Troubleshooting the T4030 failure in SimonPJ's early-inline branch, Reid and I discovered that the problem showed up when the argument to forkIO was obviously bottom. He came up with a tiny test case that fails without Simon's changes:

main = forkIO undefined >> threadDelay 1000000

It looks like the trouble is in the definition of forkIO:

forkIO :: IO () -> IO ThreadId
forkIO action = IO $ \ s ->
   case (fork# action_plus s) of (# s1, tid #) -> (# s1, ThreadId tid #)
 where
  action_plus = catchException action childHandler

This seems to run into the trouble with catchException and strictness explained in GHC.IO. It would appear that the conservative fix would be to replace catchException with catch. Personally, I find it a bit surprising that forkIO doesn't force its IO argument before forking, but changing that behavior could break working code.