Thanks Don. That is the issue. I just added this comment/request to trac #1168:
This just came up on the mailing list again. The program was slow unless (-fno-state-hack) was used.
Is it possible to have GHC emit a warning (as part of -Wall) to let the use know that GHC is applying a potentially risky simplification?
I had tried -Wall before you told me about -fno-state-hack. If -Wall had warned me about the hack then I could have understood the problem without a frequently asked question to the list. Cheers, Chris Don Stewart wrote:
haskell:
I can confirm this performance bug with GHC version 6.6.1 on OS X PCC (G4) Leopard.
It performs correctly, lazily keeping 'primes' for all 4 repetitions, when using "ghc -Onot" but the the doSlow fails to retain 'primes' when compiling with "ghc -O" and "ghc -O2"
Sounds like the known issue of the simplifier eta-expanding `State#` lambdas on purpose, which is good for IO, but risks re-computation. The 'calendar' nofib program exhibits this, iirc.
See, http://hackage.haskell.org/trac/ghc/ticket/1168 http://hackage.haskell.org/trac/ghc/ticket/1957
Try compiling with:
-fno-state-hack
-- Don