perf: reads of an Array grown by `a[i] = v` or `push` are 54× slower than Node (binding keeps the forwarded old head; every inline read bails out)

[proggeramlug]

Found by the package performance audit (real npm packages compiled from source, profiled against Node 26.5.1) and
re-measured on Perry 7661bc0 (v0.5.1589), Linux x64. When a plain Array outgrows its storage, Perry leaves a
growth-forwarding stub at the old address. Bindings and fields that still hold that address are never updated. Every
inline element read tests GC_FLAG_FORWARDED and leaves the fast path when it is set, so reads through such a binding
go out of line for the rest of the program. An 80-element sum loop costs 5.4× the instructions of the same loop over a
presized array, 468 vs 87 instructions per read. That makes it 54× slower than Node. An explicit gc() changes nothing.

Reproduction

bench.ts (16 lines):

const variant = process.argv[2] || "grown"; const N = Number(process.argv[3] || "1000000");
// Read-only loop over an 80-element plain Array; the variants differ ONLY in how the array was built.
function sum(d: any, n: number): number { let s = 0; for (let j = 0; j < n; j++) s += d[j]; return s; }
function sumT(d: number[], n: number): number { let s = 0; for (let j = 0; j < n; j++) s += d[j]; return s; }
let d: any;
if (variant === "presized" || variant === "presized_typed") d = new Array(80).fill(0);  // CONTROL: never grows
else if (variant === "pushed") { d = []; for (let j = 0; j < 80; j++) d.push(0); }     // grown by push
else d = [];                                                                             // grown by d[j] = v
for (let j = 0; j < 80; j++) d[j] = (j * 7) & 0xfffff;
if (variant === "grown_gc" && typeof (globalThis as any).gc === "function") (globalThis as any).gc(); // does not help
function run(n: number): number {
  let acc = 0;
  for (let r = 0; r < n; r++) acc = (acc + (variant.endsWith("_typed") ? sumT(d, 80) : sum(d, 80))) % 1000000007;
  return acc;
}
run(N / 5 | 0); const t0 = performance.now(); const cs = run(N); console.log(`${variant} checksum=${cs} ms=${(performance.now() - t0).toFixed(1)}`);

PERRY_NO_AUTO_OPTIMIZE=1 perry compile bench.ts -o bench
for v in presized grown pushed grown_gc presized_typed grown_typed; do node bench.ts $v 1000000; ./bench $v 1000000; done

Measurements

These are medians of 3 runs on a shared, loaded host. Instruction counts are the load-independent figure. They are
whole-process instructions:u and include the N/5 warm-up. N = 1,000,000 iterations of 80 reads, so the per-read column
divides by 96 M.

variant	Node loop ms	Perry loop ms	ratio	Perry instructions (per read)	Node wall	Perry wall
presized (control)	63.1	583.9	9.3×	8.33 G (87)	214 ms	746 ms
grown (d[j] = v from [])	65.4	3,541	54×	44.98 G (468)	232 ms	4,660 ms
pushed	61.5	3,755	61×	45.00 G (468)	212 ms	4,571 ms
grown_gc (gc() after growth)	61.5	3,636	59×	45.02 G (469)	213 ms	4,413 ms
presized_typed (d: number[], control)	87.7	104.2	1.2×	1.46 G (15)	247 ms	163 ms
grown_typed	82.3	587.3	7.1×	9.04 G (94)	227 ms	746 ms

Checksums are identical in every variant.

• Untyped receiver: growth adds about 380 instructions to every read.
• number[] receiver: the same growth multiplies read cost 6×.

perf record on grown:

symbol	share
js_array_get_f64	31.4 %
try_read_tracked_gc_header	29.8 %
js_packed_arraylike_index_get	17.7 %
compiled loop	13.8 %

On presized, 90 % of samples are in the compiled loop.

gdb breakpoint on js_packed_arraylike_index_get for grown, pushed and grown_gc: the receiver header is
obj_type=0x01 (ARRAY) with gc_flags=0x82 (FORWARDED | ARENA). For grown_gc this is after the gc() call.

Impact

• node-forge 1.4.0, the RSA/bignum path (numeric-group audit profile, v0.5.1587):
  • Every BigInteger starts as this.data = [] (jsbn.js:62) and is filled by index, so its digit arrays are grown
    arrays read through this.data / w.data.
  • The audit's gdb check in the forge loop found receiver flags 0x82.
  • Removing growth from its fill/sum microbenchmark (presizing the array) cut Perry's time by 38 %.
  • The "array ops" bucket is 29–31 % of forge sign and keygen time. This issue is its read half; perf: arr[i] = v on an untyped plain Array is 52× slower than Node (no inline store: one js_dyn_index_set_strict call per element; the same loop with a number[] annotation is 1.7×) #10513 is the
    store half.
• General: any code that builds an array with push or index writes and keeps reading it through the original
  reference (object fields, closure captures, module globals).

Mechanism

Verified unless marked (inferred).

• The stub is installed and retained on purpose. crates/perry-runtime/src/array/push_pop.rs:250-270 installs a
  forwarding stub on the old head when the array reallocates (install_array_growth_forwarding_with, line 265). The
  comment there says the stubs are retained "because stale array references rely on clean_arr_ptr following this
  chain".
• Every inline read lane rejects a forwarded header.
  • The dynamic element read tests gc_flags & 0x80 and branches to the miss exit
    (crates/perry-codegen/src/expr/index_get/inline_dyn_typed_array.rs:185-189). That exit calls
    js_packed_arraylike_index_get → js_array_get_f64, which classifies the address again
    (try_read_tracked_gc_header) and follows the chain on every read.
  • The bounded-index lane does the same (crates/perry-codegen/src/expr/index_get.rs:1050-1063), as does
    index_get/guarded_array.rs:303,341.
• Nothing repairs the stale binding in a compute loop. A loop that does not allocate never triggers a collection.
  In this benchmark an explicit gc() did not rewrite the module-global binding either: after it, the receiver still
  had flags 0x82 (gdb).
• Typed receivers suffer too (inferred). For a number[] receiver the loop stays in compiled code, but costs 6×.
  The likely cause is the loop-version guard rejecting the forwarded head and taking the slower clone.

What fast looks like

Any of these would work:

• Heal in place. Follow one forwarding hop inline (load the forwarding word from the stub, re-check the header),
  instead of bailing to the runtime.
• Write the resolved head back. The miss path returns the resolved head, and the read site writes it back into the
  local or field that held the stub.
• Rewrite at collection. Have every collection, including explicit gc(), rewrite slots that point at growth stubs.

Target: grown and pushed within 10 % of presized instructions per read (≤ ~95 now, ~15 for the number[]
receiver), and grown_typed ≈ presized_typed.

Notes

• perf: arr[i] = v on an untyped plain Array is 52× slower than Node (no inline store: one js_dyn_index_set_strict call per element; the same loop with a number[] annotation is 1.7×) #10513 is the store-side cost for the same receivers. The presized control's own 9.3× is spent inside the compiled
  loop, most likely the guarded dynamic += accumulator described in perf: accumulate-over-reads loops pay a per-element rooting diamond when the read isn't served by a proven tier #9361 (inferred).
• Growth forwarding exists for correctness (array growth forwarding: build()-and-return leaves a stale head in the binding (minimal N = 17) #7661, release blocker: stable packed loop reads stale forwarded array after cross-call growth #9117, both closed). This issue is about its steady-state read cost.

[proggeramlug]

Progress: #11588 landed. The inline dynamic read now follows one growth-forwarding hop, and the packed loop tier heals a forwarded head before its guard. Grown and pushed arrays drop from ~565 → ~73 instructions per read, on par with the presized control (68). The grown typed case goes from 116 → 23.

Still open: the presized control itself is ~69 instructions per read vs Node's handful, so the remaining gap is the general untyped read path rather than growth. Keeping this open for that.