Initial revision

[fw/sdcc] / support / gc / README

Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
Copyright (c) 1991-1996 by Xerox Corporation.  All rights reserved.
Copyright (c) 1996-1999 by Silicon Graphics.  All rights reserved.

THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.

Permission is hereby granted to use or copy this program
for any purpose,  provided the above notices are retained on all copies.
Permission to modify the code and to distribute modified code is granted,
provided the above notices are retained, and a notice that the code was
modified is included with the above copyright notice.

This is version 5.0alpha3 of a conservative garbage collector for C and C++.

You might find a more recent version of this at

http://www.hpl.hp.com/personal/Hans_Boehm/gc

HISTORY -

  Early versions of this collector were developed as a part of research
projects supported in part by the National Science Foundation
and the Defense Advance Research Projects Agency.
Much of the code was rewritten by Hans-J. Boehm (boehm@acm.org) at Xerox PARC
and at SGI.

Some other contributors:  

More recent contributors are mentioned in the modification history at the
end of this file.  My apologies for any omissions.

The SPARC specific code was contributed by Mark Weiser
(weiser@parc.xerox.com).  The Encore Multimax modifications were supplied by
Kevin Kenny (kenny@m.cs.uiuc.edu).  The adaptation to the RT is largely due
to Vernon Lee (scorpion@rice.edu), on machines made available by IBM.
Much of the HP specific code and a number of good suggestions for improving the
generic code are due to Walter Underwood (wunder@hp-ses.sde.hp.com).
Robert Brazile (brazile@diamond.bbn.com) originally supplied the ULTRIX code.
Al Dosser (dosser@src.dec.com) and Regis Cridlig (Regis.Cridlig@cl.cam.ac.uk)
subsequently provided updates and information on variation between ULTRIX
systems.  Parag Patel (parag@netcom.com) supplied the A/UX code.
Jesper Peterson(jep@mtiame.mtia.oz.au), Michel Schinz, and
Martin Tauchmann (martintauchmann@bigfoot.com) supplied the Amiga port.
Thomas Funke (thf@zelator.in-berlin.de(?)) and
Brian D.Carlstrom (bdc@clark.lcs.mit.edu) supplied the NeXT ports.
Douglas Steel (doug@wg.icl.co.uk) provided ICL DRS6000 code.
Bill Janssen (janssen@parc.xerox.com) supplied the SunOS dynamic loader
specific code. Manuel Serrano (serrano@cornas.inria.fr) supplied linux and
Sony News specific code.  Al Dosser provided Alpha/OSF/1 code.  He and
Dave Detlefs(detlefs@src.dec.com) also provided several generic bug fixes.
Alistair G. Crooks(agc@uts.amdahl.com) supplied the NetBSD and 386BSD ports.
Jeffrey Hsu (hsu@soda.berkeley.edu) provided the FreeBSD port.
Brent Benson (brent@jade.ssd.csd.harris.com) ported the collector to
a Motorola 88K processor running CX/UX (Harris NightHawk).
Ari Huttunen (Ari.Huttunen@hut.fi) generalized the OS/2 port to
nonIBM development environments (a nontrivial task).
Patrick Beard (beard@cs.ucdavis.edu) provided the initial MacOS port.
David Chase, then at Olivetti Research, suggested several improvements.
Scott Schwartz (schwartz@groucho.cse.psu.edu) supplied some of the
code to save and print call stacks for leak detection on a SPARC.
Jesse Hull and John Ellis supplied the C++ interface code.
Zhong Shao performed much of the experimentation that led to the
current typed allocation facility.  (His dynamic type inference code hasn't
made it into the released version of the collector, yet.)
(Blame for misinstallation of these modifications goes to the first author,
however.)

OVERVIEW

    This is intended to be a general purpose, garbage collecting storage
allocator.  The algorithms used are described in:

Boehm, H., and M. Weiser, "Garbage Collection in an Uncooperative Environment",
Software Practice & Experience, September 1988, pp. 807-820.

Boehm, H., A. Demers, and S. Shenker, "Mostly Parallel Garbage Collection",
Proceedings of the ACM SIGPLAN '91 Conference on Programming Language Design
and Implementation, SIGPLAN Notices 26, 6 (June 1991), pp. 157-164.

Boehm, H., "Space Efficient Conservative Garbage Collection", Proceedings
of the ACM SIGPLAN '91 Conference on Programming Language Design and
Implementation, SIGPLAN Notices 28, 6 (June 1993), pp. 197-206.

  Possible interactions between the collector and optimizing compilers are
discussed in

Boehm, H., and D. Chase, "A Proposal for GC-safe C Compilation",
The Journal of C Language Translation 4, 2 (December 1992).

and

Boehm H., "Simple GC-safe Compilation", Proceedings
of the ACM SIGPLAN '96 Conference on Programming Language Design and
Implementation.

(Both are also available from
http://reality.sgi.com/boehm/papers/, among other places.)

  Unlike the collector described in the second reference, this collector
operates either with the mutator stopped during the entire collection
(default) or incrementally during allocations.  (The latter is supported
on only a few machines.)  It does not rely on threads, but is intended
to be thread-safe.

  Some of the ideas underlying the collector have previously been explored
by others.  (Doug McIlroy wrote a vaguely similar collector that is part of
version 8 UNIX (tm).)  However none of this work appears to have been widely
disseminated.

  Rudimentary tools for use of the collector as a leak detector are included, as
is a fairly sophisticated string package "cord" that makes use of the collector.
(See cord/README.)


GENERAL DESCRIPTION

  This is a garbage collecting storage allocator that is intended to be
used as a plug-in replacement for C's malloc.

  Since the collector does not require pointers to be tagged, it does not
attempt to ensure that all inaccessible storage is reclaimed.  However,
in our experience, it is typically more successful at reclaiming unused
memory than most C programs using explicit deallocation.  Unlike manually
introduced leaks, the amount of unreclaimed memory typically stays
bounded.

  In the following, an "object" is defined to be a region of memory allocated
by the routines described below.  

  Any objects not intended to be collected must be pointed to either
from other such accessible objects, or from the registers,
stack, data, or statically allocated bss segments.  Pointers from
the stack or registers may point to anywhere inside an object.
The same is true for heap pointers if the collector is compiled with
 ALL_INTERIOR_POINTERS defined, as is now the default.

Compiling without ALL_INTERIOR_POINTERS may reduce accidental retention
of garbage objects, by requiring pointers from the heap to to the beginning
of an object.  But this no longer appears to be a significant
issue for most programs.

There are a number of routines which modify the pointer recognition
algorithm.  GC_register_displacement allows certain interior pointers
to be recognized even if ALL_INTERIOR_POINTERS is nor defined.
GC_malloc_ignore_off_page allows some pointers into the middle of large objects
to be disregarded, greatly reducing the probablility of accidental
retention of large objects.  For most purposes it seems best to compile
with ALL_INTERIOR_POINTERS and to use GC_malloc_ignore_off_page if
you get collector warnings from allocations of very large objects.
See README.debugging for details.

  Note that pointers inside memory allocated by the standard "malloc" are not
seen by the garbage collector.  Thus objects pointed to only from such a
region may be prematurely deallocated.  It is thus suggested that the
standard "malloc" be used only for memory regions, such as I/O buffers, that
are guaranteed not to contain pointers to garbage collectable memory.
Pointers in C language automatic, static, or register variables,
are correctly recognized.  (Note that GC_malloc_uncollectable has semantics
similar to standard malloc, but allocates objects that are traced by the
collector.)

  The collector does not always know how to find pointers in data
areas that are associated with dynamic libraries.  This is easy to
remedy IF you know how to find those data areas on your operating
system (see GC_add_roots).  Code for doing this under SunOS, IRIX 5.X and 6.X,
HP/UX, Alpha OSF/1, Linux, and win32 is included and used by default.  (See
README.win32 for win32 details.)  On other systems pointers from dynamic
library data areas may not be considered by the collector.

  Note that the garbage collector does not need to be informed of shared
read-only data.  However if the shared library mechanism can introduce
discontiguous data areas that may contain pointers, then the collector does
need to be informed.

  Signal processing for most signals may be deferred during collection,
and during uninterruptible parts of the allocation process.  Unlike
standard ANSI C mallocs, it can be safe to invoke malloc
from a signal handler while another malloc is in progress, provided
the original malloc is not restarted.  (Empirically, many UNIX
applications already assume this.)  To obtain this level  of signal
safety, remove the definition of -DNO_SIGNALS in Makefile.  This incurs
a minor performance penalty, and hence is no longer the default.

  The allocator/collector can also be configured for thread-safe operation.
(Full signal safety can also be achieved, but only at the cost of two system
calls per malloc, which is usually unacceptable.)

INSTALLATION AND PORTABILITY

  As distributed, the macro SILENT is defined in Makefile.
In the event of problems, this can be removed to obtain a moderate
amount of descriptive output for each collection.
(The given statistics exhibit a few peculiarities.
Things don't appear to add up for a variety of reasons, most notably
fragmentation losses.  These are probably much more significant for the
contrived program "test.c" than for your application.)

