On 17/02/13 16:41, Nicolas Frisby wrote:

I've noticed some suspicious code in base:GHC.IO.Encoding.

setLocaleEncoding, setFileSystemEncoding, setForeignEncoding :: TextEncoding -> IO () (getLocaleEncoding, setLocaleEncoding) = mkGlobal initLocaleEncoding (getFileSystemEncoding, setFileSystemEncoding) = mkGlobal initFileSystemEncoding

(getForeignEncoding, setForeignEncoding) = mkGlobal initForeignEncoding

mkGlobal :: a -> (IO a, a -> IO ()) mkGlobal x = unsafePerformIO $ do x_ref <- newIORef x return (readIORef x_ref, writeIORef x_ref)

I've not elided any pragmas and there's no relevant LANGUAGE or OPTIONS pragmas. With HEAD from a couple weeks ago, mkGlobal is not getting inlined. But with some of my experimental alterations of sizeExpr, it is getting inlined. So this code should probably have some of the guards for faking global variables, right?

I suspect that the NOINLINE is not really needed on global variables, since GHC will never unshare something that it doesn't know the cost of, and unsafePerformIO is opaque. There should be a -fno-cse here though. (if NOINLINE really is needed, then the code will need to be rewritten to get rid of the pattern bindings, though) Cheers, Simon