Toggle menu
Toggle preferences menu
Toggle personal menu
Not logged in
Your IP address will be publicly visible if you make any edits.

Module:String2

From the Star Citizen Wiki, the fidelity™ encyclopedia
Module documentation[view][edit][history][purge]
This documentation is transcluded from Module:String2/doc. Changes can be proposed in the talk page.
Function list
L 37 — prefix_strip
L 171 — v2p
L 253 — p._escapePattern
L 275 — p.strip
L 305 — p.matchAny
L 334 — p.hyphen_to_dash
L 371 — p.hyphen2dash
L 379 — p.startswith

This module is a helper module to be used by other modules; it may not designed to be invoked directly. See Star Citizen:Lua/Helper modules for a full list and more information. For a full list of modules using this helper click here

FunctionTypeUse
trimTrims whitespace characters from the start and end of the string.
titleCapitalizes the first letter of each word in the text, apart from a number of short words: a, an, the, at, by, for, in, of, on, to, up, and, as, but, or, and nor.
sentenceFinds the first letter and capitalises it, then renders the rest of the text in lower case.
ucfirstFinds the first letter and capitalises it, then renders the rest of the text in lower case.
findlastFinds the last item in a list
splitSplits text at boundaries specified by separator and returns the chunk for the index idx (starting at 1).
stripZeroesFinds the first number in a string of text and strips leading zeros, but retains a zero which is followed by a decimal point. For example: "0940" → "940"; "Year: 0023" → "Year: 23"; "00.12" → "0.12"
nowikiEnsures that a string of text is treated by the MediaWiki software as just a string, not code. It trims leading and trailing whitespace.
val2percentScans through a string, passed as either the first unnamed parameter or
one2aScans through a string, passed as either the first unnamed parameter or
findpagetextReturns the position of a piece of text in the wikitext source of a page.
stripStrips the first positional parameter of the characters or pattern supplied in the second positional parameter.
matchAnyReturns the index of the first positional parameter to match the source parameter.
hyphen2dashConverts a hyphen to a dash under certain conditions. The hyphen must separate like items; unlike items are returned unmodified.
startswithMatch strings start with a pattern

require ('strict');
local p = {}

p.trim = function(frame)
	return mw.text.trim(frame.args[1] or "")
end

p.sentence = function (frame)
	-- {{lc:}} is strip-marker safe, string.lower is not.
	frame.args[1] = frame:callParserFunction('lc', frame.args[1])
	return p.ucfirst(frame)
end

p.ucfirst = function (frame )
	local s = frame.args[1];
	if not s or '' == s or s:match ('^%s+$') then								-- when <s> is nil, empty, or only whitespace
		return s;																-- abandon because nothing to do
	end

	s =  mw.text.trim( frame.args[1] or "" )
	local s1 = ""

	local prefix_patterns_t = {													-- sequence of prefix patterns
		'^\127[^\127]*UNIQ%-%-%a+%-%x+%-QINU[^\127]*\127',						-- stripmarker
		'^([%*;:#]+)',															-- various list markup
		'^(\'\'\'*)',															-- bold / italic markup
		'^(%b<>)',																-- html-like tags because some templates render these
		'^(&%a+;)',																-- html character entities because some templates render these
		'^(&#%d+;)',															-- html numeric (decimal) entities because some templates render these
		'^(&#x%x+;)',															-- html numeric (hexadecimal) entities because some templates render these
		'^(%s+)',																-- any whitespace characters
		'^([%(%)%-%+%?%.%%!~!@%$%^&_={}/`,‘’„“”ʻ|\"\'\\]+)',					-- miscellaneous punctuation
		}
	
	local prefixes_t = {};														-- list, bold/italic, and html-like markup, & whitespace saved here

	local function prefix_strip (s)												-- local function to strip prefixes from <s>
		for _, pattern in ipairs (prefix_patterns_t) do							-- spin through <prefix_patterns_t> 
			if s:match (pattern) then											-- when there is a match
				local prefix = s:match (pattern);								-- get a copy of the matched prefix
				table.insert (prefixes_t, prefix);								-- save it
				s = s:sub (prefix:len() + 1);									-- remove the prefix from <s>
				return s, true;													-- return <s> without prefix and flag; force restart at top of sequence because misc punct removal can break stripmarker
			end
		end
		return s;																-- no prefix found; return <s> with nil flag
	end

	local prefix_removed;														-- flag; boolean true as long as prefix_strip() finds and removes a prefix
	
	repeat																		-- one by one remove list, bold/italic, html-like markup, whitespace, etc from start of <s>
		s, prefix_removed = prefix_strip (s);
	until (not prefix_removed);													-- until <prefix_removed> is nil

	s1 = table.concat (prefixes_t);												-- recreate the prefix string for later reattachment

	local first_text = string.match (s, '^%[%[[^%]]+%]%]');					-- extract wikilink at start of string if present; TODO: this can be string.match()?

	local upcased;
	if first_text then
		if first_text:match ('^%[%[[^|]+|[^%]]+%]%]') then						-- if <first_text> is a piped link
			upcased = string.match (s, '^%[%[[^|]+|%W*(%w)');				-- get first letter character
			upcased = string.upper (upcased);								-- upcase first letter character
			s = string.gsub (s, '^(%[%[[^|]+|%W*)%w', '%1' .. upcased);		-- replace
		else																	-- here when <first_text> is a wikilink but not a piped link
			upcased = string.match (s, '^%[%[%W*%w');						-- get '[[' and first letter
			upcased = string.upper (upcased);								-- upcase first letter character
			s = string.gsub (s, '^%[%[%W*%w', upcased);						-- replace; no capture needed here
		end

	elseif s:match ('^%[%S+%s+[^%]]+%]') then									-- if <s> is a ext link of some sort; must have label text
		upcased = string.match (s, '^%[%S+%s+%W*(%w)');						-- get first letter character
		upcased = string.upper (upcased);									-- upcase first letter character
		s = string.gsub (s, '^(%[%S+%s+%W*)%w', '%1' .. upcased);			-- replace
	
	elseif s:match ('^%[%S+%s*%]') then											-- if <s> is a ext link without label text; nothing to do
		return s1 .. s;															-- reattach prefix string (if present) and done

	else																		-- <s> is not a wikilink or ext link; assume plain text
		upcased = string.match (s, '^%W*%w');								-- get the first letter character
		upcased = string.upper (upcased);									-- upcase first letter character
		s = string.gsub (s, '^%W*%w', upcased);								-- replace; no capture needed here
	end

	return s1 .. s;																-- reattach prefix string (if present) and done
end


p.title = function (frame )
	-- http://grammar.yourdictionary.com/capitalization/rules-for-capitalization-in-titles.html
	-- recommended by The U.S. Government Printing Office Style Manual:
	-- "Capitalize all words in titles of publications and documents,
	-- except a, an, the, at, by, for, in, of, on, to, up, and, as, but, or, and nor."
	local alwayslower = {['a'] = 1, ['an'] = 1, ['the'] = 1,
		['and'] = 1, ['but'] = 1, ['or'] = 1, ['for'] = 1,
		['nor'] = 1, ['on'] = 1, ['in'] = 1, ['at'] = 1, ['to'] = 1,
		['from'] = 1, ['by'] = 1, ['of'] = 1, ['up'] = 1 }
	local res = ''
	local s =  mw.text.trim( frame.args[1] or "" )
	local words = mw.text.split( s, " ")
	for i, s in ipairs(words) do
		-- {{lc:}} is strip-marker safe, string.lower is not.
		s = frame:callParserFunction('lc', s)
		if i == 1 or alwayslower[s] ~= 1 then
			s = mw.getContentLanguage():ucfirst(s)
		end
		words[i] = s
	end
	return table.concat(words, " ")
end

-- findlast finds the last item in a list
-- the first unnamed parameter is the list
-- the second, optional unnamed parameter is the list separator (default = comma space)
-- returns the whole list if separator not found
p.findlast = function(frame)
	local s =  mw.text.trim( frame.args[1] or "" )
	local sep = frame.args[2] or ""
	if sep == "" then sep = ", " end
	local pattern = ".*" .. sep .. "(.*)"
	local a, b, last = s:find(pattern)
	if a then
		return last
	else
		return s
	end
end

-- stripZeros finds the first number and strips leading zeros (apart from units)
-- e.g "0940" -> "940"; "Year: 0023" -> "Year: 23"; "00.12" -> "0.12"
p.stripZeros = function(frame)
	local s = mw.text.trim(frame.args[1] or "")
	local n = tonumber( string.match( s, "%d+" ) ) or ""
	s = string.gsub( s, "%d+", n, 1 )
	return s
end

-- nowiki ensures that a string of text is treated by the MediaWiki software as just a string
-- it takes an unnamed parameter and trims whitespace, then removes any wikicode
p.nowiki = function(frame)
	local str = mw.text.trim(frame.args[1] or "")
	return mw.text.nowiki(str)
end

