os_dep.c

/*
 * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
 * Copyright (c) 1991-1995 by Xerox Corporation.  All rights reserved.
 * Copyright (c) 1996-1999 by Silicon Graphics.  All rights reserved.
 * Copyright (c) 1999 by Hewlett-Packard Company.  All rights reserved.
 *
 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
 * OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
 *
 * Permission is hereby granted to use or copy this program
 * for any purpose,  provided the above notices are retained on all copies.
 * Permission to modify the code and to distribute modified code is granted,
 * provided the above notices are retained, and a notice that the code was
 * modified is included with the above copyright notice.
 */

# include "private/gc_priv.h"

# if defined(THREADS) && defined(MPROTECT_VDB)
#   include "atomic_ops.h"
# endif

# if defined(LINUX) && !defined(POWERPC)
#   include <linux/version.h>
#   if (LINUX_VERSION_CODE <= 0x10400)
      /* Ugly hack to get struct sigcontext_struct definition.  Required      */
      /* for some early 1.3.X releases.  Will hopefully go away soon. */
      /* in some later Linux releases, asm/sigcontext.h may have to   */
      /* be included instead.                                         */
#     define __KERNEL__
#     include <asm/signal.h>
#     undef __KERNEL__
#   else
      /* Kernels prior to 2.1.1 defined struct sigcontext_struct instead of */
      /* struct sigcontext.  libc6 (glibc2) uses "struct sigcontext" in     */
      /* prototypes, so we have to include the top-level sigcontext.h to    */
      /* make sure the former gets defined to be the latter if appropriate. */
#     include <features.h>
#     if 2 <= __GLIBC__
#       if 2 == __GLIBC__ && 0 == __GLIBC_MINOR__
          /* glibc 2.1 no longer has sigcontext.h.  But signal.h        */
          /* has the right declaration for glibc 2.1.                   */
#         include <sigcontext.h>
#       endif /* 0 == __GLIBC_MINOR__ */
#     else /* not 2 <= __GLIBC__ */
        /* libc5 doesn't have <sigcontext.h>: go directly with the kernel   */
        /* one.  Check LINUX_VERSION_CODE to see which we should reference. */
#       include <asm/sigcontext.h>
#     endif /* 2 <= __GLIBC__ */
#   endif
# endif
# if !defined(OS2) && !defined(PCR) && !defined(AMIGA) && !defined(MACOS) \
    && !defined(MSWINCE)
#   include <sys/types.h>
#   if !defined(MSWIN32)
#       include <unistd.h>
#   endif
# endif

# include <stdio.h>
# if defined(MSWINCE)
#   define SIGSEGV 0 /* value is irrelevant */
# else
#   include <signal.h>
# endif

#if defined(UNIX_LIKE) || defined(CYGWIN32)
# include <fcntl.h>
#endif

#if defined(LINUX) || defined(LINUX_STACKBOTTOM)
# include <ctype.h>
#endif

/* Blatantly OS dependent routines, except for those that are related   */
/* to dynamic loading.                                                  */

#ifdef AMIGA
# define GC_AMIGA_DEF
# include "extra/AmigaOS.c"
# undef GC_AMIGA_DEF
#endif

#if defined(MSWIN32) || defined(MSWINCE) || defined(CYGWIN32)
# define WIN32_LEAN_AND_MEAN
# define NOSERVICE
# include <windows.h>
  /* It's not clear this is completely kosher under Cygwin.  But it     */
  /* allows us to get a working GC_get_stack_base.                      */
#endif

#ifdef MACOS
# include <Processes.h>
#endif

#ifdef IRIX5
# include <sys/uio.h>
# include <malloc.h>   /* for locking */
#endif

#if defined(LINUX) || defined(FREEBSD) || defined(SOLARIS) || defined(IRIX5) \
        || ((defined(USE_MMAP) || defined(USE_MUNMAP)) \
        && !defined(MSWIN32) && !defined(MSWINCE))
# define MMAP_SUPPORTED
#endif

#if defined(MMAP_SUPPORTED) || defined(ADD_HEAP_GUARD_PAGES)
# if defined(USE_MUNMAP) && !defined(USE_MMAP)
#   error "invalid config - USE_MUNMAP requires USE_MMAP"
# endif
# include <sys/types.h>
# include <sys/mman.h>
# include <sys/stat.h>
# include <errno.h>
#endif

#ifdef DARWIN
/* for get_etext and friends */
#include <mach-o/getsect.h>
#endif

#ifdef DJGPP
  /* Apparently necessary for djgpp 2.01.  May cause problems with      */
  /* other versions.                                                    */
  typedef long unsigned int caddr_t;
#endif

#ifdef PCR
# include "il/PCR_IL.h"
# include "th/PCR_ThCtl.h"
# include "mm/PCR_MM.h"
#endif

#if !defined(NO_EXECUTE_PERMISSION)
# define OPT_PROT_EXEC PROT_EXEC
#else
# define OPT_PROT_EXEC 0
#endif

#if defined(LINUX) && \
    (defined(USE_PROC_FOR_LIBRARIES) || defined(IA64) || !defined(SMALL_CONFIG))
# define NEED_PROC_MAPS
#endif

#ifdef NEED_PROC_MAPS
/* We need to parse /proc/self/maps, either to find dynamic libraries,  */
/* and/or to find the register backing store base (IA64).  Do it once   */
/* here.                                                                */

#define READ read

/* Repeatedly perform a read call until the buffer is filled or */
/* we encounter EOF.                                            */
ssize_t GC_repeat_read(int fd, char *buf, size_t count)
{
    ssize_t num_read = 0;
    ssize_t result;

    ASSERT_CANCEL_DISABLED();
    while (num_read < count) {
        result = READ(fd, buf + num_read, count - num_read);
        if (result < 0) return result;
        if (result == 0) break;
        num_read += result;
    }
    return num_read;
}

#ifdef THREADS
/* Determine the length of a file by incrementally reading it into a    */
/* This would be silly to use on a file supporting lseek, but Linux     */
/* /proc files usually do not.                                          */
STATIC size_t GC_get_file_len(int f)
{
    size_t total = 0;
    ssize_t result;
#   define GET_FILE_LEN_BUF_SZ 500
    char buf[GET_FILE_LEN_BUF_SZ];

    do {
        result = read(f, buf, GET_FILE_LEN_BUF_SZ);
        if (result == -1) return 0;
        total += result;
    } while (result > 0);
    return total;
}

STATIC size_t GC_get_maps_len(void)
{
    int f = open("/proc/self/maps", O_RDONLY);
    size_t result = GC_get_file_len(f);
    close(f);
    return result;
}
#endif

/*
 * Copy the contents of /proc/self/maps to a buffer in our address space.
 * Return the address of the buffer, or zero on failure.
 * This code could be simplified if we could determine its size
 * ahead of time.
 */
char * GC_get_maps(void)
{
    int f;
    int result;
    static char init_buf[1];
    static char *maps_buf = init_buf;
    static size_t maps_buf_sz = 1;
    size_t maps_size, old_maps_size = 0;

    /* The buffer is essentially static, so there must be a single client. */
    GC_ASSERT(I_HOLD_LOCK());

    /* Note that in the presence of threads, the maps file can  */
    /* essentially shrink asynchronously and unexpectedly as    */
    /* threads that we already think of as dead release their   */
    /* stacks.  And there is no easy way to read the entire     */
    /* file atomically.  This is arguably a misfeature of the   */
    /* /proc/.../maps interface.                                */

    /* Since we dont believe the file can grow                  */
    /* asynchronously, it should suffice to first determine     */
    /* the size (using lseek or read), and then to reread the   */
    /* file.  If the size is inconsistent we have to retry.     */
    /* This only matters with threads enabled, and if we use    */
    /* this to locate roots (not the default).                  */

    /* Determine the initial size of /proc/self/maps.           */
    /* Note that lseek doesn't work, at least as of 2.6.15.     */
#   ifdef THREADS
        maps_size = GC_get_maps_len();
        if (0 == maps_size) return 0;
#   else
        maps_size = 4000;       /* Guess */
#   endif

    /* Read /proc/self/maps, growing maps_buf as necessary.     */
    /* Note that we may not allocate conventionally, and        */
    /* thus can't use stdio.                                    */
        do {
            while (maps_size >= maps_buf_sz) {
              /* Grow only by powers of 2, since we leak "too small" buffers. */
              while (maps_size >= maps_buf_sz) maps_buf_sz *= 2;
              maps_buf = GC_scratch_alloc(maps_buf_sz);
#             ifdef THREADS
                /* Recompute initial length, since we allocated.        */
                /* This can only happen a few times per program         */
                /* execution.                                           */
                maps_size = GC_get_maps_len();
                if (0 == maps_size) return 0;
#             endif
              if (maps_buf == 0) return 0;
            }
            GC_ASSERT(maps_buf_sz >= maps_size + 1);
            f = open("/proc/self/maps", O_RDONLY);
            if (-1 == f) return 0;
#           ifdef THREADS
              old_maps_size = maps_size;
#           endif
            maps_size = 0;
            do {
                result = GC_repeat_read(f, maps_buf, maps_buf_sz-1);
                if (result <= 0) return 0;
                maps_size += result;
            } while (result == maps_buf_sz-1);
            close(f);
#           ifdef THREADS
              if (maps_size > old_maps_size) {
                GC_err_printf("Old maps size = %lu, new maps size = %lu\n",
                              (unsigned long)old_maps_size,
                              (unsigned long)maps_size);
                ABORT("Unexpected asynchronous /proc/self/maps growth: "
                      "Unregistered thread?");
              }
#           endif
        } while (maps_size >= maps_buf_sz || maps_size < old_maps_size);
                /* In the single-threaded case, the second clause is false.     */
        maps_buf[maps_size] = '\0';

    /* Apply fn to result. */
        return maps_buf;
}

/*
 *  GC_parse_map_entry parses an entry from /proc/self/maps so we can
 *  locate all writable data segments that belong to shared libraries.
 *  The format of one of these entries and the fields we care about
 *  is as follows:
 *  XXXXXXXX-XXXXXXXX r-xp 00000000 30:05 260537     name of mapping...\n
 *  ^^^^^^^^ ^^^^^^^^ ^^^^          ^^
 *  start    end      prot          maj_dev
 *
 *  Note that since about august 2003 kernels, the columns no longer have
 *  fixed offsets on 64-bit kernels.  Hence we no longer rely on fixed offsets
 *  anywhere, which is safer anyway.
 */

/*
 * Assign various fields of the first line in buf_ptr to *start, *end,
 * *prot, *maj_dev and *mapping_name.  Mapping_name may be NULL.
 * *prot and *mapping_name are assigned pointers into the original
 * buffer.
 */
char *GC_parse_map_entry(char *buf_ptr, ptr_t *start, ptr_t *end,
                                char **prot, unsigned int *maj_dev,
                                char **mapping_name)
{
    char *start_start, *end_start, *maj_dev_start;
    char *p;
    char *endp;

    if (buf_ptr == NULL || *buf_ptr == '\0') {
        return NULL;
    }

    p = buf_ptr;
    while (isspace(*p)) ++p;
    start_start = p;
    GC_ASSERT(isxdigit(*start_start));
    *start = (ptr_t)strtoul(start_start, &endp, 16); p = endp;
    GC_ASSERT(*p=='-');

    ++p;
    end_start = p;
    GC_ASSERT(isxdigit(*end_start));
    *end = (ptr_t)strtoul(end_start, &endp, 16); p = endp;
    GC_ASSERT(isspace(*p));

    while (isspace(*p)) ++p;
    GC_ASSERT(*p == 'r' || *p == '-');
    *prot = p;
    /* Skip past protection field to offset field */
       while (!isspace(*p)) ++p; while (isspace(*p)) ++p;
    GC_ASSERT(isxdigit(*p));
    /* Skip past offset field, which we ignore */
          while (!isspace(*p)) ++p; while (isspace(*p)) ++p;
    maj_dev_start = p;
    GC_ASSERT(isxdigit(*maj_dev_start));
    *maj_dev = strtoul(maj_dev_start, NULL, 16);

    if (mapping_name == 0) {
      while (*p && *p++ != '\n');
    } else {
      while (*p && *p != '\n' && *p != '/' && *p != '[') p++;
      *mapping_name = p;
      while (*p && *p++ != '\n');
    }

    return p;
}

/* Try to read the backing store base from /proc/self/maps.             */
/* Return the bounds of the writable mapping with a 0 major device,     */
/* which includes the address passed as data.                           */
/* Return FALSE if there is no such mapping.                            */
GC_bool GC_enclosing_mapping(ptr_t addr, ptr_t *startp, ptr_t *endp)
{
  char *prot;
  ptr_t my_start, my_end;
  unsigned int maj_dev;
  char *maps = GC_get_maps();
  char *buf_ptr = maps;

  if (0 == maps) return(FALSE);
  for (;;) {
    buf_ptr = GC_parse_map_entry(buf_ptr, &my_start, &my_end,
                                 &prot, &maj_dev, 0);

    if (buf_ptr == NULL) return FALSE;
    if (prot[1] == 'w' && maj_dev == 0) {
        if (my_end > addr && my_start <= addr) {
          *startp = my_start;
          *endp = my_end;
          return TRUE;
        }
    }
  }
  return FALSE;
}

#if defined(REDIRECT_MALLOC)
/* Find the text(code) mapping for the library whose name, after        */
/* stripping the directory part, starts with nm.                        */
GC_bool GC_text_mapping(char *nm, ptr_t *startp, ptr_t *endp)
{
  size_t nm_len = strlen(nm);
  char *prot;
  char *map_path;
  ptr_t my_start, my_end;
  unsigned int maj_dev;
  char *maps = GC_get_maps();
  char *buf_ptr = maps;

  if (0 == maps) return(FALSE);
  for (;;) {
    buf_ptr = GC_parse_map_entry(buf_ptr, &my_start, &my_end,
                                 &prot, &maj_dev, &map_path);

    if (buf_ptr == NULL) return FALSE;
    if (prot[0] == 'r' && prot[1] == '-' && prot[2] == 'x') {
        char *p = map_path;
        /* Set p to point just past last slash, if any. */
          while (*p != '\0' && *p != '\n' && *p != ' ' && *p != '\t') ++p;
          while (*p != '/' && p >= map_path) --p;
          ++p;
        if (strncmp(nm, p, nm_len) == 0) {
          *startp = my_start;
          *endp = my_end;
          return TRUE;
        }
    }
  }
  return FALSE;
}
#endif /* REDIRECT_MALLOC */

#ifdef IA64
static ptr_t backing_store_base_from_proc(void)
{
    ptr_t my_start, my_end;
    if (!GC_enclosing_mapping(GC_save_regs_in_stack(), &my_start, &my_end)) {
        if (GC_print_stats) {
            GC_log_printf("Failed to find backing store base from /proc\n");
        }
        return 0;
    }
    return my_start;
}
#endif

#endif /* NEED_PROC_MAPS */

#if defined(SEARCH_FOR_DATA_START)
  /* The I386 case can be handled without a search.  The Alpha case     */
  /* used to be handled differently as well, but the rules changed      */
  /* for recent Linux versions.  This seems to be the easiest way to    */
  /* cover all versions.                                                */

# if defined(LINUX) || defined(HURD)
    /* Some Linux distributions arrange to define __data_start.  Some   */
    /* define data_start as a weak symbol.  The latter is technically   */
    /* broken, since the user program may define data_start, in which   */
    /* case we lose.  Nonetheless, we try both, preferring __data_start.*/
    /* We assume gcc-compatible pragmas.        */
#   pragma weak __data_start
    extern int __data_start[];
#   pragma weak data_start
    extern int data_start[];
# endif /* LINUX */
  extern int _end[];

  ptr_t GC_data_start = NULL;

  ptr_t GC_find_limit(ptr_t, GC_bool);

  void GC_init_linux_data_start(void)
  {

#   if defined(LINUX) || defined(HURD)
      /* Try the easy approaches first: */
      if ((ptr_t)__data_start != 0) {
          GC_data_start = (ptr_t)(__data_start);
          return;
      }
      if ((ptr_t)data_start != 0) {
          GC_data_start = (ptr_t)(data_start);
          return;
      }
#   endif /* LINUX */
    GC_data_start = GC_find_limit((ptr_t)(_end), FALSE);
  }
#endif

# ifdef ECOS

# ifndef ECOS_GC_MEMORY_SIZE
# define ECOS_GC_MEMORY_SIZE (448 * 1024)
# endif /* ECOS_GC_MEMORY_SIZE */

/* FIXME: This is a simple way of allocating memory which is            */
/* compatible with ECOS early releases.  Later releases use a more      */
/* sophisticated means of allocating memory than this simple static     */
/* allocator, but this method is at least bound to work.                */
static char ecos_gc_memory[ECOS_GC_MEMORY_SIZE];
static char *brk = ecos_gc_memory;

static void *tiny_sbrk(ptrdiff_t increment)
{
  void *p = brk;

  brk += increment;

  if (brk > ecos_gc_memory + sizeof(ecos_gc_memory))
    {
      brk -= increment;
      return NULL;
    }

  return p;
}
#define sbrk tiny_sbrk
# endif /* ECOS */

#if (defined(NETBSD) || defined(OPENBSD)) && defined(__ELF__)
  ptr_t GC_data_start = NULL;
  ptr_t GC_find_limit(ptr_t, GC_bool);
  extern char **environ;

  void GC_init_netbsd_elf(void)
  {
        /* This may need to be environ, without the underscore, for     */
        /* some versions.                                               */
    GC_data_start = GC_find_limit((ptr_t)&environ, FALSE);
  }
#endif

# ifdef OS2

# include <stddef.h>

# if !defined(__IBMC__) && !defined(__WATCOMC__) /* e.g. EMX */

struct exe_hdr {
    unsigned short      magic_number;
    unsigned short      padding[29];
    long                new_exe_offset;
};

#define E_MAGIC(x)      (x).magic_number
#define EMAGIC          0x5A4D
#define E_LFANEW(x)     (x).new_exe_offset

struct e32_exe {
    unsigned char       magic_number[2];
    unsigned char       byte_order;
    unsigned char       word_order;
    unsigned long       exe_format_level;
    unsigned short      cpu;
    unsigned short      os;
    unsigned long       padding1[13];
    unsigned long       object_table_offset;
    unsigned long       object_count;
    unsigned long       padding2[31];
};

#define E32_MAGIC1(x)   (x).magic_number[0]
#define E32MAGIC1       'L'
#define E32_MAGIC2(x)   (x).magic_number[1]
#define E32MAGIC2       'X'
#define E32_BORDER(x)   (x).byte_order
#define E32LEBO         0
#define E32_WORDER(x)   (x).word_order
#define E32LEWO         0
#define E32_CPU(x)      (x).cpu
#define E32CPU286       1
#define E32_OBJTAB(x)   (x).object_table_offset
#define E32_OBJCNT(x)   (x).object_count

struct o32_obj {
    unsigned long       size;
    unsigned long       base;
    unsigned long       flags;
    unsigned long       pagemap;
    unsigned long       mapsize;
    unsigned long       reserved;
};

#define O32_FLAGS(x)    (x).flags
#define OBJREAD         0x0001L
#define OBJWRITE        0x0002L
#define OBJINVALID      0x0080L
#define O32_SIZE(x)     (x).size
#define O32_BASE(x)     (x).base

# else  /* IBM's compiler */

/* A kludge to get around what appears to be a header file bug */
# ifndef WORD
#   define WORD unsigned short
# endif
# ifndef DWORD
#   define DWORD unsigned long
# endif

# define EXE386 1
# include <newexe.h>
# include <exe386.h>

# endif  /* __IBMC__ */

# define INCL_DOSEXCEPTIONS
# define INCL_DOSPROCESS
# define INCL_DOSERRORS
# define INCL_DOSMODULEMGR
# define INCL_DOSMEMMGR
# include <os2.h>

# endif /* OS/2 */

/* Find the page size */
word GC_page_size = 0;

# if defined(MSWIN32) || defined(MSWINCE)

#   ifndef VER_PLATFORM_WIN32_CE
#     define VER_PLATFORM_WIN32_CE 3
#   endif

#   if defined(MSWINCE) && defined(THREADS)
      GC_bool GC_dont_query_stack_min = FALSE;
#   endif

    void GC_setpagesize(void)
    {
      GetSystemInfo(&GC_sysinfo);
      GC_page_size = GC_sysinfo.dwPageSize;
#     if defined(MSWINCE) && !defined(_WIN32_WCE_EMULATION)
        {
          OSVERSIONINFO verInfo;
          /* Check the current WinCE version. */
          verInfo.dwOSVersionInfoSize = sizeof(OSVERSIONINFO);
          if (!GetVersionEx(&verInfo))
            ABORT("GetVersionEx failed");
          if (verInfo.dwPlatformId == VER_PLATFORM_WIN32_CE &&
              verInfo.dwMajorVersion < 6) {
            /* Only the first 32 MB of address space belongs to the     */
            /* current process (unless WinCE 6.0+ or emulation).        */
            GC_sysinfo.lpMaximumApplicationAddress = (LPVOID)((word)32 << 20);
#           ifdef THREADS
              /* On some old WinCE versions, it's observed that         */
              /* VirtualQuery calls don't work properly when used to    */
              /* get thread current stack committed minimum.            */
              if (verInfo.dwMajorVersion < 5)
                GC_dont_query_stack_min = TRUE;
#           endif
          }
        }
#     endif
    }

# else
    void GC_setpagesize(void)
    {
#     if defined(MPROTECT_VDB) || defined(PROC_VDB) || defined(USE_MMAP)
        GC_page_size = GETPAGESIZE();
#     else
        /* It's acceptable to fake it. */
        GC_page_size = HBLKSIZE;
#     endif
    }
# endif

# if defined(MSWIN32) || defined(MSWINCE) || defined(CYGWIN32)

#ifndef CYGWIN32

# define is_writable(prot) ((prot) == PAGE_READWRITE \
                            || (prot) == PAGE_WRITECOPY \
                            || (prot) == PAGE_EXECUTE_READWRITE \
                            || (prot) == PAGE_EXECUTE_WRITECOPY)
/* Return the number of bytes that are writable starting at p.  */
/* The pointer p is assumed to be page aligned.                 */
/* If base is not 0, *base becomes the beginning of the         */
/* allocation region containing p.                              */
STATIC word GC_get_writable_length(ptr_t p, ptr_t *base)
{
    MEMORY_BASIC_INFORMATION buf;
    word result;
    word protect;

    result = VirtualQuery(p, &buf, sizeof(buf));
    if (result != sizeof(buf)) ABORT("Weird VirtualQuery result");
    if (base != 0) *base = (ptr_t)(buf.AllocationBase);
    protect = (buf.Protect & ~(PAGE_GUARD | PAGE_NOCACHE));
    if (!is_writable(protect)) {
        return(0);
    }
    if (buf.State != MEM_COMMIT) return(0);
    return(buf.RegionSize);
}

GC_API int GC_CALL GC_get_stack_base(struct GC_stack_base *sb)
{
    int dummy;
    ptr_t sp = (ptr_t)(&dummy);
    ptr_t trunc_sp = (ptr_t)((word)sp & ~(GC_page_size - 1));
    /* FIXME: This won't work if called from a deeply recursive */
    /* client code (and the committed stack space has grown).   */
    word size = GC_get_writable_length(trunc_sp, 0);
    GC_ASSERT(size != 0);
    sb -> mem_base = trunc_sp + size;
    return GC_SUCCESS;
}

#else /* CYGWIN32 */

/* An alternate version for Cygwin (adapted from Dave Korn's    */
/* gcc version of boehm-gc).                                    */
  GC_API int GC_CALL GC_get_stack_base(struct GC_stack_base *sb)
  {
    extern void * _tlsbase __asm__ ("%fs:4");
    sb -> mem_base = _tlsbase;
    return GC_SUCCESS;
  }

#endif /* CYGWIN32 */


#define HAVE_GET_STACK_BASE

/* This is always called from the main thread.  */
ptr_t GC_get_main_stack_base(void)
{
    struct GC_stack_base sb;

    GC_get_stack_base(&sb);
    GC_ASSERT((void *)&sb HOTTER_THAN sb.mem_base);
    return (ptr_t)sb.mem_base;
}

# endif /* MS Windows */

# ifdef BEOS
#   include <kernel/OS.h>
    ptr_t GC_get_main_stack_base(void)
    {
      thread_info th;
      get_thread_info(find_thread(NULL),&th);
      return th.stack_end;
    }
# endif /* BEOS */


# ifdef OS2

ptr_t GC_get_main_stack_base(void)
{
    PTIB ptib;
    PPIB ppib;

    if (DosGetInfoBlocks(&ptib, &ppib) != NO_ERROR) {
        GC_err_printf("DosGetInfoBlocks failed\n");
        ABORT("DosGetInfoBlocks failed\n");
    }
    return((ptr_t)(ptib -> tib_pstacklimit));
}

# endif /* OS2 */

# ifdef AMIGA
#   define GC_AMIGA_SB
#   include "extra/AmigaOS.c"
#   undef GC_AMIGA_SB
# endif /* AMIGA */

# if defined(NEED_FIND_LIMIT) || defined(UNIX_LIKE)

    typedef void (*handler)(int);

#   if defined(SUNOS5SIGS) || defined(IRIX5) || defined(OSF1) \
    || defined(HURD) || defined(NETBSD)
        static struct sigaction old_segv_act;
#       if defined(_sigargs) /* !Irix6.x */ || defined(HPUX) \
        || defined(HURD) || defined(NETBSD) || defined(FREEBSD)
            static struct sigaction old_bus_act;
#       endif
#   else
        static handler old_segv_handler, old_bus_handler;
#   endif

    void GC_set_and_save_fault_handler(handler h)
    {
#       if defined(SUNOS5SIGS) || defined(IRIX5)  \
        || defined(OSF1) || defined(HURD) || defined(NETBSD)
          struct sigaction      act;

          act.sa_handler        = h;
#         if 0 /* Was necessary for Solaris 2.3 and very temporary      */
               /* NetBSD bugs.                                          */
            act.sa_flags          = SA_RESTART | SA_NODEFER;
#         else
            act.sa_flags          = SA_RESTART;
#         endif

          (void) sigemptyset(&act.sa_mask);
#         ifdef GC_IRIX_THREADS
                /* Older versions have a bug related to retrieving and  */
                /* and setting a handler at the same time.              */
                (void) sigaction(SIGSEGV, 0, &old_segv_act);
                (void) sigaction(SIGSEGV, &act, 0);
#         else
                (void) sigaction(SIGSEGV, &act, &old_segv_act);
#               if defined(IRIX5) && defined(_sigargs) /* Irix 5.x, not 6.x */ \
                   || defined(HPUX) || defined(HURD) || defined(NETBSD) \
                   || defined(FREEBSD)
                    /* Under Irix 5.x or HP/UX, we may get SIGBUS.      */
                    /* Pthreads doesn't exist under Irix 5.x, so we     */
                    /* don't have to worry in the threads case.         */
                    (void) sigaction(SIGBUS, &act, &old_bus_act);
#               endif
#         endif /* GC_IRIX_THREADS */
#       else
          old_segv_handler = signal(SIGSEGV, h);
#         ifdef SIGBUS
            old_bus_handler = signal(SIGBUS, h);
#         endif
#       endif
    }
# endif /* NEED_FIND_LIMIT || UNIX_LIKE */

# if defined(NEED_FIND_LIMIT) || \
     defined(USE_PROC_FOR_LIBRARIES) && defined(THREADS)
  /* Some tools to implement HEURISTIC2 */
#   define MIN_PAGE_SIZE 256    /* Smallest conceivable page size, bytes */

    /*ARGSUSED*/
    STATIC void GC_fault_handler(int sig)
    {
        LONGJMP(GC_jmp_buf, 1);
    }

    void GC_setup_temporary_fault_handler(void)
    {
        /* Handler is process-wide, so this should only happen in */
        /* one thread at a time.                                  */
        GC_ASSERT(I_HOLD_LOCK());
        GC_set_and_save_fault_handler(GC_fault_handler);
    }

    void GC_reset_fault_handler(void)
    {
#       if defined(SUNOS5SIGS) || defined(IRIX5) \
           || defined(OSF1) || defined(HURD) || defined(NETBSD)
          (void) sigaction(SIGSEGV, &old_segv_act, 0);
#         if defined(IRIX5) && defined(_sigargs) /* Irix 5.x, not 6.x */ \
             || defined(HPUX) || defined(HURD) || defined(NETBSD) \
             || defined(FREEBSD)
              (void) sigaction(SIGBUS, &old_bus_act, 0);
#         endif
#       else
          (void) signal(SIGSEGV, old_segv_handler);
#         ifdef SIGBUS
            (void) signal(SIGBUS, old_bus_handler);
#         endif
#       endif
    }

    /* Return the first non-addressable location > p (up) or    */
    /* the smallest location q s.t. [q,p) is addressable (!up). */
    /* We assume that p (up) or p-1 (!up) is addressable.       */
    /* Requires allocation lock.                                */
    STATIC ptr_t GC_find_limit_with_bound(ptr_t p, GC_bool up, ptr_t bound)
    {
        static volatile ptr_t result;
                /* Safer if static, since otherwise it may not be       */
                /* preserved across the longjmp.  Can safely be         */
                /* static since it's only called with the               */
                /* allocation lock held.                                */

        GC_ASSERT(I_HOLD_LOCK());
        GC_setup_temporary_fault_handler();
        if (SETJMP(GC_jmp_buf) == 0) {
            result = (ptr_t)(((word)(p))
                              & ~(MIN_PAGE_SIZE-1));
            for (;;) {
                if (up) {
                    result += MIN_PAGE_SIZE;
                    if (result >= bound) return bound;
                } else {
                    result -= MIN_PAGE_SIZE;
                    if (result <= bound) return bound;
                }
                GC_noop1((word)(*result));
            }
        }
        GC_reset_fault_handler();
        if (!up) {
            result += MIN_PAGE_SIZE;
        }
        return(result);
    }

    ptr_t GC_find_limit(ptr_t p, GC_bool up)
    {
        return GC_find_limit_with_bound(p, up, up ? (ptr_t)(word)(-1) : 0);
    }
# endif

#if defined(ECOS) || defined(NOSYS)
  ptr_t GC_get_main_stack_base(void)
  {
    return STACKBOTTOM;
  }
#endif

#ifdef HPUX_STACKBOTTOM

#include <sys/param.h>
#include <sys/pstat.h>

  ptr_t GC_get_register_stack_base(void)
  {
    struct pst_vm_status vm_status;

    int i = 0;
    while (pstat_getprocvm(&vm_status, sizeof(vm_status), 0, i++) == 1) {
      if (vm_status.pst_type == PS_RSESTACK) {
        return (ptr_t) vm_status.pst_vaddr;
      }
    }

    /* old way to get the register stackbottom */
    return (ptr_t)(((word)GC_stackbottom - BACKING_STORE_DISPLACEMENT - 1)
                   & ~(BACKING_STORE_ALIGNMENT - 1));
  }

#endif /* HPUX_STACK_BOTTOM */

#ifdef LINUX_STACKBOTTOM

#include <sys/types.h>
#include <sys/stat.h>

# define STAT_SKIP 27   /* Number of fields preceding startstack        */
                        /* field in /proc/self/stat                     */

#ifdef USE_LIBC_PRIVATES
# pragma weak __libc_stack_end
  extern ptr_t __libc_stack_end;
#endif

# ifdef IA64
#   ifdef USE_LIBC_PRIVATES
#     pragma weak __libc_ia64_register_backing_store_base
      extern ptr_t __libc_ia64_register_backing_store_base;
#   endif

    ptr_t GC_get_register_stack_base(void)
    {
      ptr_t result;

#     ifdef USE_LIBC_PRIVATES
        if (0 != &__libc_ia64_register_backing_store_base
            && 0 != __libc_ia64_register_backing_store_base) {
          /* Glibc 2.2.4 has a bug such that for dynamically linked     */
          /* executables __libc_ia64_register_backing_store_base is     */
          /* defined but uninitialized during constructor calls.        */
          /* Hence we check for both nonzero address and value.         */
          return __libc_ia64_register_backing_store_base;
        }
#     endif
      result = backing_store_base_from_proc();
      if (0 == result) {
          result = GC_find_limit(GC_save_regs_in_stack(), FALSE);
          /* Now seems to work better than constant displacement        */
          /* heuristic used in 6.X versions.  The latter seems to       */
          /* fail for 2.6 kernels.                                      */
      }
      return result;
    }
# endif

  STATIC ptr_t GC_linux_stack_base(void)
  {
    /* We read the stack base value from /proc/self/stat.  We do this   */
    /* using direct I/O system calls in order to avoid calling malloc   */
    /* in case REDIRECT_MALLOC is defined.                              */
#   define STAT_BUF_SIZE 4096
#   define STAT_READ read
          /* Should probably call the real read, if read is wrapped.    */
    char stat_buf[STAT_BUF_SIZE];
    int f;
    char c;
    word result = 0;
    size_t i, buf_offset = 0;

    /* First try the easy way.  This should work for glibc 2.2  */
    /* This fails in a prelinked ("prelink" command) executable */
    /* since the correct value of __libc_stack_end never        */
    /* becomes visible to us.  The second test works around     */
    /* this.                                                    */
#   ifdef USE_LIBC_PRIVATES
      if (0 != &__libc_stack_end && 0 != __libc_stack_end ) {
#       if defined(IA64)
          /* Some versions of glibc set the address 16 bytes too        */
          /* low while the initialization code is running.              */
          if (((word)__libc_stack_end & 0xfff) + 0x10 < 0x1000) {
            return __libc_stack_end + 0x10;
          } /* Otherwise it's not safe to add 16 bytes and we fall      */
            /* back to using /proc.                                     */
#       elif defined(SPARC)
          /* Older versions of glibc for 64-bit Sparc do not set
           * this variable correctly, it gets set to either zero
           * or one.
           */
          if (__libc_stack_end != (ptr_t) (unsigned long)0x1)
            return __libc_stack_end;
#       else
          return __libc_stack_end;
#       endif
      }
#   endif
    f = open("/proc/self/stat", O_RDONLY);
    if (f < 0 || STAT_READ(f, stat_buf, STAT_BUF_SIZE) < 2 * STAT_SKIP) {
        ABORT("Couldn't read /proc/self/stat");
    }
    c = stat_buf[buf_offset++];
    /* Skip the required number of fields.  This number is hopefully    */
    /* constant across all Linux implementations.                       */
      for (i = 0; i < STAT_SKIP; ++i) {
        while (isspace(c)) c = stat_buf[buf_offset++];
        while (!isspace(c)) c = stat_buf[buf_offset++];
      }
    while (isspace(c)) c = stat_buf[buf_offset++];
    while (isdigit(c)) {
      result *= 10;
      result += c - '0';
      c = stat_buf[buf_offset++];
    }
    close(f);
    if (result < 0x100000) ABORT("Absurd stack bottom value");
    return (ptr_t)result;
  }

#endif /* LINUX_STACKBOTTOM */

#ifdef FREEBSD_STACKBOTTOM

/* This uses an undocumented sysctl call, but at least one expert       */
/* believes it will stay.                                               */

#include <unistd.h>
#include <sys/types.h>
#include <sys/sysctl.h>

  STATIC ptr_t GC_freebsd_stack_base(void)
  {
    int nm[2] = {CTL_KERN, KERN_USRSTACK};
    ptr_t base;
    size_t len = sizeof(ptr_t);
    int r = sysctl(nm, 2, &base, &len, NULL, 0);

    if (r) ABORT("Error getting stack base");

    return base;
  }

#endif /* FREEBSD_STACKBOTTOM */

#if !defined(BEOS) && !defined(AMIGA) && !defined(MSWIN32) \
    && !defined(MSWINCE) && !defined(OS2) && !defined(NOSYS) && !defined(ECOS) \
    && !defined(CYGWIN32)

ptr_t GC_get_main_stack_base(void)
{
#   ifdef STACKBOTTOM
        return(STACKBOTTOM);
#   else
#       if defined(HEURISTIC1) || defined(HEURISTIC2)
          word dummy;
#       endif
        ptr_t result;
#       define STACKBOTTOM_ALIGNMENT_M1 ((word)STACK_GRAN - 1)
#       ifdef HEURISTIC1
#          ifdef STACK_GROWS_DOWN
             result = (ptr_t)((((word)(&dummy))
                               + STACKBOTTOM_ALIGNMENT_M1)
                              & ~STACKBOTTOM_ALIGNMENT_M1);
#          else
             result = (ptr_t)(((word)(&dummy))
                              & ~STACKBOTTOM_ALIGNMENT_M1);
#          endif
#       endif /* HEURISTIC1 */
#       ifdef LINUX_STACKBOTTOM
           result = GC_linux_stack_base();
#       endif
#       ifdef FREEBSD_STACKBOTTOM
           result = GC_freebsd_stack_base();
#       endif
#       ifdef HEURISTIC2
#           ifdef STACK_GROWS_DOWN
                result = GC_find_limit((ptr_t)(&dummy), TRUE);
#               ifdef HEURISTIC2_LIMIT
                    if (result > HEURISTIC2_LIMIT
                        && (ptr_t)(&dummy) < HEURISTIC2_LIMIT) {
                            result = HEURISTIC2_LIMIT;
                    }
#               endif
#           else
                result = GC_find_limit((ptr_t)(&dummy), FALSE);
#               ifdef HEURISTIC2_LIMIT
                    if (result < HEURISTIC2_LIMIT
                        && (ptr_t)(&dummy) > HEURISTIC2_LIMIT) {
                            result = HEURISTIC2_LIMIT;
                    }
#               endif
#           endif

#       endif /* HEURISTIC2 */
#       ifdef STACK_GROWS_DOWN
            if (result == 0) result = (ptr_t)(signed_word)(-sizeof(ptr_t));
#       endif
        return(result);
#   endif /* STACKBOTTOM */
}

# endif /* ! AMIGA, !OS 2, ! MS Windows, !BEOS, !NOSYS, !ECOS */

#if defined(GC_LINUX_THREADS) && !defined(HAVE_GET_STACK_BASE)

#include <pthread.h>
/* extern int pthread_getattr_np(pthread_t, pthread_attr_t *); */

#ifdef IA64
  ptr_t GC_greatest_stack_base_below(ptr_t bound);
        /* From pthread_support.c */
#endif

GC_API int GC_CALL GC_get_stack_base(struct GC_stack_base *b)
{
    pthread_attr_t attr;
    size_t size;

    if (pthread_getattr_np(pthread_self(), &attr) != 0) {
        WARN("pthread_getattr_np failed\n", 0);
        return GC_UNIMPLEMENTED;
    }
    if (pthread_attr_getstack(&attr, &(b -> mem_base), &size) != 0) {
        ABORT("pthread_attr_getstack failed");
    }
    pthread_attr_destroy(&attr);
#   ifdef STACK_GROWS_DOWN
        b -> mem_base = (char *)(b -> mem_base) + size;
#   endif
#   ifdef IA64
      /* We could try backing_store_base_from_proc, but that's safe     */
      /* only if no mappings are being asynchronously created.          */
      /* Subtracting the size from the stack base doesn't work for at   */
      /* least the main thread.                                         */
      LOCK();
      {
        IF_CANCEL(int cancel_state;)
        ptr_t bsp;
        ptr_t next_stack;

        DISABLE_CANCEL(cancel_state);
        bsp = GC_save_regs_in_stack();
        next_stack = GC_greatest_stack_base_below(bsp);
        if (0 == next_stack) {
          b -> reg_base = GC_find_limit(bsp, FALSE);
        } else {
          /* Avoid walking backwards into preceding memory stack and    */
          /* growing it.                                                */
          b -> reg_base = GC_find_limit_with_bound(bsp, FALSE, next_stack);
        }
        RESTORE_CANCEL(cancel_state);
      }
      UNLOCK();
#   endif
    return GC_SUCCESS;
}

#define HAVE_GET_STACK_BASE

#endif /* GC_LINUX_THREADS */

#ifndef HAVE_GET_STACK_BASE
/* Retrieve stack base.                                         */
/* Using the GC_find_limit version is risky.                    */
/* On IA64, for example, there is no guard page between the     */
/* stack of one thread and the register backing store of the    */
/* next.  Thus this is likely to identify way too large a       */
/* "stack" and thus at least result in disastrous performance.  */
/* FIXME - Implement better strategies here.                    */
GC_API int GC_CALL GC_get_stack_base(struct GC_stack_base *b)
{
#   ifdef NEED_FIND_LIMIT
      int dummy;
      IF_CANCEL(int cancel_state;)
      DISABLE_CANCEL(cancel_state);  /* May be unnecessary? */
#     ifdef STACK_GROWS_DOWN
        b -> mem_base = GC_find_limit((ptr_t)(&dummy), TRUE);
#       ifdef IA64
          b -> reg_base = GC_find_limit(GC_save_regs_in_stack(), FALSE);
#       endif
#     else
        b -> mem_base = GC_find_limit(&dummy, FALSE);
#     endif
      RESTORE_CANCEL(cancel_state);
      return GC_SUCCESS;
#   else
      return GC_UNIMPLEMENTED;
#   endif
}
#endif

/*
 * Register static data segment(s) as roots.
 * If more data segments are added later then they need to be registered
 * add that point (as we do with SunOS dynamic loading),
 * or GC_mark_roots needs to check for them (as we do with PCR).
 * Called with allocator lock held.
 */

# ifdef OS2

void GC_register_data_segments(void)
{
    PTIB ptib;
    PPIB ppib;
    HMODULE module_handle;
#   define PBUFSIZ 512
    UCHAR path[PBUFSIZ];
    FILE * myexefile;
    struct exe_hdr hdrdos;      /* MSDOS header.        */
    struct e32_exe hdr386;      /* Real header for my executable */
    struct o32_obj seg; /* Currrent segment */
    int nsegs;


    if (DosGetInfoBlocks(&ptib, &ppib) != NO_ERROR) {
        GC_err_printf("DosGetInfoBlocks failed\n");
        ABORT("DosGetInfoBlocks failed\n");
    }
    module_handle = ppib -> pib_hmte;
    if (DosQueryModuleName(module_handle, PBUFSIZ, path) != NO_ERROR) {
        GC_err_printf("DosQueryModuleName failed\n");
        ABORT("DosGetInfoBlocks failed\n");
    }
    myexefile = fopen(path, "rb");
    if (myexefile == 0) {
        GC_err_puts("Couldn't open executable ");
        GC_err_puts(path); GC_err_puts("\n");
        ABORT("Failed to open executable\n");
    }
    if (fread((char *)(&hdrdos), 1, sizeof hdrdos, myexefile) < sizeof hdrdos) {
        GC_err_puts("Couldn't read MSDOS header from ");
        GC_err_puts(path); GC_err_puts("\n");
        ABORT("Couldn't read MSDOS header");
    }
    if (E_MAGIC(hdrdos) != EMAGIC) {
        GC_err_puts("Executable has wrong DOS magic number: ");
        GC_err_puts(path); GC_err_puts("\n");
        ABORT("Bad DOS magic number");
    }
    if (fseek(myexefile, E_LFANEW(hdrdos), SEEK_SET) != 0) {
        GC_err_puts("Seek to new header failed in ");
        GC_err_puts(path); GC_err_puts("\n");
        ABORT("Bad DOS magic number");
    }
    if (fread((char *)(&hdr386), 1, sizeof hdr386, myexefile) < sizeof hdr386) {
        GC_err_puts("Couldn't read MSDOS header from ");
        GC_err_puts(path); GC_err_puts("\n");
        ABORT("Couldn't read OS/2 header");
    }
    if (E32_MAGIC1(hdr386) != E32MAGIC1 || E32_MAGIC2(hdr386) != E32MAGIC2) {
        GC_err_puts("Executable has wrong OS/2 magic number:");
        GC_err_puts(path); GC_err_puts("\n");
        ABORT("Bad OS/2 magic number");
    }
    if ( E32_BORDER(hdr386) != E32LEBO || E32_WORDER(hdr386) != E32LEWO) {
        GC_err_puts("Executable %s has wrong byte order: ");
        GC_err_puts(path); GC_err_puts("\n");
        ABORT("Bad byte order");
    }
    if ( E32_CPU(hdr386) == E32CPU286) {
        GC_err_puts("GC can't handle 80286 executables: ");
        GC_err_puts(path); GC_err_puts("\n");
        EXIT();
    }
    if (fseek(myexefile, E_LFANEW(hdrdos) + E32_OBJTAB(hdr386),
              SEEK_SET) != 0) {
        GC_err_puts("Seek to object table failed: ");
        GC_err_puts(path); GC_err_puts("\n");
        ABORT("Seek to object table failed");
    }
    for (nsegs = E32_OBJCNT(hdr386); nsegs > 0; nsegs--) {
      int flags;
      if (fread((char *)(&seg), 1, sizeof seg, myexefile) < sizeof seg) {
        GC_err_puts("Couldn't read obj table entry from ");
        GC_err_puts(path); GC_err_puts("\n");
        ABORT("Couldn't read obj table entry");
      }
      flags = O32_FLAGS(seg);
      if (!(flags & OBJWRITE)) continue;
      if (!(flags & OBJREAD)) continue;
      if (flags & OBJINVALID) {
          GC_err_printf("Object with invalid pages?\n");
          continue;
      }
      GC_add_roots_inner((ptr_t)O32_BASE(seg),
                         (ptr_t)(O32_BASE(seg)+O32_SIZE(seg)), FALSE);
    }
}

# else /* !OS2 */

# if defined(GWW_VDB)

#   ifndef MEM_WRITE_WATCH
#     define MEM_WRITE_WATCH 0x200000
#   endif

#   ifndef WRITE_WATCH_FLAG_RESET
#     define WRITE_WATCH_FLAG_RESET 1
#   endif

#   if !defined(_BASETSD_H_) && !defined(_BASETSD_H)
#     ifdef _WIN64
        typedef unsigned __int64 ULONG_PTR;
#     else
        typedef unsigned long ULONG_PTR;
#     endif
      typedef ULONG_PTR SIZE_T;
      typedef ULONG_PTR * PULONG_PTR;
#   endif

    typedef UINT (WINAPI * GetWriteWatch_type)(
      DWORD, PVOID, SIZE_T, PVOID*, PULONG_PTR, PULONG);
    static GetWriteWatch_type GetWriteWatch_func;
    static DWORD GetWriteWatch_alloc_flag;

#   define GC_GWW_AVAILABLE() (GetWriteWatch_func != NULL)

    static void detect_GetWriteWatch(void)
    {
      static GC_bool done;
      HMODULE hK32;
      if (done)
        return;

#     if defined(MPROTECT_VDB)
        {
          char * str = GETENV("GC_USE_GETWRITEWATCH");
#         if defined(GC_PREFER_MPROTECT_VDB)
            if (str == NULL || (*str == '0' && *(str + 1) == '\0')) {
              /* GC_USE_GETWRITEWATCH is unset or set to "0".           */
              done = TRUE; /* falling back to MPROTECT_VDB strategy.    */
              /* This should work as if GWW_VDB is undefined. */
              return;
            }
#         else
            if (str != NULL && *str == '0' && *(str + 1) == '\0') {
              /* GC_USE_GETWRITEWATCH is set "0".                       */
              done = TRUE; /* falling back to MPROTECT_VDB strategy.    */
              return;
            }
#         endif
        }
#     endif

      hK32 = GetModuleHandle(TEXT("kernel32.dll"));
      if (hK32 != (HMODULE)0 &&
          (GetWriteWatch_func = (GetWriteWatch_type)GetProcAddress(hK32,
                                                "GetWriteWatch")) != NULL) {
        /* Also check whether VirtualAlloc accepts MEM_WRITE_WATCH,   */
        /* as some versions of kernel32.dll have one but not the      */
        /* other, making the feature completely broken.               */
        void * page = VirtualAlloc(NULL, GC_page_size,
                                    MEM_WRITE_WATCH | MEM_RESERVE,
                                    PAGE_READWRITE);
        if (page != NULL) {
          PVOID pages[16];
          ULONG_PTR count = 16;
          DWORD page_size;
          /* Check that it actually works.  In spite of some            */
          /* documentation it actually seems to exist on W2K.           */
          /* This test may be unnecessary, but ...                      */
          if (GetWriteWatch_func(WRITE_WATCH_FLAG_RESET,
                                 page, GC_page_size,
                                 pages,
                                 &count,
                                 &page_size) != 0) {
            /* GetWriteWatch always fails. */
            GetWriteWatch_func = NULL;
          } else {
            GetWriteWatch_alloc_flag = MEM_WRITE_WATCH;
          }
          VirtualFree(page, GC_page_size, MEM_RELEASE);
        } else {
          /* GetWriteWatch will be useless. */
          GetWriteWatch_func = NULL;
        }
      }
      if (GC_print_stats) {
        if (GetWriteWatch_func == NULL) {
          GC_log_printf("Did not find a usable GetWriteWatch()\n");
        } else {
          GC_log_printf("Using GetWriteWatch()\n");
        }
      }
      done = TRUE;
    }

# else /* !GWW_VDB */
#   define GetWriteWatch_alloc_flag 0
# endif /* GWW_VDB */

# if defined(MSWIN32) || defined(MSWINCE)

# ifdef MSWIN32
  /* Unfortunately, we have to handle win32s very differently from NT,  */
  /* Since VirtualQuery has very different semantics.  In particular,   */
  /* under win32s a VirtualQuery call on an unmapped page returns an    */
  /* invalid result.  Under NT, GC_register_data_segments is a no-op    */
  /* and all real work is done by GC_register_dynamic_libraries.  Under */
  /* win32s, we cannot find the data segments associated with dll's.    */
  /* We register the main data segment here.                            */
  GC_bool GC_no_win32_dlls = FALSE;
        /* This used to be set for gcc, to avoid dealing with           */
        /* the structured exception handling issues.  But we now have   */
        /* assembly code to do that right.                              */

  GC_bool GC_wnt = FALSE;
         /* This is a Windows NT derivative, i.e. NT, W2K, XP or later.  */

# ifdef USE_MUNMAP
    extern int GC_unmap_threshold; /* defined in allchblk.c */
# endif

  void GC_init_win32(void)
  {
    /* Set GC_wnt.                                                       */
    /* If we're running under win32s, assume that no DLLs will be loaded */
    /* I doubt anyone still runs win32s, but ...                         */
    DWORD v = GetVersion();
    GC_wnt = !(v & 0x80000000);
    GC_no_win32_dlls |= ((!GC_wnt) && (v & 0xff) <= 3);
#   ifdef USE_MUNMAP
      if (GC_no_win32_dlls) {
        /* Turn off unmapping for safety (since may not work well with  */
        /* GlobalAlloc).                                                */
        GC_unmap_threshold = 0;
      }
#   endif
  }

  /* Return the smallest address a such that VirtualQuery               */
  /* returns correct results for all addresses between a and start.     */
  /* Assumes VirtualQuery returns correct information for start.        */
  STATIC ptr_t GC_least_described_address(ptr_t start)
  {
    MEMORY_BASIC_INFORMATION buf;
    size_t result;
    LPVOID limit;
    ptr_t p;
    LPVOID q;

    limit = GC_sysinfo.lpMinimumApplicationAddress;
    p = (ptr_t)((word)start & ~(GC_page_size - 1));
    for (;;) {
        q = (LPVOID)(p - GC_page_size);
        if ((ptr_t)q > (ptr_t)p /* underflow */ || q < limit) break;
        result = VirtualQuery(q, &buf, sizeof(buf));
        if (result != sizeof(buf) || buf.AllocationBase == 0) break;
        p = (ptr_t)(buf.AllocationBase);
    }
    return p;
  }
# endif

# ifndef REDIRECT_MALLOC
  /* We maintain a linked list of AllocationBase values that we know    */
  /* correspond to malloc heap sections.  Currently this is only called */
  /* during a GC.  But there is some hope that for long running         */
  /* programs we will eventually see most heap sections.                */

  /* In the long run, it would be more reliable to occasionally walk    */
  /* the malloc heap with HeapWalk on the default heap.  But that       */
  /* apparently works only for NT-based Windows.                        */

  /* In the long run, a better data structure would also be nice ...    */
  STATIC struct GC_malloc_heap_list {
    void * allocation_base;
    struct GC_malloc_heap_list *next;
  } *GC_malloc_heap_l = 0;

  /* Is p the base of one of the malloc heap sections we already know   */
  /* about?                                                             */
  STATIC GC_bool GC_is_malloc_heap_base(ptr_t p)
  {
    struct GC_malloc_heap_list *q = GC_malloc_heap_l;

    while (0 != q) {
      if (q -> allocation_base == p) return TRUE;
      q = q -> next;
    }
    return FALSE;
  }

  STATIC void *GC_get_allocation_base(void *p)
  {
    MEMORY_BASIC_INFORMATION buf;
    size_t result = VirtualQuery(p, &buf, sizeof(buf));
    if (result != sizeof(buf)) {
      ABORT("Weird VirtualQuery result");
    }
    return buf.AllocationBase;
  }

  STATIC size_t GC_max_root_size = 100000;      /* Appr. largest root size.     */

  void GC_add_current_malloc_heap(void)
  {
    struct GC_malloc_heap_list *new_l =
                 malloc(sizeof(struct GC_malloc_heap_list));
    void * candidate = GC_get_allocation_base(new_l);

    if (new_l == 0) return;
    if (GC_is_malloc_heap_base(candidate)) {
      /* Try a little harder to find malloc heap.                       */
        size_t req_size = 10000;
        do {
          void *p = malloc(req_size);
          if (0 == p) { free(new_l); return; }
          candidate = GC_get_allocation_base(p);
          free(p);
          req_size *= 2;
        } while (GC_is_malloc_heap_base(candidate)
                 && req_size < GC_max_root_size/10 && req_size < 500000);
        if (GC_is_malloc_heap_base(candidate)) {
          free(new_l); return;
        }
    }
    if (GC_print_stats)
          GC_log_printf("Found new system malloc AllocationBase at %p\n",
                        candidate);
    new_l -> allocation_base = candidate;
    new_l -> next = GC_malloc_heap_l;
    GC_malloc_heap_l = new_l;
  }
# endif /* REDIRECT_MALLOC */

  STATIC word GC_n_heap_bases = 0;      /* See GC_heap_bases.   */

  /* Is p the start of either the malloc heap, or of one of our */
  /* heap sections?                                             */
  GC_bool GC_is_heap_base (ptr_t p)
  {

     unsigned i;

#    ifndef REDIRECT_MALLOC
       if (GC_root_size > GC_max_root_size) GC_max_root_size = GC_root_size;
       if (GC_is_malloc_heap_base(p)) return TRUE;
#    endif
     for (i = 0; i < GC_n_heap_bases; i++) {
         if (GC_heap_bases[i] == p) return TRUE;
     }
     return FALSE ;
  }

# ifdef MSWIN32
  STATIC void GC_register_root_section(ptr_t static_root)
  {
      MEMORY_BASIC_INFORMATION buf;
      size_t result;
      DWORD protect;
      LPVOID p;
      char * base;
      char * limit, * new_limit;

      if (!GC_no_win32_dlls) return;
      p = base = limit = GC_least_described_address(static_root);
      while (p < GC_sysinfo.lpMaximumApplicationAddress) {
        result = VirtualQuery(p, &buf, sizeof(buf));
        if (result != sizeof(buf) || buf.AllocationBase == 0
            || GC_is_heap_base(buf.AllocationBase)) break;
        new_limit = (char *)p + buf.RegionSize;
        protect = buf.Protect;
        if (buf.State == MEM_COMMIT
            && is_writable(protect)) {
            if ((char *)p == limit) {
                limit = new_limit;
            } else {
                if (base != limit) GC_add_roots_inner(base, limit, FALSE);
                base = p;
                limit = new_limit;
            }
        }
        if (p > (LPVOID)new_limit /* overflow */) break;
        p = (LPVOID)new_limit;
      }
      if (base != limit) GC_add_roots_inner(base, limit, FALSE);
  }
#endif

  void GC_register_data_segments(void)
  {
#   ifdef MSWIN32
      static char dummy;
      GC_register_root_section((ptr_t)(&dummy));
#   endif
  }

# else /* !OS2 && !Windows */

# if (defined(SVR4) || defined(AUX) || defined(DGUX) \
      || (defined(LINUX) && defined(SPARC))) && !defined(PCR)
ptr_t GC_SysVGetDataStart(size_t max_page_size, ptr_t etext_addr)
{
    word text_end = ((word)(etext_addr) + sizeof(word) - 1)
                    & ~(sizeof(word) - 1);
        /* etext rounded to word boundary       */
    word next_page = ((text_end + (word)max_page_size - 1)
                      & ~((word)max_page_size - 1));
    word page_offset = (text_end & ((word)max_page_size - 1));
    volatile char * result = (char *)(next_page + page_offset);
    /* Note that this isnt equivalent to just adding            */
    /* max_page_size to &etext if &etext is at a page boundary  */

    GC_setup_temporary_fault_handler();
    if (SETJMP(GC_jmp_buf) == 0) {
        /* Try writing to the address.  */
        *result = *result;
        GC_reset_fault_handler();
    } else {
        GC_reset_fault_handler();
        /* We got here via a longjmp.  The address is not readable.     */
        /* This is known to happen under Solaris 2.4 + gcc, which place */
        /* string constants in the text segment, but after etext.       */
        /* Use plan B.  Note that we now know there is a gap between    */
        /* text and data segments, so plan A bought us something.       */
        result = (char *)GC_find_limit((ptr_t)(DATAEND), FALSE);
    }
    return((ptr_t)result);
}
# endif

# if defined(FREEBSD) && (defined(I386) || defined(X86_64) || defined(powerpc) || defined(__powerpc__)) && !defined(PCR)
/* Its unclear whether this should be identical to the above, or        */
/* whether it should apply to non-X86 architectures.                    */
/* For now we don't assume that there is always an empty page after     */
/* etext.  But in some cases there actually seems to be slightly more.  */
/* This also deals with holes between read-only data and writable data. */
ptr_t GC_FreeBSDGetDataStart(size_t max_page_size, ptr_t etext_addr)
{
    word text_end = ((word)(etext_addr) + sizeof(word) - 1)
                     & ~(sizeof(word) - 1);
        /* etext rounded to word boundary       */
    volatile word next_page = (text_end + (word)max_page_size - 1)
                              & ~((word)max_page_size - 1);
    volatile ptr_t result = (ptr_t)text_end;
    GC_setup_temporary_fault_handler();
    if (SETJMP(GC_jmp_buf) == 0) {
        /* Try reading at the address.                          */
        /* This should happen before there is another thread.   */
        for (; next_page < (word)(DATAEND); next_page += (word)max_page_size)
            *(volatile char *)next_page;
        GC_reset_fault_handler();
    } else {
        GC_reset_fault_handler();
        /* As above, we go to plan B    */
        result = GC_find_limit((ptr_t)(DATAEND), FALSE);
    }
    return(result);
}

# endif


#ifdef AMIGA

#  define GC_AMIGA_DS
#  include "extra/AmigaOS.c"
#  undef GC_AMIGA_DS

#else /* !OS2 && !Windows && !AMIGA */

void GC_register_data_segments(void)
{
#   if !defined(PCR) && !defined(MACOS)
#     if defined(REDIRECT_MALLOC) && defined(GC_SOLARIS_THREADS)
        /* As of Solaris 2.3, the Solaris threads implementation        */
        /* allocates the data structure for the initial thread with     */
        /* sbrk at process startup.  It needs to be scanned, so that    */
        /* we don't lose some malloc allocated data structures          */
        /* hanging from it.  We're on thin ice here ...                 */
        extern caddr_t sbrk(int);

