 4, 2026

View reviewed changes

nekevss merged commit 819cc7e into boa-dev:main Apr 4, 2026
4 checks passed

jedel1043 mentioned this pull request Apr 4, 2026

docs(gc): add precise-tracing API redesign plan #55

Closed

shruti2522 mentioned this pull request Apr 5, 2026

prototype: enforce single-realm invariants for roots and weak pointers #73

Merged

Flamki mentioned this pull request Apr 5, 2026

prototype(gc): explore static root/context binding with scoped roots #74

Closed

		collector_roots: Rc<RefCell<RootList>>,
		node: RootListNode, // Intrusive list node

		// `T: Sized` ensures `gc_ptr` is always a thin (one-word) pointer, so
		// `link` sits at the same byte offset in `Root<T>` for every sized `T`.

Conversation

shruti2522 commented Mar 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

shruti2522 commented Mar 28, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jedel1043 left a comment

Choose a reason for hiding this comment

Uh oh!

shruti2522 commented Mar 29, 2026

Uh oh!

nekevss left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

shruti2522 commented Mar 27, 2026 •

edited

Loading