[GHC] #16004: Vector performance regression in GHC 8.6

[GHC]

#16004: Vector performance regression in GHC 8.6
-------------------------------------+-------------------------------------
        Reporter:  guibou            |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:  8.6.3
       Component:  Compiler          |              Version:  8.6.2
      Resolution:                    |             Keywords:
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 simonpj):

 > with full-laziness we float out the creation of the [0..n] list,

 This particular thing is quite delicate.  Consider
 {{{
 f xs = map (\x -> foo x [1..100]) xs
 }}}
 Should we float that `[1..100]` out to top level, and share it among all
 the calls that `map` makes?  Or should we reconstruct it on every call.
 In the latter case we might be able to fuse it with the consumer (perhaps
 we inline `foo`).  But even if we fuse away the list we might still
 allocate heap objects for `I# 1`, `I# 2`, ... `I# 100`, for each element
 of `xs` rather sharing those values between all those calls.  Or, perhaps,
 we get to fuse away those `I#` boxes too!  It all depends on `foo`.

 I don't really have a good answer here.  Sometimes full laziness is a win,
 sometimes not.

 There is more discussion in #7206.

 I'd love someone to work on this a bit.  Somehow we should do better than
 we are doing.

 PS: I have no idea why the particular commit that made this test worse did
 make it worse; that might be worth investigation.

 Another thing that would be v useful is to distil the example into a
 standalone test case.  `vector` is a complicated library, but David's
 analysis above isolates the issue well. Maybe someone can turn that into a
 repro case?

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/16004#comment:5>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler