This cleans up the API in general, removing a bunch of deprecated stuff,
cleaning up circular imports, etc.
But the biggest change is switching to an optional XML format for the
regex lexer.
Having lexers defined only in Go is not ideal for a couple of reasons.
Firstly, it impedes a significant portion of contributors who use Chroma
in Hugo, but don't know Go. Secondly, it bloats the binary size of any
project that imports Chroma.
Why XML? YAML is an abomination and JSON is not human editable. XML
also compresses very well (eg. Go template lexer XML compresses from
3239 bytes to 718).
Why a new syntax format? All major existing formats rely on the
Oniguruma regex engine, which is extremely complex and for which there
is no Go port.
Why not earlier? Prior to the existence of fs.FS this was not a viable
option.
Benchmarks:
$ hyperfine --warmup 3 \
'./chroma.master --version' \
'./chroma.xml-pre-opt --version' \
'./chroma.xml --version'
Benchmark 1: ./chroma.master --version
Time (mean ± σ): 5.3 ms ± 0.5 ms [User: 3.6 ms, System: 1.4 ms]
Range (min … max): 4.2 ms … 6.6 ms 233 runs
Benchmark 2: ./chroma.xml-pre-opt --version
Time (mean ± σ): 50.6 ms ± 0.5 ms [User: 52.4 ms, System: 3.6 ms]
Range (min … max): 49.2 ms … 51.5 ms 51 runs
Benchmark 3: ./chroma.xml --version
Time (mean ± σ): 6.9 ms ± 1.1 ms [User: 5.1 ms, System: 1.5 ms]
Range (min … max): 5.7 ms … 19.9 ms 196 runs
Summary
'./chroma.master --version' ran
1.30 ± 0.23 times faster than './chroma.xml --version'
9.56 ± 0.83 times faster than './chroma.xml-pre-opt --version'
A slight increase in init time, but I think this is okay given the
increase in flexibility.
And binary size difference:
$ du -h lexers.test*
$ du -sh chroma* 951371ms
8.8M chroma.master
7.8M chroma.xml
7.8M chroma.xml-pre-opt
Benchmarks:
$ hyperfine --warmup 3 \
'./chroma.master --version' \
'./chroma.xml-pre-opt --version' \
'./chroma.xml --version'
Benchmark 1: ./chroma.master --version
Time (mean ± σ): 5.3 ms ± 0.5 ms [User: 3.6 ms, System: 1.4 ms]
Range (min … max): 4.2 ms … 6.6 ms 233 runs
Benchmark 2: ./chroma.xml-pre-opt --version
Time (mean ± σ): 50.6 ms ± 0.5 ms [User: 52.4 ms, System: 3.6 ms]
Range (min … max): 49.2 ms … 51.5 ms 51 runs
Benchmark 3: ./chroma.xml --version
Time (mean ± σ): 6.9 ms ± 1.1 ms [User: 5.1 ms, System: 1.5 ms]
Range (min … max): 5.7 ms … 19.9 ms 196 runs
Summary
'./chroma.master --version' ran
1.30 ± 0.23 times faster than './chroma.xml --version'
9.56 ± 0.83 times faster than './chroma.xml-pre-opt --version'
Incompatible changes:
- (*RegexLexer).SetAnalyser: changed from func(func(text string) float32) *RegexLexer to func(func(text string) float32) Lexer
- (*TokenType).UnmarshalJSON: removed
- Lexer.AnalyseText: added
- Lexer.SetAnalyser: added
- Lexer.SetRegistry: added
- MustNewLazyLexer: removed
- MustNewLexer: changed from func(*Config, Rules) *RegexLexer to func(*Config, func() Rules) *RegexLexer
- Mutators: changed from func(...Mutator) MutatorFunc to func(...Mutator) Mutator
- NewLazyLexer: removed
- NewLexer: changed from func(*Config, Rules) (*RegexLexer, error) to func(*Config, func() Rules) (*RegexLexer, error)
- Pop: changed from func(int) MutatorFunc to func(int) Mutator
- Push: changed from func(...string) MutatorFunc to func(...string) Mutator
- TokenType.MarshalJSON: removed
- Using: changed from func(Lexer) Emitter to func(string) Emitter
- UsingByGroup: changed from func(func(string) Lexer, int, int, ...Emitter) Emitter to func(int, int, ...Emitter) Emitter
The bashsession lexer did only recognized command prompts starting with
one of the characters '#$%>'. This commit adds support for a prefixed
block of the form `[.*@.*]`, which is a common layout for command prompts.
Apparently only "phrasing content" are permitted inside `pre` and
`code`, and `div` is not one of them, so replace it with `span`, `div`
wasn't necessary anyway.
Also it makes sense to put code inside `code` tag(except for line
numbers).
Removed `overflow:auto` and `width: auto` as they didn't seem to be
necessary and actually prevented horizontal scroll bar to appear when content
didn't fit in the viewport.
This package appears to be old and unmaintained, and it gets pulled into the
go.sum of every single package importing Chroma. Also fnmatch doesnt seem to do
any error checking on the pattern, while the standard package does. Also, the
standard package path/filepath is already imported, so it appears not to cost
anything to change it.
JSON spec doesn't allow comments, but many implementation support them anyway. Having comments is very convenient, for example, when showing code in blog posts and articles.
Huge hardcoded character lists have a cost. In some current lexers, such
as Haskell and JavaScript, this bloats lexer init time to 60-80ms on
some current systems, as opposed to sub-ms for many others.
Replacing them with character classes such as `\p{L}` seems to
eliminate this cost, reducing lexer init time to the norm (around 1ms).
In addition, this significantly reduces and simplifies the code.
The current tests pass, but there may be inaccuracies not covered by
tests. This requires a review.
This change is likely to cause edge case regressions, as the sets of
characters considered "letters" vary between languages. However, Chroma
lexers don't aim to be perfectly accurate. Performance should be just as
much a goal as accuracy. I believe this tradeoff to be justified.
This commit leaves at least two lexers unfixed: Julia and Kotlin.
Judging by the code, they might have the same issue, and should also be
addressed.
