On 12 March 2010 13:13, Roman Beslik
Thanks for the answer. Sorry, I can not follow all of your thoughts because my knowledge of strictness analysis and GHC optimizations are very basic. :( I looked into GHC code once several years ago. BTW there are a lot of papers about strictness analysis, but I do not know which is relevant for GHC. Can you suggest something?
There is nothing *published* (Simon has a half-written one lying around though), but the general approach is similar to that shown in "Projections for strictness analysis" at http://homepages.inf.ed.ac.uk/wadler/topics/strictness-analysis.html. Unfortunately the weird behaviour you are seeing is due to the "demand transformer" technology of GHC, which is one of the unpublished bits...
So GHC records strictness information for lambda-abstractions, not for function types? Eta-expansion changes strictness because it adds lambda-abstractions.
It records strictness info on *binders*, and it only records strictness info about lambdas that are syntactically manifest at the binder. So you get: let f = \z. bar e_1 g = foo e_2 e_3 in e_3 (\y. baz e_4) Then f gets strictness info about one argument, g about no arguments and the (\y. baz e_4) just doesn't stand a chance of getting improved at all. Eta expansion moves lambdas towards the binders, so it makes this approximation more effective, as you say.
2) GHC does not seem to be eta-expanding as much as it could get away with. Generally eta expansion has the following effects
I think it is better to go without eta-expansion. Eta-expansion (if not optimized when compiled) do absolutely useless job. I discovered its effect by an accident.
Eta-expansion happens quite a bit in GHC right now, though I'm not sure how important it is pragmatically. Probably quite important though - you might want to look up the "state hack" for a situation where it seems to be quite necessary. Sorry that I don't have an easy answer to your problem except "eta-expand by hand"! Cheers, Max