Bryan O'Sullivan wrote:
In my profile results, I find that simply converting words to lower case accounts for a whopping 40% of time and allocation (see the attachment for my definition of the train function).
COST CENTRE MODULE %time %alloc
lower Spell 40.5 41.2 train Spell 26.3 14.3 mkWords Spell 21.9 24.1
A little more instrumentation says this (using the darcs head of fps
built with -auto-all):
loopU NewData.ByteString.Fusion 25.4 28.8
splitWith NewData.ByteString 15.4 17.2
train Spell 10.2 6.1
isNotAlpha Spell 9.4 12.2
compareBytes NewData.ByteString 8.8 9.6
compareBytes NewData.ByteString.Lazy 7.4 0.4
inlinePerformIO NewData.ByteString.Base 6.6 0.0
(At Stefan's suggestion, I renamed the modules in fps to NewData.*, in
order to get around the name clashes when I try to concurrently use fps
and base.)