Luke Palmer wrote:
In attempting to devise a variant of cycle which did not keep its argument alive (for the purpose of cycle [1::Int..]), I came across this peculiar behavior:
import Debug.Trace
cycle' :: (a -> [b]) -> [b] cycle' xs = xs undefined ++ cycle' xs
take 20 $ cycle' (const $ 1:2:3:4:trace "x" 5:[]) [1,2,3,4,x 5,1,2,3,4,5,1,2,3,4,5,1,2,3,4,5]
Nuts. Oh, but wait:
take 20 $ cycle' (\_ -> 1:2:3:4:trace "x" 5:[]) [1,2,3,4,x 5,1,2,3,4,x 5,1,2,3,4,x 5,1,2,3,4,x 5]
Hey, it worked!
Can someone explain what the heck is going on here?
Luke
(\_ -> 1:2:3:4:trace "x" 5:[]) literally could mean your second program, but... the 1:2:3:4:trace "x" 5:[] does not depend on the _ argument, and so it can be lifted outside the (\_ -> ... ) and lazily evaluated once and shared between calls. Optimization in ghc do this for you. The definition "const x = (\_ -> x)" binds 'x' outside of the _ argument, so 'x' is obviously outside (\_ -> ...) and will be lazily evaluated once and shared. I see that making the binding and sharing explicit in
take 20 $ cycle' (let x = 1:2:3:4:trace "x" 5:[] in (\_ -> x)) [1,2,3,4,x 5,1,2,3,4,5,1,2,3,4,5,1,2,3,4,5]
behaves like const. And pushing the binding inside the (\_ -> ...)
take 20 $ cycle' (\_ -> let x = 1:2:3:4:trace "x" 5:[] in x) [1,2,3,4,x 5,1,2,3,4,x 5,1,2,3,4,x 5,1,2,3,4,x 5]
behaves like your second example. -- Chris