🚀 go-pugleaf

I recently ported my Fhourstones integer benchmark, previously
available
in both C and Java, to Haskell.

This allowed for significantly clearer and more concise code,
with much improved modularity.

Unfortunately, performance went down the drain:( On my desktop PC, the
java version
examines over 1700K positions per second, while Haskell strains to get
25 Kpos/sec.

Source of all versions may be found on
http://www.cwi.nl/~tromp/c4/fhour.html

Curiously, profiling shows all 3 versions spend between 1/5 and 1/3
of their time on the following function for testing 4 in a row in 64bit
bitboards:

  isWon :: Word64 -> Bool
  isWon bb = dir 1 /= 0 || dir height /= 0 ||
             dir height' /= 0 || dir height'' /= 0 where
    dir d = let t = bb .&. (bb `shiftR` d) in t .&. (t `shiftR` (2*d))

So this simple strict function alone appears to be dozens times slower
in Haskell.

I think I've already added strictness annotations in all the relevant
places.

Is there any hope of getting say, within a factor of 4 of java
performance,
without sacrificing (too much of) the clarity of the source code?

regards,
-John

PS: the haskell version has slightly different move ordering and so
doesn't reproduce the test results of the other versions exactly,
but this shouldn't affect the speed as measured in pos/sec.