/*
* Copyright (c) 1991-1994 by Xerox Corporation.  All rights reserved.
* Copyright (c) 1997 by Silicon Graphics.  All rights reserved.
* Copyright (c) 2009-2021 Ivan Maidanski
*
* 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"

/*
* This is incredibly OS specific code for tracking down data sections in
* dynamic libraries.  There appears to be no way of doing this quickly
* without groveling through undocumented data structures.  We would argue
* that this is a bug in the design of the dlopen interface.  THIS CODE
* MAY BREAK IN FUTURE OS RELEASES.  If this matters to you, don't hesitate
* to let your vendor know ...
*
* None of this is safe with dlclose and incremental collection.
* But then not much of anything is safe in the presence of dlclose.
*/

#if !defined(MACOS) && !defined(GC_NO_TYPES) && !defined(SN_TARGET_PSP2) \
   && !defined(_WIN32_WCE) && !defined(__CC_ARM)
# include <sys/types.h>
#endif

/* BTL: avoid circular redefinition of dlopen if GC_SOLARIS_THREADS defined */
#undef GC_MUST_RESTORE_REDEFINED_DLOPEN
#if defined(GC_PTHREADS) && !defined(GC_NO_DLOPEN) \
   && !defined(GC_NO_THREAD_REDIRECTS) && !defined(GC_USE_LD_WRAP)
 /* To support threads in Solaris, gc.h interposes on dlopen by        */
 /* defining "dlopen" to be "GC_dlopen", which is implemented below.   */
 /* However, both GC_FirstDLOpenedLinkMap() and GC_dlopen() use the    */
 /* real system dlopen() in their implementation. We first remove      */
 /* gc.h's dlopen definition and restore it later, after GC_dlopen().  */
# undef dlopen
# define GC_MUST_RESTORE_REDEFINED_DLOPEN
#endif /* !GC_NO_DLOPEN */

/* A user-supplied routine (custom filter) that might be called to      */
/* determine whether a DSO really needs to be scanned by the GC.        */
/* 0 means no filter installed.  May be unused on some platforms.       */
/* FIXME: Add filter support for more platforms.                        */
STATIC GC_has_static_roots_func GC_has_static_roots = 0;

#if (defined(DYNAMIC_LOADING) || defined(MSWIN32) || defined(MSWINCE) \
   || defined(CYGWIN32)) && !defined(PCR)

#if !defined(DARWIN) && !defined(SCO_ELF) && !defined(SOLARISDL) \
   && !defined(AIX) && !defined(DGUX) && !defined(IRIX5) && !defined(HPUX) \
   && !defined(CYGWIN32) && !defined(MSWIN32) && !defined(MSWINCE) \
   && !(defined(ALPHA) && defined(OSF1)) \
   && !(defined(FREEBSD) && defined(__ELF__)) \
   && !(defined(LINUX) && defined(__ELF__)) \
   && !(defined(NETBSD) && defined(__ELF__)) \
   && !(defined(OPENBSD) && (defined(__ELF__) || defined(M68K))) \
   && !defined(HAIKU) && !defined(HURD) && !defined(NACL) \
   && !defined(CPPCHECK)
# error We only know how to find data segments of dynamic libraries for above.
# error Additional SVR4 variants might not be too hard to add.
#endif

#include <stdio.h>
#ifdef SOLARISDL
#   include <sys/elf.h>
#   include <dlfcn.h>
#   include <link.h>
#endif

#if defined(NETBSD)
#   include <sys/param.h>
#   include <dlfcn.h>
#   include <machine/elf_machdep.h>
#   define ELFSIZE ARCH_ELFSIZE
#endif

#if defined(OPENBSD)
# include <sys/param.h>
# if (OpenBSD >= 200519) && !defined(HAVE_DL_ITERATE_PHDR)
#   define HAVE_DL_ITERATE_PHDR
# endif
#endif /* OPENBSD */

#if defined(SCO_ELF) || defined(DGUX) || defined(HURD) || defined(NACL) \
   || (defined(__ELF__) && (defined(LINUX) || defined(FREEBSD) \
                            || defined(NETBSD) || defined(OPENBSD)))
# include <stddef.h>
# if !defined(OPENBSD) && !defined(HOST_ANDROID)
   /* OpenBSD does not have elf.h file; link.h below is sufficient.    */
   /* Exclude Android because linker.h below includes its own version. */
#   include <elf.h>
# endif
# ifdef HOST_ANDROID
   /* If you don't need the "dynamic loading" feature, you may build   */
   /* the collector with -D IGNORE_DYNAMIC_LOADING.                    */
#   ifdef BIONIC_ELFDATA_REDEF_BUG
     /* Workaround a problem in Bionic (as of Android 4.2) which has   */
     /* mismatching ELF_DATA definitions in sys/exec_elf.h and         */
     /* asm/elf.h included from linker.h file (similar to EM_ALPHA).   */
#     include <asm/elf.h>
#     include <linux/elf-em.h>
#     undef ELF_DATA
#     undef EM_ALPHA
#   endif
#   include <link.h>
#   if !defined(GC_DONT_DEFINE_LINK_MAP) && !(__ANDROID_API__ >= 21)
     /* link_map and r_debug are defined in link.h of NDK r10+.        */
     /* bionic/linker/linker.h defines them too but the header         */
     /* itself is a C++ one starting from Android 4.3.                 */
     struct link_map {
       uintptr_t l_addr;
       char* l_name;
       uintptr_t l_ld;
       struct link_map* l_next;
       struct link_map* l_prev;
     };
     struct r_debug {
       int32_t r_version;
       struct link_map* r_map;
       void (*r_brk)(void);
       int32_t r_state;
       uintptr_t r_ldbase;
     };
#   endif
# else
   EXTERN_C_BEGIN      /* Workaround missing extern "C" around _DYNAMIC */
                       /* symbol in link.h of some Linux hosts.         */
#   include <link.h>
   EXTERN_C_END
# endif
#endif

/* Newer versions of GNU/Linux define this macro.  We
* define it similarly for any ELF systems that don't.  */
#  ifndef ElfW
#    if defined(FREEBSD)
#      if __ELF_WORD_SIZE == 32
#        define ElfW(type) Elf32_##type
#      else
#        define ElfW(type) Elf64_##type
#      endif
#    elif defined(NETBSD) || defined(OPENBSD)
#      if ELFSIZE == 32
#        define ElfW(type) Elf32_##type
#      elif ELFSIZE == 64
#        define ElfW(type) Elf64_##type
#      else
#        error Missing ELFSIZE define
#      endif
#    else
#      if !defined(ELF_CLASS) || ELF_CLASS == ELFCLASS32
#        define ElfW(type) Elf32_##type
#      else
#        define ElfW(type) Elf64_##type
#      endif
#    endif
#  endif

#if defined(SOLARISDL) && !defined(USE_PROC_FOR_LIBRARIES)

 EXTERN_C_BEGIN
 extern ElfW(Dyn) _DYNAMIC;
 EXTERN_C_END

 STATIC struct link_map *
 GC_FirstDLOpenedLinkMap(void)
 {
   ElfW(Dyn) *dp;
   static struct link_map * cachedResult = 0;
   static ElfW(Dyn) *dynStructureAddr = 0;
               /* BTL: added to avoid Solaris 5.3 ld.so _DYNAMIC bug   */

#   ifdef SUNOS53_SHARED_LIB
       /* BTL: Avoid the Solaris 5.3 bug that _DYNAMIC isn't being set */
       /* up properly in dynamically linked .so's. This means we have  */
       /* to use its value in the set of original object files loaded  */
       /* at program startup.                                          */
       if( dynStructureAddr == 0 ) {
         void* startupSyms = dlopen(0, RTLD_LAZY);
         dynStructureAddr = (ElfW(Dyn)*)(word)dlsym(startupSyms, "_DYNAMIC");
       }
#   else
       dynStructureAddr = &_DYNAMIC;
#   endif

   if (0 == COVERT_DATAFLOW(dynStructureAddr)) {
       /* _DYNAMIC symbol not resolved. */
       return(0);
   }
   if (cachedResult == 0) {
       int tag;
       for( dp = ((ElfW(Dyn) *)(&_DYNAMIC)); (tag = dp->d_tag) != 0; dp++ ) {
           if (tag == DT_DEBUG) {
               struct r_debug *rd = (struct r_debug *)dp->d_un.d_ptr;
               if (rd != NULL) {
                   struct link_map *lm = rd->r_map;
                   if (lm != NULL)
                       cachedResult = lm->l_next; /* might be NULL */
               }
               break;
           }
       }
   }
   return cachedResult;
 }

#endif /* SOLARISDL ... */

/* BTL: added to fix circular dlopen definition if GC_SOLARIS_THREADS defined */
# ifdef GC_MUST_RESTORE_REDEFINED_DLOPEN
#   define dlopen GC_dlopen
# endif

# if defined(SOLARISDL)

/* Add dynamic library data sections to the root set.           */
# if !defined(PCR) && !defined(GC_SOLARIS_THREADS) && defined(THREADS) \
    && !defined(CPPCHECK)
#   error Fix mutual exclusion with dlopen
# endif

# ifndef USE_PROC_FOR_LIBRARIES
GC_INNER void GC_register_dynamic_libraries(void)
{
 struct link_map *lm;

 for (lm = GC_FirstDLOpenedLinkMap(); lm != 0; lm = lm->l_next) {
       ElfW(Ehdr) * e;
       ElfW(Phdr) * p;
       unsigned long offset;
       char * start;
       int i;

       e = (ElfW(Ehdr) *) lm->l_addr;
       p = ((ElfW(Phdr) *)(((char *)(e)) + e->e_phoff));
       offset = ((unsigned long)(lm->l_addr));
       for( i = 0; i < (int)e->e_phnum; i++, p++ ) {
         switch( p->p_type ) {
           case PT_LOAD:
             {
               if( !(p->p_flags & PF_W) ) break;
               start = ((char *)(p->p_vaddr)) + offset;
               GC_add_roots_inner(start, start + p->p_memsz, TRUE);
             }
             break;
           default:
             break;
         }
       }
   }
}

# endif /* !USE_PROC_FOR_LIBRARIES */
# endif /* SOLARISDL */

#if defined(SCO_ELF) || defined(DGUX) || defined(HURD) || defined(NACL) \
   || (defined(__ELF__) && (defined(LINUX) || defined(FREEBSD) \
                            || defined(NETBSD) || defined(OPENBSD)))

#ifdef USE_PROC_FOR_LIBRARIES

#include <string.h>

#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>

#define MAPS_BUF_SIZE (32*1024)

/* Sort an array of HeapSects by start address.                         */
/* Unfortunately at least some versions of                              */
/* Linux qsort end up calling malloc by way of sysconf, and hence can't */
/* be used in the collector.  Hence we roll our own.  Should be         */
/* reasonably fast if the array is already mostly sorted, as we expect  */
/* it to be.                                                            */
static void sort_heap_sects(struct HeapSect *base, size_t number_of_elements)
{
   signed_word n = (signed_word)number_of_elements;
   signed_word nsorted = 1;

   while (nsorted < n) {
     signed_word i;

     while (nsorted < n &&
            (word)base[nsorted-1].hs_start < (word)base[nsorted].hs_start)
         ++nsorted;
     if (nsorted == n) break;
     GC_ASSERT((word)base[nsorted-1].hs_start > (word)base[nsorted].hs_start);
     i = nsorted - 1;
     while (i >= 0 && (word)base[i].hs_start > (word)base[i+1].hs_start) {
       struct HeapSect tmp = base[i];
       base[i] = base[i+1];
       base[i+1] = tmp;
       --i;
     }
     GC_ASSERT((word)base[nsorted-1].hs_start < (word)base[nsorted].hs_start);
     ++nsorted;
   }
}

STATIC void GC_register_map_entries(const char *maps)
{
   const char *prot;
   ptr_t start, end;
   unsigned int maj_dev;
   ptr_t least_ha, greatest_ha;
   unsigned i;

   GC_ASSERT(I_HOLD_LOCK());
   sort_heap_sects(GC_our_memory, GC_n_memory);
   least_ha = GC_our_memory[0].hs_start;
   greatest_ha = GC_our_memory[GC_n_memory-1].hs_start
                 + GC_our_memory[GC_n_memory-1].hs_bytes;

   for (;;) {
       maps = GC_parse_map_entry(maps, &start, &end, &prot, &maj_dev, 0);
       if (NULL == maps) break;

       if (prot[1] == 'w') {
           /* This is a writable mapping.  Add it to           */
           /* the root set unless it is already otherwise      */
           /* accounted for.                                   */
           if ((word)start <= (word)GC_stackbottom
               && (word)end >= (word)GC_stackbottom) {
               /* Stack mapping; discard       */
               continue;
           }
#           ifdef THREADS
             /* This may fail, since a thread may already be           */
             /* unregistered, but its thread stack may still be there. */
             /* That can fail because the stack may disappear while    */
             /* we're marking.  Thus the marker is, and has to be      */
             /* prepared to recover from segmentation faults.          */

             if (GC_segment_is_thread_stack(start, end)) continue;

             /* FIXME: NPTL squirrels                                  */
             /* away pointers in pieces of the stack segment that we   */
             /* don't scan.  We work around this                       */
             /* by treating anything allocated by libpthread as        */
             /* uncollectible, as we do in some other cases.           */
             /* A specifically identified problem is that              */
             /* thread stacks contain pointers to dynamic thread       */
             /* vectors, which may be reused due to thread caching.    */
             /* They may not be marked if the thread is still live.    */
             /* This specific instance should be addressed by          */
             /* INCLUDE_LINUX_THREAD_DESCR, but that doesn't quite     */
             /* seem to suffice.                                       */
             /* We currently trace entire thread stacks, if they are   */
             /* are currently cached but unused.  This is              */
             /* very suboptimal for performance reasons.               */
#           endif
           /* We no longer exclude the main data segment.              */
           if ((word)end <= (word)least_ha
               || (word)start >= (word)greatest_ha) {
             /* The easy case; just trace entire segment */
             GC_add_roots_inner(start, end, TRUE);
             continue;
           }
           /* Add sections that don't belong to us. */
             i = 0;
             while ((word)(GC_our_memory[i].hs_start
                               + GC_our_memory[i].hs_bytes) < (word)start)
                 ++i;
             GC_ASSERT(i < GC_n_memory);
             if ((word)GC_our_memory[i].hs_start <= (word)start) {
                 start = GC_our_memory[i].hs_start
                         + GC_our_memory[i].hs_bytes;
                 ++i;
             }
             while (i < GC_n_memory
                    && (word)GC_our_memory[i].hs_start < (word)end
                    && (word)start < (word)end) {
                 if ((word)start < (word)GC_our_memory[i].hs_start)
                   GC_add_roots_inner(start,
                                      GC_our_memory[i].hs_start, TRUE);
                 start = GC_our_memory[i].hs_start
                         + GC_our_memory[i].hs_bytes;
                 ++i;
             }
             if ((word)start < (word)end)
                 GC_add_roots_inner(start, end, TRUE);
       } else if (prot[0] == '-' && prot[1] == '-' && prot[2] == '-') {
           /* Even roots added statically might disappear partially    */
           /* (e.g. the roots added by INCLUDE_LINUX_THREAD_DESCR).    */
           GC_remove_roots_subregion(start, end);
       }
   }
}

