fix(gc): route the census thread-locals through hot TLS

[proggeramlug]

The break

scripts/check_thread_locals.py fails on main, which takes the self-test-checkers
job of the TLS Budget workflow down — on main itself (run 33945210479) and on every open
PR, including unrelated ones:

thread-local policy check FAILED:

  crates/perry-runtime/src/gc/census.rs: 2 raw `thread_local!` block(s), none allowed.
    Use `crate::perry_thread_local!` — same syntax, same `.with()` at every call site,
    and the address lands in this thread's hot cache instead of costing a
    `_tlv_get_addr` call (#7469).

In that workflow the expensive tls-budget job passes (13m19s); only this 12-second
hermetic checker fails.

Why it is a real finding, not just a ratchet

gc::census is compiled unconditionally — crates/perry-runtime/src/gc/mod.rs:247 declares
pub(crate) mod census; with no cfg — so both blocks are in shipping builds and pay
_tlv_get_addr per access on Darwin, which is exactly the cost #7469's policy exists to
prevent. (The census is runtime-gated by PERRY_GC_CENSUS, but that gate is read through
ARMED, one of these very thread-locals.)

The file's third thread_local! is #[cfg(test)] and is out of scope by construction, as
the checker's own header documents — hence "2 raw blocks", not 3.

This is the same treatment 12efed122 ("perf(runtime): route four recent thread_local
declarations through hot TLS") gave four other recent declarations; census.rs was missed.

The change

Both blocks become crate::perry_thread_local!. The macro takes the same syntax — multiple
statics per block, doc comments, const { … } initialisers — and yields the same .with()
API. All nine call sites of ARMED / SEQ / LABEL / PASS1_MARKED already go through
.with() (verified: zero non-.with() uses), so nothing else changes.

Verification

Checker, run locally against the tree:

before:  thread-local policy check FAILED: ... 2 raw `thread_local!` block(s)   rc=1
after:   thread-local policy OK: 304 hot declarations, 122 raw blocks in
         87 recorded cold files, capacity 768                                   rc=0

Compile, on linux-x86_64 — cargo check -p perry-runtime --all-targets:

ASSERT converted: 2 (expect 2)
RC=0
real check of perry-runtime: 1        # a genuine `Checking perry-runtime v…`, not a cache hit
census.rs diagnostics: 0
Finished `dev` profile in 30.68s

Note: with RUSTFLAGS=-D warnings this branch does not build, but for an unrelated
reason — the three clashing_extern_declarations errors fixed by #9776, none of which
come from census.rs. That is why the compile check above is run without -D warnings;
once #9776 lands, both hold together.

Scope

One of several independent breakages on main, which is currently red across six workflows.
Also open from this sweep: #9776 (warnings), #9777 (cargo-test), #9780
(check + ext-link).

https://claude.ai/code/session_01YPfnmWZmSpSWpmnoXvH8z2

Summary by CodeRabbit

• Performance

  • Reduced runtime overhead for garbage-collection census tracking, particularly on Apple platforms.

• Reliability

  • Restored automated checks that help detect thread-local storage regressions before release.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Team

Run ID: cc6681ff-9db8-49c0-a048-ea9467f859e2

📥 Commits

Reviewing files that changed from the base of the PR and between 12efed1 and 4367aad.

📒 Files selected for processing (2)

• changelog.d/9781-census-hot-tls.md
• crates/perry-runtime/src/gc/census.rs

---

📝 Walkthrough

Walkthrough

The census module replaces two standard thread-local declarations with crate::perry_thread_local!. The changelog records the Darwin _tlv_get_addr optimization and the restored self-test-checkers job.

Changes

Census thread-local declarations

Layer / File(s)	Summary
Replace census thread-local macros crates/perry-runtime/src/gc/census.rs, changelog.d/9781-census-hot-tls.md	The ARMED, SEQ, LABEL, and PASS1_MARKED declarations now use crate::perry_thread_local!. The changelog documents the optimization and restored checker job.

Estimated code review effort: 1 (Trivial) | ~5 minutes

✨ 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.}

[proggeramlug]

Closing as a duplicate of #9774, which fixes the same gc/census.rs raw thread_local! breakage and does it more completely — converting the blocks makes gc_runtime_root_holders.py enumerate them for the first time, and #9774 gives PASS1_MARKED an honest verdict (untraced_in_nonmoving_window) plus a gate that verifies the window against the holder's own source. My independent reproduction agrees with the diagnosis, so nothing is lost here.

https://claude.ai/code/session_01YPfnmWZmSpSWpmnoXvH8z2