        GC_add_roots_inner(DATASTART, (ptr_t)sbrk(0), FALSE);
#     else
        GC_add_roots_inner(DATASTART, (ptr_t)(DATAEND), FALSE);
#       if defined(DATASTART2)
          GC_add_roots_inner(DATASTART2, (ptr_t)(DATAEND2), FALSE);
#       endif
#     endif
#   endif
#   if defined(MACOS)
    {
#   if defined(THINK_C)
        extern void* GC_MacGetDataStart(void);
        /* globals begin above stack and end at a5. */
        GC_add_roots_inner((ptr_t)GC_MacGetDataStart(),
                           (ptr_t)LMGetCurrentA5(), FALSE);
#   else
#     if defined(__MWERKS__)
#       if !__POWERPC__
          extern void* GC_MacGetDataStart(void);
          /* MATTHEW: Function to handle Far Globals (CW Pro 3) */
#         if __option(far_data)
          extern void* GC_MacGetDataEnd(void);
#         endif
          /* globals begin above stack and end at a5. */
          GC_add_roots_inner((ptr_t)GC_MacGetDataStart(),
                             (ptr_t)LMGetCurrentA5(), FALSE);
          /* MATTHEW: Handle Far Globals */
#         if __option(far_data)
      /* Far globals follow he QD globals: */
          GC_add_roots_inner((ptr_t)LMGetCurrentA5(),
                             (ptr_t)GC_MacGetDataEnd(), FALSE);
#         endif
#       else
          extern char __data_start__[], __data_end__[];
          GC_add_roots_inner((ptr_t)&__data_start__,
                             (ptr_t)&__data_end__, FALSE);
#       endif /* __POWERPC__ */
#     endif /* __MWERKS__ */
#   endif /* !THINK_C */
    }
#   endif /* MACOS */

    /* Dynamic libraries are added at every collection, since they may  */
    /* change.                                                          */
}

# endif  /* ! AMIGA */
# endif  /* ! MSWIN32 && ! MSWINCE*/
# endif  /* ! OS2 */

/*
 * Auxiliary routines for obtaining memory from OS.
 */

# if !defined(OS2) && !defined(PCR) && !defined(AMIGA) \
        && !defined(MSWIN32) && !defined(MSWINCE) \
        && !defined(MACOS) && !defined(DOS4GW) && !defined(NONSTOP)

# define SBRK_ARG_T ptrdiff_t

#if defined(MMAP_SUPPORTED)

#ifdef USE_MMAP_FIXED
#   define GC_MMAP_FLAGS MAP_FIXED | MAP_PRIVATE
        /* Seems to yield better performance on Solaris 2, but can      */
        /* be unreliable if something is already mapped at the address. */
#else
#   define GC_MMAP_FLAGS MAP_PRIVATE
#endif

#ifdef USE_MMAP_ANON
# define zero_fd -1
# if defined(MAP_ANONYMOUS)
#   define OPT_MAP_ANON MAP_ANONYMOUS
# else
#   define OPT_MAP_ANON MAP_ANON
# endif
#else
  static int zero_fd;
# define OPT_MAP_ANON 0
#endif

#ifndef HEAP_START
#   define HEAP_START ((ptr_t)0)
#endif

STATIC ptr_t GC_unix_mmap_get_mem(word bytes)
{
    void *result;
    static ptr_t last_addr = HEAP_START;

#   ifndef USE_MMAP_ANON
      static GC_bool initialized = FALSE;

      if (!initialized) {
          zero_fd = open("/dev/zero", O_RDONLY);
          fcntl(zero_fd, F_SETFD, FD_CLOEXEC);
          initialized = TRUE;
      }
#   endif

    if (bytes & (GC_page_size - 1)) ABORT("Bad GET_MEM arg");
    result = mmap(last_addr, bytes, PROT_READ | PROT_WRITE | OPT_PROT_EXEC,
                  GC_MMAP_FLAGS | OPT_MAP_ANON, zero_fd, 0/* offset */);
    if (result == MAP_FAILED) return(0);
    last_addr = (ptr_t)result + bytes + GC_page_size - 1;
    last_addr = (ptr_t)((word)last_addr & ~(GC_page_size - 1));
#   if !defined(LINUX)
      if (last_addr == 0) {
        /* Oops.  We got the end of the address space.  This isn't      */
        /* usable by arbitrary C code, since one-past-end pointers      */
        /* don't work, so we discard it and try again.                  */
        munmap(result, (size_t)(-GC_page_size) - (size_t)result);
                        /* Leave last page mapped, so we can't repeat. */
        return GC_unix_mmap_get_mem(bytes);
      }
#   else
      GC_ASSERT(last_addr != 0);
#   endif
    return((ptr_t)result);
}

# endif  /* MMAP_SUPPORTED */

#if defined(USE_MMAP)

ptr_t GC_unix_get_mem(word bytes)
{
    return GC_unix_mmap_get_mem(bytes);
}

#else /* Not USE_MMAP */

STATIC ptr_t GC_unix_sbrk_get_mem(word bytes)
{
  ptr_t result;
# ifdef IRIX5
    /* Bare sbrk isn't thread safe.  Play by malloc rules.      */
    /* The equivalent may be needed on other systems as well.   */
    __LOCK_MALLOC();
# endif
  {
    ptr_t cur_brk = (ptr_t)sbrk(0);
    SBRK_ARG_T lsbs = (word)cur_brk & (GC_page_size-1);

    if ((SBRK_ARG_T)bytes < 0) {
        result = 0; /* too big */
        goto out;
    }
    if (lsbs != 0) {
        if((ptr_t)sbrk(GC_page_size - lsbs) == (ptr_t)(-1)) {
            result = 0;
            goto out;
        }
    }
#   ifdef ADD_HEAP_GUARD_PAGES
      /* This is useful for catching severe memory overwrite problems that */
      /* span heap sections.  It shouldn't otherwise be turned on.         */
      {
        ptr_t guard = (ptr_t)sbrk((SBRK_ARG_T)GC_page_size);
        if (mprotect(guard, GC_page_size, PROT_NONE) != 0)
            ABORT("ADD_HEAP_GUARD_PAGES: mprotect failed");
      }
#   endif /* ADD_HEAP_GUARD_PAGES */
    result = (ptr_t)sbrk((SBRK_ARG_T)bytes);
    if (result == (ptr_t)(-1)) result = 0;
  }
 out:
# ifdef IRIX5
    __UNLOCK_MALLOC();
# endif
  return(result);
}

#if defined(MMAP_SUPPORTED)

/* By default, we try both sbrk and mmap, in that order. */
ptr_t GC_unix_get_mem(word bytes)
{
    static GC_bool sbrk_failed = FALSE;
    ptr_t result = 0;

    if (!sbrk_failed) result = GC_unix_sbrk_get_mem(bytes);
    if (0 == result) {
        sbrk_failed = TRUE;
        result = GC_unix_mmap_get_mem(bytes);
    }
    if (0 == result) {
        /* Try sbrk again, in case sbrk memory became available. */
        result = GC_unix_sbrk_get_mem(bytes);
    }
    return result;
}

#else /* !MMAP_SUPPORTED */

ptr_t GC_unix_get_mem(word bytes)
{
    return GC_unix_sbrk_get_mem(bytes);
}

#endif

#endif /* Not USE_MMAP */

# endif /* UN*X */

# ifdef OS2

void * os2_alloc(size_t bytes)
{
    void * result;

    if (DosAllocMem(&result, bytes, PAG_EXECUTE | PAG_READ |
                                    PAG_WRITE | PAG_COMMIT)
                    != NO_ERROR) {
        return(0);
    }
    if (result == 0) return(os2_alloc(bytes));
    return(result);
}

# endif /* OS2 */


# if defined(MSWIN32) || defined(MSWINCE)
    SYSTEM_INFO GC_sysinfo;
# endif

# ifdef MSWIN32

# ifdef USE_GLOBAL_ALLOC
#   define GLOBAL_ALLOC_TEST 1
# else
#   define GLOBAL_ALLOC_TEST GC_no_win32_dlls
# endif

#ifdef GC_USE_MEM_TOP_DOWN
  STATIC DWORD GC_mem_top_down = MEM_TOP_DOWN;
                           /* Use GC_USE_MEM_TOP_DOWN for better 64-bit */
                           /* testing.  Otherwise all addresses tend to */
                           /* end up in first 4GB, hiding bugs.         */
#else
  STATIC DWORD GC_mem_top_down = 0;
#endif

ptr_t GC_win32_get_mem(word bytes)
{
    ptr_t result;

    if (GLOBAL_ALLOC_TEST) {
        /* VirtualAlloc doesn't like PAGE_EXECUTE_READWRITE.    */
        /* There are also unconfirmed rumors of other           */
        /* problems, so we dodge the issue.                     */
        result = (ptr_t) GlobalAlloc(0, bytes + HBLKSIZE);
        result = (ptr_t)(((word)result + HBLKSIZE - 1) & ~(HBLKSIZE-1));
    } else {
        /* VirtualProtect only works on regions returned by a   */
        /* single VirtualAlloc call.  Thus we allocate one      */
        /* extra page, which will prevent merging of blocks     */
        /* in separate regions, and eliminate any temptation    */
        /* to call VirtualProtect on a range spanning regions.  */
        /* This wastes a small amount of memory, and risks      */
        /* increased fragmentation.  But better alternatives    */
        /* would require effort.                                */
#       ifdef MPROTECT_VDB
          /* We can't check for GC_incremental here (because    */
          /* GC_enable_incremental() might be called some time  */
          /* later after the GC initialization).                */
#         ifdef GWW_VDB
#           define VIRTUAL_ALLOC_PAD (GC_GWW_AVAILABLE() ? 0 : 1)
#         else
#           define VIRTUAL_ALLOC_PAD 1
#         endif
#       else
#         define VIRTUAL_ALLOC_PAD 0
#       endif
        /* Pass the MEM_WRITE_WATCH only if GetWriteWatch-based */
        /* VDBs are enabled and the GetWriteWatch function is   */
        /* available.  Otherwise we waste resources or possibly */
        /* cause VirtualAlloc to fail (observed in Windows 2000 */
        /* SP2).                                                */
        result = (ptr_t) VirtualAlloc(NULL, bytes + VIRTUAL_ALLOC_PAD,
                                      GetWriteWatch_alloc_flag |
                                      MEM_COMMIT | MEM_RESERVE
                                      | GC_mem_top_down,
                                      PAGE_EXECUTE_READWRITE);
    }
    if (HBLKDISPL(result) != 0) ABORT("Bad VirtualAlloc result");
        /* If I read the documentation correctly, this can      */
        /* only happen if HBLKSIZE > 64k or not a power of 2.   */
    if (GC_n_heap_bases >= MAX_HEAP_SECTS) ABORT("Too many heap sections");
    if (0 != result) GC_heap_bases[GC_n_heap_bases++] = result;
    return(result);
}

GC_API void GC_CALL GC_win32_free_heap(void)
{
    if (GC_no_win32_dlls) {
        while (GC_n_heap_bases > 0) {
            GlobalFree (GC_heap_bases[--GC_n_heap_bases]);
            GC_heap_bases[GC_n_heap_bases] = 0;
        }
    }
}
# endif

#ifdef AMIGA
# define GC_AMIGA_AM
# include "extra/AmigaOS.c"
# undef GC_AMIGA_AM
#endif


# ifdef MSWINCE
ptr_t GC_wince_get_mem(word bytes)
{
    ptr_t result = 0; /* initialized to prevent warning. */
    word i;

    /* Round up allocation size to multiple of page size */
    bytes = (bytes + GC_page_size-1) & ~(GC_page_size-1);

    /* Try to find reserved, uncommitted pages */
    for (i = 0; i < GC_n_heap_bases; i++) {
        if (((word)(-(signed_word)GC_heap_lengths[i])
             & (GC_sysinfo.dwAllocationGranularity-1))
            >= bytes) {
            result = GC_heap_bases[i] + GC_heap_lengths[i];
            break;
        }
    }

    if (i == GC_n_heap_bases) {
        /* Reserve more pages */
        word res_bytes = (bytes + GC_sysinfo.dwAllocationGranularity-1)
                         & ~(GC_sysinfo.dwAllocationGranularity-1);
        /* If we ever support MPROTECT_VDB here, we will probably need to    */
        /* ensure that res_bytes is strictly > bytes, so that VirtualProtect */
        /* never spans regions.  It seems to be OK for a VirtualFree         */
        /* argument to span regions, so we should be OK for now.             */
        result = (ptr_t) VirtualAlloc(NULL, res_bytes,
                                      MEM_RESERVE | MEM_TOP_DOWN,
                                      PAGE_EXECUTE_READWRITE);
        if (HBLKDISPL(result) != 0) ABORT("Bad VirtualAlloc result");
            /* If I read the documentation correctly, this can  */
            /* only happen if HBLKSIZE > 64k or not a power of 2.       */
        if (GC_n_heap_bases >= MAX_HEAP_SECTS) ABORT("Too many heap sections");
        if (result == NULL) return NULL;
        GC_heap_bases[GC_n_heap_bases] = result;
        GC_heap_lengths[GC_n_heap_bases] = 0;
        GC_n_heap_bases++;
    }

    /* Commit pages */
    result = (ptr_t) VirtualAlloc(result, bytes,
                                  MEM_COMMIT,
                                  PAGE_EXECUTE_READWRITE);
    if (result != NULL) {
        if (HBLKDISPL(result) != 0) ABORT("Bad VirtualAlloc result");
        GC_heap_lengths[i] += bytes;
    }

    return(result);
}
# endif

#ifdef USE_MUNMAP

/* For now, this only works on Win32/WinCE and some Unix-like   */
/* systems.  If you have something else, don't define           */
/* USE_MUNMAP.                                                  */

#if !defined(MSWIN32) && !defined(MSWINCE)

#include <unistd.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <sys/types.h>

#endif

/* Compute a page aligned starting address for the unmap        */
/* operation on a block of size bytes starting at start.        */
/* Return 0 if the block is too small to make this feasible.    */
STATIC ptr_t GC_unmap_start(ptr_t start, size_t bytes)
{
    ptr_t result;
    /* Round start to next page boundary.       */
    result = (ptr_t)((word)(start + GC_page_size - 1) & ~(GC_page_size - 1));
    if (result + GC_page_size > start + bytes) return 0;
    return result;
}

/* Compute end address for an unmap operation on the indicated  */
/* block.                                                       */
STATIC ptr_t GC_unmap_end(ptr_t start, size_t bytes)
{
    return (ptr_t)((word)(start + bytes) & ~(GC_page_size - 1));
}

/* Under Win32/WinCE we commit (map) and decommit (unmap)       */
/* memory using VirtualAlloc and VirtualFree.  These functions  */
/* work on individual allocations of virtual memory, made       */
/* previously using VirtualAlloc with the MEM_RESERVE flag.     */
/* The ranges we need to (de)commit may span several of these   */
/* allocations; therefore we use VirtualQuery to check          */
/* allocation lengths, and split up the range as necessary.     */

/* We assume that GC_remap is called on exactly the same range  */
/* as a previous call to GC_unmap.  It is safe to consistently  */
/* round the endpoints in both places.                          */
void GC_unmap(ptr_t start, size_t bytes)
{
    ptr_t start_addr = GC_unmap_start(start, bytes);
    ptr_t end_addr = GC_unmap_end(start, bytes);
    word len = end_addr - start_addr;
    if (0 == start_addr) return;
#   if defined(MSWIN32) || defined(MSWINCE)
      while (len != 0) {
          MEMORY_BASIC_INFORMATION mem_info;
          GC_word free_len;
          if (VirtualQuery(start_addr, &mem_info, sizeof(mem_info))
              != sizeof(mem_info))
              ABORT("Weird VirtualQuery result");
          free_len = (len < mem_info.RegionSize) ? len : mem_info.RegionSize;
          if (!VirtualFree(start_addr, free_len, MEM_DECOMMIT))
              ABORT("VirtualFree failed");
          GC_unmapped_bytes += free_len;
          start_addr += free_len;
          len -= free_len;
      }
#   else
      /* We immediately remap it to prevent an intervening mmap from    */
      /* accidentally grabbing the same address space.                  */
      {
        void * result;
        result = mmap(start_addr, len, PROT_NONE,
                      MAP_PRIVATE | MAP_FIXED | OPT_MAP_ANON,
                      zero_fd, 0/* offset */);
        if (result != (void *)start_addr) ABORT("mmap(...PROT_NONE...) failed");
      }
      GC_unmapped_bytes += len;
#   endif
}


void GC_remap(ptr_t start, size_t bytes)
{
    ptr_t start_addr = GC_unmap_start(start, bytes);
    ptr_t end_addr = GC_unmap_end(start, bytes);
    word len = end_addr - start_addr;

    /* FIXME: Handle out-of-memory correctly (at least for Win32)       */
#   if defined(MSWIN32) || defined(MSWINCE)
      ptr_t result;

      if (0 == start_addr) return;
      while (len != 0) {
          MEMORY_BASIC_INFORMATION mem_info;
          GC_word alloc_len;
          if (VirtualQuery(start_addr, &mem_info, sizeof(mem_info))
              != sizeof(mem_info))
              ABORT("Weird VirtualQuery result");
          alloc_len = (len < mem_info.RegionSize) ? len : mem_info.RegionSize;
          result = VirtualAlloc(start_addr, alloc_len,
                                MEM_COMMIT,
                                PAGE_EXECUTE_READWRITE);
          if (result != start_addr) {
              if (GetLastError() == ERROR_NOT_ENOUGH_MEMORY ||
                  GetLastError() == ERROR_OUTOFMEMORY) {
                  ABORT("Not enough memory to process remapping");
              } else {
                  ABORT("VirtualAlloc remapping failed");
              }
          }
          GC_unmapped_bytes -= alloc_len;
          start_addr += alloc_len;
          len -= alloc_len;
      }
#   else
      /* It was already remapped with PROT_NONE. */
      int result;

      if (0 == start_addr) return;
      result = mprotect(start_addr, len,
                        PROT_READ | PROT_WRITE | OPT_PROT_EXEC);
      if (result != 0) {
          GC_err_printf(
                "Mprotect failed at %p (length %ld) with errno %d\n",
                start_addr, (unsigned long)len, errno);
          ABORT("Mprotect remapping failed");
      }
      GC_unmapped_bytes -= len;
#   endif
}

/* Two adjacent blocks have already been unmapped and are about to      */
/* be merged.  Unmap the whole block.  This typically requires          */
/* that we unmap a small section in the middle that was not previously  */
/* unmapped due to alignment constraints.                               */
void GC_unmap_gap(ptr_t start1, size_t bytes1, ptr_t start2, size_t bytes2)
{
    ptr_t start1_addr = GC_unmap_start(start1, bytes1);
    ptr_t end1_addr = GC_unmap_end(start1, bytes1);
    ptr_t start2_addr = GC_unmap_start(start2, bytes2);
    ptr_t start_addr = end1_addr;
    ptr_t end_addr = start2_addr;
    size_t len;
    GC_ASSERT(start1 + bytes1 == start2);
    if (0 == start1_addr) start_addr = GC_unmap_start(start1, bytes1 + bytes2);
    if (0 == start2_addr) end_addr = GC_unmap_end(start1, bytes1 + bytes2);
    if (0 == start_addr) return;
    len = end_addr - start_addr;
#   if defined(MSWIN32) || defined(MSWINCE)
      while (len != 0) {
          MEMORY_BASIC_INFORMATION mem_info;
          GC_word free_len;
          if (VirtualQuery(start_addr, &mem_info, sizeof(mem_info))
              != sizeof(mem_info))
              ABORT("Weird VirtualQuery result");
          free_len = (len < mem_info.RegionSize) ? len : mem_info.RegionSize;
          if (!VirtualFree(start_addr, free_len, MEM_DECOMMIT))
              ABORT("VirtualFree failed");
          GC_unmapped_bytes += free_len;
          start_addr += free_len;
          len -= free_len;
      }
#   else
      if (len != 0) {
        /* Immediately remap as above. */
        void * result;
        result = mmap(start_addr, len, PROT_NONE,
                      MAP_PRIVATE | MAP_FIXED | OPT_MAP_ANON,
                      zero_fd, 0/* offset */);
        if (result != (void *)start_addr) ABORT("mmap(...PROT_NONE...) failed");
      }
      GC_unmapped_bytes += len;
#   endif
}

#endif /* USE_MUNMAP */

/* Routine for pushing any additional roots.  In THREADS        */
/* environment, this is also responsible for marking from       */
/* thread stacks.                                               */
#ifndef THREADS
void (*GC_push_other_roots)(void) = 0;
#else /* THREADS */

# ifdef PCR
PCR_ERes GC_push_thread_stack(PCR_Th_T *t, PCR_Any dummy)
{
    struct PCR_ThCtl_TInfoRep info;
    PCR_ERes result;

    info.ti_stkLow = info.ti_stkHi = 0;
    result = PCR_ThCtl_GetInfo(t, &info);
    GC_push_all_stack((ptr_t)(info.ti_stkLow), (ptr_t)(info.ti_stkHi));
    return(result);
}

/* Push the contents of an old object. We treat this as stack   */
/* data only because that makes it robust against mark stack    */
/* overflow.                                                    */
PCR_ERes GC_push_old_obj(void *p, size_t size, PCR_Any data)
{
    GC_push_all_stack((ptr_t)p, (ptr_t)p + size);
    return(PCR_ERes_okay);
}

extern struct PCR_MM_ProcsRep * GC_old_allocator;
                                        /* defined in pcr_interface.c.  */

void GC_default_push_other_roots(void)
{
    /* Traverse data allocated by previous memory managers.             */
          if ((*(GC_old_allocator->mmp_enumerate))(PCR_Bool_false,
                                                   GC_push_old_obj, 0)
              != PCR_ERes_okay) {
              ABORT("Old object enumeration failed");
          }
    /* Traverse all thread stacks. */
        if (PCR_ERes_IsErr(
                PCR_ThCtl_ApplyToAllOtherThreads(GC_push_thread_stack,0))
              || PCR_ERes_IsErr(GC_push_thread_stack(PCR_Th_CurrThread(), 0))) {
              ABORT("Thread stack marking failed\n");
        }
}

# endif /* PCR */


# if defined(GC_PTHREADS) || defined(GC_WIN32_THREADS)

void GC_push_all_stacks(void);

STATIC void GC_default_push_other_roots(void)
{
    GC_push_all_stacks();
}

# endif /* GC_WIN32_THREADS || GC_PTHREADS */

void (*GC_push_other_roots)(void) = GC_default_push_other_roots;

#endif /* THREADS */

/*
 * Routines for accessing dirty bits on virtual pages.
 * There are six ways to maintain this information:
 * DEFAULT_VDB: A simple dummy implementation that treats every page
 *              as possibly dirty.  This makes incremental collection
 *              useless, but the implementation is still correct.
 * MANUAL_VDB:  Stacks and static data are always considered dirty.
 *              Heap pages are considered dirty if GC_dirty(p) has been
 *              called on some pointer p pointing to somewhere inside
 *              an object on that page.  A GC_dirty() call on a large
 *              object directly dirties only a single page, but for
 *              MANUAL_VDB we are careful to treat an object with a dirty
 *              page as completely dirty.
 *              In order to avoid races, an object must be marked dirty
 *              after it is written, and a reference to the object
 *              must be kept on a stack or in a register in the interim.
 *              With threads enabled, an object directly reachable from the
 *              stack at the time of a collection is treated as dirty.
 *              In single-threaded mode, it suffices to ensure that no
 *              collection can take place between the pointer assignment
 *              and the GC_dirty() call.
 * PCR_VDB:     Use PPCRs virtual dirty bit facility.
 * PROC_VDB:    Use the /proc facility for reading dirty bits.  Only
 *              works under some SVR4 variants.  Even then, it may be
 *              too slow to be entirely satisfactory.  Requires reading
 *              dirty bits for entire address space.  Implementations tend
 *              to assume that the client is a (slow) debugger.
 * MPROTECT_VDB:Protect pages and then catch the faults to keep track of
 *              dirtied pages.  The implementation (and implementability)
 *              is highly system dependent.  This usually fails when system
 *              calls write to a protected page.  We prevent the read system
 *              call from doing so.  It is the clients responsibility to
 *              make sure that other system calls are similarly protected
 *              or write only to the stack.
 * GWW_VDB:     Use the Win32 GetWriteWatch functions, if available, to
 *              read dirty bits.  In case it is not available (because we
 *              are running on Windows 95, Windows 2000 or earlier),
 *              MPROTECT_VDB may be defined as a fallback strategy.
 */
GC_bool GC_dirty_maintained = FALSE;

#if defined(PROC_VDB) || defined(GWW_VDB)

/* Add all pages in pht2 to pht1 */
STATIC void GC_or_pages(page_hash_table pht1, page_hash_table pht2)
{
    register int i;

    for (i = 0; i < PHT_SIZE; i++) pht1[i] |= pht2[i];
}

#endif

#ifdef GWW_VDB

# define GC_GWW_BUF_LEN (MAXHINCR * HBLKSIZE / 4096 /* X86 page size */)
  /* Still susceptible to overflow, if there are very large allocations, */
  /* and everything is dirty.                                            */
  static PVOID gww_buf[GC_GWW_BUF_LEN];

# ifdef MPROTECT_VDB
    GC_bool GC_gww_dirty_init(void)
    {
      detect_GetWriteWatch();
      return GC_GWW_AVAILABLE();
    }
# else
    void GC_dirty_init(void)
    {
      detect_GetWriteWatch();
      GC_dirty_maintained = GC_GWW_AVAILABLE();
    }
# endif

# ifdef MPROTECT_VDB
    STATIC void GC_gww_read_dirty(void)
# else
    void GC_read_dirty(void)
# endif
  {
    word i;

    BZERO(GC_grungy_pages, sizeof(GC_grungy_pages));

    for (i = 0; i != GC_n_heap_sects; ++i) {
      ULONG_PTR count;

      do {
        PVOID * pages, * pages_end;
        DWORD page_size;

        pages = gww_buf;
        count = GC_GWW_BUF_LEN;
        /*
        * GetWriteWatch is documented as returning non-zero when it fails,
        * but the documentation doesn't explicitly say why it would fail or
        * what its behaviour will be if it fails.
        * It does appear to fail, at least on recent W2K instances, if
        * the underlying memory was not allocated with the appropriate
        * flag.  This is common if GC_enable_incremental is called
        * shortly after GC initialization.  To avoid modifying the
        * interface, we silently work around such a failure, it it only
        * affects the initial (small) heap allocation.
        * If there are more dirty
        * pages than will fit in the buffer, this is not treated as a
        * failure; we must check the page count in the loop condition.
        * Since each partial call will reset the status of some
        * pages, this should eventually terminate even in the overflow
        * case.
        */
        if (GetWriteWatch_func(WRITE_WATCH_FLAG_RESET,
                               GC_heap_sects[i].hs_start,
                               GC_heap_sects[i].hs_bytes,
                               pages,
                               &count,
                               &page_size) != 0) {
          static int warn_count = 0;
          unsigned j;
          struct hblk * start = (struct hblk *)GC_heap_sects[i].hs_start;
          static struct hblk *last_warned = 0;
          size_t nblocks = divHBLKSZ(GC_heap_sects[i].hs_bytes);

          if ( i != 0 && last_warned != start && warn_count++ < 5) {
            last_warned = start;
            WARN(
              "GC_gww_read_dirty unexpectedly failed at %p: "
              "Falling back to marking all pages dirty\n", start);
          }
          for (j = 0; j < nblocks; ++j) {
              word hash = PHT_HASH(start + j);
              set_pht_entry_from_index(GC_grungy_pages, hash);
          }
          count = 1;  /* Done with this section. */
        } else /* succeeded */{
          pages_end = pages + count;
          while (pages != pages_end) {
            struct hblk * h = (struct hblk *) *pages++;
            struct hblk * h_end = (struct hblk *) ((char *) h + page_size);
            do
              set_pht_entry_from_index(GC_grungy_pages, PHT_HASH(h));
            while (++h < h_end);
          }
        }
      } while (count == GC_GWW_BUF_LEN);
      /* FIXME: It's unclear from Microsoft's documentation if this loop  */
      /* is useful.  We suspect the call just fails if the buffer fills   */
      /* up.  But that should still be handled correctly.                 */
    }

    GC_or_pages(GC_written_pages, GC_grungy_pages);
  }

# ifdef MPROTECT_VDB
    STATIC GC_bool GC_gww_page_was_dirty(struct hblk * h)
# else
    GC_bool GC_page_was_dirty(struct hblk * h)
# endif
  {
    return HDR(h) == 0 ||
            get_pht_entry_from_index(GC_grungy_pages, PHT_HASH(h));
  }

# ifdef MPROTECT_VDB
    STATIC GC_bool GC_gww_page_was_ever_dirty(struct hblk * h)
# else
    GC_bool GC_page_was_ever_dirty(struct hblk * h)
# endif
  {
    return HDR(h) == 0 ||
            get_pht_entry_from_index(GC_written_pages, PHT_HASH(h));
  }

# ifndef MPROTECT_VDB
    /*ARGSUSED*/
    void GC_remove_protection(struct hblk *h, word nblocks,
                              GC_bool is_ptrfree) {}
# endif

# endif /* GWW_VDB */

# ifdef DEFAULT_VDB

/* All of the following assume the allocation lock is held.     */

/* The client asserts that unallocated pages in the heap are never      */
/* written.                                                             */

/* Initialize virtual dirty bit implementation.                 */
void GC_dirty_init(void)
{
    if (GC_print_stats == VERBOSE)
      GC_log_printf("Initializing DEFAULT_VDB...\n");
    GC_dirty_maintained = TRUE;
}

/* Retrieve system dirty bits for heap to a local buffer.       */
/* Restore the systems notion of which pages are dirty.         */
void GC_read_dirty(void) {}

/* Is the HBLKSIZE sized page at h marked dirty in the local buffer?    */
/* If the actual page size is different, this returns TRUE if any       */
/* of the pages overlapping h are dirty.  This routine may err on the   */
/* side of labeling pages as dirty (and this implementation does).      */
/*ARGSUSED*/
GC_bool GC_page_was_dirty(struct hblk *h)
{
    return(TRUE);
}

/*
 * The following two routines are typically less crucial.  They matter
 * most with large dynamic libraries, or if we can't accurately identify
 * stacks, e.g. under Solaris 2.X.  Otherwise the following default
 * versions are adequate.
 */

/* Could any valid GC heap pointer ever have been written to this page? */
/*ARGSUSED*/
GC_bool GC_page_was_ever_dirty(struct hblk *h)
{
    return(TRUE);
}

/* A call that:                                         */
/* I) hints that [h, h+nblocks) is about to be written. */
/* II) guarantees that protection is removed.           */
/* (I) may speed up some dirty bit implementations.     */
/* (II) may be essential if we need to ensure that      */
/* pointer-free system call buffers in the heap are     */
/* not protected.                                       */
/*ARGSUSED*/
void GC_remove_protection(struct hblk *h, word nblocks, GC_bool is_ptrfree)
{
}

# endif /* DEFAULT_VDB */

# ifdef MANUAL_VDB

/* Initialize virtual dirty bit implementation.                 */
void GC_dirty_init(void)
{
    if (GC_print_stats == VERBOSE)
      GC_log_printf("Initializing MANUAL_VDB...\n");
    /* GC_dirty_pages and GC_grungy_pages are already cleared. */
    GC_dirty_maintained = TRUE;
}

/* Retrieve system dirty bits for heap to a local buffer.       */
/* Restore the systems notion of which pages are dirty.         */
void GC_read_dirty(void)
{
    BCOPY((word *)GC_dirty_pages, GC_grungy_pages,
          (sizeof GC_dirty_pages));
    BZERO((word *)GC_dirty_pages, (sizeof GC_dirty_pages));
}

/* Is the HBLKSIZE sized page at h marked dirty in the local buffer?    */
/* If the actual page size is different, this returns TRUE if any       */
/* of the pages overlapping h are dirty.  This routine may err on the   */
/* side of labeling pages as dirty (and this implementation does).      */
GC_bool GC_page_was_dirty(struct hblk *h)
{
    register word index;

    index = PHT_HASH(h);
    return(HDR(h) == 0 || get_pht_entry_from_index(GC_grungy_pages, index));
}

/* Could any valid GC heap pointer ever have been written to this page? */
/*ARGSUSED*/
GC_bool GC_page_was_ever_dirty(struct hblk *h)
{
    /* FIXME - implement me.    */
    return(TRUE);
}

/* Mark the page containing p as dirty.  Logically, this dirties the    */
/* entire object.                                                       */
void GC_dirty(ptr_t p)
{
    word index = PHT_HASH(p);
    async_set_pht_entry_from_index(GC_dirty_pages, index);
}

/*ARGSUSED*/
void GC_remove_protection(struct hblk *h, word nblocks, GC_bool is_ptrfree)
{
}

# endif /* MANUAL_VDB */


# ifdef MPROTECT_VDB

/*
 * See DEFAULT_VDB for interface descriptions.
 */

/*
 * This implementation maintains dirty bits itself by catching write
 * faults and keeping track of them.  We assume nobody else catches
 * SIGBUS or SIGSEGV.  We assume no write faults occur in system calls.
 * This means that clients must ensure that system calls don't write
 * to the write-protected heap.  Probably the best way to do this is to
 * ensure that system calls write at most to POINTERFREE objects in the
 * heap, and do even that only if we are on a platform on which those
 * are not protected.  Another alternative is to wrap system calls
 * (see example for read below), but the current implementation holds
 * applications.
 * We assume the page size is a multiple of HBLKSIZE.
 * We prefer them to be the same.  We avoid protecting POINTERFREE
 * objects only if they are the same.
 */

# if !defined(MSWIN32) && !defined(MSWINCE) && !defined(DARWIN)

#   include <sys/mman.h>
#   include <signal.h>
#   include <sys/syscall.h>

#   define PROTECT(addr, len) \
          if (mprotect((caddr_t)(addr), (size_t)(len), \
                       PROT_READ | OPT_PROT_EXEC) < 0) { \
            ABORT("mprotect failed"); \
          }
#   define UNPROTECT(addr, len) \
          if (mprotect((caddr_t)(addr), (size_t)(len), \
                       PROT_WRITE | PROT_READ | OPT_PROT_EXEC ) < 0) { \
            ABORT("un-mprotect failed"); \
          }

# else

# ifdef DARWIN
    /* Using vm_protect (mach syscall) over mprotect (BSD syscall) seems to
       decrease the likelihood of some of the problems described below. */
#   include <mach/vm_map.h>
    STATIC mach_port_t GC_task_self = 0;
#   define PROTECT(addr,len) \
        if(vm_protect(GC_task_self,(vm_address_t)(addr),(vm_size_t)(len), \
                FALSE,VM_PROT_READ) != KERN_SUCCESS) { \
            ABORT("vm_protect (PROTECT) failed"); \
        }
#   define UNPROTECT(addr,len) \
        if(vm_protect(GC_task_self,(vm_address_t)(addr),(vm_size_t)(len), \
                FALSE,VM_PROT_READ|VM_PROT_WRITE) != KERN_SUCCESS) { \
            ABORT("vm_protect (UNPROTECT) failed"); \
        }
# else

#   ifndef MSWINCE
#     include <signal.h>
#   endif

    static DWORD protect_junk;
#   define PROTECT(addr, len) \
          if (!VirtualProtect((addr), (len), PAGE_EXECUTE_READ, \
                              &protect_junk)) { \
            GC_printf("Last error code: %lx\n", (long)GetLastError()); \
            ABORT("VirtualProtect failed"); \
          }
#   define UNPROTECT(addr, len) \
          if (!VirtualProtect((addr), (len), PAGE_EXECUTE_READWRITE, \
                              &protect_junk)) { \
            ABORT("un-VirtualProtect failed"); \
          }
# endif /* !DARWIN */
# endif /* MSWIN32 || MSWINCE || DARWIN */

#if defined(MSWIN32)
    typedef LPTOP_LEVEL_EXCEPTION_FILTER SIG_HNDLR_PTR;
#   undef SIG_DFL
#   define SIG_DFL (LPTOP_LEVEL_EXCEPTION_FILTER)((signed_word)-1)
#elif defined(MSWINCE)
    typedef LONG (WINAPI *SIG_HNDLR_PTR)(struct _EXCEPTION_POINTERS *);
#   undef SIG_DFL
#   define SIG_DFL (SIG_HNDLR_PTR) (-1)
#elif defined(DARWIN)
    typedef void (* SIG_HNDLR_PTR)();
#else
    typedef void (* SIG_HNDLR_PTR)(int, siginfo_t *, void *);
    typedef void (* PLAIN_HNDLR_PTR)(int);
#endif

#if defined(__GLIBC__)
#   if __GLIBC__ < 2 || __GLIBC__ == 2 && __GLIBC_MINOR__ < 2
#       error glibc too old?
#   endif
#endif

#ifndef DARWIN
STATIC SIG_HNDLR_PTR GC_old_segv_handler = 0;
                        /* Also old MSWIN32 ACCESS_VIOLATION filter */
#if !defined(MSWIN32) && !defined(MSWINCE)
STATIC SIG_HNDLR_PTR GC_old_bus_handler = 0;
STATIC GC_bool GC_old_bus_handler_used_si = FALSE;
STATIC GC_bool GC_old_segv_handler_used_si = FALSE;
#endif
#endif /* !DARWIN */

#if defined(THREADS)
/* We need to lock around the bitmap update in the write fault handler  */
/* in order to avoid the risk of losing a bit.  We do this with a       */
/* test-and-set spin lock if we know how to do that.  Otherwise we      */
/* check whether we are already in the handler and use the dumb but     */
/* safe fallback algorithm of setting all bits in the word.             */
/* Contention should be very rare, so we do the minimum to handle it    */
/* correctly.                                                           */
#ifdef AO_HAVE_test_and_set_acquire
  volatile AO_TS_t GC_fault_handler_lock = 0;
  static void async_set_pht_entry_from_index(volatile page_hash_table db,
                                             size_t index)
  {
    while (AO_test_and_set_acquire(&GC_fault_handler_lock) == AO_TS_SET) {
      /* empty */
    }
    /* Could also revert to set_pht_entry_from_index_safe if initial    */
    /* GC_test_and_set fails.                                           */
    set_pht_entry_from_index(db, index);
    AO_CLEAR(&GC_fault_handler_lock);
  }
#else /* !AO_HAVE_test_and_set_acquire */
# error No test_and_set operation: Introduces a race.
  /* THIS WOULD BE INCORRECT!                                           */
  /* The dirty bit vector may be temporarily wrong,                     */
  /* just before we notice the conflict and correct it. We may end up   */
  /* looking at it while it's wrong.  But this requires contention      */
  /* exactly when a GC is triggered, which seems far less likely to     */
  /* fail than the old code, which had no reported failures.  Thus we   */
  /* leave it this way while we think of something better, or support   */
  /* GC_test_and_set on the remaining platforms.                        */
  static volatile word currently_updating = 0;
  static void async_set_pht_entry_from_index(volatile page_hash_table db,
                                             size_t index)
  {
    unsigned int update_dummy;
    currently_updating = (word)(&update_dummy);
    set_pht_entry_from_index(db, index);
    /* If we get contention in the 10 or so instruction window here,    */
    /* and we get stopped by a GC between the two updates, we lose!     */
    if (currently_updating != (word)(&update_dummy)) {
        set_pht_entry_from_index_safe(db, index);
        /* We claim that if two threads concurrently try to update the  */
        /* dirty bit vector, the first one to execute UPDATE_START      */
        /* will see it changed when UPDATE_END is executed.  (Note that */
        /* &update_dummy must differ in two distinct threads.)  It      */
        /* will then execute set_pht_entry_from_index_safe, thus        */
        /* returning us to a safe state, though not soon enough.        */
    }
  }
#endif /* !AO_HAVE_test_and_set_acquire */
#else /* !THREADS */
# define async_set_pht_entry_from_index(db, index) \
                        set_pht_entry_from_index(db, index)
#endif /* !THREADS */

