[Haskell-cafe] Type constraint with RankNTypes gradually slows down function?

[Alexander Eyers-Taylor]

Hello

I have no idea why this happens but GHC eta expands before passing it to 
the recursive case. i.e.

run1 f state

de-sugars to

run1 (\a dictNum -> f a dictNum) state

So it allocates a new closure each time. The time increases as each time 
f is called we must descend through the stack of closure. As to why this 
happens I have no idea but as it is there after de-sugaring I presume it 
has something to so with passing dictionaries.

Alex

On 15/06/15 23:38, Ben Gunton wrote:
> When I write a function with the type constraint "embedded" in the 
> function's type declaration, instead of at the beginning, the function 
> takes longer each time it is run. The simplest example I could write 
> is this: http://lpaste.net/134563. "run1" is near instantaneous the 
> first time, but gets marginally slower after each iteration... after 
> thousands of iterations it takes about a millisecond to run (and keeps 
> getting worse). "run2" is always near instantaneous. What is happening 
> in "run1" that makes it slow down?
>
> PS: The example is fairly contrived, I know "run1"'s type declaration 
> wouldn't be used in practice.
>
> Thanks for any guidance!
>
> {-# LANGUAGE RankNTypes #-}
> {-# LANGUAGE FlexibleContexts #-}
>
> module Main where
>
> import Control.Monad.State.Strict
> import qualified Data.Time.Clock as Clock
> import Control.Exception
>
> run1 :: (Int -> (Num Int => State Int Bool)) -> Int -> IO ()
> run1 f state = do
>     t1 <- Clock.getCurrentTime
>     evaluate $ runState (f 1) state
>     t2 <- Clock.getCurrentTime
>     print $ Clock.diffUTCTime t2 t1
>     run1 f state
>
> run2 :: Num s => (Int -> State s Bool) -> s -> IO ()
> run2 f state = do
>     t1 <- Clock.getCurrentTime
>     evaluate $ runState (f 1) state
>     t2 <- Clock.getCurrentTime
>     print $ Clock.diffUTCTime t2 t1
>     run2 f state
>
> main :: IO ()
> main = run1 (const $ return False) 1
> --main = run2 (const $ return False) 1
>
>
>
> _______________________________________________
> Haskell-Cafe mailing list
> Haskell-Cafe at haskell.org
> http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/haskell-cafe/attachments/20150616/db8665a5/attachment.html>