[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):

 I've done some profiling today and I've optimized `piResultTys` a bit.

 Here are the results (this is built with prof):
 {{{
 Benchmark name          baseline        with-inscope    without-inscope
 tests/alloc/T5030       1084862416      1198020320      1084861344
 tests/alloc/T9872a      5426538936      5576679896      5426504344
 tests/alloc/T9872b      7807323280      8022344352      7728326496
 tests/alloc/T9872c      7113767856      7405191768      7113729216
 }}}

 `baseline` is the current incorrect implementation that doesn't compute
 the in-scope sets.

 `with-inscope` is https://phabricator.haskell.org/P97 - an implementation
 where the `coreView` that was hidden in `splitPiTy_maybe` and in
 `getTyVar_maybe` is done only once and both are inlined. This is the hint
 that I got from profiling, there was a significant difference in bytes
 allocated by `splitPiTy_maybe`.

 `without-inscope` is https://phabricator.haskell.org/P95 - a version of
 `with-inscope` that's doesn't compute the in-scope sets.

 So `without-inscope` is a strict, but small improvement over `baseline`
 performance wise, but the cost of computing the `in-scope` sets is non-
 trivial, so the correct version has a performance penalty.

 Are we willing to accept it in the name of correctness? The correct
 version has to do more work.


 The hot path is `cmpTypeX -> typeKind -> piResultTys`, I believe
 `cmpTypeX` already has the relevant free vars in the `env`, so passing
 them down is what I'll try to do next.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/11371#comment:30>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler