Cabal: Rewrite the cabal file parser using parsec

[phadej]

The existing parser has no formal grammar, has terrible error messages and is sometimes very slow and memory hungry. In large part this is because ReadP is a terrible parser. Now that the ghc library no longer depends on the Cabal library, Cabal is now allowed to depend on parsec.

The Cabal parser is two-stage, outline then individual fields. This approach will remain. There is a prototype new parser (with a grammar!) using parsec for the outline phase which has had some significant testing (the new grammar only rejects a very small number of old quirky .cabal files). This new parser needs to be integrated and tested further. Additionally the infrastructure for parsing the individual fields needs to be rewritten to use parsec.

[23Skidoo]

This ticket should be either closed or updated to reflect the current state of affairs.

[phadej]

@23Skidoo, briefly the next steps are

• Setup parsec (mtl and text) into GHC tree
• Rip-off non-parsec stuff from Cabal
• ...
• Profit

But I don't want to bother @bgamari on this before GHC-8.2.1 is out.

[BardurArantsson]

... and then upgrade to Megaparsec? :)

(Sorry, couldn't resist.)

[hvr]

@BardurArantsson that's a fair point :-) the main reason to stick with parsec right now is IMO that its dependency footprint is lighter, and given it's going to become a GHC bundled library, it's a good thing that parsec is so mature by now that its development has slowed down (while megaparsec is actively evolving). ;-)