I have the following Haskell program: ```haskell {-# LANGUAGE MonoLocalBinds #-} module Main (main) where class C a where op :: a -> String instance {-# OVERLAPPABLE #-} C a where op _ = "Generic" {-# NOINLINE op #-} instance {-# INCOHERENT #-} C () where op _ = "Specific" {-# NOINLINE op #-} {-# NOINLINE large #-} -- | Inhibit inlining, but keep specialize-ability large :: a -> a large x = x bar :: C a => a -> String bar x = large (large (large (large (large (large (large (large (large (large (large (large (large (large (op x)))))))))))))) spec :: () -> String spec = bar gen :: a -> String gen = bar main :: IO () main = do putStrLn $ "spec () == " <> spec () putStrLn $ "gen () == " <> gen () ``` Because `gen` has no typeclass constraint on `C a`, the only way to resolve the constraint on `bar` is to choose the generic instance. On the other hand, in `spec`, we have a more specific instance for `C ()`, and so that one should be picked. Compiling and running without optimizations, this is indeed the behaviour one can observe: ``` $ ./_build/stage1/bin/ghc -fforce-recomp --make input/overlap3.hs && ./input/overlap3 [1 of 2] Compiling Main ( input/overlap3.hs, input/overlap3.o ) [2 of 2] Linking input/overlap3 [Objects changed] spec () == Specific gen () == Generic ``` However, if we then enable specialisation, something goes wrong. How exactly, depends on the version of GHC. On and before 8d2dbe2db4cc7c8b6d39b1ea64b0508304a3273c, we get: ``` $ ./_build/stage1/bin/ghc -fforce-recomp -fspecialise -fenable-rewrite-rules --make input/overlap3.hs && ./input/overlap3 [1 of 2] Compiling Main ( input/overlap3.hs, input/overlap3.o ) [2 of 2] Linking input/overlap3 [Objects changed] spec () == Generic gen () == Generic ``` But from 5a997e16cb2079c52f980d59af3de176922fa320 onwards, we get the, perhaps even more baffling: ``` $ ./_build/stage1/bin/ghc -fforce-recomp -fspecialise -fenable-rewrite-rules --make input/overlap3.hs && ./input/overlap3 [1 of 2] Compiling Main ( input/overlap3.hs, input/overlap3.o ) [2 of 2] Linking input/overlap3 [Objects changed] spec () == Specific gen () == Specific ``` The change between these two commits is an overhaul of the specialiser implementation by @simonpj , so it's at least understandable that the behaviour changed at this point.