  Note that typing "make test" will automatically build the collector
and then run setjmp_test and gctest. Setjmp_test will give you information
about configuring the collector, which is useful primarily if you have
a machine that's not already supported.  Gctest is a somewhat superficial
test of collector functionality.  Failure is indicated by a core dump or
a message to the effect that the collector is broken.  Gctest takes about 
35 seconds to run on a SPARCstation 2. On a slower machine,
expect it to take a while.  It may use up to 8 MB of memory.  (The
multi-threaded version will use more.)  "Make test" will also, as
its last step, attempt to build and test the "cord" string library.
This will fail without an ANSI C compiler.

  The Makefile will generate a library gc.a which you should link against.
Typing "make cords" will add the cord library to gc.a.
Note that this requires an ANSI C compiler.

  It is suggested that if you need to replace a piece of the collector
(e.g. GC_mark_rts.c) you simply list your version ahead of gc.a on the
                work.)
ld command line, rather than replacing the one in gc.a.  (This will
generate numerous warnings under some versions of AIX, but it still
works.)

  All include files that need to be used by clients will be put in the
include subdirectory.  (Normally this is just gc.h.  "Make cords" adds
"cord.h" and "ec.h".)

  The collector currently is designed to run essentially unmodified on
machines that use a flat 32-bit or 64-bit address space.
That includes the vast majority of Workstations and X86 (X >= 3) PCs.
(The list here was deleted because it was getting too long and constantly
out of date.)
  It does NOT run under plain 16-bit DOS or Windows 3.X.  There are however
various packages (e.g. win32s, djgpp) that allow flat 32-bit address
applications to run under those systemsif the have at least an 80386 processor,
and several of those are compatible with the collector.

  In a few cases (Amiga, OS/2, Win32, MacOS) a separate makefile
or equivalent is supplied.  Many of these have separate README.system
files.

  Dynamic libraries are completely supported only under SunOS
(and even that support is not functional on the last Sun 3 release),
IRIX 5&6, HP-PA, Win32 (not Win32S) and OSF/1 on DEC AXP machines.
On other machines we recommend that you do one of the following:

  1) Add dynamic library support (and send us the code).
  2) Use static versions of the libraries.
  3) Arrange for dynamic libraries to use the standard malloc.
     This is still dangerous if the library stores a pointer to a
     garbage collected object.  But nearly all standard interfaces
     prohibit this, because they deal correctly with pointers
     to stack allocated objects.  (Strtok is an exception.  Don't
     use it.)

  In all cases we assume that pointer alignment is consistent with that
enforced by the standard C compilers.  If you use a nonstandard compiler
you may have to adjust the alignment parameters defined in gc_priv.h.

  A port to a machine that is not byte addressed, or does not use 32 bit
or 64 bit addresses will require a major effort.  A port to plain MSDOS
or win16 is hard.

  For machines not already mentioned, or for nonstandard compilers, the
following are likely to require change:

1.  The parameters in gcconfig.h.
      The parameters that will usually require adjustment are
   STACKBOTTOM,  ALIGNMENT and DATASTART.  Setjmp_test
   prints its guesses of the first two.
      DATASTART should be an expression for computing the
   address of the beginning of the data segment.  This can often be
   &etext.  But some memory management units require that there be
   some unmapped space between the text and the data segment.  Thus
   it may be more complicated.   On UNIX systems, this is rarely
   documented.  But the adb "$m" command may be helpful.  (Note
   that DATASTART will usually be a function of &etext.  Thus a
   single experiment is usually insufficient.)
     STACKBOTTOM is used to initialize GC_stackbottom, which
   should be a sufficient approximation to the coldest stack address.
   On some machines, it is difficult to obtain such a value that is
   valid across a variety of MMUs, OS releases, etc.  A number of
   alternatives exist for using the collector in spite of this.  See the
   discussion in gcconfig.h immediately preceding the various
   definitions of STACKBOTTOM.
   
2.  mach_dep.c.
      The most important routine here is one to mark from registers.
    The distributed file includes a generic hack (based on setjmp) that
    happens to work on many machines, and may work on yours.  Try
    compiling and running setjmp_t.c to see whether it has a chance of
    working.  (This is not correct C, so don't blame your compiler if it
    doesn't work.  Based on limited experience, register window machines
    are likely to cause trouble.  If your version of setjmp claims that
    all accessible variables, including registers, have the value they
    had at the time of the longjmp, it also will not work.  Vanilla 4.2 BSD
    on Vaxen makes such a claim.  SunOS does not.)
      If your compiler does not allow in-line assembly code, or if you prefer
    not to use such a facility, mach_dep.c may be replaced by a .s file
    (as we did for the MIPS machine and the PC/RT).
      At this point enough architectures are supported by mach_dep.c
    that you will rarely need to do more than adjust for assembler
    syntax.

3.  os_dep.c (and gc_priv.h).
          Several kinds of operating system dependent routines reside here.
        Many are optional.  Several are invoked only through corresponding
        macros in gc_priv.h, which may also be redefined as appropriate.
      The routine GC_register_data_segments is crucial.  It registers static
    data areas that must be traversed by the collector. (User calls to
    GC_add_roots may sometimes be used for similar effect.)
      Routines to obtain memory from the OS also reside here.
    Alternatively this can be done entirely by the macro GET_MEM
    defined in gc_priv.h.  Routines to disable and reenable signals
    also reside here if they are need by the macros DISABLE_SIGNALS
    and ENABLE_SIGNALS defined in gc_priv.h.
      In a multithreaded environment, the macros LOCK and UNLOCK
    in gc_priv.h will need to be suitably redefined.
      The incremental collector requires page dirty information, which
    is acquired through routines defined in os_dep.c.  Unless directed
    otherwise by gcconfig.h, these are implemented as stubs that simply
    treat all pages as dirty.  (This of course makes the incremental
    collector much less useful.)

4.  dyn_load.c
        This provides a routine that allows the collector to scan data
        segments associated with dynamic libraries.  Often it is not
        necessary to provide this routine unless user-written dynamic
        libraries are used.

  For a different version of UN*X or different machines using the
Motorola 68000, Vax, SPARC, 80386, NS 32000, PC/RT, or MIPS architecture,
it should frequently suffice to change definitions in gcconfig.h.


THE C INTERFACE TO THE ALLOCATOR

  The following routines are intended to be directly called by the user.
Note that usually only GC_malloc is necessary.  GC_clear_roots and GC_add_roots
calls may be required if the collector has to trace from nonstandard places
(e.g. from dynamic library data areas on a machine on which the 
collector doesn't already understand them.)  On some machines, it may
be desirable to set GC_stacktop to a good approximation of the stack base. 
(This enhances code portability on HP PA machines, since there is no
good way for the collector to compute this value.)  Client code may include
"gc.h", which defines all of the following, plus many others.

1)  GC_malloc(nbytes)
    - allocate an object of size nbytes.  Unlike malloc, the object is
      cleared before being returned to the user.  Gc_malloc will
      invoke the garbage collector when it determines this to be appropriate.
      GC_malloc may return 0 if it is unable to acquire sufficient
      space from the operating system.  This is the most probable
      consequence of running out of space.  Other possible consequences
      are that a function call will fail due to lack of stack space,
      or that the collector will fail in other ways because it cannot
      maintain its internal data structures, or that a crucial system
      process will fail and take down the machine.  Most of these
      possibilities are independent of the malloc implementation.

2)  GC_malloc_atomic(nbytes)
    - allocate an object of size nbytes that is guaranteed not to contain any
      pointers.  The returned object is not guaranteed to be cleared.
      (Can always be replaced by GC_malloc, but results in faster collection
      times.  The collector will probably run faster if large character
      arrays, etc. are allocated with GC_malloc_atomic than if they are
      statically allocated.)

3)  GC_realloc(object, new_size)
    - change the size of object to be new_size.  Returns a pointer to the
      new object, which may, or may not, be the same as the pointer to
      the old object.  The new object is taken to be atomic iff the old one
      was.  If the new object is composite and larger than the original object,
      then the newly added bytes are cleared (we hope).  This is very likely
      to allocate a new object, unless MERGE_SIZES is defined in gc_priv.h.
      Even then, it is likely to recycle the old object only if the object
      is grown in small additive increments (which, we claim, is generally bad
      coding practice.)

4)  GC_free(object)
    - explicitly deallocate an object returned by GC_malloc or
      GC_malloc_atomic.  Not necessary, but can be used to minimize
      collections if performance is critical.  Probably a performance
      loss for very small objects (<= 8 bytes).

5)  GC_expand_hp(bytes)
    - Explicitly increase the heap size.  (This is normally done automatically
      if a garbage collection failed to GC_reclaim enough memory.  Explicit
      calls to GC_expand_hp may prevent unnecessarily frequent collections at
      program startup.)

6)  GC_malloc_ignore_off_page(bytes)
        - identical to GC_malloc, but the client promises to keep a pointer to
          the somewhere within the first 256 bytes of the object while it is
          live.  (This pointer should nortmally be declared volatile to prevent
          interference from compiler optimizations.)  This is the recommended
          way to allocate anything that is likely to be larger than 100Kbytes
          or so.  (GC_malloc may result in failure to reclaim such objects.)

7)  GC_set_warn_proc(proc)
        - Can be used to redirect warnings from the collector.  Such warnings
          should be rare, and should not be ignored during code development.
      
8) GC_enable_incremental()
    - Enables generational and incremental collection.  Useful for large
      heaps on machines that provide access to page dirty information.
      Some dirty bit implementations may interfere with debugging
      (by catching address faults) and place restrictions on heap arguments
      to system calls (since write faults inside a system call may not be
      handled well).

9) Several routines to allow for registration of finalization code.
   User supplied finalization code may be invoked when an object becomes
   unreachable.  To call (*f)(obj, x) when obj becomes inaccessible, use
        GC_register_finalizer(obj, f, x, 0, 0);
   For more sophisticated uses, and for finalization ordering issues,
   see gc.h.

  The global variable GC_free_space_divisor may be adjusted up from its
default value of 4 to use less space and more collection time, or down for
the opposite effect.  Setting it to 1 or 0 will effectively disable collections
and cause all allocations to simply grow the heap.

  The variable GC_non_gc_bytes, which is normally 0, may be changed to reflect
the amount of memory allocated by the above routines that should not be
considered as a candidate for collection.  Careless use may, of course, result
in excessive memory consumption.

  Some additional tuning is possible through the parameters defined
near the top of gc_priv.h.
  
  If only GC_malloc is intended to be used, it might be appropriate to define:

#define malloc(n) GC_malloc(n)
#define calloc(m,n) GC_malloc((m)*(n))

  For small pieces of VERY allocation intensive code, gc_inl.h
includes some allocation macros that may be used in place of GC_malloc
and friends.

  All externally visible names in the garbage collector start with "GC_".
To avoid name conflicts, client code should avoid this prefix, except when
accessing garbage collector routines or variables.

  There are provisions for allocation with explicit type information.
This is rarely necessary.  Details can be found in gc_typed.h.

THE C++ INTERFACE TO THE ALLOCATOR:

  The Ellis-Hull C++ interface to the collector is included in
the collector distribution.  If you intend to use this, type
"make c++" after the initial build of the collector is complete.
See gc_cpp.h for the definition of the interface.  This interface
tries to approximate the Ellis-Detlefs C++ garbage collection
proposal without compiler changes.

Cautions:
1. Arrays allocated without new placement syntax are
allocated as uncollectable objects.  They are traced by the
collector, but will not be reclaimed.

2. Failure to use "make c++" in combination with (1) will
result in arrays allocated using the default new operator.
This is likely to result in disaster without linker warnings.

3. If your compiler supports an overloaded new[] operator,
then gc_cpp.cc and gc_cpp.h should be suitably modified.

4. Many current C++ compilers have deficiencies that
break some of the functionality.  See the comments in gc_cpp.h
for suggested workarounds.

USE AS LEAK DETECTOR:

  The collector may be used to track down leaks in C programs that are
intended to run with malloc/free (e.g. code with extreme real-time or
portability constraints).  To do so define FIND_LEAK in Makefile
This will cause the collector to invoke the report_leak
routine defined near the top of reclaim.c whenever an inaccessible
object is found that has not been explicitly freed.  The collector will
no longer reclaim inaccessible memory; in this form it is purely a
debugging tool.
  Productive use of this facility normally involves redefining report_leak
to do something more intelligent.  This typically requires annotating
objects with additional information (e.g. creation time stack trace) that
identifies their origin.  Such code is typically not very portable, and is
not included here, except on SPARC machines.
  If all objects are allocated with GC_DEBUG_MALLOC (see next section),
then the default version of report_leak will report the source file
and line number at which the leaked object was allocated.  This may
sometimes be sufficient.  (On SPARC/SUNOS4 machines, it will also report
a cryptic stack trace.  This can often be turned into a sympolic stack
trace by invoking program "foo" with "callprocs foo".  Callprocs is
a short shell script that invokes adb to expand program counter values
to symbolic addresses.  It was largely supplied by Scott Schwartz.)
  Note that the debugging facilities described in the next section can
sometimes be slightly LESS effective in leak finding mode, since in
leak finding mode, GC_debug_free actually results in reuse of the object.
(Otherwise the object is simply marked invalid.)  Also note that the test
program is not designed to run meaningfully in FIND_LEAK mode.
Use "make gc.a" to build the collector.

DEBUGGING FACILITIES:

  The routines GC_debug_malloc, GC_debug_malloc_atomic, GC_debug_realloc,
and GC_debug_free provide an alternate interface to the collector, which
provides some help with memory overwrite errors, and the like.
Objects allocated in this way are annotated with additional
information.  Some of this information is checked during garbage
collections, and detected inconsistencies are reported to stderr.

  Simple cases of writing past the end of an allocated object should
be caught if the object is explicitly deallocated, or if the
collector is invoked while the object is live.  The first deallocation
of an object will clear the debugging info associated with an
object, so accidentally repeated calls to GC_debug_free will report the
deallocation of an object without debugging information.  Out of
memory errors will be reported to stderr, in addition to returning
NIL.

  GC_debug_malloc checking  during garbage collection is enabled
with the first call to GC_debug_malloc.  This will result in some
slowdown during collections.  If frequent heap checks are desired,
this can be achieved by explicitly invoking GC_gcollect, e.g. from
the debugger.

  GC_debug_malloc allocated objects should not be passed to GC_realloc
or GC_free, and conversely.  It is however acceptable to allocate only
some objects with GC_debug_malloc, and to use GC_malloc for other objects,
provided the two pools are kept distinct.  In this case, there is a very
low probablility that GC_malloc allocated objects may be misidentified as
having been overwritten.  This should happen with probability at most
one in 2**32.  This probability is zero if GC_debug_malloc is never called.

  GC_debug_malloc, GC_malloc_atomic, and GC_debug_realloc take two
additional trailing arguments, a string and an integer.  These are not
interpreted by the allocator.  They are stored in the object (the string is
not copied).  If an error involving the object is detected, they are printed.

  The macros GC_MALLOC, GC_MALLOC_ATOMIC, GC_REALLOC, GC_FREE, and
GC_REGISTER_FINALIZER are also provided.  These require the same arguments
as the corresponding (nondebugging) routines.  If gc.h is included
with GC_DEBUG defined, they call the debugging versions of these
functions, passing the current file name and line number as the two
extra arguments, where appropriate.  If gc.h is included without GC_DEBUG
defined, then all these macros will instead be defined to their nondebugging
equivalents.  (GC_REGISTER_FINALIZER is necessary, since pointers to
objects with debugging information are really pointers to a displacement
of 16 bytes form the object beginning, and some translation is necessary
when finalization routines are invoked.  For details, about what's stored
in the header, see the definition of the type oh in debug_malloc.c)

INCREMENTAL/GENERATIONAL COLLECTION:

The collector normally interrupts client code for the duration of 
a garbage collection mark phase.  This may be unacceptable if interactive
response is needed for programs with large heaps.  The collector
can also run in a "generational" mode, in which it usually attempts to
collect only objects allocated since the last garbage collection.
Furthermore, in this mode, garbage collections run mostly incrementally,
with a small amount of work performed in response to each of a large number of
GC_malloc requests.

This mode is enabled by a call to GC_enable_incremental().

Incremental and generational collection is effective in reducing
pause times only if the collector has some way to tell which objects
or pages have been recently modified.  The collector uses two sources
of information:

1. Information provided by the VM system.  This may be provided in
one of several forms.  Under Solaris 2.X (and potentially under other
similar systems) information on dirty pages can be read from the
/proc file system.  Under other systems (currently SunOS4.X) it is
possible to write-protect the heap, and catch the resulting faults.
On these systems we require that system calls writing to the heap
(other than read) be handled specially by client code.
See os_dep.c for details.

2. Information supplied by the programmer.  We define "stubborn"
objects to be objects that are rarely changed.  Such an object
can be allocated (and enabled for writing) with GC_malloc_stubborn.
Once it has been initialized, the collector should be informed with
a call to GC_end_stubborn_change.  Subsequent writes that store
pointers into the object must be preceded by a call to
GC_change_stubborn.

This mechanism performs best for objects that are written only for
initialization, and such that only one stubborn object is writable
at once.  It is typically not worth using for short-lived
objects.  Stubborn objects are treated less efficiently than pointerfree
(atomic) objects.

A rough rule of thumb is that, in the absence of VM information, garbage
collection pauses are proportional to the amount of pointerful storage
plus the amount of modified "stubborn" storage that is reachable during
the collection.  

Initial allocation of stubborn objects takes longer than allocation
of other objects, since other data structures need to be maintained.

We recommend against random use of stubborn objects in client
code, since bugs caused by inappropriate writes to stubborn objects
are likely to be very infrequently observed and hard to trace.  
However, their use may be appropriate in a few carefully written
library routines that do not make the objects themselves available
for writing by client code.


BUGS:

  Any memory that does not have a recognizable pointer to it will be
reclaimed.  Exclusive-or'ing forward and backward links in a list
doesn't cut it.
  Some C optimizers may lose the last undisguised pointer to a memory
object as a consequence of clever optimizations.  This has almost
never been observed in practice.  Send mail to boehm@acm.org
for suggestions on how to fix your compiler.
  This is not a real-time collector.  In the standard configuration,
percentage of time required for collection should be constant across
heap sizes.  But collection pauses will increase for larger heaps.
(On SPARCstation 2s collection times will be on the order of 300 msecs
per MB of accessible memory that needs to be scanned.  Your mileage
may vary.)  The incremental/generational collection facility helps,
but is portable only if "stubborn" allocation is used.
  Please address bug reports to boehm@acm.org.  If you are
contemplating a major addition, you might also send mail to ask whether
it's already been done (or whether we tried and discarded it).

RECENT VERSIONS:

  Version 1.3 and immediately preceding versions contained spurious
assembly language assignments to TMP_SP.  Only the assignment in the PC/RT
code is necessary.  On other machines, with certain compiler options,
the assignments can lead to an unsaved register being overwritten.
Known to cause problems under SunOS 3.5 WITHOUT the -O option.  (With
-O the compiler recognizes it as dead code.  It probably shouldn't,
but that's another story.)

  Version 1.4 and earlier versions used compile time determined values
for the stack base.  This no longer works on Sun 3s, since Sun 3/80s use
a different stack base.  We now use a straightforward heuristic on all
machines on which it is known to work (incl. Sun 3s) and compile-time
determined values for the rest.  There should really be library calls
to determine such values.

  Version 1.5 and earlier did not ensure 8 byte alignment for objects
allocated on a sparc based machine.

  Version 1.8 added ULTRIX support in gc_private.h.
  
  Version 1.9 fixed a major bug in gc_realloc.
  
  Version 2.0 introduced a consistent naming convention for collector
routines and added support for registering dynamic library data segments
in the standard mark_roots.c.  Most of the data structures were revamped.
The treatment of interior pointers was completely changed.  Finalization
was added.  Support for locking was added.  Object kinds were added.
We added a black listing facility to avoid allocating at addresses known
to occur as integers somewhere in the address space.  Much of this
was accomplished by adapting ideas and code from the PCR collector.
The test program was changed and expanded.

  Version 2.1 was the first stable version since 1.9, and added support
for PPCR.

  Version 2.2 added debugging allocation, and fixed various bugs.  Among them:
- GC_realloc could fail to extend the size of the object for certain large object sizes.
- A blatant subscript range error in GC_printf, which unfortunately
  wasn't exercised on machines with sufficient stack alignment constraints.
- GC_register_displacement did the wrong thing if it was called after
  any allocation had taken place.
- The leak finding code would eventually break after 2048 byte
  byte objects leaked.
- interface.c didn't compile.
- The heap size remained much too small for large stacks.
- The stack clearing code behaved badly for large stacks, and perhaps
  on HP/PA machines.

  Version 2.3 added ALL_INTERIOR_POINTERS and fixed the following bugs:
- Missing declaration of etext in the A/UX version.
- Some PCR root-finding problems.
- Blacklisting was not 100% effective, because the plausible future
  heap bounds were being miscalculated.
- GC_realloc didn't handle out-of-memory correctly.
- GC_base could return a nonzero value for addresses inside free blocks.
- test.c wasn't really thread safe, and could erroneously report failure
  in a multithreaded environment.  (The locking primitives need to be
  replaced for other threads packages.)
- GC_CONS was thoroughly broken.
- On a SPARC with dynamic linking, signals stayed diabled while the
  client code was running.
  (Thanks to Manuel Serrano at INRIA for reporting the last two.)
  
  Version 2.4 added GC_free_space_divisor as a tuning knob, added
  support for OS/2 and linux, and fixed the following bugs:
- On machines with unaligned pointers (e.g. Sun 3), every 128th word could
  fail to be considered for marking.
- Dynamic_load.c erroneously added 4 bytes to the length of the data and
  bss sections of the dynamic library.  This could result in a bad memory
  reference if the actual length was a multiple of a page.  (Observed on
  Sun 3.  Can probably also happen on a Sun 4.)
  (Thanks to Robert Brazile for pointing out that the Sun 3 version
  was broken.  Dynamic library handling is still broken on Sun 3s
  under 4.1.1U1, but apparently not 4.1.1.  If you have such a machine,
  use -Bstatic.)
  
  Version 2.5 fixed the following bugs:
- Removed an explicit call to exit(1)
- Fixed calls to GC_printf and GC_err_printf, so the correct number of
  arguments are always supplied.  The OS/2 C compiler gets confused if
  the number of actuals and the number of formals differ.  (ANSI C
  doesn't require this to work.  The ANSI sanctioned way of doing things
  causes too many compatibility problems.)
  
  Version 3.0  added generational/incremental collection and stubborn
  objects.

  Version 3.1 added the following features:
- A workaround for a SunOS 4.X SPARC C compiler
  misfeature that caused problems when the collector was turned into
  a dynamic library.  
- A fix for a bug in GC_base that could result in a memory fault.
- A fix for a performance bug (and several other misfeatures) pointed
  out by Dave Detlefs and Al Dosser.
- Use of dirty bit information for static data under Solaris 2.X.
- DEC Alpha/OSF1 support (thanks to Al Dosser).
- Incremental collection on more platforms.
- A more refined heap expansion policy.  Less space usage by default.
- Various minor enhancements to reduce space usage, and to reduce
  the amount of memory scanned by the collector.
- Uncollectable allocation without per object overhead.
- More conscientious handling of out-of-memory conditions.
- Fixed a bug in debugging stubborn allocation.
- Fixed a bug that resulted in occasional erroneous reporting of smashed
  objects with debugging allocation.
- Fixed bogus leak reports of size 4096 blocks with FIND_LEAK.

  Version 3.2 fixed a serious and not entirely repeatable bug in
  the incremental collector.  It appeared only when dirty bit info
  on the roots was available, which is normally only under Solaris.
  It also added GC_general_register_disappearing_link, and some
  testing code.  Interface.c disappeared.

  Version 3.3 fixes several bugs and adds new ports:
- PCR-specific bugs.
- Missing locking in GC_free, redundant FASTUNLOCK
  in GC_malloc_stubborn, and 2 bugs in
  GC_unregister_disappearing_link.
  All of the above were pointed out by Neil Sharman
  (neil@cs.mu.oz.au).
- Common symbols allocated by the SunOS4.X dynamic loader
  were not included in the root set.
- Bug in GC_finalize (reported by Brian Beuning and Al Dosser)
- Merged Amiga port from Jesper Peterson (untested)
- Merged NeXT port from Thomas Funke (significantly
  modified and untested)

  Version 3.4:
- Fixed a performance bug in GC_realloc.
- Updated the amiga port.
- Added NetBSD and 386BSD ports.
- Added cord library.
- Added trivial performance enhancement for
  ALL_INTERIOR_POINTERS.  (Don't scan last word.)
  
  Version 3.5
- Minor collections now mark from roots only once, if that
  doesn't cause an excessive pause.
- The stack clearing heuristic was refined to prevent anomalies
  with very heavily recursive programs and sparse stacks.
- Fixed a bug that prevented mark stack growth in some cases.
  GC_objects_are_marked should be set to TRUE after a call
  to GC_push_roots and as part of GC_push_marked, since
  both can now set mark bits.  I think this is only a performance
  bug, but I wouldn't bet on it.  It's certainly very hard to argue
  that the old version was correct.
- Fixed an incremental collection bug that prevented it from
  working at all when HBLKSIZE != getpagesize()
- Changed dynamic_loading.c to include gc_priv.h before testing
  DYNAMIC_LOADING.  SunOS dynamic library scanning
  must have been broken in 3.4.
- Object size rounding now adapts to program behavior.
- Added a workaround (provided by Manuel Serrano and
  colleagues) to a long-standing SunOS 4.X (and 3.X?) ld bug
  that I had incorrectly assumed to have been squished.
  The collector was broken if the text segment size was within
  32 bytes of a multiple of 8K bytes, and if the beginning of
  the data segment contained interesting roots.  The workaround
  assumes a demand-loadable executable.  The original may have
  have "worked" in some other cases.
- Added dynamic library support under IRIX5.
- Added support for EMX under OS/2 (thanks to Ari Huttunen).
  
Version 3.6:
- fixed a bug in the mark stack growth code that was introduced
  in 3.4.
- fixed Makefile to work around DEC AXP compiler tail recursion
  bug.

Version 3.7:
- Added a workaround for an HP/UX compiler bug.
- Fixed another stack clearing performance bug.  Reworked
  that code once more.
  
Version 4.0:
- Added support for Solaris threads (which was possible
  only by reimplementing some fraction of Solaris threads,
  since Sun doesn't currently make the thread debugging
  interface available).
- Added non-threads win32 and win32S support.
- (Grudgingly, with suitable muttering of obscenities) renamed
  files so that the collector distribution could live on a FAT
  file system.  Files that are guaranteed to be useless on
  a PC still have long names.  Gc_inline.h and gc_private.h
  still exist, but now just include  gc_inl.h and gc_priv.h.
- Fixed a really obscure bug in finalization that could cause
  undetected mark stack overflows.  (I would be surprised if
  any real code ever tickled this one.)
- Changed finalization code to dynamically resize the hash
  tables it maintains.  (This probably does not matter for well-
  -written code.  It no doubt does for C++ code that overuses
  destructors.)
- Added typed allocation primitives.  Rewrote the marker to
  accommodate them with more reasonable efficiency.  This
  change should also speed up marking for GC_malloc allocated
  objects a little.  See gc_typed.h for new primitives.
- Improved debugging facilities slightly.  Allocation time
  stack traces are now kept by default on SPARC/SUNOS4.
  (Thanks to Scott Schwartz.)
- Added better support for small heap applications.
- Significantly extended cord package.  Fixed a bug in the
  implementation of lazily read files.  Printf and friends now
  have cord variants.  Cord traversals are a bit faster.
- Made ALL_INTERIOR_POINTERS recognition the default.
- Fixed de so that it can run in constant space, independent
  of file size.  Added simple string searching to cords and de.
- Added the Hull-Ellis C++ interface.
- Added dynamic library support for OSF/1.
  (Thanks to Al Dosser and Tim Bingham at DEC.)
- Changed argument to GC_expand_hp to be expressed
  in units of bytes instead of heap blocks.  (Necessary
  since the heap block size now varies depending on
  configuration.  The old version was never very clean.)
- Added GC_get_heap_size().  The previous "equivalent"
  was broken.
- Restructured the Makefile a bit.  

Since version 4.0:
- Changed finalization implementation to guarantee that
  finalization procedures are called outside of the allocation
  lock, making direct use of the interface a little less dangerous.
  MAY BREAK EXISTING CLIENTS that assume finalizers
  are protected by a lock.  Since there seem to be few multithreaded
  clients that use finalization, this is hopefully not much of
  a problem.
- Fixed a gross bug in CORD_prev.
- Fixed a bug in blacklst.c that could result in unbounded
  heap growth during startup on machines that do not clear
  memory obtained from the OS (e.g. win32S).
- Ported de editor to win32/win32S.  (This is now the only
  version with a mouse-sensitive UI.)
- Added GC_malloc_ignore_off_page to allocate large arrays
  in the presence of ALL_INTERIOR_POINTERS.
- Changed GC_call_with_alloc_lock to not disable signals in
  the single-threaded case.
- Reduced retry count in GC_collect_or_expand for garbage
  collecting when out of memory.
- Made uncollectable allocations bypass black-listing, as they
  should.
- Fixed a bug in typed_test in test.c that could cause (legitimate)
  GC crashes.
- Fixed some potential synchronization problems in finalize.c
- Fixed a real locking problem in typd_mlc.c.
- Worked around an AIX 3.2 compiler feature that results in
  out of bounds memory references.
- Partially worked around an IRIX5.2 beta problem (which may
  or may not persist to the final release).
- Fixed a bug in the heap integrity checking code that could
  result in explicitly deallocated objects being identified as
  smashed.  Fixed a bug in the dbg_mlc stack saving code
  that caused old argument pointers to be considered live.
- Fixed a bug in CORD_ncmp (and hence CORD_str).
- Repaired the OS2 port, which had suffered from bit rot
  in 4.0.  Worked around what appears to be CSet/2 V1.0
  optimizer bug.
- Fixed a Makefile bug for target "c++".

Since version 4.1:
- Multiple bug fixes/workarounds in the Solaris threads version.
  (It occasionally failed to locate some register contents for
  marking.  It also turns out that thr_suspend and friends are
  unreliable in Solaris 2.3.  Dirty bit reads appear
  to be unreliable under some weird 
  circumstances.  My stack marking code
  contained a serious performance bug.  The new code is
  extremely defensive, and has not failed in several cpu
  hours of testing.  But  no guarantees ...)
- Added MacOS support (thanks to Patrick Beard.)
- Fixed several syntactic bugs in gc_c++.h and friends.  (These
  didn't bother g++, but did bother most other compilers.)
  Fixed gc_c++.h finalization interface.  (It didn't.)
- 64 bit alignment for allocated objects was not guaranteed in a
  few cases in which it should have been.
- Added GC_malloc_atomic_ignore_off_page.
- Added GC_collect_a_little.
- Added some prototypes to gc.h.
- Some other minor bug fixes (notably in Makefile).
- Fixed OS/2 / EMX port (thanks to Ari Huttunen).
- Fixed AmigaDOS port. (thanks to Michel Schinz).
- Fixed the DATASTART definition under Solaris.  There
  was a 1 in 16K chance of the collector missing the first
  64K of static data (and thus crashing).
- Fixed some blatant anachronisms in the README file.
- Fixed PCR-Makefile for upcoming PPCR release.

Since version 4.2:
- Fixed SPARC alignment problem with GC_DEBUG.
- Fixed Solaris threads /proc workaround.  The real
  problem was an interaction with mprotect.
- Incorporated fix from Patrick Beard for gc_c++.h (now gc_cpp.h).
- Slightly improved allocator space utilization by
  fixing the GC_size_map mechanism.
- Integrated some Sony News and MIPS RISCos 4.51
  patches.  (Thanks to Nobuyuki Hikichi of
  Software Research Associates, Inc. Japan)
- Fixed HP_PA alignment problem.  (Thanks to
  xjam@cork.cs.berkeley.edu.)
- Added GC_same_obj and friends.  Changed GC_base
  to return 0 for pointers past the end of large objects.
  Improved GC_base performance with ALL_INTERIOR_POINTERS
  on machines with a slow integer mod operation.
  Added GC_PTR_ADD, GC_PTR_STORE, etc. to prepare
  for preprocessor.
- changed the default on most UNIX machines to be that
  signals are not disabled during critical GC operations.
  This is still ANSI-conforming, though somewhat dangerous
  in the presence of signal handlers. But the performance
  cost of the alternative is sometimes problematic.
  Can be changed back with a minor Makefile edit.
- renamed IS_STRING in gc.h, to CORD_IS_STRING, thus
  following my own naming convention.  Added the function
  CORD_to_const_char_star.
- Fixed a gross bug in GC_finalize.  Symptom: occasional
  address faults in that function.  (Thanks to Anselm
  Baird-Smith (Anselm.BairdSmith@inria.fr)
- Added port to ICL DRS6000 running DRS/NX.  Restructured
  things a bit to factor out common code, and remove obsolete
  code.  Collector should now run under SUNOS5 with either
  mprotect or /proc dirty bits.  (Thanks to Douglas Steel
  (doug@wg.icl.co.uk)).
- More bug fixes and workarounds for Solaris 2.X.  (These were
  mostly related to putting the collector in a dynamic library,
  which didn't really work before.  Also SOLARIS_THREADS
  didn't interact well with dl_open.)  Thanks to btlewis@eng.sun.com.
- Fixed a serious performance bug on the DEC Alpha.  The text
  segment was getting registered as part of the root set.
  (Amazingly, the result was still fast enough that the bug
  was not conspicuous.) The fix works on OSF/1, version 1.3.
  Hopefully it also works on other versions of OSF/1 ...
- Fixed a bug in GC_clear_roots.
- Fixed a bug in GC_generic_malloc_words_small that broke
  gc_inl.h.  (Reported by Antoine de Maricourt.  I broke it
  in trying to tweak the Mac port.) 
- Fixed some problems with cord/de under Linux.
- Fixed some cord problems, notably with CORD_riter4.
- Added DG/UX port.
  Thanks to Ben A. Mesander (ben@piglet.cr.usgs.gov)
- Added finalization registration routines with weaker ordering
  constraints.  (This is necessary for C++ finalization with
  multiple inheritance, since the compiler often adds self-cycles.)
- Filled the holes in the SCO port. (Thanks to Michael Arnoldus
  <chime@proinf.dk>.)
- John Ellis' additions to the C++ support:  From John:

* I completely rewrote the documentation in the interface gc_c++.h
(later renamed gc_cpp.h).  I've tried to make it both clearer and more
precise.

* The definition of accessibility now ignores pointers from an
finalizable object (an object with a clean-up function) to itself.
This allows objects with virtual base classes to be finalizable by the
collector.  Compilers typically implement virtual base classes using
pointers from an object to itself, which under the old definition of
accessibility prevented objects with virtual base classes from ever
being collected or finalized.

* gc_cleanup now includes gc as a virtual base.  This was enabled by
the change in the definition of accessibility.

* I added support for operator new[].  Since most (all?) compilers
don't yet support operator new[], it is conditionalized on
-DOPERATOR_NEW_ARRAY.  The code is untested, but its trivial and looks
correct.

* The test program test_gc_c++ (later renamed test_cpp.cc)
tries to test for the C++-specific functionality not tested by the
other programs.
- Added <unistd.h> include to misc.c.  (Needed for ppcr.)
- Added PowerMac port. (Thanks to Patrick Beard again.)
- Fixed "srcdir"-related Makefile problems.  Changed things so
  that all externally visible include files always appear in the
  include subdirectory of the source.  Made gc.h directly
  includable from C++ code.  (These were at Per
  Bothner's suggestion.)
- Changed Intel code to also mark from ebp (Kevin Warne's
  suggestion).
- Renamed C++ related files so they could live in a FAT
  file system. (Charles Fiterman's suggestion.)
- Changed Windows NT Makefile to include C++ support in
  gc.lib.  Added C++ test as Makefile target.
  
