The generator

This is the Hakyll program for generating (see my Hakyll tutorial). Look at this Git repository for the data files as well. The most defining trait of the site is the tree menu at the top, which contains every content page on the site. Apart from that, I also do a lot of small hacks to generate various bits of the site. There is also a simple blog system, with one file per post.

{-# LANGUAGE OverloadedStrings #-}
module Main(main) where
import Control.Arrow (first, (&&&))
import Control.Monad
import Data.Char
import Data.List hiding (group)
import Data.Ord
import Data.Maybe
import Data.Monoid
import System.FilePath
import System.Directory (removeFile)
import Text.Blaze.Internal (preEscapedString)
import Text.Blaze.Html5 ((!))
import qualified Text.Blaze.Html5 as H
import qualified Text.Blaze.Html5.Attributes as A
import Text.Blaze.Html.Renderer.String (renderHtml)
import Text.Pandoc
import Text.Pandoc.Walk
import Hakyll.Web.Pandoc (renderPandocWith, renderPandocWithTransform)
import Hakyll

Hierarchical menu

We are going to define a data type and associated helper functions for generating a menu. Conceptually, the site is a directory tree, with a page being a leaf of the tree. The menu for a given page will illustrate the path taken from the root to the page, namely which intermediary directories were entered.

A level (or “line”, if you look at its actual visual appearance) of the menu consists of two lists: the elements preceding and succeeding the focused element. The focused element itself is the first element of the aftItems list. This definition ensures that we have at most a single focused element per menu level. Each element is a pair consisting of an URL and a name.

data MenuLevel = MenuLevel { prevItems :: [(FilePath,String)]
                           , aftItems  :: [(FilePath,String)]

allItems :: MenuLevel -> [(FilePath, String)]
allItems l = prevItems l ++ aftItems l
emptyMenuLevel :: MenuLevel
emptyMenuLevel = MenuLevel [] []

First, let us define a function for inserting an element into a sorted list, returning the original list if the element is already there.

insertUniq :: Ord a => a -> [a] -> [a]
insertUniq x xs | x `elem` xs = xs
                | otherwise = insert x xs

We can use this function to insert a non-focused element into a MenuLevel. We take care to put the new element in its proper sorted position relative to the focused element, if any.

insertItem :: MenuLevel -> (FilePath, String) -> MenuLevel
insertItem l v = case aftItems l of
                   []     -> atPrev
                   (x:xs) | v < x     -> atPrev
                          | otherwise -> l { aftItems = x:insertUniq v xs }
  where atPrev = l { prevItems = insertUniq v (prevItems l) }

When inserting a focused element, we have to split the elements into those that go before and those that come after the focused element.

insertFocused :: MenuLevel -> (FilePath, String) -> MenuLevel
insertFocused l v = MenuLevel bef (v:aft)
  where (bef, aft) = partition (<v) (delete v $ allItems l)

Finally, a menu is just a list of menu levels.

newtype Menu = Menu { menuLevels :: [MenuLevel] }

emptyMenu :: Menu
emptyMenu = Menu []

I am using the BlazeHTML library for HTML generation, so the result of rendering a menu is an H.Html value. The rendering will consist of one HTML <ul> block per menu level, each with the CSS class menuN, where N is the number of the level.

showMenu :: Menu -> H.Html
showMenu = zipWithM_ showMenuLevel [0..] . menuLevels

The focus element is tagged with the CSS class thisPage.

showMenuLevel :: Int -> MenuLevel -> H.Html
showMenuLevel d m =
  H.ul (mapM_ elems) ! A.class_ (H.toValue $ "menu" ++ show d)
  where showElem (p,k) = H.a (H.toHtml k) ! A.href (H.toValue p)
        showFocusElem (p,k) = showElem (p,k) ! A.class_ "thisPage"
        elems = map showElem (prevItems m) ++
                case aftItems m of []     -> []
                                   (l:ls) -> showFocusElem l :
                                             map showElem ls

Building the menu

Recall that the directory structure of the site is a tree. To construct a menu, we are given the current node (page) and a list of all possible nodes of the tree (all pages on the site), and we then construct the minimum tree that contains all nodes on the path from the root to the current node, as well as all siblings of those nodes. In file system terms, we show the files contained in each directory traversed from the root to the current page (as well as any children of the current page, if it is a directory).

