On 12/15/07, Roman Leshchinskiy
While this is true in general, loop optimisations are not a particularly good example, IMO. For them to be effective, you often have to consider things like instruction scheduling and register pressure which you can't really do on the Core level.
Fair enough for loop optimizations in general, but I think the point about loop unrolling from the "Secrets of the GHC Inliner" paper that I was referring to gives a counterexample to that. In imperative languages, loop unrolling and inlining would be thought of as distinct and very different optimizations -- in a pure functional language, they can be unified. It seems to me that working in a pure functional language makes it easy to write high-level optimizations that can be specified very simply on a level like the level of Core that can then be amplified by a smart backend that takes things like instruction scheduling into account. Cheers, Tim -- Tim Chevalier * catamorphism.org * Often in error, never in doubt "Stupidity combined with arrogance and a huge ego will get you a long way." -- Chris Lowe