On Tue, 8 Jan 2008, Roman Leshchinskiy wrote:
Henning Thielemann wrote:
Anyway, I tried to wrap Prelude lists in a newtype and thus got GHC (still 6.4.1) to invoke my rules instead of the Prelude rules. But I encountered the following problem: I define something like
nonFusable x y = fusable (aux x y)
where fusion rules are defined for 'fusable', but not for 'nonFusable'. I hoped that 'nonFusable' will be inlined and then 'fusable' is fused with other expressions. This does not happen. If I state the function definition also as rule, then GHC fuses eagerly.
I suspect that fusable and/or aux are inlined into nonFusable when the latter is compiled. That's too early - you want nonFusable (with the simple definition above) to be inlined into the client code first. Adding
{-# INLINE nonFusable #-}
should take care of this.
I forget to mention, that I already declared {-# INLINE nonFusable #-} As you guessed, it seems that 'fusable' was inlined and thus was no longer available for fusion. So I set INLINE for nonFusable and NOINLINE for fusable. Now I see the foldr/build counter decreases in the simplifier statistics, where the counter for my custom rules increases.
Analogously I observed that usage of ($) and (.) blocks fusion, and when I add the rules
"unfold-dollar" forall f x. f $ x = f x ;
"unfold-dot" forall f g. f . g = \x -> f (g x) ;
then fusion takes place as expected.
That shouldn't be necessary, these two ought to be inlined. Do you have a concrete example where this happens?
When I constructed a small example I encountered the behaviour you mentioned. However there are still several cases where I expected that fusion took place but it didn't. Maybe I can track them down to very simple cases. Since the functions on the chunky structure can be faster than the list functions by a factor of 30, I'm very keen on getting the list functions eliminated completely.