Summary
Escape analysis scalar-replaces a class instance that "never escapes", but the promotion is blind to own properties installed at runtime by the native base's subclass-init. Reading a declared field alongside a native method makes the method read undefined.
Confirmed on clean origin/main. No override, no indirection — a direct extends EventEmitter with one field is enough.
Repro
import { EventEmitter } from "node:events";
class X extends EventEmitter {
a = 1; // <- a declared field is load-bearing
}
const x = new X();
console.log("field:", x.a, "typeof emit:", typeof x.emit, "typeof on:", typeof x.on);
|
output |
| node 26 |
field: 1 typeof emit: function typeof on: function |
perry main |
field: 1 typeof emit: undefined typeof on: undefined |
Drop the field (class X extends EventEmitter {}) and the methods reappear — the instance then escapes / isn't promoted.
Mechanism
EventEmitter (and the node:stream classes) install their method surface as own properties on the instance at subclass-init time (install_methods_on_existing_object -> js_object_set_field_by_name(obj, "emit", <native closure>)), rather than on a prototype.
When the instance is proven non-escaping, scalar replacement promotes its declared fields to scalars — and a runtime-installed own property has no declared slot, so it simply isn't in the promoted set. The subsequent x.emit read resolves against the scalarized object and finds nothing.
Relationship to other issues
Fix direction
The escape/scalar-replacement decision must treat "this class's chain reaches a native base that installs an instance surface" as a disqualifier (or the promoted set must include runtime-installed own keys). Note PR #6324's guard is precise rather than a blanket disable — follow that shape; do not simply switch scalar replacement off, which would give back a real perf win.
The scripts/run_issue_945_scalar_method_ir_guard.sh IR guard and test-files/test_issue_945_scalar_method_guards.ts are the existing harness for this analysis.
Summary
Escape analysis scalar-replaces a class instance that "never escapes", but the promotion is blind to own properties installed at runtime by the native base's subclass-init. Reading a declared field alongside a native method makes the method read
undefined.Confirmed on clean
origin/main. No override, no indirection — a directextends EventEmitterwith one field is enough.Repro
field: 1 typeof emit: function typeof on: functionmainfield: 1 typeof emit: undefined typeof on: undefinedDrop the field (
class X extends EventEmitter {}) and the methods reappear — the instance then escapes / isn't promoted.Mechanism
EventEmitter(and thenode:streamclasses) install their method surface as own properties on the instance at subclass-init time (install_methods_on_existing_object->js_object_set_field_by_name(obj, "emit", <native closure>)), rather than on a prototype.When the instance is proven non-escaping, scalar replacement promotes its declared fields to scalars — and a runtime-installed own property has no declared slot, so it simply isn't in the promoted set. The subsequent
x.emitread resolves against the scalarized object and finds nothing.Relationship to other issues
Fix direction
The escape/scalar-replacement decision must treat "this class's chain reaches a native base that installs an instance surface" as a disqualifier (or the promoted set must include runtime-installed own keys). Note PR #6324's guard is precise rather than a blanket disable — follow that shape; do not simply switch scalar replacement off, which would give back a real perf win.
The
scripts/run_issue_945_scalar_method_ir_guard.shIR guard andtest-files/test_issue_945_scalar_method_guards.tsare the existing harness for this analysis.