[Haskell-cafe] Wrestling the inliner beast
bos at serpentine.com
Thu Nov 12 02:44:59 EST 2009
I did a little bit of Data.Text benchmarking the other day, and I was
shocked to find that decoding a UTF-8 stream proceeded at a sedate 3MB/sec.
On investigation, the culprit was that I'd marked both the outer function
and the inner (a loop) as INLINE. Because the inner loop was marked as
INLINE, GHC wasn't inlining critically important leaf functions into it, so
I was getting clobbered to death by boxing and unboxing. Leaving the outer
function marked as INLINE, but taking the INLINE off the inner function,
seems to cause *both* to get inlined as I originally hoped. This behaviour
is all rather mysterious to me. The old "Secrets of the Inliner" paper is
very much out of date now, but short of reading the source, I don't know
where else to look to turn my voodoo folk intuition into something more
solid. Is this all going to change in 6.12 anyway?
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Haskell-Cafe