[GHC] #14644: Improve cmm/assembly for pattern matches with two constants.

[GHC]

#14644: Improve cmm/assembly for pattern matches with two constants.
-------------------------------------+-------------------------------------
        Reporter:  AndreasK          |                Owner:  (none)
            Type:  task              |               Status:  new
        Priority:  normal            |            Milestone:
       Component:  Compiler          |              Version:  8.2.2
  (CodeGen)                          |             Keywords:  Codegen, CMM,
      Resolution:                    |  Patterns, Pattern Matching
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 AndreasK):

 Replying to [comment:3 svenpanne]:
 > I am not sure if things are that easy: To do these kind of switches in a
 really good way, one would need a probability distribution how often the
 individual cases actually happen.

 This is by no means meant as a "this is optimal" idea.

 As far as I'm aware there is currently no way to propagate probabilities
 through GHC's various stages. So even in cases where we can make a
 judgement on the probability we can't really use it at the Cmm level.

 In the end I suggested it because it:
 * I think it has a good chance of being faster
 * Leads to smaller code
 * Seems to be what gcc/clang do (using conditional moves instead of jumps
 but still).

 > Additional things to consider: Performance in tight loops is often
 vastly different, because branch prediction/caching will most likely kick
 in visibly. Correctly predicted branches will cost you almost nothing,
 while unknown/incorrectly predicted branches will be much more costly. In
 the absence of more information from their branch predictor, quite a few
 processors assume that backward branches are taken and forward branches
 are assumed to be not taken. So code layout has a non-trivial performance
 impact.
 >
 > Instruction counts are quite misleading nowadays...

 Indeed :(

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/14644#comment:4>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler