in lib/ace/mode/rst_highlight_rules.js [38:292]
var RSTHighlightRules = function() {
var tokens = {
title: "markup.heading",
list: "markup.heading",
table: "constant",
directive: "keyword.operator",
entity: "string",
link: "markup.underline.list",
bold: "markup.bold",
italic: "markup.italic",
literal: "support.function",
comment: "comment"
};
var startStringPrefix = "(^|\\s|[\"'(<\\[{\\-/:])";
var endStringSuffix = "(?:$|(?=\\s|[\\\\.,;!?\\-/:\"')>\\]}]))";
this.$rules = {
"start": [
/* NB: Only the underline of the heading is highlighted.
* ACE tokenizer does not allow backtracking, the underlined text of the
* heading cannot be highlighted. */
{
token : tokens.title,
regex : "(^)([\\=\\-`:\\.'\"~\\^_\\*\\+#])(\\2{2,}\\s*$)"
},
/* Generic directive syntax (e.g. .. code-block:: c)
* All of the directive body is highlighted as a code block. */
{
token : ["text", tokens.directive, tokens.literal],
regex : "(^\\s*\\.\\. )([^: ]+::)(.*$)",
next : "codeblock"
},
{
token : tokens.directive,
regex : "::$",
next : "codeblock"
},
/* Link/anchor definitions */
{
token : [tokens.entity, tokens.link],
regex : "(^\\.\\. _[^:]+:)(.*$)"
},
{
token : [tokens.entity, tokens.link],
regex : "(^__ )(https?://.*$)"
},
/* Footnote definition */
{
token : tokens.entity,
regex : "^\\.\\. \\[[^\\]]+\\] "
},
/* Comment block start */
{
token : tokens.comment,
regex : "^\\.\\. .*$",
next : "comment"
},
/* List items */
{
token : tokens.list,
regex : "^\\s*[\\*\\+-] "
},
{
token : tokens.list,
regex : "^\\s*(?:[A-Za-z]|[0-9]+|[ivxlcdmIVXLCDM]+)\\. "
},
{
token : tokens.list,
regex : "^\\s*\\(?(?:[A-Za-z]|[0-9]+|[ivxlcdmIVXLCDM]+)\\) "
},
/* "Simple" tables */
{
token : tokens.table,
regex : "^={2,}(?: +={2,})+$"
},
/* "Grid" tables */
{
token : tokens.table,
regex : "^\\+-{2,}(?:\\+-{2,})+\\+$"
},
{
token : tokens.table,
regex : "^\\+={2,}(?:\\+={2,})+\\+$"
},
/* Inline markup */
{
token : ["text", tokens.literal],
regex : startStringPrefix + "(``)(?=\\S)",
next : "code"
},
{
token : ["text", tokens.bold],
regex : startStringPrefix + "(\\*\\*)(?=\\S)",
next : "bold"
},
{
token : ["text", tokens.italic],
regex : startStringPrefix + "(\\*)(?=\\S)",
next : "italic"
},
/* Substitution reference */
{
token : tokens.entity,
regex : "\\|[\\w\\-]+?\\|"
},
/* Link/footnote references */
{
token : tokens.entity,
regex : ":[\\w-:]+:`\\S",
next : "entity"
},
{
token : ["text", tokens.entity],
regex : startStringPrefix + "(_`)(?=\\S)",
next : "entity"
},
{
token : tokens.entity,
regex : "_[A-Za-z0-9\\-]+?"
},
{
token : ["text", tokens.link],
regex : startStringPrefix + "(`)(?=\\S)",
next : "link"
},
{
token : tokens.link,
regex : "[A-Za-z0-9\\-]+?__?"
},
{
token : tokens.link,
regex : "\\[[^\\]]+?\\]_"
},
{
token : tokens.link,
regex : "https?://\\S+"
},
/* "Grid" tables column separator
* This is at the end to make it lower priority over all other rules. */
{
token : tokens.table,
regex : "\\|"
}
],
/* This state is used for all directive bodies and literal blocks.
* The parser returns to the "start" state when reaching the first
* non-empty line that does not start with at least one space. */
"codeblock": [
{
token : tokens.literal,
regex : "^ +.+$",
next : "codeblock"
},
{
token : tokens.literal,
regex : '^$',
next: "codeblock"
},
{
token : "empty",
regex : "",
next : "start"
}
],
/* Inline code
* The parser returns to the "start" state when reaching "``" */
"code": [
{
token : tokens.literal,
regex : "\\S``" + endStringSuffix,
next : "start"
},
{
defaultToken: tokens.literal
}
],
/* Bold (strong) text
* The parser returns to the "start" state when reaching "**" */
"bold": [
{
token : tokens.bold,
regex : "\\S\\*\\*" + endStringSuffix,
next : "start"
},
{
defaultToken: tokens.bold
}
],
/* Italic (emphasis) text
* The parser returns to the "start" state when reaching "*" */
"italic": [
{
token : tokens.italic,
regex : "\\S\\*" + endStringSuffix,
next : "start"
},
{
defaultToken: tokens.italic
}
],
/* Explicit role/class text or link anchor definition
* The parser returns to the "start" state when reaching "`" */
"entity": [
{
token : tokens.entity,
regex : "\\S`" + endStringSuffix,
next : "start"
},
{
defaultToken: tokens.entity
}
],
/* Link reference
* The parser returns to the "start" state when reaching "`_" or "`__" */
"link": [
{
token : tokens.link,
regex : "\\S`__?" + endStringSuffix,
next : "start"
},
{
defaultToken: tokens.link
}
],
/* Comment block.
* The parser returns to the "start" state when reaching the first
* non-empty line that does not start with at least one space. */
"comment": [
{
token : tokens.comment,
regex : "^ +.+$",
next : "comment"
},
{
token : tokens.comment,
regex : '^$',
next: "comment"
},
{
token : "empty",
regex : "",
next : "start"
}
]
};
};