perf(codegen): route % on i64-range integer locals to srem, not fmod

[proggeramlug]

Closes #7404. bench_bitwise: 20.4× slower than Node → 1.39×. 55108ms → 3763ms, a 14.6× improvement, with the Node-verified CHECKSUM:525000000 unchanged.

The bug

frem is not an aarch64 instruction — it lowers to an fmod library call. The hot loop emitted four per iteration (1754 profile samples in fmod).

A guarded srem fast path already existed. Every gate in front of it asked the wrong question: is_integer_valued_expr resolves a LocalGet through integer_locals, which is an i32-range property that also gates i32 shadow slots. Widening that would have put an i32-overflowing value into an i32 slot — a miscompile. The % path converts to i64 and only needs integer-valued-within-i64.

collect_int_valued_i64_locals supplies that weaker property as a magnitude lattice, read only by the % gate. integer_locals is untouched.

Three holes closed, each silently-wrong arithmetic

hole	why it mattered
Mul blowup	a pure integrality predicate let (a*b*c) % n reach fptosi … to i64 with a 93-bit product — poison. The lattice gates at 2^62, which also closes the same pre-existing hole for i32 locals
Zero divisor	an early version routed 1000 % d where d decrements through zero. srem(x,0) is UB; JS needs NaN. Divisor now restricted to a non-zero integer literal
The saturation argument	"±constant can't leave i64 in finite time" is false — a = a + 1e18 escapes in ~10 iterations. Replaced with an IEEE-754 bound: once ulp(v) ≥ 4D, v ± d rounds back to v exactly, so |L| ≤ 2^(55+log₂D) forever

That third one was my error in the task brief, caught and corrected during implementation.

The gate that actually decides this

Not binary.rs:588. There are six frem emission sites, and the owner here is expr/mod.rs::lower_numeric_binary_value, which intercepts numeric binary ops before binary::lower and only handed Mod off when the dividend had an i32 counter slot. Widening the predicate alone changed nothing — the IR acceptance test caught that and stayed red.

Verification

• IR: hot fn frem 4 → 0, srem 0 → 4 (baseline had zero srem, so the check could genuinely fail)
• Correctness: edge-case differential incl. both -0 cases via Object.is (not ===, which cannot distinguish them) — byte-identical to Node 26.5.1
• Oracle: CHECKSUM:525000000 on all runs
• Speed: min 55108 → 3763; Node 2708. Load 1.95–2.8
• Regressions: cargo test -p perry-codegen --lib 632 passed / 0 failed; 26/27 math/number/int gap tests, the one failure (process.setTraceSigInt arg validation) pre-existing and byte-identical between arms
• cargo fmt --check clean; file-size cap and addr-class ratchet pass

No env knob added, per CLAUDE.md's kill-policy.

Summary by CodeRabbit

• Bug Fixes

  • Corrected JavaScript remainder (%) calculations for integer-valued values outside the 32-bit range.
  • Prevented unsafe integer remainder optimizations when the divisor may be zero.
  • Improved numerical correctness while preserving optimized execution for proven-safe integer cases.

• Tests

  • Added regression coverage for large integer locals and potentially zero divisors.
  • Refined garbage-collection evacuation checks to rely on liveness results without requiring diagnostic output.

[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: 38ef5e8a-6d8d-4fda-a3cd-c0e5085f7100

📥 Commits

Reviewing files that changed from the base of the PR and between 9f63aa1 and 63c7791.

📒 Files selected for processing (17)

• .github/workflows/gc-native-roots.yml
• changelog.d/7414-macos-rs4gc-inprocess-gc-diag.md
• changelog.d/7415-dominance-corpus.md
• changelog.d/7416-i64-integer-locals.md
• crates/perry-codegen/src/codegen/closure.rs
• crates/perry-codegen/src/codegen/entry.rs
• crates/perry-codegen/src/codegen/function.rs
• crates/perry-codegen/src/codegen/method.rs
• crates/perry-codegen/src/collectors/hir_facts.rs
• crates/perry-codegen/src/collectors/int_valued_i64_locals.rs
• crates/perry-codegen/src/collectors/mod.rs
• crates/perry-codegen/src/expr/binary.rs
• crates/perry-codegen/src/expr/mod.rs
• crates/perry-codegen/src/type_analysis.rs
• crates/perry-codegen/src/type_analysis/numeric.rs
• crates/perry-codegen/src/type_analysis/numeric/tests.rs
• scripts/gc_evacuation_liveness_assert.py

---

📝 Walkthrough

Walkthrough

The PR adds bounded i64 local analysis for safe integer % lowering and propagates these facts through code generation. It also removes collector-diagnostic requirements from native-roots GC evacuation validation.

Changes

Integer modulo lowering

Layer / File(s)	Summary
Bounded i64 local analysis crates/perry-codegen/src/collectors/int_valued_i64_locals.rs, crates/perry-codegen/src/collectors/hir_facts.rs, crates/perry-codegen/src/collectors/mod.rs	The collector identifies eligible i64-range locals, records magnitude bounds, rejects unsupported writes, and validates nested control flow with tests.
Fact graph and codegen context wiring crates/perry-codegen/src/codegen/*.rs, crates/perry-codegen/src/expr/mod.rs	Native facts populate FnCtx::int_valued_i64_locals for functions, closures, entry modules, and methods.
Numeric analysis and modulo lowering crates/perry-codegen/src/type_analysis/numeric.rs, crates/perry-codegen/src/type_analysis.rs, crates/perry-codegen/src/expr/binary.rs, crates/perry-codegen/src/expr/mod.rs, crates/perry-codegen/src/type_analysis/numeric/tests.rs, changelog.d/7416-i64-integer-locals.md	Magnitude analysis admits safe i64 values. Divisor analysis excludes potentially zero locals. Eligible modulo expressions lower to srem; unsafe cases retain floating-point remainder lowering.

GC evacuation liveness validation

Layer / File(s)	Summary
Evacuation diagnostic requirement removal .github/workflows/gc-native-roots.yml, scripts/gc_evacuation_liveness_assert.py	The workflow no longer sets PERRY_GC_DIAG, and the assertion script no longer rejects traces without collector diagnostics.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
  participant HIRFacts
  participant TypeFacts
  participant FnCtx
  participant NumericAnalysis
  participant ModuloLowering
  HIRFacts->>TypeFacts: store bounded i64 local facts
  TypeFacts->>FnCtx: initialize int_valued_i64_locals
  FnCtx->>NumericAnalysis: analyze modulo operands
  NumericAnalysis->>ModuloLowering: approve safe integer remainder
  ModuloLowering->>ModuloLowering: emit srem i64

Loading

Possibly related issues

• GC backlog re-verification sweep, 2026-08-02 #7262 — The GC evacuation diagnostic and liveness changes address related GC verification behavior.
• GC testing: PERRY_GC_FORCE_EVACUATE is inert for gc()-driven tests (full mark-sweep + forced conservative scan) — stress claims may be unsupported #6946 — The GC evacuation diagnostic and liveness changes modify a related validation area.

Possibly related PRs

• PerryTS/perry#7338 — It introduced the GC diagnostic requirements removed here.
• PerryTS/perry#7414 — It directly covers the same GC workflow and diagnostic-handling changes.
• PerryTS/perry#6898 — Both changes extend flow-analysis-derived integer-local facts through the code-generation pipeline.

Suggested reviewers: thehypnoo, andrewtdiz

✨ 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 perf/7404-i64-integer-locals

---

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.}