[mrousavy]

Construct heap-buffered JSI values with reserve() and emplace_back() instead of default-initializing and then assigning every slot. Keep the stack path for up to eight elements and one bulk push into Hermes.

The shared helper covers C++ vectors, Swift arrays, and JNI primitive/object arrays. Tuples already construct their fixed staging values directly; their direct-insertion alternative is measured below.

Measured heap-backed numeric arrays take about 19–23% less time. Direct indexed insertion is slower for multi-element arrays and tuples, so bulk insertion stays here.

Payload	Elements	Current resize/assign + bulk push	Reserve/emplace + bulk push	Preallocated array + indexed stores
Numbers	0	11.7	11.6 (0.989×; 0.984–0.994)	10.5 (0.895×; 0.890–0.901)
Numbers	8	75.2	72.7 (0.965×; 0.958–0.971)	249.0 (3.309×; 3.287–3.320)
Numbers	9	117.8	91.5 (0.774×; 0.767–0.793)	278.8 (2.371×; 2.306–2.391)
Numbers	64	506.0	409.9 (0.811×; 0.803–0.820)	1,908.0 (3.769×; 3.737–3.792)
Numbers	16,384	111,860	89,698 (0.803×; 0.797–0.809)	484,276 (4.325×; 4.312–4.345)
Object references	64	665.6	572.1 (0.859×; 0.844–0.865)	2,049.9 (3.077×; 3.016–3.104)
Fresh UTF-8 strings	64	1,370.9	1,293.4 (0.943×; 0.938–0.949)	2,836.1 (2.065×; 2.056–2.074)

Tuple	Bulk push (current and reserve PR)	Direct indexed stores
Empty	20.1	11.1 (0.551×; 0.548–0.555)
Four numbers	37.9	128.3 (3.416×; 3.368–3.443)
Number, boolean, string, number	53.6	143.6 (2.679×; 2.649–2.709)

Reproducible benchmark source, raw samples and build commands: https://gist.github.com/mrousavy/039d0c2ee0624790396f84deda84dd80

Benchmark method and environment

Measurements are ns per produced array/tuple. Candidate cells show median time, median paired ratio to the current implementation, and the paired ratio's P10–P90 range; lower is faster. Each case has 35 samples across five fresh processes, seven rounds per process. All three methods used the same calibrated iteration count per case; option order and case order were shuffled. Each option warmed up before timing; a full Hermes GC ran before each batch, outside the timer. Every produced array escaped through a compiler barrier; result lengths, values and reference identity were checked outside timers. Allocation, conversion, stores, push and result destruction are inside the measured loop.

Apple M2 Pro (arm64), macOS 27.2 (26B123), Apple clang 21.0.0 (clang-2100.4.1.1), benchmark -std=c++20 -O3 -DNDEBUG. Hermes pinned commit 26fbb82fad39d9d06901adb1d6d630dacc84c877, rebuilt from its clean checkout using the repository's cached Release configuration: -O3 -DNDEBUG, Hades GC, HEAP_HV_64, Unicode-lite, debugger/Intl disabled. The benchmark is C++ → JSI → Hermes, with no JNI, Swift or React Native bridge. Object/string-reference cases clone preexisting rooted JS values; fresh strings use String::createFromUtf8 on 22–26-byte ASCII inputs. One process ran at a time during an exclusive no-build window.