#ifdef CHECKSUMS
  void GC_record_fault(struct hblk * h);
        /* From checksums.c */
#endif

#if !defined(DARWIN)
#   include <errno.h>
#   if defined(FREEBSD)
#     define SIG_OK TRUE
#     define CODE_OK (si -> si_code == BUS_PAGE_FAULT)
#   elif defined(OSF1)
#     define SIG_OK (sig == SIGSEGV)
#     define CODE_OK (si -> si_code == 2 /* experimentally determined */)
#   elif defined(IRIX5)
#     define SIG_OK (sig == SIGSEGV)
#     define CODE_OK (si -> si_code == EACCES)
#   elif defined(HURD)
#     define SIG_OK (sig == SIGBUS || sig == SIGSEGV)
#     define CODE_OK  TRUE
#   elif defined(LINUX)
#     define SIG_OK (sig == SIGSEGV)
#     define CODE_OK TRUE
        /* Empirically c.trapno == 14, on IA32, but is that useful?     */
        /* Should probably consider alignment issues on other           */
        /* architectures.                                               */
#   elif defined(HPUX)
#     define SIG_OK (sig == SIGSEGV || sig == SIGBUS)
#     define CODE_OK (si -> si_code == SEGV_ACCERR) \
                     || (si -> si_code == BUS_ADRERR) \
                     || (si -> si_code == BUS_UNKNOWN) \
                     || (si -> si_code == SEGV_UNKNOWN) \
                     || (si -> si_code == BUS_OBJERR)
#   elif defined(SUNOS5SIGS)
#     define SIG_OK (sig == SIGSEGV)
#     define CODE_OK (si -> si_code == SEGV_ACCERR)
#   elif defined(MSWIN32) || defined(MSWINCE)
#     define SIG_OK (exc_info -> ExceptionRecord -> ExceptionCode \
                     == STATUS_ACCESS_VIOLATION)
#     define CODE_OK (exc_info -> ExceptionRecord -> ExceptionInformation[0] \
                      == 1) /* Write fault */
#   endif

# if defined(MSWIN32) || defined(MSWINCE)
    LONG WINAPI GC_write_fault_handler(struct _EXCEPTION_POINTERS *exc_info)
# else
#   include <ucontext.h>
    /*ARGSUSED*/
    STATIC void GC_write_fault_handler(int sig, siginfo_t *si, void *raw_sc)
# endif /* MSWIN32 || MSWINCE */
{
#   if !defined(MSWIN32) && !defined(MSWINCE)
        char *addr = si -> si_addr;
#   else
        char * addr = (char *) (exc_info -> ExceptionRecord
                                -> ExceptionInformation[1]);
#   endif
    unsigned i;

    if (SIG_OK && CODE_OK) {
        register struct hblk * h =
                        (struct hblk *)((word)addr & ~(GC_page_size-1));
        GC_bool in_allocd_block;
#       ifdef CHECKSUMS
          GC_record_fault(h);
#       endif /* CHECKSUMS */

#       ifdef SUNOS5SIGS
            /* Address is only within the correct physical page.        */
            in_allocd_block = FALSE;
            for (i = 0; i < divHBLKSZ(GC_page_size); i++) {
              if (HDR(h+i) != 0) {
                in_allocd_block = TRUE;
              }
            }
#       else
            in_allocd_block = (HDR(addr) != 0);
#       endif
        if (!in_allocd_block) {
            /* FIXME - We should make sure that we invoke the   */
            /* old handler with the appropriate calling         */
            /* sequence, which often depends on SA_SIGINFO.     */

            /* Heap blocks now begin and end on page boundaries */
            SIG_HNDLR_PTR old_handler;

#           if defined(MSWIN32) || defined(MSWINCE)
                old_handler = GC_old_segv_handler;
#           else
                GC_bool used_si;

                if (sig == SIGSEGV) {
                   old_handler = GC_old_segv_handler;
                   used_si = GC_old_segv_handler_used_si;
                } else {
                   old_handler = GC_old_bus_handler;
                   used_si = GC_old_bus_handler_used_si;
                }
#           endif

            if (old_handler == (SIG_HNDLR_PTR)SIG_DFL) {
#               if !defined(MSWIN32) && !defined(MSWINCE)
                    GC_err_printf("Segfault at %p\n", addr);
                    ABORT("Unexpected bus error or segmentation fault");
#               else
                    return(EXCEPTION_CONTINUE_SEARCH);
#               endif
            } else {
                /*
                 * FIXME: This code should probably check if the
                 * old signal handler used the traditional style and
                 * if so call it using that style.
                 */
#               if defined(MSWIN32) || defined(MSWINCE)
                    return((*old_handler)(exc_info));
#               else
                    if (used_si)
                      ((SIG_HNDLR_PTR)old_handler) (sig, si, raw_sc);
                    else
                      /* FIXME: should pass nonstandard args as well. */
                      ((PLAIN_HNDLR_PTR)old_handler) (sig);
                    return;
#               endif
            }
        }
        UNPROTECT(h, GC_page_size);
        /* We need to make sure that no collection occurs between       */
        /* the UNPROTECT and the setting of the dirty bit.  Otherwise   */
        /* a write by a third thread might go unnoticed.  Reversing     */
        /* the order is just as bad, since we would end up unprotecting */
        /* a page in a GC cycle during which it's not marked.           */
        /* Currently we do this by disabling the thread stopping        */
        /* signals while this handler is running.  An alternative might */
        /* be to record the fact that we're about to unprotect, or      */
        /* have just unprotected a page in the GC's thread structure,   */
        /* and then to have the thread stopping code set the dirty      */
        /* flag, if necessary.                                          */
        for (i = 0; i < divHBLKSZ(GC_page_size); i++) {
            size_t index = PHT_HASH(h+i);

            async_set_pht_entry_from_index(GC_dirty_pages, index);
        }
        /* The write may not take place before dirty bits are read.     */
        /* But then we'll fault again ...                               */
#       if defined(MSWIN32) || defined(MSWINCE)
            return(EXCEPTION_CONTINUE_EXECUTION);
#       else
            return;
#       endif
    }
#if defined(MSWIN32) || defined(MSWINCE)
    return EXCEPTION_CONTINUE_SEARCH;
#else
    GC_err_printf("Segfault at %p\n", addr);
    ABORT("Unexpected bus error or segmentation fault");
#endif
}
#endif /* !DARWIN */

/*
 * We hold the allocation lock.  We expect block h to be written
 * shortly.  Ensure that all pages containing any part of the n hblks
 * starting at h are no longer protected.  If is_ptrfree is false,
 * also ensure that they will subsequently appear to be dirty.
 */
void GC_remove_protection(struct hblk *h, word nblocks, GC_bool is_ptrfree)
{
    struct hblk * h_trunc;  /* Truncated to page boundary */
    struct hblk * h_end;    /* Page boundary following block end */
    struct hblk * current;

#   if defined(GWW_VDB)
      if (GC_GWW_AVAILABLE()) return;
#   endif
    if (!GC_dirty_maintained) return;
    h_trunc = (struct hblk *)((word)h & ~(GC_page_size-1));
    h_end = (struct hblk *)(((word)(h + nblocks) + GC_page_size-1)
                            & ~(GC_page_size-1));
    if (h_end == h_trunc + 1 &&
        get_pht_entry_from_index(GC_dirty_pages, PHT_HASH(h_trunc))) {
        /* already marked dirty, and hence unprotected. */
        return;
    }
    for (current = h_trunc; current < h_end; ++current) {
        size_t index = PHT_HASH(current);
        if (!is_ptrfree || current < h || current >= h + nblocks) {
            async_set_pht_entry_from_index(GC_dirty_pages, index);
        }
    }
    UNPROTECT(h_trunc, (ptr_t)h_end - (ptr_t)h_trunc);
}

#if !defined(DARWIN)
void GC_dirty_init(void)
{
#   if !defined(MSWIN32) && !defined(MSWINCE)
      struct sigaction  act, oldact;
      act.sa_flags      = SA_RESTART | SA_SIGINFO;
      act.sa_sigaction = GC_write_fault_handler;
      (void)sigemptyset(&act.sa_mask);
#     ifdef SIG_SUSPEND
        /* Arrange to postpone SIG_SUSPEND while we're in a write fault */
        /* handler.  This effectively makes the handler atomic w.r.t.   */
        /* stopping the world for GC.                                   */
        (void)sigaddset(&act.sa_mask, SIG_SUSPEND);
#     endif /* SIG_SUSPEND */
#   endif
    if (GC_print_stats == VERBOSE)
        GC_log_printf(
                "Initializing mprotect virtual dirty bit implementation\n");
    GC_dirty_maintained = TRUE;
    if (GC_page_size % HBLKSIZE != 0) {
        GC_err_printf("Page size not multiple of HBLKSIZE\n");
        ABORT("Page size not multiple of HBLKSIZE");
    }
#   if !defined(MSWIN32) && !defined(MSWINCE)
#     if defined(GC_IRIX_THREADS)
        sigaction(SIGSEGV, 0, &oldact);
        sigaction(SIGSEGV, &act, 0);
#     else
        {
          int res = sigaction(SIGSEGV, &act, &oldact);
          if (res != 0) ABORT("Sigaction failed");
        }
#     endif
      if (oldact.sa_flags & SA_SIGINFO) {
        GC_old_segv_handler = oldact.sa_sigaction;
        GC_old_segv_handler_used_si = TRUE;
      } else {
        GC_old_segv_handler = (SIG_HNDLR_PTR)oldact.sa_handler;
        GC_old_segv_handler_used_si = FALSE;
      }
      if (GC_old_segv_handler == (SIG_HNDLR_PTR)SIG_IGN) {
        GC_err_printf("Previously ignored segmentation violation!?\n");
        GC_old_segv_handler = (SIG_HNDLR_PTR)SIG_DFL;
      }
      if (GC_old_segv_handler != (SIG_HNDLR_PTR)SIG_DFL) {
        if (GC_print_stats == VERBOSE)
          GC_log_printf("Replaced other SIGSEGV handler\n");
      }
#   if defined(HPUX) || defined(LINUX) || defined(HURD) \
      || (defined(FREEBSD) && defined(SUNOS5SIGS))
      sigaction(SIGBUS, &act, &oldact);
      if (oldact.sa_flags & SA_SIGINFO) {
        GC_old_bus_handler = oldact.sa_sigaction;
        GC_old_bus_handler_used_si = TRUE;
      } else {
        GC_old_bus_handler = (SIG_HNDLR_PTR)oldact.sa_handler;
        GC_old_bus_handler_used_si = FALSE;
      }
      if (GC_old_bus_handler == (SIG_HNDLR_PTR)SIG_IGN) {
             GC_err_printf("Previously ignored bus error!?\n");
             GC_old_bus_handler = (SIG_HNDLR_PTR)SIG_DFL;
      }
      if (GC_old_bus_handler != (SIG_HNDLR_PTR)SIG_DFL) {
        if (GC_print_stats == VERBOSE)
          GC_log_printf("Replaced other SIGBUS handler\n");
      }
#   endif /* HPUX || LINUX || HURD || (FREEBSD && SUNOS5SIGS) */
#   endif /* ! MS windows */
#   if defined(GWW_VDB)
      if (GC_gww_dirty_init())
        return;
#   endif
#   if defined(MSWIN32)
      GC_old_segv_handler = SetUnhandledExceptionFilter(GC_write_fault_handler);
      if (GC_old_segv_handler != NULL) {
        if (GC_print_stats)
          GC_log_printf("Replaced other UnhandledExceptionFilter\n");
      } else {
          GC_old_segv_handler = SIG_DFL;
      }
#   elif defined(MSWINCE)
      /* MPROTECT_VDB is unsupported for WinCE at present.      */
      /* FIXME: implement it (if possible). */
#   endif
}
#endif /* !DARWIN */

GC_API int GC_CALL GC_incremental_protection_needs(void)
{
    if (GC_page_size == HBLKSIZE) {
        return GC_PROTECTS_POINTER_HEAP;
    } else {
        return GC_PROTECTS_POINTER_HEAP | GC_PROTECTS_PTRFREE_HEAP;
    }
}

#define HAVE_INCREMENTAL_PROTECTION_NEEDS

#define IS_PTRFREE(hhdr) ((hhdr)->hb_descr == 0)

#define PAGE_ALIGNED(x) !((word)(x) & (GC_page_size - 1))
STATIC void GC_protect_heap(void)
{
    ptr_t start;
    size_t len;
    struct hblk * current;
    struct hblk * current_start;  /* Start of block to be protected. */
    struct hblk * limit;
    unsigned i;
    GC_bool protect_all =
          (0 != (GC_incremental_protection_needs() & GC_PROTECTS_PTRFREE_HEAP));
    for (i = 0; i < GC_n_heap_sects; i++) {
        start = GC_heap_sects[i].hs_start;
        len = GC_heap_sects[i].hs_bytes;
        if (protect_all) {
          PROTECT(start, len);
        } else {
          GC_ASSERT(PAGE_ALIGNED(len))
          GC_ASSERT(PAGE_ALIGNED(start))
          current_start = current = (struct hblk *)start;
          limit = (struct hblk *)(start + len);
          while (current < limit) {
            hdr * hhdr;
            word nhblks;
            GC_bool is_ptrfree;

            GC_ASSERT(PAGE_ALIGNED(current));
            GET_HDR(current, hhdr);
            if (IS_FORWARDING_ADDR_OR_NIL(hhdr)) {
              /* This can happen only if we're at the beginning of a    */
              /* heap segment, and a block spans heap segments.         */
              /* We will handle that block as part of the preceding     */
              /* segment.                                               */
              GC_ASSERT(current_start == current);
              current_start = ++current;
              continue;
            }
            if (HBLK_IS_FREE(hhdr)) {
              GC_ASSERT(PAGE_ALIGNED(hhdr -> hb_sz));
              nhblks = divHBLKSZ(hhdr -> hb_sz);
              is_ptrfree = TRUE;        /* dirty on alloc */
            } else {
              nhblks = OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz);
              is_ptrfree = IS_PTRFREE(hhdr);
            }
            if (is_ptrfree) {
              if (current_start < current) {
                PROTECT(current_start, (ptr_t)current - (ptr_t)current_start);
              }
              current_start = (current += nhblks);
            } else {
              current += nhblks;
            }
          }
          if (current_start < current) {
            PROTECT(current_start, (ptr_t)current - (ptr_t)current_start);
          }
        }
    }
}

/* We assume that either the world is stopped or its OK to lose dirty   */
/* bits while this is happenning (as in GC_enable_incremental).         */
void GC_read_dirty(void)
{
#   if defined(GWW_VDB)
      if (GC_GWW_AVAILABLE()) {
        GC_gww_read_dirty();
        return;
      }
#   endif
    BCOPY((word *)GC_dirty_pages, GC_grungy_pages,
          (sizeof GC_dirty_pages));
    BZERO((word *)GC_dirty_pages, (sizeof GC_dirty_pages));
    GC_protect_heap();
}

GC_bool GC_page_was_dirty(struct hblk *h)
{
    register word index;

#   if defined(GWW_VDB)
      if (GC_GWW_AVAILABLE())
        return GC_gww_page_was_dirty(h);
#   endif

    index = PHT_HASH(h);
    return(HDR(h) == 0 || get_pht_entry_from_index(GC_grungy_pages, index));
}

/*
 * Acquiring the allocation lock here is dangerous, since this
 * can be called from within GC_call_with_alloc_lock, and the cord
 * package does so.  On systems that allow nested lock acquisition, this
 * happens to work.
 * On other systems, SET_LOCK_HOLDER and friends must be suitably defined.
 */

#if 0
static GC_bool syscall_acquired_lock = FALSE;   /* Protected by GC lock. */

void GC_begin_syscall(void)
{
    /* FIXME: Resurrecting this code would require fixing the   */
    /* test, which can spuriously return TRUE.                  */
    if (!I_HOLD_LOCK()) {
        LOCK();
        syscall_acquired_lock = TRUE;
    }
}

void GC_end_syscall(void)
{
    if (syscall_acquired_lock) {
        syscall_acquired_lock = FALSE;
        UNLOCK();
    }
}

void GC_unprotect_range(ptr_t addr, word len)
{
    struct hblk * start_block;
    struct hblk * end_block;
    register struct hblk *h;
    ptr_t obj_start;

    if (!GC_dirty_maintained) return;
    obj_start = GC_base(addr);
    if (obj_start == 0) return;
    if (GC_base(addr + len - 1) != obj_start) {
        ABORT("GC_unprotect_range(range bigger than object)");
    }
    start_block = (struct hblk *)((word)addr & ~(GC_page_size - 1));
    end_block = (struct hblk *)((word)(addr + len - 1) & ~(GC_page_size - 1));
    end_block += GC_page_size/HBLKSIZE - 1;
    for (h = start_block; h <= end_block; h++) {
        register word index = PHT_HASH(h);

        async_set_pht_entry_from_index(GC_dirty_pages, index);
    }
    UNPROTECT(start_block,
              ((ptr_t)end_block - (ptr_t)start_block) + HBLKSIZE);
}


/* We no longer wrap read by default, since that was causing too many   */
/* problems.  It is preferred that the client instead avoids writing    */
/* to the write-protected heap with a system call.                      */
/* This still serves as sample code if you do want to wrap system calls.*/

#if !defined(MSWIN32) && !defined(MSWINCE) && !defined(GC_USE_LD_WRAP)
/* Replacement for UNIX system call.                                      */
/* Other calls that write to the heap should be handled similarly.        */
/* Note that this doesn't work well for blocking reads:  It will hold     */
/* the allocation lock for the entire duration of the call. Multithreaded */
/* clients should really ensure that it won't block, either by setting    */
/* the descriptor nonblocking, or by calling select or poll first, to     */
/* make sure that input is available.                                     */
/* Another, preferred alternative is to ensure that system calls never    */
/* write to the protected heap (see above).                               */
# include <unistd.h>
# include <sys/uio.h>
ssize_t read(int fd, void *buf, size_t nbyte)
{
    int result;

    GC_begin_syscall();
    GC_unprotect_range(buf, (word)nbyte);
#   if defined(IRIX5) || defined(GC_LINUX_THREADS)
        /* Indirect system call may not always be easily available.     */
        /* We could call _read, but that would interfere with the       */
        /* libpthread interception of read.                             */
        /* On Linux, we have to be careful with the linuxthreads        */
        /* read interception.                                           */
        {
            struct iovec iov;

            iov.iov_base = buf;
            iov.iov_len = nbyte;
            result = readv(fd, &iov, 1);
        }
#   else
#     if defined(HURD)
        result = __read(fd, buf, nbyte);
#     else
        /* The two zero args at the end of this list are because one
           IA-64 syscall() implementation actually requires six args
           to be passed, even though they aren't always used. */
        result = syscall(SYS_read, fd, buf, nbyte, 0, 0);
#     endif /* !HURD */
#   endif
    GC_end_syscall();
    return(result);
}
#endif /* !MSWIN32 && !MSWINCE && !GC_LINUX_THREADS */

#if defined(GC_USE_LD_WRAP) && !defined(THREADS)
    /* We use the GNU ld call wrapping facility.                        */
    /* This requires that the linker be invoked with "--wrap read".     */
    /* This can be done by passing -Wl,"--wrap read" to gcc.            */
    /* I'm not sure that this actually wraps whatever version of read   */
    /* is called by stdio.  That code also mentions __read.             */
#   include <unistd.h>
    ssize_t __wrap_read(int fd, void *buf, size_t nbyte)
    {
        int result;

        GC_begin_syscall();
        GC_unprotect_range(buf, (word)nbyte);
        result = __real_read(fd, buf, nbyte);
        GC_end_syscall();
        return(result);
    }

    /* We should probably also do this for __read, or whatever stdio    */
    /* actually calls.                                                  */
#endif

#endif /* 0 */

/*ARGSUSED*/
GC_bool GC_page_was_ever_dirty(struct hblk *h)
{
#   if defined(GWW_VDB)
      if (GC_GWW_AVAILABLE())
        return GC_gww_page_was_ever_dirty(h);
#   endif
    return(TRUE);
}

# endif /* MPROTECT_VDB */

# ifdef PROC_VDB

/*
 * See DEFAULT_VDB for interface descriptions.
 */

/*
 * This implementation assumes a Solaris 2.X like /proc pseudo-file-system
 * from which we can read page modified bits.  This facility is far from
 * optimal (e.g. we would like to get the info for only some of the
 * address space), but it avoids intercepting system calls.
 */

#include <errno.h>
#include <sys/types.h>
#include <sys/signal.h>
#include <sys/fault.h>
#include <sys/syscall.h>
#include <sys/procfs.h>
#include <sys/stat.h>

#define INITIAL_BUF_SZ 16384
STATIC word GC_proc_buf_size = INITIAL_BUF_SZ;
STATIC char *GC_proc_buf = NULL;

STATIC int GC_proc_fd = 0;

void GC_dirty_init(void)
{
    int fd;
    char buf[30];

    GC_dirty_maintained = TRUE;
    if (GC_bytes_allocd != 0 || GC_bytes_allocd_before_gc != 0) {
        register int i;

        for (i = 0; i < PHT_SIZE; i++) GC_written_pages[i] = (word)(-1);
        if (GC_print_stats == VERBOSE)
            GC_log_printf(
                      "Allocated bytes:%lu:all pages may have been written\n",
                      (unsigned long)
                                (GC_bytes_allocd + GC_bytes_allocd_before_gc));
    }
    sprintf(buf, "/proc/%ld", (long)getpid());
    fd = open(buf, O_RDONLY);
    if (fd < 0) {
        ABORT("/proc open failed");
    }
    GC_proc_fd = syscall(SYS_ioctl, fd, PIOCOPENPD, 0);
    close(fd);
    syscall(SYS_fcntl, GC_proc_fd, F_SETFD, FD_CLOEXEC);
    if (GC_proc_fd < 0) {
        ABORT("/proc ioctl failed");
    }
    GC_proc_buf = GC_scratch_alloc(GC_proc_buf_size);
}

/* Ignore write hints. They don't help us here. */
/*ARGSUSED*/
void GC_remove_protection(struct hblk *h, word nblocks, GC_bool is_ptrfree)
{
}

# define READ(fd,buf,nbytes) read(fd, buf, nbytes)

void GC_read_dirty(void)
{
    unsigned long ps, np;
    int nmaps;
    ptr_t vaddr;
    struct prasmap * map;
    char * bufp;
    ptr_t current_addr, limit;
    int i;

    BZERO(GC_grungy_pages, (sizeof GC_grungy_pages));

    bufp = GC_proc_buf;
    if (READ(GC_proc_fd, bufp, GC_proc_buf_size) <= 0) {
        if (GC_print_stats)
            GC_log_printf("/proc read failed: GC_proc_buf_size = %lu\n",
                          (unsigned long)GC_proc_buf_size);
        {
            /* Retry with larger buffer. */
            word new_size = 2 * GC_proc_buf_size;
            char * new_buf = GC_scratch_alloc(new_size);

            if (new_buf != 0) {
                GC_proc_buf = bufp = new_buf;
                GC_proc_buf_size = new_size;
            }
            if (READ(GC_proc_fd, bufp, GC_proc_buf_size) <= 0) {
                WARN("Insufficient space for /proc read\n", 0);
                /* Punt:        */
                memset(GC_grungy_pages, 0xff, sizeof (page_hash_table));
                memset(GC_written_pages, 0xff, sizeof(page_hash_table));
                return;
            }
        }
    }
    /* Copy dirty bits into GC_grungy_pages */
        nmaps = ((struct prpageheader *)bufp) -> pr_nmap;
        /* printf( "nmaps = %d, PG_REFERENCED = %d, PG_MODIFIED = %d\n",
                     nmaps, PG_REFERENCED, PG_MODIFIED); */
        bufp = bufp + sizeof(struct prpageheader);
        for (i = 0; i < nmaps; i++) {
            map = (struct prasmap *)bufp;
            vaddr = (ptr_t)(map -> pr_vaddr);
            ps = map -> pr_pagesize;
            np = map -> pr_npage;
            /* printf("vaddr = 0x%X, ps = 0x%X, np = 0x%X\n", vaddr, ps, np); */
            limit = vaddr + ps * np;
            bufp += sizeof (struct prasmap);
            for (current_addr = vaddr;
                 current_addr < limit; current_addr += ps) {
                if ((*bufp++) & PG_MODIFIED) {
                    register struct hblk * h = (struct hblk *) current_addr;

                    while ((ptr_t)h < current_addr + ps) {
                        register word index = PHT_HASH(h);

                        set_pht_entry_from_index(GC_grungy_pages, index);
                        h++;
                    }
                }
            }
            bufp += sizeof(long) - 1;
            bufp = (char *)((unsigned long)bufp & ~(sizeof(long)-1));
        }
    /* Update GC_written_pages. */
        GC_or_pages(GC_written_pages, GC_grungy_pages);
}

#undef READ

GC_bool GC_page_was_dirty(struct hblk *h)
{
    register word index = PHT_HASH(h);

    return get_pht_entry_from_index(GC_grungy_pages, index);
}

GC_bool GC_page_was_ever_dirty(struct hblk *h)
{
    register word index = PHT_HASH(h);

    return get_pht_entry_from_index(GC_written_pages, index);
}

# endif /* PROC_VDB */


# ifdef PCR_VDB

# include "vd/PCR_VD.h"

# define NPAGES (32*1024)       /* 128 MB */

PCR_VD_DB GC_grungy_bits[NPAGES];

STATIC ptr_t GC_vd_base = NULL;
                        /* Address corresponding to GC_grungy_bits[0]   */
                        /* HBLKSIZE aligned.                            */

void GC_dirty_init(void)
{
    GC_dirty_maintained = TRUE;
    /* For the time being, we assume the heap generally grows up */
    GC_vd_base = GC_heap_sects[0].hs_start;
    if (GC_vd_base == 0) {
        ABORT("Bad initial heap segment");
    }
    if (PCR_VD_Start(HBLKSIZE, GC_vd_base, NPAGES*HBLKSIZE)
        != PCR_ERes_okay) {
        ABORT("dirty bit initialization failed");
    }
}

void GC_read_dirty(void)
{
    /* lazily enable dirty bits on newly added heap sects */
    {
        static int onhs = 0;
        int nhs = GC_n_heap_sects;
        for( ; onhs < nhs; onhs++ ) {
            PCR_VD_WriteProtectEnable(
                    GC_heap_sects[onhs].hs_start,
                    GC_heap_sects[onhs].hs_bytes );
        }
    }


    if (PCR_VD_Clear(GC_vd_base, NPAGES*HBLKSIZE, GC_grungy_bits)
        != PCR_ERes_okay) {
        ABORT("dirty bit read failed");
    }
}

GC_bool GC_page_was_dirty(struct hblk *h)
{
    if((ptr_t)h < GC_vd_base || (ptr_t)h >= GC_vd_base + NPAGES*HBLKSIZE) {
        return(TRUE);
    }
    return(GC_grungy_bits[h - (struct hblk *)GC_vd_base] & PCR_VD_DB_dirtyBit);
}

/*ARGSUSED*/
void GC_remove_protection(struct hblk *h, word nblocks, GC_bool is_ptrfree)
{
    PCR_VD_WriteProtectDisable(h, nblocks*HBLKSIZE);
    PCR_VD_WriteProtectEnable(h, nblocks*HBLKSIZE);
}

# endif /* PCR_VDB */

#if defined(MPROTECT_VDB) && defined(DARWIN)
/* The following sources were used as a *reference* for this exception handling
   code:
      1. Apple's mach/xnu documentation
      2. Timothy J. Wood's "Mach Exception Handlers 101" post to the
         omnigroup's macosx-dev list.
         www.omnigroup.com/mailman/archive/macosx-dev/2000-June/014178.html
      3. macosx-nat.c from Apple's GDB source code.
*/

/* The bug that caused all this trouble should now be fixed. This should
   eventually be removed if all goes well. */

/* #define BROKEN_EXCEPTION_HANDLING */

#include <mach/mach.h>
#include <mach/mach_error.h>
#include <mach/thread_status.h>
#include <mach/exception.h>
#include <mach/task.h>
#include <pthread.h>

void GC_darwin_register_mach_handler_thread(mach_port_t);

/* These are not defined in any header, although they are documented */
extern boolean_t
exc_server(mach_msg_header_t *, mach_msg_header_t *);

extern kern_return_t
exception_raise(mach_port_t, mach_port_t, mach_port_t, exception_type_t,
                exception_data_t, mach_msg_type_number_t);

extern kern_return_t
exception_raise_state(mach_port_t, mach_port_t, mach_port_t, exception_type_t,
                      exception_data_t, mach_msg_type_number_t,
                      thread_state_flavor_t*, thread_state_t,
                      mach_msg_type_number_t, thread_state_t,
                      mach_msg_type_number_t*);

extern kern_return_t
exception_raise_state_identity(mach_port_t, mach_port_t, mach_port_t,
                               exception_type_t, exception_data_t,
                               mach_msg_type_number_t, thread_state_flavor_t*,
                               thread_state_t, mach_msg_type_number_t,
                               thread_state_t, mach_msg_type_number_t*);


#define MAX_EXCEPTION_PORTS 16

static struct {
    mach_msg_type_number_t count;
    exception_mask_t      masks[MAX_EXCEPTION_PORTS];
    exception_handler_t   ports[MAX_EXCEPTION_PORTS];
    exception_behavior_t  behaviors[MAX_EXCEPTION_PORTS];
    thread_state_flavor_t flavors[MAX_EXCEPTION_PORTS];
} GC_old_exc_ports;

STATIC struct {
    mach_port_t exception;
#   if defined(THREADS)
      mach_port_t reply;
#   endif
} GC_ports = {0};

typedef struct {
    mach_msg_header_t head;
} GC_msg_t;

typedef enum {
    GC_MP_NORMAL, GC_MP_DISCARDING, GC_MP_STOPPED
} GC_mprotect_state_t;

/* FIXME: 1 and 2 seem to be safe to use in the msgh_id field,
   but it isn't documented. Use the source and see if they
   should be ok. */
#define ID_STOP 1
#define ID_RESUME 2

/* These values are only used on the reply port */
#define ID_ACK 3

#if defined(THREADS)

STATIC GC_mprotect_state_t GC_mprotect_state = 0;

/* The following should ONLY be called when the world is stopped  */
STATIC void GC_mprotect_thread_notify(mach_msg_id_t id)
{

  struct {
    GC_msg_t msg;
    mach_msg_trailer_t trailer;
  } buf;

  mach_msg_return_t r;
  /* remote, local */
  buf.msg.head.msgh_bits = MACH_MSGH_BITS(MACH_MSG_TYPE_MAKE_SEND, 0);
  buf.msg.head.msgh_size = sizeof(buf.msg);
  buf.msg.head.msgh_remote_port = GC_ports.exception;
  buf.msg.head.msgh_local_port = MACH_PORT_NULL;
  buf.msg.head.msgh_id = id;

  r = mach_msg(&buf.msg.head, MACH_SEND_MSG | MACH_RCV_MSG | MACH_RCV_LARGE,
               sizeof(buf.msg), sizeof(buf), GC_ports.reply,
               MACH_MSG_TIMEOUT_NONE, MACH_PORT_NULL);
  if(r != MACH_MSG_SUCCESS)
    ABORT("mach_msg failed in GC_mprotect_thread_notify");
  if(buf.msg.head.msgh_id != ID_ACK)
    ABORT("invalid ack in GC_mprotect_thread_notify");
}

/* Should only be called by the mprotect thread */
STATIC void GC_mprotect_thread_reply(void)
{
  GC_msg_t msg;
  mach_msg_return_t r;
  /* remote, local */
  msg.head.msgh_bits = MACH_MSGH_BITS(MACH_MSG_TYPE_MAKE_SEND, 0);
  msg.head.msgh_size = sizeof(msg);
  msg.head.msgh_remote_port = GC_ports.reply;
  msg.head.msgh_local_port = MACH_PORT_NULL;
  msg.head.msgh_id = ID_ACK;

  r = mach_msg(&msg.head, MACH_SEND_MSG, sizeof(msg), 0, MACH_PORT_NULL,
               MACH_MSG_TIMEOUT_NONE, MACH_PORT_NULL);
  if(r != MACH_MSG_SUCCESS)
    ABORT("mach_msg failed in GC_mprotect_thread_reply");
}

void GC_mprotect_stop(void)
{
  GC_mprotect_thread_notify(ID_STOP);
}
void GC_mprotect_resume(void)
{
  GC_mprotect_thread_notify(ID_RESUME);
}

#else /* !THREADS */
/* The compiler should optimize away any GC_mprotect_state computations */
#define GC_mprotect_state GC_MP_NORMAL
#endif

STATIC void *GC_mprotect_thread(void *arg)
{
  mach_msg_return_t r;
  /* These two structures contain some private kernel data. We don't need to
     access any of it so we don't bother defining a proper struct. The
     correct definitions are in the xnu source code. */
  struct {
    mach_msg_header_t head;
    char data[256];
  } reply;
  struct {
    mach_msg_header_t head;
    mach_msg_body_t msgh_body;
    char data[1024];
  } msg;

  mach_msg_id_t id;

  GC_darwin_register_mach_handler_thread(mach_thread_self());

  for(;;) {
    r = mach_msg(&msg.head, MACH_RCV_MSG | MACH_RCV_LARGE |
                 (GC_mprotect_state == GC_MP_DISCARDING ? MACH_RCV_TIMEOUT : 0),
                 0, sizeof(msg), GC_ports.exception,
                 GC_mprotect_state == GC_MP_DISCARDING ? 0
                 : MACH_MSG_TIMEOUT_NONE, MACH_PORT_NULL);

    id = r == MACH_MSG_SUCCESS ? msg.head.msgh_id : -1;

#   if defined(THREADS)
      if(GC_mprotect_state == GC_MP_DISCARDING) {
        if(r == MACH_RCV_TIMED_OUT) {
          GC_mprotect_state = GC_MP_STOPPED;
          GC_mprotect_thread_reply();
          continue;
        }
        if(r == MACH_MSG_SUCCESS && (id == ID_STOP || id == ID_RESUME))
          ABORT("out of order mprotect thread request");
      }
#   endif /* THREADS */

    if(r != MACH_MSG_SUCCESS) {
      GC_err_printf("mach_msg failed with %d %s\n", (int)r,
                    mach_error_string(r));
      ABORT("mach_msg failed");
    }

    switch(id) {
#     if defined(THREADS)
        case ID_STOP:
          if(GC_mprotect_state != GC_MP_NORMAL)
            ABORT("Called mprotect_stop when state wasn't normal");
          GC_mprotect_state = GC_MP_DISCARDING;
          break;
        case ID_RESUME:
          if(GC_mprotect_state != GC_MP_STOPPED)
            ABORT("Called mprotect_resume when state wasn't stopped");
          GC_mprotect_state = GC_MP_NORMAL;
          GC_mprotect_thread_reply();
          break;
#     endif /* THREADS */
        default:
          /* Handle the message (calls catch_exception_raise) */
          if(!exc_server(&msg.head, &reply.head))
            ABORT("exc_server failed");
          /* Send the reply */
          r = mach_msg(&reply.head, MACH_SEND_MSG, reply.head.msgh_size, 0,
                       MACH_PORT_NULL, MACH_MSG_TIMEOUT_NONE,
                       MACH_PORT_NULL);
          if(r != MACH_MSG_SUCCESS) {
            /* This will fail if the thread dies, but the thread */
            /* shouldn't die... */
#           ifdef BROKEN_EXCEPTION_HANDLING
              GC_err_printf("mach_msg failed with %d %s while sending "
                            "exc reply\n", (int)r,mach_error_string(r));
#           else
              ABORT("mach_msg failed while sending exception reply");
#           endif
          }
    } /* switch */
  } /* for(;;) */
    /* NOT REACHED */
  return NULL;
}

/* All this SIGBUS code shouldn't be necessary. All protection faults should
   be going through the mach exception handler. However, it seems a SIGBUS is
   occasionally sent for some unknown reason. Even more odd, it seems to be
   meaningless and safe to ignore. */
#ifdef BROKEN_EXCEPTION_HANDLING

/* Updates to this aren't atomic, but the SIGBUSs seem pretty rare.
   Even if this doesn't get updated property, it isn't really a problem */
STATIC int GC_sigbus_count = 0;

STATIC void GC_darwin_sigbus(int num, siginfo_t *sip, void *context)
{
  if(num != SIGBUS)
    ABORT("Got a non-sigbus signal in the sigbus handler");

  /* Ugh... some seem safe to ignore, but too many in a row probably means
     trouble. GC_sigbus_count is reset for each mach exception that is
     handled */
  if(GC_sigbus_count >= 8) {
    ABORT("Got more than 8 SIGBUSs in a row!");
  } else {
    GC_sigbus_count++;
    WARN("Ignoring SIGBUS.\n", 0);
  }
}
#endif /* BROKEN_EXCEPTION_HANDLING */

void GC_dirty_init(void)
{
  kern_return_t r;
  mach_port_t me;
  pthread_t thread;
  pthread_attr_t attr;
  exception_mask_t mask;

  if (GC_print_stats == VERBOSE)
    GC_log_printf("Initializing mach/darwin mprotect virtual dirty bit "
                  "implementation\n");
# ifdef BROKEN_EXCEPTION_HANDLING
    WARN("Enabling workarounds for various darwin "
         "exception handling bugs.\n", 0);
# endif
  GC_dirty_maintained = TRUE;
  if (GC_page_size % HBLKSIZE != 0) {
    GC_err_printf("Page size not multiple of HBLKSIZE\n");
    ABORT("Page size not multiple of HBLKSIZE");
  }

  GC_task_self = me = mach_task_self();

  r = mach_port_allocate(me, MACH_PORT_RIGHT_RECEIVE, &GC_ports.exception);
  if(r != KERN_SUCCESS)
    ABORT("mach_port_allocate failed (exception port)");

  r = mach_port_insert_right(me, GC_ports.exception, GC_ports.exception,
                             MACH_MSG_TYPE_MAKE_SEND);
  if(r != KERN_SUCCESS)
    ABORT("mach_port_insert_right failed (exception port)");

#  if defined(THREADS)
     r = mach_port_allocate(me, MACH_PORT_RIGHT_RECEIVE, &GC_ports.reply);
     if(r != KERN_SUCCESS)
       ABORT("mach_port_allocate failed (reply port)");
#  endif

  /* The exceptions we want to catch */
  mask = EXC_MASK_BAD_ACCESS;

  r = task_get_exception_ports(me, mask, GC_old_exc_ports.masks,
                               &GC_old_exc_ports.count, GC_old_exc_ports.ports,
                               GC_old_exc_ports.behaviors,
                               GC_old_exc_ports.flavors);
  if(r != KERN_SUCCESS)
    ABORT("task_get_exception_ports failed");

  r = task_set_exception_ports(me, mask, GC_ports.exception, EXCEPTION_DEFAULT,
                               GC_MACH_THREAD_STATE);
  if(r != KERN_SUCCESS)
    ABORT("task_set_exception_ports failed");
  if(pthread_attr_init(&attr) != 0)
    ABORT("pthread_attr_init failed");
  if(pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED) != 0)
    ABORT("pthread_attr_setdetachedstate failed");

# undef pthread_create
  /* This will call the real pthread function, not our wrapper */
  if(pthread_create(&thread, &attr, GC_mprotect_thread, NULL) != 0)
    ABORT("pthread_create failed");
  pthread_attr_destroy(&attr);

  /* Setup the sigbus handler for ignoring the meaningless SIGBUSs */
# ifdef BROKEN_EXCEPTION_HANDLING
    {
      struct sigaction sa, oldsa;
      sa.sa_handler = (SIG_HNDLR_PTR)GC_darwin_sigbus;
      sigemptyset(&sa.sa_mask);
      sa.sa_flags = SA_RESTART|SA_SIGINFO;
      if(sigaction(SIGBUS, &sa, &oldsa) < 0)
        ABORT("sigaction");
      if ((SIG_HNDLR_PTR)oldsa.sa_handler != SIG_DFL) {
        if (GC_print_stats == VERBOSE)
          GC_err_printf("Replaced other SIGBUS handler\n");
      }
    }
#  endif /* BROKEN_EXCEPTION_HANDLING  */
}

/* The source code for Apple's GDB was used as a reference for the exception
   forwarding code. This code is similar to be GDB code only because there is
   only one way to do it. */
STATIC kern_return_t GC_forward_exception(mach_port_t thread, mach_port_t task,
                                          exception_type_t exception,
                                          exception_data_t data,
                                          mach_msg_type_number_t data_count)
{
  unsigned int i;
  kern_return_t r;
  mach_port_t port;
  exception_behavior_t behavior;
  thread_state_flavor_t flavor;

  thread_state_data_t thread_state;
  mach_msg_type_number_t thread_state_count = THREAD_STATE_MAX;

  for(i=0; i < GC_old_exc_ports.count; i++)
    if(GC_old_exc_ports.masks[i] & (1 << exception))
      break;
  if(i==GC_old_exc_ports.count)
    ABORT("No handler for exception!");

  port = GC_old_exc_ports.ports[i];
  behavior = GC_old_exc_ports.behaviors[i];
  flavor = GC_old_exc_ports.flavors[i];

  if(behavior == EXCEPTION_STATE || behavior == EXCEPTION_STATE_IDENTITY) {
    r = thread_get_state(thread, flavor, thread_state, &thread_state_count);
    if(r != KERN_SUCCESS)
      ABORT("thread_get_state failed in forward_exception");
    }

  switch(behavior) {
    case EXCEPTION_STATE:
      r = exception_raise_state(port, thread, task, exception, data, data_count,
                                &flavor, thread_state, thread_state_count,
                                thread_state, &thread_state_count);
      break;
    case EXCEPTION_STATE_IDENTITY:
      r = exception_raise_state_identity(port, thread, task, exception, data,
                                         data_count, &flavor, thread_state,
                                         thread_state_count, thread_state,
                                         &thread_state_count);
      break;
    /* case EXCEPTION_DEFAULT: */ /* default signal handlers */
    default: /* user-supplied signal handlers */
      r = exception_raise(port, thread, task, exception, data, data_count);
  }

  if(behavior == EXCEPTION_STATE || behavior == EXCEPTION_STATE_IDENTITY) {
    r = thread_set_state(thread, flavor, thread_state, thread_state_count);
    if(r != KERN_SUCCESS)
      ABORT("thread_set_state failed in forward_exception");
  }

  return r;
}

#define FWD() GC_forward_exception(thread, task, exception, code, code_count)

/* This violates the namespace rules but there isn't anything that can be done
   about it.  The exception handling stuff is hard coded to call this. */
kern_return_t
catch_exception_raise(mach_port_t exception_port, mach_port_t thread,
                      mach_port_t task, exception_type_t exception,
                      exception_data_t code, mach_msg_type_number_t code_count)
{
  kern_return_t r;
  char *addr;
  struct hblk *h;
  unsigned int i;
# if defined(POWERPC)
#   if CPP_WORDSZ == 32
      thread_state_flavor_t flavor = PPC_EXCEPTION_STATE;
      mach_msg_type_number_t exc_state_count = PPC_EXCEPTION_STATE_COUNT;
      ppc_exception_state_t exc_state;
#   else
      thread_state_flavor_t flavor = PPC_EXCEPTION_STATE64;
      mach_msg_type_number_t exc_state_count = PPC_EXCEPTION_STATE64_COUNT;
      ppc_exception_state64_t exc_state;
#   endif
# elif defined(I386) || defined(X86_64)
#   if CPP_WORDSZ == 32
      thread_state_flavor_t flavor = x86_EXCEPTION_STATE32;
      mach_msg_type_number_t exc_state_count = x86_EXCEPTION_STATE32_COUNT;
      x86_exception_state32_t exc_state;
#   else
      thread_state_flavor_t flavor = x86_EXCEPTION_STATE64;
      mach_msg_type_number_t exc_state_count = x86_EXCEPTION_STATE64_COUNT;
      x86_exception_state64_t exc_state;
#   endif
# else
#   error FIXME for non-ppc/x86 darwin
# endif


  if(exception != EXC_BAD_ACCESS || code[0] != KERN_PROTECTION_FAILURE) {
#   ifdef DEBUG_EXCEPTION_HANDLING
      /* We aren't interested, pass it on to the old handler */
      GC_printf("Exception: 0x%x Code: 0x%x 0x%x in catch....\n", exception,
                code_count > 0 ? code[0] : -1, code_count > 1 ? code[1] : -1);
#   endif
    return FWD();
  }

  r = thread_get_state(thread, flavor, (natural_t*)&exc_state,
                       &exc_state_count);
  if(r != KERN_SUCCESS) {
    /* The thread is supposed to be suspended while the exception handler
       is called. This shouldn't fail. */
#   ifdef BROKEN_EXCEPTION_HANDLING
      GC_err_printf("thread_get_state failed in catch_exception_raise\n");
      return KERN_SUCCESS;
#   else
      ABORT("thread_get_state failed in catch_exception_raise");
#   endif
  }

    /* This is the address that caused the fault */
# if defined(POWERPC)
    addr = (char*) exc_state. THREAD_FLD(dar);
# elif defined (I386) || defined (X86_64)
    addr = (char*) exc_state. THREAD_FLD(faultvaddr);
# else
#   error FIXME for non POWERPC/I386
# endif

    if((HDR(addr)) == 0) {
      /* Ugh... just like the SIGBUS problem above, it seems we get a bogus
         KERN_PROTECTION_FAILURE every once and a while. We wait till we get
         a bunch in a row before doing anything about it. If a "real" fault
         ever occurs it'll just keep faulting over and over and we'll hit
         the limit pretty quickly. */
#     ifdef BROKEN_EXCEPTION_HANDLING
        static char *last_fault;
        static int last_fault_count;

        if(addr != last_fault) {
          last_fault = addr;
          last_fault_count = 0;
        }
        if(++last_fault_count < 32) {
          if(last_fault_count == 1)
            WARN("Ignoring KERN_PROTECTION_FAILURE at %p\n", addr);
          return KERN_SUCCESS;
        }

        GC_err_printf("Unexpected KERN_PROTECTION_FAILURE at %p\n",addr);
        /* Can't pass it along to the signal handler because that is
           ignoring SIGBUS signals. We also shouldn't call ABORT here as
           signals don't always work too well from the exception handler. */
        GC_err_printf("Aborting\n");
        exit(EXIT_FAILURE);
#     else /* BROKEN_EXCEPTION_HANDLING */
        /* Pass it along to the next exception handler
           (which should call SIGBUS/SIGSEGV) */
        return FWD();
#     endif /* !BROKEN_EXCEPTION_HANDLING */
    }

#   ifdef BROKEN_EXCEPTION_HANDLING
      /* Reset the number of consecutive SIGBUSs */
      GC_sigbus_count = 0;
#   endif

    if(GC_mprotect_state == GC_MP_NORMAL) { /* common case */
      h = (struct hblk*)((word)addr & ~(GC_page_size-1));
      UNPROTECT(h, GC_page_size);
      for (i = 0; i < divHBLKSZ(GC_page_size); i++) {
        register int index = PHT_HASH(h+i);
        async_set_pht_entry_from_index(GC_dirty_pages, index);
      }
    } else if(GC_mprotect_state == GC_MP_DISCARDING) {
      /* Lie to the thread for now. No sense UNPROTECT()ing the memory
         when we're just going to PROTECT() it again later. The thread
         will just fault again once it resumes */
    } else {
      /* Shouldn't happen, i don't think */
      GC_printf("KERN_PROTECTION_FAILURE while world is stopped\n");
      return FWD();
    }
    return KERN_SUCCESS;
}
#undef FWD

/* These should never be called, but just in case...  */
kern_return_t
catch_exception_raise_state(mach_port_name_t exception_port, int exception,
                            exception_data_t code,
                            mach_msg_type_number_t codeCnt, int flavor,
                            thread_state_t old_state, int old_stateCnt,
                            thread_state_t new_state, int new_stateCnt)
{
  ABORT("catch_exception_raise_state");
  return(KERN_INVALID_ARGUMENT);
}

kern_return_t
catch_exception_raise_state_identity(mach_port_name_t exception_port,
                                     mach_port_t thread, mach_port_t task,
                                     int exception, exception_data_t code,
                                     mach_msg_type_number_t codeCnt, int flavor,
                                     thread_state_t old_state, int old_stateCnt,
                                     thread_state_t new_state, int new_stateCnt)
{
  ABORT("catch_exception_raise_state_identity");
  return(KERN_INVALID_ARGUMENT);
}


#endif /* DARWIN && MPROTECT_VDB */

# ifndef HAVE_INCREMENTAL_PROTECTION_NEEDS
  GC_API int GC_CALL GC_incremental_protection_needs(void)
  {
    return GC_PROTECTS_NONE;
  }
# endif /* !HAVE_INCREMENTAL_PROTECTION_NEEDS */

/*
 * Call stack save code for debugging.
 * Should probably be in mach_dep.c, but that requires reorganization.
 */

/* I suspect the following works for most X86 *nix variants, so         */
/* long as the frame pointer is explicitly stored.  In the case of gcc, */
/* compiler flags (e.g. -fomit-frame-pointer) determine whether it is.  */
#if defined(I386) && defined(LINUX) && defined(SAVE_CALL_CHAIN)
#   include <features.h>

    struct frame {
        struct frame *fr_savfp;
        long    fr_savpc;
        long    fr_arg[NARGS];  /* All the arguments go here.   */
    };
#endif

#if defined(SPARC)
#  if defined(LINUX)
#    include <features.h>

     struct frame {
        long    fr_local[8];
        long    fr_arg[6];
        struct frame *fr_savfp;
        long    fr_savpc;
#       ifndef __arch64__
          char  *fr_stret;
#       endif
        long    fr_argd[6];
        long    fr_argx[0];
     };
#  elif defined (DRSNX)
#    include <sys/sparc/frame.h>
#  elif defined(OPENBSD)
#    include <frame.h>
#  elif defined(FREEBSD) || defined(NETBSD)
#    include <machine/frame.h>
#  else
#    include <sys/frame.h>
#  endif
#  if NARGS > 6
#    error We only know how to to get the first 6 arguments
#  endif
#endif /* SPARC */

#ifdef  NEED_CALLINFO
/* Fill in the pc and argument information for up to NFRAMES of my      */
/* callers.  Ignore my frame and my callers frame.                      */

#ifdef LINUX
#   include <unistd.h>
#endif

#endif /* NEED_CALLINFO */

#if defined(GC_HAVE_BUILTIN_BACKTRACE)
# ifdef _MSC_VER
#  include "private/msvc_dbg.h"
# else
#  include <execinfo.h>
# endif
#endif

#ifdef SAVE_CALL_CHAIN

#if NARGS == 0 && NFRAMES % 2 == 0 /* No padding */ \
    && defined(GC_HAVE_BUILTIN_BACKTRACE)

#ifdef REDIRECT_MALLOC
  /* Deal with possible malloc calls in backtrace by omitting   */
  /* the infinitely recursing backtrace.                        */
# ifdef THREADS
    __thread    /* If your compiler doesn't understand this */
                /* you could use something like pthread_getspecific.    */
# endif
  GC_in_save_callers = FALSE;
#endif

void GC_save_callers (struct callinfo info[NFRAMES])
{
  void * tmp_info[NFRAMES + 1];
  int npcs, i;
# define IGNORE_FRAMES 1

  /* We retrieve NFRAMES+1 pc values, but discard the first, since it   */
  /* points to our own frame.                                           */
# ifdef REDIRECT_MALLOC
    if (GC_in_save_callers) {
      info[0].ci_pc = (word)(&GC_save_callers);
      for (i = 1; i < NFRAMES; ++i) info[i].ci_pc = 0;
      return;
    }
    GC_in_save_callers = TRUE;
# endif
  GC_STATIC_ASSERT(sizeof(struct callinfo) == sizeof(void *));
  npcs = backtrace((void **)tmp_info, NFRAMES + IGNORE_FRAMES);
  BCOPY(tmp_info+IGNORE_FRAMES, info, (npcs - IGNORE_FRAMES) * sizeof(void *));
  for (i = npcs - IGNORE_FRAMES; i < NFRAMES; ++i) info[i].ci_pc = 0;
# ifdef REDIRECT_MALLOC
    GC_in_save_callers = FALSE;
# endif
}

#else /* No builtin backtrace; do it ourselves */

#if (defined(OPENBSD) || defined(NETBSD) || defined(FREEBSD)) && defined(SPARC)
#  define FR_SAVFP fr_fp
#  define FR_SAVPC fr_pc
#else
#  define FR_SAVFP fr_savfp
#  define FR_SAVPC fr_savpc
#endif

#if defined(SPARC) && (defined(__arch64__) || defined(__sparcv9))
#   define BIAS 2047
#else
#   define BIAS 0
#endif

void GC_save_callers (struct callinfo info[NFRAMES])
{
  struct frame *frame;
  struct frame *fp;
  int nframes = 0;
# ifdef I386
    /* We assume this is turned on only with gcc as the compiler. */
    asm("movl %%ebp,%0" : "=r"(frame));
    fp = frame;
# else
    frame = (struct frame *) GC_save_regs_in_stack ();
    fp = (struct frame *)((long) frame -> FR_SAVFP + BIAS);
#endif

   for (; (!(fp HOTTER_THAN frame) && !(GC_stackbottom HOTTER_THAN (ptr_t)fp)
           && (nframes < NFRAMES));
       fp = (struct frame *)((long) fp -> FR_SAVFP + BIAS), nframes++) {
      register int i;

      info[nframes].ci_pc = fp->FR_SAVPC;
#     if NARGS > 0
        for (i = 0; i < NARGS; i++) {
          info[nframes].ci_arg[i] = ~(fp->fr_arg[i]);
        }
#     endif /* NARGS > 0 */
  }
  if (nframes < NFRAMES) info[nframes].ci_pc = 0;
}

#endif /* No builtin backtrace */

#endif /* SAVE_CALL_CHAIN */

#ifdef NEED_CALLINFO

/* Print info to stderr.  We do NOT hold the allocation lock */
void GC_print_callers (struct callinfo info[NFRAMES])
{
    int i;
    static int reentry_count = 0;
    GC_bool stop = FALSE;

    /* FIXME: This should probably use a different lock, so that we     */
    /* become callable with or without the allocation lock.             */
    LOCK();
      ++reentry_count;
    UNLOCK();

#   if NFRAMES == 1
      GC_err_printf("\tCaller at allocation:\n");
#   else
      GC_err_printf("\tCall chain at allocation:\n");
#   endif
    for (i = 0; i < NFRAMES && !stop ; i++) {
        if (info[i].ci_pc == 0) break;
#       if NARGS > 0
        {
          int j;

          GC_err_printf("\t\targs: ");
          for (j = 0; j < NARGS; j++) {
            if (j != 0) GC_err_printf(", ");
            GC_err_printf("%d (0x%X)", ~(info[i].ci_arg[j]),
                                        ~(info[i].ci_arg[j]));
          }
          GC_err_printf("\n");
        }
#       endif
        if (reentry_count > 1) {
            /* We were called during an allocation during       */
            /* a previous GC_print_callers call; punt.          */
            GC_err_printf("\t\t##PC##= 0x%lx\n", info[i].ci_pc);
            continue;
        }
        {
#         ifdef LINUX
            FILE *pipe;
#         endif
#         if defined(GC_HAVE_BUILTIN_BACKTRACE) \
             && !defined(GC_BACKTRACE_SYMBOLS_BROKEN)
            char **sym_name =
              backtrace_symbols((void **)(&(info[i].ci_pc)), 1);
            char *name = sym_name[0];
#         else
            char buf[40];
            char *name = buf;
            sprintf(buf, "##PC##= 0x%lx", info[i].ci_pc);
#         endif
#         if defined(LINUX) && !defined(SMALL_CONFIG)
            /* Try for a line number. */
            {
#               define EXE_SZ 100
                static char exe_name[EXE_SZ];
#               define CMD_SZ 200
                char cmd_buf[CMD_SZ];
#               define RESULT_SZ 200
                static char result_buf[RESULT_SZ];
                size_t result_len;
                char *old_preload;
#               define PRELOAD_SZ 200
                char preload_buf[PRELOAD_SZ];
                static GC_bool found_exe_name = FALSE;
                static GC_bool will_fail = FALSE;
                int ret_code;
                /* Try to get it via a hairy and expensive scheme.      */
                /* First we get the name of the executable:             */
                if (will_fail) goto out;
                if (!found_exe_name) {
                  ret_code = readlink("/proc/self/exe", exe_name, EXE_SZ);
                  if (ret_code < 0 || ret_code >= EXE_SZ
                      || exe_name[0] != '/') {
                    will_fail = TRUE;   /* Dont try again. */
                    goto out;
                  }
                  exe_name[ret_code] = '\0';
                  found_exe_name = TRUE;
                }
                /* Then we use popen to start addr2line -e <exe> <addr> */
                /* There are faster ways to do this, but hopefully this */
                /* isn't time critical.                                 */
                sprintf(cmd_buf, "/usr/bin/addr2line -f -e %s 0x%lx", exe_name,
                                 (unsigned long)info[i].ci_pc);
                old_preload = GETENV("LD_PRELOAD");
                if (0 != old_preload) {
                  if (strlen (old_preload) >= PRELOAD_SZ) {
                    will_fail = TRUE;
                    goto out;
                  }
                  strcpy (preload_buf, old_preload);
                  unsetenv ("LD_PRELOAD");
                }
                pipe = popen(cmd_buf, "r");
                if (0 != old_preload
                    && 0 != setenv ("LD_PRELOAD", preload_buf, 0)) {
                  WARN("Failed to reset LD_PRELOAD\n", 0);
                }
                if (pipe == NULL
                    || (result_len = fread(result_buf, 1, RESULT_SZ - 1, pipe))
                       == 0) {
                  if (pipe != NULL) pclose(pipe);
                  will_fail = TRUE;
                  goto out;
                }
                if (result_buf[result_len - 1] == '\n') --result_len;
                result_buf[result_len] = 0;
                if (result_buf[0] == '?'
                    || (result_buf[result_len-2] == ':'
                        && result_buf[result_len-1] == '0')) {
                    pclose(pipe);
                    goto out;
                }
                /* Get rid of embedded newline, if any.  Test for "main" */
                {
                   char * nl = strchr(result_buf, '\n');
                   if (nl != NULL && nl < result_buf + result_len) {
                     *nl = ':';
                   }
                   if (strncmp(result_buf, "main", nl - result_buf) == 0) {
                     stop = TRUE;
                   }
                }
                if (result_len < RESULT_SZ - 25) {
                  /* Add in hex address */
                    sprintf(result_buf + result_len, " [0x%lx]",
                          (unsigned long)info[i].ci_pc);
                }
                name = result_buf;
                pclose(pipe);
                out:;
            }
#         endif /* LINUX */
          GC_err_printf("\t\t%s\n", name);
#         if defined(GC_HAVE_BUILTIN_BACKTRACE) \
             && !defined(GC_BACKTRACE_SYMBOLS_BROKEN)
            free(sym_name);  /* May call GC_free; that's OK */
#         endif
        }
    }
    LOCK();
      --reentry_count;
    UNLOCK();
}

#endif /* NEED_CALLINFO */

#if defined(LINUX) && defined(__ELF__) && !defined(SMALL_CONFIG)

/* Dump /proc/self/maps to GC_stderr, to enable looking up names for
   addresses in FIND_LEAK output. */

static word dump_maps(char *maps)
{
    GC_err_write(maps, strlen(maps));
    return 1;
}

void GC_print_address_map(void)
{
    GC_err_printf("---------- Begin address map ----------\n");
    dump_maps(GC_get_maps());
    GC_err_printf("---------- End address map ----------\n");
}

#endif