#11254: GHC panic ---------------------------------+---------------------------------------- Reporter: Iceland_jack | Owner: goldfire Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.11 Resolution: | Keywords: Operating System: Linux | Architecture: Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | ---------------------------------+---------------------------------------- Comment (by goldfire): From TcErrors: {{{ Note [Deferred errors for coercion holes] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Suppose we need to defer a type error where the destination for the evidence is a coercion hole. We can't just put the error in the hole, because we can't make an erroneous coercion. (Remember that coercions are erased for runtime.) Instead, we invent a new EvVar, bind it to an error and then make a coercion from that EvVar, filling the hole with that coercion. Because coercions' types are unlifted, the error is guaranteed to be hit before we get to the coercion. }}} But that's admittedly not the whole story. This also matters, from `TcUnify.buildImplication`: {{{ -- But with the solver producing unlifted equalities, we need -- to have an EvBindsVar for them when they might be deferred to -- runtime. Otherwise, they end up as top-level unlifted bindings, -- which are verboten. See also Note [Deferred errors for coercion holes] -- in TcErrors. }}} The comment describes a check to make sure that we always build an implication when `-fdefer-type-errors` is on (and we're at top-level). I noticed an unexpected `pushTcLevelM` at !TcInstDecls:811 that may throw this check off. Do you have a recommendation for how to improve the documentation? I'll look into this. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/11254#comment:2 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler