fast-normalize-spaces
$ yarn add @shelf/fast-normalize-spaces
const {normalizeSpaces} = require('@shelf/fast-normalize-spaces');
normalizeSpaces(' hello \n\n\n \n \n \t world ');
// 'hello world'Benchmarks reuse the same pool of 45 worst-case scenarios that cover multilingual text, surrogate pairs, HTML-like tokens, and the full 2018 Unicode whitespace set.
| Scenario | normalize-space-x | @shelf/fast-normalize-spaces | Speedup |
|---|---|---|---|
| ~33 kb | 2,772 ops/s, ±0.22% | 15,880 ops/s, ±0.34% | ~5.7x |
| ~330 kb | 270 ops/s, ±0.50% | 1,539 ops/s, ±1.49% | ~5.7x |
| ~3.3 mb | 20 ops/s, ±1.62% | 152 ops/s, ±0.36% | 7.6x |
| ~33 mb | 2 ops/s, ±6.05% | 16 ops/s, ±0.76% | 8.0x |
You can run yarn benchmark:speed to test on your own.
| Text size (UTF-8) | normalize-space-x | @shelf/fast-normalize-spaces | Improvement |
|---|---|---|---|
| ~33 mb (34,603,010 bytes) | 74.69mb | 25.31mb | 2.95x less |
The larger the string, the bigger the gap. Memory usage stays close to the size of the input buffer.
September 2025 improvements were delivered autonomously by the gpt-5-codex model. We treated the normalization routine like any critical path service and tightened the slowest sections:
- Smarter lookup table – precomputes the Unicode whitespace bitmap by iterating only over the relevant code points, keeping startup cost small and lookups cache-friendly.
- Single-pass whitespace collapse – streams over the text once and writes normalized characters immediately, eliminating the prior buffer-wide fill and cutting per-call writes by roughly half.
- Early return for clean inputs – detects unchanged strings and returns them as-is, removing allocations when input already meets expectations.
- Lean buffer management – trims trailing whitespace in place, which dropped peak RSS from ~44 MB to ~25 MB on the 33 MB payloads.
The result is a jump from ~10k ops/s to 15.8k ops/s on 33 KB payloads and 5.7–8.0× gains over normalize-space-x, with memory use reduced nearly threefold.
Set a custom payload by exporting TEXT_SIZE (in bytes) when running the benchmark:
TEXT_SIZE=$((10 * 1024 * 1024)) yarn benchmark:memory$ git checkout master
$ yarn version
$ yarn publish
$ git push origin master --tagsMIT © Shelf