GC_INNER void GC_register_dynamic_libraries(void)
{
   GC_register_map_entries(GC_get_maps());
}

/* We now take care of the main data segment ourselves: */
GC_INNER GC_bool GC_register_main_static_data(void)
{
   return FALSE;
}

# define HAVE_REGISTER_MAIN_STATIC_DATA

#else /* !USE_PROC_FOR_LIBRARIES */

/* The following is the preferred way to walk dynamic libraries */
/* for glibc 2.2.4+.  Unfortunately, it doesn't work for older  */
/* versions.  Thanks to Jakub Jelinek for most of the code.     */

#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 2) \
   || (__GLIBC__ == 2 && __GLIBC_MINOR__ == 2 && defined(DT_CONFIG)) \
   || defined(HOST_ANDROID) /* Are others OK here, too? */
# ifndef HAVE_DL_ITERATE_PHDR
#   define HAVE_DL_ITERATE_PHDR
# endif
# ifdef HOST_ANDROID
   /* Android headers might have no such definition for some targets.  */
   EXTERN_C_BEGIN
   extern int dl_iterate_phdr(int (*cb)(struct dl_phdr_info *,
                                        size_t, void *),
                              void *data);
   EXTERN_C_END
# endif
#endif /* __GLIBC__ >= 2 || HOST_ANDROID */

#if defined(__DragonFly__) || defined(__FreeBSD_kernel__) \
   || (defined(FREEBSD) && __FreeBSD__ >= 7)
 /* On the FreeBSD system, any target system at major version 7 shall   */
 /* have dl_iterate_phdr; therefore, we need not make it weak as below. */
# ifndef HAVE_DL_ITERATE_PHDR
#   define HAVE_DL_ITERATE_PHDR
# endif
# define DL_ITERATE_PHDR_STRONG
#elif defined(HAVE_DL_ITERATE_PHDR)
 /* We have the header files for a glibc that includes dl_iterate_phdr.*/
 /* It may still not be available in the library on the target system. */
 /* Thus we also treat it as a weak symbol.                            */
 EXTERN_C_BEGIN
# pragma weak dl_iterate_phdr
 EXTERN_C_END
#endif

#if defined(HAVE_DL_ITERATE_PHDR)

# ifdef PT_GNU_RELRO
/* Instead of registering PT_LOAD sections directly, we keep them       */
/* in a temporary list, and filter them by excluding PT_GNU_RELRO       */
/* segments.  Processing PT_GNU_RELRO sections with                     */
/* GC_exclude_static_roots instead would be superficially cleaner.  But */
/* it runs into trouble if a client registers an overlapping segment,   */
/* which unfortunately seems quite possible.                            */

#   define MAX_LOAD_SEGS MAX_ROOT_SETS

   static struct load_segment {
     ptr_t start;
     ptr_t end;
     /* Room for a second segment if we remove a RELRO segment */
     /* from the middle.                                       */
     ptr_t start2;
     ptr_t end2;
   } load_segs[MAX_LOAD_SEGS];

   static int n_load_segs;
   static GC_bool load_segs_overflow;
# endif /* PT_GNU_RELRO */

STATIC int GC_register_dynlib_callback(struct dl_phdr_info * info,
                                      size_t size, void * ptr)
{
 const ElfW(Phdr) * p;
 ptr_t start, end;
 int i;

 /* Make sure struct dl_phdr_info is at least as big as we need.  */
 if (size < offsetof (struct dl_phdr_info, dlpi_phnum)
     + sizeof (info->dlpi_phnum))
   return -1;

 p = info->dlpi_phdr;
 for (i = 0; i < (int)info->dlpi_phnum; i++, p++) {
   if (p->p_type == PT_LOAD) {
     GC_has_static_roots_func callback = GC_has_static_roots;
     if ((p->p_flags & PF_W) == 0) continue;

     start = (ptr_t)p->p_vaddr + info->dlpi_addr;
     end = start + p->p_memsz;
     if (callback != 0 && !callback(info->dlpi_name, start, p->p_memsz))
       continue;
#     ifdef PT_GNU_RELRO
#       if CPP_WORDSZ == 64
         /* TODO: GC_push_all eventually does the correct          */
         /* rounding to the next multiple of ALIGNMENT, so, most   */
         /* probably, we should remove the corresponding assertion */
         /* check in GC_add_roots_inner along with this code line. */
         /* start pointer value may require aligning.              */
         start = (ptr_t)((word)start & ~(word)(sizeof(word) - 1));
#       endif
       if (n_load_segs >= MAX_LOAD_SEGS) {
         if (!load_segs_overflow) {
           WARN("Too many PT_LOAD segments;"
                " registering as roots directly...\n", 0);
           load_segs_overflow = TRUE;
         }
         GC_add_roots_inner(start, end, TRUE);
       } else {
         load_segs[n_load_segs].start = start;
         load_segs[n_load_segs].end = end;
         load_segs[n_load_segs].start2 = 0;
         load_segs[n_load_segs].end2 = 0;
         ++n_load_segs;
       }
#     else
       GC_add_roots_inner(start, end, TRUE);
#     endif /* !PT_GNU_RELRO */
   }
 }

# ifdef PT_GNU_RELRO
   p = info->dlpi_phdr;
   for (i = 0; i < (int)info->dlpi_phnum; i++, p++) {
     if (p->p_type == PT_GNU_RELRO) {
       /* This entry is known to be constant and will eventually be    */
       /* remapped as read-only.  However, the address range covered   */
       /* by this entry is typically a subset of a previously          */
       /* encountered "LOAD" segment, so we need to exclude it.        */
       int j;

       start = (ptr_t)p->p_vaddr + info->dlpi_addr;
       end = start + p->p_memsz;
       for (j = n_load_segs; --j >= 0; ) {
         if ((word)start >= (word)load_segs[j].start
             && (word)start < (word)load_segs[j].end) {
           if (load_segs[j].start2 != 0) {
             WARN("More than one GNU_RELRO segment per load one\n",0);
           } else {
             GC_ASSERT((word)end <=
                           (((word)load_segs[j].end + GC_page_size - 1) &
                            ~(word)(GC_page_size - 1)));
             /* Remove from the existing load segment */
             load_segs[j].end2 = load_segs[j].end;
             load_segs[j].end = start;
             load_segs[j].start2 = end;
             /* Note that start2 may be greater than end2 because of   */
             /* p->p_memsz value multiple of page size.                */
           }
           break;
         }
         if (0 == j && 0 == GC_has_static_roots)
           WARN("Failed to find PT_GNU_RELRO segment"
                " inside PT_LOAD region\n", 0);
           /* No warning reported in case of the callback is present   */
           /* because most likely the segment has been excluded.       */
       }
     }
   }
# endif

