perf(gc): layout side tables are 34% of object construction — the construction/death half of #5094 (allocation is 7.7%)

[proggeramlug]

Summary

Granular member of the #5094 family. #5094's table covers the access paths
(this.field get/set, array element writes, dynamic property writes). This ticket is
the construction and death path, which that table does not mention and which is now
the single largest cost in allocation-heavy code.

On gc-handoff/bench/churn_alloc.ts — 20M {v: number, w: number} literals pushed into
a 1000-element array — a symbolicated profile attributes 33.6% of self time to
gc::layout per-object side-table maintenance, against 7.7% for the actual
allocation. layout_forget_object alone is 14.5%, nearly twice the allocator.

#6893 closed the memory half of this subsystem (per-class layout interning took the
record footprint 289 B → 133 B). The CPU half was never touched.

Evidence

Symbolicated profile (PERRY_DEBUG_SYMBOLS=1, sample 2 s, 1500 leaf samples, quiet M1
mini at load 1.4):

group	share
gc::layout side tables	33.6%
— layout_forget_object (gc/layout.rs:336)	14.5%
— layout_note_slot (gc/layout.rs:685)	7.9%
— js_gc_init_typed_shape_layout (gc/layout.rs:962)	7.7%
— shape_install_*	2.2%
— LayoutSlot*	1.3%
_tlv_get_addr	17.0%
write barriers (see sibling ticket)	16.1%
typed-feedback guards	9.2%
array (js_array_length, js_array_note_numeric_write)	6.2%
the allocation itself (js_object_alloc_class_inline_keys 6.3%, arena_alloc* 1.4%)	7.7%
user code	3.7%

Note the interaction with the TLS row: TYPED_LAYOUTS / LAYOUT_SLOT_MASKS are
thread-locals, so an unknown but material share of the 17% _tlv_get_addr is also this
subsystem. #7474 already cached the hot TLS addresses and TLS still costs 17%; removing
the layout ops removes their share of it too. One fix, two rows.

js_gc_init_typed_shape_layout is emitted by lower_call/new.rs on every
construction, and layout_forget_object runs on every object death — so the cost
scales with allocation rate, not with live set. That is why it dominates churn and not
the retain/tree benchmarks.

Workload decomposition (quiet mini, best-of-3)

variant	Perry	node	ratio
churn (full)	2.72 s	0.17 s	16.0×
churn_alloc — object literal + push	2.44 s	0.14 s	17.4×
churn_read — element reads only	0.35 s	0.08 s	4.3×
push_num — numbers into array, no object	0.30 s	0.11 s	2.7×

push_num at 2.7× shows the array machinery is not the problem. Subtracting it, object
construction is ~2.14 s, 79% of churn, and ~76% of that is bookkeeping rather than
allocation.

Repro

cd gc-handoff/bench
export PERRY_RUNTIME_DIR=<repo>/target/release
PERRY_DEBUG_SYMBOLS=1 PERRY_NO_AUTO_OPTIMIZE=1 <repo>/target/release/perry \
    churn_alloc.ts -o sym_churn_alloc     # PERRY_DEBUG_SYMBOLS=1 is what makes the profile readable

./sym_churn_alloc >/dev/null & P=$!; sleep 0.7
sample $P 2 -mayDie -f /tmp/sym.sample; wait $P
# read the "Sort by top of stack" section; demangle with rustfilt or read the raw _RNv… names

churn_alloc.ts is in gc-handoff/bench/ alongside push_num.ts, push_cls.ts and
churn_read.ts (the decomposition set).

Task

Apply #5094's "layout is canonical" design to the construction/death path specifically:

1. Construction — js_gc_init_typed_shape_layout per new should become a header
   bit-set, not a side-table insert. The mask is already a compile-time constant per class
   (codegen emits perry_typed_shape_raw_f64_mask_<class>); TYPED_LAYOUTS stores per-class-constant layout masks per OBJECT — O(objects) memory (272MB on churn bench) + hashmap insert on every new #6893 established the
   per-class interning this can hang off.
2. Death — layout_forget_object should be free when the object never diverged from
   its canonical shape. Today it runs unconditionally per dead object; with a canonical
   bit it becomes a bit-test that is almost always false. gc/layout.rs:938 already notes
   the map is usually empty and skips the hash — but the call, the TLS resolve and the
   branch still happen per object.
3. Slot notes — layout_note_slot on stores into a slot that still matches the
   canonical mask should compile away entirely.

The fallback path (object diverged → today's TYPED_LAYOUTS behaviour) stays exactly as
it is. This is about not paying for it on the overwhelmingly common case.

Acceptance criteria

1. gc::layout symbols fall below 8% of self time on the churn_alloc.ts profile
   (from 33.6%).
2. churn_alloc.ts improves by ≥1.5× (2.44 s → ≤1.6 s) on a quiet host.
3. churn.ts improves correspondingly; ratio vs Node drops from 16.0× to ≤11×.
4. No GC regressions: PERRY_GC_TRACE=1 on churn still shows ~105 cycles, ~0.004 GB
   copied, positive reclamation every cycle, max pause in the low milliseconds; tree.ts
   stays at ~43 cycles / ~0.017 GB.
5. No regression on the perf(GC): make per-object layout O(1)-loadable — kill per-operation thread-local layout tracking (umbrella: method_calls/array-downgrade/object-property) #5094 access-path benchmarks (method_calls,
   bench_numeric_array_downgrade, bench_object_property) — this change should help
   them, but confirm.
6. The 8 gc_ratchet probes plus 12_large_live_set hold.
7. cargo test workspace sweep green (exclude cross-host UI crates on macOS).

Traps

• Do not benchmark on the dev Mac while builds run — it sits at load 15–140 from other
  agents. Use the dedicated quiet host: ssh perry@perry-macos.local (M1, 8 cores, idles
  ~1.5). Perry binaries are statically linked arm64 — just ship them over
  (tar czf - n_* | ssh perry@… 'tar xzf - -C ~/benchX'), no toolchain needed there.
  Fallback on a loaded host: best-of-N user CPU time, which tracks the quiet host's
  wall clock within ~5%. Peak RSS and PERRY_GC_TRACE are load-independent.
• Default perry output is stripped; PERRY_DEBUG_SYMBOLS=1 is required or the profile
  is all ???.
• Rebuild runtime and stdlib — perry-runtime is rlib-only, the .a comes from the
  -static wrappers.
• PERRY_NO_AUTO_OPTIMIZE=1 on ad-hoc compiles; rm -rf node_modules/.cache/perry after
  switching compilers.
• Never CARGO_PROFILE_RELEASE_CODEGEN_UNITS=16 for anything measured — it miscompiles
  the release runtime.
• PERRY_WRITE_BARRIERS=0 makes this benchmark slower (2.44 → 5.21 s) because it also
  switches the collector out of evacuating mode. It cannot be used to isolate costs.

Context

Parent: #5094 (umbrella — access paths). Sibling tickets from the same profile: #7511 (write barriers on
non-pointer stores, 16.1%) and #7512 (new Klass() slower than the equivalent object
literal). Predecessor: #6893 (memory half of this subsystem, closed). Related: #7469
(the remaining _tlv_get_addr share), #6759 (V8-style object model).

[proggeramlug]

Folding in the residual half of #7512, per the diagnosis there (#7515 merged the other half).

js_gc_init_typed_shape_layout is emitted AFTER the constructor call (lower_call/new.rs; the ordering is visible in a --trace llvm dump). The consequence is that no raw-f64-declared class-field store inside any constructor can pass its guard — neither the inline and i16 %r, 4096 INTACT test nor class_field_fast_contract's layout_typed_raw_f64_slot_for_user. Every such store falls back to js_put_value_set.

The user-visible shape of this is worth stating plainly, because it inverts the usual assumption:

Declaring the fields number is what makes the class slower. More type information selects a representation whose guard the construction path has made unsatisfiable.

That is the #6377 lesson in a new costume — a "more type visibility" change un-gating a path that cannot actually fire — and it is why #7512's remaining cost belongs here rather than in a repsel ticket.

The obvious fix does not work. Moving the call before the ctor fails because init_typed_shape_layout (gc/layout.rs:897) validates that each raw-f64 slot holds a plain double, and fresh slots hold undefined. So this needs the construction path to establish the layout in a way that tolerates the pre-initialization state — i.e. it is the construction-path item of this issue, not a one-line reorder.

Concrete task wording for whoever picks this up: make the typed-shape layout observable to stores inside the constructor body, without asserting slot contents that the prologue has not written yet. Acceptance is an IR census showing the two real field stores in constructor(v, w) { this.v = v; this.w = w } with v: number; w: number no longer routing through js_put_value_set.

Cross-refs: #7512 (diagnosis + census), #7515 (the dead-store half, merged), #7511 is not implicated — the object literal emits more barrier work per store and is still the faster form.

[proggeramlug]

Item 1 (construction) is up as #7532, together with the residual half of #7512 folded in above. 1.553× on push_cls on the pinned host, and the collector result is strictly better rather than merely equal — promotion 210,488 → 64 bytes, peak RSS 29.6 → 24.1 MB, cycles unchanged. Those are exactly the numbers the equivalent object literal already had, so #7512's inversion is closed on both axes.

The mechanism, since it is not the obvious one: js_gc_declare_typed_shape_layout skips the slot validation that makes the existing call impossible to move earlier, and codegen carries the proof instead — every raw-f64 field prologue-assigned from a plain parameter (#7486's predicate), and an empty pointer mask so the declared state is POINTER_FREE, byte-identical to what birth already sets. A mistyped store still downgrades through the ordinary guard → boxed setter → layout_note_slot path; there is a 20,000-instance witness under GC zeal.

Two findings for this ticket, both of which change what is left to do.

1. churn_alloc cannot be fixed by item 1, because object literals carry no types. HIR rewrites a closed-shape literal to new __AnonShape_<hash>(…) whose synthesized constructor is a qualifying prologue — but the minted class's field types come out as Any, not Number. So there was never a raw-f64 store path there to unlock. The corollary is worse than the missed optimisation: since Any is pointer-bearing, {v: number, w: number} is currently declared to the collector as two pointer slots, and js_gc_init_typed_shape_layout spends its 13.6% installing that. The literal path's construction cost is a type-propagation gap in mint_anon_shape_class's field_shapes, not an ordering one. Worth its own ticket; I have not filed one.

2. This ticket's acceptance criteria were written against a stale profile and should be restated. Measured on main at d3472a529, layout_forget_object is 3.0% of churn_alloc, not the 14.5% quoted in the description — the #7469/#7474/#7486/#7487/#7501 wave already took it. Removing the side tables entirely could never have delivered ≥1.5× on churn_alloc, so criteria 1–3 were unreachable from items 1–3 as scoped. The current leaf profile is:

symbol	share
_tlv_get_addr	17.8%
js_gc_init_typed_shape_layout	13.6%
barrier_child_prologue	9.3%
js_typed_feedback_class_field_set_guard	7.5%
js_object_alloc_class_inline_keys	6.6%
js_array_length	6.4%
layout_note_slot	6.2%
user code	4.0%
layout_forget_object	3.0%
shape_install_shared	2.5%

Remaining on this ticket: item 3 (layout_note_slot compiling away on a store that still matches the canonical mask, ~6%). Item 2 shipped in #7525, item 1 in #7532.

Not yet verified by me, and I would not claim them: criterion 5 (the #5094 access-path benchmarks — I could not find method_calls / bench_numeric_array_downgrade / bench_object_property under those names in benchmarks/), criterion 6 (the 8 gc_ratchet probes plus 12_large_live_set), and criterion 7 (the workspace cargo test sweep — I ran perry-runtime and perry-codegen only, the latter with a failure set identical to origin/main's 22 pre-existing).