A bug of linebuffering

Kazu Yamamoto ( 山本和彦 ) kazu at iij.ad.jp
Mon Sep 9 12:25:15 UTC 2013


While I support GHC head for "doctest", I encountered the following

"doctest" uses a GHCi subprocess to evaluate an expression represented
in String. Stderr from GHCi is merged into stdout by hDuplicateTo in
the GHCi side. Even evaluating an error expression, for instance "1
`div` 0", the line buffering does not work. "doctest" waits for output
from GHCi forever. This does not happen if stderr is not merged into

The following code demonstrates this bug. Running it with GHC head
waits forever. Running it with GHC 7.6.3 prints:
"*** Exception: divide by zero"

If you change "1 `div` 0" into "1 `div` 0\nprint 10", this code run by
GHC head prints:
"*** Exception: divide by zero"

This is a serious behavior change for "doctest". I hope this will be


module Main where

import System.Process
import System.IO

myProc :: CreateProcess
myProc = (proc "ghc" ["-v0", "--interactive", "-ignore-dot-ghci"]) {
            std_in  = CreatePipe
          , std_out = CreatePipe
	      , std_err = Inherit

setMode :: Handle -> IO ()
setMode hdl = do
    hSetBinaryMode hdl False
    hSetBuffering hdl LineBuffering

newInterpreter :: IO (Handle, Handle)
newInterpreter = do
    (Just stdin_, Just stdout_, _, _) <- createProcess myProc
    setMode stdin_
    setMode stdout_
    hPutStrLn stdin_ "import System.IO"
    hPutStrLn stdin_ "import GHC.IO.Handle"
    hPutStrLn stdin_ "hDuplicateTo stdout stderr"
    hFlush stdin_
    return (stdin_, stdout_)

eval :: Handle -> Handle -> String -> IO String
eval hin hout expr = do
    hPutStrLn hin expr
    hFlush hin
    hGetLine hout

main :: IO ()
main = do
    (stdin_, stdout_) <- newInterpreter
    eval stdin_ stdout_ "1 `div` 0" >>= print
    eval stdin_ stdout_ "1 + 2" >>= print
    return ()

More information about the ghc-devs mailing list