 *(int *)ptr = 1;     /* Signal that we were called */
 return 0;
}

/* Do we need to separately register the main static data segment? */
GC_INNER GC_bool GC_register_main_static_data(void)
{
# ifdef DL_ITERATE_PHDR_STRONG
   /* If dl_iterate_phdr is not a weak symbol then don't test against  */
   /* zero (otherwise a compiler might issue a warning).               */
   return FALSE;
# else
   return 0 == COVERT_DATAFLOW(dl_iterate_phdr);
# endif
}

/* Return TRUE if we succeed, FALSE if dl_iterate_phdr wasn't there. */
STATIC GC_bool GC_register_dynamic_libraries_dl_iterate_phdr(void)
{
 int did_something;
 if (GC_register_main_static_data())
   return FALSE;

# ifdef PT_GNU_RELRO
   {
     static GC_bool excluded_segs = FALSE;
     n_load_segs = 0;
     load_segs_overflow = FALSE;
     if (!EXPECT(excluded_segs, TRUE)) {
       GC_exclude_static_roots_inner((ptr_t)load_segs,
                                     (ptr_t)load_segs + sizeof(load_segs));
       excluded_segs = TRUE;
     }
   }
# endif

 did_something = 0;
 dl_iterate_phdr(GC_register_dynlib_callback, &did_something);
 if (did_something) {
#   ifdef PT_GNU_RELRO
     int i;

     for (i = 0; i < n_load_segs; ++i) {
       if ((word)load_segs[i].end > (word)load_segs[i].start) {
         GC_add_roots_inner(load_segs[i].start, load_segs[i].end, TRUE);
       }
       if ((word)load_segs[i].end2 > (word)load_segs[i].start2) {
         GC_add_roots_inner(load_segs[i].start2, load_segs[i].end2, TRUE);
       }
     }
#   endif
 } else {
     ptr_t datastart, dataend;
#     ifdef DATASTART_IS_FUNC
       static ptr_t datastart_cached = (ptr_t)GC_WORD_MAX;

       /* Evaluate DATASTART only once.  */
       if (datastart_cached == (ptr_t)GC_WORD_MAX) {
         datastart_cached = DATASTART;
       }
       datastart = datastart_cached;
#     else
       datastart = DATASTART;
#     endif
#     ifdef DATAEND_IS_FUNC
       {
         static ptr_t dataend_cached = 0;
         /* Evaluate DATAEND only once. */
         if (dataend_cached == 0) {
           dataend_cached = DATAEND;
         }
         dataend = dataend_cached;
       }
#     else
       dataend = DATAEND;
#     endif
     if (NULL == *(char * volatile *)&datastart
         || (word)datastart > (word)dataend)
       ABORT_ARG2("Wrong DATASTART/END pair",
                  ": %p .. %p", (void *)datastart, (void *)dataend);

     /* dl_iterate_phdr may forget the static data segment in  */
     /* statically linked executables.                         */
     GC_add_roots_inner(datastart, dataend, TRUE);
#     ifdef GC_HAVE_DATAREGION2
       if ((word)DATASTART2 - 1U >= (word)DATAEND2) {
                       /* Subtract one to check also for NULL  */
                       /* without a compiler warning.          */
         ABORT_ARG2("Wrong DATASTART/END2 pair",
                    ": %p .. %p", (void *)DATASTART2, (void *)DATAEND2);
       }
       GC_add_roots_inner(DATASTART2, DATAEND2, TRUE);
#     endif
 }
 return TRUE;
}

# define HAVE_REGISTER_MAIN_STATIC_DATA

#else /* !HAVE_DL_ITERATE_PHDR */

/* Dynamic loading code for Linux running ELF. Somewhat tested on
* Linux/x86, untested but hopefully should work on Linux/Alpha.
* This code was derived from the Solaris/ELF support. Thanks to
* whatever kind soul wrote that.  - Patrick Bridges */

/* This doesn't necessarily work in all cases, e.g. with preloaded
* dynamic libraries.                                           */

# if defined(NETBSD) || defined(OPENBSD)
#   include <sys/exec_elf.h>
  /* for compatibility with 1.4.x */
#   ifndef DT_DEBUG
#     define DT_DEBUG   21
#   endif
#   ifndef PT_LOAD
#     define PT_LOAD    1
#   endif
#   ifndef PF_W
#     define PF_W       2
#   endif
# elif !defined(HOST_ANDROID)
#  include <elf.h>
# endif

# ifndef HOST_ANDROID
#   include <link.h>
# endif

#endif /* !HAVE_DL_ITERATE_PHDR */

EXTERN_C_BEGIN
#ifdef __GNUC__
# pragma weak _DYNAMIC
#endif
extern ElfW(Dyn) _DYNAMIC[];
EXTERN_C_END

STATIC struct link_map *
GC_FirstDLOpenedLinkMap(void)
{
   static struct link_map *cachedResult = 0;

   if (0 == COVERT_DATAFLOW(_DYNAMIC)) {
       /* _DYNAMIC symbol not resolved. */
       return(0);
   }
   if( cachedResult == 0 ) {
#     if defined(NETBSD) && defined(RTLD_DI_LINKMAP)
#       if defined(CPPCHECK)
#         define GC_RTLD_DI_LINKMAP 2
#       else
#         define GC_RTLD_DI_LINKMAP RTLD_DI_LINKMAP
#       endif
       struct link_map *lm = NULL;
       if (!dlinfo(RTLD_SELF, GC_RTLD_DI_LINKMAP, &lm) && lm != NULL) {
           /* Now lm points link_map object of libgc.  Since it    */
           /* might not be the first dynamically linked object,    */
           /* try to find it (object next to the main object).     */
           while (lm->l_prev != NULL) {
               lm = lm->l_prev;
           }
           cachedResult = lm->l_next;
       }
#     else
       ElfW(Dyn) *dp;
       int tag;

       for( dp = _DYNAMIC; (tag = dp->d_tag) != 0; dp++ ) {
           if (tag == DT_DEBUG) {
               struct r_debug *rd = (struct r_debug *)dp->d_un.d_ptr;
               /* d_ptr could be null if libs are linked statically. */
               if (rd != NULL) {
                   struct link_map *lm = rd->r_map;
                   if (lm != NULL)
                       cachedResult = lm->l_next; /* might be NULL */
               }
               break;
           }
       }
#     endif /* !NETBSD || !RTLD_DI_LINKMAP */
   }
   return cachedResult;
}

GC_INNER void GC_register_dynamic_libraries(void)
{
 struct link_map *lm;

# ifdef HAVE_DL_ITERATE_PHDR
   if (GC_register_dynamic_libraries_dl_iterate_phdr()) {
       return;
   }
# endif
 for (lm = GC_FirstDLOpenedLinkMap(); lm != 0; lm = lm->l_next)
   {
       ElfW(Ehdr) * e;
       ElfW(Phdr) * p;
       unsigned long offset;
       char * start;
       int i;

       e = (ElfW(Ehdr) *) lm->l_addr;
#       ifdef HOST_ANDROID
         if (e == NULL)
           continue;
#       endif
       p = ((ElfW(Phdr) *)(((char *)(e)) + e->e_phoff));
       offset = ((unsigned long)(lm->l_addr));
       for( i = 0; i < (int)e->e_phnum; i++, p++ ) {
         switch( p->p_type ) {
           case PT_LOAD:
             {
               if( !(p->p_flags & PF_W) ) break;
               start = ((char *)(p->p_vaddr)) + offset;
               GC_add_roots_inner(start, start + p->p_memsz, TRUE);
             }
             break;
           default:
             break;
         }
       }
   }
}

#endif /* !USE_PROC_FOR_LIBRARIES */

#endif /* LINUX */

#if defined(IRIX5) || (defined(USE_PROC_FOR_LIBRARIES) && !defined(LINUX))

#include <sys/procfs.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <elf.h>
#include <errno.h>
#include <signal.h>  /* Only for the following test. */
#ifndef _sigargs
# define IRIX6
#endif

/* We use /proc to track down all parts of the address space that are   */
/* mapped by the process, and throw out regions we know we shouldn't    */
/* worry about.  This may also work under other SVR4 variants.          */
GC_INNER void GC_register_dynamic_libraries(void)
{
   static int fd = -1;
   char buf[30];
   static prmap_t * addr_map = 0;
   static int current_sz = 0;  /* Number of records currently in addr_map */
   int needed_sz = 0;          /* Required size of addr_map            */
   int i;
   long flags;
   ptr_t start;
   ptr_t limit;
   ptr_t heap_start = HEAP_START;
   ptr_t heap_end = heap_start;

#   ifdef SOLARISDL
#     define MA_PHYS 0
#   endif /* SOLARISDL */

   if (fd < 0) {
     (void)snprintf(buf, sizeof(buf), "/proc/%ld", (long)getpid());
     buf[sizeof(buf) - 1] = '\0';
     fd = open(buf, O_RDONLY);
     if (fd < 0) {
       ABORT("/proc open failed");
     }
   }
   if (ioctl(fd, PIOCNMAP, &needed_sz) < 0) {
       ABORT_ARG2("/proc PIOCNMAP ioctl failed",
                  ": fd= %d, errno= %d", fd, errno);
   }
   if (needed_sz >= current_sz) {
       GC_scratch_recycle_no_gww(addr_map,
                                 (size_t)current_sz * sizeof(prmap_t));
       current_sz = needed_sz * 2 + 1;
                       /* Expansion, plus room for 0 record */
       addr_map = (prmap_t *)GC_scratch_alloc(
                               (size_t)current_sz * sizeof(prmap_t));
       if (addr_map == NULL)
         ABORT("Insufficient memory for address map");
   }
   if (ioctl(fd, PIOCMAP, addr_map) < 0) {
       ABORT_ARG3("/proc PIOCMAP ioctl failed",
                  ": errcode= %d, needed_sz= %d, addr_map= %p",
                  errno, needed_sz, (void *)addr_map);
   }
   if (GC_n_heap_sects > 0) {
       heap_end = GC_heap_sects[GC_n_heap_sects-1].hs_start
                       + GC_heap_sects[GC_n_heap_sects-1].hs_bytes;
       if ((word)heap_end < (word)GC_scratch_last_end_ptr)
         heap_end = GC_scratch_last_end_ptr;
   }
   for (i = 0; i < needed_sz; i++) {
       flags = addr_map[i].pr_mflags;
       if ((flags & (MA_BREAK | MA_STACK | MA_PHYS
                     | MA_FETCHOP | MA_NOTCACHED)) != 0) goto irrelevant;
       if ((flags & (MA_READ | MA_WRITE)) != (MA_READ | MA_WRITE))
           goto irrelevant;
         /* The latter test is empirically useless in very old Irix    */
         /* versions.  Other than the                                  */
         /* main data and stack segments, everything appears to be     */
         /* mapped readable, writable, executable, and shared(!!).     */
         /* This makes no sense to me. - HB                            */
       start = (ptr_t)(addr_map[i].pr_vaddr);
       if (GC_roots_present(start)) goto irrelevant;
       if ((word)start < (word)heap_end && (word)start >= (word)heap_start)
               goto irrelevant;

       limit = start + addr_map[i].pr_size;
       /* The following seemed to be necessary for very old versions   */
       /* of Irix, but it has been reported to discard relevant        */
       /* segments under Irix 6.5.                                     */
#       ifndef IRIX6
         if (addr_map[i].pr_off == 0 && strncmp(start, ELFMAG, 4) == 0) {
           /* Discard text segments, i.e. 0-offset mappings against    */
           /* executable files which appear to have ELF headers.       */
           caddr_t arg;
           int obj;
#           define MAP_IRR_SZ 10
           static ptr_t map_irr[MAP_IRR_SZ];
                                       /* Known irrelevant map entries */
           static int n_irr = 0;
           struct stat buf;
           int j;

           for (j = 0; j < n_irr; j++) {
               if (map_irr[j] == start) goto irrelevant;
           }
           arg = (caddr_t)start;
           obj = ioctl(fd, PIOCOPENM, &arg);
           if (obj >= 0) {
               fstat(obj, &buf);
               close(obj);
               if ((buf.st_mode & 0111) != 0) {
                   if (n_irr < MAP_IRR_SZ) {
                       map_irr[n_irr++] = start;
                   }
                   goto irrelevant;
               }
           }
         }
#       endif /* !IRIX6 */
       GC_add_roots_inner(start, limit, TRUE);
     irrelevant: ;
   }
   /* Don't keep cached descriptor, for now.  Some kernels don't like us */
   /* to keep a /proc file descriptor around during kill -9.             */
   /* Otherwise, it should also require FD_CLOEXEC and proper handling   */
   /* at fork (i.e. close because of the pid change).                    */
       if (close(fd) < 0) ABORT("Couldn't close /proc file");
       fd = -1;
}

# endif /* USE_PROC_FOR_LIBRARIES || IRIX5 */

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

# include <stdlib.h>

 /* We traverse the entire address space and register all segments     */
 /* that could possibly have been written to.                          */
 STATIC void GC_cond_add_roots(char *base, char * limit)
 {
#   ifdef GC_WIN32_THREADS
     char * curr_base = base;
     char * next_stack_lo;
     char * next_stack_hi;

     if (base == limit) return;
     for(;;) {
         GC_get_next_stack(curr_base, limit, &next_stack_lo, &next_stack_hi);
         if ((word)next_stack_lo >= (word)limit) break;
         if ((word)next_stack_lo > (word)curr_base)
           GC_add_roots_inner(curr_base, next_stack_lo, TRUE);
         curr_base = next_stack_hi;
     }
     if ((word)curr_base < (word)limit)
       GC_add_roots_inner(curr_base, limit, TRUE);
#   else
     char * stack_top
        = (char *)((word)GC_approx_sp() &
                   ~(word)(GC_sysinfo.dwAllocationGranularity - 1));

     if (base == limit) return;
     if ((word)limit > (word)stack_top
         && (word)base < (word)GC_stackbottom) {
         /* Part of the stack; ignore it. */
         return;
     }
     GC_add_roots_inner(base, limit, TRUE);
#   endif
 }

#ifdef DYNAMIC_LOADING
 /* GC_register_main_static_data is not needed unless DYNAMIC_LOADING. */
 GC_INNER GC_bool GC_register_main_static_data(void)
 {
#   if defined(MSWINCE) || defined(CYGWIN32)
     /* Do we need to separately register the main static data segment? */
     return FALSE;
#   else
     return GC_no_win32_dlls;
#   endif
 }
# define HAVE_REGISTER_MAIN_STATIC_DATA
#endif /* DYNAMIC_LOADING */

# ifdef DEBUG_VIRTUALQUERY
 void GC_dump_meminfo(MEMORY_BASIC_INFORMATION *buf)
 {
   GC_printf("BaseAddress= 0x%lx, AllocationBase= 0x%lx,"
             " RegionSize= 0x%lx(%lu)\n",
             buf -> BaseAddress, buf -> AllocationBase,
             buf -> RegionSize, buf -> RegionSize);
   GC_printf("\tAllocationProtect= 0x%lx, State= 0x%lx, Protect= 0x%lx, "
             "Type= 0x%lx\n", buf -> AllocationProtect, buf -> State,
             buf -> Protect, buf -> Type);
 }
# endif /* DEBUG_VIRTUALQUERY */

# if defined(MSWINCE) || defined(CYGWIN32)
   /* FIXME: Should we really need to scan MEM_PRIVATE sections?       */
   /* For now, we don't add MEM_PRIVATE sections to the data roots for */
   /* WinCE because otherwise SEGV fault sometimes happens to occur in */
   /* GC_mark_from() (and, even if we use WRAP_MARK_SOME, WinCE prints */
   /* a "Data Abort" message to the debugging console).                */
   /* To workaround that, use -DGC_REGISTER_MEM_PRIVATE.               */
#   define GC_wnt TRUE
# endif

 GC_INNER void GC_register_dynamic_libraries(void)
 {
   MEMORY_BASIC_INFORMATION buf;
   DWORD protect;
   LPVOID p;
   char * base;
   char * limit;

#   ifdef MSWIN32
     if (GC_no_win32_dlls) return;
#   endif
   p = GC_sysinfo.lpMinimumApplicationAddress;
   base = limit = (char *)p;
   while ((word)p < (word)GC_sysinfo.lpMaximumApplicationAddress) {
       size_t result = VirtualQuery(p, &buf, sizeof(buf));

#       ifdef MSWINCE
         if (result == 0) {
           if ((word)p > GC_WORD_MAX - GC_sysinfo.dwAllocationGranularity)
             break; /* overflow */
           /* Page is free; advance to the next possible allocation base. */
           p = (LPVOID)(((DWORD)p + GC_sysinfo.dwAllocationGranularity)
                        & ~(GC_sysinfo.dwAllocationGranularity-1));
         } else
#       endif
       /* else */ {
           if (result != sizeof(buf)) {
               ABORT("Weird VirtualQuery result");
           }
           if ((word)p > GC_WORD_MAX - buf.RegionSize) break; /* overflow */

           protect = buf.Protect;
           if (buf.State == MEM_COMMIT
               && (protect == PAGE_EXECUTE_READWRITE
                   || protect == PAGE_EXECUTE_WRITECOPY
                   || protect == PAGE_READWRITE
                   || protect == PAGE_WRITECOPY)
               && (buf.Type == MEM_IMAGE
#                   ifdef GC_REGISTER_MEM_PRIVATE
                     || (protect == PAGE_READWRITE && buf.Type == MEM_PRIVATE)
#                   else
                     /* There is some evidence that we cannot always   */
                     /* ignore MEM_PRIVATE sections under Windows ME   */
                     /* and predecessors.  Hence we now also check for */
                     /* that case.                                     */
                     || (!GC_wnt && buf.Type == MEM_PRIVATE)
#                   endif
                  )
               && !GC_is_heap_base(buf.AllocationBase)) {
#               ifdef DEBUG_VIRTUALQUERY
                 GC_dump_meminfo(&buf);
#               endif
               if ((char *)p != limit) {
                   GC_cond_add_roots(base, limit);
                   base = (char *)p;
               }
               limit = (char *)p + buf.RegionSize;
           }
           p = (char *)p + buf.RegionSize;
       }
   }
   GC_cond_add_roots(base, limit);
 }

#endif /* MSWIN32 || MSWINCE || CYGWIN32 */

#if defined(ALPHA) && defined(OSF1)

#include <loader.h>

EXTERN_C_BEGIN
extern char *sys_errlist[];
extern int sys_nerr;
extern int errno;
EXTERN_C_END

GC_INNER void GC_register_dynamic_libraries(void)
{
 ldr_module_t moduleid = LDR_NULL_MODULE;
 ldr_process_t mypid = ldr_my_process(); /* obtain id of this process */

 /* For each module */
   while (TRUE) {
     ldr_module_info_t moduleinfo;
     size_t modulereturnsize;
     ldr_region_t region;
     ldr_region_info_t regioninfo;
     size_t regionreturnsize;
     int status = ldr_next_module(mypid, &moduleid);
                               /* Get the next (first) module */

     /* Any more modules? */
       if (moduleid == LDR_NULL_MODULE)
           break;    /* No more modules */

     /* Check status AFTER checking moduleid because       */
     /* of a bug in the non-shared ldr_next_module stub.   */
       if (status != 0) {
         ABORT_ARG3("ldr_next_module failed",
                    ": status= %d, errcode= %d (%s)", status, errno,
                    errno < sys_nerr ? sys_errlist[errno] : "");
       }

     /* Get the module information */
       status = ldr_inq_module(mypid, moduleid, &moduleinfo,
                               sizeof(moduleinfo), &modulereturnsize);
       if (status != 0 )
           ABORT("ldr_inq_module failed");

     /* is module for the main program (i.e. nonshared portion)? */
         if (moduleinfo.lmi_flags & LDR_MAIN)
             continue;    /* skip the main module */

#     ifdef DL_VERBOSE
       GC_log_printf("---Module---\n");
       GC_log_printf("Module ID: %ld\n", moduleinfo.lmi_modid);
       GC_log_printf("Count of regions: %d\n", moduleinfo.lmi_nregion);
       GC_log_printf("Flags for module: %016lx\n", moduleinfo.lmi_flags);
       GC_log_printf("Module pathname: \"%s\"\n", moduleinfo.lmi_name);
#     endif

     /* For each region in this module */
       for (region = 0; region < moduleinfo.lmi_nregion; region++) {
         /* Get the region information */
           status = ldr_inq_region(mypid, moduleid, region, &regioninfo,
                                   sizeof(regioninfo), &regionreturnsize);
           if (status != 0 )
               ABORT("ldr_inq_region failed");

         /* only process writable (data) regions */
           if (! (regioninfo.lri_prot & LDR_W))
               continue;

#         ifdef DL_VERBOSE
           GC_log_printf("--- Region ---\n");
           GC_log_printf("Region number: %ld\n", regioninfo.lri_region_no);
           GC_log_printf("Protection flags: %016x\n", regioninfo.lri_prot);
           GC_log_printf("Virtual address: %p\n", regioninfo.lri_vaddr);
           GC_log_printf("Mapped address: %p\n", regioninfo.lri_mapaddr);
           GC_log_printf("Region size: %ld\n", regioninfo.lri_size);
           GC_log_printf("Region name: \"%s\"\n", regioninfo.lri_name);
#         endif

         /* register region as a garbage collection root */
         GC_add_roots_inner((char *)regioninfo.lri_mapaddr,
                       (char *)regioninfo.lri_mapaddr + regioninfo.lri_size,
                       TRUE);

       }
   }
}
#endif

#if defined(HPUX)

#include <errno.h>
#include <dl.h>

EXTERN_C_BEGIN
extern char *sys_errlist[];
extern int sys_nerr;
EXTERN_C_END

GC_INNER void GC_register_dynamic_libraries(void)
{
 int index = 1; /* Ordinal position in shared library search list */

 /* For each dynamic library loaded */
   while (TRUE) {
     struct shl_descriptor *shl_desc; /* Shared library info, see dl.h */
     int status = shl_get(index, &shl_desc);
                               /* Get info about next shared library   */

     /* Check if this is the end of the list or if some error occurred */
       if (status != 0) {
#        ifdef GC_HPUX_THREADS
          /* I've seen errno values of 0.  The man page is not clear   */
          /* as to whether errno should get set on a -1 return.        */
          break;
#        else
         if (errno == EINVAL) {
           break; /* Moved past end of shared library list --> finished */
         } else {
           ABORT_ARG3("shl_get failed",
                      ": status= %d, errcode= %d (%s)", status, errno,
                      errno < sys_nerr ? sys_errlist[errno] : "");
         }
#        endif
       }

#     ifdef DL_VERBOSE
       GC_log_printf("---Shared library---\n");
       GC_log_printf("filename= \"%s\"\n", shl_desc->filename);
       GC_log_printf("index= %d\n", index);
       GC_log_printf("handle= %08x\n", (unsigned long) shl_desc->handle);
       GC_log_printf("text seg.start= %08x\n", shl_desc->tstart);
       GC_log_printf("text seg.end= %08x\n", shl_desc->tend);
       GC_log_printf("data seg.start= %08x\n", shl_desc->dstart);
       GC_log_printf("data seg.end= %08x\n", shl_desc->dend);
       GC_log_printf("ref.count= %lu\n", shl_desc->ref_count);
#     endif

     /* register shared library's data segment as a garbage collection root */
       GC_add_roots_inner((char *) shl_desc->dstart,
                          (char *) shl_desc->dend, TRUE);

       index++;
   }
}
#endif /* HPUX */

#ifdef AIX
# include <alloca.h>
# include <sys/ldr.h>
# include <sys/errno.h>
 GC_INNER void GC_register_dynamic_libraries(void)
 {
     int ldibuflen = 8192;

     for (;;) {
       int len;
       struct ld_info *ldi;
#       if defined(CPPCHECK)
         char ldibuf[ldibuflen];
#       else
         char *ldibuf = alloca(ldibuflen);
#       endif

       len = loadquery(L_GETINFO, ldibuf, ldibuflen);
       if (len < 0) {
               if (errno != ENOMEM) {
                       ABORT("loadquery failed");
               }
               ldibuflen *= 2;
               continue;
       }

       ldi = (struct ld_info *)ldibuf;
       while (ldi) {
               len = ldi->ldinfo_next;
               GC_add_roots_inner(
                               ldi->ldinfo_dataorg,
                               (ptr_t)(unsigned long)ldi->ldinfo_dataorg
                               + ldi->ldinfo_datasize,
                               TRUE);
               ldi = len ? (struct ld_info *)((char *)ldi + len) : 0;
       }
       break;
     }
 }
#endif /* AIX */

#ifdef DARWIN

/* __private_extern__ hack required for pre-3.4 gcc versions.   */
#ifndef __private_extern__
# define __private_extern__ extern
# include <mach-o/dyld.h>
# undef __private_extern__
#else
# include <mach-o/dyld.h>
#endif
#include <mach-o/getsect.h>

/*#define DARWIN_DEBUG*/

/* Writable sections generally available on Darwin.     */
STATIC const struct dyld_sections_s {
   const char *seg;
   const char *sect;
} GC_dyld_sections[] = {
   { SEG_DATA, SECT_DATA },
   /* Used by FSF GCC, but not by OS X system tools, so far.   */
   { SEG_DATA, "__static_data" },
   { SEG_DATA, SECT_BSS },
   { SEG_DATA, SECT_COMMON },
   /* FSF GCC - zero-sized object sections for targets         */
   /*supporting section anchors.                               */
   { SEG_DATA, "__zobj_data" },
   { SEG_DATA, "__zobj_bss" }
};

/* Additional writable sections:                                */
/* GCC on Darwin constructs aligned sections "on demand", where */
/* the alignment size is embedded in the section name.          */
/* Furthermore, there are distinctions between sections         */
/* containing private vs. public symbols.  It also constructs   */
/* sections specifically for zero-sized objects, when the       */
/* target supports section anchors.                             */
STATIC const char * const GC_dyld_add_sect_fmts[] = {
 "__bss%u",
 "__pu_bss%u",
 "__zo_bss%u",
 "__zo_pu_bss%u"
};

/* Currently, mach-o will allow up to the max of 2^15 alignment */
/* in an object file.                                           */
#ifndef L2_MAX_OFILE_ALIGNMENT
# define L2_MAX_OFILE_ALIGNMENT 15
#endif

STATIC const char *GC_dyld_name_for_hdr(const struct GC_MACH_HEADER *hdr)
{
   unsigned long i, c;
   c = _dyld_image_count();
   for (i = 0; i < c; i++)
     if ((const struct GC_MACH_HEADER *)_dyld_get_image_header(i) == hdr)
       return _dyld_get_image_name(i);
   return NULL;
}

/* This should never be called by a thread holding the lock.    */
STATIC void GC_dyld_image_add(const struct GC_MACH_HEADER *hdr,
                             intptr_t slide)
{
 unsigned long start, end;
 unsigned i, j;
 const struct GC_MACH_SECTION *sec;
 const char *name;
 GC_has_static_roots_func callback = GC_has_static_roots;
 DCL_LOCK_STATE;

 if (GC_no_dls) return;
# ifdef DARWIN_DEBUG
   name = GC_dyld_name_for_hdr(hdr);
# else
   name = callback != 0 ? GC_dyld_name_for_hdr(hdr) : NULL;
# endif
 for (i = 0; i < sizeof(GC_dyld_sections)/sizeof(GC_dyld_sections[0]); i++) {
   sec = GC_GETSECTBYNAME(hdr, GC_dyld_sections[i].seg,
                          GC_dyld_sections[i].sect);
   if (sec == NULL || sec->size < sizeof(word))
     continue;
   start = slide + sec->addr;
   end = start + sec->size;
   LOCK();
   /* The user callback is called holding the lock.    */
   if (callback == 0 || callback(name, (void*)start, (size_t)sec->size)) {
#     ifdef DARWIN_DEBUG
       GC_log_printf(
             "Adding section __DATA,%s at %p-%p (%lu bytes) from image %s\n",
              GC_dyld_sections[i].sect, (void*)start, (void*)end,
              (unsigned long)sec->size, name);
#     endif
     GC_add_roots_inner((ptr_t)start, (ptr_t)end, FALSE);
   }
   UNLOCK();
 }

 /* Sections constructed on demand.    */
 for (j = 0; j < sizeof(GC_dyld_add_sect_fmts) / sizeof(char *); j++) {
   const char *fmt = GC_dyld_add_sect_fmts[j];

   /* Add our manufactured aligned BSS sections.       */
   for (i = 0; i <= L2_MAX_OFILE_ALIGNMENT; i++) {
     char secnam[16];

     (void)snprintf(secnam, sizeof(secnam), fmt, (unsigned)i);
     secnam[sizeof(secnam) - 1] = '\0';
     sec = GC_GETSECTBYNAME(hdr, SEG_DATA, secnam);
     if (sec == NULL || sec->size == 0)
       continue;
     start = slide + sec->addr;
     end = start + sec->size;
#     ifdef DARWIN_DEBUG
       GC_log_printf("Adding on-demand section __DATA,%s at"
                     " %p-%p (%lu bytes) from image %s\n",
                     secnam, (void*)start, (void*)end,
                     (unsigned long)sec->size, name);
#     endif
     GC_add_roots((char*)start, (char*)end);
   }
 }

# if defined(DARWIN_DEBUG) && !defined(NO_DEBUGGING)
   LOCK();
   GC_print_static_roots();
   UNLOCK();
# endif
}

/* This should never be called by a thread holding the lock.    */
STATIC void GC_dyld_image_remove(const struct GC_MACH_HEADER *hdr,
                                intptr_t slide)
{
 unsigned long start, end;
 unsigned i, j;
 const struct GC_MACH_SECTION *sec;
# if defined(DARWIN_DEBUG) && !defined(NO_DEBUGGING)
   DCL_LOCK_STATE;
# endif

 for (i = 0; i < sizeof(GC_dyld_sections)/sizeof(GC_dyld_sections[0]); i++) {
   sec = GC_GETSECTBYNAME(hdr, GC_dyld_sections[i].seg,
                          GC_dyld_sections[i].sect);
   if (sec == NULL || sec->size == 0)
     continue;
   start = slide + sec->addr;
   end = start + sec->size;
#   ifdef DARWIN_DEBUG
     GC_log_printf(
           "Removing section __DATA,%s at %p-%p (%lu bytes) from image %s\n",
           GC_dyld_sections[i].sect, (void*)start, (void*)end,
           (unsigned long)sec->size, GC_dyld_name_for_hdr(hdr));
#   endif
   GC_remove_roots((char*)start, (char*)end);
 }

 /* Remove our on-demand sections.     */
 for (j = 0; j < sizeof(GC_dyld_add_sect_fmts) / sizeof(char *); j++) {
   const char *fmt = GC_dyld_add_sect_fmts[j];

   for (i = 0; i <= L2_MAX_OFILE_ALIGNMENT; i++) {
     char secnam[16];

     (void)snprintf(secnam, sizeof(secnam), fmt, (unsigned)i);
     secnam[sizeof(secnam) - 1] = '\0';
     sec = GC_GETSECTBYNAME(hdr, SEG_DATA, secnam);
     if (sec == NULL || sec->size == 0)
       continue;
     start = slide + sec->addr;
     end = start + sec->size;
#     ifdef DARWIN_DEBUG
       GC_log_printf("Removing on-demand section __DATA,%s at"
                     " %p-%p (%lu bytes) from image %s\n", secnam,
                     (void*)start, (void*)end, (unsigned long)sec->size,
                     GC_dyld_name_for_hdr(hdr));
#     endif
     GC_remove_roots((char*)start, (char*)end);
   }
 }

# if defined(DARWIN_DEBUG) && !defined(NO_DEBUGGING)
   LOCK();
   GC_print_static_roots();
   UNLOCK();
# endif
}

GC_INNER void GC_register_dynamic_libraries(void)
{
   /* Currently does nothing. The callbacks are setup by GC_init_dyld()
   The dyld library takes it from there. */
}

/* The _dyld_* functions have an internal lock so no _dyld functions
  can be called while the world is stopped without the risk of a deadlock.
  Because of this we MUST setup callbacks BEFORE we ever stop the world.
  This should be called BEFORE any thread is created and WITHOUT the
  allocation lock held. */

GC_INNER void GC_init_dyld(void)
{
 static GC_bool initialized = FALSE;

 if (initialized) return;

# ifdef DARWIN_DEBUG
   GC_log_printf("Registering dyld callbacks...\n");
# endif

 /* Apple's Documentation:
    When you call _dyld_register_func_for_add_image, the dynamic linker
    runtime calls the specified callback (func) once for each of the images
    that is currently loaded into the program. When a new image is added to
    the program, your callback is called again with the mach_header for the
    new image, and the virtual memory slide amount of the new image.

    This WILL properly register already linked libraries and libraries
    linked in the future.
 */
 _dyld_register_func_for_add_image(
       (void (*)(const struct mach_header*, intptr_t))GC_dyld_image_add);
 _dyld_register_func_for_remove_image(
       (void (*)(const struct mach_header*, intptr_t))GC_dyld_image_remove);
                       /* Structure mach_header_64 has the same fields */
                       /* as mach_header except for the reserved one   */
                       /* at the end, so these casts are OK.           */

 /* Set this early to avoid reentrancy issues. */
 initialized = TRUE;

# ifdef NO_DYLD_BIND_FULLY_IMAGE
   /* FIXME: What should we do in this case?   */
# else
   if (GC_no_dls) return; /* skip main data segment registration */

   /* When the environment variable is set, the dynamic linker binds   */
   /* all undefined symbols the application needs at launch time.      */
   /* This includes function symbols that are normally bound lazily at */
   /* the time of their first invocation.                              */
   if (GETENV("DYLD_BIND_AT_LAUNCH") == 0) {
     /* The environment variable is unset, so we should bind manually. */
#     ifdef DARWIN_DEBUG
       GC_log_printf("Forcing full bind of GC code...\n");
#     endif
     /* FIXME: '_dyld_bind_fully_image_containing_address' is deprecated. */
     if (!_dyld_bind_fully_image_containing_address(
                                                 (unsigned long *)GC_malloc))
       ABORT("_dyld_bind_fully_image_containing_address failed");
   }
# endif
}

#define HAVE_REGISTER_MAIN_STATIC_DATA
GC_INNER GC_bool GC_register_main_static_data(void)
{
 /* Already done through dyld callbacks */
 return FALSE;
}

#endif /* DARWIN */

#if defined(HAIKU)
# include <kernel/image.h>

 GC_INNER void GC_register_dynamic_libraries(void)
 {
   image_info info;
   int32 cookie = 0;

   while (get_next_image_info(0, &cookie, &info) == B_OK) {
     ptr_t data = (ptr_t)info.data;
     GC_add_roots_inner(data, data + info.data_size, TRUE);
   }
 }
#endif /* HAIKU */

#elif defined(PCR)

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

 GC_INNER void GC_register_dynamic_libraries(void)
 {
   /* Add new static data areas of dynamically loaded modules. */
   PCR_IL_LoadedFile * p = PCR_IL_GetLastLoadedFile();
   PCR_IL_LoadedSegment * q;

   /* Skip uncommitted files */
   while (p != NIL && !(p -> lf_commitPoint)) {
       /* The loading of this file has not yet been committed    */
       /* Hence its description could be inconsistent.           */
       /* Furthermore, it hasn't yet been run.  Hence its data   */
       /* segments can't possibly reference heap allocated       */
       /* objects.                                               */
       p = p -> lf_prev;
   }
   for (; p != NIL; p = p -> lf_prev) {
     for (q = p -> lf_ls; q != NIL; q = q -> ls_next) {
       if ((q -> ls_flags & PCR_IL_SegFlags_Traced_MASK)
           == PCR_IL_SegFlags_Traced_on) {
         GC_add_roots_inner((ptr_t)q->ls_addr,
                            (ptr_t)q->ls_addr + q->ls_bytes, TRUE);
       }
     }
   }
 }
#endif /* PCR && !DYNAMIC_LOADING && !MSWIN32 */

#if !defined(HAVE_REGISTER_MAIN_STATIC_DATA) && defined(DYNAMIC_LOADING)
 /* Do we need to separately register the main static data segment? */
 GC_INNER GC_bool GC_register_main_static_data(void)
 {
   return TRUE;
 }
#endif /* HAVE_REGISTER_MAIN_STATIC_DATA */

/* Register a routine to filter dynamic library registration.  */
GC_API void GC_CALL GC_register_has_static_roots_callback(
                                       GC_has_static_roots_func callback)
{
   GC_has_static_roots = callback;
}