fix(runtime): register the per-thread globalThis cache as a GC root

[proggeramlug]

Problem

js_get_global_this caches this thread's globalThis in a thread-local raw-pointer cell (THREAD_GLOBAL_THIS) but — unlike its sibling js_module_top_this — never registers that cache slot as a GC root.

The authoritative GLOBAL_THIS_PTR static is scanned and relocated by scan_object_cache_roots_mut. So when a copying minor evacuates globalThis, that static and the object's overflow side-table are rekeyed to the moved address — but the thread-local cache still points at the stale from-space copy.

Subsequent js_get_global_this() calls hand back that dead pointer. Reading an overflow-stored builtin such as globalThis.Error then returns undefined, because the overflow entry was rekeyed onto the moved object. Downstream this surfaces as Class extends value is not a constructor in a program that constructs many classes during startup and happens to trigger a collection between two global reads (e.g. class X extends Error {} where globalThis.Error reads undefined).

Fix

Register the THREAD_GLOBAL_THIS cache slot as a mutable global root (via js_gc_register_global_root), exactly as js_module_top_this already does for its own thread-local cache. The collector then rewrites the raw pointer to the forwarding address on every move.

Raw-pointer global-root slots are already handled by mark_global_root_bits / rewrite_value_bits (the same path js_module_top_this relies on), and their rewrite-on-move is covered by test_rewrite_mutable_root_slots_updates_shadow_and_global_roots.

Validation

• Confirmed against a large esbuild-bundled CLI app whose class-heavy setup triggered a collection between two globalThis reads: before the fix, globalThis.Error read undefined and the app threw Class extends value is not a constructor; after the fix that throw is gone deterministically across repeated runs and the app proceeds further into its pipeline.
• The rewrite mechanism this fix uses is already unit-tested (test_rewrite_mutable_root_slots_updates_shadow_and_global_roots).

Summary by CodeRabbit

• Bug Fixes
  • Improved stability when accessing the app’s shared runtime state, reducing the risk of crashes or invalid data after memory management events.
  • Fixed an issue where thread-local cached state could become outdated during cleanup, helping ensure consistent behavior across requests.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 3a25b0fe-1f84-41b4-94a8-fbb456cfdefa

📥 Commits

Reviewing files that changed from the base of the PR and between 496620e and 4f91209.

📒 Files selected for processing (1)

• crates/perry-runtime/src/object/global_this/fetch_globals.rs

---

📝 Walkthrough

Walkthrough

In js_get_global_this(), the first-time allocation path for the per-thread globalThis now registers the THREAD_GLOBAL_THIS cache slot as a mutable GC global root, ensuring the collector rewrites this cached pointer during copying/evacuating collections instead of leaving it stale.

Changes

GC Root Registration Fix

Layer / File(s)	Summary
Register thread-local globalThis cache as GC root crates/perry-runtime/src/object/global_this/fetch_globals.rs	After allocating and caching new_ptr in THREAD_GLOBAL_THIS, the code now registers the cache slot address with the GC via runtime_write_barrier_root_heap_word and js_gc_register_global_root, so the pointer is updated on collection instead of becoming stale.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Related PRs: None specified.

Suggested labels: bug, gc, runtime

Suggested reviewers: None specified.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is concise and accurately names the main runtime GC fix.
Description check	✅ Passed	The PR explains the problem, fix, and validation, though it doesn’t follow the repo’s template sections.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}