[GHC] #11735: Optimize coercionKind

[GHC]

#11735: Optimize coercionKind
-------------------------------------+-------------------------------------
        Reporter:  goldfire          |                Owner:  (none)
            Type:  task              |               Status:  new
        Priority:  normal            |            Milestone:
       Component:  Compiler          |              Version:  7.10.3
      Resolution:                    |             Keywords:
Operating System:  Unknown/Multiple  |         Architecture:
 Type of failure:  Compile-time      |  Unknown/Multiple
  performance bug                    |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by tdammers):

 I ran a selection of commits against two test inputs: Grammar.hs, as
 above, and the `test_rules2.hs` file from #5631. Here's a list of
 execution times for both:

 {{{
 --- ./cases/Grammar.hs ---
 452dee3ff4: 245.51
 4572849929: 19.62
 4eb140f564: 12.25
 8a6aa5030d: 12.26
 d74b37d565: 16.14
 --- ./cases/test_rules2.hs ---
 452dee3ff4: 1.20
 4572849929: 1.19
 4eb140f564: 1.19
 8a6aa5030d: 1.23
 d74b37d565: 1.18
 }}}

 The commits are:

 {{{
 452dee3ff4: GHC `master` before branching off on this issue (should be the
 same as GHC HEAD)
 4572849929: Simon's patch from [comment:5] applied.
 4eb140f564: After "un-refactoring"
 8a6aa5030d: With coercion roles cached, and mkNthCo taking an extra Role
 argument
 d74b37d565: current HEAD of wip/tdammers/T11735
 }}}

 (Note that the execution times are a bit faster overall in this run
 because I didn't pass any `-ddump` flags, so GHC spends no time pretty-
 printing the dumps).

 So, conclusions:

 - The [comment:5] patch makes a huge difference for `Grammar.hs` (down to
 less than 10% in execution time)
 - "un-refactoring", strangely enough, improves performance on `Grammar.hs`
 by roughly another 25% or so
 - caching coercion rules in NthCo doesn't seem to make a difference at
 all, and even makes things slightly worse (this one is truly baffling IMO)
 - improvements from [comment:41] make the `Grammar.hs` test case worse
 (this, too, is unexpected)
 - none of the changes here seem to affect performance for the
 `test_rules2.hs` case much, if at all

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/11735#comment:46>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler