Which then fails the consistentStgInfo check in CoreToStg for reasons
#11158: Combine exprIsTrivial and cpe_ExprIsTrivial -------------------------------------+------------------------------------- Reporter: simonpj | Owner: Type: task | Status: closed Priority: highest | Milestone: 8.2.1 Component: Compiler | Version: 7.10.2 Resolution: fixed | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Phab:D2666 Wiki Page: | -------------------------------------+------------------------------------- Comment (by simonpj): that I am still trying to work out. Here's why (I think): * As the commit message says, if we case-bind the literal string, something that was a static data structure (predicted as `NoCafRefs` by `CorePrep` becomes a thunk. This prediction is made by `rhsIsStatic`, called in `CorePrep`. * Since it is marked as `NoCafRefs` it won't appear in anyone's SRT * But when it is evaluated, the thunk will allocate its result in the heap, and update the static closure to point to it. * Later the garbage collector will collect that thunk; the only pointer to it is from a static closure that is in no SRT. * The space will be re-used * Later someone uses the thing again, but the pointer now points to garbage. Seg fault. It's FATAL for something marked as `NoCafRefs` to turn into a thunk. Arrangig to panic on that would be good. It should not be just a warning, and I think it'd be worth the tiny perf hit to test it even in non-DEBUG code. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/11158#comment:37 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler