[GHC] #13458: Panic with unsafeCoerce and -dcore-lint

[GHC]

#13458: Panic with unsafeCoerce and -dcore-lint
-------------------------------------+-------------------------------------
        Reporter:  dfeuer            |                Owner:  goldfire
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:  8.2.1
       Component:  Compiler (Type    |              Version:  8.1
  checker)                           |             Keywords:
      Resolution:                    |  LevityPolymorphism
Operating System:  Unknown/Multiple  |         Architecture:
 Type of failure:  Compile-time      |  Unknown/Multiple
  crash or panic                     |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by goldfire):

 I'm on it. Indeed, I believe I've fixed this, but need to rebuild to be
 sure.

 The problem is the checks from #9122 that try to make sure we're not
 `unsafeCoerce`ing between types with incompatible representations. But
 when we look to see about the representations involved here, the types
 don't have representations, so GHC falls over.

 What's annoying is that the user-written program is safe in this regard.
 It's the coercion optimizer that rewrites an `unsafeCoerce` from a lifted
 type to a lifted type (`Any -> Any` to `a -> b`) to two `unsafeCoerce`s
 involving `a` and `b` separately. Thus, an alternate fix would be to have
 the coercion optimizer look for precisely this scenario and avoid pushing
 the unsafe coercion through the `->`. I haven't done this for two reasons:
  1. Failing the checks from #9122 (which I've never fully agreed with) is
 a sometimes-unavoidable consequence of a sketchy `unsafeCoerce`. The
 workaround if you're in this scenario is not to use `-dcore-lint`. The
 original program is pretty sketchy here, and so requiring that user to
 avoid `-dcore-lint` may be reasonable.
  2. It's quite annoying (and potentially computationally expensive) to do
 this check in the coercion optimizer. The check is necessary only when the
 user is willfully misbehaving, and the better citizens shouldn't have to
 pay for it every time.

 Regardless, GHC shouldn't panic.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/13458#comment:9>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler