Haddock comment in between constructor and fields puts comment in illegal position

[chiroptical]

Describe the bug

If you put a Haddock comment above the record braces and format, ormolu moves the haddock comment into an illegal position.

To Reproduce

[nix-shell:~/mercury/mercury-web-backend]$ ormolu --version
ormolu 0.7.2.0
using ghc-lib-parser 9.6.6.20240701

[nix-shell:~/mercury/mercury-web-backend]$ cat src/Example.hs 
module Example where

data Hello = Hello
  -- | hello world
  {hello :: String}

[nix-shell:~/mercury/mercury-web-backend]$ ormolu --unsafe src/Example.hs
module Example where

data Hello = Hello
  {-- | hello world
   hello :: String}

Compiler yells at me,

[1 of 1] Compiling Example          ( src/Example.hs, interpreted ) [Source file changed]
src/Example.hs:(4,3)-(6,0): error: [GHC-21231]
    unterminated `{-' at end of input
  |
4 |   {-- | hello world
  |   ^^^^^^^^^^^^^^^^^...

Expected behavior

I don't think this is a valid Haddock comment? If it isn't a valid Haddock comment then the comment shouldn't move at all. If it is considered valid, then I guess the output should be,

module Example where

data Hello = Hello
  { -- | hello world
   hello :: String}

Another option might be,

module Example where

 -- | hello world
data Hello = Hello
  {hello :: String}

I am not a Haddock expert.

Environment

• OS name + version: PopOS, 22.04
• Version of the code: 0.7.2.0 (from nixpkgs)

Additional context
Add any other context about the problem here.

[amesgen]

Thanks for the report!

The problem here is that LConDeclField (the AST element representing hello :: String) has a single-line span, so we use single line layout for formatting it. However, we put the Haddock comment inside of the braces, so we would need to use multi-line formatting instead.

It is easy to write an ad-hoc fix for this, but it feels unprincipled, see the similar discussion in #752/#757, where a Haddock comment was also moved "into" another construct requiring similar changes (and a very similar fix would also work here).

---

Maybe we could/should consider such cases to be upstream bugs in the GHC logic for attaching Haddock comments? WDYT @mrkkrp?

[mrkkrp]

@amesgen If I understand correctly, your point is that if there is a Haddock attached to an AST element the span of that element should actually encompass the Haddock?

[amesgen]

Yeah, either that, or GHC could consider this comment to be an invalid Haddock comment (in which case we format it using the normal non-Haddock logic, where everything is fine in the example above).

[mrkkrp]

This does make sense 👍