Re: [GHC] #11731: Simplifier: Inlining trivial let can lose sharing

30 Mar 2016

      #11731: Simplifier: Inlining trivial let can lose sharing
-------------------------------------+-------------------------------------
        Reporter:  nomeata           |                Owner:
            Type:  bug               |               Status:  patch
        Priority:  normal            |            Milestone:
       Component:  Compiler          |              Version:  8.1
      Resolution:                    |             Keywords:
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):  Phab:D2064
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by simonpj):

 For the record, here is the proposed fix (currently on a wip branch):
 {{{
 commit 7ac5610606e8f338cd2eb92eb5d711e054d9d55a
 Author: Joachim Breitner 
 Date:   Wed Mar 30 12:55:10 2016 +0200

     Used-once variables are not trivial

     The specification for exprIsTrivial demands that we are
 unconditionally
     happy to duplicate the expression. This is not true for variables
 where
     we would like to exploit (or already have exploited) that they are
 used
     at most once. In order to preserve this property, they must not be
     duplicated nilly-willy. This fixes #11731 and comes with a test case.
...
---------------------------------------------------------------
7ac5610606e8f338cd2eb92eb5d711e054d9d55a
  compiler/coreSyn/CoreUtils.hs                      | 12 +++++++-
  testsuite/.gitignore                               |  1 +
  testsuite/tests/simplCore/should_run/T11731.hs     | 36
 ++++++++++++++++++++++
  testsuite/tests/simplCore/should_run/T11731.stderr |  1 +
  testsuite/tests/simplCore/should_run/all.T         |  1 +
  5 files changed, 50 insertions(+), 1 deletion(-)

 diff --git a/compiler/coreSyn/CoreUtils.hs b/compiler/coreSyn/CoreUtils.hs
 index 1d9b83b..d02b934 100644
 --- a/compiler/coreSyn/CoreUtils.hs
 +++ b/compiler/coreSyn/CoreUtils.hs
 @@ -62,6 +62,7 @@ import DataCon
  import PrimOp
  import Id
  import IdInfo
 +import Demand ( isUsedOnce )
  import Type
  import Coercion
  import TyCon
 @@ -755,6 +756,13 @@ Note [exprIsTrivial]

  Note [Variables are trivial]
  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 +Variables are usually trivial.
 +
 +Except if 'isUsedOnce (idDemandInfo v) == True':
 +In this case we have previously determined that a variable is used at
 +most once, and we likely rely on this information, e.g. during code
 +generation. In this case, we are not unconditionally happy to duplicate,
 so we don’t.  See #11731.
 +
  There used to be a gruesome test for (hasNoBinding v) in the  Var case:
          exprIsTrivial (Var v) | hasNoBinding v = idArity v == 0 @@ -793,7
 +801,9 @@ it off at source.
  -}

  exprIsTrivial :: CoreExpr -> Bool
 -exprIsTrivial (Var _)          = True        -- See Note [Variables are
 trivial]
 +exprIsTrivial (Var v) -- See Note [Variables are trivial]
 +  | isUsedOnce (idDemandInfo v) = False
 +  | otherwise                  = True
  exprIsTrivial (Type _)         = True
  exprIsTrivial (Coercion _)     = True
  exprIsTrivial (Lit lit)        = litIsTrivial lit
 }}}

--
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/11731#comment:16
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler