- Quick start: MMark vs GitHub-flavored markdown
- MMark and Common Mark
- About MMark-specific extensions
- Related packages
MMark (read “em-mark”) is a strict markdown processor for writers. “Strict” means that not every input is considered valid markdown document and parse errors are possible and even desirable, because they allow us to spot markup issues without searching for them in rendered document. If a markdown document passes MMark parser, then it’ll likely produce HTML without quirks. This feature makes it a good choice for writers and bloggers.
MMark in its current state features:
A parser that produces high-quality error messages and does not choke on first parse error. It is capable of reporting many parse errors where makes sense.
An extension system allowing to create extensions that alter parsed markdown document in some way.
There is also a blog post announcing the project:
Quick start: MMark vs GitHub-flavored markdown
It’s easy to start using MMark if you’re used to GitHub-flavored markdown. There are four main differences:
URIs are not automatically recognized, you must enclose them in
Block quotes require only one
>and they continue as long as the inner content is indented.
This is OK:
> Here goes my block quote. And this is the second line of the quote.
This produces two block quotes:
> Here goes my block quote. > And this is another block quote!
HTML blocks and inline HTML are not supported.
MMark and Common Mark
MMark mostly tries to follow the Common Mark specification as given here:
However, due to the fact that we do not allow inputs that do not make sense,
and also try to guard against common mistakes (like writing
and having it rendered as a paragraph starting with hashes) MMark obviously
can’t follow the specification precisely. In particular, parsing of inlines
differs considerably from Common Mark (see below).
Another difference between Common Mark and MMark is that the latter supports more (pun alert) common markdown extensions out-of-the-box. In particular, MMark supports:
- parsing of an optional YAML block
- strikeout using
- superscript using
- subscript using
- automatic assignment of ids to headers
- pipe tables (as on GitHub)
One does not need to enable or tweak anything for these to work, they are built-in features.
Differences in inline parsing
Emphasis and strong emphasis is an especially hairy topic in the Common Mark
specification. There are 17 ad-hoc rules defining interaction between
_ -based emphasis and more than an half of all Common Mark examples
(that’s about 300) test just this tricky logic.
Not only it is hard to implement, it’s hard to understand for humans too. For example, this input:
results in the following HTML:
(Note the nested emphasis.)
Could it produce something like this instead?
Well, why not? Without remembering those 17 ad-hoc rules, there going to be a lot of tricky cases when a user won’t be able to tell how markdown will be parsed.
I decided to make parsing of emphasis, strong emphasis, and similar constructs like strikethrough, subscript, and superscript more symmetric and less ad-hoc. In 99% of practical cases it is identical to Common Mark, and normal markdown intuitions will work OK for the users.
Let’s start by dividing all characters into four groups:
Space characters, including space, tab, newline, carriage return, and other characters like non-breaking space.
Markup characters, including the following:
]. These are used for markup and whenever they appear in a document, they must form valid markup constructions. To be used as ordinary punctuation characters they must be backslash escaped.
Punctuation characters, which include all punctuation characters that are not markup characters.
Other characters, which include all characters not falling into the three groups described above.
Next, let’s assign levels to all groups but markup characters:
- Space characters—level 0
- Punctuation characters—level 1
- Other characters—level 2
When markup characters or punctuation characters are escaped with backslash they become other characters.
We’ll call markdown characters placed between a character of level
and a character of level
R left-flanking delimiter run if and only if:
level(L) < level(R)
These markup characters sort of hang on the left hand side of a word.
Similarly we’ll call markdown characters placed between a character of
L and a character of level
R right-flanking delimiter run if and
level(L) > level (R)
These markup characters hang on the right hand side of a word.
Emphasis markup (and other similar things like strikethrough, which we won’t mention explicitly anymore for brevity) can start only as left-flanking delimiter run and end only as right-flanking delimiter run.
This produces a parse error:
*Something * is not right. Something __is __ not right.
And this too:
This means that inter-word emphasis is not supported by this approach.
The next example is OK because
s is an other character and
. is a
punctuation character, so
level('s') > level('.').
Here it *goes*.
In some rare cases backslash escaping can help get the right result:
Here goes *(something\)*.
We escaped the closing parenthesis
) so it becomes an other character
with level 2 and so its level is greater than the level of plain punctuation
- If a line starts with hash signs it is expected to be a valid non-empty header (level 1–6 inclusive). If you want to start a paragraph with hashes, just escape the first hash with backslash and that will be enough.
- Setext headings are not supported for the sake of simplicity.
- Fenced code blocks must be explicitly closed by a closing fence. They are not closed by the end of document or by start of another block.
- Lists and block quotes are defined by column at which their content starts. Content belonging to a particular list or block quote should start at the same column (or greater column, up to the column where indented code blocks start). As a consequence of this, block quotes do not feature “laziness”.
- Block quotes are started by a single
>character, it’s not necessary to put a
>character at beginning of every line belonging to a quote (in fact, this would make every line a separate block quote).
- Paragraphs can be interrupted by unordered and ordered lists with any valid starting index.
