[GHC] #16346: Lint error on master: Occurrence is GlobalId, but binding is LocalId

[GHC]

#16346: Lint error on master: Occurrence is GlobalId, but binding is LocalId
-------------------------------------+-------------------------------------
        Reporter:  simonpj           |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:
       Component:  Compiler          |              Version:  8.6.3
      Resolution:                    |             Keywords:
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by simonpj):

 Here's why the second crash happens.  We start with
 {{{
 case M.f_2dr[GblId] of f' { W# y ->
 let f_2dr[LclId] = f' in
 case M.f_2dr[GblId] of { W# z ->
 blah
 }}}
 (All this is in the body of `$s$fGBinaryGetTYPE:+:1_sfOT` in
 `Distribution.Types.ComponentRequestedSpec`.)
 Then
 * It turns out that `M.f_2dr` is a HNF so the first case can vanish.  We
 extend the substitution with `f' :-> M.f_2dr[GblId]`.
 * We simplify the RHS of the `let` to get `M.f_2dr[GblId]`
 * We simplify the let-binder.  It's not in scope, so we don't mess with
 its unique, and we extend the in-scope set with `M.f_2dr[LclId]`.
 * Now post-inline-unconditionally fires, so we extend the substitution
 with `f_2dr :-> M.f_2dr[GblId]`
 * Now at the occurrence of `M.f_2dr[GblId]` we find it in the
 substitution, mapping to, well, `M.f_2dr[GblId]`.
 * Then we look that up in the in-scope set, and get `M.f_2dr[LclId]`
 * And that is not in scope.

 What a mess.

 My conclusion: it's a total disaster to have a local binding for a
 `LocalId` that shares a unique with a `GlobalId`.
 We have multiple hacks to avoid trouble, and they don't even work.
 Let's stop doing it.  See #16296.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/16346#comment:4>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler