* The procedure OS_set_working_dir_pathname in "os2env.c" had a

* The procedure OS_set_working_dir_pathname in "os2env.c" had a
  serious bug: passing it a string like "d:\\" caused it to signal a
  Scheme error.  What was serious was that this could occur while
  spawning a subprocess, which for some as-yet unexplained reason
  would crash Scheme with no error message.  To compound matters, the
  bug appeared to be unreproduceable under a debugger (not true).  The
  remaining changes were side effects of the blind debugging process:

* Added OS2_initialize_early to "boot.c" and "os2top.c".  This allows
  the microcode to do the initialization necessary to put up dialog
  boxes.  Which in turn is needed to report boot-time errors, such as
  non-existent band file names.

* "option.c" calls outf but did not include "outf.h".  Under ANSI C,
  since outf takes a variable number of arguments, this results in
  unspecified behavior.  Under OS/2 with IBM's compiler, this causes a
  mismatch in calling conventions, which can result in bizarre errors.

* Modified "os2.c" to have an alternate implementation of OS_malloc
  and related procedures.  This was generated while attempting to
  track down a nasty bug that appeared to be a memory-allocation
  error.  The alternate implementation is #ifdef'ed out, but left in
  place should it be useful in the future.

* Improved the definition of OS2_logic_error_1 so that it treats
  errors in non-essential subthreads as non-fatal.  Also improved the
  wording of the error messages produced, and included more
  information for future reference.

* All subthreads now have exception handlers and PM message queues, so
  they can catch and report any exceptions that occur in them.
  Previously, an exception occurring in a subthread would crash Scheme
  with no explanation.

* Enhanced the acknowledgement messages sent to a channel readahead
  thread to carry a flag indicating whether the thread is being asked
  to read some more or to kill itself.  This eliminates a race
  condition associated with closing of the channel.

* Changed the input-pipe readahead thread to accept its arguments in a
  freshly malloc'ed structure that the thread is responsible for
  freeing.  Previously, the arguments were accepted in a channel,
  which contained a context structure which in turn contained the
  arguments.  The previous implementation had a race condition in that
  the argument data structures were owned by the Scheme thread, and
  could be deallocated before the readahead thread had a chance to
  dereference them.

* In "os2proc.c", moved the allocation of the structure representing
  the child process into the locked region of the process
  initialization.  This is necessary because the child-wait thread
  accesses the same structures as the allocator.

* In "os2proc.c", the working directory was being saved and later
  restored.  But this was a static string, and there weren't good
  guarantees as to it retaining its value -- so now it is copied
  before being stored.

* When putting up a message box for a fatal error, the buffer used to
  compose the message was static.  This caused garbling when the two
  threads attempted to produce fatal errors at the same time (this was
  occurring regularly during the debugging process).  Now the message
  box uses a per-thread static buffer.