Full benchmark cases

Payload	Elements	Current resize/assign + bulk push	Reserve/emplace + bulk push	Preallocated array + indexed stores
Numbers	0	11.7	11.6 (0.989×; 0.984–0.994)	10.5 (0.895×; 0.890–0.901)
Numbers	1	38.7	38.5 (0.992×; 0.985–1.003)	39.0 (1.008×; 1.001–1.026)
Numbers	8	75.2	72.7 (0.965×; 0.958–0.971)	249.0 (3.309×; 3.287–3.320)
Numbers	9	117.8	91.5 (0.774×; 0.767–0.793)	278.8 (2.371×; 2.306–2.391)
Numbers	16	166.5	135.8 (0.815×; 0.800–0.827)	486.8 (2.932×; 2.871–2.965)
Numbers	64	506.0	409.9 (0.811×; 0.803–0.820)	1,908.0 (3.769×; 3.737–3.792)
Numbers	256	1,813.0	1,468.8 (0.810×; 0.805–0.815)	7,667.0 (4.224×; 4.171–4.248)
Numbers	1,024	7,040.0	5,708.6 (0.810×; 0.805–0.825)	30,371 (4.319×; 4.277–4.336)
Numbers	16,384	111,860	89,698 (0.803×; 0.797–0.809)	484,276 (4.325×; 4.312–4.345)
Object references	0	12.0	11.6 (0.969×; 0.954–1.008)	10.8 (0.898×; 0.883–0.916)
Object references	1	40.2	39.4 (0.980×; 0.975–0.987)	40.0 (0.996×; 0.990–1.004)
Object references	8	103.8	97.8 (0.941×; 0.933–0.952)	267.6 (2.583×; 2.560–2.606)
Object references	9	143.3	121.9 (0.848×; 0.841–0.864)	299.2 (2.085×; 2.053–2.098)
Object references	16	208.6	178.5 (0.853×; 0.846–0.864)	521.4 (2.499×; 2.472–2.518)
Object references	64	665.6	572.1 (0.859×; 0.844–0.865)	2,049.9 (3.077×; 3.016–3.104)
Object references	256	2,499.8	2,159.5 (0.863×; 0.855–0.870)	8,225.5 (3.292×; 3.257–3.298)
Object references	1,024	9,718.7	8,384.0 (0.863×; 0.857–0.871)	32,679 (3.367×; 3.318–3.384)
Object references	16,384	160,729	136,143 (0.849×; 0.842–0.854)	519,944 (3.241×; 3.210–3.256)
String references	0	11.8	11.4 (0.964×; 0.953–0.974)	10.6 (0.896×; 0.888–0.906)
String references	1	40.4	39.6 (0.980×; 0.975–0.990)	39.7 (0.985×; 0.979–0.991)
String references	8	105.7	99.1 (0.936×; 0.927–0.942)	269.1 (2.546×; 2.452–2.561)
String references	9	146.0	124.1 (0.852×; 0.845–0.862)	300.8 (2.059×; 2.024–2.076)
String references	16	211.6	180.8 (0.856×; 0.850–0.864)	525.8 (2.484×; 2.450–2.502)
String references	64	670.7	581.3 (0.867×; 0.862–0.874)	2,065.6 (3.083×; 3.057–3.097)
String references	256	2,507.9	2,166.8 (0.865×; 0.858–0.871)	8,287.6 (3.302×; 3.281–3.324)
String references	1,024	9,694.9	8,365.8 (0.862×; 0.859–0.867)	32,962 (3.400×; 3.374–3.412)
String references	16,384	162,515	138,390 (0.854×; 0.843–0.876)	525,235 (3.239×; 3.176–3.264)
Fresh UTF-8 strings	0	11.9	11.6 (0.984×; 0.969–0.991)	11.7 (0.979×; 0.965–0.985)
Fresh UTF-8 strings	1	54.3	53.5 (0.985×; 0.976–0.992)	55.4 (1.020×; 1.010–1.030)
Fresh UTF-8 strings	8	192.7	185.3 (0.963×; 0.953–0.972)	359.7 (1.866×; 1.853–1.883)
Fresh UTF-8 strings	9	245.1	226.2 (0.924×; 0.917–0.931)	401.3 (1.635×; 1.621–1.649)
Fresh UTF-8 strings	16	386.0	359.4 (0.932×; 0.928–0.939)	732.6 (1.899×; 1.888–1.914)
Fresh UTF-8 strings	64	1,370.9	1,293.4 (0.943×; 0.938–0.949)	2,836.1 (2.065×; 2.056–2.074)
Fresh UTF-8 strings	256	4,908.5	4,634.2 (0.944×; 0.940–0.948)	10,951 (2.231×; 2.220–2.241)
Fresh UTF-8 strings	1,024	18,862	17,823 (0.944×; 0.941–0.950)	42,875 (2.275×; 2.265–2.282)
Fresh UTF-8 strings	16,384	358,936	341,608 (0.954×; 0.936–0.965)	718,175 (2.008×; 1.981–2.024)

Validation: the caller and template implementations pass ASan+UBSan checks for empty/small/large arrays, heterogeneous values, reference identity, conversion order, conversion exceptions and forced GC. Hermes itself is an uninstrumented Release build. All 127 generated Android C++ translation units pass NDK syntax checks. Prototype-setter behavior matches the current implementation. Benchmark fixtures remain outside the repository.

Draft direct-insertion follow-up: #2038.

Small-array follow-up: sizes 1–8

Fresh five-process run on the same Release Hermes/M2 Pro setup, with seven paired randomized rounds per case (35 samples per option). Times are ns per numeric array; ratio ranges are paired P10–P90. Baseline means the original resize/assignment implementation, reserve/emplace is #2037, and direct insertion is #2038.

Elements	Baseline ns	Reserve/emplace ns	Reserve/baseline	Direct index ns	Direct/baseline
1	39.8	39.0	0.990× (0.972–1.011)	40.3	1.010× (0.983–1.035)
2	46.5	45.5	0.976× (0.938–1.004)	71.1	1.523× (1.481–1.572)
3	50.8	49.6	0.980× (0.921–1.005)	104.7	2.058× (1.957–2.098)
4	56.5	54.7	0.967× (0.944–1.038)	134.9	2.390× (2.274–2.469)
5	61.3	60.9	0.978× (0.941–1.011)	164.8	2.688× (2.620–2.749)
6	69.1	66.4	0.971× (0.937–0.996)	200.2	2.952× (2.848–3.040)
7	72.3	69.5	0.969× (0.921–0.984)	224.9	3.131× (3.021–3.298)
8	77.4	75.2	0.971× (0.945–1.005)	257.6	3.322× (3.245–3.412)

Direct insertion is effectively tied at one element and becomes slower at two. All four measured payload kinds show a regression from two elements. The small bulk paths use stack storage, so this range has no temporary heap-buffer allocation to eliminate.

Source, raw samples, build commands and full payload table: https://gist.github.com/mrousavy/baf71995556d27d5cf3f61921f8db4d3

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
nitro-docs	Skipped		Nov 2, 2026 1:03pm UTC