fix shebang line parsing

[kokobd]

This will fix #2724.

The PR is created with a test case for this, without any actual code fix currently.

[kokobd]

In hls, we are not using the parseModule in ghc-exactprint directly. Instead, we manually call GHC's api and construct Anns using relativiseApiAnns. So something is different between the parsing in hls and ghc-exactprint. Let's take a look at ghc-exactprint-0.6.4:

-- | Internal part of 'parseModuleFromString'.
parseModuleFromStringInternal :: Parser GHC.ParsedSource
parseModuleFromStringInternal dflags fileName str =
  let (str1, lp) = stripLinePragmas str
      res        = case runParser GHC.parseModule dflags fileName str1 of
#if __GLASGOW_HASKELL__ > 808
        GHC.PFailed pst     -> Left (GHC.getErrorMessages pst dflags)
#elif __GLASGOW_HASKELL__ >= 804
        GHC.PFailed _ ss m  -> Left (ss, GHC.showSDoc dflags m)
#else
        GHC.PFailed ss m    -> Left (ss, GHC.showSDoc dflags m)
#endif
        GHC.POk     x  pmod -> Right (mkApiAnns x, lp, dflags, pmod)
  in  postParseTransform res normalLayout

It calls stripLinePragmas which itself handles shebang specially, as below (in another function, checkLine)

checkLine :: Int -> String -> (String, Maybe Comment)
checkLine line s
  |  "{-# LINE" `isPrefixOf` s =
       let (pragma, res) = getPragma s
           size   = length pragma
           mSrcLoc = mkSrcLoc (mkFastString "LINE")
           ss     = mkSrcSpan (mSrcLoc line 1) (mSrcLoc line (size+1))
       in (res, Just $ mkComment pragma (rs ss))
  -- Deal with shebang/cpp directives too
  -- x |  "#" `isPrefixOf` s = ("",Just $ Comment ((line, 1), (line, length s)) s)
  |  "#!" `isPrefixOf` s =
    let mSrcLoc = mkSrcLoc (mkFastString "SHEBANG")
        ss = mkSrcSpan (mSrcLoc line 1) (mSrcLoc line (length s))

In ghcide, I can't find code handling shebang like this, in getParsedModuleWithCommentsRule. And the test case shows it is not handled properly.

To fix this, I'm going to add a call stripLinePragmas in ghcide, like what has been done in ghc-exactprint

[michaelpj]

In hls, we are not using the parseModule in ghc-exactprint directly.

Do you know why? Could we start using it?

[kokobd]

In hls, we are not using the parseModule in ghc-exactprint directly.

That's probably because we need pass our own DynFlags and some other options to GHC's parser.

Moreover, I recently discovered that, using ghc-9.2.1 with ghc-exactprint-1.5.0 together will NOT exactprint shebang lines.

[kokobd]

This is harder than I thought. Closing.