Module:Wikitext Parsing: Difference between revisions
Jump to navigation
Jump to search
Template:Track listing>Pppery m (Changed protection settings for "Module:Wikitext Parsing": High-risk template or module: This is, among other things, an indirect dependency of Module:Pagetype and hence Template:Short description, and hence used on 26% of all pages ([Edit=Require administrator access] (indefinite) [Move=Require administrator access] (indefinite))) |
m (1 revision imported) |
(No difference)
|
Latest revision as of 02:20, 27 March 2024
Documentation for this module may be created at Module:Wikitext Parsing/doc
require("strict") --Helper functions local function startswith(text, subtext) return string.sub(text, 1, #subtext) == subtext end local function endswith(text, subtext) return string.sub(text, -#subtext, -1) == subtext end local function allcases(s) return s:gsub("%a", function(c) return "["..c:upper()..c:lower().."]" end) end local trimcache = {} local whitespace = {[" "]=1, ["\n"]=1, ["\t"]=1, ["\r"]=1} local function cheaptrim(str) --mw.text.trim is surprisingly expensive, so here's an alternative approach local quick = trimcache[str] if quick then return quick else -- local out = string.gsub(str, "^%s*(.-)%s*$", "%1") local lowEnd for i = 1,#str do if not whitespace[string.sub(str, i, i)] then lowEnd = i break end end if not lowEnd then trimcache[str] = "" return "" end for i = #str,1,-1 do if not whitespace[string.sub(str, i, i)] then local out = string.sub(str, lowEnd, i) trimcache[str] = out return out end end end end --[=[ Implementation notes ---- NORMAL HTML TAGS ---- Tags are very strict on how they want to start, but loose on how they end. The start must strictly follow <[tAgNaMe](%s|>) with no room for whitespace in the tag's name, but may then flow as they want afterwards, making <div\nclass\n=\n"\nerror\n"\n> valid There's no sense of escaping < or > E.g. <div class="error\>"> will end at \> despite it being inside a quote <div class="<span class="error">error</span>"> will not process the larger div If a tag has no end, it will consume all text instead of not processing ---- NOPROCESSING TAGS (nowiki, pre, syntaxhighlight, source, etc.) ---- (In most comments, <source> will not be mentioned. This is because it is the deprecated version of <syntaxhighlight>) No-Processing tags have some interesting differences to the above rules. For example, their syntax is a lot stricter. While an opening tag appears to follow the same set of rules, A closing tag can't have any sort of extra formatting period. While </div a/a> is valid, </nowiki a/a> isn't - only newlines and spaces/tabs are allowed in closing tags. Note that, even though <pre> tags cause a visual change when the ending tag has extra formatting, it won't cause the no-processing effects. For some reason, the format must be strict for that to apply. Both the content inside the tag pair and the content inside each side of the pair is not processed. E.g. <nowiki |}}>|}}</nowiki> would have both of the |}} escaped in practice. When something in the code is referenced to as a "Nowiki Tag", it means a tag which causes wiki text to not be processed, which includes <nowiki>, <pre>, and <syntaxhighlight> Since we only care about these tags, we can ignore the idea of an intercepting tag preventing processing, and just go straight for the first ending we can find If there is no ending to find, the tag will NOT consume the rest of the text in terms of processing behaviour (though <pre> will appear to have an effect). Even if there is no end of the tag, the content inside the opening half will still be unprocessed, meaning {{X20|<nowiki }}>}} wouldn't end at the first }} despite there being no ending to the tag. Note that there are some tags, like <math>, which also function like <nowiki> which are included in this aswell. Some other tags, like <ref>, have far too unpredictable behaviour to be handled currently (they'd have to be split and processed as something seperate - its complicated, but maybe not impossible.) I suspect that every tag listed in [[Special:Version]] may behave somewhat like this, but that's far too many cases worth checking for rarely used tags that may not even have a good reason to contain {{ or }} anyways, so we leave them alone. ---- HTML COMMENTS AND INCLUDEONLY ---- HTML Comments are about as basic as it could get for this Start at <!--, end at -->, no extra conditions. Simple enough If a comment has no end, it will eat all text instead of not being processed includeonly tags function mostly like a regular nowiki tag, with the exception that the tag will actually consume all future text if not given an ending as opposed to simply giving up and not changing anything. Due to complications and the fact that this is far less likely to be present on a page, aswell as being something that may not want to be escaped, includeonly tags are ignored during our processing --]=] local validtags = {nowiki=1, pre=1, syntaxhighlight=1, source=1, math=1} --This function expects the string to start with the tag local function TestForNowikiTag(text, scanPosition) local tagName = (string.match(text, "^<([^\n />]+)", scanPosition) or ""):lower() if not validtags[tagName] then return nil end local nextOpener = string.find(text, "<", scanPosition+1) or -1 local nextCloser = string.find(text, ">", scanPosition+1) or -1 if nextCloser > -1 and (nextOpener == -1 or nextCloser < nextOpener) then local startingTag = string.sub(text, scanPosition, nextCloser) --We have our starting tag (E.g. '<pre style="color:red">') --Now find our ending... if endswith(startingTag, "/>") then --self-closing tag (we are our own ending) return { Tag = tagName, Start = startingTag, Content = "", End = "", Length = #startingTag } else local endingTagStart, endingTagEnd = string.find(text, "</"..allcases(tagName).."[ \t\n]*>", scanPosition) if endingTagStart then --Regular tag formation local endingTag = string.sub(text, endingTagStart, endingTagEnd) local tagContent = string.sub(text, nextCloser+1, endingTagStart-1) return { Tag = tagName, Start = startingTag, Content = tagContent, End = endingTag, Length = #startingTag + #tagContent + #endingTag } else --Content inside still needs escaping (also linter error!) return { Tag = tagName, Start = startingTag, Content = "", End = "", Length = #startingTag } end end end return nil end local function TestForComment(text, scanPosition) --Like TestForNowikiTag but for <!-- --> if string.match(text, "^<!%-%-", scanPosition) then local commentEnd = string.find(text, "-->", scanPosition+4, true) if commentEnd then return { Start = "<!--", End = "-->", Content = string.sub(text, scanPosition+4, commentEnd-1), Length = commentEnd-scanPosition+3 } else --Consumes all text if not given an ending return { Start = "<!--", End = "", Content = string.sub(text, scanPosition+4), Length = #text-scanPosition+1 } end end return nil end --[[ Implementation notes The goal of this function is to escape all text that wouldn't be parsed if it was preprocessed (see above implementation notes). Using keepComments will keep all HTML comments instead of removing them. They will still be escaped regardless to avoid processing errors --]] local function PrepareText(text, keepComments) local newtext = {} local scanPosition = 1 while true do local NextCheck = string.find(text, "<[NnSsPpMm!]", scanPosition) --Advance to the next potential tag we care about if not NextCheck then --Done newtext[#newtext+1] = string.sub(text,scanPosition) break end newtext[#newtext+1] = string.sub(text,scanPosition,NextCheck-1) scanPosition = NextCheck local Comment = TestForComment(text, scanPosition) if Comment then if keepComments then newtext[#newtext+1] = Comment.Start .. mw.text.nowiki(Comment.Content) .. Comment.End end scanPosition = scanPosition + Comment.Length else local Tag = TestForNowikiTag(text, scanPosition) if Tag then local newTagStart = "<" .. mw.text.nowiki(string.sub(Tag.Start,2,-2)) .. ">" local newTagEnd = Tag.End == "" and "" or --Respect no tag ending "</" .. mw.text.nowiki(string.sub(Tag.End,3,-2)) .. ">" local newContent = mw.text.nowiki(Tag.Content) newtext[#newtext+1] = newTagStart .. newContent .. newTagEnd scanPosition = scanPosition + Tag.Length else --Nothing special, move on... newtext[#newtext+1] = string.sub(text, scanPosition, scanPosition) scanPosition = scanPosition + 1 end end end return table.concat(newtext, "") end --[=[ Implementation notes This function is an alternative to Transcluder's getParameters which considers the potential for a singular { or } or other odd syntax that %b doesn't like to be in a parameter's value. When handling the difference between {{ and {{{, mediawiki will attempt to match as many sequences of {{{ as possible before matching a {{ E.g. {{{{A}}}} -> { {{{A}}} } {{{{{{{{Text|A}}}}}}}} -> {{ {{{ {{{Text|A}}} }}} }} If there aren't enough triple braces on both sides, the parser will compromise for a template interpretation. E.g. {{{{A}} }} -> {{ {{ A }} }} While there are technically concerns about things such as wikilinks breaking template processing (E.g. {{[[}}]]}} doesn't stop at the first }}), it shouldn't be our job to process inputs perfectly when the input has garbage ({ / } isn't legal in titles anyways, so if something's unmatched in a wikilink, it's guaranteed GIGO) Setting dontEscape will prevent running the input text through EET. Avoid setting this to true if you don't have to set it. Returned values: A table of all templates. Template data goes as follows: Text: The raw text of the template Name: The name of the template Args: A list of arguments Children: A list of immediate template children --]=] --Helper functions local function boundlen(pair) return pair.End-pair.Start+1 end --Main function local function ParseTemplates(InputText, dontEscape) --Setup if not dontEscape then InputText = PrepareText(InputText) end local function finalise(text) if not dontEscape then return mw.text.decode(text) else return text end end local function CreateContainerObj(Container) Container.Text = {} Container.Args = {} Container.ArgOrder = {} Container.Children = {} -- Container.Name = nil -- Container.Value = nil -- Container.Key = nil Container.BeyondStart = false Container.LastIndex = 1 Container.finalise = finalise function Container:HandleArgInput(character, internalcall) if not internalcall then self.Text[#self.Text+1] = character end if character == "=" then if self.Key then self.Value[#self.Value+1] = character else self.Key = cheaptrim(self.Value and table.concat(self.Value, "") or "") self.Value = {} end else --"|" or "}" if not self.Name then self.Name = cheaptrim(self.Value and table.concat(self.Value, "") or "") self.Value = nil else self.Value = self.finalise(self.Value and table.concat(self.Value, "") or "") if self.Key then self.Key = self.finalise(self.Key) self.Args[self.Key] = cheaptrim(self.Value) self.ArgOrder[#self.ArgOrder+1] = self.Key else local Key = tostring(self.LastIndex) self.Args[Key] = self.Value self.ArgOrder[#self.ArgOrder+1] = Key self.LastIndex = self.LastIndex + 1 end self.Key = nil self.Value = nil end end end function Container:AppendText(text, ftext) self.Text[#self.Text+1] = (ftext or text) if not self.Value then self.Value = {} end self.BeyondStart = self.BeyondStart or (#table.concat(self.Text, "") > 2) if self.BeyondStart then self.Value[#self.Value+1] = text end end function Container:Clean(IsTemplate) self.Text = table.concat(self.Text, "") if self.Value and IsTemplate then self.Value = {string.sub(table.concat(self.Value, ""), 1, -3)} --Trim ending }} self:HandleArgInput("|", true) --Simulate ending end self.Value = nil self.Key = nil self.BeyondStart = nil self.LastIndex = nil self.finalise = nil self.HandleArgInput = nil self.AppendText = nil self.Clean = nil end return Container end --Step 1: Find and escape the content of all wikilinks on the page, which are stronger than templates (see implementation notes) local scannerPosition = 1 local wikilinks = {} local openWikilinks = {} while true do local Position, _, Character = string.find(InputText, "([%[%]])%1", scannerPosition) if not Position then --Done break end scannerPosition = Position+2 --+2 to pass the [[ / ]] if Character == "[" then --Add a [[ to the pending wikilink queue openWikilinks[#openWikilinks+1] = Position else --Pair up the ]] to any available [[ if #openWikilinks >= 1 then local start = table.remove(openWikilinks) --Pop the latest [[ wikilinks[start] = {Start=start, End=Position+1, Type="Wikilink"} --Note the pair end end end --Step 2: Find the bounds of every valid template and variable ({{ and {{{) local scannerPosition = 1 local templates = {} local variables = {} local openBrackets = {} while true do local Start, _, Character = string.find(InputText, "([{}])%1", scannerPosition) if not Start then --Done (both 9e9) break end local _, End = string.find(InputText, "^"..Character.."+", Start) scannerPosition = Start --Get to the {{ / }} set if Character == "{" then --Add the {{+ set to the queue openBrackets[#openBrackets+1] = {Start=Start, End=End} else --Pair up the }} to any available {{, accounting for {{{ / }}} local BracketCount = End-Start+1 while BracketCount >= 2 and #openBrackets >= 1 do local OpenSet = table.remove(openBrackets) if boundlen(OpenSet) >= 3 and BracketCount >= 3 then --We have a {{{variable}}} (both sides have 3 spare) variables[OpenSet.End-2] = {Start=OpenSet.End-2, End=scannerPosition+2, Type="Variable"} --Done like this to ensure chronological order BracketCount = BracketCount - 3 OpenSet.End = OpenSet.End - 3 scannerPosition = scannerPosition + 3 else --We have a {{template}} (both sides have 2 spare, but at least one side doesn't have 3 spare) templates[OpenSet.End-1] = {Start=OpenSet.End-1, End=scannerPosition+1, Type="Template"} --Done like this to ensure chronological order BracketCount = BracketCount - 2 OpenSet.End = OpenSet.End - 2 scannerPosition = scannerPosition + 2 end if boundlen(OpenSet) >= 2 then --Still has enough data left, leave it in openBrackets[#openBrackets+1] = OpenSet end end end scannerPosition = End --Now move past the bracket set end --Step 3: Re-trace every object using their known bounds, collecting our parameters with (slight) ease local scannerPosition = 1 local activeObjects = {} local finalObjects = {} while true do local LatestObject = activeObjects[#activeObjects] --Commonly needed object local NNC, _, Character --NNC = NextNotableCharacter if LatestObject then NNC, _, Character = string.find(InputText, "([{}%[%]|=])", scannerPosition) else NNC, _, Character = string.find(InputText, "([{}])", scannerPosition) --We are only after templates right now end if not NNC then break end if NNC > scannerPosition and LatestObject then local scannedContent = string.sub(InputText, scannerPosition, NNC-1) LatestObject:AppendText(scannedContent, finalise(scannedContent)) end scannerPosition = NNC+1 if Character == "{" or Character == "[" then local Container = templates[NNC] or variables[NNC] or wikilinks[NNC] if Container then CreateContainerObj(Container) if Container.Type == "Template" then Container:AppendText("{{") scannerPosition = NNC+2 elseif Container.Type == "Variable" then Container:AppendText("{{{") scannerPosition = NNC+3 else --Wikilink Container:AppendText("[[") scannerPosition = NNC+2 end if LatestObject and Container.Type == "Template" then --Only templates count as children LatestObject.Children[#LatestObject.Children+1] = Container end activeObjects[#activeObjects+1] = Container elseif LatestObject then LatestObject:AppendText(Character) end elseif Character == "}" or Character == "]" then if LatestObject then LatestObject:AppendText(Character) if LatestObject.End == NNC then if LatestObject.Type == "Template" then LatestObject:Clean(true) finalObjects[#finalObjects+1] = LatestObject else LatestObject:Clean(false) end activeObjects[#activeObjects] = nil local NewLatest = activeObjects[#activeObjects] if NewLatest then NewLatest:AppendText(LatestObject.Text) --Append to new latest end end end else --| or = if LatestObject then LatestObject:HandleArgInput(Character) end end end --Step 4: Fix the order local FixedOrder = {} local SortableReference = {} for _,Object in next,finalObjects do SortableReference[#SortableReference+1] = Object.Start end table.sort(SortableReference) for i = 1,#SortableReference do local start = SortableReference[i] for n,Object in next,finalObjects do if Object.Start == start then finalObjects[n] = nil Object.Start = nil --Final cleanup Object.End = nil Object.Type = nil FixedOrder[#FixedOrder+1] = Object break end end end --Finished, return return FixedOrder end local p = {} --Main entry points p.PrepareText = PrepareText p.ParseTemplates = ParseTemplates --Extra entry points, not really required p.TestForNowikiTag = TestForNowikiTag p.TestForComment = TestForComment return p --[==[ console tests local s = [=[Hey!{{Text|<nowiki | ||> Hey! }} A</nowiki>|<!--AAAAA|AAA-->Should see|Shouldn't see}}]=] local out = p.PrepareText(s) mw.logObject(out) local s = [=[B<!-- Hey! -->A]=] local out = p.TestForComment(s, 2) mw.logObject(out); mw.log(string.sub(s, 2, out.Length)) local a = p.ParseTemplates([=[ {{User:Aidan9382/templates/dummy |A|B|C {{{A|B}}} { } } { |<nowiki>D</nowiki> |<pre>E |F</pre> |G|=|a=|A = [[{{PAGENAME}}|A=B]]{{Text|1==<nowiki>}}</nowiki>}}|A B=Success}} ]=]) mw.logObject(a) ]==]