Since version 4.3:
 - ASM_CLEAR_CODE was erroneously defined for HP
   PA machines, resulting in a compile error.
 - Fixed OS/2 Makefile to create a library.  (Thanks to
   Mark Boulter (mboulter@vnet.ibm.com)).
 - Gc_cleanup objects didn't work if they were created on
   the stack.  Fixed.
 - One copy of Gc_cpp.h in the distribution was out of 
   synch, and failed to document some known compiler
   problems with explicit destructor invocation.  Partially
   fixed.  There are probably other compilers on which
   gc_cleanup is miscompiled.
 - Fixed Makefile to pass C compiler flags to C++ compiler.
 - Added Mac fixes.
 - Fixed os_dep.c to work around what appears to be
   a new and different VirtualQuery bug under newer
   versions of win32S.
 - GC_non_gc_bytes was not correctly maintained by
   GC_free.  Fixed.  Thanks to James Clark (jjc@jclark.com).
 - Added GC_set_max_heap_size.
 - Changed allocation code to ignore blacklisting if it is preventing
   use of a very large block of memory.  This has the advantage
   that naive code allocating very large objects is much more
   likely to work.  The downside is you might no
   longer find out that such code should really use
   GC_malloc_ignore_off_page.
 - Changed GC_printf under win32 to close and reopen the file
   between calls.  FAT file systems otherwise make the log file
   useless for debugging.
 - Added GC_try_to_collect and GC_get_bytes_since_gc.  These
   allow starting an abortable collection during idle times. 
   This facility does not require special OS support.  (Thanks to
   Michael Spertus of Geodesic Systems for suggesting this.  It was
   actually an easy addition.  Kumar Srikantan previously added a similar
   facility to a now ancient version of the collector.  At the time
   this was much harder, and the result was less convincing.)
 - Added some support for the Borland development environment.  (Thanks
   to John Ellis and Michael Spertus.)
 - Removed a misfeature from checksums.c that caused unexpected 
   heap growth.  (Thanks to Scott Schwartz.)
 - Changed finalize.c to call WARN if it encounters a finalization cycle.
   WARN is defined in gc_priv.h to write a message, usually to stdout.
   In many environments, this may be inappropriate.
 - Renamed NO_PARAMS in gc.h to GC_NO_PARAMS, thus adhering to my own
   naming convention.
 - Added GC_set_warn_proc to intercept warnings.
 - Fixed Amiga port. (Thanks to Michel Schinz (schinz@alphanet.ch).)
 - Fixed a bug in mark.c that could result in an access to unmapped
   memory from GC_mark_from_mark_stack on machines with unaligned
   pointers.
 - Fixed a win32 specific performance bug that could result in scanning of
   objects allocated with the system malloc.
 - Added REDIRECT_MALLOC.

Since version 4.4:
 - Fixed many minor and one major README bugs. (Thanks to Franklin Chen
   (chen@adi.com) for pointing out many of them.)
 - Fixed ALPHA/OSF/1 dynamic library support. (Thanks to Jonathan Bachrach
   (jonathan@harlequin.com)).
 - Added incremental GC support (MPROTECT_VDB) for Linux (with some
   help from Bruno Haible).
 - Altered SPARC recognition tests in gc.h and config.h (mostly as
   suggested by Fergus Henderson).
 - Added basic incremental GC support for win32, as implemented by
   Windows NT and Windows 95.  GC_enable_incremental is a noop
   under win32s, which doesn't implement enough of the VM interface.
 - Added -DLARGE_CONFIG.
 - Fixed GC_..._ignore_off_page to also function without
   -DALL_INTERIOR_POINTERS.
 - (Hopefully) fixed RS/6000 port.  (Only the test was broken.)
 - Fixed a performance bug in the nonincremental collector running
   on machines supporting incremental collection with MPROTECT_VDB
   (e.g. SunOS 4, DEC AXP).  This turned into a correctness bug under
   win32s with win32 incremental collection.  (Not all memory protection
   was disabled.)
 - Fixed some ppcr related bit rot.
 - Caused dynamic libraries to be unregistered before reregistering.
   The old way turned out to be a performance bug on some machines.
 - GC_root_size was not properly maintained under MSWIN32.
 - Added -DNO_DEBUGGING and GC_dump.
 - Fixed a couple of bugs arising with SOLARIS_THREADS +
   REDIRECT_MALLOC.
 - Added NetBSD/M68K port.  (Thanks to Peter Seebach
   <seebs@taniemarie.solon.com>.)
 - Fixed a serious realloc bug.  For certain object sizes, the collector
   wouldn't scan the expanded part of the object.  (Thanks to Clay Spence
   (cds@peanut.sarnoff.com) for noticing the problem, and helping me to
   track it down.)
   
Since version 4.5:
 - Added Linux ELF support.  (Thanks to Arrigo Triulzi <arrigo@ic.ac.uk>.)
 - GC_base crashed if it was called before any other GC_ routines.
   This could happen if a gc_cleanup object was allocated outside the heap
   before any heap allocation.
 - The heap expansion heuristic was not stable if all objects had finalization
   enabled.  Fixed finalize.c to count memory in finalization queue and
   avoid explicit deallocation.  Changed alloc.c to also consider this count.
   (This is still not recommended.  It's expensive if nothing else.)  Thanks
   to John Ellis for pointing this out.
 - GC_malloc_uncollectable(0) was broken.  Thanks to Phong Vo for pointing
   this out.
 - The collector didn't compile under Linux 1.3.X.  (Thanks to Fred Gilham for
   pointing this out.)  The current workaround is ugly, but expected to be
   temporary.
 - Fixed a formatting problem for SPARC stack traces.
 - Fixed some '=='s in os_dep.c that should have been assignments.
   Fortunately these were in code that should never be executed anyway.
   (Thanks to Fergus Henderson.)
 - Fixed the heap block allocator to only drop blacklisted blocks in small
   chunks.  Made BL_LIMIT self adjusting.  (Both of these were in response
   to heap growth observed by Paul Graham.)
 - Fixed the Metrowerks/68K Mac code to also mark from a6.  (Thanks
   to Patrick Beard.)
 - Significantly updated README.debugging.
 - Fixed some problems with longjmps out of signal handlers, especially under
   Solaris.  Added a workaround for the fact that siglongjmp doesn't appear to
   do the right thing with -lthread under Solaris.
 - Added MSDOS/djgpp port.  (Thanks to Mitch Harris  (maharri@uiuc.edu).)
 - Added "make reserved_namespace" and "make user_namespace".  The
   first renames ALL "GC_xxx" identifiers as "_GC_xxx".  The second is the
   inverse transformation.  Note that doing this is guaranteed to break all
   clients written for the other names.
 - descriptor field for kind NORMAL in GC_obj_kinds with ADD_BYTE_AT_END
   defined should be -ALIGNMENT not WORDS_TO_BYTES(-1).  This is
   a serious bug on machines with pointer alignment of less than a word.
 - GC_ignore_self_finalize_mark_proc didn't handle pointers to very near the
   end of the object correctly.  Caused failures of the C++ test on a DEC Alpha
   with g++.
 - gc_inl.h still had problems.  Partially fixed.  Added warnings at the
   beginning to hopefully specify the remaining dangers.
 - Added DATAEND definition to config.h.
 - Fixed some of the .h file organization.  Fixed "make floppy".
 
Since version 4.6:
 - Fixed some compilation problems with -DCHECKSUMS (thanks to Ian Searle)
 - Updated some Mac specific files to synchronize with Patrick Beard.
 - Fixed a serious bug for machines with non-word-aligned pointers.
   (Thanks to Patrick Beard for pointing out the problem.  The collector
   should fail almost any conceivable test immediately on such machines.)

Since version 4.7:
 - Changed a "comment" in a MacOS specific part of mach-dep.c that caused
   gcc to fail on other platforms.

Since version 4.8
 - More README.debugging fixes.
 - Objects ready for finalization, but not finalized in the same GC
   cycle, could be prematurely collected.  This occasionally happened
   in test_cpp.
 - Too little memory was obtained from the system for very large
   objects.  That could cause a heap explosion if these objects were
   not contiguous (e.g. under PCR), and too much of them was blacklisted.
 - Due to an improper initialization, the collector was too hesitant to
   allocate blacklisted objects immediately after system startup.
 - Moved GC_arrays from the data into the bss segment by not explicitly
   initializing it to zero.  This significantly
   reduces the size of executables, and probably avoids some disk accesses
   on program startup.  It's conceivable that it might break a port that I
   didn't test.
 - Fixed EMX_MAKEFILE to reflect the gc_c++.h to gc_cpp.h renaming which
   occurred a while ago.

Since 4.9:
 - Fixed a typo around a call to GC_collect_or_expand in alloc.c.  It broke
   handling of out of memory.  (Thanks to Patrick Beard for noticing.)

Since 4.10:
 - Rationalized (hopefully) GC_try_to_collect in an incremental collection
   environment.  It appeared to not handle a call while a collection was in
   progress, and was otherwise too conservative.
 - Merged GC_reclaim_or_delete_all into GC_reclaim_all to get rid of some
   code.
 - Added Patrick Beard's Mac fixes, with substantial completely untested
   modifications.
 - Fixed the MPROTECT_VDB code to deal with large pages and imprecise
   fault addresses (as on an UltraSPARC running Solaris 2.5).  Note that this
   was not a problem in the default configuration, which uses PROC_VDB.
 - The DEC Alpha assembly code needed to restore $gp between calls.
   Thanks to Fergus Henderson for tracking this down and supplying a
   patch.
 - The write command for "de" was completely broken for large files.
   I used the easiest portable fix, which involved changing the semantics
   so that f.new is written instead of overwriting f.  That's safer anyway.
 - Added README.solaris2 with a discussion of the possible problems of
   mixing the collector's sbrk allocation with malloc/realloc.
 - Changed the data segment starting address for SGI machines.  The
   old code failed under IRIX6.
 - Required double word alignment for MIPS.
 - Various minor fixes to remove warnings.
 - Attempted to fix some Solaris threads problems reported by Zhiying Chen.
   In particular, the collector could try to fork a thread with the
   world stopped as part of GC_thr_init.  It also failed to deal with
   the case in which the original thread terminated before the whole
   process did.
 - Added -DNO_EXECUTE_PERMISSION.  This has a major performance impact
   on the incremental collector under Irix, and perhaps under other
   operating systems.
 - Added some code to support allocating the heap with mmap.  This may
   be preferable under some circumstances.
 - Integrated dynamic library support for HP.
   (Thanks to Knut Tvedten <knuttv@ifi.uio.no>.)
 - Integrated James Clark's win32 threads support, and made a number
   of changes to it, many of which were suggested by Pontus Rydin.
   This is still not 100% solid.
 - Integrated Alistair Crooks' support for UTS4 running on an Amdahl
   370-class machine.
 - Fixed a serious bug in explicitly typed allocation.  Objects requiring
   large descriptors where handled in a way that usually resulted in
   a segmentation fault in the marker.  (Thanks to Jeremy Fitzhardinge
   for helping to track this down.)
 - Added partial support for GNU win32 development.  (Thanks to Fergus
   Henderson.)
 - Added optional support for Java-style finalization semantics.  (Thanks
   to Patrick Bridges.)  This is recommended only for Java implementations.
 - GC_malloc_uncollectable faulted instead of returning 0 when out of
   memory.  (Thanks to dan@math.uiuc.edu for noticing.)
 - Calls to GC_base before the collector was initialized failed on a
   DEC Alpha.  (Thanks to Matthew Flatt.)
 - Added base pointer checking to GC_REGISTER_FINALIZER in debugging
   mode, at the suggestion of Jeremy Fitzhardinge.
 - GC_debug_realloc failed for uncollectable objects.  (Thanks to
   Jeremy Fitzhardinge.)
 - Explicitly typed allocation could crash if it ran out of memory.
   (Thanks to Jeremy Fitzhardinge.)
 - Added minimal support for a DEC Alpha running Linux.
 - Fixed a problem with allocation of objects whose size overflowed
   ptrdiff_t.  (This now fails unconditionally, as it should.)
 - Added the beginning of Irix pthread support.
 - Integrated Xiaokun Zhu's fixes for djgpp 2.01.
 - Added SGI-style STL allocator support (gc_alloc.h).
 - Fixed a serious bug in README.solaris2.  Multithreaded programs must include
   gc.h with SOLARIS_THREADS defined.
 - Changed GC_free so it actually deallocates uncollectable objects.
   (Thanks to Peter Chubb for pointing out the problem.)
 - Added Linux ELF support for dynamic libararies.  (Thanks again to
   Patrick Bridges.)
 - Changed the Borland cc configuration so that the assembler is not
   required.
 - Fixed a bug in the C++ test that caused it to fail in 64-bit
   environments.

Since 4.11:
 - Fixed ElfW definition in dyn_load.c. (Thanks to Fergus Henderson.)
   This prevented the dynamic library support from compiling on some
   older ELF Linux systems.
 - Fixed UTS4 port (which I apparently mangled during the integration)
   (Thanks to again to Alistair Crooks.)
 - "Make C++" failed on Suns with SC4.0, due to a problem with "bool".
   Fixed in gc_priv.h.
 - Added more pieces for GNU win32.  (Thanks to Timothy N. Newsham.)
   The current state of things should suffice for at least some
   applications.
 - Changed the out of memory retry count handling as suggested by
   Kenjiro Taura.  (This matters only if GC_max_retries > 0, which
   is no longer the default.)
 - If a /proc read failed repeatedly, GC_written_pages was not updated
   correctly.  (Thanks to Peter Chubb for diagnosing this.)
 - Under unlikely circumstances, the allocator could infinite loop in
   an out of memory situation.  (Thanks again to Kenjiro Taura for
   identifying the problem and supplying a fix.)
 - Fixed a syntactic error in the DJGPP code.  (Thanks to Fergus
   Henderson for finding this by inspection.)  Also fixed a test program
   problem with DJGPP (Thanks to Peter Monks.)
 - Atomic uncollectable objects were not treated correctly by the
   incremental collector.  This resulted in weird log statistics and
   occasional performance problems.  (Thanks to Peter Chubb for pointing
   this out.)
 - Fixed some problems resulting from compilers that dont define
   __STDC__.  In this case void * and char * were used inconsistently
   in some cases.  (Void * should not have been used at all.  If
   you have an ANSI superset compiler that does not define __STDC__,
   please compile with -D__STDC__=0. Thanks to Manuel Serrano and others
   for pointing out the problem.)
 - Fixed a compilation problem on Irix with -n32 and -DIRIX_THREADS.
   Also fixed some other IRIX_THREADS problems which may or may not have
   had observable symptoms.
 - Fixed an HP PA compilation problem in dyn_load.c.  (Thanks to
   Philippe Queinnec.)
 - SEGV fault handlers sometimes did not get reset correctly.  (Thanks
   to David Pickens.)
 - Added a fix for SOLARIS_THREADS on Intel.  (Thanks again to David
   Pickens.)  This probably needs more work to become functional.
 - Fixed struct sigcontext_struct in os_dep.c for compilation under
   Linux 2.1.X. (Thanks to Fergus Henderson.)
 - Changed the DJGPP STACKBOTTOM and DATASTART values to those suggested
   by Kristian Kristensen.  These may still not be right, but it is
   it is likely to work more often than what was there before.  They may
   even be exactly right.
 - Added a #include <string.h> to test_cpp.cc.  This appears to help
   with HP/UX and gcc.  (Thanks to assar@sics.se.)
 - Version 4.11 failed to run in incremental mode on recent 64-bit Irix
   kernels.  This was a problem related to page unaligned heap segments.
   Changed the code to page align heap sections on all platforms.
   (I had mistakenly identified this as a kernel problem earlier.
   It was not.)
 - Version 4.11 did not make allocated storage executable, except on
   one or two platforms, due to a bug in a #if test.  (Thanks to Dave
   Grove for pointing this out.)
 - Added sparc_sunos4_mach_dep.s to support Sun's compilers under SunOS4.
 - Added GC_exclude_static_roots.
 - Fixed the object size mapping algorithm.  This shouldn't matter,
   but the old code was ugly.
 - Heap checking code could die if one of the allocated objects was
   larger than its base address.  (Unsigned underflow problem.  Thanks
   to Clay Spence for isolating the problem.)
 - Added RS6000 (AIX) dynamic library support and fixed STACK_BOTTOM.
   (Thanks to Fred Stearns.)
 - Added Fergus Henderson's patches for improved robustness with large
   heaps and lots of blacklisting.
 - Added Peter Chubb's changes to support Solaris Pthreads, to support
   MMAP allocation in Solaris, to allow Solaris to find dynamic libraries
   through /proc, to add malloc_typed_ignore_off_page, and a few other
   minor features and bug fixes.
 - The Solaris 2 port should not use sbrk.  I received confirmation from
   Sun that the use of sbrk and malloc in the same program is not
   supported.  The collector now defines USE_MMAP by default on Solaris.
 - Replaced the djgpp makefile with Gary Leavens' version.
 - Fixed MSWIN32 detection test.
 - Added Fergus Henderson's patches to allow putting the collector into
   a DLL under GNU win32.
 - Added Ivan V. Demakov's port to Watcom C on X86.
 - Added Ian Piumarta's Linux/PowerPC port.
 - On Brian Burton's suggestion added PointerFreeGC to the placement
   options in gc_cpp.h.  This is of course unsafe, and may be controversial.
   On the other hand, it seems to be needed often enough that it's worth
   adding as a standard facility.

Since 4.12:
 - Fixed a crucial bug in the Watcom port.  There was a redundant decl
   of GC_push_one in gc_priv.h.
 - Added FINALIZE_ON_DEMAND.
 - Fixed some pre-ANSI cc problems in test.c.
 - Removed getpagesize() use for Solaris.  It seems to be missing in one
   or two versions.
 - Fixed bool handling for SPARCCompiler version 4.2.
 - Fixed some files in include that had gotten unlinked from the main
   copy.
 - Some RS/6000 fixes (missing casts).  Thanks to Toralf Foerster.
 - Fixed several problems in GC_debug_realloc, affecting mostly the
   FIND_LEAK case.
 - GC_exclude_static_roots contained a buggy unsigned comparison to
   terminate a loop.  (Thanks to Wilson Ho.)
 - CORD_str failed if the substring occurred at the last possible position.
   (Only affects cord users.)
 - Fixed Linux code to deal with RedHat 5.0 and integrated Peter Bigot's
   os_dep.c code for dealing with various Linux versions.
 - Added workaround for Irix pthreads sigaction bug and possible signal
   misdirection problems.
Since alpha1:
 - Changed RS6000 STACKBOTTOM.
 - Integrated Patrick Beard's Mac changes.
 - Alpha1 didn't compile on Irix m.n, m < 6.
 - Replaced Makefile.dj with a new one from Gary Leavens.
 - Added Andrew Stitcher's changes to support SCO OpenServer.
 - Added PRINT_BLACK_LIST, to allow debugging of high densities of false
   pointers.
 - Added code to debug allocator to keep track of return address
   in GC_malloc caller, thus giving a bit more context.
 - Changed default behavior of large block allocator to more
   aggressively avoid fragmentation.  This is likely to slow down the
   collector when it succeeds at reducing space cost.
 - Integrated Fergus Henderson's CYGWIN32 changes.  They are untested,
   but needed for newer versions.
 - USE_MMAP had some serious bugs.  This caused the collector to fail
   consistently on Solaris with -DSMALL_CONFIG.
 - Added Linux threads support, thanks largely to Fergus Henderson.
Since alpha2:
 - Fixed more Linux threads problems.
 - Changed default GC_free_space_divisor to 3 with new large block allocation.
   (Thanks to Matthew Flatt for some measurements that suggest the old
   value sometimes favors space too much over time.)
 - More CYGWIN32 fixes.
 - Integrated Tyson-Dowd's Linux-M68K port.
 - Minor HP PA and DEC UNIX fixes from Fergus Henderson.
 - Integrated Christoffe Raffali's Linux-SPARC changes.
 - Allowed for one more GC fixup iteration after a full GC in incremental
   mode.  Some quick measurements suggested that this significantly
   reduces pause times even with smaller GC_RATE values.
 - Moved some more GC data structures into GC_arrays.  This decreases
   pause times and GC overhead, but makes debugging slightly less convenient.
 - Fixed namespace pollution problem ("excl_table").
 - Made GC_incremental a constant for -DSMALL_CONFIG, hopefully shrinking
   that slightly.
 - Added some win32 threads fixes.
 - Integrated Ivan Demakov and David Stes' Watcom fixes.
 - Various other minor fixes contributed by many people.
 - Renamed config.h to gcconfig.h, since config.h tends to be used for
   many other things.
 - Integrated Matthew Flatt's support for 68K MacOS "far globals".
 - Fixed up some of the dynamic library Makefile targets for consistency
   across platforms.
 - Fixed a USE_MMAP typo that caused out-of-memory handling to fail
   on Solaris.
 - Added code to test.c to test thread creation a bit more.
 - Integrated GC_win32_free_heap, as suggested by Ivan Demakov.
 - Fixed Solaris 2.7 stack base finding problem.  (This may actually
   have been done in an earlier alpha release.)
Since alpha3:
 - Fixed MSWIN32 recognition test, which interfered with cygwin.
 - Removed unnecessary gc_watcom.asm from distribution.  Removed
   some obsolete README.win32 text.
 - Added Alpha Linux incremental GC support.  (Thanks to Philipp Tomsich
   for code for retrieving the fault address in a signal handler.)
   Changed Linux signal handler context argument to be a pointer.
 - Took care of some new warnings generated by the 7.3 SGI compiler.
 - Integrated Phillip Musumeci's FreeBSD/ELF fixes.
 - -DIRIX_THREADS was broken with the -o32 ABI (typo in gc_priv.h>

Since 4.13:
 - Fixed GC_print_source_ptr to not use a prototype.
 - generalized CYGWIN test.
 - gc::new did the wrong thing with PointerFreeGC placement.
   (Thanks to Rauli Ruohonen.)
 - In the ALL_INTERIOR_POINTERS (default) case, some callee-save register
   values could fail to be scanned if the register was saved and
   reused in a GC frame.  This showed up in verbose mode with gctest
   compiled with an unreleased SGI compiler.  I vaguely recall an old
   bug report that may have been related.  The bug was probably quite old.
   (The problem was that the stack scanning could be deferred until
   after the relevant frame was overwritten, and the new save location
   might be outside the scanned area.  Fixed by more eager stack scanning.)
 - PRINT_BLACK_LIST had some problems.  A few source addresses were garbage.
 - Replaced Makefile.dj and added -I flags to cord make targets.
   (Thanks to Gary Leavens.)
 - GC_try_to_collect was broken with the nonincremental collector.
 - gc_cleanup destructors could pass the wrong address to
   GC_register_finalizer_ignore_self in the presence of multiple
   inheritance.  (Thanks to Darrell Schiebel.)
 - Changed PowerPC Linux stack finding code.

Since 4.14alpha1
 - -DSMALL_CONFIG did not work reliably with large (> 4K) pages.
   Recycling the mark stack during expansion could result in a size
   zero heap segment, which confused things.  (This was probably also an
   issue with the normal config and huge pages.)
 - Did more work to make sure that callee-save registers were scanned
   completely, even with the setjmp-based code.  Added USE_GENERIC_PUSH_REGS
   macro to facilitate testing on machines I have access to.
 - Added code to explicitly push register contents for win32 threads.
   This seems to be necessary.  (Thanks to Pierre de Rop.)

Since 4.14alpha2
 - changed STACKBOTTOM for DJGPP (Thanks to Salvador Eduardo Tropea).
 
Since 4.14
 - Reworked large block allocator.  Now uses multiple doubly linked free
   lists to approximate best fit.
 - Changed heap expansion heuristic.  Entirely free blocks are no longer
   counted towards the heap size.  This seems to have a major impact on
   heap size stability; the old version could expand the heap way too
   much in the presence of large block fragmentation.
 - added -DGC_ASSERTIONS and some simple assertions inside the collector.
   This is mainlyt for collector debugging.
 - added -DUSE_MUNMAP to allow the heap to shrink.  Suupported on only
   a few UNIX-like platforms for now.
 - added GC_dump_regions() for debugging of fragmentation issues.
 - Changed PowerPC pointer alignment under Linux to 4.  (This needs
   checking by someone who has one.  The suggestions came to me via a
   rather circuitous path.)
 - Changed the Linux/Alpha port to walk the data segment backwards until
   it encounters a SIGSEGV.  The old way to find the start of the data
   segment broke with a recent release.
 - cordxtra.c needed to call GC_REGISTER_FINALIZER instead of
   GC_register_finalizer, so that it would continue to work with GC_DEBUG.
 - allochblk sometimes cleared the wrong block for debugging purposes
   when it dropped blacklisted blocks.  This could result in spurious
   error reports with GC_DEBUG.
 - added MACOS X Server support.  (Thanks to Andrew Stone.)
 - Changed the Solaris threads code to ignore stack limits > 8 MB with
   a warning.  Empirically, it is not safe to access arbitrary pages
   in such large stacks.  And the dirty bit implementation does not
   guarantee that none of them will be accessed.
 - Integrated Martin Tauchmann's Amiga changes.
 - Integrated James Dominy's OpenBSD/SPARC port.

Since 5.0alpha1
 - Fixed bugs introduced in alpha1 (OpenBSD & large block initialization).
 - Added -DKEEP_BACK_PTRS and backptr.h interface.  (The implementation
   idea came from Al Demers.)

Since 5.0alpha2
 - Added some highly incomplete code to support a copied young generation.
   Comments on nursery.h are appreciated.
 - Changed -DFIND_LEAK, -DJAVA_FINALIZATION, and -DFINALIZE_ON_DEMAND,
   so the same effect could be obtained with a runtime switch.   This is
   a step towards standardizing on a single dynamic GC library.
 - Significantly changed the way leak detection is handled, as a consequence
   of the above.

To do:
 - Very large root set sizes (> 16 MB or so) could cause the collector
   to abort with an unexpected mark stack overflow.  (Thanks again to
   Peter Chubb.)  NOT YET FIXED.  Workaround is to increase the initial
   size.
 - The SGI version of the collector marks from mmapped pages, even
   if they are not part of dynamic library static data areas.  This
   causes performance problems with some SGI libraries that use mmap
   as a bitmap allocator.  NOT YET FIXED.  It may be possible to turn
   off DYNAMIC_LOADING in the collector as a workaround.  It may also
   be possible to conditionally intercept mmap and use GC_exclude_static_roots.
   The real fix is to walk rld data structures, which looks possible.
 - Integrate MIT and DEC pthreads ports.
 - Deal with very uneven black-listing distributions.  If all the black listed
   blocks reside in the newly allocated heap section, the heuristic for
   temporarily ignoring black-listing fails, and the heap grows too much.
   (This was observed in only one case, and could be worked around, but ...)
 - Some platform specific updates are waiting for 4.15alpha1.