[GHC] #2154: Stack overflow due to unsafePerformIO

[GHC]

#2154: Stack overflow due to unsafePerformIO
------------------------+---------------------------------------------------
    Reporter:  dons     |       Owner:                
        Type:  bug      |      Status:  new           
    Priority:  normal   |   Component:  Runtime System
     Version:  6.8.2    |    Severity:  normal        
    Keywords:           |    Testcase:                
Architecture:  Unknown  |          Os:  Unknown       
------------------------+---------------------------------------------------
 The following program, extracted from a utf8 decoder for ByteStrings,
 overflows the stack when run with rules on. Speaking to SimonM, this is
 due to the stack squeezing optimisation not working for unsafePerformIO,
 meaning the
 resulting loop actually uses up stack.

 {{{
 import qualified Data.ByteString.Char8 as S

 main = putStrLn $ filter (\x -> enc x /= [x]) list
   where list = ['\0'..'\x10ffff'] ++ ['\0'..'\10']

 {-# NOINLINE enc #-}
 enc :: Char -> String
 enc x = S.singleton x `seq` [x]

 }}}

 After fusing and optimising, S.singleton here is allocated using
 unsafePerformIO to wrap the allocation, and seems to run out of stack.

 However, I've confirmed this code does work if its changed to use
 unsafeDupablePerformIO.

 I'm filing this ticket to get documentation on what is going on here, so
 we
 can have something to point to if it comes up again, and to clarify that
 unsafeDupablePerformIO
 is the correct fix.

 SimonM, could you summarise what the issue is?

-- 
Ticket URL: <http://hackage.haskell.org/trac/ghc/ticket/2154>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler