The results of the GRIN experiment are written up in Urban Boquist's
PhD thesis from last year, which you can fetch from his home page:
http://www.cs.chalmers.se/~boquist/
I found the results very exciting. It's true that GRIN uses a whole-program
analysis, but in practice that turned out not to be particularly expensive
(usually no more than 2% of compile time). If I remember rightly, Urban
compiled programs up to around 10,000 lines without a problem, although the
benchmarks in the thesis are much smaller. The code generated was on average
three-and-a-half times faster than code from GHC, and in the best case 40
times faster! That extra order of magnitude seems to come from conservative
unfolding transformations which are effectively disabled in larger programs,
which raises the possibility that, with more aggressive unfolding, a further
substantial improvement in the running times of large programs might be
achievable.
I recommend the thesis heartily. I enjoyed reading it very much.
John Hughes
