Shape-backed object literals inside generator/async frames when values cannot suspend

[lahma]

Shape-backed object literals inside generator/async frames when values cannot suspend

Follow-up to #2571/#2593: with the per-yield handler rebuild fixed, the dominant remaining allocation in generator bodies that build object literals is the dictionary build path itself — one PropertyDescriptor + a PropertyDictionary (plus its ListDictionary head) per constructed object.

That path exists because a literal evaluated inside a suspendable frame can suspend mid-build if a property value contains yield/await, and the shape-backed fast build (slot fill, no descriptors) has no resume support. But the exclusion was frame-based, not literal-based: yield { a: i, b: i + 1 } — a literal that can never suspend its own build — paid the dictionary path merely for being inside a generator.

Change

At handler construction, prove per literal that no property value contains a lexical yield/await:

• descent stops at function boundaries (function/arrow) — suspensions there belong to the nested frame;
• class nodes are descended conservatively (computed member keys evaluate in the enclosing frame); in practice class-expression values already take the dictionary path via the existing function-definition exclusion;
• proof only computed for literals already eligible for the fast build (static distinct keys, no spread/accessors/function shorthand).

Proven-suspend-free literals then take the existing shape-backed build (BuildObjectFastShapeBacked) even when ExecutionContext.Suspendable is set. Literals whose values do contain yield/await keep the dictionary path and its suspend/resume machinery unchanged.

Results

Census workload (generator yielding { a: i, b: i + 1, c: i + 2 } 100k times, fresh engine per iter):

	main	this PR
allocated/iter	53.0 MB	23.1 MB (−56%)
PropertyDescriptor share	18.3%	eliminated
HybridDictionary/ListDictionary share	12.3%	eliminated

Remaining allocations are user-visible: the literal objects themselves (45%), IteratorResult from explicit .next() calls (40%), boxed numbers (13%).

GeneratorBenchmark and AsyncFunctionExitBenchmark are neutral with byte-identical allocations (their bodies build no literals — the census script is the allocation authority for this pattern).

Testing

• Full test262: 99,260 passed, 0 failed, 133 skipped (yield-inside-literal suspension/resume is heavily covered there)
• Jint.Tests + Jint.Tests.PublicInterface green (net10.0 + net472)
• Smoke-verified: mid-build suspension ({ first: 1, second: yield "mid", third: 3 }) resumes with correct values and key order; nested deep yield keeps the outer literal on the resumable path

🤖 Generated with Claude Code

https://claude.ai/code/session_01S2U2y9voMQV9rvCkYeQbka