[GHC] #9098: Don't use unsafe coerce when desugaring lazy pattersn

12 May 2014

      #9098: Don't use unsafe coerce when desugaring lazy pattersn
------------------------------------+-------------------------------------
       Reporter:  simonpj           |             Owner:
           Type:  bug               |            Status:  new
       Priority:  normal            |         Milestone:
      Component:  Compiler          |           Version:  7.8.2
       Keywords:                    |  Operating System:  Unknown/Multiple
   Architecture:  Unknown/Multiple  |   Type of failure:  None/Unknown
     Difficulty:  Unknown           |         Test Case:
     Blocked By:                    |          Blocking:
Related Tickets:                    |
------------------------------------+-------------------------------------
 Christiaan Baaij asks (on ghc-devs): when I ask for the desugaring of:
 {{{
 module PatError where

 paterror :: Maybe Int -> Int
 paterror (Just i) = i
 }}}
 I get the following:
 {{{
 PatError.paterror =
   \ (ds_dIS :: Data.Maybe.Maybe GHC.Types.Int) ->
     break<1>()
     case ds_dIS of _ [Occ=Dead] {
       __DEFAULT ->
         (\ _ [Occ=Dead, OS=OneShot] ->
            Control.Exception.Base.patError
              @ GHC.Types.Int "PatError.hs:4:1-21|function paterror"#)
           GHC.Prim.void#;
       Data.Maybe.Just i_aqG -> break<0>(i_aqG) i_aqG
     }
 }}}
 Where `Control.Exception.Base.patError` is applied to `Int`.

 However, when I compile the almost identical code with an irrefutable
 pattern:
 {{{
 module PatError2 where

 patError :: Maybe Int -> Int
 patError ~(Just i) = i
 }}}
 I get the following core:
 {{{
 PatError2.paterror =
   \ (ds_dIT :: Data.Maybe.Maybe GHC.Types.Int) ->
     break<1>()
     let {
       i_aqG :: GHC.Types.Int
       [LclId, Str=DmdType]
       i_aqG =
         case ds_dIT of _ [Occ=Dead] {
           __DEFAULT ->
             (\ _ [Occ=Dead, OS=OneShot] ->
                (Control.Exception.Base.irrefutPatError
                   @ () "PatError2.hs:4:1-22|(Data.Maybe.Just i)"#)
                `cast` (UnivCo representational () GHC.Types.Int
                        :: () ~# GHC.Types.Int))
               GHC.Prim.void#;
           Data.Maybe.Just i_aqG -> i_aqG
         } } in
     break<0>(i_aqG) i_aqG
 }}}
 Where `Control.Exception.Base.patError` is first applied to `()` after
 which it is casted to `GHC.Types.Int`.
 Out of the `[Note]` for the `UnivCo` coercion:
 {{{
 The UnivCo ("universal coercion") serves two rather separate functions:
  - the implementation for unsafeCoerce#
  - placeholder for phantom parameters in a TyConAppCo
 }}}

 It seems to be used in the `unsafeCoerce#` role here. My question is: why
 can't the irrefutable pattern be translated to
 `Control.Exception.Base.irrefutPatError' applied to 'Int'?  Instead of
 using this 'cast'?  I find the cast-less version much more readable.

 Greetings,

 Christiaan

--
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/9098
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler