[GHC] #9374: Investigate Static Argument Transformation

[GHC]

#9374: Investigate Static Argument Transformation
-------------------------------------+-------------------------------------
        Reporter:  jstolarek         |                Owner:  (none)
            Type:  task              |               Status:  new
        Priority:  lowest            |            Milestone:
       Component:  Compiler          |              Version:  7.9
      Resolution:                    |             Keywords:
                                     |  StaticArgumentTransformation
Operating System:  Unknown/Multiple  |         Architecture:
 Type of failure:  Runtime           |  Unknown/Multiple
  performance bug                    |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by mpickering):

 I looked at all the examples where allocations changed a significant
 amount. Essentially it seems to me the transformation is bad for
 allocations if it doesn't lead to more simplification but potentially very
 good if it does.

 There were particular situations where overloaded functions were satted as
 they had a normal static argument and a known dictionary argument.
 Normally the specialiser would be expected to deal with these but self-
 recursive functions are only specialised across modules if `-fspecialise-
 aggresively` is turned on or they are marked with an `INLINABLE` pragma.
 (For example `CSD`, `gg`).

 In situations where the definition was not further simplified, allocations
 tended to increase by a few %. So I'm not convinced that the example in
 comment:13 would be a win. There were also some situations where the
 unfolding for the satted definition was not exported as it was too big so
 there was never any potential for improvement.

 To sum up, there were some wins where SAT acted across modules where
 normal specialisation failed. Satting static function arguments could also
 be very good but only if there are actually some cases where the function
 is called with a static function argument. (For example #14211) In general
 though there was very little benefit satting as it seems that static
 values are not usually consumed by functions (and so no further
 simplification takes place after the sat).

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9374#comment:24>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler