On 11/12/10 4:53 PM, Ryan Ingram wrote:
From http://www.haskell.org/ghc/docs/6.12.2/html/users_guide/bugs.html#bugs-ghc:
GHC's inliner can be persuaded into non-termination using the standard
way to encode recursion via a data type:
More specifically, since your bad2 does not look recursive, the inliner
will get inline happy and apply the definition of bad1, but then it'll
see a constant case match that it can reduce statically, and <loop>
{ okay, let's start with bad1 }
bad1 b@(C f) = f b
{ mmm, tasty tasty sugar }
bad1 = \b -> case b of C f -> f b
{ ha, dare you to optimize that further }
{ okay, now let's co compile bad2 }
bad2 = bad1 (C bad1)
{ oh look, it's nonrecursive so let's start inlining }
bad2 = (\b -> case b of C f -> f b) (C bad1)
{ oh look, an application we can evaluate statically }
bad2 = let b = (C bad1) in case b of C f -> f b
{ meh, let's ignore sharing for now }
bad2 = case (C bad1) of C f -> f (C bad1)
{ oh look, a case analysis we can evaluate statically }
bad2 = let f = bad1 in f (C bad1)
{ oh look, f is only used once so we can substitute the let }
{ and if we didn't ignore sharing before, then we'd see that b is only
used once now, so we could substitute that let too }
bad2 = bad1 (C bad1)
{ oh look, it's nonrecursive so let's start inlining }
--
Live well,
~wren
_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe@haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
