Tentatively for the 2.23 release:

  * updated Travis CI to work again
  * tweaked bounds in the `.cabal` files
  * adjusted `extra-source-files` to properly identify test files

Tentatively adjust bounds and changelogs for the release to be bundled
with GHC 8.8.1.

When possible, associated types with promoted lists should use the
promoted list literal syntax (instead of repeated applications of
': and '[]). This was fixed in 2122de54.

Closes #466,

At this point, Haddock depended on Cabal-the-library solely for a
verbosity parser (which misleadingly accepts all sorts of verbosity
options that Haddock never uses). Now, the only dependency on Cabal
is for `haddock-test` (which uses Cabal to locate the Haddock interface
files of a couple boot libraries).

The only other change in html/hoogle/hyperlinker output for the boot
libraries that this caused is a fix to some Hoogle output for implicit
params.

```
$ diff -r _build/docs/ old_docs
diff -r _build/docs/html/libraries/base/base.txt old_docs/html/libraries/base/base.txt
13296c13296
< assertError :: (?callStack :: CallStack) => Bool -> a -> a
---
> assertError :: ?callStack :: CallStack => Bool -> a -> a
```

`fail` is no longer part of `Monad`.

Now that GHC is hosted on Gitlab, the arcanist files don't make sense
anymore. The STYLE file contains nothing more than a dead link too.

The `.ghci` files are actively annoying when trying to `cabal v2-repl`.
As for the `scripts`, the distribution workflow is completely different.

Alan Zimmerman authored 6 years ago and Alec Theriault committed 6 years ago

(cherry picked from commit 3ee6526d)

Before LLVM 6.0.1 (or 10.0 on Apple LLVM), there was a bug where
lines that started with an octothorpe but turned out not
to lex like pragmas would have an extra line added after them.

Since this bug has been fixed upstream and that it doesn't have dire
consequences anyways, the workaround is not really worth it
anymore - we can just tell people to update their clang version (or re-structure
their pragma code).

After this commit, we can run with `--latex` on all boot libraries
without crashing (although the generated LaTeX still fails to compile in
a handful of larger packages like `ghc` and `base`).

 * Add newlines after all block elements in LaTeX. This is important to
   prevent the final output from being more an more indented. See the
   `latext-test/src/Example` test case for a sample of this.

 * Support associated types in class declarations (but not yet defaults)

 * Several small issues for producing compiling LaTeX;
     - avoid empy `\haddockbeginargs` lists (ex: `type family Any`)
     - properly escape identifiers depending on context (ex: `Int#`)
     - add `vbox` around `itemize`/`enumerate` (so they can be in tables)

 * Several spacing fixes:
     - limit the width of `Pretty`-arranged monospaced code
     - cut out extra space characters in export lists
     - only escape spaces if there are _multiple_ spaces
     - allow type signatures to be multiline (even without docs)

 * Remove uninteresting and repetitive `main.tex`/`haddock.sty` files
   from `latex-test` test reference output.

Fixes #935, #929 (LaTeX docs for `text` build & compile)
Fixes #727, #930 (I think both are really about type families...)

`markupWarning` often processes inputs which span across paragraphs.
Unfortunately, LaTeX's `emph` is not made to handle this (and will
crash).

Fixes #936.

  * default methods now get rendered differently
  * default associated types get rendered
  * fix a forgotten `s/TypeSig/ClassOpSig/` refactor in LaTeX backend
  * LaTeX backend now renders default method signatures

NB: there is still no way to document default class members and the
NB: LaTeX backend still crashes on associated types

Fixes #1030.

gershomb authored 6 years ago and Alec Theriault committed 6 years ago

This is to address the concern that, on less nice and older screens,
some of the shades of grey blend in too easily with the white
background.

 * darken the font slightly
 * darken slightly the grey behind type signatures and such
 * add a border and round the corners on code blocks
 * knock the font down by one point

Fixes #864.

Xia Li-yao authored 6 years ago and Alec Theriault committed 6 years ago

Adds a menu item (like "Quick Jump") for options related to displaying
instances. This provides functionality for:

  * expanding/collapsing all instances on the currently opened page
  * controlling whether instances are expanded/collapsed by default
  * controlling whether the state of instances should be "remembered"

This new functionality is implemented in Typescript in `details-helper`.
The built-in-themes style switcher also got a revamp so that all three
of QuickJump, the style switcher, and instance preferences now have
the same style and implementation structure.

See also: https://mail.haskell.org/pipermail/haskell-cafe/2019-January/130495.html



Fixes #698.

Co-authored-by: Lysxia <lysxia@gmail.com>
Co-authored-by: Nathan Collins <conathan@galois.com>

This avoids a situation in which an identifier would get defaulted to
a completely different identifier. Prior to this commit, the 'Bug1035'
test case would hyperlink 'Foo' into 'Bar'!

Fixes #1035.

Docs on standalone deriving decls for classes with associated types
should be associated with the class instance, not the associated type
instance.

Fixes #1033

  * '(<|>)' and '`elem`' now get parsed and rendered properly as links
  * 'DbModule'/'DbUnitId' now properly get split apart into two links
  * tuple names now get parsed properly
  * some more small niceties...

The identifier parsing code is more precise and more efficient (although to be
fair: it is also longer and in its own module). On the rendering side, we need
to pipe through information about backticks/parens/neither all the way through
from renaming to the backends.

In terms of impact: a total of 35 modules in the entirety of the bootlib + ghc
lib docs change. The only "regression" is things like '\0'. These should be
changed to @\\0@ (the path by which this previously worked seems accidental).

Identifier links can be prefixed with a 'v' or 't' to indicate the value or
type namespace of the desired identifier. For example:

-- | Some link to a value: v'Data.Functor.Identity'
--
-- Some link to a type: t'Data.Functor.Identity'

The default is still the type (with a warning about the ambiguity)

Ryan Scott authored 6 years ago and Alec Theriault committed 6 years ago

The central trick in this patch is to use `dataConUserTyVars` instead of
`univ_tvs ++ ex_tvs`, which displays the foralls in a GADT constructor in
a way that's more faithful to how the user originally wrote it.

Fixes #1015.

Oleg Grenrus authored 6 years ago and Alec Theriault committed 6 years ago

The second example is interesting.
If there's a list directly after the header, and that list has
deeper structure, the parser is confused: It finds two lists:

 - One with the first nested element,
 - everything after it

I'm not trying to fix this, as I'm not even sure this is a bug,
and not a feature.

Now that Haddock is moving towards working entirely over `.hi` and `.hie` files,
all declarations and types are going to be synthesized via the `Convert` module.
In preparation for this change, here are a bunch of fixes to this module:

  * Add kind annotations to type variables in `forall`'s whose kind is not `Type`,
    unless the kind can be inferred from some later use of the variable. See
    `implicitForAll` and `noKindTyVars` in particular if you wish to dive into this.

  * Properly detect `HsQualTy` in `synifyType`. This is done by following suit with
    what GHC's `toIfaceTypeX` does and checking the first argument of
    `FunTy{} :: Type` to see if it classified as a given/wanted in the typechecker
    (see `isPredTy`). 

  * Beef up the logic around figuring out when an explicit `forall` is needed. This
    includes: observing if any of the type variables will need kind signatures, if the
    inferred type variable order _without_ a forall will still match the one GHC
    claims, and some other small things.

  * Add some (not yet used) functionality for default levity polymorphic type
    signatures. This functionality similar to `fprint-explicit-runtime-reps`.

Couple other smaller fixes only worth mentioning:

  * Show the family result signature only when it isn't `Type`
  * Fix rendering of implicit parameters in the LaTeX and Hoogle backends
  * Better handling of the return kind of polykinded H98 data declarations
  * Class decls produced by `tyThingToLHsDecl` now contain associated type
    defaults and default method signatures when appropriate
  * Filter out more `forall`'s in pattern synonyms

The circumstances under which this module appeared are completely gone.
The Hyperlinker backend no longer needs this module (it uses the more
efficient `Encoding` module from `ghc`).

Why no deprecation? Because this module really shouldn't exist!

  - It isn't used in `haddock-library`/`haddock-api` anymore
  - It was copy pasted directly from `utf8-string`
  - Folks seeking a boot-lib only solution can use `ghc`'s `Encoding`

Nathan Collins authored 6 years ago and Alec Theriault committed 6 years ago

No need to `stack install` Haddock to test it. Indeed, `stack install` changes the `haddock` on user's `PATH` if `~/.local/bin` is on user's `PATH` which may not be desirable when hacking on Haddock.

* Amend `ParserSpec` to match new Hyperlinker API
    - pass in compiler info
    - strip out null tokens

* Make `hypsrc-test` pass reliably
    - strip out `local-*` ids
    - strip out `line-*` ids from the `ClangCppBug` test
    - re-accept output

Should help make CI be less broken

# Summary

This is a large architectural change to the Hyperlinker.

  * extract link (and now also type) information from `.hie` instead
    of doing ad-hoc SYB traversals of the `RenamedSource`. Also
    adds a superb type-on-hover feature (#715).

 * re-engineer the lexer to avoid needless string conversions. By going
    directly through GHC's `P` monad and taking bytestring slices, we
    avoid a ton of allocation and have better handling of position
    pragmas and CPP.

In terms of performance, the Haddock side of things has gotten _much_
more efficient. Unfortunately, much of this is cancelled out by the
increased GHC workload for generating `.hie` files. For the full set of
boot libs (including `ghc`-the-library)

  * the sum of total time went down by 9-10% overall
  * the sum of total allocations went down by 6-7%

# Motivation

Haddock is moving towards working entirely over `.hi` and `.hie` files.
This change means we no longer need the `RenamedSource` from
`TypecheckedModule` (something which is _not_ in `.hi` files).

# Details

Along the way a bunch of things were fixed:

 * Cross package (and other) links are now more reliable (#496)
 * The lexer tries to recover from errors on every line (instead of at CPP
    boundaries)
 * `LINE`/`COLUMN` pragmas are taken into account
 * filter out zero length tokens before rendering
 * avoid recomputing the `ModuleName`-based `SrcMap`
 * remove the last use of `Documentation.Haddock.Utf8` (see  #998)
 * restructure temporary folder logic for `.hi`/`.hie` model

It looks like the new versions don't cause any breakage
and loosening the bounds helps deps fit in one stack resolver.