-- split splits text at boundaries specified by separator
-- and returns the chunk for the index idx (starting at 1)
-- #invoke:String2 |split |text |separator |index |true/false
-- #invoke:String2 |split |txt=text |sep=separator |idx=index |plain=true/false
-- if plain is false/no/0 then separator is treated as a Lua pattern - defaults to plain=true
p.split = function(frame)
	local args = frame.args
	if not(args[1] or args.txt) then args = frame:getParent().args end
	local txt = args[1] or args.txt or ""
	if txt == "" then return nil end
	local sep = (args[2] or args.sep or ""):gsub('"', '')
	local idx = tonumber(args[3] or args.idx) or 1
	local plain = (args[4] or args.plain or "true"):sub(1,1)
	plain = (plain ~= "f" and plain ~= "n" and plain ~= "0")
	local splittbl = mw.text.split( txt, sep, plain )
	if idx < 0 then idx = #splittbl + idx + 1 end
	return splittbl[idx]
end

-- val2percent scans through a string, passed as either the first unnamed parameter or |txt=
-- it converts each number it finds into a percentage and returns the resultant string.
p.val2percent = function(frame)
	local args = frame.args
	if not(args[1] or args.txt) then args = frame:getParent().args end
	local txt = mw.text.trim(args[1] or args.txt or "")
	if txt == "" then return nil end
	local function v2p (x)
		x = (tonumber(x) or 0) * 100
		if x == math.floor(x) then x = math.floor(x) end
		return x .. "%"
	end
	txt = txt:gsub("%d[%d%.]*", v2p) -- store just the string
	return txt
end

-- one2a scans through a string, passed as either the first unnamed parameter or |txt=
-- it converts each occurrence of 'one ' into either 'a ' or 'an ' and returns the resultant string.
p.one2a = function(frame)
	local args = frame.args
	if not(args[1] or args.txt) then args = frame:getParent().args end
	local txt = mw.text.trim(args[1] or args.txt or "")
	if txt == "" then return nil end
	txt = txt:gsub(" one ", " a "):gsub("^one", "a"):gsub("One ", "A "):gsub("a ([aeiou])", "an %1"):gsub("A ([aeiou])", "An %1")
	return txt
end

-- findpagetext returns the position of a piece of text in a page
-- First positional parameter or |text is the search text
-- Optional parameter |title is the page title, defaults to current page
-- Optional parameter |plain is either true for plain search (default) or false for Lua pattern search
-- Optional parameter |nomatch is the return value when no match is found; default is nil
p._findpagetext = function(args)
	-- process parameters
	local nomatch = args.nomatch or ""
	if nomatch == "" then nomatch = nil end
	--
	local text = mw.text.trim(args[1] or args.text or "")
	if text == "" then return nil end
	--
	local title = args.title or ""
	local titleobj
	if title == "" then
		titleobj = mw.title.getCurrentTitle()
	else
		titleobj = mw.title.new(title)
	end
	--
	local plain = args.plain or ""
	if plain:sub(1, 1) == "f" then plain = false else plain = true end
	-- get the page content and look for 'text' - return position or nomatch
	local content = titleobj and titleobj:getContent()
	return content and string.find(content, text, 1, plain) or nomatch
end
p.findpagetext = function(frame)
	local args = frame.args
	local pargs = frame:getParent().args
	for k, v in pairs(pargs) do
		args[k] = v
	end
	if not (args[1] or args.text) then return nil end
	-- just the first value
	return (p._findpagetext(args))
end

-- returns the decoded url. Inverse of parser function {{urlencode:val|TYPE}}
-- Type is:
-- QUERY decodes + to space (default)
-- PATH does no extra decoding
-- WIKI decodes _ to space
p._urldecode = function(url, type)
	url = url or ""
	type = (type == "PATH" or type == "WIKI") and type
	return mw.uri.decode( url, type )
end
-- {{#invoke:String2|urldecode|url=url|type=type}}
p.urldecode = function(frame)
	return mw.uri.decode( frame.args.url, frame.args.type )
end

-- what follows was merged from Module:StringFunc

-- helper functions
p._GetParameters = require('Module:GetParameters')

-- Argument list helper function, as per Module:String
p._getParameters = p._GetParameters.getParameters

-- Escape Pattern helper function so that all characters are treated as plain text, as per Module:String
function p._escapePattern( pattern_str )
	return string.gsub( pattern_str, "([%(%)%.%%%+%-%*%?%[%^%$%]])", "%%%1" )
end

-- Helper Function to interpret boolean strings, as per Module:String
p._getBoolean = p._GetParameters.getBoolean

--[[
Strip

This function Strips characters from string

Usage:
{{#invoke:String2|strip|source_string|characters_to_strip|plain_flag}}

Parameters
	source: The string to strip
	chars:  The pattern or list of characters to strip from string, replaced with ''
	plain:  A flag indicating that the chars should be understood as plain text. defaults to true.

