Re: [GHC] #13429: Optimizer produces Core with an infinite <<loop>>

1 May 2017

      #13429: Optimizer produces Core with an infinite <<loop>>
-------------------------------------+-------------------------------------
        Reporter:  lehins            |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  high              |            Milestone:  7.10.4
       Component:  Compiler          |              Version:  8.0.2
      Resolution:                    |             Keywords:
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  Runtime crash     |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by dfeuer):

 I've reduced it once more. I believe the fundamental problem here happens
 in specialization. We get

 {{{
 Rec {
 -- RHS size: {terms: 2, types: 1, coercions: 0, joins: 0/0}
 $dArray_s1RC :: Array X Word8
 [LclId]
 $dArray_s1RC = Loop.$fArrayXe @ Word8 GHC.Word.$fNumWord8

 -- RHS size: {terms: 2, types: 1, coercions: 0, joins: 0/0}
 $dArray1_s1RG [Occ=OnceL] :: Array X Word8
 [LclId]
 $dArray1_s1RG = Loop.$fArrayXe @ Word8 $dNum_s1RH

 -- RHS size: {terms: 2, types: 2, coercions: 0, joins: 0/0}
 $dNum_s1RH :: Num Word8
 [LclId]
 $dNum_s1RH = Loop.$p1Array @ X @ Word8 $dArray_s1RC

 -- RHS size: {terms: 7, types: 5, coercions: 0, joins: 0/0}
 $s$fArrayXe_s1RS [InlPrag=CONLIKE] :: Array X Word8
 [LclId,
  Unf=DFun: \ ->
        Loop.C:Array TYPE: X
                     TYPE: Word8
                     Loop.$fArrayXe_$cp1Array @ Word8 $dNum_s1RH
                     Loop.$fArrayXe_$cpromote @ Word8 $dNum_s1RH
                     Loop.$fArrayXe_$cmakeImage @ Word8 $dNum_s1RH]
 $s$fArrayXe_s1RS =
   Loop.C:Array
     @ X
     @ Word8
     (Loop.$fArrayXe_$cp1Array @ Word8 $dNum_s1RH)
     (Loop.$fArrayXe_$cpromote @ Word8 $dNum_s1RH)
     (Loop.$fArrayXe_$cmakeImage @ Word8 $dNum_s1RH)

 ...
 }}}

 which doesn't seem ''inherently'' bad (although it's not very clean), but
 then we also get this rather awful specialization rule:

 {{{
 "SPEC/Main $fArrayXe @ Word8"
     forall (v_s1RJ :: Num Word8).
       Loop.$fArrayXe @ Word8 v_s1RJ
       = $s$fArrayXe_s1RS
 }}}

 If this rule fires in the RHS of `$dArray_s1RC`, then we get

 {{{
 $dArray_s1RC = $s$fArrayXe_s1RS
 }}}

 Inlining that into `$dNum_s1RH` gives

 {{{
 $dNum_s1RH = Loop.$p1Array @ X @ Word8 $s$fArrayXe_s1RS
 }}}

 But `$dNum_s1RH` is used in the definition of `$s$fArrayXe_s1RS`, so we
 have a loop. I don't know if I've told exactly the right story here, but I
 think the real story is probably pretty similar. The results will be
 excruciatingly fragile, depending on whether the right or wrong
 specialization rule fires.

--
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/13429#comment:14
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler