Functionality
#############

- Kill (deadline) timeouts if a thread is terminated.
- Fix Receiver::sender_enqueue (see ktest: "unfair scheduling").
- Long IPC: Reserve first 4MB page after User_max as ``not available'' to
  allow user to use the last 4MB page for long IPC (keyword IPC window)
- Lazy IPC windowing for long IPC (similar to lazy FPU). The kernel should
  store the IPC owner per address space (e.g. the last thread that called
  setup_ipc_window()). When switching to a thread which performs long IPC
  only flush the IPC window if it doesn't own the IPC window.
  * Udo: It is also possible to setup the IPC window for the other thread
         (saves pagefaults) but don't do this in switch_to()
- Compare version numbers in IPC (and not only there).
- Add receiver lists (similar to sender lists). A thread A stays in closed
  wait to thread B. The IPC is not aborted if thread B is killed.
  * Udo: Receiver lists are probably also needed for time donation.
- Enhance the mapping database to support I/O flexpages for x86.
- Enhance the mapping database to support Memory-Mapped flexpages.
- Add scatter/gather as defined in the X.0 specification.
- Replace fixed addresses of TCBs with e.g. a slab allocator and a hash table
  for mapping Threadid->TCB (even for V2?)
- IPC via chiefs. Does that still make sense?
- The current profiling implementation produces the old BSD-GPROF4 format
  which don't includes a magic number and the sample frequency. Either use
  the 4.4BSD format (which isn't recognized by the binutils by default) or
  use the GNU format (take the libc6 sources).
- Vkey IRQ support for Fiasco-UX.
- UTCBs for all architectures.


Code cosmetics
##############

- Use the minilibc as GlibC replacement for Fiasco-UX. This would reduce the
  stack footprint of printf() & colleagues and would allow to use these
  functions from kernel thread contexts without switching the stack.
- Remove all 32 bit values of -1 (0xffffffff). Replace them by a constant
  defined as static_cast<Mword>(-1).
- Remove multiple definitions of offsetof macro.
- Unify long IPC pathes of all architectures.
- More unification between the kernel debugger of IA32 and UX.
- Fiasco-UX platform/subarchitecture separation.
- Class Kmem should be related to Space. Kmem should be a ``real'' object.
- ``Signal'' should be renamed to ``transactional message''.
- Rename object variables ``my_foo'' to ``_foo''.
- Simplify the IPC path by reducing the number of possible states.
- Unify the Thread stuff in ARM and IA32 V4.


Performance
###########

- The most important members of the TCB should have an offset less than 0x80
  to achive smaller code (can save up to 3 bytes per instruction).
- TCB cache line optimization.
- Selective assertions, differ between ``normal'' assertions and ``expensive
  assertions''.
  * Frank: Assertions are quite cheap because they are marked as EXPECT_FALSE
           (at least with gcc-3).

