Repository navigation
Date is a value type — setter mutations don't propagate through aliasing/function/closure boundaries (breaks effect DateTime.add) #2089
Description
Summary
Perry models Date as a value type (a raw NaN-boxed f64 timestamp), but a JavaScript Date is a mutable heap object passed by reference. Mutating a Date through any aliasing boundary — a function parameter, a closure parameter, or even a plain const b = a — does not propagate back, because each binding holds an independent copy of the f64.
This surfaced while bringing up effect (#321): DateTime.add(d, { days: 1 }) is a no-op, because effect implements it as mutate(self, (date) => { date.setUTCDate(date.getUTCDate() + parts.days) }) — mutating a Date inside a closure and reading it back. DateTime.addDuration works because it uses pure arithmetic (mapEpochMillis).
Minimal repro (no effect)
const a = new Date(1704067200000);
const b = a; // alias — same object in JS
b.setUTCDate(b.getUTCDate() + 1);
console.log(a.getTime() - 1704067200000); // Node: 86400000 Perry: 0 ✗
console.log(b.getTime() - 1704067200000); // Node: 86400000 Perry: 86400000b sees its own mutation; a does not — they are not the same object. Same failure when the mutation happens through a function/closure parameter (the mutate(self, f) shape). A Date created as a top-level const and mutated via a param happens to work by accident (the inliner rewrites the setter writeback against the caller's local); a function-local Date passed to a mutating fn/closure fails.
Mechanism
- Representation —
crates/perry-runtime/src/date.rs(js_date_new*): returns the f64 timestamp itself and records its bits in a thread-localDATE_REGISTRY(forinstanceof). There is no heap cell. - Setters —
crates/perry-runtime/src/date.rs(js_date_set_*(timestamp, value) -> new_timestamp): each returns a fresh f64; nothing is mutated in place. - Writeback lowering —
crates/perry-hir/src/lower/expr_call/url_date_instance.rs: aLocalGet(id)receiver wraps the setter asLocalSet(id, setter_call), which only updates the local in the current function; through a call/closure boundary the receiver is the callee's own copy, so the new timestamp never reaches the caller's variable.
Fix direction
Promote Date to a reference type — box the timestamp in a 1-slot mutable heap cell so all aliases/params share state. This is invasive (≈8 files, cross-crate, regression-prone): all ~20 getters/setters in date.rs, every site keyed on the f64 bit pattern (numeric coercion in builtins/arithmetic.rs, json/stringify.rs, object/instanceof.rs, object/util_types.rs, object/native_module_dispatch.rs, object/field_get_set.rs, promise/combinators.rs), the Date-producing codegen exprs currently typed DOUBLE (expr/arrays_finds.rs, expr/misc_methods.rs, expr/os_uri_dates.rs), and GC scanning for the new cell.
Acceptance
The repro above: a.getTime() reflects a mutation made through b / a function param / a closure. And effect's DateTime.add(d, { days: 1 }) adds 86400000 ms.
Refs #321.
Activity
proggeramlug commentedon May 27, 2026
Deeper analysis (second investigation) — this is an architectural spike, not a single-pass fix
Confirmed the bug + that the rest of the Date matrix is byte-perfect today on the value-type representation (
typeof,instanceof,+date,getTime,date - date,date < date2,toISOString,JSON.stringify,Date.parse, Invalid Date). So a reference-type change risks regressing many working paths to fix the aliasing one.Two hard sub-problems gate a regression-free landing:
crates/perry-codegen/src/expr/compare.rs:370lowersdate1 < date2/>/<=/>=as inline LLVMfcmpdirectly on operand values. Today operands ARE the timestamps so it works; a NaN-boxed pointer Date would makefcmpunordered → silentlyfalse. Operands are usuallyany-typed → can't statically detect Date here. Fix needs either a runtime relational helper on the Any path (perf regression for all numeric</>) or accepts breakage. (Arithmetic IS hookable viajs_number_coerce'sis_pointer()branch,builtins/numbers.rs:304— only comparison is the gap.)is_pointer()consumer surface. A POINTER-tagged Date cell is seen by everyis_pointer()consumer (typeof, equality,js_to_primitive, field-get, native dispatch, JSON, thread serialization, conservative stack scan). Thebox.rsregistry-scanner keeps cells alive without shadow rooting but pins them forever → every Date leaks (bad for long-running servers). Leak-free needs GC-arena allocation + teaching codegen pointer/escape analysis (collectors/pointer_locals.rs,shadow_slots.rs) to treat Date-producing exprs as pointers.Recommended representation (keeps codegen churn minimal)
Box the Date as a POINTER_TAG value that is itself an
f64pointing at a 1-slotDateCell { ts: f64 }. All codegen that types Date exprs asDOUBLEand passes(DOUBLE, &v)to getters/setters stays unchanged; getters deref, setters mutate in place, and theLocalSetwriteback hack inlower/expr_call/url_date_instance.rs:350is dropped. Implement as a spike behind problems (1) and (2); validate the full Date gap matrix + effectDateTime.addbefore expanding to all ~30 getters/setters; resolve the leak question (arena+rooting vs registry-pin) explicitly.Recommend tackling as a planned, phased effort rather than a single PR.
fix(date): model Date as a reference type so mutations propagate (#2089)
fix(date): model Date as a reference type so mutations propagate (#2089)
fix(date): model Date as a reference type so mutations propagate (#2089)
fix(date): model Date as a reference type so mutations propagate (#2089)
fix(date): model Date as a reference type so mutations propagate (#2089)
fix(date): model Date as a reference type so mutations propagate (#2089…
test(date): regression coverage for console.log(new Date(NaN)) (#2371)
test(date): regression coverage for console.log(new Date(NaN)) (#2371)
fix(console): revert Date-as-ISO inspect — it regressed Date.getTime(…
fix(console): revert Date-as-ISO inspect — it regressed Date.getTime(…
test(#2398): lock in console.log/util.inspect Date-as-ISO parity (#2414)
fix(runtime): Object.prototype.toString.call(date) is [object Date] (#…