rustybuzz

rustybuzz is a complete harfbuzz's shaping algorithm port to Rust.

Matches harfbuzz v9.0.0

Why?

Because you can add rustybuzz = "*" to your project and it just works. No need for a C++ compiler. No need to configure anything. No need to link to system libraries.

Conformance

rustybuzz passes nearly all of harfbuzz shaping tests (2221 out of 2252 to be more precise). So it's mostly identical, but there are still some tiny edge-cases which are not implemented yet or cannot be implemented at all.

Major changes

• Subsetting removed.
• TrueType parsing is completely handled by the ttf-parser. And while the parsing algorithm is very different, it's not better or worse, just different.
• Malformed fonts will cause an error. HarfBuzz uses fallback/dummy shaper in this case.
• No font size property. Shaping is always using UnitsPerEm. You should scale the result manually.
• Most of the TrueType and Unicode handling code was moved into separate crates.
• rustybuzz doesn't interact with any system libraries and must produce exactly the same results on all OS'es and targets.
• mort table is not supported, since it's deprecated by Apple.
• No Arabic fallback shaper, since it requires subsetting.
• No graphite library support.
• avar2 as well as other parts of the boring-expansion-spec are not supported yet.

Performance

At the moment, performance isn't that great. We're 1.5-2x slower than harfbuzz. Also, rustybuzz doesn't support shaping plan caching at the moment.

See benches/README.md for details.

Notes about the port

rustybuzz is not a faithful port.

harfbuzz can roughly be split into 6 parts: shaping, subsetting, TrueType parsing, Unicode routines, custom containers and utilities (harfbuzz doesn't use C++ std) and glue for system/3rd party libraries. In the mean time, rustybuzz contains only shaping. All of the TrueType parsing was moved to the ttf-parser. Subsetting was removed. Unicode code was mostly moved to external crates. We don't need custom containers because Rust's std is good enough. And we do not use any non Rust libraries, so no glue code either.

In the end, we still have around 23 KLOC. While harfbuzz is around 80 KLOC.

Lines of code

As mentioned above, rustybuzz has around 23 KLOC. But this is not strictly true, because there are a lot of auto-generated data tables.

You can find the "real" code size using:

tokei --exclude hb/unicode_norm.rs --exclude hb/ot_shaper_vowel_constraints.rs \
      --exclude '*_machine.rs' --exclude '*_table.rs' src

Which gives us around 17 KLOC, which is still a lot.

Future work

Since the port is finished, there is not much to do other than syncing it with a new harfbuzz releases. However, there is still lots of potential areas of improvement:

• Better benchmarking: Currently, our benchmarking harness is very barebones and only contains 3 different tests to compare performance against harfbuzz. Ideally, this benchmark suite should be extended further with both, more benching functions and also a wider variety of fonts/cases, so that we can have more insight into how much worse performance is in certain cases, compared to harfbuzz.
• Wider test coverage: Currently, we only test the result of the positioned glyphs in the shaping output. We should add tests so that we can test other parts of the API as well, such as glyph extents and glyph flags.
• Custom tests: Currently, we only rely on the harfbuzz test suite to assess the correctness of rustybuzz. While this is already very good, we should extend our current testing mechanism with infrastructure so that bugs specific to rustybuzz can be tested as well.
• Fuzzing against harfbuzz: While rustybuzz passes the whole harfbuzz test suite, this does not mean that output will always be 100% identical to harfbuzz. Given the complexity of the code base, there are bound to be other bugs that just have not been discovered yet. One potential way of addressing this issue could be to create a fuzzer that takes random fonts, and shapes them with a random set of Unicode codepoints as well as input settings. In case of a discovered discrepancy, this test case could then be investigated and once the bug has been identified, added to our custom test suite. On the one hand we could use the Google Fonts font collection for this so that the fonts can be added to the repository, but we could also just use MacOS/Windows system fonts and only test them in CI, similarly to how it's currently done for AAT in harfbuzz.
• Performance: harfbuzz contains tons of optimization structures (accelerators and caches) which have not been included in rustybuzz. As a result of this, performance is much worse in many cases, as mentioned above (although in the grand scheme of things rustybuzz is still very performant), but the upside of excluding all those optimizations is that the code base is much simpler and straightforward. This makes it a lot easier to backport new changes (which already is a very difficult task). Now that we are back in sync with harfbuzz, we can consider attempting to port some of the major optimization improvements from harfbuzz, but we should do so carefully to not make it even harder to keep the code bases in sync.
• Code alignment: rustybuzz tries its best to make the code base look like a 1:1 C++ to Rust translation, which is the case for most parts of the code, but there also are many variations (most notable in AAT) that arise from the fact that a lot of the C++ concepts are not straightforward to port. Nevertheless, there probably still are parts of the code that probably could be made more similar, given that someone puts time into looking into that.

All of this is a lot of work, so contributions are more than welcome.

Safety

The library is completely safe.

We do have one unsafe to cast between two POD structures, which is perfectly safe. But except that, there are no unsafe in this library and in most of its dependencies (excluding bytemuck).

Alternatives

• harfbuzz_rs - bindings to the actual harfbuzz library. As of v2 doesn't expose subsetting and glyph outlining, which harfbuzz supports.
• allsorts - shaper and subsetter. As of v0.6 doesn't support variable fonts and Apple Advanced Typography. Relies on some unsafe code.
• swash - Supports variable fonts, text layout and rendering. No subsetting. Relies on some unsafe code. As of v0.1.4 has zero tests.

License

rustybuzz is licensed under the MIT.

harfbuzz is licensed under the Old MIT