- Implement setenv / unsetenv

- Implement other stuff missing in node.cc/process
  Like getuid, getgid, setgid, kill etc.

- Implement missing `net` methods
  Are unix sockets similar to windows named pipes? If so, should they be 
  supported? -> currently: no. Complication: they block.

- New libev backend
  The current libev backend supports sockets only. This complicates stuff like
  child processes, stdio. Best would be if node_net switched from exposing
  readyness notifications to using completion notifications, so on windows we
  could use IOCP for sockets. Experts tell me that is really the fastest way
  to work with sockets on windows.

- Child process issues
  * Communication between parent and child is slow; it uses a socketpair
    where a pipe would be much faster. Replace it by a pipe when there
    is a libev backend that supports waiting for a pipe.
  * When a child process spawns the pid is not available straightaway.
    On linux the pid is available immediately because fork() doesn't
    block; on windows a libeio thread is used to call CreateProcess.
    So this can't really be fixed, but it could be worked around by adding a
    'spawn' or 'pid' event.
  * kill() doesn't work when the pid is not available yet. All the plumbing
    is there to make it work, but lib/child_process.js just doesn't call
    ChildProcess::Kill() as long as the pid is not known.
  * passing socket custom_fds is not supported
  * child_process.exec() only works on systems with msys installed.
    It's because it relies on the 'sh' shell. The default windows shell
    is 'cmd' and it works a little differently. Maybe add an option to
    specify the shell to exec()?

- Stdio (make TTY's / repl / readline work)
  This will be hard: there is no ANSI escape code support in windows.
  Select() doesn't work on TTYs -- use a dedicated `getchar()` thread 
  that relays everything to an internal socket?
  Also verify writeError and isStdoutBlocking correctness.

- Think about exposing the platform through the process object
  It sucks but it may be necessary to know which platfom you're running on, e.g.
  you can't do spawn('grep') on windows (unless there's msys).
  Something like process.os or process.platform?

- Skip/fix tests that can never pass on windows

- Find a solution for fs.symlink / fs.lstat / fs.chown
  Windows has different symlink types: file symlinks (vista+),
  directory symlinks (vista+), junction points (xp+) 

- Handle _open_osfhandle failures
  E.g. currently we're using the construct _open_osfhandle(socket/open/accept(...)).
  Now socket() can fail by itself and _open_osfhandle can fail by itself too.
  If socket() fails it returns -1 so _open_osfhandle fails as well, but and we'll always return/throw EBADF.
  If _open_osfhandle fails but socket doesn't, a stray handle is left open. It should be fixed.

- Check error number mappings.
  Winsock errnos are sometimes different. Subtracting WSABASEERR from errnos works in most cases.

- Think about `make install`

- Extensions
  Should be DLLs on windows.

- Link pthreads-w32 statically by default

- Link Mingw libraries statically by default
  Like libstdc++.dll, more maybe.
  Microsoft libs are always there, no static linkage required (e.g. msvcrt, winsock2).

- Make (open?)SSL work

- Support using shared libs (libeio, v8, c-ares)
  Need to link with with a stub library. Libraries should use `dllexport`,
  headers must have `dllimport`.

- V8: push MINGW32 build fixes upstream
  (mostly done, V8 3.0.0 introduces some new issues)

- Work with the V8 team to get the stack corruption bug fixed
  (fixed in 3.0.0)

- Work around missing pread/pwrite more elegantly
  Currently it's exported from libeio, while it wasn't intended to be exported.
  The libeio workaround implementation sucks, it uses a global mutex.
  It should be possible to implement pread and pwrite using winapi's ReadFile/Writefile
  directly, passing an OVERLAPPED structure while not associating with an completion port.

- Work around missing inet_pton/inet_ntop more elegantly
  Currently it's exported from from c-ares, while it wasn't intended to be exported.
  It prevents linking c-ares dynamically. 

- See what libev/libeio changes can be pushed upstream

- 64-bit build
  Should be possible with MinGW-w64, it's pretty good.

- ... much more probably