[GHC] #7367: Optimiser / Linker Problem on amd64

[GHC]

#7367: Optimiser / Linker Problem on amd64
--------------------------------------------+------------------------------
        Reporter:  wurmli                   |            Owner:
            Type:  bug                      |           Status:  new
        Priority:  normal                   |        Milestone:  7.8.1
       Component:  Build System             |          Version:  7.6.1
      Resolution:                           |         Keywords:
Operating System:  Linux                    |     Architecture:  x86_64
 Type of failure:  Runtime performance bug  |  (amd64)
       Test Case:                           |       Difficulty:  Unknown
        Blocking:                           |       Blocked By:
                                            |  Related Tickets:
--------------------------------------------+------------------------------

Comment (by rwbarton):

 Actually, `flopS` uses its "continuation" `f` linearly, like the CPS
 translation of an ordinary imperative function.  Teaching GHC to recognize
 this sounds like quite an undertaking, though.

 This seems like a typical instance of how it is hard to decide whether to
 float out a let.

 The trouble with let-floating from the user's perspective is that when GHC
 wrongly decides to float out a let, it is virtually impossible to stop it
 from doing so.  One way out is `-fno-full-laziness`, but GHC probably was
 correctly floating out a lot of other things, and it's difficult and
 tedious to locate and transform all those parts of the code.

 One suggestion is to identify situations in which let-floating can never
 be harmful, and allow `-fno-full-laziness` (or perhaps a new flag,
 `-fsome-full-laziness`?) to perform those floatings.  For example, I don't
 think it can ever cost to float a binding whose RHS is a lambda to the top
 level.  That rule would float out `flopp` and `swap` from `flop` here, and
 compiling with `-fno-full-laziness` would then eliminate all allocations.

 It would also be very useful to be able to control let-floating locally
 with some kind of pragma or magic built-in function.  I'm sure this idea
 has been explored before and I also have some ideas of how to go about it,
 but I want to get a little more familiar with the structure of GHC before
 proposing anything.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/7367#comment:7>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler