gcc/boehm-gc/doc/README

Copyright (c) 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.
Copyright (c) 1999-2001 by Hewlett-Packard Company. All rights reserved.

The file linux_threads.c is also
Copyright (c) 1998 by Fergus Henderson.  All rights reserved.

The files Makefile.am, and configure.in are
Copyright (c) 2001 by Red Hat Inc. All rights reserved.

The files config.guess and a few others are copyrighted by the Free
Software Foundation.

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.

A few of the files needed to use the GNU-style build procedure come with
slightly different licenses, though they are all similar in spirit.  A few
are GPL'ed, but with an exception that should cover all uses in the
collector.  (If you are concerned about such things, I recommend you look
at the notice in config.guess or ltmain.sh.)

This is version 6.1alpha3 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

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.

Boehm H., "Reducing Garbage Collector Cache Misses", Proceedings of the
2000 International Symposium on Memory Management.

  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.

(Some of these are also available from
http://www.hpl.hp.com/personal/Hans_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.)  On the most common platforms, it can be built
with or without thread support.  On a few platforms, it can take advantage
of a multiprocessor to speed up garbage collection.

  Many of the ideas underlying the collector have previously been explored
by others.  Notably, some of the run-time systems developed at Xerox PARC
in the early 1980s conservatively scanned thread stacks to locate possible
pointers (cf. Paul Rovner, "On Adding Garbage Collection and Runtime Types
to a Strongly-Typed Statically Checked, Concurrent Language"  Xerox PARC
CSL 84-7).  Doug McIlroy wrote a simpler fully conservative collector that
was part of version 8 UNIX (tm), but appears to not have received
widespread use.

  Rudimentary tools for use of the collector as a leak detector are included
(see http://www.hpl.hp.com/personal/Hans_Boehm/gc/leak.html),
as is a fairly sophisticated string package "cord" that makes use of the
collector.  (See doc/README.cords and H.-J. Boehm, R. Atkinson, and M. Plass,
"Ropes: An Alternative to Strings", Software Practice and Experience 25, 12
(December 1995), pp. 1315-1330.  This is very similar to the "rope" package
in Xerox Cedar, or the "rope" package in the SGI STL or the g++ distribution.)

Further collector documantation can be found at

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


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.

  WARNING: 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.)

  WARNING: 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.
If you're writing a program that depends on the collector scanning
dynamic library data areas, it may be a good idea to include at least
one call to GC_is_visible() to ensure that those areas are visible
to 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.
Like standard ANSI C mallocs, by default it is unsafe to invoke
malloc (and other GC routines) from a signal handler while another
malloc call may be in progress. Removing -DNO_SIGNALS from Makefile
attempts to remedy that.  But that may not be reliable with a compiler that
substantially reorders memory operations inside GC_malloc.

  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.)
WARNING: the collector does not guarantee to scan thread-local storage
(e.g. of the kind accessed with pthread_getspecific()).  The collector
does scan thread stacks, though, so generally the best solution is to
ensure that any pointers stored in thread-local storage are also
stored on the thread's stack for the duration of their lifetime.
(This is arguably a longstanding bug, but it hasn't been fixed yet.)

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. It may use up to 8 MB of memory.  (The
multi-threaded version will use more.  64-bit versions may 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, but
the garbage collector itself should still be usable.

  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
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),
Linux, 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.  Such objects will
also be automatically reclaimed.
  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).