[GHC] #15333: Weird cachegrind results for binary-trees

[GHC]

#15333: Weird cachegrind results for binary-trees
-------------------------------------+-------------------------------------
        Reporter:  sgraf             |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:  Research
       Component:  NoFib benchmark   |  needed
  suite                              |              Version:  8.5
      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 sgraf):

 I'm still not really sure what's going on, but I feel like alignment
 instructions might skew cachegrinds instruction counts according to
 [http://valgrind.org/docs/manual/cg-manual.html section 5.2.10 of the
 valgrind manual]. I've had a few other cases in the meantime. Passing
 `-fllvm -optlo -Os` to use the LLVM backend optimising for code size
 helps.

 Is there some equivalent for the NCG? That would be far more reliable for
 the counted instruction metric that our benchmark CI relies on.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/15333#comment:2>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler