fix(codegen): sloppy-mode class-field number stores keep the inline raw store

[proggeramlug]

Fixes #7288. 3762ms → 81ms outside a Perry checkout, matching the in-checkout number exactly.

Root cause: strict mode, from an upward package.json walk

perry_parser::file_is_in_esm_package_context walks up for the nearest package.json. "type": "module" ⇒ ES module ⇒ strict. Perry's root package.json is "type": "module", so everything inside the checkout compiles strict and everything outside compiles sloppy.

Three if !strict { return None; } bails in expr/proxy_reflect.rs (added by #6542) then discard the fast class-field-store path. The HIR diff between arms is one bit:

PutValueSet { … strict: true  }   ← in-checkout → class_field_set.fast → store double
PutValueSet { … strict: false }   ← outside     → put.dynic.* + js_put_value_set_dyn_ic per iteration

benchmarks/results/public-node-bun-v1.json reports the fast arm because compare.sh does cd benchmarks/suite. A user compiling the same file in their own project got the slow one.

6/6 controls, including the decisive one: inside the checkout with a nested {"name":"x"} → slow, because Node stops the walk at the nearest package scope. A bare {"name":"x"} with no perry key reproduces the flip exactly.

The fix

The bail discarded the fast arm in order to fix the fallback arm — but only the fallback needed strict-awareness, and js_put_value_set(…, strict=0) already existed.

The #5093 precheck already rejects OBJ_FLAG_FROZEN, OBJ_FLAG_HAS_DESCRIPTORS, mismatched class-id/keys, cleared layout-intact, and every non-plain-finite value — so a store reaching the raw slot could not have been rejected in either mode. No runtime change. Scope: raw-f64 fields, receiver == target.

Verification

• Independently re-measured after rebuild: outside 81–83ms, inside 81ms, fresh PERRY_CACHE_DIR on both, at load 25 (the 46× gap is far outside any noise band)
• 20-case sloppy differential — inheritance, shadowed fields, accessors, non-number stores, boundary numbers, preventExtensions, delete+re-add, aliasing, 100k megamorphic loop, Object.freeze mid-loop, enumeration order — byte-identical to Node 26.5.1 and to the pre-change compiler, with 150 class_field_sloppy_set blocks proving the arm is live
• 24/24 test_gap_class*/test_gap_object* match Node; --lib 633 pass; native_proof_regressions identical 4 pre-existing failures before and after
• New test verified non-vacuous (fails on origin/main)

Two things worth knowing

Today's two perf wins are unaffected — verified by IR diff, not timing: the only difference between arms for bench_bitwise and bench_array_ops is one startup-time js_register_closure_strict_function call; hot code byte-identical. #7416 and #7421 stand.

The gap suite structurally cannot exercise the sloppy arm. Every test-files/*.ts sits under the root "type": "module", so the entire corpus compiles strict. run_parity_tests.sh already acknowledges this for Node (retrying globals fixtures as .cts) but there is no sloppy arm on the compiler side — so any codegen predicate keyed on strict is tested in one state only. Worth its own issue.

Summary by CodeRabbit

• Performance

  • Improved performance for sloppy-mode assignments to eligible numeric class fields.
  • Optimized behavior is now consistent inside and outside ESM package scopes.

• Bug Fixes

  • Preserved correct silent-failure behavior when class-field assignment checks do not match.
  • Retained existing handling for strict mode and unsupported field types.

• Tests

  • Added regression coverage for optimized class-field writes and fallback behavior.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 3d4c16ee-349f-4a85-9c0f-5a4ae1c6ca9b

📥 Commits

Reviewing files that changed from the base of the PR and between 4143c2a and 8365ded.

📒 Files selected for processing (4)

• changelog.d/7288-path-dependent-class-field-store.md
• crates/perry-codegen/src/expr/property_set.rs
• crates/perry-codegen/src/expr/proxy_reflect.rs
• crates/perry-codegen/tests/native_proof_regressions.rs

---

📝 Walkthrough

Walkthrough

The compiler adds sloppy-mode lowering for eligible numeric class-field stores. It emits an inline guard and direct raw-slot write on hits, and calls js_put_value_set with strict = 0 on misses. Integration and regression coverage validate the new path.

Changes

Sloppy class-field store

Layer / File(s)	Summary
Raw-store eligibility and helper crates/perry-codegen/src/expr/property_set.rs	Adds try_lower_sloppy_class_field_raw_store with checks for module shape, computed properties, setters, class metadata, and raw-f64 fields.
Guarded lowering and integration crates/perry-codegen/src/expr/property_set.rs, crates/perry-codegen/src/expr/proxy_reflect.rs	Evaluates assignment operands in order, emits the class-field precheck and direct canonicalized slot store, and routes misses through js_put_value_set with strict = 0.
Regression coverage and changelog crates/perry-codegen/tests/native_proof_regressions.rs, changelog.d/7288-path-dependent-class-field-store.md	Tests sloppy hit and miss lowering and confirms strict-mode behavior remains unchanged. Documents scope and validation results.

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

Sequence Diagram(s)

sequenceDiagram
  participant PutValueSet
  participant ClassFieldLowering
  participant RawFieldSlot
  participant RuntimeSetter
  PutValueSet->>ClassFieldLowering: attempt eligible sloppy class-field store
  ClassFieldLowering->>RawFieldSlot: run inline guard and store canonicalized f64
  ClassFieldLowering->>RuntimeSetter: call js_put_value_set with strict = 0 on guard miss

Loading

Possibly related PRs

• PerryTS/perry#6811: Uses related PutValueSet and guarded raw/static object-write lowering paths.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/7288-path-dependent-codegen

---

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.}