Neil says: I just tried compiling the following program:
foo = (1 :: Int) == (2 :: Int)
with ghc --ddump-simpl, and no optimisation flags, in GHC 6.6.1
The resultant code is:
Test.foo =
case GHC.Base.$f2 of tpl_X8 { GHC.Base.:DEq tpl1_B2 tpl2_B3 ->
tpl1_B2 (GHC.Base.I# 1) (GHC.Base.I# 2)
}
GHC has introduced dictionaries in this example. In comparison, Yhc
and nhc wouldn't introduce dictionaries here, as the type class
desugaring knows the type of the item is fixed, and makes a direct
call to the appropriate function.
If the desugaring was changed, it would make programs in GHCi run
faster, would reduce the size of programs, and would speed up the
optimiser, as it would have less to do. I am not sure how much
additional work this would require in the simplifier, but if it was
minimal, the gains would probably be worth it.
The reason nothing happens is that without -0 GHC does not load the unfoldings for anything from interface files, including GHC.Base. Idea: even without -O, load the unfoldings from GHC.* modules. That would pick up a lot of very fundamental stuff, and might, as Neil says, make compilation faster.
Relatively easy to try. But would need a good nofib run to test the effects. Any takers? I'd point you at what to change.
Simon
