Files
crosspoint-reader-mod/docs/hyphenation-trie-format.md
Xuan-Son Nguyen 0508bfc1f7 perf: apply (micro) optimization on SerializedHyphenationPatterns (#689)
## Summary

This PR applies a micro optimization on `SerializedHyphenationPatterns`,
which allow reading `rootOffset` directly without having to parse then
cache it.

It should not affect storage space since no new bytes are added.

This also gets rid of the linear cache search whenever
`liangBreakIndexes` is called. In theory, the performance should be
improved a bit, although it may be too small to be noticeable in
practice.

## Testing

master branch:

```
english: 99.1023%
french: 100%
german: 97.7289%
russian: 97.2167%
spanish: 99.0236%
```

This PR:

```
english: 99.1023%
french: 100%
german: 97.7289%
russian: 97.2167%
spanish: 99.0236%
```

---

### AI Usage

While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.

Did you use AI tools to help write this code? PARTIALLY - mostly IDE
tab-autocompletions
2026-02-16 20:27:43 +11:00

2.0 KiB
Raw Permalink Blame History

Hypher Binary Tries

CrossPoint embeds the exact binary automata produced by Typst's hypher.

File layout

Each .bin blob is a single self-contained automaton:

uint32_t root_addr_be;  // big-endian offset of the root node
uint8_t  levels[];      // shared "levels" tape (dist/score pairs)
uint8_t  nodes[];       // node records packed back-to-back

The size of the levels tape is implicit. Individual nodes reference slices inside that tape via 12-bit offsets, so no additional pointers are required.

Node encoding

Every node starts with a single control byte:

  • Bit 7 set when the node stores scores (levels).
  • Bits 5-6 stride of the target deltas (1, 2, or 3 bytes, big-endian).
  • Bits 0-4 transition count (values ≥ 31 spill into an extra byte).

If the levels flag is set, two more bytes follow. Together they encode a 12-bit offset into the global levels tape and a 4-bit length. Each byte in the levels tape packs a distance/score pair as dist * 10 + score, where dist counts how many UTF-8 bytes we advanced since the previous digit.

After the optional levels header come the transition labels (one byte per edge) followed by the signed target deltas. Targets are stored as relative offsets from the current node address. Deltas up to ±128 fit in a single byte, larger distances grow to 2 or 3 bytes. The runtime walks the transitions with a simple linear scan and materializes the absolute address by adding the decoded delta to the current nodes base.

Embedding blobs into the firmware

The helper script scripts/generate_hyphenation_trie.py acts as a thin wrapper: it reads the hypher-generated .bin files, formats them as constexpr byte arrays, and emits headers under lib/Epub/Epub/hyphenation/generated/. Each header defines the raw data plus a SerializedHyphenationPatterns descriptor so the reader can keep the automaton in flash.

A convenient script update_hyphenation.sh is used to update all languages. To use it, run:

./scripts/update_hypenation.sh