On Tue, Jan 05, 2010 at 04:40:55PM -0800, Bryan O'Sullivan wrote:
You've got an extra level of indirection there due to the use of data instead of newtype, so you're paying an additional boxing penalty for everything except your first case. Are you compiling with -funbox-strict-fields?
I've changed those data's to newtype's but using words still seems better, unless mapping and folding over bytes is more important. In that case maybe storing the bytes separately might be better. Maybe a more complex/realistic benchmark? I've also rerun the benchmark in a x86-32 chroot. In this environment Word32 seems to win over Word64. But, who cares about 32 bits anyway? ;) I'm attaching the source code and the summary results. Everything was compiled with 'ghc -fforce-recomp -O3'. -- Felipe.