Opt target 2 — `iso_miss` (2.11x) and `interp` (1.48x): tree-walking interpreters

[proggeramlug]

Opt target 2 — iso_miss (2.11x) and interp (1.48x): tree-walking interpreters

Treat these as one task. They share their shape, and a win on one plausibly moves the
other — check whether it does and report both either way.

Measurement — quiet M1 mini, 5 shuffled interleaved repeats, verdict CLEAN, perry fefdc367b

	node	perry	ratio
iso_miss wall	0.338 s	0.714 s	2.11x
— instructions	6.264 G	12.462 G	1.99x
— cycles	1.122 G	2.281 G	2.03x
— IPC	5.58	5.46	—
— peak RSS	89.4 MB	30.9 MB	0.35x
interp wall	0.324 s	0.481 s	1.48x
— instructions	5.973 G	8.643 G	1.45x
— cycles	1.079 G	1.530 G	1.42x
— IPC	5.54	5.65	—
— peak RSS	94.7 MB	31.3 MB	0.33x

Both are genuinely CPU-bound and the diagnosis is unambiguous: instruction ratio, cycle
ratio and wall ratio all agree to within a few percent, and IPC is essentially identical to
Node's. Perry is executing the same quality of code, just ~2x and ~1.45x more of it. Fewer
instructions is the only lever; there is no scheduling or stall story to find.

Workloads: gc-handoff/apps/iso_miss.ts and apps/interp.ts — tree-walking interpreters for
small functional languages. Recursive union allocation, closures capturing environments,
polymorphic dispatch, string building, deep recursion, string-keyed Maps.

History worth knowing

• iso_miss regressed +24.8% at perf(codegen): retain string accumulators in concat chains #8417 (a codegen string-accumulator change). perf(string): bypass empty accumulator copies #8487
  recovered most of it (14.788 G → 12.470 G instructions, −15.7%), and perf(runtime): resolve #8434 string-builder round trips — reconciling #8568 and #8571 #8581's join/repeat work
  landed since. Roughly +5% of that original regression was never accounted for.
• interp improved −4.61% instructions at perf(gc): skip tiny per-object layout masks #8539 (per-object layout-mask cutoff 2 → 4). That
  came from profiling, which found evalNode at 61.1% of self samples across 30 runs — not
  the dynamic-numeric-envelope problem the brief had guessed. Profile before assuming.
• An unattributed step in the 08-21 window moved iso_miss +7.9%; it has since been flat, so it
  was a one-time change rather than a trend. Attribution is still open if you want it.

Method

Profile first, with PERRY_KEEP_SYMBOLS=1 PERRY_DEBUG_SYMBOLS=1, aggregated over many runs —
these programs are 300–700 ms and a single sample yields almost nothing.

Because IPC already matches Node's, the question is what work exists that Node does not do:
redundant guards, allocation Node avoids, dispatch Node caches, or a per-node cost in the
interpreter loop that should be hoisted.

Acceptance

• Instruction reduction on either row, with the effect on both reported.
• All 20 corpus programs byte-exact; whole-corpus timing before and after, every row moving
  more than 1% reported.
• RSS is already 0.33–0.35x Node. Do not trade it for speed; quantify any cost.
• Instructions retired is the primary signal — the bench host is contended and wall ranges
  overlap.

[proggeramlug]

Refreshed against the 2026-08-22 sweep at adbebbe9d (verdict CLEAN, quiet mini, 5 shuffled interleaved repeats). Both rows improved but the diagnosis is unchanged, so this stays open.

	wall	prev	instructions
iso_miss	2.07x	2.11x	1.96x
interp	1.42x	1.48x	1.41x

The instruction ratio still tracks the wall ratio almost exactly on both — 2.07 vs 1.96, and 1.42 vs 1.41. These remain genuinely CPU-bound with IPC comparable to Node's, so fewer instructions is still the only lever and there is no stall or scheduling story to find.

#8597 landed since (inline equality for short heap strings, targeting exactly the identifier comparisons these interpreters do) and is part of the −2% / −4% above. Worth knowing it is already applied, so it is not available as a fresh idea.

For contrast, both of the other rows from the original ticket set have now changed character: typed_array fell 6.29x → 3.87x with instructions 8.52x → 2.24x, so its wall ratio now exceeds its instruction ratio and it needs a stall-oriented look rather than more work-deletion. pipeline remains the one row where instructions (1.74x) far exceed cycles — scheduling-bound, not CPU-bound. This ticket is the only one of the three still cleanly in the "delete executed work" category.