now I understand. I've created a Trac ticket. Shouldn't be hard.
http://hackage.haskell.org/trac/ghc/ticket/4201
Thanks
Simon
From: [email protected]
[mailto:[email protected]] On Behalf Of Louis Wasserman
Sent: 13 July 2010 17:46
To: Simon Peyton-Jones
Cc: [email protected]
Subject: Re: Casting + eta reduction
Or a different way:
I want -fdo-lambda-eta-expansion (which, if I understand correctly, actually
triggers eta *reduction*) to eliminate argument casts, as well.
My motivation: I'm working on a generalized trie library, and due to
http://hackage.haskell.org/trac/ghc/ticket/4185, I can't use
GeneralizedNewtypeDeriving to minimize the overhead of a stack of about 20
newtypes, each with their own class instance, even if I wasn't trying to use
Template Haskell, which currently has no syntax for doing
GeneralizedNewtypeDeriving on multi-parameter type classes (it only derives
single-argument type classes.) As it stands, I have a lot of methods that
compile to
foo f = bar (\ x -> f (x `cast` a))
and they stack 20 deep, which means I have to do 20 allocations (\ x -> f (x
`cast` a)) for even the most simple methods. What I'd like to see is this
getting reduced to
foo = bar `cast` (...)
which would reduce my overhead significantly.
Louis Wasserman
[email protected]<mailto:[email protected]>
http://profiles.google.com/wasserman.louis
On Tue, Jul 13, 2010 at 10:06 AM, Louis Wasserman
<[email protected]<mailto:[email protected]>> wrote:
Mmmm, let's give a slightly different example:
foo :: Foo -> Int
foo (Foo a) = a + 1
bar :: Int -> Int
bar = foo . Foo
and I'd expect bar to be replaced with (foo `cast` (Int -> Int)) and inlined,
eliminating an allocation. In general, we'd get the equivalent of the
no-allocation versions of GeneralizedNewtypeDeriving instances, so long as we
could write them out for ourselves.
Louis Wasserman
[email protected]<mailto:[email protected]>
http://profiles.google.com/wasserman.louis
On Tue, Jul 13, 2010 at 9:09 AM, Simon Peyton-Jones
<[email protected]<mailto:[email protected]>> wrote:
It compiles to
lift f d = f (d `cast` blah)
which seems fine to me. Are you unhappy with that?
Simon
From:
[email protected]<mailto:[email protected]>
[mailto:[email protected]<mailto:[email protected]>]
On Behalf Of Louis Wasserman
Sent: 09 July 2010 03:30
To: [email protected]<mailto:[email protected]>
Subject: Casting + eta reduction
Consider
newtype Foo = Foo Int
lift :: (Int -> a) -> Foo -> a
lift f (Foo x) = f x
Now, I'd expect this to compile with -O2 down to something like
lift f = f `cast` (Foo -> a)
but it doesn't.
It seems that GeneralizedNewtypeDeriving assumes that these two things *are*
equivalent, and it just directly casts the class dictionary. The implication
would be that that GeneralizedNewtypeDeriving gives more efficient instances
than you could *possibly* get if you wrote them by hand, which is very sad.
Louis Wasserman
[email protected]<mailto:[email protected]>
http://profiles.google.com/wasserman.louis
_______________________________________________
Glasgow-haskell-users mailing list
[email protected]
http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
