[Haskell-cafe] benchmarking a function that returns an unlifted type

[David Feuer]

The critical feature of whnf (as I understand it) is that GHC does not use
the static argument transformation. The function passed to "go" therefore
doesn't inline into it. So you'll have to write your version in some
fashion that similarly prevents undesirable inlining. Could you just copy
the code and specialize it to the desired type?

On Mar 18, 2017 1:32 PM, "Richard Eisenberg" <rae at cs.brynmawr.edu> wrote:

Apologies for a newbie question, but I am seeking advice on how to
benchmark a pure function that returns an unlifted type. I've heard tell
that `criterion` is good for benchmarking, and so I'm using that library
(but have no particular allegiance to it, if something else is better).

Specifically, I'm comparing a function over unlifted types with the same
function written over lifted types, trying to observe how much we pay for
the boxing/unboxing. To benchmark the lifted version, I just use
criterion's whnf function. That seems to be working splendidly. Of course,
whnf doesn't work over unlifted types, so I found its source code and
inlined it. This compiles. But it doesn't work! No matter what I do, the
unlifted version measures in the nanoseconds. I'm guessing that GHC
"cleverly" is avoiding recomputation. How can I stop this?

Here is my code:

> main :: IO ()
> main = do
>   initializeTime
>   let lifted   = whnf fastSumPrimes 100000
>
>       unlifted = Benchmarkable go
>         where
>           go n
>             | n <= 0    = return ()
>             | otherwise = let x = fastSumPrimes# 100000# in go (n-1)
>
>   benchmark lifted
>   benchmark unlifted

I've tried increasing the argument in the unlifted case, to no avail. I've
also tried using `case` instead of `let`. -ddump-simpl suggests that
fastSumPrimes# is really being called, but I'm dubious.

Can anyone offer advice?

Thanks!
Richard
_______________________________________________
Haskell-Cafe mailing list
To (un)subscribe, modify options or view archives go to:
http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe
Only members subscribed via the mailman list are allowed to post.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/haskell-cafe/attachments/20170318/ad527d6e/attachment.html>