Re: [GHC] #16357: Add `oneShot` to the implementation of foldlM

25 Feb 2019

      #16357: Add `oneShot` to the implementation of foldlM
-------------------------------------+-------------------------------------
        Reporter:  autotaker         |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:
       Component:  libraries/base    |              Version:  8.9
      Resolution:                    |             Keywords:
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by autotaker):

 This bug is caused by INLINE annotation for `c` which is introduced while
 fixing #8763.

 After list fusion, the example program would be:
 {{{#!hs
 f (I# n) =
   letrec go i =
         case i `remInt#` 2 of
            1# -> case i ==# n of
              0# -> go (i +# 1)
              1# -> pure
            0# -> c (I# i) (case i ==# n of
              0# -> go (i +# 1)
              1# -> pure)
   in go 0# (I# 0)
 c x k z = k $! (z + x)
 {-# INLINE c #-}
 }}}
 We can see function `c` is lifted to the **top-level definition**.
 Besides, the continuation `case i ==# n of { 0# -> go (i +# 1); 1# -> pure
 }`
 is passed as a **higher-order** argument of `c`.
 Therefore, CallArity analyzer cannot find the correct call-arity of
 function `go`.

 Without INLINE pragma for `c`, we have:
 {{{#!hs
 f (I# n) =
   letrec go i =
         case i `remInt#` 2 of
            1# -> case i ==# n of
              0# -> go (i +# 1)
              1# -> pure
            0# ->
              let k = case i ==# n of
                      0# -> go (i +# 1)
                      1# -> pure in
              \z -> k $! I# (z + (I# i))
   in go 0# (I# 0)
 }}}
 We can see the continuation `k` is defined inside `go`.
 In this case, CallArity works good and finds the call-arity of `go` as 3
 (not 2 because it includes `State# Realworld`).

 If we define `c x k = \z -> f z x >>= k` (instead of `c x k z = f z x >>=
 k`) and add an INLINE pragma for `c`,
 CallArity works good because the continuation is defined inside `go`.

-- 
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/16357#comment:2
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler