[BACK]Return to README.changes CVS log [TXT][DIR] Up to [local] / OpenXM_contrib2 / asir2000 / gc / doc

Annotation of OpenXM_contrib2/asir2000/gc/doc/README.changes, Revision 1.3

1.1       noro        1: This is a rough history of garbage collector bugs and versions.
                      2:
                      3: This has been maintained with varying diligence over the years.
                      4:
                      5: I made an attempt to include recent contributors here.  I apologize for any
                      6: omissions.
                      7:
                      8: -------------------------
                      9:
                     10:   Version 1.3 and immediately preceding versions contained spurious
                     11: assembly language assignments to TMP_SP.  Only the assignment in the PC/RT
                     12: code is necessary.  On other machines, with certain compiler options,
                     13: the assignments can lead to an unsaved register being overwritten.
                     14: Known to cause problems under SunOS 3.5 WITHOUT the -O option.  (With
                     15: -O the compiler recognizes it as dead code.  It probably shouldn't,
                     16: but that's another story.)
                     17:
                     18:   Version 1.4 and earlier versions used compile time determined values
                     19: for the stack base.  This no longer works on Sun 3s, since Sun 3/80s use
                     20: a different stack base.  We now use a straightforward heuristic on all
                     21: machines on which it is known to work (incl. Sun 3s) and compile-time
                     22: determined values for the rest.  There should really be library calls
                     23: to determine such values.
                     24:
                     25:   Version 1.5 and earlier did not ensure 8 byte alignment for objects
                     26: allocated on a sparc based machine.
                     27:
                     28:   Version 1.8 added ULTRIX support in gc_private.h.
                     29:
                     30:   Version 1.9 fixed a major bug in gc_realloc.
                     31:
                     32:   Version 2.0 introduced a consistent naming convention for collector
                     33: routines and added support for registering dynamic library data segments
                     34: in the standard mark_roots.c.  Most of the data structures were revamped.
                     35: The treatment of interior pointers was completely changed.  Finalization
                     36: was added.  Support for locking was added.  Object kinds were added.
                     37: We added a black listing facility to avoid allocating at addresses known
                     38: to occur as integers somewhere in the address space.  Much of this
                     39: was accomplished by adapting ideas and code from the PCR collector.
                     40: The test program was changed and expanded.
                     41:
                     42:   Version 2.1 was the first stable version since 1.9, and added support
                     43: for PPCR.
                     44:
                     45:   Version 2.2 added debugging allocation, and fixed various bugs.  Among them:
                     46: - GC_realloc could fail to extend the size of the object for certain large object sizes.
                     47: - A blatant subscript range error in GC_printf, which unfortunately
                     48:   wasn't exercised on machines with sufficient stack alignment constraints.
                     49: - GC_register_displacement did the wrong thing if it was called after
                     50:   any allocation had taken place.
                     51: - The leak finding code would eventually break after 2048 byte
                     52:   byte objects leaked.
                     53: - interface.c didn't compile.
                     54: - The heap size remained much too small for large stacks.
                     55: - The stack clearing code behaved badly for large stacks, and perhaps
                     56:   on HP/PA machines.
                     57:
                     58:   Version 2.3 added ALL_INTERIOR_POINTERS and fixed the following bugs:
                     59: - Missing declaration of etext in the A/UX version.
                     60: - Some PCR root-finding problems.
                     61: - Blacklisting was not 100% effective, because the plausible future
                     62:   heap bounds were being miscalculated.
                     63: - GC_realloc didn't handle out-of-memory correctly.
                     64: - GC_base could return a nonzero value for addresses inside free blocks.
                     65: - test.c wasn't really thread safe, and could erroneously report failure
                     66:   in a multithreaded environment.  (The locking primitives need to be
                     67:   replaced for other threads packages.)
                     68: - GC_CONS was thoroughly broken.
                     69: - On a SPARC with dynamic linking, signals stayed diabled while the
                     70:   client code was running.
                     71:   (Thanks to Manuel Serrano at INRIA for reporting the last two.)
                     72:
                     73:   Version 2.4 added GC_free_space_divisor as a tuning knob, added
                     74:   support for OS/2 and linux, and fixed the following bugs:
                     75: - On machines with unaligned pointers (e.g. Sun 3), every 128th word could
                     76:   fail to be considered for marking.
                     77: - Dynamic_load.c erroneously added 4 bytes to the length of the data and
                     78:   bss sections of the dynamic library.  This could result in a bad memory
                     79:   reference if the actual length was a multiple of a page.  (Observed on
                     80:   Sun 3.  Can probably also happen on a Sun 4.)
                     81:   (Thanks to Robert Brazile for pointing out that the Sun 3 version
                     82:   was broken.  Dynamic library handling is still broken on Sun 3s
                     83:   under 4.1.1U1, but apparently not 4.1.1.  If you have such a machine,
                     84:   use -Bstatic.)
                     85:
                     86:   Version 2.5 fixed the following bugs:
                     87: - Removed an explicit call to exit(1)
                     88: - Fixed calls to GC_printf and GC_err_printf, so the correct number of
                     89:   arguments are always supplied.  The OS/2 C compiler gets confused if
                     90:   the number of actuals and the number of formals differ.  (ANSI C
                     91:   doesn't require this to work.  The ANSI sanctioned way of doing things
                     92:   causes too many compatibility problems.)
                     93:
                     94:   Version 3.0  added generational/incremental collection and stubborn
                     95:   objects.
                     96:
                     97:   Version 3.1 added the following features:
                     98: - A workaround for a SunOS 4.X SPARC C compiler
                     99:   misfeature that caused problems when the collector was turned into
                    100:   a dynamic library.
                    101: - A fix for a bug in GC_base that could result in a memory fault.
                    102: - A fix for a performance bug (and several other misfeatures) pointed
                    103:   out by Dave Detlefs and Al Dosser.
                    104: - Use of dirty bit information for static data under Solaris 2.X.
                    105: - DEC Alpha/OSF1 support (thanks to Al Dosser).
                    106: - Incremental collection on more platforms.
                    107: - A more refined heap expansion policy.  Less space usage by default.
                    108: - Various minor enhancements to reduce space usage, and to reduce
                    109:   the amount of memory scanned by the collector.
                    110: - Uncollectable allocation without per object overhead.
                    111: - More conscientious handling of out-of-memory conditions.
                    112: - Fixed a bug in debugging stubborn allocation.
                    113: - Fixed a bug that resulted in occasional erroneous reporting of smashed
                    114:   objects with debugging allocation.
                    115: - Fixed bogus leak reports of size 4096 blocks with FIND_LEAK.
                    116:
                    117:   Version 3.2 fixed a serious and not entirely repeatable bug in
                    118:   the incremental collector.  It appeared only when dirty bit info
                    119:   on the roots was available, which is normally only under Solaris.
                    120:   It also added GC_general_register_disappearing_link, and some
                    121:   testing code.  Interface.c disappeared.
                    122:
                    123:   Version 3.3 fixes several bugs and adds new ports:
                    124: - PCR-specific bugs.
                    125: - Missing locking in GC_free, redundant FASTUNLOCK
                    126:   in GC_malloc_stubborn, and 2 bugs in
                    127:   GC_unregister_disappearing_link.
                    128:   All of the above were pointed out by Neil Sharman
                    129:   (neil@cs.mu.oz.au).
                    130: - Common symbols allocated by the SunOS4.X dynamic loader
                    131:   were not included in the root set.
                    132: - Bug in GC_finalize (reported by Brian Beuning and Al Dosser)
                    133: - Merged Amiga port from Jesper Peterson (untested)
                    134: - Merged NeXT port from Thomas Funke (significantly
                    135:   modified and untested)
                    136:
                    137:   Version 3.4:
                    138: - Fixed a performance bug in GC_realloc.
                    139: - Updated the amiga port.
                    140: - Added NetBSD and 386BSD ports.
                    141: - Added cord library.
                    142: - Added trivial performance enhancement for
                    143:   ALL_INTERIOR_POINTERS.  (Don't scan last word.)
                    144:
                    145:   Version 3.5
                    146: - Minor collections now mark from roots only once, if that
                    147:   doesn't cause an excessive pause.
                    148: - The stack clearing heuristic was refined to prevent anomalies
                    149:   with very heavily recursive programs and sparse stacks.
                    150: - Fixed a bug that prevented mark stack growth in some cases.
                    151:   GC_objects_are_marked should be set to TRUE after a call
                    152:   to GC_push_roots and as part of GC_push_marked, since
                    153:   both can now set mark bits.  I think this is only a performance
                    154:   bug, but I wouldn't bet on it.  It's certainly very hard to argue
                    155:   that the old version was correct.
                    156: - Fixed an incremental collection bug that prevented it from
                    157:   working at all when HBLKSIZE != getpagesize()
                    158: - Changed dynamic_loading.c to include gc_priv.h before testing
                    159:   DYNAMIC_LOADING.  SunOS dynamic library scanning
                    160:   must have been broken in 3.4.
                    161: - Object size rounding now adapts to program behavior.
                    162: - Added a workaround (provided by Manuel Serrano and
                    163:   colleagues) to a long-standing SunOS 4.X (and 3.X?) ld bug
                    164:   that I had incorrectly assumed to have been squished.
                    165:   The collector was broken if the text segment size was within
                    166:   32 bytes of a multiple of 8K bytes, and if the beginning of
                    167:   the data segment contained interesting roots.  The workaround
                    168:   assumes a demand-loadable executable.  The original may have
                    169:   have "worked" in some other cases.
                    170: - Added dynamic library support under IRIX5.
                    171: - Added support for EMX under OS/2 (thanks to Ari Huttunen).
                    172:
                    173: Version 3.6:
                    174: - fixed a bug in the mark stack growth code that was introduced
                    175:   in 3.4.
                    176: - fixed Makefile to work around DEC AXP compiler tail recursion
                    177:   bug.
                    178:
                    179: Version 3.7:
                    180: - Added a workaround for an HP/UX compiler bug.
                    181: - Fixed another stack clearing performance bug.  Reworked
                    182:   that code once more.
                    183:
                    184: Version 4.0:
                    185: - Added support for Solaris threads (which was possible
                    186:   only by reimplementing some fraction of Solaris threads,
                    187:   since Sun doesn't currently make the thread debugging
                    188:   interface available).
                    189: - Added non-threads win32 and win32S support.
                    190: - (Grudgingly, with suitable muttering of obscenities) renamed
                    191:   files so that the collector distribution could live on a FAT
                    192:   file system.  Files that are guaranteed to be useless on
                    193:   a PC still have long names.  Gc_inline.h and gc_private.h
                    194:   still exist, but now just include  gc_inl.h and gc_priv.h.
                    195: - Fixed a really obscure bug in finalization that could cause
                    196:   undetected mark stack overflows.  (I would be surprised if
                    197:   any real code ever tickled this one.)
                    198: - Changed finalization code to dynamically resize the hash
                    199:   tables it maintains.  (This probably does not matter for well-
                    200:   -written code.  It no doubt does for C++ code that overuses
                    201:   destructors.)
                    202: - Added typed allocation primitives.  Rewrote the marker to
                    203:   accommodate them with more reasonable efficiency.  This
                    204:   change should also speed up marking for GC_malloc allocated
                    205:   objects a little.  See gc_typed.h for new primitives.
                    206: - Improved debugging facilities slightly.  Allocation time
                    207:   stack traces are now kept by default on SPARC/SUNOS4.
                    208:   (Thanks to Scott Schwartz.)
                    209: - Added better support for small heap applications.
                    210: - Significantly extended cord package.  Fixed a bug in the
                    211:   implementation of lazily read files.  Printf and friends now
                    212:   have cord variants.  Cord traversals are a bit faster.
                    213: - Made ALL_INTERIOR_POINTERS recognition the default.
                    214: - Fixed de so that it can run in constant space, independent
                    215:   of file size.  Added simple string searching to cords and de.
                    216: - Added the Hull-Ellis C++ interface.
                    217: - Added dynamic library support for OSF/1.
                    218:   (Thanks to Al Dosser and Tim Bingham at DEC.)
                    219: - Changed argument to GC_expand_hp to be expressed
                    220:   in units of bytes instead of heap blocks.  (Necessary
                    221:   since the heap block size now varies depending on
                    222:   configuration.  The old version was never very clean.)
                    223: - Added GC_get_heap_size().  The previous "equivalent"
                    224:   was broken.
                    225: - Restructured the Makefile a bit.
                    226:
                    227: Since version 4.0:
                    228: - Changed finalization implementation to guarantee that
                    229:   finalization procedures are called outside of the allocation
                    230:   lock, making direct use of the interface a little less dangerous.
                    231:   MAY BREAK EXISTING CLIENTS that assume finalizers
                    232:   are protected by a lock.  Since there seem to be few multithreaded
                    233:   clients that use finalization, this is hopefully not much of
                    234:   a problem.
                    235: - Fixed a gross bug in CORD_prev.
                    236: - Fixed a bug in blacklst.c that could result in unbounded
                    237:   heap growth during startup on machines that do not clear
                    238:   memory obtained from the OS (e.g. win32S).
                    239: - Ported de editor to win32/win32S.  (This is now the only
                    240:   version with a mouse-sensitive UI.)
                    241: - Added GC_malloc_ignore_off_page to allocate large arrays
                    242:   in the presence of ALL_INTERIOR_POINTERS.
                    243: - Changed GC_call_with_alloc_lock to not disable signals in
                    244:   the single-threaded case.
                    245: - Reduced retry count in GC_collect_or_expand for garbage
                    246:   collecting when out of memory.
                    247: - Made uncollectable allocations bypass black-listing, as they
                    248:   should.
                    249: - Fixed a bug in typed_test in test.c that could cause (legitimate)
                    250:   GC crashes.
                    251: - Fixed some potential synchronization problems in finalize.c
                    252: - Fixed a real locking problem in typd_mlc.c.
                    253: - Worked around an AIX 3.2 compiler feature that results in
                    254:   out of bounds memory references.
                    255: - Partially worked around an IRIX5.2 beta problem (which may
                    256:   or may not persist to the final release).
                    257: - Fixed a bug in the heap integrity checking code that could
                    258:   result in explicitly deallocated objects being identified as
                    259:   smashed.  Fixed a bug in the dbg_mlc stack saving code
                    260:   that caused old argument pointers to be considered live.
                    261: - Fixed a bug in CORD_ncmp (and hence CORD_str).
                    262: - Repaired the OS2 port, which had suffered from bit rot
                    263:   in 4.0.  Worked around what appears to be CSet/2 V1.0
                    264:   optimizer bug.
                    265: - Fixed a Makefile bug for target "c++".
                    266:
                    267: Since version 4.1:
                    268: - Multiple bug fixes/workarounds in the Solaris threads version.
                    269:   (It occasionally failed to locate some register contents for
                    270:   marking.  It also turns out that thr_suspend and friends are
                    271:   unreliable in Solaris 2.3.  Dirty bit reads appear
                    272:   to be unreliable under some weird
                    273:   circumstances.  My stack marking code
                    274:   contained a serious performance bug.  The new code is
                    275:   extremely defensive, and has not failed in several cpu
                    276:   hours of testing.  But  no guarantees ...)
                    277: - Added MacOS support (thanks to Patrick Beard.)
                    278: - Fixed several syntactic bugs in gc_c++.h and friends.  (These
                    279:   didn't bother g++, but did bother most other compilers.)
                    280:   Fixed gc_c++.h finalization interface.  (It didn't.)
                    281: - 64 bit alignment for allocated objects was not guaranteed in a
                    282:   few cases in which it should have been.
                    283: - Added GC_malloc_atomic_ignore_off_page.
                    284: - Added GC_collect_a_little.
                    285: - Added some prototypes to gc.h.
                    286: - Some other minor bug fixes (notably in Makefile).
                    287: - Fixed OS/2 / EMX port (thanks to Ari Huttunen).
                    288: - Fixed AmigaDOS port. (thanks to Michel Schinz).
                    289: - Fixed the DATASTART definition under Solaris.  There
                    290:   was a 1 in 16K chance of the collector missing the first
                    291:   64K of static data (and thus crashing).
                    292: - Fixed some blatant anachronisms in the README file.
                    293: - Fixed PCR-Makefile for upcoming PPCR release.
                    294:
                    295: Since version 4.2:
                    296: - Fixed SPARC alignment problem with GC_DEBUG.
                    297: - Fixed Solaris threads /proc workaround.  The real
                    298:   problem was an interaction with mprotect.
                    299: - Incorporated fix from Patrick Beard for gc_c++.h (now gc_cpp.h).
                    300: - Slightly improved allocator space utilization by
                    301:   fixing the GC_size_map mechanism.
                    302: - Integrated some Sony News and MIPS RISCos 4.51
                    303:   patches.  (Thanks to Nobuyuki Hikichi of
                    304:   Software Research Associates, Inc. Japan)
                    305: - Fixed HP_PA alignment problem.  (Thanks to
                    306:   xjam@cork.cs.berkeley.edu.)
                    307: - Added GC_same_obj and friends.  Changed GC_base
                    308:   to return 0 for pointers past the end of large objects.
                    309:   Improved GC_base performance with ALL_INTERIOR_POINTERS
                    310:   on machines with a slow integer mod operation.
                    311:   Added GC_PTR_ADD, GC_PTR_STORE, etc. to prepare
                    312:   for preprocessor.
                    313: - changed the default on most UNIX machines to be that
                    314:   signals are not disabled during critical GC operations.
                    315:   This is still ANSI-conforming, though somewhat dangerous
                    316:   in the presence of signal handlers. But the performance
                    317:   cost of the alternative is sometimes problematic.
                    318:   Can be changed back with a minor Makefile edit.
                    319: - renamed IS_STRING in gc.h, to CORD_IS_STRING, thus
                    320:   following my own naming convention.  Added the function
                    321:   CORD_to_const_char_star.
                    322: - Fixed a gross bug in GC_finalize.  Symptom: occasional
                    323:   address faults in that function.  (Thanks to Anselm
                    324:   Baird-Smith (Anselm.BairdSmith@inria.fr)
                    325: - Added port to ICL DRS6000 running DRS/NX.  Restructured
                    326:   things a bit to factor out common code, and remove obsolete
                    327:   code.  Collector should now run under SUNOS5 with either
                    328:   mprotect or /proc dirty bits.  (Thanks to Douglas Steel
                    329:   (doug@wg.icl.co.uk)).
                    330: - More bug fixes and workarounds for Solaris 2.X.  (These were
                    331:   mostly related to putting the collector in a dynamic library,
                    332:   which didn't really work before.  Also SOLARIS_THREADS
                    333:   didn't interact well with dl_open.)  Thanks to btlewis@eng.sun.com.
                    334: - Fixed a serious performance bug on the DEC Alpha.  The text
                    335:   segment was getting registered as part of the root set.
                    336:   (Amazingly, the result was still fast enough that the bug
                    337:   was not conspicuous.) The fix works on OSF/1, version 1.3.
                    338:   Hopefully it also works on other versions of OSF/1 ...
                    339: - Fixed a bug in GC_clear_roots.
                    340: - Fixed a bug in GC_generic_malloc_words_small that broke
                    341:   gc_inl.h.  (Reported by Antoine de Maricourt.  I broke it
                    342:   in trying to tweak the Mac port.)
                    343: - Fixed some problems with cord/de under Linux.
                    344: - Fixed some cord problems, notably with CORD_riter4.
                    345: - Added DG/UX port.
                    346:   Thanks to Ben A. Mesander (ben@piglet.cr.usgs.gov)
                    347: - Added finalization registration routines with weaker ordering
                    348:   constraints.  (This is necessary for C++ finalization with
                    349:   multiple inheritance, since the compiler often adds self-cycles.)
                    350: - Filled the holes in the SCO port. (Thanks to Michael Arnoldus
                    351:   <chime@proinf.dk>.)
                    352: - John Ellis' additions to the C++ support:  From John:
                    353:
                    354: * I completely rewrote the documentation in the interface gc_c++.h
                    355: (later renamed gc_cpp.h).  I've tried to make it both clearer and more
                    356: precise.
                    357:
                    358: * The definition of accessibility now ignores pointers from an
                    359: finalizable object (an object with a clean-up function) to itself.
                    360: This allows objects with virtual base classes to be finalizable by the
                    361: collector.  Compilers typically implement virtual base classes using
                    362: pointers from an object to itself, which under the old definition of
                    363: accessibility prevented objects with virtual base classes from ever
                    364: being collected or finalized.
                    365:
                    366: * gc_cleanup now includes gc as a virtual base.  This was enabled by
                    367: the change in the definition of accessibility.
                    368:
                    369: * I added support for operator new[].  Since most (all?) compilers
                    370: don't yet support operator new[], it is conditionalized on
                    371: -DOPERATOR_NEW_ARRAY.  The code is untested, but its trivial and looks
                    372: correct.
                    373:
                    374: * The test program test_gc_c++ (later renamed test_cpp.cc)
                    375: tries to test for the C++-specific functionality not tested by the
                    376: other programs.
                    377: - Added <unistd.h> include to misc.c.  (Needed for ppcr.)
                    378: - Added PowerMac port. (Thanks to Patrick Beard again.)
                    379: - Fixed "srcdir"-related Makefile problems.  Changed things so
                    380:   that all externally visible include files always appear in the
                    381:   include subdirectory of the source.  Made gc.h directly
                    382:   includable from C++ code.  (These were at Per
                    383:   Bothner's suggestion.)
                    384: - Changed Intel code to also mark from ebp (Kevin Warne's
                    385:   suggestion).
                    386: - Renamed C++ related files so they could live in a FAT
                    387:   file system. (Charles Fiterman's suggestion.)
                    388: - Changed Windows NT Makefile to include C++ support in
                    389:   gc.lib.  Added C++ test as Makefile target.
                    390:
                    391: Since version 4.3:
                    392:  - ASM_CLEAR_CODE was erroneously defined for HP
                    393:    PA machines, resulting in a compile error.
                    394:  - Fixed OS/2 Makefile to create a library.  (Thanks to
                    395:    Mark Boulter (mboulter@vnet.ibm.com)).
                    396:  - Gc_cleanup objects didn't work if they were created on
                    397:    the stack.  Fixed.
                    398:  - One copy of Gc_cpp.h in the distribution was out of
                    399:    synch, and failed to document some known compiler
                    400:    problems with explicit destructor invocation.  Partially
                    401:    fixed.  There are probably other compilers on which
                    402:    gc_cleanup is miscompiled.
                    403:  - Fixed Makefile to pass C compiler flags to C++ compiler.
                    404:  - Added Mac fixes.
                    405:  - Fixed os_dep.c to work around what appears to be
                    406:    a new and different VirtualQuery bug under newer
                    407:    versions of win32S.
                    408:  - GC_non_gc_bytes was not correctly maintained by
                    409:    GC_free.  Fixed.  Thanks to James Clark (jjc@jclark.com).
                    410:  - Added GC_set_max_heap_size.
                    411:  - Changed allocation code to ignore blacklisting if it is preventing
                    412:    use of a very large block of memory.  This has the advantage
                    413:    that naive code allocating very large objects is much more
                    414:    likely to work.  The downside is you might no
                    415:    longer find out that such code should really use
                    416:    GC_malloc_ignore_off_page.
                    417:  - Changed GC_printf under win32 to close and reopen the file
                    418:    between calls.  FAT file systems otherwise make the log file
                    419:    useless for debugging.
                    420:  - Added GC_try_to_collect and GC_get_bytes_since_gc.  These
                    421:    allow starting an abortable collection during idle times.
                    422:    This facility does not require special OS support.  (Thanks to
                    423:    Michael Spertus of Geodesic Systems for suggesting this.  It was
                    424:    actually an easy addition.  Kumar Srikantan previously added a similar
                    425:    facility to a now ancient version of the collector.  At the time
                    426:    this was much harder, and the result was less convincing.)
                    427:  - Added some support for the Borland development environment.  (Thanks
                    428:    to John Ellis and Michael Spertus.)
                    429:  - Removed a misfeature from checksums.c that caused unexpected
                    430:    heap growth.  (Thanks to Scott Schwartz.)
                    431:  - Changed finalize.c to call WARN if it encounters a finalization cycle.
                    432:    WARN is defined in gc_priv.h to write a message, usually to stdout.
                    433:    In many environments, this may be inappropriate.
                    434:  - Renamed NO_PARAMS in gc.h to GC_NO_PARAMS, thus adhering to my own
                    435:    naming convention.
                    436:  - Added GC_set_warn_proc to intercept warnings.
                    437:  - Fixed Amiga port. (Thanks to Michel Schinz (schinz@alphanet.ch).)
                    438:  - Fixed a bug in mark.c that could result in an access to unmapped
                    439:    memory from GC_mark_from_mark_stack on machines with unaligned
                    440:    pointers.
                    441:  - Fixed a win32 specific performance bug that could result in scanning of
                    442:    objects allocated with the system malloc.
                    443:  - Added REDIRECT_MALLOC.
                    444:
                    445: Since version 4.4:
                    446:  - Fixed many minor and one major README bugs. (Thanks to Franklin Chen
                    447:    (chen@adi.com) for pointing out many of them.)
                    448:  - Fixed ALPHA/OSF/1 dynamic library support. (Thanks to Jonathan Bachrach
                    449:    (jonathan@harlequin.com)).
                    450:  - Added incremental GC support (MPROTECT_VDB) for Linux (with some
                    451:    help from Bruno Haible).
                    452:  - Altered SPARC recognition tests in gc.h and config.h (mostly as
                    453:    suggested by Fergus Henderson).
                    454:  - Added basic incremental GC support for win32, as implemented by
                    455:    Windows NT and Windows 95.  GC_enable_incremental is a noop
                    456:    under win32s, which doesn't implement enough of the VM interface.
                    457:  - Added -DLARGE_CONFIG.
                    458:  - Fixed GC_..._ignore_off_page to also function without
                    459:    -DALL_INTERIOR_POINTERS.
                    460:  - (Hopefully) fixed RS/6000 port.  (Only the test was broken.)
                    461:  - Fixed a performance bug in the nonincremental collector running
                    462:    on machines supporting incremental collection with MPROTECT_VDB
                    463:    (e.g. SunOS 4, DEC AXP).  This turned into a correctness bug under
                    464:    win32s with win32 incremental collection.  (Not all memory protection
                    465:    was disabled.)
                    466:  - Fixed some ppcr related bit rot.
                    467:  - Caused dynamic libraries to be unregistered before reregistering.
                    468:    The old way turned out to be a performance bug on some machines.
                    469:  - GC_root_size was not properly maintained under MSWIN32.
                    470:  - Added -DNO_DEBUGGING and GC_dump.
                    471:  - Fixed a couple of bugs arising with SOLARIS_THREADS +
                    472:    REDIRECT_MALLOC.
                    473:  - Added NetBSD/M68K port.  (Thanks to Peter Seebach
                    474:    <seebs@taniemarie.solon.com>.)
                    475:  - Fixed a serious realloc bug.  For certain object sizes, the collector
                    476:    wouldn't scan the expanded part of the object.  (Thanks to Clay Spence
                    477:    (cds@peanut.sarnoff.com) for noticing the problem, and helping me to
                    478:    track it down.)
                    479:
                    480: Since version 4.5:
                    481:  - Added Linux ELF support.  (Thanks to Arrigo Triulzi <arrigo@ic.ac.uk>.)
                    482:  - GC_base crashed if it was called before any other GC_ routines.
                    483:    This could happen if a gc_cleanup object was allocated outside the heap
                    484:    before any heap allocation.
                    485:  - The heap expansion heuristic was not stable if all objects had finalization
                    486:    enabled.  Fixed finalize.c to count memory in finalization queue and
                    487:    avoid explicit deallocation.  Changed alloc.c to also consider this count.
                    488:    (This is still not recommended.  It's expensive if nothing else.)  Thanks
                    489:    to John Ellis for pointing this out.
                    490:  - GC_malloc_uncollectable(0) was broken.  Thanks to Phong Vo for pointing
                    491:    this out.
                    492:  - The collector didn't compile under Linux 1.3.X.  (Thanks to Fred Gilham for
                    493:    pointing this out.)  The current workaround is ugly, but expected to be
                    494:    temporary.
                    495:  - Fixed a formatting problem for SPARC stack traces.
                    496:  - Fixed some '=='s in os_dep.c that should have been assignments.
                    497:    Fortunately these were in code that should never be executed anyway.
                    498:    (Thanks to Fergus Henderson.)
                    499:  - Fixed the heap block allocator to only drop blacklisted blocks in small
                    500:    chunks.  Made BL_LIMIT self adjusting.  (Both of these were in response
                    501:    to heap growth observed by Paul Graham.)
                    502:  - Fixed the Metrowerks/68K Mac code to also mark from a6.  (Thanks
                    503:    to Patrick Beard.)
                    504:  - Significantly updated README.debugging.
                    505:  - Fixed some problems with longjmps out of signal handlers, especially under
                    506:    Solaris.  Added a workaround for the fact that siglongjmp doesn't appear to
                    507:    do the right thing with -lthread under Solaris.
                    508:  - Added MSDOS/djgpp port.  (Thanks to Mitch Harris  (maharri@uiuc.edu).)
                    509:  - Added "make reserved_namespace" and "make user_namespace".  The
                    510:    first renames ALL "GC_xxx" identifiers as "_GC_xxx".  The second is the
                    511:    inverse transformation.  Note that doing this is guaranteed to break all
                    512:    clients written for the other names.
                    513:  - descriptor field for kind NORMAL in GC_obj_kinds with ADD_BYTE_AT_END
                    514:    defined should be -ALIGNMENT not WORDS_TO_BYTES(-1).  This is
                    515:    a serious bug on machines with pointer alignment of less than a word.
                    516:  - GC_ignore_self_finalize_mark_proc didn't handle pointers to very near the
                    517:    end of the object correctly.  Caused failures of the C++ test on a DEC Alpha
                    518:    with g++.
                    519:  - gc_inl.h still had problems.  Partially fixed.  Added warnings at the
                    520:    beginning to hopefully specify the remaining dangers.
                    521:  - Added DATAEND definition to config.h.
                    522:  - Fixed some of the .h file organization.  Fixed "make floppy".
                    523:
                    524: Since version 4.6:
                    525:  - Fixed some compilation problems with -DCHECKSUMS (thanks to Ian Searle)
                    526:  - Updated some Mac specific files to synchronize with Patrick Beard.
                    527:  - Fixed a serious bug for machines with non-word-aligned pointers.
                    528:    (Thanks to Patrick Beard for pointing out the problem.  The collector
                    529:    should fail almost any conceivable test immediately on such machines.)
                    530:
                    531: Since version 4.7:
                    532:  - Changed a "comment" in a MacOS specific part of mach-dep.c that caused
                    533:    gcc to fail on other platforms.
                    534:
                    535: Since version 4.8
                    536:  - More README.debugging fixes.
                    537:  - Objects ready for finalization, but not finalized in the same GC
                    538:    cycle, could be prematurely collected.  This occasionally happened
                    539:    in test_cpp.
                    540:  - Too little memory was obtained from the system for very large
                    541:    objects.  That could cause a heap explosion if these objects were
                    542:    not contiguous (e.g. under PCR), and too much of them was blacklisted.
                    543:  - Due to an improper initialization, the collector was too hesitant to
                    544:    allocate blacklisted objects immediately after system startup.
                    545:  - Moved GC_arrays from the data into the bss segment by not explicitly
                    546:    initializing it to zero.  This significantly
                    547:    reduces the size of executables, and probably avoids some disk accesses
                    548:    on program startup.  It's conceivable that it might break a port that I
                    549:    didn't test.
                    550:  - Fixed EMX_MAKEFILE to reflect the gc_c++.h to gc_cpp.h renaming which
                    551:    occurred a while ago.
                    552:
                    553: Since 4.9:
                    554:  - Fixed a typo around a call to GC_collect_or_expand in alloc.c.  It broke
                    555:    handling of out of memory.  (Thanks to Patrick Beard for noticing.)
                    556:
                    557: Since 4.10:
                    558:  - Rationalized (hopefully) GC_try_to_collect in an incremental collection
                    559:    environment.  It appeared to not handle a call while a collection was in
                    560:    progress, and was otherwise too conservative.
                    561:  - Merged GC_reclaim_or_delete_all into GC_reclaim_all to get rid of some
                    562:    code.
                    563:  - Added Patrick Beard's Mac fixes, with substantial completely untested
                    564:    modifications.
                    565:  - Fixed the MPROTECT_VDB code to deal with large pages and imprecise
                    566:    fault addresses (as on an UltraSPARC running Solaris 2.5).  Note that this
                    567:    was not a problem in the default configuration, which uses PROC_VDB.
                    568:  - The DEC Alpha assembly code needed to restore $gp between calls.
                    569:    Thanks to Fergus Henderson for tracking this down and supplying a
                    570:    patch.
                    571:  - The write command for "de" was completely broken for large files.
                    572:    I used the easiest portable fix, which involved changing the semantics
                    573:    so that f.new is written instead of overwriting f.  That's safer anyway.
                    574:  - Added README.solaris2 with a discussion of the possible problems of
                    575:    mixing the collector's sbrk allocation with malloc/realloc.
                    576:  - Changed the data segment starting address for SGI machines.  The
                    577:    old code failed under IRIX6.
                    578:  - Required double word alignment for MIPS.
                    579:  - Various minor fixes to remove warnings.
                    580:  - Attempted to fix some Solaris threads problems reported by Zhiying Chen.
                    581:    In particular, the collector could try to fork a thread with the
                    582:    world stopped as part of GC_thr_init.  It also failed to deal with
                    583:    the case in which the original thread terminated before the whole
                    584:    process did.
                    585:  - Added -DNO_EXECUTE_PERMISSION.  This has a major performance impact
                    586:    on the incremental collector under Irix, and perhaps under other
                    587:    operating systems.
                    588:  - Added some code to support allocating the heap with mmap.  This may
                    589:    be preferable under some circumstances.
                    590:  - Integrated dynamic library support for HP.
                    591:    (Thanks to Knut Tvedten <knuttv@ifi.uio.no>.)
                    592:  - Integrated James Clark's win32 threads support, and made a number
                    593:    of changes to it, many of which were suggested by Pontus Rydin.
                    594:    This is still not 100% solid.
                    595:  - Integrated Alistair Crooks' support for UTS4 running on an Amdahl
                    596:    370-class machine.
                    597:  - Fixed a serious bug in explicitly typed allocation.  Objects requiring
                    598:    large descriptors where handled in a way that usually resulted in
                    599:    a segmentation fault in the marker.  (Thanks to Jeremy Fitzhardinge
                    600:    for helping to track this down.)
                    601:  - Added partial support for GNU win32 development.  (Thanks to Fergus
                    602:    Henderson.)
                    603:  - Added optional support for Java-style finalization semantics.  (Thanks
                    604:    to Patrick Bridges.)  This is recommended only for Java implementations.
                    605:  - GC_malloc_uncollectable faulted instead of returning 0 when out of
                    606:    memory.  (Thanks to dan@math.uiuc.edu for noticing.)
                    607:  - Calls to GC_base before the collector was initialized failed on a
                    608:    DEC Alpha.  (Thanks to Matthew Flatt.)
                    609:  - Added base pointer checking to GC_REGISTER_FINALIZER in debugging
                    610:    mode, at the suggestion of Jeremy Fitzhardinge.
                    611:  - GC_debug_realloc failed for uncollectable objects.  (Thanks to
                    612:    Jeremy Fitzhardinge.)
                    613:  - Explicitly typed allocation could crash if it ran out of memory.
                    614:    (Thanks to Jeremy Fitzhardinge.)
                    615:  - Added minimal support for a DEC Alpha running Linux.
                    616:  - Fixed a problem with allocation of objects whose size overflowed
                    617:    ptrdiff_t.  (This now fails unconditionally, as it should.)
                    618:  - Added the beginning of Irix pthread support.
                    619:  - Integrated Xiaokun Zhu's fixes for djgpp 2.01.
                    620:  - Added SGI-style STL allocator support (gc_alloc.h).
                    621:  - Fixed a serious bug in README.solaris2.  Multithreaded programs must include
                    622:    gc.h with SOLARIS_THREADS defined.
                    623:  - Changed GC_free so it actually deallocates uncollectable objects.
                    624:    (Thanks to Peter Chubb for pointing out the problem.)
                    625:  - Added Linux ELF support for dynamic libararies.  (Thanks again to
                    626:    Patrick Bridges.)
                    627:  - Changed the Borland cc configuration so that the assembler is not
                    628:    required.
                    629:  - Fixed a bug in the C++ test that caused it to fail in 64-bit
                    630:    environments.
                    631:
                    632: Since 4.11:
                    633:  - Fixed ElfW definition in dyn_load.c. (Thanks to Fergus Henderson.)
                    634:    This prevented the dynamic library support from compiling on some
                    635:    older ELF Linux systems.
                    636:  - Fixed UTS4 port (which I apparently mangled during the integration)
                    637:    (Thanks to again to Alistair Crooks.)
                    638:  - "Make C++" failed on Suns with SC4.0, due to a problem with "bool".
                    639:    Fixed in gc_priv.h.
                    640:  - Added more pieces for GNU win32.  (Thanks to Timothy N. Newsham.)
                    641:    The current state of things should suffice for at least some
                    642:    applications.
                    643:  - Changed the out of memory retry count handling as suggested by
                    644:    Kenjiro Taura.  (This matters only if GC_max_retries > 0, which
                    645:    is no longer the default.)
                    646:  - If a /proc read failed repeatedly, GC_written_pages was not updated
                    647:    correctly.  (Thanks to Peter Chubb for diagnosing this.)
                    648:  - Under unlikely circumstances, the allocator could infinite loop in
                    649:    an out of memory situation.  (Thanks again to Kenjiro Taura for
                    650:    identifying the problem and supplying a fix.)
                    651:  - Fixed a syntactic error in the DJGPP code.  (Thanks to Fergus
                    652:    Henderson for finding this by inspection.)  Also fixed a test program
                    653:    problem with DJGPP (Thanks to Peter Monks.)
                    654:  - Atomic uncollectable objects were not treated correctly by the
                    655:    incremental collector.  This resulted in weird log statistics and
                    656:    occasional performance problems.  (Thanks to Peter Chubb for pointing
                    657:    this out.)
                    658:  - Fixed some problems resulting from compilers that dont define
                    659:    __STDC__.  In this case void * and char * were used inconsistently
                    660:    in some cases.  (Void * should not have been used at all.  If
                    661:    you have an ANSI superset compiler that does not define __STDC__,
                    662:    please compile with -D__STDC__=0. Thanks to Manuel Serrano and others
                    663:    for pointing out the problem.)
                    664:  - Fixed a compilation problem on Irix with -n32 and -DIRIX_THREADS.
                    665:    Also fixed some other IRIX_THREADS problems which may or may not have
                    666:    had observable symptoms.
                    667:  - Fixed an HP PA compilation problem in dyn_load.c.  (Thanks to
                    668:    Philippe Queinnec.)
                    669:  - SEGV fault handlers sometimes did not get reset correctly.  (Thanks
                    670:    to David Pickens.)
                    671:  - Added a fix for SOLARIS_THREADS on Intel.  (Thanks again to David
                    672:    Pickens.)  This probably needs more work to become functional.
                    673:  - Fixed struct sigcontext_struct in os_dep.c for compilation under
                    674:    Linux 2.1.X.        (Thanks to Fergus Henderson.)
                    675:  - Changed the DJGPP STACKBOTTOM and DATASTART values to those suggested
                    676:    by Kristian Kristensen.  These may still not be right, but it is
                    677:    it is likely to work more often than what was there before.  They may
                    678:    even be exactly right.
                    679:  - Added a #include <string.h> to test_cpp.cc.  This appears to help
                    680:    with HP/UX and gcc.  (Thanks to assar@sics.se.)
                    681:  - Version 4.11 failed to run in incremental mode on recent 64-bit Irix
                    682:    kernels.  This was a problem related to page unaligned heap segments.
                    683:    Changed the code to page align heap sections on all platforms.
                    684:    (I had mistakenly identified this as a kernel problem earlier.
                    685:    It was not.)
                    686:  - Version 4.11 did not make allocated storage executable, except on
                    687:    one or two platforms, due to a bug in a #if test.  (Thanks to Dave
                    688:    Grove for pointing this out.)
                    689:  - Added sparc_sunos4_mach_dep.s to support Sun's compilers under SunOS4.
                    690:  - Added GC_exclude_static_roots.
                    691:  - Fixed the object size mapping algorithm.  This shouldn't matter,
                    692:    but the old code was ugly.
                    693:  - Heap checking code could die if one of the allocated objects was
                    694:    larger than its base address.  (Unsigned underflow problem.  Thanks
                    695:    to Clay Spence for isolating the problem.)
                    696:  - Added RS6000 (AIX) dynamic library support and fixed STACK_BOTTOM.
                    697:    (Thanks to Fred Stearns.)
                    698:  - Added Fergus Henderson's patches for improved robustness with large
                    699:    heaps and lots of blacklisting.
                    700:  - Added Peter Chubb's changes to support Solaris Pthreads, to support
                    701:    MMAP allocation in Solaris, to allow Solaris to find dynamic libraries
                    702:    through /proc, to add malloc_typed_ignore_off_page, and a few other
                    703:    minor features and bug fixes.
                    704:  - The Solaris 2 port should not use sbrk.  I received confirmation from
                    705:    Sun that the use of sbrk and malloc in the same program is not
                    706:    supported.  The collector now defines USE_MMAP by default on Solaris.
                    707:  - Replaced the djgpp makefile with Gary Leavens' version.
                    708:  - Fixed MSWIN32 detection test.
                    709:  - Added Fergus Henderson's patches to allow putting the collector into
                    710:    a DLL under GNU win32.
                    711:  - Added Ivan V. Demakov's port to Watcom C on X86.
                    712:  - Added Ian Piumarta's Linux/PowerPC port.
                    713:  - On Brian Burton's suggestion added PointerFreeGC to the placement
                    714:    options in gc_cpp.h.  This is of course unsafe, and may be controversial.
                    715:    On the other hand, it seems to be needed often enough that it's worth
                    716:    adding as a standard facility.
                    717:
                    718: Since 4.12:
                    719:  - Fixed a crucial bug in the Watcom port.  There was a redundant decl
                    720:    of GC_push_one in gc_priv.h.
                    721:  - Added FINALIZE_ON_DEMAND.
                    722:  - Fixed some pre-ANSI cc problems in test.c.
                    723:  - Removed getpagesize() use for Solaris.  It seems to be missing in one
                    724:    or two versions.
                    725:  - Fixed bool handling for SPARCCompiler version 4.2.
                    726:  - Fixed some files in include that had gotten unlinked from the main
                    727:    copy.
                    728:  - Some RS/6000 fixes (missing casts).  Thanks to Toralf Foerster.
                    729:  - Fixed several problems in GC_debug_realloc, affecting mostly the
                    730:    FIND_LEAK case.
                    731:  - GC_exclude_static_roots contained a buggy unsigned comparison to
                    732:    terminate a loop.  (Thanks to Wilson Ho.)
                    733:  - CORD_str failed if the substring occurred at the last possible position.
                    734:    (Only affects cord users.)
                    735:  - Fixed Linux code to deal with RedHat 5.0 and integrated Peter Bigot's
                    736:    os_dep.c code for dealing with various Linux versions.
                    737:  - Added workaround for Irix pthreads sigaction bug and possible signal
                    738:    misdirection problems.
                    739: Since alpha1:
                    740:  - Changed RS6000 STACKBOTTOM.
                    741:  - Integrated Patrick Beard's Mac changes.
                    742:  - Alpha1 didn't compile on Irix m.n, m < 6.
                    743:  - Replaced Makefile.dj with a new one from Gary Leavens.
                    744:  - Added Andrew Stitcher's changes to support SCO OpenServer.
                    745:  - Added PRINT_BLACK_LIST, to allow debugging of high densities of false
                    746:    pointers.
                    747:  - Added code to debug allocator to keep track of return address
                    748:    in GC_malloc caller, thus giving a bit more context.
                    749:  - Changed default behavior of large block allocator to more
                    750:    aggressively avoid fragmentation.  This is likely to slow down the
                    751:    collector when it succeeds at reducing space cost.
                    752:  - Integrated Fergus Henderson's CYGWIN32 changes.  They are untested,
                    753:    but needed for newer versions.
                    754:  - USE_MMAP had some serious bugs.  This caused the collector to fail
                    755:    consistently on Solaris with -DSMALL_CONFIG.
                    756:  - Added Linux threads support, thanks largely to Fergus Henderson.
                    757: Since alpha2:
                    758:  - Fixed more Linux threads problems.
                    759:  - Changed default GC_free_space_divisor to 3 with new large block allocation.
                    760:    (Thanks to Matthew Flatt for some measurements that suggest the old
                    761:    value sometimes favors space too much over time.)
                    762:  - More CYGWIN32 fixes.
                    763:  - Integrated Tyson-Dowd's Linux-M68K port.
                    764:  - Minor HP PA and DEC UNIX fixes from Fergus Henderson.
                    765:  - Integrated Christoffe Raffali's Linux-SPARC changes.
                    766:  - Allowed for one more GC fixup iteration after a full GC in incremental
                    767:    mode.  Some quick measurements suggested that this significantly
                    768:    reduces pause times even with smaller GC_RATE values.
                    769:  - Moved some more GC data structures into GC_arrays.  This decreases
                    770:    pause times and GC overhead, but makes debugging slightly less convenient.
                    771:  - Fixed namespace pollution problem ("excl_table").
                    772:  - Made GC_incremental a constant for -DSMALL_CONFIG, hopefully shrinking
                    773:    that slightly.
                    774:  - Added some win32 threads fixes.
                    775:  - Integrated Ivan Demakov and David Stes' Watcom fixes.
                    776:  - Various other minor fixes contributed by many people.
                    777:  - Renamed config.h to gcconfig.h, since config.h tends to be used for
                    778:    many other things.
                    779:  - Integrated Matthew Flatt's support for 68K MacOS "far globals".
                    780:  - Fixed up some of the dynamic library Makefile targets for consistency
                    781:    across platforms.
                    782:  - Fixed a USE_MMAP typo that caused out-of-memory handling to fail
                    783:    on Solaris.
                    784:  - Added code to test.c to test thread creation a bit more.
                    785:  - Integrated GC_win32_free_heap, as suggested by Ivan Demakov.
                    786:  - Fixed Solaris 2.7 stack base finding problem.  (This may actually
                    787:    have been done in an earlier alpha release.)
                    788: Since alpha3:
                    789:  - Fixed MSWIN32 recognition test, which interfered with cygwin.
                    790:  - Removed unnecessary gc_watcom.asm from distribution.  Removed
                    791:    some obsolete README.win32 text.
                    792:  - Added Alpha Linux incremental GC support.  (Thanks to Philipp Tomsich
                    793:    for code for retrieving the fault address in a signal handler.)
                    794:    Changed Linux signal handler context argument to be a pointer.
                    795:  - Took care of some new warnings generated by the 7.3 SGI compiler.
                    796:  - Integrated Phillip Musumeci's FreeBSD/ELF fixes.
                    797:  - -DIRIX_THREADS was broken with the -o32 ABI (typo in gc_priv.h>
                    798:
                    799: Since 4.13:
                    800:  - Fixed GC_print_source_ptr to not use a prototype.
                    801:  - generalized CYGWIN test.
                    802:  - gc::new did the wrong thing with PointerFreeGC placement.
                    803:    (Thanks to Rauli Ruohonen.)
                    804:  - In the ALL_INTERIOR_POINTERS (default) case, some callee-save register
                    805:    values could fail to be scanned if the register was saved and
                    806:    reused in a GC frame.  This showed up in verbose mode with gctest
                    807:    compiled with an unreleased SGI compiler.  I vaguely recall an old
                    808:    bug report that may have been related.  The bug was probably quite old.
                    809:    (The problem was that the stack scanning could be deferred until
                    810:    after the relevant frame was overwritten, and the new save location
                    811:    might be outside the scanned area.  Fixed by more eager stack scanning.)
                    812:  - PRINT_BLACK_LIST had some problems.  A few source addresses were garbage.
                    813:  - Replaced Makefile.dj and added -I flags to cord make targets.
                    814:    (Thanks to Gary Leavens.)
                    815:  - GC_try_to_collect was broken with the nonincremental collector.
                    816:  - gc_cleanup destructors could pass the wrong address to
                    817:    GC_register_finalizer_ignore_self in the presence of multiple
                    818:    inheritance.  (Thanks to Darrell Schiebel.)
                    819:  - Changed PowerPC Linux stack finding code.
                    820:
                    821: Since 4.14alpha1
                    822:  - -DSMALL_CONFIG did not work reliably with large (> 4K) pages.
                    823:    Recycling the mark stack during expansion could result in a size
                    824:    zero heap segment, which confused things.  (This was probably also an
                    825:    issue with the normal config and huge pages.)
                    826:  - Did more work to make sure that callee-save registers were scanned
                    827:    completely, even with the setjmp-based code.  Added USE_GENERIC_PUSH_REGS
                    828:    macro to facilitate testing on machines I have access to.
                    829:  - Added code to explicitly push register contents for win32 threads.
                    830:    This seems to be necessary.  (Thanks to Pierre de Rop.)
                    831:
                    832: Since 4.14alpha2
                    833:  - changed STACKBOTTOM for DJGPP (Thanks to Salvador Eduardo Tropea).
                    834:
                    835: Since 4.14
                    836:  - Reworked large block allocator.  Now uses multiple doubly linked free
                    837:    lists to approximate best fit.
                    838:  - Changed heap expansion heuristic.  Entirely free blocks are no longer
                    839:    counted towards the heap size.  This seems to have a major impact on
                    840:    heap size stability; the old version could expand the heap way too
                    841:    much in the presence of large block fragmentation.
                    842:  - added -DGC_ASSERTIONS and some simple assertions inside the collector.
                    843:    This is mainlyt for collector debugging.
                    844:  - added -DUSE_MUNMAP to allow the heap to shrink.  Suupported on only
                    845:    a few UNIX-like platforms for now.
                    846:  - added GC_dump_regions() for debugging of fragmentation issues.
                    847:  - Changed PowerPC pointer alignment under Linux to 4.  (This needs
                    848:    checking by someone who has one.  The suggestions came to me via a
                    849:    rather circuitous path.)
                    850:  - Changed the Linux/Alpha port to walk the data segment backwards until
                    851:    it encounters a SIGSEGV.  The old way to find the start of the data
                    852:    segment broke with a recent release.
                    853:  - cordxtra.c needed to call GC_REGISTER_FINALIZER instead of
                    854:    GC_register_finalizer, so that it would continue to work with GC_DEBUG.
                    855:  - allochblk sometimes cleared the wrong block for debugging purposes
                    856:    when it dropped blacklisted blocks.  This could result in spurious
                    857:    error reports with GC_DEBUG.
                    858:  - added MACOS X Server support.  (Thanks to Andrew Stone.)
                    859:  - Changed the Solaris threads code to ignore stack limits > 8 MB with
                    860:    a warning.  Empirically, it is not safe to access arbitrary pages
                    861:    in such large stacks.  And the dirty bit implementation does not
                    862:    guarantee that none of them will be accessed.
                    863:  - Integrated Martin Tauchmann's Amiga changes.
                    864:  - Integrated James Dominy's OpenBSD/SPARC port.
                    865:
                    866: Since 5.0alpha1
                    867:  - Fixed bugs introduced in alpha1 (OpenBSD & large block initialization).
                    868:  - Added -DKEEP_BACK_PTRS and backptr.h interface.  (The implementation
                    869:    idea came from Al Demers.)
                    870:
                    871: Since 5.0alpha2
                    872:  - Added some highly incomplete code to support a copied young generation.
                    873:    Comments on nursery.h are appreciated.
                    874:  - Changed -DFIND_LEAK, -DJAVA_FINALIZATION, and -DFINALIZE_ON_DEMAND,
                    875:    so the same effect could be obtained with a runtime switch.   This is
                    876:    a step towards standardizing on a single dynamic GC library.
                    877:  - Significantly changed the way leak detection is handled, as a consequence
                    878:    of the above.
                    879:
                    880: Since 5.0 alpha3
                    881:  - Added protection fault handling patch for Linux/M68K from Fergus
                    882:    Henderson and Roman Hodek.
                    883:  - Removed the tests for SGI_SOURCE in new_gc_alloc.h.  This was causing that
                    884:    interface to fail on nonSGI platforms.
                    885:  - Changed the Linux stack finding code to use /proc, after changing it
                    886:    to use HEURISTIC1.  (Thanks to David Mossberger for pointing out the
                    887:    /proc hook.)
                    888:  - Added HP/UX incremental GC support and HP/UX 11 thread support.
                    889:    Thread support is currently still flakey.
                    890:  - Added basic Linux/IA64 support.
                    891:  - Integrated Anthony Green's PicoJava support.
                    892:  - Integrated Scott Ananian's StrongARM/NetBSD support.
                    893:  - Fixed some fairly serious performance bugs in the incremental
                    894:    collector.  These have probably been there essentially forever.
                    895:    (Mark bits were sometimes set before scanning dirty pages.
                    896:    The reclaim phase unnecessarily dirtied full small object pages.)
                    897:  - Changed the reclaim phase to ignore nearly full pages to avoid
                    898:    touching them.
                    899:  - Limited GC_black_list_spacing to roughly the heap growth increment.
                    900:  - Changed full collection triggering heuristic to decrease full GC
                    901:    frequency by default, but to explicitly trigger full GCs during
                    902:    heap growth.  This doesn't always improve things, but on average it's
                    903:    probably a win.
                    904:  - GC_debug_free(0, ...) failed.  Thanks to Fergus Henderson for the
                    905:    bug report and fix.
                    906:
                    907: Since 5.0 alpha4
                    908:  - GC_malloc_explicitly_typed and friends sometimes failed to
                    909:    initialize first word.
                    910:  - Added allocation routines and support in the marker for mark descriptors
                    911:    in a type structure referenced by the first word of an object.  This was
                    912:    introduced to support gcj, but hopefully in a way that makes it
                    913:    generically useful.
                    914:  - Added GC_requested_heapsize, and inhibited collections in nonincremental
                    915:    mode if the actual used heap size is less than what was explicitly
                    916:    requested.
                    917:  - The Solaris pthreads version of GC_pthread_create didn't handle a NULL
                    918:    attribute pointer.  Solaris thread support used the wrong default thread
                    919:    stack size.  (Thanks to Melissa O'Neill for the patch.)
                    920:  - Changed PUSH_CONTENTS macro to no longer modify first parameter.
                    921:    This usually doesn't matter, but it was certainly an accident waiting
                    922:    to happen ...
                    923:  - Added GC_register_finalizer_no_order and friends to gc.h.  They're
                    924:    needed by Java implementations.
                    925:  - Integrated a fix for a win32 deadlock resulting from clock() calling
                    926:    malloc.  (Thanks to Chris Dodd.)
                    927:  - Integrated Hiroshi Kawashima's port to Linux/MIPS.  This was designed
                    928:    for a handheld platform, and may or may not be sufficient for other
                    929:    machines.
                    930:  - Fixed a va_arg problem with the %c specifier in cordprnt.c.  It appears
                    931:    that this was always broken, but recent versions of gcc are the first to
                    932:    report the (statically detectable) bug.
                    933:  - Added an attempt at a more general solution to dlopen races/deadlocks.
                    934:    GC_dlopen now temporarily disables collection.  Still not ideal, but ...
                    935:  - Added -DUSE_I686_PREFETCH, -DUSE_3DNOW_PREFETCH, and support for IA64
                    936:    prefetch instructions.  May improve performance measurably, but I'm not
                    937:    sure the code will run correctly on processors that don't support the
                    938:    instruction.  Won't build except with very recent gcc.
                    939:  - Added caching for header lookups in the marker.  This seems to result
                    940:    in a barely measurable performance gain.  Added support for interleaved
                    941:    lookups of two pointers, but unconfigured that since the performance
                    942:    gain is currently near zero, and it adds to code size.
                    943:  - Changed Linux DATA_START definition to check both data_start and
                    944:    __data_start, since nothing else seems to be portable.
                    945:  - Added -DUSE_LD_WRAP to optionally take advantage of the GNU ld function
                    946:    wrapping mechanism.  Probably currently useful only on Linux.
                    947:  - Moved some variables for the scratch allocator into GC_arrays, on
                    948:    Martin Hirzel's suggestion.
                    949:  - Fixed a win32 threads bug that caused the collector to not look for
                    950:    interior pointers from one of the thread stacks without
                    951:    ALL_INTERIOR_POINTERS.  (Thanks to Jeff Sturm.)
                    952:  - Added Mingw32 support.  (Thanks again to Jeff Sturm for the patch.)
                    953:  - Changed the alpha port to use the generic register scanning code instead
                    954:    of alpha_mach_dep.s.  Alpha_mach_dep.s doesn't look for pointers in fp
                    955:    registers, but gcc sometimes spills pointers there.  (Thanks to Manuel
                    956:    Serrano for helping me debug this by email.)  Changed the IA64 code to
                    957:    do something similar for similar reasons.
                    958:
                    959: [5.0alpha5 doesn't really exist, but it may have escaped.]
                    960:
                    961: Since 5.0alpha6:
                    962:  - -DREDIRECT_MALLOC was broken in alpha6. Fixed.
                    963:  - Cleaned up gc_ccp.h slightly, thus also causing the HP C++ compiler to
                    964:    accept it.
                    965:  - Removed accidental reference to dbg_mlc.c, which caused dbg_mlc.o to be
                    966:    linked into every executable.
                    967:  - Added PREFETCH to bitmap marker.  Changed it to use the header cache.
                    968:  - GC_push_marked sometimes pushed one object too many, resulting in a
                    969:    segmentation fault in GC_mark_from_mark_stack.  This was probably an old
                    970:    bug.  It finally showed up in gctest on win32.
                    971:  - Gc_priv.h erroneously #defined GC_incremental to be TRUE instead of FALSE
                    972:    when SMALL_CONFIG was defined.  This was no doubt a major performance bug for
                    973:    the default win32 configuration.
                    974:  - Removed -DSMALL_CONFIG from NT_MAKEFILE.  It seemed like an anchronism now
                    975:    that the average PC has 64MB or so.
                    976:  - Integrated Bryce McKinley's patches for linux threads and dynamic loading
                    977:    from the libgcj tree.  Turned on dynamic loading support for Linux/PPC.
                    978:  - Changed the stack finding code to use environ on HP/UX.  (Thanks
                    979:    to Gustavo Rodriguez-Rivera for the suggestion.)  This should probably
                    980:    be done on other platforms, too.  Since I can't test those, that'll
                    981:    wait until after 5.0.
                    982:
                    983: Since 5.0alpha7:
                    984:  - Fixed threadlibs.c for linux threads.  -DUSE_LD_WRAP was broken and
                    985:    -ldl was omitted.  Fixed Linux stack finding code to handle
                    986:    -DUSE_LD_WRAP correctly.
                    987:  - Added MSWIN32 exception handler around marker, so that the collector
                    988:    can recover from root segments that are unmapped during the collection.
                    989:    This caused occasional failures under Windows 98, and may also be
                    990:    an issue under Windows NT/2000.
                    991:
                    992: Since 5.0
                    993:  - Fixed a gc.h header bug which showed up under Irix.  (Thanks to
                    994:    Dan Sullivan.)
                    995:  - Fixed a typo in GC_double_descr in typd_mlc.c.
                    996:    This probably could result in objects described by array descriptors not
                    997:    getting traced correctly.  (Thanks to Ben Hutchings for pointing this out.)
                    998:  - The block nearly full tests in reclaim.c were not correct for 64 bit
                    999:    environments.  This could result in unnecessary heap growth under unlikely
                   1000:    conditions.
                   1001:
                   1002: Since 5.1
                   1003:  - dyn_load.c declared GC_scratch_last_end_ptr as an extern even if it
                   1004:    was defined as a macro.  This prevented the collector from building on
                   1005:    Irix.
                   1006:  - We quietly assumed that indirect mark descriptors were never 0.
                   1007:    Our own typed allocation interface violated that.  This could result
                   1008:    in segmentation faults in the marker with typed allocation.
                   1009:  - Fixed a _DUSE_MUNMAP bug in the heap block allocation code.
                   1010:    (Thanks to Ben Hutchings for the patch.)
                   1011:  - Taught the collector about VC++ handling array operator new.
                   1012:    (Thanks again to Ben Hutchings for the patch.)
                   1013:  - The two copies of gc_hdrs.h had diverged.  Made one a link to the other
                   1014:    again.
                   1015:
                   1016: Since 5.2  (A few 5.2 patches are not in 6.0alpha1)
                   1017:  - Fixed _end declaration for OSF1.
                   1018:  - There were lots of spurious leak reports in leak detection mode, caused
                   1019:    by the fact that some pages were not being swept, and hence unmarked
                   1020:    objects weren't making it onto free lists.  (This bug dated back to 5.0.)
                   1021:  - Fixed a typo in the liblinuxgc.so Makefile rule.
                   1022:  - Added the GetExitCodeThread to Win32 GC_stop_world to (mostly) work
                   1023:    around a Windows 95 GetOpenFileName problem.  (Thanks to Jacob Navia.)
                   1024:
                   1025: Since 5.3
                   1026:  - Fixed a typo that prevented compilation with -DUSE_3DNOW_PREFETCH.
                   1027:    (Thanks to Shawn Wagner for actually testing this.)
                   1028:  - Fixed GC_is_thread_stack in solaris_threads.c.  It forgot to return a value
                   1029:    in the common case.  I wonder why nobody noticed?
                   1030:  - Fixed another silly syntax problem in GC_double_descr.  (Thanks to
                   1031:    Fergus Henderson for finding it.)
                   1032:  - Fixed a GC_gcj_malloc bug: It tended to release the allocator lock twice.
                   1033:
                   1034: Since 5.4  (A few 5.3 patches are not in 6.0alpha2)
                   1035:  - Added HP/PA prefetch support.
                   1036:  - Added -DDBG_HDRS_ALL and -DSHORT_DBG_HDRS to reduce the cost and improve
                   1037:    the reliability of generating pointer backtrace information, e.g. in
                   1038:    the Bigloo environment.
                   1039:  - Added parallel marking support (-DPARALLEL_MARK).  This currently
                   1040:    works only under IA32 and IA64 Linux, but it shouldn't be hard to adapt
                   1041:    to other platforms.  This is intended to be a lighter-weight (less
                   1042:    new code, probably not as scalable) solution than the work by Toshio Endo
                   1043:    et al, at the University of Tokyo.  A number of their ideas were
                   1044:    reused, though the code wasn't, and the underlying data structure
                   1045:    is significantly different.  In particular, we keep the global mark
                   1046:    stack as a single shared data structure, but most of the work is done
                   1047:    on smaller thread-local mark stacks.
                   1048:  - Changed GC_malloc_many to be cheaper, and to require less mutual exclusion
                   1049:    with -DPARALLEL_MARK.
                   1050:  - Added full support for thread local allocation under Linux
                   1051:    (-DTHREAD_LOCAL_ALLOC).  This is a thin veneer on GC_malloc_many, and
                   1052:    should be easily portable to other platforms, especially those that
                   1053:    support pthreads.
                   1054:  - CLEAR_DOUBLE was not always getting invoked when it should have been.
                   1055:  - GC_gcj_malloc and friends used different out of memory handling than
                   1056:    everything else, probably because I forgot about one when I implemented
                   1057:    the other.  They now both call GC_oom_fn(), not GC_oom_action().
                   1058:  - Integrated Jakub Jelinek's fixes for Linux/SPARC.
                   1059:  - Moved GC_objfreelist, GC_aobjfreelist, and GC_words_allocd out of
                   1060:    GC_arrays, and separately registered the first two as excluded roots.
                   1061:    This makes code compiled with gc_inl.h less dependent on the
                   1062:    collector version.  (It would be nice to remove the inclusion of
                   1063:    gc_priv.h by gc_inl.h completely, but we're not there yet.  The
                   1064:    locking definitions in gc_priv.h are still referenced.)
                   1065:    This change was later coniditoned on SEPARATE_GLOBALS, which
                   1066:    is not defined by default, since it involves a performance hit.
                   1067:  - Register GC_obj_kinds separately as an excluded root region.  The
                   1068:    attempt to register it with GC_arrays was usually failing.  (This wasn't
                   1069:    serious, but seemed to generate some confusion.)
                   1070:  - Moved backptr.h to gc_backptr.h.
                   1071:
                   1072: Since 6.0alpha1
                   1073:  - Added USE_MARK_BYTES to reduce the need for compare-and-swap on platforms
                   1074:    for which that's expensive.
                   1075:  - Fixed a locking bug ib GC_gcj_malloc and some locking assertion problems.
                   1076:  - Added a missing volatile to OR_WORD and renamed the parameter to
                   1077:    GC_compare_and_swap so it's not a C++ reserved word.  (Thanks to
                   1078:    Toshio Endo for pointing out both of those.)
                   1079:  - Changed Linux dynamic library registration code to look at /proc/self/maps
                   1080:    instead of the rld data structures when REDIRECT_MALLOC is defined.
                   1081:    Otherwise some of the rld data data structures may be prematurely garbage
                   1082:    collected.  (Thanks to Eric Benson for helping to track this down.)
                   1083:  - Fixed USE_LD_WRAP a bit more, so it should now work without threads.
                   1084:  - Renamed XXX_THREADS macros to GC_XXX_THREADS for namespace correctness.
                   1085:    Tomporarily added some backward compatibility definitions.  Renamed
                   1086:    USE_LD_WRAP to GC_USE_LD_WRAP.
                   1087:  - Many MACOSX POWERPC changes, some additions to the gctest output, and
                   1088:    a few minor generic bug fixes.  (Thanks to Dietmar Planitzer.)
                   1089:
                   1090: Since 6.0 alpha2
                   1091:  - Fixed the /proc/self/maps code to not seek, since that apparently is not
                   1092:    reliable across all interesting kernels.
                   1093:  - Fixed some compilation problems in the absence of PARALLEL_MARK
                   1094:    (introduced in alpha2).
                   1095:  - Fixed an algorithmic problem with PARALLEL_MARK.  If work needs to
                   1096:    be given back to the main mark "stack", the BOTTOM entries of the local
                   1097:    stack should be given away, not the top ones.  This has substantial
                   1098:    performance impact, especially for > 2 processors, from what I can tell.
                   1099:  - Extracted gc_lock.h from gc_priv.h.  This should eventually make it a
                   1100:    bit easier to avoid including gc_priv.h in clients.
                   1101:  - Moved all include files to include/ and removed duplicate links to the
                   1102:    same file.  The old scheme was a bad idea because it was too easy to get the
                   1103:    copies out of sync, and many systems don't support hard links.
                   1104:    Unfortunately, it's likely that I broke some of the non-Unix Makefiles in
                   1105:    the process, although I tried to update them appropriately.
                   1106:  - Removed the partial support for a copied nursery.  It's not clear that
                   1107:    this would be a tremendous win, since we don't consistently lose to
                   1108:    generational copying collectors.  And it would significantly complicate
                   1109:    many things.  May be reintroduced if/when it really turns out to win.
                   1110:  - Removed references to IRIX_JDK_THREADS, since I believe there never
                   1111:    were and never will be any clients.
                   1112:  - Added some code to linux_threads.c to possibly support HPUX threads
                   1113:    using the Linux code.  Unfortunately, it doesn't work yet, and is
                   1114:    currently disabled.
                   1115:  - Added support under Linux/X86 for saving the call chain, both in (debug)
                   1116:    objects for client debugging, and in GC_arrays._last_stack for GC
                   1117:    debugging.  This was previously supported only under Solaris.  It is
                   1118:    not enabled by default under X86, since it requires that code be compiled
                   1119:    to explicitly dave frame pointers on the call stack.  (With gcc this
                   1120:    currently happens by default, but is often turned off explicitly.)
                   1121:    To turn it on, define SAVE_CALL_CHAIN.
                   1122:
                   1123: Since 6.0 alpha3
                   1124:  - Moved up the detection of mostly full blocks to the initiatiation of the
                   1125:    sweep phase.  This eliminates some lock conention in the PARALLEL_MARK case,
                   1126:    as multiple threads try to look at mostly full blocks concurrently.
                   1127:  - Restored the code in GC_malloc_many that grabs a prefix of the global
                   1128:    free list.  This avoids the case in which every GC_malloc_many call
                   1129:    tries and fails to allocate a new heap block, and the returns a single
                   1130:    object from the global free list.
                   1131:  - Some minor fixes in new_hblk.c.  (Attempted to build free lists in order
                   1132:    of increasing addresses instead of decreasing addresses for cache performance
                   1133:    reasons.  But this seems to be only a very minor gain with -DEAGER_SWEEP,
                   1134:    and a loss in other cases.  So the change was backed out.)
                   1135:  - Fixed some of the documentation.  (Thanks in large part to Fergus
                   1136:    Henderson.)
                   1137:  - Fixed the Linux USE_PROC_FOR_LIBRARIES code to deal with apps that perform
                   1138:    large numbers of mmaps.  (Thanks to Eric Benson.)  Also fixed that code to
                   1139:    deal with short reads.
                   1140:  - Added GC_get_total_bytes().
                   1141:  - Fixed leak detection mode to avoid spurious messages under linuxthreads.
                   1142:    (This should also now be easy for the other supported threads packages.
                   1143:    But the code is tricky enough that I'm hesitant to do it without being able
                   1144:    to test.  Everything allocated in the GC thread support itself should be
                   1145:    explicitly deallocated.)
                   1146:  - Made it possible (with luck) to redirect malloc to GC_local_malloc.
                   1147:
                   1148: Since 6.0 alpha4
                   1149:  - Changed the definition of GC_pause in linux_threads.c to use a volatile
                   1150:    asm.  Some versions of gcc apparently optimize away writes to local volatile
                   1151:    variables.  This caused poor locking behaviour starting at about
                   1152:    4 processors.
                   1153:  - Added GC_start_blocking(), GC_end_blocking() calls and wrapper for sleep
                   1154:    to linux_threads.c.
                   1155:    The first two calls could be used to generally avoid sending GC signals to
                   1156:    blocked threads, avoiding both premature wakeups and unnecessary overhead.
                   1157:  - Fixed a serious bug in thread-local allocation.  At thread termination,
                   1158:    GC_free could get called on small integers.  Changed the code for thread
                   1159:    termination to more efficiently return left-over free-lists.
                   1160:  - Integrated Kjetil Matheussen's BeOS support.
                   1161:  - Rearranged the directory structure to create the doc and tests
                   1162:    subdirectories.
                   1163:  - Sort of integrated Eric Benson's patch for OSF1.  This provided basic
                   1164:    OSF1 thread support by suitably extending hpux_irix_threads.c.  Based
                   1165:    on earlier email conversations with David Butenhof, I suspect that it
                   1166:    will be more reliable in the long run to base this on linux_threads.c
                   1167:    instead.  Thus I attempted to patch up linux_threads.c based on Eric's code.
                   1168:    The result is almost certainly broken, but hopefully close enough that
                   1169:    someone with access to a machine can pick it up.
                   1170:  - Integrated lots of minor changes from the NetBSD distribution.  (These
                   1171:    were supplied by David Brownlee.  I'm not sure about the original
                   1172:    authors.)
                   1173:  - Hacked a bit more on the HP/UX thread-support in linux_threads.c.  It
                   1174:    now appears to work in the absence of incremental collection.  Renamed
                   1175:    hpux_irix_threads.c back to irix_threads.c, and removed the attempt to
                   1176:    support HPUX there.
                   1177:  - Changed gc.h to define _REENTRANT in cases in which it should already
                   1178:    have been defined. It is still safer to also define it on the command
                   1179:    line.
                   1180:
                   1181: Since 6.0alpha5:
                   1182:  - Changed the definition of DATASTART on ALPHA and IA64, where data_start
                   1183:    and __data_start are not defined by earlier versions of glibc.  This might
                   1184:    need to be fixed on other platforms as well.
                   1185:  - Changed the way the stack base and backing store base are found on IA64.
                   1186:    This should now remain reliable on future kernels.  But since it relies
                   1187:    on /proc, it will no longer work in the simulated NUE environment.
                   1188:  - Made the call to random() in dbg_mlc.c with -DKEEP_BACK_PTRS dependent
                   1189:    on the OS.  On non-Unix systems, rand() should be used instead.  Handled
                   1190:    small RAND_MAX.  (Thanks to Peter Ross for pointing this out.)
                   1191:  - Fixed the cord make rules to create the cord subdirectory, if necessary.
                   1192:    (Thanks to Doug Moen.)
                   1193:  - Changed fo_object_size calculation in finalize.c.  Turned finalization
                   1194:    of nonheap object into a no-op.  Removed anachronism from GC_size()
                   1195:    implementation.
                   1196:  - Changed GC_push_dirty call in solaris_threads.c to GC_push_selected.
                   1197:    It was missed in a previous renaming. (Thanks to Vladimir Tsichevski
                   1198:    for pointing this out.)
                   1199:  - Arranged to not not mask SIGABRT in linux_threads.c.  (Thanks to Bryce
                   1200:    McKinlay.)
                   1201:  - Added GC_no_dls hook for applications that want to register their own
                   1202:    roots.
                   1203:  - Integrated Kjetil Matheussen's Amiga changes.
                   1204:  - Added FREEBSD_STACKBOTTOM.  Changed the X86/FreeBSD port to use it.
                   1205:    (Thanks to Matthew Flatt.)
                   1206:  - Added pthread_detach interception for platforms supported by linux_threads.c
                   1207:    and irix_threads.c.  Should also be added for Solaris?
                   1208:  - Changed the USE_MMAP code to check for the case in which we got the
                   1209:    high end of the address space, i.e. mem_ptr + mem_sz == 0.  It appears
                   1210:    that this can happen under Solaris 7.  It seems to be allowed by what
                   1211:    I would claim is an oversight in the mmap specification.  (Thanks to Toshio
                   1212:    Endo for pointing out the problem.)
                   1213:  - Cleanup of linux_threads.c.  Some code was originally cloned from
                   1214:    irix_threads.c and now unnecessary.  Some comments were obviously wrong.
                   1215:  - (Mostly) fixed a longstanding problem with setting of dirty bits from
                   1216:    a signal handler.  In the presence of threads, dirty bits could get lost,
                   1217:    since the etting of a bit in the bit vector was not atomic with respect
                   1218:    to other updates.  The fix is 100% correct only for platforms for which
                   1219:    GC_test_and_set is defined.  The goal is to make that all platforms with
                   1220:    thread support.  Matters only if incremental GC and threads are both
                   1221:    enabled.
                   1222:  - made GC_all_interior_pointers (a.k.a. ALL_INTERIOR_POINTERS) an
                   1223:    initialization time, instead of build-time option.  This is a
                   1224:    nontrivial, high risk change.  It should slow down the code measurably
                   1225:    only if MERGE_SIZES is not defined, which is a very nonstandard
                   1226:    configuration.
                   1227:  - Added doc/README.environment, and implemented what it describes.  This
                   1228:    allows a number of additional configuration options to be set through
                   1229:    the environment.  It documents a few previously undocumented options.
                   1230:  - Integrated Eric Benson's leak testing improvements.
                   1231:  - Removed the option to throw away the beginning of each page (DISCARD_WORDS).
                   1232:    This became less and less useful as processors enforce stricter alignment.
                   1233:    And it hadn't been tested in ages, and was thus probably broken anyway.
                   1234:
                   1235: Since 6.0alpha6:
                   1236:  - Added GC_finalizer_notifier.  Fixed GC_finalize_on_demand.  (The variable
                   1237:    actually wasn't being tested at the right points.  The build-time flag
                   1238:    was.)
                   1239:  - Added Tom Tromey's S390 Linux patch.
                   1240:  - Added code to push GC_finalize_now in GC_push_finalizer_structures.
                   1241:    (Thanks to Matthew Flatt.)
                   1242:  - Added GC_push_gc_structures() to push all GC internal roots.
                   1243:  - Integrated some FreeBSD changes from Matthew Flatt.
                   1244:  - It looks like USRSTACK is not always correctly defined under Solaris.
                   1245:    Hacked gcconfig.h to attempt to work around the problem.  The result
                   1246:    is not well tested.  (Thanks again to Matthew Flatt for pointing this
                   1247:    out.  The gross hack is mine. - HB)
                   1248:  - Added Ji-Yong Chung's win32 threads and C++ fixes.
                   1249:  - Arranged for hpux_test_and_clear.s to no longer be needed or built.
                   1250:    It was causing build problems with gas, and it's not clear this is
                   1251:    better than the pthreads alternative on this platform.
                   1252:  - Some MINGW32 fixes from Hubert Garavel.
                   1253:  - Added Initial Hitachi SH4 port from Kaz Kojima.
                   1254:  - Ported thread-local allocation and parallel mark code to HP/UX on PA_RISC.
                   1255:  - Made include/gc_mark.h more public and separated out the really private
                   1256:    pieces.  This is probably still not quite sufficient for clients that
                   1257:    want to supply their own kind of type information.  But it's a start.
                   1258:    This involved lots of identifier renaming to make it namespace clean.
                   1259:  - Added GC_dont_precollect for clients that need complete control over
                   1260:    the root set.
                   1261:  - GC_is_visible didn't do the right thing with gcj objects.  (Not that
                   1262:    many people are likely to care, but ...)
                   1263:  - Don't redefine read with GC_USE_LD_WRAP.
                   1264:  - Initial port to LINUX/HP_PA.  Incremental collection and threads are not
                   1265:    yet supported.  (Incremental collection should work if you have the
                   1266:    right kernel.  Threads may work with a sufficiently patched pthread
                   1267:    library.)
                   1268:  - Changed gcconfig.h to recognize __i386__ as an alternative to i386 in
                   1269:    many places.  (Thanks to Benjamin Lerman.)
                   1270:  - Made win32_threads.c more tolerant of detaching a thread that it didn't
                   1271:    know about.  (Thanks to Paul Nash.)
                   1272:  - Added Makefile.am and configure.in from gcc to the distribution, with
                   1273:    minimal changes.  For the moment, those are just placeholders.  In the
                   1274:    future, we're planning to switch to a GNU-style build environment for
                   1275:    Un*x-like systems, though the old Makefile will remain as a backup.
                   1276:  - Turned off STUBBORN_ALLOC by default, and added it back as a Makefile
                   1277:    option.
                   1278:  - Redistributed some functions between malloc.c and mallocx.c, so that
                   1279:    simple statically linked apps no longer pull in mallocx.o.
                   1280:  - Changed large object allocation to clear the first and last few words
                   1281:    of each block before releassing the lock.  Otherwise the marker could see
                   1282:    objects with nonsensical type descriptors.
                   1283:  - Fixed a couple of subtle problems that could result in not recognizing
                   1284:    interior pointers from the stack.  (I believe these were introduced
                   1285:    in 6.0alpha6.)
                   1286:  - GC_debug_free_inner called GC_free, which tried to reacquire the
                   1287:    allocator lock, and hence deadlocked.  (DBG_HDRS_ALL probably never worked
                   1288:    with threads?)
                   1289:  - Fixed several problems with back traces.  Accidental references to a free
                   1290:    list could cause the free list pointer to be overwritten by a back pointer.
                   1291:    There seemed to be some problems with the encoding of root and finalizer
                   1292:    references.
                   1293:
                   1294: Since 6.0alpha7:
                   1295:  - Changed GC_debug_malloc_replacement and GC_debug_realloc_replacement
                   1296:    so that they compile under Irix.  (Thanks to Dave Love.)
                   1297:  - Updated powerpc_macosx_mach_dep.s so that it works if the collector
                   1298:    is in a dynamic library.  (Thanks to Andrew Begel.)
                   1299:  - Transformed README.debugging into debugging.html, updating and
                   1300:    expanding it in the process.  Added gcdescr.html and tree.html
                   1301:    from the web site to the GC distribution.
                   1302:  - Fixed several problems related to PRINT_BLACK_LIST. This involved
                   1303:    restructuring some of the marker macros.
                   1304:  - Fixed some problems with the sizing of objects with debug information.
                   1305:    Finalization was broken KEEP_BACK_PTRS or PRINT_BLACK_LIST.  Reduced the
                   1306:    object size with SHORT_DEBUG_HDRS by another word.
                   1307:  - The "Needed to allocate blacklisted ..." warning had inadvertently
                   1308:    been turned off by default, due to a buggy test in allchblk.c.  Turned
                   1309:    it back on.
                   1310:  - Removed the marker macros to deal with 2 pointers in interleaved fashion.
                   1311:    They were messy and the performance improvement seemed minimal.  We'll
                   1312:    leave such scheduling issues to the compiler.
                   1313:  - Changed Linux/PowerPC test to also check for __powerpc__ in response
                   1314:    to a discussion on the gcc mailing list.
                   1315:  - On Matthew Flatt's suggestion removed the "static" from the jmp_buf
                   1316:    declaration in GC_generic_push_regs.  This was causing problems in
                   1317:    systems that register all of their own roots.  It looks far more correct
                   1318:    to me without the "static" anyway.
                   1319:  - Fixed several problems with thread local allocation of pointerfree or
                   1320:    typed objects.  The collector was reclaiming thread-local free lists, since
                   1321:    it wasn't following the link fields.
                   1322:  - There was apparently a long-standing race condition related to multithreaded
                   1323:    incremental collection.  A collection could be started and a thread stopped
                   1324:    between the memory unprotect system call and the setting of the
                   1325:    corresponding dirt bit.  I believe this did not affect Solaris or PCR, which
                   1326:    use a different dirty-bit implementation.  Fixed this by installing
                   1327:    signal handlers with sigaction instead of signal, and disabling the thread
                   1328:    suspend signal while in the write-protect handler.  (It is unclear
                   1329:    whether this scenario ever actually occurred.  I found it while tracking
                   1330:    down the following:)
                   1331:  - Incremental collection did not cooperate correctly with the PARALLEL_MARK
                   1332:    implementation of GC_malloc_many or the local_malloc primitves.  It still
                   1333:    doesn't work well, but it shouldn't lose memory anymore.
                   1334:  - Integrated some changes from the gcc source tree that I had previously
                   1335:    missed.  (Thanks to Bryce McKinley for the reminder/diff.)
                   1336:  - Added Makefile.direct as a copy of the default Makefile, which would
                   1337:    normally be overwritten if configure is run.
                   1338:  - Changed the gc.tar target in Makefile.direct to embed the version number
                   1339:    in the gc directory name.  This will affect future tar file distributions.
                   1340:  - Changed the Irix dynamic library finding code to no longer try to
                   1341:    eliminate writable text segments under Irix6.x, since that is probably no
                   1342:    longer necessary, and can apparently be unsafe on occasion.  (Thanks to
                   1343:    Shiro Kawai for pointing this out.)
                   1344:  - GC_cleanup with GC_DEBUG enabled passed a real object base address to
                   1345:    GC_debug_register_finalizer_ignore_self, which expected a pointer past the
                   1346:    debug header.  Call GC_register_finalizer_ignore_self instead, even with
                   1347:    debugging enabled.  (Thanks to Jean-Daniel Fekete for catching this.)
                   1348:  - The collector didn't build with call chain saving enabled but NARGS=0.
                   1349:    (Thanks to Maarten Thibaut.)
                   1350:  - Fixed up the GNU-style build files enough so that they work in some
                   1351:    obvious cases.
                   1352:  - Added initial port to Digital Mars compiler for win32. (Thanks to Walter
                   1353:    Bright.)
                   1354:
                   1355: Since 6.0alpha8:
                   1356:  - added README.macros.
                   1357:  - Made gc.mak a symbolic link to work around winzip's tendency to ignore
                   1358:    hard links.
                   1359:  - Simplified the setting of NEED_FIND_LIMIT in os_dep.c, possibly breaking
                   1360:    it on untested platforms.
                   1361:  - Integrated initial GNU HURD port. (Thanks to Chris Lingard and Igor
                   1362:    Khavkine.)
                   1363:  - A few more fixes for Digital Mars compiler (Walter Bright).
                   1364:  - Fixed gcc version recognition.  Renamed OPERATOR_NEW_ARRAY to
                   1365:    GC_OPERATOR_NEW_ARRAY.  Changed GC_OPERATOR_NEW_ARRAY to be the default.
                   1366:    It can be overridden with -DGC_NO_OPERATOR_NEW_ARRAY.  (Thanks to
                   1367:    Cesar Eduardo Barros.)
                   1368:  - Changed the byte size to free-list mapping in thread local allocation
                   1369:    so that size 0 allocations are handled correctly.
                   1370:  - Fixed Linux/MIPS stackbottom for new toolchain. (Thanks to Ryan Murray.)
                   1371:  - Changed finalization registration to invoke GC_oom_fn when it runs out
                   1372:    of memory.
                   1373:  - Removed lvalue cast in finalize.c.  This caused some debug configurations
                   1374:    not to build with some non-gcc compilers.
                   1375:
                   1376: Since 6.0alpha9:
                   1377:  - Two more bug fixes for KEEP_BACK_PTRS and DBG_HDRS_ALL.
                   1378:  - Fixed a stack clearing problem that resulted in SIGILL with a
                   1379:    misaligned stack pointer for multithreaded SPARC builds.
                   1380:  - Integrated another HURD patch (thanks to Igor Khavkine).
                   1381:
                   1382: Since 6.0:
                   1383:  - Non-debug, atomic allocations could result in bogus smashed object
                   1384:    reports with debugging on.  (Thanks to Patrick Doyle for the small
                   1385:    test case.)
                   1386:  - Fixed GC_get_register_stack_base (Itanium only) to work around a glibc
                   1387:    2.2.4 bug.
                   1388:  - Initial port to HP/UX on Itanium.  Thread support and both 32 and 64
                   1389:    bit ABIs appear to work.  Parallel mark support doesn't yet, due to
                   1390:    some inline assembly code issues.  Thread local allocation does appear
                   1391:    to work.
                   1392:  - ifdef'ed out glibc2.1/Itanium workaround.  I suspect nobody is using
                   1393:    that combination anymore.
                   1394:  - Added a patch to make new_gc_alloc.h usable with gcc3.0.  (Thanks to
                   1395:    Dimitris Vyzovitis for the patch.)
                   1396:  - Debugged 64-bit support on HP/UX PA-RISC.
                   1397:  - Turned on dynamic loading support for FreeBSD/ELF.  (Thanks to Peter
                   1398:    Housel.)
                   1399:  - Unregistering of finalizers with debugging allocation was broken.
                   1400:    (Thanks to Jani Kajala for the test case.)
                   1401:  - Old finalizers were not returned correctly from GC_debug_register_finalizer.
                   1402:  - Disabled MPROTECT_VDB for Linux/M68K based on a report that it doesn't work.
                   1403:  - Cleaned up some statistics gathering code in reclaim.c (Thanks to Walter
                   1404:    Bright.)
                   1405:  - Added some support for OpenBSD/ELF/Linux.  (Thanks to Suzuki Toshiya.)
                   1406:  - Added Jakub Jelinek's patch to use dl_iterate_phdr for dynamic library
                   1407:    traversal to dyn_load.c.  Changed it to weakly reference dl_iterate_phdr,
                   1408:    so that the old code is stilll used with old versions of glibc.
                   1409:  - Cleaned up feature test macros for various threads packages and
                   1410:    integrated (partially functional) FreeBSD threads code from Loren Rittle.
                   1411:    It's likely that the cleanup broke something, since it touched lots of
                   1412:    code.  It's also likelly that it fixed some unreported bugs in the
                   1413:    less common thread implementations, since some of the original code
                   1414:    didn't stand up to close scrutiny.  Support for the next pthreads
                   1415:    implementation should be easier to add.
                   1416:
                   1417: Since 6.1alpha1:
                   1418:  - No longer wrap read by default in multithreaded applications.  It was
                   1419:    pointed out on the libgcj list that this holds the allocation lock for
                   1420:    way too long if the read blocks.  For now, reads into the heap are
                   1421:    broken with incremental collection.  It's possible to turn this back on
                   1422:    if you make sure that read calls don't block (e.g. by calling select
                   1423:    first).
                   1424:  - Fix ifdef in Solaris_threads.h to refer to GC_SOLARIS_THREADS.
                   1425:  - Added check for environment variable GC_IGNORE_GCJ_INFO.
                   1426:  - Added printing of stop-the-world GC times if GC_PRINT_STATS environment
                   1427:    variable is set.
                   1428:  - The calloc definition in leak_detector.h was missing parentheses, and
                   1429:    realloc was missing a second argument to GC_REALLOC.
                   1430:    (Thanks to Elrond (elrond<at>samba-tng.org).)
                   1431:  - Added GC_PRINT_BACK_HEIGHT environment variable and associated
                   1432:    code, mostly in the new file backgraph.c.  See doc/README.environment.
                   1433:  - Added -DUSE_GLOBAL_ALLOC to work around a Windows NT issue.  (Thanks to
                   1434:    Jonathan Clark.)
                   1435:  - Integrated port to NEC EWS4800 (MIPS-based workstation, with somewhat
                   1436:    different address-space layout). This may help for other machines with
                   1437:    holes in the data segment.  (Thanks to Hironori Sakamoto.)
                   1438:  - Changed the order in which GC_push_roots and friends push things onto
                   1439:    the mark stack.  GC_push_all calls need to come first, since we can't
                   1440:    necessarily recovere if those overflow the mark stack.  (Thanks to
                   1441:    Matthew Flatt for tracking down the problem.)
                   1442:  - Some minor cleanups to mostly support the Intel compiler on Linux/IA64.
                   1443:
                   1444: Since 6.1 alpha2:
                   1445:  - Minor cleanup on the gcconfig.h section for SPARC.
                   1446:  - Minor fix to support Intel compiler for I386/Linux. (Thanks to Sven
                   1447:    Hartrumpf.)
                   1448:  - Added SPARC V9 (64-bit) support.  (Thanks to Jeff Sturm.)
                   1449:  - Restructured the way in which we determine whether or not to keep
                   1450:    call stacks for debug allocation.  By default SAVE_CALL_COUNT is
                   1451:    now zero on all platforms.  Added SAVE_CALL_NARGS parameters.
                   1452:    If possible, use execinfo.h to capture call stack.  (This should
                   1453:    add support for a number of new platforms, though often at
                   1454:    considerable runtime expense.)
                   1455:  - Try to print symbolic information for call stacks.  On Linux, we
                   1456:    do this with a combination of execinfo.h and running addr2line in
                   1457:    a separate process.  This is both much more expensive and much more
                   1458:    useful.  Amazingly, it seems to be fast enough for most purposes.
                   1459:  - Redefined strdup if -DREDIRECT_MALLOC is given.
                   1460:  - Changed incremental collector and MPROTECT_VDB implementation so that,
                   1461:    under favorable conditions, pointerfree objects are not protected.
                   1462:    Added GC_incremental_protection_needs() to determine ahead of time whether
                   1463:    pointerfree objects may be protected.  Replaced GC_write_hint() with
                   1464:    GC_remove_protection().
                   1465:  - Added test for GC_ENABLE_INCREMENTAL environment variable.
                   1466:  - Made GC_time_limit runtime configurable.  Added GC_PAUSE_TIME_TARGET
                   1467:    environment variable.
                   1468:  - Eliminated GC_page_sz, a duplicate of GC_page_size.
                   1469:  - Caused the Solaris and Irix thread creation primitives to call
                   1470:    GC_init_inner().
                   1471:
                   1472: Since 6.1alpha3:
                   1473:  - Fixed typo in sparc_mach_dep.S, preventing the 64-bit version from
                   1474:    building.  Increased 64-bit heap size limit in test.c slightly, since
                   1475:    a functional SPARC collector seems to slightly exceed the old limits.
                   1476:    (Thanks again to Jeff Sturm.)
                   1477:  - Use NPRGREG in solaris_threads.c, thus printing all registers if things
                   1478:    go wrong.
                   1479:  - Added GC_MARKERS environment variable to allow use of a single marker
                   1480:    thread on an MP without confusing the lock implementation.
                   1481:  - Collect much less aggressively in incremental mode with GC_TIME_UNLIMITED.
                   1482:    This is really a purely generational mode, and we can afford to
                   1483:    postpone the collection until the heap is (nearly) full.
                   1484:  - Remove read() wrapper for MPROTECT_VDB.  It was causing more harm than
                   1485:    good.  It is often no longer needed if system calls avoid writing to
                   1486:    pointerfull heap objects.
                   1487:  - Fix MACOSX test in gcconfig.h. (Thanks to John Clements.)
                   1488:  - Change GC_test_and_set so that it consistently has one argument.
                   1489:    Add spaces to ::: in powerpc assembly code in gc_locks.h.
                   1490:    (Thanks to Ryan Murray.)
                   1491:  - Fixed a formatting error in dbg_mlc.c.  Added prototype to GC_abort()
                   1492:    declaration.   (Thanks to Michael Smith.)
                   1493:  - Removed "source" argument to GC_find_start().  Eliminate GC_FIND_START().
                   1494:  - Added win32 recognition code in configure.in.  Changed some of the
                   1495:    dllimport/export defines in gc.h.  (Thanks to Adam Megacz.)
                   1496:  - GC_malloc_many didn't set hb_last_reclaimed when it called
                   1497:    GC_reclaim_generic.  (I'm not sure this matters much, but ...)
                   1498:  - Allocating uncollectable objects with debug information sometimes
                   1499:    allocated objects that were one byte too small, since uncollectable
                   1500:    objects don't have the extra byte added at the end.  (Thanks to
                   1501:    Wink Saville for pointing this out.)
                   1502:  - Added a bit more assertion checking to make sure that gcj objects
                   1503:    on free lists never have a nonzero second word.
                   1504:  - Replaced BCC_MAKEFILE with an up-to-date one.  (Thanks to
                   1505:    Andre Leiradella.)
                   1506:  - Upgraded libtool, cinfigure.in and some related files to hopefully
                   1507:    support NetBSD/SPARC.  (Thanks to Adrian Bunk.)  Unfortunately,
                   1508:    libtool 1.4.2 seemed to be buggy due to missing quotes in several
                   1509:    "test" invocations.  Fixed those in the ltmain.sh script.
                   1510:  - Some win32-specific patches, including the introduction of
                   1511:    GC_CreateThread.  (Thanks to Adam Megacz.)
                   1512:  - Merged in gcj changes from Anthony Green to support embedded systems.
                   1513:  - Tried to consistently rename preprocessed assembly files with a capital
                   1514:    .S extension.
                   1515:  - Use alpha_mach_dep.S on ALPHA again.  It doesn't really matter, but this
                   1516:    makes our distribution consistent with the gcc one, avoiding future merge
                   1517:    problems.
                   1518:  - Move GET_MEM definition into gcconfig.h.  Include gcconfig.h slightly
                   1519:    later in gc_priv.h to avoid forward references to ptr_t.
                   1520:  - Add some testing of local allocation to test.c.
                   1521:  - Change definition of INVALID_QTID in specific.h.  The -1 value was used
                   1522:    inconsistently, and too likely to collide with a valid stack address.
                   1523:    Some general clean-up of specific.[ch].  Added assertions.  (Thanks
                   1524:    to Michael Smith for tracking down an intermittent bug to this
                   1525:    general area.  I'm not sure it has been squashed yet, however.)
                   1526:  - On Pthread systems it was not safe to call GC_malloc() between fork()
                   1527:    and exec().  According to the applicable standards, it doesn't appear
                   1528:    to be safe to call malloc() or many other libc functions either, thus
                   1529:    it's not clear this is fixable.  Added experimental support for
                   1530:    -DHANDLE_FORK in linux_threads.c which tries to support it.  It may
                   1531:    succeed if libc does the right thing.  I'm not sure whether it does.
                   1532:    (Thanks to Kenneth Schalk for pointing out this issue.)
                   1533:  - Documented thread local allocation primitives to require an
                   1534:    explicit GC_init call.  GC_init_parallel is no longer declared to
                   1535:    be a constructor function, since that isn't portable and often
                   1536:    seems to lead to initialization order problems.
                   1537:  - Changed gc_cpp.cc and gc_cpp.h in one more attempt to make them
                   1538:    compatible with Visual C++ 6.  (Thanks to Wink Saville for the
                   1539:    patch.)
                   1540:  - Some more patches for Linux on HP PA-RISC.
                   1541:  - Added include/gc_allocator.h.  It implements (hopefully) standard
                   1542:    conforming (as opposed to SGI-style) allocators that allocate
                   1543:    collectable (gc_allocator) or GC-traceable, but not collectable
                   1544:    (traceable_allocator) objects.  This borrows heavily from libstc++,
                   1545:    which borrows heavily from the SGI implementation, this part of
                   1546:    which was written by Matt Austern.  Changed test_cpp.cc to very
                   1547:    minimally test this.
                   1548:  - On Linux/X86, retry mmap with a different start argument.  That should
                   1549:    allow the collector to use more (closer to 3GB) of the address space.
                   1550:    (Thanks to Jeffrey Mark Siskind for tracking this down.)
                   1551:  - Force 64 bit alignment with GCJ support.  (Reflects Bryce McKinley's
                   1552:    patch to the gcc tree.)
                   1553:  - Refined the choice of sa_handler vs. sa_sigaction in GC_dirty_init
                   1554:    to accomodate some glibc5 systems.  (Thanks to Dan Fandrich for the
                   1555:    patch.)
                   1556:  - Compensated for the fact that current versions of glibc set
                   1557:    __libc_stack_end incorrectly on Linux/IA64 while initialization code
                   1558:    is running.  This could cause the collector to miss 16 bytes of
                   1559:    the memory stack if GC_malloc or friends where called before main().
                   1560:  - Mostly integrated Takis Psarogiannakopoulos' port to DG/UX Inix 86.
                   1561:    This will probably take another iteration to work, since his
                   1562:    patch conflicted with the libtool upgrade.
                   1563:  - Added README.arm.cross containing some information about cross-
                   1564:    compiling to an ARM processor from Margaret Fleck.
                   1565:
                   1566: Since 6.1alpha4:
                   1567:  - Added GC_finalizer_mem_freed, and changed some of the code that
                   1568:    decided on heap expansion to look at it.  Memory explicitly
                   1569:    deallocated by finalizers essentially needs to be counted as reclaimed
                   1570:    by the GC.  Otherwise there are cases in which the heap can grow
                   1571:    unboundedly.  (Thanks to Mark Reichert for the test case.)
                   1572:  - Integrated Adam Megacz patches to not scan dynamic libraries if
                   1573:    we are compiling with gcc on win32.  Otherwise we need structured
                   1574:    exception handling to deal with asynchronously unmapped root
                   1575:    segments, and gcc doesn't directly support that.
                   1576:  - Integrated Anthony Green's patch to support Wine.
                   1577:  - GC_OPERATOR_NEW_ARRAY was misspelled OPERATOR_NEW_ARRAY in several
                   1578:    places, including gc_cpp.cc.  (Thanks to Wink Saville for pointing
                   1579:    this out.)
                   1580:  - Integrated Loren James Rittle's Alpha FreeBSD patches.  In
                   1581:    response to Richard Henderson's suggestion, these also
                   1582:    changed the declarations of symbols like _end on many platforms to
                   1583:    that they wouldn't mistakenly be declared as short data symbols.
                   1584:  - Integrated changes from the Debian distribution.  (Thanks to Ryan Murray
                   1585:    for pointing these out.)  Fix C++ comments in POWERPC port.  Add ARM32
                   1586:    incremental GC support.  Get rid of USE_GENERIC_PUSH_REGS for alpha/Linux,
                   1587:    this time for real.  Use va_copy to get rid of cord printf problems
                   1588:    (finally).
                   1589:  - Close file descriptor used to count cpus.  Thanks to Jeff Sturm for
                   1590:    pointing out the omission.
                   1591:  - Don't just drop gcj free lists in GC_start_reclaim, since that can
                   1592:    eventually cause the marker to see a bogus mark descriptor in the
                   1593:    dropped objects.  The usual symptom was a very intermittent segmentation
                   1594:    fault in the marker.  This mattered only if one of the GC_gcj_malloc
                   1595:    variants was used.  (Thanks to Michael Smith, Jeff Sturm, Bryce
                   1596:    McKinley and Tom Tromey for helping to track this down.)
                   1597:  - Fixed Linux and Solaris/64 SPARC configuration.  (Thanks to David Miller,
                   1598:    Jeff Sturm, Tom Tromey, and Christian Joensson.)
                   1599:  - Fixed a typo in strdup definition.  (Thanks to Gerard A Allan.)
                   1600:  - Changed Makefile.direct to invoke $(CC) to assemble alpha_mach_dep.S.
                   1601:    This is needed on Linux.  I'm not sure whether it's better or worse
                   1602:    on Tru64.
                   1603:  - Changed gc_cpp.h once more to declare operator new and friends only in
                   1604:    a Microsoft environment.  This may need further fine tuning.  (Thanks to
                   1605:    Johannes Schmidt for pointing out that the older code breaks on gcc3.0.4.)
                   1606:  - Don't ever override strdup if it's already macro defined.  (Thanks to
                   1607:    Adnan Ali for pointing out the problem.)
                   1608:  - Changed gc_cpp.h yet again to also overload placement new.  Due to the
                   1609:    C++ overloading rules, the other overloaded new operations otherwise hide
                   1610:    placement new, which causes many STL uses to break.  (Thanks to Reza
                   1611:    Shahidi for reporting this, and to Matt Austern for proposing a fix.)
                   1612:  - Integrated cygwin pthreads support from Dan Bonachea.
                   1613:  - Turn on DYNAMIC_LOADING for NetBSD.  (Thanks to Krister Walfridsson.)
                   1614:  - Changed printing code to print more complete GC times.
                   1615:  - Applied Mark Mitchell's Irix patch to correct some bitrot.
                   1616:  - Clarified which object-printing routines in dbg_mlc.c should hold
                   1617:    the allocation lock.  Restructured the code to allow reasonable object
                   1618:    printing with -DREDIRECT_MALLOC.
                   1619:  - Fix the Linux mmap code to always start with 0x1000 as the initial hint.
                   1620:    Minor patches for 64-bit AIX, particularly to STACKBOTTOM.
                   1621:    (Thanks again to Jeffrey Mark Siskind.)
                   1622:  - Renamed "SUSPENDED" flag for Solaris threads support to avoid a conflict
                   1623:    with a system header. (Thanks to Philp Brown.)
1.2       noro     1624:  - Cause win32_threads.c to handle an out of range stack pointer correctly,
                   1625:    though currently with a warning.  (Thanks to Jonathan Clark for
                   1626:    observing that win32 applications may temporarily use the stack
                   1627:    pointer for other purposes, and suggesting a fix.  Unfortunately, it's
                   1628:    not clear that there is a complete solution to this problem.)
                   1629:
                   1630: Since 6.1alpha5:
                   1631:  - Added GC_MAXIMUM_HEAP_SIZE environment variable.
                   1632:  - Fix configure.in for MIPS/LINUX. (Thanks to H.J. Lu.)
                   1633:  - Double page hash table size for -DLARGE_CONFIG.
                   1634:  - Integrated Bo Thorsen's X86-64 support.
                   1635:  - STACKBOTTOM definition for LINUX/MIPS was partially changed back.
                   1636:    (Thanks to H.J. Lu and Hiroshi Kawashima for resolving this.)
                   1637:  - Replaced all occurrences of LINUX_DATA_START in gcconfig.h with
                   1638:    SEARCH_FOR_DATA_START.  It doesn't hurt to falll back to a search.
                   1639:    And __data_start doesn't seem to get defined correctly of the GC
                   1640:    library is loaded with LD_PRELOAD, e.g. for leak detection.
                   1641:  - If the GC_find_leak environment variable is set, do a
                   1642:    atexit(GC_gcollect) to give us at least one chance to detect leaks.
                   1643:    This may report some very benign leaks, but ...
                   1644:  - Addeded REDIRECT_FREE.  It's necessary if we want leak detection with
                   1645:    LD_PRELOAD.
                   1646:  - Defer printing of leaked objects, as for smashed objects.
                   1647:  - Fixed process and descriptor leak in GC_print_callers.  Try for
                   1648:    line number even if we got function name.)
                   1649:  - Ported parallel GC support and thread local allocation to Alpha.
                   1650:    Not yet well-tested.
                   1651:  - Added GC_DUMP_REGULARLY and added finalization statistics to GC_dump().
                   1652:  - Fixed Makefile.am to mention alpha_mach_dep.S instead of the defunct
                   1653:    alpha_mach_dep.s.  (Thanks to Fergus Henderson.)
                   1654:  - Incorporated a change to new_gc_alloc.h, suggested by Johannes Schmidt,
                   1655:    which should make it work with gcc3.1.  (I would still like to encourage
                   1656:    use of gc_allocator.h instead.)
                   1657:  - Use alpha_mach_dep.S only on Linux.  (It's not clear that this is
                   1658:    optimal, but it otherwise didn't build on Tru64.  Thanks to Fergus
                   1659:    Henderson.)
                   1660:  - Added ifdef to guard free() in os_dep.c.  Otherwise we get a
                   1661:    compilation error on Irix.  (Thanks to Dai Sato.)
                   1662:  - Added an experimental version of GC_memalign to mallocx.c.  This can't
                   1663:    always work, since we don't handle alignment requests in the hblk-level
                   1664:    allocator, and we can't handle arbitrary pointer displacements unless
                   1665:    GC_all_interior_pointers is enabled.  But it should work for alignment
                   1666:    requests up to HBLKSIZE.  This is not yet documented in the standard
                   1667:    places.
                   1668:  - Finally debugged the OSF1/Tru64 thread support.  This needs more testing,
                   1669:    since I needed to add a somewhat unconvincing workaround for signal
                   1670:    delivery issues that I don't yet completely understand.  But it does
                   1671:    pass my tests, even in parallel GC mode.  Incremental GC support is
                   1672:    disabled if thread support is enabled, due to the signal issues.
                   1673:  - Eliminated name-space-incorrect definition of _cdecl from gc_cpp.h.
                   1674:  - Added GC_debug_malloc_replacement and GC_debug_realloc_replacement
                   1675:    declarations to gc.h.  On IA64, this is required for REDIRECT_MALLOC
                   1676:    to work correctly with these.
                   1677:  - Fixed Linux USE_PROC_FOR_LIBRARIES to work with a 64-bit /proc format.
                   1678:
                   1679: Since 6.1:
                   1680:  - Guard the test for GC_DUMP_REGULARLY in misc.c with
                   1681:    "#ifndef NO_DEBUGGING".  Otherwise it fails to build with NO_DEBUGGING
                   1682:    defined.  (Thanks to Manuel Serrano.)
                   1683:  - Message about retrying suspend signals was incorrectly generated even when
                   1684:    flag was not set.
                   1685:  - Cleaned up MACOSX/NEXT root registration code.  There was apparently a
                   1686:    separate ifdef case in GC_register_data_segments() for no reason.
                   1687:  - Removed MPROTECT_VDB for MACOSX port, based on one negative report.
                   1688:  - Arrange for gc.h and friends to be correctly installed with GNU-style
                   1689:    "make install".
                   1690:  - Enable the GNU-style build facility include C++ support in the library
                   1691:    with --enable-cplusplus. (Thanks to Thomas Maier for some of the patch.)
                   1692:  - Mark from GC_thread_key in linux_threads.c, in case that's allocated
                   1693:    from the garbage collected heap, as it is with our own thread-specific
                   1694:    storage implementation.  (Thanks to Jeff Sturm.)
                   1695:  - Mark all free list header blocks if they are heap allocated.  This avoids
                   1696:    some unnecessary tracing.  And it remains correct if we clear the
                   1697:    root set. (Thanks to Jeff Sturm for identifying the bug.)
                   1698:  - Improved S390/Linux support.  Add S390/Linux 64-bit support.  (Thanks
                   1699:    to Ulrich Weigand.)
                   1700:  - Corrected the spelling of GC_{M,C}ALLOC_EXPLICTLY_TYPED to
                   1701:    GC_{M,C}ALLOC_EXPLICITLY_TYPED in gc_typed.h.  This is technically
                   1702:    an interface change.  Based on the fact that nobody reported this,
                   1703:    I suspect/hope there were no clients.
                   1704:  - Cleaned up gc_typed.h so that (1) it adds an extern "C" declaration
                   1705:    when appropriate, (2) doesn't generate references to undefined internal
                   1706:    macros, and (3) allows easier manual construction of descriptors.
                   1707:  - Close the file descriptor used by GC_print_address_map().
                   1708:  - Set the "close-on-exec" bit for various file descriptors maintained
                   1709:    for the collector's internal use.
                   1710:  - Added a hack to find memory segments owned by the system allocator
                   1711:    under win32.  Based on my tests, this tends to eventually find all
                   1712:    segments, though it may take a while.  There appear to be cleaner,
                   1713:    but slower solutions under NT/XP.  But they rely on an API that's
                   1714:    unsupported under 9X.
                   1715:  - Changed Linux PowerPC stack finding to LINUX_STACKBOTTOM.  (Thanks
                   1716:    to Akira Tagoh for pointing out that HEURISTIC1 doesn't work on
                   1717:    64-bit kernels.)
                   1718:  - Added GC_set_free_space_divisor to avoid some Windows dll issues.
                   1719:  - Added FIXUP_POINTER, POINTER_SHIFT, POINTER_MASK to allow preprocessing
                   1720:    of candidate pointers for tagging, etc.
                   1721:  - Always lock around GC_notify_full_gc().  Simplified code for
                   1722:    invoking GC_notify_full_gc().
                   1723:  - Changed the way DATASTART is defined on FreeBSD to be robust against
                   1724:    an unmapped page after etext.  (Thanks to Hironori Sakamoto for
                   1725:    tracking down the intermittent failure.)
                   1726:  - Made GC_enable() and GC_disable() official.  Deprecated direct update
                   1727:    of GC_dont_gc.  Changed GC_gcollect to be a noop when garbage collection
                   1728:    is disabled.
                   1729:  - Call GC_register_dynamic_libraries before stopping the world on Linux,
                   1730:    in order to avoid a potential deadlock due to the dl_iterate_phdr lock.
                   1731:  - Introduced a more general mechanism for platform-dependent code to
                   1732:    decide whether the main data segment should be handled separately
                   1733:    from dynamic libraries, or registered by GC_register_dynamic_libraries.
                   1734:    The latter is more reliable and easier on Linux with dl_iterate_phdr.
                   1735:
                   1736: Since 6.2alpha1:
                   1737:  - Fixed the completely broken FreeBSD code in 6.2alpha1.  (Thanks to
                   1738:    Hironori Sakamoto for the patch.)
                   1739:  - Changed IRIX reference in dbg_mlc.c to IRIX5. (Thanks to Marcus Herbert.)
                   1740:  - Attempted to work around the problems with .S filenames and the SGI
                   1741:    compiler.  (Reported by several people. Untested.)
                   1742:  - Worked around an HP/UX make issue with the GNU-style build process.
                   1743:  - Fixed the --enable-cplusplus build machinery to allow builds without
                   1744:    a C++ compiler.  (That was always the intent ...)
                   1745:  - Changed the debugging allocation macros to explicitly pass the return
                   1746:    address for Linux and XXXBSD on hardware for which we can't get stack
                   1747:    traces.  Use __builtin_return_address(0) to generate it when possible.
                   1748:    Some of the configuration work was cleaned up (good) and moved to gc.h
                   1749:    (bad, but necessary).  This should make leak detection more useful
                   1750:    on a number of platforms.  (Thanks to Fabian Thylman for the suggestion.)
                   1751:  - Fixed compilation problems in dbg_mlc.c with GC_ADD_CALLER.
                   1752:  - Bumped revision number for dynamic library.
                   1753:
                   1754: Since 6.2alpha2:
                   1755:  - Don't include execinfo.h in os_dep.c when it's not needed, and may not exist.
                   1756:
                   1757: Since 6.2alpha3:
                   1758:  - Use LINUX_STACKBOTTOM for >= glibc2.2 on Linux/MIPS.  (See Debian bug
                   1759:    # 177204)
                   1760:  - Integrated Jeff Sturm and Jesse Rosenstock's MACOSX threads patches.
                   1761:  - Integrated Grzegorz Jakacki's substantial GNU build patch.  "Make dist"
                   1762:    should now work for the GNU build process.  Documentation files
                   1763:    are installed under share/gc.
                   1764:  - Tweaked gc_cpp.h to again support the Borland compiler.  (Thanks to
                   1765:    Rene Girard for pointing out the problems.)
                   1766:  - Updated BCC_MAKEFILE (thanks to Rene Girard).
                   1767:  - Added GC_ASSERT check for minimum thread stack size.
                   1768:  - Added --enable-gc-assertions.
                   1769:  - Added some web documentation to the distribution.  Updated it in the
                   1770:    process.
                   1771:  - Separate gc_conf_macros.h from gc.h.
                   1772:  - Added generic GC_THREADS client-defined macro to set the appropriate
                   1773:    GC_XXX_THREADS internal macro.  (gc_config_macros.h.)
                   1774:  - Add debugging versions of _ignore_off_page allocation primitves.
                   1775:  - Moved declarations of GC_make_closure and GC_debug_invoke_finalizer
                   1776:    from gc.h to gc_priv.h.
                   1777:  - Reset GC_fail_count even if only a small allocation succeeds.
                   1778:  - Integrated Brian Alliet's patch for dynamic library support on Darwin.
                   1779:  - gc_cpp.h's gc_cleanup destructor called GC_REGISTER_FINALIZER_IGNORE_SELF
                   1780:    when it should have called the lower case version, since it was
                   1781:    explicitly computing a base pointer.
                   1782:
                   1783: Since 6.2alpha4:
                   1784:  - GC_invoke_finalizers could, under rare conditions, set
                   1785:    GC_finalizer_mem_freed to an essentially random value.  This could
                   1786:    possibly cause unbounded heap growth for long-running applications
                   1787:    under some conditions.  (The bug was introduced in 6.1alpha5, and
                   1788:    is not in gcc3.3.  Thanks to Ben Hutchings for finding it.)
                   1789:  - Attempted to sanitize the various DLL macros.  GC_USE_DLL disappeared.
                   1790:    GC_DLL is used instead.  All internal tests are now on GC_DLL.
                   1791:    README.macros is now more precise about the intended meaning.
                   1792:  - Include DllMain in the multithreaded win32 version only if the
                   1793:    collector is actually built as a dll.  (Thanks to Mohan Embar for
                   1794:    a version of the patch.)
                   1795:  - Hide the cygwin threadAttach/Detach functions.  They were violating our
                   1796:    namespace rules.
                   1797:  - Fixed an assertion in GC_check_heap_proc.  Added GC_STATIC_ASSERT.
                   1798:    (Thanks again to Ben Hutchings.)
                   1799:  - Removed some obsolete definitions for Linux/PowerPC in gcconfig.h.
                   1800:  - CORD_cat was not rebalancing unbalanced trees in some cases, violating
                   1801:    a CORD invariant.  Also tweaked the rebalancing rule for
                   1802:    CORD_cat_char_star.  (Thanks to Alexandr Petrosian for the bug report
                   1803:    and patch.)
                   1804:  - Added hand-coded structured exception handling support to mark.c.
                   1805:    This should enable support of dynamic libraries under win32 with
                   1806:    gcc-compiled code.  (Thanks to Ranjit Mathew for the patch.)
                   1807:    Turned on dynamic library scanning for win32/gcc.
                   1808:  - Removed some remnants of read wrapping.  (Thanks to Kenneth Schalk.)
                   1809:    GC_USE_LD_WRAP ws probably broken in recent versions.
                   1810:  - The build could fail on some platforms since gcconfig.h could include
                   1811:    declarations mentioning ptr_t, which was not defined, e.g. when if_mach
                   1812:    was built.  (Thanks to Yann Dirson for pointing this out.)  Also
                   1813:    cleaned up tests for GC_PRIVATE_H in gcconfig.h a bit.
                   1814:  - The GC_LOOP_ON_ABORT environment variable interfered with incremental
                   1815:    collection, since the write fault handler was erroneously overridden.
                   1816:    Handlers are now set up in the correct order.
                   1817:  - It used to be possible to call GC_mark_thread_local_free_lists() while
                   1818:    the world was not stopped during an incremental GC.  This was not safe.
                   1819:    Fortunately, it was also unnecessary.  Added GC_world_stopped flag
                   1820:    to avoid it.  (This caused occasional crashes in GC_set_fl_marks
                   1821:    with thread local allocation and incremental GC.  This probably happened
                   1822:    primarily on old, slow multiprocessors.)
                   1823:  - Allowed overriding of MAX_THREADS in win32_threads.c from the build
                   1824:    command line.  (Patch from Yannis Bres.)
                   1825:  - Taught the IA64/linux code to determine the register backing store base from
                   1826:    /proc/self/maps after checking the __libc symbol, but before guessing.
                   1827:    (__libc symbols are on the endangered list, and the guess is likely to not
                   1828:    always be right for 2.6 kernels.)  Restructured the code to read and parse
                   1829:    /proc/self/maps so it only exists in one place (all platforms).
                   1830:  - The -DUSE_PROC_FOR_LIBRARIES code was broken on Linux.  It claimed that it
                   1831:    also registered the main data segment, but didn't actually do so.  (I don't
                   1832:    think anyone actually uses this configuration, but ...)
                   1833:  - Made another attempt to get --enablecplusplus to do the right thing.
                   1834:    Since there are unavoidable problems with C programs linking against a
                   1835:    dynamic library that includes C++ code, I separated out the c++ code into
                   1836:    libgccpp.
                   1837:
                   1838: Since 6.2alpha5:
                   1839:  - There was extra underscore in the name of GC_save_registers_in_stack
                   1840:    for NetBSD/SPARC.  (Thanks to Jaap Boender for the patch.)
                   1841:  - Integrated Brian Alliet's patch for Darwin.  This restructured the
                   1842:    linuxthreads/pthreads support to separate generic pthreads support
                   1843:    from more the system-dependent thread-stopping code.  I believe this
                   1844:    should make it easier to eliminate the code duplication between
                   1845:    pthreads platforms in the future.  The patch included some other
                   1846:    code cleanups.
                   1847:  - Integrated Dan Bonachea's patch to support AIX threads.  This required
                   1848:    substantial manual integration, mostly due to conflicts with other
                   1849:    recent threads changes.  It may take another iteration to
                   1850:    get it to work.
                   1851:  - Removed HPUX/PA-RISC support from aix_irix_threads.c.  It wasn't used
                   1852:    anyway and it cluttered up the code.  And anything we can do to migrate
                   1853:    towards generic pthreads support is a good thing.
                   1854:  - Added a more explicit test for tracing of function arguments to test.c.
                   1855:    (Thanks to Dan Grayson.)
                   1856:  - Added Akira Tagoh's PowerPC64 patch.
                   1857:  - Fixed some bit rot in the Cygwin port.  (Thanks to Dan Bonachea for
                   1858:    pointing it out.)  Gc.h now includes just windows.h, not winbase.h.
                   1859:  - Declared GC_save_regs_in_stack() in gc_priv.h.  Remove other declarations.
                   1860:  - Changed --enable-cplusplus to use automake consitionals.  The old way
                   1861:    confused libtool.  "Make install" didn't work correctly for the old version.
                   1862:    Previously --enable-cplusplus was broken on cygwin.
                   1863:  - Changed the C version of GC_push_regs to fail at compile time if it is
                   1864:    generated with an empty body.  This seems to have been the cause of one
                   1865:    or two subtle failures on unusual platforms.  Those failures should
                   1866:    now occur at build time and be easily fixable.
1.3     ! noro     1867:
        !          1868: Since 6.2alpha6:
        !          1869:  - Integrated a second round of Irix/AIX patches from Dan Bonachea.
        !          1870:    Renamed mips_sgi_mach_dep.S back to mips_sgi_mach_dep.s, since it requires
        !          1871:    the Irix assembler to do the C preprocessing; gcc -E doesn't work.
        !          1872:  - Fixed Makefile.direct for DARWIN.  (Thanks to Manuel Serrano.)
        !          1873:  - There was a race between GC_pthread_detach and thread exit that could
        !          1874:    result in a thread structure being deallocated by GC_pthread_detach
        !          1875:    eventhough it was still needed by the thread exit code.  (Thanks to
        !          1876:    Dick Porter for the small test case that allowed this to be debugged.)
        !          1877:  - Fixed version parsing for non-alpha versions in acinclude.m4 and
        !          1878:    version checking in version.h.
1.2       noro     1879:
1.1       noro     1880:
                   1881: To do:
1.2       noro     1882:  - A dynamic libgc.so references dlopen unconditionally, but doesn't link
                   1883:    against libdl.
                   1884:  - GC_proc_fd for Solaris is not correctly updated in response to a
                   1885:    fork() call.  Thus incremental collection in the child won't work
                   1886:    correctly.  (Thanks to Ben Cottrell for pointing this out.)
1.1       noro     1887:  - --enable-redirect-malloc is mostly untested and known not to work
                   1888:    on some platforms.
                   1889:  - There seem to be outstanding issues on Solaris/X86, possibly with
                   1890:    finding the data segment starting address.  Information/patches would
                   1891:    be appreciated.
                   1892:  - Very large root set sizes (> 16 MB or so) could cause the collector
                   1893:    to abort with an unexpected mark stack overflow.  (Thanks again to
                   1894:    Peter Chubb.)  NOT YET FIXED.  Workaround is to increase the initial
                   1895:    size.
                   1896:  - The SGI version of the collector marks from mmapped pages, even
                   1897:    if they are not part of dynamic library static data areas.  This
                   1898:    causes performance problems with some SGI libraries that use mmap
                   1899:    as a bitmap allocator.  NOT YET FIXED.  It may be possible to turn
                   1900:    off DYNAMIC_LOADING in the collector as a workaround.  It may also
                   1901:    be possible to conditionally intercept mmap and use GC_exclude_static_roots.
                   1902:    The real fix is to walk rld data structures, which looks possible.
                   1903:  - Incremental collector should handle large objects better.  Currently,
                   1904:    it looks like the whole object is treated as dirty if any part of it
                   1905:    is.
                   1906:

FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>