Re: [GHC] #10665: INLINE breaks rewrite rules when '-g' is used

23 Jul 2015

      #10665: INLINE breaks rewrite rules when '-g' is used
-------------------------------------+-------------------------------------
        Reporter:  slyfox            |                   Owner:
            Type:  bug               |                  Status:  new
        Priority:  normal            |               Milestone:
       Component:  Compiler          |                 Version:  7.10.2-rc2
  (CodeGen)                          |
      Resolution:                    |                Keywords:
Operating System:  Unknown/Multiple  |            Architecture:
 Type of failure:  Compile-time      |  Unknown/Multiple
  crash                              |               Test Case:
      Blocked By:                    |                Blocking:
 Related Tickets:                    |  Differential Revisions:
-------------------------------------+-------------------------------------

Comment (by slyfox):

 And here goes conduit sample: http://code.haskell.org/~slyfox/when-rules-
 tick-in-conduit-T10665.tar.gz (3MB of haskell code). Tarball contains
 original hackage packages with 2-3
 language pragmas added.

 It does not require external depends and can be built with inplace/bin
 /ghc-stage2.
 Fails thusly on 7.10.2:
 {{{
 $ ./trigger-a-bug.bash

 [97 of 99] Compiling Data.Conduit.Extra.ZipConduitSpec (
 conduit-1.2.4.3/test/Data/Conduit/Extra/ZipConduitSpec.hs,
 conduit-1.2.4.3/test/Data/Conduit/Extra/ZipConduitSpec.o )
 ghc: panic! (the 'impossible' happened)
   (GHC version 7.10.2 for x86_64-unknown-linux):
         Tick in rule
   unstream
     @ a21_a4C66
     @ Void
     @ m_a4C64
     @ b_a4C65
     (ConduitWithStream
        @ a21_a4C66
        @ Void
        @ m_a4C64
        @ b_a4C65
        (let {
           $dApplicative_a4BDD
             :: Applicative (ConduitM a21_a4C66 Void m_a4C64)
           [LclId, Str=DmdType]
           $dApplicative_a4BDD =
             $fApplicativeConduitM
               @ a21_a4C66
               @ Void
               @ m_a4C64
               ($fFunctorConduitM @ a21_a4C66 @ Void @ m_a4C64) } in
 ...
 }}}

 To follow by source code:
 {{{#!hs
 --| rewrite rule is triggered by '$$' operator at
 conduit-1.2.4.3/test/Data/Conduit/Extra/ZipConduitSpec.hs:
 res <- src $$ conduit =$ sink

 --| '''conduit-1.2.4.3/Data/Conduit/List.hs''' refines rewrite rule as:
 {-# RULES "conduit: $$ fold" forall src f b. src $$ fold f b = connectFold
 src f b #-}

 --| '''fold''' is defined at conduit-1.2.4.3/Data/Conduit/List.hs:
 fold, foldC :: Monad m
             => (b -> a -> b)
             -> b
             -> Consumer a m b
 foldC f =
     loop
   where
     loop !accum = await >>= maybe (return accum) (loop . f accum)
 {-# INLINE foldC #-}
 STREAMING(fold, foldC, foldS, f accum)

 --| STREAMING is a macro in conduit-1.2.4.3/fusion-macros.h that inlines
 'fold' lately:
 #define STREAMING(name, nameC, nameS, vars)                             ;\
     name = nameC                                                        ;\
     {-# INLINE [0] name #-}                                             ;\
     {-# RULES "unstream name" forall vars.                              \
       name vars = unstream (streamConduit (nameC vars) (nameS vars))    \
       #-}
 }}}

 Here delaying inline does not help as rewrite rule becomes broken
 when compliling one module (where it's defined) and used in another.

 That's the reson why original reproducer so carefully setups phases.

 To be able to compile it under windows you might need to change
 {{{
 -D'UNIX=1' \
 }}}
 to
 {{{
 -D'WINDOWS=1' \
 }}}
 in '''trigger-a-bug.bash''' script.

--
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/10665#comment:3
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler