Native Code gen miscompiles double2Int# / float2Int# on x86_32

Brad Clow: I upgraded from GHC 6.6.1 to 6.8.1 and around that time I noticed that the output from an app I am working on changed. I have distilled the code down to the following example that produces different output depending on whether it is compiled with -O2 or not:

main = do
    let (T x) = read "T 3"
    print $ f x
    print $ g x

data T = T Int deriving (Read, Show)

f x = truncate $ 2000 / fromIntegral ((x * 25) + 10)

g :: Int -> Int
g x = f x

(See http://www.haskell.org/pipermail/haskell-cafe/2007-November/034930.html)

The reason is that the code generated for double2Int# by the native code generator does not round correctly. (It's not 100% predictable, I believe, but it'll usually round.)

Trac metadata

Trac field	Value
Version	6.8.1
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system	Unknown
Architecture