- HTML blocks are not supported because the syntax conflicts with autolinks and the feature is a hack to compensate for the lack of extensibility and customization in the original markdown.
- MMark does not support hard line breaks represented as double space before newline. Nevertheless, hard line breaks in the form of backslash before newline are supported (these are more explicit too).
- All URI references (in links, images, autolinks, etc.) are parsed as per
RFC 3986, no support for escaping or support for entity and numeric
character references is provided. In addition to that, when a URI
reference in not enclosed with
>, then closing parenthesis character
)is not considered part of URI (use
<uri>syntax if you want a closing parenthesis as part of a URI). Since the empty string is a valid URI and it may be confusing in some cases, we also force the user to write
<>to represent the empty URI.
- Putting links in text of another link is not allowed, i.e. no nested links is possible.
- Putting images in description of other images is not allowed (similarly to the situation with links).
- HTML inlines are not supported for the same reason why HTML blocks are not supported.
About MMark-specific extensions
- YAML block must start with three hyphens
---and end with three hyphens
---. It can only be placed at the beginning of a markdown document. Trailing white space after the
---sequences is allowed.
I have compared speed and memory consumption of various Haskell markdown libraries by running them on an identical, big-enough markdown document and by rendering it as HTML:
|Library||Parsing library||Execution time||Allocated||Max residency|
||Custom C code||323.4 μs||228,440||9,608|
||Custom Haskell code||10.76 ms||44,686,272||799,200|
Results are ordered from fastest to slowest.
markdown library is sloppy and parses markdown incorrectly. For
example, it parses the following
*My * text as an inline containing
emphasis, while in reality both asterisks must form flanking delimiter runs
to create emphasis, like so
*My* text. This allowed
markdown to get away
with a far simpler approach to parsing at the price that it’s not really a
valid markdown implementation.
mmark-extcontains some commonly useful MMark extensions.
mmark-cliis a command line interface to MMark.
flycheck-mmarkis a way to check markdown documents against MMark parser interactively from Emacs.
Issues, bugs, and questions may be reported in the GitHub issue tracker for this project.
Pull requests are also welcome.
Copyright © 2017–present Mark Karpov
Distributed under BSD 3 clause license.
Uses Megaparsec 8.0.0.
Dropped suppot for GHC 8.2.
Dropped support for GHC 8.0 and older.
- Added GHCJS support by making
yamldependency optional. With GHCJS a yaml block simply always returns the empty object.
- Fixed setting offset after parsing of collapsed reference links. Previously offset in parser state was restored incorrectly and errors that would happen after such links would be reported two characters before their real position.
Uses Megaparsec 7. The
parsefunction now returns
- Improved parse errors related to the optional YAML block.
inlineTransare applied to deeply nested elements too, not only top-level elements.
- Fixed the bug in parser which signalled a parse error when YAML block was followed by more than one newline without markdown content after it.
- Empty autolinks are now disallowed.
<>will result in literal
<>in resulting HTML.
- Now HTML is escaped properly inside inline code spans.
- Fixed the bug that prevented application of rendering extensions to sub-blocks (blocks contained inside other blocks) and sub-inlines (inlines contained inside other inlines).
The parser can now recover from block-level parse errors in tables and continue parsing.
Pipes in code spans in table cells are not considered as table cell delimiters anymore.
Table sub-parser now faster rejects inputs that do not look like a table, this improves overall performance.
Better handling of the cases when a block can be interpreted as a list and as a table at the same time.
Added a dummy
Showinstance for the
- Compiles with
- Made parsing of emphasis-like markup more flexible and forgiving, see
README.mdfor more information.
- This version uses
parser-combinators-0.4.0and has improved performance.
Added support for pipe tables (like on GitHub).
Fixed a nasty space leak in the parser, made it faster too.
Empty strings are not parsed as URIs anymore (even though a valid URI may be represented as the empty string). Instead, it’s now possible to write an empty URI using the
<>syntax (which previously was not recognized as a URI in some contexts).
Improved parse errors related to parsing of titles in links, images, and reference definitions.
Parsing of reference definitions now can recover from failures, so the parser doesn’t choke on malformed reference definitions anymore.
Reduced allocations and improved speed of the parser significantly.
Fixed a couple of bugs in the parser for reference definitions.
Now link and image titles may contain newline character as per the Common Mark spec.
Code can interrupt paragraphs now, as per Common Mark spec.
Implemented parsing of reference-links (including collapsed and shortcut-style links).
Implemented parsing of reference-style images (including collapsed and shortcut-style images).
Added support for entity and numeric references (section 6.2 of the Common Mark spec).
Improved quality of parse errors.
Improved performance of the parser. Mainly the inline-level parser to be precise. The result is that now there are 3× less allocations and the code runs about 3× faster on paragraphs and block quotes (it’s about 2.5× faster for a big realistic document).
Improved quality of parse errors.
Now punctuation is stripped from header ids in
Added support for block quotes.
Added support for unordered and ordered lists.
- Fixed a bug in skipping of headers (only one newline after the header line was picked, not all white space up to next block).
- Initial release.