[GHC] #12990: Partially applied constructors with unpacked fields simplified badly

15 Dec 2016

      #12990: Partially applied constructors with unpacked fields simplified badly
-------------------------------------+-------------------------------------
           Reporter:  dfeuer         |             Owner:
               Type:  bug            |            Status:  new
           Priority:  normal         |         Milestone:
          Component:  Compiler       |           Version:  8.0.1
           Keywords:                 |  Operating System:  Unknown/Multiple
       Architecture:                 |   Type of failure:  Runtime
  Unknown/Multiple                   |  performance bug
          Test Case:                 |        Blocked By:
           Blocking:                 |   Related Tickets:
Differential Rev(s):                 |         Wiki Page:
-------------------------------------+-------------------------------------
 When a constructor is partially applied to at least one unpacked field,
 the simplifier produces pretty lousy results.

 {{{#!hs
 data Foo a = Foo !Int a

 silly :: ((a -> Foo a) -> r) -> r
 silly f = f (Foo 3)
 }}}

 compiles to

 {{{
 -- RHS size: {terms: 9, types: 7, coercions: 0}
 $WFoo :: forall a. Int -> a -> Foo a
 $WFoo =
   \ (@ a_aqI) (dt_ayl :: Int) (dt_aym :: a_aqI[sk:1]) ->
     case dt_ayl of { I# dt_ayn -> Foo dt_ayn dt_aym }

 -- RHS size: {terms: 2, types: 0, coercions: 0}
 silly2 :: Int
 silly2 = I# 3#

 -- RHS size: {terms: 3, types: 3, coercions: 0}
 silly1 :: forall a. a -> Foo a
 silly1 = \ (@ a_ayG) -> $WFoo silly2

 -- RHS size: {terms: 5, types: 9, coercions: 0}
 silly :: forall a r. ((a -> Foo a) -> r) -> r
 silly =
   \ (@ a_ayG) (@ r_ayH) (f_axY :: (a -> Foo a) -> r) -> f_axY silly1
 }}}

 That is, GHC represents `Foo 3` as a closure containing a ''boxed'' `Int`.
 Manually eta-expanding would fix it.

 {{{#!hs
 silly :: ((a -> Foo a) -> r) -> r
 silly f = f (\x -> Foo 3 x)
 }}}

 compiles to

 {{{#!hs
 -- RHS size: {terms: 5, types: 4, coercions: 0}
 silly1 :: forall a. a -> Foo a
 silly1 = \ (@ a_ayO) (x_ay9 :: a) -> Foo 3# x_ay9

 -- RHS size: {terms: 5, types: 9, coercions: 0}
 silly :: forall a r. ((a -> Foo a) -> r) -> r
 silly =
   \ (@ a_ayO) (@ r_ayP) (f_ay8 :: (a -> Foo a) -> r) -> f_ay8 silly1
 }}}

 in which the `Int` is unpacked into the closure. This transformation is
 valid whenever the value to be stored unpacked is known not to be bottom,
 and is certainly a good idea if it's known to have been forced.

--
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/12990
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler