SIMD and wasm‑opt in WebAssembly builds: November 2026

Reference machine: Apple M4 Pro (arm64) · macOS 26.5.2
Browsers: Chrome 156, Firefox 151 and Safari 26
Samples: 15 rounds per row, two captures

Only the Apple M4 Pro (arm64) has been measured, not x86-64.

Measure on this machine

1. Main points

On arm64, simd128 cut zlib-rs deflate L1's time by up to 18.4%, but slowed zstd decode by up to 7.6% in Firefox (Figure 1).
Skipping wasm-opt cut brotli encode's time by up to 18.5% in Chrome and Safari, but slowed 3 workloads by up to 9.6% (Figure 2).
wasm-opt -Os shrank all 5 modules by up to 33.1 kB after compression (Figure 3).

2. SIMD and speed

Figure 1. simd128 cut zlib‑rs deflate L1's time by 6.3%–18.4% on 2 of 3 inputs, in all 3 browsers

simd128 build's time ÷ the scalar build's in the same round, both at wasm-opt -Os; 15 rounds.

Claimed in both captures
Not claimed
95% CI
Min–max
±3% threshold

Notes: 10 of 22 rows, with no claim and browsers agreeing, are in Table 1 only. zstd decode in Safari: A/A control failed, so grey (Figure 4). 3 claims held in one capture only, so not counted.

Why the processor matters

zlib-rs's compare256, deflate's match-length loop, ends each 16-byte step in i8x16.bitmask, one instruction on x86-64 but a sequence on arm64.

This may give SIMD a larger gain in zlib-rs deflate on x86-64, which has not been measured here.

;; zlib-rs compare256, wasm32 + simd128, per 16 bytes
(i8x16.eq a b)        ;; 0xff where bytes match
(i8x16.bitmask ...)   ;; one bit per byte

;; x86-64 (V8): one instruction each
pcmpeqb   xmm0, xmm1
pmovmskb  eax, xmm0

;; arm64 (V8): no single equivalent for i8x16.bitmask;
;; V8 emits a multi-instruction sequence (which one
;; depends on the CPU's features)

Source: zlib-rs 0.6.8 and V8 at commit 63481c19 (Measurement notes).

3. wasm-opt and speed

Figure 2. Skipping wasm‑opt cut brotli encode's time by 4.3%–18.5% on all 3 inputs, in Chrome and Safari

Scalar build without wasm-opt: its time ÷ the shipped build's (wasm-opt -Os) in the same round; 15 rounds.

Claimed in both captures
Not claimed
95% CI
Min–max
±3% threshold

Notes: 11 of 22 rows, with no claim and browsers agreeing, are in Table 1 only. zstd decode in Safari: A/A control failed, so grey (Figure 4). 1 claim held in one capture only, so not counted.

Without SIMD, wasm-opt -O3 instead of -Os cut lz4_flex decode's time up to 5.5% on samba in Chrome and on x-ray and samba in Safari; no other workload changed in both captures (Table 1).

4. wasm-opt, SIMD and download size

Figure 3. wasm‑opt ‑Os shrank all 5 modules; simd128 grew all 5

Change in brotli-compressed bytes per module; exact and the same in every browser.

Source: Build manifests of the modules this page runs (scripts/build-variants.ts). rustc 1.99.1 (58cf8f9c3 2026-10-13); wasm-opt version 133 (version_133)

Notes: Raw and gzip bytes, and the -O3 builds: Table 2.

5. The A/A control

Figure 4. A/A control: within ±2.2% in Chrome and Firefox; failed for the zstd decoder in Safari

Time of a second, byte-identical instance of the shipped build ÷ the first's, pooling each workload's inputs and both captures. Where it fails, nothing is claimed for that workload in that browser.

Control failed
Passed
95% CI
Min–max
±3% threshold

6. All configurations

Chrome 156Firefox 151Safari 26

**Table 1.** Each build's time ÷ the shipped build's in the same round (Chrome 156, capture 1): median of 15 rounds and 95% CI. Bold clears the claim rule in this capture.
Workload · input	Shipped, per run	scalar no wasm-opt	scalar -O3	simd128 no wasm-opt	simd128 -Os	simd128 -O3	A/A
zstd decodedickens	10.1 ms	1.0371.02–1.04	1.0611.05–1.07	1.0381.02–1.05	1.0091.00–1.03	1.0111.01–1.02	0.9990.99–1.01
x-ray	11.0 ms	1.0241.01–1.03	1.0040.99–1.01	1.0040.99–1.01	0.9960.99–1.01	1.0030.99–1.01	0.9940.99–1.00
samba	13.6 ms	1.0221.00–1.03	0.9880.98–1.01	0.9960.98–1.01	0.9620.95–0.97	0.9860.97–0.99	0.9940.99–1.01
zlib-rs inflatedickens	15.9 ms	1.0141.00–1.03	1.0030.99–1.02	0.9590.95–0.97	0.9550.95–0.96	0.9500.95–0.97	1.0010.99–1.01
x-ray	24.1 ms	1.0171.01–1.02	0.9991.00–1.00	1.0201.02–1.03	1.0121.00–1.02	1.0121.00–1.02	1.0051.00–1.01
samba	24.7 ms	1.0171.00–1.02	1.0081.00–1.02	0.9240.92–0.93	0.9130.91–0.92	0.9160.91–0.93	1.0031.00–1.01
zlib-rs deflate L1dickens	50.6 ms	1.0091.00–1.02	1.0070.99–1.01	0.8630.85–0.87	0.8590.85–0.87	0.8620.85–0.87	1.0060.99–1.02
x-ray	57.6 ms	0.9740.94–0.99	0.9740.95–1.01	0.9730.95–0.99	0.9600.93–0.98	0.9670.94–0.98	1.0000.99–1.01
samba	68.3 ms	0.9930.99–1.00	1.0040.99–1.01	0.8210.81–0.83	0.8160.81–0.84	0.8250.81–0.83	0.9970.99–1.01
zlib-rs deflate L6dickens	151 ms	1.0131.01–1.02	0.9950.99–1.00	0.9940.99–1.00	0.9790.98–0.99	0.9820.98–0.98	0.9991.00–1.00
x-ray	126 ms	1.0961.09–1.10	0.9990.99–1.00	1.0611.06–1.07	0.9720.97–0.98	0.9730.96–0.98	0.9980.99–1.00
samba	194 ms	1.0251.02–1.03	1.0011.00–1.00	0.9740.97–0.98	0.9480.95–0.95	0.9530.95–0.96	1.0031.00–1.01
zlib-rs deflate L9dickens	294 ms	0.9950.99–1.00	0.9930.99–1.00	0.9760.97–0.98	0.9820.98–0.98	0.9780.98–0.98	1.0021.00–1.01
x-ray	194 ms	0.9790.98–0.98	0.9790.98–0.98	0.9580.95–0.96	0.9900.99–0.99	0.9720.97–0.98	1.0001.00–1.00
samba	392 ms	0.9890.99–0.99	0.9880.99–0.99	0.9530.95–0.96	0.9630.96–0.97	0.9600.96–0.96	1.0041.00–1.01
lz4_flex decodedickens	4.65 ms	0.9860.98–0.99	0.9770.97–0.98	0.9870.98–0.99	0.9810.98–0.99	0.9830.97–0.98	1.0030.99–1.01
x-ray	2.11 ms	0.9850.97–1.00	0.9760.96–0.98	0.9840.97–1.00	0.9870.98–1.00	0.9660.95–0.98	1.0041.00–1.01
samba	7.57 ms	0.9870.98–0.99	0.9650.96–0.97	0.9770.97–0.98	0.9810.97–0.98	0.9520.95–0.97	1.0010.99–1.01
brotli encodedickens	149 ms	0.9570.95–0.96	0.9900.99–1.00	0.9550.95–0.96	0.9860.98–0.99	0.9870.98–0.99	1.0000.99–1.00
x-ray	149 ms	0.9300.93–0.93	0.9880.99–0.99	0.9210.92–0.92	0.9700.97–0.97	0.9680.96–0.98	0.9991.00–1.00
samba	172 ms	0.9490.94–0.95	0.9860.98–0.99	0.9410.94–0.94	0.9860.98–0.99	0.9840.98–0.99	0.9990.99–1.00
XML (dlmalloc)xml	46.8 ms	0.9970.99–1.00	1.0011.00–1.01	0.9970.99–1.00	0.9980.99–1.00	0.9970.99–1.00	1.0071.00–1.01

Export Table 1 for every browser and capture (with min–max):

**Table 2.** Brotli bytes of each build; under each, raw · gzip. Bold is the shipped build; SIMD instructions are a static count.
Module · SIMD	No wasm-opt	wasm-opt -Os	wasm-opt -O3	SIMD instructions none · -Os · -O3
zstd decoderscalar	43,227145,543 · 53,932	41,465137,334 · 51,179	41,265137,379 · 51,131	0 · 0 · 0
simd128	44,653148,212 · 55,859	42,612140,050 · 52,524	42,496140,173 · 52,634	658 · 664 · 666
zlib-rsscalar	47,980130,845 · 57,481	47,217123,119 · 56,415	47,124123,255 · 56,382	0 · 0 · 0
simd128	48,466130,831 · 57,655	47,670123,194 · 56,427	47,474123,537 · 56,376	523 · 533 · 547
lz4_flexscalar	23,31066,313 · 27,512	22,40459,668 · 26,487	22,42360,129 · 26,496	0 · 0 · 0
simd128	23,62766,978 · 27,960	22,56260,601 · 26,624	22,54061,056 · 26,716	184 · 183 · 183
brotli encoderscalar	378,6541,007,336 · 518,206	345,521945,132 · 527,136	345,923945,156 · 527,080	0 · 0 · 0
simd128	379,7461,008,484 · 518,301	347,733949,621 · 526,455	394,572950,835 · 527,290	2,180 · 2,377 · 2,387
XML processingscalar	16,95848,763 · 20,560	16,04445,905 · 19,315	15,95645,984 · 19,282	0 · 0 · 0
simd128	17,03048,614 · 20,568	16,05445,923 · 19,272	16,02246,000 · 19,259	104 · 115 · 115

Measurement notes

Conditions: Chrome 156.0.8078.80 (headless) on macOS 26.5.2 (25F84); Apple M4 Pro, arm64, 14 logical cores; on AC power, Low Power Mode off. Captured 8 November 2026, 21:47–21:53 UTC. Capture 1 of 2.
Clock: Cross-origin isolated. Timer resolution in a worker 5 µs (median step; finest 5 µs).
Method: 15 samples per row, each a batch of at least 50 ms divided by its repetitions. Untimed warm-up of at least 250 ms and 2 rounds (at most 20). Order rotated by one each round; ratios are paired by round.
Workloads: zstd decode: libzstd 1.5.7 (zstd-sys) decompressing a level-3 frame.
zlib-rs inflate: zlib-rs (via flate2) inflating a level-6 raw deflate stream; explicit simd128 chunk copies.
zlib-rs deflate L1: zlib-rs (via flate2) deflating at level 1 (deflate_quick); explicit simd128 compare256, slide_hash.
zlib-rs deflate L6: zlib-rs (via flate2) deflating at level 6 (deflate_medium); explicit simd128 compare256, slide_hash.
zlib-rs deflate L9: zlib-rs (via flate2) deflating at level 9 (deflate_slow); explicit simd128 compare256, slide_hash.
lz4_flex decode: lz4_flex frame decoder; no explicit SIMD, autovectorization only.
brotli encode: the brotli crate compressing at quality 4 (22-bit window); autovectorization only.
XML document processing (dlmalloc): Application code: the allocation page's workload on Rust's default allocator, parsing, walking and dropping each of the Silesia xml file's 21 XML documents. No explicit SIMD; the simd128 build autovectorizes some loops.
Builds: Six builds of each module: simd128 off or on, then no wasm-opt, -Os or -O3. Every one is Rust's release profile (opt-level 3 (default), lto = true, codegen-units = 1). SIMD on: -C target-feature=+simd128 and, for the C code, -msimd128. Then wasm-bindgen --target web --remove-name-section and wasm-opt -Os or -O3. The shipped build, which every other page runs, is scalar with wasm-opt -Os.
Pairing: Each round runs all seven variants (the six builds and the A/A instance) in a rotating order, so every comparison is a ratio from the same round. Figures and main points count a claim only where both captures make it.
Sizes: Raw bytes of each .wasm after wasm-bindgen and wasm-opt, then gzip level 9 and brotli quality 11 of the same file. Exact, so they need no interval, and the same in every browser. simd128 made every module 0.1%–2.8% larger at wasm-opt -Os (brotli bytes). SIMD instructions are a static count.
Mechanism: zlib-rs 0.6.8, src/deflate/compare256.rs (the wasm32 and NEON paths); V8 at commit 63481c19 (7 November 2026), kX64IBitMask in src/compiler/backend/x64/code-generator-x64.cc and MacroAssembler::I8x16BitMask in src/codegen/arm64/macro-assembler-arm64.cc. The lowering is V8's (Chrome); Firefox and Safari compile wasm with their own compilers, not covered here.
Recaptured: XML document processing (dlmalloc) was recaptured on 9 November 2026, 01:52–01:53 UTC, in Chrome, Firefox and Safari, and replaced the original rows. Its two captures per browser ran back to back in that one session, about 40 seconds apart, so their agreement shows back-to-back reproduction, not independent sessions.
Raw data: Chrome 156, Apple M4 Pro (arm64), capture 1 of 2 (JSON)
Chrome 156, Apple M4 Pro (arm64), capture 2 of 2 (JSON)
Firefox 151, Apple M4 Pro (arm64), capture 1 of 2 (JSON)
Firefox 151, Apple M4 Pro (arm64), capture 2 of 2 (JSON)
Safari 26, Apple M4 Pro (arm64), capture 1 of 2 (JSON)
Safari 26, Apple M4 Pro (arm64), capture 2 of 2 (JSON)
Build: rustc 1.99.1 (58cf8f9c3 2026-10-13); wasi-sdk-29; wasm-opt version 133 (version_133); wasm-bindgen 0.2.130. Release profile: opt-level 3, LTO true, codegen-units 1, target features none (no simd128), wasm-opt -Os.
Libraries: zstd 0.14.3 (libzstd 1.5.7), zlib-rs 0.6.8 (via flate2 1.1.11), lz4_flex 0.14.0, brotli 9.1.1 (decoder brotli-decompressor 6.3.0).
Commit: ef16b9543512
Artifacts: 32 wasm modules, SHA-256
corpus_bg.wasm 139 kB f35f70fcf55f73022df9ef8a7695abe6be8c2eaccd95237e02a3c1ca5eee7f1b
variants/alloc/off-none/alloc_bg.wasm 48.8 kB ab2887573b311b69ff8536abd581ec3a560815a4c547ae6d2060588ccaa262b8
variants/alloc/off-O3/alloc_bg.wasm 46.0 kB a04a127b213d470ce9573c05606d9af933a25a2102a60dc1144d330e541cd3b1
variants/alloc/off-Os/alloc_bg.wasm 45.9 kB 507c0dd00c71371c4fbcd5a45aa96fca5728eae2ae20d598dfa3475a7148c6c3
variants/alloc/on-none/alloc_bg.wasm 48.6 kB a76f3b94028727f68808b0f468261ecc7ba56833bd1619600867042d7331eb60
variants/alloc/on-O3/alloc_bg.wasm 46.0 kB 64db27d174fcbb18f60734f3dde24ce8af7fa9ffa83c4a6943e2769fad2d80d7
variants/alloc/on-Os/alloc_bg.wasm 45.9 kB 19f91020ca38978c64dddb84e420771a8277123da9bb5996d4f22689f68f3be2
variants/br_write/off-none/br_write_bg.wasm 1.01 MB b67075e0d5b787b9d7298bc0cf8656981a771c3507975b450fb0b9f3658ca803
variants/br_write/off-O3/br_write_bg.wasm 945 kB 5e7d92a860070f4ce7cb054db55bc0d96e1847b567baf4a02d2cab26982bac14
variants/br_write/off-Os/br_write_bg.wasm 945 kB 5ca1c6f2cdfff4f1b0c79bec77176c14545c9dddb17e4ec94d5eaa413522ce2e
variants/br_write/on-none/br_write_bg.wasm 1.01 MB 334e8d106fefeda519ec915265b9d2e52dac0391b97152d204f5d8997b796af9
variants/br_write/on-O3/br_write_bg.wasm 951 kB 2d041090b38371dee7097b29182387dd438544235c2c16a3abddc0332ef21e76
variants/br_write/on-Os/br_write_bg.wasm 949 kB d0852a8ec562a1adc8faa16247787dc5d2a3a1ef516b484d745c8dad6f1cea81
variants/lz4/off-none/lz4_bg.wasm 66.2 kB 6217191dda9baeb80f2714bc330c7fa63b541464ceb9fb14d01cf50be41e2244
variants/lz4/off-O3/lz4_bg.wasm 60.1 kB 470263fb17ff3b89fef5fcf7f5ff7d1a7763b9109a967c1982b4e0f1c0618a83
variants/lz4/off-Os/lz4_bg.wasm 59.6 kB ebe01820ec888b26467fd0303cfaed85cea6c29f6c7980d4e088525a632646dd
variants/lz4/on-none/lz4_bg.wasm 66.9 kB 6751223b53f1c4b501418d71d5cf765315c37cf256d7d2e849339bce8c64ba16
variants/lz4/on-O3/lz4_bg.wasm 61.0 kB 9bc4dfd68a3de30d0d29abb2e3df0a79430b5a1e12532bf33721bfc3cae89450
variants/lz4/on-Os/lz4_bg.wasm 60.5 kB f2a2b9037cf9fb8e3c52b6bd32cd23de17929572dd8fc99096557a082b8ac0a5
variants/zlib_rs/off-none/zlib_rs_bg.wasm 131 kB e1c5d4788c99c4dcbb1b6467b7853f6598564c2a420d8cd509c816f27d8d3b51
variants/zlib_rs/off-O3/zlib_rs_bg.wasm 123 kB b4a9b727c2bd23ddb3a062b4ce4fd1b26c1e7b1c87bb98a78acb1991aedbc7b1
variants/zlib_rs/off-Os/zlib_rs_bg.wasm 123 kB a738fdcf343c7046d34c98896be663ac15df54b9a711bd4b20513db39981d69c
variants/zlib_rs/on-none/zlib_rs_bg.wasm 131 kB 812730d272912bdd676845b07ca51b978a40ef2c2910887ce5cf36d0804b38d9
variants/zlib_rs/on-O3/zlib_rs_bg.wasm 123 kB 1b2964f7dffe1502aa0f419534ef990b97c9baf3304e7f5093f2b9a9fd02cce9
variants/zlib_rs/on-Os/zlib_rs_bg.wasm 123 kB ca511fcb65faa6cbc87fcb35cbeb2bd7f4939f84371bf7c9963570821e6e9da1
variants/zstd_read/off-none/zstd_read_bg.wasm 146 kB 4157ae9ede9414c99a8ecc7470719a9003624917575aee86e5cd3dc3ceac828b
variants/zstd_read/off-O3/zstd_read_bg.wasm 137 kB 612669c809a6da58f075454a3fe4f8162f75fc34c8e55c21a51062d690f2500f
variants/zstd_read/off-Os/zstd_read_bg.wasm 137 kB 29f5cc2a376640aec12ba67b5cf7ad25c344f42688f552ac3759975f101eb4cc
variants/zstd_read/on-none/zstd_read_bg.wasm 148 kB dfd55dc429050955e70286a86b5dfd372acd702f110b6a5cc8b2c2b5bbeab17b
variants/zstd_read/on-O3/zstd_read_bg.wasm 140 kB f2784feb7b2b1b3e1ae89033bd393658eceb355c38e2aa70f6bb9b80df58a538
variants/zstd_read/on-Os/zstd_read_bg.wasm 140 kB 73c1ae63834110234945de0a4f5462fc46bea1a2052f3b4ed34a4cf2a5e85328
zstd_write_bg.wasm 464 kB b9dcb6036757d7b1ac36f4516beabf0cdf25f0c1439e63af7839d3f4b9b337ff
Corpus: Silesia corpus subset (3 of 12 files), checked against pinned SHA-256 hashes when packed and by length when loaded.
dickens 10,192,446 bytes b24c37886142e11d0ee687db6ab06f936207aa7f2ea1fd1d9a36763c7a507e6a
x-ray 8,474,240 bytes 7de9fce1405dc44ae5e6813ed21cd5751e761bd4265655a005d39b9685d1c9ad
samba 21,606,400 bytes 93ba07bc44d8267789c1d911992f40b089ffa2140b4a160fac11ccae9a40e7b2
Raw data: Raw per-round data (JSON)

Reference edition 2026-11-08· Apple M4 Pro (arm64), macOS 26.5.2 (25F84)· Chrome 156, Firefox 151 and Safari 26· 15 rounds per row, two independent captures each· XML (dlmalloc) captured back to back· Method· Raw data (JSON)