a compiled program is slower than byte code

[Simon Marlow]

On 09/02/2011 04:35, Kazu Yamamoto (山本和彦) wrote:
> Hello Simon,
>
>> $ ghc-nightly2 ./kazu.hs -O2 -fforce-recomp; time ./kazu
>> [1 of 1] Compiling Main             ( kazu.hs, kazu.o )
>> Linking kazu ...
>> 4.17s real   4.16s user   0.01s system   99% ./kazu
>
> OK. I ran it on 32bit Linux.
>
> 6.12.3  runghc   -- 2.22s user 0.40s system 96% cpu 2.724 total
>          ghc      -- 1.96s user 0.14s system 97% cpu 2.151 total
>          ghc -O   -- 2.18s user 0.10s system 97% cpu 2.333 total
>          ghc -O2  -- 2.27s user 0.07s system 97% cpu 2.393 total
>
> ghc-7.0 runghc   -- 3.43s user 0.35s system 97% cpu 3.861 total
>          ghc      -- 5.11s user 0.07s system 97% cpu 5.299 total
>          ghc -O   -- 5.38s user 0.03s system 97% cpu 5.534 total
>          ghc -O2  -- 5.54s user 0.10s system 97% cpu 5.783 total

Thanks, I do see the problem now (I was inadvertently running the 
compiled code with runghc, because I used "runghc foo.hs 
-fforce-recomp", which doesn't do what I thought it did).

I created a ticket: http://hackage.haskell.org/trac/ghc/ticket/4951

I think the difference between interpreted and compiled is due to
http://hackage.haskell.org/trac/ghc/ticket/917.  What I'm more worried 
about is the slowdown in pre-7.0.2 relative to other versions.


>> HEAD is more than twice as fast on this program.  Fusion not working,
>> perhaps?
>
> Here is the results of "ghc -O -ddump-simpl-stats":
>
> ghc-7.0:
>
> 19 RuleFired
>      6 ++
>      2<=#
>      1 fold/build
>      5 foldr/app
>      1 foldr/augment
>      1 foldr/single
>      1 map
>      1 unpack
>      1 unpack-list
>
> 6.12.3:
>
> 25 RuleFired
>      6 ++
>      2<=#
>      3 fold/build
>      6 foldr/app
>      1 foldr/augment
>      1 map
>      2 repeat
>      2 take
>      1 unpack
>      1 unpack-list

My stats look very different.

6 RuleFired
     1 ++
     2 <=#
     1 foldr/app
     1 unpack
     1 unpack-list

Are your libraries compiled with -O2?

Cheers,
	Simon