[GHC] #10711: Defining mapM_ in terms of traverse_ causes substantial blow-up in ByteCodeAsm

[GHC]

#10711: Defining mapM_ in terms of traverse_ causes substantial blow-up in
ByteCodeAsm
-------------------------------------+-------------------------------------
              Reporter:  bgamari     |             Owner:  bgamari
                  Type:  bug         |            Status:  new
              Priority:  normal      |         Milestone:
             Component:  Compiler    |           Version:
              Keywords:              |  Operating System:  Unknown/Multiple
          Architecture:              |   Type of failure:  Compile-time
  Unknown/Multiple                   |  performance bug
             Test Case:  ghcirun004  |        Blocked By:
              Blocking:              |   Related Tickets:
Differential Revisions:              |
-------------------------------------+-------------------------------------
 Phab:D924 has proposed that we redefine `mapM_`, currently,
 {{{
 mapM_ = foldr ((>>) . f) (return ())
 }}}
 as,
 {{{
 mapM_ = foldr ((*>) . f) (return ()) = traverse_
 }}}
 as part of the AMP proposal.

 However, this appears to have severe effects on the performance
 characteristics of the `Assembler` monad defined in `ByteCodeAsm`. In
 particular, the `mapM_` use in `ByteCodeAsm.assembleBCO` blows up
 severely, increasing the runtime of the `ghcirun004` testcase from 4
 seconds to over 5 minutes.

 Intriguingly, defining `(*>) = (>>)` in `Assembler`'s `Applicative`
 instance (as done in Phab:D1097) restores reasonable runtime.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10711>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler