Here are the changes from version 20010806 to version 20011006.

Summary:
   + Added MLton.Exn.history, which is similar to SMLofNJ.exnHistory.
   + Support for #line directives of the form (*#line line.col "file"*).
   + Performance improvements in native codegenerator.
   + Bug fixes in front-end, optimizer, register allocator,
     Real.{maxFinite, minPos, toManExp}, and in heap save and restore.

* 2001-10-5
  - Fixed a bug in polymorphic layered patterns, like
	val 'a a as b = []
    These would always fail due to the variable "a" not being handled correctly.
  - Fixed the syntax of "val rec" so that a pattern is allowed on the left-hand
    side of the =.  Thus, we used to reject, but now accept, the following.
	val rec a as b as c = fn _ => ()
	val rec a : unit -> unit : unit -> unit = fn () => ()
    Thanks again to Andreas Rossberg's test files.  This is now tested for in
    valrec.sml.
  - Fixed dynamic semantics of "val rec" so that if "val rec" is used to
    override constructor status, then at run time, the Bind exception is raised
    as per rule 126 of the Definition.  So, for example, the following program
    type checks and compiles, but raises Bind at run time.
	val rec NONE = fn () => ()
	val _ = NONE ()
    Again, this is checked in valrec.sml.
  - Added '\r\n' to ml.lex so that Windows style newlines are acceptable in 
    input files.

* 2001-10-4
  - Fixed bug in the implementation of "open" declarations, which in the case of
    "open A B" had opened A and then looked up B in the resulting environment.  
    The correct behaviour (see rule 22 of the Definition) is to lookup each
    longstrid in the current environment, and then open them all in sequence. 
    This is now checked for in the open.sml regression test.  Thanks to Andreas
    Rossberg for pointing this bug out.
  - Fixed bug that caused tyvars of length 1 (i.e. ') to be rejected.  This is
    now checked in the id.sml regression test.  Again, thanks to Andreas
    Rossberg for the test.

* 2001-10-2
  - Fixed bugs in Real.toManExp (which always returned the wrong result because
    the call to frexp was not treated as side-effecting by the optimizer) and
    in Real.minPos, which was zero because of a mistake with extra precision
    bits.

* 2001-10-1
  - Added MLton.Exn.history.
  - Fixed register allocation bug with fucom instruction.  Was
    allowing fucomp when the first source was not removable.
  - Changed Real.isFinite to use the C math.h finite function.  This fixed the
    nontermination bug which occurred in any program that used Real.maxFinite.

* 2001-9-22
  - Bug fixes found from Ramsey's lrtl in contify.fun and unused-args.fun, both
    of which caused compile-time exceptions to be raised.

* 2001-9-21
  - Fixed MLton.World.{load,save} so that the saved world does not store the
    max heap size.  Instead, the max heap size is computed upon load world in
    exactly the same way as at program startup.  This fixes a long standing (but
    only recently noticed) problem in which mlton (which uses a saved world)
    would attempt to use as much memory as was on the machine used to build
    world.mlton.

* 2001-8-22
  - Added support for #line directives of the form
	(*#line line.col "file"*)
    These directives only affect error messages produced by the parser and 
    elaborator.   

* 2001-8-17
  - Fixed bug in removeUnused optimzation that caused the following program to
    fail to compile.
	fun f l = case l of [] => f l | _ :: l => f l
	val _ = f [13]

* 2001-8-14
  - New x86-codegen infrastructure.
    o support for tracking liveness of stack slots and carrying them 
       in registers across basic blocks
    o more specific Entry and Transfer datatypes to make calling convention
       distinctions more explicit
    o new heuristic for carrying values in registers across basic blocks
       (look Ma, no Overflows!)
    o new "predict" model for generating register allocation hints
    o additional bug fixes

* 2001-8-7
  - MLton.Socket.shutdownWrite flushes the outstream.

--------------------------------------------------------------------------------

Here are the changes from version 20010706 to version 20010806.

Summary:
    + Word.andb (w, 0xFF) now works correctly
    + MLton.Rusage.rusage has a patch to work around a linux kernel bug
    + Programs of the form "exp ; program" are now accepted
    + Added the "MLton.Rlimit" structure
    + Added the "-keep dot" flag, which produces call graphs,
      intraprocedural control-flow graphs, and dominator trees

* 2001-8-6
  - Added simple common block elimination CPS optimization.

* 2001-8-2
  - Took out -keep il.

* 2001-7-31
  - Performance improvements to TextIO.{input, output, output1}

* 2001-7-25
  - Added redundant-test elimination CPS optimization.

* 2001-7-21
  - Added common-subexpression elimination CPS optimization.

* 2001-7-20
  - Bug fix to x86 codegen.  The commuteBinALMD peephole optimization would
    rewrite  mov 2,Y; add Y,Y  as  mov Y,Y; add 2,Y.  Now the appropriate
    interference checks are made.
  - Added intraprocedural unused argument removal.
  - Added intraprocedural flattener.  This avoids some stupid tuple
    allocations in loops.  Decent speedup on a few benchmarks
    (count-graphs, psdes-random, wc-scanStream) and no noticeable
    slowdowns.
  - Added -keep dot flag.

* 2001-7-17
  - Modified grammar to properly handle val rec.  There were several problems.
    o MLton had accepted "val rec 'a ..." instead of "val 'a rec ..."
    o MLton had not accepted "val x = 13 and rec f = fn () => ()"
    o MLton had not accepted "val rec rec f = fn () => ()"
    o MLton had not accepted "val rec f = fn () => () and rec g = fn () => ()"

* 2001-7-16
  - Workaround for Linux kernel bug that can cause getrusage to return a wrong
    system time value (low by one second).  See fixedGetrusage in gc.c.
  - Bug fix to x86 codegen.  The register allocator could get confused when
    doing comparisons of floating point numbers and use the wrong operand. 
    The bug seems to have never been detected because it only happens when both
    of the operands are already on the floating point stack, which is rare,
    since one is almost always in memory since we don't carry floating point
    values in the stack across basic blocks.
  - Added production to the grammar on page 58 of the Definition that had been
    missing from MLton since day one.
		program ::= exp ; <program>
    Also updated docs to reflect change.
  - Modified grammar to accept the empty program.
  - Added -type-check expert flag to turn on type checking in ILs.

* 2001-7-15
  - Bug fix to the algebraic simplifier.  It had been rewriting
		Word32.andb (w, 0wxFF) to w
    instead of  Word32.andb (w, 0wxFFFFFFFF) to w.

* 2001-7-13
  - Improved CPS shrinker so that if-tests where the then and else branch jump
    to the same label is turned into a direct jump.
  - Improved CPS shrinker (Prim.apply) to handle constructors
	A = A --> true
 	A = B --> false
	A x = B y --> false
  - Rewrote a lot of loops in the basis library to use inequalities instead of
    equality for the loop termination test so that the (forthcoming) overflow
    detection elimination will work on the loop index variable.

* 2001-7-11
  - Fixed minor bugs in Array2.{array,tabulate}, Substring.{slice} that caused 
    the Overflow exception to be raised instead of Size or Subscript
  - Fixed bug in Pack32Big.update that caused the wrong location to be updated.
  - Fixed several bugs in Pack32{Big,Little}.{subArr,subVec,update} that caused
    Overflow to be raised instead of Subscript.  Also, improved the
    implementation so that bounds checking only occurs once per call (instead of
    four times, which was sometimes happening.
  - Fixed bugs in Time.{toMilliseconds,toMicroseconds} that could cause 
    a spurious Overflow exception.
  - Fixed bugs in Time.{fromMilliseconds,fromMicroseconds} that could cause
    a spurious Time exception.
  - Improved Pack32.sub* by reordering the orbs.
  - Improved {Int,IntInf}.mod to increase chances of constant folding.
  - Switched many uses of +, -, * in basis library to the non-overflow checked
    versions.  Modules changed were: Array, Array2, Byte, Char, Int, IntInf, 
    List, Pack32{Big,Little}, Util, String, StringCvt, Substring, TextIO, Time,
    Vector.
  - Added regression tests for Array2, Int (overflow checking), Pack32, 
    Substring, Time.
  - Changed CPS output so that it includes a dot graph for each CPS function.

* 2001-7-9
  - Change OS.Process.exit so that it raises an exception if the exit status
    is not in [0, 256).
  - Added MLton.Rlimit to provide access to getrlimit and setrlimit.

--------------------------------------------------------------------------------

Here are the changes from the 20000906 version to the 20010706 version.

Summary:
    + Native X86 code generator (instead of using gcc)
    + Significantly improved compile times 
    + Significantly improved run times for generated executables
    + Many bug fixes
    + Correct raising of the Overflow exception for integer arithmetic
    + New modules in the MLton structure

* 2001-7-6
  - GC mods from Henry.  Mostly adding inline declarations.

* 2001-7-5
  - Fixed several runtime bugs involving threads, critical sections, and
    signals.

* 2001-6-29
  - Fixed performance bug in cps/two-point-lattice.fun that caused quadratic
    behavior.  This affects the raise-to-jump and useless analayses.  In
    particular, the useless analysis was blowing up when compiling fxp.

* 2001-6-27
  - Henry improved wordAlign -- this sped up GC by 27% (during a self compile).

* 2001-6-20
  - Moved MLton.random to MLton.Random.rand and added other stuff to
    MLton.Random
  - Added MLton.TextIO.mkstemp.
  - Made Int.{div,quot} respect the -detect-overflow switch.

* 2001-6-20
  - Added MLton.Syslog.

* 2001-6-7
  - Fixed bug in MLton.Socket.accept that was in the runtime implementation
    Socket_accept.  It did a setsockopt SO_REUSEADDR after the accept.  It
    should have been after the call to socket in Socket_listen.  Thanks to
    Doug Bagley for the fix.

* 2001-5-30
  - Fixed bug in remove-unused that caused polymorphic equality to return 
    true sometimes when constructors were never used in a pattern match.
    For example, the following (in which A and B are not used as patterns):
	datatype t = A | B
	datatype u = C of t
	val _ = if C A = C B then raise Fail "bug" else ()

* 2001-3-27
  - Fixed bug that caused all of the following to fail
	{LargeWord,Word,SysWord}.{toLargeInt,toLargeIntX,fromLargeInt}
    The problem was the basis library file integer/patch.sml which fixed Word32
    but not the other structures that are the same.

* 2001-2-12
  - Fixed bug in match compiler that caused it to spend a lot of extra time in
	deep patterns.  It still could be exponential however.  Hopefully this
	will get fixed in the release after next.
	This bug could cause very slow compile times in some cases.
	Anyways, this fix cut the "finish infer" time of a self compile down
	from 22 to under 4 seconds.  I.E. most of the time used to be spent due
	to this bug.

* 2001-2-6
  - Fixed bug in frontend that caused the wrong file and line number to be
	reported with errors in functor bodys.

* 2001-1-3 - 2000-2-5
  - Changes to CoreML, XML, SXML, and CPS ILs to replace lists by vectors in
	order to decrease space usage.

* 2001-1-16
  - Fixed a bug in constant propagation where the length of vectors was not
	propagated properly.

* 2000-12-11 - 2001-1-3
  - Major rewrite of elaborator to use a single hash table for each namespace
	instead of a hash table for every environment.

* 2000-12-20
  - Fixed some bugs in the SML/NJ compatibility library,
	src/lib/mlton-subs-in-smlnj.

* 2000-12-8
  - More careful removal of tracing code when compiling MLton_debug=0.
    This cut down self compile data size by 100k and compile time by a few 
    seconds.
  - Added built in character and word cases propagated throughout all ILs.

* 2000-12-6
  - Added max stack size information to gc-summary.

* 2000-12-5
  - Added src/benchmark, which contains an SML program that benchmarks all of
    the SML compilers I have my hands on.  The script has lots of hardwired
    paths for now.

* 2000-12-4
  - Fixed bug in Posix.ProcEnv.environ, which did not work correctly in a saved
    world (the original environ was saved).  In fact, it did not work at all
    because the ML primitive expected a constant and the C was a nullary
    function. This caused a segfault with any program using
    Posix.ProcEnv.environ. 
  - Added MLton.ProcEnv.setenv, since there doesn't seem to be any setenv in
    the basis library.

* 2000-11-29
  - Changed backend so that it should no longer generate machine programs with
    void operands.
  - Added -detect-overflow and -safe flags.

* 2000-11-27 - 2000-11-28
  - Changes in many places to use List.revMap instead of List.map to cut down
    on allocation.

* 2000-11-21
  - Added MLton.Word.~ and MLton.Word8.~ to the MLton basis library structure.

* 2000-11-20
  - Fixed a bug in the CPS shrinker that could cause a compile-time failure.
    It was maintaining occurrence counts incorrectly.

* 2000-11-15
  - Fixed a (performance) bug in constant propagation that caused the hashing
	to be bad.
  - Improved translation to XML so that the match compiler isn't called on
	tuple or if expressions.  This should speed up the translation and
	make the output smaller.
  - Fixed a bug in the match compiler that caused it to not generate integer
	case statements.  This should speed up the mlyacc benchmark and the 
	MLton front end.

* 2000-11-9
  - Added IntInf_equal and IntInf_compare primitives.
  - Took out the automatic -keep c when compiling -g.

* 2000-11-8
  - Added a whole bunch of algebraic laws to the CPS shrinker, including
    some specifically targeted to IntInf primitives.

* 2000-11-3
  - Improved implementation of properties so that sets don't allocate.
  - Improved implementation of type homomorphism in type inference.  What
    was there before appears to have been a bug -- it didn't use the property
    on types.

* 2000-11-2
  - Fixed timers used with -v option to use user + sys time.

* 2000-10-27
  - Split the runtime basis library C files into many separate files so that
    only the needed code would be included by the linker.
  - Fixed several bugs in the front end grammar and elaborator that caused
    type specifications to be handled incorrectly.  The following three programs
    used to be handled incorrectly, but are now handled correctly.
      signature S = sig type t and u = int end (* reject *)
      signature S = sig type t = int and u = t end (* accept *)
      signature S = sig eqtype t and u = int end (* reject *)

* 2000-10-25
  - Changes to main.sml to run complete compiles with -native switch.

* 2000-10-24
  - Removed defunctorizer.

* 2000-10-20
  - Fixed bug in cps-tree.fun PrimExp.maySideEffect.  This bug could cause
    no operand failures in the backend.
  - Fixed bug in the runtime implementation of MLton.size.  The size for stack
    objects was using the used instead of reserved, and so was too low.

* 2000-10-19
  - Replaced automatically generated dependencies in src/runtime/Makefile with
    hand generated ones.  Took out make depend from src/Makefile.  makedepend
    was behaving really badly on RHAT 7.0.
  - Tweaked compiler to shorten width of C output lines to work around
	bug in RHAT 7.0 cpp which silently truncates (very) long lines.
  - Fixed bug in grammar that didn't allow "op" to occur in datatype and 
    exception bindings, causing the following to fail
	datatype t = op T
	exception op E = op Fail
  - Improved error messages in CM processor.  Fixed bug in CM Alias handling.

* 2000-10-18
  - Fixed two bugs in the gc that did comparisons with (s->limit - s->frontier),
    which of course doesn't work if frontier is beyond limit, since these are
    unsigned.  This could have caused segfaults, except that the mutator 
    checks the frontier upon return from the GC.

* 2000-10-17
  - Fixed bug in backend in the calculation of maxFrameSize.  It could be 
    wrong (low) in some situations.
  - Improved CPS inliner's estimate of function sizes.  The size of a function 
    now takes into account other inlined functions that the function calls. 
    This also changed the meaning of the size argument to the -inline switch.
    It now corresponds (roughly) to the product of the size of the function
    and the number of calls.  In general, it should be larger than before.
 
* 2000-10-13
  - Made some calls to Array.sub unsafe in the implementation of Array2.
  - Integrated Matthew's new x86 backend with floating point support.

* 2000-10-9
  - Fixed CM file processor so that MLton works if it is run from a different
    directory than the main CM file.

* 2000-10-4
  - Changed LimitCheck so it loops on the frontier > limit check.  This fixed
    a potential bug in threads caused when there is enough space available for
    a thread, t, before switching to another thread but not enough space when it
    resumes.  This could have caused a segfault.

* 2000-10-3
  - More rewrites of TextIO.StreamIO to improve speed.
  - Changed TextIO so that only TextIO.stdErr is unbuffered.
  - Changed TextIO so that FIFOs and sockets are buffered.

* 2000-10-2
  - Combined remove-unused-constructors, remove-unused-functions, and
    remove-unused-globals into a single pass that runs to fixed-point and
    produces results at least as good as running the previous three in (any)
    sequence. 

* 2000-9-29
  - Added GC_FIRST_CHECK, which does a gc at each limit check the first time it
    reached.
  - Reimplemented TextIO.StreamIO (from 2000-9-12) to use lists of strings
    instead of lists of characters so that the per char space overhead is small.

* 2000-9-21
  - Fixed bug in profiling labels in C code.  The label was always the basic 
    block label instead of the cps function label.
  - Added -b switch to mlprof to gather data at the basic block level.
  - Improved performance of TextIO.input1 by about 3X.

* 2000-9-15 - 2000-9-19
  - Added overflow exceptions to CPS and Machine ILs.

* 2000-9-12
  - Fixed TextIO.scanStream.  It was very broken.
  - Added TextIO.{getInstream,mkInstream,setInstream}
          TextIO.StreamIO.{canInput,closeIn,endOfStream,input1,input,inputAll,
                           inputLine,inputN}

* 2000-9-11
  - Fixed Real_qequal in mlton-lib.h.  It was missing a paren that caused
    code using it to not even compile.  It was also semantically incorrect.
  - Noted that Real_{equal,lt,le,gt,ge} may not follow basis library spec, since 
    ANSI does not require IEEE compliance, and hence these could return wrong
    results when nans are involved.

--------------------------------------------------------------------------------

Here are the changes from the 20000712 version to the 20000906 version.

    Version 20000906 is mostly a bugfix release over 20000712.  The other major
    changes are that <tt>mllex</tt> and <tt>mlyacc</tt> are now included and
    that <tt>mlton</tt> can now process a limited subset of CM files as input.

* 2000-9-6
  - Fixed Socket_listen in mlton-lib.c so that it closes the socket if the 
	bind, listen, or getsockname fails.  This could have caused a file
	descriptor leak.

* 2000-9-5
  - Added -static commandline switch.
  - Changed default max heap size to .85 RAM from .95 RAM.
  - Added PackRealLittle structure to basis library.

* 2000-8-25
  - Added cases on integers to ILs (instead of using sequences of tests) so that
    backend can emit more efficient test (jump table, binary tree, ...).

* 2000-8-24
  - Fixed bug in gc.c.  dfsInitializeStack would smummap a NULL pointer whenver
    toSpace was NULL.  This could gause MLton.size to segfault.
  - Fixed bug in Popt that caused -k to fail with no keeps.

* 2000-8-22 - 2000-8-23
  - Ported mllex and mlyacc from SML/NJ

* 2000-8-20 - 2000-8-21
  - Added ability to us a .cm file as input to MLton.

* 2000-8-16
  - Ported mlprof to SML.
  - Fixed bug in library/basic/assert.sml that caused asserts to be run even
    when MLton.debug = false.

* 2000-8-15
  - Fixed bug in backend -- computation of maxFrameSize was wrong.  It didn't
    count slots in frames that didn't make nontail calls.  This could lead to
    the stack being overwritten because a stack limit check didn't guarantee 
    enough space, and lead to a seg fault.
  - Fixed bug in gc.c newThreadOfSize.  If the thread allocation caused a gc,
    then the stack wasn't forwarded, leading to a seg fault.  The solution was
    to ensure enough memory all at once, and then fill in both objects.

* 2000-8-14
  - Changed limit checks so that checks < 512 bytes are replaced by a check for
    0.  The runtime also moves the limit down by 512.  This is done so that the
    common case, a small limit check, has less code and is faster.
  - Fixed bug in cps/cps-tree.fun Program.hasPrim returned true for any program
    that had *any* primapp, not just programs satisfying the predicate.  This
    caused cps/once.fun to be overly conservative, since it thought that every
    program used continuations.

* 2000-8-10
  - Fixed bug in CPS typechecker.  It didn't enforce that handlers should be
    defined before any reference to them -- including implicit references in 
    HandlerPops.  This caused an evil bug in the liveness analysis where a 
    variable that was only live in the handler was missed in a continuation
    because the liveness for the handler wasn't computed yet.
  - Limited the size for moving up limit checks for arrays whose size is known
    at compile time to avoid huge limit checks getting moved into loops.
  - added -indent, -kp, -show-types switches.
  - Put optimization in CPS IL suggested by Neal Glew.  It determines for each
    toplevel function if it can raise an exception to its caller.  Also, it
    removes HanderPush and HandlerPop for handlers that are not on top of the 
    stack for any nontail call.

* 2000-8-8
  - Changed register allocator so that continuation formals can be allocated in
    pseudo registers -- they aren't necessarily forced to the stack.

* 2000-8-3
  - Fixed bug in constant folding.  
	Word8.>> had been used to implement Word8.~>>.
  - Fixed bug in allocate registers that was not forcing the size argument to
    Primitive.Array.array to be a stack slot.  This could cause problems if
    there was a thread switch in the limit check, since upon return the size
    pseudo register would have a bogus value.

* 2000-8-1
  - Turned back on Xml simplification after monomorphisation.

* 2000-7-31
  - Fixed bug in MLton.Itimer.set that caused the time to be doubled.
  - Fixed bug in MLton.Thread that made it look like asynchronous exceptions
    were allowed by throw'ing an exception raising thunk to an interrupted
    thread obtained via a signal handler.  Attempting asynchronous exceptions
    will now cause process death, with a helpful error message.

* 2000-7-27
  - Updated docs to include structure World: MLTON_WORLD in MLton structure.
  - Added toplevel signatures MLTON_{CONT, ..., WORLD} to basis library.
  - Fixed broken link in docs to CM in cmcat section.

* 2000-7-26
  - Eliminated GC_switchToThread and Thread_switchTo1, since the inlined
    version Thread_switchTo is all that's needed, and Matt's X86 backend 
    now handles it.
  - Added MLton.Signal.vtalrm, needed for 
    Itimer.Set{which = Itimer.Virtual, ...}

* 2000-7-25
  Added MLton.Socket.shutdownWrite.

* 2000-7-21
  Updated mlton-lib.c MLton_bug with new email (MLton@sourcelight.com).

* 2000-7-19
  Fixed Posix.Process.kill to check for errors.

* 2000-7-18
  Fixed the following Posix.ProcEnv functions to check for errors:
    setgid, setpgid, setsid, setuid.
  Fixed doc/examples/callcc.sml.

--------------------------------------------------------------------------------

Here are the changes from the 1999-7-12 to the 20000712 version.

* 2000-6-10 - 2000-7-12
  Too many changes to count: bug fixes, new basis library modules,
	optimizer improvements.
* 2000-6-30
  Fixed bug in monomorphiser that caused programs with non-value
  carrying exception declarations in polymorphic functions to have a
  compile-time error because of a duplicate label.  The problem was
  that the exception constructor wasn't duplicated.
* 2000-5-22 - 2000-6-10
  Finished the changes for the new Cps IL.
* 2000-1-1
  Fixed some errors in the basis library.
	Real.copySign
	Posix.FileSys.fpathconf
	Posix.IO.{lseek, getlk, setlk, setlkw}
	Posix.ProcEnv.setpgid
	Posix.TTY.getattr
	System.FileSys.realPath
* 1999-12-22
  Fixed bug in src/closure-convert/abstract-value.fun that caused a
  compiler failure whenever a program had a vector where the element
  type contained an ->.
* 1999-12-10
  - Changed dead code elimination in core-ml/dead-code.fun so that
    wildcard declarations (val _ = ...) in the basis are kept.
    Changed places in the basis library to take advantage of this
  - Added setTopLevelHander primitive so that the basis library code
    can define the toplevel handler.
  - Changed basis-library/misc/suffix.sml to call OS.Process.exit.
    Took out Halt transfer from Cps, since the program never should
    reach it.
  - Cleaned up basis-library/system/{process.sml, unix.sml} to use
    the new signal handling stuff.
* 1999-11-28 - 1999-12-20
  Added support for threads and cleaned up signal handling.
  This involved a number of changes.
  - The stack is now allocated as just another kind of heap object.
  - Limit checks are inserted at all loop headers, whether or not there
    is any allocation.  This is to ensure that the signal handler
    always has a chance to get called.
  - The register allocator puts more variables in stack slots.  The new
    rule is that a variable goes in a stack slot if it is ever live
    across a nontail call, in a handler, or (this is the new part)
    across a limit check.
  - Arguments are passed on the stack, with the convention determined by 
    argument types.
  - The "locals" array of pointers that was copied to/from for GC is now 
    gone, because no registers (in particular no pointer valued
    registers) can be live at a limit check point.
* 1999-11-21
  - Runtime system
    o Fixed a bug introduced by the signal code
	(presumably on 1999-8-9) that caused a gc to *not* be performed when
	doing a save world.  This caused the heaps created by save world to
	be the same size as the heap -- not the live data.  This was quite
	bad. 
    o Cleaned up the Makefile.  Add make depend.
    o Added max gc pause to gc-summary info.
    o Move heap translation variables that had been file statics into
	the GC_state.
  - Made structure Position available at toplevel.
  - Basis Library
    o Added MLton.loadWorld
  - Added Primitive.usesCallcc
  - Added Primitive.safe
  - Removed special size functions from cps/save-world -- they are no
    longer necessary since size doesn't do a gc.
  - Fixed another (sigh) bug in cps/simplify-types.fun that could
    cause it to not terminate.
* 1999-11-16
  - Cleaned up backend/machine.fun a bit so that it spits out macros
    for allocation of objects and bumping of frontier.  Added macros
    MLTON_object and MLTON_incFrontier to include/mlton-lib.h.
  - Fixed a bug in backend/limit-check.fun that caused loops to not be 
    detected if they were only reached by a case branch.  This could 
    cause there to be loop that allocates with no limit check.
    Needless to say, this could cause a segfault if the loop ran for
    long enough.
* 1999-10-18
  Added basis library function Array2.copy.
* 1999-8-15
  Turned off globalization of ref cells
  (closure-convert/globalize.fun) because it interacts badly with
  serialization.
* 1999-8-13
  Fixed bug in mlton-lib.h in MLTON_allocArrayNoPointers that was
  triggered when bytesPerElt == 0.  The problem was that it wasn't
  reserving space for the forwarding pointer.  This could cause a seg
  fault. 
* 1999-8-8 and 1999-8-9
  Added support for signal handling.
* 1999-8-7
  Fixed bugs in Array.tabulate (and other tabulate variants) caused if
  the function argument used callcc.
* 1999-8-1
  Added serialization, which was mostly code in src/runtime/gc.c.
  GC_serialize converts an object to a Word8Vector.vector.
  GC_deserialize undoes the conversion.
  (de)Serialization should work for all objects except for functions,
  because I haven't yet added the support in the flow analysis.
* 1999-7-31
  - Cleaned up the GC.  Changed headers, by stealing a bit from the
    number of non pointers and making it a mark bit (used in GC_size).
  - Rewrote GC_size so that it runs in time proportional to the number 
    of pointers in the object.  It does a depth-first-search now,
    using toSpace to hold the stack.
* 1999-7-30
  - Fixed bug in SUBSTRING.  getc had the wrong type.  This bug wasn't
    noticed because MLton doesn't do enough type checking.
  - Fixed bug (seg fault) caused when a GC immediately followed a throw.
* 1999-7-29
  Fixed bug in Date.fmt (basis-library/system/date.sml).  It was not
  setting Tm.buf, and hence the time was always 0 unless there had
  been a previous call to setTmBuf.
* 1999-7-28
  - Fixed bugs in Posix.IO.FLock.{getlk,setlk,setlkw}, which would
    cause compilation to fail because FLock.toInt was defined as the C 
    castInt, which no longer exists.  Instead, expand FLock.toInt to
    MLTON_pointerToInt, which was added to include/mlton-lib.h.
  - Changed Posix.Primitive.Flock to Posix.Primitive.FLock.
  - Added MLTON_chown, MLTON_ftruncate to include/mlton-posix.h.  They were
    missing.  This would cause compilation of any program using
    Posix.FileSys.{chown,ftruncate} to fail.  Also made it so all of the
    primitives in basis-library/posix/primitive.sml use MLTON_ versions
    of functions, even if a wrapper is unnecessary.
* 1999-7-25
  Added some other missing signature definitions to toplevel.
* 1999-7-24
  Added missing OS_* signature definitions to 
  basis-library/top-level/top-level.sml.
* 1999-7-19
  Fixed bug in basis-library/arrays-and-vectors/mono-array.sml.  Used
  :> instead of : so that the monomorphic array types are abstract.

--------------------------------------------------------------------------------

Here are the changes from the 1999-3-19 version to the 1999-7-12 version.

* 1999-7-12
  Changed src/backend/machine.fun so that the 'pointer locals' array is
  only as large as neccessary in order to copy all pointer-valued
  locals, not as large as the number of pointer-valued locals.
* 1999-7-11
  Rewrote src/backend/allocate-registers.fun so that it does a better
  job of sharing "registers" (i.e. C local variables) and stack slots.
  This should cut down on the amount of copying that has to happen
  before and after a gc.  It should also cut down on the size of stack 
  slots.
* 1999-7-10
  Fixed a bug in src/backend/parallel-move.fun that should have been
  triggered on most any parallel move.  I guess parallel moves almost
  never happened due to the old register allocation strategy -- but, 
  with the new one (see note for 1999-7-12) parallel moves will be 
  frequent.
* 1999-6-27
  Fixed src/main.sml so that when compiling -p, the .c file is compiled 
	-g and the .o is linked -p. 
  In bakend/machine.fun, added profiling comments before chunkswitches 
	and put in an optimization to avoid printing repeated profiling
	comments.  Also, profiling comments are only output when
	compiling -p.
* 1999-6-17
  Changed -i to -inline, -f to -flatten, -np to -no-polyvariance, 
	-u to -unsafe.
  Added -i, -I, -l, -L flags for includes and libraries.
  Updated documentation for these options and for ffi.
* 1999-6-16
  Hardwired version number in src/control/control.sml.  As it stood,
  the version number was computed when MLton was built after someone
  downloaded it, which was clearly wrong.
* 1999-6-16
  Fixed undefined variable time in GC_done in src/runtime/gc.c.
* 1999-6-8
  in include/mlton-lib.h, 
     removed #include <huge_val.h>
     added #include <math.h>
     and deleted all of the function signatures I had copied from math.h
  Changed Real.{minNormalPos, minPos, maxFinite} so that they are
	computed in real.sml instead of appearing as constants in the C.
* 1999-6-7
  IntInf.pow added to basis library.
* 1999-6-4
  bin/mlton changed to use .arch-n-opsys if it exists. 
* 1999-6-3
  src/Makefile changed to use sml-cm instead of sml
* 1999-5-10
  Patch to src/atoms/small-int-inf.fun to work around a bug in the SML/NJ 
  implementation of bignums.  This bug was causing some hex bignum
  constants to be lexed incorrectly.
* 1999-4-15
  Comments emitted in C code for profiling.  The comments identify the 
  Cps function responsible for each C statement.
* 1999-4-15
  callcc and throw added.
* 1999-4-15
  Bug in src/cps/simplify-types fixed.  The bug caused nontermination
   whenever there was a circular datatype with a vector on the rhs.
   E.g.  datatype t = T of t vector

--------------------------------------------------------------------------------

Here are the changes from the 1998-8-26 version to the 1999-3-19 version.

* Compile time and code size have decreased.
* Runtime performance of executables has improved.
* Large programs can now be compiled.
* MLton is self hosting.
* The basis library is mostly complete and many bugs have been fixed.
* The monomorphiser (-m) is no longer available.
* The heap and stack are automatically resized.
* There are now facilities for heap checkpointing (MLton.saveWorld)
  and object size computation (MLton.size).
* MLton uses the GNU multiprecision library to provide a fast
  implemenation of IntInf. 