Leading and trailing whitespace is also automatically stripped from the string.
]]
function p.strip( frame )
	local new_args = p._getParameters( frame.args,  {'source', 'chars', 'plain'} )
	local source_str = new_args['source'] or ''
	local chars = new_args['chars'] or '' or 'characters'
	source_str = mw.text.trim(source_str)
	if source_str == '' or chars == '' then
		return source_str
	end
	local l_plain = p._getBoolean( new_args['plain'] or true )
	if l_plain then
		chars = p._escapePattern( chars )
	end
	local result
	result = string.gsub(source_str, "["..chars.."]", '')
	return result
end

--[[
Match any
Returns the index of the first given pattern to match the input. Patterns must be consecutively numbered.
Returns the empty string if nothing matches for use in {{#if:}}

Usage:
	{{#invoke:String2|matchAll|source=123 abc|456|abc}} returns '2'.

Parameters:
	source: the string to search
	plain:  A flag indicating that the patterns should be understood as plain text. defaults to true.
	1, 2, 3, ...: the patterns to search for
]]
function p.matchAny(frame)
	local source_str = frame.args['source'] or error('The source parameter is mandatory.')
	local l_plain = p._getBoolean( frame.args['plain'] or true )
	for i = 1, math.huge do
		local pattern = frame.args[i]
		if not pattern then return '' end
		if string.find(source_str, pattern, 1, l_plain) then
			return tostring(i)
		end
	end
end

--[[--------------------------< H Y P H E N _ T O _ D A S H >--------------------------------------------------

Converts a hyphen to a dash under certain conditions.  The hyphen must separate
like items; unlike items are returned unmodified.  These forms are modified:
	letter - letter (A - B)
	digit - digit (4-5)
	digit separator digit - digit separator digit (4.1-4.5 or 4-1-4-5)
	letterdigit - letterdigit (A1-A5) (an optional separator between letter and
		digit is supported – a.1-a.5 or a-1-a-5)
	digitletter - digitletter (5a - 5d) (an optional separator between letter and
		digit is supported – 5.a-5.d or 5-a-5-d)

any other forms are returned unmodified.

str may be a comma- or semicolon-separated list

]]
function p.hyphen_to_dash( str, spacing )
	if (str == nil or str == '') then
		return str
	end

	local accept

	str = mw.text.decode(str, true )											-- replace html entities with their characters; semicolon mucks up the text.split

	local out = {}
	local list = mw.text.split (str, '%s*[,;]%s*')								-- split str at comma or semicolon separators if there are any

	for _, item in ipairs (list) do												-- for each item in the list
		item = mw.text.trim(item)												-- trim whitespace
		item, accept = item:gsub ('^%(%((.+)%)%)$', '%1')
		if accept == 0 and string.match (item, '^%w*[%.%-]?%w+%s*[%-–—]%s*%w*[%.%-]?%w+$') then	-- if a hyphenated range or has endash or emdash separators
			if item:match ('^%a+[%.%-]?%d+%s*%-%s*%a+[%.%-]?%d+$') or			-- letterdigit hyphen letterdigit (optional separator between letter and digit)
				item:match ('^%d+[%.%-]?%a+%s*%-%s*%d+[%.%-]?%a+$') or			-- digitletter hyphen digitletter (optional separator between digit and letter)
				item:match ('^%d+[%.%-]%d+%s*%-%s*%d+[%.%-]%d+$') or			-- digit separator digit hyphen digit separator digit
				item:match ('^%d+%s*%-%s*%d+$') or								-- digit hyphen digit
				item:match ('^%a+%s*%-%s*%a+$') then							-- letter hyphen letter
					item = item:gsub ('(%w*[%.%-]?%w+)%s*%-%s*(%w*[%.%-]?%w+)', '%1–%2')	-- replace hyphen, remove extraneous space characters
			else
				item = string.gsub (item, '%s*[–—]%s*', '–')				-- for endash or emdash separated ranges, replace em with en, remove extraneous whitespace
			end
		end
		table.insert (out, item)												-- add the (possibly modified) item to the output table
	end

	local temp_str = table.concat (out, ',' .. spacing)							-- concatenate the output table into a comma separated string
	temp_str, accept = temp_str:gsub ('^%(%((.+)%)%)$', '%1')					-- remove accept-this-as-written markup when it wraps all of concatenated out
	if accept ~= 0 then
		temp_str = str:gsub ('^%(%((.+)%)%)$', '%1')							-- when global markup removed, return original str; do it this way to suppress boolean second return value
	end
	return temp_str
end

function p.hyphen2dash( frame )
	local str = frame.args[1] or ''
	local spacing = frame.args[2] or ' ' -- space is part of the standard separator for normal spacing (but in conjunction with templates r/rp/ran we may need a narrower spacing

	return p.hyphen_to_dash(str, spacing)
end

-- Similar to [[Module:String#endswith]]
function p.startswith(frame)
	return (frame.args[1]:sub(1, frame.args[2]:len()) == frame.args[2]) and 'yes' or ''
end

return p