Since GCC 3.0 I have decided to switch to the mainstream version of GCC rather
than supporting PGCC further. The reasons:

 - Since EGCS project started PGCC become an rather experimental branch of
   gcc than a separate gcc version.

 - Most of the features originally available only in pgcc are now available
   as well in the mainstream gcc.

The following small changes (apart from those in ChangeLog's) have been made
relatively to the original GCC+EMX 2.7.2.1 (unordered):

o PGCC has NO bound checking.

o The symbol __OS2__ is defined if -Zomf is used.

o I've removed some of Eberhard Mattes' patches that were related to 8.3
  naming convention. FAT is ugly, besides you mostly won't have any problems
  with 8.3 names even without these patches. They were related mostly to
  GCC-developer related things, like .cse2, .flow, .combine, .sched and
  other file name extensions. If you need these files, you should use
  a normal filesystem :-)

o The -ZC++-comments switch is gone now (hopefully forever) since GCC
  always enable C++-style comments (unless -traditional or -lang-c89 is
  specified)

o The -Zexe switch now works correctly (before it worked only in couple
  with -o switch). ld and emxomfld were patched too for this. Also now
  the -Zexe option not only creates a fresh file without extension
  to keep "make" happy; the behaviour has been changed to copy a file
  called ldstub.bin that should be located in same directory where
  ld or emxomfld is into the file with same name as executable but without
  extension: this avoids two problems at once: first, many "configure"
  scripts expects after compiling a test program to obtain a *non-empty*
  file; and second, they expect the file to be runnable, otherwise C compiler
  is considered a cross-compiler. ldstub determines the name under which
  it has been launched and launches in turn the executable that has same
  name but with ".exe" appended. So, for example, if you copy ldstub.bin
  into "sh" then when you do from, say, bash:

  [c:/] ./sh

  you will launch "sh" which in turn will look for "sh.exe" (located in
  same directory as "sh" itself) and will run it.

o All compilers were compiled with -Zcrtdll, so you will need an
  installed EMX runtime to run them. Anyway, I think everybody reading
  this already have it :-)

o The -mprobe switch is retained for backward compatibility, however it
  is highly recommended that you use -mstack-arg-probe and -mno-stack-arg-probe
  instead since these options are new from mainstream GCC version, and maybe
  someday -mprobe will disappear.
  Another switch that has to do with stack probing is -f{no-}stack-check.
  These switches have same functionality as -m{no-}stack-arg-probe but uses
  inline instructions instead of alloca call, so they're faster. Here is how
  alloca(20000) is compiled with different switches:

  without any switches:

	subl $20000,%esp

  with -mstack-arg-probe:

	movl $20000,%eax
	call __alloca

  with -fstack-check options:

	movl $0,-4392(%esp)
	movl $0,-8488(%esp)
	movl $0,-12584(%esp)
	movl $0,-16680(%esp)
	movl $0,-24392(%esp)
	addl $-20000,%esp

o Now alloca() is stack-probe safe, i.e. if you`re compiling with
  -mstack-arg-probe switch, alloca() will do stack probes for you.
  However, the previous limitation still applies, i.e. if you will do
  alloca(4095) twice, this can lead to a crash. The best workaround for this
  is to not use threads with non-commited stack at all :-)

o The provided binaries of GCC will not work in DOS (with EMX or RSX extenders) since<BR>
  &gt; They were packed with LXLITE (no OS/2 2.x too)<BR>
  If you need either to run GCC in DOS (note that you still can <B><I>produce</I></B>
  executables that run in DOS), you can try to recompile it yourself.
  If you want to run it under OS/2 2.x you should unpack executables
  (<CODE>lxLite -x -r *</CODE>). If you can't do the listed above yourself,
  mail me - possibly if I'll receive enough requests, I'll do a unpacked version.
  But personally me for DOS am using DJGPP - it is better suited for DOS.

o 64-bit ([unsigned] long long) division is now FAST! :-) Its written
  in assembly now, and I`ve optimized it by hand. It looks almost ~ten
  times faster as before. Signed division is somewhat slower than unsigned, so
  if you need speed, use unsigned when possible.
  I should say that same routine, coded in plain C (take a look into libgcc2.c)
  and compiled with PGCC -O6 is only 10-20% slower (!!!). This makes me think
  this routine was the last in my life which I've assembled manually ;-)

o -mepilogue did not worked in C++ and possibly other languages which mangles
  names. Fixed.

o _set_new_handler now is called set_new_handler. However, _set_new_handler
  will work too as an alias.

o PGCC has command-line help. Try:

      gcc --help
      gcc --help=*
      gcc --help=-c
      gcc -h-m
      gcc -h13.15

  There is a new optional environment variable: CLH_TERM. If it is "mono",
  the help will use only black/white colors; if it is "-", the help won't
  use highlighting at all; if it is anything other (or is not set) the help
  text will default to colored syntax.

o Libgcc has been put in a separate DLL. I did this since libgcc changes
  often - with each new major gcc release, sometimes even in minor releases.
  To avoid conflicts with future/previous libgcc versions, libgcc will have
  a name consisting of 'gcc' and gcc's internal version number without dots.
  For example, gcc version 2.90.27 will have its runtime in gcc29027.dll.
  Do not forget to include this dll with your programs compiled with -Zcrtdll
  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  If you want to use -Zcrtdll and not depend on gcc29027.dll (par example) you
  should link with libgcc by using -lgcc (or -lgpp for C++ programs with
  exceptions). From version 2.95.2 the libgcc library is provided in two
  flavours: single-threaded (e.g. gcc2952s.dll) and multithreaded
  (gcc2952m.dll).

o PGCC has multithreaded support for exceptions. GCC has not.

o PGCC has multithreaded libstdc++. GCC has not.

o Since gcc 3.0 I have removed the support for "obj" object file extension.
  You can still get .obj files if you specify the "-o" option explicitly,
  but gcc itself always defaults to .o (even with -Zomf). The reason was that
  with the progress of gcc it becomes harder to incorporate the old kludges
  used for this (notably in the new C preprocessor related to the -M* family
  of switchs).

o Since 3.0 the -mepilogue switch won't generate epilogue labels for static
  functions. This eliminates problems with duplicate symbols in OMF version
  of libgcc (gcc.lib).

Known bugs and limitations:

o emxomf now traps (sometimes?) when compiling with -g (debug) switch.
  This is due to new format of debugging info (DWARF2). I do not know
  how to fix this. Use A.OUT format (no -Zomf) for debugging and pmgdb.

  Solution: This happens because pgcc uses new stabs format (-gstabs+)
  while emxomf understands only the standard stabs debugging info.
  I've partialy fixed this by adding a tiny command-line preprocessor to gcc.
  It scans argv[] and looks for both -Zomf and -g# where # is a number. If it
  finds one, it replaces -g# by -gstabs#. This doesn't always help, however.

o f77 compiler does not incorporate the changes made for the `main' ported
  version for OS/2. I do not know whenever there are other changes other than
  those needed `just to make it work'. Fortran compiler compiles though and
  compiled programs works - thats all users need, I think.

o pgcc/c++ complains on superfluous "__asm__ __volatile__ (...)" operators.
  For example (taken from /emx/include/386/builtin.h):

    static __inline__ void __enable (void)
    {
      __asm__ __volatile__ ("sti");
    }

  Solution: remove __volatile__'s from enable() and disable() functions in
  include/386/builtin.h; they're not needed since no registers are changed
  inside __asm__ so there is no need to declare it __volatile__:
  a better solution for EM would be to declare in __asm__ instead
  of volatile the list of registers that gets changed).
