perf: declared-class construction is 63% slower than an equivalent object literal (28.5x vs 17.4x vs node)

[proggeramlug]

Symptom

Decomposed churn on the quiet mini (best-of-3), Perry vs node:

variant	Perry	node	ratio
churn (full)	2.72 s	0.17 s	16.0×
churn_alloc — object literal + push	2.44 s	0.14 s	17.4×
push_cls — new Node(v,w) + push	3.99 s	0.14 s	28.5×
push_num — numbers into array	0.30 s	0.11 s	2.7×
churn_read — element reads only	0.35 s	0.08 s	4.3×

A declared class is 63% slower to construct than an equivalent object literal.

That ordering is backwards and is the reason this is filed as a defect rather
than a missing optimization. new Node(v, w) on a fixed-shape declared class
is the most statically-known construction form in the language: the field set,
their order, the class id and the shape are all known at compile time. The
object literal has to mint an anon shape (__AnonShape_…) and still wins.

Whatever the declared-class path is doing extra, it is doing it in spite of
having strictly more static information — so the likely finding is a fast path
that the literal reaches and the class does not, or per-construction work that
the class path repeats and the literal amortizes.

Where to look first

• perf(codegen): elide provably-dead per-store bookkeeping on class-field stores #7486's elision may not reach declared classes. It elides dead
  default-undefined field inits when the ctor opens with a run of
  this.f = <param> statements, proven by ctor_prologue_param_assigned_fields
  (lower_call/field_init.rs). It was measured on the synthesized anon-shape
  ctor. A user-declared class Node { constructor(v, w) { this.v = v; this.w = w } }
  should qualify — verify it actually does, and if it does not, why.
• js_gc_init_typed_shape_layout (7.7% of churn_alloc) — is it emitted
  per construction on the class path but hoisted/absent on the literal path?
• Typed-feedback guards (9.2%) — does the class path instrument sites the
  literal path does not?

Method

Compile both shapes with --trace llvm and diff the per-construction call
census (that is how #7501's admission analysis was verified). The difference
should be visible as concrete calls, not inferred from timings — and it is
cheap to get, which is why this may be a quick bisect rather than a project.

Provenance

Measured on the pinned quiet host, symbolicated with PERRY_DEBUG_SYMBOLS=1,
1500 leaf samples. Part of the #7469 mutator-cost campaign; sibling findings
are the layout side-table cost (commented on #5094) and write-barrier cost.

[proggeramlug]

Duplicate of #7512, which was filed concurrently with fuller cross-linking (an explicit off-ramp to #7510/#7511 if the cause turns out to live in the layout or barrier subsystems, plus the measurement-discipline notes). Closing this one; #7512 is the tracker.

My apologies for the collision — I filed without checking whether the investigation that produced the numbers had already opened its own ticket.