GHC only “looks through” *value* bindings, because (as you note) you can get arbitrary loss of sharing otherwise.  And fint isn’t a value binding, since it has work to do.  (Not much, I grant you -- maybe we could take account of that.)

Simon

From: glasgow-haskell-users-bounces@haskell.org [mailto:glasgow-haskell-users-bounces@haskell.org] On Behalf Of Conal Elliott
Sent: 09 June 2008 16:28
To: Simon Peyton-Jones
Cc: glasgow-haskell-users@haskell.org
Subject: Re: desperately seeking RULES help

How does method sharing interact with the ability of the rules engine to "look through" lets?  Wouldn't an f rule kick in when fint is seen, by looking through the fint binding?

I've been wondering: will pattern matching look through a let even when the let-bound variable is used more than once?  I chose "yes" in Pan, though somewhat nervously, since all but one of the uses are free anyway.

Cheers,  - Conal

On Mon, Jun 9, 2008 at 2:38 AM, Simon Peyton-Jones <simonpj@microsoft.com> wrote:

The -fno-method-sharing flag was supposed to be a bit experimental, which is why it takes the cheap-and-cheerful route of being a static flag.  (Only dynamic flags can go in OPTIONS_GHC.)

What it does is this. When you call an overloaded function f :: C a => a -> a, in a function

g = ...f...f...

you normally get something like this

fint :: Int -> Int

fint = f Int dCInt

g = ...fint...fint...

That is, 'fint' extracts the 'f' method from dCInt::C Int, and it's then used repeatedly. 

With -fno-method-sharing you get

g =  ...(f Int dCInt) ... (f Int dCInt)...

So the record selection is duplicated.  It shouldn't make much difference, but of course it *does* when rules are involved, because there are no rules for fint (it's a fresh, local function). 

Simon

From: glasgow-haskell-users-bounces@haskell.org [mailto:glasgow-haskell-users-bounces@haskell.org] On Behalf Of Conal Elliott
Sent: 07 June 2008 17:26
To: glasgow-haskell-users@haskell.org
Subject: Re: desperately seeking RULES help

Is it by intention that -fno-method-sharing works only from the command line, not in an OPTIONS_GHC pragma? 

On Sat, Jun 7, 2008 at 9:23 AM, Conal Elliott <conal@conal.net> wrote:

Thanks a million, Lennart!  -fno-method-sharing was the missing piece.  - Conal

On Sat, Jun 7, 2008 at 5:07 AM, Lennart Augustsson <lennart@augustsson.net> wrote:

Here's something that actually works.  You need to pass
-fno-method-sharing on the command line.
Instead of using rules on methods it uses rules on global functions,
and these global functions don't get inlined until late (after the
rule has fired).

 -- Lennart

module F where

-- | Domain of a linear map.
class AsInt a where
 toInt'   :: a -> Int
 fromInt' :: Int -> a

{-# INLINE[1] toInt #-}
toInt :: (AsInt a) => a -> Int
toInt = toInt'

{-# INLINE[1] fromInt #-}
fromInt :: (AsInt a) => Int -> a
fromInt = fromInt'

{-# RULES
"toInt/fromInt"   forall m . toInt (fromInt m) = m
 #-}

{-# INLINE onInt #-}
onInt :: AsInt a => (Int -> Int) -> (a -> a)

onInt f x = fromInt (f (toInt x))

test :: AsInt a => (Int -> Int) -> (Int -> Int) -> (a -> a)
test h g = onInt h . onInt g

2008/6/7 Conal Elliott <conal@conal.net>:

> I'm trying to do some fusion in ghc, and I'd greatly appreciate help with
> the code below (which is simplified from fusion on linear maps).  I've tried
> every variation I can think of, and always something prevents the fusion.
>
> Help, please!  Thanks, - Conal
>
>
> {-# OPTIONS_GHC -O2 -Wall -fglasgow-exts -ddump-simpl -ddump-simpl-stats #-}
> -- {-# OPTIONS_GHC -ddump-simpl-iterations #-}
>
> module F where
>
> -- | Domain of a linear map.
> class AsInt a where
>   toInt   :: a -> Int
>   fromInt :: Int -> a
>
> {-# RULES
> "toInt/fromInt"   forall m. toInt (fromInt m) = m
>  #-}
>
> {-# INLINE onInt #-}
> onInt :: AsInt a => (Int -> Int) -> (a -> a)
> onInt f = fromInt . f . toInt
>
> test :: AsInt a => (Int -> Int) -> (Int -> Int) -> (a -> a)
> test h g = onInt h . onInt g
>
> -- The desired result:
> --
> --   test h g
> --     == onInt h . onInt g
> --     == (fromInt . h . toInt) . (fromInt . g . toInt)
> --     == \ a -> (fromInt . h . toInt) ((fromInt . g . toInt) a)
> --     == \ a -> (fromInt . h . toInt) (fromInt (g (toInt a)))
> --     == \ a -> fromInt (h (toInt (fromInt (g (toInt a)))))
> --     == \ a -> fromInt (h (g (toInt a)))
>
>
>

> _______________________________________________
> Glasgow-haskell-users mailing list
> Glasgow-haskell-users@haskell.org
> http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
>
>