To begin, we define a function that given the current path, decomposes some other path into the part that should be visible. For example:

relevant "foo/bar/baz" "foo/bar/quux" = ["foo/","bar/","quux"]
relevant "foo/bar/baz" "foo/bar/quux/" = ["foo/","bar/","quux/"]
relevant "foo/bar/baz" "foo/bar/quux/zog" = ["foo/","bar/","quux/"]
relevant "foo/bar/baz" "quux/zog" = ["quux/"]
relevant :: FilePath -> FilePath -> [FilePath]
relevant this other = relevant' (splitPath this) (splitPath other)
  where relevant' (x:xs) (y:ys) = y : if x == y then relevant' xs ys else []
        relevant' [] (y:_) = [y]
        relevant' _ _ = []

To construct a full menu given the current path and a list of all paths, we repeatedly extend it by a single path. Recall that menu elements are pairs of names and paths - we generate those names by taking the file name and dropping the extension of the path, also dropping any trailing “index.html” from paths.

buildMenu :: FilePath -> [FilePath] -> Menu
buildMenu this = foldl (extendMenu this) emptyMenu
                 . map (first dropIndex . (id &&& dropExtension . takeFileName))

dropIndex :: FilePath -> FilePath
dropIndex p | takeBaseName p == "index" = dropFileName p
            | otherwise                 = p
extendMenu :: FilePath -> Menu -> (FilePath, String) -> Menu
extendMenu this m (path, name) =
  if path' `elem` ["./", "/", ""] then m else
    Menu $ add (menuLevels m) (relevant this' path') "/"
  where add ls [] _ = ls
        add ls (x:xs) p
          | x `elem` focused = insertFocused l (p++x,name') : add ls' xs (p++x)
          | otherwise        = insertItem l (p++x,name') : add ls' xs (p++x)
          where (l,ls') = case ls of []  -> (emptyMenuLevel, [])
                                     k:ks -> (k,ks)
                name' = if hasTrailingPathSeparator x then x else name
        focused = splitPath this'
        path' = normalise path
        this' = normalise this

For convenience, we define a Hakyll rule that adds anything currently matched to the menu. To do this, we first need two convenience functions. The first checks whether the identifier of the current compilation is defined with some other version (recall that a version is identified by a Maybe String, not just a String), and if so, returns the route of that identifier.

routeWithVersion :: Maybe String -> Compiler (Maybe FilePath)
routeWithVersion v = getRoute =<< setVersion v <$> getUnderlying

The second extracts the route for the current identifier with no version. As a matter of convenience, we return an empty path if the identifier has no associated route. This should never occur in practice.

normalRoute :: Compiler FilePath
normalRoute = fromMaybe "" <$> routeWithVersion Nothing

The "menu" version will contain an identifier for every page that should show up in the site menu, with the compiler for each identifier generating a pathname.

addToMenu :: Rules ()
addToMenu = version "menu" $ compile $ makeItem =<< normalRoute

To generate the menu for a given page, we use loadAll to obtain a list of everything with the version “menu” (the pathnames) and use it to build the menu, which is immediately rendered to HTML. If a compiler has been defined for these identifiers that creates anything but a FilePath, Hakyll will signal a run-time type error.

getMenu :: Compiler String
getMenu = do
  menu <- map itemBody <$> loadAll (fromVersion $ Just "menu")
  myRoute <- getRoute =<< getUnderlying
  return $ renderHtml $ showMenu $ case myRoute of
    Nothing -> buildMenu "" menu
    Just me -> buildMenu me menu

Extracting descriptive texts from small programs.

I have a number of small programs and scripts of my site, and I want to automatically generate a list and description for each of them. Each program starts with a descriptive comment containing Markdown markup, so the challenge becomes extracting that comment. I define functions for extracting the leading comment from shell, C and Haskell, respectively.

For shell scripts, we take all leading lines that have a comment character in the first column, excepting the hashbang (#!). This also works for many other languages.

shDocstring :: String -> String
shDocstring = unlines
              . map (drop 2)
              . takeWhile ("#" `isPrefixOf`)
              . dropWhile (all (`elem` ['#', ' ']))
              . dropWhile ("#!" `isPrefixOf`)
              . lines

For C, we extract the first multi-line comment. At this point we should probably have used a regular expression library.

cDocstring :: String -> String
cDocstring = unlines
             . map (dropWhile (==' ')
                    . dropWhile (=='*')
                    . dropWhile (==' '))
             . maybe [] lines
             . (return . reverse . cut . reverse
                <=< find ("*/" `isSuffixOf`) . inits
                <=< return . cut
                <=< find ("/*" `isPrefixOf`) . tails)
  where cut s | "/*" `isPrefixOf` s = cut $ drop 2 s
              | otherwise = dropWhile isSpace s

Haskell is processed much like shell script: We extract the leading line comments.

hsDocstring :: String -> String
hsDocstring = unlines
              . map (drop 3)
              . takeWhile ("--" `isPrefixOf`)
              . dropWhile ("#!" `isPrefixOf`)
              . lines

A hack compiler produces, from a script file, a String containing its name and docstring in HTML format, based on an HTML template. The docstring is assumed to be in Markdown format, so we pass the entire thing through a Markdown-to-HTML compiler. Unfortunately, it seems that renderPandoc inspects the pathname given Item to figure out the input format, so we force Markdown interpretation by pretending the docstring is in a file of name ".md".

hackCompiler :: Compiler (Item String)
hackCompiler = do
  ext <- getUnderlyingExtension
  src <- itemBody <$> getResourceString
  desc <- return $ case ext of
                    ".c"  -> cDocstring src
                    ".hs" -> hsDocstring src
                    _     -> shDocstring src
  dest <- normalRoute
  desc' <- renderPandoc $ Item ".md" desc
  let name = takeFileName dest
      ctx = constField "name" name <>
            constField "script" dest <>
            constField "description" (itemBody desc')
  loadAndApplyTemplate "templates/hack.html" ctx =<< getResourceString

Adding the hacks to a page is now just loading everything with version "hacks", and wrapping it in an HTML list.

addHacks :: Item String -> Compiler (Item String)
addHacks item = do
  hacks <- loadAll (fromVersion $ Just "hacks")
  return item { itemBody = itemBody item <> renderHtml (mapM_ hackHtml hacks) }
  where hackHtml = preEscapedString .  itemBody

Extracting update times

We extract a timestamp by invoking git log on the given file, passing a bunch of options that ultimately results in the ISO 8601 date (YY-MM-DD) for the most recent commit for the file.

getUpdateTime :: FilePath -> Compiler String
getUpdateTime f =
  unixFilter "git" ["log", "-1", "--format=%ad", "--date=format:%F", "--", f] ""

Putting it all together

First, we define three convenience compilers. The first selects between different options based on the underlying identifier.

byPattern :: a -> [(Pattern, a)] -> Compiler a
byPattern def options = do
  ident <- getUnderlying
  return $ fromMaybe def (snd <$> find ((`matches` ident) . fst) options)

This can be used to create a compiler that performs initial creation.

createByPattern :: Compiler a -> [(Pattern, Compiler a)] -> Compiler a
createByPattern def options = join $ byPattern def options

It can also be used to create a compiler that modifies the result of another compiler.

modifyByPattern :: (b -> Compiler a) -> [(Pattern, b -> Compiler a)] -> b -> Compiler a
modifyByPattern def options x = join $ byPattern def options <*> pure x

When we instantiate the final page template, we will need to provide a suitable context. Apart from the default fields, my website makes use of two others: a "menu" field containing the menu, and a "source" field that contains a link to the raw (“source”) file for the current page.

contentContext :: Compiler (Context String)
contentContext = do
  menu <- getMenu
  source <- getResourceFilePath
  updated <- getUpdateTime source
  return $
    defaultContext <>
    constField "menu" menu <>
    constField "source" source <>
    constField "updated" updated

Furthermore, blog entries need a "date" field, which is extracted from the file name of the post.

postContext :: Compiler (Context String)
postContext = do
  ctx <- contentContext
  return $ dateField "date" "%B %e, %Y" `mappend` ctx

postCtx :: Context String
postCtx = mconcat
  [ modificationTimeField "mtime" "%U"
  , dateField "date" "%B %e, %Y"
  , defaultContext

I extend the default Hakyll configuration with information on how to use rsync to copy the site contents to the server.

config :: Configuration
config = defaultConfiguration
  { deployCommand = "rsync --chmod=Do+rx,Fo+r --checksum -ave 'ssh -p 22' \
                     \_site/* --exclude pub"

Now we’re ready to describe the entire site.

main :: IO ()
main = hakyllWith config $ do

CSS files are compressed, data files, my public key, and the robots.txt are copied verbatim.

  match "css/*" $ do
    route   idRoute
    compile compressCssCompiler
  match "files/**" static
  match "pubkey.asc" static
  match "" static
  match "robots.txt" static

One of our primary objectives is the ability to write content for the site without having to modify this generator program. Therefore, we define content as a non-hidden file contained in any of the directories me, writings, hacks, programs or projects, as well as any .md file in the root directory. This property is checked by the content pattern.

We divide the content into two sets: content pages, which is all content of types .md, .lhs, .fut, and .man, and content data, which is the rest.

  let inContentDir = "me/**" .||. "writings/**" .||. "hacks/**" .||.
                     "programs/**" .||. "projects/**" .||. "*.md"
      nothidden = complement "**/.**" .&&. complement ".*/**" .&&. complement "**flycheck**"
      content = inContentDir .&&. nothidden
      contentPages = content .&&. fromRegex "\\.(md|lhs|fut|man)$"
      contentData = content .&&. complement contentPages

Content data is copied verbatim, as it is expected to be images and similar non-processable data.

  match contentData static

Content pages will end up as HTML files. This is conceptually a simple process: they are added to list of pages contained in the menu, processed by a content compiler, which is manCompiler for manpages, futCompiler for Futhark programs, and contentCompiler (which we will see later) for all other files, although everything gets instantiated with the same template in the end. Some pages also need special generated content: the hacks-page needs a list of hacks. These are special-cased in an intermediate step.

  match contentPages $ do
    route $ setExtension "html"
    compile $ do
      context <- contentContext
      createByPattern contentCompiler [("**.fut", futCompiler),
                                       ("**.man", manCompiler)]
        >>= modifyByPattern return [("hacks/", addHacks)]
        >>= loadAndApplyTemplate "templates/default.html" context
        >>= relativizeUrls

The group "source" contains the source files of all literate programs. This is practical, as the processed literate program will be an HTML file, and thus probably no longer compilable by the literate system. While the user could in some cases copy the text from the browser into a source file (this is possible for literate Haskell), it is more convenient to have the original source file available. Source files are, of course, copied verbatim.

  match contentPages $ version "source" static

The group "hacks" contains small program descriptions generated by hackCompiler. It is included by addHacks.

  match "hacks/scripts/*" $ version "hacks" $ compile hackCompiler

For the blog, there are two main tasks: first, we must create a page for every post; second, we must create an overview page. A blog post is simply a Markdown file in the "blog/" subdirectory. Note that this is not matched by contentPages.

  let blogArticles = "blog/*-*-*-*.md" .&&. hasNoVersion

Each post post gives rise to a corresponding HTML file, which uses the post template. The template expects a date field, whose value we extract from the file name. Importantly, individual blog articles are not added to the menu - there would quickly be far too many.

  match blogArticles $ version "source" static

  match blogArticles $ do
    route $ setExtension "html"
    compile $ do
      postCtx <- postContext
        >>= loadAndApplyTemplate "templates/post.html"    postCtx
        >>= saveSnapshot "content"
        >>= loadAndApplyTemplate "templates/default.html" postCtx
        >>= relativizeUrls

The posts list is created by a template. The blog posts are sorted by date, with the date encoded into the filename of the blog entry. The created page is "blog/index.html" rather than "blog.html" in order for blog entries to be considered children of the blog entry in the menu.

  create ["blog/index.html"] $ do
    route idRoute
    compile $ do
      ctx <- contentContext
      posts <- recentFirst =<< loadAll blogArticles
      let ctx' = constField "title" "Blog" <>
                 listField "posts" postCtx (return posts) <>
      makeItem ""
          >>= loadAndApplyTemplate "templates/posts.html" ctx'
          >>= loadAndApplyTemplate "templates/default.html" ctx'
          >>= relativizeUrls

As the final touch on the blog, we produce an Atom feed. I have no particular reason for choosing Atom over the alternatives, except that it seems slightly more modern. Hakyll seems to support most formats, so I can add more if I feel like it.

  create ["atom.xml"] $ do
    route idRoute
    compile $ do
      let feedCtx = postCtx `mappend` bodyField "description"
      posts <- fmap (take 10) . recentFirst =<<
               loadAllSnapshots blogArticles "content"
      let feedConfiguration = FeedConfiguration
            { feedTitle       = "Troels Henriksen's blog"
            , feedDescription = "What some hacker has written"
            , feedAuthorName  = "Troels Henriksen"
            , feedAuthorEmail = ""
            , feedRoot        = ""
      renderAtom feedConfiguration feedCtx posts

Finally, HTML templates are, of course, handled by the default template compiler.

  match "templates/*" $ compile templateCompiler

We’re done with the main function. All we need to do now is some fleshing out. To start with, static files are merely copied into position.

static :: Rules ()
static = route idRoute >> compile copyFileCompiler >> return ()

Compiling content pages is done with the default pandocCompiler, but we extend it slightly with a transformation that makes every headline link to itself. First, we define the function that transforms a Header block into a Header block with a self-link.

selfLinkHeader :: Block -> Block
selfLinkHeader (Header n (ident, classes, kvs) b) =
  Header n (ident, classes, kvs) [b']
  where b' = Link (ident <> "-link", ["titlelink"], []) b ("#" <> ident, ident)
selfLinkHeader x = x

Then we can define our contentCompiler as a pandocCompiler with an additional transformation step.

contentCompiler :: Compiler (Item String)
contentCompiler = pandocCompilerWithTransform
                  defaultHakyllWriterOptions $
                  walk selfLinkHeader

Compiling man pages is done using the system groff program. The output from groff will contain control characters (notably backspaces), which we process with col -b to generate plain text. Finally, we insert the text into an HTML pre element to preserve the whitespace formatting.

manCompiler :: Compiler (Item String)
manCompiler = getResourceString
              >>= withItemBody (unixFilter "groff" (words "-m mandoc -T utf8")
                                >=> unixFilter "col" ["-b"]
                                >=> return . renderHtml . H.pre . H.toHtml)

Futhark programs (with the .fut extension) are taken to be literate Futhark programs, and processed with futhark literate. We need to play tricks with the itemIdentifier because renderPandocWith expects an .md extension.

futCompiler :: Compiler (Item String)
futCompiler = do
  source <- getResourceFilePath
  void $ unixFilter "futhark" ["literate", source] mempty
  let mdfile = source `replaceExtension` "md"
  item <- makeItem =<< unsafeCompiler (readFile mdfile)
  let oldident = itemIdentifier item
  unsafeCompiler $ removeFile mdfile
  item' <- renderPandocWithTransform defaultHakyllReaderOptions defaultHakyllWriterOptions
           (walk selfLinkHeader)
           item { itemIdentifier = fromFilePath mdfile }
  pure item' { itemIdentifier = oldident }

And that’s it.