[GHC] #11371: Bogus in-scope set in substitutions

[GHC]

#11371: Bogus in-scope set in substitutions
-------------------------------------+-------------------------------------
        Reporter:  simonpj           |                Owner:  niteria
            Type:  bug               |               Status:  new
        Priority:  high              |            Milestone:  8.2.1
       Component:  Compiler          |              Version:  7.10.3
      Resolution:                    |             Keywords:
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:  #11360            |  Differential Rev(s):  phab:D1792,
       Wiki Page:                    |  phab:D1801, phab:D1802
-------------------------------------+-------------------------------------

Comment (by niteria):

 Btw this is a subset of tests from #11196 which got 30-60% worse after
 TypeInType.

 > I do not expect the in-scope set to be evaluated at all (that is, it
 should remain as a thunk) when substituting into for-all-free types.

 I've experimented a bit and you're right. The in-scope set is never forced
 for T5030. Why do we see a difference then?

 I've compared:

 {{{
 empty_subst = mkEmptyTCvSubst (pprTrace "forced" (ppr (ty,args)) $
   mkInScopeSet $ tyCoVarsOfTypes (ty:args))
 }}}

 with

 {{{
 empty_subst = mkEmptyTCvSubst (pprTrace "forced" (ppr (ty,args)) $
   mkInScopeSet $ tyCoVarsOfTypes [])
 }}}

 and they performed exactly the same! They were different without
 `pprTrace`. So I've tried:

 {{{
 empty_subst = mkEmptyTCvSubst (pprTrace "forced" (ppr ()) $
   mkInScopeSet $ tyCoVarsOfTypes [])
 }}}

 and it was better, back to the same amount of allocations as `baseline`.

 It appears that the difference in allocations comes from the thunks that
 get created or `empty_subst` being constant changes the generated code
 significantly. I will take a look at the core.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/11371#comment:32>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler