Re: Optimizing Haskell programs is hard

[Simon L Peyton Jones]

You are right to be "bugged" (see your last para).  Fortunately,
ghc 2.09 (i.e. our current working copy) gives identical runtimes for
all three. 

(Might be true of 2.08; I haven't tried.)

Simon

> I'm benchmarking MVar's and other shared memory abstractions, e.g. by
> accessing a variable a number of times:
> 
> 
> main :: IO ()
> main = do {
>       v <- newMVar 0;
>       access v 10000000
> } where access v 0 = print "done"
> -- 1) access v n = do {swapMVar v n; access v (n-1)}
> -- 2) access v n = do {takeMVar v; putMVar v n; access v (n-1)}
> -- 3) access v n = do {setVar v n; access v (n-1)}
> 
> I'm accessing the MVar's according to 3 different schemes: one using swapMVar
> (1), the other using takeMVar and putMVar (2) and the third by calling a method of 
>class 
> 
>       Variable var where
>               setVar :: Var a -> a -> IO ()
>               
>       instance Variable MVar where
>               setVar mvar val = do {takeMVar mvar; putMVar mvar val}
> 
> These are the results that I get when running the programs:
> 
>       1) 42 secs
>       2) 16 secs
>       3) 32 secs
> 
> It's quite surprising to see how much time swapMVar costs: I actually
> expected it to be faster than a take followed by a put. I have now
> spend some time getting rid of swapMVar wherever performance matters.
> 
> One thing bugs me though: what is causing the overhead when I call the
> class operation, and - how can I get around it! This is quite
> important to the efficency of some of my basic abstractions such as
> reentrant monitors and the like.