[GHC] #13964: Pattern-match warnings for datatypes with COMPLETE sets break abstraction

[GHC]

#13964: Pattern-match warnings for datatypes with COMPLETE sets break abstraction
-------------------------------------+-------------------------------------
        Reporter:  RyanGlScott       |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:
       Component:  Compiler          |              Version:  8.2.1-rc2
      Resolution:                    |             Keywords:
                                     |  PatternSynonyms,
                                     |  PatternMatchWarnings
Operating System:  Unknown/Multiple  |         Architecture:
 Type of failure:  Poor/confusing    |  Unknown/Multiple
  error message                      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by RyanGlScott):

 Perhaps I'm fundamentally misunderstanding something about `COMPLETE`
 sets, but this doesn't feel like a very satisfying explanation. My
 intuition is that specifying a `COMPLETE` set for a data type is
 tantamount to overriding the pattern match exhaustiveness checker's
 behavior w.r.t. to that particular data type, and as such, I would any
 warnings that are emitted under the `-Wincomplete-patterns` label to be
 adjusted according to whatever `COMPLETE` sets are in scope. For example,
 in my original example, I would simply expect that in this warning:

 {{{
 Foo.hs:6:1: warning: [-Wincomplete-patterns]
     Pattern match(es) are non-exhaustive
     In an equation for ‘catchAll2’: Patterns not matched: Bug.T
 }}}

 It would warn about `TooGoodToBeTrue`, not `Bug.T`. That's all.

 Replying to [comment:2 mpickering]:
 > This is nothing unique to `COMPLETE` pragmas, the exhaustiveness checker
 already breaks abstraction in this way.
 >
 > For example:
 >
 > {{{
 > module T (D(A)) where
 >
 > data D = A | B
 > }}}
 >
 > {{{
 > module M where
 >
 > import T
 >
 > foo A = ()
 > }}}
 >
 > warns with
 >
 > {{{
 > M.hs:5:1: warning: [-Wincomplete-patterns]
 >     Pattern match(es) are non-exhaustive
 >     In an equation for ‘foo’: Patterns not matched: T.B
 >   |
 > 5 | foo A = ()
 >   |
 > }}}

 That seems fine to me, since the user didn't specify `{-# COMPLETE A #-}`
 for the data type `D`. If she did, I wouldn't expect a warning for `foo`.

 > and after all, to retain safety it has to. A function exported by `T`
 could mention `B` and then be used in `M`.

 I don't understand what you mean. A function exported from `T` surely
 doesn't have any effect on whether a function defined in `M` is
 exhaustive, right?

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/13964#comment:3>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler