Clark Gaebel
Oh yeah, and if anyone wants to help me figure out why it's 1.3x slower than my hand-rolled instances, that would be really helpful.
[...] I've taken a look at the bench/Bench.hs benchmark[1]: The generated Core looks almost[2] the same as your 'HandRolled'; but the 1.3x slow-down factor seems to be caused by the way the 'bigGenericRolledDS' CAF is defined in the test-harness: if I define it explicitly (i.e. just as 'bigHandRolledDS' is defined, and not as an isomorphic transformation of the 'bigHandRolledDS' value) the benchmark results in both versions having more or less equal performance as would be expected. [1]: https://github.com/wowus/hashable-generics/blob/master/bench/Bench.hs [2]: with the following change, it would look exactly the same (modulo alpha renamings): --8<---------------cut here---------------start------------->8--- --- a/bench/Bench.hs +++ b/bench/Bench.hs @@ -18,7 +18,7 @@ data GenericRolled = GR0 deriving Generic instance Hashable HandRolled where - hashWithSalt salt HR0 = hashWithSalt salt $ (Left () :: Either () ()) + hashWithSalt salt HR0 = hashWithSalt salt $ () hashWithSalt salt (HR1 mi) = hashWithSalt salt $ (Right $ Left mi :: Either () (Either (Maybe Int) ())) hashWithSalt salt (HR2 x y) = hashWithSalt salt $ (Right $ Right (x, y) :: Either () (Either () (HandRolled, HandRolled))) --8<---------------cut here---------------end--------------->8--- hth, hvr