[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 simonpj):

 This looks terribly suspicious (in `TyCoRep`):
 {{{
 extendSubstEnvs :: (TvSubstEnv, CvSubstEnv) -> Var -> Type
                 -> (TvSubstEnv, CvSubstEnv)
 extendSubstEnvs (tenv, cenv) v ty
   | isTyVar v
   = ASSERT( not $ isCoercionTy ty )
     (extendVarEnv tenv v ty, cenv)

     -- NB: v might *not* be a proper covar, because it might be lifted.
     -- This happens in tcCoercionToCoercion
   | CoercionTy co <- ty
   = (tenv, extendVarEnv cenv v co)
   | otherwise
   = pprPanic "extendSubstEnvs" (ppr v <+> text "|->" <+> ppr ty)
 }}}

 I think that the `Var` passed to `extendSubsEnv` should be a proper
 `TyVar` or a `CoVar`. Moreover `tcCoercionToCoercion` doesn’t exist.   Is
 the comment now out of date?  I hope so.

 Moreover, it seems suspicious to pattern match on `CoercionTy`.  What if
 there was a type synonym in the way?  Or a cast?


 HOWEVER, we do need to get the coercion out to put into `cenv`.

 But that makes me suspicious too.  I think that almost all calls to
 `extendSubstEnvs` provide only `TyVars`, not `CoVars`.  The only call I
 can find that doesn’t is in `TcMType.instSkolTyCoVarX`, which tests the
 variable, and makes a `CoercionTy` – only for `extendTCvSubst` to unpack
 it again.
 Concretely, I suggest that

 * we rename `extendTCvSubst` to be `extendTvSubst`, and work only over
 `TyVars`.  (Have an assertion of course.)
 * have a similar one for `extendCvSubst`
 and see how that goes.  It’s be in line with the new `zipTvSubst`,
 `zipCvSubst` etc.

 Richard, can you see anything wrong with that?
 Bartosz, might you execute?

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/11371#comment:16>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler